Skip to main content

bamboo_engine/external_agents/
actor_adapter.rs

1//! Actor external child runner.
2//!
3//! Runs a child session as an independent **actor**: a separate OS process with its own
4//! isolated context, speaking the `bamboo-subagent` WebSocket protocol. This is the
5//! engine-side adapter on the `wants_external` seam: it spawns the worker binary, waits for
6//! it to self-register into the Tier-1 file fabric, connects, sends the assignment, and
7//! forwards the child's `AgentEvent`s back onto the parent's `event_tx`.
8//!
9//! The built-in **local actor** instance of this runner is the default runtime for
10//! every sub-agent (the in-process runtime was removed). The expert `externalAgents`
11//! tables can additionally route specific roles to other actor/a2a agents.
12
13use std::collections::{BTreeSet, HashMap, VecDeque};
14use std::path::{Path, PathBuf};
15use std::sync::atomic::{AtomicBool, AtomicU64, AtomicUsize, Ordering};
16use std::sync::Arc;
17use std::time::{Duration, Instant};
18
19use crate::placement_scheduler::PlacementScheduler;
20use async_trait::async_trait;
21use bamboo_agent_core::storage::Storage;
22use bamboo_agent_core::tools::tool_start_arguments_for_display;
23use bamboo_agent_core::{AgentError, AgentEvent, Role, Session};
24use bamboo_domain::poison::PoisonRecover;
25use bamboo_domain::{
26    ActorActivationClaim, ActorActivationFence, ActorActivationFinish, ActorDirectoryPort,
27    ActorLogicalState, ActorSnapshotLimits, ActorSnapshotPort, ActorSnapshotPrincipal, HookResult,
28    ParentQuestion, ParentQuestionCheckpointV1, SessionInboxClaim, SessionInboxConsumerId,
29    SessionInboxFailureReport, SessionInboxLeaseRequest, SessionInboxOwnedClaim, SessionInboxPort,
30    PARENT_QUESTION_CHECKPOINT_ACTION,
31};
32use bamboo_domain::{HostPlacementIntent, HostPlacementRequest, WorkerSlotLease};
33use sha2::{Digest, Sha256};
34use tokio::sync::mpsc;
35use tokio_util::sync::CancellationToken;
36
37use bamboo_subagent::fleet::{spawn_worker_on_bus, SpawnedChild};
38use bamboo_subagent::proto::{
39    ActorEventBatch, AgentRecord, ChildFrame, LogicalSessionIdentity, ParentFrame,
40    PermissionPolicyContext, RunSpec, SessionMessageDelivery, TerminalStatus,
41};
42use bamboo_subagent::provision::{
43    ChildIdentity, ExecutorSpec, ModelRefSpec, Placement, ProvisionSpec, ScopedCredential,
44};
45
46use super::actor_event_router::{
47    ActorEventClass, ActorEventRoute, ActorEventRouteError, ActorEventRouter,
48};
49use super::actor_event_stream::{ActorEventObserver, PublicActorEvent};
50use crate::runtime::execution::{ExternalChildRunner, SessionInboxRuntimeBinding, SpawnJob};
51
52/// Default cap on simultaneously running actor activations. The event and
53/// broker transports are bounded independently, so 200 active sub-agents are a
54/// supported operating point rather than an opt-in escape hatch.
55pub const DEFAULT_MAX_CONCURRENT_ACTORS: usize = 200;
56
57/// Compatibility name for the default. Runtime creation uses Host config.
58pub use bamboo_config::DEFAULT_MAX_SPAWN_DEPTH as MAX_SPAWN_DEPTH;
59
60/// Default cap on idle pooled (warm, reusable) workers kept per fingerprint.
61const DEFAULT_MAX_IDLE_PER_KEY: usize = 4;
62
63/// Process-wide-per-runner cap across every reuse fingerprint. Without this,
64/// workloads that continually change role/model/workspace can leave one parked
65/// worker in an unbounded number of otherwise-small buckets.
66const DEFAULT_MAX_IDLE_TOTAL: usize = 16;
67
68/// How long a pooled worker waits for its next assignment before reclaiming
69/// itself (must comfortably exceed the gap between sibling spawns).
70const POOLED_IDLE_TIMEOUT_SECS: u64 = 300;
71
72/// Sweep often enough to retire an expired/dead pool entry even when no later
73/// child checks out the same fingerprint. The worker's own idle deadline is the
74/// primary process-side guard; this task closes the parent handle and pool entry.
75const POOLED_REAPER_INTERVAL: Duration = Duration::from_secs(30);
76
77/// Deadline for a local worker's FIRST frame after a Run is dispatched. A warm
78/// worker answers in seconds; a cold spawn within tens. Total silence past this
79/// means the worker is dead (e.g. a pooled worker that exited right after its
80/// liveness check) and its Run is queued with nobody to serve it — trip it so the
81/// runner respawns once instead of hanging forever. Generous, to never false-trip
82/// a slow-but-healthy cold start.
83const WORKER_FIRST_FRAME_TIMEOUT: Duration = Duration::from_secs(60);
84
85fn active_scoped_session_deny_count(
86    config: &bamboo_tools::permission::PermissionConfig,
87    session_id: &str,
88) -> usize {
89    config
90        .temporary_grants()
91        .into_iter()
92        .filter(|grant| {
93            grant.scope == bamboo_tools::permission::TemporaryPermissionGrantScope::Session
94                && grant.effect == bamboo_tools::permission::TemporaryPermissionGrantEffect::Deny
95                && grant.session_id.as_deref() == Some(session_id)
96        })
97        .count()
98}
99
100fn ensure_no_active_scoped_session_denies(
101    config: &bamboo_tools::permission::PermissionConfig,
102    session_id: &str,
103) -> Result<(), AgentError> {
104    let count = active_scoped_session_deny_count(config, session_id);
105    if count == 0 {
106        Ok(())
107    } else {
108        Err(AgentError::LLM(format!(
109            "external executor activation blocked by {count} active session-scoped explicit deny rule(s)"
110        )))
111    }
112}
113
114/// Plaintext token returned once by the server authority for one Codex run.
115/// `token_id` is non-secret and is the handle used for guaranteed revocation.
116pub struct IssuedCodexRunToken {
117    pub token_id: String,
118    pub token: String,
119}
120
121/// Server-owned authority for Bamboo-as-provider Codex credentials. The engine
122/// only needs mint/revoke; verification remains inside the HTTP server.
123pub trait CodexRunTokenAuthority: Send + Sync + 'static {
124    fn issue(&self, session_id: &str) -> Result<IssuedCodexRunToken, String>;
125    fn revoke(&self, token_id: &str);
126}
127
128struct CodexRunTokenGuard {
129    authority: Arc<dyn CodexRunTokenAuthority>,
130    token_id: String,
131}
132
133impl Drop for CodexRunTokenGuard {
134    fn drop(&mut self) {
135        self.authority.revoke(&self.token_id);
136    }
137}
138
139fn executor_uses_bamboo_codex(executor: &ExecutorSpec) -> bool {
140    matches!(
141        executor,
142        ExecutorSpec::Codex {
143            auth_mode: Some(mode),
144            ..
145        } if mode == "bamboo"
146    ) || matches!(
147        executor,
148        ExecutorSpec::Codex {
149            auth_mode: None,
150            inherit_user_config,
151            ..
152        } if !inherit_user_config.unwrap_or(false)
153    )
154}
155
156fn executor_has_read_only_permission_profile(executor: &ExecutorSpec) -> bool {
157    match executor {
158        ExecutorSpec::ClaudeCode {
159            permission_mode, ..
160        } => permission_mode
161            .as_deref()
162            .is_some_and(|mode| mode.eq_ignore_ascii_case("plan")),
163        ExecutorSpec::Codex {
164            permission_profile,
165            sandbox,
166            ..
167        } => {
168            permission_profile
169                .as_deref()
170                .is_some_and(|profile| profile.eq_ignore_ascii_case("read-only"))
171                || sandbox
172                    .as_deref()
173                    .is_some_and(|value| value.eq_ignore_ascii_case("read-only"))
174        }
175        _ => false,
176    }
177}
178
179/// Exact non-secret executor posture the host expects for one typed activation.
180///
181/// Echo is intentionally a transport-only smoke executor and CliAdapter is not
182/// implemented by the production worker. Neither claims the typed permission
183/// contract, so legacy/custom frame handling remains available only for those
184/// two variants. Every executable permission-aware variant must prove the
185/// mapping derived from its provisioned spec before any execution event.
186fn expected_permission_executor_mapping(
187    executor: &ExecutorSpec,
188    resolution: bamboo_domain::PermissionModeResolution,
189    has_explicit_deny: bool,
190) -> Result<Option<String>, AgentError> {
191    let mapping = match executor {
192        ExecutorSpec::Echo | ExecutorSpec::CliAdapter { .. } => return Ok(None),
193        ExecutorSpec::BambooRuntime => {
194            format!("bamboo_runtime:{}", resolution.effective.as_str())
195        }
196        ExecutorSpec::ClaudeCode {
197            permission_mode, ..
198        } => {
199            if has_explicit_deny {
200                "claude_code:blocked_explicit_deny".to_string()
201            } else {
202                let mode = match resolution.effective {
203                    bamboo_domain::PermissionMode::Plan => "plan",
204                    bamboo_domain::PermissionMode::Auto => "bypassPermissions",
205                    bamboo_domain::PermissionMode::AcceptEdits => "acceptEdits",
206                    bamboo_domain::PermissionMode::DontAsk => "dontAsk",
207                    bamboo_domain::PermissionMode::Default
208                    | bamboo_domain::PermissionMode::BypassPermissions => {
209                        permission_mode.as_deref().unwrap_or("default")
210                    }
211                };
212                format!("claude_code:permission_mode={mode}")
213            }
214        }
215        ExecutorSpec::Codex {
216            mode,
217            sandbox,
218            approval_policy,
219            allow_danger_bypass,
220            ..
221        } => match mode.as_deref().unwrap_or("exec") {
222            "exec" => {
223                let approval_policy = expected_codex_exec_approval_policy(
224                    sandbox.as_deref(),
225                    approval_policy.as_deref(),
226                    allow_danger_bypass.unwrap_or(false),
227                    resolution,
228                )?;
229                if has_explicit_deny {
230                    "codex_exec:blocked_explicit_deny".to_string()
231                } else {
232                    format!("codex_exec:approval_policy={approval_policy}")
233                }
234            }
235            "app_server" => {
236                if !matches!(approval_policy.as_deref(), None | Some("on-request")) {
237                    return Err(AgentError::LLM(
238                        "invalid Codex app-server permission posture configuration".to_string(),
239                    ));
240                }
241                if has_explicit_deny {
242                    "codex_app_server:blocked_explicit_deny".to_string()
243                } else {
244                    let approval_policy = if resolution.suppress_approval_prompts()
245                        || resolution.effective == bamboo_domain::PermissionMode::Plan
246                    {
247                        "never"
248                    } else {
249                        "on-request"
250                    };
251                    format!("codex_app_server:approvalPolicy={approval_policy}")
252                }
253            }
254            _ => {
255                return Err(AgentError::LLM(
256                    "unsupported Codex executor mode for permission posture contract".to_string(),
257                ));
258            }
259        },
260    };
261    Ok(Some(mapping))
262}
263
264fn expected_codex_exec_approval_policy(
265    sandbox: Option<&str>,
266    approval_policy: Option<&str>,
267    allow_danger_bypass: bool,
268    resolution: bamboo_domain::PermissionModeResolution,
269) -> Result<&'static str, AgentError> {
270    let configured = match approval_policy {
271        None | Some("never") => "never",
272        Some("on-failure") => "on-failure",
273        Some(_) => {
274            return Err(AgentError::LLM(
275                "invalid Codex exec permission posture configuration".to_string(),
276            ));
277        }
278    };
279    if resolution.suppress_approval_prompts()
280        || resolution.effective == bamboo_domain::PermissionMode::Plan
281    {
282        return Ok("never");
283    }
284    match sandbox {
285        Some("danger-full-access") => Ok("never"),
286        Some("read-only") | Some("workspace-write") => Ok(configured),
287        None if allow_danger_bypass || resolution.bypass_permissions() => Ok("never"),
288        None => Ok(configured),
289        Some(_) => Err(AgentError::LLM(
290            "invalid Codex exec permission posture configuration".to_string(),
291        )),
292    }
293}
294
295fn workspace_is_bamboo_owned(raw: &str) -> bool {
296    let workspace = std::fs::canonicalize(raw).unwrap_or_else(|_| PathBuf::from(raw));
297    let configured_root = bamboo_config::paths::resolve_workspace_root();
298    let configured_root = std::fs::canonicalize(&configured_root).unwrap_or(configured_root);
299    if workspace.starts_with(&configured_root) {
300        return true;
301    }
302
303    // Project worktrees created by Bamboo live under
304    // `<project>/.bamboo/worktree/<name>` and carry the ownership marker used
305    // by the project-worktree lifecycle. A path that merely imitates the
306    // directory shape is not sufficient to bypass Codex's git guard.
307    workspace.ancestors().any(|candidate| {
308        let Some(name) = candidate.file_name().and_then(|name| name.to_str()) else {
309            return false;
310        };
311        let Some(worktree_root) = candidate.parent() else {
312            return false;
313        };
314        if worktree_root.file_name() != Some(std::ffi::OsStr::new("worktree"))
315            || worktree_root.parent().and_then(Path::file_name)
316                != Some(std::ffi::OsStr::new(".bamboo"))
317        {
318            return false;
319        }
320        let marker = worktree_root.join(".bamboo-owned").join(name);
321        std::fs::read_to_string(marker).is_ok_and(|branch| branch == format!("bamboo/{name}"))
322    })
323}
324
325fn build_codex_run_secrets(
326    executor: &ExecutorSpec,
327    authority: Option<Arc<dyn CodexRunTokenAuthority>>,
328    child_session_id: &str,
329) -> Result<
330    (
331        bamboo_subagent::proto::RunSecrets,
332        Option<CodexRunTokenGuard>,
333    ),
334    AgentError,
335> {
336    if !executor_uses_bamboo_codex(executor) {
337        return Ok((bamboo_subagent::proto::RunSecrets::default(), None));
338    }
339
340    let authority = authority.ok_or_else(|| {
341        AgentError::LLM(
342            "Codex auth mode 'bamboo' requires the server per-run token authority".to_string(),
343        )
344    })?;
345    let issued = authority
346        .issue(child_session_id)
347        .map_err(|error| AgentError::LLM(format!("mint Codex per-run provider token: {error}")))?;
348    let guard = CodexRunTokenGuard {
349        authority,
350        token_id: issued.token_id,
351    };
352    Ok((
353        bamboo_subagent::proto::RunSecrets {
354            codex_provider_token: Some(bamboo_subagent::proto::SecretValue::new(issued.token)),
355        },
356        Some(guard),
357    ))
358}
359
360/// A warm worker on the mailbox bus, parked for reuse between runs. It stays
361/// dialed-in + subscribed to `mailbox_id`; the next interchangeable child
362/// delivers its `Run` there instead of spawning a fresh process. Dropping it
363/// kills a local kill-on-drop subprocess; a remote / schedulable handle is
364/// process-less (`kill()` is a no-op — it self-manages via its idle timeout).
365struct PooledWorker {
366    worker: SpawnedChild,
367    /// The bus mailbox this worker subscribes to (where its `Run`s are delivered).
368    mailbox_id: String,
369    /// Set only while the worker is parked. Checked both at checkout and by the
370    /// background pool sweep; cleared before the worker is handed to a run.
371    parked_at: Option<Instant>,
372}
373
374type WorkerPool = HashMap<String, Vec<PooledWorker>>;
375
376#[derive(Debug, Clone, Copy, PartialEq, Eq)]
377enum PoolReapReason {
378    ProcessExited,
379    IdleTimeout,
380}
381
382impl PoolReapReason {
383    fn as_str(self) -> &'static str {
384        match self {
385            Self::ProcessExited => "process_exited_while_parked",
386            Self::IdleTimeout => "pool_idle_timeout",
387        }
388    }
389}
390
391fn parked_worker_expired(parked_at: Option<Instant>, now: Instant, timeout: Duration) -> bool {
392    parked_at.is_some_and(|parked_at| now.saturating_duration_since(parked_at) >= timeout)
393}
394
395fn idle_pool_limit_reason(
396    pool: &WorkerPool,
397    key: &str,
398    max_per_key: usize,
399    max_total: usize,
400) -> Option<&'static str> {
401    if pool.get(key).map_or(0, Vec::len) >= max_per_key {
402        Some("pool_per_key_limit")
403    } else if pool.values().map(Vec::len).sum::<usize>() >= max_total {
404        Some("pool_global_limit")
405    } else {
406        None
407    }
408}
409
410fn take_reapable_workers(
411    pool: &mut WorkerPool,
412    now: Instant,
413    timeout: Duration,
414) -> Vec<(String, PoolReapReason, PooledWorker)> {
415    let mut reaped = Vec::new();
416    for (key, bucket) in pool.iter_mut() {
417        let mut retained = Vec::with_capacity(bucket.len());
418        for mut worker in std::mem::take(bucket) {
419            let reason = if !worker.worker.is_alive() {
420                Some(PoolReapReason::ProcessExited)
421            } else if parked_worker_expired(worker.parked_at, now, timeout) {
422                Some(PoolReapReason::IdleTimeout)
423            } else {
424                None
425            };
426            if let Some(reason) = reason {
427                reaped.push((key.clone(), reason, worker));
428            } else {
429                retained.push(worker);
430            }
431        }
432        *bucket = retained;
433    }
434    pool.retain(|_, bucket| !bucket.is_empty());
435    reaped
436}
437
438async fn reap_worker_pool_once(pool: &Arc<tokio::sync::Mutex<WorkerPool>>) -> usize {
439    let now = Instant::now();
440    let timeout = Duration::from_secs(POOLED_IDLE_TIMEOUT_SECS);
441    let reaped = {
442        let mut pool = pool.lock().await;
443        take_reapable_workers(&mut pool, now, timeout)
444    };
445    let count = reaped.len();
446    for (pool_key, reason, worker) in reaped {
447        let idle_ms = worker
448            .parked_at
449            .map(|parked_at| now.saturating_duration_since(parked_at).as_millis() as u64);
450        tracing::info!(
451            pool_key,
452            mailbox_id = %worker.mailbox_id,
453            worker_pid = ?worker.worker.pid(),
454            worker_idle_ms = ?idle_ms,
455            shutdown_reason = reason.as_str(),
456            "reaping parked sub-agent worker"
457        );
458        worker.worker.kill().await;
459    }
460    count
461}
462
463/// A role pinned to a remote resident worker (remote-actor-plan §3.4 / P1.5,
464/// #193), resolved at runner-build time from `SubagentsConfig.remote_placements`:
465/// the env-named broker bearer is already READ into `token` here (the raw token
466/// never rides the config), and `ca_cert_file` pins the broker's TLS cert.
467#[derive(Clone, Default)]
468pub struct ResolvedRemotePlacement {
469    pub endpoint: String,
470    pub token: Option<String>,
471    pub ca_cert_file: Option<PathBuf>,
472    /// Display name for the machine this role runs on — the matching cluster
473    /// node's `label`/host, surfaced on the UI placement badge. `None` ⇒ derive
474    /// from the endpoint host.
475    pub host_label: Option<String>,
476    /// Some(Err) preserves explicit unavailable selection; never falls back.
477    pub broker_peer: Option<Result<bamboo_config::RemoteBrokerPeer, ()>>,
478    pub requirements: Option<bamboo_config::OperatorPlacementRequirements>,
479}
480impl std::fmt::Debug for ResolvedRemotePlacement {
481    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
482        f.debug_struct("ResolvedRemotePlacement")
483            .field("broker_peer", &self.broker_peer.is_some())
484            .finish_non_exhaustive()
485    }
486}
487
488/// A role routed to a SCHEDULED worker (remote-actor-plan §3.4 / P2b, #181),
489/// resolved at runner-build time from `SubagentsConfig.schedulable_placements`.
490/// Names the logical `pool` (= the bus role) whose LIVE connected workers are the
491/// scheduling candidates — the runner picks one via the bus presence query
492/// (`BrokerClient::list_connected`). Phase 3 retired the old HTTP registry, so a
493/// pool is now just a role on the bus.
494#[derive(Debug, Clone, Default)]
495pub struct ResolvedSchedulablePlacement {
496    pub pool: String,
497    /// Display name for the machine this pool's workers run on — the matching
498    /// cluster node's `label`/host, surfaced on the UI placement badge. `None` ⇒
499    /// fall back to the pool name.
500    pub host_label: Option<String>,
501    pub endpoint: String,
502    pub token: Option<String>,
503    pub ca_cert_file: Option<PathBuf>,
504    pub broker_parent: Option<bamboo_config::SchedulableBrokerParent>,
505    pub requirements: Option<bamboo_config::OperatorPlacementRequirements>,
506}
507
508#[derive(Clone)]
509struct ScopedPlacementRoute {
510    endpoint: String,
511    token: String,
512    ca_cert_file: PathBuf,
513    parent: bamboo_subagent::AgentRef,
514    worker_role: String,
515    pinned_mailbox: Option<String>,
516    requirements: bamboo_config::OperatorPlacementRequirements,
517}
518
519impl ScopedPlacementRoute {
520    async fn query(&self) -> Result<bamboo_broker::BrokerClient, AgentError> {
521        let tls = bamboo_broker::client_config_trusting_cert(&self.ca_cert_file)
522            .map_err(|_| remote_broker_unavailable())?;
523        bamboo_broker::BrokerClient::connect_with_tls(
524            &self.endpoint,
525            self.parent.clone(),
526            &self.token,
527            Some(tls),
528        )
529        .await
530        .map_err(|_| remote_broker_unavailable())
531    }
532}
533
534/// Holds one atomic HostRegistry slot for the entire remote Run. A fresh
535/// authenticated broker observation extends the Host lease before the slot is
536/// renewed. Losing either identity cancels the Run; Drop releases the exact
537/// generation/epoch and cannot free a replacement worker's slot.
538struct ActivePlacementLease {
539    registry: bamboo_storage::v2::FileHostRegistry,
540    actor_store: Arc<bamboo_storage::SessionStoreV2>,
541    route: ScopedPlacementRoute,
542    mailbox: String,
543    host_ref: String,
544    fence: ActorActivationFence,
545    placement_ref: bamboo_domain::ActorPlacementRef,
546    lease: Arc<tokio::sync::Mutex<WorkerSlotLease>>,
547    heartbeat: tokio::task::JoinHandle<()>,
548    run_cancel: CancellationToken,
549    lost: Arc<AtomicBool>,
550    release_on_drop: AtomicBool,
551}
552
553/// No Run has reached a worker yet, so a failed placement preflight may end
554/// only the activation it claimed. A replacement activation has a different
555/// fence and must remain untouched.
556async fn fail_pre_dispatch_activation(
557    actor_store: &bamboo_storage::SessionStoreV2,
558    fence: &ActorActivationFence,
559) -> Result<(), bamboo_domain::ActorDirectoryError> {
560    actor_store
561        .finish_activation(fence, chrono::Utc::now(), ActorActivationFinish::Failed)
562        .await
563        .map(|_| ())
564}
565
566/// Once the Run send is invoked, an error may mean that the broker persisted
567/// the FencedRun but its delivery receipt was lost. Only errors before this
568/// call can safely roll back the Actor activation and its reserved slot.
569async fn send_actor_run(
570    client: &mut dyn bamboo_subagent::ChildLink,
571    spec: RunSpec,
572    send_invoked: &mut bool,
573) -> Result<(), AgentError> {
574    *send_invoked = true;
575    client
576        .send(ParentFrame::Run(spec))
577        .await
578        .map_err(|error| AgentError::LLM(format!("actor run dispatch failed: {error}")))
579}
580
581fn remote_activation_finish(
582    result: &Result<Option<String>, AgentError>,
583    accepted_terminal: Option<TerminalStatus>,
584) -> ActorActivationFinish {
585    if result.is_ok() {
586        ActorActivationFinish::Succeeded
587    } else if matches!(result, Err(AgentError::Cancelled))
588        && accepted_terminal == Some(TerminalStatus::Cancelled)
589    {
590        ActorActivationFinish::Cancelled
591    } else {
592        ActorActivationFinish::Failed
593    }
594}
595
596/// Commit a remote reply through the current Actor fence before marking the
597/// activation terminal. The SDK's later status save may then carry exactly
598/// the canonical main transcript instead of trying an unfenced append.
599async fn append_fenced_remote_reply(
600    actor_store: &bamboo_storage::SessionStoreV2,
601    fence: &ActorActivationFence,
602    session: &mut Session,
603    text: &str,
604) -> Result<(), AgentError> {
605    let message = bamboo_agent_core::Message::assistant(text, None);
606    let committed = actor_store
607        .append_actor_transcript(bamboo_storage::ActorTranscriptAppend {
608            fence: fence.clone(),
609            expected_created_at: session.created_at,
610            expected_conversation_generation: session.conversation_generation_observation(),
611            expected_messages: session.messages.clone(),
612            expected_provider_transcript: session.provider_transcript.clone(),
613            messages: vec![message],
614            native_groups: Vec::new(),
615        })
616        .await
617        .map_err(|error| AgentError::LLM(format!("remote actor reply commit failed: {error}")))?;
618    session.messages = committed.messages;
619    session.provider_transcript = committed.provider_transcript;
620    session.updated_at = committed.updated_at;
621    Ok(())
622}
623
624impl ActivePlacementLease {
625    async fn reserve(
626        registry: bamboo_storage::v2::FileHostRegistry,
627        actor_store: Arc<bamboo_storage::SessionStoreV2>,
628        route: ScopedPlacementRoute,
629        mailbox: String,
630        actor_id: String,
631        run_id: String,
632        project_id: Option<bamboo_domain::ProjectId>,
633        required_tools: BTreeSet<String>,
634        parent_cancel: &CancellationToken,
635    ) -> Result<Self, AgentError> {
636        let scheduler = PlacementScheduler::new(registry.clone());
637        let mut query = route.query().await?;
638        let observed = query
639            .observe_host_capacity(&mailbox, &route.worker_role)
640            .await
641            .map_err(|_| remote_broker_unavailable())?
642            .ok_or_else(remote_broker_unavailable)?;
643        let host = scheduler
644            .observe_scoped_host(
645                &mut query,
646                &observed.host_ref,
647                &mailbox,
648                &route.worker_role,
649                chrono::Duration::seconds(110),
650            )
651            .await
652            .map_err(|_| remote_broker_unavailable())?;
653        if host.capabilities.placement_class != bamboo_domain::ActorPlacementClass::Remote
654            || host.connection_generation != observed.connection_generation
655            || route
656                .pinned_mailbox
657                .as_deref()
658                .is_some_and(|pinned| pinned != mailbox)
659        {
660            return Err(remote_broker_unavailable());
661        }
662        let now = chrono::Utc::now();
663        let request = HostPlacementRequest {
664            intent: HostPlacementIntent::PinnedHost {
665                host_ref: host.host_ref.clone(),
666            },
667            actor_id,
668            run_id,
669            project_id: project_id.clone(),
670            trust_zone: route.requirements.trust_zone.clone(),
671            workspace_label: route.requirements.workspace_label.clone(),
672            executor: "bamboo-runtime".into(),
673            required_tools,
674            network_zone: route.requirements.network_zone.clone(),
675            require_network_isolation: route.requirements.require_network_isolation,
676            preferred_host_ref: None,
677            now,
678            lease_expires_at: now + chrono::Duration::seconds(80),
679        };
680        if !request.eligible(&host) {
681            return Err(remote_broker_unavailable());
682        }
683        let lease = scheduler
684            .reserve(request)
685            .await
686            .map_err(|_| remote_broker_unavailable())?;
687        let placement_ref = bamboo_domain::ActorPlacementRef {
688            class: bamboo_domain::ActorPlacementClass::Remote,
689            lease_id: lease.lease_id.clone(),
690            slot_epoch: Some(lease.epoch),
691        };
692        let mut claimed_fence = None;
693        let actor_binding = async {
694            let actor = actor_store
695                .ensure_actor(&lease.actor_id)
696                .await
697                .map_err(|_| remote_broker_unavailable())?;
698            if actor.actor.project_id != project_id.as_ref().map(|id| id.to_string()) {
699                return Err(remote_broker_unavailable());
700            }
701            let placement_intent = if route.pinned_mailbox.is_some() {
702                bamboo_domain::ActorPlacementIntent::PinnedHost {
703                    host_ref: host.host_ref.clone(),
704                }
705            } else {
706                bamboo_domain::ActorPlacementIntent::Pool {
707                    pool_id: route.worker_role.clone(),
708                }
709            };
710            actor_store
711                .bind_control_plane(
712                    &lease.actor_id,
713                    bamboo_domain::ActorControlPlaneBinding {
714                        policy_revision: None,
715                        placement_intent: Some(placement_intent),
716                    },
717                )
718                .await
719                .map_err(|_| {
720                    tracing::warn!(stage = "actor_bind", "remote placement admission failed");
721                    remote_broker_unavailable()
722                })?;
723            let activation = actor_store
724                .claim_activation(&ActorActivationClaim {
725                    actor_id: lease.actor_id.clone(),
726                    run_id: lease.run_id.clone(),
727                    lease_owner: format!("worker-host-{}", uuid::Uuid::new_v4()),
728                    lease_expires_at: lease.expires_at,
729                    inbox_generation: 0,
730                    placement_ref: Some(placement_ref.clone()),
731                    now: chrono::Utc::now(),
732                })
733                .await
734                .map_err(|_| {
735                    tracing::warn!(stage = "actor_claim", "remote placement admission failed");
736                    remote_broker_unavailable()
737                })?;
738            let fence = activation.fence();
739            claimed_fence = Some(fence.clone());
740            actor_store
741                .start_activation(&fence, chrono::Utc::now())
742                .await
743                .map_err(|_| {
744                    tracing::warn!(stage = "actor_start", "remote placement admission failed");
745                    remote_broker_unavailable()
746                })?;
747            let reservation = registry
748                .validate_run_reservation(bamboo_domain::WorkerRunReservationCheck {
749                    placement_ref: placement_ref.clone(),
750                    actor_id: lease.actor_id.clone(),
751                    run_id: lease.run_id.clone(),
752                    observed_host_ref: host.host_ref.clone(),
753                    observed_connection_generation: host.connection_generation.clone(),
754                    now: chrono::Utc::now(),
755                })
756                .await
757                .map_err(|error| {
758                    tracing::warn!(stage = "slot_preflight", reason = %error, "remote placement admission failed");
759                    remote_broker_unavailable()
760                })?;
761            if reservation.mailbox != mailbox || reservation.lease != lease {
762                tracing::warn!(stage = "slot_binding", "remote placement admission failed");
763                return Err(remote_broker_unavailable());
764            }
765            Ok::<_, AgentError>(fence)
766        }
767        .await;
768        let fence = match actor_binding {
769            Ok(fence) => fence,
770            Err(error) => {
771                if let Some(fence) = claimed_fence.as_ref() {
772                    if let Err(rollback_error) =
773                        fail_pre_dispatch_activation(actor_store.as_ref(), fence).await
774                    {
775                        tracing::warn!(
776                            stage = "actor_pre_dispatch_rollback",
777                            reason = %rollback_error,
778                            "remote placement activation rollback failed"
779                        );
780                    }
781                }
782                let _ = registry.release_slot(&lease).await;
783                return Err(error);
784            }
785        };
786        let lease = Arc::new(tokio::sync::Mutex::new(lease));
787        let host_ref = host.host_ref.clone();
788        let lost = Arc::new(AtomicBool::new(false));
789        let run_cancel = parent_cancel.child_token();
790        let heartbeat = {
791            let registry = registry.clone();
792            let actor_store = actor_store.clone();
793            let route = route.clone();
794            let host_ref = host_ref.clone();
795            let mailbox = mailbox.clone();
796            let fence = fence.clone();
797            let placement_ref = placement_ref.clone();
798            let lease = lease.clone();
799            let lost = lost.clone();
800            let run_cancel = run_cancel.clone();
801            tokio::spawn(async move {
802                let mut interval = tokio::time::interval(Duration::from_secs(30));
803                interval.tick().await;
804                loop {
805                    tokio::select! {
806                        _ = run_cancel.cancelled() => break,
807                        _ = interval.tick() => {}
808                    }
809                    let refreshed = async {
810                        let mut query = route.query().await?;
811                        let host = PlacementScheduler::new(registry.clone())
812                            .observe_scoped_host(
813                                &mut query,
814                                &host_ref,
815                                &mailbox,
816                                &route.worker_role,
817                                chrono::Duration::seconds(110),
818                            )
819                            .await
820                            .map_err(|_| remote_broker_unavailable())?;
821                        let current = lease.lock().await.clone();
822                        if host.connection_generation != current.connection_generation {
823                            return Err(remote_broker_unavailable());
824                        }
825                        let entry = actor_store
826                            .inspect_actor(&current.actor_id)
827                            .await
828                            .map_err(|_| remote_broker_unavailable())?;
829                        if entry.activation.as_ref().is_none_or(|activation| {
830                            activation.fence() != fence
831                                || activation.placement_ref.as_ref() != Some(&placement_ref)
832                        }) {
833                            return Err(remote_broker_unavailable());
834                        }
835                        let now = chrono::Utc::now();
836                        registry
837                            .validate_run_reservation(bamboo_domain::WorkerRunReservationCheck {
838                                placement_ref: placement_ref.clone(),
839                                actor_id: current.actor_id.clone(),
840                                run_id: current.run_id.clone(),
841                                observed_host_ref: host.host_ref.clone(),
842                                observed_connection_generation: host.connection_generation.clone(),
843                                now,
844                            })
845                            .await
846                            .map_err(|_| remote_broker_unavailable())?;
847                        let renewed = registry
848                            .renew_slot(&current, now, now + chrono::Duration::seconds(80))
849                            .await
850                            .map_err(|_| remote_broker_unavailable())?;
851                        *lease.lock().await = renewed;
852                        let expiry = lease.lock().await.expires_at;
853                        actor_store
854                            .renew_activation(&fence, now, expiry)
855                            .await
856                            .map_err(|_| remote_broker_unavailable())?;
857                        Ok::<(), AgentError>(())
858                    }
859                    .await;
860                    if refreshed.is_err() {
861                        lost.store(true, Ordering::SeqCst);
862                        run_cancel.cancel();
863                        break;
864                    }
865                }
866            })
867        };
868        Ok(Self {
869            registry,
870            actor_store,
871            route,
872            mailbox,
873            host_ref,
874            fence,
875            placement_ref,
876            lease,
877            heartbeat,
878            run_cancel,
879            lost,
880            release_on_drop: AtomicBool::new(false),
881        })
882    }
883
884    fn lost(&self) -> bool {
885        self.lost.load(Ordering::SeqCst)
886    }
887
888    async fn abort_if_run_unsent(&mut self, send_invoked: bool) -> Result<bool, AgentError> {
889        if send_invoked {
890            return Ok(false);
891        }
892        self.abort_before_dispatch().await?;
893        Ok(true)
894    }
895
896    async fn abort_before_dispatch(&mut self) -> Result<(), AgentError> {
897        self.heartbeat.abort();
898        let _ = (&mut self.heartbeat).await;
899        let current = self.lease.lock().await.clone();
900        let failed = fail_pre_dispatch_activation(self.actor_store.as_ref(), &self.fence).await;
901        let released = self.registry.release_slot(&current).await;
902        if let Err(error) = failed {
903            tracing::warn!(stage = "actor_pre_dispatch_rollback", reason = %error, "remote activation failure could not be recorded");
904            return Err(remote_broker_unavailable());
905        }
906        if let Err(error) = released {
907            if !matches!(error, bamboo_domain::HostRegistryError::StaleLease) {
908                tracing::warn!(stage = "slot_pre_dispatch_rollback", reason = %error, "remote slot release failed");
909                return Err(remote_broker_unavailable());
910            }
911        }
912        Ok(())
913    }
914
915    async fn complete(
916        &mut self,
917        outcome: ActorActivationFinish,
918        terminal_proven: bool,
919    ) -> Result<(), AgentError> {
920        if !terminal_proven {
921            // Keep the activation live until its lease expires. A silent Worker
922            // may still be executing; this blocks an immediate ActorId retry.
923            return if outcome == ActorActivationFinish::Succeeded {
924                Err(remote_broker_unavailable())
925            } else {
926                Ok(())
927            };
928        }
929        // Wait for the heartbeat to stop before reading the exact lease. It
930        // must not renew the slot after terminal verification or release.
931        self.heartbeat.abort();
932        let _ = (&mut self.heartbeat).await;
933        let current = self.lease.lock().await.clone();
934        // A correlated Outcome from the selected mailbox alone cannot release
935        // a new generation's slot. Reobserve the scoped broker identity at
936        // terminal time, including fast reconnects between heartbeat ticks.
937        let mut query = self.route.query().await?;
938        let observed = query
939            .observe_host_capacity(&self.mailbox, &self.route.worker_role)
940            .await
941            .map_err(|_| remote_broker_unavailable())?
942            .ok_or_else(remote_broker_unavailable)?;
943        if observed.host_ref != self.host_ref
944            || observed.connection_generation != current.connection_generation
945            || observed.mailbox != self.mailbox
946            || observed.role.as_deref() != Some(self.route.worker_role.as_str())
947        {
948            return Err(remote_broker_unavailable());
949        }
950        let entry = self
951            .actor_store
952            .inspect_actor(&current.actor_id)
953            .await
954            .map_err(|_| remote_broker_unavailable())?;
955        if entry.activation.as_ref().is_none_or(|activation| {
956            activation.fence() != self.fence
957                || activation.placement_ref.as_ref() != Some(&self.placement_ref)
958        }) {
959            return Err(remote_broker_unavailable());
960        }
961        self.registry
962            .validate_slot(&current, chrono::Utc::now())
963            .await
964            .map_err(|_| remote_broker_unavailable())?;
965        self.actor_store
966            .finish_activation(&self.fence, chrono::Utc::now(), outcome)
967            .await
968            .map_err(|_| remote_broker_unavailable())?;
969        // A terminal ActorDirectory record permits an immediate retry. Free
970        // its exact slot before returning rather than racing a Drop task.
971        self.release_on_drop.store(true, Ordering::SeqCst);
972        match self.registry.release_slot(&current).await {
973            Ok(()) | Err(bamboo_domain::HostRegistryError::StaleLease) => {
974                self.release_on_drop.store(false, Ordering::SeqCst);
975                Ok(())
976            }
977            Err(error) => {
978                tracing::warn!(stage = "slot_terminal_release", reason = %error, "remote slot release failed");
979                Err(remote_broker_unavailable())
980            }
981        }
982    }
983}
984
985impl Drop for ActivePlacementLease {
986    fn drop(&mut self) {
987        self.run_cancel.cancel();
988        self.heartbeat.abort();
989        if !self.release_on_drop.load(Ordering::SeqCst) {
990            return;
991        }
992        let registry = self.registry.clone();
993        let actor_store = self.actor_store.clone();
994        let fence = self.fence.clone();
995        let placement_ref = self.placement_ref.clone();
996        let lease = self.lease.clone();
997        if let Ok(runtime) = tokio::runtime::Handle::try_current() {
998            runtime.spawn(async move {
999                let current = lease.lock().await.clone();
1000                if actor_store
1001                    .inspect_actor(&current.actor_id)
1002                    .await
1003                    .ok()
1004                    .and_then(|entry| entry.activation)
1005                    .is_some_and(|activation| {
1006                        activation.fence() == fence
1007                            && activation.placement_ref.as_ref() == Some(&placement_ref)
1008                    })
1009                {
1010                    let _ = registry.release_slot(&current).await;
1011                }
1012            });
1013        }
1014    }
1015}
1016
1017/// Host-only receipt repair uses the same immutable boot configuration as the
1018/// actor runner. It never infers a broker route from Child display metadata or
1019/// a Worker claim. A role whose fixed remote route is ambiguous or incomplete
1020/// stays blocked until an operator restores a provable route.
1021pub struct BrokerTerminalReceiptReconciler {
1022    store: Arc<bamboo_storage::SessionStoreV2>,
1023    bus: Option<bamboo_subagent::BusEndpoint>,
1024    remote: HashMap<String, Option<ReceiptRemoteRoute>>,
1025    ticket_service: Option<Arc<bamboo_tickets::TicketService>>,
1026}
1027
1028struct ReceiptRemoteRoute {
1029    endpoint: String,
1030    token: String,
1031    ca_cert_file: PathBuf,
1032    parent: bamboo_subagent::AgentRef,
1033}
1034
1035#[derive(Debug, Default, Clone, Copy, PartialEq, Eq)]
1036pub struct BrokerReceiptRepairReport {
1037    pub candidates: usize,
1038    pub repaired: usize,
1039    pub blocked: usize,
1040}
1041
1042impl BrokerTerminalReceiptReconciler {
1043    pub fn new(store: Arc<bamboo_storage::SessionStoreV2>, config: &bamboo_config::Config) -> Self {
1044        let subagents = config.subagents();
1045        let mut counts = HashMap::<&str, usize>::new();
1046        for placement in &subagents.remote_placements {
1047            *counts.entry(placement.role.as_str()).or_default() += 1;
1048        }
1049        let mut remote = HashMap::new();
1050        for placement in &subagents.remote_placements {
1051            let route = (counts.get(placement.role.as_str()) == Some(&1))
1052                .then(|| receipt_remote_route(placement))
1053                .flatten();
1054            remote.insert(placement.role.clone(), route);
1055        }
1056        Self {
1057            store,
1058            ticket_service: None,
1059            bus: subagents.broker.clone().and_then(|broker| {
1060                (!broker.endpoint.trim().is_empty() && !broker.token.is_empty()).then_some(
1061                    bamboo_subagent::BusEndpoint {
1062                        endpoint: broker.endpoint,
1063                        token: broker.token,
1064                    },
1065                )
1066            }),
1067            remote,
1068        }
1069    }
1070
1071    pub fn with_ticket_service(
1072        mut self,
1073        service: Option<Arc<bamboo_tickets::TicketService>>,
1074    ) -> Self {
1075        self.ticket_service = service;
1076        self
1077    }
1078
1079    /// One physical-tree scan. Every Child is independent; a failed route,
1080    /// broker, or ACK retains its Host receipt and never blocks another Child.
1081    pub async fn reconcile_once(&self) -> BrokerReceiptRepairReport {
1082        let candidates = match self
1083            .store
1084            .discover_unconfirmed_broker_terminal_children()
1085            .await
1086        {
1087            Ok(candidates) => candidates,
1088            Err(error) => {
1089                tracing::warn!(%error, "broker terminal receipt scan blocked; no ACK attempted");
1090                return BrokerReceiptRepairReport {
1091                    blocked: 1,
1092                    ..BrokerReceiptRepairReport::default()
1093                };
1094            }
1095        };
1096        let mut report = BrokerReceiptRepairReport {
1097            candidates: candidates.len(),
1098            ..BrokerReceiptRepairReport::default()
1099        };
1100        for child in candidates {
1101            let result =
1102                tokio::time::timeout(Duration::from_secs(15), self.reconcile_child(&child)).await;
1103            match result {
1104                Ok(Ok(())) => report.repaired += 1,
1105                Ok(Err(reason)) => {
1106                    report.blocked += 1;
1107                    tracing::warn!(child_id = %child.id, reason, "broker terminal receipt retained for retry");
1108                }
1109                Err(_) => {
1110                    report.blocked += 1;
1111                    tracing::warn!(child_id = %child.id, "broker terminal receipt repair timed out; retained for retry");
1112                }
1113            }
1114        }
1115        report
1116    }
1117
1118    async fn reconcile_child(&self, child: &Session) -> Result<(), &'static str> {
1119        if child.metadata.get("runtime.kind").map(String::as_str) != Some("external")
1120            || child.metadata.get("external.protocol").map(String::as_str) != Some("actor")
1121            || child.metadata.get("external.agent_id").map(String::as_str)
1122                != Some(super::config::LOCAL_ACTOR_AGENT_ID)
1123        {
1124            return Err("Child has no trusted built-in actor recovery route");
1125        }
1126        let current = self
1127            .store
1128            .load_session(&child.id)
1129            .await
1130            .map_err(|_| "canonical Child reload failed")?
1131            .ok_or("canonical Child disappeared")?;
1132        if current.kind != bamboo_domain::SessionKind::Child
1133            || current.id != child.id
1134            || current.created_at != child.created_at
1135            || current.parent_session_id != child.parent_session_id
1136            || current.root_session_id != child.root_session_id
1137            || current.project_id_meta() != child.project_id_meta()
1138            || current.metadata.get("runtime.kind").map(String::as_str) != Some("external")
1139            || current
1140                .metadata
1141                .get("external.protocol")
1142                .map(String::as_str)
1143                != Some("actor")
1144            || current
1145                .metadata
1146                .get("external.agent_id")
1147                .map(String::as_str)
1148                != Some(super::config::LOCAL_ACTOR_AGENT_ID)
1149            || bamboo_domain::ActorSession::from_session(&current).is_err()
1150        {
1151            return Err("canonical Child birth, Root, or Project changed");
1152        }
1153        let root = self
1154            .store
1155            .load_session(&child.root_session_id)
1156            .await
1157            .map_err(|_| "canonical Root reload failed")?
1158            .ok_or("canonical Root disappeared")?;
1159        if root.kind != bamboo_domain::SessionKind::Root
1160            || root.id != child.root_session_id
1161            || root.root_session_id != child.root_session_id
1162            || root.project_id_meta() != child.project_id_meta()
1163        {
1164            return Err("canonical Root identity changed");
1165        }
1166        let parent_id = child
1167            .parent_session_id
1168            .as_deref()
1169            .ok_or("Child parent identity missing")?;
1170        let parent = self
1171            .store
1172            .load_session(parent_id)
1173            .await
1174            .map_err(|_| "canonical parent reload failed")?
1175            .ok_or("canonical parent disappeared")?;
1176        if parent.id != parent_id
1177            || !matches!(
1178                parent.kind,
1179                bamboo_domain::SessionKind::Root | bamboo_domain::SessionKind::Child
1180            )
1181            || parent.root_session_id != child.root_session_id
1182            || parent.project_id_meta() != child.project_id_meta()
1183            || parent.spawn_depth.checked_add(1) != Some(current.spawn_depth)
1184        {
1185            return Err("canonical parent lineage changed");
1186        }
1187
1188        // Validate the Host checkpoint before opening a recovery connection.
1189        // This rejects an uncommitted terminal, stale transcript, or changed
1190        // Child identity before touching the broker.
1191        let receipts = self
1192            .store
1193            .recover_broker_terminal_receipts(&current)
1194            .await
1195            .map_err(|_| "Host broker terminal checkpoint unverified")?;
1196        if receipts.is_empty() {
1197            return Ok(());
1198        }
1199        let role = current
1200            .metadata
1201            .get("subagent_type")
1202            .map(String::as_str)
1203            .unwrap_or("worker");
1204        for receipt in receipts {
1205            publish_recovered_completion_source(&self.store, &receipt)
1206                .await
1207                .map_err(|_| "committed Child completion source remains pending")?;
1208            if current
1209                .metadata
1210                .contains_key(crate::ticket_worker_plan::TICKET_LOCAL_PLAN_KEY)
1211            {
1212                let service = self
1213                    .ticket_service
1214                    .as_ref()
1215                    .ok_or("Ticket authority unavailable; no result ACK")?;
1216                crate::ticket_runtime::checkpoint_ticket_result(
1217                    service,
1218                    self.store.as_ref(),
1219                    &current,
1220                    &receipt.activation_run_id,
1221                )
1222                .await
1223                .map_err(|_| "Ticket result not checkpointed; broker receipt retained")?;
1224            }
1225            let local_mailbox = format!("p-{}", child.id);
1226            // The committed receipt records the route used by this run. A
1227            // later role configuration change must not redirect its ACK.
1228            let mut link = if receipt.parent_mailbox == local_mailbox {
1229                let bus = self
1230                    .bus
1231                    .as_ref()
1232                    .ok_or("local broker receipt bus unavailable")?;
1233                bamboo_broker::BrokerChildLink::connect_receipt_recovery(
1234                    &bus.endpoint,
1235                    bamboo_subagent::AgentRef {
1236                        session_id: local_mailbox,
1237                        role: None,
1238                    },
1239                    &bus.token,
1240                )
1241                .await
1242                .map_err(|_| "local broker receipt connection unavailable")?
1243            } else if let Some(remote) = self.remote.get(role) {
1244                let route = remote
1245                    .as_ref()
1246                    .ok_or("fixed remote broker route is unavailable or ambiguous")?;
1247                if receipt.parent_mailbox != route.parent.session_id {
1248                    return Err("fixed remote parent mailbox changed since Host checkpoint");
1249                }
1250                let tls = bamboo_broker::client_config_trusting_cert(&route.ca_cert_file)
1251                    .map_err(|_| "fixed remote broker TLS pin unavailable")?;
1252                bamboo_broker::BrokerChildLink::connect_receipt_recovery_with_tls(
1253                    &route.endpoint,
1254                    route.parent.clone(),
1255                    &route.token,
1256                    tls,
1257                )
1258                .await
1259                .map_err(|_| "fixed remote broker receipt connection unavailable")?
1260            } else {
1261                return Err("committed broker receipt route is unavailable");
1262            };
1263            if link.parent_mailbox() != receipt.parent_mailbox {
1264                return Err("broker receipt link mailbox differs from Host checkpoint");
1265            }
1266            bamboo_subagent::ChildLink::acknowledge_recovered_durable_frames(
1267                &mut link,
1268                &receipt.broker_identity,
1269                &receipt.message_ids,
1270            )
1271            .await
1272            .map_err(|_| "broker identity changed or exact ACK unconfirmed")?;
1273            self.store
1274                .clear_acknowledged_broker_terminal_receipt(
1275                    &receipt.session_id,
1276                    receipt.created_at,
1277                    &receipt.activation_run_id,
1278                    &receipt.parent_mailbox,
1279                )
1280                .await
1281                .map_err(|_| "broker ACK succeeded but Host receipt cleanup failed")?;
1282        }
1283        Ok(())
1284    }
1285}
1286
1287/// Finish only the interrupted publication of an already committed Host
1288/// receipt. The pending source stays in the existing protected source record;
1289/// no provider, event admission or transcript reconstruction is involved.
1290async fn publish_recovered_completion_source(
1291    store: &Arc<bamboo_storage::SessionStoreV2>,
1292    receipt: &bamboo_storage::v2::BrokerTerminalReceipt,
1293) -> Result<(), String> {
1294    let expected = store
1295        .load_session(&receipt.session_id)
1296        .await
1297        .map_err(|error| format!("Child completion recovery reload failed: {error}"))?
1298        .ok_or("Child completion recovery snapshot missing")?;
1299    if expected.id != receipt.session_id
1300        || expected.created_at != receipt.created_at
1301        || expected.parent_session_id.as_deref() != Some(receipt.parent_session_id.as_str())
1302        || expected.root_session_id != receipt.root_session_id
1303        || expected.project_id_meta() != receipt.project_id
1304    {
1305        return Err("Child completion recovery identity changed".into());
1306    }
1307    let mut candidate = expected.clone();
1308    if candidate.last_run_status().as_deref() == Some("running") {
1309        candidate.set_last_run_status(receipt.terminal_status.clone());
1310        if let Some(error) = &receipt.terminal_error {
1311            candidate.set_last_run_error(error.clone());
1312        } else {
1313            candidate.clear_last_run_error();
1314        }
1315    }
1316    if crate::execution::ChildCompletionSource::publish_after_durable_delivery(
1317        &mut candidate,
1318        &receipt.activation_run_id,
1319    )? {
1320        let source = crate::execution::ChildCompletionSource::from_committed_session(&candidate)
1321            .ok_or("Child completion recovery source invalid")?;
1322        let published = candidate
1323            .metadata
1324            .get("runtime.child_completion_source_v1")
1325            .ok_or("Child completion recovery source missing")?;
1326        // CAS reloads under the physical Session lock and mutates only the
1327        // source/status fields. Another reconciler or successor cannot be
1328        // overwritten by this earlier snapshot.
1329        let saved = store
1330            .publish_broker_terminal_source(receipt, &expected, published)
1331            .await
1332            .map_err(|error| format!("Child completion recovery publication failed: {error}"))?;
1333        if crate::execution::ChildCompletionSource::from_committed_session(&saved).as_ref()
1334            != Some(&source)
1335        {
1336            return Err("Child completion recovery source changed".into());
1337        }
1338    }
1339    Ok(())
1340}
1341
1342fn receipt_remote_route(
1343    placement: &bamboo_config::RemoteActorPlacement,
1344) -> Option<ReceiptRemoteRoute> {
1345    let peer = placement.broker_peer.as_ref().filter(|peer| peer.valid())?;
1346    let endpoint = url::Url::parse(&placement.endpoint).ok()?;
1347    if placement.endpoint.len() > 2048
1348        || endpoint.scheme() != "wss"
1349        || endpoint.host_str().is_none()
1350        || !endpoint.username().is_empty()
1351        || endpoint.password().is_some()
1352        || endpoint.query().is_some()
1353        || endpoint.fragment().is_some()
1354    {
1355        return None;
1356    }
1357    let token_env = placement
1358        .token_env
1359        .as_deref()
1360        .filter(|name| !name.is_empty() && name.len() <= 256)?;
1361    let token = std::env::var(token_env).ok()?;
1362    if !(32..=256).contains(&token.len()) || !token.bytes().all(|byte| byte.is_ascii_graphic()) {
1363        return None;
1364    }
1365    let ca_cert_file = PathBuf::from(placement.ca_cert_file.as_deref()?);
1366    bamboo_broker::client_config_trusting_cert(&ca_cert_file).ok()?;
1367    Some(ReceiptRemoteRoute {
1368        endpoint: placement.endpoint.clone(),
1369        token,
1370        ca_cert_file,
1371        parent: bamboo_subagent::AgentRef {
1372            session_id: peer.parent_mailbox.clone(),
1373            role: peer.parent_role.clone(),
1374        },
1375    })
1376}
1377
1378/// How `execute_external_child` should obtain its worker connection, decided
1379/// once from `spec.placement`. Splits the divergent acquire/connect + retire
1380/// logic three ways while the shared middle (Run dispatch, live registration,
1381/// drive, close) stays identical. Remote uses a scoped broker peer route.
1382enum PlacementKind {
1383    Local,
1384    Remote,
1385    Schedulable,
1386}
1387
1388/// Ticket execution needs the local process-owned stop checkpoint before ACK.
1389fn ensure_ticket_placement_is_enforceable(
1390    session: &Session,
1391    spec: &ProvisionSpec,
1392) -> Result<(), AgentError> {
1393    let ticket = session
1394        .metadata
1395        .contains_key(crate::ticket_runtime::TICKET_DISPATCH_KEY)
1396        || session
1397            .metadata
1398            .contains_key(crate::ticket_worker_plan::TICKET_LOCAL_PLAN_KEY);
1399    if ticket && !matches!(spec.placement, Placement::Local) {
1400        return Err(AgentError::LLM(
1401            "Ticket execution requires local placement and owned stop proof".into(),
1402        ));
1403    }
1404    Ok(())
1405}
1406
1407/// A read-only activation needs an executor and worker whose tool surface and
1408/// no-shell checker enforce the typed capability. Codex's read-only sandbox
1409/// prevents writes but still permits command execution, so neither Codex mode
1410/// can currently satisfy the no-shell contract. Local Bamboo/Claude workers are
1411/// fingerprinted by their baked fields, while remote/schedulable placements
1412/// connect to an already-provisioned resident and [`RunSpec`] currently carries
1413/// no equivalent per-activation authority. Fail closed before provisioning or
1414/// dispatch instead of treating either posture as read-only enforcement.
1415fn ensure_read_only_activation_is_enforceable(spec: &ProvisionSpec) -> Result<(), AgentError> {
1416    if !spec.capabilities.read_only_enforced() {
1417        return Ok(());
1418    }
1419
1420    if matches!(spec.executor, ExecutorSpec::Codex { .. }) {
1421        return Err(AgentError::LLM(format!(
1422            "read-only actor role '{}' cannot use the Codex executor because its read-only sandbox still permits command execution; use bamboo_runtime or claude_code",
1423            spec.identity.role
1424        )));
1425    }
1426
1427    if !matches!(spec.placement, Placement::Local) {
1428        return Err(AgentError::LLM(format!(
1429            "read-only actor role '{}' requires local placement; remote and schedulable resident workers cannot verify this activation's read-only tool boundary",
1430            spec.identity.role
1431        )));
1432    }
1433
1434    Ok(())
1435}
1436
1437/// Spawns and drives a child session as an independent actor: a `bamboo-subagent` worker process.
1438pub struct ActorChildRunner {
1439    approval_registry: Option<super::approval_registry::SharedApprovalRegistry>,
1440    permission_config: Option<Arc<bamboo_tools::permission::PermissionConfig>>,
1441    agent_id: String,
1442    worker_bin: PathBuf,
1443    worker_args: Vec<String>,
1444    builtin_required_context_route: bool,
1445    native_tool_ceiling: Option<Arc<dyn super::runtime::NativeToolCeilingSource>>,
1446    fabric_dir: PathBuf,
1447    executor: ExecutorSpec,
1448    /// Per-provider credentials snapshotted from the parent config at build
1449    /// time; the spec carries only the ONE the child's provider needs. Server
1450    /// runtimes install `live_provider_config` below so a provider hot reload
1451    /// cannot strand this fallback snapshot for the rest of the process.
1452    credentials: Vec<ScopedCredential>,
1453    /// Parent's default provider (used when the child has no explicit one).
1454    default_provider: String,
1455    /// Startup fallback for embeddings without a live Host configuration.
1456    max_spawn_depth: u32,
1457    /// Live server configuration used to resolve provider credentials at child
1458    /// activation time, including the depth cap for new worker provisioning.
1459    /// Executor/placement policy remains the immutable runner configuration.
1460    live_provider_config: Option<Arc<tokio::sync::RwLock<bamboo_llm::Config>>>,
1461    /// The mailbox bus to run local children over (the unified transport). Local
1462    /// sub-agents require it; `None` only when no broker could be embedded.
1463    bus: Option<bamboo_subagent::BusEndpoint>,
1464    /// Backpressure: bounds the number of concurrently *running* actors; further
1465    /// runs wait for a slot instead of exploding the process table. (Idle pooled
1466    /// workers do not hold a slot.)
1467    concurrency: std::sync::Arc<tokio::sync::Semaphore>,
1468    /// Warm-worker pool keyed by a reuse fingerprint
1469    /// (role/provider/model/workspace/disabled-tools/baked-caps). A finished run
1470    /// parks its bus worker here so the next interchangeable child reuses it
1471    /// (delivers its `Run` to the same mailbox) instead of spawning a fresh
1472    /// process — collapsing N sibling sub-agents onto a few warm workers.
1473    pool: Arc<tokio::sync::Mutex<WorkerPool>>,
1474    max_idle_per_key: usize,
1475    max_idle_total: usize,
1476    pool_reaper_started: AtomicBool,
1477    /// Monotonic fencing token for concrete worker attempts in this runtime.
1478    /// `activation_run_id` fences across runtime restarts; this epoch also
1479    /// fences a late first worker when an activation is retried in-place.
1480    next_execution_epoch: AtomicU64,
1481    /// Host-side decision for a child's gated-tool approval request (Phase 2).
1482    /// `None` ⇒ fail-closed DENY (the safe default). A wired decider (policy or
1483    /// human-routing bridge) returns approve/deny over the actor WS.
1484    approval_decider: Option<Arc<dyn ChildApprovalDecider>>,
1485    /// Off-loop parent-agent reviewer for forced-ask requests. The root server
1486    /// wires a session-aware reviewer; nested workers wire their owning model
1487    /// reviewer directly into the per-run runner.
1488    approval_reviewer: Option<Arc<dyn ChildApprovalReviewer>>,
1489    /// Per-run escalation host bridge for non-bypass child-approval routing (#68;
1490    /// Phase 6, Part B). The owning worker's `run()` installs its OWN host bridge
1491    /// here via `set_escalation_bridge`; `execute_external_child` CAPTURES it at
1492    /// grandchild-spawn time and hands the owned value to `drive()`, which uses it
1493    /// to RE-PROXY a child's approval request UP to the parent run — chaining up
1494    /// every level until a bypass level (model-review) or the top orchestrator
1495    /// (human) decides, then relaying the reply back down. Was a process-global
1496    /// slot; now per-runner so a fire-and-forget grandchild that OUTLIVES the run
1497    /// that spawned it keeps that run's bridge for its whole lifetime instead of
1498    /// reading a stale/overwritten global at approval time (→ fail-closed deny).
1499    escalation_bridge: Arc<std::sync::Mutex<Option<bamboo_subagent::executor::HostBridge>>>,
1500    /// Roles pinned to a REMOTE resident worker (#193), keyed by sub-agent role
1501    /// (the child's `subagent_type`). A role present here routes through the
1502    /// dedicated remote branch in `execute_external_child` (Bearer-authenticated
1503    /// `wss://` connect, no spawn, no pool, no kill) instead of the local
1504    /// subprocess + warm-pool path. Empty (the default) = all-local behavior.
1505    remote_placements: HashMap<String, ResolvedRemotePlacement>,
1506    /// Fixed parent mailbox subscription has exactly one in-process owner.
1507    strict_remote_subscription: tokio::sync::Mutex<()>,
1508    /// Roles routed to a REGISTRY-SCHEDULED worker (#181, P2b), keyed by sub-agent
1509    /// role. A role present here (AND not already in `remote_placements`, which
1510    /// wins) routes through the dedicated SCHEDULABLE branch in
1511    /// `execute_external_child`: query the registry for live workers in the pool,
1512    /// pick one (round-robin), connect over `wss://` — no spawn, no pool, no kill,
1513    /// and NO local-subprocess fallback (no live worker ⇒ a clear error). Empty
1514    /// (the default) = all-local behavior.
1515    schedulable_placements: HashMap<String, ResolvedSchedulablePlacement>,
1516    /// Per-pool round-robin cursor for schedulable scheduling (#181, P2b). Bumped
1517    /// once per pick so successive sibling spawns SPREAD across a pool's live
1518    /// workers instead of all landing on the first candidate. Best-effort spread,
1519    /// not a load balancer — the registry's live set can change between picks.
1520    /// Optional server authority used only by `Codex` in `bamboo` auth mode.
1521    codex_run_tokens: Option<Arc<dyn CodexRunTokenAuthority>>,
1522    /// Canonical logical-session inbox resources, late-bound by each owning
1523    /// runtime. Kept per runner/runtime; never process-global.
1524    session_inbox_runtime: Arc<std::sync::Mutex<Option<SessionInboxRuntimeBinding>>>,
1525    actor_directory_store: std::sync::Mutex<Option<Arc<bamboo_storage::SessionStoreV2>>>,
1526    canonical_subagent_tool: std::sync::Mutex<Option<Arc<dyn bamboo_agent_core::tools::Tool>>>,
1527    ticket_service: std::sync::Mutex<Option<Arc<bamboo_tickets::TicketService>>>,
1528    actor_event_observer: std::sync::Mutex<Option<Arc<dyn ActorEventObserver>>>,
1529    /// A processed broker terminal still needs the Host's final Child Session
1530    /// save before its exact Event/Outcome mailbox receipts can be ACKed.
1531    pending_durable_links: tokio::sync::Mutex<
1532        HashMap<(String, chrono::DateTime<chrono::Utc>, String), PendingDurableChildLink>,
1533    >,
1534}
1535
1536struct PendingDurableChildLink {
1537    link: Box<dyn bamboo_subagent::ChildLink>,
1538    /// Captured from this Host frame pump after canonical nested-wait checks.
1539    /// The map key binds it to the exact Child birth and activation Run.
1540    host_nested_wait_handoff: bool,
1541    // Never reconstructed from Worker JSON or ordinary Session metadata.
1542    strict_history: Option<StrictTerminalHistory>,
1543    strict_history_required: bool,
1544    strict_history_committed: bool,
1545}
1546
1547#[derive(Clone)]
1548struct StrictTerminalHistory {
1549    proof: bamboo_storage::v2::HostTerminalCompleteness,
1550    watermark: bamboo_subagent::ActorEventWatermark,
1551}
1552
1553fn effective_broker_terminal_status(
1554    raw: TerminalStatus,
1555    host_nested_wait_handoff: bool,
1556) -> Result<&'static str, String> {
1557    if host_nested_wait_handoff {
1558        if raw != TerminalStatus::Completed {
1559            return Err("nested wait cannot project a failed broker terminal".into());
1560        }
1561        Ok("suspended")
1562    } else {
1563        Ok(broker_terminal_status_label(raw))
1564    }
1565}
1566
1567fn broker_terminal_status_label(status: TerminalStatus) -> &'static str {
1568    match status {
1569        TerminalStatus::Completed => "completed",
1570        TerminalStatus::Suspended => "suspended",
1571        TerminalStatus::Error => "error",
1572        TerminalStatus::Cancelled => "cancelled",
1573    }
1574}
1575
1576/// Decides how the host answers a child worker's gated-tool approval request
1577/// (Phase 2: child → parent approval delegation). Async so an implementation
1578/// can consult a policy. With no decider wired the host replies with a
1579/// fail-closed DENY.
1580///
1581/// NOTE: `decide` is awaited inside the per-child frame pump, so an
1582/// implementation must resolve promptly (e.g. a policy lookup). Model-based
1583/// review belongs in [`ChildApprovalReviewer`], which runs off-loop and returns
1584/// through the live steering channel without stalling the frame pump.
1585#[async_trait]
1586pub trait ChildApprovalDecider: Send + Sync {
1587    /// Decide whether `child_session_id` may perform the gated action described
1588    /// by `request` (`{tool_name, permission, resource}`).
1589    async fn decide(&self, child_session_id: &str, request: &serde_json::Value) -> bool;
1590}
1591
1592/// Resolve a child approval request to approve/deny. Fail-closed (DENY) when no
1593/// decider is wired — the single, testable seam for the host-side decision.
1594async fn decide_child_approval(
1595    decider: Option<&Arc<dyn ChildApprovalDecider>>,
1596    child_session_id: &str,
1597    request: &serde_json::Value,
1598) -> bool {
1599    match decider {
1600        Some(decider) => decider.decide(child_session_id, request).await,
1601        None => false,
1602    }
1603}
1604
1605/// How long a chained parent-agent review may take before the child's gated
1606/// tool fails closed (DENY). Bounds an unanswered request so it cannot hang the
1607/// worker indefinitely.
1608const CHILD_APPROVAL_TIMEOUT: Duration = Duration::from_secs(300);
1609
1610/// Off-loop reviewer for a child's gated-tool approval request (Phase 6, Part B).
1611///
1612/// Installed (process-global) by a BYPASSED self-orchestrating worker so its
1613/// children's forced-ask (dangerous) gated actions — which still raise
1614/// `ConfirmationRequired` even under bypass — get an LLM reasonableness check
1615/// rather than a blind pass. `review` is an LLM call: `drive()` invokes it in a
1616/// SPAWNED task (NEVER in the frame pump) and delivers the verdict async via the
1617/// live channel, so the agent loop is never blocked.
1618/// Process-local Host scope. Durable data can compare its stamp but cannot
1619/// deserialize a new live authority from it.
1620#[derive(Clone)]
1621pub struct ChildApprovalScope {
1622    stamp: serde_json::Value,
1623    deadline: chrono::DateTime<chrono::Utc>,
1624    parent: String,
1625    child: String,
1626    run: String,
1627    epoch: u64,
1628    current_epoch: Arc<AtomicU64>,
1629    cancel: CancellationToken,
1630    router: Arc<crate::SessionActivationRouter>,
1631    admission: Arc<AtomicUsize>,
1632}
1633
1634pub struct ChildApprovalAdmission(Arc<AtomicUsize>);
1635impl Drop for ChildApprovalAdmission {
1636    fn drop(&mut self) {
1637        self.0.store(2, Ordering::Release);
1638    }
1639}
1640
1641impl ChildApprovalScope {
1642    /// Called by the actual Host frame pump, never from a wire payload.
1643    pub fn new(
1644        parent: &str,
1645        child: &str,
1646        transport: (u32, &str, u64, &str, chrono::DateTime<chrono::Utc>),
1647        router: Arc<crate::SessionActivationRouter>,
1648        cancel: CancellationToken,
1649        current_epoch: Arc<AtomicU64>,
1650    ) -> Self {
1651        let (attempt, run, epoch, reply, deadline) = transport;
1652        let deadline = deadline.min(chrono::Utc::now() + chrono::Duration::seconds(240));
1653        Self {
1654            stamp: serde_json::json!({"host_scope": uuid::Uuid::new_v4().to_string(),
1655                "attempt": attempt, "run": run, "epoch": epoch, "reply": reply,
1656                "deadline": deadline}),
1657            deadline,
1658            parent: parent.into(),
1659            child: child.into(),
1660            run: run.into(),
1661            epoch,
1662            current_epoch,
1663            cancel,
1664            router,
1665            admission: Arc::new(AtomicUsize::new(0)),
1666        }
1667    }
1668    pub fn stamp(&self) -> &serde_json::Value {
1669        &self.stamp
1670    }
1671    pub fn deadline(&self) -> chrono::DateTime<chrono::Utc> {
1672        self.deadline
1673    }
1674    pub fn first_admission(&self) -> Option<ChildApprovalAdmission> {
1675        self.admission
1676            .compare_exchange(0, 1, Ordering::AcqRel, Ordering::Acquire)
1677            .ok()
1678            .map(|_| ChildApprovalAdmission(self.admission.clone()))
1679    }
1680    pub fn is_admitting(&self) -> bool {
1681        self.admission.load(Ordering::Acquire) == 1
1682    }
1683    pub async fn is_current(&self, parent: &str, child: &str) -> bool {
1684        self.parent == parent
1685            && self.child == child
1686            && !self.cancel.is_cancelled()
1687            && self.current_epoch.load(Ordering::Acquire) == self.epoch
1688            && self.router.owns_run(child, &self.run).await
1689    }
1690}
1691
1692#[derive(Debug, Clone, Copy, PartialEq, Eq)]
1693pub enum ChildApprovalReview {
1694    Reply(bool),
1695    NoReply,
1696}
1697
1698#[async_trait]
1699pub trait ChildApprovalReviewer: Send + Sync {
1700    /// Judge whether the gated action `request` (`{tool_name, permission,
1701    /// resource}`) is reasonable for `child_session_id`'s task. `true` = approve.
1702    async fn review(
1703        &self,
1704        parent_session_id: &str,
1705        child_session_id: &str,
1706        request: &serde_json::Value,
1707    ) -> bool;
1708
1709    async fn review_scoped(
1710        &self,
1711        parent: &str,
1712        child: &str,
1713        request: &serde_json::Value,
1714        _scope: &ChildApprovalScope,
1715    ) -> ChildApprovalReview {
1716        ChildApprovalReview::Reply(self.review(parent, child, request).await)
1717    }
1718}
1719
1720fn child_approval_reviewer_slot() -> &'static std::sync::OnceLock<Arc<dyn ChildApprovalReviewer>> {
1721    static SLOT: std::sync::OnceLock<Arc<dyn ChildApprovalReviewer>> = std::sync::OnceLock::new();
1722    &SLOT
1723}
1724
1725/// Install the process-global child-approval reviewer (idempotent; first wins).
1726pub fn set_child_approval_reviewer(reviewer: Arc<dyn ChildApprovalReviewer>) {
1727    let _ = child_approval_reviewer_slot().set(reviewer);
1728}
1729
1730/// The process-global child-approval reviewer, if installed.
1731pub fn child_approval_reviewer() -> Option<Arc<dyn ChildApprovalReviewer>> {
1732    child_approval_reviewer_slot().get().cloned()
1733}
1734
1735impl ActorChildRunner {
1736    #[allow(clippy::too_many_arguments)]
1737    pub fn new(
1738        agent_id: String,
1739        worker_bin: PathBuf,
1740        worker_args: Vec<String>,
1741        fabric_dir: PathBuf,
1742        executor: ExecutorSpec,
1743        credentials: Vec<ScopedCredential>,
1744        default_provider: String,
1745        max_concurrent: usize,
1746    ) -> Self {
1747        Self {
1748            approval_registry: None,
1749            permission_config: None,
1750            agent_id,
1751            worker_bin,
1752            worker_args,
1753            builtin_required_context_route: false,
1754            fabric_dir,
1755            executor,
1756            credentials,
1757            default_provider,
1758            max_spawn_depth: MAX_SPAWN_DEPTH,
1759            live_provider_config: None,
1760            native_tool_ceiling: None,
1761            bus: None,
1762            concurrency: std::sync::Arc::new(tokio::sync::Semaphore::new(max_concurrent.max(1))),
1763            pool: Arc::new(tokio::sync::Mutex::new(HashMap::new())),
1764            max_idle_per_key: DEFAULT_MAX_IDLE_PER_KEY,
1765            max_idle_total: DEFAULT_MAX_IDLE_TOTAL,
1766            pool_reaper_started: AtomicBool::new(false),
1767            next_execution_epoch: AtomicU64::new(0),
1768            approval_decider: None,
1769            approval_reviewer: None,
1770            escalation_bridge: Arc::new(std::sync::Mutex::new(None)),
1771            remote_placements: HashMap::new(),
1772            strict_remote_subscription: tokio::sync::Mutex::new(()),
1773            schedulable_placements: HashMap::new(),
1774            codex_run_tokens: None,
1775            session_inbox_runtime: Arc::new(std::sync::Mutex::new(None)),
1776            actor_directory_store: std::sync::Mutex::new(None),
1777            canonical_subagent_tool: std::sync::Mutex::new(None),
1778            ticket_service: std::sync::Mutex::new(None),
1779            actor_event_observer: std::sync::Mutex::new(None),
1780            pending_durable_links: tokio::sync::Mutex::new(HashMap::new()),
1781        }
1782    }
1783
1784    /// Bind the AppState-owned live configuration. Child activations read this
1785    /// after a successful provider reload, so newly added/rotated provider
1786    /// credentials take effect without restarting Bamboo.
1787    /// Only the built-in factory sets this from its immutable launch config.
1788    pub(crate) fn with_builtin_required_context_route(mut self, supported: bool) -> Self {
1789        self.builtin_required_context_route = supported;
1790        self
1791    }
1792
1793    pub(crate) fn with_native_tool_ceiling_source(
1794        mut self,
1795        source: Option<Arc<dyn super::runtime::NativeToolCeilingSource>>,
1796    ) -> Self {
1797        self.native_tool_ceiling = source;
1798        self
1799    }
1800
1801    async fn require_initial_input_release_worker(&self) -> crate::runtime::runner::Result<()> {
1802        bamboo_subagent::fleet::require_worker_capability(
1803            &self.worker_bin,
1804            &self.worker_args,
1805            bamboo_subagent::provision::INITIAL_INPUT_RELEASE_WORKER_CAPABILITY,
1806        )
1807        .await
1808        .map_err(|_| AgentError::LLM("owned initial release worker capability unconfirmed".into()))
1809    }
1810
1811    pub fn with_live_provider_config(
1812        mut self,
1813        config: Arc<tokio::sync::RwLock<bamboo_llm::Config>>,
1814    ) -> Self {
1815        self.live_provider_config = Some(config);
1816        self
1817    }
1818
1819    pub fn with_max_spawn_depth(mut self, max_spawn_depth: u32) -> Self {
1820        self.max_spawn_depth = max_spawn_depth;
1821        self
1822    }
1823
1824    pub fn with_approval_registry(
1825        mut self,
1826        registry: super::approval_registry::SharedApprovalRegistry,
1827    ) -> Self {
1828        self.approval_registry = Some(registry);
1829        self
1830    }
1831
1832    pub fn with_permission_config(
1833        mut self,
1834        config: Arc<bamboo_tools::permission::PermissionConfig>,
1835    ) -> Self {
1836        self.permission_config = Some(config);
1837        self
1838    }
1839
1840    /// Run children over the mailbox bus (the unified actor+mailbox transport).
1841    /// When set, local children dial this bus and are driven by mailbox id; when
1842    /// unset they use the legacy direct-WS path. The server passes its in-process
1843    /// broker here (`subagents.broker`); tests without a broker leave it unset.
1844    pub fn with_bus(mut self, bus: Option<bamboo_subagent::BusEndpoint>) -> Self {
1845        self.bus = bus.filter(|b| !b.endpoint.trim().is_empty());
1846        self
1847    }
1848
1849    /// Wire the host-side decider for child gated-tool approval requests
1850    /// (Phase 2). Without this the host fail-closed DENYs every request.
1851    pub fn with_approval_decider(mut self, decider: Arc<dyn ChildApprovalDecider>) -> Self {
1852        self.approval_decider = Some(decider);
1853        self
1854    }
1855
1856    pub fn with_approval_reviewer(mut self, reviewer: Arc<dyn ChildApprovalReviewer>) -> Self {
1857        self.approval_reviewer = Some(reviewer);
1858        self
1859    }
1860
1861    pub fn with_codex_run_tokens(
1862        mut self,
1863        authority: Option<Arc<dyn CodexRunTokenAuthority>>,
1864    ) -> Self {
1865        self.codex_run_tokens = authority;
1866        self
1867    }
1868
1869    /// Pin specific sub-agent roles to remote resident workers (#193). The map
1870    /// is keyed by role (`subagent_type`); a child whose role is present connects
1871    /// over `wss://` to the resolved endpoint instead of spawning a local
1872    /// subprocess. Default (empty) keeps every role on the local path — exactly
1873    /// today's behavior.
1874    pub fn with_remote_placements(
1875        mut self,
1876        placements: HashMap<String, ResolvedRemotePlacement>,
1877    ) -> Self {
1878        self.remote_placements = placements;
1879        self
1880    }
1881
1882    /// Route specific sub-agent roles to a registry-SCHEDULED worker (#181, P2b).
1883    /// The map is keyed by role (`subagent_type`); a child whose role is present
1884    /// (and NOT already pinned by `remote_placements`, which takes precedence) is
1885    /// run on a live worker discovered from the registry instead of a local
1886    /// subprocess. Default (empty) keeps every role on the local path.
1887    pub fn with_schedulable_placements(
1888        mut self,
1889        placements: HashMap<String, ResolvedSchedulablePlacement>,
1890    ) -> Self {
1891        self.schedulable_placements = placements;
1892        self
1893    }
1894
1895    /// Reuse fingerprint: two children are interchangeable on one warm worker iff
1896    /// they share role, provider, model, provider-credential revision,
1897    /// workspace, disabled-tool set, AND every
1898    /// capability the worker BAKES at provision time (`BambooRuntimeExecutor`
1899    /// stamps these once and reuses them across runs): nesting depth, nested-spawn
1900    /// stack, requested/effective permission modes, legacy bypass/auto flags,
1901    /// permission enforcement, and the depth cap. Omitting any
1902    /// of these lets the pool hand a run a worker baked for a DIFFERENT posture —
1903    /// e.g. a depth-1 worker (with its own spawn stack) reused for a depth-4
1904    /// child would re-stamp `spawn_depth=1` and pass the depth-cap check, breaking
1905    /// the recursion bound; or a bypass worker reused for a non-bypass child. So
1906    /// these MUST split the pool bucket. Everything else (assignment, history) is
1907    /// shipped per-run in the `RunSpec` and does not affect the fingerprint.
1908    /// Reuse fingerprint (role/provider/model/workspace/disabled-tools/baked
1909    /// caps): two children with the same fingerprint are interchangeable on one
1910    /// warm worker, so they share a pool bucket. Any axis the worker bakes ONCE
1911    /// at provision time MUST be in here, else a worker baked for one posture
1912    /// gets reused for another (see the `fingerprint_*` tests).
1913    fn fingerprint(spec: &ProvisionSpec) -> String {
1914        let role = spec.identity.role.as_str();
1915        let (provider, model) = spec
1916            .model
1917            .as_ref()
1918            .map(|m| (m.provider.as_str(), m.model.as_str()))
1919            .unwrap_or(("", ""));
1920        let workspace = spec.workspace.as_deref().unwrap_or("");
1921        let mut tools = spec.disabled_tools.clone().unwrap_or_default();
1922        tools.sort();
1923        let caps = &spec.capabilities;
1924        // Provider credentials are baked into a worker at provision time. Hash
1925        // the scoped envelope so a hot-reloaded key/base URL cannot reuse a
1926        // worker provisioned with stale authority. The key itself never enters
1927        // the pool key or logs.
1928        let credential_revision = if spec.secrets.provider_credentials.is_empty() {
1929            "none".to_string()
1930        } else {
1931            let serialized = serde_json::to_vec(&spec.secrets.provider_credentials)
1932                .expect("scoped provider credentials are serializable");
1933            hex::encode(Sha256::digest(serialized))
1934        };
1935        // The worker constructs its executor exactly once. In particular,
1936        // Codex exec and app-server workers are not interchangeable.
1937        let executor = serde_json::to_string(&spec.executor).unwrap_or_default();
1938        format!(
1939            "{role}\u{1}{provider}\u{1}{model}\u{1}cred={credential_revision}\u{1}{workspace}\u{1}{}\u{1}d={}\u{1}ns={}\u{1}pr={}\u{1}pe={}\u{1}by={}\u{1}auto={}\u{1}ep={}\u{1}md={}\u{1}nha={}\u{1}ro={}\u{1}gro={}\u{1}ci={}\u{1}executor={executor}",
1940            tools.join(","),
1941            spec.identity.depth,
1942            caps.nested_spawn,
1943            caps.permission_requested_mode,
1944            caps.permission_effective_mode,
1945            caps.bypass,
1946            caps.auto_approve_permissions,
1947            caps.enforce_permissions,
1948            caps.max_spawn_depth.unwrap_or(0),
1949            // #73 review (P1): a worker bakes `no_human_review` ONCE from this flag
1950            // at build() and never re-reads it per run, so the pool MUST NOT hand a
1951            // worker baked for one approval posture to a run of the opposite one —
1952            // else a scheduled-root worker reused for an interactive child would
1953            // silently model-review instead of asking the human (and vice-versa,
1954            // reintroducing the 300s-deny). Split the bucket on it.
1955            caps.no_human_approver,
1956            // The read-only no-shell checker is baked once at build(), so a planner
1957            // or Guardian worker must NOT be reused for an ordinary child.
1958            caps.read_only,
1959            // Preserve the legacy Guardian bit as a distinct fingerprint axis
1960            // during rolling upgrades even though both bits enforce read-only.
1961            caps.guardian_read_only,
1962            caps.child_creation_identity,
1963        )
1964    }
1965
1966    /// Start one weakly-owned sweep task for this runner. It does not keep the
1967    /// runner/pool alive, and it never holds the pool lock while killing a child.
1968    fn ensure_pool_reaper(&self) {
1969        if self
1970            .pool_reaper_started
1971            .compare_exchange(false, true, Ordering::AcqRel, Ordering::Acquire)
1972            .is_err()
1973        {
1974            return;
1975        }
1976        let pool = Arc::downgrade(&self.pool);
1977        tokio::spawn(async move {
1978            let mut interval = tokio::time::interval(POOLED_REAPER_INTERVAL);
1979            interval.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Skip);
1980            // Tokio intervals tick immediately once; consume that tick so a new
1981            // runner does not scan an empty pool before its first worker parks.
1982            interval.tick().await;
1983            loop {
1984                interval.tick().await;
1985                let Some(pool) = pool.upgrade() else {
1986                    break;
1987                };
1988                let _ = reap_worker_pool_once(&pool).await;
1989            }
1990        });
1991    }
1992
1993    fn fresh_required_worker_spec(spec: &ProvisionSpec) -> ProvisionSpec {
1994        let mut physical = spec.clone();
1995        // A killed one-shot worker may leave an unACKed Run in its mailbox.
1996        // A fresh subscriber must not replay that Run. Logical birth, storage
1997        // namespace, typed input targets and authority remain on the RunSpec.
1998        physical.identity.child_id = format!("required-worker-{}", uuid::Uuid::new_v4());
1999        physical
2000    }
2001
2002    /// Check out a warm bus worker for `key`, reusing a live parked one if any,
2003    /// else spawning a fresh one that dials the bus. The returned worker is OWNED
2004    /// by the caller for the run's duration (checkout removes it from the pool, so
2005    /// a concurrent sibling gets a different worker or spawns its own — one run per
2006    /// worker at a time, matching the pre-bus pool semantics).
2007    async fn acquire_bus_worker(
2008        &self,
2009        key: &str,
2010        spec: &ProvisionSpec,
2011    ) -> crate::runtime::runner::Result<PooledWorker> {
2012        if spec.capabilities.required_child_context {
2013            let physical = Self::fresh_required_worker_spec(spec);
2014            let worker = spawn_worker_on_bus(&self.worker_bin, &self.worker_args, &physical)
2015                .await
2016                .map_err(|error| {
2017                    AgentError::LLM(format!("required-context worker spawn failed: {error}"))
2018                })?;
2019            let mailbox_id = worker.record.agent_id.clone();
2020            return Ok(PooledWorker {
2021                worker,
2022                mailbox_id,
2023                parked_at: None,
2024            });
2025        }
2026        self.ensure_pool_reaper();
2027        // Drain the bucket, skipping (and reaping) any worker whose process exited
2028        // or crossed its idle deadline while parked. A live, fresh one is handed
2029        // straight out for reuse.
2030        loop {
2031            let candidate = {
2032                let mut pool = self.pool.lock().await;
2033                let candidate = pool.get_mut(key).and_then(|bucket| bucket.pop());
2034                if pool.get(key).is_some_and(Vec::is_empty) {
2035                    pool.remove(key);
2036                }
2037                candidate
2038            };
2039            let Some(mut candidate) = candidate else {
2040                break;
2041            };
2042            let now = Instant::now();
2043            let expired = parked_worker_expired(
2044                candidate.parked_at,
2045                now,
2046                Duration::from_secs(POOLED_IDLE_TIMEOUT_SECS),
2047            );
2048            if candidate.worker.is_alive() && !expired {
2049                candidate.parked_at = None;
2050                return Ok(candidate);
2051            }
2052            tracing::info!(
2053                pool_key = key,
2054                mailbox_id = %candidate.mailbox_id,
2055                worker_pid = ?candidate.worker.pid(),
2056                shutdown_reason = if expired {
2057                    PoolReapReason::IdleTimeout.as_str()
2058                } else {
2059                    PoolReapReason::ProcessExited.as_str()
2060                },
2061                "discarding unusable parked sub-agent worker during checkout"
2062            );
2063            candidate.worker.kill().await;
2064        }
2065
2066        let spawned = spawn_worker_on_bus(&self.worker_bin, &self.worker_args, spec)
2067            .await
2068            .map_err(|e| AgentError::LLM(format!("actor spawn (bus) failed: {e}")))?;
2069        let mailbox_id = spawned.record.agent_id.clone();
2070        Ok(PooledWorker {
2071            worker: spawned,
2072            mailbox_id,
2073            parked_at: None,
2074        })
2075    }
2076
2077    /// Park a warm bus worker for reuse after a clean run; if its bucket is full
2078    /// (or it died), kill it instead. The worker stays dialed-in + subscribed
2079    /// while parked, so a reusing child just delivers a new `Run` to its mailbox.
2080    async fn release_bus_worker(&self, key: &str, mut worker: PooledWorker) {
2081        self.ensure_pool_reaper();
2082        if !worker.worker.is_alive() {
2083            tracing::info!(
2084                pool_key = key,
2085                mailbox_id = %worker.mailbox_id,
2086                worker_pid = ?worker.worker.pid(),
2087                shutdown_reason = "process_exited_before_park",
2088                "discarding sub-agent worker instead of parking"
2089            );
2090            worker.worker.kill().await;
2091            return;
2092        }
2093        worker.parked_at = Some(Instant::now());
2094        let mut pool = self.pool.lock().await;
2095        let key_count = pool.get(key).map_or(0, Vec::len);
2096        let total_count = pool.values().map(Vec::len).sum::<usize>();
2097        let limit_reason =
2098            idle_pool_limit_reason(&pool, key, self.max_idle_per_key, self.max_idle_total);
2099        if let Some(shutdown_reason) = limit_reason {
2100            drop(pool);
2101            tracing::info!(
2102                pool_key = key,
2103                mailbox_id = %worker.mailbox_id,
2104                worker_pid = ?worker.worker.pid(),
2105                idle_workers_for_key = key_count,
2106                idle_workers_total = total_count,
2107                max_idle_per_key = self.max_idle_per_key,
2108                max_idle_total = self.max_idle_total,
2109                shutdown_reason,
2110                "discarding sub-agent worker because the warm pool is full"
2111            );
2112            worker.worker.kill().await;
2113            return;
2114        }
2115        pool.entry(key.to_string()).or_default().push(worker);
2116    }
2117
2118    /// Assemble the parent-resolved provisioning document for this child using
2119    /// one coherent provider snapshot.
2120    fn build_spec_with_provider_config(
2121        &self,
2122        session: &Session,
2123        job: &SpawnJob,
2124        credentials: &[ScopedCredential],
2125        default_provider: &str,
2126        max_spawn_depth: u32,
2127    ) -> ProvisionSpec {
2128        let mut spec = ProvisionSpec::new(
2129            ChildIdentity {
2130                child_id: job.child_session_id.clone(),
2131                parent_id: Some(job.parent_session_id.clone()),
2132                project_key: None,
2133                role: session
2134                    .metadata
2135                    .get("subagent_type")
2136                    .cloned()
2137                    .unwrap_or_else(|| "worker".to_string()),
2138                // The child session already carries the correct depth
2139                // (create_child_action's new_child_of did parent.spawn_depth+1);
2140                // stamp it so the worker can re-establish it on its run session
2141                // and enforce the max-depth cap across the actor boundary.
2142                depth: session.spawn_depth,
2143            },
2144            self.executor.clone(),
2145            self.fabric_dir.to_string_lossy().into_owned(),
2146        );
2147        spec.workspace = session.workspace.clone();
2148        if let ExecutorSpec::Codex {
2149            workspace_owned, ..
2150        } = &mut spec.executor
2151        {
2152            *workspace_owned = Some(
2153                spec.workspace
2154                    .as_deref()
2155                    .is_some_and(workspace_is_bamboo_owned),
2156            );
2157        }
2158        // Unified transport: when a bus is configured, the child dials it (no
2159        // listen socket / file discovery) and the parent drives it by mailbox id.
2160        spec.bus = self.bus.clone();
2161        // Final model: the session's pinned model_ref (create.model / routing already applied),
2162        // falling back to the job's bare model on the parent's default provider.
2163        spec.model = session
2164            .model_ref
2165            .as_ref()
2166            .map(|r| ModelRefSpec {
2167                provider: r.provider.clone(),
2168                model: r.model.clone(),
2169            })
2170            .or_else(|| {
2171                let m = job.model.trim();
2172                (!m.is_empty()).then(|| ModelRefSpec {
2173                    provider: default_provider.to_string(),
2174                    model: m.to_string(),
2175                })
2176            });
2177        spec.disabled_tools = job.disabled_tools.clone();
2178        match &spec.executor {
2179            // Codex auth is independent of the session's normal Bamboo model
2180            // provider. Inherit/API-key/Bamboo modes need no credential-store
2181            // secret at provisioning; custom mode gets exactly its referenced
2182            // key. This prevents an unrelated upstream provider key from
2183            // reaching a Codex worker that only needs a per-run bcx1_ token.
2184            ExecutorSpec::Codex {
2185                auth_mode,
2186                provider_key_ref,
2187                ..
2188            } => {
2189                if auth_mode.as_deref() == Some("custom") {
2190                    if let Some(reference) = provider_key_ref {
2191                        if let Some(credential) = credentials.iter().find(|credential| {
2192                            credential.credential_ref.as_deref() == Some(reference)
2193                        }) {
2194                            spec.secrets.provider_credentials.push(credential.clone());
2195                        } else {
2196                            tracing::warn!(
2197                                "actor child {}: custom Codex credential reference '{}' did not resolve",
2198                                job.child_session_id,
2199                                reference
2200                            );
2201                        }
2202                    } else {
2203                        tracing::warn!(
2204                            "actor child {}: custom Codex executor has no credential reference",
2205                            job.child_session_id
2206                        );
2207                    }
2208                }
2209            }
2210            // Other executors keep the existing least-privilege contract: only
2211            // the credential for the child session's selected provider.
2212            _ => {
2213                let provider = spec
2214                    .model
2215                    .as_ref()
2216                    .map(|model| model.provider.as_str())
2217                    .filter(|provider| !provider.trim().is_empty())
2218                    .unwrap_or(default_provider);
2219                if let Some(credential) = credentials
2220                    .iter()
2221                    .find(|credential| credential.provider == provider)
2222                {
2223                    spec.secrets.provider_credentials.push(credential.clone());
2224                } else {
2225                    tracing::warn!(
2226                        "actor child {}: no credential found for provider '{}'",
2227                        job.child_session_id,
2228                        provider
2229                    );
2230                }
2231            }
2232        }
2233        // Tool exposure is baked at provision time. Every actual nested create
2234        // is also checked against the canonical Host's current depth policy.
2235        spec.capabilities.nested_spawn = session.spawn_depth < max_spawn_depth;
2236        spec.capabilities.max_spawn_depth = Some(max_spawn_depth);
2237        // #69: activate child-approval review. Sub-agents enforce permissions so
2238        // their DANGEROUS actions (the worker uses a HIGH threshold) reach the
2239        // parent for review — escalated to the human, or model-reviewed off-loop
2240        // when the parent is in bypass. The worker installs no checker without
2241        // this, so the whole review chain would otherwise stay dormant.
2242        spec.capabilities.enforce_permissions = true;
2243        // Propagate "bypass permissions" so a self-orchestrating worker knows it
2244        // is a bypassed parent and installs the off-loop model-reviewer for its
2245        // children's forced-ask actions (Phase 6, Part B). The child session
2246        // already carries the inherited flag (create_child_action seeds it).
2247        let requested_permission_mode = session
2248            .agent_runtime_state
2249            .as_ref()
2250            .map(|state| state.effective_permission_mode())
2251            .unwrap_or_default();
2252        let configured_permission_mode = self
2253            .permission_config
2254            .as_ref()
2255            .map(|config| config.mode())
2256            .unwrap_or_default();
2257        // #73: propagate "no interactive human approver" (headless / scheduled /
2258        // deployed root, inherited by the child session). When set, the worker's
2259        // per-run approval proxy model-reviews a gated action locally instead of
2260        // escalating to a human who will never answer (which would 300s-deny).
2261        spec.capabilities.no_human_approver = session
2262            .agent_runtime_state
2263            .as_ref()
2264            .is_some_and(|s| s.no_human_approver);
2265        // Read-only authority comes from the child's typed runtime state, never
2266        // its cosmetic role label. The sole legacy recovery path additionally
2267        // requires the old Guardian role AND its host-owned mutation denylist,
2268        // preserving in-flight pre-upgrade reviewers without making a label an
2269        // authority boundary.
2270        let typed_read_only = session
2271            .agent_runtime_state
2272            .as_ref()
2273            .is_some_and(|state| state.read_only);
2274        let legacy_guardian_read_only = spec.identity.role == "guardian"
2275            && spec.disabled_tools.as_ref().is_some_and(|disabled| {
2276                ["Edit", "Write", "SubAgent", "WebFetch"]
2277                    .iter()
2278                    .all(|name| disabled.iter().any(|disabled_name| disabled_name == *name))
2279            });
2280        let read_only = typed_read_only || legacy_guardian_read_only;
2281        spec.capabilities.read_only = read_only;
2282        // Emit the old field for every read-only child. Older workers ignore
2283        // `read_only` but understand this bit and will install the strict
2284        // no-shell checker, which prevents a rolling upgrade from weakening Plan.
2285        spec.capabilities.guardian_read_only = read_only;
2286        if spec.capabilities.read_only_enforced() {
2287            // Do not even construct nested delegation machinery for a read-only
2288            // inspection child; the denylist below remains a second boundary.
2289            spec.capabilities.nested_spawn = false;
2290            // Treat the capability as the authority boundary, not as a hint
2291            // that every caller must remember to pair with a schema filter.
2292            // Preserve caller-specific disables while always removing Bamboo's
2293            // shell, mutating, persistent, interactive, and delegation surfaces.
2294            spec.disabled_tools
2295                .get_or_insert_with(Default::default)
2296                .extend(crate::runtime::guardian_state::read_only_child_disabled_tools());
2297            match &mut spec.executor {
2298                ExecutorSpec::ClaudeCode {
2299                    permission_mode,
2300                    inherit_user_config,
2301                    ..
2302                } => {
2303                    *permission_mode = Some("plan".to_string());
2304                    // User config may add MCP tools/hooks outside the provisioned
2305                    // surface, so it is never inherited by a read-only child.
2306                    *inherit_user_config = Some(false);
2307                }
2308                ExecutorSpec::Codex {
2309                    mode,
2310                    sandbox,
2311                    inherit_user_config,
2312                    approval_policy,
2313                    network_access,
2314                    allow_danger_bypass,
2315                    permission_profile,
2316                    ..
2317                } => {
2318                    *sandbox = Some("read-only".to_string());
2319                    *inherit_user_config = Some(false);
2320                    // app-server provisioning accepts `on-request`; the typed
2321                    // Plan activation maps that to `approvalPolicy=never`.
2322                    *approval_policy = Some(
2323                        if mode.as_deref() == Some("app_server") {
2324                            "on-request"
2325                        } else {
2326                            "never"
2327                        }
2328                        .to_string(),
2329                    );
2330                    *network_access = Some(false);
2331                    *allow_danger_bypass = Some(false);
2332                    *permission_profile = Some("read-only".to_string());
2333                }
2334                _ => {}
2335            }
2336        }
2337        let read_only_overlay = session
2338            .agent_runtime_state
2339            .as_ref()
2340            .is_some_and(|state| state.plan_mode.is_some())
2341            || spec.capabilities.read_only_enforced()
2342            || executor_has_read_only_permission_profile(&spec.executor);
2343        let permission_resolution = bamboo_domain::resolve_permission_mode_with_read_only(
2344            requested_permission_mode,
2345            configured_permission_mode,
2346            read_only_overlay,
2347        );
2348        spec.capabilities.bypass = permission_resolution.bypass_permissions();
2349        spec.capabilities.auto_approve_permissions =
2350            permission_resolution.suppress_approval_prompts();
2351        spec.capabilities.permission_requested_mode =
2352            permission_resolution.requested.as_str().to_string();
2353        spec.capabilities.permission_effective_mode =
2354            permission_resolution.effective.as_str().to_string();
2355        // #193: route this role to a REMOTE resident worker when one is pinned.
2356        // `spec.identity.role` was just computed from `subagent_type` above; a
2357        // match flips the placement to Remote. The scoped broker credential stays
2358        // on the Host transport and never enters the child RunSpec. No match
2359        // leaves the default `Placement::Local`,
2360        // so the local path is byte-for-byte unchanged for every non-pinned role.
2361        if let Some(placement) = self.remote_placements.get(spec.identity.role.as_str()) {
2362            spec.placement = Placement::Remote {
2363                endpoint: placement.endpoint.clone(),
2364            };
2365        } else if let Some(placement) = self.schedulable_placements.get(spec.identity.role.as_str())
2366        {
2367            // #181 (P2b): route this role to a SCHEDULED worker — ONLY when it is
2368            // NOT already pinned to a fixed remote endpoint (the `else if` makes
2369            // remote_placements take precedence for a role in both). The concrete
2370            // worker is picked at run time in `execute_external_child` from the bus
2371            // (a live connected worker of the pool role). No per-placement bearer
2372            // now — the bus connection uses the bus token. No match in either map
2373            // leaves the default `Placement::Local`.
2374            spec.placement = Placement::Schedulable {
2375                pool: placement.pool.clone(),
2376            };
2377        }
2378        spec
2379    }
2380
2381    /// Assemble from the startup snapshot. Kept synchronous for focused unit
2382    /// tests and non-server callers that do not install a live config source.
2383    fn build_spec(&self, session: &Session, job: &SpawnJob) -> ProvisionSpec {
2384        self.build_spec_with_provider_config(
2385            session,
2386            job,
2387            &self.credentials,
2388            &self.default_provider,
2389            self.max_spawn_depth,
2390        )
2391    }
2392
2393    /// Assemble from the latest AppState config when available. Taking the
2394    /// async read lock here gives every new activation a post-reload credential
2395    /// view without copying secrets into a second independently refreshed
2396    /// runtime store.
2397    async fn build_live_spec(&self, session: &Session, job: &SpawnJob) -> ProvisionSpec {
2398        let Some(config) = self.live_provider_config.as_ref() else {
2399            return self.build_spec(session, job);
2400        };
2401
2402        let config = config.read().await;
2403        let credentials = super::runtime::extract_provider_credentials(&config);
2404        let default_provider = config.effective_default_provider().to_string();
2405        self.build_spec_with_provider_config(
2406            session,
2407            job,
2408            &credentials,
2409            &default_provider,
2410            config.subagents().effective_max_spawn_depth(),
2411        )
2412    }
2413
2414    /// The `metadata["placement"]` JSON to stamp on a child from its resolved
2415    /// placement, preferring the matching cluster node's `host_label` (its
2416    /// operator label/host) over the raw endpoint/pool. `None` for a Local child
2417    /// (the DTO defaults it to the backend's own host). Split out of
2418    /// `execute_external_child` so the role→placement→host resolution is unit-testable.
2419    fn placement_stamp_for(&self, spec: &ProvisionSpec) -> Option<String> {
2420        let host_label = match &spec.placement {
2421            Placement::Remote { .. } => self
2422                .remote_placements
2423                .get(spec.identity.role.as_str())
2424                .and_then(|p| p.host_label.as_deref()),
2425            Placement::Schedulable { .. } => self
2426                .schedulable_placements
2427                .get(spec.identity.role.as_str())
2428                .and_then(|p| p.host_label.as_deref()),
2429            Placement::Local => None,
2430        };
2431        placement_metadata(&spec.placement, host_label)
2432    }
2433
2434    fn scoped_placement_route(
2435        &self,
2436        spec: &ProvisionSpec,
2437    ) -> Result<Option<ScopedPlacementRoute>, AgentError> {
2438        match &spec.placement {
2439            Placement::Local => Ok(None),
2440            Placement::Remote { .. } => {
2441                let placement = self
2442                    .remote_placements
2443                    .get(&spec.identity.role)
2444                    .ok_or_else(remote_broker_unavailable)?;
2445                let peer = placement
2446                    .broker_peer
2447                    .as_ref()
2448                    .and_then(|peer| peer.as_ref().ok())
2449                    .ok_or_else(remote_broker_unavailable)?;
2450                let worker_role = peer
2451                    .worker_role
2452                    .clone()
2453                    .ok_or_else(remote_broker_unavailable)?;
2454                Ok(Some(ScopedPlacementRoute {
2455                    endpoint: placement.endpoint.clone(),
2456                    token: placement
2457                        .token
2458                        .clone()
2459                        .ok_or_else(remote_broker_unavailable)?,
2460                    ca_cert_file: placement
2461                        .ca_cert_file
2462                        .clone()
2463                        .ok_or_else(remote_broker_unavailable)?,
2464                    parent: bamboo_subagent::AgentRef {
2465                        session_id: peer.parent_mailbox.clone(),
2466                        role: peer.parent_role.clone(),
2467                    },
2468                    worker_role,
2469                    pinned_mailbox: Some(peer.worker_mailbox.clone()),
2470                    requirements: placement
2471                        .requirements
2472                        .clone()
2473                        .ok_or_else(remote_broker_unavailable)?,
2474                }))
2475            }
2476            Placement::Schedulable { .. } => {
2477                let placement = self
2478                    .schedulable_placements
2479                    .get(&spec.identity.role)
2480                    .ok_or_else(remote_broker_unavailable)?;
2481                let parent = placement
2482                    .broker_parent
2483                    .as_ref()
2484                    .ok_or_else(remote_broker_unavailable)?;
2485                Ok(Some(ScopedPlacementRoute {
2486                    endpoint: placement.endpoint.clone(),
2487                    token: placement
2488                        .token
2489                        .clone()
2490                        .ok_or_else(remote_broker_unavailable)?,
2491                    ca_cert_file: placement
2492                        .ca_cert_file
2493                        .clone()
2494                        .ok_or_else(remote_broker_unavailable)?,
2495                    parent: bamboo_subagent::AgentRef {
2496                        session_id: parent.parent_mailbox.clone(),
2497                        role: Some(parent.parent_role.clone()),
2498                    },
2499                    worker_role: placement.pool.clone(),
2500                    pinned_mailbox: None,
2501                    requirements: placement
2502                        .requirements
2503                        .clone()
2504                        .ok_or_else(remote_broker_unavailable)?,
2505                }))
2506            }
2507        }
2508    }
2509
2510    async fn select_scoped_worker(
2511        &self,
2512        route: &ScopedPlacementRoute,
2513        registry: &bamboo_storage::v2::FileHostRegistry,
2514        actor_id: &str,
2515        project_id: Option<bamboo_domain::ProjectId>,
2516        required_tools: &BTreeSet<String>,
2517    ) -> Result<String, AgentError> {
2518        if let Some(mailbox) = &route.pinned_mailbox {
2519            return Ok(mailbox.clone());
2520        }
2521        let mut query = route.query().await?;
2522        let candidates = query
2523            .list_connected(&route.worker_role)
2524            .await
2525            .map_err(|_| remote_broker_unavailable())?;
2526        let scheduler = PlacementScheduler::new(registry.clone());
2527        let mut selected: Option<(String, usize, u16)> = None;
2528        for mailbox in candidates {
2529            let Some(observed) = query
2530                .observe_host_capacity(&mailbox, &route.worker_role)
2531                .await
2532                .map_err(|_| remote_broker_unavailable())?
2533            else {
2534                continue;
2535            };
2536            let Ok(host) = scheduler
2537                .observe_scoped_host(
2538                    &mut query,
2539                    &observed.host_ref,
2540                    &mailbox,
2541                    &route.worker_role,
2542                    chrono::Duration::seconds(110),
2543                )
2544                .await
2545            else {
2546                continue;
2547            };
2548            let now = chrono::Utc::now();
2549            let request = HostPlacementRequest {
2550                intent: HostPlacementIntent::Pool {
2551                    pool_id: route.worker_role.clone(),
2552                },
2553                actor_id: actor_id.to_owned(),
2554                run_id: "selection".into(),
2555                project_id: project_id.clone(),
2556                trust_zone: route.requirements.trust_zone.clone(),
2557                workspace_label: route.requirements.workspace_label.clone(),
2558                executor: "bamboo-runtime".into(),
2559                required_tools: required_tools.clone(),
2560                network_zone: route.requirements.network_zone.clone(),
2561                require_network_isolation: route.requirements.require_network_isolation,
2562                preferred_host_ref: None,
2563                now,
2564                lease_expires_at: now + chrono::Duration::seconds(80),
2565            };
2566            if host.capabilities.placement_class != bamboo_domain::ActorPlacementClass::Remote
2567                || !request.eligible(&host)
2568            {
2569                continue;
2570            }
2571            let load = host.active_slots(now);
2572            if selected.as_ref().is_none_or(|(_, old_load, old_slots)| {
2573                load * usize::from(*old_slots) < *old_load * usize::from(host.max_slots)
2574            }) {
2575                selected = Some((mailbox, load, host.max_slots));
2576            }
2577        }
2578        selected
2579            .map(|(mailbox, _, _)| mailbox)
2580            .ok_or_else(remote_broker_unavailable)
2581    }
2582}
2583
2584#[async_trait]
2585impl ExternalChildRunner for ActorChildRunner {
2586    async fn validate_required_child_context_route(&self, session: &Session) -> Result<(), String> {
2587        let role = session
2588            .metadata
2589            .get("subagent_type")
2590            .map(String::as_str)
2591            .unwrap_or("worker");
2592        let current = std::env::current_exe()
2593            .ok()
2594            .and_then(|path| path.canonicalize().ok());
2595        if self.native_tool_ceiling.is_none()
2596            || !self.builtin_required_context_route
2597            || !self.should_handle(session).await
2598            || current.is_none()
2599            || self.worker_bin.canonicalize().ok() != current
2600            || self.worker_args != ["subagent-worker".to_string()]
2601            || !matches!(self.executor, ExecutorSpec::BambooRuntime)
2602            || self.bus.is_none()
2603            || self.remote_placements.contains_key(role)
2604            || self.schedulable_placements.contains_key(role)
2605        {
2606            return Err(
2607                "required_child_context_unsupported: built-in fresh local Bamboo worker required"
2608                    .into(),
2609            );
2610        }
2611        bamboo_subagent::fleet::require_worker_capability(
2612            &self.worker_bin,
2613            &self.worker_args,
2614            bamboo_subagent::provision::REQUIRED_CHILD_CONTEXT_WORKER_CAPABILITY,
2615        )
2616        .await
2617        .map_err(|_| {
2618            "required_child_context_unsupported: worker capability unconfirmed".to_string()
2619        })?;
2620        bamboo_subagent::fleet::require_worker_capability(
2621            &self.worker_bin,
2622            &self.worker_args,
2623            bamboo_subagent::provision::NATIVE_TOOL_CEILING_WORKER_CAPABILITY,
2624        )
2625        .await
2626        .map_err(|_| {
2627            "native_tool_ceiling_unsupported: worker capability unconfirmed".to_string()
2628        })?;
2629        bamboo_subagent::fleet::require_worker_capability(
2630            &self.worker_bin,
2631            &self.worker_args,
2632            bamboo_subagent::provision::CHILD_CREATION_IDENTITY_WORKER_CAPABILITY,
2633        )
2634        .await
2635        .map_err(|_| {
2636            "required_child_context_unsupported: worker birth capability unconfirmed".into()
2637        })
2638    }
2639    async fn should_handle(&self, session: &Session) -> bool {
2640        session.metadata.get("runtime.kind") == Some(&"external".to_string())
2641            && session.metadata.get("external.protocol") == Some(&"actor".to_string())
2642            && session.metadata.get("external.agent_id") == Some(&self.agent_id)
2643    }
2644
2645    fn set_escalation_bridge(&self, bridge: Option<bamboo_subagent::executor::HostBridge>) {
2646        *self.escalation_bridge.lock().recover_poison() = bridge;
2647    }
2648
2649    fn set_session_inbox_runtime(&self, binding: Option<SessionInboxRuntimeBinding>) {
2650        *self.session_inbox_runtime.lock().recover_poison() = binding;
2651    }
2652
2653    fn set_actor_directory_store(&self, store: Option<Arc<bamboo_storage::SessionStoreV2>>) {
2654        *self.actor_directory_store.lock().recover_poison() = store;
2655    }
2656
2657    fn set_canonical_subagent_tool(&self, tool: Option<Arc<dyn bamboo_agent_core::tools::Tool>>) {
2658        *self.canonical_subagent_tool.lock().recover_poison() = tool;
2659    }
2660
2661    fn set_ticket_service(&self, service: Option<Arc<bamboo_tickets::TicketService>>) {
2662        *self.ticket_service.lock().recover_poison() = service;
2663    }
2664
2665    fn set_actor_event_observer(&self, observer: Option<Arc<dyn ActorEventObserver>>) {
2666        *self.actor_event_observer.lock().recover_poison() = observer;
2667    }
2668
2669    async fn prepare_durable_child_delivery(
2670        &self,
2671        session: &Session,
2672        activation_run_id: &str,
2673    ) -> Result<bool, String> {
2674        let key = (
2675            session.id.clone(),
2676            session.created_at,
2677            activation_run_id.to_owned(),
2678        );
2679        let (delivery, host_nested_wait_handoff, strict_history) = {
2680            let pending = self.pending_durable_links.lock().await;
2681            let Some(pending) = pending.get(&key) else {
2682                return Ok(false);
2683            };
2684            if pending.strict_history_required && pending.strict_history.is_none() {
2685                return Err("Host strict terminal completeness proof missing".into());
2686            }
2687            (
2688                pending
2689                    .link
2690                    .durable_delivery_receipt()
2691                    .ok_or_else(|| "broker terminal has no exact durable receipt".to_string())?,
2692                pending.host_nested_wait_handoff,
2693                pending.strict_history.clone(),
2694            )
2695        };
2696        let expected_status =
2697            effective_broker_terminal_status(delivery.terminal_status, host_nested_wait_handoff)?;
2698        if session.last_run_status().as_deref() != Some(expected_status) {
2699            tracing::warn!(
2700                child_session_id = %session.id,
2701                activation_run_id,
2702                expected_status,
2703                actual_status = ?session.last_run_status(),
2704                "Host final status differs from accepted broker terminal"
2705            );
2706            return Err("Host final status differs from accepted broker terminal".into());
2707        }
2708        let store = self
2709            .actor_directory_store
2710            .lock()
2711            .recover_poison()
2712            .clone()
2713            .ok_or_else(|| "Host broker receipt store unavailable".to_string())?;
2714        let binding = self
2715            .session_inbox_runtime
2716            .lock()
2717            .recover_poison()
2718            .clone()
2719            .ok_or_else(|| "Host activation fence unavailable for broker receipt".to_string())?;
2720        if !binding
2721            .router
2722            .owns_run(&session.id, activation_run_id)
2723            .await
2724        {
2725            return Err("stale Child activation cannot prepare broker receipt".into());
2726        }
2727        let ticket_service = self.ticket_service.lock().recover_poison().clone();
2728        let yielded_call = if expected_status == "completed"
2729            && session
2730                .metadata
2731                .contains_key(crate::ticket_worker_plan::TICKET_LOCAL_PLAN_KEY)
2732        {
2733            let service = ticket_service
2734                .as_deref()
2735                .ok_or("Ticket authority unavailable for broker question")?;
2736            crate::ticket_runtime::question_terminal_call_id(service, session, activation_run_id)
2737                .map_err(|e| e.to_string())?
2738        } else {
2739            None
2740        };
2741        let route = bamboo_storage::v2::BrokerTerminalRoute {
2742            activation_run_id,
2743            broker_identity: &delivery.broker_identity,
2744            parent_mailbox: &delivery.parent_mailbox,
2745            broker_correlation_id: &delivery.correlation_id,
2746            message_ids: &delivery.message_ids,
2747        };
2748        if let Some(call_id) = yielded_call {
2749            let proof = bamboo_storage::v2::HostToolYield::from_verified_host(
2750                session,
2751                activation_run_id,
2752                &call_id,
2753            )
2754            .map_err(|error| error.to_string())?;
2755            if let Some(history) = strict_history.as_ref() {
2756                store
2757                    .prepare_broker_tool_yield_receipt_with_completeness(
2758                        session,
2759                        route,
2760                        &proof,
2761                        &history.proof,
2762                    )
2763                    .await
2764            } else {
2765                store
2766                    .prepare_broker_tool_yield_receipt(session, route, &proof)
2767                    .await
2768            }
2769            .map_err(|error| format!("Host broker receipt prepare failed: {error}"))?;
2770        } else {
2771            if let Some(history) = strict_history.as_ref() {
2772                store
2773                    .prepare_broker_terminal_receipt_with_completeness(
2774                        session,
2775                        route,
2776                        &history.proof,
2777                    )
2778                    .await
2779            } else {
2780                store
2781                    .prepare_broker_terminal_receipt(
2782                        session,
2783                        activation_run_id,
2784                        &delivery.broker_identity,
2785                        &delivery.parent_mailbox,
2786                        &delivery.correlation_id,
2787                        &delivery.message_ids,
2788                    )
2789                    .await
2790            }
2791            .map_err(|error| format!("Host broker receipt prepare failed: {error}"))?;
2792        }
2793        Ok(true)
2794    }
2795
2796    async fn commit_durable_child_delivery(
2797        &self,
2798        session: &Session,
2799        activation_run_id: &str,
2800    ) -> Result<(), String> {
2801        let key = (
2802            session.id.clone(),
2803            session.created_at,
2804            activation_run_id.to_owned(),
2805        );
2806        let mut pending_links = self.pending_durable_links.lock().await;
2807        let Some(pending) = pending_links.get_mut(&key) else {
2808            return Ok(());
2809        };
2810        let delivery = pending
2811            .link
2812            .durable_delivery_receipt()
2813            .ok_or_else(|| "broker terminal lost its exact durable receipt".to_string())?;
2814        let expected_status = effective_broker_terminal_status(
2815            delivery.terminal_status,
2816            pending.host_nested_wait_handoff,
2817        )?;
2818        if session.last_run_status().as_deref() != Some(expected_status) {
2819            return Err("Host final status differs from accepted broker terminal".into());
2820        }
2821        if pending.strict_history_required && pending.strict_history.is_none() {
2822            return Err("Host strict terminal completeness proof missing".into());
2823        }
2824        let store = self
2825            .actor_directory_store
2826            .lock()
2827            .recover_poison()
2828            .clone()
2829            .ok_or_else(|| "Host broker receipt store unavailable".to_string())?;
2830        let committed = if let Some(history) = pending.strict_history.as_ref() {
2831            store
2832                .commit_broker_terminal_receipt_with_completeness(
2833                    session,
2834                    activation_run_id,
2835                    &history.proof,
2836                )
2837                .await
2838        } else {
2839            store
2840                .commit_broker_terminal_receipt(session, activation_run_id)
2841                .await
2842        }
2843        .map_err(|error| format!("Host broker receipt checkpoint unconfirmed: {error}"))?;
2844        if committed.broker_identity != delivery.broker_identity
2845            || committed.parent_mailbox != delivery.parent_mailbox
2846            || committed.broker_correlation_id != delivery.correlation_id
2847            || committed.message_ids != delivery.message_ids
2848            || committed.terminal_status != expected_status
2849        {
2850            return Err("Host broker receipt changed before completion".into());
2851        }
2852        if let Some(history) = pending.strict_history.as_ref() {
2853            let proof = committed
2854                .terminal_completeness
2855                .as_ref()
2856                .ok_or_else(|| "Host strict terminal completeness readback missing".to_string())?;
2857            let creation = bamboo_subagent::proto::ChildCreationIdentity {
2858                created_at: session.created_at,
2859                spawn_depth: session.spawn_depth,
2860            };
2861            validate_strict_history_terminal(
2862                Some(&history.watermark),
2863                session,
2864                session
2865                    .parent_session_id
2866                    .as_deref()
2867                    .ok_or("Child parent missing")?,
2868                Some(activation_run_id),
2869                proof.execution_epoch,
2870                Some(&creation),
2871                proof.contiguous_applied_seq,
2872            )
2873            .map_err(|error| error.to_string())?;
2874        }
2875        // The same exact-snapshot CAS serves live and cold publication. An
2876        // append-safe generic save could expose Completed after merging new
2877        // input that invalidates the source; CAS rejects before that write.
2878        publish_recovered_completion_source(&store, &committed).await?;
2879        pending.strict_history_committed = true;
2880        Ok(())
2881    }
2882
2883    async fn confirm_durable_child_delivery(
2884        &self,
2885        session: &Session,
2886        activation_run_id: &str,
2887        save_succeeded: bool,
2888    ) -> Result<(), String> {
2889        if save_succeeded
2890            && session
2891                .metadata
2892                .contains_key(crate::ticket_worker_plan::TICKET_LOCAL_PLAN_KEY)
2893        {
2894            let service = self
2895                .ticket_service
2896                .lock()
2897                .recover_poison()
2898                .clone()
2899                .ok_or_else(|| {
2900                    "Ticket authority unavailable for terminal checkpoint".to_string()
2901                })?;
2902            let runtime = self
2903                .session_inbox_runtime
2904                .lock()
2905                .recover_poison()
2906                .clone()
2907                .ok_or_else(|| "Ticket runtime unavailable for terminal checkpoint".to_string())?;
2908            crate::ticket_runtime::checkpoint_ticket_result(
2909                &service,
2910                runtime.storage.as_ref(),
2911                session,
2912                activation_run_id,
2913            )
2914            .await
2915            .map_err(|e| e.to_string())?;
2916        }
2917        let key = (
2918            session.id.clone(),
2919            session.created_at,
2920            activation_run_id.to_owned(),
2921        );
2922        let Some(mut pending) = self.pending_durable_links.lock().await.remove(&key) else {
2923            return Ok(());
2924        };
2925        // A later SDK timeout or Host postprocessing error cannot convert an
2926        // accepted Worker terminal into a different receipt status.
2927        let delivery = pending
2928            .link
2929            .durable_delivery_receipt()
2930            .ok_or_else(|| "broker terminal lost its exact durable receipt".to_string())?;
2931        let expected_status = effective_broker_terminal_status(
2932            delivery.terminal_status,
2933            pending.host_nested_wait_handoff,
2934        )?;
2935        if !save_succeeded || session.last_run_status().as_deref() != Some(expected_status) {
2936            return Ok(());
2937        }
2938        let store = self
2939            .actor_directory_store
2940            .lock()
2941            .recover_poison()
2942            .clone()
2943            .ok_or_else(|| "Host broker receipt store unavailable".to_string())?;
2944        let committed = if pending.strict_history_required {
2945            if !pending.strict_history_committed {
2946                return Err("Host strict completeness was not committed before ACK".into());
2947            }
2948            let history = pending
2949                .strict_history
2950                .as_ref()
2951                .ok_or_else(|| "Host strict terminal completeness proof missing".to_string())?;
2952            store
2953                .commit_broker_terminal_receipt_with_completeness(
2954                    session,
2955                    activation_run_id,
2956                    &history.proof,
2957                )
2958                .await
2959        } else {
2960            store
2961                .commit_broker_terminal_receipt(session, activation_run_id)
2962                .await
2963        }
2964        .map_err(|error| format!("Host broker receipt checkpoint unconfirmed: {error}"))?;
2965        if committed.broker_identity != delivery.broker_identity
2966            || committed.parent_mailbox != delivery.parent_mailbox
2967            || committed.broker_correlation_id != delivery.correlation_id
2968            || committed.message_ids != delivery.message_ids
2969            || committed.terminal_status != expected_status
2970        {
2971            return Err("Host broker receipt changed before ACK".into());
2972        }
2973        pending
2974            .link
2975            .acknowledge_durable_frames()
2976            .await
2977            .map_err(|error| format!("child broker durable ACK unconfirmed: {error}"))?;
2978        store
2979            .clear_acknowledged_broker_terminal_receipt(
2980                &session.id,
2981                session.created_at,
2982                activation_run_id,
2983                &delivery.parent_mailbox,
2984            )
2985            .await
2986            .map_err(|error| {
2987                format!("broker ACK succeeded but Host receipt cleanup failed: {error}")
2988            })
2989    }
2990
2991    #[tracing::instrument(
2992        name = "actor_child_run",
2993        level = "warn",
2994        skip_all,
2995        fields(child_session_id = %job.child_session_id, activation_run_id = tracing::field::Empty)
2996    )]
2997    async fn execute_external_child(
2998        &self,
2999        session: &mut Session,
3000        job: &SpawnJob,
3001        event_tx: mpsc::Sender<AgentEvent>,
3002        cancel_token: CancellationToken,
3003    ) -> crate::runtime::runner::Result<()> {
3004        crate::session_app::child_session::named_profile::validate_named_profile(session)
3005            .map_err(|error| AgentError::LLM(error.to_string()))?;
3006        // #68 CORRECTNESS CRUX: capture the per-run escalation bridge HERE, at the
3007        // moment this grandchild is spawned — while the parent run's bridge is
3008        // still in our slot — into an owned local handed to `drive()` for this
3009        // grandchild's whole lifetime. A fire-and-forget grandchild that OUTLIVES
3010        // the run that spawned it must NOT re-read `self.escalation_bridge` at
3011        // approval time: by then `run()` may have cleared/overwritten it (a worker
3012        // serves runs sequentially), and re-proxying through a closed bridge
3013        // fail-closed denies. Capturing at spawn pins the right bridge per run.
3014        let escalation = self.escalation_bridge.lock().recover_poison().clone();
3015        let session_inbox_runtime = self.session_inbox_runtime.lock().recover_poison().clone();
3016        let actor_directory_store = self.actor_directory_store.lock().recover_poison().clone();
3017        let canonical_subagent_tool = self.canonical_subagent_tool.lock().recover_poison().clone();
3018        let ticket_service = self.ticket_service.lock().recover_poison().clone();
3019        let actor_event_observer = self.actor_event_observer.lock().recover_poison().clone();
3020        let required_context = bamboo_domain::ChildContextBinding::from_session(session)
3021            .map_err(|error| AgentError::Budget(error.to_string()))?;
3022        if crate::session_app::child_session::named_profile::has_named_profile(session)
3023            && required_context.is_none()
3024        {
3025            return Err(AgentError::LLM(
3026                "named_profile_requires_strict_context_route".into(),
3027            ));
3028        }
3029        let assignment = required_context
3030            .as_ref()
3031            .map(|binding| binding.payload.required_assignment.clone())
3032            .unwrap_or_else(|| extract_assignment(session));
3033        let mut spec = self.build_live_spec(session, job).await;
3034        // Reject before remote routing, Actor claims, provisioning or Ticket
3035        // Run admission: resident workers cannot produce local owned stop proof.
3036        ensure_ticket_placement_is_enforceable(session, &spec)?;
3037        let scoped_route = self.scoped_placement_route(&spec)?;
3038        let strict_remote = scoped_route.as_ref();
3039        if strict_remote.is_some() && !matches!(spec.executor, ExecutorSpec::BambooRuntime) {
3040            return Err(remote_broker_unavailable());
3041        }
3042        let lineage = if strict_remote.is_some() {
3043            Some(
3044                remote_canonical_lineage(
3045                    session_inbox_runtime
3046                        .as_ref()
3047                        .ok_or_else(remote_broker_unavailable)?,
3048                    session,
3049                )
3050                .await?,
3051            )
3052        } else {
3053            None
3054        };
3055        let creation = if (matches!(spec.placement, Placement::Local) || strict_remote.is_some())
3056            && matches!(spec.executor, ExecutorSpec::BambooRuntime)
3057        {
3058            let binding = session_inbox_runtime.as_ref().ok_or_else(|| {
3059                AgentError::LLM(
3060                    "child creation identity unavailable: canonical host Storage required".into(),
3061                )
3062            })?;
3063            spec.capabilities.child_creation_identity = true;
3064            Some(canonical_child_creation(binding, session, job).await?)
3065        } else {
3066            None
3067        };
3068        // Mark the worker reusable + give it an idle timeout so it self-reaps if
3069        // orphaned. Warm bus workers are pooled per fingerprint and reused.
3070        spec.reusable = required_context.is_none();
3071        spec.capabilities.required_child_context = required_context.is_some();
3072        if required_context.is_some() {
3073            self.validate_required_child_context_route(session)
3074                .await
3075                .map_err(AgentError::LLM)?;
3076        }
3077        if required_context.is_some() {
3078            let names = self
3079                .native_tool_ceiling
3080                .as_ref()
3081                .unwrap()
3082                .observe(session)
3083                .await
3084                .map_err(AgentError::LLM)?;
3085            let birth = creation
3086                .as_ref()
3087                .ok_or_else(|| AgentError::LLM("native_tool_ceiling_birth_missing".into()))?;
3088            spec.capabilities.native_tool_ceiling_required = true;
3089            spec.capabilities.native_tool_ceiling =
3090                Some(bamboo_subagent::proto::NativeToolCeiling {
3091                    version: 1,
3092                    child_session_id: session.id.clone(),
3093                    parent_session_id: job.parent_session_id.clone(),
3094                    root_session_id: session.root_session_id.clone(),
3095                    created_at: birth.created_at,
3096                    spawn_depth: birth.spawn_depth,
3097                    project_id: project_id_for_actor_run(session)?,
3098                    tools: names,
3099                });
3100            // Positive native names are a leaf surface, never an expansion path.
3101            spec.capabilities.mcp = None;
3102            spec.capabilities.mcp_proxy = None;
3103            spec.capabilities.skills_dir = None;
3104            spec.capabilities.nested_spawn = false;
3105            spec.validate()
3106                .map_err(|_| AgentError::LLM("native_tool_ceiling_invalid".into()))?;
3107        }
3108        // Eligibility uses the actual Host callable ceiling, after the strict
3109        // built-in route and birth capability checks. A role label is not a grant.
3110        let first_reply_store = actor_directory_store.clone().filter(|_| {
3111            required_context.is_some()
3112                && crate::session_app::child_session::named_profile::has_named_profile(session)
3113                && matches!(spec.placement, Placement::Local)
3114                && matches!(spec.executor, ExecutorSpec::BambooRuntime)
3115                && spec.capabilities.native_tool_ceiling_required
3116                && spec
3117                    .capabilities
3118                    .native_tool_ceiling
3119                    .as_ref()
3120                    .is_some_and(|ceiling| {
3121                        ceiling.tools.is_empty()
3122                            || (spec.capabilities.read_only_enforced() && ceiling.tools == ["Glob"])
3123                    })
3124        });
3125        let plain_actor_store = if let Some(store) = first_reply_store {
3126            let root = bamboo_agent_core::storage::Storage::load_session(
3127                store.as_ref(),
3128                &session.root_session_id,
3129            )
3130            .await
3131            .map_err(|_| plain_actor_unsupported())?
3132            .ok_or_else(|| plain_actor_unsupported())?;
3133            if root.id != session.root_session_id || root.kind != bamboo_domain::SessionKind::Root {
3134                return Err(plain_actor_unsupported());
3135            }
3136            if root.root_orchestration_only_enabled() {
3137                if !plain_initial_history(session)
3138                    && (spec
3139                        .capabilities
3140                        .native_tool_ceiling
3141                        .as_ref()
3142                        .is_none_or(|ceiling| !ceiling.tools.is_empty())
3143                        || (!plain_completed_history(session) && pre_ack_input(session)?.is_none()))
3144                {
3145                    return Err(plain_actor_unsupported());
3146                }
3147                Some(store)
3148            } else if spec
3149                .capabilities
3150                .native_tool_ceiling
3151                .as_ref()
3152                .is_some_and(|ceiling| ceiling.tools == ["Glob"])
3153            {
3154                // Ordinary Glob remains legacy. Observe only its own mapped
3155                // initialization signals; never require a whole-tree snapshot.
3156                require_unowned_glob(store.as_ref(), session).await?;
3157                None
3158            } else {
3159                // Public bounded observation is only a no-replay check. It is
3160                // never an activation grant. Ordinary first runs stay legacy.
3161                let observed = store
3162                    .actor_subtree_snapshot(
3163                        ActorSnapshotPrincipal::host_owner(),
3164                        &session.root_session_id,
3165                        &session.id,
3166                        ActorSnapshotLimits::default(),
3167                    )
3168                    .await
3169                    .map_err(|_| plain_actor_unsupported())?;
3170                if observed
3171                    .nodes
3172                    .iter()
3173                    .any(|node| node.actor_id == session.id && node.activation.is_some())
3174                {
3175                    return Err(plain_actor_unsupported());
3176                }
3177                None
3178            }
3179        } else {
3180            None
3181        };
3182        let readonly_actor = plain_actor_store.is_some()
3183            && spec
3184                .capabilities
3185                .native_tool_ceiling
3186                .as_ref()
3187                .is_some_and(|ceiling| ceiling.tools == ["Glob"]);
3188        let local_history_tools = (plain_actor_store.is_none()
3189            && required_context.is_some()
3190            && (crate::session_app::child_session::named_profile::has_named_profile(session)
3191                || session
3192                    .metadata
3193                    .contains_key(crate::ticket_worker_plan::TICKET_LOCAL_PLAN_KEY))
3194            && matches!(spec.placement, Placement::Local)
3195            && matches!(spec.executor, ExecutorSpec::BambooRuntime))
3196        .then_some(spec.capabilities.native_tool_ceiling.as_ref())
3197        .flatten()
3198        .filter(|ceiling| {
3199            bamboo_subagent::proto::LocalToolMessages::supports_tools(
3200                &ceiling.tools,
3201                spec.capabilities.read_only_enforced(),
3202            )
3203        })
3204        .map(|ceiling| ceiling.tools.clone());
3205        if plain_actor_store.is_some() && !readonly_actor {
3206            self.require_initial_input_release_worker().await?;
3207            spec.capabilities.initial_input_release_required = true;
3208            spec.validate().map_err(|_| plain_actor_unsupported())?;
3209        }
3210        let mut readonly_messages = Vec::new();
3211        let mut plain_activation = None;
3212        if spec.limits.idle_timeout_secs.is_none() {
3213            spec.limits.idle_timeout_secs = Some(POOLED_IDLE_TIMEOUT_SECS);
3214        }
3215        ensure_read_only_activation_is_enforceable(&spec)?;
3216        let pool_key = Self::fingerprint(&spec);
3217
3218        // The recommended provider URL is deliberately parent-loopback. A
3219        // resident remote worker would interpret 127.0.0.1 as itself, not this
3220        // server, so reject that ambiguous deployment instead of minting a
3221        // credential that can never authenticate to the intended parent.
3222        if executor_uses_bamboo_codex(&spec.executor) && !matches!(spec.placement, Placement::Local)
3223        {
3224            return Err(AgentError::LLM(
3225                "Codex auth mode 'bamboo' requires local actor placement; use custom mode with a reachable URL for remote workers"
3226                    .to_string(),
3227            ));
3228        }
3229        let project_id = project_id_for_actor_run(session)?;
3230        let requested_permission_mode = session
3231            .agent_runtime_state
3232            .as_ref()
3233            .map(|state| state.effective_permission_mode())
3234            .unwrap_or_default();
3235        // Policy is captured per activation (not only when a worker is
3236        // provisioned), so reused local workers and resident remote/broker
3237        // workers observe the latest durable revision and bypass flag at the
3238        // next run boundary. Session grants are intentionally not inherited.
3239        let permission_policy = if let Some(config) = self.permission_config.as_ref() {
3240            ensure_no_active_scoped_session_denies(config, &session.id)?;
3241            let read_only_overlay = session
3242                .agent_runtime_state
3243                .as_ref()
3244                .is_some_and(|state| state.plan_mode.is_some())
3245                || spec.capabilities.read_only_enforced()
3246                || executor_has_read_only_permission_profile(&spec.executor);
3247            let resolution = bamboo_domain::resolve_permission_mode_with_read_only(
3248                requested_permission_mode,
3249                config.mode(),
3250                read_only_overlay,
3251            );
3252            let policy = serde_json::to_value(config.to_serializable()).map_err(|error| {
3253                AgentError::LLM(format!(
3254                    "failed to serialize permission policy for external executor: {error}"
3255                ))
3256            })?;
3257            Some(PermissionPolicyContext {
3258                revision: config.policy_revision(),
3259                requested_mode: resolution.requested.as_str().to_string(),
3260                effective_mode: resolution.effective.as_str().to_string(),
3261                bypass_permissions: resolution.bypass_permissions(),
3262                auto_approve_permissions: resolution.suppress_approval_prompts(),
3263                session_id: session.id.clone(),
3264                workspace_path: session.workspace.clone(),
3265                environment_lease: None,
3266                inherit_session_grants: false,
3267                policy,
3268            })
3269        } else {
3270            None
3271        };
3272        let provisioned_permission =
3273            spec.capabilities.permission_resolution().map_err(|error| {
3274                AgentError::LLM(format!("invalid provisioned permission posture: {error}"))
3275            })?;
3276        let policy_resolution = permission_policy
3277            .as_ref()
3278            .map(|context| {
3279                context.resolved_modes().map(|(requested, effective)| {
3280                    bamboo_domain::PermissionModeResolution {
3281                        requested,
3282                        effective,
3283                    }
3284                })
3285            })
3286            .transpose()
3287            .map_err(|error| {
3288                AgentError::LLM(format!("invalid host permission posture: {error}"))
3289            })?;
3290        let host_audit = bamboo_domain::PermissionAuditSnapshot::from_metadata(&session.metadata);
3291        let has_explicit_deny = self.permission_config.as_ref().is_some_and(|config| {
3292            bamboo_tools::permission::explicit_deny_policy_reason(&config.to_serializable())
3293                .is_some()
3294        });
3295        let expected_executor_mapping = expected_permission_executor_mapping(
3296            &spec.executor,
3297            policy_resolution.unwrap_or(provisioned_permission),
3298            has_explicit_deny,
3299        )?;
3300        let mut expected_permission_posture =
3301            expected_executor_mapping.map(|executor_mapping| ExpectedPermissionPosture {
3302                policy_revision: permission_policy
3303                    .as_ref()
3304                    .map(|context| context.revision)
3305                    .or_else(|| host_audit.as_ref().map(|audit| audit.policy_revision))
3306                    .unwrap_or_default(),
3307                resolution: policy_resolution.unwrap_or(provisioned_permission),
3308                expected_audit_revision: host_audit.as_ref().map(|audit| audit.audit_revision),
3309                executor_mapping,
3310            });
3311
3312        // Backpressure: hold a concurrency slot for the lifetime of the *run*
3313        // (cancellation still proceeds — the cancel branch in drive() runs while
3314        // we hold the permit). Released when this fn returns, i.e. once the worker
3315        // is parked back into the pool, so idle workers don't pin slots.
3316        let _slot = if strict_remote.is_some() {
3317            tokio::select! { biased;
3318                _ = cancel_token.cancelled() => return Err(AgentError::Cancelled),
3319                permit = self.concurrency.acquire() => permit.map_err(|_| remote_broker_unavailable())?,
3320            }
3321        } else {
3322            self.concurrency
3323                .acquire()
3324                .await
3325                .map_err(|_| AgentError::LLM("actor concurrency limiter closed".into()))?
3326        };
3327        let _remote_subscription = if strict_remote.is_some() {
3328            Some(tokio::select! { biased;
3329                _ = cancel_token.cancelled() => return Err(AgentError::Cancelled),
3330                guard = self.strict_remote_subscription.lock() => guard,
3331            })
3332        } else {
3333            None
3334        };
3335
3336        // Bamboo-as-provider credentials are minted at the activation boundary,
3337        // after backpressure admits the run and never at worker provisioning.
3338        // This is load-bearing for warm workers: a parked process must never
3339        // retain a token from its previous run. The guard revokes on every return
3340        // path (success, error, cancellation, dispatch failure, or first-frame
3341        // retry exhaustion).
3342        let (run_secrets, _codex_token_guard) = build_codex_run_secrets(
3343            &spec.executor,
3344            self.codex_run_tokens.clone(),
3345            &job.child_session_id,
3346        )?;
3347
3348        // Split LOCAL (spawn + warm-pool) from the two process-less remote paths
3349        // ONLY at the divergent spots — acquire/connect here and the park/retire at
3350        // the end. Everything between (Run dispatch, live-actor registration,
3351        // drive, the close) is identical for all three. `kind` is the single guard.
3352        //   - Local       (#0):  byte-for-byte the pre-#193 reuse-or-spawn path.
3353        //   - Remote       (#1311): connect to a scoped broker peer, no spawn.
3354        //   - Schedulable  (#181): resolve a live worker from the registry, connect.
3355        let kind = match spec.placement {
3356            Placement::Remote { .. } => PlacementKind::Remote,
3357            Placement::Schedulable { .. } => PlacementKind::Schedulable,
3358            Placement::Local => PlacementKind::Local,
3359        };
3360        let remote = !matches!(kind, PlacementKind::Local);
3361        let scoped_registry = if remote {
3362            let store = actor_directory_store
3363                .as_ref()
3364                .ok_or_else(remote_broker_unavailable)?;
3365            Some(
3366                bamboo_storage::v2::FileHostRegistry::new(store.bamboo_home_dir().to_path_buf())
3367                    .await
3368                    .map_err(|_| remote_broker_unavailable())?,
3369            )
3370        } else {
3371            None
3372        };
3373        let scoped_tools = scoped_route.as_ref().map(|route| {
3374            let mut tools = route.requirements.required_tools.clone();
3375            if let Some(ceiling) = &spec.capabilities.native_tool_ceiling {
3376                tools.extend(ceiling.tools.iter().cloned());
3377            }
3378            tools
3379        });
3380
3381        // Stamp WHICH machine this child runs on onto its session metadata, so the
3382        // UI can show it (mirrored into the session index → SessionSummary.placement).
3383        // Only remote/scheduled placements need a stamp — a Local child falls through
3384        // to the DTO default (this backend's own host). Persisted by the caller with
3385        // the rest of the child session after we return.
3386        if let Some(placement_meta) = self.placement_stamp_for(&spec) {
3387            session
3388                .metadata
3389                .insert("placement".to_string(), placement_meta);
3390        }
3391
3392        // Retry-once loop: a pooled local worker can die between its liveness
3393        // check and handling the Run (a tiny TOCTOU window) — its Run then sits
3394        // queued with no server. The first-frame watchdog in `drive` surfaces that
3395        // as `WorkerUnresponsive`; we reap the dead worker and re-acquire ONCE
3396        // (which spawns fresh / reuses the next live one). Remote/schedulable have
3397        // no spawn fallback, so they never retry.
3398        let mut attempt = 0u8;
3399        let (result, actor) = loop {
3400            let execution_epoch = self.next_execution_epoch.fetch_add(1, Ordering::Relaxed) + 1;
3401            let (actor, mut client) = match kind {
3402                PlacementKind::Remote => {
3403                    let route = scoped_route
3404                        .as_ref()
3405                        .ok_or_else(remote_broker_unavailable)?;
3406                    let mailbox_id = route
3407                        .pinned_mailbox
3408                        .clone()
3409                        .ok_or_else(remote_broker_unavailable)?;
3410                    let tls = bamboo_broker::client_config_trusting_cert(&route.ca_cert_file)
3411                        .map_err(|_| remote_broker_unavailable())?;
3412                    let connect =
3413                        bamboo_broker::BrokerChildLink::connect_strict_with_tls_environment_lease(
3414                            &route.endpoint,
3415                            route.parent.clone(),
3416                            &route.token,
3417                            bamboo_subagent::AgentRef {
3418                                session_id: mailbox_id.clone(),
3419                                role: Some(route.worker_role.clone()),
3420                            },
3421                            tls,
3422                        );
3423                    let link = tokio::select! { biased;
3424                        _ = cancel_token.cancelled() => return Err(AgentError::Cancelled),
3425                        result = tokio::time::timeout(Duration::from_secs(30), connect) =>
3426                            result.map_err(|_| remote_broker_unavailable())?.map_err(|_| remote_broker_unavailable())?,
3427                    };
3428                    let record = AgentRecord {
3429                        agent_id: job.child_session_id.clone(),
3430                        role: spec.identity.role.clone(),
3431                        labels: Vec::new(),
3432                        endpoint: String::new(),
3433                        pid: 0,
3434                        version: String::new(),
3435                        started_at: chrono::Utc::now(),
3436                        lease_expires_at: chrono::Utc::now(),
3437                    };
3438                    (
3439                        PooledWorker {
3440                            worker: SpawnedChild::remote(record),
3441                            mailbox_id,
3442                            parked_at: None,
3443                        },
3444                        Box::new(link) as Box<dyn bamboo_subagent::ChildLink>,
3445                    )
3446                }
3447                PlacementKind::Schedulable => {
3448                    let route = scoped_route
3449                        .as_ref()
3450                        .ok_or_else(remote_broker_unavailable)?;
3451                    let registry = scoped_registry
3452                        .as_ref()
3453                        .ok_or_else(remote_broker_unavailable)?;
3454                    let mailbox_id = self
3455                        .select_scoped_worker(
3456                            route,
3457                            registry,
3458                            &session.id,
3459                            project_id.clone(),
3460                            scoped_tools
3461                                .as_ref()
3462                                .ok_or_else(remote_broker_unavailable)?,
3463                        )
3464                        .await?;
3465                    let tls = bamboo_broker::client_config_trusting_cert(&route.ca_cert_file)
3466                        .map_err(|_| remote_broker_unavailable())?;
3467                    let connect =
3468                        bamboo_broker::BrokerChildLink::connect_strict_with_tls_environment_lease(
3469                            &route.endpoint,
3470                            route.parent.clone(),
3471                            &route.token,
3472                            bamboo_subagent::AgentRef {
3473                                session_id: mailbox_id.clone(),
3474                                role: Some(route.worker_role.clone()),
3475                            },
3476                            tls,
3477                        );
3478                    let link = tokio::select! { biased;
3479                        _ = cancel_token.cancelled() => return Err(AgentError::Cancelled),
3480                        result = tokio::time::timeout(Duration::from_secs(30), connect) =>
3481                            result.map_err(|_| remote_broker_unavailable())?.map_err(|_| remote_broker_unavailable())?,
3482                    };
3483                    let actor = PooledWorker {
3484                        worker: SpawnedChild::remote(AgentRecord {
3485                            agent_id: mailbox_id.clone(),
3486                            role: spec.identity.role.clone(),
3487                            labels: Vec::new(),
3488                            endpoint: route.endpoint.clone(),
3489                            pid: 0,
3490                            version: String::new(),
3491                            started_at: chrono::Utc::now(),
3492                            lease_expires_at: chrono::Utc::now(),
3493                        }),
3494                        mailbox_id,
3495                        parked_at: None,
3496                    };
3497                    let client: Box<dyn bamboo_subagent::ChildLink> = Box::new(link);
3498                    (actor, client)
3499                }
3500                PlacementKind::Local => {
3501                    // LOCAL = the mailbox bus (the unified transport): check out a warm
3502                    // pooled worker (reuse a live parked one, else spawn fresh) and
3503                    // drive it by mailbox id — no listen socket, no file discovery, no
3504                    // respawn-on-connect-miss (the broker queues the Run until the
3505                    // worker handles it). The legacy direct-WS path was retired; the bus
3506                    // is required.
3507                    let bus = self.bus.as_ref().ok_or_else(|| {
3508                        AgentError::LLM(
3509                            "local sub-agents require a mailbox bus (subagents.broker); none is \
3510                         configured and the bus could not be embedded"
3511                                .to_string(),
3512                        )
3513                    })?;
3514                    let actor = self.acquire_bus_worker(&pool_key, &spec).await?;
3515                    let parent = bamboo_subagent::AgentRef {
3516                        session_id: format!("p-{}", job.child_session_id),
3517                        role: None,
3518                    };
3519                    let link = bamboo_broker::BrokerChildLink::connect(
3520                        &bus.endpoint,
3521                        parent,
3522                        &bus.token,
3523                        actor.mailbox_id.clone(),
3524                    )
3525                    .await
3526                    .map_err(|e| {
3527                        AgentError::LLM(format!("broker child link connect failed: {e}"))
3528                    })?;
3529                    let client: Box<dyn bamboo_subagent::ChildLink> = Box::new(link);
3530                    (actor, client)
3531                }
3532            };
3533
3534            // Recover an older Run's already Host-checkpointed Event/Outcome
3535            // receipts before dispatching a successor. The private V2 ledger
3536            // binds exact Child birth, Run, transcript prefix and broker
3537            // Maildir identity. A different broker or an unproven final save
3538            // blocks this activation instead of deleting the wrong mailbox.
3539            if let Some(store) = actor_directory_store.as_ref() {
3540                let recovery = async {
3541                    let receipts = store
3542                        .recover_broker_terminal_receipts(session)
3543                        .await
3544                        .map_err(|error| {
3545                            format!("Host broker receipt recovery blocked: {error}")
3546                        })?;
3547                    for receipt in receipts {
3548                        // This path already owns a successor Run and its public
3549                        // running marker. Old committed receipts authorize ACK
3550                        // only; never publish the predecessor over that owner.
3551                        if client.broker_parent_mailbox() != Some(receipt.parent_mailbox.as_str()) {
3552                            return Err(
3553                                "old Child broker parent mailbox changed before ACK".to_string()
3554                            );
3555                        }
3556                        client
3557                            .acknowledge_recovered_durable_frames(
3558                                &receipt.broker_identity,
3559                                &receipt.message_ids,
3560                            )
3561                            .await
3562                            .map_err(|error| {
3563                                format!("old Child broker ACK unconfirmed: {error}")
3564                            })?;
3565                        store
3566                            .clear_acknowledged_broker_terminal_receipt(
3567                                &receipt.session_id,
3568                                receipt.created_at,
3569                                &receipt.activation_run_id,
3570                                &receipt.parent_mailbox,
3571                            )
3572                            .await
3573                            .map_err(|error| {
3574                                format!("old Child broker ACK cleanup failed: {error}")
3575                            })?;
3576                    }
3577                    Ok::<(), String>(())
3578                }
3579                .await;
3580                if let Err(error) = recovery {
3581                    if !remote {
3582                        actor.worker.kill().await;
3583                    }
3584                    return Err(AgentError::LLM(error));
3585                }
3586            }
3587
3588            // Publish the actor delivery owner and claim the complete bounded
3589            // authorized prefix before dispatching Run. These deliveries ride
3590            // inside RunSpec, so the worker durably enqueues them before its
3591            // first provider boundary rather than racing a later steer frame.
3592            let (delivery_tx, mut delivery_rx) = mpsc::unbounded_channel::<u64>();
3593            let bound_activation_run_id = match session_inbox_runtime.as_ref() {
3594                Some(binding) => {
3595                    let run_id = binding
3596                        .router
3597                        .attach_delivery_sink(&job.child_session_id, delivery_tx.clone())
3598                        .await;
3599                    if run_id.is_none() {
3600                        tracing::debug!(
3601                            session_id = %job.child_session_id,
3602                            "actor driver had no current SessionInbox activation owner to bind"
3603                        );
3604                    }
3605                    run_id
3606                }
3607                None => None,
3608            };
3609            tracing::Span::current().record(
3610                "activation_run_id",
3611                bound_activation_run_id.as_deref().unwrap_or("unbound"),
3612            );
3613            drop(delivery_tx);
3614            let initial_pairs = match (
3615                session_inbox_runtime.as_ref(),
3616                bound_activation_run_id.as_deref(),
3617            ) {
3618                (Some(binding), Some(run_id)) if plain_actor_store.is_none() && !remote => {
3619                    match claim_canonical_deliveries(binding, session, run_id, usize::MAX, None)
3620                        .await
3621                    {
3622                        Ok(deliveries) => deliveries,
3623                        Err(error) => {
3624                            binding
3625                                .router
3626                                .detach_delivery_sink(&job.child_session_id, run_id)
3627                                .await;
3628                            if !remote {
3629                                actor.worker.kill().await;
3630                            }
3631                            return Err(error);
3632                        }
3633                    }
3634                }
3635                _ => Vec::new(),
3636            };
3637            let mut initial_session_messages = initial_pairs
3638                .iter()
3639                .map(|(_, delivery)| delivery.clone())
3640                .collect::<Vec<_>>();
3641            let mut initial_inflight_claims = initial_pairs
3642                .into_iter()
3643                .map(|(claim, _)| claim)
3644                .collect::<VecDeque<_>>();
3645            // Recompute after claim reconciliation: a warm retry may have had a
3646            // canonical receipt whose transcript proof was restored above.
3647            let mut messages = if let Some(binding) = &required_context {
3648                binding
3649                    .validate_messages(&session.id, &session.messages)
3650                    .map_err(|error| AgentError::Budget(error.to_string()))?;
3651                session
3652                    .messages
3653                    .iter()
3654                    .map(serde_json::to_value)
3655                    .collect::<std::result::Result<Vec<_>, _>>()
3656                    .map_err(|_| {
3657                        AgentError::Budget(
3658                            bamboo_domain::ChildContextPacketError::Invalid.to_string(),
3659                        )
3660                    })?
3661            } else {
3662                session
3663                    .messages
3664                    .iter()
3665                    .filter_map(|message| serde_json::to_value(message).ok())
3666                    .collect()
3667            };
3668
3669            if strict_remote.is_some() {
3670                let current =
3671                    remote_canonical_lineage(session_inbox_runtime.as_ref().unwrap(), session)
3672                        .await;
3673                if current.as_ref().ok() != lineage.as_ref()
3674                    || bound_activation_run_id.as_deref().is_none_or(str::is_empty)
3675                    || cancel_token.is_cancelled()
3676                {
3677                    if let (Some(binding), Some(run_id)) = (
3678                        session_inbox_runtime.as_ref(),
3679                        bound_activation_run_id.as_deref(),
3680                    ) {
3681                        binding
3682                            .router
3683                            .detach_delivery_sink(&job.child_session_id, run_id)
3684                            .await;
3685                    }
3686                    return Err(if cancel_token.is_cancelled() {
3687                        AgentError::Cancelled
3688                    } else {
3689                        remote_broker_unavailable()
3690                    });
3691                }
3692            }
3693            if let Some(expected) = creation.as_ref() {
3694                let checked =
3695                    canonical_child_creation(session_inbox_runtime.as_ref().unwrap(), session, job)
3696                        .await;
3697                if checked.as_ref().ok() != Some(expected) {
3698                    if let (Some(binding), Some(run_id)) = (
3699                        session_inbox_runtime.as_ref(),
3700                        bound_activation_run_id.as_deref(),
3701                    ) {
3702                        binding
3703                            .router
3704                            .detach_delivery_sink(&job.child_session_id, run_id)
3705                            .await;
3706                    }
3707                    actor.worker.kill().await;
3708                    return Err(AgentError::LLM(
3709                        "child creation identity changed before dispatch".into(),
3710                    ));
3711                }
3712            }
3713            if let Some(store) = &plain_actor_store {
3714                let started = async {
3715                    // The Run posture was resolved from Host configuration
3716                    // before provisioning. Bind that exact revision and the
3717                    // Host-selected local placement before claiming authority;
3718                    // a worker event or display placement cannot supply either.
3719                    let policy = permission_policy
3720                        .as_ref()
3721                        .ok_or_else(|| plain_actor_unsupported())?;
3722                    let config = self
3723                        .permission_config
3724                        .as_ref()
3725                        .ok_or_else(|| plain_actor_unsupported())?;
3726                    if config.policy_revision() != policy.revision
3727                        || !matches!(spec.placement, Placement::Local)
3728                    {
3729                        return Err(plain_actor_unsupported());
3730                    }
3731                    store
3732                        .bind_control_plane(
3733                            &session.id,
3734                            bamboo_domain::ActorControlPlaneBinding {
3735                                policy_revision: Some(policy.revision),
3736                                placement_intent: Some(bamboo_domain::ActorPlacementIntent::Local),
3737                            },
3738                        )
3739                        .await
3740                        .map_err(|error| AgentError::LLM(error.to_string()))?;
3741                    PlainActorActivation::start_and_prepare(
3742                        store.clone(),
3743                        session,
3744                        session_inbox_runtime.as_ref().unwrap(),
3745                        bound_activation_run_id.as_deref(),
3746                        self.permission_config.clone(),
3747                        !readonly_actor,
3748                        spec.capabilities
3749                            .initial_input_release_required
3750                            .then_some(actor.mailbox_id.as_str()),
3751                    )
3752                    .await
3753                }
3754                .await;
3755                match started {
3756                    Ok((activation, prepared)) => {
3757                        if let Some((prefix, deliveries)) = prepared {
3758                            messages = prefix;
3759                            initial_session_messages = deliveries;
3760                        }
3761                        plain_activation = Some(activation);
3762                    }
3763                    Err(error) => {
3764                        if let Some(run_id) = bound_activation_run_id.as_deref() {
3765                            session_inbox_runtime
3766                                .as_ref()
3767                                .unwrap()
3768                                .router
3769                                .detach_delivery_sink(&job.child_session_id, run_id)
3770                                .await;
3771                        }
3772                        actor.worker.kill().await;
3773                        return Err(error);
3774                    }
3775                }
3776            }
3777            let mut logical_identity = logical_identity_for_actor_run(session, job);
3778            logical_identity.creation = creation.clone();
3779            let run_permission_policy = if strict_remote.is_some() {
3780                let prepared = async {
3781                    let mut policy = permission_policy.clone().ok_or_else(|| {
3782                        AgentError::LLM("remote_environment_permission_policy_missing".into())
3783                    })?;
3784                    let workspace = session.workspace.as_deref().ok_or_else(|| {
3785                        AgentError::LLM("remote_environment_workspace_missing".into())
3786                    })?;
3787                    let run_id = bound_activation_run_id.as_deref().ok_or_else(|| {
3788                        AgentError::LLM("remote_environment_activation_missing".into())
3789                    })?;
3790                    let lease = bamboo_subagent::environment::EnvironmentLease::capture(
3791                        workspace,
3792                        &session.id,
3793                        run_id,
3794                        execution_epoch,
3795                    )
3796                    .await
3797                    .map_err(|error| AgentError::LLM(error.into()))?;
3798                    policy.workspace_path = None;
3799                    policy.environment_lease = Some(lease);
3800                    Ok::<_, AgentError>(Some(policy))
3801                }
3802                .await;
3803                match prepared {
3804                    Ok(policy) => policy,
3805                    Err(error) => {
3806                        if let (Some(binding), Some(run_id)) = (
3807                            session_inbox_runtime.as_ref(),
3808                            bound_activation_run_id.as_deref(),
3809                        ) {
3810                            binding
3811                                .router
3812                                .detach_delivery_sink(&job.child_session_id, run_id)
3813                                .await;
3814                        }
3815                        return Err(error);
3816                    }
3817                }
3818            } else {
3819                permission_policy.clone()
3820            };
3821            if session
3822                .metadata
3823                .contains_key(crate::ticket_worker_plan::TICKET_LOCAL_PLAN_KEY)
3824            {
3825                let admission = async {
3826                    let service = ticket_service
3827                        .as_ref()
3828                        .ok_or_else(|| AgentError::LLM("Ticket authority unavailable".into()))?;
3829                    let runtime = session_inbox_runtime.as_ref().ok_or_else(|| {
3830                        AgentError::LLM("Ticket activation runtime unavailable".into())
3831                    })?;
3832                    let run_id = bound_activation_run_id
3833                        .as_deref()
3834                        .ok_or_else(|| AgentError::LLM("Ticket activation owner missing".into()))?;
3835                    crate::ticket_runtime::admit_ticket_run(service, runtime, session, run_id)
3836                        .await
3837                        .map_err(|e| AgentError::LLM(e.to_string()))
3838                }
3839                .await;
3840                if let Err(error) = admission {
3841                    if let (Some(runtime), Some(run_id)) = (
3842                        session_inbox_runtime.as_ref(),
3843                        bound_activation_run_id.as_deref(),
3844                    ) {
3845                        runtime
3846                            .router
3847                            .detach_delivery_sink(&job.child_session_id, run_id)
3848                            .await;
3849                    }
3850                    actor.worker.kill().await;
3851                    return Err(error);
3852                }
3853            }
3854            let mut run_spec = RunSpec {
3855                // Cloned (not moved) so a retry can re-dispatch to a fresh worker.
3856                assignment: assignment.clone(),
3857                logical_session: Some(logical_identity),
3858                project_id: project_id.clone(),
3859                // Creation/update already resolves the child's explicit
3860                // selection and own role preference into this field. Do
3861                // not infer effort from a possibly inherited model_ref.
3862                reasoning_effort: if !remote && matches!(self.executor, ExecutorSpec::BambooRuntime)
3863                {
3864                    session
3865                        .reasoning_effort
3866                        .map(|effort| effort.as_str().to_owned())
3867                } else {
3868                    None
3869                },
3870                permission_policy: run_permission_policy,
3871                messages,
3872                activation_run_id: bound_activation_run_id.clone(),
3873                execution_epoch,
3874                initial_session_messages,
3875                secrets: run_secrets.clone(),
3876            };
3877            let mut active_lease = if remote {
3878                let reserve = ActivePlacementLease::reserve(
3879                    scoped_registry
3880                        .as_ref()
3881                        .ok_or_else(remote_broker_unavailable)?
3882                        .clone(),
3883                    actor_directory_store
3884                        .as_ref()
3885                        .ok_or_else(remote_broker_unavailable)?
3886                        .clone(),
3887                    scoped_route
3888                        .as_ref()
3889                        .ok_or_else(remote_broker_unavailable)?
3890                        .clone(),
3891                    actor.mailbox_id.clone(),
3892                    session.id.clone(),
3893                    bound_activation_run_id
3894                        .clone()
3895                        .ok_or_else(remote_broker_unavailable)?,
3896                    project_id.clone(),
3897                    scoped_tools.clone().ok_or_else(remote_broker_unavailable)?,
3898                    &cancel_token,
3899                )
3900                .await;
3901                match reserve {
3902                    Ok(lease) => Some(lease),
3903                    Err(error) => {
3904                        if let (Some(binding), Some(run_id)) = (
3905                            session_inbox_runtime.as_ref(),
3906                            bound_activation_run_id.as_deref(),
3907                        ) {
3908                            binding
3909                                .router
3910                                .detach_delivery_sink(&job.child_session_id, run_id)
3911                                .await;
3912                        }
3913                        return Err(error);
3914                    }
3915                }
3916            } else {
3917                None
3918            };
3919            if let Some(lease) = active_lease.as_mut() {
3920                let seeded = async {
3921                    let binding = session_inbox_runtime
3922                        .as_ref()
3923                        .ok_or_else(remote_broker_unavailable)?;
3924                    let run_id = bound_activation_run_id
3925                        .as_deref()
3926                        .ok_or_else(remote_broker_unavailable)?;
3927                    let pairs = claim_canonical_deliveries(
3928                        binding,
3929                        session,
3930                        run_id,
3931                        usize::MAX,
3932                        Some((lease.actor_store.as_ref(), &lease.fence)),
3933                    )
3934                    .await?;
3935                    if let Some(binding) = &required_context {
3936                        binding
3937                            .validate_messages(&session.id, &session.messages)
3938                            .map_err(|error| AgentError::Budget(error.to_string()))?;
3939                    }
3940                    run_spec.messages = session
3941                        .messages
3942                        .iter()
3943                        .map(serde_json::to_value)
3944                        .collect::<std::result::Result<Vec<_>, _>>()
3945                        .map_err(|_| remote_broker_unavailable())?;
3946                    run_spec.initial_session_messages =
3947                        pairs.iter().map(|(_, delivery)| delivery.clone()).collect();
3948                    initial_inflight_claims = pairs
3949                        .into_iter()
3950                        .map(|(claim, _)| claim)
3951                        .collect::<VecDeque<_>>();
3952                    Ok::<(), AgentError>(())
3953                }
3954                .await;
3955                if let Err(error) = seeded {
3956                    if let Err(rollback_error) = lease.abort_before_dispatch().await {
3957                        tracing::warn!(stage = "claim_pre_dispatch_rollback", reason = %rollback_error, "remote claim rollback failed");
3958                    }
3959                    if let (Some(binding), Some(run_id)) = (
3960                        session_inbox_runtime.as_ref(),
3961                        bound_activation_run_id.as_deref(),
3962                    ) {
3963                        binding
3964                            .router
3965                            .detach_delivery_sink(&job.child_session_id, run_id)
3966                            .await;
3967                    }
3968                    return Err(error);
3969                }
3970            }
3971            let mut run_send_invoked = false;
3972            let dispatch = async {
3973                if let Some(activation) = plain_activation
3974                    .as_ref()
3975                    .filter(|activation| !activation.initial_inputs.is_empty())
3976                {
3977                    activation.refresh_continuation_posture(
3978                        session,
3979                        &run_spec,
3980                        &mut expected_permission_posture,
3981                    )?;
3982                }
3983                send_actor_run(&mut *client, run_spec.clone(), &mut run_send_invoked).await
3984            }
3985            .await;
3986            if let Err(e) = dispatch {
3987                if remote {
3988                    tracing::warn!(stage = "run_dispatch", "remote placement admission failed");
3989                }
3990                if let (Some(binding), Some(run_id)) = (
3991                    session_inbox_runtime.as_ref(),
3992                    bound_activation_run_id.as_deref(),
3993                ) {
3994                    binding
3995                        .router
3996                        .detach_delivery_sink(&job.child_session_id, run_id)
3997                        .await;
3998                }
3999                if !remote {
4000                    actor.worker.kill().await;
4001                }
4002                if let Some(activation) = &plain_activation {
4003                    activation.finish(ActorActivationFinish::Failed).await?;
4004                }
4005                if let Some(lease) = active_lease.as_mut() {
4006                    match lease.abort_if_run_unsent(run_send_invoked).await {
4007                        Ok(false) => {
4008                            tracing::warn!(stage = "run_dispatch_unconfirmed", "remote Run send failed after publication became possible; retaining Actor activation and slot until terminal proof or expiry");
4009                        }
4010                        Ok(true) => {}
4011                        Err(rollback_error) => {
4012                            tracing::warn!(stage = "run_pre_dispatch_rollback", reason = %rollback_error, "remote Run rollback failed");
4013                        }
4014                    }
4015                }
4016                return Err(e);
4017            }
4018
4019            // Register as a live actor so send_message (running, no interrupt) can
4020            // steer this child in-band over the existing WS connection. The guard
4021            // unregisters on every exit path.
4022            let (live_tx, mut live_rx) = mpsc::unbounded_channel::<ParentFrame>();
4023            let live_guard = super::live::register(
4024                &job.child_session_id,
4025                live_tx,
4026                attempt as u32,
4027                self.approval_registry.clone(),
4028            );
4029
4030            let mut strict_history = None;
4031            let result = drive(ActorDriveContext {
4032                client: &mut *client,
4033                parent_session_id: &job.parent_session_id,
4034                child_session_id: &job.child_session_id,
4035                child_attempt: attempt as u32,
4036                approval_registry: self.approval_registry.as_ref(),
4037                approval_decider: self.approval_decider.as_ref(),
4038                approval_reviewer: self.approval_reviewer.as_ref(),
4039                escalation_bridge: escalation.clone(),
4040                event_tx: &event_tx,
4041                cancel_token: active_lease
4042                    .as_ref()
4043                    .map_or(&cancel_token, |lease| &lease.run_cancel),
4044                live_rx: &mut live_rx,
4045                delivery_rx: &mut delivery_rx,
4046                logical_session: session,
4047                expected_permission_posture: expected_permission_posture.clone(),
4048                expected_creation: creation.as_ref(),
4049                session_inbox_runtime: session_inbox_runtime.as_ref(),
4050                actor_directory_store: actor_directory_store.as_deref(),
4051                canonical_subagent_tool: canonical_subagent_tool.clone(),
4052                ticket_service: ticket_service.clone(),
4053                activation_run_id: bound_activation_run_id.as_deref(),
4054                execution_epoch,
4055                expected_source_actor_id: &actor.mailbox_id,
4056                initial_inflight_claims,
4057                // First-frame watchdog for EVERY placement: a wedged-but-connected
4058                // worker (subscribed ≠ serving — e.g. stuck on a prior LLM call) emits
4059                // no first frame; without a deadline drive() blocks forever. Bounding it
4060                // turns the "running-but-unresponsive" hang into a recoverable
4061                // WorkerUnresponsive (reap+respawn local / re-pick schedulable / error
4062                // on a fixed remote endpoint).
4063                plain_actor: plain_activation.is_some(),
4064                remote_environment_lease: strict_remote.is_some(),
4065                readonly_output: if readonly_actor {
4066                    Some(&mut readonly_messages)
4067                } else {
4068                    None
4069                },
4070                local_history_tools: local_history_tools.as_deref(),
4071                local_history_read_only: spec.capabilities.read_only_enforced(),
4072                strict_history_output: Some(&mut strict_history),
4073                plain_input: plain_activation.as_ref(),
4074                canonical_activation: active_lease.as_ref().map(|lease| &lease.fence),
4075                canonical_placement_ref: active_lease.as_ref().map(|lease| &lease.placement_ref),
4076                actor_event_observer: actor_event_observer.as_deref(),
4077                // BrokerChildLink replaces the actual correlation for each Run;
4078                // direct legacy WS Terminal frames carry no such identity.
4079                plain_run: self
4080                    .bus
4081                    .as_ref()
4082                    .map(|_| (&run_spec, &self.next_execution_epoch)),
4083                first_frame_timeout: Some(WORKER_FIRST_FRAME_TIMEOUT),
4084            })
4085            .await;
4086            let mut result = if active_lease
4087                .as_ref()
4088                .is_some_and(ActivePlacementLease::lost)
4089            {
4090                Err(remote_broker_unavailable())
4091            } else {
4092                result
4093            };
4094            if local_history_tools.is_none() {
4095                let reply = result
4096                    .as_ref()
4097                    .ok()
4098                    .and_then(|reply| reply.as_deref())
4099                    .filter(|reply| !reply.is_empty())
4100                    .map(str::to_owned);
4101                if let (Some(lease), Some(reply)) = (active_lease.as_ref(), reply) {
4102                    if let Err(error) = append_fenced_remote_reply(
4103                        lease.actor_store.as_ref(),
4104                        &lease.fence,
4105                        session,
4106                        &reply,
4107                    )
4108                    .await
4109                    {
4110                        result = Err(error);
4111                    }
4112                }
4113            }
4114            if remote {
4115                if let Err(error) = &result {
4116                    let code = match error {
4117                        AgentError::LLM(message) if message.starts_with("actor event") => {
4118                            "actor_event"
4119                        }
4120                        AgentError::LLM(message) if message.starts_with("remote_actor") => {
4121                            "remote_actor"
4122                        }
4123                        AgentError::LLM(message) if message.starts_with("remote_environment") => {
4124                            if message.len() <= 128
4125                                && message.bytes().all(|byte| {
4126                                    byte.is_ascii_lowercase()
4127                                        || byte.is_ascii_digit()
4128                                        || byte == b'_'
4129                                })
4130                            {
4131                                message.as_str()
4132                            } else {
4133                                "environment"
4134                            }
4135                        }
4136                        AgentError::LLM(message) if message.contains("permission posture") => {
4137                            "permission_posture"
4138                        }
4139                        AgentError::LLM(message) if message.contains("broker") => "broker",
4140                        AgentError::LLM(_) => "other_llm",
4141                        AgentError::WorkerUnresponsive(_) => "unresponsive",
4142                        AgentError::Cancelled => "cancelled",
4143                        _ => "other",
4144                    };
4145                    tracing::warn!(stage = "actor_drive", code, "remote actor run failed");
4146                }
4147            }
4148            if let Some(lease) = active_lease.as_mut() {
4149                let accepted_terminal = client
4150                    .durable_delivery_receipt()
4151                    .map(|receipt| receipt.terminal_status);
4152                if let Err(error) = lease
4153                    .complete(
4154                        remote_activation_finish(&result, accepted_terminal),
4155                        client.has_pending_durable_terminal(),
4156                    )
4157                    .await
4158                {
4159                    tracing::warn!(
4160                        stage = "actor_complete",
4161                        "remote placement admission failed"
4162                    );
4163                    if let (Some(binding), Some(run_id)) = (
4164                        session_inbox_runtime.as_ref(),
4165                        bound_activation_run_id.as_deref(),
4166                    ) {
4167                        binding
4168                            .router
4169                            .detach_delivery_sink(&job.child_session_id, run_id)
4170                            .await;
4171                    }
4172                    drop(live_guard);
4173                    drop(client); // No Host checkpoint proof means no broker ACK.
4174                    drop(active_lease);
4175                    return Err(error);
4176                }
4177            }
4178            drop(active_lease);
4179            if let (Some(binding), Some(run_id)) = (
4180                session_inbox_runtime.as_ref(),
4181                bound_activation_run_id.as_deref(),
4182            ) {
4183                binding
4184                    .router
4185                    .detach_delivery_sink(&job.child_session_id, run_id)
4186                    .await;
4187            }
4188            // Unregister IMMEDIATELY: after drive returns nobody consumes live_rx,
4189            // so a send_message landing in the close/park window below must see
4190            // "not live" and take the durable-queue fallback instead of vanishing.
4191            // (Even if one slipped in earlier, send_message also appends it to the
4192            // durable transcript, so the next activation still rehydrates it.)
4193            drop(live_guard);
4194            // Keep only a fully processed broker terminal's exact mailbox
4195            // receipts alive until the SDK has saved this Child's canonical
4196            // transcript and final status. No run fence means no Host proof.
4197            if client.has_pending_durable_terminal() {
4198                if let Some(run_id) = bound_activation_run_id.as_deref() {
4199                    let key = (session.id.clone(), session.created_at, run_id.to_owned());
4200                    let host_nested_wait_handoff = session
4201                        .metadata
4202                        .get("runtime.actor_nested_wait_handoff")
4203                        .is_some_and(|value| value == "true");
4204                    let mut pending = self.pending_durable_links.lock().await;
4205                    if pending.contains_key(&key) {
4206                        return Err(AgentError::LLM(
4207                            "child broker terminal receipt already awaits Host checkpoint".into(),
4208                        ));
4209                    }
4210                    pending.insert(
4211                        key,
4212                        PendingDurableChildLink {
4213                            link: client,
4214                            host_nested_wait_handoff,
4215                            strict_history,
4216                            strict_history_required: local_history_tools.is_some(),
4217                            strict_history_committed: false,
4218                        },
4219                    );
4220                } else {
4221                    // A Worker-local transcript is not Host checkpoint proof.
4222                    drop(client);
4223                }
4224            } else {
4225                drop(client);
4226            }
4227
4228            // Retry only a local worker that we can kill. A silent remote
4229            // Worker may still be executing its first Run; starting the same
4230            // Actor on another Host before a terminal fence would duplicate it.
4231            if required_context.is_none()
4232                && attempt == 0
4233                && matches!(kind, PlacementKind::Local)
4234                && matches!(result, Err(AgentError::WorkerUnresponsive(_)))
4235            {
4236                tracing::warn!(
4237                    "actor child {} got no first frame; reaping the worker and respawning once",
4238                    job.child_session_id
4239                );
4240                actor.worker.kill().await;
4241                attempt += 1;
4242                continue;
4243            }
4244            break (result, actor);
4245        };
4246
4247        let result = if strict_remote.is_some() {
4248            result.map_err(|error| match &error {
4249                AgentError::Cancelled => error,
4250                AgentError::LLM(message) => remote_environment_error_code(message)
4251                    .map(|code| AgentError::LLM(code.to_owned()))
4252                    .unwrap_or_else(remote_broker_unavailable),
4253                _ => remote_broker_unavailable(),
4254            })
4255        } else {
4256            result
4257        };
4258
4259        // Park the warm worker for reuse on a clean run, or kill it on
4260        // error/cancel (a wedged worker must not be reused). Remote / schedulable
4261        // workers are registry-managed — never ours to pool/kill, just drop.
4262        if remote {
4263            drop(actor);
4264        } else if required_context.is_some() {
4265            if session
4266                .metadata
4267                .contains_key(crate::ticket_worker_plan::TICKET_LOCAL_PLAN_KEY)
4268            {
4269                let service = self.ticket_service.lock().recover_poison().clone();
4270                let runtime = self.session_inbox_runtime.lock().recover_poison().clone();
4271                let owned_run_id = crate::ticket_runtime::read_dispatch(session)
4272                    .ok()
4273                    .flatten()
4274                    .and_then(|dispatch| dispatch.receipt.map(|receipt| receipt.run_id));
4275                match actor.worker.kill_confirmed().await {
4276                    Ok(proof) => {
4277                        if let (Some(service), Some(runtime), Some(run_id)) =
4278                            (service, runtime, owned_run_id.as_deref())
4279                        {
4280                            if let Err(error) = crate::ticket_runtime::checkpoint_owned_ticket_stop(
4281                                &service,
4282                                &runtime,
4283                                session,
4284                                run_id,
4285                                &proof,
4286                                result.is_ok(),
4287                            )
4288                            .await
4289                            {
4290                                // The canonical terminal save can retain output
4291                                // and retry this stop fact before broker ACK.
4292                                tracing::warn!(child_id = %session.id, %error, "Ticket owned-process stop publication retained for reconciliation");
4293                            }
4294                        }
4295                    }
4296                    Err(error) => {
4297                        tracing::warn!(child_id = %session.id, %error, "Ticket process stop remains unconfirmed; resources retained");
4298                    }
4299                }
4300            } else {
4301                actor.worker.kill().await;
4302            }
4303        } else {
4304            match &result {
4305                Ok(_) => self.release_bus_worker(&pool_key, actor).await,
4306                Err(_) => actor.worker.kill().await,
4307            }
4308        }
4309
4310        if let Some(activation) = plain_activation {
4311            return match result {
4312                Ok(Some(text)) if !text.is_empty() => {
4313                    if readonly_actor {
4314                        let messages = std::mem::take(&mut readonly_messages);
4315                        let committed = activation
4316                            .store
4317                            .append_actor_transcript(bamboo_storage::ActorTranscriptAppend {
4318                                fence: activation.fence.clone(),
4319                                expected_created_at: activation.created_at,
4320                                expected_conversation_generation: activation
4321                                    .conversation_generation
4322                                    .clone(),
4323                                expected_messages: session.messages.clone(),
4324                                expected_provider_transcript: activation
4325                                    .provider_transcript
4326                                    .clone(),
4327                                messages: messages.clone(),
4328                                native_groups: Vec::new(),
4329                            })
4330                            .await;
4331                        return match committed {
4332                            Ok(committed) => {
4333                                *session = committed;
4334                                for message in messages {
4335                                    let _ = event_tx
4336                                        .send(AgentEvent::message_appended(&session.id, &message))
4337                                        .await;
4338                                }
4339                                activation
4340                                    .finish_with_observer(
4341                                        ActorActivationFinish::Succeeded,
4342                                        actor_event_observer.as_deref(),
4343                                    )
4344                                    .await
4345                            }
4346                            Err(error) => {
4347                                let _ = activation.finish(ActorActivationFinish::Failed).await;
4348                                Err(AgentError::LLM(format!(
4349                                    "actor reply commit failed: {error}"
4350                                )))
4351                            }
4352                        };
4353                    }
4354                    match activation.append_reply(session, text, &event_tx).await {
4355                        Ok(()) => {
4356                            activation
4357                                .finish_with_observer(
4358                                    ActorActivationFinish::Succeeded,
4359                                    actor_event_observer.as_deref(),
4360                                )
4361                                .await
4362                        }
4363                        Err(error) => {
4364                            // OutcomeUnconfirmed is an error, never ordinary-save fallback.
4365                            let _ = activation.finish(ActorActivationFinish::Failed).await;
4366                            Err(AgentError::LLM(format!(
4367                                "actor reply commit failed: {error}"
4368                            )))
4369                        }
4370                    }
4371                }
4372                other => {
4373                    let outcome = if matches!(&other, Err(AgentError::Cancelled)) {
4374                        ActorActivationFinish::Cancelled
4375                    } else {
4376                        ActorActivationFinish::Failed
4377                    };
4378                    activation.finish(outcome).await?;
4379                    match other {
4380                        Err(error) => Err(error),
4381                        _ => Err(plain_actor_unsupported()),
4382                    }
4383                }
4384            };
4385        }
4386
4387        // Write-back: persist the actor's final reply onto the child session so
4388        // the transcript survives and the NEXT activation sees it as history.
4389        // (run_child_spawn saves the session right after we return.)
4390        match result {
4391            Ok(Some(text)) => {
4392                if !remote && local_history_tools.is_none() && !text.is_empty() {
4393                    session.add_message(bamboo_agent_core::Message::assistant(text, None));
4394                }
4395                Ok(())
4396            }
4397            Ok(None) => Ok(()),
4398            Err(e) => Err(e),
4399        }
4400    }
4401}
4402
4403// This is a conservative no-replay observation, not Actor authority. The
4404// canonical mapping comes from Storage; record/marker names match its writer.
4405async fn require_unowned_glob(
4406    store: &bamboo_storage::SessionStoreV2,
4407    session: &Session,
4408) -> Result<(), AgentError> {
4409    let rel = store
4410        .resolve_rel_path(&session.id)
4411        .await
4412        .ok_or_else(|| plain_actor_unsupported())?;
4413    if !store.bamboo_home_dir().is_absolute()
4414        || rel
4415            != format!(
4416                "sessions/{}/children/{}",
4417                session.root_session_id, session.id
4418            )
4419    {
4420        return Err(plain_actor_unsupported());
4421    }
4422    let directory = store.bamboo_home_dir().join(rel);
4423    let mut checked = PathBuf::new();
4424    for component in directory.components() {
4425        if !matches!(
4426            component,
4427            std::path::Component::RootDir | std::path::Component::Normal(_)
4428        ) {
4429            return Err(plain_actor_unsupported());
4430        }
4431        checked.push(component.as_os_str());
4432        if !tokio::fs::symlink_metadata(&checked)
4433            .await
4434            .map_err(|_| plain_actor_unsupported())?
4435            .is_dir()
4436        {
4437            return Err(plain_actor_unsupported());
4438        }
4439    }
4440    for name in ["actor-authority.json", "actor-authority.initialized.json"] {
4441        match tokio::fs::symlink_metadata(directory.join(name)).await {
4442            Err(error) if error.kind() == std::io::ErrorKind::NotFound => {}
4443            _ => return Err(plain_actor_unsupported()),
4444        }
4445    }
4446    Ok(())
4447}
4448
4449// Zero-tool or read-only Glob named Child of a durable Ultra Root.
4450// Glob uses its typed tail; zero-tool supports new input after a plain completion
4451// as well as two bounded corrections within a Running activation.
4452// An expired initial owner can be replaced with one new input through the same
4453// checkpoint/release boundary. No lease renewal, raw steering or remote activation.
4454struct PlainActorActivation {
4455    store: Arc<bamboo_storage::SessionStoreV2>,
4456    fence: ActorActivationFence,
4457    inbox_generation: u64,
4458    created_at: chrono::DateTime<chrono::Utc>,
4459    conversation_generation: bamboo_domain::ConversationGenerationObservation,
4460    provider_transcript: bamboo_domain::ProviderTranscriptState,
4461    input_consumer: SessionInboxConsumerId,
4462    input_deadline: chrono::DateTime<chrono::Utc>,
4463    actor_deadline: chrono::DateTime<chrono::Utc>,
4464    initial_release_required: bool,
4465    input_inbox: bamboo_storage::FileSessionInbox,
4466    permission_config: Option<Arc<bamboo_tools::permission::PermissionConfig>>,
4467    initial_inputs: Vec<SessionInboxOwnedClaim>,
4468    initial_input_count: usize,
4469    recovering_pre_ack: bool,
4470    completion_router: std::sync::Mutex<Option<ActorEventRouter>>,
4471}
4472type PlainInitialDelivery = (Vec<serde_json::Value>, SessionMessageDelivery);
4473type PlainInitialBatch = (Vec<serde_json::Value>, Vec<SessionMessageDelivery>);
4474#[track_caller]
4475fn plain_actor_unsupported() -> AgentError {
4476    let source = std::panic::Location::caller();
4477    tracing::warn!(
4478        process_id = std::process::id(),
4479        source_file = source.file(),
4480        source_line = source.line(),
4481        source_column = source.column(),
4482        "plain Actor continuation unsupported"
4483    );
4484    AgentError::LLM("This Actor Child supports a fresh plain activation, a completed plain Child with a bounded batch of new inputs, two bounded corrections while Running, a Failed retry with one new input, an expired initial Local owner with one new input, or verified expired pre-ACK input recovery through run(reset_to_last_user=false). Other continuation is unsupported; durable history is preserved.".into())
4485}
4486fn plain_initial_history(session: &Session) -> bool {
4487    !session.messages.iter().any(|message| {
4488        matches!(message.role, Role::Assistant | Role::Tool)
4489            || message
4490                .metadata
4491                .as_ref()
4492                .is_some_and(|metadata| metadata.get("_bamboo_owned_input_checkpoint").is_some())
4493    }) && session.provider_transcript.is_empty()
4494        && session.session_inbox_admission().is_none()
4495        && session.pending_injected_messages().is_none()
4496}
4497// Selection only. Canonical terminal status/new input are checked at start;
4498// the existing fenced input checkpoint validates the complete durable prefix.
4499fn plain_completed_history(session: &Session) -> bool {
4500    session.provider_transcript.is_empty()
4501        && session.pending_injected_messages().is_none()
4502        && session
4503            .messages
4504            .last()
4505            .is_some_and(|message| message.role == Role::Assistant && !message.content.is_empty())
4506        && !session
4507            .messages
4508            .iter()
4509            .any(|message| message.role == Role::Tool || message.tool_calls.is_some())
4510}
4511// Selection only. Storage's closed raw bookkeeper/current-prefix validator is
4512// still mandatory after the actual replacement claim; this metadata is no grant.
4513fn pre_ack_input(
4514    session: &Session,
4515) -> Result<Option<(bamboo_domain::SessionMessageId, u64)>, AgentError> {
4516    let Some(message) = session.messages.last() else {
4517        return Ok(None);
4518    };
4519    let Some(proof) = message
4520        .metadata
4521        .as_ref()
4522        .and_then(|v| v.get("_bamboo_owned_input_checkpoint"))
4523    else {
4524        return Ok(None);
4525    };
4526    let id = bamboo_domain::SessionMessageId::parse(message.id.clone())
4527        .map_err(|_| plain_actor_unsupported())?;
4528    let generation = proof
4529        .get("generation")
4530        .and_then(serde_json::Value::as_u64)
4531        .filter(|g| *g > 0)
4532        .ok_or_else(|| plain_actor_unsupported())?;
4533    if message.role != Role::User
4534        || session
4535            .messages
4536            .iter()
4537            .filter(|m| m.id == message.id)
4538            .count()
4539            != 1
4540        || !session
4541            .session_inbox_admission()
4542            .is_some_and(|cursor| cursor.contains(&id))
4543        || !session.provider_transcript.is_empty()
4544        || session.pending_injected_messages().is_some()
4545        || session
4546            .messages
4547            .iter()
4548            .any(|m| m.role == Role::Tool || m.tool_calls.is_some())
4549    {
4550        return Err(plain_actor_unsupported());
4551    }
4552    Ok(Some((id, generation)))
4553}
4554fn qualified_pre_ack_owner(
4555    entry: &bamboo_domain::ActorDirectoryEntry,
4556    now: chrono::DateTime<chrono::Utc>,
4557) -> bool {
4558    let Some(old) = entry.activation.as_ref() else {
4559        return false;
4560    };
4561    let recoverable = (entry.actor.state == ActorLogicalState::Failed
4562        && old.status == bamboo_domain::ActorActivationStatus::Failed)
4563        || (entry.actor.state == ActorLogicalState::Active
4564            && old.status.is_live()
4565            && old.lease_expires_at <= now);
4566    let trusted = old.placement_ref.as_ref().is_some_and(|placement| {
4567        placement.class == bamboo_domain::ActorPlacementClass::Local
4568            && placement
4569                .lease_id
4570                .strip_prefix("owned-initial-release-v1:")
4571                .is_some_and(|worker| {
4572                    !worker.is_empty()
4573                        && worker.len() <= 256
4574                        && !worker.contains("..")
4575                        && worker.bytes().all(|b| {
4576                            b.is_ascii_alphanumeric() || b == b'-' || b == b'_' || b == b'.'
4577                        })
4578                })
4579    });
4580    recoverable && trusted
4581}
4582async fn expired_pre_ack_input(
4583    inbox: &bamboo_storage::FileSessionInbox,
4584    session: &Session,
4585) -> Result<Option<(bamboo_domain::SessionMessageId, u64)>, AgentError> {
4586    let Some((id, generation)) = pre_ack_input(session)? else {
4587        return Ok(None);
4588    };
4589    let backlog = inbox
4590        .inspect(&session.id)
4591        .await
4592        .map_err(|_| plain_actor_unsupported())?;
4593    let leases = inbox
4594        .inspect_owned_leases(&session.id, 2, chrono::Utc::now())
4595        .await
4596        .map_err(|_| plain_actor_unsupported())?;
4597    if backlog.pending != 0
4598        || backlog.claimed != 1
4599        || backlog.generation != generation
4600        || backlog.oldest_generation != Some(generation)
4601        || !backlog.activation_pending()
4602        || leases.len() != 1
4603        || !leases[0].expired
4604        || leases[0].generation != generation
4605        || inbox
4606            .was_admitted(&session.id, &id)
4607            .await
4608            .map_err(|_| plain_actor_unsupported())?
4609    {
4610        return Err(plain_actor_unsupported());
4611    }
4612    Ok(Some((id, generation)))
4613}
4614impl PlainActorActivation {
4615    // Callee-side boxing keeps this storage chain out of the already large
4616    // external-runner future without changing its cancellation semantics.
4617    fn start<'a>(
4618        store: Arc<bamboo_storage::SessionStoreV2>,
4619        session: &'a Session,
4620        binding: &'a SessionInboxRuntimeBinding,
4621        run_id: Option<&'a str>,
4622        permission_config: Option<Arc<bamboo_tools::permission::PermissionConfig>>,
4623        allow_failed_retry: bool,
4624        release_worker: Option<&'a str>,
4625    ) -> std::pin::Pin<Box<impl std::future::Future<Output = Result<Self, AgentError>> + Send + 'a>>
4626    {
4627        Box::pin(async move {
4628            let run_id = run_id.ok_or_else(|| plain_actor_unsupported())?;
4629            let backlog = binding
4630                .inbox
4631                .inspect(&session.id)
4632                .await
4633                .map_err(|_| plain_actor_unsupported())?;
4634            let entry = store
4635                .ensure_actor(&session.id)
4636                .await
4637                .map_err(|error| AgentError::LLM(error.to_string()))?;
4638            let config = permission_config
4639                .as_ref()
4640                .ok_or_else(|| plain_actor_unsupported())?;
4641            if entry.actor.policy_revision != Some(config.policy_revision())
4642                || entry.actor.placement_intent != Some(bamboo_domain::ActorPlacementIntent::Local)
4643            {
4644                return Err(plain_actor_unsupported());
4645            }
4646            let fresh = entry.actor.current_attempt == 0
4647                && entry.actor.state == ActorLogicalState::Cold
4648                && entry.activation.is_none()
4649                && backlog.pending == 0
4650                && backlog.claimed == 0
4651                && backlog.generation == 0;
4652            let new_input = backlog.pending == 1
4653                && backlog.claimed == 0
4654                && backlog.oldest_generation == Some(backlog.generation)
4655                && backlog.activation_pending();
4656            let retry = allow_failed_retry
4657                && entry.actor.state == ActorLogicalState::Failed
4658                && entry.actor.current_attempt > 0
4659                && entry.activation.as_ref().is_some_and(|old| {
4660                    old.status == bamboo_domain::ActorActivationStatus::Failed
4661                        && backlog.generation > old.inbox_generation
4662                })
4663                && new_input;
4664            // Only completed plain local history gains a bounded initial batch.
4665            // Retry, expired owner and pre-ACK recovery retain their single-input
4666            // contracts. QueueOnly alone still cannot authorize a cold start.
4667            let new_batch = backlog.pending > 1
4668                && backlog.pending <= bamboo_domain::SessionInboxLimits::default().max_claim_batch
4669                && backlog.claimed == 0
4670                && backlog
4671                    .oldest_generation
4672                    .is_some_and(|oldest| oldest < backlog.generation)
4673                && backlog.activation_pending();
4674            let continuation = allow_failed_retry
4675                && release_worker.is_some()
4676                && entry.actor.state == ActorLogicalState::Cold
4677                && entry.actor.current_attempt > 0
4678                && entry.activation.as_ref().is_some_and(|old| {
4679                    old.status == bamboo_domain::ActorActivationStatus::Succeeded
4680                        && backlog
4681                            .oldest_generation
4682                            .is_some_and(|oldest| oldest > old.inbox_generation)
4683                })
4684                && (new_input || new_batch)
4685                && plain_completed_history(session);
4686            let input_inbox = bamboo_storage::FileSessionInbox::new(
4687                store.clone(),
4688                bamboo_domain::SessionInboxLimits::default(),
4689            );
4690            let now = chrono::Utc::now();
4691            // The old initial Run may still be executing physically. Its final
4692            // append remains fenced; only a new owned input can release this
4693            // replacement worker. Previously admitted history is not replayed.
4694            let replacement = allow_failed_retry
4695                && release_worker.is_some()
4696                && entry.actor.state == ActorLogicalState::Active
4697                && qualified_pre_ack_owner(&entry, now)
4698                && entry.activation.as_ref().is_some_and(|old| {
4699                    old.status.is_live() && old.lease_expires_at <= now && old.inbox_generation == 0
4700                })
4701                && new_input
4702                && plain_initial_history(session);
4703            let recovery = if allow_failed_retry
4704                && release_worker.is_some()
4705                && qualified_pre_ack_owner(&entry, now)
4706            {
4707                expired_pre_ack_input(&input_inbox, session).await?
4708            } else {
4709                None
4710            };
4711            if (recovery.is_some() || continuation || replacement)
4712                && (!entry.actor.matches_session(session)
4713                    || entry.actor.project_id
4714                        != project_id_for_actor_run(session)?.map(|id| id.to_string()))
4715            {
4716                return Err(plain_actor_unsupported());
4717            }
4718            if recovery.is_none()
4719                && !continuation
4720                && !replacement
4721                && ((!fresh && !retry) || !plain_initial_history(session))
4722            {
4723                return Err(plain_actor_unsupported());
4724            }
4725            let duration = crate::runtime::execution::spawn::watchdog_policy_for_session(session)
4726                .max_total_secs
4727                .checked_add(60)
4728                .and_then(chrono::Duration::try_seconds)
4729                .ok_or_else(|| plain_actor_unsupported())?;
4730            let expires = now
4731                .checked_add_signed(duration)
4732                .ok_or_else(|| plain_actor_unsupported())?;
4733            let input_consumer = SessionInboxConsumerId::new();
4734            let input_deadline = expires.min(now + chrono::Duration::hours(1));
4735            let recovered = if let Some((id, generation)) = &recovery {
4736                if !binding.router.owns_run(&session.id, run_id).await {
4737                    return Err(plain_actor_unsupported());
4738                }
4739                let lease = SessionInboxLeaseRequest {
4740                    consumer: input_consumer.clone(),
4741                    now,
4742                    duration: input_deadline - now,
4743                };
4744                let mut claims = input_inbox
4745                    .claim_owned(&session.id, 1, Some(run_id), &lease)
4746                    .await
4747                    .map_err(|_| plain_actor_unsupported())?;
4748                let claim = claims.pop().ok_or_else(|| plain_actor_unsupported())?;
4749                if claim.claim.envelope.id != *id
4750                    || claim.claim.generation != *generation
4751                    || !matches!(&claim.claim.envelope.body, bamboo_domain::SessionMessageBody::Content(content)
4752                        if content.parts.is_empty() && !content.text.trim().is_empty() && content.text.len() <= 8192)
4753                {
4754                    return Err(plain_actor_unsupported());
4755                }
4756                Some(claim)
4757            } else {
4758                None
4759            };
4760            let claimed = store
4761                .claim_activation(&ActorActivationClaim {
4762                    actor_id: session.id.clone(),
4763                    run_id: run_id.into(),
4764                    lease_owner: input_consumer.as_str().into(),
4765                    lease_expires_at: expires,
4766                    inbox_generation: recovered
4767                        .as_ref()
4768                        .map_or(backlog.generation, |claim| claim.claim.generation),
4769                    placement_ref: release_worker.map(|worker| bamboo_domain::ActorPlacementRef {
4770                        class: bamboo_domain::ActorPlacementClass::Local,
4771                        // Written only after the real worker capability probe
4772                        // and spawn. Opaque provenance, never a live-owner grant.
4773                        lease_id: format!("owned-initial-release-v1:{worker}"),
4774                        slot_epoch: None,
4775                    }),
4776                    now,
4777                })
4778                .await
4779                .map_err(|error| AgentError::LLM(error.to_string()))?;
4780            let activation = Self {
4781                store,
4782                fence: claimed.fence(),
4783                inbox_generation: claimed.inbox_generation,
4784                created_at: session.created_at,
4785                conversation_generation: session.conversation_generation_observation(),
4786                provider_transcript: session.provider_transcript.clone(),
4787                input_consumer,
4788                input_deadline,
4789                actor_deadline: claimed.lease_expires_at,
4790                initial_release_required: release_worker.is_some(),
4791                input_inbox,
4792                permission_config,
4793                initial_inputs: recovered.into_iter().collect(),
4794                initial_input_count: if continuation {
4795                    backlog.pending
4796                } else {
4797                    usize::from(claimed.inbox_generation > 0)
4798                },
4799                recovering_pre_ack: recovery.is_some(),
4800                completion_router: std::sync::Mutex::new(None),
4801            };
4802            if let Err(error) = activation
4803                .store
4804                .start_activation(&activation.fence, chrono::Utc::now())
4805                .await
4806            {
4807                let _ = activation.finish(ActorActivationFinish::Failed).await;
4808                return Err(AgentError::LLM(error.to_string()));
4809            }
4810            Ok(activation)
4811        })
4812    }
4813    fn start_and_prepare<'a>(
4814        store: Arc<bamboo_storage::SessionStoreV2>,
4815        session: &'a mut Session,
4816        binding: &'a SessionInboxRuntimeBinding,
4817        run_id: Option<&'a str>,
4818        permission_config: Option<Arc<bamboo_tools::permission::PermissionConfig>>,
4819        allow_failed_retry: bool,
4820        release_worker: Option<&'a str>,
4821    ) -> std::pin::Pin<
4822        Box<
4823            impl std::future::Future<Output = Result<(Self, Option<PlainInitialBatch>), AgentError>>
4824                + Send
4825                + 'a,
4826        >,
4827    > {
4828        Box::pin(async move {
4829            let mut activation = Self::start(
4830                store,
4831                session,
4832                binding,
4833                run_id,
4834                permission_config,
4835                allow_failed_retry,
4836                release_worker,
4837            )
4838            .await?;
4839            if activation.inbox_generation == 0 {
4840                return Ok((activation, None));
4841            }
4842            let prepared = async {
4843                let run_id = run_id.ok_or_else(|| plain_actor_unsupported())?;
4844                let backlog = activation
4845                    .input_inbox
4846                    .inspect(&session.id)
4847                    .await
4848                    .map_err(|_| plain_actor_unsupported())?;
4849                let expected = if activation.recovering_pre_ack {
4850                    (0, 1)
4851                } else {
4852                    (activation.initial_input_count, 0)
4853                };
4854                if (backlog.pending, backlog.claimed) != expected
4855                    || backlog.generation != activation.inbox_generation
4856                {
4857                    return Err(plain_actor_unsupported());
4858                }
4859                let claims = if activation.initial_inputs.is_empty() {
4860                    // claim_owned is idempotent for this consumer. Repeating
4861                    // limit=1 would return the same first claim, not the next.
4862                    // Select once under the existing bounded queue transaction.
4863                    activation
4864                        .claim_inputs(binding, session, run_id, activation.initial_input_count)
4865                        .await?
4866                } else {
4867                    activation.initial_inputs.clone()
4868                };
4869                if claims.len() != activation.initial_input_count
4870                    || claims.last().map(|claim| claim.claim.generation)
4871                        != Some(activation.inbox_generation)
4872                    || claims
4873                        .windows(2)
4874                        .any(|pair| pair[0].claim.generation >= pair[1].claim.generation)
4875                {
4876                    return Err(plain_actor_unsupported());
4877                }
4878                // Validate/claim the entire bounded selection before publishing
4879                // any input. Each checkpoint still uses its own physical lease,
4880                // complete current prefix and actual Actor fence. A partial or
4881                // uncertain checkpoint never dispatches a Run or ACKs a sibling.
4882                let mut prefix = None;
4883                let mut deliveries = Vec::with_capacity(claims.len());
4884                for claim in &claims {
4885                    let (messages, delivery) =
4886                        activation.checkpoint_input(session, run_id, claim).await?;
4887                    prefix.get_or_insert(messages);
4888                    deliveries.push(delivery);
4889                }
4890                activation.initial_inputs = claims;
4891                Ok((prefix.ok_or_else(|| plain_actor_unsupported())?, deliveries))
4892            }
4893            .await;
4894            match prepared {
4895                Ok(delivery) => Ok((activation, Some(delivery))),
4896                Err(error) => {
4897                    activation.finish(ActorActivationFinish::Failed).await?;
4898                    Err(error)
4899                }
4900            }
4901        })
4902    }
4903    fn claim_input<'a>(
4904        &'a self,
4905        binding: &'a SessionInboxRuntimeBinding,
4906        session: &'a Session,
4907        run_id: &'a str,
4908    ) -> std::pin::Pin<
4909        Box<
4910            impl std::future::Future<Output = Result<Option<SessionInboxOwnedClaim>, AgentError>>
4911                + Send
4912                + 'a,
4913        >,
4914    > {
4915        Box::pin(async move { Ok(self.claim_inputs(binding, session, run_id, 1).await?.pop()) })
4916    }
4917    fn claim_inputs<'a>(
4918        &'a self,
4919        binding: &'a SessionInboxRuntimeBinding,
4920        session: &'a Session,
4921        run_id: &'a str,
4922        limit: usize,
4923    ) -> std::pin::Pin<
4924        Box<
4925            impl std::future::Future<Output = Result<Vec<SessionInboxOwnedClaim>, AgentError>>
4926                + Send
4927                + 'a,
4928        >,
4929    > {
4930        Box::pin(async move {
4931            if limit == 0 || limit > bamboo_domain::SessionInboxLimits::default().max_claim_batch {
4932                return Err(plain_actor_unsupported());
4933            }
4934            if !self
4935                .input_inbox
4936                .inspect(&session.id)
4937                .await
4938                .map_err(|_| plain_actor_unsupported())?
4939                .activation_pending()
4940            {
4941                return Ok(Vec::new());
4942            }
4943            if run_id != self.fence.run_id || !binding.router.owns_run(&session.id, run_id).await {
4944                return Err(plain_actor_unsupported());
4945            }
4946            let now = chrono::Utc::now();
4947            if now >= self.input_deadline {
4948                return Err(plain_actor_unsupported());
4949            }
4950            let lease = SessionInboxLeaseRequest {
4951                consumer: self.input_consumer.clone(),
4952                now,
4953                duration: self.input_deadline - now,
4954            };
4955            let claims = self
4956                .input_inbox
4957                .claim_owned(&session.id, limit, Some(run_id), &lease)
4958                .await
4959                .map_err(|error| {
4960                    tracing::warn!(%error, "Actor correction claim unconfirmed");
4961                    AgentError::LLM(
4962                        "Actor correction claim unconfirmed; durable input is preserved".into(),
4963                    )
4964                })?;
4965            for claim in &claims {
4966                if !matches!(&claim.claim.envelope.body, bamboo_domain::SessionMessageBody::Content(content)
4967                    if content.parts.is_empty() && !content.text.trim().is_empty() && content.text.len() <= 8192)
4968                {
4969                    self.input_inbox
4970                        .fail_owned(
4971                            &session.id,
4972                            claim,
4973                            &SessionInboxFailureReport {
4974                                now: chrono::Utc::now(),
4975                                error_code: "unsupported_actor_input".into(),
4976                            },
4977                        )
4978                        .await
4979                        .map_err(|error| {
4980                            tracing::warn!(%error, "Actor correction failure report unconfirmed");
4981                            AgentError::LLM(
4982                                "Actor correction failure report unconfirmed; durable input is preserved".into(),
4983                            )
4984                        })?;
4985                    return Err(plain_actor_unsupported());
4986                }
4987            }
4988            Ok(claims)
4989        })
4990    }
4991    fn checkpoint_input<'a>(
4992        &'a self,
4993        session: &'a mut Session,
4994        run_id: &'a str,
4995        claim: &'a SessionInboxOwnedClaim,
4996    ) -> std::pin::Pin<
4997        Box<
4998            impl std::future::Future<Output = Result<PlainInitialDelivery, AgentError>> + Send + 'a,
4999        >,
5000    > {
5001        Box::pin(async move {
5002            // The worker receives the committed pre-input prefix. Only typed
5003            // delivery adds the User; the Host-private bookkeeper is not a grant.
5004            let request = bamboo_storage::ActorInputCheckpoint {
5005                fence: self.fence.clone(),
5006                expected_created_at: self.created_at,
5007                expected_conversation_generation: self.conversation_generation.clone(),
5008                claim: claim.clone(),
5009                expected_messages: session.messages.clone(),
5010                expected_provider_transcript: session.provider_transcript.clone(),
5011                expected_admission: session.session_inbox_admission().cloned(),
5012            };
5013            let mut retries = 0;
5014            let committed = loop {
5015                let result = self.input_inbox.checkpoint_actor_input(request.clone()).await;
5016                if matches!(&result, Err(bamboo_storage::ActorInputCheckpointError::PrefixConflict))
5017                    && retries < 3
5018                {
5019                    // A Parent can save ordinary tool/wait state while this
5020                    // Child checkpoints. PrefixConflict guarantees no publish;
5021                    // retry the SAME Child prefix, claim and fence, with every
5022                    // storage authority check repeated. Never adopt a changed
5023                    // Child transcript or retry an unconfirmed publication.
5024                    tokio::time::sleep(Duration::from_millis(10 << retries)).await;
5025                    retries += 1;
5026                    continue;
5027                }
5028                break result;
5029            }.map_err(|error| {
5030                tracing::warn!(%error, "Actor correction checkpoint rejected or unconfirmed; no worker dispatch or ACK");
5031                AgentError::LLM("Actor correction checkpoint unconfirmed; durable Inbox and history are preserved".into())
5032            })?;
5033            if self.recovering_pre_ack
5034                && committed.status
5035                    != bamboo_storage::ActorInputCheckpointStatus::AlreadyCheckpointed
5036            {
5037                return Err(plain_actor_unsupported());
5038            }
5039            let messages = session
5040                .messages
5041                .iter()
5042                .map(|message| {
5043                    let mut worker = message.clone();
5044                    if committed.status
5045                        == bamboo_storage::ActorInputCheckpointStatus::AlreadyCheckpointed
5046                        && worker.id == committed.envelope.id.as_str()
5047                    {
5048                        worker
5049                            .metadata
5050                            .as_mut()
5051                            .and_then(serde_json::Value::as_object_mut)
5052                            .ok_or_else(|| plain_actor_unsupported())?
5053                            .remove("_bamboo_owned_input_checkpoint");
5054                        if !bamboo_domain::is_matching_session_message(&worker, &committed.envelope)
5055                        {
5056                            return Err(plain_actor_unsupported());
5057                        }
5058                    }
5059                    serde_json::to_value(worker).map_err(|_| plain_actor_unsupported())
5060                })
5061                .collect::<Result<Vec<_>, _>>()?;
5062            *session = committed.session;
5063            Ok((
5064                messages,
5065                SessionMessageDelivery {
5066                    target_session_id: session.id.clone(),
5067                    envelope: committed.envelope,
5068                    canonical_claim_generation: committed.generation,
5069                    activation_run_id: run_id.into(),
5070                    activation_policy: claim.claim.activation_policy,
5071                },
5072            ))
5073        })
5074    }
5075    fn append_reply<'a>(
5076        &'a self,
5077        session: &'a mut Session,
5078        text: String,
5079        event_tx: &'a mpsc::Sender<AgentEvent>,
5080    ) -> std::pin::Pin<Box<impl std::future::Future<Output = Result<(), AgentError>> + Send + 'a>>
5081    {
5082        Box::pin(async move {
5083            let message = bamboo_agent_core::Message::assistant(text, None);
5084            let request = bamboo_storage::ActorTranscriptAppend {
5085                fence: self.fence.clone(),
5086                expected_created_at: self.created_at,
5087                expected_conversation_generation: self.conversation_generation.clone(),
5088                expected_messages: session.messages.clone(),
5089                expected_provider_transcript: self.provider_transcript.clone(),
5090                messages: vec![message.clone()],
5091                native_groups: Vec::new(),
5092            };
5093            let mut retries = 0;
5094            *session = loop {
5095                let result = self.store.append_actor_transcript(request.clone()).await;
5096                if matches!(
5097                    &result,
5098                    Err(bamboo_storage::ActorTranscriptAppendError::PrefixConflict)
5099                ) && retries < 3
5100                {
5101                    // Parent writes may change the ancestor witness before
5102                    // publication. Retry this exact reply and Child prefix;
5103                    // never reload/adopt a different transcript or replay an
5104                    // unconfirmed write. Storage rechecks the same live fence.
5105                    tokio::time::sleep(Duration::from_millis(10 << retries)).await;
5106                    retries += 1;
5107                    continue;
5108                }
5109                break result;
5110            }
5111            .map_err(|error| {
5112                tracing::warn!(
5113                    %error,
5114                    "Actor reply checkpoint rejected or unconfirmed; durable history is preserved"
5115                );
5116                AgentError::LLM(format!("actor reply commit failed: {error}"))
5117            })?;
5118            let _ = event_tx
5119                .send(AgentEvent::message_appended(&session.id, &message))
5120                .await;
5121            Ok(())
5122        })
5123    }
5124    // First Terminal is a real completed plain reply, not a cache event. Commit
5125    // it before appending the typed correction, then use the existing startup
5126    // admission barrier on a second correlated Run under the SAME Host fence.
5127    fn continue_input<'a>(
5128        &'a self,
5129        client: &'a mut dyn bamboo_subagent::ChildLink,
5130        binding: &'a SessionInboxRuntimeBinding,
5131        session: &'a mut Session,
5132        run_id: &'a str,
5133        run: Option<(&'a RunSpec, &'a AtomicU64)>,
5134        expected_posture: &'a mut Option<ExpectedPermissionPosture>,
5135        text: String,
5136        event_tx: &'a mpsc::Sender<AgentEvent>,
5137    ) -> std::pin::Pin<
5138        Box<
5139            impl std::future::Future<
5140                    Output = Result<Option<(SessionInboxOwnedClaim, u64)>, AgentError>,
5141                > + Send
5142                + 'a,
5143        >,
5144    > {
5145        Box::pin(async move {
5146            let Some(claim) = self.claim_input(binding, session, run_id).await? else {
5147                return Ok(None);
5148            };
5149            let (run, epochs) = run.ok_or_else(|| plain_actor_unsupported())?;
5150            self.append_reply(session, text, event_tx).await?;
5151            let (messages, delivery) = self.checkpoint_input(session, run_id, &claim).await?;
5152            // Fresh actual Host audit, unchanged permission/ceiling checks.
5153            self.refresh_continuation_posture(session, run, expected_posture)?;
5154            let epoch = epochs
5155                .fetch_add(1, Ordering::Relaxed)
5156                .checked_add(1)
5157                .ok_or_else(|| plain_actor_unsupported())?;
5158            let mut next = run.clone();
5159            next.messages = messages;
5160            next.initial_session_messages = vec![delivery];
5161            next.execution_epoch = epoch;
5162            client.send(ParentFrame::Run(next)).await.map_err(|_| {
5163                AgentError::LLM(
5164                    "Actor correction dispatch unconfirmed; durable input is preserved".into(),
5165                )
5166            })?;
5167            Ok(Some((claim, epoch)))
5168        })
5169    }
5170    fn refresh_continuation_posture(
5171        &self,
5172        session: &Session,
5173        run: &RunSpec,
5174        expected: &mut Option<ExpectedPermissionPosture>,
5175    ) -> Result<(), AgentError> {
5176        let Some(expected) = expected.as_mut() else {
5177            return Err(plain_actor_unsupported());
5178        };
5179        let audit = bamboo_domain::PermissionAuditSnapshot::from_metadata(&session.metadata)
5180            .ok_or_else(|| plain_actor_unsupported())?;
5181        if audit.policy_revision != expected.policy_revision
5182            || audit.resolution != expected.resolution
5183            || audit.executor_mapping != expected.executor_mapping
5184        {
5185            return Err(plain_actor_unsupported());
5186        }
5187        match (&self.permission_config, &run.permission_policy) {
5188            (Some(config), Some(policy)) => {
5189                ensure_no_active_scoped_session_denies(config, &session.id)?;
5190                if config.policy_revision() != policy.revision
5191                    || policy.revision != audit.policy_revision
5192                    || policy.session_id != session.id
5193                    || policy.workspace_path != session.workspace
5194                    || policy.inherit_session_grants
5195                    || serde_json::to_value(config.to_serializable())
5196                        .map_err(|_| plain_actor_unsupported())?
5197                        != policy.policy
5198                {
5199                    return Err(plain_actor_unsupported());
5200                }
5201            }
5202            (None, None) => {}
5203            _ => return Err(plain_actor_unsupported()),
5204        }
5205        expected.expected_audit_revision = Some(audit.audit_revision);
5206        Ok(())
5207    }
5208    fn release_initial_input<'a>(
5209        &'a self,
5210        binding: &'a SessionInboxRuntimeBinding,
5211        session: &'a Session,
5212        run: &'a RunSpec,
5213        epoch: u64,
5214        claim: &'a SessionInboxOwnedClaim,
5215        request: &'a bamboo_subagent::proto::InitialInputReleaseRequest,
5216        expected: &'a mut Option<ExpectedPermissionPosture>,
5217        previous: Option<&'a bamboo_subagent::proto::InitialInputRelease>,
5218    ) -> std::pin::Pin<
5219        Box<
5220            impl std::future::Future<
5221                    Output = Result<bamboo_subagent::proto::InitialInputRelease, AgentError>,
5222                > + Send
5223                + 'a,
5224        >,
5225    > {
5226        Box::pin(async move {
5227            if !self.initial_release_required {
5228                return Err(plain_actor_unsupported());
5229            }
5230            let delivery = SessionMessageDelivery {
5231                target_session_id: session.id.clone(),
5232                envelope: claim.claim.envelope.clone(),
5233                canonical_claim_generation: claim.claim.generation,
5234                activation_run_id: self.fence.run_id.clone(),
5235                activation_policy: claim.claim.activation_policy,
5236            };
5237            let mut current_run = run.clone();
5238            current_run.execution_epoch = epoch;
5239            let exact = bamboo_subagent::proto::InitialInputReleaseRequest::from_run(
5240                &current_run,
5241                &delivery,
5242                request.nonce.clone(),
5243            )
5244            .map_err(|_| plain_actor_unsupported())?;
5245            if &exact != request
5246                || request.child_id != session.id
5247                || request.created_at != session.created_at
5248                || request.spawn_depth != session.spawn_depth
5249                || request.parent_id != session.parent_session_id.as_deref().unwrap_or_default()
5250                || request.root_id != session.root_session_id
5251                || request.project_id != project_id_for_actor_run(session)?
5252                || !binding
5253                    .router
5254                    .owns_run(&session.id, &self.fence.run_id)
5255                    .await
5256            {
5257                return Err(plain_actor_unsupported());
5258            }
5259            let deadline = self
5260                .actor_deadline
5261                .min(self.input_deadline)
5262                .min(claim.lease.expires_at);
5263            if chrono::Utc::now() >= deadline {
5264                return Err(plain_actor_unsupported());
5265            }
5266            self.store
5267                .validate_fence(&self.fence, chrono::Utc::now())
5268                .await
5269                .map_err(|_| plain_actor_unsupported())?;
5270            let actual =
5271                bamboo_agent_core::storage::Storage::load_session(self.store.as_ref(), &session.id)
5272                    .await
5273                    .map_err(|_| plain_actor_unsupported())?
5274                    .ok_or_else(|| plain_actor_unsupported())?;
5275            if actual.created_at != session.created_at
5276                || actual.root_session_id != session.root_session_id
5277                || actual.parent_session_id != session.parent_session_id
5278                || actual.spawn_depth != session.spawn_depth
5279                || project_id_for_actor_run(&actual)? != request.project_id
5280                || serde_json::to_value(&actual.messages).map_err(|_| plain_actor_unsupported())?
5281                    != serde_json::to_value(&session.messages)
5282                        .map_err(|_| plain_actor_unsupported())?
5283                || serde_json::to_value(&actual.provider_transcript)
5284                    .map_err(|_| plain_actor_unsupported())?
5285                    != serde_json::to_value(&session.provider_transcript)
5286                        .map_err(|_| plain_actor_unsupported())?
5287                || serde_json::to_value(actual.session_inbox_admission())
5288                    .map_err(|_| plain_actor_unsupported())?
5289                    != serde_json::to_value(session.session_inbox_admission())
5290                        .map_err(|_| plain_actor_unsupported())?
5291            {
5292                return Err(plain_actor_unsupported());
5293            }
5294            self.refresh_continuation_posture(&actual, &current_run, expected)?;
5295            if let Some(previous) = previous {
5296                if !previous.permits(request, chrono::Utc::now())
5297                    || previous.expires_at != deadline
5298                    || !self
5299                        .input_inbox
5300                        .was_admitted(&session.id, &claim.claim.envelope.id)
5301                        .await
5302                        .map_err(|_| plain_actor_unsupported())?
5303                {
5304                    return Err(plain_actor_unsupported());
5305                }
5306            } else {
5307                // Full current-prefix AlreadyCheckpointed readback still checks
5308                // the actual physical owner. A local worker receipt is insufficient.
5309                let checkpoint = bamboo_storage::ActorInputCheckpoint {
5310                    fence: self.fence.clone(),
5311                    expected_created_at: self.created_at,
5312                    expected_conversation_generation: self.conversation_generation.clone(),
5313                    claim: claim.clone(),
5314                    expected_messages: actual.messages.clone(),
5315                    expected_provider_transcript: actual.provider_transcript.clone(),
5316                    expected_admission: actual.session_inbox_admission().cloned(),
5317                };
5318                let mut retries = 0;
5319                loop {
5320                    let result = self
5321                        .input_inbox
5322                        .checkpoint_actor_input(checkpoint.clone())
5323                        .await;
5324                    if matches!(
5325                        &result,
5326                        Err(bamboo_storage::ActorInputCheckpointError::PrefixConflict)
5327                    ) && retries < 3
5328                    {
5329                        // As at initial checkpoint, this error guarantees no
5330                        // publication. Repeat the identical prefix, claim and
5331                        // fence; a Parent's ordinary save grants no authority.
5332                        tracing::warn!(retries, "Actor initial release prefix changed before publication; retrying same claim");
5333                        tokio::time::sleep(Duration::from_millis(10 << retries)).await;
5334                        retries += 1;
5335                        continue;
5336                    }
5337                    result.map_err(|error| {
5338                        tracing::warn!(%error, "Actor initial release checkpoint rejected or unconfirmed");
5339                        plain_actor_unsupported()
5340                    })?;
5341                    break;
5342                }
5343                self.confirm_input(
5344                    binding,
5345                    &actual,
5346                    &self.fence.run_id,
5347                    claim,
5348                    &bamboo_subagent::proto::SessionMessageAdmissionConfirmation {
5349                        target_session_id: request.child_id.clone(),
5350                        envelope_id: request.envelope_id.clone(),
5351                        canonical_claim_generation: request.generation,
5352                        activation_run_id: request.activation_run_id.clone(),
5353                    },
5354                )
5355                .await?;
5356            }
5357            // ACK's terminal receipt is idempotent even after expiry. It cannot
5358            // replace these final current-owner/deadline/policy checks.
5359            if chrono::Utc::now() >= deadline
5360                || !binding
5361                    .router
5362                    .owns_run(&session.id, &self.fence.run_id)
5363                    .await
5364            {
5365                return Err(plain_actor_unsupported());
5366            }
5367            self.store
5368                .validate_fence(&self.fence, chrono::Utc::now())
5369                .await
5370                .map_err(|_| plain_actor_unsupported())?;
5371            let final_session =
5372                bamboo_agent_core::storage::Storage::load_session(self.store.as_ref(), &session.id)
5373                    .await
5374                    .map_err(|_| plain_actor_unsupported())?
5375                    .ok_or_else(|| plain_actor_unsupported())?;
5376            self.store
5377                .validate_fence(&self.fence, chrono::Utc::now())
5378                .await
5379                .map_err(|_| plain_actor_unsupported())?;
5380            self.refresh_continuation_posture(&final_session, &current_run, expected)?;
5381            if chrono::Utc::now() >= deadline {
5382                return Err(plain_actor_unsupported());
5383            }
5384            Ok(bamboo_subagent::proto::InitialInputRelease {
5385                request: request.clone(),
5386                expires_at: deadline,
5387            })
5388        })
5389    }
5390    async fn confirm_input(
5391        &self,
5392        binding: &SessionInboxRuntimeBinding,
5393        session: &Session,
5394        run_id: &str,
5395        claim: &SessionInboxOwnedClaim,
5396        confirmation: &bamboo_subagent::proto::SessionMessageAdmissionConfirmation,
5397    ) -> Result<(), AgentError> {
5398        if confirmation.target_session_id != session.id
5399            || confirmation.envelope_id != claim.claim.envelope.id.as_str()
5400            || confirmation.canonical_claim_generation != claim.claim.generation
5401            || confirmation.activation_run_id != run_id
5402            || run_id != self.fence.run_id
5403            || !binding.router.owns_run(&session.id, run_id).await
5404        {
5405            return Err(plain_actor_unsupported());
5406        }
5407        self.input_inbox
5408            .ack_owned(&session.id, claim, chrono::Utc::now())
5409            .await
5410            .map_err(|error| {
5411                tracing::warn!(%error, "Actor correction ACK unresolved");
5412                AgentError::LLM(
5413                    "Actor correction ACK unresolved; durable input is preserved".into(),
5414                )
5415            })
5416    }
5417    async fn finish(&self, outcome: ActorActivationFinish) -> Result<(), AgentError> {
5418        self.finish_with_observer(outcome, None).await
5419    }
5420
5421    async fn finish_with_observer(
5422        &self,
5423        outcome: ActorActivationFinish,
5424        observer: Option<&dyn ActorEventObserver>,
5425    ) -> Result<(), AgentError> {
5426        let committed = self
5427            .store
5428            .finish_activation(&self.fence, chrono::Utc::now(), outcome)
5429            .await
5430            .map_err(|error| AgentError::LLM(format!("Actor completion unconfirmed: {error}")))?;
5431        // A validated final Terminal hands off this one router. Failed writes
5432        // never publish, and neither a retry nor another finish can reuse it.
5433        let router = self.completion_router.lock().recover_poison().take();
5434        if let (Some(observer), Some(router)) = (observer, router) {
5435            // The Directory may already hold a successor by this point. Never
5436            // relabel the old completion, and never turn an already committed
5437            // success into an execution failure if best-effort delivery is lost.
5438            match self.store.inspect_actor(&self.fence.actor_id).await {
5439                Ok(entry) => match router.into_host_completion(&entry, &committed) {
5440                    Ok(event) => observer.publish(event),
5441                    Err(_) => tracing::debug!(actor_id = %self.fence.actor_id,
5442                        "Actor completion notification superseded"),
5443                },
5444                Err(error) => tracing::warn!(actor_id = %self.fence.actor_id, %error,
5445                    "Actor completion committed; notification snapshot unavailable"),
5446            }
5447        }
5448        Ok(())
5449    }
5450}
5451fn local_tool_history_unsupported() -> AgentError {
5452    AgentError::LLM("local_tool_history_unsupported".into())
5453}
5454
5455#[derive(Clone, Copy)]
5456enum LocalToolTerminal<'a> {
5457    Completed(Option<&'a str>),
5458    // Only the selected Cancelled terminal may finish a validated Read/Glob
5459    // prefix. This is Host control flow, never a Worker DATA authority flag.
5460    Cancelled,
5461}
5462
5463impl LocalToolTerminal<'_> {
5464    fn permits_checkpoint(self, cancel: &CancellationToken) -> bool {
5465        match self {
5466            Self::Completed(_) => !cancel.is_cancelled(),
5467            Self::Cancelled => cancel.is_cancelled(),
5468        }
5469    }
5470}
5471
5472#[derive(Default)]
5473struct LocalToolCollector {
5474    messages: Option<Vec<bamboo_agent_core::Message>>,
5475    starts: HashMap<String, (String, serde_json::Value)>,
5476    outcomes: HashMap<String, (bool, String)>,
5477    // Installed only by the Host's exact Task callback receipt. Worker event
5478    // JSON and transcript metadata cannot grant this terminal exception.
5479    ticket_yield: Option<(String, bamboo_tickets::PendingRequest)>,
5480}
5481
5482impl LocalToolCollector {
5483    fn ticket_event(
5484        &self,
5485        value: &serde_json::Value,
5486        committed: &HashMap<String, serde_json::Value>,
5487    ) -> Result<bool, AgentError> {
5488        if value["type"] == bamboo_subagent::proto::LocalToolMessages::TYPE {
5489            return Ok(true);
5490        }
5491        match serde_json::from_value::<AgentEvent>(value.clone())
5492            .map_err(|_| local_tool_history_unsupported())?
5493        {
5494            AgentEvent::ToolStart { tool_name, .. }
5495                if !matches!(tool_name.as_str(), "Task" | "Read" | "Write") =>
5496            {
5497                Err(local_tool_history_unsupported())
5498            }
5499            AgentEvent::ToolComplete {
5500                tool_call_id,
5501                result,
5502            } if result.success => {
5503                let Some((name, args)) = self.starts.get(&tool_call_id) else {
5504                    return Err(local_tool_history_unsupported());
5505                };
5506                let valid = if name == "Task" {
5507                    committed.get(&tool_call_id) == Some(args)
5508                } else {
5509                    committed.get(&tool_call_id).is_some_and(|proof| {
5510                        proof["file_tool"] == *name
5511                            && proof["arguments"] == *args
5512                            && proof["result"] == result.result
5513                    })
5514                };
5515                if !valid {
5516                    return Err(AgentError::LLM(
5517                        "LocalPlan success has no exact Host receipt".into(),
5518                    ));
5519                }
5520                Ok(true)
5521            }
5522            // Only the committed Host projection may publish a plan update.
5523            AgentEvent::TaskListUpdated { .. } => Ok(false),
5524            _ => Ok(true),
5525        }
5526    }
5527
5528    fn event(&mut self, value: &serde_json::Value) -> Result<bool, AgentError> {
5529        use bamboo_subagent::proto::LocalToolMessages;
5530        if self.messages.is_some() {
5531            return Err(local_tool_history_unsupported());
5532        }
5533        if value["type"] == LocalToolMessages::TYPE {
5534            let observation: LocalToolMessages = serde_json::from_value(value.clone())
5535                .map_err(|_| local_tool_history_unsupported())?;
5536            self.messages = Some(
5537                observation
5538                    .validate()
5539                    .map_err(|_| local_tool_history_unsupported())?,
5540            );
5541            return Ok(false);
5542        }
5543        let event: AgentEvent =
5544            serde_json::from_value(value.clone()).map_err(|_| local_tool_history_unsupported())?;
5545        match event {
5546            AgentEvent::ToolStart {
5547                tool_call_id,
5548                tool_name,
5549                arguments,
5550            } => {
5551                if !arguments.is_object()
5552                    || !matches!(tool_name.as_str(), "Read" | "Glob" | "Write" | "Task")
5553                    || tool_call_id.is_empty()
5554                    || tool_call_id.len() > 128
5555                    || self.starts.len() >= LocalToolMessages::MAX_PAIRS
5556                    || self.outcomes.contains_key(&tool_call_id)
5557                    || self
5558                        .starts
5559                        .insert(tool_call_id, (tool_name, arguments))
5560                        .is_some()
5561                {
5562                    return Err(local_tool_history_unsupported());
5563                }
5564            }
5565            AgentEvent::ToolComplete {
5566                tool_call_id,
5567                result,
5568            } => {
5569                if !result.images.is_empty() {
5570                    return Err(local_tool_history_unsupported());
5571                }
5572                self.outcome(tool_call_id, result.success, result.result)?;
5573            }
5574            AgentEvent::ToolError {
5575                tool_call_id,
5576                error,
5577            } => {
5578                self.outcome(tool_call_id, false, format!("Error: {error}"))?;
5579            }
5580            AgentEvent::ToolToken { tool_call_id, .. } => {
5581                if !self.starts.contains_key(&tool_call_id)
5582                    || self.outcomes.contains_key(&tool_call_id)
5583                {
5584                    return Err(local_tool_history_unsupported());
5585                }
5586            }
5587            AgentEvent::ToolLifecycle {
5588                tool_call_id,
5589                tool_name,
5590                is_mutating,
5591                auto_approved,
5592                phase,
5593                ..
5594            } => {
5595                if self.starts.get(&tool_call_id).map(|(name, _)| name) != Some(&tool_name)
5596                    || is_mutating != matches!(tool_name.as_str(), "Write" | "Task")
5597                    || auto_approved != !matches!(tool_name.as_str(), "Write" | "Task")
5598                    || !matches!(phase.as_str(), "begin" | "finished" | "error" | "cancelled")
5599                {
5600                    return Err(local_tool_history_unsupported());
5601                }
5602            }
5603            AgentEvent::ReasoningToken { .. } => {
5604                // Text reasoning is a stream observation. Its complete value
5605                // is preserved by the bounded typed completion, which still
5606                // rejects signatures and unsupported provider state.
5607            }
5608            AgentEvent::ContextSummarized { .. }
5609            | AgentEvent::ContextArchived { .. }
5610            | AgentEvent::ContextCompressionStatus { .. }
5611            | AgentEvent::NeedClarification { .. }
5612            | AgentEvent::SubAgentStarted { .. }
5613            | AgentEvent::SubAgentEvent { .. } => return Err(local_tool_history_unsupported()),
5614            // Worker commit events are observations, not canonical Host commits.
5615            AgentEvent::SessionHistoryCommitted { .. } | AgentEvent::MessageAppended { .. } => {
5616                return Ok(false)
5617            }
5618            _ => {}
5619        }
5620        Ok(true)
5621    }
5622
5623    fn outcome(&mut self, id: String, success: bool, text: String) -> Result<(), AgentError> {
5624        if id.is_empty()
5625            || id.len() > 128
5626            || text.len() > bamboo_subagent::proto::LocalToolMessages::MAX_BYTES
5627            || self.outcomes.len() >= bamboo_subagent::proto::LocalToolMessages::MAX_PAIRS
5628            || self.outcomes.insert(id, (success, text)).is_some()
5629        {
5630            return Err(local_tool_history_unsupported());
5631        }
5632        Ok(())
5633    }
5634
5635    fn suffix(
5636        self,
5637        host: &Session,
5638        tools: &[String],
5639        read_only: bool,
5640        terminal: Option<&str>,
5641    ) -> Result<Vec<bamboo_agent_core::Message>, AgentError> {
5642        self.suffix_for_terminal(
5643            host,
5644            tools,
5645            read_only,
5646            LocalToolTerminal::Completed(terminal),
5647        )
5648    }
5649
5650    fn suffix_for_terminal(
5651        self,
5652        host: &Session,
5653        tools: &[String],
5654        read_only: bool,
5655        terminal: LocalToolTerminal<'_>,
5656    ) -> Result<Vec<bamboo_agent_core::Message>, AgentError> {
5657        use bamboo_subagent::proto::LocalToolMessages;
5658        let cancelled = matches!(terminal, LocalToolTerminal::Cancelled);
5659        if cancelled
5660            && (!read_only
5661                || tools.is_empty()
5662                || tools
5663                    .iter()
5664                    .any(|name| !matches!(name.as_str(), "Read" | "Glob"))
5665                || self.ticket_yield.is_some())
5666        {
5667            return Err(local_tool_history_unsupported());
5668        }
5669        let data = self.messages.ok_or_else(local_tool_history_unsupported)?;
5670        let candidate: Vec<_> = data.iter().skip_while(|m| m.role == Role::System).collect();
5671        let prefix: Vec<_> = host
5672            .messages
5673            .iter()
5674            .filter(|m| m.role != Role::System)
5675            .collect();
5676        if !host.provider_transcript.groups().is_empty()
5677            || candidate.len() < prefix.len()
5678            || (!cancelled && candidate.len() == prefix.len())
5679            || serde_json::to_value(&candidate[..prefix.len()])
5680                .map_err(|_| local_tool_history_unsupported())?
5681                != serde_json::to_value(&prefix).map_err(|_| local_tool_history_unsupported())?
5682        {
5683            return Err(local_tool_history_unsupported());
5684        }
5685        let mut ids = std::collections::HashSet::new();
5686        let mut calls = std::collections::HashSet::new();
5687        for message in &host.messages {
5688            if !ids.insert(message.id.clone()) {
5689                return Err(local_tool_history_unsupported());
5690            }
5691            for call in message.tool_calls.iter().flatten() {
5692                calls.insert(call.id.clone());
5693            }
5694            if let Some(id) = &message.tool_call_id {
5695                calls.insert(id.clone());
5696            }
5697        }
5698        let suffix: Vec<_> = candidate[prefix.len()..]
5699            .iter()
5700            .map(|message| (**message).clone())
5701            .collect();
5702        let mut pending = HashMap::new();
5703        let mut used = std::collections::HashSet::new();
5704        for message in &suffix {
5705            if !ids.insert(message.id.clone()) || message.never_compress {
5706                return Err(local_tool_history_unsupported());
5707            }
5708            match message.role {
5709                Role::Assistant => {
5710                    if !pending.is_empty()
5711                        || message.metadata.is_some()
5712                        || message.tool_call_id.is_some()
5713                        || message.tool_success.is_some()
5714                    {
5715                        return Err(local_tool_history_unsupported());
5716                    }
5717                    if let Some(batch) = &message.tool_calls {
5718                        if batch.is_empty() {
5719                            return Err(local_tool_history_unsupported());
5720                        }
5721                        for call in batch {
5722                            let arguments =
5723                                serde_json::from_str::<serde_json::Value>(&call.function.arguments)
5724                                    .ok()
5725                                    .filter(|arguments| arguments.is_object())
5726                                    .ok_or_else(local_tool_history_unsupported)?;
5727                            if call.id.is_empty()
5728                                || call.id.len() > 128
5729                                || call.tool_type != "function"
5730                                || !calls.insert(call.id.clone())
5731                                || calls.len() > LocalToolMessages::MAX_MESSAGES
5732                                || !matches!(
5733                                    call.function.name.as_str(),
5734                                    "Read" | "Glob" | "Write" | "Task"
5735                                )
5736                                || (cancelled
5737                                    && !matches!(call.function.name.as_str(), "Read" | "Glob"))
5738                                || self.starts.get(&call.id).is_some_and(|(name, actual)| {
5739                                    name != &call.function.name || actual != &arguments
5740                                })
5741                            {
5742                                return Err(local_tool_history_unsupported());
5743                            }
5744                            pending.insert(call.id.clone(), call.function.name.clone());
5745                        }
5746                    } else if message.content.trim().is_empty() {
5747                        return Err(local_tool_history_unsupported());
5748                    }
5749                }
5750                Role::Tool => {
5751                    let id = message
5752                        .tool_call_id
5753                        .as_ref()
5754                        .ok_or_else(local_tool_history_unsupported)?;
5755                    let name = pending
5756                        .remove(id)
5757                        .ok_or_else(local_tool_history_unsupported)?;
5758                    let success = message
5759                        .tool_success
5760                        .ok_or_else(local_tool_history_unsupported)?;
5761                    if message.tool_calls.is_some()
5762                        || message.phase.is_some()
5763                        || self.outcomes.get(id) != Some(&(success, message.content.clone()))
5764                        || !used.insert(id.clone())
5765                        || used.len() > LocalToolMessages::MAX_PAIRS
5766                    {
5767                        return Err(local_tool_history_unsupported());
5768                    }
5769                    let denied_write = name == "Write" && (read_only || !tools.contains(&name));
5770                    if denied_write {
5771                        if success
5772                            || !(message.content.contains("native_tool_ceiling_denied")
5773                                || message
5774                                    .content
5775                                    .contains("not callable at the current conversation position"))
5776                        {
5777                            return Err(local_tool_history_unsupported());
5778                        }
5779                    } else if !tools.contains(&name)
5780                        || self.starts.get(id).map(|(actual, _)| actual) != Some(&name)
5781                    {
5782                        return Err(local_tool_history_unsupported());
5783                    }
5784                    if let Some(metadata) = &message.metadata {
5785                        let object = metadata
5786                            .as_object()
5787                            .ok_or_else(local_tool_history_unsupported)?;
5788                        if object.len() != 5
5789                            || metadata["elapsed_ms"].as_u64().is_none()
5790                            || metadata["tool_name"].as_str() != Some(name.as_str())
5791                            || metadata["is_mutating"].as_bool()
5792                                != Some(matches!(name.as_str(), "Write" | "Task"))
5793                            || metadata["auto_approved"].as_bool()
5794                                != Some(!matches!(name.as_str(), "Write" | "Task"))
5795                            || metadata["success"].as_bool() != Some(success)
5796                        {
5797                            return Err(local_tool_history_unsupported());
5798                        }
5799                    }
5800                }
5801                _ => return Err(local_tool_history_unsupported()),
5802            }
5803        }
5804        let last = suffix.last();
5805        let completed_text = match terminal {
5806            LocalToolTerminal::Completed(text) => text,
5807            LocalToolTerminal::Cancelled => None,
5808        };
5809        let question_terminal =
5810            self.ticket_yield.as_ref().is_some_and(|(id, question)| {
5811                last.is_some_and(|last| last.role == Role::Tool
5812                && last.tool_call_id.as_ref() == Some(id)
5813                && last.tool_success == Some(true)
5814                && self.starts.get(id).is_some_and(|(name, _)| name == "Task")
5815                && completed_text.is_none_or(str::is_empty)
5816                && serde_json::from_str::<serde_json::Value>(&last.content).ok()
5817                    == Some(serde_json::json!({"status":"waiting_for_answer", "request":question})))
5818            });
5819        let report_terminal = last.is_some_and(|last| {
5820            last.role == Role::Assistant
5821                && last.tool_calls.is_none()
5822                && last.phase != Some(bamboo_domain::MessagePhase::Commentary)
5823                && completed_text == Some(last.content.as_str())
5824        });
5825        if !pending.is_empty()
5826            || used.len() != self.outcomes.len()
5827            || self.starts.keys().any(|id| !used.contains(id))
5828            || !(cancelled || report_terminal || question_terminal)
5829        {
5830            return Err(local_tool_history_unsupported());
5831        }
5832        Ok(suffix)
5833    }
5834}
5835
5836async fn validate_local_tool_checkpoint(
5837    session: &Session,
5838    binding: &SessionInboxRuntimeBinding,
5839    run_id: &str,
5840    terminal: LocalToolTerminal<'_>,
5841    cancel: &CancellationToken,
5842) -> Result<(), AgentError> {
5843    if !terminal.permits_checkpoint(cancel) || !binding.router.owns_run(&session.id, run_id).await {
5844        return Err(local_tool_history_unsupported());
5845    }
5846    if matches!(terminal, LocalToolTerminal::Cancelled) {
5847        let backlog = binding
5848            .inbox
5849            .inspect(&session.id)
5850            .await
5851            .map_err(|_| local_tool_history_unsupported())?;
5852        if backlog.pending != 0 || backlog.claimed != 0 || backlog.activation_pending() {
5853            // An unclaimed correction can race its delivery notification.
5854            // Inspect only: neither admission nor ACK belongs to cancellation.
5855            return Err(local_tool_history_unsupported());
5856        }
5857    }
5858    Ok(())
5859}
5860
5861async fn commit_local_tool_history(
5862    collector: LocalToolCollector,
5863    session: &mut Session,
5864    binding: &SessionInboxRuntimeBinding,
5865    run_id: &str,
5866    selection: (&[String], bool),
5867    terminal: LocalToolTerminal<'_>,
5868    cancel: &CancellationToken,
5869) -> Result<(), AgentError> {
5870    validate_local_tool_checkpoint(session, binding, run_id, terminal, cancel).await?;
5871    let latest = binding
5872        .storage
5873        .load_session(&session.id)
5874        .await
5875        .map_err(|_| local_tool_history_unsupported())?
5876        .ok_or_else(local_tool_history_unsupported)?;
5877    if latest.id != session.id
5878        || latest.kind != session.kind
5879        || latest.created_at != session.created_at
5880        || latest.root_session_id != session.root_session_id
5881        || latest.parent_session_id != session.parent_session_id
5882        || latest.spawn_depth != session.spawn_depth
5883        || latest.project_id_meta() != session.project_id_meta()
5884        || session.check_conversation_generation(&latest).is_err()
5885        || serde_json::to_value(&latest.provider_transcript)
5886            .map_err(|_| local_tool_history_unsupported())?
5887            != serde_json::to_value(&session.provider_transcript)
5888                .map_err(|_| local_tool_history_unsupported())?
5889    {
5890        return Err(local_tool_history_unsupported());
5891    }
5892    let suffix = match terminal {
5893        LocalToolTerminal::Completed(text) => {
5894            collector.suffix(&latest, selection.0, selection.1, text)?
5895        }
5896        LocalToolTerminal::Cancelled => {
5897            collector.suffix_for_terminal(&latest, selection.0, selection.1, terminal)?
5898        }
5899    };
5900    let before = session.clone();
5901    let mut staged = session.clone();
5902    staged.messages = latest.messages.clone();
5903    staged.messages.extend(suffix.iter().cloned());
5904    staged.updated_at = chrono::Utc::now();
5905    let cancelled_messages = if matches!(terminal, LocalToolTerminal::Cancelled) {
5906        Some(serde_json::to_value(&staged.messages).map_err(|_| local_tool_history_unsupported())?)
5907    } else {
5908        None
5909    };
5910    validate_local_tool_checkpoint(session, binding, run_id, terminal, cancel).await?;
5911    binding
5912        .persistence
5913        .checkpoint_runtime_session(&mut staged)
5914        .await
5915        .map_err(|_| local_tool_history_unsupported())?;
5916    let saved = binding
5917        .storage
5918        .load_session(&session.id)
5919        .await
5920        .map_err(|_| local_tool_history_unsupported())?
5921        .ok_or_else(local_tool_history_unsupported)?;
5922    if session.check_conversation_generation(&saved).is_err()
5923        || serde_json::to_value(&saved.provider_transcript)
5924            .map_err(|_| local_tool_history_unsupported())?
5925            != serde_json::to_value(&latest.provider_transcript)
5926                .map_err(|_| local_tool_history_unsupported())?
5927    {
5928        return Err(local_tool_history_unsupported());
5929    }
5930    let start = match suffix.first() {
5931        Some(first) => saved
5932            .messages
5933            .iter()
5934            .position(|message| message.id == first.id)
5935            .ok_or_else(local_tool_history_unsupported)?,
5936        None => latest.messages.len(),
5937    };
5938    if cancelled_messages.as_ref().is_some_and(|expected| {
5939        serde_json::to_value(&saved.messages).ok().as_ref() != Some(expected)
5940    }) {
5941        // A new canonical input during cancellation has not been applied by
5942        // this Worker. Preserve it, but do not certify it with the old terminal.
5943        return Err(local_tool_history_unsupported());
5944    }
5945    if saved.messages.len() < latest.messages.len()
5946        || serde_json::to_value(&saved.messages[..latest.messages.len()])
5947            .map_err(|_| local_tool_history_unsupported())?
5948            != serde_json::to_value(&latest.messages)
5949                .map_err(|_| local_tool_history_unsupported())?
5950        || serde_json::to_value(saved.messages.get(start..start + suffix.len()))
5951            .map_err(|_| local_tool_history_unsupported())?
5952            != serde_json::to_value(Some(&suffix[..]))
5953                .map_err(|_| local_tool_history_unsupported())?
5954    {
5955        return Err(local_tool_history_unsupported());
5956    }
5957    // Append-safe reconciliation may add a concurrent Host User; it may never
5958    // rewrite the validated Host prefix or any worker row being committed.
5959    let expected = latest.messages.iter().chain(&suffix);
5960    for message in expected {
5961        let matches: Vec<_> = saved
5962            .messages
5963            .iter()
5964            .filter(|current| current.id == message.id)
5965            .collect();
5966        if matches.len() != 1
5967            || serde_json::to_value(matches[0]).map_err(|_| local_tool_history_unsupported())?
5968                != serde_json::to_value(message).map_err(|_| local_tool_history_unsupported())?
5969        {
5970            *session = before;
5971            return Err(local_tool_history_unsupported());
5972        }
5973    }
5974    validate_local_tool_checkpoint(session, binding, run_id, terminal, cancel).await?;
5975    staged.messages = saved.messages;
5976    *session = staged;
5977    Ok(())
5978}
5979
5980#[derive(Default)]
5981struct ReadOnlyActorCollector {
5982    start: Option<(String, serde_json::Value)>,
5983    result: Option<String>,
5984    messages: Option<Vec<bamboo_agent_core::Message>>,
5985}
5986impl ReadOnlyActorCollector {
5987    fn event(&mut self, value: &serde_json::Value) -> Result<bool, AgentError> {
5988        if self.messages.is_some() {
5989            return Err(plain_actor_unsupported());
5990        }
5991        if value["type"] == "owned_readonly_transcript" {
5992            let tail: bamboo_subagent::proto::ReadOnlyActorTranscript =
5993                serde_json::from_value(value.clone()).map_err(|_| plain_actor_unsupported())?;
5994            tail.validate().map_err(|_| plain_actor_unsupported())?;
5995            let bamboo_subagent::proto::ReadOnlyActorTranscript::Complete { messages } = tail;
5996            let typed: Vec<bamboo_agent_core::Message> = messages
5997                .iter()
5998                .map(|raw| {
5999                    let message: bamboo_agent_core::Message =
6000                        serde_json::from_value(raw.clone())
6001                            .map_err(|_| plain_actor_unsupported())?;
6002                    if serde_json::to_value(&message).map_err(|_| plain_actor_unsupported())?
6003                        != *raw
6004                    {
6005                        return Err(plain_actor_unsupported());
6006                    }
6007                    Ok(message)
6008                })
6009                .collect::<Result<_, _>>()?;
6010            let (id, args) = self
6011                .start
6012                .as_ref()
6013                .ok_or_else(|| plain_actor_unsupported())?;
6014            let call = typed[0]
6015                .tool_calls
6016                .as_ref()
6017                .and_then(|calls| {
6018                    if calls.len() == 1 {
6019                        calls.first()
6020                    } else {
6021                        None
6022                    }
6023                })
6024                .ok_or_else(|| plain_actor_unsupported())?;
6025            if call.id != *id
6026                || call.function.name != "Glob"
6027                || serde_json::from_str::<serde_json::Value>(&call.function.arguments)
6028                    .map_err(|_| plain_actor_unsupported())?
6029                    != *args
6030                || typed[1].tool_call_id.as_ref() != Some(id)
6031                || typed[1].tool_success != Some(true)
6032                || self.result.as_ref() != Some(&typed[1].content)
6033            {
6034                return Err(plain_actor_unsupported());
6035            }
6036            self.messages = Some(typed);
6037            return Ok(false);
6038        }
6039        let event: AgentEvent =
6040            serde_json::from_value(value.clone()).map_err(|_| plain_actor_unsupported())?;
6041        match event {
6042            AgentEvent::ToolStart {
6043                tool_call_id,
6044                tool_name,
6045                arguments,
6046            } => {
6047                if tool_name != "Glob"
6048                    || tool_call_id.trim().is_empty()
6049                    || self.start.is_some()
6050                    || !arguments.is_object()
6051                {
6052                    return Err(plain_actor_unsupported());
6053                }
6054                self.start = Some((tool_call_id, arguments));
6055                Ok(true)
6056            }
6057            AgentEvent::ToolComplete {
6058                tool_call_id,
6059                result,
6060            } => {
6061                if self.start.as_ref().map(|s| &s.0) != Some(&tool_call_id)
6062                    || self.result.is_some()
6063                    || !result.success
6064                    || !result.images.is_empty()
6065                {
6066                    return Err(plain_actor_unsupported());
6067                }
6068                self.result = Some(result.result);
6069                Ok(true)
6070            }
6071            AgentEvent::ToolToken { tool_call_id, .. } => {
6072                if self.start.as_ref().map(|s| &s.0) != Some(&tool_call_id) || self.result.is_some()
6073                {
6074                    return Err(plain_actor_unsupported());
6075                }
6076                Ok(true)
6077            }
6078            AgentEvent::ToolLifecycle {
6079                tool_call_id,
6080                tool_name,
6081                is_mutating,
6082                phase,
6083                error,
6084                ..
6085            } => {
6086                if tool_name != "Glob"
6087                    || is_mutating
6088                    || !matches!(phase.as_str(), "begin" | "finished")
6089                    || error.is_some()
6090                    || self.start.as_ref().map(|s| &s.0) != Some(&tool_call_id)
6091                {
6092                    return Err(plain_actor_unsupported());
6093                }
6094                Ok(true)
6095            }
6096            _ => plain_actor_event(value),
6097        }
6098    }
6099    fn finish(self, terminal: Option<&str>) -> Result<Vec<bamboo_agent_core::Message>, AgentError> {
6100        let messages = self.messages.ok_or_else(|| plain_actor_unsupported())?;
6101        if terminal != Some(messages[2].content.as_str()) || messages[2].content.trim().is_empty() {
6102            return Err(plain_actor_unsupported());
6103        }
6104        Ok(messages)
6105    }
6106}
6107
6108fn plain_actor_event(value: &serde_json::Value) -> Result<bool, AgentError> {
6109    let event: AgentEvent =
6110        serde_json::from_value(value.clone()).map_err(|_| plain_actor_unsupported())?;
6111    match event {
6112        AgentEvent::ReasoningToken { .. }
6113        | AgentEvent::ToolToken { .. }
6114        | AgentEvent::ToolStart { .. }
6115        | AgentEvent::ToolComplete { .. }
6116        | AgentEvent::ToolError { .. }
6117        | AgentEvent::ToolLifecycle { .. }
6118        | AgentEvent::ContextSummarized { .. }
6119        | AgentEvent::ContextArchived { .. }
6120        | AgentEvent::ContextCompressionStatus { .. }
6121        | AgentEvent::NeedClarification { .. }
6122        | AgentEvent::SubAgentStarted { .. }
6123        | AgentEvent::SubAgentEvent { .. }
6124        | AgentEvent::ChildApprovalRequested { .. } => Err(plain_actor_unsupported()),
6125        // Worker cache commits are not host Store commits. Publish only the
6126        // actual guarded host append after it returns a committed snapshot.
6127        AgentEvent::MessageAppended { .. }
6128        | AgentEvent::SessionHistoryCommitted { .. }
6129        | AgentEvent::VisibleMessageStart { .. }
6130        | AgentEvent::VisibleMessageDiscard { .. } => Ok(false),
6131        event if event.is_durable_change() => Ok(false),
6132        _ => Ok(true),
6133    }
6134}
6135
6136/// The `{kind,host}` placement descriptor stamped onto a child session's metadata
6137/// under `"placement"` — read back by the storage index → `SessionSummary.placement`
6138/// → the UI's machine badge. `None` for `Local` (those fall through to the DTO's
6139/// default of this backend's own host). The value is a JSON string matching
6140/// `bamboo_storage::SessionPlacement { kind, host }`.
6141fn placement_metadata(placement: &Placement, host_label: Option<&str>) -> Option<String> {
6142    // Prefer the cluster node's own label/host (its metadata) when the placement
6143    // maps to a node; else fall back to the raw endpoint host / pool name.
6144    let value = match placement {
6145        Placement::Local => return None,
6146        Placement::Remote { endpoint } => serde_json::json!({
6147            "kind": "remote",
6148            "host": host_label.map(str::to_string).unwrap_or_else(|| host_of_endpoint(endpoint)),
6149        }),
6150        Placement::Schedulable { pool } => serde_json::json!({
6151            "kind": "remote",
6152            "host": host_label.unwrap_or(pool),
6153        }),
6154    };
6155    serde_json::to_string(&value).ok()
6156}
6157
6158/// Extract the host from a `ws[s]://host:port[/path]` bus endpoint, for display.
6159fn host_of_endpoint(endpoint: &str) -> String {
6160    endpoint
6161        .trim()
6162        .trim_start_matches("wss://")
6163        .trim_start_matches("ws://")
6164        .split(['/', ':'])
6165        .next()
6166        .unwrap_or(endpoint)
6167        .to_string()
6168}
6169
6170async fn reconcile_already_admitted_claim(
6171    binding: &SessionInboxRuntimeBinding,
6172    session: &mut Session,
6173    claim: &SessionInboxClaim,
6174) -> crate::runtime::runner::Result<()> {
6175    let latest = binding
6176        .storage
6177        .load_session(&session.id)
6178        .await
6179        .map_err(|error| {
6180            AgentError::LLM(format!(
6181                "load canonical SessionInbox checkpoint for {}: {error}",
6182                session.id
6183            ))
6184        })?
6185        .ok_or_else(|| {
6186            AgentError::LLM(format!(
6187                "canonical SessionInbox target disappeared: {}",
6188                session.id
6189            ))
6190        })?;
6191    session
6192        .check_conversation_generation(&latest)
6193        .map_err(|error| {
6194            AgentError::LLM(format!(
6195                "canonical SessionInbox conversation changed: {error}"
6196            ))
6197        })?;
6198    let Some(message) = latest
6199        .messages
6200        .iter()
6201        .find(|message| bamboo_domain::is_matching_session_message(message, &claim.envelope))
6202        .cloned()
6203    else {
6204        return Err(AgentError::LLM(format!(
6205            "canonical admitted receipt for {} exists without transcript message {}",
6206            session.id, claim.envelope.id
6207        )));
6208    };
6209    if let Some(existing) = session
6210        .messages
6211        .iter_mut()
6212        .find(|existing| existing.id == message.id)
6213    {
6214        *existing = message;
6215    } else {
6216        session.add_message(message);
6217    }
6218    bamboo_domain::merge_session_inbox_admission(session, &latest);
6219    binding
6220        .inbox
6221        .ack(&session.id, claim)
6222        .await
6223        .map_err(|error| {
6224            AgentError::LLM(format!(
6225                "ack recovered canonical SessionInbox claim {}: {error}",
6226                claim.envelope.id
6227            ))
6228        })
6229}
6230
6231/// Checkpoint a worker-confirmed envelope into the canonical logical Session,
6232/// then create the permanent host receipt/remove its exact claim. This order is
6233/// the actor-side equivalent of the local state_bridge crash boundary.
6234async fn checkpoint_and_ack_canonical_claim(
6235    binding: &SessionInboxRuntimeBinding,
6236    session: &mut Session,
6237    claim: &SessionInboxClaim,
6238    fenced: Option<(&bamboo_storage::SessionStoreV2, &ActorActivationFence)>,
6239) -> crate::runtime::runner::Result<()> {
6240    if claim.envelope.target_session_id != session.id {
6241        return Err(AgentError::LLM(format!(
6242            "canonical SessionInbox claim target {} does not match active logical session {}",
6243            claim.envelope.target_session_id, session.id
6244        )));
6245    }
6246    if binding
6247        .inbox
6248        .was_admitted(&session.id, &claim.envelope.id)
6249        .await
6250        .map_err(|error| {
6251            AgentError::LLM(format!(
6252                "inspect canonical admitted receipt {}: {error}",
6253                claim.envelope.id
6254            ))
6255        })?
6256    {
6257        return reconcile_already_admitted_claim(binding, session, claim).await;
6258    }
6259
6260    if let Some((store, fence)) = fenced {
6261        let committed = store
6262            .confirm_actor_claim_context(
6263                concrete_file_inbox(binding)?,
6264                bamboo_storage::ActorClaimContextSeed {
6265                    fence: fence.clone(),
6266                    expected_created_at: session.created_at,
6267                    expected_conversation_generation: session.conversation_generation_observation(),
6268                    expected_messages: session.messages.clone(),
6269                    expected_provider_transcript: session.provider_transcript.clone(),
6270                    expected_admission: session.session_inbox_admission().cloned(),
6271                    claims: vec![claim.clone()],
6272                },
6273            )
6274            .await;
6275        let committed = match committed {
6276            Ok(committed) => committed,
6277            Err(error) => {
6278                // An error after atomic replacement may still have advanced
6279                // Main. Keep the claim unacked and use that canonical context
6280                // for the SDK's terminal status save.
6281                adopt_latest_actor_context(store, session).await;
6282                return Err(AgentError::LLM(format!(
6283                    "confirm fenced Actor claim context unconfirmed: {error}"
6284                )));
6285            }
6286        };
6287        bamboo_domain::merge_session_inbox_admission(session, &committed);
6288        session.messages = committed.messages;
6289        session.provider_transcript = committed.provider_transcript;
6290        session.updated_at = committed.updated_at;
6291        if !session
6292            .messages
6293            .iter()
6294            .any(|message| bamboo_domain::is_matching_session_message(message, &claim.envelope))
6295            || !session
6296                .session_inbox_admission()
6297                .is_some_and(|cursor| cursor.contains(&claim.envelope.id))
6298        {
6299            return Err(AgentError::LLM(
6300                "fenced Actor claim confirmation lost durable proof".into(),
6301            ));
6302        }
6303        return binding
6304            .inbox
6305            .ack(&session.id, claim)
6306            .await
6307            .map_err(|error| {
6308                AgentError::LLM(format!(
6309                    "ack fenced Actor claim {} after confirmation: {error}",
6310                    claim.envelope.id
6311                ))
6312            });
6313    }
6314
6315    let transcript_has_id = session
6316        .messages
6317        .iter()
6318        .any(|message| bamboo_domain::is_matching_session_message(message, &claim.envelope));
6319    if session
6320        .messages
6321        .iter()
6322        .any(|message| message.id == claim.envelope.id.as_str())
6323        && !transcript_has_id
6324    {
6325        return Err(AgentError::LLM(format!(
6326            "canonical SessionInbox id {} collides with a non-matching transcript message",
6327            claim.envelope.id
6328        )));
6329    }
6330    let cursor_has_id = session
6331        .session_inbox_admission()
6332        .is_some_and(|state| state.contains(&claim.envelope.id));
6333    if cursor_has_id && !transcript_has_id {
6334        return Err(AgentError::LLM(format!(
6335            "canonical SessionInbox cursor exists without transcript message {}",
6336            claim.envelope.id
6337        )));
6338    }
6339    let before = session.clone();
6340    if !transcript_has_id {
6341        let message = claim.envelope.to_provider_message().map_err(|error| {
6342            AgentError::LLM(format!(
6343                "translate canonical SessionInbox envelope {}: {error}",
6344                claim.envelope.id
6345            ))
6346        })?;
6347        session.add_message(message);
6348    }
6349    session
6350        .session_inbox_admission_mut()
6351        .record(claim.envelope.id.clone(), claim.generation);
6352    session.updated_at = chrono::Utc::now();
6353
6354    if let Err(error) = binding
6355        .persistence
6356        .checkpoint_runtime_session(session)
6357        .await
6358    {
6359        *session = before;
6360        return Err(AgentError::LLM(format!(
6361            "checkpoint canonical SessionInbox claim {}: {error}",
6362            claim.envelope.id
6363        )));
6364    }
6365    if !session
6366        .messages
6367        .iter()
6368        .any(|message| bamboo_domain::is_matching_session_message(message, &claim.envelope))
6369    {
6370        *session = before;
6371        return Err(AgentError::LLM(format!(
6372            "canonical SessionInbox checkpoint lost typed transcript proof for {}",
6373            claim.envelope.id
6374        )));
6375    }
6376    binding
6377        .inbox
6378        .ack(&session.id, claim)
6379        .await
6380        .map_err(|error| {
6381            AgentError::LLM(format!(
6382                "ack canonical SessionInbox claim {} after checkpoint: {error}",
6383                claim.envelope.id
6384            ))
6385        })
6386}
6387
6388fn concrete_file_inbox(
6389    binding: &SessionInboxRuntimeBinding,
6390) -> Result<bamboo_storage::FileSessionInbox, AgentError> {
6391    binding
6392        .inbox
6393        .as_any()
6394        .and_then(|inbox| inbox.downcast_ref::<bamboo_storage::FileSessionInbox>())
6395        .cloned()
6396        .ok_or_else(|| AgentError::LLM("remote Actor requires canonical FileSessionInbox".into()))
6397}
6398
6399async fn adopt_latest_actor_context(store: &bamboo_storage::SessionStoreV2, session: &mut Session) {
6400    if let Ok(Some(latest)) = store.load_session(&session.id).await {
6401        if latest.created_at == session.created_at
6402            && session.check_conversation_generation(&latest).is_ok()
6403        {
6404            bamboo_domain::merge_session_inbox_admission(session, &latest);
6405            session.messages = latest.messages;
6406            session.provider_transcript = latest.provider_transcript;
6407            session.updated_at = latest.updated_at;
6408        }
6409    }
6410}
6411
6412/// Durably seed claimed typed messages into the canonical host transcript
6413/// before dispatching them to any actor worker, while deliberately leaving the
6414/// admission cursor and `cur/` claims untouched.
6415///
6416/// This is the cross-placement lost-confirmation invariant: if worker A admits
6417/// and reasons over the batch but its confirmations are lost, a retry on worker
6418/// B receives a host snapshot already containing each stable typed message
6419/// exactly once. Worker B's local safe boundary then records/acks the same ids
6420/// without appending duplicates. Only an exact worker confirmation advances the
6421/// host cursor; only a durable cursor checkpoint precedes host ack.
6422async fn checkpoint_claim_context_before_dispatch(
6423    binding: &SessionInboxRuntimeBinding,
6424    session: &mut Session,
6425    claims: &[SessionInboxClaim],
6426    fenced: Option<(&bamboo_storage::SessionStoreV2, &ActorActivationFence)>,
6427) -> crate::runtime::runner::Result<()> {
6428    if claims.is_empty() {
6429        return Ok(());
6430    }
6431    let before = session.clone();
6432    let mut additions = Vec::new();
6433    for claim in claims {
6434        if claim.envelope.target_session_id != session.id {
6435            return Err(AgentError::LLM(format!(
6436                "canonical SessionInbox claim target {} does not match actor session {}",
6437                claim.envelope.target_session_id, session.id
6438            )));
6439        }
6440        let matching = session
6441            .messages
6442            .iter()
6443            .any(|message| bamboo_domain::is_matching_session_message(message, &claim.envelope));
6444        if session
6445            .messages
6446            .iter()
6447            .any(|message| message.id == claim.envelope.id.as_str())
6448            && !matching
6449        {
6450            return Err(AgentError::LLM(format!(
6451                "canonical SessionInbox id {} collides before actor dispatch",
6452                claim.envelope.id
6453            )));
6454        }
6455        if session
6456            .session_inbox_admission()
6457            .is_some_and(|state| state.contains(&claim.envelope.id))
6458            && !matching
6459        {
6460            return Err(AgentError::LLM(format!(
6461                "canonical SessionInbox cursor exists without transcript proof for {}",
6462                claim.envelope.id
6463            )));
6464        }
6465        if !matching {
6466            let message = claim.envelope.to_provider_message().map_err(|error| {
6467                AgentError::LLM(format!(
6468                    "translate canonical SessionInbox envelope {} before actor dispatch: {error}",
6469                    claim.envelope.id
6470                ))
6471            })?;
6472            additions.push(message);
6473        }
6474    }
6475    let fenced_seed = fenced.is_some();
6476    if let Some((store, fence)) = fenced {
6477        let committed = store
6478            .seed_actor_claim_context(
6479                concrete_file_inbox(binding)?,
6480                bamboo_storage::ActorClaimContextSeed {
6481                    fence: fence.clone(),
6482                    expected_created_at: session.created_at,
6483                    expected_conversation_generation: session.conversation_generation_observation(),
6484                    expected_messages: session.messages.clone(),
6485                    expected_provider_transcript: session.provider_transcript.clone(),
6486                    expected_admission: session.session_inbox_admission().cloned(),
6487                    claims: claims.to_vec(),
6488                },
6489            )
6490            .await;
6491        let committed = match committed {
6492            Ok(committed) => committed,
6493            Err(error) => {
6494                // Publication may have occurred before an unconfirmed
6495                // readback. Do not dispatch or ACK; adopt only the current
6496                // canonical context so SDK can still save an error status.
6497                adopt_latest_actor_context(store, session).await;
6498                return Err(AgentError::LLM(format!(
6499                    "seed fenced Actor claim context unconfirmed: {error}"
6500                )));
6501            }
6502        };
6503        bamboo_domain::merge_session_inbox_admission(session, &committed);
6504        session.messages = committed.messages;
6505        session.provider_transcript = committed.provider_transcript;
6506        session.updated_at = committed.updated_at;
6507    } else {
6508        for message in additions {
6509            session.add_message(message);
6510        }
6511        session.updated_at = chrono::Utc::now();
6512        if let Err(error) = binding
6513            .persistence
6514            .checkpoint_runtime_session(session)
6515            .await
6516        {
6517            *session = before;
6518            return Err(AgentError::LLM(format!(
6519                "checkpoint canonical SessionInbox actor context: {error}"
6520            )));
6521        }
6522    }
6523    for claim in claims {
6524        if !session
6525            .messages
6526            .iter()
6527            .any(|message| bamboo_domain::is_matching_session_message(message, &claim.envelope))
6528        {
6529            if !fenced_seed {
6530                *session = before;
6531            }
6532            return Err(AgentError::LLM(format!(
6533                "actor context checkpoint lost typed transcript proof for {}",
6534                claim.envelope.id
6535            )));
6536        }
6537    }
6538    Ok(())
6539}
6540
6541async fn claim_canonical_deliveries(
6542    binding: &SessionInboxRuntimeBinding,
6543    session: &mut Session,
6544    activation_run_id: &str,
6545    limit: usize,
6546    fenced: Option<(&bamboo_storage::SessionStoreV2, &ActorActivationFence)>,
6547) -> crate::runtime::runner::Result<Vec<(SessionInboxClaim, SessionMessageDelivery)>> {
6548    let claims = binding
6549        .inbox
6550        .claim_for_turn(&session.id, limit, Some(activation_run_id))
6551        .await
6552        .map_err(|error| {
6553            AgentError::LLM(format!(
6554                "claim canonical SessionInbox for active actor {}: {error}",
6555                session.id
6556            ))
6557        })?;
6558    if claims.is_empty() {
6559        return Ok(Vec::new());
6560    }
6561    let mut unconfirmed = Vec::with_capacity(claims.len());
6562    for claim in claims {
6563        if binding
6564            .inbox
6565            .was_admitted(&session.id, &claim.envelope.id)
6566            .await
6567            .map_err(|error| {
6568                AgentError::LLM(format!(
6569                    "inspect canonical SessionInbox claim {}: {error}",
6570                    claim.envelope.id
6571                ))
6572            })?
6573        {
6574            reconcile_already_admitted_claim(binding, session, &claim).await?;
6575            continue;
6576        }
6577        unconfirmed.push(claim);
6578    }
6579    checkpoint_claim_context_before_dispatch(binding, session, &unconfirmed, fenced).await?;
6580
6581    let mut deliveries = Vec::with_capacity(unconfirmed.len());
6582    for claim in unconfirmed {
6583        // Cursor+transcript without the permanent tombstone is the recoverable
6584        // crash window after an exact worker confirmation was checkpointed but
6585        // before host ack removed `cur/`. Finish that ack without exposing the
6586        // message to another provider run.
6587        if session
6588            .session_inbox_admission()
6589            .is_some_and(|state| state.contains(&claim.envelope.id))
6590        {
6591            binding
6592                .inbox
6593                .ack(&session.id, &claim)
6594                .await
6595                .map_err(|error| {
6596                    AgentError::LLM(format!(
6597                        "finish confirmed canonical SessionInbox ack {}: {error}",
6598                        claim.envelope.id
6599                    ))
6600                })?;
6601            continue;
6602        }
6603        let delivery = SessionMessageDelivery {
6604            target_session_id: session.id.clone(),
6605            envelope: claim.envelope.clone(),
6606            canonical_claim_generation: claim.generation,
6607            activation_run_id: activation_run_id.to_string(),
6608            activation_policy: claim.activation_policy,
6609        };
6610        deliveries.push((claim, delivery));
6611    }
6612    Ok(deliveries)
6613}
6614
6615async fn forward_next_canonical_claim(
6616    client: &mut dyn bamboo_subagent::ChildLink,
6617    binding: &SessionInboxRuntimeBinding,
6618    session: &mut Session,
6619    activation_run_id: &str,
6620    inflight: &mut VecDeque<SessionInboxClaim>,
6621    fenced: Option<(&bamboo_storage::SessionStoreV2, &ActorActivationFence)>,
6622) -> crate::runtime::runner::Result<()> {
6623    if !inflight.is_empty() {
6624        return Ok(());
6625    }
6626    let Some((claim, delivery)) =
6627        claim_canonical_deliveries(binding, session, activation_run_id, 1, fenced)
6628            .await?
6629            .pop()
6630    else {
6631        return Ok(());
6632    };
6633    client
6634        .send(ParentFrame::SessionMessage { delivery })
6635        .await
6636        .map_err(|error| {
6637            AgentError::LLM(format!(
6638                "forward canonical SessionInbox claim {} to active actor: {error}",
6639                claim.envelope.id
6640            ))
6641        })?;
6642    inflight.push_back(claim);
6643    Ok(())
6644}
6645
6646/// Borrowed and per-run-owned inputs for one actor frame pump.
6647struct ActorDriveContext<'a> {
6648    client: &'a mut dyn bamboo_subagent::ChildLink,
6649    parent_session_id: &'a str,
6650    child_session_id: &'a str,
6651    child_attempt: u32,
6652    approval_registry: Option<&'a super::approval_registry::SharedApprovalRegistry>,
6653    approval_decider: Option<&'a Arc<dyn ChildApprovalDecider>>,
6654    approval_reviewer: Option<&'a Arc<dyn ChildApprovalReviewer>>,
6655    escalation_bridge: Option<bamboo_subagent::executor::HostBridge>,
6656    event_tx: &'a mpsc::Sender<AgentEvent>,
6657    cancel_token: &'a CancellationToken,
6658    live_rx: &'a mut mpsc::UnboundedReceiver<ParentFrame>,
6659    delivery_rx: &'a mut mpsc::UnboundedReceiver<u64>,
6660    logical_session: &'a mut Session,
6661    expected_permission_posture: Option<ExpectedPermissionPosture>,
6662    expected_creation: Option<&'a bamboo_subagent::proto::ChildCreationIdentity>,
6663    session_inbox_runtime: Option<&'a SessionInboxRuntimeBinding>,
6664    actor_directory_store: Option<&'a bamboo_storage::SessionStoreV2>,
6665    canonical_subagent_tool: Option<Arc<dyn bamboo_agent_core::tools::Tool>>,
6666    ticket_service: Option<Arc<bamboo_tickets::TicketService>>,
6667    activation_run_id: Option<&'a str>,
6668    execution_epoch: u64,
6669    expected_source_actor_id: &'a str,
6670    initial_inflight_claims: VecDeque<SessionInboxClaim>,
6671    plain_actor: bool,
6672    remote_environment_lease: bool,
6673    readonly_output: Option<&'a mut Vec<bamboo_agent_core::Message>>,
6674    local_history_tools: Option<&'a [String]>,
6675    local_history_read_only: bool,
6676    strict_history_output: Option<&'a mut Option<StrictTerminalHistory>>,
6677    plain_input: Option<&'a PlainActorActivation>,
6678    canonical_activation: Option<&'a ActorActivationFence>,
6679    canonical_placement_ref: Option<&'a bamboo_domain::ActorPlacementRef>,
6680    actor_event_observer: Option<&'a dyn ActorEventObserver>,
6681    plain_run: Option<(&'a RunSpec, &'a AtomicU64)>,
6682    first_frame_timeout: Option<Duration>,
6683}
6684
6685#[derive(Debug, Clone, PartialEq, Eq)]
6686struct ExpectedPermissionPosture {
6687    policy_revision: u64,
6688    resolution: bamboo_domain::PermissionModeResolution,
6689    expected_audit_revision: Option<u64>,
6690    executor_mapping: String,
6691}
6692
6693#[derive(Debug, Clone, Copy, PartialEq, Eq)]
6694enum PermissionPostureHandshake {
6695    /// Legacy/custom actors were dispatched without a typed posture contract.
6696    NotRequired,
6697    /// The host dispatched an exact posture and no matching, durably recorded
6698    /// worker activation has arrived yet.
6699    Awaiting,
6700    /// One worker posture matched and its host-owned audit write succeeded.
6701    Confirmed,
6702}
6703
6704impl PermissionPostureHandshake {
6705    fn new(expected: Option<&ExpectedPermissionPosture>) -> Self {
6706        if expected.is_some() {
6707            Self::Awaiting
6708        } else {
6709            Self::NotRequired
6710        }
6711    }
6712
6713    fn is_awaiting(self) -> bool {
6714        self == Self::Awaiting
6715    }
6716
6717    fn posture_was_confirmed(self) -> bool {
6718        self == Self::Confirmed
6719    }
6720}
6721
6722fn permission_posture_seed_from_event(
6723    session: &Session,
6724    event: &AgentEvent,
6725) -> Result<Option<bamboo_domain::PermissionAuditSeed>, String> {
6726    let AgentEvent::PermissionPostureActivated {
6727        session_id,
6728        policy_revision,
6729        requested_mode,
6730        effective_mode,
6731        executor_mapping,
6732    } = event
6733    else {
6734        return Ok(None);
6735    };
6736    if session_id != &session.id {
6737        return Err("permission posture event targets a different logical session".to_string());
6738    }
6739    let requested = bamboo_domain::SessionPermissionMode::from_audit_str(requested_mode)
6740        .ok_or_else(|| "permission posture event has an invalid requested mode".to_string())?;
6741    let effective = bamboo_domain::PermissionMode::from_audit_str(effective_mode)
6742        .ok_or_else(|| "permission posture event has an invalid effective mode".to_string())?;
6743    let resolution = bamboo_domain::PermissionModeResolution {
6744        requested,
6745        effective,
6746    };
6747    if !resolution.is_consistent() {
6748        return Err("permission posture event has an inconsistent mode pair".to_string());
6749    }
6750    let current_requested = session
6751        .agent_runtime_state
6752        .as_ref()
6753        .map(|state| state.effective_permission_mode())
6754        .unwrap_or_default();
6755    if current_requested != requested {
6756        return Err("permission posture event is stale for the host typed mode".to_string());
6757    }
6758    let mapping_chars = executor_mapping.chars().count();
6759    if mapping_chars == 0 || mapping_chars > bamboo_domain::MAX_PERMISSION_EXECUTOR_MAPPING_CHARS {
6760        return Err("permission posture event has an invalid executor mapping".to_string());
6761    }
6762    Ok(Some(bamboo_domain::PermissionAuditSeed::new(
6763        *policy_revision,
6764        resolution,
6765        executor_mapping,
6766    )))
6767}
6768
6769fn actor_event_logical_identity(
6770    session: &Session,
6771    parent_session_id: &str,
6772    creation: Option<&bamboo_subagent::proto::ChildCreationIdentity>,
6773) -> LogicalSessionIdentity {
6774    LogicalSessionIdentity {
6775        creation: creation.cloned(),
6776        session_id: session.id.clone(),
6777        parent_session_id: session
6778            .parent_session_id
6779            .clone()
6780            .or_else(|| Some(parent_session_id.to_string())),
6781        root_session_id: if session.root_session_id.trim().is_empty() {
6782            parent_session_id.to_string()
6783        } else {
6784            session.root_session_id.clone()
6785        },
6786    }
6787}
6788
6789fn validate_strict_history_terminal(
6790    watermark: Option<&bamboo_subagent::ActorEventWatermark>,
6791    session: &Session,
6792    parent_session_id: &str,
6793    activation_run_id: Option<&str>,
6794    execution_epoch: u64,
6795    creation: Option<&bamboo_subagent::proto::ChildCreationIdentity>,
6796    consumed_seq: u64,
6797) -> Result<(), AgentError> {
6798    let identity = actor_event_logical_identity(session, parent_session_id, creation);
6799    if watermark
6800        .zip(activation_run_id)
6801        .is_some_and(|(watermark, run_id)| {
6802            watermark.matches_consumed_run(&identity, run_id, execution_epoch, consumed_seq)
6803        })
6804    {
6805        Ok(())
6806    } else {
6807        Err(AgentError::LLM("actor_history_tail_incomplete".into()))
6808    }
6809}
6810
6811fn validate_actor_event_batch(
6812    batch: &ActorEventBatch,
6813    logical_session: &Session,
6814    parent_session_id: &str,
6815    activation_run_id: Option<&str>,
6816    execution_epoch: u64,
6817    expected_source_actor_id: &str,
6818    expected_creation: Option<&bamboo_subagent::proto::ChildCreationIdentity>,
6819) -> Result<(), AgentError> {
6820    batch
6821        .validate()
6822        .map_err(|error| AgentError::LLM(format!("actor emitted invalid event batch: {error}")))?;
6823    let identity = batch.logical_session.as_ref().ok_or_else(|| {
6824        AgentError::LLM("actor event batch omitted its logical session identity".to_string())
6825    })?;
6826    // Fence against the exact canonical identity used to build RunSpec. Some
6827    // migrated/test sessions have not yet materialized parent ancestry on the
6828    // in-memory Session, so comparing the batch directly with those raw fields
6829    // can reject the identity the host itself just dispatched.
6830    let expected_identity =
6831        actor_event_logical_identity(logical_session, parent_session_id, expected_creation);
6832    if identity != &expected_identity {
6833        return Err(AgentError::LLM(
6834            "actor event batch targets a different logical session".to_string(),
6835        ));
6836    }
6837    if batch.activation_id.as_deref() != activation_run_id {
6838        return Err(AgentError::LLM(
6839            "actor event batch belongs to a stale activation".to_string(),
6840        ));
6841    }
6842    if batch.execution_epoch != execution_epoch {
6843        return Err(AgentError::LLM(
6844            "actor event batch belongs to a stale execution epoch".to_string(),
6845        ));
6846    }
6847    if batch
6848        .source_actor_id
6849        .as_deref()
6850        .is_some_and(|source| source != expected_source_actor_id)
6851    {
6852        return Err(AgentError::LLM(
6853            "actor event batch declares a different physical actor".to_string(),
6854        ));
6855    }
6856    Ok(())
6857}
6858
6859/// The plain Actor path already owns a durable Directory activation. Check that
6860/// exact owner before admitting either batched or rolling-upgrade live frames;
6861/// worker-provided batch identity alone is never an activation authority.
6862async fn validate_directory_actor_event(
6863    store: &bamboo_storage::SessionStoreV2,
6864    fence: &ActorActivationFence,
6865    session: &Session,
6866    run_id: Option<&str>,
6867) -> Result<bamboo_domain::ActorDirectoryEntry, AgentError> {
6868    store
6869        .validate_fence(fence, chrono::Utc::now())
6870        .await
6871        .map_err(|_| AgentError::LLM("actor event owner is stale".into()))?;
6872    let entry = store
6873        .inspect_actor(&fence.actor_id)
6874        .await
6875        .map_err(|_| AgentError::LLM("actor event authority is unavailable".into()))?;
6876    if !entry.actor.matches_session(session)
6877        || entry.actor.project_id != project_id_for_actor_run(session)?.map(|id| id.to_string())
6878        || !entry.activation.as_ref().is_some_and(|current| {
6879            current.matches_fence(fence)
6880                && current.status == bamboo_domain::ActorActivationStatus::Running
6881        })
6882        || run_id.is_some_and(|id| id != fence.run_id)
6883    {
6884        return Err(AgentError::LLM("actor event authority changed".into()));
6885    }
6886    Ok(entry)
6887}
6888
6889/// The Worker supplies only a cursor. All identity and scope inputs come from
6890/// this Host's active drive and its canonical durable Session store.
6891async fn inspect_active_owned_tree(
6892    store: &bamboo_storage::SessionStoreV2,
6893    binding: &SessionInboxRuntimeBinding,
6894    run_id: &str,
6895    child_id: &str,
6896    parent_id: &str,
6897    logical_session: &Session,
6898    creation: &bamboo_subagent::proto::ChildCreationIdentity,
6899    plain_input: Option<&PlainActorActivation>,
6900    cursor: Option<&str>,
6901) -> Option<serde_json::Value> {
6902    use bamboo_agent_core::storage::Storage;
6903    let current = match store.load_session(child_id).await {
6904        Ok(Some(current)) => current,
6905        _ => {
6906            tracing::warn!("owned tree denied: canonical caller unavailable");
6907            return None;
6908        }
6909    };
6910    let current_project = match project_id_for_actor_run(&current) {
6911        Ok(project) => project,
6912        Err(_) => {
6913            tracing::warn!("owned tree denied: invalid canonical Project");
6914            return None;
6915        }
6916    };
6917    let logical_project = match project_id_for_actor_run(logical_session) {
6918        Ok(project) => project,
6919        Err(_) => {
6920            tracing::warn!("owned tree denied: invalid logical Project");
6921            return None;
6922        }
6923    };
6924    if current.id != child_id
6925        || current.kind != bamboo_domain::SessionKind::Child
6926        || current.parent_session_id.as_deref() != Some(parent_id)
6927        || current.parent_session_id != logical_session.parent_session_id
6928        || current.root_session_id != logical_session.root_session_id
6929        || current.created_at != creation.created_at
6930        || current.created_at != logical_session.created_at
6931        || current.spawn_depth != creation.spawn_depth
6932        || current.spawn_depth != logical_session.spawn_depth
6933        || current_project != logical_project
6934    {
6935        tracing::warn!("owned tree denied: caller lifetime or Project changed");
6936        return None;
6937    }
6938    if let Some(activation) = plain_input {
6939        validate_directory_actor_event(
6940            activation.store.as_ref(),
6941            &activation.fence,
6942            &current,
6943            Some(run_id),
6944        )
6945        .await
6946        .ok()?;
6947    } else if !binding.router.owns_run(child_id, run_id).await {
6948        tracing::warn!("owned tree denied: activation owner changed");
6949        return None;
6950    }
6951    let page = match crate::session_app::child_session::owned_tree::inspect_owned_tree(
6952        store, child_id, cursor,
6953    )
6954    .await
6955    {
6956        Ok(page) => page,
6957        Err(error) => {
6958            tracing::warn!(?error, "owned tree denied: canonical tree read failed");
6959            return None;
6960        }
6961    };
6962    let again = match store.load_session(child_id).await {
6963        Ok(Some(again)) => again,
6964        _ => {
6965            tracing::warn!("owned tree denied: canonical caller disappeared");
6966            return None;
6967        }
6968    };
6969    if again.created_at != current.created_at
6970        || again.parent_session_id != current.parent_session_id
6971        || again.root_session_id != current.root_session_id
6972        || again.spawn_depth != current.spawn_depth
6973        || project_id_for_actor_run(&again).ok()? != current_project
6974    {
6975        tracing::warn!("owned tree denied: caller changed during inspection");
6976        return None;
6977    }
6978    if let Some(activation) = plain_input {
6979        validate_directory_actor_event(
6980            activation.store.as_ref(),
6981            &activation.fence,
6982            &again,
6983            Some(run_id),
6984        )
6985        .await
6986        .ok()?;
6987    } else if !binding.router.owns_run(child_id, run_id).await {
6988        tracing::warn!("owned tree denied: activation owner changed after inspection");
6989        return None;
6990    }
6991    serde_json::to_vec(&page)
6992        .ok()
6993        .filter(|bytes| bytes.len() <= 8192)?;
6994    Some(page)
6995}
6996
6997/// Resolve the caller of a mutating nested SubAgent RPC entirely from the
6998/// Host's active drive and canonical store. The frame never supplies identity,
6999/// Project, ancestry, or a run fence.
7000async fn load_active_subagent_caller(
7001    store: &bamboo_storage::SessionStoreV2,
7002    binding: &SessionInboxRuntimeBinding,
7003    run_id: &str,
7004    child_id: &str,
7005    parent_id: &str,
7006    logical_session: &Session,
7007    creation: &bamboo_subagent::proto::ChildCreationIdentity,
7008    plain_input: Option<&PlainActorActivation>,
7009) -> Option<Session> {
7010    use bamboo_agent_core::storage::Storage;
7011    let current = store.load_session(child_id).await.ok().flatten()?;
7012    let current_project = project_id_for_actor_run(&current).ok()?;
7013    let logical_project = project_id_for_actor_run(logical_session).ok()?;
7014    if current.id != child_id
7015        || current.kind != bamboo_domain::SessionKind::Child
7016        || current.parent_session_id.as_deref() != Some(parent_id)
7017        || current.parent_session_id != logical_session.parent_session_id
7018        || current.root_session_id != logical_session.root_session_id
7019        || current.created_at != creation.created_at
7020        || current.created_at != logical_session.created_at
7021        || current.spawn_depth != creation.spawn_depth
7022        || current.spawn_depth != logical_session.spawn_depth
7023        || current_project != logical_project
7024        || bamboo_domain::ActorSession::from_session(&current).is_err()
7025    {
7026        return None;
7027    }
7028    if let Some(activation) = plain_input {
7029        validate_directory_actor_event(
7030            activation.store.as_ref(),
7031            &activation.fence,
7032            &current,
7033            Some(run_id),
7034        )
7035        .await
7036        .ok()?;
7037    } else if !binding.router.owns_run(child_id, run_id).await {
7038        return None;
7039    }
7040    Some(current)
7041}
7042
7043async fn invoke_active_subagent(
7044    tool: &dyn bamboo_agent_core::tools::Tool,
7045    caller: &Session,
7046    args: serde_json::Value,
7047    tool_call_id: &str,
7048) -> serde_json::Value {
7049    use bamboo_agent_core::tools::{ToolCtx, ToolOutcome};
7050    let mut ctx = ToolCtx::none(tool_call_id.to_owned());
7051    ctx.session_id = Some(Arc::from(caller.id.as_str()));
7052    match tool.invoke(args, ctx).await {
7053        Ok(ToolOutcome::Completed(result))
7054            if result.images.is_empty()
7055                && serde_json::to_vec(&result).is_ok_and(|bytes| bytes.len() <= 16 * 1024) =>
7056        {
7057            serde_json::json!({"result":result})
7058        }
7059        _ => serde_json::json!({
7060            "error":"SubAgent operation failed or its result exceeded the limit; inspect logical child state before retrying"
7061        }),
7062    }
7063}
7064
7065/// Commit one already-persisted Worker clarification under the Host's current
7066/// Child birth and Run fence. Only the Host supplies caller/parent identity.
7067#[allow(clippy::too_many_arguments)]
7068async fn checkpoint_active_parent_question(
7069    store: &bamboo_storage::SessionStoreV2,
7070    binding: &SessionInboxRuntimeBinding,
7071    run_id: &str,
7072    child_id: &str,
7073    parent_id: &str,
7074    logical_session: &mut Session,
7075    creation: &bamboo_subagent::proto::ChildCreationIdentity,
7076    plain_input: Option<&PlainActorActivation>,
7077    observation: ParentQuestionCheckpointV1,
7078) -> Option<String> {
7079    use bamboo_agent_core::storage::Storage;
7080    if plain_input.is_some() || observation.validate_shape().is_err() {
7081        return None;
7082    }
7083    // SessionRepository's final save and the direct-parent answer resolver
7084    // use this same Host-owned lock. Hold it across the V2 read/CAS/save so
7085    // neither can read an older Child and overwrite the checkpoint afterward.
7086    let _session_guard = binding
7087        .parent_question_lock
7088        .as_ref()?
7089        .acquire_lock(child_id)
7090        .await;
7091    let caller = load_active_subagent_caller(
7092        store,
7093        binding,
7094        run_id,
7095        child_id,
7096        parent_id,
7097        logical_session,
7098        creation,
7099        plain_input,
7100    )
7101    .await?;
7102    let parent = store.load_session(parent_id).await.ok().flatten()?;
7103    if parent.created_at > caller.created_at || !binding.router.owns_run(child_id, run_id).await {
7104        return None;
7105    }
7106    let (saved, question) = store
7107        .checkpoint_parent_question(&caller, &parent, &observation)
7108        .await
7109        .ok()
7110        .flatten()?;
7111    // The current Run's local snapshot must carry the committed canonical
7112    // prefix into finalization; sdk/spawn rechecks disk before its last save.
7113    logical_session.messages = saved.messages;
7114    logical_session.pending_question = saved.pending_question;
7115    if let Some(reason) = saved.metadata.get("runtime.suspend_reason") {
7116        logical_session
7117            .metadata
7118            .insert("runtime.suspend_reason".into(), reason.clone());
7119    }
7120    if let Some(request) = saved
7121        .metadata
7122        .get(bamboo_domain::PARENT_QUESTION_REQUEST_KEY)
7123    {
7124        logical_session.metadata.insert(
7125            bamboo_domain::PARENT_QUESTION_REQUEST_KEY.into(),
7126            request.clone(),
7127        );
7128    }
7129    logical_session.metadata.insert(
7130        "runtime.actor_parent_question_handoff".into(),
7131        "true".into(),
7132    );
7133    Some(question.id.as_str().to_owned())
7134}
7135
7136/// An actor Worker cannot see the Host's canonical descendants. At the end of
7137/// an otherwise completed Child turn, apply the same orphan-child safety net
7138/// used by the in-process root loop before publishing this Child as terminal.
7139/// The existing Host wait, including an explicit `any` policy, always wins.
7140async fn ensure_active_nested_wait(
7141    tool: &dyn bamboo_agent_core::tools::Tool,
7142    store: &bamboo_storage::SessionStoreV2,
7143    binding: &SessionInboxRuntimeBinding,
7144    run_id: &str,
7145    child_id: &str,
7146    parent_id: &str,
7147    logical_session: &Session,
7148    creation: &bamboo_subagent::proto::ChildCreationIdentity,
7149    plain_input: Option<&PlainActorActivation>,
7150) -> Result<bool, AgentError> {
7151    fn is_terminal_status(status: &str) -> bool {
7152        matches!(
7153            status,
7154            "completed" | "error" | "timeout" | "cancelled" | "skipped"
7155        )
7156    }
7157    use bamboo_agent_core::storage::Storage;
7158    use bamboo_agent_core::tools::{ToolCtx, ToolOutcome};
7159
7160    let caller = load_active_subagent_caller(
7161        store,
7162        binding,
7163        run_id,
7164        child_id,
7165        parent_id,
7166        logical_session,
7167        creation,
7168        plain_input,
7169    )
7170    .await
7171    .ok_or_else(|| AgentError::LLM("nested wait caller is no longer active".into()))?;
7172    if caller
7173        .agent_runtime_state
7174        .as_ref()
7175        .and_then(|state| state.waiting_for_children.as_ref())
7176        .is_some()
7177    {
7178        return Ok(true);
7179    }
7180
7181    let statuses = store
7182        .list_child_run_statuses(child_id)
7183        .await
7184        .map_err(|error| AgentError::LLM(format!("nested child status scan failed: {error}")))?;
7185    let mut active = Vec::new();
7186    for (id, status) in statuses {
7187        if !status.as_deref().is_some_and(is_terminal_status)
7188            && !store
7189                .is_supervisor_observer_child(&caller, &id)
7190                .await
7191                .map_err(|error| {
7192                    AgentError::LLM(format!("nested Observer classification failed: {error}"))
7193                })?
7194        {
7195            active.push(id);
7196        }
7197    }
7198    if active.is_empty() {
7199        return Ok(false);
7200    }
7201
7202    let mut ctx = ToolCtx::none(format!("host-nested-orphan-wait:{run_id}"));
7203    ctx.session_id = Some(Arc::from(caller.id.as_str()));
7204    let outcome = tokio::time::timeout(
7205        Duration::from_secs(30),
7206        tool.invoke(
7207            serde_json::json!({"action":"wait", "child_session_ids":active}),
7208            ctx,
7209        ),
7210    )
7211    .await
7212    .map_err(|_| AgentError::LLM("nested wait registration timed out".into()))?
7213    .map_err(|error| AgentError::LLM(format!("nested wait registration failed: {error}")))?;
7214    let wait_registered = match outcome {
7215        ToolOutcome::Completed(result) if result.success => {
7216            let payload: serde_json::Value = serde_json::from_str(&result.result)
7217                .map_err(|_| AgentError::LLM("nested wait result is invalid".into()))?;
7218            payload["status"] == "waiting"
7219        }
7220        _ => {
7221            return Err(AgentError::LLM(
7222                "nested wait registration was not completed".into(),
7223            ))
7224        }
7225    };
7226
7227    // A grandchild may finish while the wait is being registered and clear it.
7228    // A clear wait is safe only when no canonical descendants remain active.
7229    let latest = load_active_subagent_caller(
7230        store,
7231        binding,
7232        run_id,
7233        child_id,
7234        parent_id,
7235        logical_session,
7236        creation,
7237        plain_input,
7238    )
7239    .await
7240    .ok_or_else(|| AgentError::LLM("nested wait caller changed during registration".into()))?;
7241    if latest
7242        .agent_runtime_state
7243        .as_ref()
7244        .and_then(|state| state.waiting_for_children.as_ref())
7245        .is_some()
7246    {
7247        return Ok(true);
7248    }
7249    let statuses = store
7250        .list_child_run_statuses(child_id)
7251        .await
7252        .map_err(|error| {
7253            AgentError::LLM(format!("nested child wait verification failed: {error}"))
7254        })?;
7255    let mut still_active = false;
7256    for (id, status) in statuses {
7257        if !status.as_deref().is_some_and(is_terminal_status)
7258            && !store
7259                .is_supervisor_observer_child(&latest, &id)
7260                .await
7261                .map_err(|error| {
7262                    AgentError::LLM(format!("nested Observer classification failed: {error}"))
7263                })?
7264        {
7265            still_active = true;
7266            break;
7267        }
7268    }
7269    if still_active {
7270        return Err(AgentError::LLM(
7271            "nested child remains active without a durable parent wait".into(),
7272        ));
7273    }
7274    // Keep the current Run's suspend intent even if the last grandchild
7275    // finished and cleared the Host wait before SDK finalization. Its outcome
7276    // has already admitted a successor; publishing this Run as completed
7277    // would release the grandparent on the Child's interim response.
7278    Ok(wait_registered)
7279}
7280
7281fn actor_event_route_error(error: ActorEventRouteError) -> AgentError {
7282    let message = match error {
7283        ActorEventRouteError::SequenceGap { .. } => {
7284            "actor event sequence gap; authoritative session snapshot is required"
7285        }
7286        ActorEventRouteError::LifecycleMissing => {
7287            "actor emitted live content before its activation lifecycle event"
7288        }
7289        ActorEventRouteError::ForgedSession => {
7290            "actor event payload targets a different canonical session"
7291        }
7292        ActorEventRouteError::ConflictingReplay => {
7293            "actor replay reused an event sequence with different content"
7294        }
7295        ActorEventRouteError::StaleAuthority => "actor event owner is stale",
7296        ActorEventRouteError::SequenceExhausted => "actor event sequence is exhausted",
7297    };
7298    AgentError::LLM(message.into())
7299}
7300
7301const MAX_DISPLAY_CALLS: usize = 4_096;
7302const MAX_NESTED_DISPLAY_DEPTH: usize = 8;
7303const MAX_DISPLAY_ID_BYTES: usize = 256;
7304const MAX_DISPLAY_HOOK_RESULT_DEPTH: usize = 16;
7305
7306fn hook_result_for_display(result: HookResult, depth: usize) -> Option<HookResult> {
7307    if depth >= MAX_DISPLAY_HOOK_RESULT_DEPTH {
7308        return None;
7309    }
7310    Some(match result {
7311        HookResult::Continue => HookResult::Continue,
7312        HookResult::Mutated => HookResult::Mutated,
7313        HookResult::Allow => HookResult::Allow,
7314        HookResult::Ask => HookResult::Ask,
7315        HookResult::Deny { .. } => HookResult::Deny {
7316            reason: "Hook denied".into(),
7317        },
7318        HookResult::InjectContext { .. } => HookResult::InjectContext {
7319            text: "Hook context hidden".into(),
7320        },
7321        HookResult::WithContext { result, .. } => HookResult::WithContext {
7322            result: Box::new(hook_result_for_display(*result, depth + 1)?),
7323            text: "Hook context hidden".into(),
7324        },
7325        HookResult::Suspend { .. } => HookResult::Suspend {
7326            reason: "Hook suspended".into(),
7327        },
7328        HookResult::Abort { .. } => HookResult::Abort {
7329            reason: "Hook aborted".into(),
7330        },
7331    })
7332}
7333
7334#[derive(Clone, Copy, PartialEq, Eq)]
7335enum DisplayCallKind {
7336    Ordinary,
7337    PrivateBrowser,
7338    Reused,
7339}
7340
7341#[derive(Clone, Copy)]
7342struct DisplayCall {
7343    kind: DisplayCallKind,
7344    started: bool,
7345}
7346
7347#[derive(Default)]
7348struct ActorEventDisplay {
7349    calls: HashMap<String, DisplayCall>,
7350    hook_names: HashMap<String, usize>,
7351    children: HashMap<(String, String), Box<ActorEventDisplay>>,
7352    // Shared by every nested child: a per-level cap alone permits an
7353    // exponential number of identities across an actor event tree.
7354    identities: Arc<AtomicUsize>,
7355    overflowed: bool,
7356    nested_overflowed: bool,
7357}
7358
7359fn is_browser_tool_name(tool_name: &str) -> bool {
7360    tool_name
7361        .trim()
7362        .rsplit("::")
7363        .next()
7364        .is_some_and(|name| name.eq_ignore_ascii_case("browser"))
7365}
7366
7367fn known_browser_action(args: &serde_json::Value) -> bool {
7368    matches!(
7369        args.get("action").and_then(serde_json::Value::as_str),
7370        Some(
7371            "tabs"
7372                | "new_tab"
7373                | "activate_tab"
7374                | "close_tab"
7375                | "navigate"
7376                | "history"
7377                | "viewport"
7378                | "snapshot"
7379                | "click"
7380                | "click_at"
7381                | "hover"
7382                | "drag"
7383                | "fill"
7384                | "select_option"
7385                | "set_file_input"
7386                | "type"
7387                | "press"
7388                | "key"
7389                | "scroll"
7390                | "download"
7391                | "screenshot"
7392                | "dialog_respond"
7393        )
7394    )
7395}
7396
7397fn is_private_browser_call(tool_name: &str, args: &serde_json::Value) -> bool {
7398    is_browser_tool_name(tool_name)
7399        && (!known_browser_action(args)
7400            || args
7401                .get("action")
7402                .and_then(serde_json::Value::as_str)
7403                .is_some_and(|action| {
7404                    action.eq_ignore_ascii_case("download") || action == "set_file_input"
7405                })
7406            || args.get("data_base64").is_some())
7407}
7408
7409fn actor_arguments_for_display(tool_name: &str, args: &serde_json::Value) -> serde_json::Value {
7410    if is_browser_tool_name(tool_name) {
7411        // A malformed or unknown action has no reliable privacy class. A
7412        // mixed-version actor may emit it before its schema rejects it.
7413        if !known_browser_action(args) {
7414            return serde_json::json!({"details":"[redacted]"});
7415        }
7416    }
7417    tool_start_arguments_for_display(tool_name, args)
7418}
7419
7420impl ActorEventDisplay {
7421    fn fail_closed(&mut self) {
7422        self.identities
7423            .store(MAX_DISPLAY_CALLS + 1, Ordering::Relaxed);
7424        self.overflowed = true;
7425        self.calls.clear();
7426        self.hook_names.clear();
7427        self.nested_overflowed = true;
7428        self.children.clear();
7429    }
7430
7431    fn reserve_identity(&mut self) -> bool {
7432        if self.identities.fetch_add(1, Ordering::Relaxed) < MAX_DISPLAY_CALLS {
7433            true
7434        } else {
7435            self.fail_closed();
7436            false
7437        }
7438    }
7439
7440    fn remember_start(&mut self, id: &str, private: bool) -> DisplayCallKind {
7441        if self.overflowed || id.is_empty() || id.len() > MAX_DISPLAY_ID_BYTES {
7442            return DisplayCallKind::Reused;
7443        }
7444        let kind = if private {
7445            DisplayCallKind::PrivateBrowser
7446        } else {
7447            DisplayCallKind::Ordinary
7448        };
7449        if let Some(call) = self.calls.get_mut(id) {
7450            if call.started || call.kind != kind {
7451                call.kind = DisplayCallKind::Reused;
7452            }
7453            call.started = true;
7454            return call.kind;
7455        }
7456        if !self.reserve_identity() {
7457            return DisplayCallKind::Reused;
7458        }
7459        self.calls.insert(
7460            id.to_string(),
7461            DisplayCall {
7462                kind,
7463                started: true,
7464            },
7465        );
7466        kind
7467    }
7468
7469    fn remember_approval(&mut self, id: &str, private: bool) -> DisplayCallKind {
7470        if self.overflowed || id.is_empty() || id.len() > MAX_DISPLAY_ID_BYTES {
7471            return DisplayCallKind::Reused;
7472        }
7473        let kind = if private {
7474            DisplayCallKind::PrivateBrowser
7475        } else {
7476            DisplayCallKind::Ordinary
7477        };
7478        if let Some(call) = self.calls.get_mut(id) {
7479            if call.kind != kind {
7480                call.kind = DisplayCallKind::Reused;
7481            }
7482            return call.kind;
7483        }
7484        if !self.reserve_identity() {
7485            return DisplayCallKind::Reused;
7486        }
7487        self.calls.insert(
7488            id.to_string(),
7489            DisplayCall {
7490                kind,
7491                started: false,
7492            },
7493        );
7494        kind
7495    }
7496
7497    fn known_kind(&self, id: &str) -> Option<DisplayCallKind> {
7498        if self.overflowed || self.identities.load(Ordering::Relaxed) > MAX_DISPLAY_CALLS {
7499            return None;
7500        }
7501        self.calls
7502            .get(id)
7503            .filter(|call| call.started)
7504            .map(|call| call.kind)
7505    }
7506
7507    fn display_hook_name(&mut self, hook_name: &str) -> Option<String> {
7508        if self.overflowed || hook_name.is_empty() || hook_name.len() > MAX_DISPLAY_ID_BYTES {
7509            self.fail_closed();
7510            return None;
7511        }
7512        if let Some(index) = self.hook_names.get(hook_name) {
7513            return Some(format!("actor-hook-{index}"));
7514        }
7515        if !self.reserve_identity() {
7516            return None;
7517        }
7518        let index = self.hook_names.len() + 1;
7519        self.hook_names.insert(hook_name.to_string(), index);
7520        Some(format!("actor-hook-{index}"))
7521    }
7522
7523    fn project(&mut self, event: AgentEvent, depth: usize) -> Option<AgentEvent> {
7524        let call_id = match &event {
7525            AgentEvent::ToolStart { tool_call_id, .. }
7526            | AgentEvent::ToolApprovalRequested { tool_call_id, .. }
7527            | AgentEvent::ToolToken { tool_call_id, .. }
7528            | AgentEvent::ToolComplete { tool_call_id, .. }
7529            | AgentEvent::ToolError { tool_call_id, .. }
7530            | AgentEvent::ToolLifecycle { tool_call_id, .. } => Some(tool_call_id),
7531            _ => None,
7532        };
7533        if call_id.is_some_and(|id| id.is_empty() || id.len() > MAX_DISPLAY_ID_BYTES) {
7534            self.fail_closed();
7535            return None;
7536        }
7537        if self.identities.load(Ordering::Relaxed) > MAX_DISPLAY_CALLS {
7538            self.fail_closed();
7539        }
7540        Some(match event {
7541            AgentEvent::ToolStart {
7542                tool_call_id,
7543                tool_name,
7544                arguments,
7545            } => {
7546                let private = is_private_browser_call(&tool_name, &arguments);
7547                let kind = self.remember_start(&tool_call_id, private);
7548                let arguments = if kind == DisplayCallKind::Reused {
7549                    serde_json::json!({"details":"[redacted]"})
7550                } else {
7551                    actor_arguments_for_display(&tool_name, &arguments)
7552                };
7553                AgentEvent::ToolStart {
7554                    tool_call_id,
7555                    tool_name: match kind {
7556                        DisplayCallKind::Reused => "unknown".into(),
7557                        _ if is_browser_tool_name(&tool_name) => "browser".into(),
7558                        _ => tool_name,
7559                    },
7560                    arguments,
7561                }
7562            }
7563            AgentEvent::ToolApprovalRequested {
7564                tool_call_id,
7565                tool_name,
7566                parameters,
7567            } => {
7568                let private = is_private_browser_call(&tool_name, &parameters);
7569                let kind = self.remember_approval(&tool_call_id, private);
7570                let parameters = if kind == DisplayCallKind::Reused {
7571                    serde_json::json!({"details":"[redacted]"})
7572                } else {
7573                    actor_arguments_for_display(&tool_name, &parameters)
7574                };
7575                AgentEvent::ToolApprovalRequested {
7576                    tool_call_id,
7577                    tool_name: match kind {
7578                        DisplayCallKind::Reused => "unknown".into(),
7579                        _ if is_browser_tool_name(&tool_name) => "browser".into(),
7580                        _ => tool_name,
7581                    },
7582                    parameters,
7583                }
7584            }
7585            AgentEvent::ToolToken {
7586                tool_call_id,
7587                content,
7588            } => {
7589                let content = match self.known_kind(&tool_call_id) {
7590                    Some(DisplayCallKind::Ordinary) => content,
7591                    Some(DisplayCallKind::PrivateBrowser) => "Browser tool output hidden".into(),
7592                    _ => "Tool output hidden".into(),
7593                };
7594                AgentEvent::ToolToken {
7595                    tool_call_id,
7596                    content,
7597                }
7598            }
7599            AgentEvent::ToolComplete {
7600                tool_call_id,
7601                result,
7602            } => {
7603                let result = match self.known_kind(&tool_call_id) {
7604                    Some(DisplayCallKind::Ordinary) => result,
7605                    Some(DisplayCallKind::PrivateBrowser) => {
7606                        bamboo_agent_core::tools::ToolResult::text(
7607                            result.success,
7608                            "Browser tool result hidden",
7609                        )
7610                    }
7611                    _ => bamboo_agent_core::tools::ToolResult::text(
7612                        result.success,
7613                        "Tool result hidden",
7614                    ),
7615                };
7616                AgentEvent::ToolComplete {
7617                    tool_call_id,
7618                    result,
7619                }
7620            }
7621            AgentEvent::ToolError {
7622                tool_call_id,
7623                error,
7624            } => {
7625                let error = match self.known_kind(&tool_call_id) {
7626                    Some(DisplayCallKind::Ordinary) => error,
7627                    Some(DisplayCallKind::PrivateBrowser) => "Browser tool failed".into(),
7628                    _ => "Tool failed".into(),
7629                };
7630                AgentEvent::ToolError {
7631                    tool_call_id,
7632                    error,
7633                }
7634            }
7635            AgentEvent::ToolLifecycle {
7636                tool_call_id,
7637                tool_name,
7638                phase,
7639                elapsed_ms,
7640                is_mutating,
7641                auto_approved,
7642                summary,
7643                error,
7644            } => {
7645                let kind = self.known_kind(&tool_call_id);
7646                let (summary, error) = match kind {
7647                    Some(DisplayCallKind::Ordinary) => (summary, error),
7648                    Some(DisplayCallKind::PrivateBrowser) => (
7649                        summary.map(|_| "Browser tool activity hidden".into()),
7650                        error.map(|_| "Browser tool failed".into()),
7651                    ),
7652                    _ => (
7653                        summary.map(|_| "Tool activity hidden".into()),
7654                        error.map(|_| "Tool failed".into()),
7655                    ),
7656                };
7657                AgentEvent::ToolLifecycle {
7658                    tool_call_id,
7659                    tool_name: match kind {
7660                        Some(DisplayCallKind::Ordinary) => tool_name,
7661                        Some(DisplayCallKind::PrivateBrowser) => "browser".into(),
7662                        _ => "unknown".into(),
7663                    },
7664                    phase: if kind == Some(DisplayCallKind::Ordinary)
7665                        || matches!(phase.as_str(), "begin" | "finished" | "error" | "cancelled")
7666                    {
7667                        phase
7668                    } else {
7669                        "event".into()
7670                    },
7671                    elapsed_ms,
7672                    is_mutating,
7673                    auto_approved,
7674                    summary,
7675                    error,
7676                }
7677            }
7678            AgentEvent::HookLifecycle {
7679                hook_name,
7680                point,
7681                phase,
7682                duration_ms,
7683                decision,
7684            } => {
7685                let Some(decision) = hook_result_for_display(decision, 0) else {
7686                    self.fail_closed();
7687                    return None;
7688                };
7689                let hook_name = self.display_hook_name(&hook_name)?;
7690                let phase = match phase.as_str() {
7691                    "completed" | "started" | "error" => phase,
7692                    _ => "event".into(),
7693                };
7694                AgentEvent::HookLifecycle {
7695                    hook_name,
7696                    point,
7697                    phase,
7698                    duration_ms,
7699                    decision,
7700                }
7701            }
7702            AgentEvent::ChildApprovalRequested {
7703                child_session_id,
7704                request_id,
7705                tool_name,
7706                permission,
7707                resource,
7708            } => {
7709                let (tool_name, permission, resource) = if is_browser_tool_name(&tool_name) {
7710                    (
7711                        "browser".into(),
7712                        "Browser interaction approval".into(),
7713                        "[redacted]".into(),
7714                    )
7715                } else {
7716                    (tool_name, permission, resource)
7717                };
7718                AgentEvent::ChildApprovalRequested {
7719                    child_session_id,
7720                    request_id,
7721                    tool_name,
7722                    permission,
7723                    resource,
7724                }
7725            }
7726            AgentEvent::ChildApprovalChanged { tool_name, .. }
7727                if is_browser_tool_name(&tool_name) =>
7728            {
7729                // This durable approval delta is host-owned. An actor-supplied
7730                // copy is neither authoritative nor safe to mirror to the
7731                // account feed; the host publishes its own record separately.
7732                return None;
7733            }
7734            AgentEvent::SubAgentEvent {
7735                parent_session_id,
7736                child_session_id,
7737                event,
7738            } => {
7739                if depth >= MAX_NESTED_DISPLAY_DEPTH || self.nested_overflowed {
7740                    return None;
7741                }
7742                if parent_session_id.is_empty()
7743                    || child_session_id.is_empty()
7744                    || parent_session_id.len() > MAX_DISPLAY_ID_BYTES
7745                    || child_session_id.len() > MAX_DISPLAY_ID_BYTES
7746                {
7747                    self.fail_closed();
7748                    return None;
7749                }
7750                let key = (parent_session_id.clone(), child_session_id.clone());
7751                if !self.children.contains_key(&key) && !self.reserve_identity() {
7752                    return None;
7753                }
7754                let identities = self.identities.clone();
7755                let child = self.children.entry(key).or_insert_with(|| {
7756                    Box::new(ActorEventDisplay {
7757                        identities,
7758                        ..Default::default()
7759                    })
7760                });
7761                let event = child.project(*event, depth + 1)?;
7762                if self.identities.load(Ordering::Relaxed) > MAX_DISPLAY_CALLS {
7763                    self.fail_closed();
7764                    return None;
7765                }
7766                AgentEvent::SubAgentEvent {
7767                    parent_session_id,
7768                    child_session_id,
7769                    event: Box::new(event),
7770                }
7771            }
7772            other => other,
7773        })
7774    }
7775}
7776
7777#[allow(clippy::too_many_arguments)]
7778async fn process_actor_event(
7779    event: serde_json::Value,
7780    strict_permission_events: bool,
7781    permission_handshake: &mut PermissionPostureHandshake,
7782    expected_permission_posture: Option<&ExpectedPermissionPosture>,
7783    session_inbox_runtime: Option<&SessionInboxRuntimeBinding>,
7784    logical_session: &mut Session,
7785    event_tx: &mpsc::Sender<AgentEvent>,
7786    display: &mut ActorEventDisplay,
7787) -> crate::runtime::runner::Result<()> {
7788    let event = match serde_json::from_value::<AgentEvent>(event) {
7789        Ok(event) => event,
7790        Err(_) if strict_permission_events => {
7791            // Serde errors can quote an unrecognized variant or field value.
7792            // The actor event is untrusted display input, including browser
7793            // download metadata, so never echo its decode details to SSE.
7794            return Err(AgentError::LLM(
7795                "actor emitted malformed AgentEvent under a typed permission posture contract"
7796                    .to_string(),
7797            ));
7798        }
7799        Err(_) => {
7800            // A legacy actor can omit a malformed ToolStart that reuses an
7801            // earlier public call ID. Its later result must not inherit the
7802            // stale public display classification.
7803            display.fail_closed();
7804            return Ok(());
7805        }
7806    };
7807    process_typed_actor_event(
7808        event,
7809        permission_handshake,
7810        expected_permission_posture,
7811        session_inbox_runtime,
7812        logical_session,
7813        event_tx,
7814        display,
7815    )
7816    .await
7817}
7818
7819#[allow(clippy::too_many_arguments)]
7820async fn process_typed_actor_event(
7821    event: AgentEvent,
7822    permission_handshake: &mut PermissionPostureHandshake,
7823    expected_permission_posture: Option<&ExpectedPermissionPosture>,
7824    session_inbox_runtime: Option<&SessionInboxRuntimeBinding>,
7825    logical_session: &mut Session,
7826    event_tx: &mpsc::Sender<AgentEvent>,
7827    display: &mut ActorEventDisplay,
7828) -> crate::runtime::runner::Result<()> {
7829    if matches!(&event, AgentEvent::PermissionPostureActivated { .. }) {
7830        if permission_handshake.posture_was_confirmed() {
7831            return Err(AgentError::LLM(
7832                "actor emitted a duplicate permission posture activation".to_string(),
7833            ));
7834        }
7835        let seed = permission_posture_seed_from_event(logical_session, &event)
7836            .map_err(AgentError::LLM)?
7837            .ok_or_else(|| {
7838                AgentError::LLM(
7839                    "actor permission posture event did not decode as a posture".to_string(),
7840                )
7841            })?;
7842        if let Some(expected) = expected_permission_posture {
7843            if seed.policy_revision != expected.policy_revision
7844                || seed.resolution != expected.resolution
7845            {
7846                return Err(AgentError::LLM(
7847                    "permission posture event does not match the host-dispatched policy"
7848                        .to_string(),
7849                ));
7850            }
7851            if seed.executor_mapping() != expected.executor_mapping {
7852                return Err(AgentError::LLM(
7853                    "permission posture event does not match the host-dispatched executor mapping"
7854                        .to_string(),
7855                ));
7856            }
7857        }
7858        if let Some(binding) = session_inbox_runtime {
7859            let saved = binding
7860                .persistence
7861                .record_permission_posture_activation(
7862                    &logical_session.id,
7863                    expected_permission_posture
7864                        .and_then(|expected| expected.expected_audit_revision),
7865                    &seed,
7866                )
7867                .await
7868                .map_err(|error| {
7869                    AgentError::LLM(format!(
7870                        "persist child permission posture bootstrap: {error}"
7871                    ))
7872                })?
7873                .ok_or_else(|| {
7874                    AgentError::LLM(
7875                        "persist child permission posture bootstrap: session not found".to_string(),
7876                    )
7877                })?;
7878            let snapshot = bamboo_domain::PermissionAuditSnapshot::from_metadata(&saved.metadata)
7879                .ok_or_else(|| {
7880                AgentError::LLM(
7881                    "persisted child permission posture audit is incomplete".to_string(),
7882                )
7883            })?;
7884            snapshot.write_to(&mut logical_session.metadata);
7885        } else {
7886            // In-memory/custom actor embeddings still use a host-owned clock.
7887            // Durable server paths always take the atomic branch.
7888            bamboo_domain::record_permission_audit(&mut logical_session.metadata, &seed, None)
7889                .map_err(|error| {
7890                    AgentError::LLM(format!(
7891                        "record in-memory child permission posture: {error}"
7892                    ))
7893                })?;
7894        }
7895        // Confirmation is deliberately last: matching alone is not enough.
7896        // The host-owned audit write must succeed first.
7897        *permission_handshake = PermissionPostureHandshake::Confirmed;
7898    } else if permission_handshake.is_awaiting() {
7899        return Err(AgentError::LLM(
7900            "actor emitted an execution event before permission posture confirmation".to_string(),
7901        ));
7902    }
7903    if let Some(event) = display.project(event, 0) {
7904        let _ = event_tx.send(event).await;
7905    }
7906    Ok(())
7907}
7908
7909/// Pump child frames -> parent events until a terminal frame (or cancellation).
7910/// On success, yields the actor's final result text (for session write-back).
7911/// `live_rx` carries in-band frames (steering messages) from the live registry.
7912///
7913/// `escalation_bridge` (#68) is the per-run escalation host bridge CAPTURED BY
7914/// VALUE at spawn time in `execute_external_child` (NOT read live here): when a
7915/// non-bypass child re-proxies an approval request, this owned bridge routes it
7916/// UP to the parent run. Owning it for the call's lifetime is what lets a
7917/// fire-and-forget grandchild that outlives its spawning run still escalate to
7918/// the correct (then-current) parent bridge rather than a stale/overwritten one.
7919async fn drive(context: ActorDriveContext<'_>) -> crate::runtime::runner::Result<Option<String>> {
7920    let ActorDriveContext {
7921        client,
7922        parent_session_id,
7923        child_session_id,
7924        child_attempt,
7925        approval_registry,
7926        approval_decider,
7927        approval_reviewer,
7928        escalation_bridge,
7929        event_tx,
7930        cancel_token,
7931        live_rx,
7932        delivery_rx,
7933        logical_session,
7934        mut expected_permission_posture,
7935        expected_creation,
7936        session_inbox_runtime,
7937        actor_directory_store,
7938        canonical_subagent_tool,
7939        ticket_service,
7940        activation_run_id,
7941        execution_epoch,
7942        expected_source_actor_id,
7943        initial_inflight_claims,
7944        first_frame_timeout,
7945        plain_actor,
7946        remote_environment_lease,
7947        mut readonly_output,
7948        local_history_tools,
7949        local_history_read_only,
7950        mut strict_history_output,
7951        plain_input,
7952        canonical_activation,
7953        canonical_placement_ref,
7954        actor_event_observer,
7955        plain_run,
7956    } = context;
7957
7958    // A projection belongs only to this Host-validated Run. A loaded Session
7959    // or a prior attempt may carry a stale marker; only the canonical wait
7960    // checks in the terminal branch below may set it again.
7961    logical_session
7962        .metadata
7963        .remove("runtime.actor_nested_wait_handoff");
7964
7965    // First-frame watchdog: a live worker emits its first frame (run-started /
7966    // first token) within seconds; total silence past the deadline means the
7967    // worker is dead (e.g. a pooled worker that exited right after checkout), so
7968    // its Run sits queued forever. We trip ONLY before the first frame — once any
7969    // frame arrives the worker is proven live and a legitimately long run (a slow
7970    // tool between tokens) never trips it.
7971    let mut got_first_frame = false;
7972    let mut first_frame_watch = first_frame_timeout.map(|d| Box::pin(tokio::time::sleep(d)));
7973    let mut inflight_claims = initial_inflight_claims;
7974    let mut owned_inputs = plain_input
7975        .map(|activation| VecDeque::from(activation.initial_inputs.clone()))
7976        .unwrap_or_default();
7977    // A retry/recovery or completed initial batch consumes its cold allowance;
7978    // only a fresh activation can continue twice. No lease/watchdog restarts.
7979    let continuation_limit = if owned_inputs.is_empty() { 2 } else { 1 };
7980    let mut continuations = usize::from(!owned_inputs.is_empty());
7981    let mut released_inputs: Vec<(
7982        SessionInboxOwnedClaim,
7983        bamboo_subagent::proto::InitialInputRelease,
7984    )> = Vec::new();
7985    let mut current_epoch = execution_epoch;
7986    let approval_epoch = Arc::new(AtomicU64::new(execution_epoch));
7987    let approval_cancel = cancel_token.child_token();
7988    let _approval_lifetime = approval_cancel.clone().drop_guard();
7989    let mut approval_scopes: HashMap<String, (serde_json::Value, ChildApprovalScope)> =
7990        HashMap::new();
7991    let mut owned_tree_requests = 0usize;
7992    let mut canonical_subagent_requests = 0usize;
7993    let mut committed_plan_calls = HashMap::new();
7994    // Once this Run registers a canonical Child wait, a fast completion may
7995    // clear it before the Worker sends Terminal. Keep the current Run
7996    // nonterminal so its already admitted outcome belongs to the successor.
7997    let mut nested_wait_registered_this_run = false;
7998    let mut parent_question_checkpoint_id: Option<String> = None;
7999    let strict_permission_events = expected_permission_posture.is_some();
8000    let mut permission_handshake =
8001        PermissionPostureHandshake::new(expected_permission_posture.as_ref());
8002    let mut next_actor_event_seq = 1u64;
8003    let mut canonical_router: Option<ActorEventRouter> = None;
8004    let canonical_fence = plain_input
8005        .map(|activation| &activation.fence)
8006        .or(canonical_activation);
8007    let canonical_store = plain_input
8008        .map(|activation| activation.store.as_ref())
8009        .or(actor_directory_store);
8010    let remote_claim_fence = actor_directory_store.zip(canonical_activation);
8011    let workflow_usage_requested = logical_session
8012        .metadata
8013        .get(bamboo_subagent::proto::WORKFLOW_USAGE_REQUESTED_KEY)
8014        .is_some_and(|v| v == "true");
8015    if workflow_usage_requested {
8016        logical_session
8017            .metadata
8018            .remove(bamboo_subagent::proto::WORKFLOW_USAGE_OBSERVATION_KEY);
8019        logical_session
8020            .metadata
8021            .remove(bamboo_subagent::proto::WORKFLOW_TERMINAL_OBSERVATION_KEY);
8022    }
8023    let mut remote_cancel_deadline: Option<tokio::time::Instant> = None;
8024    let mut display = ActorEventDisplay::default();
8025    let mut readonly = readonly_output
8026        .as_ref()
8027        .map(|_| ReadOnlyActorCollector::default());
8028    let mut local_history = local_history_tools.map(|_| LocalToolCollector::default());
8029    loop {
8030        tokio::select! {
8031            _ = cancel_token.cancelled(), if remote_cancel_deadline.is_none() => {
8032                if remote_claim_fence.is_some() || workflow_usage_requested {
8033                    // A remote Worker may finish after its subscriber cancels.
8034                    // Keep the same correlated frame pump alive briefly so an
8035                    // exact Cancelled Outcome can terminalize the Actor fence.
8036                    // Without that proof the activation stays live until its
8037                    // lease expires and a retry must remain blocked.
8038                    let _ = client.send(ParentFrame::Cancel).await;
8039                    remote_cancel_deadline = Some(
8040                        tokio::time::Instant::now() + Duration::from_secs(10)
8041                    );
8042                    continue;
8043                }
8044                // Direct/local links retain their existing cancel behavior.
8045                break;
8046            }
8047            _ = async {
8048                if let Some(deadline) = remote_cancel_deadline {
8049                    tokio::time::sleep_until(deadline).await;
8050                } else {
8051                    std::future::pending::<()>().await;
8052                }
8053            } => {
8054                return Err(AgentError::Cancelled);
8055            }
8056            _ = async {
8057                match first_frame_watch.as_mut() {
8058                    Some(s) => s.as_mut().await,
8059                    None => std::future::pending::<()>().await,
8060                }
8061            }, if !got_first_frame => {
8062                return Err(AgentError::WorkerUnresponsive(format!(
8063                    "child {child_session_id} produced no frame within {:?}",
8064                    first_frame_timeout.unwrap_or_default()
8065                )));
8066            }
8067            Some(_generation) = delivery_rx.recv(),
8068                if session_inbox_runtime.is_some() && activation_run_id.is_some() =>
8069            {
8070                if plain_input.is_some() {
8071                    // Keep the real Inbox pending until this Run's completed
8072                    // reply can be committed BEFORE its correction User.
8073                    continue;
8074                }
8075                forward_next_canonical_claim(
8076                    client,
8077                    session_inbox_runtime.expect("guarded"),
8078                    logical_session,
8079                    activation_run_id.expect("guarded"),
8080                    &mut inflight_claims,
8081                    remote_claim_fence,
8082                )
8083                .await?;
8084            }
8085            Some(frame) = live_rx.recv() => {
8086                if plain_actor { return Err(plain_actor_unsupported()); }
8087                // Forward in-band steering to the worker over the existing WS.
8088                if client.send(frame).await.is_err() {
8089                    tracing::warn!("live steering frame could not be sent; connection failing");
8090                }
8091            }
8092            frame = client.next_frame() => {
8093                // Any frame (event / approval / terminal / close / error) proves
8094                // the worker responded — disarm the first-frame watchdog.
8095                got_first_frame = true;
8096                first_frame_watch = None;
8097                match frame {
8098                    Ok(Some(ChildFrame::Event { event })) => {
8099                        if canonical_activation.is_some() {
8100                            return Err(AgentError::LLM("remote_actor_event_batch_required".into()));
8101                        }
8102                        if local_history.is_some() {
8103                            return Err(local_tool_history_unsupported());
8104                        }
8105                        if event.get("initial_input_control").is_some() {
8106                            let bamboo_subagent::proto::InitialInputControl::Request { request } =
8107                                bamboo_subagent::proto::InitialInputControl::decode(event).map_err(|_| plain_actor_unsupported())?
8108                            else { return Err(plain_actor_unsupported()); };
8109                            if !permission_handshake.posture_was_confirmed() || cancel_token.is_cancelled() { return Err(plain_actor_unsupported()); }
8110                            let activation = plain_input.ok_or_else(|| plain_actor_unsupported())?;
8111                            let binding = session_inbox_runtime.ok_or_else(|| plain_actor_unsupported())?;
8112                            let (run, _) = plain_run.ok_or_else(|| plain_actor_unsupported())?;
8113                            // Exact retransmission may refer to an earlier
8114                            // release while the next input is still waiting.
8115                            // Every new request must match the ordered front.
8116                            let previous = released_inputs.iter()
8117                                .find(|(_, release)| release.request == request);
8118                            let (claim, previous) = match previous {
8119                                Some((claim, release)) => (claim.clone(), Some(release.clone())),
8120                                None => (owned_inputs.front().cloned()
8121                                    .ok_or_else(|| plain_actor_unsupported())?, None),
8122                            };
8123                            let release = activation.release_initial_input(binding, logical_session, run, current_epoch, &claim,
8124                                &request, &mut expected_permission_posture, previous.as_ref()).await?;
8125                            if cancel_token.is_cancelled() { return Err(AgentError::Cancelled); }
8126                            client.send(ParentFrame::InitialInputRelease { release: release.clone() }).await
8127                                .map_err(|_| AgentError::LLM("Actor initial release unconfirmed; durable input is preserved".into()))?;
8128                            if previous.is_none() {
8129                                let _ = owned_inputs.pop_front();
8130                                released_inputs.push((claim, release));
8131                            }
8132                            continue;
8133                        }
8134                        if expected_creation.is_some() {
8135                            return Err(AgentError::LLM("worker omitted required Child creation event identity".into()));
8136                        }
8137                        if event["type"] == bamboo_subagent::proto::LocalToolMessages::TYPE { continue; }
8138                        if event["type"] == "owned_readonly_transcript" {
8139                            if plain_actor { return Err(plain_actor_unsupported()); }
8140                            continue;
8141                        }
8142                        if plain_actor && !plain_actor_event(&event)? { continue; }
8143                        if let Some(activation) = plain_input {
8144                            validate_directory_actor_event(
8145                                activation.store.as_ref(),
8146                                &activation.fence,
8147                                logical_session,
8148                                None,
8149                            ).await?;
8150                        }
8151                        // Rolling-upgrade compatibility: old actors have no
8152                        // route/sequence metadata, but retain the same typed
8153                        // permission handshake and event validation.
8154                        process_actor_event(
8155                            event,
8156                            strict_permission_events,
8157                            &mut permission_handshake,
8158                            expected_permission_posture.as_ref(),
8159                            session_inbox_runtime,
8160                            logical_session,
8161                            event_tx,
8162                            &mut display,
8163                        )
8164                        .await?;
8165                    }
8166                    Ok(Some(ChildFrame::EventBatch { batch })) => {
8167                        validate_actor_event_batch(
8168                            &batch,
8169                            logical_session,
8170                            parent_session_id,
8171                            activation_run_id,
8172                            current_epoch,
8173                            expected_source_actor_id,
8174                            expected_creation,
8175                        )?;
8176                        let directory_entry = if let (Some(store), Some(fence)) = (canonical_store, canonical_fence) {
8177                            let entry = validate_directory_actor_event(
8178                                store, fence, logical_session, batch.activation_id.as_deref(),
8179                            ).await?;
8180                            if canonical_placement_ref.is_some_and(|expected| {
8181                                entry.activation.as_ref().and_then(|activation| activation.placement_ref.as_ref()) != Some(expected)
8182                            }) {
8183                                return Err(AgentError::LLM("remote_actor_placement_ref_changed".into()));
8184                            }
8185                            Some(entry)
8186                        } else { None };
8187                        if let (Some(entry), Some(fence)) = (&directory_entry, canonical_fence) {
8188                            if canonical_router.is_none() {
8189                                canonical_router = Some(
8190                                    ActorEventRouter::new(
8191                                        entry,
8192                                        fence,
8193                                        current_epoch,
8194                                        strict_permission_events && permission_handshake.is_awaiting(),
8195                                    )
8196                                    .map_err(actor_event_route_error)?,
8197                                );
8198                            }
8199                        }
8200                        if local_history_tools.is_some()
8201                            && (batch.first_seq != next_actor_event_seq
8202                                || batch.qos != bamboo_subagent::ActorEventQos::Durable
8203                                || batch.last_seq == u64::MAX) {
8204                            return Err(local_tool_history_unsupported());
8205                        }
8206                        if batch.last_seq < next_actor_event_seq {
8207                            continue;
8208                        }
8209                        if batch.first_seq > next_actor_event_seq {
8210                            if plain_actor {
8211                                return Err(AgentError::LLM(
8212                                    "actor event sequence gap; authoritative session snapshot is required".into(),
8213                                ));
8214                            }
8215                            tracing::warn!(
8216                                child_session_id,
8217                                expected_seq = next_actor_event_seq,
8218                                received_seq = batch.first_seq,
8219                                "actor event sequence gap; authoritative session snapshot is required"
8220                            );
8221                            // A missing ToolStart may have reused an earlier
8222                            // call ID for a private browser download. No prior
8223                            // identity remains safe for display after a gap.
8224                            display.fail_closed();
8225                        }
8226                        let skip = next_actor_event_seq
8227                            .saturating_sub(batch.first_seq)
8228                            .min(batch.events.len() as u64) as usize;
8229                        for (offset, event) in batch.events.into_iter().enumerate().skip(skip) {
8230                            let seq = batch.first_seq + offset as u64;
8231                            let mut publish = true;
8232                            if event["type"] == bamboo_subagent::proto::WorkflowAgentUsage::TYPE {
8233                                if workflow_usage_requested {
8234                                    if permission_handshake.is_awaiting() { return Err(plain_actor_unsupported()); }
8235                                    let usage: bamboo_subagent::proto::WorkflowAgentUsage = serde_json::from_value(event.clone()).map_err(|_| plain_actor_unsupported())?;
8236                                    let activation = activation_run_id.ok_or_else(plain_actor_unsupported)?;
8237                                    if !usage.matches(logical_session, activation) { return Err(plain_actor_unsupported()); }
8238                                    if logical_session.metadata.contains_key(bamboo_subagent::proto::WORKFLOW_USAGE_OBSERVATION_KEY) { return Err(plain_actor_unsupported()); }
8239                                    logical_session.metadata.insert(bamboo_subagent::proto::WORKFLOW_USAGE_OBSERVATION_KEY.into(), serde_json::to_string(&usage).map_err(|_| plain_actor_unsupported())?);
8240                                }
8241                                publish = false;
8242                            }
8243                            if publish { if let Some(collector) = local_history.as_mut() {
8244                                if permission_handshake.is_awaiting() && event["type"] != "permission_posture_activated" {
8245                                    return Err(local_tool_history_unsupported());
8246                                }
8247                                if logical_session.metadata.contains_key(crate::ticket_worker_plan::TICKET_LOCAL_PLAN_KEY) {
8248                                    publish = collector.ticket_event(&event, &committed_plan_calls)?;
8249                                }
8250                                publish &= collector.event(&event)?;
8251                            } else if event["type"] == bamboo_subagent::proto::LocalToolMessages::TYPE {
8252                                // Completion DATA is private and grants nothing on an unselected route.
8253                                publish = false;
8254                            }}
8255                            if publish { publish = if let Some(collector) = readonly.as_mut() {
8256                                if permission_handshake.is_awaiting()
8257                                    && event["type"] != "permission_posture_activated" {
8258                                    return Err(plain_actor_unsupported());
8259                                }
8260                                collector.event(&event)?
8261                            } else if event["type"] == "owned_readonly_transcript" {
8262                                if plain_actor { return Err(plain_actor_unsupported()); }
8263                                false // A worker cache observation is not a legacy Host commit.
8264                            } else if plain_actor { plain_actor_event(&event)? } else { true };
8265                            }
8266                            if let (Some(router), Some(entry), Some(fence)) =
8267                                (canonical_router.as_mut(), directory_entry.as_ref(), canonical_fence)
8268                            {
8269                                let typed = if publish {
8270                                    Some(serde_json::from_value::<AgentEvent>(event.clone())
8271                                        .map_err(|_| plain_actor_unsupported())?)
8272                                } else {
8273                                    None
8274                                };
8275                                match router.route(entry, fence, current_epoch, seq, typed)
8276                                    .map_err(actor_event_route_error)?
8277                                {
8278                                    ActorEventRoute::Publish(envelope) => {
8279                                        let public_event = actor_event_observer
8280                                            .filter(|_| envelope.class != ActorEventClass::Ephemeral)
8281                                            .map(|_| PublicActorEvent::from(envelope.as_ref()));
8282                                        tracing::trace!(
8283                                            actor_id = %envelope.actor_id,
8284                                            event_id = %envelope.event_id,
8285                                            class = ?envelope.class,
8286                                            "publishing canonical Actor event"
8287                                        );
8288                                        process_typed_actor_event(
8289                                            envelope.payload,
8290                                            &mut permission_handshake,
8291                                            expected_permission_posture.as_ref(),
8292                                            session_inbox_runtime,
8293                                            logical_session,
8294                                            event_tx,
8295                                            &mut display,
8296                                        ).await?;
8297                                        if let (Some(observer), Some(public_event)) =
8298                                            (actor_event_observer, public_event)
8299                                        {
8300                                            observer.publish(public_event);
8301                                        }
8302                                    }
8303                                    ActorEventRoute::Duplicate | ActorEventRoute::Suppressed => {}
8304                                }
8305                                next_actor_event_seq = seq.saturating_add(1);
8306                                continue;
8307                            }
8308                            if !publish {
8309                                next_actor_event_seq = seq.saturating_add(1);
8310                                continue;
8311                            }
8312                            process_actor_event(
8313                                event,
8314                                strict_permission_events,
8315                                &mut permission_handshake,
8316                                expected_permission_posture.as_ref(),
8317                                session_inbox_runtime,
8318                                logical_session,
8319                                event_tx,
8320                                &mut display,
8321                            )
8322                            .await?;
8323                            next_actor_event_seq = seq.saturating_add(1);
8324                        }
8325                    }
8326                    Ok(Some(ChildFrame::OwnedTreeRequest { id, cursor })) => {
8327                        owned_tree_requests += 1;
8328                        if owned_tree_requests > 16 {
8329                            return Err(AgentError::LLM("owned tree request limit exceeded".into()));
8330                        }
8331                        let page = if id.len() <= 128
8332                            && !logical_session.metadata.contains_key(crate::ticket_worker_plan::TICKET_LOCAL_PLAN_KEY)
8333                            && cursor.as_ref().is_none_or(|value| !value.is_empty() && value.len() <= 128)
8334                            && !cancel_token.is_cancelled()
8335                        {
8336                            match (actor_directory_store, session_inbox_runtime,
8337                                activation_run_id, expected_creation) {
8338                                (Some(store), Some(binding), Some(run_id), Some(creation)) => {
8339                                    tokio::time::timeout(
8340                                        Duration::from_secs(8),
8341                                        inspect_active_owned_tree(
8342                                            store, binding, run_id, child_session_id,
8343                                            parent_session_id, logical_session, creation,
8344                                            plain_input, cursor.as_deref(),
8345                                        ),
8346                                    ).await.ok().flatten()
8347                                }
8348                                _ => {
8349                                    tracing::warn!("owned tree denied: canonical Host scope unavailable");
8350                                    None
8351                                },
8352                            }
8353                        } else {
8354                            tracing::warn!("owned tree denied: request bounds or cancellation");
8355                            None
8356                        };
8357                        if cancel_token.is_cancelled() {
8358                            return Err(AgentError::Cancelled);
8359                        }
8360                        client.send(ParentFrame::OwnedTreeReply { id, page }).await
8361                            .map_err(|_| AgentError::LLM("owned tree reply transport closed".into()))?;
8362                    }
8363                    Ok(Some(ChildFrame::SubAgentRequest { id, tool_call_id, args })) => {
8364                        canonical_subagent_requests += 1;
8365                        if canonical_subagent_requests > 64 {
8366                            return Err(AgentError::LLM("canonical SubAgent request limit exceeded".into()));
8367                        }
8368                        let valid_shape = !id.is_empty() && id.len() <= 128
8369                            && !tool_call_id.is_empty() && tool_call_id.len() <= 128
8370                            && args.is_object()
8371                            && serde_json::to_vec(&args)
8372                                .is_ok_and(|bytes| bytes.len() <= 64 * 1024);
8373                        let registers_wait = matches!(args.get("action").and_then(|value| value.as_str()),
8374                            Some("wait"))
8375                            || (args.get("action").and_then(|value| value.as_str()) == Some("create")
8376                                && args.get("wait").and_then(|value| value.as_bool()) == Some(true));
8377                        let checkpoint_payload = args.get(PARENT_QUESTION_CHECKPOINT_ACTION).cloned();
8378                        let ticket_bootstrap_pending = valid_shape
8379                            && logical_session.metadata.contains_key(crate::ticket_worker_plan::TICKET_LOCAL_PLAN_KEY)
8380                            && args == serde_json::json!({(crate::ticket_worker_plan::TICKET_PLAN_ACTION): {"read":true}})
8381                            && permission_handshake.is_awaiting()
8382                            && !cancel_token.is_cancelled();
8383                        let result = if ticket_bootstrap_pending {
8384                            // A read control can overtake the queued posture
8385                            // event on the separate lanes. Keep the handshake
8386                            // fence; the exact bootstrap read may retry briefly.
8387                            serde_json::json!({"error":crate::ticket_worker_plan::remote::TICKET_BOOTSTRAP_POSTURE_PENDING})
8388                        } else if valid_shape
8389                            && (!logical_session.metadata.contains_key(crate::ticket_worker_plan::TICKET_LOCAL_PLAN_KEY)
8390                                || args.get(crate::ticket_worker_plan::TICKET_PLAN_ACTION).is_some()
8391                                || args.get(crate::ticket_worker_plan::files::TICKET_FILE_ACTION).is_some())
8392                            && !cancel_token.is_cancelled()
8393                            && !permission_handshake.is_awaiting()
8394                        {
8395                            if args.get(crate::ticket_worker_plan::TICKET_PLAN_ACTION).is_some() {
8396                                match (ticket_service.as_deref(), actor_directory_store,
8397                                    session_inbox_runtime, activation_run_id, expected_creation) {
8398                                    (Some(service), Some(store), Some(binding), Some(run_id), Some(creation)) => {
8399                                        if let Some(caller) = load_active_subagent_caller(
8400                                            store, binding, run_id, child_session_id,
8401                                            parent_session_id, logical_session, creation, plain_input,
8402                                        ).await {
8403                                            match crate::ticket_worker_plan::remote::apply_host_plan_request(service, &caller, run_id, &args, &tool_call_id) {
8404                                                Ok(projection) => {
8405                                                    logical_session.task_list = projection.task_list.clone();
8406                                                    logical_session.set_task_list_version_meta(projection.plan_revision.to_string());
8407                                                    if let Some(question) = &projection.question {
8408                                                        let raw = serde_json::to_string(question).map_err(|_| local_tool_history_unsupported())?;
8409                                                        logical_session.metadata.insert(crate::ticket_worker_plan::TICKET_QUESTION_YIELD_KEY.into(), raw);
8410                                                        if let Some(collector) = local_history.as_mut() {
8411                                                            collector.ticket_yield = Some((tool_call_id.clone(), question.clone()));
8412                                                        }
8413                                                    }
8414                                                    if let Some(task) = args[crate::ticket_worker_plan::TICKET_PLAN_ACTION].get("task") {
8415                                                        committed_plan_calls.insert(tool_call_id.clone(), task.clone());
8416                                                        if let Some(task_list) = projection.task_list.clone() {
8417                                                            let _ = event_tx.send(AgentEvent::TaskListUpdated {
8418                                                                task_list,
8419                                                                version: Some(projection.plan_revision),
8420                                                            }).await;
8421                                                        }
8422                                                    }
8423                                                    serde_json::json!({"result":projection})
8424                                                }
8425                                                Err(error) => serde_json::json!({"error":error.to_string()}),
8426                                            }
8427                                        } else { serde_json::json!({"error":"LocalPlan Child Run is no longer current"}) }
8428                                    }
8429                                    _ => serde_json::json!({"error":"LocalPlan Host authority unavailable"}),
8430                                }
8431                            } else if args.get(crate::ticket_worker_plan::files::TICKET_FILE_ACTION).is_some() {
8432                                match (ticket_service.as_deref(), actor_directory_store, session_inbox_runtime,
8433                                    activation_run_id, expected_creation, local_history_tools) {
8434                                    (Some(service),Some(store),Some(binding),Some(run_id),Some(creation),Some(ceiling)) => {
8435                                        if let Some(caller)=load_active_subagent_caller(store,binding,run_id,child_session_id,parent_session_id,logical_session,creation,plain_input).await {
8436                                            match crate::ticket_worker_plan::files::apply_host_file_request(service,&caller,run_id,&args,&tool_call_id,ceiling) {
8437                                                Ok(reply)=>{
8438                                                    let body=&args[crate::ticket_worker_plan::files::TICKET_FILE_ACTION];
8439                                                    committed_plan_calls.insert(tool_call_id.clone(),serde_json::json!({"file_tool":body["tool"],"arguments":body["arguments"],"result":serde_json::to_string(&reply).map_err(|_|local_tool_history_unsupported())?}));
8440                                                    serde_json::json!({"result":reply})
8441                                                },
8442                                                Err(error)=>serde_json::json!({"error":error.to_string()}),
8443                                            }
8444                                        } else { serde_json::json!({"error":"Ticket file Child Run is no longer current"}) }
8445                                    },
8446                                    _=>serde_json::json!({"error":"Ticket file Host authority unavailable"}),
8447                                }
8448                            } else if let Some(payload) = checkpoint_payload {
8449                                let parsed = (args.as_object().is_some_and(|object| object.len() == 1))
8450                                    .then(|| serde_json::from_value::<ParentQuestionCheckpointV1>(payload).ok())
8451                                    .flatten()
8452                                    .filter(|observation| observation.tool_call_id == tool_call_id);
8453                                match (parsed, actor_directory_store, session_inbox_runtime,
8454                                    activation_run_id, expected_creation) {
8455                                    (Some(observation), Some(store), Some(binding), Some(run_id), Some(creation)) => {
8456                                        let receipt = tokio::time::timeout(Duration::from_secs(25),
8457                                            checkpoint_active_parent_question(
8458                                                store, binding, run_id, child_session_id,
8459                                                parent_session_id, logical_session, creation,
8460                                                plain_input, observation,
8461                                            )).await.ok().flatten();
8462                                        if let Some(request_id) = receipt {
8463                                            parent_question_checkpoint_id = Some(request_id.clone());
8464                                            serde_json::json!({"result":{"request_id":request_id}})
8465                                        } else {
8466                                            serde_json::json!({"error":"canonical Child question checkpoint rejected; persisted worker question is preserved"})
8467                                        }
8468                                    }
8469                                    _ => serde_json::json!({"error":"canonical Child question checkpoint unavailable"}),
8470                                }
8471                            } else {
8472                            match (canonical_subagent_tool.as_deref(), actor_directory_store,
8473                                session_inbox_runtime, activation_run_id, expected_creation) {
8474                                (Some(tool), Some(store), Some(binding), Some(run_id), Some(creation)) => {
8475                                    if let Some(caller) = load_active_subagent_caller(
8476                                        store, binding, run_id, child_session_id,
8477                                        parent_session_id, logical_session, creation,
8478                                        plain_input,
8479                                    ).await {
8480                                        tokio::time::timeout(
8481                                            Duration::from_secs(30),
8482                                            invoke_active_subagent(tool, &caller, args, &tool_call_id),
8483                                        ).await.unwrap_or_else(|_| serde_json::json!({
8484                                            "error":"canonical SubAgent operation timed out; inspect logical child state before retrying"
8485                                        }))
8486                                    } else {
8487                                        serde_json::json!({"error":"canonical SubAgent caller is no longer active"})
8488                                    }
8489                                }
8490                                _ => serde_json::json!({"error":"canonical SubAgent Host authority unavailable"}),
8491                            }
8492                            }
8493                        } else {
8494                            serde_json::json!({"error":"canonical SubAgent request rejected"})
8495                        };
8496                        if registers_wait
8497                            && result["result"]["success"] == true
8498                            && result["result"]["result"]
8499                                .as_str()
8500                                .and_then(|value| serde_json::from_str::<serde_json::Value>(value).ok())
8501                                .is_some_and(|value| value["status"] == "waiting")
8502                        {
8503                            nested_wait_registered_this_run = true;
8504                        }
8505                        if cancel_token.is_cancelled() {
8506                            return Err(AgentError::Cancelled);
8507                        }
8508                        client.send(ParentFrame::SubAgentReply { id, result }).await
8509                            .map_err(|_| AgentError::LLM("canonical SubAgent reply transport closed".into()))?;
8510                    }
8511                    Ok(Some(ChildFrame::ApprovalRequest { id, body })) => {
8512                        if logical_session.metadata.contains_key(crate::ticket_worker_plan::TICKET_LOCAL_PLAN_KEY) {
8513                            return Err(AgentError::LLM("Ticket Task-only Worker cannot request unrelated action authority".into()));
8514                        }
8515                        if plain_actor { return Err(plain_actor_unsupported()); }
8516                        if permission_handshake.is_awaiting() {
8517                            return Err(AgentError::LLM(
8518                                "actor requested approval before permission posture confirmation"
8519                                    .to_string(),
8520                            ));
8521                        }
8522                        // Phase 2: a worker proxied a gated-tool approval back to
8523                        // the host. The WORKER side is live — its executor installs
8524                        // a per-run task-local `ApprovalProxy` (subagent_worker.rs)
8525                        // that calls `host.approval_call`, so this frame arrives
8526                        // when a child hits `ConfirmationRequired`.
8527                        if let Some(reviewer) = approval_reviewer
8528                            .cloned()
8529                            .or_else(child_approval_reviewer)
8530                        {
8531                            // Phase 6, Part B: a BYPASSED parent worker
8532                            // model-reviews its children's forced-ask (dangerous)
8533                            // actions. The review is an LLM call, so run it OFF
8534                            // the frame pump in a spawned task and deliver the
8535                            // verdict async via the live channel — the pump keeps
8536                            // forwarding events and the agent loop never blocks. A
8537                            // timeout denies a hung review so the child can't hang.
8538                            let child = child_session_id.to_string();
8539                            let parent = parent_session_id.to_string();
8540                            let req_id = id.clone();
8541                            let body = body.clone();
8542                            let registry = approval_registry.cloned();
8543                            let Some((binding, run)) = session_inbox_runtime.zip(activation_run_id) else {
8544                                // Legacy/custom reviewers retain their original boolean seam;
8545                                // the canonical server reviewer denies without a live scope.
8546                                tokio::spawn(async move {
8547                                    let approved = tokio::time::timeout(CHILD_APPROVAL_TIMEOUT,
8548                                        reviewer.review(&parent, &child, &body)).await.unwrap_or(false);
8549                                    super::live::deliver_approval_scoped(registry.as_ref(), &child,
8550                                        child_attempt, &req_id, approved);
8551                                });
8552                                continue;
8553                            };
8554                            let generation = body.pointer("/permission_request/request_generation")
8555                                .and_then(serde_json::Value::as_str).unwrap_or("").to_string();
8556                            let scope = if let Some((original, scope)) = approval_scopes.get(&generation) {
8557                                if original != &body || scope.stamp()["reply"] != req_id {
8558                                    return Err(AgentError::LLM("approval generation changed in live scope".into()));
8559                                }
8560                                scope.clone()
8561                            } else {
8562                                if approval_scopes.len() >= 64 {
8563                                    return Err(AgentError::LLM("approval live scope limit exceeded".into()));
8564                                }
8565                                let scope = ChildApprovalScope::new(&parent, &child,
8566                                    (child_attempt, run, current_epoch, &req_id, chrono::Utc::now() + chrono::Duration::seconds(240)), binding.router.clone(),
8567                                    approval_cancel.clone(), approval_epoch.clone());
8568                                approval_scopes.insert(generation, (body.clone(), scope.clone()));
8569                                scope
8570                            };
8571                            tokio::spawn(async move {
8572                                let result = tokio::time::timeout(
8573                                    CHILD_APPROVAL_TIMEOUT,
8574                                    reviewer.review_scoped(&parent, &child, &body, &scope),
8575                                ).await.unwrap_or(ChildApprovalReview::Reply(false));
8576                                if let ChildApprovalReview::Reply(approved) = result {
8577                                    if scope.is_current(&parent, &child).await {
8578                                        super::live::deliver_approval_scoped(registry.as_ref(), &child,
8579                                            child_attempt, &req_id, approved && chrono::Utc::now() < scope.deadline());
8580                                    }
8581                                }
8582                            });
8583                        } else if approval_decider.is_some() {
8584                            // A decider is wired (policy / auto): decide promptly
8585                            // and reply inline. (Must not block the pump — see the
8586                            // `ChildApprovalDecider` doc.)
8587                            let approved =
8588                                decide_child_approval(approval_decider, child_session_id, &body)
8589                                    .await;
8590                            if client
8591                                .send(ParentFrame::ApprovalReply { id, approved })
8592                                .await
8593                                .is_err()
8594                            {
8595                                tracing::warn!(
8596                                    "failed to answer approval_request; connection failing"
8597                                );
8598                            }
8599                        } else if let Some(host) = escalation_bridge.clone() {
8600                            // Non-bypass WORKER: ESCALATE up our own actor link
8601                            // (re-proxy) so the request chains to our parent — and
8602                            // up every level until a bypass level or the top
8603                            // orchestrator's model reviewer decides. With no such
8604                            // reviewer the top level fails closed. Off-loop so the
8605                            // pump never blocks; relay the reply down to the child.
8606                            let child = child_session_id.to_string();
8607                            let req_id = id.clone();
8608                            let body = body.clone();
8609                            let registry = approval_registry.cloned();
8610                            tokio::spawn(async move {
8611                                let approved = match tokio::time::timeout(
8612                                    CHILD_APPROVAL_TIMEOUT,
8613                                    host.approval_call(body),
8614                                )
8615                                .await
8616                                {
8617                                    Ok(Ok(reply)) => reply
8618                                        .get("approved")
8619                                        .and_then(|v| v.as_bool())
8620                                        .unwrap_or(false),
8621                                    // Transport error or timeout ⇒ fail closed.
8622                                    _ => false,
8623                                };
8624                                super::live::deliver_approval_scoped(
8625                                    registry.as_ref(),
8626                                    &child,
8627                                    child_attempt,
8628                                    &req_id,
8629                                    approved,
8630                                );
8631                            });
8632                        } else {
8633                            // There is no parent-agent reviewer or upstream actor
8634                            // to own this decision. Never open a manual/UI approval
8635                            // path: forced-ask is parent-reviewed or fail-closed.
8636                            tracing::warn!(
8637                                parent_session_id,
8638                                child_session_id,
8639                                request_id = %id,
8640                                "forced-ask request has no parent-agent reviewer; denying"
8641                            );
8642                            if client
8643                                .send(ParentFrame::ApprovalReply {
8644                                    id,
8645                                    approved: false,
8646                                })
8647                                .await
8648                                .is_err()
8649                            {
8650                                tracing::warn!(
8651                                    "failed to send fail-closed approval reply; connection failing"
8652                                );
8653                            }
8654                        }
8655                    }
8656                    Ok(Some(ChildFrame::SessionMessageAdmitted { confirmation })) => {
8657                        if let Some(activation) = plain_input {
8658                            if activation.initial_release_required { return Err(plain_actor_unsupported()); }
8659                            let binding = session_inbox_runtime.ok_or_else(|| plain_actor_unsupported())?;
8660                            let run_id = activation_run_id.ok_or_else(|| plain_actor_unsupported())?;
8661                            let claim = owned_inputs.front().ok_or_else(|| plain_actor_unsupported())?;
8662                            activation.confirm_input(binding, logical_session, run_id, claim, &confirmation).await?;
8663                            let _ = owned_inputs.pop_front();
8664                            continue;
8665                        }
8666                        let Some(binding) = session_inbox_runtime else {
8667                            tracing::warn!(
8668                                child_session_id,
8669                                "ignoring SessionInbox confirmation without a runtime binding"
8670                            );
8671                            continue;
8672                        };
8673                        let Some(bound_run_id) = activation_run_id else {
8674                            tracing::warn!(
8675                                child_session_id,
8676                                "ignoring SessionInbox confirmation without an activation owner"
8677                            );
8678                            continue;
8679                        };
8680                        let Some(claim) = inflight_claims.front() else {
8681                            tracing::warn!(
8682                                child_session_id,
8683                                envelope_id = %confirmation.envelope_id,
8684                                "rejecting stale SessionInbox confirmation with no in-flight canonical claim"
8685                            );
8686                            continue;
8687                        };
8688                        let exact = confirmation.target_session_id == logical_session.id
8689                            && confirmation.envelope_id == claim.envelope.id.as_str()
8690                            && confirmation.canonical_claim_generation == claim.generation
8691                            && confirmation.activation_run_id == bound_run_id;
8692                        if !exact
8693                            || !binding
8694                                .router
8695                                .owns_run(&logical_session.id, bound_run_id)
8696                                .await
8697                        {
8698                            tracing::warn!(
8699                                child_session_id,
8700                                expected_target = %logical_session.id,
8701                                received_target = %confirmation.target_session_id,
8702                                expected_envelope_id = %claim.envelope.id,
8703                                received_envelope_id = %confirmation.envelope_id,
8704                                expected_generation = claim.generation,
8705                                received_generation = confirmation.canonical_claim_generation,
8706                                expected_run_id = bound_run_id,
8707                                received_run_id = %confirmation.activation_run_id,
8708                                "rejecting stale or mismatched SessionInbox admission confirmation"
8709                            );
8710                            continue;
8711                        }
8712                        let claim = inflight_claims
8713                            .pop_front()
8714                            .expect("validated in-flight canonical claim");
8715                        // On failure the durable canonical cur file remains
8716                        // recoverable for the next owner.
8717                        checkpoint_and_ack_canonical_claim(
8718                            binding,
8719                            logical_session,
8720                            &claim,
8721                            remote_claim_fence,
8722                        )
8723                            .await?;
8724                        // Ordered single-consumer: only after the exact prior
8725                        // claim is checkpointed+acked may the driver claim and
8726                        // forward the next envelope.
8727                        if inflight_claims.is_empty() {
8728                            forward_next_canonical_claim(
8729                                client,
8730                                binding,
8731                                logical_session,
8732                                bound_run_id,
8733                                &mut inflight_claims,
8734                                remote_claim_fence,
8735                            )
8736                            .await?;
8737                        }
8738                    }
8739                    Ok(Some(ChildFrame::Terminal { status, result, error, final_event_watermark, .. })) => {
8740                        if local_history_tools.is_some() {
8741                            // Check before history commit or terminal receipt acceptance.
8742                            // Normal lossy observations do not promise a complete trace.
8743                            validate_strict_history_terminal(
8744                                final_event_watermark.as_ref(), logical_session, parent_session_id,
8745                                activation_run_id, current_epoch, expected_creation,
8746                                next_actor_event_seq - 1,
8747                            )?;
8748                        }
8749                        if remote_cancel_deadline.is_some() && status == TerminalStatus::Cancelled {
8750                            if let Some(tools) = local_history_tools {
8751                                // The watermark proves transport completeness, not
8752                                // persistence. Certify only an entirely validated
8753                                // local Read/Glob prefix with no pending handoff.
8754                                if permission_handshake.is_awaiting()
8755                                    || !owned_inputs.is_empty()
8756                                    || !inflight_claims.is_empty()
8757                                    || parent_question_checkpoint_id.is_some()
8758                                    || nested_wait_registered_this_run
8759                                    || strict_history_output.is_none()
8760                                {
8761                                    return Err(local_tool_history_unsupported());
8762                                }
8763                                commit_local_tool_history(
8764                                    local_history.take().ok_or_else(local_tool_history_unsupported)?,
8765                                    logical_session,
8766                                    session_inbox_runtime.ok_or_else(local_tool_history_unsupported)?,
8767                                    activation_run_id.ok_or_else(local_tool_history_unsupported)?,
8768                                    (tools, local_history_read_only),
8769                                    LocalToolTerminal::Cancelled,
8770                                    cancel_token,
8771                                ).await?;
8772                                let proof = bamboo_storage::v2::HostTerminalCompleteness::from_verified_host(
8773                                    logical_session,
8774                                    activation_run_id.ok_or_else(local_tool_history_unsupported)?,
8775                                    current_epoch,
8776                                    next_actor_event_seq - 1,
8777                                ).map_err(|_| local_tool_history_unsupported())?;
8778                                *strict_history_output.as_deref_mut()
8779                                    .ok_or_else(local_tool_history_unsupported)? = Some(StrictTerminalHistory {
8780                                        proof,
8781                                        watermark: final_event_watermark.ok_or_else(local_tool_history_unsupported)?,
8782                                    });
8783                            }
8784                            // A selected Worker terminal proves that its Run
8785                            // ended even if permission posture or an Inbox
8786                            // correction was still pending. Claims remain
8787                            // unacked for retry; only the placement fence can
8788                            // now be released after SDK saves Cancelled.
8789                            client.accept_durable_terminal(status);
8790                            if workflow_usage_requested {
8791                                logical_session.metadata.insert(bamboo_subagent::proto::WORKFLOW_TERMINAL_OBSERVATION_KEY.into(), activation_run_id.ok_or_else(plain_actor_unsupported)?.into());
8792                            }
8793                            return Err(AgentError::Cancelled);
8794                        }
8795                        if parent_question_checkpoint_id.is_some()
8796                            && status != TerminalStatus::Suspended
8797                        {
8798                            return Err(AgentError::LLM(
8799                                "Child question checkpoint must end in a suspended Run".into(),
8800                            ));
8801                        }
8802                        if permission_handshake.is_awaiting() {
8803                            if remote_environment_lease && status == TerminalStatus::Error {
8804                                if let Some(code) = error.as_deref().and_then(remote_environment_error_code) {
8805                                    // The Worker rejected the EnvironmentLease before it
8806                                    // could confirm its permission posture. This exact
8807                                    // correlated Outcome still proves the Run ended, so
8808                                    // retain its durable receipt and release the slot after
8809                                    // the Host records the failure. Other early terminals
8810                                    // remain unproven and keep their placement fence.
8811                                    client.accept_durable_terminal(status);
8812                            if workflow_usage_requested {
8813                                logical_session.metadata.insert(bamboo_subagent::proto::WORKFLOW_TERMINAL_OBSERVATION_KEY.into(), activation_run_id.ok_or_else(plain_actor_unsupported)?.into());
8814                            }
8815                                    return Err(AgentError::LLM(code.to_owned()));
8816                                }
8817                            }
8818                            return Err(AgentError::LLM(
8819                                "actor terminated before permission posture confirmation"
8820                                    .to_string(),
8821                            ));
8822                        }
8823                        if !owned_inputs.is_empty() {
8824                            return Err(AgentError::LLM("Actor terminated before confirming its durable correction; Inbox claim is retained".into()));
8825                        }
8826                        if let Some(claim) = inflight_claims.front() {
8827                            return Err(AgentError::LLM(format!(
8828                                "actor terminated before durably admitting SessionInbox message {}; canonical claim remains recoverable",
8829                                claim.envelope.id
8830                            )));
8831                        }
8832                        if status == TerminalStatus::Completed {
8833                            if let Some(collector) = local_history.take() {
8834                                commit_local_tool_history(
8835                                    collector, logical_session,
8836                                    session_inbox_runtime.ok_or_else(local_tool_history_unsupported)?,
8837                                    activation_run_id.ok_or_else(local_tool_history_unsupported)?,
8838                                    (local_history_tools.ok_or_else(local_tool_history_unsupported)?, local_history_read_only),
8839                                    LocalToolTerminal::Completed(result.as_deref()), cancel_token,
8840                                ).await?;
8841                            }
8842                        }
8843                        if let Some(collector) = readonly.take() {
8844                            if status == TerminalStatus::Completed {
8845                                let activation = plain_input.ok_or_else(|| plain_actor_unsupported())?;
8846                                let backlog = activation.input_inbox.inspect(&logical_session.id).await
8847                                    .map_err(|_| plain_actor_unsupported())?;
8848                                if backlog.pending != 0 || backlog.claimed != 0 || backlog.activation_pending() {
8849                                    return Err(AgentError::LLM("Glob Actor continuation is unsupported; pending Inbox input is preserved".into()));
8850                                }
8851                                *readonly_output.as_deref_mut().ok_or_else(|| plain_actor_unsupported())? =
8852                                    collector.finish(result.as_deref())?;
8853                            }
8854                        } else if let Some(activation) = plain_input {
8855                            if status == TerminalStatus::Completed {
8856                                let binding = session_inbox_runtime.ok_or_else(|| plain_actor_unsupported())?;
8857                                let run_id = activation_run_id.ok_or_else(|| plain_actor_unsupported())?;
8858                                if continuations < continuation_limit {
8859                                    if let Some((claim, epoch)) = activation.continue_input(
8860                                        client, binding, logical_session, run_id, plain_run, &mut expected_permission_posture,
8861                                        result.clone().filter(|text| !text.is_empty()).ok_or_else(|| plain_actor_unsupported())?, event_tx,
8862                                    ).await? {
8863                                        owned_inputs = VecDeque::from([claim]);
8864                                        released_inputs.clear();
8865                                        continuations += 1;
8866                                        current_epoch = epoch;
8867                                        approval_epoch.store(epoch, Ordering::Release);
8868                                        // These are native coordinates scoped to the newly
8869                                        // correlated epoch. Public Host feed seq is untouched.
8870                                        next_actor_event_seq = 1;
8871                                        canonical_router = None;
8872                                        permission_handshake = PermissionPostureHandshake::new(expected_permission_posture.as_ref());
8873                                        display = ActorEventDisplay::default();
8874                                        continue;
8875                                    }
8876                                } else if activation.input_inbox.inspect(&logical_session.id).await
8877                                    .map_err(|_| plain_actor_unsupported())?.activation_pending() {
8878                                    if continuation_limit == 1 {
8879                                        return Err(AgentError::LLM("Only one Actor correction continuation is supported; remaining Inbox input is preserved".into()));
8880                                    }
8881                                    // The outer success path appends the last reply. An
8882                                    // explicit cap refusal must preserve it here instead,
8883                                    // without claiming/acking the excess input or saving
8884                                    // through the ordinary writer on this owned Actor.
8885                                    activation.append_reply(logical_session,
8886                                        result.clone().filter(|text| !text.is_empty()).ok_or_else(|| plain_actor_unsupported())?,
8887                                        event_tx).await?;
8888                                    return Err(AgentError::LLM("Actor correction continuation limit reached; completed reply committed and remaining Inbox input preserved".into()));
8889                                }
8890                            }
8891                        }
8892                        if status == TerminalStatus::Completed {
8893                            // Only an activation with a canonical Child creation
8894                            // identity can invoke the Host SubAgent route. Legacy
8895                            // Echo and other executors may share this runner's
8896                            // installed tool but have no nested wait to fence.
8897                            if let (Some(tool), Some(creation)) =
8898                                (canonical_subagent_tool.as_deref(), expected_creation)
8899                            {
8900                                let (Some(store), Some(binding), Some(run_id)) =
8901                                    (actor_directory_store, session_inbox_runtime,
8902                                        activation_run_id)
8903                                else {
8904                                    return Err(AgentError::LLM(
8905                                        "nested SubAgent Host authority unavailable at Child terminal"
8906                                            .into(),
8907                                    ));
8908                                };
8909                                let must_suspend = ensure_active_nested_wait(
8910                                    tool, store, binding, run_id, child_session_id,
8911                                    parent_session_id, logical_session, creation, plain_input,
8912                                ).await?;
8913                                if must_suspend || nested_wait_registered_this_run {
8914                                    // This is only a current-Run handoff fence.
8915                                    // sdk/spawn re-reads the durable Host wait;
8916                                    // copying a stale wait into this Session
8917                                    // would re-arm it after a fast grandchild
8918                                    // completion, while copying no wait would
8919                                    // erase an active registration on final save.
8920                                    logical_session.metadata.insert(
8921                                        "runtime.actor_nested_wait_handoff".into(),
8922                                        "true".into(),
8923                                    );
8924                                }
8925                            }
8926                        }
8927                        let terminal_result = match status {
8928                            TerminalStatus::Completed => Ok(result),
8929                            TerminalStatus::Cancelled => Err(AgentError::Cancelled),
8930                            TerminalStatus::Error => Err(AgentError::LLM(
8931                                error.unwrap_or_else(|| "actor child errored".to_string()),
8932                            )),
8933                            TerminalStatus::Suspended => {
8934                                use bamboo_agent_core::storage::Storage;
8935                                let (Some(request_id), Some(store), Some(binding), Some(run_id), Some(creation)) =
8936                                    (parent_question_checkpoint_id.as_ref(), actor_directory_store,
8937                                        session_inbox_runtime, activation_run_id, expected_creation)
8938                                else {
8939                                    return Err(AgentError::LLM("uncheckpointed Child suspension rejected".into()));
8940                                };
8941                                if load_active_subagent_caller(
8942                                    store, binding, run_id, child_session_id,
8943                                    parent_session_id, logical_session, creation, plain_input,
8944                                ).await.is_none() {
8945                                    return Err(AgentError::LLM("stale Child question suspension rejected".into()));
8946                                }
8947                                let id = bamboo_domain::SessionMessageId::parse(request_id.clone())
8948                                    .map_err(|_| AgentError::LLM("invalid Child question receipt".into()))?;
8949                                let child = store.load_session(child_session_id).await
8950                                    .map_err(|_| AgentError::LLM("canonical Child question unavailable".into()))?
8951                                    .ok_or_else(|| AgentError::LLM("canonical Child question missing".into()))?;
8952                                let parent = store.load_session(parent_session_id).await
8953                                    .map_err(|_| AgentError::LLM("canonical direct parent unavailable".into()))?
8954                                    .ok_or_else(|| AgentError::LLM("canonical direct parent missing".into()))?;
8955                                let pending = ParentQuestion::for_pending(&parent, &child)
8956                                    .is_some_and(|question| question.id == id);
8957                                let answered = bamboo_domain::ParentQuestionResolution::from_child(
8958                                    &parent, &child, &id,
8959                                ).is_some();
8960                                if !pending && !answered {
8961                                    return Err(AgentError::LLM("canonical Child question proof changed".into()));
8962                                }
8963                                logical_session.metadata.insert(
8964                                    "runtime.actor_parent_question_handoff".into(), "true".into(),
8965                                );
8966                                Ok(None)
8967                            }
8968                        };
8969                        if local_history_tools.is_some() {
8970                            // Failed strict Runs discard their collected suffix. A
8971                            // consumed watermark cannot certify those unapplied
8972                            // effects; retain every broker frame for recovery.
8973                            if status != TerminalStatus::Completed {
8974                                return terminal_result;
8975                            }
8976                            let proof = bamboo_storage::v2::HostTerminalCompleteness::from_verified_host(
8977                                logical_session,
8978                                activation_run_id.ok_or_else(local_tool_history_unsupported)?,
8979                                current_epoch,
8980                                next_actor_event_seq - 1,
8981                            ).map_err(|_| local_tool_history_unsupported())?;
8982                            let history = StrictTerminalHistory {
8983                                proof,
8984                                watermark: final_event_watermark.ok_or_else(local_tool_history_unsupported)?,
8985                            };
8986                            if let Some(output) = strict_history_output.as_deref_mut() {
8987                                *output = Some(history);
8988                            }
8989                        }
8990                        // This marker is set only after every Host terminal
8991                        // validation above succeeded. A surfaced broker Outcome
8992                        // that failed permission, Inbox, or question checks must
8993                        // never receive a durable ACK merely because SDK saved
8994                        // an error status. Genuine Worker Error/Cancelled frames
8995                        // are accepted terminals even though drive returns Err.
8996                        client.accept_durable_terminal(status);
8997                            if workflow_usage_requested {
8998                                logical_session.metadata.insert(bamboo_subagent::proto::WORKFLOW_TERMINAL_OBSERVATION_KEY.into(), activation_run_id.ok_or_else(plain_actor_unsupported)?.into());
8999                            }
9000                        if status == TerminalStatus::Completed {
9001                            if let Some(activation) = plain_input {
9002                                // Only the final epoch reaches here: continued Runs
9003                                // above reset their router and stay in the pump.
9004                                *activation.completion_router.lock().recover_poison() =
9005                                    canonical_router.take();
9006                            }
9007                        }
9008                        return terminal_result;
9009                    }
9010                    Ok(None) => {
9011                        return Err(AgentError::LLM(
9012                            "actor child closed before terminal".to_string(),
9013                        ));
9014                    }
9015                    Err(e) => {
9016                        return Err(AgentError::LLM(format!("actor transport error: {e}")));
9017                    }
9018                }
9019            }
9020        }
9021    }
9022
9023    // Only reached on cancellation: ask the child to stop (best-effort), then report cancelled.
9024    let _ = client.send(ParentFrame::Cancel).await;
9025    Err(AgentError::Cancelled)
9026}
9027
9028/// The assignment text = the child session's latest user message (falls back to its title).
9029fn project_id_for_actor_run(
9030    session: &Session,
9031) -> Result<Option<bamboo_domain::ProjectId>, AgentError> {
9032    match crate::project_context::ProjectContextResolver::session_project_identity(session) {
9033        crate::project_context::SessionProjectIdentity::Assigned(project_id) => {
9034            Ok(Some(project_id))
9035        }
9036        crate::project_context::SessionProjectIdentity::Unassigned => Ok(None),
9037        crate::project_context::SessionProjectIdentity::Invalid { raw, message } => {
9038            Err(AgentError::LLM(format!(
9039                "child session carries an invalid Project identity '{raw}': {message}"
9040            )))
9041        }
9042    }
9043}
9044
9045fn remote_broker_unavailable() -> AgentError {
9046    AgentError::LLM("remote_broker_activation_unavailable".into())
9047}
9048
9049/// Only these bounded worker admission errors may cross the fixed Remote
9050/// diagnostic boundary before permission posture confirmation. Transport and
9051/// arbitrary worker text retain the generic remote error.
9052fn remote_environment_error_code(value: &str) -> Option<&'static str> {
9053    const CODES: &[&str] = &[
9054        "remote_environment_lease_invalid",
9055        "remote_environment_snapshot_mismatch",
9056        "remote_environment_checkout_not_clean",
9057        "remote_environment_checkout_unsupported",
9058        "remote_environment_workspace_unavailable",
9059        "remote_environment_workspace_outside_git_root",
9060        "remote_environment_path_unsupported",
9061        "remote_environment_git_unavailable",
9062        "remote_environment_git_inspection_failed",
9063        "remote_environment_git_inspection_timed_out",
9064        "remote_environment_git_output_too_large",
9065        "remote_environment_snapshot_too_large",
9066        "remote_environment_snapshot_changed",
9067        "remote_environment_index_invalid",
9068        "remote_environment_index_unsupported",
9069        "remote_environment_host_path_or_grants_rejected",
9070        "remote_environment_worker_workspace_missing",
9071        "remote_environment_runspec_invalid",
9072        "remote_environment_lease_missing_or_host_path_present",
9073    ];
9074    CODES.iter().copied().find(|code| *code == value)
9075}
9076
9077// Read only: no ensure/repair/Actor initialization or subtree budget. Capture
9078// exactly the canonical chain and compare it again immediately before Run.
9079async fn remote_canonical_lineage(
9080    binding: &SessionInboxRuntimeBinding,
9081    session: &Session,
9082) -> Result<Vec<bamboo_domain::ActorSession>, AgentError> {
9083    let mut current = bamboo_domain::ActorSession::from_session(session)
9084        .map_err(|_| remote_broker_unavailable())?;
9085    let mut chain = Vec::new();
9086    loop {
9087        if chain.len() > session.spawn_depth as usize {
9088            return Err(remote_broker_unavailable());
9089        }
9090        let saved = binding
9091            .storage
9092            .load_session(&current.actor_id)
9093            .await
9094            .map_err(|_| remote_broker_unavailable())?
9095            .ok_or_else(remote_broker_unavailable)?;
9096        let observed = bamboo_domain::ActorSession::from_session(&saved)
9097            .map_err(|_| remote_broker_unavailable())?;
9098        if observed.actor_id != current.actor_id
9099            || observed.session_created_at != current.session_created_at
9100            || observed.parent_actor_id != current.parent_actor_id
9101            || observed.root_actor_id != current.root_actor_id
9102            || observed.spawn_depth != current.spawn_depth
9103            || observed.project_id != current.project_id
9104        {
9105            return Err(remote_broker_unavailable());
9106        }
9107        chain.push(observed.clone());
9108        let Some(parent_id) = observed.parent_actor_id.as_ref() else {
9109            if observed.actor_id != session.root_session_id || observed.spawn_depth != 0 {
9110                return Err(remote_broker_unavailable());
9111            }
9112            return Ok(chain);
9113        };
9114        let parent = binding
9115            .storage
9116            .load_session(parent_id)
9117            .await
9118            .map_err(|_| remote_broker_unavailable())?
9119            .ok_or_else(remote_broker_unavailable)?;
9120        let parent = bamboo_domain::ActorSession::from_session(&parent)
9121            .map_err(|_| remote_broker_unavailable())?;
9122        if parent.root_actor_id != observed.root_actor_id
9123            || parent.project_id != observed.project_id
9124            || parent.spawn_depth.checked_add(1) != Some(observed.spawn_depth)
9125            || parent.session_created_at > observed.session_created_at
9126        {
9127            return Err(remote_broker_unavailable());
9128        }
9129        current = parent;
9130    }
9131}
9132
9133async fn canonical_child_creation(
9134    binding: &SessionInboxRuntimeBinding,
9135    session: &Session,
9136    job: &SpawnJob,
9137) -> Result<bamboo_subagent::proto::ChildCreationIdentity, AgentError> {
9138    let conflict = || {
9139        AgentError::LLM(
9140            "child creation identity unavailable or changed in canonical host Storage".into(),
9141        )
9142    };
9143    let current = binding
9144        .storage
9145        .load_session(&job.child_session_id)
9146        .await
9147        .map_err(|_| conflict())?
9148        .ok_or_else(conflict)?;
9149    if bamboo_domain::ActorSession::from_session(&current).is_err()
9150        || bamboo_domain::ActorSession::from_session(session).is_err()
9151        || current.kind != bamboo_domain::SessionKind::Child
9152        || session.kind != current.kind
9153        || current.id != session.id
9154        || current.id != job.child_session_id
9155        || current.parent_session_id.as_deref() != Some(job.parent_session_id.as_str())
9156        || current.parent_session_id != session.parent_session_id
9157        || current.root_session_id.trim().is_empty()
9158        || current.root_session_id != session.root_session_id
9159        || current.created_at != session.created_at
9160        || current.spawn_depth != session.spawn_depth
9161        || current.spawn_depth == 0
9162        || project_id_for_actor_run(&current)? != project_id_for_actor_run(session)?
9163    {
9164        return Err(conflict());
9165    }
9166    Ok(bamboo_subagent::proto::ChildCreationIdentity {
9167        created_at: current.created_at,
9168        spawn_depth: current.spawn_depth,
9169    })
9170}
9171
9172fn logical_identity_for_actor_run(session: &Session, job: &SpawnJob) -> LogicalSessionIdentity {
9173    LogicalSessionIdentity {
9174        creation: None,
9175        session_id: session.id.clone(),
9176        parent_session_id: session
9177            .parent_session_id
9178            .clone()
9179            .or_else(|| Some(job.parent_session_id.clone())),
9180        root_session_id: if session.root_session_id.trim().is_empty() {
9181            job.parent_session_id.clone()
9182        } else {
9183            session.root_session_id.clone()
9184        },
9185    }
9186}
9187
9188fn extract_assignment(session: &Session) -> String {
9189    session
9190        .messages
9191        .iter()
9192        .rev()
9193        .find(|m| matches!(m.role, Role::User))
9194        .map(|m| m.content.clone())
9195        .unwrap_or_else(|| {
9196            session
9197                .metadata
9198                .get("title")
9199                .cloned()
9200                .unwrap_or_else(|| "Execute task".to_string())
9201        })
9202}
9203
9204#[cfg(test)]
9205mod tests {
9206    use super::*;
9207    use crate::SessionActivationRouter;
9208    use bamboo_domain::{RuntimeSessionPersistence, SessionInboxPort, Storage};
9209
9210    include!("expired_plain_owner_tests.rs");
9211
9212    include!("owned_cold_input_tests.rs");
9213
9214    #[test]
9215    fn plain_rejection_logs_branch_and_run_without_changing_error() {
9216        // Keep shared callsites consulting the current subscriber during scoped capture.
9217        let capture_interest_guard = tracing::Dispatch::new(
9218            tracing_subscriber::fmt()
9219                .with_max_level(tracing_subscriber::filter::LevelFilter::OFF)
9220                .with_writer(std::io::sink)
9221                .finish(),
9222        );
9223        let output = tempfile::NamedTempFile::new().unwrap();
9224        let subscriber = tracing_subscriber::fmt()
9225            .with_ansi(false)
9226            .without_time()
9227            .with_writer(output.reopen().unwrap())
9228            .finish();
9229        let branch_line = std::cell::Cell::new(0);
9230        let error = tracing::subscriber::with_default(subscriber, || {
9231            // Exercise the shared warning callsite without this thread's scoped subscriber.
9232            std::thread::spawn(|| drop(plain_actor_unsupported()))
9233                .join()
9234                .expect("unscoped rejection warning thread panicked");
9235            let span = tracing::warn_span!(
9236                "actor_child_run",
9237                child_session_id = %"diagnostic-child",
9238                activation_run_id = tracing::field::Empty
9239            );
9240            let _entered = span.enter();
9241            span.record("activation_run_id", "diagnostic-run");
9242            None::<()>
9243                .ok_or_else(|| {
9244                    let caller_line = line!() + 1;
9245                    let error = plain_actor_unsupported();
9246                    branch_line.set(caller_line);
9247                    error
9248                })
9249                .unwrap_err()
9250        });
9251        assert!(matches!(&error, AgentError::LLM(message)
9252            if message.starts_with("This Actor Child supports a fresh plain activation,")
9253                && message.ends_with("Other continuation is unsupported; durable history is preserved.")));
9254        let log = std::fs::read_to_string(output.path()).unwrap();
9255        assert!(log.contains("child_session_id=diagnostic-child"), "{log}");
9256        assert!(
9257            log.contains("activation_run_id=\"diagnostic-run\""),
9258            "{log}"
9259        );
9260        assert!(
9261            log.contains(&format!("process_id={}", std::process::id())),
9262            "{log}"
9263        );
9264        assert!(log.contains(&format!("source_file={:?}", file!())), "{log}");
9265        assert!(
9266            log.contains(&format!("source_line={}", branch_line.get())),
9267            "{log}"
9268        );
9269        assert!(!error.to_string().contains("diagnostic-run"));
9270        drop(capture_interest_guard);
9271    }
9272
9273    #[tokio::test]
9274    async fn completed_plain_continuation_requires_bounded_inputs_and_current_identity() {
9275        for case in [
9276            "valid",
9277            "over-batch-limit",
9278            "tool-history",
9279            "stale-lineage",
9280            "no-release",
9281        ] {
9282            let temp = tempfile::tempdir().unwrap();
9283            let store = Arc::new(
9284                bamboo_storage::SessionStoreV2::new(temp.path().into())
9285                    .await
9286                    .unwrap(),
9287            );
9288            let parent = Session::new("continuation-parent", "model");
9289            store.save_session(&parent).await.unwrap();
9290            let mut child =
9291                Session::new_child_of("continuation-child", &parent, "model", "first task");
9292            child.add_message(bamboo_agent_core::Message::assistant("first result", None));
9293            if case == "tool-history" {
9294                let mut tool = bamboo_agent_core::Message::assistant("old tool result", None);
9295                tool.role = Role::Tool;
9296                child.add_message(tool);
9297            }
9298            child.set_last_run_status("completed");
9299            store.save_session(&child).await.unwrap();
9300            store.ensure_actor(&child.id).await.unwrap();
9301            let policy = Arc::new(bamboo_tools::permission::PermissionConfig::new());
9302            bind_local_control_plane(store.as_ref(), &child.id, policy.as_ref()).await;
9303            let now = chrono::Utc::now();
9304            let old = store
9305                .claim_activation(&ActorActivationClaim {
9306                    actor_id: child.id.clone(),
9307                    run_id: "first-run".into(),
9308                    lease_owner: "first-owner".into(),
9309                    lease_expires_at: now + chrono::Duration::minutes(1),
9310                    inbox_generation: 0,
9311                    placement_ref: None,
9312                    now,
9313                })
9314                .await
9315                .unwrap();
9316            store.start_activation(&old.fence(), now).await.unwrap();
9317            store
9318                .finish_activation(&old.fence(), now, ActorActivationFinish::Succeeded)
9319                .await
9320                .unwrap();
9321            let inbox: Arc<dyn SessionInboxPort> = Arc::new(bamboo_storage::FileSessionInbox::new(
9322                store.clone(),
9323                Default::default(),
9324            ));
9325            for slot in 0..if case == "over-batch-limit" {
9326                bamboo_domain::SessionInboxLimits::default().max_claim_batch + 1
9327            } else {
9328                1
9329            } {
9330                let envelope = bamboo_domain::SessionMessageEnvelope::user_input(
9331                    &child.id,
9332                    format!("new task {slot}"),
9333                );
9334                let receipt = inbox.deliver(&envelope).await.unwrap();
9335                inbox
9336                    .mark_activation_eligible(
9337                        &child.id,
9338                        receipt.generation,
9339                        bamboo_domain::SessionActivationPolicy::InterruptSpecificWait,
9340                    )
9341                    .await
9342                    .unwrap();
9343            }
9344            let storage: Arc<dyn Storage> = store.clone();
9345            let persistence = Arc::new(bamboo_storage::LockedSessionStore::new(storage));
9346            let binding = actor_binding(store.clone(), inbox, persistence);
9347            let _registration = binding
9348                .router
9349                .register_run(&child.id, "second-run")
9350                .await
9351                .unwrap();
9352            if case == "stale-lineage" {
9353                child.parent_session_id = Some("foreign-parent".into());
9354            }
9355            let result = PlainActorActivation::start(
9356                store.clone(),
9357                &child,
9358                &binding,
9359                Some("second-run"),
9360                Some(policy),
9361                true,
9362                (case != "no-release").then_some("second-worker"),
9363            )
9364            .await;
9365            if case == "valid" {
9366                let activation = result.expect("one new input on a completed plain Child");
9367                assert_eq!(activation.fence.attempt, old.attempt + 1);
9368                assert_eq!(activation.fence.actor_id, child.id);
9369                assert_eq!(activation.inbox_generation, 1);
9370                activation
9371                    .finish(ActorActivationFinish::Failed)
9372                    .await
9373                    .unwrap();
9374            } else {
9375                assert!(
9376                    result.is_err(),
9377                    "unsupported continuation {case} must fail before claim"
9378                );
9379                assert_eq!(
9380                    store
9381                        .inspect_actor(&child.id)
9382                        .await
9383                        .unwrap()
9384                        .actor
9385                        .current_attempt,
9386                    old.attempt
9387                );
9388            }
9389            let canonical = store.load_session(&child.id).await.unwrap().unwrap();
9390            assert_eq!(
9391                canonical.parent_session_id.as_deref(),
9392                Some(parent.id.as_str())
9393            );
9394            assert!(canonical
9395                .messages
9396                .iter()
9397                .any(|message| message.content == "first result"));
9398        }
9399    }
9400
9401    #[tokio::test]
9402    async fn pre_dispatch_placement_failure_fails_only_claimed_activation() {
9403        let temp = tempfile::tempdir().unwrap();
9404        let store = bamboo_storage::SessionStoreV2::new(temp.path().into())
9405            .await
9406            .unwrap();
9407        let parent = Session::new("placement-parent", "model");
9408        let child = Session::new_child_of("placement-child", &parent, "model", "task");
9409        store.save_session(&parent).await.unwrap();
9410        store.save_session(&child).await.unwrap();
9411        store.ensure_actor(&child.id).await.unwrap();
9412        store
9413            .bind_control_plane(
9414                &child.id,
9415                bamboo_domain::ActorControlPlaneBinding {
9416                    policy_revision: None,
9417                    placement_intent: Some(bamboo_domain::ActorPlacementIntent::Pool {
9418                        pool_id: "remote-worker".into(),
9419                    }),
9420                },
9421            )
9422            .await
9423            .unwrap();
9424
9425        let claim = |run_id: &str| ActorActivationClaim {
9426            actor_id: child.id.clone(),
9427            run_id: run_id.into(),
9428            lease_owner: format!("host-{run_id}"),
9429            lease_expires_at: chrono::Utc::now() + chrono::Duration::seconds(80),
9430            inbox_generation: 0,
9431            placement_ref: Some(bamboo_domain::ActorPlacementRef {
9432                class: bamboo_domain::ActorPlacementClass::Remote,
9433                lease_id: format!("slot-{run_id}"),
9434                slot_epoch: Some(1),
9435            }),
9436            now: chrono::Utc::now(),
9437        };
9438        let first = store.claim_activation(&claim("first")).await.unwrap();
9439        store
9440            .start_activation(&first.fence(), chrono::Utc::now())
9441            .await
9442            .unwrap();
9443        assert!(store.claim_activation(&claim("retry")).await.is_err());
9444
9445        // This is the post-claim placement preflight failure path. No Run has
9446        // been dispatched, so the same ActorId can retry before lease expiry.
9447        fail_pre_dispatch_activation(&store, &first.fence())
9448            .await
9449            .unwrap();
9450        let failed = store.inspect_actor(&child.id).await.unwrap();
9451        assert_eq!(
9452            failed.activation.unwrap().status,
9453            bamboo_domain::ActorActivationStatus::Failed
9454        );
9455        let retry = store.claim_activation(&claim("retry")).await.unwrap();
9456        assert_eq!(retry.attempt, first.attempt + 1);
9457        assert!(fail_pre_dispatch_activation(&store, &first.fence())
9458            .await
9459            .is_err());
9460        assert_eq!(
9461            store.inspect_actor(&child.id).await.unwrap().activation,
9462            Some(retry)
9463        );
9464    }
9465
9466    #[tokio::test]
9467    async fn lost_run_delivery_receipt_keeps_actor_activation_live() {
9468        struct AcceptedWithoutReceipt {
9469            published: bool,
9470        }
9471
9472        #[async_trait]
9473        impl bamboo_subagent::ChildLink for AcceptedWithoutReceipt {
9474            async fn send(&mut self, frame: ParentFrame) -> bamboo_subagent::TransportResult<()> {
9475                assert!(matches!(frame, ParentFrame::Run(_)));
9476                self.published = true;
9477                Err(bamboo_subagent::TransportError::Protocol(
9478                    "broker delivery receipt lost after publication".into(),
9479                ))
9480            }
9481
9482            async fn next_frame(&mut self) -> bamboo_subagent::TransportResult<Option<ChildFrame>> {
9483                std::future::pending().await
9484            }
9485        }
9486
9487        let temp = tempfile::tempdir().unwrap();
9488        let store = Arc::new(
9489            bamboo_storage::SessionStoreV2::new(temp.path().into())
9490                .await
9491                .unwrap(),
9492        );
9493        let parent = Session::new("dispatch-parent", "model");
9494        let child = Session::new_child_of("dispatch-child", &parent, "model", "task");
9495        store.save_session(&parent).await.unwrap();
9496        store.save_session(&child).await.unwrap();
9497        store.ensure_actor(&child.id).await.unwrap();
9498        let now = chrono::Utc::now();
9499        let claim = ActorActivationClaim {
9500            actor_id: child.id.clone(),
9501            run_id: "dispatch-run".into(),
9502            lease_owner: "dispatch-host".into(),
9503            lease_expires_at: now + chrono::Duration::seconds(80),
9504            inbox_generation: 0,
9505            placement_ref: None,
9506            now,
9507        };
9508        let activation = store.claim_activation(&claim).await.unwrap();
9509        store
9510            .start_activation(&activation.fence(), chrono::Utc::now())
9511            .await
9512            .unwrap();
9513        let mut link = AcceptedWithoutReceipt { published: false };
9514        let mut send_invoked = false;
9515        let run = RunSpec {
9516            assignment: "task".into(),
9517            logical_session: None,
9518            project_id: None,
9519            reasoning_effort: None,
9520            permission_policy: None,
9521            messages: Vec::new(),
9522            activation_run_id: Some(claim.run_id.clone()),
9523            execution_epoch: 1,
9524            initial_session_messages: Vec::new(),
9525            secrets: Default::default(),
9526        };
9527        assert!(send_actor_run(&mut link, run, &mut send_invoked)
9528            .await
9529            .is_err());
9530        assert!(send_invoked && link.published);
9531        let slot = WorkerSlotLease {
9532            lease_id: "dispatch-slot".into(),
9533            host_ref: "dispatch-worker".into(),
9534            connection_generation: "dispatch-generation".into(),
9535            slot: 0,
9536            epoch: 1,
9537            actor_id: child.id.clone(),
9538            run_id: claim.run_id.clone(),
9539            expires_at: claim.lease_expires_at,
9540        };
9541        let mut lease = ActivePlacementLease {
9542            registry: bamboo_storage::v2::FileHostRegistry::new(temp.path().into())
9543                .await
9544                .unwrap(),
9545            actor_store: store.clone(),
9546            route: ScopedPlacementRoute {
9547                endpoint: String::new(),
9548                token: String::new(),
9549                ca_cert_file: PathBuf::new(),
9550                parent: bamboo_subagent::AgentRef {
9551                    session_id: "dispatch-parent".into(),
9552                    role: Some("host".into()),
9553                },
9554                worker_role: "worker".into(),
9555                pinned_mailbox: None,
9556                requirements: Default::default(),
9557            },
9558            mailbox: "dispatch-worker".into(),
9559            host_ref: "dispatch-worker".into(),
9560            fence: activation.fence(),
9561            placement_ref: bamboo_domain::ActorPlacementRef {
9562                class: bamboo_domain::ActorPlacementClass::Remote,
9563                lease_id: slot.lease_id.clone(),
9564                slot_epoch: Some(slot.epoch),
9565            },
9566            lease: Arc::new(tokio::sync::Mutex::new(slot)),
9567            heartbeat: tokio::spawn(std::future::pending()),
9568            run_cancel: CancellationToken::new(),
9569            lost: Arc::new(AtomicBool::new(false)),
9570            release_on_drop: AtomicBool::new(false),
9571        };
9572        assert!(!lease.abort_if_run_unsent(send_invoked).await.unwrap());
9573        assert_eq!(
9574            store
9575                .inspect_actor(&child.id)
9576                .await
9577                .unwrap()
9578                .activation
9579                .unwrap()
9580                .status,
9581            bamboo_domain::ActorActivationStatus::Running
9582        );
9583        let mut retry = claim.clone();
9584        retry.run_id = "dispatch-retry".into();
9585        retry.now = chrono::Utc::now();
9586        retry.lease_expires_at = retry.now + chrono::Duration::seconds(80);
9587        assert!(matches!(
9588            store.claim_activation(&retry).await,
9589            Err(bamboo_domain::ActorDirectoryError::Busy)
9590        ));
9591    }
9592
9593    #[test]
9594    fn accepted_cancelled_terminal_maps_to_cancelled_actor_activation() {
9595        let cancelled: Result<Option<String>, AgentError> = Err(AgentError::Cancelled);
9596        assert_eq!(
9597            remote_activation_finish(&cancelled, Some(TerminalStatus::Cancelled)),
9598            ActorActivationFinish::Cancelled
9599        );
9600        assert_eq!(
9601            remote_activation_finish(&cancelled, None),
9602            ActorActivationFinish::Failed
9603        );
9604        assert_eq!(
9605            remote_activation_finish(&cancelled, Some(TerminalStatus::Completed)),
9606            ActorActivationFinish::Failed
9607        );
9608    }
9609
9610    #[tokio::test]
9611    async fn remote_reply_is_fenced_before_terminal_status_save() {
9612        let temp = tempfile::tempdir().unwrap();
9613        let store = bamboo_storage::SessionStoreV2::new(temp.path().into())
9614            .await
9615            .unwrap();
9616        let parent = Session::new("remote-reply-parent", "model");
9617        let mut child = Session::new_child_of("remote-reply-child", &parent, "model", "task");
9618        child.add_message(bamboo_agent_core::Message::user("task"));
9619        store.save_session(&parent).await.unwrap();
9620        store.save_session(&child).await.unwrap();
9621        store.ensure_actor(&child.id).await.unwrap();
9622        store
9623            .bind_control_plane(
9624                &child.id,
9625                bamboo_domain::ActorControlPlaneBinding {
9626                    policy_revision: None,
9627                    placement_intent: Some(bamboo_domain::ActorPlacementIntent::Pool {
9628                        pool_id: "remote-worker".into(),
9629                    }),
9630                },
9631            )
9632            .await
9633            .unwrap();
9634        let now = chrono::Utc::now();
9635        let activation = store
9636            .claim_activation(&ActorActivationClaim {
9637                actor_id: child.id.clone(),
9638                run_id: "remote-reply-run".into(),
9639                lease_owner: "remote-worker".into(),
9640                lease_expires_at: now + chrono::Duration::seconds(80),
9641                inbox_generation: 0,
9642                placement_ref: Some(bamboo_domain::ActorPlacementRef {
9643                    class: bamboo_domain::ActorPlacementClass::Remote,
9644                    lease_id: "remote-slot".into(),
9645                    slot_epoch: Some(1),
9646                }),
9647                now,
9648            })
9649            .await
9650            .unwrap();
9651        let fence = activation.fence();
9652        store.start_activation(&fence, now).await.unwrap();
9653
9654        append_fenced_remote_reply(&store, &fence, &mut child, "done")
9655            .await
9656            .unwrap();
9657        store
9658            .finish_activation(&fence, chrono::Utc::now(), ActorActivationFinish::Succeeded)
9659            .await
9660            .unwrap();
9661        child.set_last_run_status("completed");
9662        store.save_session(&child).await.unwrap();
9663        let saved = store.load_session(&child.id).await.unwrap().unwrap();
9664        assert_eq!(saved.last_run_status().as_deref(), Some("completed"));
9665        assert_eq!(saved.messages.len(), 2);
9666        assert_eq!(saved.messages.last().unwrap().content, "done");
9667        assert!(
9668            append_fenced_remote_reply(&store, &fence, &mut child, "duplicate")
9669                .await
9670                .is_err()
9671        );
9672        assert_eq!(
9673            store
9674                .load_session(&child.id)
9675                .await
9676                .unwrap()
9677                .unwrap()
9678                .messages
9679                .len(),
9680            2
9681        );
9682    }
9683
9684    async fn start_bus() -> (String, tempfile::TempDir) {
9685        let dir = tempfile::tempdir().unwrap();
9686        let core = Arc::new(bamboo_broker::BrokerCore::new(dir.path()));
9687        let server = Arc::new(bamboo_broker::BrokerServer::new(core, "t"));
9688        let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
9689        let addr = listener.local_addr().unwrap();
9690        tokio::spawn(async move {
9691            let _ = server.serve(listener).await;
9692        });
9693        (format!("ws://{addr}"), dir)
9694    }
9695
9696    struct DurableAckProbe {
9697        calls: Arc<AtomicUsize>,
9698        fail: bool,
9699        terminal_status: TerminalStatus,
9700        parent_mailbox: String,
9701    }
9702
9703    #[async_trait]
9704    impl bamboo_subagent::ChildLink for DurableAckProbe {
9705        async fn send(&mut self, _frame: ParentFrame) -> bamboo_subagent::TransportResult<()> {
9706            Ok(())
9707        }
9708
9709        async fn next_frame(&mut self) -> bamboo_subagent::TransportResult<Option<ChildFrame>> {
9710            Ok(None)
9711        }
9712
9713        fn has_pending_durable_terminal(&self) -> bool {
9714            true
9715        }
9716
9717        fn durable_delivery_receipt(&self) -> Option<bamboo_subagent::DurableChildDeliveryReceipt> {
9718            Some(bamboo_subagent::DurableChildDeliveryReceipt {
9719                broker_identity: "00000000-0000-4000-8000-000000000001".into(),
9720                parent_mailbox: self.parent_mailbox.clone(),
9721                correlation_id: "broker-run-1".into(),
9722                message_ids: vec!["broker-outcome-1".into()],
9723                terminal_status: self.terminal_status,
9724            })
9725        }
9726
9727        async fn acknowledge_durable_frames(&mut self) -> bamboo_subagent::TransportResult<()> {
9728            self.calls.fetch_add(1, Ordering::SeqCst);
9729            if self.fail {
9730                Err(bamboo_subagent::TransportError::Protocol(
9731                    "injected ACK delivery failure".into(),
9732                ))
9733            } else {
9734                Ok(())
9735            }
9736        }
9737    }
9738
9739    async fn insert_ack_probe(
9740        runner: &ActorChildRunner,
9741        session: &Session,
9742        calls: &Arc<AtomicUsize>,
9743        terminal_status: TerminalStatus,
9744    ) {
9745        insert_ack_probe_with_nested_wait(runner, session, calls, terminal_status, false).await;
9746    }
9747
9748    async fn insert_ack_probe_with_nested_wait(
9749        runner: &ActorChildRunner,
9750        session: &Session,
9751        calls: &Arc<AtomicUsize>,
9752        terminal_status: TerminalStatus,
9753        host_nested_wait_handoff: bool,
9754    ) {
9755        runner.pending_durable_links.lock().await.insert(
9756            (session.id.clone(), session.created_at, "run-1".into()),
9757            PendingDurableChildLink {
9758                link: Box::new(DurableAckProbe {
9759                    calls: calls.clone(),
9760                    fail: false,
9761                    terminal_status,
9762                    parent_mailbox: format!("p-{}", session.id),
9763                }),
9764                host_nested_wait_handoff,
9765                strict_history: None,
9766                strict_history_required: false,
9767                strict_history_committed: false,
9768            },
9769        );
9770    }
9771
9772    #[tokio::test]
9773    async fn broker_receipt_reconciler_repairs_only_the_proven_local_mailbox() {
9774        let (endpoint, broker_dir) = start_bus().await;
9775        let home = tempfile::tempdir().unwrap();
9776        let store = Arc::new(
9777            bamboo_storage::SessionStoreV2::new(home.path().to_path_buf())
9778                .await
9779                .unwrap(),
9780        );
9781        let root = Session::new("receipt-root", "model");
9782        store.save_session(&root).await.unwrap();
9783        let mut sender = bamboo_broker::BrokerClient::connect(
9784            &endpoint,
9785            bamboo_subagent::AgentRef {
9786                session_id: "receipt-worker".into(),
9787                role: None,
9788            },
9789            "t",
9790        )
9791        .await
9792        .unwrap();
9793        let broker_identity = sender.durable_broker_identity().await.unwrap();
9794        let wrong_broker_identity = if broker_identity == "00000000-0000-4000-8000-000000000001" {
9795            "00000000-0000-4000-8000-000000000002"
9796        } else {
9797            "00000000-0000-4000-8000-000000000001"
9798        };
9799
9800        for (id, receipt_mailbox, receipt_broker) in [
9801            ("receipt-good", "p-receipt-good", broker_identity.as_str()),
9802            (
9803                "receipt-wrong-mailbox",
9804                "p-another-child",
9805                broker_identity.as_str(),
9806            ),
9807            (
9808                "receipt-wrong-broker",
9809                "p-receipt-wrong-broker",
9810                wrong_broker_identity,
9811            ),
9812            (
9813                "receipt-local-role-changed",
9814                "p-receipt-local-role-changed",
9815                broker_identity.as_str(),
9816            ),
9817            (
9818                "receipt-remote-unproven",
9819                "remote-parent-unproven",
9820                broker_identity.as_str(),
9821            ),
9822        ] {
9823            let mut child = Session::new_child_of(id, &root, "model", "task");
9824            child
9825                .metadata
9826                .insert("runtime.kind".into(), "external".into());
9827            child
9828                .metadata
9829                .insert("external.protocol".into(), "actor".into());
9830            child.metadata.insert(
9831                "external.agent_id".into(),
9832                crate::external_agents::config::LOCAL_ACTOR_AGENT_ID.into(),
9833            );
9834            if id == "receipt-local-role-changed" || id == "receipt-remote-unproven" {
9835                child
9836                    .metadata
9837                    .insert("subagent_type".into(), "remote-worker".into());
9838            }
9839            child.add_message(bamboo_agent_core::Message::user("work"));
9840            store.save_session(&child).await.unwrap();
9841            let message_id = bamboo_subagent::MsgId::new();
9842            sender
9843                .deliver(
9844                    &format!("p-{id}"),
9845                    bamboo_subagent::InboxMessage {
9846                        id: message_id.clone(),
9847                        from: bamboo_subagent::AgentRef {
9848                            session_id: "receipt-worker".into(),
9849                            role: None,
9850                        },
9851                        kind: bamboo_subagent::InboxKind::Outcome,
9852                        body: serde_json::to_value(bamboo_subagent::ChildOutcome::completed(
9853                            "done",
9854                        ))
9855                        .unwrap(),
9856                        created_at: chrono::Utc::now(),
9857                        correlation_id: Some(bamboo_subagent::MsgId::new()),
9858                    },
9859                )
9860                .await
9861                .unwrap();
9862            // The Host can checkpoint only a frame already surfaced from the
9863            // broker, which moves its Maildir file from new/ into cur/.
9864            let mailbox =
9865                bamboo_subagent::Mailbox::at(broker_dir.path().join(format!("mailboxes/p-{id}")));
9866            assert_eq!(mailbox.drain().await.unwrap().len(), 1);
9867            child.add_message(bamboo_agent_core::Message::assistant("done", None));
9868            child.set_last_run_status("completed");
9869            if id == "receipt-good" {
9870                let prior_ids = child.messages[..child.messages.len() - 1]
9871                    .iter()
9872                    .map(|message| message.id.clone())
9873                    .collect();
9874                crate::execution::ChildCompletionSource::prepare(&mut child, "run-1", &prior_ids);
9875                assert!(
9876                    crate::execution::ChildCompletionSource::defer_for_durable_delivery(
9877                        &mut child, "run-1"
9878                    )
9879                    .unwrap()
9880                );
9881                let proof = bamboo_storage::v2::HostTerminalCompleteness::from_verified_host(
9882                    &child, "run-1", 7, 1,
9883                )
9884                .unwrap();
9885                store
9886                    .prepare_broker_terminal_receipt_with_completeness(
9887                        &child,
9888                        bamboo_storage::v2::BrokerTerminalRoute {
9889                            activation_run_id: "run-1",
9890                            broker_identity: receipt_broker,
9891                            parent_mailbox: receipt_mailbox,
9892                            broker_correlation_id: "correlation-1",
9893                            message_ids: &[message_id.as_str().to_owned()],
9894                        },
9895                        &proof,
9896                    )
9897                    .await
9898                    .unwrap();
9899                let mut staged = child.clone();
9900                staged.set_last_run_status("running");
9901                staged.clear_last_run_error();
9902                store.save_session(&staged).await.unwrap();
9903                store
9904                    .commit_broker_terminal_receipt_with_completeness(&child, "run-1", &proof)
9905                    .await
9906                    .unwrap();
9907            } else {
9908                store
9909                    .prepare_broker_terminal_receipt(
9910                        &child,
9911                        "run-1",
9912                        receipt_broker,
9913                        receipt_mailbox,
9914                        "correlation-1",
9915                        &[message_id.as_str().to_owned()],
9916                    )
9917                    .await
9918                    .unwrap();
9919                store.save_session(&child).await.unwrap();
9920            }
9921        }
9922
9923        let mut config = bamboo_config::Config::default();
9924        config.subagents_mut().broker = Some(bamboo_config::BrokerClientConfig {
9925            endpoint,
9926            token: "t".into(),
9927            ..Default::default()
9928        });
9929        config
9930            .subagents_mut()
9931            .remote_placements
9932            .push(bamboo_config::RemoteActorPlacement {
9933                role: "remote-worker".into(),
9934                endpoint: "ws://unscoped.example.invalid".into(),
9935                token_env: None,
9936                ca_cert_file: None,
9937                broker_peer: None,
9938                placement_requirements: None,
9939            });
9940        // Recover on an independently reopened Store, with no Actor executor or
9941        // provider available. Only the committed Host proof can authorize ACK.
9942        let cold_store = Arc::new(
9943            bamboo_storage::SessionStoreV2::new(home.path().to_path_buf())
9944                .await
9945                .unwrap(),
9946        );
9947        let before = cold_store
9948            .load_session("receipt-good")
9949            .await
9950            .unwrap()
9951            .unwrap();
9952        assert!(crate::execution::ChildCompletionSource::has_source_record(
9953            &before
9954        ));
9955        assert_eq!(before.last_run_status().as_deref(), Some("running"));
9956        assert!(crate::execution::ChildCompletionSource::from_committed_session(&before).is_none());
9957        let repair = BrokerTerminalReceiptReconciler::new(cold_store.clone(), &config);
9958        let report = repair.reconcile_once().await;
9959        assert_eq!(report.candidates, 5);
9960        assert_eq!(report.repaired, 2);
9961        assert_eq!(report.blocked, 3);
9962        let after = cold_store
9963            .load_session("receipt-good")
9964            .await
9965            .unwrap()
9966            .unwrap();
9967        assert!(crate::execution::ChildCompletionSource::from_committed_session(&after).is_some());
9968        assert_eq!(
9969            serde_json::to_value(before.messages).unwrap(),
9970            serde_json::to_value(after.messages).unwrap()
9971        );
9972        let retry = repair.reconcile_once().await;
9973        assert_eq!(retry.repaired, 0, "committed strict recovery is idempotent");
9974
9975        let good = bamboo_subagent::Mailbox::at(broker_dir.path().join("mailboxes/p-receipt-good"));
9976        let wrong = bamboo_subagent::Mailbox::at(
9977            broker_dir.path().join("mailboxes/p-receipt-wrong-mailbox"),
9978        );
9979        let wrong_broker = bamboo_subagent::Mailbox::at(
9980            broker_dir.path().join("mailboxes/p-receipt-wrong-broker"),
9981        );
9982        let local_role_changed = bamboo_subagent::Mailbox::at(
9983            broker_dir
9984                .path()
9985                .join("mailboxes/p-receipt-local-role-changed"),
9986        );
9987        let remote_unproven = bamboo_subagent::Mailbox::at(
9988            broker_dir
9989                .path()
9990                .join("mailboxes/p-receipt-remote-unproven"),
9991        );
9992        assert_eq!(good.pending_count().await.unwrap(), 0);
9993        assert_eq!(wrong.pending_count().await.unwrap(), 1);
9994        assert_eq!(wrong_broker.pending_count().await.unwrap(), 1);
9995        assert_eq!(local_role_changed.pending_count().await.unwrap(), 0);
9996        assert_eq!(remote_unproven.pending_count().await.unwrap(), 1);
9997        let good_child = store.load_session("receipt-good").await.unwrap().unwrap();
9998        let wrong_child = store
9999            .load_session("receipt-wrong-mailbox")
10000            .await
10001            .unwrap()
10002            .unwrap();
10003        let wrong_broker_child = store
10004            .load_session("receipt-wrong-broker")
10005            .await
10006            .unwrap()
10007            .unwrap();
10008        let local_role_changed_child = store
10009            .load_session("receipt-local-role-changed")
10010            .await
10011            .unwrap()
10012            .unwrap();
10013        let remote_unproven_child = store
10014            .load_session("receipt-remote-unproven")
10015            .await
10016            .unwrap()
10017            .unwrap();
10018        assert!(store
10019            .recover_broker_terminal_receipts(&good_child)
10020            .await
10021            .unwrap()
10022            .is_empty());
10023        assert_eq!(
10024            store
10025                .recover_broker_terminal_receipts(&wrong_child)
10026                .await
10027                .unwrap()
10028                .len(),
10029            1
10030        );
10031        assert_eq!(
10032            store
10033                .recover_broker_terminal_receipts(&wrong_broker_child)
10034                .await
10035                .unwrap()
10036                .len(),
10037            1
10038        );
10039        assert!(store
10040            .recover_broker_terminal_receipts(&local_role_changed_child)
10041            .await
10042            .unwrap()
10043            .is_empty());
10044        assert_eq!(
10045            store
10046                .recover_broker_terminal_receipts(&remote_unproven_child)
10047                .await
10048                .unwrap()
10049                .len(),
10050            1
10051        );
10052    }
10053
10054    #[tokio::test]
10055    async fn broker_receipt_confirmation_requires_exact_birth_run_and_committed_status() {
10056        let runner = ActorChildRunner::new(
10057            "test".into(),
10058            PathBuf::new(),
10059            vec![],
10060            PathBuf::new(),
10061            ExecutorSpec::BambooRuntime,
10062            vec![],
10063            "test".into(),
10064            1,
10065        );
10066        let home = tempfile::tempdir().unwrap();
10067        let store = Arc::new(
10068            bamboo_storage::SessionStoreV2::new(home.path().to_path_buf())
10069                .await
10070                .unwrap(),
10071        );
10072        let root = Session::new("ack-parent", "model");
10073        store.save_session(&root).await.unwrap();
10074        let mut child = Session::new_child_of("ack-child", &root, "model", "task");
10075        child.add_message(bamboo_agent_core::Message::user("work"));
10076        store.save_session(&child).await.unwrap();
10077        child.add_message(bamboo_agent_core::Message::assistant("done", None));
10078        child.set_last_run_status("completed");
10079        store
10080            .prepare_broker_terminal_receipt(
10081                &child,
10082                "run-1",
10083                "00000000-0000-4000-8000-000000000001",
10084                "p-ack-child",
10085                "broker-run-1",
10086                &["broker-outcome-1".into()],
10087            )
10088            .await
10089            .unwrap();
10090        store.save_session(&child).await.unwrap();
10091        runner.set_actor_directory_store(Some(store));
10092        let calls = Arc::new(AtomicUsize::new(0));
10093        insert_ack_probe(&runner, &child, &calls, TerminalStatus::Completed).await;
10094        runner
10095            .confirm_durable_child_delivery(&child, "wrong-run", true)
10096            .await
10097            .unwrap();
10098        let mut wrong_birth = child.clone();
10099        wrong_birth.created_at += chrono::Duration::seconds(1);
10100        runner
10101            .confirm_durable_child_delivery(&wrong_birth, "run-1", true)
10102            .await
10103            .unwrap();
10104        assert_eq!(runner.pending_durable_links.lock().await.len(), 1);
10105        child.set_last_run_status("completed");
10106        runner
10107            .confirm_durable_child_delivery(&child, "run-1", false)
10108            .await
10109            .unwrap();
10110        assert_eq!(calls.load(Ordering::SeqCst), 0);
10111
10112        insert_ack_probe(&runner, &child, &calls, TerminalStatus::Completed).await;
10113        child.set_last_run_status("error");
10114        runner
10115            .confirm_durable_child_delivery(&child, "run-1", true)
10116            .await
10117            .unwrap();
10118        assert_eq!(calls.load(Ordering::SeqCst), 0);
10119
10120        insert_ack_probe(&runner, &child, &calls, TerminalStatus::Completed).await;
10121        child.set_last_run_status("completed");
10122        runner
10123            .confirm_durable_child_delivery(&child, "run-1", true)
10124            .await
10125            .unwrap();
10126        assert_eq!(calls.load(Ordering::SeqCst), 1);
10127    }
10128
10129    #[tokio::test]
10130    async fn strict_completeness_gate_precedes_ack_and_ack_failure_preserves_cold_proof() {
10131        for commit in [false, true] {
10132            let runner = ActorChildRunner::new(
10133                "test".into(),
10134                PathBuf::new(),
10135                vec![],
10136                PathBuf::new(),
10137                ExecutorSpec::BambooRuntime,
10138                vec![],
10139                "test".into(),
10140                1,
10141            );
10142            let home = tempfile::tempdir().unwrap();
10143            let store = Arc::new(
10144                bamboo_storage::SessionStoreV2::new(home.path().into())
10145                    .await
10146                    .unwrap(),
10147            );
10148            let root = Session::new("strict-ack-parent", "model");
10149            store.save_session(&root).await.unwrap();
10150            let mut child = Session::new_child_of("strict-ack-child", &root, "model", "work");
10151            child.add_message(bamboo_agent_core::Message::user("work"));
10152            store.save_session(&child).await.unwrap();
10153            child.add_message(bamboo_agent_core::Message::assistant("finished", None));
10154            child.set_last_run_status("completed");
10155            let proof = bamboo_storage::v2::HostTerminalCompleteness::from_verified_host(
10156                &child, "run-1", 7, 2,
10157            )
10158            .unwrap();
10159            let creation = bamboo_subagent::proto::ChildCreationIdentity {
10160                created_at: child.created_at,
10161                spawn_depth: child.spawn_depth,
10162            };
10163            let history = StrictTerminalHistory {
10164                proof: proof.clone(),
10165                watermark: bamboo_subagent::ActorEventWatermark {
10166                    version: 1,
10167                    logical_session: Some(actor_event_logical_identity(
10168                        &child,
10169                        &root.id,
10170                        Some(&creation),
10171                    )),
10172                    activation_id: Some("run-1".into()),
10173                    execution_epoch: 7,
10174                    final_seq: 2,
10175                },
10176            };
10177            let parent_mailbox = "p-strict-ack-child";
10178            store
10179                .prepare_broker_terminal_receipt_with_completeness(
10180                    &child,
10181                    bamboo_storage::v2::BrokerTerminalRoute {
10182                        activation_run_id: "run-1",
10183                        broker_identity: "00000000-0000-4000-8000-000000000001",
10184                        parent_mailbox,
10185                        broker_correlation_id: "broker-run-1",
10186                        message_ids: &["broker-outcome-1".into()],
10187                    },
10188                    &proof,
10189                )
10190                .await
10191                .unwrap();
10192            store.save_session(&child).await.unwrap();
10193            runner.set_actor_directory_store(Some(store.clone()));
10194            let calls = Arc::new(AtomicUsize::new(0));
10195            runner.pending_durable_links.lock().await.insert(
10196                (child.id.clone(), child.created_at, "run-1".into()),
10197                PendingDurableChildLink {
10198                    link: Box::new(DurableAckProbe {
10199                        calls: calls.clone(),
10200                        fail: true,
10201                        terminal_status: TerminalStatus::Completed,
10202                        parent_mailbox: parent_mailbox.into(),
10203                    }),
10204                    host_nested_wait_handoff: false,
10205                    strict_history: Some(history),
10206                    strict_history_required: true,
10207                    strict_history_committed: false,
10208                },
10209            );
10210            if commit {
10211                runner
10212                    .commit_durable_child_delivery(&child, "run-1")
10213                    .await
10214                    .unwrap();
10215                assert_eq!(calls.load(Ordering::SeqCst), 0, "proof commit cannot ACK");
10216            }
10217            assert!(runner
10218                .confirm_durable_child_delivery(&child, "run-1", true)
10219                .await
10220                .is_err());
10221            assert_eq!(calls.load(Ordering::SeqCst), usize::from(commit));
10222            let cold = bamboo_storage::SessionStoreV2::new(home.path().into())
10223                .await
10224                .unwrap();
10225            let recovered = cold.recover_broker_terminal_receipts(&child).await;
10226            if commit {
10227                let receipts = recovered.unwrap();
10228                assert_eq!(receipts.len(), 1);
10229                assert_eq!(
10230                    receipts[0]
10231                        .terminal_completeness
10232                        .as_ref()
10233                        .unwrap()
10234                        .contiguous_applied_seq,
10235                    2
10236                );
10237                assert_eq!(
10238                    cold.load_session(&child.id)
10239                        .await
10240                        .unwrap()
10241                        .unwrap()
10242                        .last_run_status()
10243                        .as_deref(),
10244                    Some("completed")
10245                );
10246            } else {
10247                assert!(
10248                    recovered.is_err(),
10249                    "matching canonical history is not proof commitment"
10250                );
10251            }
10252            assert_eq!(
10253                serde_json::to_value(
10254                    cold.load_session(&child.id)
10255                        .await
10256                        .unwrap()
10257                        .unwrap()
10258                        .messages
10259                )
10260                .unwrap(),
10261                serde_json::to_value(&child.messages).unwrap(),
10262            );
10263        }
10264    }
10265
10266    #[tokio::test]
10267    async fn broker_receipt_confirms_canonical_worker_error_and_cancelled_status() {
10268        for (terminal_status, status) in [
10269            (TerminalStatus::Error, "error"),
10270            (TerminalStatus::Cancelled, "cancelled"),
10271        ] {
10272            let runner = ActorChildRunner::new(
10273                "test".into(),
10274                PathBuf::new(),
10275                vec![],
10276                PathBuf::new(),
10277                ExecutorSpec::BambooRuntime,
10278                vec![],
10279                "test".into(),
10280                1,
10281            );
10282            let home = tempfile::tempdir().unwrap();
10283            let store = Arc::new(
10284                bamboo_storage::SessionStoreV2::new(home.path().to_path_buf())
10285                    .await
10286                    .unwrap(),
10287            );
10288            let root = Session::new("failed-ack-parent", "model");
10289            store.save_session(&root).await.unwrap();
10290            let mut child = Session::new_child_of("failed-ack-child", &root, "model", "task");
10291            child.add_message(bamboo_agent_core::Message::user("work"));
10292            store.save_session(&child).await.unwrap();
10293            child.set_last_run_status(status);
10294            child.set_last_run_error(format!("worker {status}"));
10295            store
10296                .prepare_broker_terminal_receipt(
10297                    &child,
10298                    "run-1",
10299                    "00000000-0000-4000-8000-000000000001",
10300                    "p-failed-ack-child",
10301                    "broker-run-1",
10302                    &["broker-outcome-1".into()],
10303                )
10304                .await
10305                .unwrap();
10306            store.save_session(&child).await.unwrap();
10307            runner.set_actor_directory_store(Some(store));
10308            let calls = Arc::new(AtomicUsize::new(0));
10309            insert_ack_probe(&runner, &child, &calls, terminal_status).await;
10310            runner
10311                .confirm_durable_child_delivery(&child, "run-1", true)
10312                .await
10313                .unwrap();
10314            assert_eq!(calls.load(Ordering::SeqCst), 1, "{status}");
10315        }
10316    }
10317
10318    #[tokio::test]
10319    async fn broker_completed_terminal_projects_only_host_verified_nested_wait() {
10320        assert_eq!(
10321            effective_broker_terminal_status(TerminalStatus::Completed, true).unwrap(),
10322            "suspended"
10323        );
10324        for raw in [TerminalStatus::Error, TerminalStatus::Cancelled] {
10325            assert!(effective_broker_terminal_status(raw, true).is_err());
10326        }
10327        let runner = ActorChildRunner::new(
10328            "test".into(),
10329            PathBuf::new(),
10330            vec![],
10331            PathBuf::new(),
10332            ExecutorSpec::BambooRuntime,
10333            vec![],
10334            "test".into(),
10335            1,
10336        );
10337        let home = tempfile::tempdir().unwrap();
10338        let store = Arc::new(
10339            bamboo_storage::SessionStoreV2::new(home.path().to_path_buf())
10340                .await
10341                .unwrap(),
10342        );
10343        let root = Session::new("nested-ack-parent", "model");
10344        store.save_session(&root).await.unwrap();
10345        let mut child = Session::new_child_of("nested-ack-child", &root, "model", "task");
10346        child.add_message(bamboo_agent_core::Message::user("spawn a grandchild"));
10347        store.save_session(&child).await.unwrap();
10348        child.set_last_run_status("suspended");
10349        store
10350            .prepare_broker_terminal_receipt(
10351                &child,
10352                "run-1",
10353                "00000000-0000-4000-8000-000000000001",
10354                "p-nested-ack-child",
10355                "broker-run-1",
10356                &["broker-outcome-1".into()],
10357            )
10358            .await
10359            .unwrap();
10360        store.save_session(&child).await.unwrap();
10361        runner.set_actor_directory_store(Some(store.clone()));
10362        let calls = Arc::new(AtomicUsize::new(0));
10363
10364        insert_ack_probe(&runner, &child, &calls, TerminalStatus::Completed).await;
10365        runner
10366            .confirm_durable_child_delivery(&child, "run-1", true)
10367            .await
10368            .unwrap();
10369        assert_eq!(calls.load(Ordering::SeqCst), 0);
10370
10371        insert_ack_probe_with_nested_wait(&runner, &child, &calls, TerminalStatus::Completed, true)
10372            .await;
10373        runner
10374            .confirm_durable_child_delivery(&child, "run-1", true)
10375            .await
10376            .unwrap();
10377        assert_eq!(calls.load(Ordering::SeqCst), 1);
10378        assert_eq!(
10379            store
10380                .load_session(&child.id)
10381                .await
10382                .unwrap()
10383                .unwrap()
10384                .last_run_status()
10385                .as_deref(),
10386            Some("suspended")
10387        );
10388        assert!(store
10389            .recover_broker_terminal_receipts(&child)
10390            .await
10391            .unwrap()
10392            .is_empty());
10393    }
10394
10395    #[tokio::test]
10396    async fn host_completion_notifies_once_only_after_reply_and_finish_commit() {
10397        struct CommittedObserver {
10398            authority: PathBuf,
10399            events: std::sync::Mutex<Vec<PublicActorEvent>>,
10400        }
10401        impl ActorEventObserver for CommittedObserver {
10402            fn publish(&self, event: PublicActorEvent) {
10403                // Observe the real durable sidecar synchronously at publication.
10404                let entry: bamboo_domain::ActorDirectoryEntry =
10405                    serde_json::from_slice(&std::fs::read(&self.authority).unwrap()).unwrap();
10406                let committed = entry.activation.unwrap();
10407                assert_eq!(
10408                    committed.status,
10409                    bamboo_domain::ActorActivationStatus::Succeeded
10410                );
10411                assert_eq!(committed.activation_id, event.activation_id);
10412                assert!(committed.finished_at.is_some());
10413                self.events.lock().unwrap().push(event);
10414            }
10415        }
10416        struct CompletionLink(VecDeque<ChildFrame>);
10417        #[async_trait]
10418        impl bamboo_subagent::ChildLink for CompletionLink {
10419            async fn send(&mut self, _: ParentFrame) -> bamboo_subagent::TransportResult<()> {
10420                Ok(())
10421            }
10422            async fn next_frame(&mut self) -> bamboo_subagent::TransportResult<Option<ChildFrame>> {
10423                Ok(self.0.pop_front())
10424            }
10425        }
10426        for cut in [
10427            "success",
10428            "no_terminal",
10429            "invalid_terminal",
10430            "stale",
10431            "storage_failure",
10432            "append_failure",
10433        ] {
10434            let temp = tempfile::tempdir().unwrap();
10435            let store = Arc::new(
10436                bamboo_storage::SessionStoreV2::new(temp.path().into())
10437                    .await
10438                    .unwrap(),
10439            );
10440            let root = Session::new("notify-root", "model");
10441            store.save_session(&root).await.unwrap();
10442            let mut child = Session::new_child_of("notify-child", &root, "model", "assignment");
10443            store.save_session(&child).await.unwrap();
10444            let inbox: Arc<dyn SessionInboxPort> = Arc::new(bamboo_storage::FileSessionInbox::new(
10445                store.clone(),
10446                bamboo_domain::SessionInboxLimits::default(),
10447            ));
10448            let binding = actor_binding(
10449                store.clone(),
10450                inbox,
10451                Arc::new(bamboo_storage::LockedSessionStore::new(store.clone())),
10452            );
10453            let policy = Arc::new(bamboo_tools::permission::PermissionConfig::new());
10454            bind_local_control_plane(store.as_ref(), &child.id, policy.as_ref()).await;
10455            let activation = PlainActorActivation::start(
10456                store.clone(),
10457                &child,
10458                &binding,
10459                Some("notify-run"),
10460                Some(policy),
10461                false,
10462                None,
10463            )
10464            .await
10465            .unwrap();
10466            let _owner = binding
10467                .router
10468                .register_run(&child.id, "notify-run")
10469                .await
10470                .unwrap();
10471            let batch = ActorEventBatch {
10472                logical_session: Some(LogicalSessionIdentity {
10473                    session_id: child.id.clone(),
10474                    parent_session_id: child.parent_session_id.clone(),
10475                    root_session_id: child.root_session_id.clone(),
10476                    creation: None,
10477                }),
10478                activation_id: Some("notify-run".into()),
10479                execution_epoch: 9,
10480                source_node_id: None,
10481                source_actor_id: Some("notify-worker".into()),
10482                first_seq: 1,
10483                last_seq: 2,
10484                qos: bamboo_subagent::ActorEventQos::Durable,
10485                events: vec![
10486                    serde_json::to_value(AgentEvent::Token {
10487                        content: "PRIVATE_PROVIDER_TOKEN".into(),
10488                    })
10489                    .unwrap(),
10490                    serde_json::to_value(AgentEvent::message_appended(
10491                        &child.id,
10492                        &bamboo_agent_core::Message::assistant("PRIVATE_WORKER_CACHE", None),
10493                    ))
10494                    .unwrap(),
10495                ],
10496            };
10497            let mut frames = VecDeque::from([ChildFrame::EventBatch { batch }]);
10498            if cut != "no_terminal" {
10499                let mut terminal = completed_actor_frame();
10500                if cut == "invalid_terminal" {
10501                    if let ChildFrame::Terminal { status, .. } = &mut terminal {
10502                        *status = TerminalStatus::Suspended;
10503                    }
10504                }
10505                frames.push_back(terminal);
10506            }
10507            let mut link = CompletionLink(frames);
10508            let (tx, mut rx) = mpsc::channel(8);
10509            let (_live, mut live_rx) = mpsc::unbounded_channel();
10510            let (_delivery, mut delivery_rx) = mpsc::unbounded_channel();
10511            let cancel = CancellationToken::new();
10512            let outcome = drive(ActorDriveContext {
10513                client: &mut link,
10514                parent_session_id: "notify-root",
10515                child_session_id: "notify-child",
10516                child_attempt: 0,
10517                approval_registry: None,
10518                approval_decider: None,
10519                approval_reviewer: None,
10520                escalation_bridge: None,
10521                event_tx: &tx,
10522                cancel_token: &cancel,
10523                live_rx: &mut live_rx,
10524                delivery_rx: &mut delivery_rx,
10525                logical_session: &mut child,
10526                expected_permission_posture: None,
10527                expected_creation: None,
10528                session_inbox_runtime: Some(&binding),
10529                actor_directory_store: Some(store.as_ref()),
10530                canonical_subagent_tool: None,
10531                ticket_service: None,
10532                activation_run_id: Some("notify-run"),
10533                execution_epoch: 9,
10534                expected_source_actor_id: "notify-worker",
10535                initial_inflight_claims: VecDeque::new(),
10536                plain_actor: true,
10537                remote_environment_lease: false,
10538                readonly_output: None,
10539                local_history_tools: None,
10540                local_history_read_only: false,
10541                strict_history_output: None,
10542                plain_input: Some(&activation),
10543                canonical_activation: None,
10544                canonical_placement_ref: None,
10545                actor_event_observer: None,
10546                plain_run: None,
10547                first_frame_timeout: Some(Duration::from_secs(1)),
10548            })
10549            .await;
10550            let admitted = !matches!(cut, "no_terminal" | "invalid_terminal");
10551            assert_eq!(outcome.is_ok(), admitted, "{cut}: {outcome:?}");
10552            assert_eq!(
10553                activation.completion_router.lock().unwrap().is_some(),
10554                admitted,
10555                "{cut}"
10556            );
10557            while rx.try_recv().is_ok() {}
10558            let authority = temp
10559                .path()
10560                .join(store.resolve_rel_path(&child.id).await.unwrap())
10561                .join("actor-authority.json");
10562            let observer = CommittedObserver {
10563                authority: authority.clone(),
10564                events: Default::default(),
10565            };
10566            if cut == "append_failure" {
10567                child.add_message(bamboo_agent_core::Message::user("uncommitted-prefix"));
10568                assert!(activation
10569                    .append_reply(&mut child, "PRIVATE_REPLY".into(), &tx)
10570                    .await
10571                    .is_err());
10572                activation
10573                    .finish(ActorActivationFinish::Failed)
10574                    .await
10575                    .unwrap();
10576                assert!(observer.events.lock().unwrap().is_empty());
10577                continue;
10578            }
10579            activation
10580                .append_reply(&mut child, "PRIVATE_REPLY".into(), &tx)
10581                .await
10582                .unwrap();
10583            assert!(
10584                rx.try_recv().is_ok(),
10585                "legacy content notification retained"
10586            );
10587            assert!(
10588                observer.events.lock().unwrap().is_empty(),
10589                "reply alone must not publish"
10590            );
10591            if cut == "stale" {
10592                store
10593                    .retire_actor(&child.id, chrono::Utc::now())
10594                    .await
10595                    .unwrap();
10596            } else if cut == "storage_failure" {
10597                // Make the canonical write path unreadable, rather than mocking
10598                // the store or allowing completion to fall back to ordinary save.
10599                std::fs::remove_file(&authority).unwrap();
10600                std::fs::create_dir(&authority).unwrap();
10601            }
10602            let result = activation
10603                .finish_with_observer(ActorActivationFinish::Succeeded, Some(&observer))
10604                .await;
10605            assert_eq!(
10606                result.is_ok(),
10607                matches!(cut, "success" | "no_terminal" | "invalid_terminal"),
10608                "{cut}"
10609            );
10610            let events = observer.events.lock().unwrap().clone();
10611            assert_eq!(events.len(), usize::from(cut == "success"), "{cut}");
10612            if let Some(event) = events.first() {
10613                assert_eq!(event.actor_id, child.id);
10614                assert_eq!(event.root_actor_id, root.id);
10615                assert_eq!(event.parent_actor_id, Some(root.id.clone()));
10616                assert_eq!(event.source_order.sequence, 3);
10617                let json = serde_json::to_string(event).unwrap();
10618                assert!(!json.contains("PRIVATE_") && !json.contains("notify-run"));
10619                let persisted = store.load_session(&child.id).await.unwrap().unwrap();
10620                assert_eq!(persisted.messages.last().unwrap().content, "PRIVATE_REPLY");
10621            }
10622            assert!(activation
10623                .finish_with_observer(ActorActivationFinish::Succeeded, Some(&observer))
10624                .await
10625                .is_err());
10626            assert_eq!(
10627                observer.events.lock().unwrap().len(),
10628                events.len(),
10629                "repeat finish cannot publish"
10630            );
10631        }
10632    }
10633
10634    async fn bind_local_control_plane(
10635        store: &bamboo_storage::SessionStoreV2,
10636        actor_id: &str,
10637        policy: &bamboo_tools::permission::PermissionConfig,
10638    ) {
10639        store
10640            .bind_control_plane(
10641                actor_id,
10642                bamboo_domain::ActorControlPlaneBinding {
10643                    policy_revision: Some(policy.policy_revision()),
10644                    placement_intent: Some(bamboo_domain::ActorPlacementIntent::Local),
10645                },
10646            )
10647            .await
10648            .unwrap();
10649    }
10650
10651    fn local_tool_sample(
10652        host: &Session,
10653        denied_write: bool,
10654    ) -> (LocalToolCollector, Vec<bamboo_agent_core::Message>) {
10655        local_tool_sample_with_reasoning(host, denied_write, None)
10656    }
10657
10658    fn local_tool_sample_with_reasoning(
10659        host: &Session,
10660        denied_write: bool,
10661        reasoning: Option<&str>,
10662    ) -> (LocalToolCollector, Vec<bamboo_agent_core::Message>) {
10663        use bamboo_agent_core::{FunctionCall, Message, ToolCall};
10664        let calls = ["Read", "Write"]
10665            .map(|name| ToolCall {
10666                id: format!("local-{name}"),
10667                tool_type: "function".into(),
10668                function: FunctionCall {
10669                    name: name.into(),
10670                    arguments: "{\"file_path\":\"answer.txt\"}".into(),
10671                },
10672            })
10673            .to_vec();
10674        let mut rows = vec![Message::assistant("checking", Some(calls))];
10675        rows[0].reasoning = reasoning.map(str::to_owned);
10676        let mut collector = LocalToolCollector::default();
10677        if let Some(reasoning) = reasoning {
10678            assert!(collector
10679                .event(
10680                    &serde_json::to_value(AgentEvent::ReasoningToken {
10681                        content: reasoning.into(),
10682                    })
10683                    .unwrap()
10684                )
10685                .unwrap());
10686        }
10687        for name in ["Read", "Write"] {
10688            let denied = name == "Write" && denied_write;
10689            let id = format!("local-{name}");
10690            if !denied {
10691                collector
10692                    .event(
10693                        &serde_json::to_value(AgentEvent::ToolStart {
10694                            tool_call_id: id.clone(),
10695                            tool_name: name.into(),
10696                            arguments: serde_json::json!({"file_path":"answer.txt"}),
10697                        })
10698                        .unwrap(),
10699                    )
10700                    .unwrap();
10701            }
10702            let text = if denied {
10703                "Error: native_tool_ceiling_denied: Write"
10704            } else {
10705                "actual file result"
10706            };
10707            let mut message = Message::tool_result_with_status(id.clone(), text, !denied);
10708            message.metadata = Some(serde_json::json!({"elapsed_ms":7,"tool_name":name,
10709                "is_mutating":name == "Write","auto_approved":name != "Write","success":!denied}));
10710            rows.push(message);
10711            let event = if denied {
10712                AgentEvent::ToolError {
10713                    tool_call_id: id,
10714                    error: "native_tool_ceiling_denied: Write".into(),
10715                }
10716            } else {
10717                AgentEvent::ToolComplete {
10718                    tool_call_id: id,
10719                    result: bamboo_agent_core::tools::ToolResult::text(true, text),
10720                }
10721            };
10722            collector
10723                .event(&serde_json::to_value(event).unwrap())
10724                .unwrap();
10725        }
10726        rows.push(Message::assistant("complete exact report", None));
10727        rows.last_mut().unwrap().reasoning = reasoning.map(str::to_owned);
10728        if let Some(reasoning) = reasoning {
10729            assert!(collector
10730                .event(
10731                    &serde_json::to_value(AgentEvent::ReasoningToken {
10732                        content: reasoning.into(),
10733                    })
10734                    .unwrap()
10735                )
10736                .unwrap());
10737        }
10738        let mut messages = host.messages.clone();
10739        if let Some(system) = messages
10740            .iter_mut()
10741            .find(|message| message.role == Role::System)
10742        {
10743            system.content = "worker-private system never imported".into();
10744        }
10745        messages.extend(rows.iter().cloned());
10746        let data = bamboo_subagent::proto::LocalToolMessages::Complete {
10747            version: 1,
10748            messages: messages
10749                .iter()
10750                .map(|message| serde_json::to_value(message).unwrap())
10751                .collect(),
10752        };
10753        assert!(!collector
10754            .event(&serde_json::to_value(data).unwrap())
10755            .unwrap());
10756        (collector, rows)
10757    }
10758
10759    #[test]
10760    fn ticket_task_event_requires_exact_host_callback_receipt() {
10761        let mut collector = LocalToolCollector::default();
10762        let args = serde_json::json!({"tasks":[{"content":"own step","status":"pending"}]});
10763        let start = serde_json::to_value(AgentEvent::ToolStart {
10764            tool_call_id: "task-call".into(),
10765            tool_name: "Task".into(),
10766            arguments: args.clone(),
10767        })
10768        .unwrap();
10769        let mut receipts = HashMap::new();
10770        assert!(collector.ticket_event(&start, &receipts).unwrap());
10771        collector.event(&start).unwrap();
10772        let complete = serde_json::to_value(AgentEvent::ToolComplete {
10773            tool_call_id: "task-call".into(),
10774            result: bamboo_agent_core::tools::ToolResult::text(true, "done"),
10775        })
10776        .unwrap();
10777        assert!(collector.ticket_event(&complete, &receipts).is_err());
10778        receipts.insert("other-call".into(), args.clone());
10779        assert!(collector.ticket_event(&complete, &receipts).is_err());
10780        receipts.insert("task-call".into(), serde_json::json!({"tasks":[]}));
10781        assert!(collector.ticket_event(&complete, &receipts).is_err());
10782        receipts.insert("task-call".into(), args);
10783        assert!(collector.ticket_event(&complete, &receipts).unwrap());
10784        let forged_root = serde_json::to_value(AgentEvent::TaskListUpdated {
10785            task_list: bamboo_tools::TaskTool::task_list_from_args(
10786                &serde_json::json!({"tasks":[{"content":"forged root","status":"completed"}]}),
10787                "root",
10788            )
10789            .unwrap(),
10790            version: Some(999),
10791        })
10792        .unwrap();
10793        assert!(!collector.ticket_event(&forged_root, &receipts).unwrap());
10794    }
10795
10796    #[test]
10797    fn ticket_question_terminal_requires_exact_host_receipt_and_tool_result() {
10798        use bamboo_agent_core::{FunctionCall, Message, ToolCall};
10799        let mut host = Session::new("question-child", "model");
10800        host.messages.push(Message::user("own assignment"));
10801        let question: bamboo_tickets::PendingRequest = serde_json::from_value(serde_json::json!({
10802            "id":"own-question", "work_id":"own-work", "assignment_id":"own-assignment",
10803            "generation":1, "contract_revision":1, "prompt_revision":1,
10804            "kind":{"kind":"question"}, "prompt":"精确问题", "status":"open",
10805            "answer":null, "consumed_attempt":null, "worker_packet_hash":"a".repeat(64), "updated_seq":3
10806        })).unwrap();
10807        let args = serde_json::json!({"tasks":[{"content":"own step","status":"blocked"}], "question":{"prompt":"精确问题"}});
10808        for case in ["exact", "no_host_receipt", "forged_request", "fake_report"] {
10809            let mut returned =
10810                serde_json::json!({"status":"waiting_for_answer", "request":question});
10811            if case == "forged_request" {
10812                returned["request"]["id"] = serde_json::json!("sibling-question");
10813            }
10814            let text = returned.to_string();
10815            let rows = [
10816                Message::assistant(
10817                    "",
10818                    Some(vec![ToolCall {
10819                        id: "own-call".into(),
10820                        tool_type: "function".into(),
10821                        function: FunctionCall {
10822                            name: "Task".into(),
10823                            arguments: args.to_string(),
10824                        },
10825                    }]),
10826                ),
10827                Message::tool_result_with_status("own-call", &text, true),
10828            ];
10829            let collector = LocalToolCollector {
10830                messages: Some(host.messages.iter().cloned().chain(rows).collect()),
10831                starts: HashMap::from([("own-call".into(), ("Task".into(), args.clone()))]),
10832                outcomes: HashMap::from([("own-call".into(), (true, text))]),
10833                ticket_yield: (case != "no_host_receipt")
10834                    .then(|| ("own-call".into(), question.clone())),
10835            };
10836            let terminal = if case == "fake_report" {
10837                Some("fake delivery")
10838            } else {
10839                Some("")
10840            };
10841            assert_eq!(
10842                collector
10843                    .suffix(&host, &["Task".into()], false, terminal)
10844                    .is_ok(),
10845                case == "exact",
10846                "{case}"
10847            );
10848        }
10849    }
10850
10851    #[test]
10852    fn local_tool_suffix_preserves_full_current_prefix_and_rejects_authority_or_collisions() {
10853        let mut host = Session::new("tool-history-prefix", "model");
10854        host.messages
10855            .push(bamboo_agent_core::Message::user("assignment"));
10856        let mut confirmed = bamboo_agent_core::Message::user("confirmed during Run");
10857        confirmed.metadata = Some(
10858            serde_json::json!({"session_message":{"id":confirmed.id,"target_session_id":host.id}}),
10859        );
10860        host.messages.push(confirmed);
10861        let tools = vec!["Read".into(), "Write".into()];
10862        let (good, rows) = local_tool_sample(&host, false);
10863        assert_eq!(
10864            serde_json::to_value(
10865                good.suffix(&host, &tools, false, Some("complete exact report"))
10866                    .unwrap()
10867            )
10868            .unwrap(),
10869            serde_json::to_value(&rows).unwrap()
10870        );
10871        for case in [
10872            "missing",
10873            "changed",
10874            "metadata",
10875            "system",
10876            "user",
10877            "collision",
10878            "dangling",
10879            "result",
10880            "terminal",
10881            "tool_metadata",
10882            "grant",
10883            "start_arguments",
10884            "ledger",
10885        ] {
10886            let (mut collector, _) = local_tool_sample(&host, false);
10887            let mut current = host.clone();
10888            let messages = collector.messages.as_mut().unwrap();
10889            match case {
10890                "missing" => {
10891                    messages.remove(host.messages.len() - 1);
10892                }
10893                "changed" => messages[0].content.push('!'),
10894                "metadata" => messages[1].metadata = None,
10895                "system" => messages.last_mut().unwrap().role = Role::System,
10896                "user" => messages.last_mut().unwrap().role = Role::User,
10897                "collision" => messages.last_mut().unwrap().id = host.messages[0].id.clone(),
10898                "dangling" => {
10899                    messages[host.messages.len() + 1].tool_call_id = Some("foreign".into())
10900                }
10901                "result" => messages[host.messages.len() + 1].content.push('!'),
10902                "tool_metadata" => {
10903                    messages[host.messages.len() + 1].metadata.as_mut().unwrap()["permission"] =
10904                        true.into()
10905                }
10906                "grant" => {
10907                    messages[host.messages.len()].tool_calls.as_mut().unwrap()[0]
10908                        .function
10909                        .name = "Bash".into()
10910                }
10911                "start_arguments" => {
10912                    collector.starts.get_mut("local-Read").unwrap().1 =
10913                        serde_json::json!({"file_path":"different.txt"})
10914                }
10915                "ledger" => {
10916                    current.append_provider_transcript_group(current.messages[0].id.clone(), None, vec![
10917                        bamboo_domain::session::provider_transcript::ProviderTranscriptItem::try_from_payload(
10918                            bamboo_domain::ProviderFamily::OpenAi,
10919                            bamboo_domain::ProviderProtocol::OpenAiResponsesV1,
10920                            bamboo_domain::session::provider_transcript::ProviderTranscriptOrigin::HostToolSearch,
10921                            bamboo_domain::session::provider_transcript::ProviderTranscriptAuthor::ToolResult,
10922                            serde_json::json!({"type":"tool_search_output","execution":"client","call_id":"search","status":"completed","tools":[]}),
10923                        ).unwrap(),
10924                    ]).unwrap();
10925                }
10926                "terminal" => {}
10927                _ => unreachable!(),
10928            }
10929            let terminal = if case == "terminal" {
10930                "different"
10931            } else {
10932                "complete exact report"
10933            };
10934            assert!(
10935                collector
10936                    .suffix(&current, &tools, false, Some(terminal))
10937                    .is_err(),
10938                "{case}"
10939            );
10940        }
10941        let (denied, _) = local_tool_sample(&host, true);
10942        assert!(denied
10943            .suffix(&host, &["Read".into()], true, Some("complete exact report"))
10944            .is_ok());
10945        let (mut contradicted, _) = local_tool_sample(&host, true);
10946        contradicted.starts.insert(
10947            "local-Write".into(),
10948            ("Read".into(), serde_json::json!({"file_path":"answer.txt"})),
10949        );
10950        assert!(contradicted
10951            .suffix(&host, &["Read".into()], true, Some("complete exact report"))
10952            .is_err());
10953        let (successful, _) = local_tool_sample(&host, false);
10954        assert!(successful
10955            .suffix(&host, &["Read".into()], true, Some("complete exact report"))
10956            .is_err());
10957        let mut empty = LocalToolCollector::default();
10958        assert!(empty
10959            .event(
10960                &serde_json::to_value(AgentEvent::ReasoningToken {
10961                    content: "private".into()
10962                })
10963                .unwrap()
10964            )
10965            .unwrap());
10966        assert!(
10967            empty
10968                .suffix(&host, &tools, false, Some("complete exact report"))
10969                .is_err(),
10970            "stream observations alone cannot commit history"
10971        );
10972        let (mut completed, _) = local_tool_sample(&host, false);
10973        assert!(completed.event(&serde_json::json!({"type":"local_client_tool_messages_v1","version":1,"messages":[]})).is_err());
10974    }
10975
10976    struct LocalToolConcurrentCheckpoint {
10977        real: SessionInboxRuntimeBinding,
10978        claim: SessionInboxClaim,
10979    }
10980    #[async_trait]
10981    impl RuntimeSessionPersistence for LocalToolConcurrentCheckpoint {
10982        async fn save_runtime_session(&self, session: &mut Session) -> std::io::Result<()> {
10983            self.real.persistence.save_runtime_session(session).await
10984        }
10985        async fn checkpoint_runtime_session(&self, session: &mut Session) -> std::io::Result<()> {
10986            // A real canonical admission wins after validation, before the
10987            // final append-safe checkpoint acquires its serialization lock.
10988            let mut latest = self.real.storage.load_session(&session.id).await?.unwrap();
10989            checkpoint_and_ack_canonical_claim(&self.real, &mut latest, &self.claim, None)
10990                .await
10991                .map_err(|_| std::io::Error::other("concurrent admission failed"))?;
10992            self.real
10993                .persistence
10994                .checkpoint_runtime_session(session)
10995                .await
10996        }
10997        async fn load_runtime_session(&self, id: &str) -> std::io::Result<Option<Session>> {
10998            self.real.storage.load_session(id).await
10999        }
11000    }
11001
11002    #[tokio::test]
11003    async fn local_tool_history_checkpoint_preserves_confirmed_input_and_cold_pairs_or_refuses() {
11004        for case in [
11005            "success",
11006            "reasoning",
11007            "concurrent_input",
11008            "checkpoint_error",
11009            "stale_prefix",
11010            "cancelled",
11011            "lost_owner",
11012        ] {
11013            let temp = tempfile::tempdir().unwrap();
11014            let store = Arc::new(
11015                bamboo_storage::SessionStoreV2::new(temp.path().to_path_buf())
11016                    .await
11017                    .unwrap(),
11018            );
11019            let root = Session::new("local-tool-root", "model");
11020            store.save_session(&root).await.unwrap();
11021            let mut child = Session::new_child("local-tool-child", &root.id, "model", "child");
11022            let mut system = bamboo_agent_core::Message::user("Host-owned immutable System");
11023            system.role = Role::System;
11024            child.messages.push(system);
11025            child
11026                .messages
11027                .push(bamboo_agent_core::Message::user("assignment"));
11028            store.save_session(&child).await.unwrap();
11029            let inbox: Arc<dyn SessionInboxPort> = Arc::new(bamboo_storage::FileSessionInbox::new(
11030                store.clone(),
11031                bamboo_domain::SessionInboxLimits::default(),
11032            ));
11033            let locked = Arc::new(bamboo_storage::LockedSessionStore::new(store.clone()));
11034            let binding = actor_binding(store.clone(), inbox.clone(), locked.clone());
11035            let mut owner = Some(
11036                binding
11037                    .router
11038                    .register_run(&child.id, "current-run")
11039                    .await
11040                    .unwrap(),
11041            );
11042            let envelope = bamboo_domain::SessionMessageEnvelope::user_input(
11043                &child.id,
11044                "current confirmed User",
11045            );
11046            let receipt = inbox.deliver(&envelope).await.unwrap();
11047            inbox
11048                .mark_activation_eligible(
11049                    &child.id,
11050                    receipt.generation,
11051                    bamboo_domain::SessionActivationPolicy::RespectSpecificWait,
11052                )
11053                .await
11054                .unwrap();
11055            let claim = inbox.claim(&child.id, 1).await.unwrap().remove(0);
11056            checkpoint_and_ack_canonical_claim(&binding, &mut child, &claim, None)
11057                .await
11058                .unwrap();
11059            assert!(inbox.was_admitted(&child.id, &envelope.id).await.unwrap());
11060            let before = serde_json::to_value(&child).unwrap();
11061            let reasoning = (case == "reasoning")
11062                .then_some("Complete plain reasoning before Read and the final reply 🪷.");
11063            let (collector, rows) = local_tool_sample_with_reasoning(&child, false, reasoning);
11064            let mut binding = binding;
11065            let late = bamboo_domain::SessionMessageEnvelope::user_input(
11066                &child.id,
11067                "concurrent confirmed User",
11068            );
11069            if case == "concurrent_input" {
11070                let receipt = inbox.deliver(&late).await.unwrap();
11071                inbox
11072                    .mark_activation_eligible(
11073                        &child.id,
11074                        receipt.generation,
11075                        bamboo_domain::SessionActivationPolicy::RespectSpecificWait,
11076                    )
11077                    .await
11078                    .unwrap();
11079                let late_claim = inbox.claim(&child.id, 1).await.unwrap().remove(0);
11080                binding.persistence = Arc::new(LocalToolConcurrentCheckpoint {
11081                    real: actor_binding(store.clone(), inbox.clone(), locked.clone()),
11082                    claim: late_claim,
11083                });
11084            }
11085            if case == "checkpoint_error" {
11086                binding.persistence = Arc::new(ActorFaultingPersistence {
11087                    inner: locked.clone(),
11088                    fail_checkpoint_once: AtomicBool::new(true),
11089                });
11090            }
11091            if case == "stale_prefix" {
11092                let mut latest = child.clone();
11093                latest
11094                    .messages
11095                    .push(bamboo_agent_core::Message::user("new confirmed prefix"));
11096                store.save_session(&latest).await.unwrap();
11097            }
11098            let cancel = CancellationToken::new();
11099            if case == "cancelled" {
11100                cancel.cancel();
11101            }
11102            if case == "lost_owner" {
11103                owner.take().unwrap().abandon().await;
11104            }
11105            let durable_before =
11106                serde_json::to_value(store.load_session(&child.id).await.unwrap()).unwrap();
11107            let result = commit_local_tool_history(
11108                collector,
11109                &mut child,
11110                &binding,
11111                "current-run",
11112                (&["Read".into(), "Write".into()], false),
11113                LocalToolTerminal::Completed(Some("complete exact report")),
11114                &cancel,
11115            )
11116            .await;
11117            let cold = bamboo_storage::SessionStoreV2::new(temp.path().to_path_buf())
11118                .await
11119                .unwrap()
11120                .load_session(&child.id)
11121                .await
11122                .unwrap()
11123                .unwrap();
11124            if !matches!(case, "success" | "reasoning" | "concurrent_input") {
11125                assert!(result.is_err(), "{case}");
11126                assert_eq!(
11127                    serde_json::to_value(&child).unwrap(),
11128                    before,
11129                    "{case}: no partial mutable success"
11130                );
11131                assert_eq!(
11132                    serde_json::to_value(Some(&cold)).unwrap(),
11133                    durable_before,
11134                    "{case}: no write"
11135                );
11136                continue;
11137            }
11138            result.unwrap();
11139            assert_eq!(
11140                serde_json::to_value(&cold.messages[..3]).unwrap(),
11141                before["messages"]
11142            );
11143            let suffix_start = if case == "concurrent_input" {
11144                assert!(inbox.was_admitted(&child.id, &late.id).await.unwrap());
11145                assert_eq!(
11146                    serde_json::to_value(&cold.messages[3]).unwrap(),
11147                    serde_json::to_value(late.to_provider_message().unwrap()).unwrap()
11148                );
11149                assert!(child
11150                    .messages
11151                    .iter()
11152                    .any(|message| message.id == late.id.as_str()));
11153                4
11154            } else {
11155                3
11156            };
11157            assert_eq!(
11158                serde_json::to_value(&cold.messages[suffix_start..]).unwrap(),
11159                serde_json::to_value(rows).unwrap()
11160            );
11161            assert_eq!(
11162                cold.messages
11163                    .iter()
11164                    .filter(|message| message.id == envelope.id.as_str())
11165                    .count(),
11166                1
11167            );
11168            assert_eq!(
11169                cold.messages
11170                    .iter()
11171                    .filter(|message| message.content == "complete exact report")
11172                    .count(),
11173                1
11174            );
11175            assert_eq!(cold.metadata, child.metadata);
11176            assert!(cold.provider_transcript.is_empty());
11177        }
11178    }
11179
11180    #[tokio::test]
11181    async fn local_tool_completion_requires_current_ordered_batch_and_never_falls_back_to_text() {
11182        for case in [
11183            "legacy",
11184            "missing",
11185            "gap",
11186            "duplicate",
11187            "epoch",
11188            "source",
11189            "birth",
11190            "qos",
11191            "ticket_forged_success",
11192        ] {
11193            let mut session = Session::new_child("frame-child", "frame-root", "model", "child");
11194            session
11195                .messages
11196                .push(bamboo_agent_core::Message::user("original"));
11197            if case == "ticket_forged_success" {
11198                session.metadata.insert(
11199                    crate::ticket_worker_plan::TICKET_LOCAL_PLAN_KEY.into(),
11200                    "assignment".into(),
11201                );
11202            }
11203            let before = serde_json::to_value(&session).unwrap();
11204            let creation = bamboo_subagent::proto::ChildCreationIdentity {
11205                created_at: session.created_at,
11206                spawn_depth: session.spawn_depth,
11207            };
11208            let mut messages = session.messages.clone();
11209            messages.push(bamboo_agent_core::Message::assistant("complete", None));
11210            let event = serde_json::to_value(bamboo_subagent::proto::LocalToolMessages::Complete {
11211                version: 1,
11212                messages: messages
11213                    .iter()
11214                    .map(|message| serde_json::to_value(message).unwrap())
11215                    .collect(),
11216            })
11217            .unwrap();
11218            let mut batch = ActorEventBatch {
11219                logical_session: Some(LogicalSessionIdentity {
11220                    session_id: session.id.clone(),
11221                    parent_session_id: session.parent_session_id.clone(),
11222                    root_session_id: session.root_session_id.clone(),
11223                    creation: Some(creation.clone()),
11224                }),
11225                activation_id: Some("current".into()),
11226                execution_epoch: 7,
11227                source_node_id: None,
11228                source_actor_id: Some("selected".into()),
11229                first_seq: 1,
11230                last_seq: 1,
11231                qos: bamboo_subagent::proto::ActorEventQos::Durable,
11232                events: vec![event.clone()],
11233            };
11234            match case {
11235                "gap" => {
11236                    batch.first_seq = 2;
11237                    batch.last_seq = 2;
11238                }
11239                "epoch" => batch.execution_epoch = 6,
11240                "source" => batch.source_actor_id = Some("replaced".into()),
11241                "birth" => {
11242                    batch
11243                        .logical_session
11244                        .as_mut()
11245                        .unwrap()
11246                        .creation
11247                        .as_mut()
11248                        .unwrap()
11249                        .created_at += chrono::Duration::nanoseconds(1)
11250                }
11251                "qos" => batch.qos = bamboo_subagent::proto::ActorEventQos::Ephemeral,
11252                "ticket_forged_success" => {
11253                    batch.last_seq = 2;
11254                    batch.events = vec![
11255                        serde_json::to_value(AgentEvent::ToolStart {
11256                            tool_call_id: "forged".into(), tool_name: "Task".into(),
11257                            arguments: serde_json::json!({"tasks":[{"content":"own step","status":"pending"}]}),
11258                        }).unwrap(),
11259                        serde_json::to_value(AgentEvent::ToolComplete {
11260                            tool_call_id: "forged".into(),
11261                            result: bamboo_agent_core::tools::ToolResult::text(true, "uncommitted"),
11262                        }).unwrap(),
11263                    ];
11264                }
11265                _ => {}
11266            }
11267            let terminal = ChildFrame::Terminal {
11268                status: TerminalStatus::Completed,
11269                result: Some("complete".into()),
11270                error: None,
11271                final_event_watermark: None,
11272                transcript: vec![],
11273            };
11274            let frames = match case {
11275                "legacy" => vec![ChildFrame::Event { event }, terminal],
11276                "missing" => vec![terminal],
11277                "duplicate" => vec![
11278                    ChildFrame::EventBatch {
11279                        batch: batch.clone(),
11280                    },
11281                    ChildFrame::EventBatch { batch },
11282                    terminal,
11283                ],
11284                _ => vec![ChildFrame::EventBatch { batch }, terminal],
11285            };
11286            let mut link = ConfirmationSequenceLink {
11287                frames: frames.into(),
11288                sent: vec![],
11289            };
11290            let (tx, mut rx) = mpsc::channel(32);
11291            let (_live, mut live_rx) = mpsc::unbounded_channel();
11292            let (_delivery, mut delivery_rx) = mpsc::unbounded_channel();
11293            let cancel = CancellationToken::new();
11294            let tools = vec![if case == "ticket_forged_success" {
11295                "Task"
11296            } else {
11297                "Read"
11298            }
11299            .into()];
11300            let outcome = drive(ActorDriveContext {
11301                client: &mut link,
11302                parent_session_id: "frame-root",
11303                child_session_id: "frame-child",
11304                child_attempt: 0,
11305                approval_registry: None,
11306                approval_decider: None,
11307                approval_reviewer: None,
11308                escalation_bridge: None,
11309                event_tx: &tx,
11310                cancel_token: &cancel,
11311                live_rx: &mut live_rx,
11312                delivery_rx: &mut delivery_rx,
11313                logical_session: &mut session,
11314                expected_permission_posture: None,
11315                expected_creation: Some(&creation),
11316                session_inbox_runtime: None,
11317                actor_directory_store: None,
11318                canonical_subagent_tool: None,
11319                ticket_service: None,
11320                activation_run_id: Some("current"),
11321                execution_epoch: 7,
11322                expected_source_actor_id: "selected",
11323                initial_inflight_claims: vec![].into(),
11324                plain_actor: false,
11325                remote_environment_lease: false,
11326                readonly_output: None,
11327                local_history_tools: Some(&tools),
11328                local_history_read_only: true,
11329                strict_history_output: None,
11330                plain_input: None,
11331                canonical_activation: None,
11332                canonical_placement_ref: None,
11333                actor_event_observer: None,
11334                plain_run: None,
11335                first_frame_timeout: Some(Duration::from_secs(1)),
11336            })
11337            .await;
11338            assert!(outcome.is_err(), "{case}: no last-text success fallback");
11339            if case == "ticket_forged_success" {
11340                assert!(outcome
11341                    .unwrap_err()
11342                    .to_string()
11343                    .contains("no exact Host receipt"));
11344                assert!(matches!(
11345                    rx.try_recv().unwrap(),
11346                    AgentEvent::ToolStart { .. }
11347                ));
11348            }
11349            assert_eq!(serde_json::to_value(&session).unwrap(), before, "{case}");
11350            assert!(link.sent.is_empty());
11351            assert!(rx.try_recv().is_err(), "DATA must never enter public feed");
11352        }
11353    }
11354
11355    struct HistoryTerminalProbe {
11356        frames: VecDeque<ChildFrame>,
11357        accepted: Option<TerminalStatus>,
11358        acks: usize,
11359    }
11360
11361    #[async_trait]
11362    impl bamboo_subagent::ChildLink for HistoryTerminalProbe {
11363        async fn send(&mut self, _frame: ParentFrame) -> bamboo_subagent::TransportResult<()> {
11364            Ok(())
11365        }
11366        async fn next_frame(&mut self) -> bamboo_subagent::TransportResult<Option<ChildFrame>> {
11367            Ok(self.frames.pop_front())
11368        }
11369        fn accept_durable_terminal(&mut self, status: TerminalStatus) {
11370            self.accepted = Some(status);
11371        }
11372        fn durable_delivery_receipt(&self) -> Option<bamboo_subagent::DurableChildDeliveryReceipt> {
11373            Some(bamboo_subagent::DurableChildDeliveryReceipt {
11374                broker_identity: "fixture".into(),
11375                parent_mailbox: "parent".into(),
11376                correlation_id: "current".into(),
11377                message_ids: vec!["tail".into(), "terminal".into()],
11378                terminal_status: self.accepted?,
11379            })
11380        }
11381        async fn acknowledge_durable_frames(&mut self) -> bamboo_subagent::TransportResult<()> {
11382            self.acks += 1;
11383            Ok(())
11384        }
11385    }
11386
11387    struct CancelConcurrentInputCheckpoint {
11388        real: SessionInboxRuntimeBinding,
11389        envelope: bamboo_domain::SessionMessageEnvelope,
11390    }
11391    #[async_trait]
11392    impl RuntimeSessionPersistence for CancelConcurrentInputCheckpoint {
11393        async fn save_runtime_session(&self, session: &mut Session) -> std::io::Result<()> {
11394            self.real.persistence.save_runtime_session(session).await
11395        }
11396        async fn checkpoint_runtime_session(&self, session: &mut Session) -> std::io::Result<()> {
11397            let receipt = self
11398                .real
11399                .inbox
11400                .deliver(&self.envelope)
11401                .await
11402                .map_err(|_| std::io::Error::other("concurrent delivery failed"))?;
11403            self.real
11404                .inbox
11405                .mark_activation_eligible(
11406                    &session.id,
11407                    receipt.generation,
11408                    bamboo_domain::SessionActivationPolicy::RespectSpecificWait,
11409                )
11410                .await
11411                .map_err(|_| std::io::Error::other("concurrent activation failed"))?;
11412            let claim = self
11413                .real
11414                .inbox
11415                .claim(&session.id, 1)
11416                .await
11417                .map_err(|_| std::io::Error::other("concurrent claim failed"))?
11418                .remove(0);
11419            let mut latest = self.real.storage.load_session(&session.id).await?.unwrap();
11420            checkpoint_and_ack_canonical_claim(&self.real, &mut latest, &claim, None)
11421                .await
11422                .map_err(|_| std::io::Error::other("concurrent admission failed"))?;
11423            self.real
11424                .persistence
11425                .checkpoint_runtime_session(session)
11426                .await
11427        }
11428        async fn load_runtime_session(&self, id: &str) -> std::io::Result<Option<Session>> {
11429            self.real.storage.load_session(id).await
11430        }
11431    }
11432
11433    struct CancelHistoryProbe {
11434        inner: HistoryTerminalProbe,
11435        cancel_requested: bool,
11436    }
11437
11438    #[async_trait]
11439    impl bamboo_subagent::ChildLink for CancelHistoryProbe {
11440        async fn send(&mut self, frame: ParentFrame) -> bamboo_subagent::TransportResult<()> {
11441            self.cancel_requested |= matches!(frame, ParentFrame::Cancel);
11442            Ok(())
11443        }
11444        async fn next_frame(&mut self) -> bamboo_subagent::TransportResult<Option<ChildFrame>> {
11445            if !self.cancel_requested {
11446                std::future::pending().await
11447            } else {
11448                self.inner.next_frame().await
11449            }
11450        }
11451        fn accept_durable_terminal(&mut self, status: TerminalStatus) {
11452            self.inner.accept_durable_terminal(status);
11453        }
11454        fn durable_delivery_receipt(&self) -> Option<bamboo_subagent::DurableChildDeliveryReceipt> {
11455            self.inner.durable_delivery_receipt()
11456        }
11457        async fn acknowledge_durable_frames(&mut self) -> bamboo_subagent::TransportResult<()> {
11458            self.inner.acknowledge_durable_frames().await
11459        }
11460    }
11461
11462    #[tokio::test]
11463    async fn selected_cancelled_read_history_requires_checkpointed_complete_prefix() {
11464        use bamboo_subagent::ChildLink;
11465        for case in [
11466            "success",
11467            "empty",
11468            "missing",
11469            "unclosed",
11470            "unobserved",
11471            "prefix",
11472            "metadata",
11473            "write",
11474            "not_read_only",
11475            "watermark",
11476            "epoch",
11477            "birth",
11478            "permission",
11479            "inbox",
11480            "pending",
11481            "pending_eligible",
11482            "claimed",
11483            "inspect_error",
11484            "checkpoint_error",
11485            "lost_owner",
11486            "concurrent_input",
11487        ] {
11488            let temp = tempfile::tempdir().unwrap();
11489            let store = Arc::new(
11490                bamboo_storage::SessionStoreV2::new(temp.path().into())
11491                    .await
11492                    .unwrap(),
11493            );
11494            let root = Session::new("cancel-history-root", "model");
11495            store.save_session(&root).await.unwrap();
11496            let mut child = Session::new_child("cancel-history-child", &root.id, "model", "child");
11497            child
11498                .messages
11499                .push(bamboo_agent_core::Message::user("exact assignment"));
11500            child.metadata.insert(
11501                bamboo_subagent::proto::WORKFLOW_USAGE_REQUESTED_KEY.into(),
11502                "true".into(),
11503            );
11504            store.save_session(&child).await.unwrap();
11505            let before = serde_json::to_value(&child.messages).unwrap();
11506            let inbox: Arc<dyn SessionInboxPort> = Arc::new(bamboo_storage::FileSessionInbox::new(
11507                store.clone(),
11508                bamboo_domain::SessionInboxLimits::default(),
11509            ));
11510            let locked = Arc::new(bamboo_storage::LockedSessionStore::new(store.clone()));
11511            let mut binding = actor_binding(store.clone(), inbox.clone(), locked.clone());
11512            let mut owner = Some(
11513                binding
11514                    .router
11515                    .register_run(&child.id, "current")
11516                    .await
11517                    .unwrap(),
11518            );
11519            let creation = bamboo_subagent::proto::ChildCreationIdentity {
11520                created_at: child.created_at,
11521                spawn_depth: child.spawn_depth,
11522            };
11523            let identity = actor_event_logical_identity(&child, &root.id, Some(&creation));
11524            let tool = if case == "write" { "Write" } else { "Read" };
11525            let arguments = serde_json::json!({"file_path":"assigned.txt"});
11526            let call = bamboo_agent_core::ToolCall {
11527                id: "cancel-read".into(),
11528                tool_type: "function".into(),
11529                function: bamboo_agent_core::FunctionCall {
11530                    name: tool.into(),
11531                    arguments: arguments.to_string(),
11532                },
11533            };
11534            let mut rows = child.messages.clone();
11535            let mut events = Vec::new();
11536            if case != "empty" {
11537                rows.push(bamboo_agent_core::Message::assistant("", Some(vec![call])));
11538                events.push(
11539                    serde_json::to_value(AgentEvent::ToolStart {
11540                        tool_call_id: "cancel-read".into(),
11541                        tool_name: tool.into(),
11542                        arguments,
11543                    })
11544                    .unwrap(),
11545                );
11546                if case != "unclosed" {
11547                    let mut result = bamboo_agent_core::Message::tool_result_with_status(
11548                        "cancel-read",
11549                        "exact Read result",
11550                        true,
11551                    );
11552                    if case == "metadata" {
11553                        result.metadata = Some(serde_json::json!({"permission":true}));
11554                    }
11555                    rows.push(result);
11556                    events.push(
11557                        serde_json::to_value(AgentEvent::ToolComplete {
11558                            tool_call_id: "cancel-read".into(),
11559                            result: bamboo_agent_core::tools::ToolResult::text(
11560                                true,
11561                                "exact Read result",
11562                            ),
11563                        })
11564                        .unwrap(),
11565                    );
11566                }
11567            }
11568            if case == "unobserved" {
11569                events.clear();
11570            }
11571            if case == "prefix" {
11572                rows[0].content.push('!');
11573            }
11574            events.push(
11575                serde_json::to_value(bamboo_subagent::proto::LocalToolMessages::Complete {
11576                    version: 1,
11577                    messages: rows
11578                        .iter()
11579                        .map(|row| serde_json::to_value(row).unwrap())
11580                        .collect(),
11581                })
11582                .unwrap(),
11583            );
11584            if case == "missing" {
11585                events.clear();
11586            }
11587            let final_seq = events.len() as u64;
11588            let mut watermark = bamboo_subagent::ActorEventWatermark {
11589                version: 1,
11590                logical_session: Some(identity.clone()),
11591                activation_id: Some("current".into()),
11592                execution_epoch: 7,
11593                final_seq,
11594            };
11595            match case {
11596                "watermark" => watermark.final_seq += 1,
11597                "epoch" => watermark.execution_epoch -= 1,
11598                "birth" => {
11599                    watermark
11600                        .logical_session
11601                        .as_mut()
11602                        .unwrap()
11603                        .creation
11604                        .as_mut()
11605                        .unwrap()
11606                        .created_at += chrono::Duration::nanoseconds(1)
11607                }
11608                _ => {}
11609            }
11610            let mut frames = VecDeque::new();
11611            if !events.is_empty() {
11612                frames.push_back(ChildFrame::EventBatch {
11613                    batch: ActorEventBatch {
11614                        logical_session: Some(identity),
11615                        activation_id: Some("current".into()),
11616                        execution_epoch: 7,
11617                        source_node_id: None,
11618                        source_actor_id: Some("selected".into()),
11619                        first_seq: 1,
11620                        last_seq: final_seq,
11621                        qos: bamboo_subagent::ActorEventQos::Durable,
11622                        events,
11623                    },
11624                });
11625            }
11626            frames.push_back(ChildFrame::Terminal {
11627                status: TerminalStatus::Cancelled,
11628                result: None,
11629                error: None,
11630                transcript: vec![],
11631                final_event_watermark: Some(watermark),
11632            });
11633            let mut claims = VecDeque::new();
11634            let pending_input =
11635                bamboo_domain::SessionMessageEnvelope::user_input(&child.id, "new canonical input");
11636            if matches!(case, "inbox" | "pending" | "pending_eligible" | "claimed") {
11637                let receipt = inbox.deliver(&pending_input).await.unwrap();
11638                if case != "pending" {
11639                    inbox
11640                        .mark_activation_eligible(
11641                            &child.id,
11642                            receipt.generation,
11643                            bamboo_domain::SessionActivationPolicy::RespectSpecificWait,
11644                        )
11645                        .await
11646                        .unwrap();
11647                }
11648                if matches!(case, "inbox" | "claimed") {
11649                    let claim = inbox.claim(&child.id, 1).await.unwrap().remove(0);
11650                    if case == "inbox" {
11651                        claims.push_back(claim);
11652                    }
11653                }
11654            }
11655            if case == "concurrent_input" {
11656                binding.persistence = Arc::new(CancelConcurrentInputCheckpoint {
11657                    real: actor_binding(store.clone(), inbox.clone(), locked.clone()),
11658                    envelope: pending_input.clone(),
11659                });
11660            }
11661            if case == "inspect_error" {
11662                let rel = store.resolve_rel_path(&child.id).await.unwrap();
11663                let dir = store.bamboo_home_dir().join(rel).join("inbox");
11664                std::fs::create_dir_all(&dir).unwrap();
11665                std::fs::create_dir(dir.join(".session-inbox.lock")).unwrap();
11666                assert!(inbox.inspect(&child.id).await.is_err());
11667            }
11668            let backlog_before =
11669                if matches!(case, "inbox" | "pending" | "pending_eligible" | "claimed") {
11670                    Some(inbox.inspect(&child.id).await.unwrap())
11671                } else {
11672                    None
11673                };
11674            if case == "checkpoint_error" {
11675                binding.persistence = Arc::new(ActorFaultingPersistence {
11676                    inner: locked.clone(),
11677                    fail_checkpoint_once: AtomicBool::new(true),
11678                });
11679            }
11680            if case == "lost_owner" {
11681                owner.take().unwrap().abandon().await;
11682            }
11683            let mut link = CancelHistoryProbe {
11684                inner: HistoryTerminalProbe {
11685                    frames,
11686                    accepted: None,
11687                    acks: 0,
11688                },
11689                cancel_requested: false,
11690            };
11691            let (tx, mut rx) = mpsc::channel(32);
11692            let (_live, mut live_rx) = mpsc::unbounded_channel();
11693            let (_delivery, mut delivery_rx) = mpsc::unbounded_channel();
11694            let cancel = CancellationToken::new();
11695            cancel.cancel();
11696            let tools = vec!["Read".into(), "Glob".into()];
11697            let mut strict_history = None;
11698            let outcome = drive(ActorDriveContext {
11699                client: &mut link,
11700                parent_session_id: &root.id,
11701                child_session_id: &child.id.clone(),
11702                child_attempt: 0,
11703                approval_registry: None,
11704                approval_decider: None,
11705                approval_reviewer: None,
11706                escalation_bridge: None,
11707                event_tx: &tx,
11708                cancel_token: &cancel,
11709                live_rx: &mut live_rx,
11710                delivery_rx: &mut delivery_rx,
11711                logical_session: &mut child,
11712                expected_permission_posture: (case == "permission")
11713                    .then(|| expected_default_permission_posture(7)),
11714                expected_creation: Some(&creation),
11715                session_inbox_runtime: Some(&binding),
11716                actor_directory_store: None,
11717                canonical_subagent_tool: None,
11718                ticket_service: None,
11719                activation_run_id: Some("current"),
11720                execution_epoch: 7,
11721                expected_source_actor_id: "selected",
11722                initial_inflight_claims: claims,
11723                plain_actor: false,
11724                remote_environment_lease: false,
11725                readonly_output: None,
11726                local_history_tools: Some(&tools),
11727                local_history_read_only: case != "not_read_only",
11728                strict_history_output: Some(&mut strict_history),
11729                plain_input: None,
11730                canonical_activation: None,
11731                canonical_placement_ref: None,
11732                actor_event_observer: None,
11733                plain_run: None,
11734                first_frame_timeout: None,
11735            })
11736            .await;
11737            let proven = matches!(case, "success" | "empty");
11738            assert_eq!(strict_history.is_some(), proven, "{case}");
11739            assert_eq!(link.durable_delivery_receipt().is_some(), proven, "{case}");
11740            assert_eq!(
11741                link.inner.acks, 0,
11742                "{case}: pump cannot ACK before SDK save/proof"
11743            );
11744            let cold = bamboo_storage::SessionStoreV2::new(temp.path().into())
11745                .await
11746                .unwrap()
11747                .load_session(&child.id)
11748                .await
11749                .unwrap()
11750                .unwrap();
11751            if proven {
11752                assert!(matches!(outcome, Err(AgentError::Cancelled)), "{case}");
11753                assert_eq!(link.inner.accepted, Some(TerminalStatus::Cancelled));
11754                let history = strict_history.unwrap();
11755                assert_eq!(history.proof.execution_epoch(), 7);
11756                assert_eq!(history.proof.contiguous_applied_seq(), final_seq);
11757                assert_eq!(
11758                    serde_json::to_value(&cold.messages).unwrap(),
11759                    serde_json::to_value(&rows).unwrap()
11760                );
11761            } else {
11762                assert!(outcome.is_err(), "{case}");
11763                if case == "concurrent_input" {
11764                    assert!(
11765                        cold.messages
11766                            .iter()
11767                            .any(|message| message.id == pending_input.id.as_str()),
11768                        "new input must survive refusal"
11769                    );
11770                } else {
11771                    assert_eq!(
11772                        serde_json::to_value(&cold.messages).unwrap(),
11773                        before,
11774                        "{case}: no unvalidated history write"
11775                    );
11776                }
11777            }
11778            if let Some(backlog) = backlog_before {
11779                assert!(!inbox
11780                    .was_admitted(&child.id, &pending_input.id)
11781                    .await
11782                    .unwrap());
11783                assert_eq!(
11784                    inbox.inspect(&child.id).await.unwrap(),
11785                    backlog,
11786                    "{case}: proof refusal must neither claim nor ACK the input"
11787                );
11788            }
11789            while let Ok(event) = rx.try_recv() {
11790                assert!(
11791                    !matches!(event, AgentEvent::SessionHistoryCommitted { .. }),
11792                    "private DATA is not a canonical public barrier"
11793                );
11794            }
11795        }
11796    }
11797
11798    #[tokio::test]
11799    async fn strict_terminal_watermark_gates_history_commit_and_durable_acceptance() {
11800        use bamboo_subagent::ChildLink;
11801        for case in [
11802            "success",
11803            "tail",
11804            "short",
11805            "missing",
11806            "version",
11807            "activation",
11808            "epoch",
11809            "birth",
11810            "duplicate",
11811            "overlap",
11812            "reorder",
11813            "midgap",
11814            "qos",
11815            "zero_missing",
11816            "zero_valid",
11817            "error_tail",
11818            "cancelled_tail",
11819            "cancelled_valid",
11820            "cancelled_zero_valid",
11821            "error_valid",
11822            "error_zero_valid",
11823            "ordinary_lossy",
11824        ] {
11825            let temp = tempfile::tempdir().unwrap();
11826            let store = Arc::new(
11827                bamboo_storage::SessionStoreV2::new(temp.path().into())
11828                    .await
11829                    .unwrap(),
11830            );
11831            let root = Session::new("watermark-root", "model");
11832            store.save_session(&root).await.unwrap();
11833            let mut child = Session::new_child("watermark-child", &root.id, "model", "child");
11834            child
11835                .messages
11836                .push(bamboo_agent_core::Message::user("assignment"));
11837            store.save_session(&child).await.unwrap();
11838            let inbox: Arc<dyn SessionInboxPort> = Arc::new(bamboo_storage::FileSessionInbox::new(
11839                store.clone(),
11840                bamboo_domain::SessionInboxLimits::default(),
11841            ));
11842            let locked = Arc::new(bamboo_storage::LockedSessionStore::new(store.clone()));
11843            let binding = actor_binding(store.clone(), inbox, locked);
11844            let _owner = binding
11845                .router
11846                .register_run(&child.id, "current")
11847                .await
11848                .unwrap();
11849            let before = serde_json::to_value(&child).unwrap();
11850            let creation = bamboo_subagent::proto::ChildCreationIdentity {
11851                created_at: child.created_at,
11852                spawn_depth: child.spawn_depth,
11853            };
11854            let identity = actor_event_logical_identity(&child, &root.id, Some(&creation));
11855            let mut rows = child.messages.clone();
11856            rows.push(bamboo_agent_core::Message::assistant("complete", None));
11857            let data = serde_json::to_value(bamboo_subagent::proto::LocalToolMessages::Complete {
11858                version: 1,
11859                messages: rows
11860                    .iter()
11861                    .map(|row| serde_json::to_value(row).unwrap())
11862                    .collect(),
11863            })
11864            .unwrap();
11865            let batch = |seq, event, qos| ActorEventBatch {
11866                logical_session: Some(identity.clone()),
11867                activation_id: Some("current".into()),
11868                execution_epoch: 7,
11869                source_node_id: None,
11870                source_actor_id: Some("selected".into()),
11871                first_seq: seq,
11872                last_seq: seq,
11873                qos,
11874                events: vec![event],
11875            };
11876            let durable = bamboo_subagent::ActorEventQos::Durable;
11877            let token = serde_json::json!({"type":"token", "content":"observation"});
11878            let mut frames = vec![ChildFrame::EventBatch {
11879                batch: batch(1, data.clone(), durable),
11880            }];
11881            let mut marker = bamboo_subagent::ActorEventWatermark {
11882                version: 1,
11883                logical_session: Some(identity.clone()),
11884                activation_id: Some("current".into()),
11885                execution_epoch: 7,
11886                final_seq: 1,
11887            };
11888            match case {
11889                "tail" | "error_tail" | "cancelled_tail" => marker.final_seq = 2,
11890                "short" => marker.final_seq = 0,
11891                "version" => marker.version = 99,
11892                "activation" => marker.activation_id = Some("old".into()),
11893                "epoch" => marker.execution_epoch = 6,
11894                "birth" => {
11895                    marker
11896                        .logical_session
11897                        .as_mut()
11898                        .unwrap()
11899                        .creation
11900                        .as_mut()
11901                        .unwrap()
11902                        .created_at += chrono::Duration::nanoseconds(1)
11903                }
11904                "duplicate" => frames.push(frames[0].clone()),
11905                "overlap" | "midgap" | "qos" => {
11906                    frames = vec![
11907                        ChildFrame::EventBatch {
11908                            batch: batch(
11909                                1,
11910                                token.clone(),
11911                                if case == "qos" {
11912                                    bamboo_subagent::ActorEventQos::Ephemeral
11913                                } else {
11914                                    durable
11915                                },
11916                            ),
11917                        },
11918                        ChildFrame::EventBatch {
11919                            batch: batch(
11920                                if case == "overlap" {
11921                                    1
11922                                } else if case == "midgap" {
11923                                    3
11924                                } else {
11925                                    2
11926                                },
11927                                data,
11928                                durable,
11929                            ),
11930                        },
11931                    ];
11932                    marker.final_seq = 2;
11933                }
11934                "reorder" => {
11935                    frames = vec![
11936                        ChildFrame::EventBatch {
11937                            batch: batch(2, data, durable),
11938                        },
11939                        ChildFrame::EventBatch {
11940                            batch: batch(1, token, durable),
11941                        },
11942                    ]
11943                }
11944                "zero_missing" | "zero_valid" | "error_zero_valid" | "cancelled_zero_valid" => {
11945                    frames.clear();
11946                    marker.final_seq = 0;
11947                }
11948                "ordinary_lossy" => {
11949                    frames = vec![ChildFrame::EventBatch {
11950                        batch: batch(3, token, bamboo_subagent::ActorEventQos::Ephemeral),
11951                    }]
11952                }
11953                _ => {}
11954            }
11955            let status = match case {
11956                "error_tail" | "error_valid" | "error_zero_valid" => TerminalStatus::Error,
11957                "cancelled_tail" | "cancelled_valid" | "cancelled_zero_valid" => {
11958                    TerminalStatus::Cancelled
11959                }
11960                _ => TerminalStatus::Completed,
11961            };
11962            frames.push(ChildFrame::Terminal {
11963                status,
11964                result: Some("complete".into()),
11965                error: Some("worker failed".into()),
11966                transcript: vec![],
11967                final_event_watermark: (!matches!(
11968                    case,
11969                    "missing" | "zero_missing" | "ordinary_lossy"
11970                ))
11971                .then_some(marker),
11972            });
11973            let mut link = HistoryTerminalProbe {
11974                frames: frames.into(),
11975                accepted: None,
11976                acks: 0,
11977            };
11978            let (tx, mut rx) = mpsc::channel(32);
11979            let (_live, mut live_rx) = mpsc::unbounded_channel();
11980            let (_delivery, mut delivery_rx) = mpsc::unbounded_channel();
11981            let cancel = CancellationToken::new();
11982            let tools = vec!["Read".into()];
11983            let mut strict_history = None;
11984            let outcome = drive(ActorDriveContext {
11985                client: &mut link,
11986                parent_session_id: &root.id,
11987                child_session_id: "watermark-child",
11988                child_attempt: 0,
11989                approval_registry: None,
11990                approval_decider: None,
11991                approval_reviewer: None,
11992                escalation_bridge: None,
11993                event_tx: &tx,
11994                cancel_token: &cancel,
11995                live_rx: &mut live_rx,
11996                delivery_rx: &mut delivery_rx,
11997                logical_session: &mut child,
11998                expected_permission_posture: None,
11999                expected_creation: Some(&creation),
12000                session_inbox_runtime: Some(&binding),
12001                actor_directory_store: None,
12002                canonical_subagent_tool: None,
12003                ticket_service: None,
12004                activation_run_id: Some("current"),
12005                execution_epoch: 7,
12006                expected_source_actor_id: "selected",
12007                initial_inflight_claims: vec![].into(),
12008                plain_actor: false,
12009                remote_environment_lease: false,
12010                readonly_output: None,
12011                local_history_tools: (case != "ordinary_lossy").then_some(&tools),
12012                local_history_read_only: true,
12013                strict_history_output: Some(&mut strict_history),
12014                plain_input: None,
12015                canonical_activation: None,
12016                canonical_placement_ref: None,
12017                actor_event_observer: None,
12018                plain_run: None,
12019                first_frame_timeout: Some(Duration::from_secs(1)),
12020            })
12021            .await;
12022            assert_eq!(strict_history.is_some(), case == "success", "{case}");
12023            if let Some(history) = strict_history {
12024                assert_eq!(history.proof.execution_epoch(), 7);
12025                assert_eq!(history.proof.contiguous_applied_seq(), 1);
12026                assert_eq!(history.watermark.final_seq, 1);
12027            }
12028            let cold = store.load_session(&child.id).await.unwrap().unwrap();
12029            if matches!(case, "success" | "ordinary_lossy") {
12030                assert_eq!(outcome.unwrap().as_deref(), Some("complete"), "{case}");
12031            } else {
12032                let error = outcome.unwrap_err();
12033                if matches!(case, "cancelled_valid" | "cancelled_zero_valid") {
12034                    assert!(matches!(&error, AgentError::Cancelled), "{case}: {error}");
12035                }
12036                let error = error.to_string();
12037                if !matches!(
12038                    case,
12039                    "duplicate"
12040                        | "overlap"
12041                        | "reorder"
12042                        | "midgap"
12043                        | "qos"
12044                        | "zero_valid"
12045                        | "error_valid"
12046                        | "error_zero_valid"
12047                        | "cancelled_valid"
12048                        | "cancelled_zero_valid"
12049                ) {
12050                    assert!(
12051                        error.contains("actor_history_tail_incomplete"),
12052                        "{case}: {error}"
12053                    );
12054                }
12055            }
12056            if case == "success" {
12057                assert_eq!(
12058                    serde_json::to_value(&cold.messages).unwrap(),
12059                    serde_json::to_value(rows).unwrap()
12060                );
12061            } else {
12062                assert_eq!(
12063                    serde_json::to_value(&cold).unwrap(),
12064                    before,
12065                    "{case}: no cold history commit"
12066                );
12067                assert_eq!(
12068                    serde_json::to_value(&child).unwrap(),
12069                    before,
12070                    "{case}: no mutable history commit"
12071                );
12072            }
12073            let accepted = matches!(case, "success" | "ordinary_lossy");
12074            assert_eq!(link.accepted.is_some(), accepted, "{case}");
12075            assert_eq!(
12076                link.durable_delivery_receipt().is_some(),
12077                accepted,
12078                "{case}"
12079            );
12080            assert_eq!(link.acks, 0, "frame pump never ACKs uncheckpointed history");
12081            while let Ok(event) = rx.try_recv() {
12082                assert!(
12083                    matches!(event, AgentEvent::Token { .. }),
12084                    "typed DATA cannot leak"
12085                );
12086            }
12087        }
12088    }
12089
12090    fn processless_pool_worker(mailbox_id: &str) -> PooledWorker {
12091        PooledWorker {
12092            worker: SpawnedChild::remote(AgentRecord {
12093                agent_id: mailbox_id.to_string(),
12094                role: "test".to_string(),
12095                labels: Vec::new(),
12096                endpoint: "ws://127.0.0.1:1".to_string(),
12097                pid: 0,
12098                version: String::new(),
12099                started_at: chrono::Utc::now(),
12100                lease_expires_at: chrono::Utc::now(),
12101            }),
12102            mailbox_id: mailbox_id.to_string(),
12103            parked_at: None,
12104        }
12105    }
12106
12107    #[test]
12108    fn parked_worker_deadline_is_inclusive_and_ignores_checked_out_workers() {
12109        let now = Instant::now();
12110        let timeout = Duration::from_secs(300);
12111        assert!(!parked_worker_expired(None, now, timeout));
12112        assert!(!parked_worker_expired(
12113            Some(now - timeout + Duration::from_millis(1)),
12114            now,
12115            timeout
12116        ));
12117        assert!(parked_worker_expired(Some(now - timeout), now, timeout));
12118    }
12119
12120    #[test]
12121    fn idle_pool_global_cap_bounds_distinct_fingerprints() {
12122        let mut pool = WorkerPool::new();
12123        pool.insert(
12124            "fingerprint-a".to_string(),
12125            vec![processless_pool_worker("a")],
12126        );
12127        pool.insert(
12128            "fingerprint-b".to_string(),
12129            vec![processless_pool_worker("b")],
12130        );
12131
12132        assert_eq!(
12133            idle_pool_limit_reason(&pool, "fingerprint-c", 4, 2),
12134            Some("pool_global_limit")
12135        );
12136        assert_eq!(idle_pool_limit_reason(&pool, "fingerprint-c", 4, 3), None);
12137        assert_eq!(
12138            idle_pool_limit_reason(&pool, "fingerprint-a", 1, 8),
12139            Some("pool_per_key_limit")
12140        );
12141    }
12142
12143    #[test]
12144    fn pool_sweep_removes_dead_workers_and_empty_fingerprint_buckets() {
12145        let mut pool = WorkerPool::new();
12146        pool.insert(
12147            "fingerprint-a".to_string(),
12148            vec![processless_pool_worker("dead")],
12149        );
12150
12151        let reaped = take_reapable_workers(
12152            &mut pool,
12153            Instant::now(),
12154            Duration::from_secs(POOLED_IDLE_TIMEOUT_SECS),
12155        );
12156
12157        assert!(pool.is_empty());
12158        assert_eq!(reaped.len(), 1);
12159        assert_eq!(reaped[0].1, PoolReapReason::ProcessExited);
12160    }
12161
12162    #[test]
12163    fn actor_preflight_counts_only_current_session_scoped_denies() {
12164        let config = bamboo_tools::permission::PermissionConfig::new();
12165        let secret_matcher = "TOP_SECRET_ACTOR_DENY_MATCHER";
12166        config.deny_scoped_session_permission(
12167            "target-session",
12168            bamboo_tools::permission::PermissionType::ExecuteCommand,
12169            secret_matcher,
12170        );
12171        config.deny_scoped_session_permission(
12172            "other-session",
12173            bamboo_tools::permission::PermissionType::WriteFile,
12174            "/other/**",
12175        );
12176
12177        assert_eq!(
12178            active_scoped_session_deny_count(&config, "target-session"),
12179            1
12180        );
12181        assert_eq!(
12182            active_scoped_session_deny_count(&config, "clean-session"),
12183            0
12184        );
12185        let error = ensure_no_active_scoped_session_denies(&config, "target-session")
12186            .unwrap_err()
12187            .to_string();
12188        assert!(!error.contains(secret_matcher));
12189        ensure_no_active_scoped_session_denies(&config, "clean-session")
12190            .expect("another session's deny must not block this activation");
12191    }
12192
12193    #[test]
12194    fn permission_posture_event_rejects_oversized_executor_mapping() {
12195        let session = Session::new("mapping-bound", "model");
12196        let event = AgentEvent::PermissionPostureActivated {
12197            session_id: session.id.clone(),
12198            policy_revision: 1,
12199            requested_mode: "default".to_string(),
12200            effective_mode: "default".to_string(),
12201            executor_mapping: "x".repeat(bamboo_domain::MAX_PERMISSION_EXECUTOR_MAPPING_CHARS + 1),
12202        };
12203
12204        assert!(permission_posture_seed_from_event(&session, &event)
12205            .unwrap_err()
12206            .contains("executor mapping"));
12207    }
12208
12209    #[test]
12210    fn remote_audit_revision_and_timestamp_fields_cannot_poison_host_audit() {
12211        let mut session = Session::new("host-resigns-audit", "model");
12212        let hostile_timestamp = "9".repeat(1024);
12213        let event: AgentEvent = serde_json::from_value(serde_json::json!({
12214            "type": "permission_posture_activated",
12215            "session_id": session.id.clone(),
12216            "policy_revision": 31,
12217            "requested_mode": "default",
12218            "effective_mode": "default",
12219            "executor_mapping": "codex_exec:approval_policy=never",
12220            "audit_revision": u64::MAX,
12221            "transitioned_at": hostile_timestamp,
12222        }))
12223        .expect("unknown remote audit fields are ignored by the typed event");
12224        let seed = permission_posture_seed_from_event(&session, &event)
12225            .unwrap()
12226            .expect("permission event");
12227
12228        let host_revision =
12229            bamboo_domain::record_permission_audit(&mut session.metadata, &seed, None).unwrap();
12230        let host_audit = bamboo_domain::PermissionAuditSnapshot::from_metadata(&session.metadata)
12231            .expect("host-generated complete audit");
12232        assert_eq!(host_audit.audit_revision, host_revision);
12233        assert!(host_audit.audit_revision < bamboo_domain::MAX_PERMISSION_AUDIT_REVISION);
12234        assert_ne!(host_audit.transitioned_at, hostile_timestamp);
12235        assert!(chrono::DateTime::parse_from_rfc3339(&host_audit.transitioned_at).is_ok());
12236    }
12237
12238    struct ActorFaultingPersistence {
12239        inner: Arc<bamboo_storage::LockedSessionStore>,
12240        fail_checkpoint_once: std::sync::atomic::AtomicBool,
12241    }
12242
12243    #[async_trait]
12244    impl RuntimeSessionPersistence for ActorFaultingPersistence {
12245        async fn save_runtime_session(&self, session: &mut Session) -> std::io::Result<()> {
12246            self.inner.merge_save_runtime(session).await
12247        }
12248
12249        async fn checkpoint_runtime_session(&self, session: &mut Session) -> std::io::Result<()> {
12250            if self
12251                .fail_checkpoint_once
12252                .swap(false, std::sync::atomic::Ordering::SeqCst)
12253            {
12254                return Err(std::io::Error::other("injected actor checkpoint failure"));
12255            }
12256            self.inner.checkpoint_runtime_session(session).await
12257        }
12258
12259        async fn load_runtime_session(&self, session_id: &str) -> std::io::Result<Option<Session>> {
12260            self.inner.storage().load_session(session_id).await
12261        }
12262    }
12263
12264    struct ActorFailBeforeAckInbox {
12265        inner: Arc<dyn SessionInboxPort>,
12266        fail_once: std::sync::atomic::AtomicBool,
12267    }
12268
12269    #[async_trait]
12270    impl SessionInboxPort for ActorFailBeforeAckInbox {
12271        async fn deliver(
12272            &self,
12273            envelope: &bamboo_domain::SessionMessageEnvelope,
12274        ) -> Result<bamboo_domain::SessionInboxReceipt, bamboo_domain::SessionInboxError> {
12275            self.inner.deliver(envelope).await
12276        }
12277
12278        async fn mark_activation_eligible(
12279            &self,
12280            target_session_id: &str,
12281            generation: u64,
12282            policy: bamboo_domain::SessionActivationPolicy,
12283        ) -> Result<(), bamboo_domain::SessionInboxError> {
12284            self.inner
12285                .mark_activation_eligible(target_session_id, generation, policy)
12286                .await
12287        }
12288
12289        async fn claim(
12290            &self,
12291            target_session_id: &str,
12292            limit: usize,
12293        ) -> Result<Vec<SessionInboxClaim>, bamboo_domain::SessionInboxError> {
12294            self.inner.claim(target_session_id, limit).await
12295        }
12296
12297        async fn was_admitted(
12298            &self,
12299            target_session_id: &str,
12300            id: &bamboo_domain::SessionMessageId,
12301        ) -> Result<bool, bamboo_domain::SessionInboxError> {
12302            self.inner.was_admitted(target_session_id, id).await
12303        }
12304
12305        async fn ack(
12306            &self,
12307            target_session_id: &str,
12308            claim: &SessionInboxClaim,
12309        ) -> Result<(), bamboo_domain::SessionInboxError> {
12310            if self
12311                .fail_once
12312                .swap(false, std::sync::atomic::Ordering::SeqCst)
12313            {
12314                return Err(bamboo_domain::SessionInboxError::Storage(
12315                    "injected actor pre-ack failure".to_string(),
12316                ));
12317            }
12318            self.inner.ack(target_session_id, claim).await
12319        }
12320
12321        async fn inspect(
12322            &self,
12323            target_session_id: &str,
12324        ) -> Result<bamboo_domain::SessionInboxBacklog, bamboo_domain::SessionInboxError> {
12325            self.inner.inspect(target_session_id).await
12326        }
12327    }
12328
12329    async fn actor_inbox_fixture(
12330        session_id: &str,
12331    ) -> (
12332        tempfile::TempDir,
12333        Arc<bamboo_storage::SessionStoreV2>,
12334        Arc<bamboo_storage::LockedSessionStore>,
12335        Arc<dyn SessionInboxPort>,
12336        Session,
12337        SessionInboxClaim,
12338    ) {
12339        let temp = tempfile::tempdir().unwrap();
12340        let store = Arc::new(
12341            bamboo_storage::SessionStoreV2::new(temp.path().to_path_buf())
12342                .await
12343                .unwrap(),
12344        );
12345        let storage: Arc<dyn Storage> = store.clone();
12346        let locked = Arc::new(bamboo_storage::LockedSessionStore::new(storage));
12347        let inbox: Arc<dyn SessionInboxPort> = Arc::new(bamboo_storage::FileSessionInbox::new(
12348            store.clone(),
12349            bamboo_domain::SessionInboxLimits::default(),
12350        ));
12351        let session = Session::new(session_id, "model");
12352        store.save_session(&session).await.unwrap();
12353        let mut envelope =
12354            bamboo_domain::SessionMessageEnvelope::user_input(session_id, "actor follow-up");
12355        envelope.id =
12356            bamboo_domain::SessionMessageId::parse(format!("{session_id}-message")).unwrap();
12357        let receipt = inbox.deliver(&envelope).await.unwrap();
12358        inbox
12359            .mark_activation_eligible(
12360                session_id,
12361                receipt.generation,
12362                bamboo_domain::SessionActivationPolicy::InterruptSpecificWait,
12363            )
12364            .await
12365            .unwrap();
12366        let claim = inbox.claim(session_id, 1).await.unwrap().remove(0);
12367        (temp, store, locked, inbox, session, claim)
12368    }
12369
12370    fn actor_binding(
12371        store: Arc<bamboo_storage::SessionStoreV2>,
12372        inbox: Arc<dyn SessionInboxPort>,
12373        persistence: Arc<dyn RuntimeSessionPersistence>,
12374    ) -> SessionInboxRuntimeBinding {
12375        let storage: Arc<dyn Storage> = store;
12376        SessionInboxRuntimeBinding {
12377            router: SessionActivationRouter::new(),
12378            inbox,
12379            storage,
12380            persistence,
12381            parent_question_lock: None,
12382        }
12383    }
12384
12385    #[tokio::test]
12386    async fn remote_claim_seed_rejects_unsupported_inbox_before_dispatch() {
12387        let (_temp, store, locked, inbox, mut session, claim) =
12388            actor_inbox_fixture("unsupported-remote-inbox").await;
12389        let wrapped: Arc<dyn SessionInboxPort> = Arc::new(ActorFailBeforeAckInbox {
12390            inner: inbox.clone(),
12391            fail_once: std::sync::atomic::AtomicBool::new(false),
12392        });
12393        let binding = actor_binding(store.clone(), wrapped, locked);
12394        let fence = ActorActivationFence {
12395            schema_version: 1,
12396            actor_id: session.id.clone(),
12397            activation_id: "unsupported-inbox-activation".into(),
12398            attempt: 1,
12399            run_id: "unsupported-inbox-run".into(),
12400            lease_owner: "unsupported-inbox-host".into(),
12401            lease_epoch: 1,
12402        };
12403        let before = session.clone();
12404
12405        let error = checkpoint_claim_context_before_dispatch(
12406            &binding,
12407            &mut session,
12408            &[claim],
12409            Some((store.as_ref(), &fence)),
12410        )
12411        .await
12412        .unwrap_err();
12413        assert!(error
12414            .to_string()
12415            .contains("requires canonical FileSessionInbox"));
12416        assert_eq!(
12417            serde_json::to_value(&session.messages).unwrap(),
12418            serde_json::to_value(&before.messages).unwrap()
12419        );
12420        assert_eq!(
12421            serde_json::to_value(
12422                &store
12423                    .load_session(&session.id)
12424                    .await
12425                    .unwrap()
12426                    .unwrap()
12427                    .messages
12428            )
12429            .unwrap(),
12430            serde_json::to_value(&before.messages).unwrap()
12431        );
12432        assert_eq!(inbox.inspect(&session.id).await.unwrap().claimed, 1);
12433    }
12434
12435    #[tokio::test]
12436    async fn remote_broker_lineage_rejects_missing_parent_and_project_drift() {
12437        let (_temp, store, locked, inbox, parent, _claim) =
12438            actor_inbox_fixture("remote-parent").await;
12439        let child = Session::new_child_of("remote-child", &parent, "model", "child");
12440        store.save_session(&child).await.unwrap();
12441        let binding = actor_binding(store.clone(), inbox, locked);
12442        let first = remote_canonical_lineage(&binding, &child).await.unwrap();
12443        assert_eq!(first.len(), 2);
12444        let mut changed = child.clone();
12445        changed.set_project_id_meta("different-project");
12446        assert!(remote_canonical_lineage(&binding, &changed).await.is_err());
12447        store.delete_session(&parent.id).await.unwrap();
12448        assert!(remote_canonical_lineage(&binding, &child).await.is_err());
12449    }
12450
12451    #[tokio::test]
12452    async fn spawn_depth_remote_lineage_accepts_existing_tree_above_default() {
12453        let (_temp, store, locked, inbox, root, _claim) =
12454            actor_inbox_fixture("deep-remote-root").await;
12455        let mut parent = root;
12456        for depth in 1..=6 {
12457            let child =
12458                Session::new_child_of(format!("remote-depth-{depth}"), &parent, "model", "child");
12459            store.save_session(&child).await.unwrap();
12460            parent = child;
12461        }
12462        let binding = actor_binding(store, inbox, locked);
12463        let chain = remote_canonical_lineage(&binding, &parent).await.unwrap();
12464        assert_eq!(chain.len(), 7);
12465        assert_eq!(chain.last().unwrap().spawn_depth, 0);
12466    }
12467
12468    #[tokio::test]
12469    async fn child_creation_requires_canonical_host_lifetime_not_cached_candidate() {
12470        let (_temp, store, locked, inbox, parent, _claim) =
12471            actor_inbox_fixture("creation-parent").await;
12472        let mut child = Session::new_child_of("creation-child", &parent, "model", "child");
12473        child.set_project_id_meta("creation-project");
12474        store.save_session(&child).await.unwrap();
12475        let binding = actor_binding(store.clone(), inbox, locked);
12476        let job = SpawnJob {
12477            parent_session_id: parent.id.clone(),
12478            child_session_id: child.id.clone(),
12479            model: "model".into(),
12480            disabled_tools: None,
12481        };
12482        let creation = canonical_child_creation(&binding, &child, &job)
12483            .await
12484            .unwrap();
12485        assert_eq!(creation.created_at, child.created_at);
12486        assert_eq!(creation.spawn_depth, child.spawn_depth);
12487        for damage in 0..7 {
12488            let mut held = child.clone();
12489            match damage {
12490                0 => held.created_at += chrono::Duration::nanoseconds(1),
12491                1 => held.parent_session_id = Some("foreign-parent".into()),
12492                2 => held.root_session_id = "foreign-root".into(),
12493                3 => held.spawn_depth += 1,
12494                4 => held.set_project_id_meta("foreign-project"),
12495                5 => held.kind = bamboo_domain::SessionKind::Root,
12496                _ => held.id = " creation-child ".into(),
12497            }
12498            assert!(
12499                canonical_child_creation(&binding, &held, &job)
12500                    .await
12501                    .is_err(),
12502                "damage {damage}"
12503            );
12504        }
12505        store.delete_session(&child.id).await.unwrap();
12506        assert!(canonical_child_creation(&binding, &child, &job)
12507            .await
12508            .is_err());
12509        let mut replacement = Session::new_child_of(&child.id, &parent, "model", "replacement");
12510        replacement.set_project_id_meta("creation-project");
12511        assert_ne!(replacement.created_at, child.created_at);
12512        store.save_session(&replacement).await.unwrap();
12513        assert!(
12514            canonical_child_creation(&binding, &child, &job)
12515                .await
12516                .is_err(),
12517            "old held spec cannot adopt replacement birth"
12518        );
12519        canonical_child_creation(&binding, &replacement, &job)
12520            .await
12521            .unwrap();
12522    }
12523
12524    #[tokio::test]
12525    async fn actor_immediate_claims_keep_exact_policy_across_mixed_order_ack_and_restart() {
12526        use bamboo_domain::{SessionActivationPolicy, SessionInboxLimits, SessionMessageEnvelope};
12527        for policies in [
12528            [
12529                SessionActivationPolicy::RespectSpecificWait,
12530                SessionActivationPolicy::InterruptSpecificWait,
12531            ],
12532            [
12533                SessionActivationPolicy::InterruptSpecificWait,
12534                SessionActivationPolicy::RespectSpecificWait,
12535            ],
12536        ] {
12537            let temp = tempfile::tempdir().unwrap();
12538            let store = Arc::new(
12539                bamboo_storage::SessionStoreV2::new(temp.path().into())
12540                    .await
12541                    .unwrap(),
12542            );
12543            let session_id = "actor-mixed-intent-policy";
12544            store
12545                .save_session(&Session::new(session_id, "model"))
12546                .await
12547                .unwrap();
12548            let inbox = bamboo_storage::FileSessionInbox::new(store, SessionInboxLimits::default());
12549            let staged = SessionMessageEnvelope::user_input(session_id, "staged sibling");
12550            let staged_receipt = inbox.deliver(&staged).await.unwrap();
12551            let mut sent = Vec::new();
12552            for policy in policies {
12553                let envelope = SessionMessageEnvelope::user_input(session_id, "immediate");
12554                let receipt = inbox
12555                    .deliver_with_activation_intent(&envelope, policy, None)
12556                    .await
12557                    .unwrap();
12558                sent.push((envelope, policy, receipt));
12559            }
12560            let restarted = Arc::new(
12561                bamboo_storage::SessionStoreV2::new(temp.path().into())
12562                    .await
12563                    .unwrap(),
12564            );
12565            let inbox: Arc<dyn SessionInboxPort> = Arc::new(bamboo_storage::FileSessionInbox::new(
12566                restarted.clone(),
12567                SessionInboxLimits::default(),
12568            ));
12569            let binding = actor_binding(
12570                restarted.clone(),
12571                inbox.clone(),
12572                Arc::new(bamboo_storage::LockedSessionStore::new(restarted.clone())),
12573            );
12574            let mut session = restarted.load_session(session_id).await.unwrap().unwrap();
12575            let pairs =
12576                claim_canonical_deliveries(&binding, &mut session, "run-1", usize::MAX, None)
12577                    .await
12578                    .unwrap();
12579            assert_eq!(pairs.len(), 2, "the staged sibling has no permission");
12580            for ((claim, delivery), (envelope, policy, receipt)) in pairs.iter().zip(&sent) {
12581                assert_eq!(claim.envelope.id, envelope.id);
12582                assert_eq!(claim.generation, receipt.generation);
12583                assert_eq!(claim.activation_policy, *policy);
12584                assert_eq!(delivery.activation_policy, *policy);
12585                checkpoint_and_ack_canonical_claim(&binding, &mut session, claim, None)
12586                    .await
12587                    .unwrap();
12588            }
12589            let restarted = Arc::new(
12590                bamboo_storage::SessionStoreV2::new(temp.path().into())
12591                    .await
12592                    .unwrap(),
12593            );
12594            let inbox: Arc<dyn SessionInboxPort> = Arc::new(bamboo_storage::FileSessionInbox::new(
12595                restarted.clone(),
12596                SessionInboxLimits::default(),
12597            ));
12598            let binding = actor_binding(
12599                restarted.clone(),
12600                inbox.clone(),
12601                Arc::new(bamboo_storage::LockedSessionStore::new(restarted.clone())),
12602            );
12603            session = restarted.load_session(session_id).await.unwrap().unwrap();
12604            for (envelope, policy, receipt) in &sent {
12605                assert_eq!(
12606                    inbox
12607                        .deliver_with_activation_intent(envelope, *policy, None)
12608                        .await
12609                        .unwrap(),
12610                    *receipt
12611                );
12612            }
12613            assert!(
12614                claim_canonical_deliveries(&binding, &mut session, "run-2", usize::MAX, None)
12615                    .await
12616                    .unwrap()
12617                    .is_empty()
12618            );
12619            inbox
12620                .mark_activation_eligible(
12621                    session_id,
12622                    staged_receipt.generation,
12623                    SessionActivationPolicy::RespectSpecificWait,
12624                )
12625                .await
12626                .unwrap();
12627            let pairs =
12628                claim_canonical_deliveries(&binding, &mut session, "run-2", usize::MAX, None)
12629                    .await
12630                    .unwrap();
12631            assert_eq!(pairs.len(), 1);
12632            assert_eq!(pairs[0].0.envelope.id, staged.id);
12633            assert_eq!(
12634                pairs[0].1.activation_policy,
12635                SessionActivationPolicy::RespectSpecificWait
12636            );
12637        }
12638    }
12639
12640    #[tokio::test]
12641    async fn actor_mismatched_typed_marker_id_collision_never_acks() {
12642        let (_temp, store, locked, inbox, mut session, claim) =
12643            actor_inbox_fixture("actor-live-id-collision").await;
12644        let mut forged = claim.envelope.to_provider_message().unwrap();
12645        forged.metadata = Some(serde_json::json!({
12646            "session_message": {
12647                "id": claim.envelope.id,
12648                "target_session_id": "different-session"
12649            }
12650        }));
12651        session.add_message(forged);
12652        let persistence: Arc<dyn RuntimeSessionPersistence> = locked;
12653        let binding = actor_binding(store, inbox.clone(), persistence);
12654
12655        assert!(
12656            checkpoint_and_ack_canonical_claim(&binding, &mut session, &claim, None)
12657                .await
12658                .is_err()
12659        );
12660        assert_eq!(inbox.inspect(&session.id).await.unwrap().claimed, 1);
12661        assert!(!inbox
12662            .was_admitted(&session.id, &claim.envelope.id)
12663            .await
12664            .unwrap());
12665    }
12666
12667    #[tokio::test]
12668    async fn actor_concurrent_durable_id_collision_after_claim_never_acks() {
12669        let (_temp, store, locked, inbox, mut session, claim) =
12670            actor_inbox_fixture("actor-durable-id-collision").await;
12671        let mut concurrent = store.load_session(&session.id).await.unwrap().unwrap();
12672        let mut forged = bamboo_agent_core::Message::user("concurrent actor collision");
12673        forged.id = claim.envelope.id.to_string();
12674        concurrent.add_message(forged);
12675        store.save_session(&concurrent).await.unwrap();
12676        let persistence: Arc<dyn RuntimeSessionPersistence> = locked;
12677        let binding = actor_binding(store.clone(), inbox.clone(), persistence);
12678
12679        assert!(
12680            checkpoint_and_ack_canonical_claim(&binding, &mut session, &claim, None)
12681                .await
12682                .is_err()
12683        );
12684        assert_eq!(inbox.inspect(&session.id).await.unwrap().claimed, 1);
12685        assert!(!inbox
12686            .was_admitted(&session.id, &claim.envelope.id)
12687            .await
12688            .unwrap());
12689        let durable = store.load_session(&session.id).await.unwrap().unwrap();
12690        assert!(!durable.messages.iter().any(|message| {
12691            bamboo_domain::is_matching_session_message(message, &claim.envelope)
12692        }));
12693    }
12694
12695    #[tokio::test]
12696    async fn actor_concurrent_durable_typed_body_mismatch_never_acks() {
12697        let (_temp, store, locked, inbox, mut session, claim) =
12698            actor_inbox_fixture("actor-durable-typed-body-collision").await;
12699        let mut different = claim.envelope.clone();
12700        different.body = bamboo_domain::SessionMessageBody::Content(
12701            bamboo_domain::SessionMessageContent::text("forged actor body"),
12702        );
12703        let mut concurrent = store.load_session(&session.id).await.unwrap().unwrap();
12704        concurrent.add_message(different.to_provider_message().unwrap());
12705        store.save_session(&concurrent).await.unwrap();
12706        let persistence: Arc<dyn RuntimeSessionPersistence> = locked;
12707        let binding = actor_binding(store.clone(), inbox.clone(), persistence);
12708
12709        assert!(
12710            checkpoint_and_ack_canonical_claim(&binding, &mut session, &claim, None)
12711                .await
12712                .is_err()
12713        );
12714        assert_eq!(inbox.inspect(&session.id).await.unwrap().claimed, 1);
12715        assert!(!inbox
12716            .was_admitted(&session.id, &claim.envelope.id)
12717            .await
12718            .unwrap());
12719        let durable = store.load_session(&session.id).await.unwrap().unwrap();
12720        assert!(!durable
12721            .messages
12722            .iter()
12723            .any(|message| bamboo_domain::is_matching_session_message(message, &claim.envelope)));
12724    }
12725
12726    #[tokio::test]
12727    async fn actor_checkpoint_failure_rolls_back_and_restart_admits_once() {
12728        let (_temp, store, locked, inbox, mut session, claim) =
12729            actor_inbox_fixture("actor-checkpoint-failure").await;
12730        let envelope_id = claim.envelope.id.clone();
12731        let fault: Arc<dyn RuntimeSessionPersistence> = Arc::new(ActorFaultingPersistence {
12732            inner: locked.clone(),
12733            fail_checkpoint_once: std::sync::atomic::AtomicBool::new(true),
12734        });
12735        let binding = actor_binding(store.clone(), inbox.clone(), fault);
12736
12737        assert!(
12738            checkpoint_and_ack_canonical_claim(&binding, &mut session, &claim, None)
12739                .await
12740                .is_err()
12741        );
12742        assert!(!session
12743            .messages
12744            .iter()
12745            .any(|message| message.id == envelope_id.as_str()));
12746        assert_eq!(inbox.inspect(&session.id).await.unwrap().claimed, 1);
12747        assert!(!inbox.was_admitted(&session.id, &envelope_id).await.unwrap());
12748
12749        let reopened: Arc<dyn SessionInboxPort> = Arc::new(bamboo_storage::FileSessionInbox::new(
12750            store.clone(),
12751            bamboo_domain::SessionInboxLimits::default(),
12752        ));
12753        let recovered = reopened.claim(&session.id, 1).await.unwrap().remove(0);
12754        let persistence: Arc<dyn RuntimeSessionPersistence> = locked;
12755        let binding = actor_binding(store.clone(), reopened.clone(), persistence);
12756        let mut restarted = store.load_session(&session.id).await.unwrap().unwrap();
12757        checkpoint_and_ack_canonical_claim(&binding, &mut restarted, &recovered, None)
12758            .await
12759            .unwrap();
12760        assert_eq!(
12761            restarted
12762                .messages
12763                .iter()
12764                .filter(|message| message.id == envelope_id.as_str())
12765                .count(),
12766            1
12767        );
12768        assert!(reopened
12769            .was_admitted(&session.id, &envelope_id)
12770            .await
12771            .unwrap());
12772        let backlog = reopened.inspect(&session.id).await.unwrap();
12773        assert_eq!(backlog.pending + backlog.claimed, 0);
12774    }
12775
12776    #[tokio::test]
12777    async fn actor_checkpoint_success_pre_ack_failure_recovers_without_duplicate() {
12778        let (_temp, store, locked, real_inbox, mut session, claim) =
12779            actor_inbox_fixture("actor-pre-ack-failure").await;
12780        let envelope_id = claim.envelope.id.clone();
12781        let faulted: Arc<dyn SessionInboxPort> = Arc::new(ActorFailBeforeAckInbox {
12782            inner: real_inbox.clone(),
12783            fail_once: std::sync::atomic::AtomicBool::new(true),
12784        });
12785        let persistence: Arc<dyn RuntimeSessionPersistence> = locked.clone();
12786        let binding = actor_binding(store.clone(), faulted, persistence);
12787
12788        assert!(
12789            checkpoint_and_ack_canonical_claim(&binding, &mut session, &claim, None)
12790                .await
12791                .is_err()
12792        );
12793        let durable = store.load_session(&session.id).await.unwrap().unwrap();
12794        assert_eq!(
12795            durable
12796                .messages
12797                .iter()
12798                .filter(|message| message.id == envelope_id.as_str())
12799                .count(),
12800            1
12801        );
12802        assert_eq!(real_inbox.inspect(&session.id).await.unwrap().claimed, 1);
12803        assert!(!real_inbox
12804            .was_admitted(&session.id, &envelope_id)
12805            .await
12806            .unwrap());
12807
12808        let reopened: Arc<dyn SessionInboxPort> = Arc::new(bamboo_storage::FileSessionInbox::new(
12809            store.clone(),
12810            bamboo_domain::SessionInboxLimits::default(),
12811        ));
12812        let recovered = reopened.claim(&session.id, 1).await.unwrap().remove(0);
12813        let persistence: Arc<dyn RuntimeSessionPersistence> = locked;
12814        let binding = actor_binding(store.clone(), reopened.clone(), persistence);
12815        let mut restarted = durable;
12816        checkpoint_and_ack_canonical_claim(&binding, &mut restarted, &recovered, None)
12817            .await
12818            .unwrap();
12819        assert_eq!(
12820            restarted
12821                .messages
12822                .iter()
12823                .filter(|message| message.id == envelope_id.as_str())
12824                .count(),
12825            1
12826        );
12827        assert!(reopened
12828            .was_admitted(&session.id, &envelope_id)
12829            .await
12830            .unwrap());
12831        let backlog = reopened.inspect(&session.id).await.unwrap();
12832        assert_eq!(backlog.pending + backlog.claimed, 0);
12833    }
12834
12835    struct ConfirmationSequenceLink {
12836        frames: VecDeque<ChildFrame>,
12837        sent: Vec<ParentFrame>,
12838    }
12839
12840    #[async_trait]
12841    impl bamboo_subagent::ChildLink for ConfirmationSequenceLink {
12842        async fn send(&mut self, frame: ParentFrame) -> bamboo_subagent::TransportResult<()> {
12843            self.sent.push(frame);
12844            Ok(())
12845        }
12846
12847        async fn next_frame(&mut self) -> bamboo_subagent::TransportResult<Option<ChildFrame>> {
12848            match self.frames.pop_front() {
12849                Some(frame) => Ok(Some(frame)),
12850                None => std::future::pending().await,
12851            }
12852        }
12853    }
12854
12855    #[tokio::test]
12856    async fn failed_actor_retry_checkpoints_new_input_and_rejects_old_or_stale_state() {
12857        for case in [
12858            "accepted",
12859            "readonly-refusal",
12860            "prefix-refusal",
12861            "old-claim",
12862            "old-checkpoint",
12863            "succeeded",
12864            "changed-birth",
12865            "oversize",
12866        ] {
12867            let temp = tempfile::tempdir().unwrap();
12868            let store = Arc::new(
12869                bamboo_storage::SessionStoreV2::new(temp.path().into())
12870                    .await
12871                    .unwrap(),
12872            );
12873            let parent = Session::new("retry-parent", "model");
12874            store.save_session(&parent).await.unwrap();
12875            let mut child =
12876                Session::new_child_of("retry-child", &parent, "model", "original assignment");
12877            child.add_message(bamboo_agent_core::Message::user("original assignment"));
12878            store.save_session(&child).await.unwrap();
12879            let policy = Arc::new(bamboo_tools::permission::PermissionConfig::new());
12880            bind_local_control_plane(store.as_ref(), &child.id, policy.as_ref()).await;
12881            let locked = Arc::new(bamboo_storage::LockedSessionStore::new(store.clone()));
12882            let inbox: Arc<dyn SessionInboxPort> = Arc::new(bamboo_storage::FileSessionInbox::new(
12883                store.clone(),
12884                bamboo_domain::SessionInboxLimits::default(),
12885            ));
12886            let binding = actor_binding(store.clone(), inbox, locked);
12887            let registration = binding
12888                .router
12889                .register_run(&child.id, "first-run")
12890                .await
12891                .unwrap();
12892            let first = PlainActorActivation::start(
12893                store.clone(),
12894                &child,
12895                &binding,
12896                Some("first-run"),
12897                Some(policy.clone()),
12898                true,
12899                None,
12900            )
12901            .await
12902            .unwrap();
12903            if matches!(case, "old-claim" | "old-checkpoint") {
12904                let old = bamboo_domain::SessionMessageEnvelope::user_input(
12905                    &child.id,
12906                    "unconfirmed old input",
12907                );
12908                let receipt = first.input_inbox.deliver(&old).await.unwrap();
12909                first
12910                    .input_inbox
12911                    .mark_activation_eligible(
12912                        &child.id,
12913                        receipt.generation,
12914                        bamboo_domain::SessionActivationPolicy::InterruptSpecificWait,
12915                    )
12916                    .await
12917                    .unwrap();
12918                let claim = first
12919                    .claim_input(&binding, &child, "first-run")
12920                    .await
12921                    .unwrap()
12922                    .unwrap();
12923                if case == "old-checkpoint" {
12924                    first
12925                        .checkpoint_input(&mut child, "first-run", &claim)
12926                        .await
12927                        .unwrap();
12928                }
12929            }
12930            first
12931                .finish(if case == "succeeded" {
12932                    ActorActivationFinish::Succeeded
12933                } else {
12934                    ActorActivationFinish::Failed
12935                })
12936                .await
12937                .unwrap();
12938            // Drop schedules cleanup; the retry needs the old owner released first.
12939            registration.abandon().await;
12940            assert!(
12941                !binding.router.owns_run(&child.id, "first-run").await,
12942                "{case}: abandoned first run must release its owner before retry"
12943            );
12944            let envelope = bamboo_domain::SessionMessageEnvelope::user_input(
12945                &child.id,
12946                if case == "oversize" {
12947                    "x".repeat(8193)
12948                } else {
12949                    "NEW_CORRECTION".into()
12950                },
12951            );
12952            let receipt = first.input_inbox.deliver(&envelope).await.unwrap();
12953            first
12954                .input_inbox
12955                .mark_activation_eligible(
12956                    &child.id,
12957                    receipt.generation,
12958                    bamboo_domain::SessionActivationPolicy::InterruptSpecificWait,
12959                )
12960                .await
12961                .unwrap();
12962            let before = store.load_session(&child.id).await.unwrap().unwrap();
12963            let expected_messages = serde_json::to_value(&before.messages).unwrap();
12964            child = before.clone();
12965            if case == "prefix-refusal" {
12966                child.messages.last_mut().unwrap().content = "stale cached prefix".into();
12967            }
12968            if case == "changed-birth" {
12969                child.created_at += chrono::Duration::milliseconds(1);
12970            }
12971            let _retry_registration = binding
12972                .router
12973                .register_run(&child.id, "retry-run")
12974                .await
12975                .unwrap();
12976            let result = PlainActorActivation::start_and_prepare(
12977                store.clone(),
12978                &mut child,
12979                &binding,
12980                Some("retry-run"),
12981                Some(policy.clone()),
12982                case != "readonly-refusal",
12983                None,
12984            )
12985            .await;
12986            if case != "accepted" {
12987                assert!(result.is_err(), "{case}: no retry Run may be dispatched");
12988                let actual = store.load_session(&before.id).await.unwrap().unwrap();
12989                assert_eq!(
12990                    serde_json::to_value(&actual.messages).unwrap(),
12991                    expected_messages,
12992                    "{case}: failed preparation preserves Main history"
12993                );
12994                assert!(
12995                    !first
12996                        .input_inbox
12997                        .was_admitted(&before.id, &envelope.id)
12998                        .await
12999                        .unwrap(),
13000                    "{case}: no ACK"
13001                );
13002                if case == "readonly-refusal" {
13003                    let entry = store.inspect_actor(&before.id).await.unwrap();
13004                    assert_eq!(entry.actor.current_attempt, 1);
13005                    assert_eq!(
13006                        entry.activation.unwrap().activation_id,
13007                        first.fence.activation_id
13008                    );
13009                    let backlog = first.input_inbox.inspect(&before.id).await.unwrap();
13010                    assert_eq!((backlog.pending, backlog.claimed), (1, 0));
13011                }
13012                if matches!(case, "prefix-refusal" | "changed-birth" | "oversize") {
13013                    assert_eq!(
13014                        first.input_inbox.inspect(&before.id).await.unwrap().claimed,
13015                        1
13016                    );
13017                    assert_eq!(
13018                        store
13019                            .inspect_actor(&before.id)
13020                            .await
13021                            .unwrap()
13022                            .activation
13023                            .unwrap()
13024                            .status,
13025                        bamboo_domain::ActorActivationStatus::Failed
13026                    );
13027                }
13028                if case == "oversize" {
13029                    let reopened = Arc::new(
13030                        bamboo_storage::SessionStoreV2::new(temp.path().into())
13031                            .await
13032                            .unwrap(),
13033                    );
13034                    let cold_inbox = bamboo_storage::FileSessionInbox::new(
13035                        reopened,
13036                        bamboo_domain::SessionInboxLimits::default(),
13037                    );
13038                    let now = chrono::Utc::now();
13039                    let leases = cold_inbox
13040                        .inspect_owned_leases(&before.id, 2, now)
13041                        .await
13042                        .unwrap();
13043                    assert_eq!(leases.len(), 1);
13044                    assert_eq!(leases[0].failure_count, 1);
13045                    assert_eq!(
13046                        leases[0].last_error_code.as_deref(),
13047                        Some("unsupported_actor_input")
13048                    );
13049                    let retry_after = leases[0].retry_after.unwrap();
13050                    assert!(retry_after > now);
13051                    let readiness = cold_inbox
13052                        .inspect_wake_readiness(&before.id, now)
13053                        .await
13054                        .unwrap();
13055                    assert!(
13056                        readiness.ready.is_none(),
13057                        "failed correction must not hot-loop"
13058                    );
13059                    assert_eq!(readiness.next_due_at, Some(retry_after));
13060                }
13061                continue;
13062            }
13063            let (retry, prepared) = result.unwrap();
13064            let (prefix, mut deliveries) = prepared.unwrap();
13065            assert_eq!(deliveries.len(), 1);
13066            let delivery = deliveries.pop().unwrap();
13067            assert_eq!(
13068                prefix,
13069                serde_json::to_value(&before.messages)
13070                    .unwrap()
13071                    .as_array()
13072                    .unwrap()
13073                    .clone()
13074            );
13075            assert_eq!(delivery.envelope, envelope);
13076            assert_eq!(delivery.activation_run_id, "retry-run");
13077            assert_eq!(delivery.canonical_claim_generation, receipt.generation);
13078            assert!(store
13079                .validate_fence(&first.fence, chrono::Utc::now())
13080                .await
13081                .is_err());
13082            let running = store.inspect_actor(&child.id).await.unwrap();
13083            assert_eq!(running.actor.current_attempt, 2);
13084            assert_eq!(
13085                running.activation.as_ref().unwrap().status,
13086                bamboo_domain::ActorActivationStatus::Running
13087            );
13088            assert!(running.activation.unwrap().lease_epoch > first.fence.lease_epoch);
13089            let claim = retry.initial_inputs.first().unwrap();
13090            assert!(child
13091                .session_inbox_admission()
13092                .unwrap()
13093                .contains(&envelope.id));
13094            assert!(
13095                !retry
13096                    .input_inbox
13097                    .was_admitted(&child.id, &envelope.id)
13098                    .await
13099                    .unwrap(),
13100                "checkpoint is not startup confirmation"
13101            );
13102            let mut confirmation = admission_confirmation(&child.id, &claim.claim, "retry-run");
13103            confirmation.activation_run_id = "first-run".into();
13104            assert!(retry
13105                .confirm_input(&binding, &child, "retry-run", claim, &confirmation)
13106                .await
13107                .is_err());
13108            assert!(!retry
13109                .input_inbox
13110                .was_admitted(&child.id, &envelope.id)
13111                .await
13112                .unwrap());
13113            confirmation.activation_run_id = "retry-run".into();
13114            retry
13115                .confirm_input(&binding, &child, "retry-run", claim, &confirmation)
13116                .await
13117                .unwrap();
13118            let (tx, mut rx) = mpsc::channel(1);
13119            retry
13120                .append_reply(&mut child, "RETRY_REPLY".into(), &tx)
13121                .await
13122                .unwrap();
13123            rx.recv().await.unwrap();
13124            retry
13125                .finish(ActorActivationFinish::Succeeded)
13126                .await
13127                .unwrap();
13128            let reopened = Arc::new(
13129                bamboo_storage::SessionStoreV2::new(temp.path().into())
13130                    .await
13131                    .unwrap(),
13132            );
13133            let cold = reopened.load_session(&child.id).await.unwrap().unwrap();
13134            assert_eq!(cold.created_at, before.created_at);
13135            assert_eq!(cold.messages.len(), before.messages.len() + 2);
13136            assert_eq!(
13137                cold.messages[before.messages.len()].id,
13138                envelope.id.as_str()
13139            );
13140            assert_eq!(cold.messages.last().unwrap().content, "RETRY_REPLY");
13141            assert_eq!(
13142                cold.messages
13143                    .iter()
13144                    .filter(|m| m.id == envelope.id.as_str())
13145                    .count(),
13146                1
13147            );
13148            let cold_inbox = bamboo_storage::FileSessionInbox::new(
13149                reopened,
13150                bamboo_domain::SessionInboxLimits::default(),
13151            );
13152            assert!(cold_inbox
13153                .was_admitted(&cold.id, &envelope.id)
13154                .await
13155                .unwrap());
13156            assert_eq!(cold_inbox.inspect(&cold.id).await.unwrap().claimed, 0);
13157        }
13158    }
13159
13160    #[tokio::test]
13161    async fn pre_ack_owned_recovery_requires_expiry_provenance_and_verified_already() {
13162        for case in [
13163            "accepted",
13164            "actor-live",
13165            "claim-live",
13166            "legacy",
13167            "ack",
13168            "readonly",
13169            "prefix",
13170            "duplicate",
13171            "missing-marker",
13172            "mixed",
13173            "project",
13174            "birth",
13175        ] {
13176            let temp = tempfile::tempdir().unwrap();
13177            let store = Arc::new(
13178                bamboo_storage::SessionStoreV2::new(temp.path().into())
13179                    .await
13180                    .unwrap(),
13181            );
13182            let parent = Session::new("recovery-parent", "model");
13183            store.save_session(&parent).await.unwrap();
13184            let mut child =
13185                Session::new_child_of("recovery-child", &parent, "model", "original task");
13186            child.add_message(bamboo_agent_core::Message::user("original task"));
13187            store.save_session(&child).await.unwrap();
13188            let policy = Arc::new(bamboo_tools::permission::PermissionConfig::new());
13189            bind_local_control_plane(store.as_ref(), &child.id, policy.as_ref()).await;
13190            let inbox = bamboo_storage::FileSessionInbox::new(
13191                store.clone(),
13192                bamboo_domain::SessionInboxLimits::default(),
13193            );
13194            // Complete unfenced persistence before starting the setup lease window.
13195            let envelope =
13196                bamboo_domain::SessionMessageEnvelope::user_input(&child.id, "RECOVER_SAME_INPUT");
13197            let receipt = inbox.deliver(&envelope).await.unwrap();
13198            inbox
13199                .mark_activation_eligible(
13200                    &child.id,
13201                    receipt.generation,
13202                    bamboo_domain::SessionActivationPolicy::InterruptSpecificWait,
13203                )
13204                .await
13205                .unwrap();
13206            let now = chrono::Utc::now();
13207            let cutoff = now + chrono::Duration::seconds(10);
13208            let consumer = SessionInboxConsumerId::new();
13209            let old = store
13210                .claim_activation(&ActorActivationClaim {
13211                    actor_id: child.id.clone(),
13212                    run_id: "old-run".into(),
13213                    lease_owner: consumer.as_str().into(),
13214                    lease_expires_at: if case == "actor-live" {
13215                        now + chrono::Duration::minutes(5)
13216                    } else {
13217                        cutoff
13218                    },
13219                    inbox_generation: 0,
13220                    placement_ref: if case == "legacy" {
13221                        None
13222                    } else {
13223                        Some(bamboo_domain::ActorPlacementRef {
13224                            class: bamboo_domain::ActorPlacementClass::Local,
13225                            lease_id: "owned-initial-release-v1:actual-old-worker".into(),
13226                            slot_epoch: None,
13227                        })
13228                    },
13229                    now,
13230                })
13231                .await
13232                .unwrap();
13233            store
13234                .start_activation(&old.fence(), chrono::Utc::now())
13235                .await
13236                .unwrap();
13237            let claim_now = chrono::Utc::now();
13238            let claim = inbox
13239                .claim_owned(
13240                    &child.id,
13241                    1,
13242                    Some("old-run"),
13243                    &SessionInboxLeaseRequest {
13244                        consumer,
13245                        now: claim_now,
13246                        duration: if case == "claim-live" {
13247                            chrono::Duration::minutes(5)
13248                        } else {
13249                            cutoff - claim_now
13250                        },
13251                    },
13252                )
13253                .await
13254                .unwrap()
13255                .pop()
13256                .unwrap();
13257            child = inbox
13258                .checkpoint_actor_input(bamboo_storage::ActorInputCheckpoint {
13259                    fence: old.fence(),
13260                    expected_created_at: child.created_at,
13261                    expected_conversation_generation: child.conversation_generation_observation(),
13262                    claim: claim.clone(),
13263                    expected_messages: child.messages.clone(),
13264                    expected_provider_transcript: child.provider_transcript.clone(),
13265                    expected_admission: None,
13266                })
13267                .await
13268                .unwrap()
13269                .session;
13270            // The real checkpoint must succeed under the original live authority.
13271            store
13272                .validate_fence(&old.fence(), chrono::Utc::now())
13273                .await
13274                .unwrap_or_else(|error| panic!("{case}: setup fence: {error:?}"));
13275            let setup_leases = inbox
13276                .inspect_owned_leases(&child.id, 2, chrono::Utc::now())
13277                .await
13278                .unwrap();
13279            assert_eq!(setup_leases.len(), 1, "{case}: original setup claim");
13280            assert_eq!(setup_leases[0].epoch, claim.lease.epoch, "{case}");
13281            assert_eq!(setup_leases[0].expires_at, claim.lease.expires_at, "{case}");
13282            assert!(!setup_leases[0].expired, "{case}: setup claim must be live");
13283            let setup_at = chrono::Utc::now();
13284            assert!(
13285                setup_at < old.lease_expires_at,
13286                "{case}: setup Actor expired"
13287            );
13288            assert!(
13289                setup_at < claim.lease.expires_at,
13290                "{case}: setup claim expired"
13291            );
13292            eprintln!("pre-ACK {case}: setup live at {setup_at}, cutoff {cutoff}");
13293            if case == "ack" {
13294                inbox
13295                    .ack_owned(&child.id, &claim, chrono::Utc::now())
13296                    .await
13297                    .unwrap();
13298            }
13299            if case == "mixed" {
13300                inbox
13301                    .deliver(&bamboo_domain::SessionMessageEnvelope::user_input(
13302                        &child.id,
13303                        "different input",
13304                    ))
13305                    .await
13306                    .unwrap();
13307            }
13308            let directory = store
13309                .bamboo_home_dir()
13310                .join(store.resolve_rel_path(&child.id).await.unwrap());
13311            let main = std::fs::read(directory.join("session.json")).unwrap();
13312            let canonical = child.clone();
13313            match case {
13314                "prefix" => child.messages[0].content.push_str(" stale"),
13315                "duplicate" => child.messages.push(child.messages.last().unwrap().clone()),
13316                "missing-marker" => {
13317                    child
13318                        .messages
13319                        .last_mut()
13320                        .unwrap()
13321                        .metadata
13322                        .as_mut()
13323                        .unwrap()
13324                        .as_object_mut()
13325                        .unwrap()
13326                        .remove("_bamboo_owned_input_checkpoint");
13327                }
13328                "project" => child.set_project_id_meta(bamboo_domain::ProjectId::new().to_string()),
13329                "birth" => child.created_at += chrono::Duration::nanoseconds(1),
13330                _ => {}
13331            }
13332            // One shared short deadline per case; never wait for the live controls.
13333            // No reclaim, renewal or authority-file edit occurs between these stages.
13334            let delay = (cutoff - chrono::Utc::now()).to_std().unwrap_or_default();
13335            tokio::time::sleep(delay + Duration::from_millis(10)).await;
13336            let recovery_at = chrono::Utc::now();
13337            assert!(recovery_at >= cutoff, "{case}: real expiry is required");
13338            let before_recovery = store.inspect_actor(&child.id).await.unwrap();
13339            let original = before_recovery.activation.unwrap();
13340            assert_eq!(original.fence(), old.fence(), "{case}: same authority");
13341            assert_eq!(original.lease_expires_at, old.lease_expires_at, "{case}");
13342            assert_eq!(
13343                store
13344                    .validate_fence(&old.fence(), recovery_at)
13345                    .await
13346                    .is_ok(),
13347                case == "actor-live",
13348                "{case}: Actor expiry control"
13349            );
13350            let recovery_leases = inbox
13351                .inspect_owned_leases(&child.id, 2, recovery_at)
13352                .await
13353                .unwrap();
13354            if case == "ack" {
13355                assert!(recovery_leases.is_empty(), "{case}: terminal ACK");
13356            } else {
13357                assert_eq!(recovery_leases.len(), 1, "{case}: original recovery claim");
13358                assert_eq!(recovery_leases[0].epoch, claim.lease.epoch, "{case}");
13359                assert_eq!(
13360                    recovery_leases[0].expires_at, claim.lease.expires_at,
13361                    "{case}"
13362                );
13363                assert_eq!(recovery_leases[0].expired, case != "claim-live", "{case}");
13364            }
13365            eprintln!("pre-ACK {case}: recovery expiry control at {recovery_at}");
13366            let second = Arc::new(
13367                bamboo_storage::SessionStoreV2::new(temp.path().into())
13368                    .await
13369                    .unwrap(),
13370            );
13371            let binding = actor_binding(
13372                second.clone(),
13373                Arc::new(bamboo_storage::FileSessionInbox::new(
13374                    second.clone(),
13375                    bamboo_domain::SessionInboxLimits::default(),
13376                )),
13377                Arc::new(bamboo_storage::LockedSessionStore::new(second.clone())),
13378            );
13379            let _owner = binding
13380                .router
13381                .register_run(&child.id, "replacement-run")
13382                .await
13383                .unwrap();
13384            let prepared = PlainActorActivation::start_and_prepare(
13385                second.clone(),
13386                &mut child,
13387                &binding,
13388                Some("replacement-run"),
13389                Some(policy.clone()),
13390                case != "readonly",
13391                Some("probed-new-worker"),
13392            )
13393            .await;
13394            assert_eq!(
13395                std::fs::read(directory.join("session.json")).unwrap(),
13396                main,
13397                "{case}: Already or refusal never rewrites Main"
13398            );
13399            if case != "accepted" {
13400                assert!(prepared.is_err(), "{case}");
13401                assert_eq!(
13402                    inbox.was_admitted(&child.id, &envelope.id).await.unwrap(),
13403                    case == "ack"
13404                );
13405                continue;
13406            }
13407            let (replacement, delivery) = prepared.unwrap();
13408            assert!(replacement.recovering_pre_ack);
13409            let (messages, mut deliveries) = delivery.unwrap();
13410            assert_eq!(deliveries.len(), 1);
13411            let delivery = deliveries.pop().unwrap();
13412            let new_claim = replacement.initial_inputs.first().unwrap();
13413            assert_eq!(new_claim.claim.envelope, claim.claim.envelope);
13414            assert_eq!(new_claim.claim.generation, claim.claim.generation);
13415            assert_eq!(
13416                new_claim.claim.activation_policy,
13417                claim.claim.activation_policy
13418            );
13419            assert!(new_claim.lease.epoch > claim.lease.epoch);
13420            assert_ne!(new_claim.lease.incarnation, claim.lease.incarnation);
13421            assert_eq!(
13422                replacement.inbox_generation, 1,
13423                "old Actor generation was zero"
13424            );
13425            assert_eq!(
13426                second
13427                    .inspect_actor(&child.id)
13428                    .await
13429                    .unwrap()
13430                    .actor
13431                    .current_attempt,
13432                2
13433            );
13434            assert!(second
13435                .validate_fence(&old.fence(), chrono::Utc::now())
13436                .await
13437                .is_err());
13438            assert!(inbox
13439                .ack_owned(&child.id, &claim, chrono::Utc::now())
13440                .await
13441                .is_err());
13442            // The original Store/claim survives the handoff. Even with the
13443            // current exact prefix and dedupe cursor, its delayed checkpoint
13444            // must fail on authority before AlreadyCheckpointed can succeed.
13445            let authority = second.inspect_actor(&child.id).await.unwrap();
13446            let leases = inbox
13447                .inspect_owned_leases(&child.id, 2, chrono::Utc::now())
13448                .await
13449                .unwrap();
13450            let delayed = inbox
13451                .checkpoint_actor_input(bamboo_storage::ActorInputCheckpoint {
13452                    fence: old.fence(),
13453                    expected_created_at: child.created_at,
13454                    expected_conversation_generation: child.conversation_generation_observation(),
13455                    claim: claim.clone(),
13456                    expected_messages: child.messages.clone(),
13457                    expected_provider_transcript: child.provider_transcript.clone(),
13458                    expected_admission: child.session_inbox_admission().cloned(),
13459                })
13460                .await;
13461            assert!(matches!(
13462                delayed,
13463                Err(bamboo_storage::ActorInputCheckpointError::Actor(
13464                    bamboo_domain::ActorDirectoryError::StaleFence
13465                ))
13466            ));
13467            assert_eq!(std::fs::read(directory.join("session.json")).unwrap(), main);
13468            assert_eq!(second.inspect_actor(&child.id).await.unwrap(), authority);
13469            assert_eq!(
13470                inbox
13471                    .inspect_owned_leases(&child.id, 2, chrono::Utc::now())
13472                    .await
13473                    .unwrap(),
13474                leases
13475            );
13476            assert!(!inbox.was_admitted(&child.id, &envelope.id).await.unwrap());
13477            assert_eq!(
13478                serde_json::to_value(&child.messages).unwrap(),
13479                serde_json::to_value(&canonical.messages).unwrap()
13480            );
13481            let mut worker: Vec<bamboo_agent_core::Message> = messages
13482                .iter()
13483                .cloned()
13484                .map(|v| serde_json::from_value(v).unwrap())
13485                .collect();
13486            let target = worker.last_mut().unwrap();
13487            assert!(target
13488                .metadata
13489                .as_ref()
13490                .unwrap()
13491                .get("_bamboo_owned_input_checkpoint")
13492                .is_none());
13493            assert!(bamboo_domain::is_matching_session_message(
13494                target, &envelope
13495            ));
13496            let mut wanted = canonical.messages.last().unwrap().clone();
13497            wanted
13498                .metadata
13499                .as_mut()
13500                .unwrap()
13501                .as_object_mut()
13502                .unwrap()
13503                .remove("_bamboo_owned_input_checkpoint");
13504            assert_eq!(
13505                serde_json::to_value(target).unwrap(),
13506                serde_json::to_value(wanted).unwrap()
13507            );
13508            let run = RunSpec {
13509                assignment: "original task".into(),
13510                logical_session: Some(LogicalSessionIdentity {
13511                    session_id: child.id.clone(),
13512                    parent_session_id: child.parent_session_id.clone(),
13513                    root_session_id: child.root_session_id.clone(),
13514                    creation: Some(bamboo_subagent::proto::ChildCreationIdentity {
13515                        created_at: child.created_at,
13516                        spawn_depth: child.spawn_depth,
13517                    }),
13518                }),
13519                project_id: None,
13520                reasoning_effort: None,
13521                permission_policy: Some(PermissionPolicyContext {
13522                    revision: policy.policy_revision(),
13523                    requested_mode: "default".into(),
13524                    effective_mode: "default".into(),
13525                    bypass_permissions: false,
13526                    auto_approve_permissions: false,
13527                    session_id: child.id.clone(),
13528                    workspace_path: child.workspace.clone(),
13529                    environment_lease: None,
13530                    inherit_session_grants: false,
13531                    policy: serde_json::to_value(policy.to_serializable()).unwrap(),
13532                }),
13533                messages,
13534                activation_run_id: Some("replacement-run".into()),
13535                execution_epoch: 2,
13536                initial_session_messages: vec![delivery.clone()],
13537                secrets: Default::default(),
13538            };
13539            let policy_revision = policy.policy_revision();
13540            let mut expected = Some(expected_default_permission_posture(policy_revision));
13541            let ChildFrame::Event { event } = permission_posture_frame(&child.id, policy_revision)
13542            else {
13543                unreachable!()
13544            };
13545            let (events, mut rx) = mpsc::channel(1);
13546            process_actor_event(
13547                event,
13548                true,
13549                &mut PermissionPostureHandshake::Awaiting,
13550                expected.as_ref(),
13551                Some(&binding),
13552                &mut child,
13553                &events,
13554                &mut ActorEventDisplay::default(),
13555            )
13556            .await
13557            .unwrap();
13558            rx.recv().await.unwrap();
13559            let request = bamboo_subagent::proto::InitialInputReleaseRequest::from_run(
13560                &run,
13561                &delivery,
13562                uuid::Uuid::new_v4().to_string(),
13563            )
13564            .unwrap();
13565            replacement
13566                .release_initial_input(
13567                    &binding,
13568                    &child,
13569                    &run,
13570                    2,
13571                    new_claim,
13572                    &request,
13573                    &mut expected,
13574                    None,
13575                )
13576                .await
13577                .unwrap();
13578            assert!(inbox.was_admitted(&child.id, &envelope.id).await.unwrap());
13579            replacement
13580                .append_reply(&mut child, "RECOVERY_REPLY".into(), &events)
13581                .await
13582                .unwrap();
13583            rx.recv().await.unwrap();
13584            replacement
13585                .finish(ActorActivationFinish::Succeeded)
13586                .await
13587                .unwrap();
13588            let cold = second.load_session(&child.id).await.unwrap().unwrap();
13589            assert_eq!(cold.created_at, canonical.created_at);
13590            assert_eq!(
13591                serde_json::to_value(&cold.messages[..canonical.messages.len()]).unwrap(),
13592                serde_json::to_value(&canonical.messages).unwrap()
13593            );
13594            assert_eq!(cold.messages.last().unwrap().content, "RECOVERY_REPLY");
13595            assert_eq!(
13596                cold.messages
13597                    .iter()
13598                    .filter(|m| m.id == envelope.id.as_str())
13599                    .count(),
13600                1
13601            );
13602        }
13603    }
13604
13605    #[cfg(unix)]
13606    #[tokio::test]
13607    async fn owned_initial_release_missing_only_new_capability_rejects_before_spawn_or_run() {
13608        use bamboo_subagent::provision::*;
13609        use std::os::unix::fs::PermissionsExt;
13610        let temp = tempfile::tempdir().unwrap();
13611        let worker = temp.path().join("mixed-worker");
13612        let spawned = temp.path().join("spawned");
13613        let run = temp.path().join("run");
13614        let probes = temp.path().join("probes");
13615        let fabric = temp.path().join("fabric");
13616        let mut report = WorkerCapabilityReport::current();
13617        report
13618            .capabilities
13619            .retain(|name| name != INITIAL_INPUT_RELEASE_WORKER_CAPABILITY);
13620        assert!(!report.supports(INITIAL_INPUT_RELEASE_WORKER_CAPABILITY));
13621        let report = serde_json::to_string(&report).unwrap();
13622        std::fs::write(&worker, format!(
13623            "#!/bin/sh\nfor arg in \"$@\"; do last=\"$arg\"; done\nif [ \"$last\" = --print-capabilities ]; then\nprintf '%s\\n' '{report}'\nprintf 'probe\\n' >> \"$3\"\nelse\n: > \"$1\"\n: > \"$2\"\ncat >/dev/null\nfi\n"
13624        )).unwrap();
13625        std::fs::set_permissions(&worker, std::fs::Permissions::from_mode(0o700)).unwrap();
13626        let args = vec![
13627            spawned.to_str().unwrap().into(),
13628            run.to_str().unwrap().into(),
13629            probes.to_str().unwrap().into(),
13630        ];
13631        // Real subprocess probes prove the rejection cannot be an earlier missing bit.
13632        for capability in [
13633            REQUIRED_CHILD_CONTEXT_WORKER_CAPABILITY,
13634            CHILD_CREATION_IDENTITY_WORKER_CAPABILITY,
13635            NATIVE_TOOL_CEILING_WORKER_CAPABILITY,
13636            TYPED_READ_ONLY_WORKER_CAPABILITY,
13637        ] {
13638            bamboo_subagent::fleet::require_worker_capability(&worker, &args, capability)
13639                .await
13640                .unwrap();
13641        }
13642        let runner = ActorChildRunner::new(
13643            "mixed".into(),
13644            worker,
13645            args,
13646            fabric.clone(),
13647            ExecutorSpec::BambooRuntime,
13648            vec![],
13649            "openai".into(),
13650            1,
13651        );
13652        // The actual Host method remains mandatory before claim/provision/Run.
13653        let error = runner
13654            .require_initial_input_release_worker()
13655            .await
13656            .unwrap_err();
13657        assert!(error
13658            .to_string()
13659            .contains("owned initial release worker capability unconfirmed"));
13660        assert_eq!(std::fs::read_to_string(probes).unwrap().lines().count(), 5);
13661        assert!(!spawned.exists() && !run.exists() && !fabric.exists());
13662        assert!(runner.pool.lock().await.is_empty());
13663    }
13664
13665    #[tokio::test]
13666    async fn ordinary_glob_only_observes_own_initialization_signals() {
13667        let temp = tempfile::tempdir().unwrap();
13668        let home = temp.path().canonicalize().unwrap();
13669        let store = bamboo_storage::SessionStoreV2::new(home.clone())
13670            .await
13671            .unwrap();
13672        let parent = Session::new("glob-standard-parent", "model");
13673        store.save_session(&parent).await.unwrap();
13674        let child = Session::new_child_of("glob-standard-child", &parent, "model", "one Glob");
13675        store.save_session(&child).await.unwrap();
13676        let directory = home.join(store.resolve_rel_path(&child.id).await.unwrap());
13677        // An unrelated bad child would reject a full tree snapshot, but is not
13678        // part of this own-session no-replay observation.
13679        let unrelated = directory.parent().unwrap().join("unrelated");
13680        std::fs::create_dir(&unrelated).unwrap();
13681        std::fs::write(unrelated.join("session.json"), b"invalid").unwrap();
13682        require_unowned_glob(&store, &child).await.unwrap();
13683        for name in ["actor-authority.json", "actor-authority.initialized.json"] {
13684            let path = directory.join(name);
13685            std::fs::write(&path, b"invalid prior signal").unwrap();
13686            assert!(require_unowned_glob(&store, &child).await.is_err());
13687            std::fs::remove_file(&path).unwrap();
13688        }
13689        #[cfg(unix)]
13690        {
13691            let marker = directory.join("actor-authority.initialized.json");
13692            std::os::unix::fs::symlink(directory.join("missing"), &marker).unwrap();
13693            assert!(require_unowned_glob(&store, &child).await.is_err());
13694        }
13695    }
13696
13697    #[test]
13698    fn owned_glob_collector_requires_actual_ordered_tool_trace_and_typed_tail() {
13699        let call = bamboo_domain::ToolCall {
13700            id: "glob".into(),
13701            tool_type: "function".into(),
13702            function: bamboo_domain::FunctionCall {
13703                name: "Glob".into(),
13704                arguments: r#"{"pattern":"marker.txt"}"#.into(),
13705            },
13706        };
13707        let messages = [
13708            bamboo_agent_core::Message::assistant("", Some(vec![call])),
13709            bamboo_agent_core::Message::tool_result("glob", "marker.txt"),
13710            bamboo_agent_core::Message::assistant("done", None),
13711        ];
13712        let start = serde_json::to_value(AgentEvent::ToolStart {
13713            tool_call_id: "glob".into(),
13714            tool_name: "Glob".into(),
13715            arguments: serde_json::json!({"pattern":"marker.txt"}),
13716        })
13717        .unwrap();
13718        let complete = serde_json::to_value(AgentEvent::ToolComplete {
13719            tool_call_id: "glob".into(),
13720            result: bamboo_domain::ToolResult::text(true, "marker.txt"),
13721        })
13722        .unwrap();
13723        let tail =
13724            serde_json::to_value(bamboo_subagent::proto::ReadOnlyActorTranscript::Complete {
13725                messages: messages
13726                    .iter()
13727                    .map(serde_json::to_value)
13728                    .collect::<Result<_, _>>()
13729                    .unwrap(),
13730            })
13731            .unwrap();
13732        let mut ordered = ReadOnlyActorCollector::default();
13733        assert!(ordered.event(&start).unwrap());
13734        assert!(ordered.event(&complete).unwrap());
13735        assert!(!ordered.event(&tail).unwrap());
13736        assert!(ordered.event(&tail).is_err(), "duplicate suffix");
13737        assert_eq!(ordered.finish(Some("done")).unwrap().len(), 3);
13738        assert!(ReadOnlyActorCollector::default()
13739            .finish(Some("done"))
13740            .is_err());
13741        let mut unfinished = ReadOnlyActorCollector::default();
13742        unfinished.event(&start).unwrap();
13743        unfinished.event(&complete).unwrap();
13744        assert!(unfinished.finish(Some("done")).is_err(), "missing suffix");
13745        let mut closed = tail.clone();
13746        closed["grant"] = true.into();
13747        assert!(
13748            serde_json::from_value::<bamboo_subagent::proto::ReadOnlyActorTranscript>(closed)
13749                .is_err()
13750        );
13751        let mut oversized = tail.clone();
13752        oversized["messages"][2]["content"] = "x".repeat(64 * 1024).into();
13753        assert!(
13754            serde_json::from_value::<bamboo_subagent::proto::ReadOnlyActorTranscript>(oversized)
13755                .unwrap()
13756                .validate()
13757                .is_err()
13758        );
13759        assert!(ReadOnlyActorCollector::default().event(&complete).is_err());
13760        assert!(ReadOnlyActorCollector::default().event(&tail).is_err());
13761        let mut bad = start.clone();
13762        bad["tool_name"] = "Bash".into();
13763        assert!(ReadOnlyActorCollector::default().event(&bad).is_err());
13764        let mut duplicate = ReadOnlyActorCollector::default();
13765        duplicate.event(&start).unwrap();
13766        assert!(duplicate.event(&start).is_err());
13767        let hidden = serde_json::to_value(AgentEvent::ReasoningToken {
13768            content: "hidden".into(),
13769        })
13770        .unwrap();
13771        assert!(ReadOnlyActorCollector::default().event(&hidden).is_err());
13772        let mut mismatched = ReadOnlyActorCollector::default();
13773        mismatched.event(&start).unwrap();
13774        let mut bad = complete.clone();
13775        bad["tool_call_id"] = "foreign".into();
13776        assert!(mismatched.event(&bad).is_err());
13777        mismatched.event(&complete).unwrap();
13778        let mut changed = tail.clone();
13779        changed["messages"][1]["content"] = "fabricated".into();
13780        assert!(mismatched.event(&changed).is_err());
13781    }
13782
13783    #[cfg(unix)]
13784    #[tokio::test]
13785    async fn owned_initial_release_requires_current_prefix_ack_and_same_receipt() {
13786        let temp = tempfile::tempdir().unwrap();
13787        let store = Arc::new(
13788            bamboo_storage::SessionStoreV2::new(temp.path().into())
13789                .await
13790                .unwrap(),
13791        );
13792        let parent = Session::new("release-host-parent", "model");
13793        store.save_session(&parent).await.unwrap();
13794        let mut child = Session::new_child_of("release-host-child", &parent, "model", "assignment");
13795        store.save_session(&child).await.unwrap();
13796        let inbox: Arc<dyn SessionInboxPort> = Arc::new(bamboo_storage::FileSessionInbox::new(
13797            store.clone(),
13798            bamboo_domain::SessionInboxLimits::default(),
13799        ));
13800        let binding = actor_binding(
13801            store.clone(),
13802            inbox,
13803            Arc::new(bamboo_storage::LockedSessionStore::new(store.clone())),
13804        );
13805        let _owner = binding
13806            .router
13807            .register_run(&child.id, "release-host-run")
13808            .await
13809            .unwrap();
13810        let policy = Arc::new(bamboo_tools::permission::PermissionConfig::new());
13811        let revision = policy.policy_revision();
13812        bind_local_control_plane(store.as_ref(), &child.id, policy.as_ref()).await;
13813        let activation = PlainActorActivation::start(
13814            store.clone(),
13815            &child,
13816            &binding,
13817            Some("release-host-run"),
13818            Some(policy.clone()),
13819            true,
13820            Some("probed-physical-worker"),
13821        )
13822        .await
13823        .unwrap();
13824        let envelope = bamboo_domain::SessionMessageEnvelope::user_input(&child.id, "RELEASE_ONCE");
13825        let receipt = activation.input_inbox.deliver(&envelope).await.unwrap();
13826        activation
13827            .input_inbox
13828            .mark_activation_eligible(
13829                &child.id,
13830                receipt.generation,
13831                bamboo_domain::SessionActivationPolicy::InterruptSpecificWait,
13832            )
13833            .await
13834            .unwrap();
13835        let claim = activation
13836            .claim_input(&binding, &child, "release-host-run")
13837            .await
13838            .unwrap()
13839            .unwrap();
13840        let (prefix, delivery) = activation
13841            .checkpoint_input(&mut child, "release-host-run", &claim)
13842            .await
13843            .unwrap();
13844        let mut expected = Some(expected_default_permission_posture(revision));
13845        let ChildFrame::Event { event } = permission_posture_frame(&child.id, revision) else {
13846            unreachable!()
13847        };
13848        let (events, mut events_rx) = mpsc::channel(1);
13849        process_actor_event(
13850            event,
13851            true,
13852            &mut PermissionPostureHandshake::Awaiting,
13853            expected.as_ref(),
13854            Some(&binding),
13855            &mut child,
13856            &events,
13857            &mut ActorEventDisplay::default(),
13858        )
13859        .await
13860        .unwrap();
13861        events_rx.recv().await.unwrap();
13862        let run = RunSpec {
13863            assignment: "assignment".into(),
13864            logical_session: Some(LogicalSessionIdentity {
13865                session_id: child.id.clone(),
13866                parent_session_id: child.parent_session_id.clone(),
13867                root_session_id: child.root_session_id.clone(),
13868                creation: Some(bamboo_subagent::proto::ChildCreationIdentity {
13869                    created_at: child.created_at,
13870                    spawn_depth: child.spawn_depth,
13871                }),
13872            }),
13873            project_id: None,
13874            reasoning_effort: None,
13875            permission_policy: Some(bamboo_subagent::proto::PermissionPolicyContext {
13876                revision,
13877                requested_mode: "default".into(),
13878                effective_mode: "default".into(),
13879                bypass_permissions: false,
13880                auto_approve_permissions: false,
13881                session_id: child.id.clone(),
13882                workspace_path: child.workspace.clone(),
13883                environment_lease: None,
13884                inherit_session_grants: false,
13885                policy: serde_json::to_value(policy.to_serializable()).unwrap(),
13886            }),
13887            messages: prefix,
13888            activation_run_id: Some("release-host-run".into()),
13889            execution_epoch: 9,
13890            initial_session_messages: vec![delivery.clone()],
13891            secrets: Default::default(),
13892        };
13893        let request = bamboo_subagent::proto::InitialInputReleaseRequest::from_run(
13894            &run,
13895            &delivery,
13896            uuid::Uuid::new_v4().to_string(),
13897        )
13898        .unwrap();
13899        let mut stale = child.clone();
13900        stale.add_message(bamboo_agent_core::Message::user("not canonical"));
13901        assert!(activation
13902            .release_initial_input(
13903                &binding,
13904                &stale,
13905                &run,
13906                9,
13907                &claim,
13908                &request,
13909                &mut expected,
13910                None
13911            )
13912            .await
13913            .is_err());
13914        assert!(!activation
13915            .input_inbox
13916            .was_admitted(&child.id, &envelope.id)
13917            .await
13918            .unwrap());
13919        // Keep receipt lookup readable so Already reaches the real ACK write.
13920        // Unix root can bypass these modes; this FS fault needs an unprivileged user.
13921        use std::os::unix::fs::PermissionsExt;
13922        struct RestorePermissions(std::path::PathBuf, std::fs::Permissions);
13923        impl Drop for RestorePermissions {
13924            fn drop(&mut self) {
13925                let _ = std::fs::set_permissions(&self.0, self.1.clone());
13926            }
13927        }
13928        let admitted = store
13929            .bamboo_home_dir()
13930            .join(store.resolve_rel_path(&child.id).await.unwrap())
13931            .join("inbox/admitted");
13932        std::fs::create_dir_all(&admitted).unwrap();
13933        let restore = RestorePermissions(
13934            admitted.clone(),
13935            std::fs::metadata(&admitted).unwrap().permissions(),
13936        );
13937        std::fs::set_permissions(&admitted, std::fs::Permissions::from_mode(0o555)).unwrap();
13938        let verified = activation
13939            .input_inbox
13940            .checkpoint_actor_input(bamboo_storage::ActorInputCheckpoint {
13941                fence: activation.fence.clone(),
13942                expected_created_at: child.created_at,
13943                expected_conversation_generation: child.conversation_generation_observation(),
13944                claim: claim.clone(),
13945                expected_messages: child.messages.clone(),
13946                expected_provider_transcript: child.provider_transcript.clone(),
13947                expected_admission: child.session_inbox_admission().cloned(),
13948            })
13949            .await
13950            .unwrap();
13951        assert_eq!(
13952            verified.status,
13953            bamboo_storage::ActorInputCheckpointStatus::AlreadyCheckpointed
13954        );
13955        let error = activation
13956            .release_initial_input(
13957                &binding,
13958                &child,
13959                &run,
13960                9,
13961                &claim,
13962                &request,
13963                &mut expected,
13964                None,
13965            )
13966            .await
13967            .unwrap_err();
13968        std::fs::set_permissions(&admitted, restore.1.clone()).unwrap();
13969        drop(restore);
13970        assert!(
13971            error.to_string().contains("ACK unresolved"),
13972            "actual release error (<=512B): {}",
13973            error.to_string().chars().take(128).collect::<String>()
13974        );
13975        let release = activation
13976            .release_initial_input(
13977                &binding,
13978                &child,
13979                &run,
13980                9,
13981                &claim,
13982                &request,
13983                &mut expected,
13984                None,
13985            )
13986            .await
13987            .unwrap();
13988        assert!(activation
13989            .input_inbox
13990            .was_admitted(&child.id, &envelope.id)
13991            .await
13992            .unwrap());
13993        assert_eq!(
13994            release.expires_at,
13995            activation
13996                .actor_deadline
13997                .min(claim.lease.expires_at)
13998                .min(activation.input_deadline)
13999        );
14000        assert_eq!(
14001            activation
14002                .release_initial_input(
14003                    &binding,
14004                    &child,
14005                    &run,
14006                    9,
14007                    &claim,
14008                    &request,
14009                    &mut expected,
14010                    Some(&release)
14011                )
14012                .await
14013                .unwrap(),
14014            release
14015        );
14016        let mut wrong = request.clone();
14017        wrong.nonce = uuid::Uuid::new_v4().to_string();
14018        assert!(activation
14019            .release_initial_input(
14020                &binding,
14021                &child,
14022                &run,
14023                9,
14024                &claim,
14025                &wrong,
14026                &mut expected,
14027                Some(&release)
14028            )
14029            .await
14030            .is_err());
14031        assert!(activation
14032            .release_initial_input(
14033                &binding,
14034                &child,
14035                &run,
14036                10,
14037                &claim,
14038                &request,
14039                &mut expected,
14040                Some(&release)
14041            )
14042            .await
14043            .is_err());
14044        // A second real checkpoint/release reuses the activation, not its old epoch or nonce.
14045        let deadlines = (
14046            activation.fence.clone(),
14047            activation.actor_deadline,
14048            activation.input_deadline,
14049        );
14050        let second = bamboo_domain::SessionMessageEnvelope::user_input(&child.id, "RELEASE_TWICE");
14051        let receipt = activation.input_inbox.deliver(&second).await.unwrap();
14052        activation
14053            .input_inbox
14054            .mark_activation_eligible(
14055                &child.id,
14056                receipt.generation,
14057                bamboo_domain::SessionActivationPolicy::InterruptSpecificWait,
14058            )
14059            .await
14060            .unwrap();
14061        let mut link = ConfirmationSequenceLink {
14062            frames: VecDeque::new(),
14063            sent: Vec::new(),
14064        };
14065        let epochs = AtomicU64::new(9);
14066        let (claim2, epoch2) = activation
14067            .continue_input(
14068                &mut link,
14069                &binding,
14070                &mut child,
14071                "release-host-run",
14072                Some((&run, &epochs)),
14073                &mut expected,
14074                "SECOND_REPLY".into(),
14075                &events,
14076            )
14077            .await
14078            .unwrap()
14079            .unwrap();
14080        events_rx.recv().await.unwrap();
14081        assert_eq!(epoch2, 10);
14082        let ParentFrame::Run(next) = &link.sent[0] else {
14083            panic!("second correlated Run")
14084        };
14085        assert_eq!(next.logical_session, run.logical_session);
14086        assert_eq!(next.activation_run_id, run.activation_run_id);
14087        assert_eq!(next.initial_session_messages.len(), 1);
14088        assert!(next
14089            .messages
14090            .iter()
14091            .any(|m| m["id"] == envelope.id.as_str()));
14092        assert!(next.messages.iter().any(|m| m["content"] == "SECOND_REPLY"));
14093        assert!(!next.messages.iter().any(|m| m["id"] == second.id.as_str()));
14094        let request2 = bamboo_subagent::proto::InitialInputReleaseRequest::from_run(
14095            next,
14096            &next.initial_session_messages[0],
14097            uuid::Uuid::new_v4().to_string(),
14098        )
14099        .unwrap();
14100        assert_ne!(request2.nonce, request.nonce);
14101        macro_rules! release_second {
14102            ($candidate:expr, $request:expr, $epoch:expr, $previous:expr) => {
14103                activation
14104                    .release_initial_input(
14105                        &binding,
14106                        $candidate,
14107                        next,
14108                        $epoch,
14109                        &claim2,
14110                        $request,
14111                        &mut expected,
14112                        $previous,
14113                    )
14114                    .await
14115            };
14116        }
14117        assert!(release_second!(&child, &request, epoch2, None).is_err());
14118        assert!(release_second!(&child, &request2, 9, None).is_err());
14119        let mut stale2 = child.clone();
14120        stale2.add_message(bamboo_agent_core::Message::user("stale second prefix"));
14121        assert!(release_second!(&stale2, &request2, epoch2, None).is_err());
14122        assert!(!activation
14123            .input_inbox
14124            .was_admitted(&child.id, &second.id)
14125            .await
14126            .unwrap());
14127        let release2 = release_second!(&child, &request2, epoch2, None).unwrap();
14128        assert_eq!(
14129            release_second!(&child, &request2, epoch2, Some(&release2)).unwrap(),
14130            release2
14131        );
14132        let mut wrong2 = request2.clone();
14133        wrong2.nonce = request.nonce.clone();
14134        assert!(release_second!(&child, &wrong2, epoch2, Some(&release2)).is_err());
14135        assert!(activation
14136            .input_inbox
14137            .was_admitted(&child.id, &second.id)
14138            .await
14139            .unwrap());
14140        let expired = bamboo_domain::SessionMessageEnvelope::user_input(&child.id, "EXPIRED_INPUT");
14141        let receipt = activation.input_inbox.deliver(&expired).await.unwrap();
14142        activation
14143            .input_inbox
14144            .mark_activation_eligible(
14145                &child.id,
14146                receipt.generation,
14147                bamboo_domain::SessionActivationPolicy::InterruptSpecificWait,
14148            )
14149            .await
14150            .unwrap();
14151        let now = chrono::Utc::now();
14152        let expired_claim = activation
14153            .input_inbox
14154            .claim_owned(
14155                &child.id,
14156                1,
14157                Some("release-host-run"),
14158                &SessionInboxLeaseRequest {
14159                    consumer: activation.input_consumer.clone(),
14160                    now,
14161                    duration: chrono::Duration::seconds(1),
14162                },
14163            )
14164            .await
14165            .unwrap()
14166            .pop()
14167            .unwrap();
14168        let (prefix, delivery3) = activation
14169            .checkpoint_input(&mut child, "release-host-run", &expired_claim)
14170            .await
14171            .unwrap();
14172        let mut expired_run = next.clone();
14173        expired_run.execution_epoch = 11;
14174        expired_run.messages = prefix;
14175        expired_run.initial_session_messages = vec![delivery3.clone()];
14176        let expired_request = bamboo_subagent::proto::InitialInputReleaseRequest::from_run(
14177            &expired_run,
14178            &delivery3,
14179            uuid::Uuid::new_v4().to_string(),
14180        )
14181        .unwrap();
14182        tokio::time::sleep(
14183            (expired_claim.lease.expires_at - chrono::Utc::now())
14184                .to_std()
14185                .unwrap_or_default(),
14186        )
14187        .await;
14188        assert!(chrono::Utc::now() >= expired_claim.lease.expires_at);
14189        assert!(activation
14190            .release_initial_input(
14191                &binding,
14192                &child,
14193                &expired_run,
14194                11,
14195                &expired_claim,
14196                &expired_request,
14197                &mut expected,
14198                None
14199            )
14200            .await
14201            .is_err());
14202        assert!(!activation
14203            .input_inbox
14204            .was_admitted(&child.id, &expired.id)
14205            .await
14206            .unwrap());
14207        assert_eq!(
14208            (
14209                activation.fence.clone(),
14210                activation.actor_deadline,
14211                activation.input_deadline
14212            ),
14213            deadlines
14214        );
14215        assert_eq!(
14216            link.sent.len(),
14217            1,
14218            "no dispatch from rejected release checks"
14219        );
14220        let cold = store.load_session(&child.id).await.unwrap().unwrap();
14221        for id in [&envelope.id, &second.id] {
14222            assert_eq!(
14223                cold.messages.iter().filter(|m| m.id == id.as_str()).count(),
14224                1
14225            );
14226            assert!(cold.session_inbox_admission().unwrap().contains(id));
14227        }
14228        // Mutate the actual captured policy Arc; cached ACK/release is not a fresh grant.
14229        policy.deny_scoped_session_permission(
14230            &child.id,
14231            bamboo_tools::permission::PermissionType::ExecuteCommand,
14232            "blocked-action",
14233        );
14234        assert!(release_second!(&child, &request2, epoch2, Some(&release2)).is_err());
14235        assert!(activation
14236            .input_inbox
14237            .was_admitted(&child.id, &second.id)
14238            .await
14239            .unwrap());
14240        activation
14241            .finish(ActorActivationFinish::Failed)
14242            .await
14243            .unwrap();
14244        // Terminal ACK can be idempotent; stopped/expired Actor cannot grant.
14245        activation
14246            .input_inbox
14247            .ack_owned(&child.id, &claim, chrono::Utc::now())
14248            .await
14249            .unwrap();
14250        assert!(activation
14251            .release_initial_input(
14252                &binding,
14253                &child,
14254                &run,
14255                9,
14256                &claim,
14257                &request,
14258                &mut expected,
14259                Some(&release)
14260            )
14261            .await
14262            .is_err());
14263        let cold = bamboo_storage::SessionStoreV2::new(temp.path().into())
14264            .await
14265            .unwrap()
14266            .load_session(&child.id)
14267            .await
14268            .unwrap()
14269            .unwrap();
14270        assert_eq!(
14271            cold.messages
14272                .iter()
14273                .filter(|m| m.id == envelope.id.as_str())
14274                .count(),
14275            1
14276        );
14277    }
14278
14279    #[tokio::test]
14280    async fn owned_actor_input_checkpoint_precedes_dispatch_and_exact_ack() {
14281        for reject_checkpoint in [false, true] {
14282            let temp = tempfile::tempdir().unwrap();
14283            let store = Arc::new(
14284                bamboo_storage::SessionStoreV2::new(temp.path().into())
14285                    .await
14286                    .unwrap(),
14287            );
14288            let parent = Session::new("owned-input-parent", "model");
14289            store.save_session(&parent).await.unwrap();
14290            let mut child =
14291                Session::new_child_of("owned-input-child", &parent, "model", "one plain task");
14292            store.save_session(&child).await.unwrap();
14293            let policy = Arc::new(bamboo_tools::permission::PermissionConfig::new());
14294            let raw_storage: Arc<dyn Storage> = store.clone();
14295            let locked = Arc::new(bamboo_storage::LockedSessionStore::new(raw_storage));
14296            let inbox: Arc<dyn SessionInboxPort> = Arc::new(bamboo_storage::FileSessionInbox::new(
14297                store.clone(),
14298                bamboo_domain::SessionInboxLimits::default(),
14299            ));
14300            let binding = actor_binding(store.clone(), inbox, locked);
14301            let run_id = "owned-input-run";
14302            let _registration = binding
14303                .router
14304                .register_run(&child.id, run_id)
14305                .await
14306                .unwrap();
14307            assert!(PlainActorActivation::start(
14308                store.clone(),
14309                &child,
14310                &binding,
14311                Some(run_id),
14312                None,
14313                true,
14314                None,
14315            )
14316            .await
14317            .is_err());
14318            assert!(PlainActorActivation::start(
14319                store.clone(),
14320                &child,
14321                &binding,
14322                Some(run_id),
14323                Some(policy.clone()),
14324                true,
14325                None,
14326            )
14327            .await
14328            .is_err());
14329            assert_eq!(
14330                store
14331                    .inspect_actor(&child.id)
14332                    .await
14333                    .unwrap()
14334                    .actor
14335                    .current_attempt,
14336                0
14337            );
14338            bind_local_control_plane(store.as_ref(), &child.id, policy.as_ref()).await;
14339            let activation = PlainActorActivation::start(
14340                store.clone(),
14341                &child,
14342                &binding,
14343                Some(run_id),
14344                Some(policy.clone()),
14345                true,
14346                None,
14347            )
14348            .await
14349            .unwrap();
14350            let envelope =
14351                bamboo_domain::SessionMessageEnvelope::user_input(&child.id, "CORRECTION_ONCE");
14352            let receipt = activation.input_inbox.deliver(&envelope).await.unwrap();
14353            activation
14354                .input_inbox
14355                .mark_activation_eligible(
14356                    &child.id,
14357                    receipt.generation,
14358                    bamboo_domain::SessionActivationPolicy::InterruptSpecificWait,
14359                )
14360                .await
14361                .unwrap();
14362            let mut link = ConfirmationSequenceLink {
14363                frames: VecDeque::new(),
14364                sent: Vec::new(),
14365            };
14366            let policy_revision = policy.policy_revision();
14367            let run = RunSpec {
14368                assignment: "one plain task".into(),
14369                logical_session: None,
14370                project_id: None,
14371                reasoning_effort: None,
14372                permission_policy: Some(PermissionPolicyContext {
14373                    revision: policy_revision,
14374                    requested_mode: "default".into(),
14375                    effective_mode: "default".into(),
14376                    bypass_permissions: false,
14377                    auto_approve_permissions: false,
14378                    session_id: child.id.clone(),
14379                    workspace_path: child.workspace.clone(),
14380                    environment_lease: None,
14381                    inherit_session_grants: false,
14382                    policy: serde_json::to_value(policy.to_serializable()).unwrap(),
14383                }),
14384                messages: vec![],
14385                activation_run_id: Some(run_id.into()),
14386                execution_epoch: 1,
14387                initial_session_messages: vec![],
14388                secrets: Default::default(),
14389            };
14390            let epochs = AtomicU64::new(1);
14391            let (tx, mut rx) = mpsc::channel(1);
14392            let mut expected = Some(expected_default_permission_posture(policy_revision));
14393            let ChildFrame::Event { event } = permission_posture_frame(&child.id, policy_revision)
14394            else {
14395                unreachable!()
14396            };
14397            process_actor_event(
14398                event,
14399                true,
14400                &mut PermissionPostureHandshake::Awaiting,
14401                expected.as_ref(),
14402                Some(&binding),
14403                &mut child,
14404                &tx,
14405                &mut ActorEventDisplay::default(),
14406            )
14407            .await
14408            .unwrap();
14409            rx.recv().await.unwrap();
14410            let first_audit =
14411                bamboo_domain::PermissionAuditSnapshot::from_metadata(&child.metadata)
14412                    .unwrap()
14413                    .audit_revision;
14414            if reject_checkpoint {
14415                // A real bounded publication channel blocks AFTER first-reply
14416                // commit, so change actual Actor authority before checkpoint.
14417                tx.send(AgentEvent::Complete {
14418                    usage: Default::default(),
14419                })
14420                .await
14421                .unwrap();
14422            }
14423            let mut operation = activation.continue_input(
14424                &mut link,
14425                &binding,
14426                &mut child,
14427                run_id,
14428                Some((&run, &epochs)),
14429                &mut expected,
14430                "FIRST_PLAIN_REPLY".into(),
14431                &tx,
14432            );
14433            if reject_checkpoint {
14434                tokio::select! {
14435                    result = &mut operation => panic!("publication barrier released early: {result:?}"),
14436                    _ = async {
14437                        tokio::time::timeout(Duration::from_secs(5), async {
14438                            loop {
14439                                let actual = store.load_session("owned-input-child").await.unwrap().unwrap();
14440                                if actual.messages.iter().any(|m| m.content == "FIRST_PLAIN_REPLY") { break; }
14441                                tokio::task::yield_now().await;
14442                            }
14443                        }).await.unwrap();
14444                        activation.finish(ActorActivationFinish::Failed).await.unwrap();
14445                        rx.recv().await.unwrap();
14446                    } => {}
14447                }
14448            }
14449            let result = operation.await;
14450            if reject_checkpoint {
14451                assert!(result
14452                    .unwrap_err()
14453                    .to_string()
14454                    .contains("checkpoint unconfirmed"));
14455                assert!(
14456                    link.sent.is_empty(),
14457                    "zero worker/provider correction dispatch"
14458                );
14459                let durable = store.load_session(&child.id).await.unwrap().unwrap();
14460                assert!(durable
14461                    .messages
14462                    .iter()
14463                    .any(|m| m.content == "FIRST_PLAIN_REPLY"));
14464                assert!(!durable
14465                    .messages
14466                    .iter()
14467                    .any(|m| m.id == envelope.id.as_str()));
14468                assert!(!activation
14469                    .input_inbox
14470                    .was_admitted(&child.id, &envelope.id)
14471                    .await
14472                    .unwrap());
14473                assert_eq!(
14474                    activation
14475                        .input_inbox
14476                        .inspect(&child.id)
14477                        .await
14478                        .unwrap()
14479                        .claimed,
14480                    1
14481                );
14482                continue;
14483            }
14484            let (claim, epoch) = result.unwrap().unwrap();
14485            rx.recv().await.unwrap(); // Actual first-reply publication.
14486            assert_eq!(epoch, 2);
14487            assert_eq!(link.sent.len(), 1);
14488            let ParentFrame::Run(next) = &link.sent[0] else {
14489                panic!("correlated continuation Run")
14490            };
14491            assert_eq!(next.execution_epoch, 2);
14492            assert_eq!(next.activation_run_id.as_deref(), Some(run_id));
14493            assert!(next
14494                .messages
14495                .iter()
14496                .any(|m| m["content"] == "FIRST_PLAIN_REPLY"));
14497            assert!(!next
14498                .messages
14499                .iter()
14500                .any(|m| m["id"] == envelope.id.as_str()));
14501            let delivery = &next.initial_session_messages[0];
14502            assert_eq!(delivery.envelope, envelope);
14503            let durable = store.load_session(&child.id).await.unwrap().unwrap();
14504            assert!(durable
14505                .session_inbox_admission()
14506                .unwrap()
14507                .contains(&envelope.id));
14508            assert_eq!(
14509                durable
14510                    .messages
14511                    .iter()
14512                    .filter(|m| m.id == envelope.id.as_str())
14513                    .count(),
14514                1
14515            );
14516            assert!(
14517                !activation
14518                    .input_inbox
14519                    .was_admitted(&child.id, &envelope.id)
14520                    .await
14521                    .unwrap(),
14522                "checkpoint is not a worker admission receipt"
14523            );
14524            let mut confirmation = admission_confirmation(&child.id, &claim.claim, run_id);
14525            confirmation.canonical_claim_generation += 1;
14526            assert!(activation
14527                .confirm_input(&binding, &child, run_id, &claim, &confirmation)
14528                .await
14529                .is_err());
14530            assert!(!activation
14531                .input_inbox
14532                .was_admitted(&child.id, &envelope.id)
14533                .await
14534                .unwrap());
14535            confirmation.canonical_claim_generation = claim.claim.generation;
14536            activation
14537                .confirm_input(&binding, &child, run_id, &claim, &confirmation)
14538                .await
14539                .unwrap();
14540            assert!(activation
14541                .input_inbox
14542                .was_admitted(&child.id, &envelope.id)
14543                .await
14544                .unwrap());
14545            assert_eq!(
14546                expected.as_ref().unwrap().expected_audit_revision,
14547                Some(first_audit)
14548            );
14549            let ChildFrame::Event { event } = permission_posture_frame(&child.id, policy_revision)
14550            else {
14551                unreachable!()
14552            };
14553            process_actor_event(
14554                event,
14555                true,
14556                &mut PermissionPostureHandshake::Awaiting,
14557                expected.as_ref(),
14558                Some(&binding),
14559                &mut child,
14560                &tx,
14561                &mut ActorEventDisplay::default(),
14562            )
14563            .await
14564            .unwrap();
14565            rx.recv().await.unwrap();
14566            assert!(
14567                bamboo_domain::PermissionAuditSnapshot::from_metadata(&child.metadata)
14568                    .unwrap()
14569                    .audit_revision
14570                    > first_audit
14571            );
14572            let reopened = bamboo_storage::SessionStoreV2::new(temp.path().into())
14573                .await
14574                .unwrap();
14575            let cold = reopened.load_session(&child.id).await.unwrap().unwrap();
14576            assert_eq!(
14577                cold.messages
14578                    .iter()
14579                    .filter(|m| m.id == envelope.id.as_str())
14580                    .count(),
14581                1
14582            );
14583            assert!(cold
14584                .session_inbox_admission()
14585                .unwrap()
14586                .contains(&envelope.id));
14587        }
14588    }
14589
14590    fn admission_confirmation(
14591        session_id: &str,
14592        claim: &SessionInboxClaim,
14593        run_id: &str,
14594    ) -> bamboo_subagent::proto::SessionMessageAdmissionConfirmation {
14595        bamboo_subagent::proto::SessionMessageAdmissionConfirmation {
14596            target_session_id: session_id.to_string(),
14597            envelope_id: claim.envelope.id.to_string(),
14598            canonical_claim_generation: claim.generation,
14599            activation_run_id: run_id.to_string(),
14600        }
14601    }
14602
14603    fn expected_default_permission_posture(policy_revision: u64) -> ExpectedPermissionPosture {
14604        ExpectedPermissionPosture {
14605            policy_revision,
14606            resolution: bamboo_domain::PermissionModeResolution {
14607                requested: bamboo_domain::SessionPermissionMode::Default,
14608                effective: bamboo_domain::PermissionMode::Default,
14609            },
14610            expected_audit_revision: None,
14611            executor_mapping: "test_actor:permission_mode=default".to_string(),
14612        }
14613    }
14614
14615    fn permission_posture_frame(session_id: &str, policy_revision: u64) -> ChildFrame {
14616        ChildFrame::Event {
14617            event: serde_json::to_value(AgentEvent::PermissionPostureActivated {
14618                session_id: session_id.to_string(),
14619                policy_revision,
14620                requested_mode: "default".to_string(),
14621                effective_mode: "default".to_string(),
14622                executor_mapping: "test_actor:permission_mode=default".to_string(),
14623            })
14624            .expect("serialize permission posture event"),
14625        }
14626    }
14627
14628    fn actor_event_frame(event: AgentEvent) -> ChildFrame {
14629        ChildFrame::Event {
14630            event: serde_json::to_value(event).expect("serialize actor event"),
14631        }
14632    }
14633
14634    #[test]
14635    fn actor_event_batch_is_fenced_by_session_activation_and_epoch() {
14636        let session = Session::new_child("logical-child", "logical-parent", "model", "child");
14637        let mut batch = ActorEventBatch {
14638            logical_session: Some(LogicalSessionIdentity {
14639                creation: None,
14640                session_id: session.id.clone(),
14641                parent_session_id: session.parent_session_id.clone(),
14642                root_session_id: session.root_session_id.clone(),
14643            }),
14644            activation_id: Some("activation-7".into()),
14645            execution_epoch: 11,
14646            source_node_id: Some("node-a".into()),
14647            source_actor_id: Some("worker-a".into()),
14648            first_seq: 1,
14649            last_seq: 1,
14650            qos: bamboo_subagent::ActorEventQos::Ephemeral,
14651            events: vec![serde_json::json!({"type":"token","content":"ok"})],
14652        };
14653        validate_actor_event_batch(
14654            &batch,
14655            &session,
14656            "logical-parent",
14657            Some("activation-7"),
14658            11,
14659            "worker-a",
14660            None,
14661        )
14662        .unwrap();
14663
14664        let creation = bamboo_subagent::proto::ChildCreationIdentity {
14665            created_at: session.created_at,
14666            spawn_depth: session.spawn_depth,
14667        };
14668        assert!(validate_actor_event_batch(
14669            &batch,
14670            &session,
14671            "logical-parent",
14672            Some("activation-7"),
14673            11,
14674            "worker-a",
14675            Some(&creation)
14676        )
14677        .is_err());
14678        batch.logical_session.as_mut().unwrap().creation = Some(creation.clone());
14679        validate_actor_event_batch(
14680            &batch,
14681            &session,
14682            "logical-parent",
14683            Some("activation-7"),
14684            11,
14685            "worker-a",
14686            Some(&creation),
14687        )
14688        .unwrap();
14689        batch
14690            .logical_session
14691            .as_mut()
14692            .unwrap()
14693            .creation
14694            .as_mut()
14695            .unwrap()
14696            .created_at += chrono::Duration::nanoseconds(1);
14697        assert!(validate_actor_event_batch(
14698            &batch,
14699            &session,
14700            "logical-parent",
14701            Some("activation-7"),
14702            11,
14703            "worker-a",
14704            Some(&creation)
14705        )
14706        .is_err());
14707        batch.logical_session.as_mut().unwrap().creation = None;
14708
14709        batch.execution_epoch = 10;
14710        assert!(validate_actor_event_batch(
14711            &batch,
14712            &session,
14713            "logical-parent",
14714            Some("activation-7"),
14715            11,
14716            "worker-a",
14717            None,
14718        )
14719        .unwrap_err()
14720        .to_string()
14721        .contains("execution epoch"));
14722        batch.execution_epoch = 11;
14723        batch.activation_id = Some("replaced-activation".into());
14724        assert!(validate_actor_event_batch(
14725            &batch,
14726            &session,
14727            "logical-parent",
14728            Some("activation-7"),
14729            11,
14730            "worker-a",
14731            None,
14732        )
14733        .unwrap_err()
14734        .to_string()
14735        .contains("stale activation"));
14736        batch.activation_id = Some("activation-7".into());
14737        batch
14738            .logical_session
14739            .as_mut()
14740            .expect("logical identity")
14741            .session_id = "foreign-child".into();
14742        assert!(validate_actor_event_batch(
14743            &batch,
14744            &session,
14745            "logical-parent",
14746            Some("activation-7"),
14747            11,
14748            "worker-a",
14749            None,
14750        )
14751        .unwrap_err()
14752        .to_string()
14753        .contains("different logical session"));
14754        batch
14755            .logical_session
14756            .as_mut()
14757            .expect("logical identity")
14758            .session_id = session.id.clone();
14759        batch.source_actor_id = Some("worker-b".into());
14760        assert!(validate_actor_event_batch(
14761            &batch,
14762            &session,
14763            "logical-parent",
14764            Some("activation-7"),
14765            11,
14766            "worker-a",
14767            None,
14768        )
14769        .unwrap_err()
14770        .to_string()
14771        .contains("different physical actor"));
14772    }
14773
14774    #[tokio::test]
14775    async fn directory_event_ingress_rejects_old_attempt_and_forged_identity() {
14776        use bamboo_domain::Storage;
14777
14778        let temp = tempfile::tempdir().unwrap();
14779        let store = bamboo_storage::SessionStoreV2::new(temp.path().into())
14780            .await
14781            .unwrap();
14782        let parent = Session::new("event-parent", "model");
14783        let child = Session::new_child_of("event-child", &parent, "model", "task");
14784        store.save_session(&parent).await.unwrap();
14785        store.save_session(&child).await.unwrap();
14786        store.ensure_actor(&child.id).await.unwrap();
14787        let now = chrono::Utc::now();
14788        let claim = |run_id: &str| ActorActivationClaim {
14789            actor_id: child.id.clone(),
14790            run_id: run_id.into(),
14791            lease_owner: "event-owner".into(),
14792            lease_expires_at: now + chrono::Duration::minutes(2),
14793            inbox_generation: 0,
14794            placement_ref: None,
14795            now,
14796        };
14797        let first = store.claim_activation(&claim("run-one")).await.unwrap();
14798        store.start_activation(&first.fence(), now).await.unwrap();
14799        let first_fence = first.fence();
14800        validate_directory_actor_event(&store, &first_fence, &child, Some("run-one"))
14801            .await
14802            .unwrap();
14803        validate_directory_actor_event(&store, &first_fence, &child, None)
14804            .await
14805            .unwrap(); // rolling-upgrade frame uses the same owner
14806        assert!(
14807            validate_directory_actor_event(&store, &first_fence, &child, Some("forged-run"))
14808                .await
14809                .is_err()
14810        );
14811        let mut forged = child.clone();
14812        forged.root_session_id = "foreign-root".into();
14813        assert!(
14814            validate_directory_actor_event(&store, &first_fence, &forged, Some("run-one"))
14815                .await
14816                .is_err()
14817        );
14818
14819        store
14820            .finish_activation(&first_fence, now, ActorActivationFinish::Failed)
14821            .await
14822            .unwrap();
14823        let second = store.claim_activation(&claim("run-two")).await.unwrap();
14824        store.start_activation(&second.fence(), now).await.unwrap();
14825        assert!(
14826            validate_directory_actor_event(&store, &first_fence, &child, Some("run-one"))
14827                .await
14828                .is_err()
14829        );
14830        assert!(
14831            validate_directory_actor_event(&store, &first_fence, &child, None)
14832                .await
14833                .is_err()
14834        );
14835        validate_directory_actor_event(&store, &second.fence(), &child, Some("run-two"))
14836            .await
14837            .unwrap();
14838    }
14839
14840    fn completed_actor_frame() -> ChildFrame {
14841        ChildFrame::Terminal {
14842            status: TerminalStatus::Completed,
14843            result: Some("done".to_string()),
14844            error: None,
14845            final_event_watermark: None,
14846            transcript: Vec::new(),
14847        }
14848    }
14849
14850    async fn drive_permission_handshake_frames(
14851        session_id: &str,
14852        frames: impl IntoIterator<Item = ChildFrame>,
14853        expected: ExpectedPermissionPosture,
14854    ) -> (
14855        crate::runtime::runner::Result<Option<String>>,
14856        Session,
14857        Vec<AgentEvent>,
14858        Vec<ParentFrame>,
14859    ) {
14860        let mut link = ConfirmationSequenceLink {
14861            frames: frames.into_iter().collect(),
14862            sent: Vec::new(),
14863        };
14864        let (event_tx, mut event_rx) = mpsc::channel(128);
14865        let cancel = CancellationToken::new();
14866        let (_live_tx, mut live_rx) = mpsc::unbounded_channel();
14867        let (_delivery_tx, mut delivery_rx) = mpsc::unbounded_channel();
14868        let mut session = Session::new(session_id, "model");
14869        let result = drive(ActorDriveContext {
14870            expected_creation: None,
14871            client: &mut link,
14872            parent_session_id: "permission-parent",
14873            child_session_id: session_id,
14874            child_attempt: 0,
14875            approval_registry: None,
14876            approval_decider: None,
14877            approval_reviewer: None,
14878            escalation_bridge: None,
14879            event_tx: &event_tx,
14880            cancel_token: &cancel,
14881            live_rx: &mut live_rx,
14882            delivery_rx: &mut delivery_rx,
14883            logical_session: &mut session,
14884            expected_permission_posture: Some(expected),
14885            session_inbox_runtime: None,
14886            actor_directory_store: None,
14887            canonical_subagent_tool: None,
14888            ticket_service: None,
14889            activation_run_id: None,
14890            execution_epoch: 0,
14891            expected_source_actor_id: session_id,
14892            initial_inflight_claims: VecDeque::new(),
14893            plain_actor: false,
14894            remote_environment_lease: false,
14895            readonly_output: None,
14896            local_history_tools: None,
14897            local_history_read_only: false,
14898            strict_history_output: None,
14899            plain_input: None,
14900            canonical_activation: None,
14901            canonical_placement_ref: None,
14902            actor_event_observer: None,
14903            plain_run: None,
14904            first_frame_timeout: Some(Duration::from_secs(1)),
14905        })
14906        .await;
14907        let events = std::iter::from_fn(|| event_rx.try_recv().ok()).collect();
14908        (result, session, events, link.sent)
14909    }
14910
14911    #[tokio::test]
14912    async fn actor_permission_handshake_rejects_terminal_without_posture() {
14913        let session_id = "permission-missing";
14914        let (result, session, events, _) = drive_permission_handshake_frames(
14915            session_id,
14916            [completed_actor_frame()],
14917            expected_default_permission_posture(7),
14918        )
14919        .await;
14920
14921        assert!(result
14922            .unwrap_err()
14923            .to_string()
14924            .contains("terminated before permission posture confirmation"));
14925        assert!(events.is_empty());
14926        assert!(bamboo_domain::PermissionAuditSnapshot::from_metadata(&session.metadata).is_none());
14927    }
14928
14929    #[tokio::test]
14930    async fn remote_environment_rejection_accepts_only_allowlisted_durable_terminal() {
14931        struct TerminalProofProbe {
14932            frame: Option<ChildFrame>,
14933            accepted: bool,
14934        }
14935
14936        #[async_trait]
14937        impl bamboo_subagent::ChildLink for TerminalProofProbe {
14938            async fn send(&mut self, _frame: ParentFrame) -> bamboo_subagent::TransportResult<()> {
14939                Ok(())
14940            }
14941
14942            async fn next_frame(&mut self) -> bamboo_subagent::TransportResult<Option<ChildFrame>> {
14943                Ok(self.frame.take())
14944            }
14945
14946            fn accept_durable_terminal(&mut self, status: TerminalStatus) {
14947                self.accepted = status == TerminalStatus::Error;
14948            }
14949
14950            fn has_pending_durable_terminal(&self) -> bool {
14951                self.accepted
14952            }
14953        }
14954
14955        for (status, error, expected_proof) in [
14956            (
14957                TerminalStatus::Error,
14958                Some("remote_environment_checkout_not_clean"),
14959                true,
14960            ),
14961            (TerminalStatus::Error, Some("untrusted worker error"), false),
14962            (
14963                TerminalStatus::Cancelled,
14964                Some("remote_environment_checkout_not_clean"),
14965                false,
14966            ),
14967        ] {
14968            let mut link = TerminalProofProbe {
14969                frame: Some(ChildFrame::Terminal {
14970                    status,
14971                    result: None,
14972                    error: error.map(str::to_owned),
14973                    final_event_watermark: None,
14974                    transcript: Vec::new(),
14975                }),
14976                accepted: false,
14977            };
14978            let (event_tx, _event_rx) = mpsc::channel(8);
14979            let cancel = CancellationToken::new();
14980            let (_live_tx, mut live_rx) = mpsc::unbounded_channel();
14981            let (_delivery_tx, mut delivery_rx) = mpsc::unbounded_channel();
14982            let mut session = Session::new("remote-terminal-proof", "model");
14983            let result = drive(ActorDriveContext {
14984                client: &mut link,
14985                parent_session_id: "proof-parent",
14986                child_session_id: "remote-terminal-proof",
14987                child_attempt: 0,
14988                approval_registry: None,
14989                approval_decider: None,
14990                approval_reviewer: None,
14991                escalation_bridge: None,
14992                event_tx: &event_tx,
14993                cancel_token: &cancel,
14994                live_rx: &mut live_rx,
14995                delivery_rx: &mut delivery_rx,
14996                logical_session: &mut session,
14997                expected_permission_posture: Some(expected_default_permission_posture(7)),
14998                expected_creation: None,
14999                session_inbox_runtime: None,
15000                actor_directory_store: None,
15001                canonical_subagent_tool: None,
15002                ticket_service: None,
15003                activation_run_id: None,
15004                execution_epoch: 0,
15005                expected_source_actor_id: "remote-terminal-proof",
15006                initial_inflight_claims: VecDeque::new(),
15007                plain_actor: false,
15008                remote_environment_lease: true,
15009                readonly_output: None,
15010                local_history_tools: None,
15011                local_history_read_only: false,
15012                strict_history_output: None,
15013                plain_input: None,
15014                canonical_activation: None,
15015                canonical_placement_ref: None,
15016                actor_event_observer: None,
15017                plain_run: None,
15018                first_frame_timeout: Some(Duration::from_secs(1)),
15019            })
15020            .await;
15021            assert!(result.is_err());
15022            assert_eq!(
15023                bamboo_subagent::ChildLink::has_pending_durable_terminal(&link),
15024                expected_proof
15025            );
15026        }
15027    }
15028
15029    #[tokio::test(start_paused = true)]
15030    async fn remote_cancel_requires_selected_worker_terminal_for_release_proof() {
15031        struct CancelOutcomeLink {
15032            cancel_sent: bool,
15033            settle: bool,
15034            accepted: bool,
15035        }
15036
15037        #[async_trait]
15038        impl bamboo_subagent::ChildLink for CancelOutcomeLink {
15039            async fn send(&mut self, frame: ParentFrame) -> bamboo_subagent::TransportResult<()> {
15040                self.cancel_sent = matches!(frame, ParentFrame::Cancel);
15041                Ok(())
15042            }
15043
15044            async fn next_frame(&mut self) -> bamboo_subagent::TransportResult<Option<ChildFrame>> {
15045                if !self.cancel_sent || !self.settle {
15046                    std::future::pending::<bamboo_subagent::TransportResult<Option<ChildFrame>>>()
15047                        .await
15048                } else {
15049                    self.cancel_sent = false;
15050                    Ok(Some(ChildFrame::Terminal {
15051                        status: TerminalStatus::Cancelled,
15052                        result: None,
15053                        error: None,
15054                        final_event_watermark: None,
15055                        transcript: Vec::new(),
15056                    }))
15057                }
15058            }
15059
15060            fn accept_durable_terminal(&mut self, status: TerminalStatus) {
15061                self.accepted = status == TerminalStatus::Cancelled;
15062            }
15063
15064            fn has_pending_durable_terminal(&self) -> bool {
15065                self.accepted
15066            }
15067        }
15068
15069        let temp = tempfile::tempdir().unwrap();
15070        let store = bamboo_storage::SessionStoreV2::new(temp.path().into())
15071            .await
15072            .unwrap();
15073        let fence = ActorActivationFence {
15074            schema_version: 2,
15075            actor_id: "cancelled-remote-child".into(),
15076            activation_id: "cancelled-remote-activation".into(),
15077            attempt: 1,
15078            run_id: "cancelled-remote-run".into(),
15079            lease_owner: "cancelled-remote-host".into(),
15080            lease_epoch: 1,
15081        };
15082        for settle in [true, false] {
15083            let mut link = CancelOutcomeLink {
15084                cancel_sent: false,
15085                settle,
15086                accepted: false,
15087            };
15088            let (event_tx, _event_rx) = mpsc::channel(8);
15089            let cancel = CancellationToken::new();
15090            cancel.cancel();
15091            let (_live_tx, mut live_rx) = mpsc::unbounded_channel();
15092            let (_delivery_tx, mut delivery_rx) = mpsc::unbounded_channel();
15093            let mut session = Session::new("cancelled-remote-child", "model");
15094            let result = drive(ActorDriveContext {
15095                expected_creation: None,
15096                client: &mut link,
15097                parent_session_id: "cancelled-remote-parent",
15098                child_session_id: "cancelled-remote-child",
15099                child_attempt: 1,
15100                approval_registry: None,
15101                approval_decider: None,
15102                approval_reviewer: None,
15103                escalation_bridge: None,
15104                event_tx: &event_tx,
15105                cancel_token: &cancel,
15106                live_rx: &mut live_rx,
15107                delivery_rx: &mut delivery_rx,
15108                logical_session: &mut session,
15109                expected_permission_posture: Some(expected_default_permission_posture(7)),
15110                session_inbox_runtime: None,
15111                actor_directory_store: Some(&store),
15112                canonical_subagent_tool: None,
15113                ticket_service: None,
15114                activation_run_id: Some("cancelled-remote-run"),
15115                execution_epoch: 0,
15116                expected_source_actor_id: "cancelled-remote-worker",
15117                initial_inflight_claims: VecDeque::new(),
15118                plain_actor: false,
15119                remote_environment_lease: true,
15120                readonly_output: None,
15121                local_history_tools: None,
15122                local_history_read_only: false,
15123                strict_history_output: None,
15124                plain_input: None,
15125                canonical_activation: Some(&fence),
15126                canonical_placement_ref: None,
15127                actor_event_observer: None,
15128                plain_run: None,
15129                first_frame_timeout: None,
15130            })
15131            .await;
15132            assert!(matches!(result, Err(AgentError::Cancelled)));
15133            assert_eq!(
15134                link.accepted, settle,
15135                "only a consumed exact terminal can release the remote fence"
15136            );
15137            assert_eq!(
15138                bamboo_subagent::ChildLink::has_pending_durable_terminal(&link),
15139                settle
15140            );
15141        }
15142    }
15143
15144    #[tokio::test]
15145    async fn actor_permission_handshake_rejects_malformed_agent_event() {
15146        let session_id = "permission-malformed";
15147        let (result, session, events, _) = drive_permission_handshake_frames(
15148            session_id,
15149            [ChildFrame::Event {
15150                event: serde_json::json!({"type": "token", "content": 42}),
15151            }],
15152            expected_default_permission_posture(7),
15153        )
15154        .await;
15155
15156        assert!(result
15157            .unwrap_err()
15158            .to_string()
15159            .contains("malformed AgentEvent"));
15160        assert!(events.is_empty());
15161        assert!(bamboo_domain::PermissionAuditSnapshot::from_metadata(&session.metadata).is_none());
15162
15163        let (result, _, events, _) = drive_permission_handshake_frames(
15164            session_id,
15165            [
15166                permission_posture_frame(session_id, 7),
15167                ChildFrame::Event {
15168                    event: serde_json::json!({"type":"private-download-secret"}),
15169                },
15170            ],
15171            expected_default_permission_posture(7),
15172        )
15173        .await;
15174        assert!(!result
15175            .unwrap_err()
15176            .to_string()
15177            .contains("private-download-secret"));
15178        assert_eq!(
15179            events.len(),
15180            1,
15181            "only the confirmed posture may be forwarded"
15182        );
15183    }
15184
15185    #[tokio::test]
15186    async fn actor_permission_handshake_rejects_execution_event_before_posture() {
15187        let session_id = "permission-early-event";
15188        let early_events = [
15189            (
15190                "progress",
15191                AgentEvent::RunnerProgress {
15192                    session_id: session_id.to_string(),
15193                    round_count: 1,
15194                },
15195            ),
15196            (
15197                "token",
15198                AgentEvent::Token {
15199                    content: "must-not-forward".to_string(),
15200                },
15201            ),
15202            (
15203                "tool",
15204                AgentEvent::ToolStart {
15205                    tool_call_id: "early-tool".to_string(),
15206                    tool_name: "Read".to_string(),
15207                    arguments: serde_json::json!({"file_path": "README.md"}),
15208                },
15209            ),
15210        ];
15211        for (kind, event) in early_events {
15212            let (result, session, events, _) = drive_permission_handshake_frames(
15213                session_id,
15214                [
15215                    actor_event_frame(event),
15216                    permission_posture_frame(session_id, 7),
15217                    completed_actor_frame(),
15218                ],
15219                expected_default_permission_posture(7),
15220            )
15221            .await;
15222
15223            assert!(
15224                result
15225                    .unwrap_err()
15226                    .to_string()
15227                    .contains("execution event before permission posture confirmation"),
15228                "{kind} must fail closed before posture"
15229            );
15230            assert!(events.is_empty(), "{kind} must not be forwarded");
15231            assert!(
15232                bamboo_domain::PermissionAuditSnapshot::from_metadata(&session.metadata).is_none(),
15233                "{kind} must not advance the permission audit"
15234            );
15235        }
15236    }
15237
15238    #[tokio::test]
15239    async fn actor_permission_handshake_rejects_approval_before_posture() {
15240        let session_id = "permission-early-approval";
15241        let (result, session, events, sent) = drive_permission_handshake_frames(
15242            session_id,
15243            [ChildFrame::ApprovalRequest {
15244                id: "approval-before-posture".to_string(),
15245                body: serde_json::json!({
15246                    "tool_name": "Bash",
15247                    "permission": "execute",
15248                    "resource": "echo must-not-run"
15249                }),
15250            }],
15251            expected_default_permission_posture(7),
15252        )
15253        .await;
15254
15255        assert!(result
15256            .unwrap_err()
15257            .to_string()
15258            .contains("requested approval before permission posture confirmation"));
15259        assert!(events.is_empty());
15260        assert!(
15261            sent.is_empty(),
15262            "an unconfirmed actor must receive no approval reply"
15263        );
15264        assert!(bamboo_domain::PermissionAuditSnapshot::from_metadata(&session.metadata).is_none());
15265    }
15266
15267    #[tokio::test]
15268    async fn actor_permission_handshake_rejects_mismatched_posture() {
15269        let session_id = "permission-mismatch";
15270        let (result, session, events, _) = drive_permission_handshake_frames(
15271            session_id,
15272            [permission_posture_frame(session_id, 8)],
15273            expected_default_permission_posture(7),
15274        )
15275        .await;
15276
15277        assert!(result
15278            .unwrap_err()
15279            .to_string()
15280            .contains("does not match the host-dispatched policy"));
15281        assert!(events.is_empty());
15282        assert!(bamboo_domain::PermissionAuditSnapshot::from_metadata(&session.metadata).is_none());
15283    }
15284
15285    #[tokio::test]
15286    async fn actor_permission_handshake_rejects_untrusted_executor_mapping() {
15287        let session_id = "permission-hostile-mapping";
15288        for hostile_mapping in [
15289            "wrong_executor:permission_mode=default",
15290            "test_actor:permission_mode=default;credential=must-not-persist",
15291        ] {
15292            let frame = ChildFrame::Event {
15293                event: serde_json::to_value(AgentEvent::PermissionPostureActivated {
15294                    session_id: session_id.to_string(),
15295                    policy_revision: 7,
15296                    requested_mode: "default".to_string(),
15297                    effective_mode: "default".to_string(),
15298                    executor_mapping: hostile_mapping.to_string(),
15299                })
15300                .expect("serialize hostile posture fixture"),
15301            };
15302            let (result, session, events, _) = drive_permission_handshake_frames(
15303                session_id,
15304                [frame],
15305                expected_default_permission_posture(7),
15306            )
15307            .await;
15308
15309            let error = result.unwrap_err().to_string();
15310            assert!(error.contains("host-dispatched executor mapping"));
15311            assert!(
15312                !error.contains(hostile_mapping),
15313                "untrusted mapping must not be reflected in host errors"
15314            );
15315            assert!(events.is_empty());
15316            assert!(
15317                bamboo_domain::PermissionAuditSnapshot::from_metadata(&session.metadata).is_none(),
15318                "untrusted mapping must not reach durable or in-memory audit state"
15319            );
15320        }
15321    }
15322
15323    #[tokio::test]
15324    async fn actor_permission_handshake_rejects_duplicate_posture() {
15325        let session_id = "permission-duplicate";
15326        let (result, session, events, _) = drive_permission_handshake_frames(
15327            session_id,
15328            [
15329                permission_posture_frame(session_id, 7),
15330                permission_posture_frame(session_id, 7),
15331                completed_actor_frame(),
15332            ],
15333            expected_default_permission_posture(7),
15334        )
15335        .await;
15336
15337        assert!(result
15338            .unwrap_err()
15339            .to_string()
15340            .contains("duplicate permission posture activation"));
15341        assert_eq!(
15342            events.len(),
15343            1,
15344            "only the confirmed posture may be forwarded"
15345        );
15346        assert!(matches!(
15347            events[0],
15348            AgentEvent::PermissionPostureActivated { .. }
15349        ));
15350        let audit = bamboo_domain::PermissionAuditSnapshot::from_metadata(&session.metadata)
15351            .expect("the first matching posture must be recorded");
15352        assert_eq!(audit.policy_revision, 7);
15353    }
15354
15355    #[tokio::test]
15356    async fn actor_permission_handshake_happy_path_persists_before_forwarding_execution() {
15357        let session_id = "permission-happy";
15358        let (result, session, events, _) = drive_permission_handshake_frames(
15359            session_id,
15360            [
15361                permission_posture_frame(session_id, 7),
15362                actor_event_frame(AgentEvent::RunnerProgress {
15363                    session_id: session_id.to_string(),
15364                    round_count: 1,
15365                }),
15366                actor_event_frame(AgentEvent::Token {
15367                    content: "working".to_string(),
15368                }),
15369                actor_event_frame(AgentEvent::ToolStart {
15370                    tool_call_id: "tool-1".to_string(),
15371                    tool_name: "Read".to_string(),
15372                    arguments: serde_json::json!({"file_path": "README.md"}),
15373                }),
15374                completed_actor_frame(),
15375            ],
15376            expected_default_permission_posture(7),
15377        )
15378        .await;
15379
15380        assert_eq!(result.unwrap().as_deref(), Some("done"));
15381        assert!(matches!(
15382            events.as_slice(),
15383            [
15384                AgentEvent::PermissionPostureActivated { .. },
15385                AgentEvent::RunnerProgress { .. },
15386                AgentEvent::Token { .. },
15387                AgentEvent::ToolStart { .. }
15388            ]
15389        ));
15390        let audit = bamboo_domain::PermissionAuditSnapshot::from_metadata(&session.metadata)
15391            .expect("matching posture must be recorded before execution events are accepted");
15392        assert_eq!(audit.policy_revision, 7);
15393        assert_eq!(audit.executor_mapping, "test_actor:permission_mode=default");
15394    }
15395
15396    #[tokio::test]
15397    async fn actor_download_tool_start_is_projected_before_parent_forwarding() {
15398        for (index, tool_name) in ["browser", "default::browser", "private-namespace::browser"]
15399            .into_iter()
15400            .enumerate()
15401        {
15402            let session_id = format!("actor-download-{index}");
15403            let args = serde_json::json!({
15404                "action":"download",
15405                "selector":"a[data-secret='private-selector']",
15406                "expected_epoch":17,
15407                "data_base64":"private-extra-bytes",
15408                "extra":{"url":"https://private.test/file"},
15409            });
15410            let original = args.clone();
15411            let read_args = serde_json::json!({"file_path":"README.md"});
15412            let (result, _session, events, _) = drive_permission_handshake_frames(
15413                &session_id,
15414                [
15415                    permission_posture_frame(&session_id, 7),
15416                    actor_event_frame(AgentEvent::ToolStart {
15417                        tool_call_id: "download-call".to_string(),
15418                        tool_name: tool_name.to_string(),
15419                        arguments: args.clone(),
15420                    }),
15421                    actor_event_frame(AgentEvent::ToolStart {
15422                        tool_call_id: "read-call".to_string(),
15423                        tool_name: "Read".to_string(),
15424                        arguments: read_args.clone(),
15425                    }),
15426                    completed_actor_frame(),
15427                ],
15428                expected_default_permission_posture(7),
15429            )
15430            .await;
15431            assert_eq!(result.expect("actor completed").as_deref(), Some("done"));
15432            assert_eq!(args, original);
15433            let starts: Vec<_> = events
15434                .iter()
15435                .filter_map(|event| match event {
15436                    AgentEvent::ToolStart {
15437                        tool_name,
15438                        arguments,
15439                        ..
15440                    } => Some((tool_name.as_str(), arguments)),
15441                    _ => None,
15442                })
15443                .collect();
15444            assert_eq!(starts.len(), 2);
15445            assert_eq!(starts[0].0, "browser");
15446            assert_eq!(
15447                starts[0].1,
15448                &serde_json::json!({"action":"download","expected_epoch":17})
15449            );
15450            assert_eq!(starts[1], ("Read", &read_args));
15451            assert!(!format!("{events:?}").contains("private"));
15452        }
15453    }
15454
15455    #[tokio::test]
15456    async fn actor_hook_lifecycle_hides_every_decision_text_across_event_batch_and_nested_stream() {
15457        let session_id = "actor-hook-display";
15458        let private = "private-selector-url-filename-response";
15459        let hook_name = format!("audit-{private}");
15460        let hook = |duration_ms, phase: &str, decision| AgentEvent::HookLifecycle {
15461            hook_name: hook_name.clone(),
15462            point: bamboo_domain::AgentHookPoint::BeforeToolExecution,
15463            phase: phase.into(),
15464            duration_ms,
15465            decision,
15466        };
15467        let decisions = [
15468            HookResult::Deny {
15469                reason: private.into(),
15470            },
15471            HookResult::InjectContext {
15472                text: private.into(),
15473            },
15474            HookResult::WithContext {
15475                result: Box::new(HookResult::WithContext {
15476                    result: Box::new(HookResult::Deny {
15477                        reason: private.into(),
15478                    }),
15479                    text: private.into(),
15480                }),
15481                text: private.into(),
15482            },
15483            HookResult::Suspend {
15484                reason: private.into(),
15485            },
15486            HookResult::Abort {
15487                reason: private.into(),
15488            },
15489        ];
15490        let originals = decisions.to_vec();
15491        let batch_events = [
15492            hook(2, "completed", decisions[1].clone()),
15493            hook(3, "completed", decisions[2].clone()),
15494            hook(4, "completed", decisions[3].clone()),
15495        ]
15496        .into_iter()
15497        .map(|event| serde_json::to_value(event).unwrap())
15498        .collect::<Vec<_>>();
15499        let batch = ActorEventBatch {
15500            logical_session: Some(LogicalSessionIdentity {
15501                creation: None,
15502                session_id: session_id.into(),
15503                parent_session_id: Some("permission-parent".into()),
15504                root_session_id: session_id.into(),
15505            }),
15506            activation_id: None,
15507            execution_epoch: 0,
15508            source_node_id: None,
15509            source_actor_id: Some(session_id.into()),
15510            first_seq: 1,
15511            last_seq: batch_events.len() as u64,
15512            qos: bamboo_subagent::ActorEventQos::classify(&batch_events[0]),
15513            events: batch_events,
15514        };
15515        let nested = AgentEvent::SubAgentEvent {
15516            parent_session_id: "parent".into(),
15517            child_session_id: "child".into(),
15518            event: Box::new(hook(5, "completed", decisions[4].clone())),
15519        };
15520        let raw = serde_json::to_string(&originals).unwrap();
15521        assert!(raw.contains(private));
15522
15523        let (outcome, _session, forwarded, _) = drive_permission_handshake_frames(
15524            session_id,
15525            [
15526                permission_posture_frame(session_id, 7),
15527                actor_event_frame(hook(1, private, decisions[0].clone())),
15528                ChildFrame::EventBatch { batch },
15529                actor_event_frame(nested),
15530                completed_actor_frame(),
15531            ],
15532            expected_default_permission_posture(7),
15533        )
15534        .await;
15535        assert_eq!(outcome.unwrap().as_deref(), Some("done"));
15536        let wire = serde_json::to_string(&forwarded).unwrap();
15537        assert!(
15538            !wire.contains(private),
15539            "private hook text reached SSE: {wire}"
15540        );
15541
15542        let hooks = forwarded
15543            .iter()
15544            .filter_map(|event| match event {
15545                AgentEvent::HookLifecycle {
15546                    hook_name,
15547                    point,
15548                    phase,
15549                    duration_ms,
15550                    decision,
15551                } => Some((hook_name, point, phase, duration_ms, decision)),
15552                _ => None,
15553            })
15554            .collect::<Vec<_>>();
15555        assert_eq!(hooks.len(), 4);
15556        assert!(hooks.iter().all(|(name, point, _, _, _)| {
15557            name == &&"actor-hook-1".to_string()
15558                && **point == bamboo_domain::AgentHookPoint::BeforeToolExecution
15559        }));
15560        assert_eq!(hooks[0].2, "event");
15561        assert_eq!(hooks[1].2, "completed");
15562        assert_eq!(*hooks[0].3, 1);
15563        assert_eq!(*hooks[3].3, 4);
15564        assert!(matches!(hooks[0].4, HookResult::Deny { reason } if reason == "Hook denied"));
15565        assert!(
15566            matches!(hooks[1].4, HookResult::InjectContext { text } if text == "Hook context hidden")
15567        );
15568        assert!(
15569            matches!(hooks[2].4, HookResult::WithContext { result, text }
15570            if text == "Hook context hidden" && matches!(result.as_ref(),
15571                HookResult::WithContext { result, text }
15572                    if text == "Hook context hidden" && matches!(result.as_ref(),
15573                        HookResult::Deny { reason } if reason == "Hook denied")))
15574        );
15575        assert!(matches!(hooks[3].4, HookResult::Suspend { reason } if reason == "Hook suspended"));
15576        assert!(forwarded.iter().any(|event| matches!(event,
15577            AgentEvent::SubAgentEvent { event, .. }
15578                if matches!(event.as_ref(), AgentEvent::HookLifecycle {
15579                    hook_name, duration_ms: 5,
15580                    decision: HookResult::Abort { reason }, ..
15581                } if hook_name == "actor-hook-1" && reason == "Hook aborted")
15582        )));
15583    }
15584
15585    #[test]
15586    fn actor_hook_display_depth_and_identity_limits_fail_closed() {
15587        let mut display = ActorEventDisplay::default();
15588        let mut deep = HookResult::Allow;
15589        for _ in 0..=MAX_DISPLAY_HOOK_RESULT_DEPTH {
15590            deep = HookResult::WithContext {
15591                result: Box::new(deep),
15592                text: "private-download-bytes".into(),
15593            };
15594        }
15595        let event = |hook_name: String, decision| AgentEvent::HookLifecycle {
15596            hook_name,
15597            point: bamboo_domain::AgentHookPoint::BeforeToolExecution,
15598            phase: "completed".into(),
15599            duration_ms: 7,
15600            decision,
15601        };
15602        assert!(display.project(event("deep".into(), deep), 0).is_none());
15603        assert!(display.overflowed);
15604        assert!(display
15605            .project(
15606                event(
15607                    "private-download-bytes".into(),
15608                    HookResult::Deny {
15609                        reason: "private-download-bytes".into(),
15610                    },
15611                ),
15612                0,
15613            )
15614            .is_none());
15615
15616        let mut oversized = ActorEventDisplay::default();
15617        assert!(oversized
15618            .project(
15619                event("x".repeat(MAX_DISPLAY_ID_BYTES + 1), HookResult::Allow),
15620                0,
15621            )
15622            .is_none());
15623        assert!(oversized.overflowed);
15624
15625        // A new actor connection gets a fresh opaque alias, never the raw
15626        // hook name. Repeated names keep the same alias until the shared
15627        // identity budget is exhausted, then the stream fails closed.
15628        let mut reconnected = ActorEventDisplay::default();
15629        let private = "private-download-bytes";
15630        for _ in 0..2 {
15631            let projected = reconnected
15632                .project(event(private.into(), HookResult::Allow), 0)
15633                .unwrap();
15634            assert!(
15635                matches!(projected, AgentEvent::HookLifecycle { hook_name, .. }
15636                if hook_name == "actor-hook-1")
15637            );
15638        }
15639        assert_eq!(reconnected.identities.load(Ordering::Relaxed), 1);
15640        for index in 1..MAX_DISPLAY_CALLS {
15641            assert!(reconnected
15642                .project(event(format!("hook-{index}"), HookResult::Allow), 0)
15643                .is_some());
15644        }
15645        assert!(reconnected
15646            .project(event("overflow-private".into(), HookResult::Allow), 0)
15647            .is_none());
15648        assert!(reconnected.overflowed);
15649        assert!(reconnected.hook_names.is_empty());
15650    }
15651
15652    #[test]
15653    fn actor_hook_display_preserves_nontext_decision_kinds() {
15654        for result in [
15655            HookResult::Continue,
15656            HookResult::Mutated,
15657            HookResult::Allow,
15658            HookResult::Ask,
15659        ] {
15660            assert_eq!(hook_result_for_display(result.clone(), 0), Some(result));
15661        }
15662    }
15663
15664    #[tokio::test]
15665    async fn malformed_actor_hook_events_never_fall_back_to_raw_display() {
15666        let session_id = "actor-malformed-hook";
15667        let private = "private-download-selector";
15668        let original = AgentEvent::HookLifecycle {
15669            hook_name: "audit".into(),
15670            point: bamboo_domain::AgentHookPoint::BeforeToolExecution,
15671            phase: "completed".into(),
15672            duration_ms: 9,
15673            decision: HookResult::Deny {
15674                reason: private.into(),
15675            },
15676        };
15677        let mut invalid = serde_json::to_value(&original).unwrap();
15678        invalid["decision"]["type"] = serde_json::json!("unknown_decision");
15679        let nested = AgentEvent::SubAgentEvent {
15680            parent_session_id: "parent".into(),
15681            child_session_id: "child".into(),
15682            event: Box::new(original),
15683        };
15684        let mut invalid_nested = serde_json::to_value(nested).unwrap();
15685        invalid_nested["event"]["decision"]["type"] = serde_json::json!("unknown_decision");
15686        let batch = ActorEventBatch {
15687            logical_session: Some(LogicalSessionIdentity {
15688                creation: None,
15689                session_id: session_id.into(),
15690                parent_session_id: Some("permission-parent".into()),
15691                root_session_id: session_id.into(),
15692            }),
15693            activation_id: None,
15694            execution_epoch: 0,
15695            source_node_id: None,
15696            source_actor_id: Some(session_id.into()),
15697            first_seq: 1,
15698            last_seq: 1,
15699            qos: bamboo_subagent::ActorEventQos::classify(&invalid),
15700            events: vec![invalid.clone()],
15701        };
15702        for frame in [
15703            ChildFrame::Event {
15704                event: invalid.clone(),
15705            },
15706            ChildFrame::EventBatch { batch },
15707            ChildFrame::Event {
15708                event: invalid_nested,
15709            },
15710        ] {
15711            let (outcome, _session, forwarded, _) = drive_permission_handshake_frames(
15712                session_id,
15713                [permission_posture_frame(session_id, 7), frame],
15714                expected_default_permission_posture(7),
15715            )
15716            .await;
15717            let error = outcome.unwrap_err().to_string();
15718            assert!(error.contains("malformed AgentEvent"), "{error}");
15719            assert!(!error.contains(private));
15720            assert!(!serde_json::to_string(&forwarded).unwrap().contains(private));
15721        }
15722
15723        let (event_tx, mut event_rx) = mpsc::channel(8);
15724        let mut session = Session::new("legacy-hook", "model");
15725        let mut handshake = PermissionPostureHandshake::NotRequired;
15726        let mut display = ActorEventDisplay::default();
15727        process_actor_event(
15728            invalid,
15729            false,
15730            &mut handshake,
15731            None,
15732            None,
15733            &mut session,
15734            &event_tx,
15735            &mut display,
15736        )
15737        .await
15738        .unwrap();
15739        assert!(display.overflowed);
15740        assert!(event_rx.try_recv().is_err());
15741    }
15742
15743    #[tokio::test]
15744    async fn actor_download_event_batch_hides_output_and_preserves_read_output() {
15745        let session_id = "actor-download-batch";
15746        let private = "private-download-bytes-and-url";
15747        let events = vec![
15748            AgentEvent::ToolStart {
15749                tool_call_id: "download-call".into(),
15750                tool_name: "default::browser".into(),
15751                arguments: serde_json::json!({
15752                    "action":"download", "selector": private, "expected_epoch": 17,
15753                }),
15754            },
15755            AgentEvent::ToolApprovalRequested {
15756                tool_call_id: "download-call".into(),
15757                tool_name: format!("{private}::browser"),
15758                parameters: serde_json::json!({
15759                    "action":"download", "selector": private, "expected_epoch": 17,
15760                }),
15761            },
15762            AgentEvent::ToolStart {
15763                tool_call_id: "read-call".into(),
15764                tool_name: "Read".into(),
15765                arguments: serde_json::json!({"file_path":"README.md"}),
15766            },
15767        ];
15768        let batch = ActorEventBatch {
15769            logical_session: Some(LogicalSessionIdentity {
15770                creation: None,
15771                session_id: session_id.into(),
15772                parent_session_id: Some("permission-parent".into()),
15773                root_session_id: session_id.into(),
15774            }),
15775            activation_id: None,
15776            execution_epoch: 0,
15777            source_node_id: None,
15778            source_actor_id: Some(session_id.into()),
15779            first_seq: 1,
15780            last_seq: events.len() as u64,
15781            qos: bamboo_subagent::ActorEventQos::Durable,
15782            events: events
15783                .into_iter()
15784                .map(|event| serde_json::to_value(event).unwrap())
15785                .collect(),
15786        };
15787        let result = bamboo_agent_core::tools::ToolResult {
15788            success: true,
15789            result: format!("{{\"data_base64\":\"{private}\"}}"),
15790            display_preference: Some(private.into()),
15791            images: vec![bamboo_agent_core::tools::ToolResultImage {
15792                mime_type: "image/png".into(),
15793                data: private.into(),
15794            }],
15795        };
15796        let (outcome, _, forwarded, _) = drive_permission_handshake_frames(
15797            session_id,
15798            [
15799                permission_posture_frame(session_id, 7),
15800                ChildFrame::EventBatch { batch },
15801                actor_event_frame(AgentEvent::ToolToken {
15802                    tool_call_id: "download-call".into(),
15803                    content: private.into(),
15804                }),
15805                actor_event_frame(AgentEvent::ToolToken {
15806                    tool_call_id: "read-call".into(),
15807                    content: "public live output".into(),
15808                }),
15809                actor_event_frame(AgentEvent::ToolComplete {
15810                    tool_call_id: "download-call".into(),
15811                    result,
15812                }),
15813                actor_event_frame(AgentEvent::ToolLifecycle {
15814                    tool_call_id: "download-call".into(),
15815                    tool_name: "browser".into(),
15816                    phase: "finished".into(),
15817                    elapsed_ms: Some(1),
15818                    is_mutating: false,
15819                    auto_approved: false,
15820                    summary: Some(private.into()),
15821                    error: Some(private.into()),
15822                }),
15823                actor_event_frame(AgentEvent::ToolError {
15824                    tool_call_id: "download-call".into(),
15825                    error: private.into(),
15826                }),
15827                actor_event_frame(AgentEvent::ToolComplete {
15828                    tool_call_id: "read-call".into(),
15829                    result: bamboo_agent_core::tools::ToolResult::text(true, "public README text"),
15830                }),
15831                completed_actor_frame(),
15832            ],
15833            expected_default_permission_posture(7),
15834        )
15835        .await;
15836        assert_eq!(outcome.unwrap().as_deref(), Some("done"));
15837        let wire = serde_json::to_string(&forwarded).unwrap();
15838        assert!(
15839            !wire.contains(private),
15840            "private content reached actor SSE: {wire}"
15841        );
15842        assert!(wire.contains("public README text"));
15843        assert!(wire.contains("public live output"));
15844        assert!(forwarded.iter().any(|event| matches!(event,
15845            AgentEvent::ToolComplete { tool_call_id, result }
15846                if tool_call_id == "download-call"
15847                && result.result == "Browser tool result hidden"
15848                && result.images.is_empty() && result.display_preference.is_none()
15849        )));
15850    }
15851
15852    #[tokio::test]
15853    async fn actor_missing_reused_and_nested_tool_identity_fails_closed() {
15854        let private = "private-download-content";
15855        let nested = |event| AgentEvent::SubAgentEvent {
15856            parent_session_id: "parent".into(),
15857            child_session_id: "nested-child".into(),
15858            event: Box::new(event),
15859        };
15860        let (outcome, _, forwarded, _) = drive_permission_handshake_frames(
15861            "actor-reorder",
15862            [
15863                permission_posture_frame("actor-reorder", 7),
15864                actor_event_frame(AgentEvent::ToolComplete {
15865                    tool_call_id: "missing".into(),
15866                    result: bamboo_agent_core::tools::ToolResult::text(true, private),
15867                }),
15868                actor_event_frame(AgentEvent::ToolLifecycle {
15869                    tool_call_id: "missing".into(),
15870                    tool_name: "browser".into(),
15871                    phase: private.into(),
15872                    elapsed_ms: None,
15873                    is_mutating: false,
15874                    auto_approved: false,
15875                    summary: Some(private.into()),
15876                    error: Some(private.into()),
15877                }),
15878                actor_event_frame(AgentEvent::ToolApprovalRequested {
15879                    tool_call_id: "approval-only".into(),
15880                    tool_name: "browser".into(),
15881                    parameters: serde_json::json!({"action":"download","selector":private}),
15882                }),
15883                actor_event_frame(AgentEvent::ToolToken {
15884                    tool_call_id: "approval-only".into(),
15885                    content: private.into(),
15886                }),
15887                actor_event_frame(AgentEvent::ToolStart {
15888                    tool_call_id: "malformed-browser".into(),
15889                    tool_name: format!("{private}::browser"),
15890                    arguments: serde_json::json!({"selector":private}),
15891                }),
15892                actor_event_frame(AgentEvent::ToolComplete {
15893                    tool_call_id: "malformed-browser".into(),
15894                    result: bamboo_agent_core::tools::ToolResult::text(true, private),
15895                }),
15896                actor_event_frame(AgentEvent::ToolStart {
15897                    tool_call_id: "reused".into(),
15898                    tool_name: "Read".into(),
15899                    arguments: serde_json::json!({"file_path":"README.md"}),
15900                }),
15901                actor_event_frame(AgentEvent::ToolStart {
15902                    tool_call_id: "reused".into(),
15903                    tool_name: "browser".into(),
15904                    arguments: serde_json::json!({"action":"download","selector":private}),
15905                }),
15906                actor_event_frame(AgentEvent::ToolToken {
15907                    tool_call_id: "reused".into(),
15908                    content: private.into(),
15909                }),
15910                actor_event_frame(nested(AgentEvent::ToolStart {
15911                    tool_call_id: "nested".into(),
15912                    tool_name: "browser".into(),
15913                    arguments: serde_json::json!({"action":"download","selector":private}),
15914                })),
15915                actor_event_frame(nested(AgentEvent::ToolComplete {
15916                    tool_call_id: "nested".into(),
15917                    result: bamboo_agent_core::tools::ToolResult::text(true, private),
15918                })),
15919                actor_event_frame(AgentEvent::ChildApprovalRequested {
15920                    child_session_id: "nested-child".into(),
15921                    request_id: "approval".into(),
15922                    tool_name: format!("{private}::browser"),
15923                    permission: private.into(),
15924                    resource: private.into(),
15925                }),
15926                actor_event_frame(AgentEvent::ChildApprovalChanged {
15927                    parent_session_id: "parent".into(),
15928                    child_session_id: "nested-child".into(),
15929                    child_attempt: 1,
15930                    request_id: "approval".into(),
15931                    version: 1,
15932                    status: "pending".into(),
15933                    reason: Some(private.into()),
15934                    tool_name: "browser".into(),
15935                    permission: private.into(),
15936                    resource: private.into(),
15937                    created_at: "2026-01-01T00:00:00Z".into(),
15938                    resolved_at: None,
15939                }),
15940                completed_actor_frame(),
15941            ],
15942            expected_default_permission_posture(7),
15943        )
15944        .await;
15945        assert_eq!(outcome.unwrap().as_deref(), Some("done"));
15946        let wire = serde_json::to_string(&forwarded).unwrap();
15947        assert!(
15948            !wire.contains(private),
15949            "private content reached actor SSE: {wire}"
15950        );
15951        assert!(wire.contains("Tool result hidden"));
15952        assert!(wire.contains("Browser tool result hidden"));
15953        assert!(wire.contains("Browser interaction approval"));
15954        assert!(!forwarded
15955            .iter()
15956            .any(|event| matches!(event, AgentEvent::ChildApprovalChanged { .. })));
15957    }
15958
15959    #[test]
15960    fn actor_display_tracking_overflow_never_reclassifies_old_private_id() {
15961        let mut display = ActorEventDisplay::default();
15962        display
15963            .project(
15964                AgentEvent::ToolStart {
15965                    tool_call_id: "private-call".into(),
15966                    tool_name: "browser".into(),
15967                    arguments: serde_json::json!({"action":"download","selector":"private"}),
15968                },
15969                0,
15970            )
15971            .unwrap();
15972        for index in 0..MAX_DISPLAY_CALLS {
15973            display
15974                .project(
15975                    AgentEvent::ToolStart {
15976                        tool_call_id: format!("ordinary-{index}"),
15977                        tool_name: "Read".into(),
15978                        arguments: serde_json::json!({"file_path":"README.md"}),
15979                    },
15980                    0,
15981                )
15982                .unwrap();
15983        }
15984        assert!(display.overflowed);
15985        let forwarded = display
15986            .project(
15987                AgentEvent::ToolComplete {
15988                    tool_call_id: "private-call".into(),
15989                    result: bamboo_agent_core::tools::ToolResult::text(true, "private-bytes"),
15990                },
15991                0,
15992            )
15993            .unwrap();
15994        assert!(!serde_json::to_string(&forwarded)
15995            .unwrap()
15996            .contains("private-bytes"));
15997    }
15998
15999    #[test]
16000    fn nested_actor_display_uses_one_identity_budget_and_bounded_keys() {
16001        let mut display = ActorEventDisplay::default();
16002        let nested = |index: usize| AgentEvent::SubAgentEvent {
16003            parent_session_id: "parent".into(),
16004            child_session_id: format!("child-{index}"),
16005            event: Box::new(AgentEvent::ToolStart {
16006                tool_call_id: format!("read-{index}"),
16007                tool_name: "Read".into(),
16008                arguments: serde_json::json!({"file_path":"README.md"}),
16009            }),
16010        };
16011        for index in 0..(MAX_DISPLAY_CALLS / 2) {
16012            assert!(display.project(nested(index), 0).is_some());
16013        }
16014        assert_eq!(
16015            display.identities.load(Ordering::Relaxed),
16016            MAX_DISPLAY_CALLS
16017        );
16018        assert!(display.project(nested(MAX_DISPLAY_CALLS / 2), 0).is_none());
16019        assert!(display.overflowed);
16020        assert!(display.children.is_empty());
16021
16022        let mut oversized = ActorEventDisplay::default();
16023        assert!(oversized
16024            .project(
16025                AgentEvent::SubAgentEvent {
16026                    parent_session_id: "parent".into(),
16027                    child_session_id: "x".repeat(MAX_DISPLAY_ID_BYTES + 1),
16028                    event: Box::new(AgentEvent::ToolComplete {
16029                        tool_call_id: "call".into(),
16030                        result: bamboo_agent_core::tools::ToolResult::text(
16031                            true,
16032                            "private-download-bytes",
16033                        ),
16034                    }),
16035                },
16036                0,
16037            )
16038            .is_none());
16039        assert!(oversized.overflowed);
16040        assert!(oversized
16041            .project(
16042                AgentEvent::ToolStart {
16043                    tool_call_id: "x".repeat(MAX_DISPLAY_ID_BYTES + 1),
16044                    tool_name: "Read".into(),
16045                    arguments: serde_json::json!({"file_path":"README.md"}),
16046                },
16047                0,
16048            )
16049            .is_none());
16050    }
16051
16052    #[tokio::test]
16053    async fn actor_batch_sequence_gap_invalidates_prior_tool_identity() {
16054        let session_id = "actor-gap";
16055        let batch = |seq, event: AgentEvent| {
16056            let event = serde_json::to_value(event).unwrap();
16057            ChildFrame::EventBatch {
16058                batch: ActorEventBatch {
16059                    logical_session: Some(LogicalSessionIdentity {
16060                        creation: None,
16061                        session_id: session_id.into(),
16062                        parent_session_id: Some("permission-parent".into()),
16063                        root_session_id: session_id.into(),
16064                    }),
16065                    activation_id: None,
16066                    execution_epoch: 0,
16067                    source_node_id: None,
16068                    source_actor_id: Some(session_id.into()),
16069                    first_seq: seq,
16070                    last_seq: seq,
16071                    qos: bamboo_subagent::ActorEventQos::classify(&event),
16072                    events: vec![event],
16073                },
16074            }
16075        };
16076        let (outcome, _, forwarded, _) = drive_permission_handshake_frames(
16077            session_id,
16078            [
16079                permission_posture_frame(session_id, 7),
16080                batch(
16081                    1,
16082                    AgentEvent::ToolStart {
16083                        tool_call_id: "reused-call".into(),
16084                        tool_name: "Read".into(),
16085                        arguments: serde_json::json!({"file_path":"README.md"}),
16086                    },
16087                ),
16088                // Sequence 2 may have replaced this ID with a private browser
16089                // call. The actor's next result must not inherit Read display.
16090                batch(
16091                    3,
16092                    AgentEvent::ToolComplete {
16093                        tool_call_id: "reused-call".into(),
16094                        result: bamboo_agent_core::tools::ToolResult::text(
16095                            true,
16096                            "private-download-bytes",
16097                        ),
16098                    },
16099                ),
16100                completed_actor_frame(),
16101            ],
16102            expected_default_permission_posture(7),
16103        )
16104        .await;
16105        assert_eq!(outcome.unwrap().as_deref(), Some("done"));
16106        let wire = serde_json::to_string(&forwarded).unwrap();
16107        assert!(!wire.contains("private-download-bytes"), "{wire}");
16108        assert!(wire.contains("Tool result hidden"));
16109    }
16110
16111    #[tokio::test]
16112    async fn legacy_malformed_actor_event_invalidates_prior_tool_identity() {
16113        let (event_tx, mut event_rx) = mpsc::channel(8);
16114        let mut session = Session::new("legacy-malformed-private", "model");
16115        let mut handshake = PermissionPostureHandshake::NotRequired;
16116        let mut display = ActorEventDisplay::default();
16117        let read_start = serde_json::to_value(AgentEvent::ToolStart {
16118            tool_call_id: "reused-call".into(),
16119            tool_name: "Read".into(),
16120            arguments: serde_json::json!({"file_path":"README.md"}),
16121        })
16122        .unwrap();
16123        process_actor_event(
16124            read_start,
16125            false,
16126            &mut handshake,
16127            None,
16128            None,
16129            &mut session,
16130            &event_tx,
16131            &mut display,
16132        )
16133        .await
16134        .unwrap();
16135
16136        // A mixed-version actor may send a browser ToolStart whose name has
16137        // the wrong wire type. Serde drops it; the reused ID is no longer safe.
16138        let mut malformed = serde_json::to_value(AgentEvent::ToolStart {
16139            tool_call_id: "reused-call".into(),
16140            tool_name: "browser".into(),
16141            arguments: serde_json::json!({
16142                "action":"download", "selector":"private-download-selector"
16143            }),
16144        })
16145        .unwrap();
16146        malformed["tool_name"] = serde_json::json!(7);
16147        process_actor_event(
16148            malformed,
16149            false,
16150            &mut handshake,
16151            None,
16152            None,
16153            &mut session,
16154            &event_tx,
16155            &mut display,
16156        )
16157        .await
16158        .unwrap();
16159        process_actor_event(
16160            serde_json::to_value(AgentEvent::ToolComplete {
16161                tool_call_id: "reused-call".into(),
16162                result: bamboo_agent_core::tools::ToolResult::text(true, "private-download-bytes"),
16163            })
16164            .unwrap(),
16165            false,
16166            &mut handshake,
16167            None,
16168            None,
16169            &mut session,
16170            &event_tx,
16171            &mut display,
16172        )
16173        .await
16174        .unwrap();
16175        let forwarded = std::iter::from_fn(|| event_rx.try_recv().ok()).collect::<Vec<_>>();
16176        assert_eq!(forwarded.len(), 2);
16177        let wire = serde_json::to_string(&forwarded).unwrap();
16178        assert!(!wire.contains("private-download"), "{wire}");
16179        assert!(wire.contains("Tool result hidden"));
16180    }
16181
16182    #[tokio::test]
16183    async fn actor_initial_batch_acks_in_order_and_rejects_stale_confirmation() {
16184        let temp = tempfile::tempdir().unwrap();
16185        let store = Arc::new(
16186            bamboo_storage::SessionStoreV2::new(temp.path().to_path_buf())
16187                .await
16188                .unwrap(),
16189        );
16190        let storage: Arc<dyn Storage> = store.clone();
16191        let locked = Arc::new(bamboo_storage::LockedSessionStore::new(storage.clone()));
16192        let inbox: Arc<dyn SessionInboxPort> = Arc::new(bamboo_storage::FileSessionInbox::new(
16193            store.clone(),
16194            bamboo_domain::SessionInboxLimits::default(),
16195        ));
16196        let session_id = "actor-confirmation-order";
16197        let run_id = "actor-run-current";
16198        let mut session = Session::new(session_id, "model");
16199        store.save_session(&session).await.unwrap();
16200        for (id, text) in [("actor-first", "first"), ("actor-second", "second")] {
16201            let mut envelope = bamboo_domain::SessionMessageEnvelope::user_input(session_id, text);
16202            envelope.id = bamboo_domain::SessionMessageId::parse(id).unwrap();
16203            inbox.deliver(&envelope).await.unwrap();
16204        }
16205        inbox
16206            .mark_activation_eligible(
16207                session_id,
16208                2,
16209                bamboo_domain::SessionActivationPolicy::InterruptSpecificWait,
16210            )
16211            .await
16212            .unwrap();
16213        let router = SessionActivationRouter::new();
16214        let mut owner_registration = router.register_run(session_id, run_id).await.unwrap();
16215        let binding = SessionInboxRuntimeBinding {
16216            router,
16217            inbox: inbox.clone(),
16218            storage,
16219            persistence: locked,
16220            parent_question_lock: None,
16221        };
16222        let pairs = claim_canonical_deliveries(&binding, &mut session, run_id, usize::MAX, None)
16223            .await
16224            .unwrap();
16225        assert_eq!(
16226            pairs
16227                .iter()
16228                .map(|(claim, _)| claim.envelope.id.as_str())
16229                .collect::<Vec<_>>(),
16230            vec!["actor-first", "actor-second"]
16231        );
16232        let seeded = store.load_session(session_id).await.unwrap().unwrap();
16233        assert_eq!(
16234            seeded
16235                .messages
16236                .iter()
16237                .filter(|message| matches!(message.id.as_str(), "actor-first" | "actor-second"))
16238                .map(|message| message.id.as_str())
16239                .collect::<Vec<_>>(),
16240            vec!["actor-first", "actor-second"],
16241            "host context must be durable before actor dispatch"
16242        );
16243        for (claim, _) in &pairs {
16244            assert_eq!(
16245                seeded
16246                    .messages
16247                    .iter()
16248                    .filter(|message| bamboo_domain::is_matching_session_message(
16249                        message,
16250                        &claim.envelope
16251                    ))
16252                    .count(),
16253                1,
16254                "pre-dispatch host checkpoint must contain exactly one canonical marker for {}",
16255                claim.envelope.id
16256            );
16257        }
16258        assert!(
16259            seeded.session_inbox_admission().is_none_or(|cursor| {
16260                !cursor.contains(&pairs[0].0.envelope.id)
16261                    && !cursor.contains(&pairs[1].0.envelope.id)
16262            }),
16263            "pre-dispatch transcript seeding must not forge worker confirmation"
16264        );
16265        assert_eq!(inbox.inspect(session_id).await.unwrap().claimed, 2);
16266        let claims = pairs
16267            .into_iter()
16268            .map(|(claim, _)| claim)
16269            .collect::<VecDeque<_>>();
16270        let first = claims[0].clone();
16271        let second = claims[1].clone();
16272        let mut stale = admission_confirmation(session_id, &first, "stale-run");
16273        stale.canonical_claim_generation = second.generation;
16274        let mut link = ConfirmationSequenceLink {
16275            frames: VecDeque::from([
16276                ChildFrame::SessionMessageAdmitted {
16277                    confirmation: stale,
16278                },
16279                ChildFrame::SessionMessageAdmitted {
16280                    confirmation: admission_confirmation(session_id, &first, run_id),
16281                },
16282                ChildFrame::SessionMessageAdmitted {
16283                    confirmation: admission_confirmation(session_id, &second, run_id),
16284                },
16285                ChildFrame::Terminal {
16286                    status: TerminalStatus::Completed,
16287                    result: Some("done".to_string()),
16288                    error: None,
16289                    final_event_watermark: None,
16290                    transcript: Vec::new(),
16291                },
16292            ]),
16293            sent: Vec::new(),
16294        };
16295        let (event_tx, _event_rx) = mpsc::channel(8);
16296        let cancel = CancellationToken::new();
16297        let (_live_tx, mut live_rx) = mpsc::unbounded_channel();
16298        let (_delivery_tx, mut delivery_rx) = mpsc::unbounded_channel();
16299        let result = drive(ActorDriveContext {
16300            expected_creation: None,
16301            client: &mut link,
16302            parent_session_id: "parent",
16303            child_session_id: session_id,
16304            child_attempt: 0,
16305            approval_registry: None,
16306            approval_decider: None,
16307            approval_reviewer: None,
16308            escalation_bridge: None,
16309            event_tx: &event_tx,
16310            cancel_token: &cancel,
16311            live_rx: &mut live_rx,
16312            delivery_rx: &mut delivery_rx,
16313            logical_session: &mut session,
16314            expected_permission_posture: None,
16315            session_inbox_runtime: Some(&binding),
16316            actor_directory_store: None,
16317            canonical_subagent_tool: None,
16318            ticket_service: None,
16319            activation_run_id: Some(run_id),
16320            execution_epoch: 0,
16321            expected_source_actor_id: session_id,
16322            initial_inflight_claims: claims,
16323            plain_actor: false,
16324            remote_environment_lease: false,
16325            readonly_output: None,
16326            local_history_tools: None,
16327            local_history_read_only: false,
16328            strict_history_output: None,
16329            plain_input: None,
16330            canonical_activation: None,
16331            canonical_placement_ref: None,
16332            actor_event_observer: None,
16333            plain_run: None,
16334            first_frame_timeout: Some(Duration::from_secs(1)),
16335        })
16336        .await
16337        .unwrap();
16338        assert_eq!(result.as_deref(), Some("done"));
16339        assert_eq!(
16340            session
16341                .messages
16342                .iter()
16343                .filter(|message| matches!(message.id.as_str(), "actor-first" | "actor-second"))
16344                .map(|message| message.id.as_str())
16345                .collect::<Vec<_>>(),
16346            vec!["actor-first", "actor-second"]
16347        );
16348        let backlog = inbox.inspect(session_id).await.unwrap();
16349        assert_eq!(backlog.pending + backlog.claimed, 0);
16350        let confirmed = store.load_session(session_id).await.unwrap().unwrap();
16351        let cursor = confirmed
16352            .session_inbox_admission()
16353            .expect("exact worker confirmation must checkpoint the admission cursor");
16354        assert!(cursor.contains(&first.envelope.id));
16355        assert!(cursor.contains(&second.envelope.id));
16356        for claim in [&first, &second] {
16357            assert_eq!(
16358                confirmed
16359                    .messages
16360                    .iter()
16361                    .filter(|message| bamboo_domain::is_matching_session_message(
16362                        message,
16363                        &claim.envelope
16364                    ))
16365                    .count(),
16366                1,
16367                "confirmation must retain one exact canonical marker for {}",
16368                claim.envelope.id
16369            );
16370        }
16371        assert!(inbox
16372            .was_admitted(session_id, &first.envelope.id)
16373            .await
16374            .unwrap());
16375        assert!(inbox
16376            .was_admitted(session_id, &second.envelope.id)
16377            .await
16378            .unwrap());
16379        owner_registration.begin_finalization().await;
16380        owner_registration.finish(2).await.unwrap();
16381    }
16382
16383    #[derive(Default)]
16384    struct RecordingCodexTokenAuthority {
16385        issued_for: std::sync::Mutex<Vec<String>>,
16386        revoked: std::sync::Mutex<Vec<String>>,
16387    }
16388
16389    impl CodexRunTokenAuthority for RecordingCodexTokenAuthority {
16390        fn issue(&self, session_id: &str) -> Result<IssuedCodexRunToken, String> {
16391            self.issued_for
16392                .lock()
16393                .expect("issued fixture lock")
16394                .push(session_id.to_string());
16395            Ok(IssuedCodexRunToken {
16396                token_id: format!("id-{session_id}"),
16397                token: format!("bcx1_secret-{session_id}"),
16398            })
16399        }
16400
16401        fn revoke(&self, token_id: &str) {
16402            self.revoked
16403                .lock()
16404                .expect("revoked fixture lock")
16405                .push(token_id.to_string());
16406        }
16407    }
16408
16409    fn codex_executor(auth_mode: Option<&str>, inherit_user_config: Option<bool>) -> ExecutorSpec {
16410        let bamboo_mode = auth_mode == Some("bamboo")
16411            || (auth_mode.is_none() && !inherit_user_config.unwrap_or(false));
16412        ExecutorSpec::Codex {
16413            binary: None,
16414            model: None,
16415            mode: None,
16416            sandbox: None,
16417            inherit_user_config,
16418            auth_mode: auth_mode.map(str::to_string),
16419            base_url: bamboo_mode.then(|| "http://127.0.0.1:9562/openai/v1".to_string()),
16420            wire_api: Some("responses".to_string()),
16421            provider_key_ref: None,
16422            forward_env: None,
16423            approval_policy: None,
16424            network_access: None,
16425            allow_danger_bypass: None,
16426            permission_profile: None,
16427            workspace_owned: None,
16428        }
16429    }
16430
16431    fn permission_resolution(
16432        requested: bamboo_domain::SessionPermissionMode,
16433        effective: bamboo_domain::PermissionMode,
16434    ) -> bamboo_domain::PermissionModeResolution {
16435        bamboo_domain::PermissionModeResolution {
16436            requested,
16437            effective,
16438        }
16439    }
16440
16441    #[test]
16442    fn permission_posture_mapping_contract_is_exact_for_supported_executors() {
16443        use bamboo_domain::{PermissionMode, SessionPermissionMode};
16444
16445        let default =
16446            permission_resolution(SessionPermissionMode::Default, PermissionMode::Default);
16447        assert_eq!(
16448            expected_permission_executor_mapping(&ExecutorSpec::BambooRuntime, default, false)
16449                .unwrap()
16450                .as_deref(),
16451            Some("bamboo_runtime:default")
16452        );
16453        assert_eq!(
16454            expected_permission_executor_mapping(&ExecutorSpec::Echo, default, false).unwrap(),
16455            None,
16456            "transport-only Echo must not claim the typed permission contract"
16457        );
16458        assert_eq!(
16459            expected_permission_executor_mapping(
16460                &ExecutorSpec::CliAdapter {
16461                    command: "must-not-appear-in-contract".to_string(),
16462                    args: vec!["credential-like-argument".to_string()],
16463                },
16464                default,
16465                false,
16466            )
16467            .unwrap(),
16468            None,
16469            "unimplemented CliAdapter must not leak command data into a contract"
16470        );
16471
16472        let claude = ExecutorSpec::ClaudeCode {
16473            binary: None,
16474            model: None,
16475            permission_mode: Some("default".to_string()),
16476            inherit_user_config: None,
16477            forward_env: None,
16478        };
16479        for (resolution, mapping) in [
16480            (
16481                permission_resolution(SessionPermissionMode::Default, PermissionMode::Plan),
16482                "claude_code:permission_mode=plan",
16483            ),
16484            (
16485                permission_resolution(SessionPermissionMode::Auto, PermissionMode::Auto),
16486                "claude_code:permission_mode=bypassPermissions",
16487            ),
16488            (
16489                permission_resolution(SessionPermissionMode::Default, PermissionMode::AcceptEdits),
16490                "claude_code:permission_mode=acceptEdits",
16491            ),
16492            (
16493                permission_resolution(SessionPermissionMode::Default, PermissionMode::DontAsk),
16494                "claude_code:permission_mode=dontAsk",
16495            ),
16496            (
16497                permission_resolution(
16498                    SessionPermissionMode::Bypass,
16499                    PermissionMode::BypassPermissions,
16500                ),
16501                "claude_code:permission_mode=default",
16502            ),
16503        ] {
16504            assert_eq!(
16505                expected_permission_executor_mapping(&claude, resolution, false)
16506                    .unwrap()
16507                    .as_deref(),
16508                Some(mapping)
16509            );
16510        }
16511        assert_eq!(
16512            expected_permission_executor_mapping(&claude, default, true)
16513                .unwrap()
16514                .as_deref(),
16515            Some("claude_code:blocked_explicit_deny")
16516        );
16517
16518        let mut codex_exec = codex_executor(Some("inherit"), Some(true));
16519        if let ExecutorSpec::Codex {
16520            approval_policy, ..
16521        } = &mut codex_exec
16522        {
16523            *approval_policy = Some("on-failure".to_string());
16524        }
16525        assert_eq!(
16526            expected_permission_executor_mapping(&codex_exec, default, false)
16527                .unwrap()
16528                .as_deref(),
16529            Some("codex_exec:approval_policy=on-failure")
16530        );
16531        assert_eq!(
16532            expected_permission_executor_mapping(
16533                &codex_exec,
16534                permission_resolution(SessionPermissionMode::Auto, PermissionMode::Auto),
16535                false,
16536            )
16537            .unwrap()
16538            .as_deref(),
16539            Some("codex_exec:approval_policy=never")
16540        );
16541        assert_eq!(
16542            expected_permission_executor_mapping(&codex_exec, default, true)
16543                .unwrap()
16544                .as_deref(),
16545            Some("codex_exec:blocked_explicit_deny")
16546        );
16547
16548        let mut codex_app_server = codex_executor(Some("inherit"), Some(true));
16549        if let ExecutorSpec::Codex {
16550            mode,
16551            approval_policy,
16552            ..
16553        } = &mut codex_app_server
16554        {
16555            *mode = Some("app_server".to_string());
16556            *approval_policy = Some("on-request".to_string());
16557        }
16558        assert_eq!(
16559            expected_permission_executor_mapping(&codex_app_server, default, false)
16560                .unwrap()
16561                .as_deref(),
16562            Some("codex_app_server:approvalPolicy=on-request")
16563        );
16564        assert_eq!(
16565            expected_permission_executor_mapping(
16566                &codex_app_server,
16567                permission_resolution(SessionPermissionMode::Auto, PermissionMode::Auto),
16568                false,
16569            )
16570            .unwrap()
16571            .as_deref(),
16572            Some("codex_app_server:approvalPolicy=never")
16573        );
16574        assert_eq!(
16575            expected_permission_executor_mapping(&codex_app_server, default, true)
16576                .unwrap()
16577                .as_deref(),
16578            Some("codex_app_server:blocked_explicit_deny")
16579        );
16580    }
16581
16582    #[test]
16583    fn only_bamboo_managed_non_git_workspaces_are_marked_owned() {
16584        let project = tempfile::tempdir().unwrap();
16585        let managed = project.path().join(".bamboo/worktree/child-571");
16586        std::fs::create_dir_all(&managed).unwrap();
16587        assert!(!workspace_is_bamboo_owned(managed.to_str().unwrap()));
16588        let marker = project
16589            .path()
16590            .join(".bamboo/worktree/.bamboo-owned/child-571");
16591        std::fs::create_dir_all(marker.parent().unwrap()).unwrap();
16592        std::fs::write(&marker, "bamboo/child-571").unwrap();
16593        assert!(workspace_is_bamboo_owned(managed.to_str().unwrap()));
16594        let nested = managed.join("nested/path");
16595        std::fs::create_dir_all(&nested).unwrap();
16596        assert!(workspace_is_bamboo_owned(nested.to_str().unwrap()));
16597
16598        let arbitrary = tempfile::tempdir().unwrap();
16599        assert!(!workspace_is_bamboo_owned(
16600            arbitrary.path().to_str().unwrap()
16601        ));
16602    }
16603
16604    #[test]
16605    fn bamboo_codex_token_is_per_run_redacted_and_revoked_on_guard_drop() {
16606        let authority = Arc::new(RecordingCodexTokenAuthority::default());
16607        let authority_dyn: Arc<dyn CodexRunTokenAuthority> = authority.clone();
16608
16609        let (secrets, guard) = build_codex_run_secrets(
16610            &codex_executor(Some("bamboo"), None),
16611            Some(authority_dyn),
16612            "child-570",
16613        )
16614        .unwrap();
16615
16616        let token = secrets
16617            .codex_provider_token
16618            .as_ref()
16619            .expect("bamboo mode mints a token");
16620        assert_eq!(token.expose(), "bcx1_secret-child-570");
16621        assert!(!format!("{token:?}").contains("secret-child-570"));
16622        assert_eq!(
16623            authority.issued_for.lock().unwrap().as_slice(),
16624            ["child-570"]
16625        );
16626        assert!(authority.revoked.lock().unwrap().is_empty());
16627
16628        drop(guard);
16629        assert_eq!(
16630            authority.revoked.lock().unwrap().as_slice(),
16631            ["id-child-570"]
16632        );
16633    }
16634
16635    #[test]
16636    fn non_bamboo_codex_never_mints_and_bamboo_fails_closed_without_authority() {
16637        let authority = Arc::new(RecordingCodexTokenAuthority::default());
16638        let authority_dyn: Arc<dyn CodexRunTokenAuthority> = authority.clone();
16639        let (secrets, guard) = build_codex_run_secrets(
16640            &codex_executor(Some("custom"), None),
16641            Some(authority_dyn),
16642            "child-custom",
16643        )
16644        .unwrap();
16645        assert!(secrets.codex_provider_token.is_none());
16646        assert!(guard.is_none());
16647        assert!(authority.issued_for.lock().unwrap().is_empty());
16648
16649        let error = build_codex_run_secrets(
16650            &codex_executor(Some("bamboo"), None),
16651            None,
16652            "child-no-authority",
16653        )
16654        .err()
16655        .expect("bamboo mode without an authority must fail closed");
16656        assert!(error.to_string().contains("per-run token authority"));
16657    }
16658
16659    #[test]
16660    fn codex_provisioning_never_leaks_the_session_provider_credential() {
16661        let credentials = vec![ScopedCredential {
16662            provider: "openai".to_string(),
16663            api_key: "upstream-secret-must-not-cross".to_string(),
16664            base_url: None,
16665            provider_type: Some("openai".to_string()),
16666            credential_ref: Some("provider.openai.api_key".to_string()),
16667        }];
16668
16669        for (mode, label) in [
16670            (Some("inherit"), "inherit"),
16671            (Some("api_key"), "api_key"),
16672            (Some("bamboo"), "bamboo"),
16673            (None, "default-bamboo"),
16674        ] {
16675            let runner = ActorChildRunner::new(
16676                format!("codex-{label}-test"),
16677                PathBuf::from("/bin/false"),
16678                Vec::new(),
16679                std::env::temp_dir().join(format!("bamboo-codex-{label}-570")),
16680                codex_executor(mode, None),
16681                credentials.clone(),
16682                "openai".to_string(),
16683                1,
16684            );
16685            let mut session = Session::new(format!("child-{label}"), "model");
16686            session.add_message(bamboo_agent_core::Message::user("test"));
16687            let spec = runner.build_spec(
16688                &session,
16689                &crate::runtime::execution::SpawnJob {
16690                    parent_session_id: "parent".to_string(),
16691                    child_session_id: format!("child-{label}"),
16692                    model: "gpt-5.4".to_string(),
16693                    disabled_tools: None,
16694                },
16695            );
16696            assert!(
16697                spec.secrets.provider_credentials.is_empty(),
16698                "{label} Codex must not receive the session provider key"
16699            );
16700        }
16701    }
16702
16703    #[test]
16704    fn non_codex_provisioning_still_receives_only_its_selected_provider_credential() {
16705        let credentials = vec![
16706            ScopedCredential {
16707                provider: "openai".to_string(),
16708                api_key: "selected-openai-secret".to_string(),
16709                base_url: None,
16710                provider_type: Some("openai".to_string()),
16711                credential_ref: Some("provider.openai.api_key".to_string()),
16712            },
16713            ScopedCredential {
16714                provider: "other".to_string(),
16715                api_key: "unrelated-secret".to_string(),
16716                base_url: None,
16717                provider_type: Some("openai".to_string()),
16718                credential_ref: Some("provider.other.api_key".to_string()),
16719            },
16720        ];
16721        let runner = ActorChildRunner::new(
16722            "echo-test".to_string(),
16723            PathBuf::from("/bin/false"),
16724            Vec::new(),
16725            std::env::temp_dir().join("bamboo-echo-provider-570"),
16726            ExecutorSpec::Echo,
16727            credentials,
16728            "openai".to_string(),
16729            1,
16730        );
16731        let spec = runner.build_spec(
16732            &Session::new("child-echo", "model"),
16733            &crate::runtime::execution::SpawnJob {
16734                parent_session_id: "parent".to_string(),
16735                child_session_id: "child-echo".to_string(),
16736                model: "gpt-5.4".to_string(),
16737                disabled_tools: None,
16738            },
16739        );
16740
16741        assert_eq!(spec.secrets.provider_credentials.len(), 1);
16742        assert_eq!(
16743            spec.secrets.provider_credentials[0].api_key,
16744            "selected-openai-secret"
16745        );
16746    }
16747
16748    #[tokio::test]
16749    async fn live_provider_config_refreshes_credentials_and_default_per_activation() {
16750        fn provider_instance(
16751            provider_type: &str,
16752            api_key: &str,
16753        ) -> bamboo_config::ProviderInstanceConfig {
16754            bamboo_config::ProviderInstanceConfig {
16755                provider_type: provider_type.to_string(),
16756                label: None,
16757                api_key: api_key.to_string(),
16758                api_key_encrypted: None,
16759                credential_ref: None,
16760                base_url: None,
16761                model: None,
16762                fast_model: None,
16763                vision_model: None,
16764                reasoning_effort: None,
16765                responses_only_models: Vec::new(),
16766                request_overrides: None,
16767                enabled: true,
16768                extra: Default::default(),
16769            }
16770        }
16771
16772        let mut initial = bamboo_llm::Config::default();
16773        initial.provider_instances.insert(
16774            "old-provider".to_string(),
16775            provider_instance("openai", "old-key"),
16776        );
16777        initial.default_provider_instance = Some("old-provider".to_string());
16778        let live = Arc::new(tokio::sync::RwLock::new(initial));
16779
16780        let runner = ActorChildRunner::new(
16781            "echo-live-config-test".to_string(),
16782            PathBuf::from("/bin/false"),
16783            Vec::new(),
16784            std::env::temp_dir().join("bamboo-echo-live-config"),
16785            ExecutorSpec::Echo,
16786            Vec::new(),
16787            "stale-provider".to_string(),
16788            1,
16789        )
16790        .with_live_provider_config(live.clone());
16791        let job = crate::runtime::execution::SpawnJob {
16792            parent_session_id: "parent".to_string(),
16793            child_session_id: "child".to_string(),
16794            model: "model".to_string(),
16795            disabled_tools: None,
16796        };
16797
16798        let first = runner
16799            .build_live_spec(&Session::new("child", "model"), &job)
16800            .await;
16801        assert_eq!(
16802            first.model.as_ref().map(|model| model.provider.as_str()),
16803            Some("old-provider")
16804        );
16805        assert_eq!(first.secrets.provider_credentials.len(), 1);
16806        assert_eq!(first.secrets.provider_credentials[0].api_key, "old-key");
16807
16808        let mut reloaded = bamboo_llm::Config::default();
16809        reloaded.provider_instances.insert(
16810            "new-provider".to_string(),
16811            provider_instance("openai", "new-key"),
16812        );
16813        reloaded.default_provider_instance = Some("new-provider".to_string());
16814        *live.write().await = reloaded;
16815
16816        let second = runner
16817            .build_live_spec(&Session::new("child", "model"), &job)
16818            .await;
16819        assert_eq!(
16820            second.model.as_ref().map(|model| model.provider.as_str()),
16821            Some("new-provider")
16822        );
16823        assert_eq!(second.secrets.provider_credentials.len(), 1);
16824        assert_eq!(second.secrets.provider_credentials[0].api_key, "new-key");
16825    }
16826
16827    #[test]
16828    fn custom_codex_provisioning_scopes_only_the_referenced_credential() {
16829        let mut executor = codex_executor(Some("custom"), None);
16830        if let ExecutorSpec::Codex {
16831            base_url,
16832            provider_key_ref,
16833            ..
16834        } = &mut executor
16835        {
16836            *base_url = Some("https://provider.example/v1".to_string());
16837            *provider_key_ref = Some("provider.custom.api_key".to_string());
16838        }
16839        let credentials = vec![
16840            ScopedCredential {
16841                provider: "openai".to_string(),
16842                api_key: "session-provider-secret".to_string(),
16843                base_url: None,
16844                provider_type: Some("openai".to_string()),
16845                credential_ref: Some("provider.openai.api_key".to_string()),
16846            },
16847            ScopedCredential {
16848                provider: "custom".to_string(),
16849                api_key: "selected-secret".to_string(),
16850                base_url: None,
16851                provider_type: Some("openai".to_string()),
16852                credential_ref: Some("provider.custom.api_key".to_string()),
16853            },
16854            ScopedCredential {
16855                provider: "other".to_string(),
16856                api_key: "unrelated-secret".to_string(),
16857                base_url: None,
16858                provider_type: Some("openai".to_string()),
16859                credential_ref: Some("provider.other.api_key".to_string()),
16860            },
16861        ];
16862        let runner = ActorChildRunner::new(
16863            "codex-test".to_string(),
16864            PathBuf::from("/bin/false"),
16865            Vec::new(),
16866            std::env::temp_dir().join("bamboo-codex-570"),
16867            executor,
16868            credentials,
16869            "openai".to_string(),
16870            1,
16871        );
16872        let mut session = Session::new("child-custom", "model");
16873        session.add_message(bamboo_agent_core::Message::user("test"));
16874        let spec = runner.build_spec(
16875            &session,
16876            &crate::runtime::execution::SpawnJob {
16877                parent_session_id: "parent".to_string(),
16878                child_session_id: "child-custom".to_string(),
16879                model: "gpt-5.4".to_string(),
16880                disabled_tools: None,
16881            },
16882        );
16883
16884        assert_eq!(spec.secrets.provider_credentials.len(), 1);
16885        assert_eq!(
16886            spec.secrets.provider_credentials[0]
16887                .credential_ref
16888                .as_deref(),
16889            Some("provider.custom.api_key")
16890        );
16891        assert_eq!(
16892            spec.secrets.provider_credentials[0].api_key,
16893            "selected-secret"
16894        );
16895    }
16896
16897    fn spec_with(
16898        role: &str,
16899        provider: &str,
16900        model: &str,
16901        workspace: Option<&str>,
16902        disabled: Option<Vec<&str>>,
16903    ) -> ProvisionSpec {
16904        let mut spec = ProvisionSpec::new(
16905            ChildIdentity {
16906                child_id: "c".into(),
16907                parent_id: None,
16908                project_key: None,
16909                role: role.into(),
16910                depth: 0,
16911            },
16912            ExecutorSpec::Echo,
16913            "/tmp/fab".into(),
16914        );
16915        spec.workspace = workspace.map(|w| w.to_string());
16916        spec.model = Some(ModelRefSpec {
16917            provider: provider.into(),
16918            model: model.into(),
16919        });
16920        spec.disabled_tools = disabled.map(|d| d.into_iter().map(String::from).collect());
16921        spec
16922    }
16923
16924    #[test]
16925    fn fingerprint_matches_interchangeable_children() {
16926        // Same role/provider/model/workspace and equal tool sets (order-insensitive)
16927        // are interchangeable on one warm worker — and differ only in child_id.
16928        let a = spec_with(
16929            "explorer",
16930            "p",
16931            "m",
16932            Some("/ws"),
16933            Some(vec!["Bash", "Edit"]),
16934        );
16935        let mut b = spec_with(
16936            "explorer",
16937            "p",
16938            "m",
16939            Some("/ws"),
16940            Some(vec!["Edit", "Bash"]),
16941        );
16942        b.identity.child_id = "other".into();
16943        assert_eq!(
16944            ActorChildRunner::fingerprint(&a),
16945            ActorChildRunner::fingerprint(&b)
16946        );
16947    }
16948
16949    #[test]
16950    fn logical_identity_is_invariant_across_local_remote_scheduled_and_warm_reuse() {
16951        let mut session =
16952            Session::new_child("logical-child-681", "logical-parent-681", "model", "child");
16953        session.root_session_id = "logical-root-681".to_string();
16954        let job = SpawnJob {
16955            parent_session_id: "logical-parent-681".to_string(),
16956            child_session_id: "logical-child-681".to_string(),
16957            model: "model".to_string(),
16958            disabled_tools: None,
16959        };
16960        let expected = LogicalSessionIdentity {
16961            creation: None,
16962            session_id: "logical-child-681".to_string(),
16963            parent_session_id: Some("logical-parent-681".to_string()),
16964            root_session_id: "logical-root-681".to_string(),
16965        };
16966
16967        let placements_and_transport_ids = [
16968            (Placement::Local, "local-mailbox-first"),
16969            (
16970                Placement::Remote {
16971                    endpoint: "wss://remote.example/actor".to_string(),
16972                },
16973                "remote-process-44",
16974            ),
16975            (
16976                Placement::Schedulable {
16977                    pool: "gpu-pool".to_string(),
16978                },
16979                "scheduled-mailbox-9",
16980            ),
16981            // Same logical child reactivated on a different pooled mailbox.
16982            (Placement::Local, "warm-mailbox-reused-77"),
16983        ];
16984        for (placement, transport_id) in placements_and_transport_ids {
16985            let mut provision = spec_with("worker", "provider", "model", None, None);
16986            provision.placement = placement;
16987            provision.identity.child_id = transport_id.to_string();
16988            assert_eq!(logical_identity_for_actor_run(&session, &job), expected);
16989            assert_ne!(
16990                provision.identity.child_id, expected.session_id,
16991                "test fixture must prove transport identity is independent"
16992            );
16993        }
16994        let mut original = spec_with("worker", "provider", "model", Some("/ws"), None);
16995        original.executor = ExecutorSpec::BambooRuntime;
16996        original.identity.depth = session.spawn_depth;
16997        original.capabilities.required_child_context = true;
16998        original.capabilities.child_creation_identity = true;
16999        original.capabilities.enforce_permissions = true;
17000        original.capabilities.native_tool_ceiling_required = true;
17001        original.capabilities.native_tool_ceiling =
17002            Some(bamboo_subagent::proto::NativeToolCeiling {
17003                version: 1,
17004                child_session_id: session.id.clone(),
17005                parent_session_id: job.parent_session_id.clone(),
17006                root_session_id: session.root_session_id.clone(),
17007                created_at: session.created_at,
17008                spawn_depth: session.spawn_depth,
17009                project_id: None,
17010                tools: vec![],
17011            });
17012        original.validate().unwrap();
17013        let first = ActorChildRunner::fresh_required_worker_spec(&original);
17014        let second = ActorChildRunner::fresh_required_worker_spec(&original);
17015        assert_ne!(first.identity.child_id, second.identity.child_id);
17016        assert_ne!(first.identity.child_id, original.identity.child_id);
17017        for physical in [first, second] {
17018            physical.validate().unwrap();
17019            let mut actual = serde_json::to_value(&physical).unwrap();
17020            actual["identity"]["child_id"] = serde_json::json!(original.identity.child_id);
17021            assert_eq!(actual, serde_json::to_value(&original).unwrap());
17022            assert_eq!(logical_identity_for_actor_run(&session, &job), expected);
17023        }
17024    }
17025
17026    #[test]
17027    fn fingerprint_separates_distinct_runtimes() {
17028        let base = spec_with("explorer", "p", "m", Some("/ws"), None);
17029        let base_fp = ActorChildRunner::fingerprint(&base);
17030        // Each axis that is baked into the worker must split the pool bucket.
17031        assert_ne!(
17032            base_fp,
17033            ActorChildRunner::fingerprint(&spec_with("writer", "p", "m", Some("/ws"), None))
17034        );
17035        assert_ne!(
17036            base_fp,
17037            ActorChildRunner::fingerprint(&spec_with("explorer", "p2", "m", Some("/ws"), None))
17038        );
17039        assert_ne!(
17040            base_fp,
17041            ActorChildRunner::fingerprint(&spec_with("explorer", "p", "m2", Some("/ws"), None))
17042        );
17043        assert_ne!(
17044            base_fp,
17045            ActorChildRunner::fingerprint(&spec_with("explorer", "p", "m", Some("/ws2"), None))
17046        );
17047        assert_ne!(
17048            base_fp,
17049            ActorChildRunner::fingerprint(&spec_with(
17050                "explorer",
17051                "p",
17052                "m",
17053                Some("/ws"),
17054                Some(vec!["Bash"])
17055            ))
17056        );
17057    }
17058
17059    #[test]
17060    fn fingerprint_splits_rotated_provider_credentials_without_exposing_them() {
17061        let mut before = spec_with("explorer", "provider", "model", Some("/ws"), None);
17062        before.secrets.provider_credentials.push(ScopedCredential {
17063            provider: "provider".to_string(),
17064            api_key: "old-secret-key".to_string(),
17065            base_url: Some("https://provider.example/v1".to_string()),
17066            provider_type: Some("openai".to_string()),
17067            credential_ref: None,
17068        });
17069        let mut after = before.clone();
17070        after.secrets.provider_credentials[0].api_key = "new-secret-key".to_string();
17071
17072        let before_fingerprint = ActorChildRunner::fingerprint(&before);
17073        let after_fingerprint = ActorChildRunner::fingerprint(&after);
17074        assert_ne!(before_fingerprint, after_fingerprint);
17075        assert!(!before_fingerprint.contains("old-secret-key"));
17076        assert!(!after_fingerprint.contains("new-secret-key"));
17077    }
17078
17079    #[test]
17080    fn fingerprint_splits_on_baked_capabilities() {
17081        // Every capability baked once at provision time must split the pool
17082        // bucket, else a worker baked for one posture gets reused for another
17083        // (e.g. a depth-1 worker re-stamping spawn_depth onto a depth-4 child,
17084        // breaking the depth cap; or a bypass worker reused for a non-bypass one).
17085        let base_fp =
17086            ActorChildRunner::fingerprint(&spec_with("explorer", "p", "m", Some("/ws"), None));
17087
17088        let mut creation = spec_with("explorer", "p", "m", Some("/ws"), None);
17089        creation.capabilities.child_creation_identity = true;
17090        let typed = ActorChildRunner::fingerprint(&creation);
17091        assert_ne!(
17092            base_fp, typed,
17093            "legacy worker pool cannot satisfy birth-required Run"
17094        );
17095        creation.identity.child_id = "different-logical-child".into();
17096        assert_eq!(
17097            typed,
17098            ActorChildRunner::fingerprint(&creation),
17099            "birth-required mode retains cross-Child pool reuse"
17100        );
17101
17102        let mut depth = spec_with("explorer", "p", "m", Some("/ws"), None);
17103        depth.identity.depth = 2;
17104        assert_ne!(
17105            base_fp,
17106            ActorChildRunner::fingerprint(&depth),
17107            "depth must split"
17108        );
17109
17110        let mut nested = spec_with("explorer", "p", "m", Some("/ws"), None);
17111        nested.capabilities.nested_spawn = true;
17112        assert_ne!(
17113            base_fp,
17114            ActorChildRunner::fingerprint(&nested),
17115            "nested_spawn must split"
17116        );
17117
17118        let mut bypass = spec_with("explorer", "p", "m", Some("/ws"), None);
17119        bypass.capabilities.bypass = true;
17120        assert_ne!(
17121            base_fp,
17122            ActorChildRunner::fingerprint(&bypass),
17123            "bypass must split"
17124        );
17125
17126        let mut auto = spec_with("explorer", "p", "m", Some("/ws"), None);
17127        auto.capabilities.auto_approve_permissions = true;
17128        assert_ne!(
17129            base_fp,
17130            ActorChildRunner::fingerprint(&auto),
17131            "auto_approve_permissions must split"
17132        );
17133
17134        let mut global_auto = spec_with("explorer", "p", "m", Some("/ws"), None);
17135        global_auto.capabilities.permission_requested_mode = "default".to_string();
17136        global_auto.capabilities.permission_effective_mode = "auto".to_string();
17137        global_auto.capabilities.auto_approve_permissions = true;
17138        let mut explicit_auto = global_auto.clone();
17139        explicit_auto.capabilities.permission_requested_mode = "auto".to_string();
17140        assert_ne!(
17141            ActorChildRunner::fingerprint(&global_auto),
17142            ActorChildRunner::fingerprint(&explicit_auto),
17143            "permission_requested_mode must split global and explicit Auto"
17144        );
17145
17146        let mut plan_overlay = explicit_auto.clone();
17147        plan_overlay.capabilities.permission_effective_mode = "plan".to_string();
17148        assert_ne!(
17149            ActorChildRunner::fingerprint(&explicit_auto),
17150            ActorChildRunner::fingerprint(&plan_overlay),
17151            "permission_effective_mode must split Plan overlay from Auto"
17152        );
17153
17154        let mut enforce = spec_with("explorer", "p", "m", Some("/ws"), None);
17155        enforce.capabilities.enforce_permissions = true;
17156        assert_ne!(
17157            base_fp,
17158            ActorChildRunner::fingerprint(&enforce),
17159            "enforce_permissions must split"
17160        );
17161
17162        let mut cap = spec_with("explorer", "p", "m", Some("/ws"), None);
17163        cap.capabilities.max_spawn_depth = Some(8);
17164        assert_ne!(
17165            base_fp,
17166            ActorChildRunner::fingerprint(&cap),
17167            "max_spawn_depth must split"
17168        );
17169
17170        // #73 (P1): the worker bakes `no_human_review` from this flag once at
17171        // build(), so it MUST split the pool or a worker baked for one approval
17172        // posture is reused for the opposite one.
17173        let mut nha = spec_with("explorer", "p", "m", Some("/ws"), None);
17174        nha.capabilities.no_human_approver = true;
17175        assert_ne!(
17176            base_fp,
17177            ActorChildRunner::fingerprint(&nha),
17178            "no_human_approver must split"
17179        );
17180
17181        // The generic read-only no-shell checker is baked once at build(), so a
17182        // planner/Guardian worker must not be reused for an ordinary child.
17183        let mut read_only = spec_with("explorer", "p", "m", Some("/ws"), None);
17184        read_only.capabilities.read_only = true;
17185        assert_ne!(
17186            base_fp,
17187            ActorChildRunner::fingerprint(&read_only),
17188            "read_only must split"
17189        );
17190
17191        // Preserve the legacy Guardian bit as its own rolling-upgrade axis.
17192        let mut gro = spec_with("explorer", "p", "m", Some("/ws"), None);
17193        gro.capabilities.guardian_read_only = true;
17194        assert_ne!(
17195            base_fp,
17196            ActorChildRunner::fingerprint(&gro),
17197            "guardian_read_only must split"
17198        );
17199    }
17200
17201    #[test]
17202    fn fingerprint_splits_codex_exec_and_app_server_workers() {
17203        let mut exec = spec_with("explorer", "p", "m", Some("/ws"), None);
17204        exec.executor = codex_executor(Some("inherit"), None);
17205        let mut app_server = exec.clone();
17206        if let ExecutorSpec::Codex { mode, .. } = &mut app_server.executor {
17207            *mode = Some("app_server".to_string());
17208        }
17209        assert_ne!(
17210            ActorChildRunner::fingerprint(&exec),
17211            ActorChildRunner::fingerprint(&app_server)
17212        );
17213    }
17214
17215    struct StaticDecider(bool);
17216
17217    #[async_trait]
17218    impl ChildApprovalDecider for StaticDecider {
17219        async fn decide(&self, _child: &str, _req: &serde_json::Value) -> bool {
17220            self.0
17221        }
17222    }
17223
17224    struct RecordingReviewer {
17225        reviewed: mpsc::UnboundedSender<(String, String, serde_json::Value)>,
17226    }
17227
17228    #[async_trait]
17229    impl ChildApprovalReviewer for RecordingReviewer {
17230        async fn review(&self, parent: &str, child: &str, request: &serde_json::Value) -> bool {
17231            let _ = self
17232                .reviewed
17233                .send((parent.to_string(), child.to_string(), request.clone()));
17234            true
17235        }
17236    }
17237
17238    // ---- first-frame watchdog (dead-pooled-worker recovery) -----------------
17239
17240    /// A link that never yields a frame — models a worker that died (or never
17241    /// subscribed) so its Run sits queued with no server.
17242    struct SilentLink;
17243    #[async_trait]
17244    impl bamboo_subagent::ChildLink for SilentLink {
17245        async fn send(&mut self, _: ParentFrame) -> bamboo_subagent::TransportResult<()> {
17246            Ok(())
17247        }
17248        async fn next_frame(&mut self) -> bamboo_subagent::TransportResult<Option<ChildFrame>> {
17249            std::future::pending().await
17250        }
17251    }
17252
17253    /// A link that immediately yields one terminal frame (a healthy fast worker).
17254    struct InstantTerminalLink {
17255        done: bool,
17256    }
17257
17258    struct ApprovalRoundTripLink {
17259        step: u8,
17260        approval_reply: Option<(String, bool)>,
17261    }
17262
17263    #[async_trait]
17264    impl bamboo_subagent::ChildLink for ApprovalRoundTripLink {
17265        async fn send(&mut self, frame: ParentFrame) -> bamboo_subagent::TransportResult<()> {
17266            if let ParentFrame::ApprovalReply { id, approved } = frame {
17267                self.approval_reply = Some((id, approved));
17268                self.step = 2;
17269            }
17270            Ok(())
17271        }
17272
17273        async fn next_frame(&mut self) -> bamboo_subagent::TransportResult<Option<ChildFrame>> {
17274            match self.step {
17275                0 => {
17276                    self.step = 1;
17277                    Ok(Some(ChildFrame::ApprovalRequest {
17278                        id: "approval-1".into(),
17279                        body: serde_json::json!({
17280                            "tool_name": "Bash",
17281                            "permission": "execute",
17282                            "resource": "rm -rf target",
17283                            "permission_request": {"reason_code": "hard_dangerous"}
17284                        }),
17285                    }))
17286                }
17287                1 => std::future::pending().await,
17288                2 => {
17289                    self.step = 3;
17290                    Ok(Some(ChildFrame::Terminal {
17291                        status: TerminalStatus::Completed,
17292                        result: Some("done".into()),
17293                        error: None,
17294                        final_event_watermark: None,
17295                        transcript: vec![],
17296                    }))
17297                }
17298                _ => std::future::pending().await,
17299            }
17300        }
17301    }
17302
17303    #[tokio::test]
17304    async fn drive_routes_forced_ask_to_parent_reviewer_without_human_event() {
17305        let (event_tx, mut event_rx) = mpsc::channel::<AgentEvent>(8);
17306        let (review_tx, mut review_rx) = mpsc::unbounded_channel();
17307        let reviewer: Arc<dyn ChildApprovalReviewer> = Arc::new(RecordingReviewer {
17308            reviewed: review_tx,
17309        });
17310        let cancel = CancellationToken::new();
17311        let (live_tx, mut live_rx) = mpsc::unbounded_channel::<ParentFrame>();
17312        let (_delivery_tx, mut delivery_rx) = mpsc::unbounded_channel();
17313        let mut logical_session = Session::new("child-reviewer", "model");
17314        let live_guard = crate::external_agents::live::register("child-reviewer", live_tx, 0, None);
17315        let mut link = ApprovalRoundTripLink {
17316            step: 0,
17317            approval_reply: None,
17318        };
17319
17320        let result = tokio::time::timeout(
17321            Duration::from_secs(1),
17322            drive(ActorDriveContext {
17323                expected_creation: None,
17324                client: &mut link,
17325                parent_session_id: "parent-reviewer",
17326                child_session_id: "child-reviewer",
17327                child_attempt: 0,
17328                approval_registry: None,
17329                approval_decider: None,
17330                approval_reviewer: Some(&reviewer),
17331                escalation_bridge: None,
17332                event_tx: &event_tx,
17333                cancel_token: &cancel,
17334                live_rx: &mut live_rx,
17335                delivery_rx: &mut delivery_rx,
17336                logical_session: &mut logical_session,
17337                expected_permission_posture: None,
17338                session_inbox_runtime: None,
17339                actor_directory_store: None,
17340                canonical_subagent_tool: None,
17341                ticket_service: None,
17342                activation_run_id: None,
17343                execution_epoch: 0,
17344                expected_source_actor_id: "child-reviewer",
17345                initial_inflight_claims: VecDeque::new(),
17346                plain_actor: false,
17347                remote_environment_lease: false,
17348                readonly_output: None,
17349                local_history_tools: None,
17350                local_history_read_only: false,
17351                strict_history_output: None,
17352                plain_input: None,
17353                canonical_activation: None,
17354                canonical_placement_ref: None,
17355                actor_event_observer: None,
17356                plain_run: None,
17357                first_frame_timeout: None,
17358            }),
17359        )
17360        .await
17361        .expect("worker must receive the reviewer verdict before terminating");
17362
17363        assert_eq!(result.ok().flatten().as_deref(), Some("done"));
17364        assert_eq!(
17365            link.approval_reply,
17366            Some(("approval-1".to_string(), true)),
17367            "reviewer verdict must traverse the live route back to the worker"
17368        );
17369        let (parent, child, body) = tokio::time::timeout(Duration::from_secs(1), review_rx.recv())
17370            .await
17371            .expect("reviewer should be invoked off-loop")
17372            .expect("review channel should remain open");
17373        assert_eq!(parent, "parent-reviewer");
17374        assert_eq!(child, "child-reviewer");
17375        assert_eq!(
17376            body.pointer("/permission_request/reason_code")
17377                .and_then(serde_json::Value::as_str),
17378            Some("hard_dangerous")
17379        );
17380        assert!(
17381            event_rx.try_recv().is_err(),
17382            "must not emit a human-review event"
17383        );
17384        drop(live_guard);
17385    }
17386
17387    #[tokio::test]
17388    async fn drive_denies_forced_ask_without_parent_reviewer_or_manual_event() {
17389        let (event_tx, mut event_rx) = mpsc::channel::<AgentEvent>(8);
17390        let cancel = CancellationToken::new();
17391        let (_live_tx, mut live_rx) = mpsc::unbounded_channel::<ParentFrame>();
17392        let (_delivery_tx, mut delivery_rx) = mpsc::unbounded_channel();
17393        let mut logical_session = Session::new("child-no-reviewer", "model");
17394        let mut link = ApprovalRoundTripLink {
17395            step: 0,
17396            approval_reply: None,
17397        };
17398
17399        let result = tokio::time::timeout(
17400            Duration::from_secs(1),
17401            drive(ActorDriveContext {
17402                expected_creation: None,
17403                client: &mut link,
17404                parent_session_id: "parent-no-reviewer",
17405                child_session_id: "child-no-reviewer",
17406                child_attempt: 0,
17407                approval_registry: None,
17408                approval_decider: None,
17409                approval_reviewer: None,
17410                escalation_bridge: None,
17411                event_tx: &event_tx,
17412                cancel_token: &cancel,
17413                live_rx: &mut live_rx,
17414                delivery_rx: &mut delivery_rx,
17415                logical_session: &mut logical_session,
17416                expected_permission_posture: None,
17417                session_inbox_runtime: None,
17418                actor_directory_store: None,
17419                canonical_subagent_tool: None,
17420                ticket_service: None,
17421                activation_run_id: None,
17422                execution_epoch: 0,
17423                expected_source_actor_id: "child-no-reviewer",
17424                initial_inflight_claims: VecDeque::new(),
17425                plain_actor: false,
17426                remote_environment_lease: false,
17427                readonly_output: None,
17428                local_history_tools: None,
17429                local_history_read_only: false,
17430                strict_history_output: None,
17431                plain_input: None,
17432                canonical_activation: None,
17433                canonical_placement_ref: None,
17434                actor_event_observer: None,
17435                plain_run: None,
17436                first_frame_timeout: None,
17437            }),
17438        )
17439        .await
17440        .expect("fail-closed reply must unblock the child immediately");
17441
17442        assert_eq!(result.ok().flatten().as_deref(), Some("done"));
17443        assert_eq!(link.approval_reply, Some(("approval-1".to_string(), false)));
17444        assert!(
17445            event_rx.try_recv().is_err(),
17446            "missing parent review must not surface a manual approval event"
17447        );
17448    }
17449    #[async_trait]
17450    impl bamboo_subagent::ChildLink for InstantTerminalLink {
17451        async fn send(&mut self, _: ParentFrame) -> bamboo_subagent::TransportResult<()> {
17452            Ok(())
17453        }
17454        async fn next_frame(&mut self) -> bamboo_subagent::TransportResult<Option<ChildFrame>> {
17455            if self.done {
17456                std::future::pending().await
17457            } else {
17458                self.done = true;
17459                Ok(Some(ChildFrame::Terminal {
17460                    status: TerminalStatus::Completed,
17461                    result: Some("done".into()),
17462                    error: None,
17463                    final_event_watermark: None,
17464                    transcript: vec![],
17465                }))
17466            }
17467        }
17468    }
17469
17470    #[tokio::test]
17471    async fn drive_trips_first_frame_watchdog_on_a_silent_worker() {
17472        let (event_tx, _rx) = mpsc::channel::<AgentEvent>(8);
17473        let cancel = CancellationToken::new();
17474        let (_live_tx, mut live_rx) = mpsc::unbounded_channel::<ParentFrame>();
17475        let (_delivery_tx, mut delivery_rx) = mpsc::unbounded_channel();
17476        let mut logical_session = Session::new("child-x", "model");
17477        let mut link = SilentLink;
17478        let r = drive(ActorDriveContext {
17479            expected_creation: None,
17480            client: &mut link,
17481            parent_session_id: "parent-x",
17482            child_session_id: "child-x",
17483            child_attempt: 0,
17484            approval_registry: None,
17485            approval_decider: None,
17486            approval_reviewer: None,
17487            escalation_bridge: None,
17488            event_tx: &event_tx,
17489            cancel_token: &cancel,
17490            live_rx: &mut live_rx,
17491            delivery_rx: &mut delivery_rx,
17492            logical_session: &mut logical_session,
17493            expected_permission_posture: None,
17494            session_inbox_runtime: None,
17495            actor_directory_store: None,
17496            canonical_subagent_tool: None,
17497            ticket_service: None,
17498            activation_run_id: None,
17499            execution_epoch: 0,
17500            expected_source_actor_id: "child-x",
17501            initial_inflight_claims: VecDeque::new(),
17502            plain_actor: false,
17503            remote_environment_lease: false,
17504            readonly_output: None,
17505            local_history_tools: None,
17506            local_history_read_only: false,
17507            strict_history_output: None,
17508            plain_input: None,
17509            canonical_activation: None,
17510            canonical_placement_ref: None,
17511            actor_event_observer: None,
17512            plain_run: None,
17513            first_frame_timeout: Some(Duration::from_millis(100)),
17514        })
17515        .await;
17516        assert!(
17517            matches!(r, Err(AgentError::WorkerUnresponsive(_))),
17518            "a silent worker must trip the first-frame watchdog, got {r:?}"
17519        );
17520    }
17521
17522    #[tokio::test]
17523    async fn drive_does_not_trip_when_a_frame_arrives() {
17524        let (event_tx, _rx) = mpsc::channel::<AgentEvent>(8);
17525        let cancel = CancellationToken::new();
17526        let (_live_tx, mut live_rx) = mpsc::unbounded_channel::<ParentFrame>();
17527        let (_delivery_tx, mut delivery_rx) = mpsc::unbounded_channel();
17528        let mut logical_session = Session::new("child-y", "model");
17529        let mut link = InstantTerminalLink { done: false };
17530        // Even a tiny timeout must NOT trip: the terminal frame arrives first and
17531        // disarms the watchdog.
17532        let r = drive(ActorDriveContext {
17533            expected_creation: None,
17534            client: &mut link,
17535            parent_session_id: "parent-y",
17536            child_session_id: "child-y",
17537            child_attempt: 0,
17538            approval_registry: None,
17539            approval_decider: None,
17540            approval_reviewer: None,
17541            escalation_bridge: None,
17542            event_tx: &event_tx,
17543            cancel_token: &cancel,
17544            live_rx: &mut live_rx,
17545            delivery_rx: &mut delivery_rx,
17546            logical_session: &mut logical_session,
17547            expected_permission_posture: None,
17548            session_inbox_runtime: None,
17549            actor_directory_store: None,
17550            canonical_subagent_tool: None,
17551            ticket_service: None,
17552            activation_run_id: None,
17553            execution_epoch: 0,
17554            expected_source_actor_id: "child-y",
17555            initial_inflight_claims: VecDeque::new(),
17556            plain_actor: false,
17557            remote_environment_lease: false,
17558            readonly_output: None,
17559            local_history_tools: None,
17560            local_history_read_only: false,
17561            strict_history_output: None,
17562            plain_input: None,
17563            canonical_activation: None,
17564            canonical_placement_ref: None,
17565            actor_event_observer: None,
17566            plain_run: None,
17567            first_frame_timeout: Some(Duration::from_millis(50)),
17568        })
17569        .await;
17570        assert_eq!(r.ok().flatten().as_deref(), Some("done"));
17571    }
17572
17573    #[tokio::test]
17574    async fn child_approval_fails_closed_without_decider() {
17575        // No decider wired ⇒ the host denies (safe default), unchanged behavior.
17576        let body = serde_json::json!({"tool_name":"Bash","permission":"run","resource":"rm -rf /"});
17577        assert!(!decide_child_approval(None, "child-1", &body).await);
17578    }
17579
17580    #[tokio::test]
17581    async fn child_approval_honors_wired_decider() {
17582        let body =
17583            serde_json::json!({"tool_name":"Write","permission":"write","resource":"/tmp/x"});
17584        let approve: Arc<dyn ChildApprovalDecider> = Arc::new(StaticDecider(true));
17585        let deny: Arc<dyn ChildApprovalDecider> = Arc::new(StaticDecider(false));
17586        assert!(decide_child_approval(Some(&approve), "child-1", &body).await);
17587        assert!(!decide_child_approval(Some(&deny), "child-1", &body).await);
17588    }
17589
17590    // ---- #193: remote placement routing -------------------------------------
17591
17592    use crate::runtime::execution::SpawnJob;
17593    use bamboo_agent_core::Session;
17594
17595    /// A runner with a BOGUS worker_bin (`/bin/false`): a local spawn here would
17596    /// FAIL, so a passing remote test proves the remote path never spawns.
17597    fn bogus_runner(placements: HashMap<String, ResolvedRemotePlacement>) -> ActorChildRunner {
17598        ActorChildRunner::new(
17599            "test-actor".into(),
17600            PathBuf::from("/bin/false"),
17601            vec![],
17602            std::env::temp_dir().join("bamboo-test-fab-193"),
17603            ExecutorSpec::Echo,
17604            vec![],
17605            "anthropic".into(),
17606            4,
17607        )
17608        .with_remote_placements(placements)
17609    }
17610
17611    /// A child session of the given role (the role rides `subagent_type`, the
17612    /// path build_spec + the remote lookup both read).
17613    fn session_of_role(role: &str, assignment: &str) -> Session {
17614        let mut s = Session::new("child-1", "test-model");
17615        s.metadata
17616            .insert("subagent_type".to_string(), role.to_string());
17617        s.add_message(bamboo_agent_core::Message::user(assignment));
17618        s
17619    }
17620
17621    fn job_for(child: &str) -> SpawnJob {
17622        SpawnJob {
17623            parent_session_id: "parent-1".into(),
17624            child_session_id: child.into(),
17625            model: String::new(),
17626            disabled_tools: None,
17627        }
17628    }
17629
17630    #[tokio::test]
17631    async fn spawn_depth_provisioning_uses_live_policy_for_local_and_remote_workers() {
17632        let live = Arc::new(tokio::sync::RwLock::new(bamboo_llm::Config::default()));
17633        live.write().await.subagents_mut().max_spawn_depth = Some(1);
17634        let mut remote = HashMap::new();
17635        remote.insert(
17636            "remote-role".into(),
17637            ResolvedRemotePlacement {
17638                endpoint: "wss://worker.example:8443".into(),
17639                token: None,
17640                ca_cert_file: None,
17641                host_label: None,
17642                broker_peer: None,
17643                requirements: None,
17644            },
17645        );
17646        let runner = bogus_runner(remote).with_live_provider_config(live.clone());
17647        let root = Session::new("provision-root", "model");
17648        for role in ["local-role", "remote-role"] {
17649            let mut child =
17650                Session::new_child_of(format!("provision-{role}"), &root, "model", "child");
17651            child.metadata.insert("subagent_type".into(), role.into());
17652            let first = runner.build_live_spec(&child, &job_for(&child.id)).await;
17653            assert_eq!(first.capabilities.max_spawn_depth, Some(1));
17654            assert!(!first.capabilities.nested_spawn);
17655            assert_eq!(
17656                matches!(first.placement, Placement::Remote { .. }),
17657                role == "remote-role"
17658            );
17659            live.write().await.subagents_mut().max_spawn_depth = Some(7);
17660            let second = runner.build_live_spec(&child, &job_for(&child.id)).await;
17661            assert_eq!(second.capabilities.max_spawn_depth, Some(7));
17662            assert!(second.capabilities.nested_spawn);
17663            assert_ne!(
17664                ActorChildRunner::fingerprint(&first),
17665                ActorChildRunner::fingerprint(&second)
17666            );
17667            assert_eq!(second.identity.depth, 1);
17668            live.write().await.subagents_mut().max_spawn_depth = Some(1);
17669        }
17670        let fallback = bogus_runner(HashMap::new()).with_max_spawn_depth(0);
17671        let spec = fallback.build_spec(&root, &job_for(&root.id));
17672        assert_eq!(spec.capabilities.max_spawn_depth, Some(0));
17673        assert!(!spec.capabilities.nested_spawn);
17674    }
17675
17676    struct RecordingChildSessionPort {
17677        parent: Session,
17678        saved: std::sync::Mutex<Option<Session>>,
17679    }
17680
17681    impl RecordingChildSessionPort {
17682        fn saved_child(&self) -> Session {
17683            self.saved
17684                .lock()
17685                .expect("saved-child fixture lock")
17686                .clone()
17687                .expect("create_child_action must save the child")
17688        }
17689    }
17690
17691    #[async_trait]
17692    impl crate::session_app::child_session::ChildSessionPort for RecordingChildSessionPort {
17693        async fn load_root_session(
17694            &self,
17695            root_id: &str,
17696        ) -> Result<Session, crate::session_app::child_session::ChildSessionError> {
17697            assert_eq!(root_id, self.parent.id);
17698            Ok(self.parent.clone())
17699        }
17700
17701        async fn load_child_for_parent(
17702            &self,
17703            _parent_id: &str,
17704            _child_id: &str,
17705        ) -> Result<Session, crate::session_app::child_session::ChildSessionError> {
17706            unreachable!("create_child_action does not reload the child")
17707        }
17708
17709        async fn save_child_session(
17710            &self,
17711            child: &mut Session,
17712        ) -> Result<(), crate::session_app::child_session::ChildSessionError> {
17713            *self.saved.lock().expect("saved-child fixture lock") = Some(child.clone());
17714            Ok(())
17715        }
17716
17717        async fn save_child_session_authoritative_flags(
17718            &self,
17719            _child: &mut Session,
17720        ) -> Result<(), crate::session_app::child_session::ChildSessionError> {
17721            unreachable!("new-child creation uses the ordinary save")
17722        }
17723
17724        async fn is_child_running(&self, _child_id: &str) -> bool {
17725            false
17726        }
17727
17728        async fn list_children(
17729            &self,
17730            _parent_id: &str,
17731        ) -> Vec<crate::session_app::child_session::ChildSessionEntry> {
17732            Vec::new()
17733        }
17734
17735        async fn enqueue_child_run(
17736            &self,
17737            _parent: &Session,
17738            _child: &Session,
17739        ) -> Result<(), crate::session_app::child_session::ChildSessionError> {
17740            unreachable!("fixture creates the child with auto_run=false")
17741        }
17742
17743        async fn cancel_child_run_and_wait(
17744            &self,
17745            _child_id: &str,
17746        ) -> Result<(), crate::session_app::child_session::ChildSessionError> {
17747            unreachable!("create_child_action does not cancel")
17748        }
17749
17750        async fn delete_child_session(
17751            &self,
17752            _parent_id: &str,
17753            _child_id: &str,
17754        ) -> Result<
17755            crate::session_app::child_session::DeleteChildResult,
17756            crate::session_app::child_session::ChildSessionError,
17757        > {
17758            unreachable!("create_child_action does not delete")
17759        }
17760
17761        async fn get_child_runner_info(
17762            &self,
17763            _child_id: &str,
17764        ) -> Option<crate::session_app::child_session::ChildRunnerInfo> {
17765            None
17766        }
17767
17768        async fn register_parent_wait_for_child(
17769            &self,
17770            _parent_session_id: &str,
17771            _child_session_id: &str,
17772            _tool_call_id: Option<&str>,
17773        ) -> Result<(), crate::session_app::child_session::ChildSessionError> {
17774            unreachable!("create_child_action does not register a wait")
17775        }
17776
17777        async fn register_parent_wait_for_children(
17778            &self,
17779            _parent_session_id: &str,
17780            _child_session_ids: &[String],
17781            _policy: bamboo_domain::session::runtime_state::ChildWaitPolicy,
17782        ) -> Result<usize, crate::session_app::child_session::ChildSessionError> {
17783            unreachable!("create_child_action does not register a wait")
17784        }
17785
17786        async fn active_child_ids(&self, _parent_session_id: &str) -> Vec<String> {
17787            Vec::new()
17788        }
17789
17790        async fn find_resident_child(
17791            &self,
17792            _root_session_id: &str,
17793            _resident_name: &str,
17794        ) -> Option<String> {
17795            None
17796        }
17797
17798        async fn ensure_child_indexed(&self, _child_session_id: &str) {}
17799    }
17800
17801    #[test]
17802    fn build_spec_sets_remote_placement_for_matching_role() {
17803        let mut placements = HashMap::new();
17804        placements.insert(
17805            "explorer".to_string(),
17806            ResolvedRemotePlacement {
17807                endpoint: "wss://gpu-host:8443".into(),
17808                token: Some("T-secret".into()),
17809                ca_cert_file: None,
17810                host_label: None,
17811                broker_peer: None,
17812                requirements: None,
17813            },
17814        );
17815        let runner = bogus_runner(placements);
17816
17817        // The broker credential never rides the child RunSpec secrets envelope.
17818        let s = session_of_role("explorer", "do the thing");
17819        let spec = runner.build_spec(&s, &job_for("child-1"));
17820        match &spec.placement {
17821            Placement::Remote { endpoint } => assert_eq!(endpoint, "wss://gpu-host:8443"),
17822            other => panic!("expected Remote, got {other:?}"),
17823        }
17824        assert!(spec.secrets.worker_auth_token.is_none());
17825    }
17826
17827    #[test]
17828    fn build_spec_leaves_local_for_unmatched_role() {
17829        let mut placements = HashMap::new();
17830        placements.insert(
17831            "explorer".to_string(),
17832            ResolvedRemotePlacement {
17833                endpoint: "wss://gpu-host:8443".into(),
17834                token: Some("T".into()),
17835                ca_cert_file: None,
17836                host_label: None,
17837                broker_peer: None,
17838                requirements: None,
17839            },
17840        );
17841        let runner = bogus_runner(placements);
17842
17843        // A DIFFERENT role keeps the default Local placement + no bearer.
17844        let s = session_of_role("writer", "do the thing");
17845        let spec = runner.build_spec(&s, &job_for("child-1"));
17846        assert_eq!(spec.placement, Placement::Local);
17847        assert!(spec.secrets.worker_auth_token.is_none());
17848    }
17849
17850    #[test]
17851    fn build_spec_local_when_no_placements() {
17852        let runner = bogus_runner(HashMap::new());
17853        let s = session_of_role("explorer", "do the thing");
17854        let spec = runner.build_spec(&s, &job_for("child-1"));
17855        assert_eq!(spec.placement, Placement::Local);
17856        assert!(spec.secrets.worker_auth_token.is_none());
17857    }
17858
17859    #[test]
17860    fn typed_child_state_is_the_read_only_authority_under_auto_and_bypass() {
17861        for mode in [
17862            bamboo_domain::SessionPermissionMode::Auto,
17863            bamboo_domain::SessionPermissionMode::Bypass,
17864        ] {
17865            let runner = bogus_runner(HashMap::new());
17866            let mut child = session_of_role("planner", "inspect and plan");
17867            let runtime = child
17868                .agent_runtime_state
17869                .get_or_insert_with(bamboo_domain::AgentRuntimeState::default);
17870            runtime.set_permission_mode(mode);
17871            runtime.read_only = true;
17872
17873            let spec = runner.build_spec(&child, &job_for("read-only-child"));
17874
17875            assert!(spec.capabilities.read_only);
17876            assert!(spec.capabilities.read_only_enforced());
17877            assert!(
17878                spec.capabilities.guardian_read_only,
17879                "the compatibility alias must protect planners on older workers"
17880            );
17881            assert_eq!(spec.capabilities.permission_requested_mode, mode.as_str());
17882            assert_eq!(spec.capabilities.permission_effective_mode, "plan");
17883            assert!(
17884                !spec.capabilities.bypass,
17885                "read-only must win over requested {mode:?}"
17886            );
17887            assert_eq!(
17888                spec.capabilities.auto_approve_permissions,
17889                mode == bamboo_domain::SessionPermissionMode::Auto,
17890                "Auto may suppress prompts but must not widen Plan authorization"
17891            );
17892            let disabled = spec
17893                .disabled_tools
17894                .as_ref()
17895                .expect("read-only state must provision a host-owned denylist");
17896            for tool_name in [
17897                "Edit",
17898                "Write",
17899                "Plan",
17900                "SubAgent",
17901                "session_control",
17902                "request_permissions",
17903            ] {
17904                assert!(
17905                    disabled.iter().any(|name| name == tool_name),
17906                    "read-only provisioning must disable {tool_name}"
17907                );
17908            }
17909        }
17910
17911        let runner = bogus_runner(HashMap::new());
17912        let cosmetic_planner = runner.build_spec(
17913            &session_of_role("planner", "label only"),
17914            &job_for("cosmetic"),
17915        );
17916        assert!(
17917            !cosmetic_planner.capabilities.read_only_enforced(),
17918            "a cosmetic role label must not change authority"
17919        );
17920        assert!(cosmetic_planner.disabled_tools.is_none());
17921
17922        let cosmetic_guardian = runner.build_spec(
17923            &session_of_role("guardian", "label only"),
17924            &job_for("cosmetic-guardian"),
17925        );
17926        assert!(
17927            !cosmetic_guardian.capabilities.read_only_enforced(),
17928            "a Guardian label without typed state or the legacy denylist is cosmetic"
17929        );
17930
17931        let legacy_guardian = session_of_role("guardian", "pre-upgrade reviewer");
17932        let mut legacy_job = job_for("legacy-guardian");
17933        legacy_job.disabled_tools = Some(vec![
17934            "Edit".to_string(),
17935            "Write".to_string(),
17936            "SubAgent".to_string(),
17937            "WebFetch".to_string(),
17938        ]);
17939        let legacy_spec = runner.build_spec(&legacy_guardian, &legacy_job);
17940        assert!(
17941            legacy_spec.capabilities.read_only_enforced(),
17942            "an in-flight pre-upgrade Guardian keeps its read-only authority"
17943        );
17944
17945        let mut differently_named = session_of_role("explorer", "typed authority");
17946        differently_named
17947            .agent_runtime_state
17948            .get_or_insert_with(bamboo_domain::AgentRuntimeState::default)
17949            .read_only = true;
17950        let typed = runner.build_spec(&differently_named, &job_for("typed"));
17951        assert!(
17952            typed.capabilities.read_only_enforced(),
17953            "typed state must enforce read-only independently of role"
17954        );
17955    }
17956
17957    #[test]
17958    fn typed_read_only_forces_native_external_executor_profiles() {
17959        let mut planner = session_of_role("planner", "inspect and plan");
17960        planner
17961            .agent_runtime_state
17962            .get_or_insert_with(bamboo_domain::AgentRuntimeState::default)
17963            .read_only = true;
17964
17965        let claude_runner = ActorChildRunner::new(
17966            "test-actor".into(),
17967            PathBuf::from("/bin/false"),
17968            vec![],
17969            std::env::temp_dir().join("bamboo-test-plan-claude"),
17970            ExecutorSpec::ClaudeCode {
17971                binary: None,
17972                model: None,
17973                permission_mode: Some("bypassPermissions".to_string()),
17974                inherit_user_config: Some(true),
17975                forward_env: None,
17976            },
17977            vec![],
17978            "anthropic".into(),
17979            4,
17980        );
17981        let claude_spec = claude_runner.build_spec(&planner, &job_for("planner-claude"));
17982        let disabled = claude_spec
17983            .disabled_tools
17984            .as_ref()
17985            .expect("typed read-only child must receive a host-owned denylist");
17986        for tool in ["Bash", "BashInput", "BashOutput", "KillShell"] {
17987            assert!(
17988                disabled.iter().any(|name| name == tool),
17989                "typed read-only child must not advertise {tool}"
17990            );
17991        }
17992        assert!(matches!(
17993            claude_spec.executor,
17994            ExecutorSpec::ClaudeCode {
17995                permission_mode: Some(ref mode),
17996                inherit_user_config: Some(false),
17997                ..
17998            } if mode == "plan"
17999        ));
18000        assert_eq!(claude_spec.capabilities.permission_effective_mode, "plan");
18001
18002        let mut codex_exec = codex_executor(Some("inherit"), Some(true));
18003        if let ExecutorSpec::Codex {
18004            sandbox,
18005            approval_policy,
18006            network_access,
18007            allow_danger_bypass,
18008            permission_profile,
18009            ..
18010        } = &mut codex_exec
18011        {
18012            *sandbox = Some("danger-full-access".to_string());
18013            *approval_policy = Some("on-failure".to_string());
18014            *network_access = Some(true);
18015            *allow_danger_bypass = Some(true);
18016            *permission_profile = Some("full-access".to_string());
18017        }
18018        let codex_runner = ActorChildRunner::new(
18019            "test-actor".into(),
18020            PathBuf::from("/bin/false"),
18021            vec![],
18022            std::env::temp_dir().join("bamboo-test-plan-codex"),
18023            codex_exec,
18024            vec![],
18025            "openai".into(),
18026            4,
18027        );
18028        let codex_spec = codex_runner.build_spec(&planner, &job_for("planner-codex"));
18029        assert_eq!(
18030            expected_permission_executor_mapping(
18031                &codex_spec.executor,
18032                codex_spec.capabilities.permission_resolution().unwrap(),
18033                false,
18034            )
18035            .unwrap()
18036            .as_deref(),
18037            Some("codex_exec:approval_policy=never")
18038        );
18039        assert!(matches!(
18040            codex_spec.executor,
18041            ExecutorSpec::Codex {
18042                sandbox: Some(ref sandbox),
18043                inherit_user_config: Some(false),
18044                approval_policy: Some(ref approval_policy),
18045                network_access: Some(false),
18046                allow_danger_bypass: Some(false),
18047                permission_profile: Some(ref profile),
18048                ..
18049            } if sandbox == "read-only"
18050                && approval_policy == "never"
18051                && profile == "read-only"
18052        ));
18053
18054        let mut app_server = codex_executor(Some("inherit"), Some(true));
18055        if let ExecutorSpec::Codex {
18056            mode,
18057            sandbox,
18058            approval_policy,
18059            network_access,
18060            allow_danger_bypass,
18061            permission_profile,
18062            ..
18063        } = &mut app_server
18064        {
18065            *mode = Some("app_server".to_string());
18066            *sandbox = Some("danger-full-access".to_string());
18067            *approval_policy = Some("on-request".to_string());
18068            *network_access = Some(true);
18069            *allow_danger_bypass = Some(true);
18070            *permission_profile = Some("full-access".to_string());
18071        }
18072        let app_server_runner = ActorChildRunner::new(
18073            "test-actor".into(),
18074            PathBuf::from("/bin/false"),
18075            vec![],
18076            std::env::temp_dir().join("bamboo-test-plan-codex-app-server"),
18077            app_server,
18078            vec![],
18079            "openai".into(),
18080            4,
18081        );
18082        let app_server_spec =
18083            app_server_runner.build_spec(&planner, &job_for("planner-codex-app-server"));
18084        assert_eq!(
18085            expected_permission_executor_mapping(
18086                &app_server_spec.executor,
18087                app_server_spec
18088                    .capabilities
18089                    .permission_resolution()
18090                    .unwrap(),
18091                false,
18092            )
18093            .unwrap()
18094            .as_deref(),
18095            Some("codex_app_server:approvalPolicy=never")
18096        );
18097        assert!(matches!(
18098            app_server_spec.executor,
18099            ExecutorSpec::Codex {
18100                mode: Some(ref mode),
18101                sandbox: Some(ref sandbox),
18102                inherit_user_config: Some(false),
18103                approval_policy: Some(ref approval_policy),
18104                network_access: Some(false),
18105                allow_danger_bypass: Some(false),
18106                permission_profile: Some(ref profile),
18107                ..
18108            } if mode == "app_server"
18109                && sandbox == "read-only"
18110                && approval_policy == "on-request"
18111                && profile == "read-only"
18112        ));
18113    }
18114
18115    #[test]
18116    fn typed_read_only_rejects_both_codex_executor_modes() {
18117        let mut planner = session_of_role("planner", "inspect and plan");
18118        planner
18119            .agent_runtime_state
18120            .get_or_insert_with(bamboo_domain::AgentRuntimeState::default)
18121            .read_only = true;
18122
18123        for mode in ["exec", "app_server"] {
18124            let mut executor = codex_executor(Some("inherit"), Some(true));
18125            if let ExecutorSpec::Codex {
18126                mode: configured_mode,
18127                ..
18128            } = &mut executor
18129            {
18130                *configured_mode = Some(mode.to_string());
18131            }
18132            let runner = ActorChildRunner::new(
18133                "test-actor".into(),
18134                PathBuf::from("/bin/false"),
18135                vec![],
18136                std::env::temp_dir().join(format!("bamboo-test-plan-codex-{mode}")),
18137                executor,
18138                vec![],
18139                "openai".into(),
18140                4,
18141            );
18142            let spec = runner.build_spec(&planner, &job_for(&format!("planner-codex-{mode}")));
18143            let error = ensure_read_only_activation_is_enforceable(&spec)
18144                .expect_err("Codex command execution must fail closed for read-only activations");
18145            let message = error.to_string();
18146            assert!(message.contains("Codex executor"), "{message}");
18147            assert!(message.contains("command execution"), "{message}");
18148        }
18149    }
18150
18151    #[tokio::test]
18152    async fn typed_read_only_codex_fails_before_worker_spawn_or_dispatch() {
18153        let mut planner = session_of_role("planner", "inspect and plan");
18154        planner
18155            .agent_runtime_state
18156            .get_or_insert_with(bamboo_domain::AgentRuntimeState::default)
18157            .read_only = true;
18158
18159        let runner = ActorChildRunner::new(
18160            "test-actor".into(),
18161            PathBuf::from("/bin/false"),
18162            vec![],
18163            std::env::temp_dir().join("bamboo-test-plan-codex-no-spawn"),
18164            codex_executor(Some("inherit"), Some(true)),
18165            vec![],
18166            "openai".into(),
18167            4,
18168        );
18169        let (event_tx, _event_rx) = mpsc::channel::<AgentEvent>(4);
18170        let error = runner
18171            .execute_external_child(
18172                &mut planner,
18173                &job_for("planner-codex-no-spawn"),
18174                event_tx,
18175                CancellationToken::new(),
18176            )
18177            .await
18178            .expect_err("typed read-only Codex must fail before invoking /bin/false");
18179        let message = error.to_string();
18180        assert!(message.contains("Codex executor"), "{message}");
18181        assert!(message.contains("command execution"), "{message}");
18182    }
18183
18184    #[test]
18185    fn ticket_placement_keeps_local_and_ordinary_remote_execution_available() {
18186        let runner = bogus_runner(HashMap::new());
18187        let ordinary = session_of_role("worker", "ordinary work");
18188        for placement in [
18189            Placement::Local,
18190            Placement::Remote {
18191                endpoint: "wss://fixture.invalid/actor".into(),
18192            },
18193            Placement::Schedulable {
18194                pool: "fixture-pool".into(),
18195            },
18196        ] {
18197            let mut spec = runner.build_spec(&ordinary, &job_for("placement"));
18198            spec.placement = placement;
18199            ensure_ticket_placement_is_enforceable(&ordinary, &spec).unwrap();
18200            for marker in [
18201                crate::ticket_runtime::TICKET_DISPATCH_KEY,
18202                crate::ticket_worker_plan::TICKET_LOCAL_PLAN_KEY,
18203            ] {
18204                let mut ticket = ordinary.clone();
18205                ticket.metadata.insert(marker.into(), "fixture".into());
18206                assert_eq!(
18207                    ensure_ticket_placement_is_enforceable(&ticket, &spec).is_ok(),
18208                    matches!(spec.placement, Placement::Local)
18209                );
18210            }
18211        }
18212    }
18213
18214    #[tokio::test]
18215    async fn ticket_placement_rejected_before_remote_route_claim_or_worker_spawn() {
18216        for marker in [
18217            crate::ticket_runtime::TICKET_DISPATCH_KEY,
18218            crate::ticket_worker_plan::TICKET_LOCAL_PLAN_KEY,
18219        ] {
18220            for schedulable in [false, true] {
18221                let mut session = session_of_role("worker", "Ticket work");
18222                session.metadata.insert(marker.into(), "fixture".into());
18223                let before = serde_json::to_value(&session).unwrap();
18224                let remote = if schedulable {
18225                    HashMap::new()
18226                } else {
18227                    HashMap::from([(
18228                        "worker".into(),
18229                        ResolvedRemotePlacement {
18230                            endpoint: "wss://fixture.invalid/actor".into(),
18231                            ..Default::default()
18232                        },
18233                    )])
18234                };
18235                let sched = if schedulable {
18236                    HashMap::from([(
18237                        "worker".into(),
18238                        sched_placement("fixture-pool", "wss://fixture.invalid"),
18239                    )])
18240                } else {
18241                    HashMap::new()
18242                };
18243                let runner = bogus_sched_runner(remote, sched);
18244                let (events, mut observations) = mpsc::channel::<AgentEvent>(4);
18245                let error = runner
18246                    .execute_external_child(
18247                        &mut session,
18248                        &job_for("ticket-no-remote"),
18249                        events,
18250                        CancellationToken::new(),
18251                    )
18252                    .await
18253                    .unwrap_err();
18254                assert!(
18255                    error
18256                        .to_string()
18257                        .contains("Ticket execution requires local placement"),
18258                    "{error}"
18259                );
18260                assert_eq!(serde_json::to_value(&session).unwrap(), before);
18261                assert!(observations.try_recv().is_err());
18262            }
18263        }
18264    }
18265
18266    #[test]
18267    fn typed_read_only_rejects_resident_actor_placements() {
18268        let runner = bogus_runner(HashMap::new());
18269        let mut planner = session_of_role("planner", "inspect and plan");
18270        planner
18271            .agent_runtime_state
18272            .get_or_insert_with(bamboo_domain::AgentRuntimeState::default)
18273            .read_only = true;
18274
18275        for placement in [
18276            Placement::Remote {
18277                endpoint: "wss://resident.example/actor".to_string(),
18278            },
18279            Placement::Schedulable {
18280                pool: "resident-pool".to_string(),
18281            },
18282        ] {
18283            let mut spec = runner.build_spec(&planner, &job_for("read-only-resident"));
18284            spec.placement = placement;
18285
18286            let error = ensure_read_only_activation_is_enforceable(&spec)
18287                .expect_err("resident placement must fail closed for read-only activation");
18288            let message = error.to_string();
18289            assert!(message.contains("requires local placement"), "{message}");
18290            assert!(message.contains("read-only tool boundary"), "{message}");
18291        }
18292
18293        let local = runner.build_spec(&planner, &job_for("read-only-local"));
18294        ensure_read_only_activation_is_enforceable(&local)
18295            .expect("a locally provisioned read-only worker is enforceable");
18296    }
18297
18298    #[tokio::test]
18299    async fn build_spec_preserves_exact_inherited_permission_mode_for_child_worker() {
18300        // Exercise the real creation path instead of pre-seeding the child by
18301        // hand: both legacy Bypass and zero-prompt Auto must survive child
18302        // creation and the actor provisioning boundary without collapsing.
18303        for (label, mode) in [
18304            ("bypass", bamboo_domain::SessionPermissionMode::Bypass),
18305            ("auto", bamboo_domain::SessionPermissionMode::Auto),
18306        ] {
18307            let runner = bogus_runner(HashMap::new());
18308            let mut parent = Session::new(format!("parent-{label}"), "test-model");
18309            parent
18310                .agent_runtime_state
18311                .get_or_insert_with(bamboo_domain::AgentRuntimeState::default)
18312                .set_permission_mode(mode);
18313            let workspace = tempfile::tempdir().expect("workspace fixture");
18314            let port = RecordingChildSessionPort {
18315                parent: parent.clone(),
18316                saved: Default::default(),
18317            };
18318            let child_id = format!("child-{label}-{}", uuid::Uuid::new_v4());
18319            crate::session_app::child_session::create_child_action(
18320                &port,
18321                crate::session_app::child_session::CreateChildInput {
18322                    parent_session: parent,
18323                    child_id: child_id.clone(),
18324                    title: format!("{label} child"),
18325                    responsibility: "Run ordinary commands".to_string(),
18326                    assignment_prompt: "run an ordinary command".to_string(),
18327                    subagent_type: "explorer".to_string(),
18328                    workspace: workspace.path().to_string_lossy().into_owned(),
18329                    workspace_source: crate::project_context::WorkspaceSource::Explicit,
18330                    model_override: None,
18331                    model_ref_override: None,
18332                    runtime_metadata: HashMap::new(),
18333                    read_only: false,
18334                    auto_run: false,
18335                    reasoning_effort: None,
18336                    lifecycle: None,
18337                    resident_name: None,
18338                    resident_context: None,
18339                    disabled_tools: None,
18340                    context_fork: None,
18341                },
18342            )
18343            .await
18344            .expect("create inherited-permission child");
18345            let child = port.saved_child();
18346
18347            assert_eq!(
18348                child
18349                    .agent_runtime_state
18350                    .as_ref()
18351                    .map(bamboo_domain::AgentRuntimeState::effective_permission_mode),
18352                Some(mode),
18353                "create_child_action must inherit {label} from the parent"
18354            );
18355
18356            let spec = runner.build_spec(&child, &job_for(&child_id));
18357
18358            assert_eq!(
18359                spec.capabilities.bypass,
18360                mode == bamboo_domain::SessionPermissionMode::Bypass
18361            );
18362            assert_eq!(
18363                spec.capabilities.auto_approve_permissions,
18364                mode == bamboo_domain::SessionPermissionMode::Auto
18365            );
18366            assert!(
18367                spec.capabilities.enforce_permissions,
18368                "policy evaluation must remain active under {label}"
18369            );
18370        }
18371    }
18372
18373    #[tokio::test]
18374    async fn child_resident_and_guardian_inherit_project_through_actor_run_spec() {
18375        let project_id = bamboo_domain::ProjectId::parse("project-inherited").expect("Project id");
18376        let workspace = tempfile::tempdir().expect("workspace fixture");
18377
18378        for (role, lifecycle, resident_name) in [
18379            ("explorer", None, None),
18380            ("resident", Some("resident"), Some("stable-reviewer")),
18381            ("guardian", None, None),
18382        ] {
18383            let mut parent = Session::new(format!("parent-{role}"), "test-model");
18384            parent.set_project_id_meta(project_id.to_string());
18385            let port = RecordingChildSessionPort {
18386                parent: parent.clone(),
18387                saved: Default::default(),
18388            };
18389            let child_id = format!("child-{role}-{}", uuid::Uuid::new_v4());
18390            crate::session_app::child_session::create_child_action(
18391                &port,
18392                crate::session_app::child_session::CreateChildInput {
18393                    parent_session: parent,
18394                    child_id: child_id.clone(),
18395                    title: format!("{role} child"),
18396                    responsibility: "Review the assigned work".to_string(),
18397                    assignment_prompt: "inspect the change".to_string(),
18398                    subagent_type: role.to_string(),
18399                    workspace: workspace.path().to_string_lossy().into_owned(),
18400                    workspace_source: crate::project_context::WorkspaceSource::Explicit,
18401                    model_override: None,
18402                    model_ref_override: None,
18403                    runtime_metadata: HashMap::new(),
18404                    read_only: false,
18405                    auto_run: false,
18406                    reasoning_effort: None,
18407                    lifecycle: lifecycle.map(str::to_string),
18408                    resident_name: resident_name.map(str::to_string),
18409                    resident_context: None,
18410                    disabled_tools: None,
18411                    context_fork: None,
18412                },
18413            )
18414            .await
18415            .expect("create Project-inheriting child");
18416            let child = port.saved_child();
18417            assert_eq!(
18418                crate::project_context::ProjectContextResolver::project_id_from_session(&child),
18419                Some(project_id.clone()),
18420                "{role} child must inherit its parent's Project"
18421            );
18422
18423            assert_eq!(
18424                project_id_for_actor_run(&child).expect("valid actor Project identity"),
18425                Some(project_id.clone()),
18426                "{role} actor RunSpec must preserve inherited Project identity"
18427            );
18428        }
18429    }
18430
18431    #[test]
18432    fn placement_metadata_stamps_remote_and_schedulable_not_local() {
18433        // Local children carry no stamp — the DTO defaults them to the backend host.
18434        assert_eq!(placement_metadata(&Placement::Local, None), None);
18435
18436        // Remote, no node label → host derived from the endpoint.
18437        let r = placement_metadata(
18438            &Placement::Remote {
18439                endpoint: "wss://10.0.0.5:8443/stream".into(),
18440            },
18441            None,
18442        )
18443        .unwrap();
18444        assert!(r.contains(r#""kind":"remote""#), "{r}");
18445        assert!(r.contains(r#""host":"10.0.0.5""#), "{r}");
18446
18447        // A cluster node's label (its metadata) OVERRIDES the raw endpoint host.
18448        let labeled = placement_metadata(
18449            &Placement::Remote {
18450                endpoint: "ws://169.254.230.101:8899".into(),
18451            },
18452            Some("mini"),
18453        )
18454        .unwrap();
18455        assert!(labeled.contains(r#""host":"mini""#), "{labeled}");
18456
18457        // Schedulable → {kind:"remote", host:<node label, else pool>}.
18458        let s = placement_metadata(
18459            &Placement::Schedulable {
18460                pool: "explorers".into(),
18461            },
18462            Some("mini"),
18463        )
18464        .unwrap();
18465        assert!(s.contains(r#""kind":"remote""#), "{s}");
18466        assert!(s.contains(r#""host":"mini""#), "{s}");
18467
18468        // The stamp round-trips through the storage placement type.
18469        let p: bamboo_storage::SessionPlacement = serde_json::from_str(&labeled).unwrap();
18470        assert_eq!(p.kind, "remote");
18471        assert_eq!(p.host, "mini");
18472    }
18473
18474    /// A legacy direct worker endpoint cannot bypass the broker activation
18475    /// contract, even when the endpoint is reachable and has a bearer token.
18476    #[tokio::test]
18477    async fn execute_external_child_rejects_direct_remote_worker() {
18478        let mut placements = HashMap::new();
18479        placements.insert(
18480            "explorer".to_string(),
18481            ResolvedRemotePlacement {
18482                endpoint: "ws://127.0.0.1:1".into(),
18483                token: Some("remote-test-token".into()),
18484                ca_cert_file: None,
18485                host_label: Some("mini-e2e".into()),
18486                broker_peer: None,
18487                requirements: None,
18488            },
18489        );
18490        let runner = bogus_runner(placements);
18491        let mut session = session_of_role("explorer", "hello remote");
18492        let job = job_for("child-1");
18493        let (event_tx, _event_rx) = mpsc::channel::<AgentEvent>(64);
18494        let error = runner
18495            .execute_external_child(&mut session, &job, event_tx, CancellationToken::new())
18496            .await
18497            .expect_err("direct remote transport must fail closed");
18498        assert!(error
18499            .to_string()
18500            .contains("remote_broker_activation_unavailable"));
18501        assert!(session
18502            .messages
18503            .iter()
18504            .all(|message| message.role != Role::Assistant));
18505    }
18506
18507    // ---- #181 (P2b): schedulable placement routing --------------------------
18508
18509    /// A bogus-worker_bin runner carrying SCHEDULABLE placements (and optionally
18510    /// remote ones, to test precedence). A local spawn here would fail on
18511    /// `/bin/false`, so a passing schedulable test proves no subprocess spawned.
18512    fn bogus_sched_runner(
18513        remote: HashMap<String, ResolvedRemotePlacement>,
18514        sched: HashMap<String, ResolvedSchedulablePlacement>,
18515    ) -> ActorChildRunner {
18516        ActorChildRunner::new(
18517            "test-actor".into(),
18518            PathBuf::from("/bin/false"),
18519            vec![],
18520            std::env::temp_dir().join("bamboo-test-fab-181"),
18521            ExecutorSpec::Echo,
18522            vec![],
18523            "anthropic".into(),
18524            4,
18525        )
18526        .with_remote_placements(remote)
18527        .with_schedulable_placements(sched)
18528    }
18529
18530    fn sched_placement(
18531        pool: &str,
18532        _registry_url: impl Into<String>,
18533    ) -> ResolvedSchedulablePlacement {
18534        ResolvedSchedulablePlacement {
18535            pool: pool.into(),
18536            host_label: None,
18537            requirements: None,
18538            ..Default::default()
18539        }
18540    }
18541
18542    #[test]
18543    fn build_spec_sets_schedulable_placement_for_matching_role() {
18544        let mut sched = HashMap::new();
18545        sched.insert(
18546            "explorer".to_string(),
18547            sched_placement("gpu-pool", "unused"),
18548        );
18549        let runner = bogus_sched_runner(HashMap::new(), sched);
18550
18551        let s = session_of_role("explorer", "do the thing");
18552        let spec = runner.build_spec(&s, &job_for("child-1"));
18553        match &spec.placement {
18554            Placement::Schedulable { pool } => assert_eq!(pool, "gpu-pool"),
18555            other => panic!("expected Schedulable, got {other:?}"),
18556        }
18557        // No per-placement bearer now — the bus connection carries the bus token.
18558        assert!(spec.secrets.worker_auth_token.is_none());
18559    }
18560
18561    #[test]
18562    fn build_spec_remote_wins_when_role_in_both_maps() {
18563        // A role present in BOTH remote_placements and schedulable_placements must
18564        // resolve to the FIXED remote placement (documented precedence).
18565        let mut remote = HashMap::new();
18566        remote.insert(
18567            "explorer".to_string(),
18568            ResolvedRemotePlacement {
18569                endpoint: "wss://fixed-host:8443".into(),
18570                token: Some("T-remote".into()),
18571                ca_cert_file: None,
18572                host_label: None,
18573                broker_peer: None,
18574                requirements: None,
18575            },
18576        );
18577        let mut sched = HashMap::new();
18578        sched.insert(
18579            "explorer".to_string(),
18580            sched_placement("gpu-pool", "https://control-plane:9562"),
18581        );
18582        let runner = bogus_sched_runner(remote, sched);
18583
18584        let s = session_of_role("explorer", "do the thing");
18585        let spec = runner.build_spec(&s, &job_for("child-1"));
18586        match &spec.placement {
18587            Placement::Remote { endpoint } => assert_eq!(endpoint, "wss://fixed-host:8443"),
18588            other => panic!("expected Remote (precedence), got {other:?}"),
18589        }
18590        assert!(spec.secrets.worker_auth_token.is_none());
18591    }
18592
18593    #[test]
18594    fn build_spec_local_for_unmatched_schedulable_role() {
18595        let mut sched = HashMap::new();
18596        sched.insert(
18597            "explorer".to_string(),
18598            sched_placement("gpu-pool", "https://control-plane:9562"),
18599        );
18600        let runner = bogus_sched_runner(HashMap::new(), sched);
18601        let s = session_of_role("writer", "do the thing");
18602        let spec = runner.build_spec(&s, &job_for("child-1"));
18603        assert_eq!(spec.placement, Placement::Local);
18604        assert!(spec.secrets.worker_auth_token.is_none());
18605    }
18606
18607    /// The full role → resolved-placement → badge-host chain: a child routed to a
18608    /// remote/schedulable placement carrying a cluster node's `host_label` stamps
18609    /// that label; without a label it falls back to the endpoint host / pool; a
18610    /// Local child gets no stamp (the DTO defaults it to the backend host).
18611    #[test]
18612    fn placement_stamp_uses_node_label_for_remote_and_schedulable() {
18613        // Remote WITH a node label → {remote, <label>}, overriding the raw IP.
18614        let mut remote = HashMap::new();
18615        remote.insert(
18616            "explorer".to_string(),
18617            ResolvedRemotePlacement {
18618                endpoint: "ws://169.254.230.101:8899".into(),
18619                token: None,
18620                ca_cert_file: None,
18621                host_label: Some("mini".into()),
18622                broker_peer: None,
18623                requirements: None,
18624            },
18625        );
18626        let runner = bogus_runner(remote);
18627        let spec = runner.build_spec(&session_of_role("explorer", "go"), &job_for("c1"));
18628        let stamp = runner
18629            .placement_stamp_for(&spec)
18630            .expect("remote child is stamped");
18631        assert!(stamp.contains(r#""kind":"remote""#), "{stamp}");
18632        assert!(stamp.contains(r#""host":"mini""#), "{stamp}");
18633
18634        // Remote WITHOUT a node label → falls back to the endpoint host.
18635        let mut remote_nolabel = HashMap::new();
18636        remote_nolabel.insert(
18637            "explorer".to_string(),
18638            ResolvedRemotePlacement {
18639                endpoint: "ws://169.254.230.101:8899".into(),
18640                token: None,
18641                ca_cert_file: None,
18642                host_label: None,
18643                broker_peer: None,
18644                requirements: None,
18645            },
18646        );
18647        let r2 = bogus_runner(remote_nolabel);
18648        let spec2 = r2.build_spec(&session_of_role("explorer", "go"), &job_for("c1"));
18649        assert!(r2
18650            .placement_stamp_for(&spec2)
18651            .unwrap()
18652            .contains(r#""host":"169.254.230.101""#));
18653
18654        // Schedulable WITH a node label → {remote, <label>} (a node, not a pool name).
18655        let mut sched = HashMap::new();
18656        sched.insert(
18657            "mac-mini-monitor".to_string(),
18658            ResolvedSchedulablePlacement {
18659                pool: "mac-mini-monitor".into(),
18660                host_label: Some("mini".into()),
18661                requirements: None,
18662                ..Default::default()
18663            },
18664        );
18665        let sr = bogus_sched_runner(HashMap::new(), sched);
18666        let spec3 = sr.build_spec(&session_of_role("mac-mini-monitor", "go"), &job_for("c1"));
18667        let stamp3 = sr
18668            .placement_stamp_for(&spec3)
18669            .expect("scheduled child is stamped");
18670        assert!(stamp3.contains(r#""kind":"remote""#), "{stamp3}");
18671        assert!(stamp3.contains(r#""host":"mini""#), "{stamp3}");
18672
18673        // A Local (unmatched) child gets NO stamp.
18674        let local = bogus_runner(HashMap::new());
18675        let spec4 = local.build_spec(&session_of_role("writer", "go"), &job_for("c1"));
18676        assert_eq!(local.placement_stamp_for(&spec4), None);
18677    }
18678
18679    #[tokio::test]
18680    async fn legacy_schedulable_route_fails_closed_without_local_fallback() {
18681        let mut sched = HashMap::new();
18682        sched.insert(
18683            "explorer".to_string(),
18684            sched_placement("gpu-pool", "unused"),
18685        );
18686        let runner = bogus_sched_runner(HashMap::new(), sched);
18687        let mut session = session_of_role("explorer", "do the thing");
18688        let (event_tx, _rx) = mpsc::channel::<AgentEvent>(8);
18689        let error = runner
18690            .execute_external_child(
18691                &mut session,
18692                &job_for("child-1"),
18693                event_tx,
18694                CancellationToken::new(),
18695            )
18696            .await
18697            .expect_err("legacy bus presence cannot grant placement");
18698        assert!(error
18699            .to_string()
18700            .contains("remote_broker_activation_unavailable"));
18701    }
18702}