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    use bamboo_agent_core::storage::Storage;
7152    use bamboo_agent_core::tools::{ToolCtx, ToolOutcome};
7153
7154    let caller = load_active_subagent_caller(
7155        store,
7156        binding,
7157        run_id,
7158        child_id,
7159        parent_id,
7160        logical_session,
7161        creation,
7162        plain_input,
7163    )
7164    .await
7165    .ok_or_else(|| AgentError::LLM("nested wait caller is no longer active".into()))?;
7166    if caller
7167        .agent_runtime_state
7168        .as_ref()
7169        .and_then(|state| state.waiting_for_children.as_ref())
7170        .is_some()
7171    {
7172        return Ok(true);
7173    }
7174
7175    let active = store
7176        .list_child_run_statuses(child_id)
7177        .await
7178        .map_err(|error| AgentError::LLM(format!("nested child status scan failed: {error}")))?
7179        .into_iter()
7180        .filter(|(_, status)| {
7181            !status.as_deref().is_some_and(|status| {
7182                matches!(
7183                    status,
7184                    "completed" | "error" | "timeout" | "cancelled" | "skipped"
7185                )
7186            })
7187        })
7188        .map(|(id, _)| id)
7189        .collect::<Vec<_>>();
7190    if active.is_empty() {
7191        return Ok(false);
7192    }
7193
7194    let mut ctx = ToolCtx::none(format!("host-nested-orphan-wait:{run_id}"));
7195    ctx.session_id = Some(Arc::from(caller.id.as_str()));
7196    let outcome = tokio::time::timeout(
7197        Duration::from_secs(30),
7198        tool.invoke(
7199            serde_json::json!({"action":"wait", "child_session_ids":active}),
7200            ctx,
7201        ),
7202    )
7203    .await
7204    .map_err(|_| AgentError::LLM("nested wait registration timed out".into()))?
7205    .map_err(|error| AgentError::LLM(format!("nested wait registration failed: {error}")))?;
7206    let wait_registered = match outcome {
7207        ToolOutcome::Completed(result) if result.success => {
7208            let payload: serde_json::Value = serde_json::from_str(&result.result)
7209                .map_err(|_| AgentError::LLM("nested wait result is invalid".into()))?;
7210            payload["status"] == "waiting"
7211        }
7212        _ => {
7213            return Err(AgentError::LLM(
7214                "nested wait registration was not completed".into(),
7215            ))
7216        }
7217    };
7218
7219    // A grandchild may finish while the wait is being registered and clear it.
7220    // A clear wait is safe only when no canonical descendants remain active.
7221    let latest = load_active_subagent_caller(
7222        store,
7223        binding,
7224        run_id,
7225        child_id,
7226        parent_id,
7227        logical_session,
7228        creation,
7229        plain_input,
7230    )
7231    .await
7232    .ok_or_else(|| AgentError::LLM("nested wait caller changed during registration".into()))?;
7233    if latest
7234        .agent_runtime_state
7235        .as_ref()
7236        .and_then(|state| state.waiting_for_children.as_ref())
7237        .is_some()
7238    {
7239        return Ok(true);
7240    }
7241    let still_active = store
7242        .list_child_run_statuses(child_id)
7243        .await
7244        .map_err(|error| {
7245            AgentError::LLM(format!("nested child wait verification failed: {error}"))
7246        })?
7247        .into_iter()
7248        .any(|(_, status)| {
7249            !status.as_deref().is_some_and(|status| {
7250                matches!(
7251                    status,
7252                    "completed" | "error" | "timeout" | "cancelled" | "skipped"
7253                )
7254            })
7255        });
7256    if still_active {
7257        return Err(AgentError::LLM(
7258            "nested child remains active without a durable parent wait".into(),
7259        ));
7260    }
7261    // Keep the current Run's suspend intent even if the last grandchild
7262    // finished and cleared the Host wait before SDK finalization. Its outcome
7263    // has already admitted a successor; publishing this Run as completed
7264    // would release the grandparent on the Child's interim response.
7265    Ok(wait_registered)
7266}
7267
7268fn actor_event_route_error(error: ActorEventRouteError) -> AgentError {
7269    let message = match error {
7270        ActorEventRouteError::SequenceGap { .. } => {
7271            "actor event sequence gap; authoritative session snapshot is required"
7272        }
7273        ActorEventRouteError::LifecycleMissing => {
7274            "actor emitted live content before its activation lifecycle event"
7275        }
7276        ActorEventRouteError::ForgedSession => {
7277            "actor event payload targets a different canonical session"
7278        }
7279        ActorEventRouteError::ConflictingReplay => {
7280            "actor replay reused an event sequence with different content"
7281        }
7282        ActorEventRouteError::StaleAuthority => "actor event owner is stale",
7283        ActorEventRouteError::SequenceExhausted => "actor event sequence is exhausted",
7284    };
7285    AgentError::LLM(message.into())
7286}
7287
7288const MAX_DISPLAY_CALLS: usize = 4_096;
7289const MAX_NESTED_DISPLAY_DEPTH: usize = 8;
7290const MAX_DISPLAY_ID_BYTES: usize = 256;
7291const MAX_DISPLAY_HOOK_RESULT_DEPTH: usize = 16;
7292
7293fn hook_result_for_display(result: HookResult, depth: usize) -> Option<HookResult> {
7294    if depth >= MAX_DISPLAY_HOOK_RESULT_DEPTH {
7295        return None;
7296    }
7297    Some(match result {
7298        HookResult::Continue => HookResult::Continue,
7299        HookResult::Mutated => HookResult::Mutated,
7300        HookResult::Allow => HookResult::Allow,
7301        HookResult::Ask => HookResult::Ask,
7302        HookResult::Deny { .. } => HookResult::Deny {
7303            reason: "Hook denied".into(),
7304        },
7305        HookResult::InjectContext { .. } => HookResult::InjectContext {
7306            text: "Hook context hidden".into(),
7307        },
7308        HookResult::WithContext { result, .. } => HookResult::WithContext {
7309            result: Box::new(hook_result_for_display(*result, depth + 1)?),
7310            text: "Hook context hidden".into(),
7311        },
7312        HookResult::Suspend { .. } => HookResult::Suspend {
7313            reason: "Hook suspended".into(),
7314        },
7315        HookResult::Abort { .. } => HookResult::Abort {
7316            reason: "Hook aborted".into(),
7317        },
7318    })
7319}
7320
7321#[derive(Clone, Copy, PartialEq, Eq)]
7322enum DisplayCallKind {
7323    Ordinary,
7324    PrivateBrowser,
7325    Reused,
7326}
7327
7328#[derive(Clone, Copy)]
7329struct DisplayCall {
7330    kind: DisplayCallKind,
7331    started: bool,
7332}
7333
7334#[derive(Default)]
7335struct ActorEventDisplay {
7336    calls: HashMap<String, DisplayCall>,
7337    hook_names: HashMap<String, usize>,
7338    children: HashMap<(String, String), Box<ActorEventDisplay>>,
7339    // Shared by every nested child: a per-level cap alone permits an
7340    // exponential number of identities across an actor event tree.
7341    identities: Arc<AtomicUsize>,
7342    overflowed: bool,
7343    nested_overflowed: bool,
7344}
7345
7346fn is_browser_tool_name(tool_name: &str) -> bool {
7347    tool_name
7348        .trim()
7349        .rsplit("::")
7350        .next()
7351        .is_some_and(|name| name.eq_ignore_ascii_case("browser"))
7352}
7353
7354fn known_browser_action(args: &serde_json::Value) -> bool {
7355    matches!(
7356        args.get("action").and_then(serde_json::Value::as_str),
7357        Some(
7358            "tabs"
7359                | "new_tab"
7360                | "activate_tab"
7361                | "close_tab"
7362                | "navigate"
7363                | "history"
7364                | "viewport"
7365                | "snapshot"
7366                | "click"
7367                | "click_at"
7368                | "hover"
7369                | "drag"
7370                | "fill"
7371                | "select_option"
7372                | "set_file_input"
7373                | "type"
7374                | "press"
7375                | "key"
7376                | "scroll"
7377                | "download"
7378                | "screenshot"
7379                | "dialog_respond"
7380        )
7381    )
7382}
7383
7384fn is_private_browser_call(tool_name: &str, args: &serde_json::Value) -> bool {
7385    is_browser_tool_name(tool_name)
7386        && (!known_browser_action(args)
7387            || args
7388                .get("action")
7389                .and_then(serde_json::Value::as_str)
7390                .is_some_and(|action| {
7391                    action.eq_ignore_ascii_case("download") || action == "set_file_input"
7392                })
7393            || args.get("data_base64").is_some())
7394}
7395
7396fn actor_arguments_for_display(tool_name: &str, args: &serde_json::Value) -> serde_json::Value {
7397    if is_browser_tool_name(tool_name) {
7398        // A malformed or unknown action has no reliable privacy class. A
7399        // mixed-version actor may emit it before its schema rejects it.
7400        if !known_browser_action(args) {
7401            return serde_json::json!({"details":"[redacted]"});
7402        }
7403    }
7404    tool_start_arguments_for_display(tool_name, args)
7405}
7406
7407impl ActorEventDisplay {
7408    fn fail_closed(&mut self) {
7409        self.identities
7410            .store(MAX_DISPLAY_CALLS + 1, Ordering::Relaxed);
7411        self.overflowed = true;
7412        self.calls.clear();
7413        self.hook_names.clear();
7414        self.nested_overflowed = true;
7415        self.children.clear();
7416    }
7417
7418    fn reserve_identity(&mut self) -> bool {
7419        if self.identities.fetch_add(1, Ordering::Relaxed) < MAX_DISPLAY_CALLS {
7420            true
7421        } else {
7422            self.fail_closed();
7423            false
7424        }
7425    }
7426
7427    fn remember_start(&mut self, id: &str, private: bool) -> DisplayCallKind {
7428        if self.overflowed || id.is_empty() || id.len() > MAX_DISPLAY_ID_BYTES {
7429            return DisplayCallKind::Reused;
7430        }
7431        let kind = if private {
7432            DisplayCallKind::PrivateBrowser
7433        } else {
7434            DisplayCallKind::Ordinary
7435        };
7436        if let Some(call) = self.calls.get_mut(id) {
7437            if call.started || call.kind != kind {
7438                call.kind = DisplayCallKind::Reused;
7439            }
7440            call.started = true;
7441            return call.kind;
7442        }
7443        if !self.reserve_identity() {
7444            return DisplayCallKind::Reused;
7445        }
7446        self.calls.insert(
7447            id.to_string(),
7448            DisplayCall {
7449                kind,
7450                started: true,
7451            },
7452        );
7453        kind
7454    }
7455
7456    fn remember_approval(&mut self, id: &str, private: bool) -> DisplayCallKind {
7457        if self.overflowed || id.is_empty() || id.len() > MAX_DISPLAY_ID_BYTES {
7458            return DisplayCallKind::Reused;
7459        }
7460        let kind = if private {
7461            DisplayCallKind::PrivateBrowser
7462        } else {
7463            DisplayCallKind::Ordinary
7464        };
7465        if let Some(call) = self.calls.get_mut(id) {
7466            if call.kind != kind {
7467                call.kind = DisplayCallKind::Reused;
7468            }
7469            return call.kind;
7470        }
7471        if !self.reserve_identity() {
7472            return DisplayCallKind::Reused;
7473        }
7474        self.calls.insert(
7475            id.to_string(),
7476            DisplayCall {
7477                kind,
7478                started: false,
7479            },
7480        );
7481        kind
7482    }
7483
7484    fn known_kind(&self, id: &str) -> Option<DisplayCallKind> {
7485        if self.overflowed || self.identities.load(Ordering::Relaxed) > MAX_DISPLAY_CALLS {
7486            return None;
7487        }
7488        self.calls
7489            .get(id)
7490            .filter(|call| call.started)
7491            .map(|call| call.kind)
7492    }
7493
7494    fn display_hook_name(&mut self, hook_name: &str) -> Option<String> {
7495        if self.overflowed || hook_name.is_empty() || hook_name.len() > MAX_DISPLAY_ID_BYTES {
7496            self.fail_closed();
7497            return None;
7498        }
7499        if let Some(index) = self.hook_names.get(hook_name) {
7500            return Some(format!("actor-hook-{index}"));
7501        }
7502        if !self.reserve_identity() {
7503            return None;
7504        }
7505        let index = self.hook_names.len() + 1;
7506        self.hook_names.insert(hook_name.to_string(), index);
7507        Some(format!("actor-hook-{index}"))
7508    }
7509
7510    fn project(&mut self, event: AgentEvent, depth: usize) -> Option<AgentEvent> {
7511        let call_id = match &event {
7512            AgentEvent::ToolStart { tool_call_id, .. }
7513            | AgentEvent::ToolApprovalRequested { tool_call_id, .. }
7514            | AgentEvent::ToolToken { tool_call_id, .. }
7515            | AgentEvent::ToolComplete { tool_call_id, .. }
7516            | AgentEvent::ToolError { tool_call_id, .. }
7517            | AgentEvent::ToolLifecycle { tool_call_id, .. } => Some(tool_call_id),
7518            _ => None,
7519        };
7520        if call_id.is_some_and(|id| id.is_empty() || id.len() > MAX_DISPLAY_ID_BYTES) {
7521            self.fail_closed();
7522            return None;
7523        }
7524        if self.identities.load(Ordering::Relaxed) > MAX_DISPLAY_CALLS {
7525            self.fail_closed();
7526        }
7527        Some(match event {
7528            AgentEvent::ToolStart {
7529                tool_call_id,
7530                tool_name,
7531                arguments,
7532            } => {
7533                let private = is_private_browser_call(&tool_name, &arguments);
7534                let kind = self.remember_start(&tool_call_id, private);
7535                let arguments = if kind == DisplayCallKind::Reused {
7536                    serde_json::json!({"details":"[redacted]"})
7537                } else {
7538                    actor_arguments_for_display(&tool_name, &arguments)
7539                };
7540                AgentEvent::ToolStart {
7541                    tool_call_id,
7542                    tool_name: match kind {
7543                        DisplayCallKind::Reused => "unknown".into(),
7544                        _ if is_browser_tool_name(&tool_name) => "browser".into(),
7545                        _ => tool_name,
7546                    },
7547                    arguments,
7548                }
7549            }
7550            AgentEvent::ToolApprovalRequested {
7551                tool_call_id,
7552                tool_name,
7553                parameters,
7554            } => {
7555                let private = is_private_browser_call(&tool_name, &parameters);
7556                let kind = self.remember_approval(&tool_call_id, private);
7557                let parameters = if kind == DisplayCallKind::Reused {
7558                    serde_json::json!({"details":"[redacted]"})
7559                } else {
7560                    actor_arguments_for_display(&tool_name, &parameters)
7561                };
7562                AgentEvent::ToolApprovalRequested {
7563                    tool_call_id,
7564                    tool_name: match kind {
7565                        DisplayCallKind::Reused => "unknown".into(),
7566                        _ if is_browser_tool_name(&tool_name) => "browser".into(),
7567                        _ => tool_name,
7568                    },
7569                    parameters,
7570                }
7571            }
7572            AgentEvent::ToolToken {
7573                tool_call_id,
7574                content,
7575            } => {
7576                let content = match self.known_kind(&tool_call_id) {
7577                    Some(DisplayCallKind::Ordinary) => content,
7578                    Some(DisplayCallKind::PrivateBrowser) => "Browser tool output hidden".into(),
7579                    _ => "Tool output hidden".into(),
7580                };
7581                AgentEvent::ToolToken {
7582                    tool_call_id,
7583                    content,
7584                }
7585            }
7586            AgentEvent::ToolComplete {
7587                tool_call_id,
7588                result,
7589            } => {
7590                let result = match self.known_kind(&tool_call_id) {
7591                    Some(DisplayCallKind::Ordinary) => result,
7592                    Some(DisplayCallKind::PrivateBrowser) => {
7593                        bamboo_agent_core::tools::ToolResult::text(
7594                            result.success,
7595                            "Browser tool result hidden",
7596                        )
7597                    }
7598                    _ => bamboo_agent_core::tools::ToolResult::text(
7599                        result.success,
7600                        "Tool result hidden",
7601                    ),
7602                };
7603                AgentEvent::ToolComplete {
7604                    tool_call_id,
7605                    result,
7606                }
7607            }
7608            AgentEvent::ToolError {
7609                tool_call_id,
7610                error,
7611            } => {
7612                let error = match self.known_kind(&tool_call_id) {
7613                    Some(DisplayCallKind::Ordinary) => error,
7614                    Some(DisplayCallKind::PrivateBrowser) => "Browser tool failed".into(),
7615                    _ => "Tool failed".into(),
7616                };
7617                AgentEvent::ToolError {
7618                    tool_call_id,
7619                    error,
7620                }
7621            }
7622            AgentEvent::ToolLifecycle {
7623                tool_call_id,
7624                tool_name,
7625                phase,
7626                elapsed_ms,
7627                is_mutating,
7628                auto_approved,
7629                summary,
7630                error,
7631            } => {
7632                let kind = self.known_kind(&tool_call_id);
7633                let (summary, error) = match kind {
7634                    Some(DisplayCallKind::Ordinary) => (summary, error),
7635                    Some(DisplayCallKind::PrivateBrowser) => (
7636                        summary.map(|_| "Browser tool activity hidden".into()),
7637                        error.map(|_| "Browser tool failed".into()),
7638                    ),
7639                    _ => (
7640                        summary.map(|_| "Tool activity hidden".into()),
7641                        error.map(|_| "Tool failed".into()),
7642                    ),
7643                };
7644                AgentEvent::ToolLifecycle {
7645                    tool_call_id,
7646                    tool_name: match kind {
7647                        Some(DisplayCallKind::Ordinary) => tool_name,
7648                        Some(DisplayCallKind::PrivateBrowser) => "browser".into(),
7649                        _ => "unknown".into(),
7650                    },
7651                    phase: if kind == Some(DisplayCallKind::Ordinary)
7652                        || matches!(phase.as_str(), "begin" | "finished" | "error" | "cancelled")
7653                    {
7654                        phase
7655                    } else {
7656                        "event".into()
7657                    },
7658                    elapsed_ms,
7659                    is_mutating,
7660                    auto_approved,
7661                    summary,
7662                    error,
7663                }
7664            }
7665            AgentEvent::HookLifecycle {
7666                hook_name,
7667                point,
7668                phase,
7669                duration_ms,
7670                decision,
7671            } => {
7672                let Some(decision) = hook_result_for_display(decision, 0) else {
7673                    self.fail_closed();
7674                    return None;
7675                };
7676                let hook_name = self.display_hook_name(&hook_name)?;
7677                let phase = match phase.as_str() {
7678                    "completed" | "started" | "error" => phase,
7679                    _ => "event".into(),
7680                };
7681                AgentEvent::HookLifecycle {
7682                    hook_name,
7683                    point,
7684                    phase,
7685                    duration_ms,
7686                    decision,
7687                }
7688            }
7689            AgentEvent::ChildApprovalRequested {
7690                child_session_id,
7691                request_id,
7692                tool_name,
7693                permission,
7694                resource,
7695            } => {
7696                let (tool_name, permission, resource) = if is_browser_tool_name(&tool_name) {
7697                    (
7698                        "browser".into(),
7699                        "Browser interaction approval".into(),
7700                        "[redacted]".into(),
7701                    )
7702                } else {
7703                    (tool_name, permission, resource)
7704                };
7705                AgentEvent::ChildApprovalRequested {
7706                    child_session_id,
7707                    request_id,
7708                    tool_name,
7709                    permission,
7710                    resource,
7711                }
7712            }
7713            AgentEvent::ChildApprovalChanged { tool_name, .. }
7714                if is_browser_tool_name(&tool_name) =>
7715            {
7716                // This durable approval delta is host-owned. An actor-supplied
7717                // copy is neither authoritative nor safe to mirror to the
7718                // account feed; the host publishes its own record separately.
7719                return None;
7720            }
7721            AgentEvent::SubAgentEvent {
7722                parent_session_id,
7723                child_session_id,
7724                event,
7725            } => {
7726                if depth >= MAX_NESTED_DISPLAY_DEPTH || self.nested_overflowed {
7727                    return None;
7728                }
7729                if parent_session_id.is_empty()
7730                    || child_session_id.is_empty()
7731                    || parent_session_id.len() > MAX_DISPLAY_ID_BYTES
7732                    || child_session_id.len() > MAX_DISPLAY_ID_BYTES
7733                {
7734                    self.fail_closed();
7735                    return None;
7736                }
7737                let key = (parent_session_id.clone(), child_session_id.clone());
7738                if !self.children.contains_key(&key) && !self.reserve_identity() {
7739                    return None;
7740                }
7741                let identities = self.identities.clone();
7742                let child = self.children.entry(key).or_insert_with(|| {
7743                    Box::new(ActorEventDisplay {
7744                        identities,
7745                        ..Default::default()
7746                    })
7747                });
7748                let event = child.project(*event, depth + 1)?;
7749                if self.identities.load(Ordering::Relaxed) > MAX_DISPLAY_CALLS {
7750                    self.fail_closed();
7751                    return None;
7752                }
7753                AgentEvent::SubAgentEvent {
7754                    parent_session_id,
7755                    child_session_id,
7756                    event: Box::new(event),
7757                }
7758            }
7759            other => other,
7760        })
7761    }
7762}
7763
7764#[allow(clippy::too_many_arguments)]
7765async fn process_actor_event(
7766    event: serde_json::Value,
7767    strict_permission_events: bool,
7768    permission_handshake: &mut PermissionPostureHandshake,
7769    expected_permission_posture: Option<&ExpectedPermissionPosture>,
7770    session_inbox_runtime: Option<&SessionInboxRuntimeBinding>,
7771    logical_session: &mut Session,
7772    event_tx: &mpsc::Sender<AgentEvent>,
7773    display: &mut ActorEventDisplay,
7774) -> crate::runtime::runner::Result<()> {
7775    let event = match serde_json::from_value::<AgentEvent>(event) {
7776        Ok(event) => event,
7777        Err(_) if strict_permission_events => {
7778            // Serde errors can quote an unrecognized variant or field value.
7779            // The actor event is untrusted display input, including browser
7780            // download metadata, so never echo its decode details to SSE.
7781            return Err(AgentError::LLM(
7782                "actor emitted malformed AgentEvent under a typed permission posture contract"
7783                    .to_string(),
7784            ));
7785        }
7786        Err(_) => {
7787            // A legacy actor can omit a malformed ToolStart that reuses an
7788            // earlier public call ID. Its later result must not inherit the
7789            // stale public display classification.
7790            display.fail_closed();
7791            return Ok(());
7792        }
7793    };
7794    process_typed_actor_event(
7795        event,
7796        permission_handshake,
7797        expected_permission_posture,
7798        session_inbox_runtime,
7799        logical_session,
7800        event_tx,
7801        display,
7802    )
7803    .await
7804}
7805
7806#[allow(clippy::too_many_arguments)]
7807async fn process_typed_actor_event(
7808    event: AgentEvent,
7809    permission_handshake: &mut PermissionPostureHandshake,
7810    expected_permission_posture: Option<&ExpectedPermissionPosture>,
7811    session_inbox_runtime: Option<&SessionInboxRuntimeBinding>,
7812    logical_session: &mut Session,
7813    event_tx: &mpsc::Sender<AgentEvent>,
7814    display: &mut ActorEventDisplay,
7815) -> crate::runtime::runner::Result<()> {
7816    if matches!(&event, AgentEvent::PermissionPostureActivated { .. }) {
7817        if permission_handshake.posture_was_confirmed() {
7818            return Err(AgentError::LLM(
7819                "actor emitted a duplicate permission posture activation".to_string(),
7820            ));
7821        }
7822        let seed = permission_posture_seed_from_event(logical_session, &event)
7823            .map_err(AgentError::LLM)?
7824            .ok_or_else(|| {
7825                AgentError::LLM(
7826                    "actor permission posture event did not decode as a posture".to_string(),
7827                )
7828            })?;
7829        if let Some(expected) = expected_permission_posture {
7830            if seed.policy_revision != expected.policy_revision
7831                || seed.resolution != expected.resolution
7832            {
7833                return Err(AgentError::LLM(
7834                    "permission posture event does not match the host-dispatched policy"
7835                        .to_string(),
7836                ));
7837            }
7838            if seed.executor_mapping() != expected.executor_mapping {
7839                return Err(AgentError::LLM(
7840                    "permission posture event does not match the host-dispatched executor mapping"
7841                        .to_string(),
7842                ));
7843            }
7844        }
7845        if let Some(binding) = session_inbox_runtime {
7846            let saved = binding
7847                .persistence
7848                .record_permission_posture_activation(
7849                    &logical_session.id,
7850                    expected_permission_posture
7851                        .and_then(|expected| expected.expected_audit_revision),
7852                    &seed,
7853                )
7854                .await
7855                .map_err(|error| {
7856                    AgentError::LLM(format!(
7857                        "persist child permission posture bootstrap: {error}"
7858                    ))
7859                })?
7860                .ok_or_else(|| {
7861                    AgentError::LLM(
7862                        "persist child permission posture bootstrap: session not found".to_string(),
7863                    )
7864                })?;
7865            let snapshot = bamboo_domain::PermissionAuditSnapshot::from_metadata(&saved.metadata)
7866                .ok_or_else(|| {
7867                AgentError::LLM(
7868                    "persisted child permission posture audit is incomplete".to_string(),
7869                )
7870            })?;
7871            snapshot.write_to(&mut logical_session.metadata);
7872        } else {
7873            // In-memory/custom actor embeddings still use a host-owned clock.
7874            // Durable server paths always take the atomic branch.
7875            bamboo_domain::record_permission_audit(&mut logical_session.metadata, &seed, None)
7876                .map_err(|error| {
7877                    AgentError::LLM(format!(
7878                        "record in-memory child permission posture: {error}"
7879                    ))
7880                })?;
7881        }
7882        // Confirmation is deliberately last: matching alone is not enough.
7883        // The host-owned audit write must succeed first.
7884        *permission_handshake = PermissionPostureHandshake::Confirmed;
7885    } else if permission_handshake.is_awaiting() {
7886        return Err(AgentError::LLM(
7887            "actor emitted an execution event before permission posture confirmation".to_string(),
7888        ));
7889    }
7890    if let Some(event) = display.project(event, 0) {
7891        let _ = event_tx.send(event).await;
7892    }
7893    Ok(())
7894}
7895
7896/// Pump child frames -> parent events until a terminal frame (or cancellation).
7897/// On success, yields the actor's final result text (for session write-back).
7898/// `live_rx` carries in-band frames (steering messages) from the live registry.
7899///
7900/// `escalation_bridge` (#68) is the per-run escalation host bridge CAPTURED BY
7901/// VALUE at spawn time in `execute_external_child` (NOT read live here): when a
7902/// non-bypass child re-proxies an approval request, this owned bridge routes it
7903/// UP to the parent run. Owning it for the call's lifetime is what lets a
7904/// fire-and-forget grandchild that outlives its spawning run still escalate to
7905/// the correct (then-current) parent bridge rather than a stale/overwritten one.
7906async fn drive(context: ActorDriveContext<'_>) -> crate::runtime::runner::Result<Option<String>> {
7907    let ActorDriveContext {
7908        client,
7909        parent_session_id,
7910        child_session_id,
7911        child_attempt,
7912        approval_registry,
7913        approval_decider,
7914        approval_reviewer,
7915        escalation_bridge,
7916        event_tx,
7917        cancel_token,
7918        live_rx,
7919        delivery_rx,
7920        logical_session,
7921        mut expected_permission_posture,
7922        expected_creation,
7923        session_inbox_runtime,
7924        actor_directory_store,
7925        canonical_subagent_tool,
7926        ticket_service,
7927        activation_run_id,
7928        execution_epoch,
7929        expected_source_actor_id,
7930        initial_inflight_claims,
7931        first_frame_timeout,
7932        plain_actor,
7933        remote_environment_lease,
7934        mut readonly_output,
7935        local_history_tools,
7936        local_history_read_only,
7937        mut strict_history_output,
7938        plain_input,
7939        canonical_activation,
7940        canonical_placement_ref,
7941        actor_event_observer,
7942        plain_run,
7943    } = context;
7944
7945    // A projection belongs only to this Host-validated Run. A loaded Session
7946    // or a prior attempt may carry a stale marker; only the canonical wait
7947    // checks in the terminal branch below may set it again.
7948    logical_session
7949        .metadata
7950        .remove("runtime.actor_nested_wait_handoff");
7951
7952    // First-frame watchdog: a live worker emits its first frame (run-started /
7953    // first token) within seconds; total silence past the deadline means the
7954    // worker is dead (e.g. a pooled worker that exited right after checkout), so
7955    // its Run sits queued forever. We trip ONLY before the first frame — once any
7956    // frame arrives the worker is proven live and a legitimately long run (a slow
7957    // tool between tokens) never trips it.
7958    let mut got_first_frame = false;
7959    let mut first_frame_watch = first_frame_timeout.map(|d| Box::pin(tokio::time::sleep(d)));
7960    let mut inflight_claims = initial_inflight_claims;
7961    let mut owned_inputs = plain_input
7962        .map(|activation| VecDeque::from(activation.initial_inputs.clone()))
7963        .unwrap_or_default();
7964    // A retry/recovery or completed initial batch consumes its cold allowance;
7965    // only a fresh activation can continue twice. No lease/watchdog restarts.
7966    let continuation_limit = if owned_inputs.is_empty() { 2 } else { 1 };
7967    let mut continuations = usize::from(!owned_inputs.is_empty());
7968    let mut released_inputs: Vec<(
7969        SessionInboxOwnedClaim,
7970        bamboo_subagent::proto::InitialInputRelease,
7971    )> = Vec::new();
7972    let mut current_epoch = execution_epoch;
7973    let approval_epoch = Arc::new(AtomicU64::new(execution_epoch));
7974    let approval_cancel = cancel_token.child_token();
7975    let _approval_lifetime = approval_cancel.clone().drop_guard();
7976    let mut approval_scopes: HashMap<String, (serde_json::Value, ChildApprovalScope)> =
7977        HashMap::new();
7978    let mut owned_tree_requests = 0usize;
7979    let mut canonical_subagent_requests = 0usize;
7980    let mut committed_plan_calls = HashMap::new();
7981    // Once this Run registers a canonical Child wait, a fast completion may
7982    // clear it before the Worker sends Terminal. Keep the current Run
7983    // nonterminal so its already admitted outcome belongs to the successor.
7984    let mut nested_wait_registered_this_run = false;
7985    let mut parent_question_checkpoint_id: Option<String> = None;
7986    let strict_permission_events = expected_permission_posture.is_some();
7987    let mut permission_handshake =
7988        PermissionPostureHandshake::new(expected_permission_posture.as_ref());
7989    let mut next_actor_event_seq = 1u64;
7990    let mut canonical_router: Option<ActorEventRouter> = None;
7991    let canonical_fence = plain_input
7992        .map(|activation| &activation.fence)
7993        .or(canonical_activation);
7994    let canonical_store = plain_input
7995        .map(|activation| activation.store.as_ref())
7996        .or(actor_directory_store);
7997    let remote_claim_fence = actor_directory_store.zip(canonical_activation);
7998    let workflow_usage_requested = logical_session
7999        .metadata
8000        .get(bamboo_subagent::proto::WORKFLOW_USAGE_REQUESTED_KEY)
8001        .is_some_and(|v| v == "true");
8002    if workflow_usage_requested {
8003        logical_session
8004            .metadata
8005            .remove(bamboo_subagent::proto::WORKFLOW_USAGE_OBSERVATION_KEY);
8006        logical_session
8007            .metadata
8008            .remove(bamboo_subagent::proto::WORKFLOW_TERMINAL_OBSERVATION_KEY);
8009    }
8010    let mut remote_cancel_deadline: Option<tokio::time::Instant> = None;
8011    let mut display = ActorEventDisplay::default();
8012    let mut readonly = readonly_output
8013        .as_ref()
8014        .map(|_| ReadOnlyActorCollector::default());
8015    let mut local_history = local_history_tools.map(|_| LocalToolCollector::default());
8016    loop {
8017        tokio::select! {
8018            _ = cancel_token.cancelled(), if remote_cancel_deadline.is_none() => {
8019                if remote_claim_fence.is_some() || workflow_usage_requested {
8020                    // A remote Worker may finish after its subscriber cancels.
8021                    // Keep the same correlated frame pump alive briefly so an
8022                    // exact Cancelled Outcome can terminalize the Actor fence.
8023                    // Without that proof the activation stays live until its
8024                    // lease expires and a retry must remain blocked.
8025                    let _ = client.send(ParentFrame::Cancel).await;
8026                    remote_cancel_deadline = Some(
8027                        tokio::time::Instant::now() + Duration::from_secs(10)
8028                    );
8029                    continue;
8030                }
8031                // Direct/local links retain their existing cancel behavior.
8032                break;
8033            }
8034            _ = async {
8035                if let Some(deadline) = remote_cancel_deadline {
8036                    tokio::time::sleep_until(deadline).await;
8037                } else {
8038                    std::future::pending::<()>().await;
8039                }
8040            } => {
8041                return Err(AgentError::Cancelled);
8042            }
8043            _ = async {
8044                match first_frame_watch.as_mut() {
8045                    Some(s) => s.as_mut().await,
8046                    None => std::future::pending::<()>().await,
8047                }
8048            }, if !got_first_frame => {
8049                return Err(AgentError::WorkerUnresponsive(format!(
8050                    "child {child_session_id} produced no frame within {:?}",
8051                    first_frame_timeout.unwrap_or_default()
8052                )));
8053            }
8054            Some(_generation) = delivery_rx.recv(),
8055                if session_inbox_runtime.is_some() && activation_run_id.is_some() =>
8056            {
8057                if plain_input.is_some() {
8058                    // Keep the real Inbox pending until this Run's completed
8059                    // reply can be committed BEFORE its correction User.
8060                    continue;
8061                }
8062                forward_next_canonical_claim(
8063                    client,
8064                    session_inbox_runtime.expect("guarded"),
8065                    logical_session,
8066                    activation_run_id.expect("guarded"),
8067                    &mut inflight_claims,
8068                    remote_claim_fence,
8069                )
8070                .await?;
8071            }
8072            Some(frame) = live_rx.recv() => {
8073                if plain_actor { return Err(plain_actor_unsupported()); }
8074                // Forward in-band steering to the worker over the existing WS.
8075                if client.send(frame).await.is_err() {
8076                    tracing::warn!("live steering frame could not be sent; connection failing");
8077                }
8078            }
8079            frame = client.next_frame() => {
8080                // Any frame (event / approval / terminal / close / error) proves
8081                // the worker responded — disarm the first-frame watchdog.
8082                got_first_frame = true;
8083                first_frame_watch = None;
8084                match frame {
8085                    Ok(Some(ChildFrame::Event { event })) => {
8086                        if canonical_activation.is_some() {
8087                            return Err(AgentError::LLM("remote_actor_event_batch_required".into()));
8088                        }
8089                        if local_history.is_some() {
8090                            return Err(local_tool_history_unsupported());
8091                        }
8092                        if event.get("initial_input_control").is_some() {
8093                            let bamboo_subagent::proto::InitialInputControl::Request { request } =
8094                                bamboo_subagent::proto::InitialInputControl::decode(event).map_err(|_| plain_actor_unsupported())?
8095                            else { return Err(plain_actor_unsupported()); };
8096                            if !permission_handshake.posture_was_confirmed() || cancel_token.is_cancelled() { return Err(plain_actor_unsupported()); }
8097                            let activation = plain_input.ok_or_else(|| plain_actor_unsupported())?;
8098                            let binding = session_inbox_runtime.ok_or_else(|| plain_actor_unsupported())?;
8099                            let (run, _) = plain_run.ok_or_else(|| plain_actor_unsupported())?;
8100                            // Exact retransmission may refer to an earlier
8101                            // release while the next input is still waiting.
8102                            // Every new request must match the ordered front.
8103                            let previous = released_inputs.iter()
8104                                .find(|(_, release)| release.request == request);
8105                            let (claim, previous) = match previous {
8106                                Some((claim, release)) => (claim.clone(), Some(release.clone())),
8107                                None => (owned_inputs.front().cloned()
8108                                    .ok_or_else(|| plain_actor_unsupported())?, None),
8109                            };
8110                            let release = activation.release_initial_input(binding, logical_session, run, current_epoch, &claim,
8111                                &request, &mut expected_permission_posture, previous.as_ref()).await?;
8112                            if cancel_token.is_cancelled() { return Err(AgentError::Cancelled); }
8113                            client.send(ParentFrame::InitialInputRelease { release: release.clone() }).await
8114                                .map_err(|_| AgentError::LLM("Actor initial release unconfirmed; durable input is preserved".into()))?;
8115                            if previous.is_none() {
8116                                let _ = owned_inputs.pop_front();
8117                                released_inputs.push((claim, release));
8118                            }
8119                            continue;
8120                        }
8121                        if expected_creation.is_some() {
8122                            return Err(AgentError::LLM("worker omitted required Child creation event identity".into()));
8123                        }
8124                        if event["type"] == bamboo_subagent::proto::LocalToolMessages::TYPE { continue; }
8125                        if event["type"] == "owned_readonly_transcript" {
8126                            if plain_actor { return Err(plain_actor_unsupported()); }
8127                            continue;
8128                        }
8129                        if plain_actor && !plain_actor_event(&event)? { continue; }
8130                        if let Some(activation) = plain_input {
8131                            validate_directory_actor_event(
8132                                activation.store.as_ref(),
8133                                &activation.fence,
8134                                logical_session,
8135                                None,
8136                            ).await?;
8137                        }
8138                        // Rolling-upgrade compatibility: old actors have no
8139                        // route/sequence metadata, but retain the same typed
8140                        // permission handshake and event validation.
8141                        process_actor_event(
8142                            event,
8143                            strict_permission_events,
8144                            &mut permission_handshake,
8145                            expected_permission_posture.as_ref(),
8146                            session_inbox_runtime,
8147                            logical_session,
8148                            event_tx,
8149                            &mut display,
8150                        )
8151                        .await?;
8152                    }
8153                    Ok(Some(ChildFrame::EventBatch { batch })) => {
8154                        validate_actor_event_batch(
8155                            &batch,
8156                            logical_session,
8157                            parent_session_id,
8158                            activation_run_id,
8159                            current_epoch,
8160                            expected_source_actor_id,
8161                            expected_creation,
8162                        )?;
8163                        let directory_entry = if let (Some(store), Some(fence)) = (canonical_store, canonical_fence) {
8164                            let entry = validate_directory_actor_event(
8165                                store, fence, logical_session, batch.activation_id.as_deref(),
8166                            ).await?;
8167                            if canonical_placement_ref.is_some_and(|expected| {
8168                                entry.activation.as_ref().and_then(|activation| activation.placement_ref.as_ref()) != Some(expected)
8169                            }) {
8170                                return Err(AgentError::LLM("remote_actor_placement_ref_changed".into()));
8171                            }
8172                            Some(entry)
8173                        } else { None };
8174                        if let (Some(entry), Some(fence)) = (&directory_entry, canonical_fence) {
8175                            if canonical_router.is_none() {
8176                                canonical_router = Some(
8177                                    ActorEventRouter::new(
8178                                        entry,
8179                                        fence,
8180                                        current_epoch,
8181                                        strict_permission_events && permission_handshake.is_awaiting(),
8182                                    )
8183                                    .map_err(actor_event_route_error)?,
8184                                );
8185                            }
8186                        }
8187                        if local_history_tools.is_some()
8188                            && (batch.first_seq != next_actor_event_seq
8189                                || batch.qos != bamboo_subagent::ActorEventQos::Durable
8190                                || batch.last_seq == u64::MAX) {
8191                            return Err(local_tool_history_unsupported());
8192                        }
8193                        if batch.last_seq < next_actor_event_seq {
8194                            continue;
8195                        }
8196                        if batch.first_seq > next_actor_event_seq {
8197                            if plain_actor {
8198                                return Err(AgentError::LLM(
8199                                    "actor event sequence gap; authoritative session snapshot is required".into(),
8200                                ));
8201                            }
8202                            tracing::warn!(
8203                                child_session_id,
8204                                expected_seq = next_actor_event_seq,
8205                                received_seq = batch.first_seq,
8206                                "actor event sequence gap; authoritative session snapshot is required"
8207                            );
8208                            // A missing ToolStart may have reused an earlier
8209                            // call ID for a private browser download. No prior
8210                            // identity remains safe for display after a gap.
8211                            display.fail_closed();
8212                        }
8213                        let skip = next_actor_event_seq
8214                            .saturating_sub(batch.first_seq)
8215                            .min(batch.events.len() as u64) as usize;
8216                        for (offset, event) in batch.events.into_iter().enumerate().skip(skip) {
8217                            let seq = batch.first_seq + offset as u64;
8218                            let mut publish = true;
8219                            if event["type"] == bamboo_subagent::proto::WorkflowAgentUsage::TYPE {
8220                                if workflow_usage_requested {
8221                                    if permission_handshake.is_awaiting() { return Err(plain_actor_unsupported()); }
8222                                    let usage: bamboo_subagent::proto::WorkflowAgentUsage = serde_json::from_value(event.clone()).map_err(|_| plain_actor_unsupported())?;
8223                                    let activation = activation_run_id.ok_or_else(plain_actor_unsupported)?;
8224                                    if !usage.matches(logical_session, activation) { return Err(plain_actor_unsupported()); }
8225                                    if logical_session.metadata.contains_key(bamboo_subagent::proto::WORKFLOW_USAGE_OBSERVATION_KEY) { return Err(plain_actor_unsupported()); }
8226                                    logical_session.metadata.insert(bamboo_subagent::proto::WORKFLOW_USAGE_OBSERVATION_KEY.into(), serde_json::to_string(&usage).map_err(|_| plain_actor_unsupported())?);
8227                                }
8228                                publish = false;
8229                            }
8230                            if publish { if let Some(collector) = local_history.as_mut() {
8231                                if permission_handshake.is_awaiting() && event["type"] != "permission_posture_activated" {
8232                                    return Err(local_tool_history_unsupported());
8233                                }
8234                                if logical_session.metadata.contains_key(crate::ticket_worker_plan::TICKET_LOCAL_PLAN_KEY) {
8235                                    publish = collector.ticket_event(&event, &committed_plan_calls)?;
8236                                }
8237                                publish &= collector.event(&event)?;
8238                            } else if event["type"] == bamboo_subagent::proto::LocalToolMessages::TYPE {
8239                                // Completion DATA is private and grants nothing on an unselected route.
8240                                publish = false;
8241                            }}
8242                            if publish { publish = if let Some(collector) = readonly.as_mut() {
8243                                if permission_handshake.is_awaiting()
8244                                    && event["type"] != "permission_posture_activated" {
8245                                    return Err(plain_actor_unsupported());
8246                                }
8247                                collector.event(&event)?
8248                            } else if event["type"] == "owned_readonly_transcript" {
8249                                if plain_actor { return Err(plain_actor_unsupported()); }
8250                                false // A worker cache observation is not a legacy Host commit.
8251                            } else if plain_actor { plain_actor_event(&event)? } else { true };
8252                            }
8253                            if let (Some(router), Some(entry), Some(fence)) =
8254                                (canonical_router.as_mut(), directory_entry.as_ref(), canonical_fence)
8255                            {
8256                                let typed = if publish {
8257                                    Some(serde_json::from_value::<AgentEvent>(event.clone())
8258                                        .map_err(|_| plain_actor_unsupported())?)
8259                                } else {
8260                                    None
8261                                };
8262                                match router.route(entry, fence, current_epoch, seq, typed)
8263                                    .map_err(actor_event_route_error)?
8264                                {
8265                                    ActorEventRoute::Publish(envelope) => {
8266                                        let public_event = actor_event_observer
8267                                            .filter(|_| envelope.class != ActorEventClass::Ephemeral)
8268                                            .map(|_| PublicActorEvent::from(envelope.as_ref()));
8269                                        tracing::trace!(
8270                                            actor_id = %envelope.actor_id,
8271                                            event_id = %envelope.event_id,
8272                                            class = ?envelope.class,
8273                                            "publishing canonical Actor event"
8274                                        );
8275                                        process_typed_actor_event(
8276                                            envelope.payload,
8277                                            &mut permission_handshake,
8278                                            expected_permission_posture.as_ref(),
8279                                            session_inbox_runtime,
8280                                            logical_session,
8281                                            event_tx,
8282                                            &mut display,
8283                                        ).await?;
8284                                        if let (Some(observer), Some(public_event)) =
8285                                            (actor_event_observer, public_event)
8286                                        {
8287                                            observer.publish(public_event);
8288                                        }
8289                                    }
8290                                    ActorEventRoute::Duplicate | ActorEventRoute::Suppressed => {}
8291                                }
8292                                next_actor_event_seq = seq.saturating_add(1);
8293                                continue;
8294                            }
8295                            if !publish {
8296                                next_actor_event_seq = seq.saturating_add(1);
8297                                continue;
8298                            }
8299                            process_actor_event(
8300                                event,
8301                                strict_permission_events,
8302                                &mut permission_handshake,
8303                                expected_permission_posture.as_ref(),
8304                                session_inbox_runtime,
8305                                logical_session,
8306                                event_tx,
8307                                &mut display,
8308                            )
8309                            .await?;
8310                            next_actor_event_seq = seq.saturating_add(1);
8311                        }
8312                    }
8313                    Ok(Some(ChildFrame::OwnedTreeRequest { id, cursor })) => {
8314                        owned_tree_requests += 1;
8315                        if owned_tree_requests > 16 {
8316                            return Err(AgentError::LLM("owned tree request limit exceeded".into()));
8317                        }
8318                        let page = if id.len() <= 128
8319                            && !logical_session.metadata.contains_key(crate::ticket_worker_plan::TICKET_LOCAL_PLAN_KEY)
8320                            && cursor.as_ref().is_none_or(|value| !value.is_empty() && value.len() <= 128)
8321                            && !cancel_token.is_cancelled()
8322                        {
8323                            match (actor_directory_store, session_inbox_runtime,
8324                                activation_run_id, expected_creation) {
8325                                (Some(store), Some(binding), Some(run_id), Some(creation)) => {
8326                                    tokio::time::timeout(
8327                                        Duration::from_secs(8),
8328                                        inspect_active_owned_tree(
8329                                            store, binding, run_id, child_session_id,
8330                                            parent_session_id, logical_session, creation,
8331                                            plain_input, cursor.as_deref(),
8332                                        ),
8333                                    ).await.ok().flatten()
8334                                }
8335                                _ => {
8336                                    tracing::warn!("owned tree denied: canonical Host scope unavailable");
8337                                    None
8338                                },
8339                            }
8340                        } else {
8341                            tracing::warn!("owned tree denied: request bounds or cancellation");
8342                            None
8343                        };
8344                        if cancel_token.is_cancelled() {
8345                            return Err(AgentError::Cancelled);
8346                        }
8347                        client.send(ParentFrame::OwnedTreeReply { id, page }).await
8348                            .map_err(|_| AgentError::LLM("owned tree reply transport closed".into()))?;
8349                    }
8350                    Ok(Some(ChildFrame::SubAgentRequest { id, tool_call_id, args })) => {
8351                        canonical_subagent_requests += 1;
8352                        if canonical_subagent_requests > 64 {
8353                            return Err(AgentError::LLM("canonical SubAgent request limit exceeded".into()));
8354                        }
8355                        let valid_shape = !id.is_empty() && id.len() <= 128
8356                            && !tool_call_id.is_empty() && tool_call_id.len() <= 128
8357                            && args.is_object()
8358                            && serde_json::to_vec(&args)
8359                                .is_ok_and(|bytes| bytes.len() <= 64 * 1024);
8360                        let registers_wait = matches!(args.get("action").and_then(|value| value.as_str()),
8361                            Some("wait"))
8362                            || (args.get("action").and_then(|value| value.as_str()) == Some("create")
8363                                && args.get("wait").and_then(|value| value.as_bool()) == Some(true));
8364                        let checkpoint_payload = args.get(PARENT_QUESTION_CHECKPOINT_ACTION).cloned();
8365                        let ticket_bootstrap_pending = valid_shape
8366                            && logical_session.metadata.contains_key(crate::ticket_worker_plan::TICKET_LOCAL_PLAN_KEY)
8367                            && args == serde_json::json!({(crate::ticket_worker_plan::TICKET_PLAN_ACTION): {"read":true}})
8368                            && permission_handshake.is_awaiting()
8369                            && !cancel_token.is_cancelled();
8370                        let result = if ticket_bootstrap_pending {
8371                            // A read control can overtake the queued posture
8372                            // event on the separate lanes. Keep the handshake
8373                            // fence; the exact bootstrap read may retry briefly.
8374                            serde_json::json!({"error":crate::ticket_worker_plan::remote::TICKET_BOOTSTRAP_POSTURE_PENDING})
8375                        } else if valid_shape
8376                            && (!logical_session.metadata.contains_key(crate::ticket_worker_plan::TICKET_LOCAL_PLAN_KEY)
8377                                || args.get(crate::ticket_worker_plan::TICKET_PLAN_ACTION).is_some()
8378                                || args.get(crate::ticket_worker_plan::files::TICKET_FILE_ACTION).is_some())
8379                            && !cancel_token.is_cancelled()
8380                            && !permission_handshake.is_awaiting()
8381                        {
8382                            if args.get(crate::ticket_worker_plan::TICKET_PLAN_ACTION).is_some() {
8383                                match (ticket_service.as_deref(), actor_directory_store,
8384                                    session_inbox_runtime, activation_run_id, expected_creation) {
8385                                    (Some(service), Some(store), Some(binding), Some(run_id), Some(creation)) => {
8386                                        if let Some(caller) = load_active_subagent_caller(
8387                                            store, binding, run_id, child_session_id,
8388                                            parent_session_id, logical_session, creation, plain_input,
8389                                        ).await {
8390                                            match crate::ticket_worker_plan::remote::apply_host_plan_request(service, &caller, run_id, &args, &tool_call_id) {
8391                                                Ok(projection) => {
8392                                                    logical_session.task_list = projection.task_list.clone();
8393                                                    logical_session.set_task_list_version_meta(projection.plan_revision.to_string());
8394                                                    if let Some(question) = &projection.question {
8395                                                        let raw = serde_json::to_string(question).map_err(|_| local_tool_history_unsupported())?;
8396                                                        logical_session.metadata.insert(crate::ticket_worker_plan::TICKET_QUESTION_YIELD_KEY.into(), raw);
8397                                                        if let Some(collector) = local_history.as_mut() {
8398                                                            collector.ticket_yield = Some((tool_call_id.clone(), question.clone()));
8399                                                        }
8400                                                    }
8401                                                    if let Some(task) = args[crate::ticket_worker_plan::TICKET_PLAN_ACTION].get("task") {
8402                                                        committed_plan_calls.insert(tool_call_id.clone(), task.clone());
8403                                                        if let Some(task_list) = projection.task_list.clone() {
8404                                                            let _ = event_tx.send(AgentEvent::TaskListUpdated {
8405                                                                task_list,
8406                                                                version: Some(projection.plan_revision),
8407                                                            }).await;
8408                                                        }
8409                                                    }
8410                                                    serde_json::json!({"result":projection})
8411                                                }
8412                                                Err(error) => serde_json::json!({"error":error.to_string()}),
8413                                            }
8414                                        } else { serde_json::json!({"error":"LocalPlan Child Run is no longer current"}) }
8415                                    }
8416                                    _ => serde_json::json!({"error":"LocalPlan Host authority unavailable"}),
8417                                }
8418                            } else if args.get(crate::ticket_worker_plan::files::TICKET_FILE_ACTION).is_some() {
8419                                match (ticket_service.as_deref(), actor_directory_store, session_inbox_runtime,
8420                                    activation_run_id, expected_creation, local_history_tools) {
8421                                    (Some(service),Some(store),Some(binding),Some(run_id),Some(creation),Some(ceiling)) => {
8422                                        if let Some(caller)=load_active_subagent_caller(store,binding,run_id,child_session_id,parent_session_id,logical_session,creation,plain_input).await {
8423                                            match crate::ticket_worker_plan::files::apply_host_file_request(service,&caller,run_id,&args,&tool_call_id,ceiling) {
8424                                                Ok(reply)=>{
8425                                                    let body=&args[crate::ticket_worker_plan::files::TICKET_FILE_ACTION];
8426                                                    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())?}));
8427                                                    serde_json::json!({"result":reply})
8428                                                },
8429                                                Err(error)=>serde_json::json!({"error":error.to_string()}),
8430                                            }
8431                                        } else { serde_json::json!({"error":"Ticket file Child Run is no longer current"}) }
8432                                    },
8433                                    _=>serde_json::json!({"error":"Ticket file Host authority unavailable"}),
8434                                }
8435                            } else if let Some(payload) = checkpoint_payload {
8436                                let parsed = (args.as_object().is_some_and(|object| object.len() == 1))
8437                                    .then(|| serde_json::from_value::<ParentQuestionCheckpointV1>(payload).ok())
8438                                    .flatten()
8439                                    .filter(|observation| observation.tool_call_id == tool_call_id);
8440                                match (parsed, actor_directory_store, session_inbox_runtime,
8441                                    activation_run_id, expected_creation) {
8442                                    (Some(observation), Some(store), Some(binding), Some(run_id), Some(creation)) => {
8443                                        let receipt = tokio::time::timeout(Duration::from_secs(25),
8444                                            checkpoint_active_parent_question(
8445                                                store, binding, run_id, child_session_id,
8446                                                parent_session_id, logical_session, creation,
8447                                                plain_input, observation,
8448                                            )).await.ok().flatten();
8449                                        if let Some(request_id) = receipt {
8450                                            parent_question_checkpoint_id = Some(request_id.clone());
8451                                            serde_json::json!({"result":{"request_id":request_id}})
8452                                        } else {
8453                                            serde_json::json!({"error":"canonical Child question checkpoint rejected; persisted worker question is preserved"})
8454                                        }
8455                                    }
8456                                    _ => serde_json::json!({"error":"canonical Child question checkpoint unavailable"}),
8457                                }
8458                            } else {
8459                            match (canonical_subagent_tool.as_deref(), actor_directory_store,
8460                                session_inbox_runtime, activation_run_id, expected_creation) {
8461                                (Some(tool), Some(store), Some(binding), Some(run_id), Some(creation)) => {
8462                                    if let Some(caller) = load_active_subagent_caller(
8463                                        store, binding, run_id, child_session_id,
8464                                        parent_session_id, logical_session, creation,
8465                                        plain_input,
8466                                    ).await {
8467                                        tokio::time::timeout(
8468                                            Duration::from_secs(30),
8469                                            invoke_active_subagent(tool, &caller, args, &tool_call_id),
8470                                        ).await.unwrap_or_else(|_| serde_json::json!({
8471                                            "error":"canonical SubAgent operation timed out; inspect logical child state before retrying"
8472                                        }))
8473                                    } else {
8474                                        serde_json::json!({"error":"canonical SubAgent caller is no longer active"})
8475                                    }
8476                                }
8477                                _ => serde_json::json!({"error":"canonical SubAgent Host authority unavailable"}),
8478                            }
8479                            }
8480                        } else {
8481                            serde_json::json!({"error":"canonical SubAgent request rejected"})
8482                        };
8483                        if registers_wait
8484                            && result["result"]["success"] == true
8485                            && result["result"]["result"]
8486                                .as_str()
8487                                .and_then(|value| serde_json::from_str::<serde_json::Value>(value).ok())
8488                                .is_some_and(|value| value["status"] == "waiting")
8489                        {
8490                            nested_wait_registered_this_run = true;
8491                        }
8492                        if cancel_token.is_cancelled() {
8493                            return Err(AgentError::Cancelled);
8494                        }
8495                        client.send(ParentFrame::SubAgentReply { id, result }).await
8496                            .map_err(|_| AgentError::LLM("canonical SubAgent reply transport closed".into()))?;
8497                    }
8498                    Ok(Some(ChildFrame::ApprovalRequest { id, body })) => {
8499                        if logical_session.metadata.contains_key(crate::ticket_worker_plan::TICKET_LOCAL_PLAN_KEY) {
8500                            return Err(AgentError::LLM("Ticket Task-only Worker cannot request unrelated action authority".into()));
8501                        }
8502                        if plain_actor { return Err(plain_actor_unsupported()); }
8503                        if permission_handshake.is_awaiting() {
8504                            return Err(AgentError::LLM(
8505                                "actor requested approval before permission posture confirmation"
8506                                    .to_string(),
8507                            ));
8508                        }
8509                        // Phase 2: a worker proxied a gated-tool approval back to
8510                        // the host. The WORKER side is live — its executor installs
8511                        // a per-run task-local `ApprovalProxy` (subagent_worker.rs)
8512                        // that calls `host.approval_call`, so this frame arrives
8513                        // when a child hits `ConfirmationRequired`.
8514                        if let Some(reviewer) = approval_reviewer
8515                            .cloned()
8516                            .or_else(child_approval_reviewer)
8517                        {
8518                            // Phase 6, Part B: a BYPASSED parent worker
8519                            // model-reviews its children's forced-ask (dangerous)
8520                            // actions. The review is an LLM call, so run it OFF
8521                            // the frame pump in a spawned task and deliver the
8522                            // verdict async via the live channel — the pump keeps
8523                            // forwarding events and the agent loop never blocks. A
8524                            // timeout denies a hung review so the child can't hang.
8525                            let child = child_session_id.to_string();
8526                            let parent = parent_session_id.to_string();
8527                            let req_id = id.clone();
8528                            let body = body.clone();
8529                            let registry = approval_registry.cloned();
8530                            let Some((binding, run)) = session_inbox_runtime.zip(activation_run_id) else {
8531                                // Legacy/custom reviewers retain their original boolean seam;
8532                                // the canonical server reviewer denies without a live scope.
8533                                tokio::spawn(async move {
8534                                    let approved = tokio::time::timeout(CHILD_APPROVAL_TIMEOUT,
8535                                        reviewer.review(&parent, &child, &body)).await.unwrap_or(false);
8536                                    super::live::deliver_approval_scoped(registry.as_ref(), &child,
8537                                        child_attempt, &req_id, approved);
8538                                });
8539                                continue;
8540                            };
8541                            let generation = body.pointer("/permission_request/request_generation")
8542                                .and_then(serde_json::Value::as_str).unwrap_or("").to_string();
8543                            let scope = if let Some((original, scope)) = approval_scopes.get(&generation) {
8544                                if original != &body || scope.stamp()["reply"] != req_id {
8545                                    return Err(AgentError::LLM("approval generation changed in live scope".into()));
8546                                }
8547                                scope.clone()
8548                            } else {
8549                                if approval_scopes.len() >= 64 {
8550                                    return Err(AgentError::LLM("approval live scope limit exceeded".into()));
8551                                }
8552                                let scope = ChildApprovalScope::new(&parent, &child,
8553                                    (child_attempt, run, current_epoch, &req_id, chrono::Utc::now() + chrono::Duration::seconds(240)), binding.router.clone(),
8554                                    approval_cancel.clone(), approval_epoch.clone());
8555                                approval_scopes.insert(generation, (body.clone(), scope.clone()));
8556                                scope
8557                            };
8558                            tokio::spawn(async move {
8559                                let result = tokio::time::timeout(
8560                                    CHILD_APPROVAL_TIMEOUT,
8561                                    reviewer.review_scoped(&parent, &child, &body, &scope),
8562                                ).await.unwrap_or(ChildApprovalReview::Reply(false));
8563                                if let ChildApprovalReview::Reply(approved) = result {
8564                                    if scope.is_current(&parent, &child).await {
8565                                        super::live::deliver_approval_scoped(registry.as_ref(), &child,
8566                                            child_attempt, &req_id, approved && chrono::Utc::now() < scope.deadline());
8567                                    }
8568                                }
8569                            });
8570                        } else if approval_decider.is_some() {
8571                            // A decider is wired (policy / auto): decide promptly
8572                            // and reply inline. (Must not block the pump — see the
8573                            // `ChildApprovalDecider` doc.)
8574                            let approved =
8575                                decide_child_approval(approval_decider, child_session_id, &body)
8576                                    .await;
8577                            if client
8578                                .send(ParentFrame::ApprovalReply { id, approved })
8579                                .await
8580                                .is_err()
8581                            {
8582                                tracing::warn!(
8583                                    "failed to answer approval_request; connection failing"
8584                                );
8585                            }
8586                        } else if let Some(host) = escalation_bridge.clone() {
8587                            // Non-bypass WORKER: ESCALATE up our own actor link
8588                            // (re-proxy) so the request chains to our parent — and
8589                            // up every level until a bypass level or the top
8590                            // orchestrator's model reviewer decides. With no such
8591                            // reviewer the top level fails closed. Off-loop so the
8592                            // pump never blocks; relay the reply down to the child.
8593                            let child = child_session_id.to_string();
8594                            let req_id = id.clone();
8595                            let body = body.clone();
8596                            let registry = approval_registry.cloned();
8597                            tokio::spawn(async move {
8598                                let approved = match tokio::time::timeout(
8599                                    CHILD_APPROVAL_TIMEOUT,
8600                                    host.approval_call(body),
8601                                )
8602                                .await
8603                                {
8604                                    Ok(Ok(reply)) => reply
8605                                        .get("approved")
8606                                        .and_then(|v| v.as_bool())
8607                                        .unwrap_or(false),
8608                                    // Transport error or timeout ⇒ fail closed.
8609                                    _ => false,
8610                                };
8611                                super::live::deliver_approval_scoped(
8612                                    registry.as_ref(),
8613                                    &child,
8614                                    child_attempt,
8615                                    &req_id,
8616                                    approved,
8617                                );
8618                            });
8619                        } else {
8620                            // There is no parent-agent reviewer or upstream actor
8621                            // to own this decision. Never open a manual/UI approval
8622                            // path: forced-ask is parent-reviewed or fail-closed.
8623                            tracing::warn!(
8624                                parent_session_id,
8625                                child_session_id,
8626                                request_id = %id,
8627                                "forced-ask request has no parent-agent reviewer; denying"
8628                            );
8629                            if client
8630                                .send(ParentFrame::ApprovalReply {
8631                                    id,
8632                                    approved: false,
8633                                })
8634                                .await
8635                                .is_err()
8636                            {
8637                                tracing::warn!(
8638                                    "failed to send fail-closed approval reply; connection failing"
8639                                );
8640                            }
8641                        }
8642                    }
8643                    Ok(Some(ChildFrame::SessionMessageAdmitted { confirmation })) => {
8644                        if let Some(activation) = plain_input {
8645                            if activation.initial_release_required { return Err(plain_actor_unsupported()); }
8646                            let binding = session_inbox_runtime.ok_or_else(|| plain_actor_unsupported())?;
8647                            let run_id = activation_run_id.ok_or_else(|| plain_actor_unsupported())?;
8648                            let claim = owned_inputs.front().ok_or_else(|| plain_actor_unsupported())?;
8649                            activation.confirm_input(binding, logical_session, run_id, claim, &confirmation).await?;
8650                            let _ = owned_inputs.pop_front();
8651                            continue;
8652                        }
8653                        let Some(binding) = session_inbox_runtime else {
8654                            tracing::warn!(
8655                                child_session_id,
8656                                "ignoring SessionInbox confirmation without a runtime binding"
8657                            );
8658                            continue;
8659                        };
8660                        let Some(bound_run_id) = activation_run_id else {
8661                            tracing::warn!(
8662                                child_session_id,
8663                                "ignoring SessionInbox confirmation without an activation owner"
8664                            );
8665                            continue;
8666                        };
8667                        let Some(claim) = inflight_claims.front() else {
8668                            tracing::warn!(
8669                                child_session_id,
8670                                envelope_id = %confirmation.envelope_id,
8671                                "rejecting stale SessionInbox confirmation with no in-flight canonical claim"
8672                            );
8673                            continue;
8674                        };
8675                        let exact = confirmation.target_session_id == logical_session.id
8676                            && confirmation.envelope_id == claim.envelope.id.as_str()
8677                            && confirmation.canonical_claim_generation == claim.generation
8678                            && confirmation.activation_run_id == bound_run_id;
8679                        if !exact
8680                            || !binding
8681                                .router
8682                                .owns_run(&logical_session.id, bound_run_id)
8683                                .await
8684                        {
8685                            tracing::warn!(
8686                                child_session_id,
8687                                expected_target = %logical_session.id,
8688                                received_target = %confirmation.target_session_id,
8689                                expected_envelope_id = %claim.envelope.id,
8690                                received_envelope_id = %confirmation.envelope_id,
8691                                expected_generation = claim.generation,
8692                                received_generation = confirmation.canonical_claim_generation,
8693                                expected_run_id = bound_run_id,
8694                                received_run_id = %confirmation.activation_run_id,
8695                                "rejecting stale or mismatched SessionInbox admission confirmation"
8696                            );
8697                            continue;
8698                        }
8699                        let claim = inflight_claims
8700                            .pop_front()
8701                            .expect("validated in-flight canonical claim");
8702                        // On failure the durable canonical cur file remains
8703                        // recoverable for the next owner.
8704                        checkpoint_and_ack_canonical_claim(
8705                            binding,
8706                            logical_session,
8707                            &claim,
8708                            remote_claim_fence,
8709                        )
8710                            .await?;
8711                        // Ordered single-consumer: only after the exact prior
8712                        // claim is checkpointed+acked may the driver claim and
8713                        // forward the next envelope.
8714                        if inflight_claims.is_empty() {
8715                            forward_next_canonical_claim(
8716                                client,
8717                                binding,
8718                                logical_session,
8719                                bound_run_id,
8720                                &mut inflight_claims,
8721                                remote_claim_fence,
8722                            )
8723                            .await?;
8724                        }
8725                    }
8726                    Ok(Some(ChildFrame::Terminal { status, result, error, final_event_watermark, .. })) => {
8727                        if local_history_tools.is_some() {
8728                            // Check before history commit or terminal receipt acceptance.
8729                            // Normal lossy observations do not promise a complete trace.
8730                            validate_strict_history_terminal(
8731                                final_event_watermark.as_ref(), logical_session, parent_session_id,
8732                                activation_run_id, current_epoch, expected_creation,
8733                                next_actor_event_seq - 1,
8734                            )?;
8735                        }
8736                        if remote_cancel_deadline.is_some() && status == TerminalStatus::Cancelled {
8737                            if let Some(tools) = local_history_tools {
8738                                // The watermark proves transport completeness, not
8739                                // persistence. Certify only an entirely validated
8740                                // local Read/Glob prefix with no pending handoff.
8741                                if permission_handshake.is_awaiting()
8742                                    || !owned_inputs.is_empty()
8743                                    || !inflight_claims.is_empty()
8744                                    || parent_question_checkpoint_id.is_some()
8745                                    || nested_wait_registered_this_run
8746                                    || strict_history_output.is_none()
8747                                {
8748                                    return Err(local_tool_history_unsupported());
8749                                }
8750                                commit_local_tool_history(
8751                                    local_history.take().ok_or_else(local_tool_history_unsupported)?,
8752                                    logical_session,
8753                                    session_inbox_runtime.ok_or_else(local_tool_history_unsupported)?,
8754                                    activation_run_id.ok_or_else(local_tool_history_unsupported)?,
8755                                    (tools, local_history_read_only),
8756                                    LocalToolTerminal::Cancelled,
8757                                    cancel_token,
8758                                ).await?;
8759                                let proof = bamboo_storage::v2::HostTerminalCompleteness::from_verified_host(
8760                                    logical_session,
8761                                    activation_run_id.ok_or_else(local_tool_history_unsupported)?,
8762                                    current_epoch,
8763                                    next_actor_event_seq - 1,
8764                                ).map_err(|_| local_tool_history_unsupported())?;
8765                                *strict_history_output.as_deref_mut()
8766                                    .ok_or_else(local_tool_history_unsupported)? = Some(StrictTerminalHistory {
8767                                        proof,
8768                                        watermark: final_event_watermark.ok_or_else(local_tool_history_unsupported)?,
8769                                    });
8770                            }
8771                            // A selected Worker terminal proves that its Run
8772                            // ended even if permission posture or an Inbox
8773                            // correction was still pending. Claims remain
8774                            // unacked for retry; only the placement fence can
8775                            // now be released after SDK saves Cancelled.
8776                            client.accept_durable_terminal(status);
8777                            if workflow_usage_requested {
8778                                logical_session.metadata.insert(bamboo_subagent::proto::WORKFLOW_TERMINAL_OBSERVATION_KEY.into(), activation_run_id.ok_or_else(plain_actor_unsupported)?.into());
8779                            }
8780                            return Err(AgentError::Cancelled);
8781                        }
8782                        if parent_question_checkpoint_id.is_some()
8783                            && status != TerminalStatus::Suspended
8784                        {
8785                            return Err(AgentError::LLM(
8786                                "Child question checkpoint must end in a suspended Run".into(),
8787                            ));
8788                        }
8789                        if permission_handshake.is_awaiting() {
8790                            if remote_environment_lease && status == TerminalStatus::Error {
8791                                if let Some(code) = error.as_deref().and_then(remote_environment_error_code) {
8792                                    // The Worker rejected the EnvironmentLease before it
8793                                    // could confirm its permission posture. This exact
8794                                    // correlated Outcome still proves the Run ended, so
8795                                    // retain its durable receipt and release the slot after
8796                                    // the Host records the failure. Other early terminals
8797                                    // remain unproven and keep their placement fence.
8798                                    client.accept_durable_terminal(status);
8799                            if workflow_usage_requested {
8800                                logical_session.metadata.insert(bamboo_subagent::proto::WORKFLOW_TERMINAL_OBSERVATION_KEY.into(), activation_run_id.ok_or_else(plain_actor_unsupported)?.into());
8801                            }
8802                                    return Err(AgentError::LLM(code.to_owned()));
8803                                }
8804                            }
8805                            return Err(AgentError::LLM(
8806                                "actor terminated before permission posture confirmation"
8807                                    .to_string(),
8808                            ));
8809                        }
8810                        if !owned_inputs.is_empty() {
8811                            return Err(AgentError::LLM("Actor terminated before confirming its durable correction; Inbox claim is retained".into()));
8812                        }
8813                        if let Some(claim) = inflight_claims.front() {
8814                            return Err(AgentError::LLM(format!(
8815                                "actor terminated before durably admitting SessionInbox message {}; canonical claim remains recoverable",
8816                                claim.envelope.id
8817                            )));
8818                        }
8819                        if status == TerminalStatus::Completed {
8820                            if let Some(collector) = local_history.take() {
8821                                commit_local_tool_history(
8822                                    collector, logical_session,
8823                                    session_inbox_runtime.ok_or_else(local_tool_history_unsupported)?,
8824                                    activation_run_id.ok_or_else(local_tool_history_unsupported)?,
8825                                    (local_history_tools.ok_or_else(local_tool_history_unsupported)?, local_history_read_only),
8826                                    LocalToolTerminal::Completed(result.as_deref()), cancel_token,
8827                                ).await?;
8828                            }
8829                        }
8830                        if let Some(collector) = readonly.take() {
8831                            if status == TerminalStatus::Completed {
8832                                let activation = plain_input.ok_or_else(|| plain_actor_unsupported())?;
8833                                let backlog = activation.input_inbox.inspect(&logical_session.id).await
8834                                    .map_err(|_| plain_actor_unsupported())?;
8835                                if backlog.pending != 0 || backlog.claimed != 0 || backlog.activation_pending() {
8836                                    return Err(AgentError::LLM("Glob Actor continuation is unsupported; pending Inbox input is preserved".into()));
8837                                }
8838                                *readonly_output.as_deref_mut().ok_or_else(|| plain_actor_unsupported())? =
8839                                    collector.finish(result.as_deref())?;
8840                            }
8841                        } else if let Some(activation) = plain_input {
8842                            if status == TerminalStatus::Completed {
8843                                let binding = session_inbox_runtime.ok_or_else(|| plain_actor_unsupported())?;
8844                                let run_id = activation_run_id.ok_or_else(|| plain_actor_unsupported())?;
8845                                if continuations < continuation_limit {
8846                                    if let Some((claim, epoch)) = activation.continue_input(
8847                                        client, binding, logical_session, run_id, plain_run, &mut expected_permission_posture,
8848                                        result.clone().filter(|text| !text.is_empty()).ok_or_else(|| plain_actor_unsupported())?, event_tx,
8849                                    ).await? {
8850                                        owned_inputs = VecDeque::from([claim]);
8851                                        released_inputs.clear();
8852                                        continuations += 1;
8853                                        current_epoch = epoch;
8854                                        approval_epoch.store(epoch, Ordering::Release);
8855                                        // These are native coordinates scoped to the newly
8856                                        // correlated epoch. Public Host feed seq is untouched.
8857                                        next_actor_event_seq = 1;
8858                                        canonical_router = None;
8859                                        permission_handshake = PermissionPostureHandshake::new(expected_permission_posture.as_ref());
8860                                        display = ActorEventDisplay::default();
8861                                        continue;
8862                                    }
8863                                } else if activation.input_inbox.inspect(&logical_session.id).await
8864                                    .map_err(|_| plain_actor_unsupported())?.activation_pending() {
8865                                    if continuation_limit == 1 {
8866                                        return Err(AgentError::LLM("Only one Actor correction continuation is supported; remaining Inbox input is preserved".into()));
8867                                    }
8868                                    // The outer success path appends the last reply. An
8869                                    // explicit cap refusal must preserve it here instead,
8870                                    // without claiming/acking the excess input or saving
8871                                    // through the ordinary writer on this owned Actor.
8872                                    activation.append_reply(logical_session,
8873                                        result.clone().filter(|text| !text.is_empty()).ok_or_else(|| plain_actor_unsupported())?,
8874                                        event_tx).await?;
8875                                    return Err(AgentError::LLM("Actor correction continuation limit reached; completed reply committed and remaining Inbox input preserved".into()));
8876                                }
8877                            }
8878                        }
8879                        if status == TerminalStatus::Completed {
8880                            // Only an activation with a canonical Child creation
8881                            // identity can invoke the Host SubAgent route. Legacy
8882                            // Echo and other executors may share this runner's
8883                            // installed tool but have no nested wait to fence.
8884                            if let (Some(tool), Some(creation)) =
8885                                (canonical_subagent_tool.as_deref(), expected_creation)
8886                            {
8887                                let (Some(store), Some(binding), Some(run_id)) =
8888                                    (actor_directory_store, session_inbox_runtime,
8889                                        activation_run_id)
8890                                else {
8891                                    return Err(AgentError::LLM(
8892                                        "nested SubAgent Host authority unavailable at Child terminal"
8893                                            .into(),
8894                                    ));
8895                                };
8896                                let must_suspend = ensure_active_nested_wait(
8897                                    tool, store, binding, run_id, child_session_id,
8898                                    parent_session_id, logical_session, creation, plain_input,
8899                                ).await?;
8900                                if must_suspend || nested_wait_registered_this_run {
8901                                    // This is only a current-Run handoff fence.
8902                                    // sdk/spawn re-reads the durable Host wait;
8903                                    // copying a stale wait into this Session
8904                                    // would re-arm it after a fast grandchild
8905                                    // completion, while copying no wait would
8906                                    // erase an active registration on final save.
8907                                    logical_session.metadata.insert(
8908                                        "runtime.actor_nested_wait_handoff".into(),
8909                                        "true".into(),
8910                                    );
8911                                }
8912                            }
8913                        }
8914                        let terminal_result = match status {
8915                            TerminalStatus::Completed => Ok(result),
8916                            TerminalStatus::Cancelled => Err(AgentError::Cancelled),
8917                            TerminalStatus::Error => Err(AgentError::LLM(
8918                                error.unwrap_or_else(|| "actor child errored".to_string()),
8919                            )),
8920                            TerminalStatus::Suspended => {
8921                                use bamboo_agent_core::storage::Storage;
8922                                let (Some(request_id), Some(store), Some(binding), Some(run_id), Some(creation)) =
8923                                    (parent_question_checkpoint_id.as_ref(), actor_directory_store,
8924                                        session_inbox_runtime, activation_run_id, expected_creation)
8925                                else {
8926                                    return Err(AgentError::LLM("uncheckpointed Child suspension rejected".into()));
8927                                };
8928                                if load_active_subagent_caller(
8929                                    store, binding, run_id, child_session_id,
8930                                    parent_session_id, logical_session, creation, plain_input,
8931                                ).await.is_none() {
8932                                    return Err(AgentError::LLM("stale Child question suspension rejected".into()));
8933                                }
8934                                let id = bamboo_domain::SessionMessageId::parse(request_id.clone())
8935                                    .map_err(|_| AgentError::LLM("invalid Child question receipt".into()))?;
8936                                let child = store.load_session(child_session_id).await
8937                                    .map_err(|_| AgentError::LLM("canonical Child question unavailable".into()))?
8938                                    .ok_or_else(|| AgentError::LLM("canonical Child question missing".into()))?;
8939                                let parent = store.load_session(parent_session_id).await
8940                                    .map_err(|_| AgentError::LLM("canonical direct parent unavailable".into()))?
8941                                    .ok_or_else(|| AgentError::LLM("canonical direct parent missing".into()))?;
8942                                let pending = ParentQuestion::for_pending(&parent, &child)
8943                                    .is_some_and(|question| question.id == id);
8944                                let answered = bamboo_domain::ParentQuestionResolution::from_child(
8945                                    &parent, &child, &id,
8946                                ).is_some();
8947                                if !pending && !answered {
8948                                    return Err(AgentError::LLM("canonical Child question proof changed".into()));
8949                                }
8950                                logical_session.metadata.insert(
8951                                    "runtime.actor_parent_question_handoff".into(), "true".into(),
8952                                );
8953                                Ok(None)
8954                            }
8955                        };
8956                        if local_history_tools.is_some() {
8957                            // Failed strict Runs discard their collected suffix. A
8958                            // consumed watermark cannot certify those unapplied
8959                            // effects; retain every broker frame for recovery.
8960                            if status != TerminalStatus::Completed {
8961                                return terminal_result;
8962                            }
8963                            let proof = bamboo_storage::v2::HostTerminalCompleteness::from_verified_host(
8964                                logical_session,
8965                                activation_run_id.ok_or_else(local_tool_history_unsupported)?,
8966                                current_epoch,
8967                                next_actor_event_seq - 1,
8968                            ).map_err(|_| local_tool_history_unsupported())?;
8969                            let history = StrictTerminalHistory {
8970                                proof,
8971                                watermark: final_event_watermark.ok_or_else(local_tool_history_unsupported)?,
8972                            };
8973                            if let Some(output) = strict_history_output.as_deref_mut() {
8974                                *output = Some(history);
8975                            }
8976                        }
8977                        // This marker is set only after every Host terminal
8978                        // validation above succeeded. A surfaced broker Outcome
8979                        // that failed permission, Inbox, or question checks must
8980                        // never receive a durable ACK merely because SDK saved
8981                        // an error status. Genuine Worker Error/Cancelled frames
8982                        // are accepted terminals even though drive returns Err.
8983                        client.accept_durable_terminal(status);
8984                            if workflow_usage_requested {
8985                                logical_session.metadata.insert(bamboo_subagent::proto::WORKFLOW_TERMINAL_OBSERVATION_KEY.into(), activation_run_id.ok_or_else(plain_actor_unsupported)?.into());
8986                            }
8987                        if status == TerminalStatus::Completed {
8988                            if let Some(activation) = plain_input {
8989                                // Only the final epoch reaches here: continued Runs
8990                                // above reset their router and stay in the pump.
8991                                *activation.completion_router.lock().recover_poison() =
8992                                    canonical_router.take();
8993                            }
8994                        }
8995                        return terminal_result;
8996                    }
8997                    Ok(None) => {
8998                        return Err(AgentError::LLM(
8999                            "actor child closed before terminal".to_string(),
9000                        ));
9001                    }
9002                    Err(e) => {
9003                        return Err(AgentError::LLM(format!("actor transport error: {e}")));
9004                    }
9005                }
9006            }
9007        }
9008    }
9009
9010    // Only reached on cancellation: ask the child to stop (best-effort), then report cancelled.
9011    let _ = client.send(ParentFrame::Cancel).await;
9012    Err(AgentError::Cancelled)
9013}
9014
9015/// The assignment text = the child session's latest user message (falls back to its title).
9016fn project_id_for_actor_run(
9017    session: &Session,
9018) -> Result<Option<bamboo_domain::ProjectId>, AgentError> {
9019    match crate::project_context::ProjectContextResolver::session_project_identity(session) {
9020        crate::project_context::SessionProjectIdentity::Assigned(project_id) => {
9021            Ok(Some(project_id))
9022        }
9023        crate::project_context::SessionProjectIdentity::Unassigned => Ok(None),
9024        crate::project_context::SessionProjectIdentity::Invalid { raw, message } => {
9025            Err(AgentError::LLM(format!(
9026                "child session carries an invalid Project identity '{raw}': {message}"
9027            )))
9028        }
9029    }
9030}
9031
9032fn remote_broker_unavailable() -> AgentError {
9033    AgentError::LLM("remote_broker_activation_unavailable".into())
9034}
9035
9036/// Only these bounded worker admission errors may cross the fixed Remote
9037/// diagnostic boundary before permission posture confirmation. Transport and
9038/// arbitrary worker text retain the generic remote error.
9039fn remote_environment_error_code(value: &str) -> Option<&'static str> {
9040    const CODES: &[&str] = &[
9041        "remote_environment_lease_invalid",
9042        "remote_environment_snapshot_mismatch",
9043        "remote_environment_checkout_not_clean",
9044        "remote_environment_checkout_unsupported",
9045        "remote_environment_workspace_unavailable",
9046        "remote_environment_workspace_outside_git_root",
9047        "remote_environment_path_unsupported",
9048        "remote_environment_git_unavailable",
9049        "remote_environment_git_inspection_failed",
9050        "remote_environment_git_inspection_timed_out",
9051        "remote_environment_git_output_too_large",
9052        "remote_environment_snapshot_too_large",
9053        "remote_environment_snapshot_changed",
9054        "remote_environment_index_invalid",
9055        "remote_environment_index_unsupported",
9056        "remote_environment_host_path_or_grants_rejected",
9057        "remote_environment_worker_workspace_missing",
9058        "remote_environment_runspec_invalid",
9059        "remote_environment_lease_missing_or_host_path_present",
9060    ];
9061    CODES.iter().copied().find(|code| *code == value)
9062}
9063
9064// Read only: no ensure/repair/Actor initialization or subtree budget. Capture
9065// exactly the canonical chain and compare it again immediately before Run.
9066async fn remote_canonical_lineage(
9067    binding: &SessionInboxRuntimeBinding,
9068    session: &Session,
9069) -> Result<Vec<bamboo_domain::ActorSession>, AgentError> {
9070    let mut current = bamboo_domain::ActorSession::from_session(session)
9071        .map_err(|_| remote_broker_unavailable())?;
9072    let mut chain = Vec::new();
9073    loop {
9074        if chain.len() > session.spawn_depth as usize {
9075            return Err(remote_broker_unavailable());
9076        }
9077        let saved = binding
9078            .storage
9079            .load_session(&current.actor_id)
9080            .await
9081            .map_err(|_| remote_broker_unavailable())?
9082            .ok_or_else(remote_broker_unavailable)?;
9083        let observed = bamboo_domain::ActorSession::from_session(&saved)
9084            .map_err(|_| remote_broker_unavailable())?;
9085        if observed.actor_id != current.actor_id
9086            || observed.session_created_at != current.session_created_at
9087            || observed.parent_actor_id != current.parent_actor_id
9088            || observed.root_actor_id != current.root_actor_id
9089            || observed.spawn_depth != current.spawn_depth
9090            || observed.project_id != current.project_id
9091        {
9092            return Err(remote_broker_unavailable());
9093        }
9094        chain.push(observed.clone());
9095        let Some(parent_id) = observed.parent_actor_id.as_ref() else {
9096            if observed.actor_id != session.root_session_id || observed.spawn_depth != 0 {
9097                return Err(remote_broker_unavailable());
9098            }
9099            return Ok(chain);
9100        };
9101        let parent = binding
9102            .storage
9103            .load_session(parent_id)
9104            .await
9105            .map_err(|_| remote_broker_unavailable())?
9106            .ok_or_else(remote_broker_unavailable)?;
9107        let parent = bamboo_domain::ActorSession::from_session(&parent)
9108            .map_err(|_| remote_broker_unavailable())?;
9109        if parent.root_actor_id != observed.root_actor_id
9110            || parent.project_id != observed.project_id
9111            || parent.spawn_depth.checked_add(1) != Some(observed.spawn_depth)
9112            || parent.session_created_at > observed.session_created_at
9113        {
9114            return Err(remote_broker_unavailable());
9115        }
9116        current = parent;
9117    }
9118}
9119
9120async fn canonical_child_creation(
9121    binding: &SessionInboxRuntimeBinding,
9122    session: &Session,
9123    job: &SpawnJob,
9124) -> Result<bamboo_subagent::proto::ChildCreationIdentity, AgentError> {
9125    let conflict = || {
9126        AgentError::LLM(
9127            "child creation identity unavailable or changed in canonical host Storage".into(),
9128        )
9129    };
9130    let current = binding
9131        .storage
9132        .load_session(&job.child_session_id)
9133        .await
9134        .map_err(|_| conflict())?
9135        .ok_or_else(conflict)?;
9136    if bamboo_domain::ActorSession::from_session(&current).is_err()
9137        || bamboo_domain::ActorSession::from_session(session).is_err()
9138        || current.kind != bamboo_domain::SessionKind::Child
9139        || session.kind != current.kind
9140        || current.id != session.id
9141        || current.id != job.child_session_id
9142        || current.parent_session_id.as_deref() != Some(job.parent_session_id.as_str())
9143        || current.parent_session_id != session.parent_session_id
9144        || current.root_session_id.trim().is_empty()
9145        || current.root_session_id != session.root_session_id
9146        || current.created_at != session.created_at
9147        || current.spawn_depth != session.spawn_depth
9148        || current.spawn_depth == 0
9149        || project_id_for_actor_run(&current)? != project_id_for_actor_run(session)?
9150    {
9151        return Err(conflict());
9152    }
9153    Ok(bamboo_subagent::proto::ChildCreationIdentity {
9154        created_at: current.created_at,
9155        spawn_depth: current.spawn_depth,
9156    })
9157}
9158
9159fn logical_identity_for_actor_run(session: &Session, job: &SpawnJob) -> LogicalSessionIdentity {
9160    LogicalSessionIdentity {
9161        creation: None,
9162        session_id: session.id.clone(),
9163        parent_session_id: session
9164            .parent_session_id
9165            .clone()
9166            .or_else(|| Some(job.parent_session_id.clone())),
9167        root_session_id: if session.root_session_id.trim().is_empty() {
9168            job.parent_session_id.clone()
9169        } else {
9170            session.root_session_id.clone()
9171        },
9172    }
9173}
9174
9175fn extract_assignment(session: &Session) -> String {
9176    session
9177        .messages
9178        .iter()
9179        .rev()
9180        .find(|m| matches!(m.role, Role::User))
9181        .map(|m| m.content.clone())
9182        .unwrap_or_else(|| {
9183            session
9184                .metadata
9185                .get("title")
9186                .cloned()
9187                .unwrap_or_else(|| "Execute task".to_string())
9188        })
9189}
9190
9191#[cfg(test)]
9192mod tests {
9193    use super::*;
9194    use crate::SessionActivationRouter;
9195    use bamboo_domain::{RuntimeSessionPersistence, SessionInboxPort, Storage};
9196
9197    include!("expired_plain_owner_tests.rs");
9198
9199    include!("owned_cold_input_tests.rs");
9200
9201    #[test]
9202    fn plain_rejection_logs_branch_and_run_without_changing_error() {
9203        // Keep shared callsites consulting the current subscriber during scoped capture.
9204        let capture_interest_guard = tracing::Dispatch::new(
9205            tracing_subscriber::fmt()
9206                .with_max_level(tracing_subscriber::filter::LevelFilter::OFF)
9207                .with_writer(std::io::sink)
9208                .finish(),
9209        );
9210        let output = tempfile::NamedTempFile::new().unwrap();
9211        let subscriber = tracing_subscriber::fmt()
9212            .with_ansi(false)
9213            .without_time()
9214            .with_writer(output.reopen().unwrap())
9215            .finish();
9216        let branch_line = std::cell::Cell::new(0);
9217        let error = tracing::subscriber::with_default(subscriber, || {
9218            // Exercise the shared warning callsite without this thread's scoped subscriber.
9219            std::thread::spawn(|| drop(plain_actor_unsupported()))
9220                .join()
9221                .expect("unscoped rejection warning thread panicked");
9222            let span = tracing::warn_span!(
9223                "actor_child_run",
9224                child_session_id = %"diagnostic-child",
9225                activation_run_id = tracing::field::Empty
9226            );
9227            let _entered = span.enter();
9228            span.record("activation_run_id", "diagnostic-run");
9229            None::<()>
9230                .ok_or_else(|| {
9231                    let caller_line = line!() + 1;
9232                    let error = plain_actor_unsupported();
9233                    branch_line.set(caller_line);
9234                    error
9235                })
9236                .unwrap_err()
9237        });
9238        assert!(matches!(&error, AgentError::LLM(message)
9239            if message.starts_with("This Actor Child supports a fresh plain activation,")
9240                && message.ends_with("Other continuation is unsupported; durable history is preserved.")));
9241        let log = std::fs::read_to_string(output.path()).unwrap();
9242        assert!(log.contains("child_session_id=diagnostic-child"), "{log}");
9243        assert!(
9244            log.contains("activation_run_id=\"diagnostic-run\""),
9245            "{log}"
9246        );
9247        assert!(
9248            log.contains(&format!("process_id={}", std::process::id())),
9249            "{log}"
9250        );
9251        assert!(log.contains(&format!("source_file={:?}", file!())), "{log}");
9252        assert!(
9253            log.contains(&format!("source_line={}", branch_line.get())),
9254            "{log}"
9255        );
9256        assert!(!error.to_string().contains("diagnostic-run"));
9257        drop(capture_interest_guard);
9258    }
9259
9260    #[tokio::test]
9261    async fn completed_plain_continuation_requires_bounded_inputs_and_current_identity() {
9262        for case in [
9263            "valid",
9264            "over-batch-limit",
9265            "tool-history",
9266            "stale-lineage",
9267            "no-release",
9268        ] {
9269            let temp = tempfile::tempdir().unwrap();
9270            let store = Arc::new(
9271                bamboo_storage::SessionStoreV2::new(temp.path().into())
9272                    .await
9273                    .unwrap(),
9274            );
9275            let parent = Session::new("continuation-parent", "model");
9276            store.save_session(&parent).await.unwrap();
9277            let mut child =
9278                Session::new_child_of("continuation-child", &parent, "model", "first task");
9279            child.add_message(bamboo_agent_core::Message::assistant("first result", None));
9280            if case == "tool-history" {
9281                let mut tool = bamboo_agent_core::Message::assistant("old tool result", None);
9282                tool.role = Role::Tool;
9283                child.add_message(tool);
9284            }
9285            child.set_last_run_status("completed");
9286            store.save_session(&child).await.unwrap();
9287            store.ensure_actor(&child.id).await.unwrap();
9288            let policy = Arc::new(bamboo_tools::permission::PermissionConfig::new());
9289            bind_local_control_plane(store.as_ref(), &child.id, policy.as_ref()).await;
9290            let now = chrono::Utc::now();
9291            let old = store
9292                .claim_activation(&ActorActivationClaim {
9293                    actor_id: child.id.clone(),
9294                    run_id: "first-run".into(),
9295                    lease_owner: "first-owner".into(),
9296                    lease_expires_at: now + chrono::Duration::minutes(1),
9297                    inbox_generation: 0,
9298                    placement_ref: None,
9299                    now,
9300                })
9301                .await
9302                .unwrap();
9303            store.start_activation(&old.fence(), now).await.unwrap();
9304            store
9305                .finish_activation(&old.fence(), now, ActorActivationFinish::Succeeded)
9306                .await
9307                .unwrap();
9308            let inbox: Arc<dyn SessionInboxPort> = Arc::new(bamboo_storage::FileSessionInbox::new(
9309                store.clone(),
9310                Default::default(),
9311            ));
9312            for slot in 0..if case == "over-batch-limit" {
9313                bamboo_domain::SessionInboxLimits::default().max_claim_batch + 1
9314            } else {
9315                1
9316            } {
9317                let envelope = bamboo_domain::SessionMessageEnvelope::user_input(
9318                    &child.id,
9319                    format!("new task {slot}"),
9320                );
9321                let receipt = inbox.deliver(&envelope).await.unwrap();
9322                inbox
9323                    .mark_activation_eligible(
9324                        &child.id,
9325                        receipt.generation,
9326                        bamboo_domain::SessionActivationPolicy::InterruptSpecificWait,
9327                    )
9328                    .await
9329                    .unwrap();
9330            }
9331            let storage: Arc<dyn Storage> = store.clone();
9332            let persistence = Arc::new(bamboo_storage::LockedSessionStore::new(storage));
9333            let binding = actor_binding(store.clone(), inbox, persistence);
9334            let _registration = binding
9335                .router
9336                .register_run(&child.id, "second-run")
9337                .await
9338                .unwrap();
9339            if case == "stale-lineage" {
9340                child.parent_session_id = Some("foreign-parent".into());
9341            }
9342            let result = PlainActorActivation::start(
9343                store.clone(),
9344                &child,
9345                &binding,
9346                Some("second-run"),
9347                Some(policy),
9348                true,
9349                (case != "no-release").then_some("second-worker"),
9350            )
9351            .await;
9352            if case == "valid" {
9353                let activation = result.expect("one new input on a completed plain Child");
9354                assert_eq!(activation.fence.attempt, old.attempt + 1);
9355                assert_eq!(activation.fence.actor_id, child.id);
9356                assert_eq!(activation.inbox_generation, 1);
9357                activation
9358                    .finish(ActorActivationFinish::Failed)
9359                    .await
9360                    .unwrap();
9361            } else {
9362                assert!(
9363                    result.is_err(),
9364                    "unsupported continuation {case} must fail before claim"
9365                );
9366                assert_eq!(
9367                    store
9368                        .inspect_actor(&child.id)
9369                        .await
9370                        .unwrap()
9371                        .actor
9372                        .current_attempt,
9373                    old.attempt
9374                );
9375            }
9376            let canonical = store.load_session(&child.id).await.unwrap().unwrap();
9377            assert_eq!(
9378                canonical.parent_session_id.as_deref(),
9379                Some(parent.id.as_str())
9380            );
9381            assert!(canonical
9382                .messages
9383                .iter()
9384                .any(|message| message.content == "first result"));
9385        }
9386    }
9387
9388    #[tokio::test]
9389    async fn pre_dispatch_placement_failure_fails_only_claimed_activation() {
9390        let temp = tempfile::tempdir().unwrap();
9391        let store = bamboo_storage::SessionStoreV2::new(temp.path().into())
9392            .await
9393            .unwrap();
9394        let parent = Session::new("placement-parent", "model");
9395        let child = Session::new_child_of("placement-child", &parent, "model", "task");
9396        store.save_session(&parent).await.unwrap();
9397        store.save_session(&child).await.unwrap();
9398        store.ensure_actor(&child.id).await.unwrap();
9399        store
9400            .bind_control_plane(
9401                &child.id,
9402                bamboo_domain::ActorControlPlaneBinding {
9403                    policy_revision: None,
9404                    placement_intent: Some(bamboo_domain::ActorPlacementIntent::Pool {
9405                        pool_id: "remote-worker".into(),
9406                    }),
9407                },
9408            )
9409            .await
9410            .unwrap();
9411
9412        let claim = |run_id: &str| ActorActivationClaim {
9413            actor_id: child.id.clone(),
9414            run_id: run_id.into(),
9415            lease_owner: format!("host-{run_id}"),
9416            lease_expires_at: chrono::Utc::now() + chrono::Duration::seconds(80),
9417            inbox_generation: 0,
9418            placement_ref: Some(bamboo_domain::ActorPlacementRef {
9419                class: bamboo_domain::ActorPlacementClass::Remote,
9420                lease_id: format!("slot-{run_id}"),
9421                slot_epoch: Some(1),
9422            }),
9423            now: chrono::Utc::now(),
9424        };
9425        let first = store.claim_activation(&claim("first")).await.unwrap();
9426        store
9427            .start_activation(&first.fence(), chrono::Utc::now())
9428            .await
9429            .unwrap();
9430        assert!(store.claim_activation(&claim("retry")).await.is_err());
9431
9432        // This is the post-claim placement preflight failure path. No Run has
9433        // been dispatched, so the same ActorId can retry before lease expiry.
9434        fail_pre_dispatch_activation(&store, &first.fence())
9435            .await
9436            .unwrap();
9437        let failed = store.inspect_actor(&child.id).await.unwrap();
9438        assert_eq!(
9439            failed.activation.unwrap().status,
9440            bamboo_domain::ActorActivationStatus::Failed
9441        );
9442        let retry = store.claim_activation(&claim("retry")).await.unwrap();
9443        assert_eq!(retry.attempt, first.attempt + 1);
9444        assert!(fail_pre_dispatch_activation(&store, &first.fence())
9445            .await
9446            .is_err());
9447        assert_eq!(
9448            store.inspect_actor(&child.id).await.unwrap().activation,
9449            Some(retry)
9450        );
9451    }
9452
9453    #[tokio::test]
9454    async fn lost_run_delivery_receipt_keeps_actor_activation_live() {
9455        struct AcceptedWithoutReceipt {
9456            published: bool,
9457        }
9458
9459        #[async_trait]
9460        impl bamboo_subagent::ChildLink for AcceptedWithoutReceipt {
9461            async fn send(&mut self, frame: ParentFrame) -> bamboo_subagent::TransportResult<()> {
9462                assert!(matches!(frame, ParentFrame::Run(_)));
9463                self.published = true;
9464                Err(bamboo_subagent::TransportError::Protocol(
9465                    "broker delivery receipt lost after publication".into(),
9466                ))
9467            }
9468
9469            async fn next_frame(&mut self) -> bamboo_subagent::TransportResult<Option<ChildFrame>> {
9470                std::future::pending().await
9471            }
9472        }
9473
9474        let temp = tempfile::tempdir().unwrap();
9475        let store = Arc::new(
9476            bamboo_storage::SessionStoreV2::new(temp.path().into())
9477                .await
9478                .unwrap(),
9479        );
9480        let parent = Session::new("dispatch-parent", "model");
9481        let child = Session::new_child_of("dispatch-child", &parent, "model", "task");
9482        store.save_session(&parent).await.unwrap();
9483        store.save_session(&child).await.unwrap();
9484        store.ensure_actor(&child.id).await.unwrap();
9485        let now = chrono::Utc::now();
9486        let claim = ActorActivationClaim {
9487            actor_id: child.id.clone(),
9488            run_id: "dispatch-run".into(),
9489            lease_owner: "dispatch-host".into(),
9490            lease_expires_at: now + chrono::Duration::seconds(80),
9491            inbox_generation: 0,
9492            placement_ref: None,
9493            now,
9494        };
9495        let activation = store.claim_activation(&claim).await.unwrap();
9496        store
9497            .start_activation(&activation.fence(), chrono::Utc::now())
9498            .await
9499            .unwrap();
9500        let mut link = AcceptedWithoutReceipt { published: false };
9501        let mut send_invoked = false;
9502        let run = RunSpec {
9503            assignment: "task".into(),
9504            logical_session: None,
9505            project_id: None,
9506            reasoning_effort: None,
9507            permission_policy: None,
9508            messages: Vec::new(),
9509            activation_run_id: Some(claim.run_id.clone()),
9510            execution_epoch: 1,
9511            initial_session_messages: Vec::new(),
9512            secrets: Default::default(),
9513        };
9514        assert!(send_actor_run(&mut link, run, &mut send_invoked)
9515            .await
9516            .is_err());
9517        assert!(send_invoked && link.published);
9518        let slot = WorkerSlotLease {
9519            lease_id: "dispatch-slot".into(),
9520            host_ref: "dispatch-worker".into(),
9521            connection_generation: "dispatch-generation".into(),
9522            slot: 0,
9523            epoch: 1,
9524            actor_id: child.id.clone(),
9525            run_id: claim.run_id.clone(),
9526            expires_at: claim.lease_expires_at,
9527        };
9528        let mut lease = ActivePlacementLease {
9529            registry: bamboo_storage::v2::FileHostRegistry::new(temp.path().into())
9530                .await
9531                .unwrap(),
9532            actor_store: store.clone(),
9533            route: ScopedPlacementRoute {
9534                endpoint: String::new(),
9535                token: String::new(),
9536                ca_cert_file: PathBuf::new(),
9537                parent: bamboo_subagent::AgentRef {
9538                    session_id: "dispatch-parent".into(),
9539                    role: Some("host".into()),
9540                },
9541                worker_role: "worker".into(),
9542                pinned_mailbox: None,
9543                requirements: Default::default(),
9544            },
9545            mailbox: "dispatch-worker".into(),
9546            host_ref: "dispatch-worker".into(),
9547            fence: activation.fence(),
9548            placement_ref: bamboo_domain::ActorPlacementRef {
9549                class: bamboo_domain::ActorPlacementClass::Remote,
9550                lease_id: slot.lease_id.clone(),
9551                slot_epoch: Some(slot.epoch),
9552            },
9553            lease: Arc::new(tokio::sync::Mutex::new(slot)),
9554            heartbeat: tokio::spawn(std::future::pending()),
9555            run_cancel: CancellationToken::new(),
9556            lost: Arc::new(AtomicBool::new(false)),
9557            release_on_drop: AtomicBool::new(false),
9558        };
9559        assert!(!lease.abort_if_run_unsent(send_invoked).await.unwrap());
9560        assert_eq!(
9561            store
9562                .inspect_actor(&child.id)
9563                .await
9564                .unwrap()
9565                .activation
9566                .unwrap()
9567                .status,
9568            bamboo_domain::ActorActivationStatus::Running
9569        );
9570        let mut retry = claim.clone();
9571        retry.run_id = "dispatch-retry".into();
9572        retry.now = chrono::Utc::now();
9573        retry.lease_expires_at = retry.now + chrono::Duration::seconds(80);
9574        assert!(matches!(
9575            store.claim_activation(&retry).await,
9576            Err(bamboo_domain::ActorDirectoryError::Busy)
9577        ));
9578    }
9579
9580    #[test]
9581    fn accepted_cancelled_terminal_maps_to_cancelled_actor_activation() {
9582        let cancelled: Result<Option<String>, AgentError> = Err(AgentError::Cancelled);
9583        assert_eq!(
9584            remote_activation_finish(&cancelled, Some(TerminalStatus::Cancelled)),
9585            ActorActivationFinish::Cancelled
9586        );
9587        assert_eq!(
9588            remote_activation_finish(&cancelled, None),
9589            ActorActivationFinish::Failed
9590        );
9591        assert_eq!(
9592            remote_activation_finish(&cancelled, Some(TerminalStatus::Completed)),
9593            ActorActivationFinish::Failed
9594        );
9595    }
9596
9597    #[tokio::test]
9598    async fn remote_reply_is_fenced_before_terminal_status_save() {
9599        let temp = tempfile::tempdir().unwrap();
9600        let store = bamboo_storage::SessionStoreV2::new(temp.path().into())
9601            .await
9602            .unwrap();
9603        let parent = Session::new("remote-reply-parent", "model");
9604        let mut child = Session::new_child_of("remote-reply-child", &parent, "model", "task");
9605        child.add_message(bamboo_agent_core::Message::user("task"));
9606        store.save_session(&parent).await.unwrap();
9607        store.save_session(&child).await.unwrap();
9608        store.ensure_actor(&child.id).await.unwrap();
9609        store
9610            .bind_control_plane(
9611                &child.id,
9612                bamboo_domain::ActorControlPlaneBinding {
9613                    policy_revision: None,
9614                    placement_intent: Some(bamboo_domain::ActorPlacementIntent::Pool {
9615                        pool_id: "remote-worker".into(),
9616                    }),
9617                },
9618            )
9619            .await
9620            .unwrap();
9621        let now = chrono::Utc::now();
9622        let activation = store
9623            .claim_activation(&ActorActivationClaim {
9624                actor_id: child.id.clone(),
9625                run_id: "remote-reply-run".into(),
9626                lease_owner: "remote-worker".into(),
9627                lease_expires_at: now + chrono::Duration::seconds(80),
9628                inbox_generation: 0,
9629                placement_ref: Some(bamboo_domain::ActorPlacementRef {
9630                    class: bamboo_domain::ActorPlacementClass::Remote,
9631                    lease_id: "remote-slot".into(),
9632                    slot_epoch: Some(1),
9633                }),
9634                now,
9635            })
9636            .await
9637            .unwrap();
9638        let fence = activation.fence();
9639        store.start_activation(&fence, now).await.unwrap();
9640
9641        append_fenced_remote_reply(&store, &fence, &mut child, "done")
9642            .await
9643            .unwrap();
9644        store
9645            .finish_activation(&fence, chrono::Utc::now(), ActorActivationFinish::Succeeded)
9646            .await
9647            .unwrap();
9648        child.set_last_run_status("completed");
9649        store.save_session(&child).await.unwrap();
9650        let saved = store.load_session(&child.id).await.unwrap().unwrap();
9651        assert_eq!(saved.last_run_status().as_deref(), Some("completed"));
9652        assert_eq!(saved.messages.len(), 2);
9653        assert_eq!(saved.messages.last().unwrap().content, "done");
9654        assert!(
9655            append_fenced_remote_reply(&store, &fence, &mut child, "duplicate")
9656                .await
9657                .is_err()
9658        );
9659        assert_eq!(
9660            store
9661                .load_session(&child.id)
9662                .await
9663                .unwrap()
9664                .unwrap()
9665                .messages
9666                .len(),
9667            2
9668        );
9669    }
9670
9671    async fn start_bus() -> (String, tempfile::TempDir) {
9672        let dir = tempfile::tempdir().unwrap();
9673        let core = Arc::new(bamboo_broker::BrokerCore::new(dir.path()));
9674        let server = Arc::new(bamboo_broker::BrokerServer::new(core, "t"));
9675        let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
9676        let addr = listener.local_addr().unwrap();
9677        tokio::spawn(async move {
9678            let _ = server.serve(listener).await;
9679        });
9680        (format!("ws://{addr}"), dir)
9681    }
9682
9683    struct DurableAckProbe {
9684        calls: Arc<AtomicUsize>,
9685        fail: bool,
9686        terminal_status: TerminalStatus,
9687        parent_mailbox: String,
9688    }
9689
9690    #[async_trait]
9691    impl bamboo_subagent::ChildLink for DurableAckProbe {
9692        async fn send(&mut self, _frame: ParentFrame) -> bamboo_subagent::TransportResult<()> {
9693            Ok(())
9694        }
9695
9696        async fn next_frame(&mut self) -> bamboo_subagent::TransportResult<Option<ChildFrame>> {
9697            Ok(None)
9698        }
9699
9700        fn has_pending_durable_terminal(&self) -> bool {
9701            true
9702        }
9703
9704        fn durable_delivery_receipt(&self) -> Option<bamboo_subagent::DurableChildDeliveryReceipt> {
9705            Some(bamboo_subagent::DurableChildDeliveryReceipt {
9706                broker_identity: "00000000-0000-4000-8000-000000000001".into(),
9707                parent_mailbox: self.parent_mailbox.clone(),
9708                correlation_id: "broker-run-1".into(),
9709                message_ids: vec!["broker-outcome-1".into()],
9710                terminal_status: self.terminal_status,
9711            })
9712        }
9713
9714        async fn acknowledge_durable_frames(&mut self) -> bamboo_subagent::TransportResult<()> {
9715            self.calls.fetch_add(1, Ordering::SeqCst);
9716            if self.fail {
9717                Err(bamboo_subagent::TransportError::Protocol(
9718                    "injected ACK delivery failure".into(),
9719                ))
9720            } else {
9721                Ok(())
9722            }
9723        }
9724    }
9725
9726    async fn insert_ack_probe(
9727        runner: &ActorChildRunner,
9728        session: &Session,
9729        calls: &Arc<AtomicUsize>,
9730        terminal_status: TerminalStatus,
9731    ) {
9732        insert_ack_probe_with_nested_wait(runner, session, calls, terminal_status, false).await;
9733    }
9734
9735    async fn insert_ack_probe_with_nested_wait(
9736        runner: &ActorChildRunner,
9737        session: &Session,
9738        calls: &Arc<AtomicUsize>,
9739        terminal_status: TerminalStatus,
9740        host_nested_wait_handoff: bool,
9741    ) {
9742        runner.pending_durable_links.lock().await.insert(
9743            (session.id.clone(), session.created_at, "run-1".into()),
9744            PendingDurableChildLink {
9745                link: Box::new(DurableAckProbe {
9746                    calls: calls.clone(),
9747                    fail: false,
9748                    terminal_status,
9749                    parent_mailbox: format!("p-{}", session.id),
9750                }),
9751                host_nested_wait_handoff,
9752                strict_history: None,
9753                strict_history_required: false,
9754                strict_history_committed: false,
9755            },
9756        );
9757    }
9758
9759    #[tokio::test]
9760    async fn broker_receipt_reconciler_repairs_only_the_proven_local_mailbox() {
9761        let (endpoint, broker_dir) = start_bus().await;
9762        let home = tempfile::tempdir().unwrap();
9763        let store = Arc::new(
9764            bamboo_storage::SessionStoreV2::new(home.path().to_path_buf())
9765                .await
9766                .unwrap(),
9767        );
9768        let root = Session::new("receipt-root", "model");
9769        store.save_session(&root).await.unwrap();
9770        let mut sender = bamboo_broker::BrokerClient::connect(
9771            &endpoint,
9772            bamboo_subagent::AgentRef {
9773                session_id: "receipt-worker".into(),
9774                role: None,
9775            },
9776            "t",
9777        )
9778        .await
9779        .unwrap();
9780        let broker_identity = sender.durable_broker_identity().await.unwrap();
9781        let wrong_broker_identity = if broker_identity == "00000000-0000-4000-8000-000000000001" {
9782            "00000000-0000-4000-8000-000000000002"
9783        } else {
9784            "00000000-0000-4000-8000-000000000001"
9785        };
9786
9787        for (id, receipt_mailbox, receipt_broker) in [
9788            ("receipt-good", "p-receipt-good", broker_identity.as_str()),
9789            (
9790                "receipt-wrong-mailbox",
9791                "p-another-child",
9792                broker_identity.as_str(),
9793            ),
9794            (
9795                "receipt-wrong-broker",
9796                "p-receipt-wrong-broker",
9797                wrong_broker_identity,
9798            ),
9799            (
9800                "receipt-local-role-changed",
9801                "p-receipt-local-role-changed",
9802                broker_identity.as_str(),
9803            ),
9804            (
9805                "receipt-remote-unproven",
9806                "remote-parent-unproven",
9807                broker_identity.as_str(),
9808            ),
9809        ] {
9810            let mut child = Session::new_child_of(id, &root, "model", "task");
9811            child
9812                .metadata
9813                .insert("runtime.kind".into(), "external".into());
9814            child
9815                .metadata
9816                .insert("external.protocol".into(), "actor".into());
9817            child.metadata.insert(
9818                "external.agent_id".into(),
9819                crate::external_agents::config::LOCAL_ACTOR_AGENT_ID.into(),
9820            );
9821            if id == "receipt-local-role-changed" || id == "receipt-remote-unproven" {
9822                child
9823                    .metadata
9824                    .insert("subagent_type".into(), "remote-worker".into());
9825            }
9826            child.add_message(bamboo_agent_core::Message::user("work"));
9827            store.save_session(&child).await.unwrap();
9828            let message_id = bamboo_subagent::MsgId::new();
9829            sender
9830                .deliver(
9831                    &format!("p-{id}"),
9832                    bamboo_subagent::InboxMessage {
9833                        id: message_id.clone(),
9834                        from: bamboo_subagent::AgentRef {
9835                            session_id: "receipt-worker".into(),
9836                            role: None,
9837                        },
9838                        kind: bamboo_subagent::InboxKind::Outcome,
9839                        body: serde_json::to_value(bamboo_subagent::ChildOutcome::completed(
9840                            "done",
9841                        ))
9842                        .unwrap(),
9843                        created_at: chrono::Utc::now(),
9844                        correlation_id: Some(bamboo_subagent::MsgId::new()),
9845                    },
9846                )
9847                .await
9848                .unwrap();
9849            // The Host can checkpoint only a frame already surfaced from the
9850            // broker, which moves its Maildir file from new/ into cur/.
9851            let mailbox =
9852                bamboo_subagent::Mailbox::at(broker_dir.path().join(format!("mailboxes/p-{id}")));
9853            assert_eq!(mailbox.drain().await.unwrap().len(), 1);
9854            child.add_message(bamboo_agent_core::Message::assistant("done", None));
9855            child.set_last_run_status("completed");
9856            if id == "receipt-good" {
9857                let prior_ids = child.messages[..child.messages.len() - 1]
9858                    .iter()
9859                    .map(|message| message.id.clone())
9860                    .collect();
9861                crate::execution::ChildCompletionSource::prepare(&mut child, "run-1", &prior_ids);
9862                assert!(
9863                    crate::execution::ChildCompletionSource::defer_for_durable_delivery(
9864                        &mut child, "run-1"
9865                    )
9866                    .unwrap()
9867                );
9868                let proof = bamboo_storage::v2::HostTerminalCompleteness::from_verified_host(
9869                    &child, "run-1", 7, 1,
9870                )
9871                .unwrap();
9872                store
9873                    .prepare_broker_terminal_receipt_with_completeness(
9874                        &child,
9875                        bamboo_storage::v2::BrokerTerminalRoute {
9876                            activation_run_id: "run-1",
9877                            broker_identity: receipt_broker,
9878                            parent_mailbox: receipt_mailbox,
9879                            broker_correlation_id: "correlation-1",
9880                            message_ids: &[message_id.as_str().to_owned()],
9881                        },
9882                        &proof,
9883                    )
9884                    .await
9885                    .unwrap();
9886                let mut staged = child.clone();
9887                staged.set_last_run_status("running");
9888                staged.clear_last_run_error();
9889                store.save_session(&staged).await.unwrap();
9890                store
9891                    .commit_broker_terminal_receipt_with_completeness(&child, "run-1", &proof)
9892                    .await
9893                    .unwrap();
9894            } else {
9895                store
9896                    .prepare_broker_terminal_receipt(
9897                        &child,
9898                        "run-1",
9899                        receipt_broker,
9900                        receipt_mailbox,
9901                        "correlation-1",
9902                        &[message_id.as_str().to_owned()],
9903                    )
9904                    .await
9905                    .unwrap();
9906                store.save_session(&child).await.unwrap();
9907            }
9908        }
9909
9910        let mut config = bamboo_config::Config::default();
9911        config.subagents_mut().broker = Some(bamboo_config::BrokerClientConfig {
9912            endpoint,
9913            token: "t".into(),
9914            ..Default::default()
9915        });
9916        config
9917            .subagents_mut()
9918            .remote_placements
9919            .push(bamboo_config::RemoteActorPlacement {
9920                role: "remote-worker".into(),
9921                endpoint: "ws://unscoped.example.invalid".into(),
9922                token_env: None,
9923                ca_cert_file: None,
9924                broker_peer: None,
9925                placement_requirements: None,
9926            });
9927        // Recover on an independently reopened Store, with no Actor executor or
9928        // provider available. Only the committed Host proof can authorize ACK.
9929        let cold_store = Arc::new(
9930            bamboo_storage::SessionStoreV2::new(home.path().to_path_buf())
9931                .await
9932                .unwrap(),
9933        );
9934        let before = cold_store
9935            .load_session("receipt-good")
9936            .await
9937            .unwrap()
9938            .unwrap();
9939        assert!(crate::execution::ChildCompletionSource::has_source_record(
9940            &before
9941        ));
9942        assert_eq!(before.last_run_status().as_deref(), Some("running"));
9943        assert!(crate::execution::ChildCompletionSource::from_committed_session(&before).is_none());
9944        let repair = BrokerTerminalReceiptReconciler::new(cold_store.clone(), &config);
9945        let report = repair.reconcile_once().await;
9946        assert_eq!(report.candidates, 5);
9947        assert_eq!(report.repaired, 2);
9948        assert_eq!(report.blocked, 3);
9949        let after = cold_store
9950            .load_session("receipt-good")
9951            .await
9952            .unwrap()
9953            .unwrap();
9954        assert!(crate::execution::ChildCompletionSource::from_committed_session(&after).is_some());
9955        assert_eq!(
9956            serde_json::to_value(before.messages).unwrap(),
9957            serde_json::to_value(after.messages).unwrap()
9958        );
9959        let retry = repair.reconcile_once().await;
9960        assert_eq!(retry.repaired, 0, "committed strict recovery is idempotent");
9961
9962        let good = bamboo_subagent::Mailbox::at(broker_dir.path().join("mailboxes/p-receipt-good"));
9963        let wrong = bamboo_subagent::Mailbox::at(
9964            broker_dir.path().join("mailboxes/p-receipt-wrong-mailbox"),
9965        );
9966        let wrong_broker = bamboo_subagent::Mailbox::at(
9967            broker_dir.path().join("mailboxes/p-receipt-wrong-broker"),
9968        );
9969        let local_role_changed = bamboo_subagent::Mailbox::at(
9970            broker_dir
9971                .path()
9972                .join("mailboxes/p-receipt-local-role-changed"),
9973        );
9974        let remote_unproven = bamboo_subagent::Mailbox::at(
9975            broker_dir
9976                .path()
9977                .join("mailboxes/p-receipt-remote-unproven"),
9978        );
9979        assert_eq!(good.pending_count().await.unwrap(), 0);
9980        assert_eq!(wrong.pending_count().await.unwrap(), 1);
9981        assert_eq!(wrong_broker.pending_count().await.unwrap(), 1);
9982        assert_eq!(local_role_changed.pending_count().await.unwrap(), 0);
9983        assert_eq!(remote_unproven.pending_count().await.unwrap(), 1);
9984        let good_child = store.load_session("receipt-good").await.unwrap().unwrap();
9985        let wrong_child = store
9986            .load_session("receipt-wrong-mailbox")
9987            .await
9988            .unwrap()
9989            .unwrap();
9990        let wrong_broker_child = store
9991            .load_session("receipt-wrong-broker")
9992            .await
9993            .unwrap()
9994            .unwrap();
9995        let local_role_changed_child = store
9996            .load_session("receipt-local-role-changed")
9997            .await
9998            .unwrap()
9999            .unwrap();
10000        let remote_unproven_child = store
10001            .load_session("receipt-remote-unproven")
10002            .await
10003            .unwrap()
10004            .unwrap();
10005        assert!(store
10006            .recover_broker_terminal_receipts(&good_child)
10007            .await
10008            .unwrap()
10009            .is_empty());
10010        assert_eq!(
10011            store
10012                .recover_broker_terminal_receipts(&wrong_child)
10013                .await
10014                .unwrap()
10015                .len(),
10016            1
10017        );
10018        assert_eq!(
10019            store
10020                .recover_broker_terminal_receipts(&wrong_broker_child)
10021                .await
10022                .unwrap()
10023                .len(),
10024            1
10025        );
10026        assert!(store
10027            .recover_broker_terminal_receipts(&local_role_changed_child)
10028            .await
10029            .unwrap()
10030            .is_empty());
10031        assert_eq!(
10032            store
10033                .recover_broker_terminal_receipts(&remote_unproven_child)
10034                .await
10035                .unwrap()
10036                .len(),
10037            1
10038        );
10039    }
10040
10041    #[tokio::test]
10042    async fn broker_receipt_confirmation_requires_exact_birth_run_and_committed_status() {
10043        let runner = ActorChildRunner::new(
10044            "test".into(),
10045            PathBuf::new(),
10046            vec![],
10047            PathBuf::new(),
10048            ExecutorSpec::BambooRuntime,
10049            vec![],
10050            "test".into(),
10051            1,
10052        );
10053        let home = tempfile::tempdir().unwrap();
10054        let store = Arc::new(
10055            bamboo_storage::SessionStoreV2::new(home.path().to_path_buf())
10056                .await
10057                .unwrap(),
10058        );
10059        let root = Session::new("ack-parent", "model");
10060        store.save_session(&root).await.unwrap();
10061        let mut child = Session::new_child_of("ack-child", &root, "model", "task");
10062        child.add_message(bamboo_agent_core::Message::user("work"));
10063        store.save_session(&child).await.unwrap();
10064        child.add_message(bamboo_agent_core::Message::assistant("done", None));
10065        child.set_last_run_status("completed");
10066        store
10067            .prepare_broker_terminal_receipt(
10068                &child,
10069                "run-1",
10070                "00000000-0000-4000-8000-000000000001",
10071                "p-ack-child",
10072                "broker-run-1",
10073                &["broker-outcome-1".into()],
10074            )
10075            .await
10076            .unwrap();
10077        store.save_session(&child).await.unwrap();
10078        runner.set_actor_directory_store(Some(store));
10079        let calls = Arc::new(AtomicUsize::new(0));
10080        insert_ack_probe(&runner, &child, &calls, TerminalStatus::Completed).await;
10081        runner
10082            .confirm_durable_child_delivery(&child, "wrong-run", true)
10083            .await
10084            .unwrap();
10085        let mut wrong_birth = child.clone();
10086        wrong_birth.created_at += chrono::Duration::seconds(1);
10087        runner
10088            .confirm_durable_child_delivery(&wrong_birth, "run-1", true)
10089            .await
10090            .unwrap();
10091        assert_eq!(runner.pending_durable_links.lock().await.len(), 1);
10092        child.set_last_run_status("completed");
10093        runner
10094            .confirm_durable_child_delivery(&child, "run-1", false)
10095            .await
10096            .unwrap();
10097        assert_eq!(calls.load(Ordering::SeqCst), 0);
10098
10099        insert_ack_probe(&runner, &child, &calls, TerminalStatus::Completed).await;
10100        child.set_last_run_status("error");
10101        runner
10102            .confirm_durable_child_delivery(&child, "run-1", true)
10103            .await
10104            .unwrap();
10105        assert_eq!(calls.load(Ordering::SeqCst), 0);
10106
10107        insert_ack_probe(&runner, &child, &calls, TerminalStatus::Completed).await;
10108        child.set_last_run_status("completed");
10109        runner
10110            .confirm_durable_child_delivery(&child, "run-1", true)
10111            .await
10112            .unwrap();
10113        assert_eq!(calls.load(Ordering::SeqCst), 1);
10114    }
10115
10116    #[tokio::test]
10117    async fn strict_completeness_gate_precedes_ack_and_ack_failure_preserves_cold_proof() {
10118        for commit in [false, true] {
10119            let runner = ActorChildRunner::new(
10120                "test".into(),
10121                PathBuf::new(),
10122                vec![],
10123                PathBuf::new(),
10124                ExecutorSpec::BambooRuntime,
10125                vec![],
10126                "test".into(),
10127                1,
10128            );
10129            let home = tempfile::tempdir().unwrap();
10130            let store = Arc::new(
10131                bamboo_storage::SessionStoreV2::new(home.path().into())
10132                    .await
10133                    .unwrap(),
10134            );
10135            let root = Session::new("strict-ack-parent", "model");
10136            store.save_session(&root).await.unwrap();
10137            let mut child = Session::new_child_of("strict-ack-child", &root, "model", "work");
10138            child.add_message(bamboo_agent_core::Message::user("work"));
10139            store.save_session(&child).await.unwrap();
10140            child.add_message(bamboo_agent_core::Message::assistant("finished", None));
10141            child.set_last_run_status("completed");
10142            let proof = bamboo_storage::v2::HostTerminalCompleteness::from_verified_host(
10143                &child, "run-1", 7, 2,
10144            )
10145            .unwrap();
10146            let creation = bamboo_subagent::proto::ChildCreationIdentity {
10147                created_at: child.created_at,
10148                spawn_depth: child.spawn_depth,
10149            };
10150            let history = StrictTerminalHistory {
10151                proof: proof.clone(),
10152                watermark: bamboo_subagent::ActorEventWatermark {
10153                    version: 1,
10154                    logical_session: Some(actor_event_logical_identity(
10155                        &child,
10156                        &root.id,
10157                        Some(&creation),
10158                    )),
10159                    activation_id: Some("run-1".into()),
10160                    execution_epoch: 7,
10161                    final_seq: 2,
10162                },
10163            };
10164            let parent_mailbox = "p-strict-ack-child";
10165            store
10166                .prepare_broker_terminal_receipt_with_completeness(
10167                    &child,
10168                    bamboo_storage::v2::BrokerTerminalRoute {
10169                        activation_run_id: "run-1",
10170                        broker_identity: "00000000-0000-4000-8000-000000000001",
10171                        parent_mailbox,
10172                        broker_correlation_id: "broker-run-1",
10173                        message_ids: &["broker-outcome-1".into()],
10174                    },
10175                    &proof,
10176                )
10177                .await
10178                .unwrap();
10179            store.save_session(&child).await.unwrap();
10180            runner.set_actor_directory_store(Some(store.clone()));
10181            let calls = Arc::new(AtomicUsize::new(0));
10182            runner.pending_durable_links.lock().await.insert(
10183                (child.id.clone(), child.created_at, "run-1".into()),
10184                PendingDurableChildLink {
10185                    link: Box::new(DurableAckProbe {
10186                        calls: calls.clone(),
10187                        fail: true,
10188                        terminal_status: TerminalStatus::Completed,
10189                        parent_mailbox: parent_mailbox.into(),
10190                    }),
10191                    host_nested_wait_handoff: false,
10192                    strict_history: Some(history),
10193                    strict_history_required: true,
10194                    strict_history_committed: false,
10195                },
10196            );
10197            if commit {
10198                runner
10199                    .commit_durable_child_delivery(&child, "run-1")
10200                    .await
10201                    .unwrap();
10202                assert_eq!(calls.load(Ordering::SeqCst), 0, "proof commit cannot ACK");
10203            }
10204            assert!(runner
10205                .confirm_durable_child_delivery(&child, "run-1", true)
10206                .await
10207                .is_err());
10208            assert_eq!(calls.load(Ordering::SeqCst), usize::from(commit));
10209            let cold = bamboo_storage::SessionStoreV2::new(home.path().into())
10210                .await
10211                .unwrap();
10212            let recovered = cold.recover_broker_terminal_receipts(&child).await;
10213            if commit {
10214                let receipts = recovered.unwrap();
10215                assert_eq!(receipts.len(), 1);
10216                assert_eq!(
10217                    receipts[0]
10218                        .terminal_completeness
10219                        .as_ref()
10220                        .unwrap()
10221                        .contiguous_applied_seq,
10222                    2
10223                );
10224                assert_eq!(
10225                    cold.load_session(&child.id)
10226                        .await
10227                        .unwrap()
10228                        .unwrap()
10229                        .last_run_status()
10230                        .as_deref(),
10231                    Some("completed")
10232                );
10233            } else {
10234                assert!(
10235                    recovered.is_err(),
10236                    "matching canonical history is not proof commitment"
10237                );
10238            }
10239            assert_eq!(
10240                serde_json::to_value(
10241                    cold.load_session(&child.id)
10242                        .await
10243                        .unwrap()
10244                        .unwrap()
10245                        .messages
10246                )
10247                .unwrap(),
10248                serde_json::to_value(&child.messages).unwrap(),
10249            );
10250        }
10251    }
10252
10253    #[tokio::test]
10254    async fn broker_receipt_confirms_canonical_worker_error_and_cancelled_status() {
10255        for (terminal_status, status) in [
10256            (TerminalStatus::Error, "error"),
10257            (TerminalStatus::Cancelled, "cancelled"),
10258        ] {
10259            let runner = ActorChildRunner::new(
10260                "test".into(),
10261                PathBuf::new(),
10262                vec![],
10263                PathBuf::new(),
10264                ExecutorSpec::BambooRuntime,
10265                vec![],
10266                "test".into(),
10267                1,
10268            );
10269            let home = tempfile::tempdir().unwrap();
10270            let store = Arc::new(
10271                bamboo_storage::SessionStoreV2::new(home.path().to_path_buf())
10272                    .await
10273                    .unwrap(),
10274            );
10275            let root = Session::new("failed-ack-parent", "model");
10276            store.save_session(&root).await.unwrap();
10277            let mut child = Session::new_child_of("failed-ack-child", &root, "model", "task");
10278            child.add_message(bamboo_agent_core::Message::user("work"));
10279            store.save_session(&child).await.unwrap();
10280            child.set_last_run_status(status);
10281            child.set_last_run_error(format!("worker {status}"));
10282            store
10283                .prepare_broker_terminal_receipt(
10284                    &child,
10285                    "run-1",
10286                    "00000000-0000-4000-8000-000000000001",
10287                    "p-failed-ack-child",
10288                    "broker-run-1",
10289                    &["broker-outcome-1".into()],
10290                )
10291                .await
10292                .unwrap();
10293            store.save_session(&child).await.unwrap();
10294            runner.set_actor_directory_store(Some(store));
10295            let calls = Arc::new(AtomicUsize::new(0));
10296            insert_ack_probe(&runner, &child, &calls, terminal_status).await;
10297            runner
10298                .confirm_durable_child_delivery(&child, "run-1", true)
10299                .await
10300                .unwrap();
10301            assert_eq!(calls.load(Ordering::SeqCst), 1, "{status}");
10302        }
10303    }
10304
10305    #[tokio::test]
10306    async fn broker_completed_terminal_projects_only_host_verified_nested_wait() {
10307        assert_eq!(
10308            effective_broker_terminal_status(TerminalStatus::Completed, true).unwrap(),
10309            "suspended"
10310        );
10311        for raw in [TerminalStatus::Error, TerminalStatus::Cancelled] {
10312            assert!(effective_broker_terminal_status(raw, true).is_err());
10313        }
10314        let runner = ActorChildRunner::new(
10315            "test".into(),
10316            PathBuf::new(),
10317            vec![],
10318            PathBuf::new(),
10319            ExecutorSpec::BambooRuntime,
10320            vec![],
10321            "test".into(),
10322            1,
10323        );
10324        let home = tempfile::tempdir().unwrap();
10325        let store = Arc::new(
10326            bamboo_storage::SessionStoreV2::new(home.path().to_path_buf())
10327                .await
10328                .unwrap(),
10329        );
10330        let root = Session::new("nested-ack-parent", "model");
10331        store.save_session(&root).await.unwrap();
10332        let mut child = Session::new_child_of("nested-ack-child", &root, "model", "task");
10333        child.add_message(bamboo_agent_core::Message::user("spawn a grandchild"));
10334        store.save_session(&child).await.unwrap();
10335        child.set_last_run_status("suspended");
10336        store
10337            .prepare_broker_terminal_receipt(
10338                &child,
10339                "run-1",
10340                "00000000-0000-4000-8000-000000000001",
10341                "p-nested-ack-child",
10342                "broker-run-1",
10343                &["broker-outcome-1".into()],
10344            )
10345            .await
10346            .unwrap();
10347        store.save_session(&child).await.unwrap();
10348        runner.set_actor_directory_store(Some(store.clone()));
10349        let calls = Arc::new(AtomicUsize::new(0));
10350
10351        insert_ack_probe(&runner, &child, &calls, TerminalStatus::Completed).await;
10352        runner
10353            .confirm_durable_child_delivery(&child, "run-1", true)
10354            .await
10355            .unwrap();
10356        assert_eq!(calls.load(Ordering::SeqCst), 0);
10357
10358        insert_ack_probe_with_nested_wait(&runner, &child, &calls, TerminalStatus::Completed, true)
10359            .await;
10360        runner
10361            .confirm_durable_child_delivery(&child, "run-1", true)
10362            .await
10363            .unwrap();
10364        assert_eq!(calls.load(Ordering::SeqCst), 1);
10365        assert_eq!(
10366            store
10367                .load_session(&child.id)
10368                .await
10369                .unwrap()
10370                .unwrap()
10371                .last_run_status()
10372                .as_deref(),
10373            Some("suspended")
10374        );
10375        assert!(store
10376            .recover_broker_terminal_receipts(&child)
10377            .await
10378            .unwrap()
10379            .is_empty());
10380    }
10381
10382    #[tokio::test]
10383    async fn host_completion_notifies_once_only_after_reply_and_finish_commit() {
10384        struct CommittedObserver {
10385            authority: PathBuf,
10386            events: std::sync::Mutex<Vec<PublicActorEvent>>,
10387        }
10388        impl ActorEventObserver for CommittedObserver {
10389            fn publish(&self, event: PublicActorEvent) {
10390                // Observe the real durable sidecar synchronously at publication.
10391                let entry: bamboo_domain::ActorDirectoryEntry =
10392                    serde_json::from_slice(&std::fs::read(&self.authority).unwrap()).unwrap();
10393                let committed = entry.activation.unwrap();
10394                assert_eq!(
10395                    committed.status,
10396                    bamboo_domain::ActorActivationStatus::Succeeded
10397                );
10398                assert_eq!(committed.activation_id, event.activation_id);
10399                assert!(committed.finished_at.is_some());
10400                self.events.lock().unwrap().push(event);
10401            }
10402        }
10403        struct CompletionLink(VecDeque<ChildFrame>);
10404        #[async_trait]
10405        impl bamboo_subagent::ChildLink for CompletionLink {
10406            async fn send(&mut self, _: ParentFrame) -> bamboo_subagent::TransportResult<()> {
10407                Ok(())
10408            }
10409            async fn next_frame(&mut self) -> bamboo_subagent::TransportResult<Option<ChildFrame>> {
10410                Ok(self.0.pop_front())
10411            }
10412        }
10413        for cut in [
10414            "success",
10415            "no_terminal",
10416            "invalid_terminal",
10417            "stale",
10418            "storage_failure",
10419            "append_failure",
10420        ] {
10421            let temp = tempfile::tempdir().unwrap();
10422            let store = Arc::new(
10423                bamboo_storage::SessionStoreV2::new(temp.path().into())
10424                    .await
10425                    .unwrap(),
10426            );
10427            let root = Session::new("notify-root", "model");
10428            store.save_session(&root).await.unwrap();
10429            let mut child = Session::new_child_of("notify-child", &root, "model", "assignment");
10430            store.save_session(&child).await.unwrap();
10431            let inbox: Arc<dyn SessionInboxPort> = Arc::new(bamboo_storage::FileSessionInbox::new(
10432                store.clone(),
10433                bamboo_domain::SessionInboxLimits::default(),
10434            ));
10435            let binding = actor_binding(
10436                store.clone(),
10437                inbox,
10438                Arc::new(bamboo_storage::LockedSessionStore::new(store.clone())),
10439            );
10440            let policy = Arc::new(bamboo_tools::permission::PermissionConfig::new());
10441            bind_local_control_plane(store.as_ref(), &child.id, policy.as_ref()).await;
10442            let activation = PlainActorActivation::start(
10443                store.clone(),
10444                &child,
10445                &binding,
10446                Some("notify-run"),
10447                Some(policy),
10448                false,
10449                None,
10450            )
10451            .await
10452            .unwrap();
10453            let _owner = binding
10454                .router
10455                .register_run(&child.id, "notify-run")
10456                .await
10457                .unwrap();
10458            let batch = ActorEventBatch {
10459                logical_session: Some(LogicalSessionIdentity {
10460                    session_id: child.id.clone(),
10461                    parent_session_id: child.parent_session_id.clone(),
10462                    root_session_id: child.root_session_id.clone(),
10463                    creation: None,
10464                }),
10465                activation_id: Some("notify-run".into()),
10466                execution_epoch: 9,
10467                source_node_id: None,
10468                source_actor_id: Some("notify-worker".into()),
10469                first_seq: 1,
10470                last_seq: 2,
10471                qos: bamboo_subagent::ActorEventQos::Durable,
10472                events: vec![
10473                    serde_json::to_value(AgentEvent::Token {
10474                        content: "PRIVATE_PROVIDER_TOKEN".into(),
10475                    })
10476                    .unwrap(),
10477                    serde_json::to_value(AgentEvent::message_appended(
10478                        &child.id,
10479                        &bamboo_agent_core::Message::assistant("PRIVATE_WORKER_CACHE", None),
10480                    ))
10481                    .unwrap(),
10482                ],
10483            };
10484            let mut frames = VecDeque::from([ChildFrame::EventBatch { batch }]);
10485            if cut != "no_terminal" {
10486                let mut terminal = completed_actor_frame();
10487                if cut == "invalid_terminal" {
10488                    if let ChildFrame::Terminal { status, .. } = &mut terminal {
10489                        *status = TerminalStatus::Suspended;
10490                    }
10491                }
10492                frames.push_back(terminal);
10493            }
10494            let mut link = CompletionLink(frames);
10495            let (tx, mut rx) = mpsc::channel(8);
10496            let (_live, mut live_rx) = mpsc::unbounded_channel();
10497            let (_delivery, mut delivery_rx) = mpsc::unbounded_channel();
10498            let cancel = CancellationToken::new();
10499            let outcome = drive(ActorDriveContext {
10500                client: &mut link,
10501                parent_session_id: "notify-root",
10502                child_session_id: "notify-child",
10503                child_attempt: 0,
10504                approval_registry: None,
10505                approval_decider: None,
10506                approval_reviewer: None,
10507                escalation_bridge: None,
10508                event_tx: &tx,
10509                cancel_token: &cancel,
10510                live_rx: &mut live_rx,
10511                delivery_rx: &mut delivery_rx,
10512                logical_session: &mut child,
10513                expected_permission_posture: None,
10514                expected_creation: None,
10515                session_inbox_runtime: Some(&binding),
10516                actor_directory_store: Some(store.as_ref()),
10517                canonical_subagent_tool: None,
10518                ticket_service: None,
10519                activation_run_id: Some("notify-run"),
10520                execution_epoch: 9,
10521                expected_source_actor_id: "notify-worker",
10522                initial_inflight_claims: VecDeque::new(),
10523                plain_actor: true,
10524                remote_environment_lease: false,
10525                readonly_output: None,
10526                local_history_tools: None,
10527                local_history_read_only: false,
10528                strict_history_output: None,
10529                plain_input: Some(&activation),
10530                canonical_activation: None,
10531                canonical_placement_ref: None,
10532                actor_event_observer: None,
10533                plain_run: None,
10534                first_frame_timeout: Some(Duration::from_secs(1)),
10535            })
10536            .await;
10537            let admitted = !matches!(cut, "no_terminal" | "invalid_terminal");
10538            assert_eq!(outcome.is_ok(), admitted, "{cut}: {outcome:?}");
10539            assert_eq!(
10540                activation.completion_router.lock().unwrap().is_some(),
10541                admitted,
10542                "{cut}"
10543            );
10544            while rx.try_recv().is_ok() {}
10545            let authority = temp
10546                .path()
10547                .join(store.resolve_rel_path(&child.id).await.unwrap())
10548                .join("actor-authority.json");
10549            let observer = CommittedObserver {
10550                authority: authority.clone(),
10551                events: Default::default(),
10552            };
10553            if cut == "append_failure" {
10554                child.add_message(bamboo_agent_core::Message::user("uncommitted-prefix"));
10555                assert!(activation
10556                    .append_reply(&mut child, "PRIVATE_REPLY".into(), &tx)
10557                    .await
10558                    .is_err());
10559                activation
10560                    .finish(ActorActivationFinish::Failed)
10561                    .await
10562                    .unwrap();
10563                assert!(observer.events.lock().unwrap().is_empty());
10564                continue;
10565            }
10566            activation
10567                .append_reply(&mut child, "PRIVATE_REPLY".into(), &tx)
10568                .await
10569                .unwrap();
10570            assert!(
10571                rx.try_recv().is_ok(),
10572                "legacy content notification retained"
10573            );
10574            assert!(
10575                observer.events.lock().unwrap().is_empty(),
10576                "reply alone must not publish"
10577            );
10578            if cut == "stale" {
10579                store
10580                    .retire_actor(&child.id, chrono::Utc::now())
10581                    .await
10582                    .unwrap();
10583            } else if cut == "storage_failure" {
10584                // Make the canonical write path unreadable, rather than mocking
10585                // the store or allowing completion to fall back to ordinary save.
10586                std::fs::remove_file(&authority).unwrap();
10587                std::fs::create_dir(&authority).unwrap();
10588            }
10589            let result = activation
10590                .finish_with_observer(ActorActivationFinish::Succeeded, Some(&observer))
10591                .await;
10592            assert_eq!(
10593                result.is_ok(),
10594                matches!(cut, "success" | "no_terminal" | "invalid_terminal"),
10595                "{cut}"
10596            );
10597            let events = observer.events.lock().unwrap().clone();
10598            assert_eq!(events.len(), usize::from(cut == "success"), "{cut}");
10599            if let Some(event) = events.first() {
10600                assert_eq!(event.actor_id, child.id);
10601                assert_eq!(event.root_actor_id, root.id);
10602                assert_eq!(event.parent_actor_id, Some(root.id.clone()));
10603                assert_eq!(event.source_order.sequence, 3);
10604                let json = serde_json::to_string(event).unwrap();
10605                assert!(!json.contains("PRIVATE_") && !json.contains("notify-run"));
10606                let persisted = store.load_session(&child.id).await.unwrap().unwrap();
10607                assert_eq!(persisted.messages.last().unwrap().content, "PRIVATE_REPLY");
10608            }
10609            assert!(activation
10610                .finish_with_observer(ActorActivationFinish::Succeeded, Some(&observer))
10611                .await
10612                .is_err());
10613            assert_eq!(
10614                observer.events.lock().unwrap().len(),
10615                events.len(),
10616                "repeat finish cannot publish"
10617            );
10618        }
10619    }
10620
10621    async fn bind_local_control_plane(
10622        store: &bamboo_storage::SessionStoreV2,
10623        actor_id: &str,
10624        policy: &bamboo_tools::permission::PermissionConfig,
10625    ) {
10626        store
10627            .bind_control_plane(
10628                actor_id,
10629                bamboo_domain::ActorControlPlaneBinding {
10630                    policy_revision: Some(policy.policy_revision()),
10631                    placement_intent: Some(bamboo_domain::ActorPlacementIntent::Local),
10632                },
10633            )
10634            .await
10635            .unwrap();
10636    }
10637
10638    fn local_tool_sample(
10639        host: &Session,
10640        denied_write: bool,
10641    ) -> (LocalToolCollector, Vec<bamboo_agent_core::Message>) {
10642        local_tool_sample_with_reasoning(host, denied_write, None)
10643    }
10644
10645    fn local_tool_sample_with_reasoning(
10646        host: &Session,
10647        denied_write: bool,
10648        reasoning: Option<&str>,
10649    ) -> (LocalToolCollector, Vec<bamboo_agent_core::Message>) {
10650        use bamboo_agent_core::{FunctionCall, Message, ToolCall};
10651        let calls = ["Read", "Write"]
10652            .map(|name| ToolCall {
10653                id: format!("local-{name}"),
10654                tool_type: "function".into(),
10655                function: FunctionCall {
10656                    name: name.into(),
10657                    arguments: "{\"file_path\":\"answer.txt\"}".into(),
10658                },
10659            })
10660            .to_vec();
10661        let mut rows = vec![Message::assistant("checking", Some(calls))];
10662        rows[0].reasoning = reasoning.map(str::to_owned);
10663        let mut collector = LocalToolCollector::default();
10664        if let Some(reasoning) = reasoning {
10665            assert!(collector
10666                .event(
10667                    &serde_json::to_value(AgentEvent::ReasoningToken {
10668                        content: reasoning.into(),
10669                    })
10670                    .unwrap()
10671                )
10672                .unwrap());
10673        }
10674        for name in ["Read", "Write"] {
10675            let denied = name == "Write" && denied_write;
10676            let id = format!("local-{name}");
10677            if !denied {
10678                collector
10679                    .event(
10680                        &serde_json::to_value(AgentEvent::ToolStart {
10681                            tool_call_id: id.clone(),
10682                            tool_name: name.into(),
10683                            arguments: serde_json::json!({"file_path":"answer.txt"}),
10684                        })
10685                        .unwrap(),
10686                    )
10687                    .unwrap();
10688            }
10689            let text = if denied {
10690                "Error: native_tool_ceiling_denied: Write"
10691            } else {
10692                "actual file result"
10693            };
10694            let mut message = Message::tool_result_with_status(id.clone(), text, !denied);
10695            message.metadata = Some(serde_json::json!({"elapsed_ms":7,"tool_name":name,
10696                "is_mutating":name == "Write","auto_approved":name != "Write","success":!denied}));
10697            rows.push(message);
10698            let event = if denied {
10699                AgentEvent::ToolError {
10700                    tool_call_id: id,
10701                    error: "native_tool_ceiling_denied: Write".into(),
10702                }
10703            } else {
10704                AgentEvent::ToolComplete {
10705                    tool_call_id: id,
10706                    result: bamboo_agent_core::tools::ToolResult::text(true, text),
10707                }
10708            };
10709            collector
10710                .event(&serde_json::to_value(event).unwrap())
10711                .unwrap();
10712        }
10713        rows.push(Message::assistant("complete exact report", None));
10714        rows.last_mut().unwrap().reasoning = reasoning.map(str::to_owned);
10715        if let Some(reasoning) = reasoning {
10716            assert!(collector
10717                .event(
10718                    &serde_json::to_value(AgentEvent::ReasoningToken {
10719                        content: reasoning.into(),
10720                    })
10721                    .unwrap()
10722                )
10723                .unwrap());
10724        }
10725        let mut messages = host.messages.clone();
10726        if let Some(system) = messages
10727            .iter_mut()
10728            .find(|message| message.role == Role::System)
10729        {
10730            system.content = "worker-private system never imported".into();
10731        }
10732        messages.extend(rows.iter().cloned());
10733        let data = bamboo_subagent::proto::LocalToolMessages::Complete {
10734            version: 1,
10735            messages: messages
10736                .iter()
10737                .map(|message| serde_json::to_value(message).unwrap())
10738                .collect(),
10739        };
10740        assert!(!collector
10741            .event(&serde_json::to_value(data).unwrap())
10742            .unwrap());
10743        (collector, rows)
10744    }
10745
10746    #[test]
10747    fn ticket_task_event_requires_exact_host_callback_receipt() {
10748        let mut collector = LocalToolCollector::default();
10749        let args = serde_json::json!({"tasks":[{"content":"own step","status":"pending"}]});
10750        let start = serde_json::to_value(AgentEvent::ToolStart {
10751            tool_call_id: "task-call".into(),
10752            tool_name: "Task".into(),
10753            arguments: args.clone(),
10754        })
10755        .unwrap();
10756        let mut receipts = HashMap::new();
10757        assert!(collector.ticket_event(&start, &receipts).unwrap());
10758        collector.event(&start).unwrap();
10759        let complete = serde_json::to_value(AgentEvent::ToolComplete {
10760            tool_call_id: "task-call".into(),
10761            result: bamboo_agent_core::tools::ToolResult::text(true, "done"),
10762        })
10763        .unwrap();
10764        assert!(collector.ticket_event(&complete, &receipts).is_err());
10765        receipts.insert("other-call".into(), args.clone());
10766        assert!(collector.ticket_event(&complete, &receipts).is_err());
10767        receipts.insert("task-call".into(), serde_json::json!({"tasks":[]}));
10768        assert!(collector.ticket_event(&complete, &receipts).is_err());
10769        receipts.insert("task-call".into(), args);
10770        assert!(collector.ticket_event(&complete, &receipts).unwrap());
10771        let forged_root = serde_json::to_value(AgentEvent::TaskListUpdated {
10772            task_list: bamboo_tools::TaskTool::task_list_from_args(
10773                &serde_json::json!({"tasks":[{"content":"forged root","status":"completed"}]}),
10774                "root",
10775            )
10776            .unwrap(),
10777            version: Some(999),
10778        })
10779        .unwrap();
10780        assert!(!collector.ticket_event(&forged_root, &receipts).unwrap());
10781    }
10782
10783    #[test]
10784    fn ticket_question_terminal_requires_exact_host_receipt_and_tool_result() {
10785        use bamboo_agent_core::{FunctionCall, Message, ToolCall};
10786        let mut host = Session::new("question-child", "model");
10787        host.messages.push(Message::user("own assignment"));
10788        let question: bamboo_tickets::PendingRequest = serde_json::from_value(serde_json::json!({
10789            "id":"own-question", "work_id":"own-work", "assignment_id":"own-assignment",
10790            "generation":1, "contract_revision":1, "prompt_revision":1,
10791            "kind":{"kind":"question"}, "prompt":"精确问题", "status":"open",
10792            "answer":null, "consumed_attempt":null, "worker_packet_hash":"a".repeat(64), "updated_seq":3
10793        })).unwrap();
10794        let args = serde_json::json!({"tasks":[{"content":"own step","status":"blocked"}], "question":{"prompt":"精确问题"}});
10795        for case in ["exact", "no_host_receipt", "forged_request", "fake_report"] {
10796            let mut returned =
10797                serde_json::json!({"status":"waiting_for_answer", "request":question});
10798            if case == "forged_request" {
10799                returned["request"]["id"] = serde_json::json!("sibling-question");
10800            }
10801            let text = returned.to_string();
10802            let rows = [
10803                Message::assistant(
10804                    "",
10805                    Some(vec![ToolCall {
10806                        id: "own-call".into(),
10807                        tool_type: "function".into(),
10808                        function: FunctionCall {
10809                            name: "Task".into(),
10810                            arguments: args.to_string(),
10811                        },
10812                    }]),
10813                ),
10814                Message::tool_result_with_status("own-call", &text, true),
10815            ];
10816            let collector = LocalToolCollector {
10817                messages: Some(host.messages.iter().cloned().chain(rows).collect()),
10818                starts: HashMap::from([("own-call".into(), ("Task".into(), args.clone()))]),
10819                outcomes: HashMap::from([("own-call".into(), (true, text))]),
10820                ticket_yield: (case != "no_host_receipt")
10821                    .then(|| ("own-call".into(), question.clone())),
10822            };
10823            let terminal = if case == "fake_report" {
10824                Some("fake delivery")
10825            } else {
10826                Some("")
10827            };
10828            assert_eq!(
10829                collector
10830                    .suffix(&host, &["Task".into()], false, terminal)
10831                    .is_ok(),
10832                case == "exact",
10833                "{case}"
10834            );
10835        }
10836    }
10837
10838    #[test]
10839    fn local_tool_suffix_preserves_full_current_prefix_and_rejects_authority_or_collisions() {
10840        let mut host = Session::new("tool-history-prefix", "model");
10841        host.messages
10842            .push(bamboo_agent_core::Message::user("assignment"));
10843        let mut confirmed = bamboo_agent_core::Message::user("confirmed during Run");
10844        confirmed.metadata = Some(
10845            serde_json::json!({"session_message":{"id":confirmed.id,"target_session_id":host.id}}),
10846        );
10847        host.messages.push(confirmed);
10848        let tools = vec!["Read".into(), "Write".into()];
10849        let (good, rows) = local_tool_sample(&host, false);
10850        assert_eq!(
10851            serde_json::to_value(
10852                good.suffix(&host, &tools, false, Some("complete exact report"))
10853                    .unwrap()
10854            )
10855            .unwrap(),
10856            serde_json::to_value(&rows).unwrap()
10857        );
10858        for case in [
10859            "missing",
10860            "changed",
10861            "metadata",
10862            "system",
10863            "user",
10864            "collision",
10865            "dangling",
10866            "result",
10867            "terminal",
10868            "tool_metadata",
10869            "grant",
10870            "start_arguments",
10871            "ledger",
10872        ] {
10873            let (mut collector, _) = local_tool_sample(&host, false);
10874            let mut current = host.clone();
10875            let messages = collector.messages.as_mut().unwrap();
10876            match case {
10877                "missing" => {
10878                    messages.remove(host.messages.len() - 1);
10879                }
10880                "changed" => messages[0].content.push('!'),
10881                "metadata" => messages[1].metadata = None,
10882                "system" => messages.last_mut().unwrap().role = Role::System,
10883                "user" => messages.last_mut().unwrap().role = Role::User,
10884                "collision" => messages.last_mut().unwrap().id = host.messages[0].id.clone(),
10885                "dangling" => {
10886                    messages[host.messages.len() + 1].tool_call_id = Some("foreign".into())
10887                }
10888                "result" => messages[host.messages.len() + 1].content.push('!'),
10889                "tool_metadata" => {
10890                    messages[host.messages.len() + 1].metadata.as_mut().unwrap()["permission"] =
10891                        true.into()
10892                }
10893                "grant" => {
10894                    messages[host.messages.len()].tool_calls.as_mut().unwrap()[0]
10895                        .function
10896                        .name = "Bash".into()
10897                }
10898                "start_arguments" => {
10899                    collector.starts.get_mut("local-Read").unwrap().1 =
10900                        serde_json::json!({"file_path":"different.txt"})
10901                }
10902                "ledger" => {
10903                    current.append_provider_transcript_group(current.messages[0].id.clone(), None, vec![
10904                        bamboo_domain::session::provider_transcript::ProviderTranscriptItem::try_from_payload(
10905                            bamboo_domain::ProviderFamily::OpenAi,
10906                            bamboo_domain::ProviderProtocol::OpenAiResponsesV1,
10907                            bamboo_domain::session::provider_transcript::ProviderTranscriptOrigin::HostToolSearch,
10908                            bamboo_domain::session::provider_transcript::ProviderTranscriptAuthor::ToolResult,
10909                            serde_json::json!({"type":"tool_search_output","execution":"client","call_id":"search","status":"completed","tools":[]}),
10910                        ).unwrap(),
10911                    ]).unwrap();
10912                }
10913                "terminal" => {}
10914                _ => unreachable!(),
10915            }
10916            let terminal = if case == "terminal" {
10917                "different"
10918            } else {
10919                "complete exact report"
10920            };
10921            assert!(
10922                collector
10923                    .suffix(&current, &tools, false, Some(terminal))
10924                    .is_err(),
10925                "{case}"
10926            );
10927        }
10928        let (denied, _) = local_tool_sample(&host, true);
10929        assert!(denied
10930            .suffix(&host, &["Read".into()], true, Some("complete exact report"))
10931            .is_ok());
10932        let (mut contradicted, _) = local_tool_sample(&host, true);
10933        contradicted.starts.insert(
10934            "local-Write".into(),
10935            ("Read".into(), serde_json::json!({"file_path":"answer.txt"})),
10936        );
10937        assert!(contradicted
10938            .suffix(&host, &["Read".into()], true, Some("complete exact report"))
10939            .is_err());
10940        let (successful, _) = local_tool_sample(&host, false);
10941        assert!(successful
10942            .suffix(&host, &["Read".into()], true, Some("complete exact report"))
10943            .is_err());
10944        let mut empty = LocalToolCollector::default();
10945        assert!(empty
10946            .event(
10947                &serde_json::to_value(AgentEvent::ReasoningToken {
10948                    content: "private".into()
10949                })
10950                .unwrap()
10951            )
10952            .unwrap());
10953        assert!(
10954            empty
10955                .suffix(&host, &tools, false, Some("complete exact report"))
10956                .is_err(),
10957            "stream observations alone cannot commit history"
10958        );
10959        let (mut completed, _) = local_tool_sample(&host, false);
10960        assert!(completed.event(&serde_json::json!({"type":"local_client_tool_messages_v1","version":1,"messages":[]})).is_err());
10961    }
10962
10963    struct LocalToolConcurrentCheckpoint {
10964        real: SessionInboxRuntimeBinding,
10965        claim: SessionInboxClaim,
10966    }
10967    #[async_trait]
10968    impl RuntimeSessionPersistence for LocalToolConcurrentCheckpoint {
10969        async fn save_runtime_session(&self, session: &mut Session) -> std::io::Result<()> {
10970            self.real.persistence.save_runtime_session(session).await
10971        }
10972        async fn checkpoint_runtime_session(&self, session: &mut Session) -> std::io::Result<()> {
10973            // A real canonical admission wins after validation, before the
10974            // final append-safe checkpoint acquires its serialization lock.
10975            let mut latest = self.real.storage.load_session(&session.id).await?.unwrap();
10976            checkpoint_and_ack_canonical_claim(&self.real, &mut latest, &self.claim, None)
10977                .await
10978                .map_err(|_| std::io::Error::other("concurrent admission failed"))?;
10979            self.real
10980                .persistence
10981                .checkpoint_runtime_session(session)
10982                .await
10983        }
10984        async fn load_runtime_session(&self, id: &str) -> std::io::Result<Option<Session>> {
10985            self.real.storage.load_session(id).await
10986        }
10987    }
10988
10989    #[tokio::test]
10990    async fn local_tool_history_checkpoint_preserves_confirmed_input_and_cold_pairs_or_refuses() {
10991        for case in [
10992            "success",
10993            "reasoning",
10994            "concurrent_input",
10995            "checkpoint_error",
10996            "stale_prefix",
10997            "cancelled",
10998            "lost_owner",
10999        ] {
11000            let temp = tempfile::tempdir().unwrap();
11001            let store = Arc::new(
11002                bamboo_storage::SessionStoreV2::new(temp.path().to_path_buf())
11003                    .await
11004                    .unwrap(),
11005            );
11006            let root = Session::new("local-tool-root", "model");
11007            store.save_session(&root).await.unwrap();
11008            let mut child = Session::new_child("local-tool-child", &root.id, "model", "child");
11009            let mut system = bamboo_agent_core::Message::user("Host-owned immutable System");
11010            system.role = Role::System;
11011            child.messages.push(system);
11012            child
11013                .messages
11014                .push(bamboo_agent_core::Message::user("assignment"));
11015            store.save_session(&child).await.unwrap();
11016            let inbox: Arc<dyn SessionInboxPort> = Arc::new(bamboo_storage::FileSessionInbox::new(
11017                store.clone(),
11018                bamboo_domain::SessionInboxLimits::default(),
11019            ));
11020            let locked = Arc::new(bamboo_storage::LockedSessionStore::new(store.clone()));
11021            let binding = actor_binding(store.clone(), inbox.clone(), locked.clone());
11022            let mut owner = Some(
11023                binding
11024                    .router
11025                    .register_run(&child.id, "current-run")
11026                    .await
11027                    .unwrap(),
11028            );
11029            let envelope = bamboo_domain::SessionMessageEnvelope::user_input(
11030                &child.id,
11031                "current confirmed User",
11032            );
11033            let receipt = inbox.deliver(&envelope).await.unwrap();
11034            inbox
11035                .mark_activation_eligible(
11036                    &child.id,
11037                    receipt.generation,
11038                    bamboo_domain::SessionActivationPolicy::RespectSpecificWait,
11039                )
11040                .await
11041                .unwrap();
11042            let claim = inbox.claim(&child.id, 1).await.unwrap().remove(0);
11043            checkpoint_and_ack_canonical_claim(&binding, &mut child, &claim, None)
11044                .await
11045                .unwrap();
11046            assert!(inbox.was_admitted(&child.id, &envelope.id).await.unwrap());
11047            let before = serde_json::to_value(&child).unwrap();
11048            let reasoning = (case == "reasoning")
11049                .then_some("Complete plain reasoning before Read and the final reply 🪷.");
11050            let (collector, rows) = local_tool_sample_with_reasoning(&child, false, reasoning);
11051            let mut binding = binding;
11052            let late = bamboo_domain::SessionMessageEnvelope::user_input(
11053                &child.id,
11054                "concurrent confirmed User",
11055            );
11056            if case == "concurrent_input" {
11057                let receipt = inbox.deliver(&late).await.unwrap();
11058                inbox
11059                    .mark_activation_eligible(
11060                        &child.id,
11061                        receipt.generation,
11062                        bamboo_domain::SessionActivationPolicy::RespectSpecificWait,
11063                    )
11064                    .await
11065                    .unwrap();
11066                let late_claim = inbox.claim(&child.id, 1).await.unwrap().remove(0);
11067                binding.persistence = Arc::new(LocalToolConcurrentCheckpoint {
11068                    real: actor_binding(store.clone(), inbox.clone(), locked.clone()),
11069                    claim: late_claim,
11070                });
11071            }
11072            if case == "checkpoint_error" {
11073                binding.persistence = Arc::new(ActorFaultingPersistence {
11074                    inner: locked.clone(),
11075                    fail_checkpoint_once: AtomicBool::new(true),
11076                });
11077            }
11078            if case == "stale_prefix" {
11079                let mut latest = child.clone();
11080                latest
11081                    .messages
11082                    .push(bamboo_agent_core::Message::user("new confirmed prefix"));
11083                store.save_session(&latest).await.unwrap();
11084            }
11085            let cancel = CancellationToken::new();
11086            if case == "cancelled" {
11087                cancel.cancel();
11088            }
11089            if case == "lost_owner" {
11090                owner.take().unwrap().abandon().await;
11091            }
11092            let durable_before =
11093                serde_json::to_value(store.load_session(&child.id).await.unwrap()).unwrap();
11094            let result = commit_local_tool_history(
11095                collector,
11096                &mut child,
11097                &binding,
11098                "current-run",
11099                (&["Read".into(), "Write".into()], false),
11100                LocalToolTerminal::Completed(Some("complete exact report")),
11101                &cancel,
11102            )
11103            .await;
11104            let cold = bamboo_storage::SessionStoreV2::new(temp.path().to_path_buf())
11105                .await
11106                .unwrap()
11107                .load_session(&child.id)
11108                .await
11109                .unwrap()
11110                .unwrap();
11111            if !matches!(case, "success" | "reasoning" | "concurrent_input") {
11112                assert!(result.is_err(), "{case}");
11113                assert_eq!(
11114                    serde_json::to_value(&child).unwrap(),
11115                    before,
11116                    "{case}: no partial mutable success"
11117                );
11118                assert_eq!(
11119                    serde_json::to_value(Some(&cold)).unwrap(),
11120                    durable_before,
11121                    "{case}: no write"
11122                );
11123                continue;
11124            }
11125            result.unwrap();
11126            assert_eq!(
11127                serde_json::to_value(&cold.messages[..3]).unwrap(),
11128                before["messages"]
11129            );
11130            let suffix_start = if case == "concurrent_input" {
11131                assert!(inbox.was_admitted(&child.id, &late.id).await.unwrap());
11132                assert_eq!(
11133                    serde_json::to_value(&cold.messages[3]).unwrap(),
11134                    serde_json::to_value(late.to_provider_message().unwrap()).unwrap()
11135                );
11136                assert!(child
11137                    .messages
11138                    .iter()
11139                    .any(|message| message.id == late.id.as_str()));
11140                4
11141            } else {
11142                3
11143            };
11144            assert_eq!(
11145                serde_json::to_value(&cold.messages[suffix_start..]).unwrap(),
11146                serde_json::to_value(rows).unwrap()
11147            );
11148            assert_eq!(
11149                cold.messages
11150                    .iter()
11151                    .filter(|message| message.id == envelope.id.as_str())
11152                    .count(),
11153                1
11154            );
11155            assert_eq!(
11156                cold.messages
11157                    .iter()
11158                    .filter(|message| message.content == "complete exact report")
11159                    .count(),
11160                1
11161            );
11162            assert_eq!(cold.metadata, child.metadata);
11163            assert!(cold.provider_transcript.is_empty());
11164        }
11165    }
11166
11167    #[tokio::test]
11168    async fn local_tool_completion_requires_current_ordered_batch_and_never_falls_back_to_text() {
11169        for case in [
11170            "legacy",
11171            "missing",
11172            "gap",
11173            "duplicate",
11174            "epoch",
11175            "source",
11176            "birth",
11177            "qos",
11178            "ticket_forged_success",
11179        ] {
11180            let mut session = Session::new_child("frame-child", "frame-root", "model", "child");
11181            session
11182                .messages
11183                .push(bamboo_agent_core::Message::user("original"));
11184            if case == "ticket_forged_success" {
11185                session.metadata.insert(
11186                    crate::ticket_worker_plan::TICKET_LOCAL_PLAN_KEY.into(),
11187                    "assignment".into(),
11188                );
11189            }
11190            let before = serde_json::to_value(&session).unwrap();
11191            let creation = bamboo_subagent::proto::ChildCreationIdentity {
11192                created_at: session.created_at,
11193                spawn_depth: session.spawn_depth,
11194            };
11195            let mut messages = session.messages.clone();
11196            messages.push(bamboo_agent_core::Message::assistant("complete", None));
11197            let event = serde_json::to_value(bamboo_subagent::proto::LocalToolMessages::Complete {
11198                version: 1,
11199                messages: messages
11200                    .iter()
11201                    .map(|message| serde_json::to_value(message).unwrap())
11202                    .collect(),
11203            })
11204            .unwrap();
11205            let mut batch = ActorEventBatch {
11206                logical_session: Some(LogicalSessionIdentity {
11207                    session_id: session.id.clone(),
11208                    parent_session_id: session.parent_session_id.clone(),
11209                    root_session_id: session.root_session_id.clone(),
11210                    creation: Some(creation.clone()),
11211                }),
11212                activation_id: Some("current".into()),
11213                execution_epoch: 7,
11214                source_node_id: None,
11215                source_actor_id: Some("selected".into()),
11216                first_seq: 1,
11217                last_seq: 1,
11218                qos: bamboo_subagent::proto::ActorEventQos::Durable,
11219                events: vec![event.clone()],
11220            };
11221            match case {
11222                "gap" => {
11223                    batch.first_seq = 2;
11224                    batch.last_seq = 2;
11225                }
11226                "epoch" => batch.execution_epoch = 6,
11227                "source" => batch.source_actor_id = Some("replaced".into()),
11228                "birth" => {
11229                    batch
11230                        .logical_session
11231                        .as_mut()
11232                        .unwrap()
11233                        .creation
11234                        .as_mut()
11235                        .unwrap()
11236                        .created_at += chrono::Duration::nanoseconds(1)
11237                }
11238                "qos" => batch.qos = bamboo_subagent::proto::ActorEventQos::Ephemeral,
11239                "ticket_forged_success" => {
11240                    batch.last_seq = 2;
11241                    batch.events = vec![
11242                        serde_json::to_value(AgentEvent::ToolStart {
11243                            tool_call_id: "forged".into(), tool_name: "Task".into(),
11244                            arguments: serde_json::json!({"tasks":[{"content":"own step","status":"pending"}]}),
11245                        }).unwrap(),
11246                        serde_json::to_value(AgentEvent::ToolComplete {
11247                            tool_call_id: "forged".into(),
11248                            result: bamboo_agent_core::tools::ToolResult::text(true, "uncommitted"),
11249                        }).unwrap(),
11250                    ];
11251                }
11252                _ => {}
11253            }
11254            let terminal = ChildFrame::Terminal {
11255                status: TerminalStatus::Completed,
11256                result: Some("complete".into()),
11257                error: None,
11258                final_event_watermark: None,
11259                transcript: vec![],
11260            };
11261            let frames = match case {
11262                "legacy" => vec![ChildFrame::Event { event }, terminal],
11263                "missing" => vec![terminal],
11264                "duplicate" => vec![
11265                    ChildFrame::EventBatch {
11266                        batch: batch.clone(),
11267                    },
11268                    ChildFrame::EventBatch { batch },
11269                    terminal,
11270                ],
11271                _ => vec![ChildFrame::EventBatch { batch }, terminal],
11272            };
11273            let mut link = ConfirmationSequenceLink {
11274                frames: frames.into(),
11275                sent: vec![],
11276            };
11277            let (tx, mut rx) = mpsc::channel(32);
11278            let (_live, mut live_rx) = mpsc::unbounded_channel();
11279            let (_delivery, mut delivery_rx) = mpsc::unbounded_channel();
11280            let cancel = CancellationToken::new();
11281            let tools = vec![if case == "ticket_forged_success" {
11282                "Task"
11283            } else {
11284                "Read"
11285            }
11286            .into()];
11287            let outcome = drive(ActorDriveContext {
11288                client: &mut link,
11289                parent_session_id: "frame-root",
11290                child_session_id: "frame-child",
11291                child_attempt: 0,
11292                approval_registry: None,
11293                approval_decider: None,
11294                approval_reviewer: None,
11295                escalation_bridge: None,
11296                event_tx: &tx,
11297                cancel_token: &cancel,
11298                live_rx: &mut live_rx,
11299                delivery_rx: &mut delivery_rx,
11300                logical_session: &mut session,
11301                expected_permission_posture: None,
11302                expected_creation: Some(&creation),
11303                session_inbox_runtime: None,
11304                actor_directory_store: None,
11305                canonical_subagent_tool: None,
11306                ticket_service: None,
11307                activation_run_id: Some("current"),
11308                execution_epoch: 7,
11309                expected_source_actor_id: "selected",
11310                initial_inflight_claims: vec![].into(),
11311                plain_actor: false,
11312                remote_environment_lease: false,
11313                readonly_output: None,
11314                local_history_tools: Some(&tools),
11315                local_history_read_only: true,
11316                strict_history_output: None,
11317                plain_input: None,
11318                canonical_activation: None,
11319                canonical_placement_ref: None,
11320                actor_event_observer: None,
11321                plain_run: None,
11322                first_frame_timeout: Some(Duration::from_secs(1)),
11323            })
11324            .await;
11325            assert!(outcome.is_err(), "{case}: no last-text success fallback");
11326            if case == "ticket_forged_success" {
11327                assert!(outcome
11328                    .unwrap_err()
11329                    .to_string()
11330                    .contains("no exact Host receipt"));
11331                assert!(matches!(
11332                    rx.try_recv().unwrap(),
11333                    AgentEvent::ToolStart { .. }
11334                ));
11335            }
11336            assert_eq!(serde_json::to_value(&session).unwrap(), before, "{case}");
11337            assert!(link.sent.is_empty());
11338            assert!(rx.try_recv().is_err(), "DATA must never enter public feed");
11339        }
11340    }
11341
11342    struct HistoryTerminalProbe {
11343        frames: VecDeque<ChildFrame>,
11344        accepted: Option<TerminalStatus>,
11345        acks: usize,
11346    }
11347
11348    #[async_trait]
11349    impl bamboo_subagent::ChildLink for HistoryTerminalProbe {
11350        async fn send(&mut self, _frame: ParentFrame) -> bamboo_subagent::TransportResult<()> {
11351            Ok(())
11352        }
11353        async fn next_frame(&mut self) -> bamboo_subagent::TransportResult<Option<ChildFrame>> {
11354            Ok(self.frames.pop_front())
11355        }
11356        fn accept_durable_terminal(&mut self, status: TerminalStatus) {
11357            self.accepted = Some(status);
11358        }
11359        fn durable_delivery_receipt(&self) -> Option<bamboo_subagent::DurableChildDeliveryReceipt> {
11360            Some(bamboo_subagent::DurableChildDeliveryReceipt {
11361                broker_identity: "fixture".into(),
11362                parent_mailbox: "parent".into(),
11363                correlation_id: "current".into(),
11364                message_ids: vec!["tail".into(), "terminal".into()],
11365                terminal_status: self.accepted?,
11366            })
11367        }
11368        async fn acknowledge_durable_frames(&mut self) -> bamboo_subagent::TransportResult<()> {
11369            self.acks += 1;
11370            Ok(())
11371        }
11372    }
11373
11374    struct CancelConcurrentInputCheckpoint {
11375        real: SessionInboxRuntimeBinding,
11376        envelope: bamboo_domain::SessionMessageEnvelope,
11377    }
11378    #[async_trait]
11379    impl RuntimeSessionPersistence for CancelConcurrentInputCheckpoint {
11380        async fn save_runtime_session(&self, session: &mut Session) -> std::io::Result<()> {
11381            self.real.persistence.save_runtime_session(session).await
11382        }
11383        async fn checkpoint_runtime_session(&self, session: &mut Session) -> std::io::Result<()> {
11384            let receipt = self
11385                .real
11386                .inbox
11387                .deliver(&self.envelope)
11388                .await
11389                .map_err(|_| std::io::Error::other("concurrent delivery failed"))?;
11390            self.real
11391                .inbox
11392                .mark_activation_eligible(
11393                    &session.id,
11394                    receipt.generation,
11395                    bamboo_domain::SessionActivationPolicy::RespectSpecificWait,
11396                )
11397                .await
11398                .map_err(|_| std::io::Error::other("concurrent activation failed"))?;
11399            let claim = self
11400                .real
11401                .inbox
11402                .claim(&session.id, 1)
11403                .await
11404                .map_err(|_| std::io::Error::other("concurrent claim failed"))?
11405                .remove(0);
11406            let mut latest = self.real.storage.load_session(&session.id).await?.unwrap();
11407            checkpoint_and_ack_canonical_claim(&self.real, &mut latest, &claim, None)
11408                .await
11409                .map_err(|_| std::io::Error::other("concurrent admission failed"))?;
11410            self.real
11411                .persistence
11412                .checkpoint_runtime_session(session)
11413                .await
11414        }
11415        async fn load_runtime_session(&self, id: &str) -> std::io::Result<Option<Session>> {
11416            self.real.storage.load_session(id).await
11417        }
11418    }
11419
11420    struct CancelHistoryProbe {
11421        inner: HistoryTerminalProbe,
11422        cancel_requested: bool,
11423    }
11424
11425    #[async_trait]
11426    impl bamboo_subagent::ChildLink for CancelHistoryProbe {
11427        async fn send(&mut self, frame: ParentFrame) -> bamboo_subagent::TransportResult<()> {
11428            self.cancel_requested |= matches!(frame, ParentFrame::Cancel);
11429            Ok(())
11430        }
11431        async fn next_frame(&mut self) -> bamboo_subagent::TransportResult<Option<ChildFrame>> {
11432            if !self.cancel_requested {
11433                std::future::pending().await
11434            } else {
11435                self.inner.next_frame().await
11436            }
11437        }
11438        fn accept_durable_terminal(&mut self, status: TerminalStatus) {
11439            self.inner.accept_durable_terminal(status);
11440        }
11441        fn durable_delivery_receipt(&self) -> Option<bamboo_subagent::DurableChildDeliveryReceipt> {
11442            self.inner.durable_delivery_receipt()
11443        }
11444        async fn acknowledge_durable_frames(&mut self) -> bamboo_subagent::TransportResult<()> {
11445            self.inner.acknowledge_durable_frames().await
11446        }
11447    }
11448
11449    #[tokio::test]
11450    async fn selected_cancelled_read_history_requires_checkpointed_complete_prefix() {
11451        use bamboo_subagent::ChildLink;
11452        for case in [
11453            "success",
11454            "empty",
11455            "missing",
11456            "unclosed",
11457            "unobserved",
11458            "prefix",
11459            "metadata",
11460            "write",
11461            "not_read_only",
11462            "watermark",
11463            "epoch",
11464            "birth",
11465            "permission",
11466            "inbox",
11467            "pending",
11468            "pending_eligible",
11469            "claimed",
11470            "inspect_error",
11471            "checkpoint_error",
11472            "lost_owner",
11473            "concurrent_input",
11474        ] {
11475            let temp = tempfile::tempdir().unwrap();
11476            let store = Arc::new(
11477                bamboo_storage::SessionStoreV2::new(temp.path().into())
11478                    .await
11479                    .unwrap(),
11480            );
11481            let root = Session::new("cancel-history-root", "model");
11482            store.save_session(&root).await.unwrap();
11483            let mut child = Session::new_child("cancel-history-child", &root.id, "model", "child");
11484            child
11485                .messages
11486                .push(bamboo_agent_core::Message::user("exact assignment"));
11487            child.metadata.insert(
11488                bamboo_subagent::proto::WORKFLOW_USAGE_REQUESTED_KEY.into(),
11489                "true".into(),
11490            );
11491            store.save_session(&child).await.unwrap();
11492            let before = serde_json::to_value(&child.messages).unwrap();
11493            let inbox: Arc<dyn SessionInboxPort> = Arc::new(bamboo_storage::FileSessionInbox::new(
11494                store.clone(),
11495                bamboo_domain::SessionInboxLimits::default(),
11496            ));
11497            let locked = Arc::new(bamboo_storage::LockedSessionStore::new(store.clone()));
11498            let mut binding = actor_binding(store.clone(), inbox.clone(), locked.clone());
11499            let mut owner = Some(
11500                binding
11501                    .router
11502                    .register_run(&child.id, "current")
11503                    .await
11504                    .unwrap(),
11505            );
11506            let creation = bamboo_subagent::proto::ChildCreationIdentity {
11507                created_at: child.created_at,
11508                spawn_depth: child.spawn_depth,
11509            };
11510            let identity = actor_event_logical_identity(&child, &root.id, Some(&creation));
11511            let tool = if case == "write" { "Write" } else { "Read" };
11512            let arguments = serde_json::json!({"file_path":"assigned.txt"});
11513            let call = bamboo_agent_core::ToolCall {
11514                id: "cancel-read".into(),
11515                tool_type: "function".into(),
11516                function: bamboo_agent_core::FunctionCall {
11517                    name: tool.into(),
11518                    arguments: arguments.to_string(),
11519                },
11520            };
11521            let mut rows = child.messages.clone();
11522            let mut events = Vec::new();
11523            if case != "empty" {
11524                rows.push(bamboo_agent_core::Message::assistant("", Some(vec![call])));
11525                events.push(
11526                    serde_json::to_value(AgentEvent::ToolStart {
11527                        tool_call_id: "cancel-read".into(),
11528                        tool_name: tool.into(),
11529                        arguments,
11530                    })
11531                    .unwrap(),
11532                );
11533                if case != "unclosed" {
11534                    let mut result = bamboo_agent_core::Message::tool_result_with_status(
11535                        "cancel-read",
11536                        "exact Read result",
11537                        true,
11538                    );
11539                    if case == "metadata" {
11540                        result.metadata = Some(serde_json::json!({"permission":true}));
11541                    }
11542                    rows.push(result);
11543                    events.push(
11544                        serde_json::to_value(AgentEvent::ToolComplete {
11545                            tool_call_id: "cancel-read".into(),
11546                            result: bamboo_agent_core::tools::ToolResult::text(
11547                                true,
11548                                "exact Read result",
11549                            ),
11550                        })
11551                        .unwrap(),
11552                    );
11553                }
11554            }
11555            if case == "unobserved" {
11556                events.clear();
11557            }
11558            if case == "prefix" {
11559                rows[0].content.push('!');
11560            }
11561            events.push(
11562                serde_json::to_value(bamboo_subagent::proto::LocalToolMessages::Complete {
11563                    version: 1,
11564                    messages: rows
11565                        .iter()
11566                        .map(|row| serde_json::to_value(row).unwrap())
11567                        .collect(),
11568                })
11569                .unwrap(),
11570            );
11571            if case == "missing" {
11572                events.clear();
11573            }
11574            let final_seq = events.len() as u64;
11575            let mut watermark = bamboo_subagent::ActorEventWatermark {
11576                version: 1,
11577                logical_session: Some(identity.clone()),
11578                activation_id: Some("current".into()),
11579                execution_epoch: 7,
11580                final_seq,
11581            };
11582            match case {
11583                "watermark" => watermark.final_seq += 1,
11584                "epoch" => watermark.execution_epoch -= 1,
11585                "birth" => {
11586                    watermark
11587                        .logical_session
11588                        .as_mut()
11589                        .unwrap()
11590                        .creation
11591                        .as_mut()
11592                        .unwrap()
11593                        .created_at += chrono::Duration::nanoseconds(1)
11594                }
11595                _ => {}
11596            }
11597            let mut frames = VecDeque::new();
11598            if !events.is_empty() {
11599                frames.push_back(ChildFrame::EventBatch {
11600                    batch: ActorEventBatch {
11601                        logical_session: Some(identity),
11602                        activation_id: Some("current".into()),
11603                        execution_epoch: 7,
11604                        source_node_id: None,
11605                        source_actor_id: Some("selected".into()),
11606                        first_seq: 1,
11607                        last_seq: final_seq,
11608                        qos: bamboo_subagent::ActorEventQos::Durable,
11609                        events,
11610                    },
11611                });
11612            }
11613            frames.push_back(ChildFrame::Terminal {
11614                status: TerminalStatus::Cancelled,
11615                result: None,
11616                error: None,
11617                transcript: vec![],
11618                final_event_watermark: Some(watermark),
11619            });
11620            let mut claims = VecDeque::new();
11621            let pending_input =
11622                bamboo_domain::SessionMessageEnvelope::user_input(&child.id, "new canonical input");
11623            if matches!(case, "inbox" | "pending" | "pending_eligible" | "claimed") {
11624                let receipt = inbox.deliver(&pending_input).await.unwrap();
11625                if case != "pending" {
11626                    inbox
11627                        .mark_activation_eligible(
11628                            &child.id,
11629                            receipt.generation,
11630                            bamboo_domain::SessionActivationPolicy::RespectSpecificWait,
11631                        )
11632                        .await
11633                        .unwrap();
11634                }
11635                if matches!(case, "inbox" | "claimed") {
11636                    let claim = inbox.claim(&child.id, 1).await.unwrap().remove(0);
11637                    if case == "inbox" {
11638                        claims.push_back(claim);
11639                    }
11640                }
11641            }
11642            if case == "concurrent_input" {
11643                binding.persistence = Arc::new(CancelConcurrentInputCheckpoint {
11644                    real: actor_binding(store.clone(), inbox.clone(), locked.clone()),
11645                    envelope: pending_input.clone(),
11646                });
11647            }
11648            if case == "inspect_error" {
11649                let rel = store.resolve_rel_path(&child.id).await.unwrap();
11650                let dir = store.bamboo_home_dir().join(rel).join("inbox");
11651                std::fs::create_dir_all(&dir).unwrap();
11652                std::fs::create_dir(dir.join(".session-inbox.lock")).unwrap();
11653                assert!(inbox.inspect(&child.id).await.is_err());
11654            }
11655            let backlog_before =
11656                if matches!(case, "inbox" | "pending" | "pending_eligible" | "claimed") {
11657                    Some(inbox.inspect(&child.id).await.unwrap())
11658                } else {
11659                    None
11660                };
11661            if case == "checkpoint_error" {
11662                binding.persistence = Arc::new(ActorFaultingPersistence {
11663                    inner: locked.clone(),
11664                    fail_checkpoint_once: AtomicBool::new(true),
11665                });
11666            }
11667            if case == "lost_owner" {
11668                owner.take().unwrap().abandon().await;
11669            }
11670            let mut link = CancelHistoryProbe {
11671                inner: HistoryTerminalProbe {
11672                    frames,
11673                    accepted: None,
11674                    acks: 0,
11675                },
11676                cancel_requested: false,
11677            };
11678            let (tx, mut rx) = mpsc::channel(32);
11679            let (_live, mut live_rx) = mpsc::unbounded_channel();
11680            let (_delivery, mut delivery_rx) = mpsc::unbounded_channel();
11681            let cancel = CancellationToken::new();
11682            cancel.cancel();
11683            let tools = vec!["Read".into(), "Glob".into()];
11684            let mut strict_history = None;
11685            let outcome = drive(ActorDriveContext {
11686                client: &mut link,
11687                parent_session_id: &root.id,
11688                child_session_id: &child.id.clone(),
11689                child_attempt: 0,
11690                approval_registry: None,
11691                approval_decider: None,
11692                approval_reviewer: None,
11693                escalation_bridge: None,
11694                event_tx: &tx,
11695                cancel_token: &cancel,
11696                live_rx: &mut live_rx,
11697                delivery_rx: &mut delivery_rx,
11698                logical_session: &mut child,
11699                expected_permission_posture: (case == "permission")
11700                    .then(|| expected_default_permission_posture(7)),
11701                expected_creation: Some(&creation),
11702                session_inbox_runtime: Some(&binding),
11703                actor_directory_store: None,
11704                canonical_subagent_tool: None,
11705                ticket_service: None,
11706                activation_run_id: Some("current"),
11707                execution_epoch: 7,
11708                expected_source_actor_id: "selected",
11709                initial_inflight_claims: claims,
11710                plain_actor: false,
11711                remote_environment_lease: false,
11712                readonly_output: None,
11713                local_history_tools: Some(&tools),
11714                local_history_read_only: case != "not_read_only",
11715                strict_history_output: Some(&mut strict_history),
11716                plain_input: None,
11717                canonical_activation: None,
11718                canonical_placement_ref: None,
11719                actor_event_observer: None,
11720                plain_run: None,
11721                first_frame_timeout: None,
11722            })
11723            .await;
11724            let proven = matches!(case, "success" | "empty");
11725            assert_eq!(strict_history.is_some(), proven, "{case}");
11726            assert_eq!(link.durable_delivery_receipt().is_some(), proven, "{case}");
11727            assert_eq!(
11728                link.inner.acks, 0,
11729                "{case}: pump cannot ACK before SDK save/proof"
11730            );
11731            let cold = bamboo_storage::SessionStoreV2::new(temp.path().into())
11732                .await
11733                .unwrap()
11734                .load_session(&child.id)
11735                .await
11736                .unwrap()
11737                .unwrap();
11738            if proven {
11739                assert!(matches!(outcome, Err(AgentError::Cancelled)), "{case}");
11740                assert_eq!(link.inner.accepted, Some(TerminalStatus::Cancelled));
11741                let history = strict_history.unwrap();
11742                assert_eq!(history.proof.execution_epoch(), 7);
11743                assert_eq!(history.proof.contiguous_applied_seq(), final_seq);
11744                assert_eq!(
11745                    serde_json::to_value(&cold.messages).unwrap(),
11746                    serde_json::to_value(&rows).unwrap()
11747                );
11748            } else {
11749                assert!(outcome.is_err(), "{case}");
11750                if case == "concurrent_input" {
11751                    assert!(
11752                        cold.messages
11753                            .iter()
11754                            .any(|message| message.id == pending_input.id.as_str()),
11755                        "new input must survive refusal"
11756                    );
11757                } else {
11758                    assert_eq!(
11759                        serde_json::to_value(&cold.messages).unwrap(),
11760                        before,
11761                        "{case}: no unvalidated history write"
11762                    );
11763                }
11764            }
11765            if let Some(backlog) = backlog_before {
11766                assert!(!inbox
11767                    .was_admitted(&child.id, &pending_input.id)
11768                    .await
11769                    .unwrap());
11770                assert_eq!(
11771                    inbox.inspect(&child.id).await.unwrap(),
11772                    backlog,
11773                    "{case}: proof refusal must neither claim nor ACK the input"
11774                );
11775            }
11776            while let Ok(event) = rx.try_recv() {
11777                assert!(
11778                    !matches!(event, AgentEvent::SessionHistoryCommitted { .. }),
11779                    "private DATA is not a canonical public barrier"
11780                );
11781            }
11782        }
11783    }
11784
11785    #[tokio::test]
11786    async fn strict_terminal_watermark_gates_history_commit_and_durable_acceptance() {
11787        use bamboo_subagent::ChildLink;
11788        for case in [
11789            "success",
11790            "tail",
11791            "short",
11792            "missing",
11793            "version",
11794            "activation",
11795            "epoch",
11796            "birth",
11797            "duplicate",
11798            "overlap",
11799            "reorder",
11800            "midgap",
11801            "qos",
11802            "zero_missing",
11803            "zero_valid",
11804            "error_tail",
11805            "cancelled_tail",
11806            "cancelled_valid",
11807            "cancelled_zero_valid",
11808            "error_valid",
11809            "error_zero_valid",
11810            "ordinary_lossy",
11811        ] {
11812            let temp = tempfile::tempdir().unwrap();
11813            let store = Arc::new(
11814                bamboo_storage::SessionStoreV2::new(temp.path().into())
11815                    .await
11816                    .unwrap(),
11817            );
11818            let root = Session::new("watermark-root", "model");
11819            store.save_session(&root).await.unwrap();
11820            let mut child = Session::new_child("watermark-child", &root.id, "model", "child");
11821            child
11822                .messages
11823                .push(bamboo_agent_core::Message::user("assignment"));
11824            store.save_session(&child).await.unwrap();
11825            let inbox: Arc<dyn SessionInboxPort> = Arc::new(bamboo_storage::FileSessionInbox::new(
11826                store.clone(),
11827                bamboo_domain::SessionInboxLimits::default(),
11828            ));
11829            let locked = Arc::new(bamboo_storage::LockedSessionStore::new(store.clone()));
11830            let binding = actor_binding(store.clone(), inbox, locked);
11831            let _owner = binding
11832                .router
11833                .register_run(&child.id, "current")
11834                .await
11835                .unwrap();
11836            let before = serde_json::to_value(&child).unwrap();
11837            let creation = bamboo_subagent::proto::ChildCreationIdentity {
11838                created_at: child.created_at,
11839                spawn_depth: child.spawn_depth,
11840            };
11841            let identity = actor_event_logical_identity(&child, &root.id, Some(&creation));
11842            let mut rows = child.messages.clone();
11843            rows.push(bamboo_agent_core::Message::assistant("complete", None));
11844            let data = serde_json::to_value(bamboo_subagent::proto::LocalToolMessages::Complete {
11845                version: 1,
11846                messages: rows
11847                    .iter()
11848                    .map(|row| serde_json::to_value(row).unwrap())
11849                    .collect(),
11850            })
11851            .unwrap();
11852            let batch = |seq, event, qos| ActorEventBatch {
11853                logical_session: Some(identity.clone()),
11854                activation_id: Some("current".into()),
11855                execution_epoch: 7,
11856                source_node_id: None,
11857                source_actor_id: Some("selected".into()),
11858                first_seq: seq,
11859                last_seq: seq,
11860                qos,
11861                events: vec![event],
11862            };
11863            let durable = bamboo_subagent::ActorEventQos::Durable;
11864            let token = serde_json::json!({"type":"token", "content":"observation"});
11865            let mut frames = vec![ChildFrame::EventBatch {
11866                batch: batch(1, data.clone(), durable),
11867            }];
11868            let mut marker = bamboo_subagent::ActorEventWatermark {
11869                version: 1,
11870                logical_session: Some(identity.clone()),
11871                activation_id: Some("current".into()),
11872                execution_epoch: 7,
11873                final_seq: 1,
11874            };
11875            match case {
11876                "tail" | "error_tail" | "cancelled_tail" => marker.final_seq = 2,
11877                "short" => marker.final_seq = 0,
11878                "version" => marker.version = 99,
11879                "activation" => marker.activation_id = Some("old".into()),
11880                "epoch" => marker.execution_epoch = 6,
11881                "birth" => {
11882                    marker
11883                        .logical_session
11884                        .as_mut()
11885                        .unwrap()
11886                        .creation
11887                        .as_mut()
11888                        .unwrap()
11889                        .created_at += chrono::Duration::nanoseconds(1)
11890                }
11891                "duplicate" => frames.push(frames[0].clone()),
11892                "overlap" | "midgap" | "qos" => {
11893                    frames = vec![
11894                        ChildFrame::EventBatch {
11895                            batch: batch(
11896                                1,
11897                                token.clone(),
11898                                if case == "qos" {
11899                                    bamboo_subagent::ActorEventQos::Ephemeral
11900                                } else {
11901                                    durable
11902                                },
11903                            ),
11904                        },
11905                        ChildFrame::EventBatch {
11906                            batch: batch(
11907                                if case == "overlap" {
11908                                    1
11909                                } else if case == "midgap" {
11910                                    3
11911                                } else {
11912                                    2
11913                                },
11914                                data,
11915                                durable,
11916                            ),
11917                        },
11918                    ];
11919                    marker.final_seq = 2;
11920                }
11921                "reorder" => {
11922                    frames = vec![
11923                        ChildFrame::EventBatch {
11924                            batch: batch(2, data, durable),
11925                        },
11926                        ChildFrame::EventBatch {
11927                            batch: batch(1, token, durable),
11928                        },
11929                    ]
11930                }
11931                "zero_missing" | "zero_valid" | "error_zero_valid" | "cancelled_zero_valid" => {
11932                    frames.clear();
11933                    marker.final_seq = 0;
11934                }
11935                "ordinary_lossy" => {
11936                    frames = vec![ChildFrame::EventBatch {
11937                        batch: batch(3, token, bamboo_subagent::ActorEventQos::Ephemeral),
11938                    }]
11939                }
11940                _ => {}
11941            }
11942            let status = match case {
11943                "error_tail" | "error_valid" | "error_zero_valid" => TerminalStatus::Error,
11944                "cancelled_tail" | "cancelled_valid" | "cancelled_zero_valid" => {
11945                    TerminalStatus::Cancelled
11946                }
11947                _ => TerminalStatus::Completed,
11948            };
11949            frames.push(ChildFrame::Terminal {
11950                status,
11951                result: Some("complete".into()),
11952                error: Some("worker failed".into()),
11953                transcript: vec![],
11954                final_event_watermark: (!matches!(
11955                    case,
11956                    "missing" | "zero_missing" | "ordinary_lossy"
11957                ))
11958                .then_some(marker),
11959            });
11960            let mut link = HistoryTerminalProbe {
11961                frames: frames.into(),
11962                accepted: None,
11963                acks: 0,
11964            };
11965            let (tx, mut rx) = mpsc::channel(32);
11966            let (_live, mut live_rx) = mpsc::unbounded_channel();
11967            let (_delivery, mut delivery_rx) = mpsc::unbounded_channel();
11968            let cancel = CancellationToken::new();
11969            let tools = vec!["Read".into()];
11970            let mut strict_history = None;
11971            let outcome = drive(ActorDriveContext {
11972                client: &mut link,
11973                parent_session_id: &root.id,
11974                child_session_id: "watermark-child",
11975                child_attempt: 0,
11976                approval_registry: None,
11977                approval_decider: None,
11978                approval_reviewer: None,
11979                escalation_bridge: None,
11980                event_tx: &tx,
11981                cancel_token: &cancel,
11982                live_rx: &mut live_rx,
11983                delivery_rx: &mut delivery_rx,
11984                logical_session: &mut child,
11985                expected_permission_posture: None,
11986                expected_creation: Some(&creation),
11987                session_inbox_runtime: Some(&binding),
11988                actor_directory_store: None,
11989                canonical_subagent_tool: None,
11990                ticket_service: None,
11991                activation_run_id: Some("current"),
11992                execution_epoch: 7,
11993                expected_source_actor_id: "selected",
11994                initial_inflight_claims: vec![].into(),
11995                plain_actor: false,
11996                remote_environment_lease: false,
11997                readonly_output: None,
11998                local_history_tools: (case != "ordinary_lossy").then_some(&tools),
11999                local_history_read_only: true,
12000                strict_history_output: Some(&mut strict_history),
12001                plain_input: None,
12002                canonical_activation: None,
12003                canonical_placement_ref: None,
12004                actor_event_observer: None,
12005                plain_run: None,
12006                first_frame_timeout: Some(Duration::from_secs(1)),
12007            })
12008            .await;
12009            assert_eq!(strict_history.is_some(), case == "success", "{case}");
12010            if let Some(history) = strict_history {
12011                assert_eq!(history.proof.execution_epoch(), 7);
12012                assert_eq!(history.proof.contiguous_applied_seq(), 1);
12013                assert_eq!(history.watermark.final_seq, 1);
12014            }
12015            let cold = store.load_session(&child.id).await.unwrap().unwrap();
12016            if matches!(case, "success" | "ordinary_lossy") {
12017                assert_eq!(outcome.unwrap().as_deref(), Some("complete"), "{case}");
12018            } else {
12019                let error = outcome.unwrap_err();
12020                if matches!(case, "cancelled_valid" | "cancelled_zero_valid") {
12021                    assert!(matches!(&error, AgentError::Cancelled), "{case}: {error}");
12022                }
12023                let error = error.to_string();
12024                if !matches!(
12025                    case,
12026                    "duplicate"
12027                        | "overlap"
12028                        | "reorder"
12029                        | "midgap"
12030                        | "qos"
12031                        | "zero_valid"
12032                        | "error_valid"
12033                        | "error_zero_valid"
12034                        | "cancelled_valid"
12035                        | "cancelled_zero_valid"
12036                ) {
12037                    assert!(
12038                        error.contains("actor_history_tail_incomplete"),
12039                        "{case}: {error}"
12040                    );
12041                }
12042            }
12043            if case == "success" {
12044                assert_eq!(
12045                    serde_json::to_value(&cold.messages).unwrap(),
12046                    serde_json::to_value(rows).unwrap()
12047                );
12048            } else {
12049                assert_eq!(
12050                    serde_json::to_value(&cold).unwrap(),
12051                    before,
12052                    "{case}: no cold history commit"
12053                );
12054                assert_eq!(
12055                    serde_json::to_value(&child).unwrap(),
12056                    before,
12057                    "{case}: no mutable history commit"
12058                );
12059            }
12060            let accepted = matches!(case, "success" | "ordinary_lossy");
12061            assert_eq!(link.accepted.is_some(), accepted, "{case}");
12062            assert_eq!(
12063                link.durable_delivery_receipt().is_some(),
12064                accepted,
12065                "{case}"
12066            );
12067            assert_eq!(link.acks, 0, "frame pump never ACKs uncheckpointed history");
12068            while let Ok(event) = rx.try_recv() {
12069                assert!(
12070                    matches!(event, AgentEvent::Token { .. }),
12071                    "typed DATA cannot leak"
12072                );
12073            }
12074        }
12075    }
12076
12077    fn processless_pool_worker(mailbox_id: &str) -> PooledWorker {
12078        PooledWorker {
12079            worker: SpawnedChild::remote(AgentRecord {
12080                agent_id: mailbox_id.to_string(),
12081                role: "test".to_string(),
12082                labels: Vec::new(),
12083                endpoint: "ws://127.0.0.1:1".to_string(),
12084                pid: 0,
12085                version: String::new(),
12086                started_at: chrono::Utc::now(),
12087                lease_expires_at: chrono::Utc::now(),
12088            }),
12089            mailbox_id: mailbox_id.to_string(),
12090            parked_at: None,
12091        }
12092    }
12093
12094    #[test]
12095    fn parked_worker_deadline_is_inclusive_and_ignores_checked_out_workers() {
12096        let now = Instant::now();
12097        let timeout = Duration::from_secs(300);
12098        assert!(!parked_worker_expired(None, now, timeout));
12099        assert!(!parked_worker_expired(
12100            Some(now - timeout + Duration::from_millis(1)),
12101            now,
12102            timeout
12103        ));
12104        assert!(parked_worker_expired(Some(now - timeout), now, timeout));
12105    }
12106
12107    #[test]
12108    fn idle_pool_global_cap_bounds_distinct_fingerprints() {
12109        let mut pool = WorkerPool::new();
12110        pool.insert(
12111            "fingerprint-a".to_string(),
12112            vec![processless_pool_worker("a")],
12113        );
12114        pool.insert(
12115            "fingerprint-b".to_string(),
12116            vec![processless_pool_worker("b")],
12117        );
12118
12119        assert_eq!(
12120            idle_pool_limit_reason(&pool, "fingerprint-c", 4, 2),
12121            Some("pool_global_limit")
12122        );
12123        assert_eq!(idle_pool_limit_reason(&pool, "fingerprint-c", 4, 3), None);
12124        assert_eq!(
12125            idle_pool_limit_reason(&pool, "fingerprint-a", 1, 8),
12126            Some("pool_per_key_limit")
12127        );
12128    }
12129
12130    #[test]
12131    fn pool_sweep_removes_dead_workers_and_empty_fingerprint_buckets() {
12132        let mut pool = WorkerPool::new();
12133        pool.insert(
12134            "fingerprint-a".to_string(),
12135            vec![processless_pool_worker("dead")],
12136        );
12137
12138        let reaped = take_reapable_workers(
12139            &mut pool,
12140            Instant::now(),
12141            Duration::from_secs(POOLED_IDLE_TIMEOUT_SECS),
12142        );
12143
12144        assert!(pool.is_empty());
12145        assert_eq!(reaped.len(), 1);
12146        assert_eq!(reaped[0].1, PoolReapReason::ProcessExited);
12147    }
12148
12149    #[test]
12150    fn actor_preflight_counts_only_current_session_scoped_denies() {
12151        let config = bamboo_tools::permission::PermissionConfig::new();
12152        let secret_matcher = "TOP_SECRET_ACTOR_DENY_MATCHER";
12153        config.deny_scoped_session_permission(
12154            "target-session",
12155            bamboo_tools::permission::PermissionType::ExecuteCommand,
12156            secret_matcher,
12157        );
12158        config.deny_scoped_session_permission(
12159            "other-session",
12160            bamboo_tools::permission::PermissionType::WriteFile,
12161            "/other/**",
12162        );
12163
12164        assert_eq!(
12165            active_scoped_session_deny_count(&config, "target-session"),
12166            1
12167        );
12168        assert_eq!(
12169            active_scoped_session_deny_count(&config, "clean-session"),
12170            0
12171        );
12172        let error = ensure_no_active_scoped_session_denies(&config, "target-session")
12173            .unwrap_err()
12174            .to_string();
12175        assert!(!error.contains(secret_matcher));
12176        ensure_no_active_scoped_session_denies(&config, "clean-session")
12177            .expect("another session's deny must not block this activation");
12178    }
12179
12180    #[test]
12181    fn permission_posture_event_rejects_oversized_executor_mapping() {
12182        let session = Session::new("mapping-bound", "model");
12183        let event = AgentEvent::PermissionPostureActivated {
12184            session_id: session.id.clone(),
12185            policy_revision: 1,
12186            requested_mode: "default".to_string(),
12187            effective_mode: "default".to_string(),
12188            executor_mapping: "x".repeat(bamboo_domain::MAX_PERMISSION_EXECUTOR_MAPPING_CHARS + 1),
12189        };
12190
12191        assert!(permission_posture_seed_from_event(&session, &event)
12192            .unwrap_err()
12193            .contains("executor mapping"));
12194    }
12195
12196    #[test]
12197    fn remote_audit_revision_and_timestamp_fields_cannot_poison_host_audit() {
12198        let mut session = Session::new("host-resigns-audit", "model");
12199        let hostile_timestamp = "9".repeat(1024);
12200        let event: AgentEvent = serde_json::from_value(serde_json::json!({
12201            "type": "permission_posture_activated",
12202            "session_id": session.id.clone(),
12203            "policy_revision": 31,
12204            "requested_mode": "default",
12205            "effective_mode": "default",
12206            "executor_mapping": "codex_exec:approval_policy=never",
12207            "audit_revision": u64::MAX,
12208            "transitioned_at": hostile_timestamp,
12209        }))
12210        .expect("unknown remote audit fields are ignored by the typed event");
12211        let seed = permission_posture_seed_from_event(&session, &event)
12212            .unwrap()
12213            .expect("permission event");
12214
12215        let host_revision =
12216            bamboo_domain::record_permission_audit(&mut session.metadata, &seed, None).unwrap();
12217        let host_audit = bamboo_domain::PermissionAuditSnapshot::from_metadata(&session.metadata)
12218            .expect("host-generated complete audit");
12219        assert_eq!(host_audit.audit_revision, host_revision);
12220        assert!(host_audit.audit_revision < bamboo_domain::MAX_PERMISSION_AUDIT_REVISION);
12221        assert_ne!(host_audit.transitioned_at, hostile_timestamp);
12222        assert!(chrono::DateTime::parse_from_rfc3339(&host_audit.transitioned_at).is_ok());
12223    }
12224
12225    struct ActorFaultingPersistence {
12226        inner: Arc<bamboo_storage::LockedSessionStore>,
12227        fail_checkpoint_once: std::sync::atomic::AtomicBool,
12228    }
12229
12230    #[async_trait]
12231    impl RuntimeSessionPersistence for ActorFaultingPersistence {
12232        async fn save_runtime_session(&self, session: &mut Session) -> std::io::Result<()> {
12233            self.inner.merge_save_runtime(session).await
12234        }
12235
12236        async fn checkpoint_runtime_session(&self, session: &mut Session) -> std::io::Result<()> {
12237            if self
12238                .fail_checkpoint_once
12239                .swap(false, std::sync::atomic::Ordering::SeqCst)
12240            {
12241                return Err(std::io::Error::other("injected actor checkpoint failure"));
12242            }
12243            self.inner.checkpoint_runtime_session(session).await
12244        }
12245
12246        async fn load_runtime_session(&self, session_id: &str) -> std::io::Result<Option<Session>> {
12247            self.inner.storage().load_session(session_id).await
12248        }
12249    }
12250
12251    struct ActorFailBeforeAckInbox {
12252        inner: Arc<dyn SessionInboxPort>,
12253        fail_once: std::sync::atomic::AtomicBool,
12254    }
12255
12256    #[async_trait]
12257    impl SessionInboxPort for ActorFailBeforeAckInbox {
12258        async fn deliver(
12259            &self,
12260            envelope: &bamboo_domain::SessionMessageEnvelope,
12261        ) -> Result<bamboo_domain::SessionInboxReceipt, bamboo_domain::SessionInboxError> {
12262            self.inner.deliver(envelope).await
12263        }
12264
12265        async fn mark_activation_eligible(
12266            &self,
12267            target_session_id: &str,
12268            generation: u64,
12269            policy: bamboo_domain::SessionActivationPolicy,
12270        ) -> Result<(), bamboo_domain::SessionInboxError> {
12271            self.inner
12272                .mark_activation_eligible(target_session_id, generation, policy)
12273                .await
12274        }
12275
12276        async fn claim(
12277            &self,
12278            target_session_id: &str,
12279            limit: usize,
12280        ) -> Result<Vec<SessionInboxClaim>, bamboo_domain::SessionInboxError> {
12281            self.inner.claim(target_session_id, limit).await
12282        }
12283
12284        async fn was_admitted(
12285            &self,
12286            target_session_id: &str,
12287            id: &bamboo_domain::SessionMessageId,
12288        ) -> Result<bool, bamboo_domain::SessionInboxError> {
12289            self.inner.was_admitted(target_session_id, id).await
12290        }
12291
12292        async fn ack(
12293            &self,
12294            target_session_id: &str,
12295            claim: &SessionInboxClaim,
12296        ) -> Result<(), bamboo_domain::SessionInboxError> {
12297            if self
12298                .fail_once
12299                .swap(false, std::sync::atomic::Ordering::SeqCst)
12300            {
12301                return Err(bamboo_domain::SessionInboxError::Storage(
12302                    "injected actor pre-ack failure".to_string(),
12303                ));
12304            }
12305            self.inner.ack(target_session_id, claim).await
12306        }
12307
12308        async fn inspect(
12309            &self,
12310            target_session_id: &str,
12311        ) -> Result<bamboo_domain::SessionInboxBacklog, bamboo_domain::SessionInboxError> {
12312            self.inner.inspect(target_session_id).await
12313        }
12314    }
12315
12316    async fn actor_inbox_fixture(
12317        session_id: &str,
12318    ) -> (
12319        tempfile::TempDir,
12320        Arc<bamboo_storage::SessionStoreV2>,
12321        Arc<bamboo_storage::LockedSessionStore>,
12322        Arc<dyn SessionInboxPort>,
12323        Session,
12324        SessionInboxClaim,
12325    ) {
12326        let temp = tempfile::tempdir().unwrap();
12327        let store = Arc::new(
12328            bamboo_storage::SessionStoreV2::new(temp.path().to_path_buf())
12329                .await
12330                .unwrap(),
12331        );
12332        let storage: Arc<dyn Storage> = store.clone();
12333        let locked = Arc::new(bamboo_storage::LockedSessionStore::new(storage));
12334        let inbox: Arc<dyn SessionInboxPort> = Arc::new(bamboo_storage::FileSessionInbox::new(
12335            store.clone(),
12336            bamboo_domain::SessionInboxLimits::default(),
12337        ));
12338        let session = Session::new(session_id, "model");
12339        store.save_session(&session).await.unwrap();
12340        let mut envelope =
12341            bamboo_domain::SessionMessageEnvelope::user_input(session_id, "actor follow-up");
12342        envelope.id =
12343            bamboo_domain::SessionMessageId::parse(format!("{session_id}-message")).unwrap();
12344        let receipt = inbox.deliver(&envelope).await.unwrap();
12345        inbox
12346            .mark_activation_eligible(
12347                session_id,
12348                receipt.generation,
12349                bamboo_domain::SessionActivationPolicy::InterruptSpecificWait,
12350            )
12351            .await
12352            .unwrap();
12353        let claim = inbox.claim(session_id, 1).await.unwrap().remove(0);
12354        (temp, store, locked, inbox, session, claim)
12355    }
12356
12357    fn actor_binding(
12358        store: Arc<bamboo_storage::SessionStoreV2>,
12359        inbox: Arc<dyn SessionInboxPort>,
12360        persistence: Arc<dyn RuntimeSessionPersistence>,
12361    ) -> SessionInboxRuntimeBinding {
12362        let storage: Arc<dyn Storage> = store;
12363        SessionInboxRuntimeBinding {
12364            router: SessionActivationRouter::new(),
12365            inbox,
12366            storage,
12367            persistence,
12368            parent_question_lock: None,
12369        }
12370    }
12371
12372    #[tokio::test]
12373    async fn remote_claim_seed_rejects_unsupported_inbox_before_dispatch() {
12374        let (_temp, store, locked, inbox, mut session, claim) =
12375            actor_inbox_fixture("unsupported-remote-inbox").await;
12376        let wrapped: Arc<dyn SessionInboxPort> = Arc::new(ActorFailBeforeAckInbox {
12377            inner: inbox.clone(),
12378            fail_once: std::sync::atomic::AtomicBool::new(false),
12379        });
12380        let binding = actor_binding(store.clone(), wrapped, locked);
12381        let fence = ActorActivationFence {
12382            schema_version: 1,
12383            actor_id: session.id.clone(),
12384            activation_id: "unsupported-inbox-activation".into(),
12385            attempt: 1,
12386            run_id: "unsupported-inbox-run".into(),
12387            lease_owner: "unsupported-inbox-host".into(),
12388            lease_epoch: 1,
12389        };
12390        let before = session.clone();
12391
12392        let error = checkpoint_claim_context_before_dispatch(
12393            &binding,
12394            &mut session,
12395            &[claim],
12396            Some((store.as_ref(), &fence)),
12397        )
12398        .await
12399        .unwrap_err();
12400        assert!(error
12401            .to_string()
12402            .contains("requires canonical FileSessionInbox"));
12403        assert_eq!(
12404            serde_json::to_value(&session.messages).unwrap(),
12405            serde_json::to_value(&before.messages).unwrap()
12406        );
12407        assert_eq!(
12408            serde_json::to_value(
12409                &store
12410                    .load_session(&session.id)
12411                    .await
12412                    .unwrap()
12413                    .unwrap()
12414                    .messages
12415            )
12416            .unwrap(),
12417            serde_json::to_value(&before.messages).unwrap()
12418        );
12419        assert_eq!(inbox.inspect(&session.id).await.unwrap().claimed, 1);
12420    }
12421
12422    #[tokio::test]
12423    async fn remote_broker_lineage_rejects_missing_parent_and_project_drift() {
12424        let (_temp, store, locked, inbox, parent, _claim) =
12425            actor_inbox_fixture("remote-parent").await;
12426        let child = Session::new_child_of("remote-child", &parent, "model", "child");
12427        store.save_session(&child).await.unwrap();
12428        let binding = actor_binding(store.clone(), inbox, locked);
12429        let first = remote_canonical_lineage(&binding, &child).await.unwrap();
12430        assert_eq!(first.len(), 2);
12431        let mut changed = child.clone();
12432        changed.set_project_id_meta("different-project");
12433        assert!(remote_canonical_lineage(&binding, &changed).await.is_err());
12434        store.delete_session(&parent.id).await.unwrap();
12435        assert!(remote_canonical_lineage(&binding, &child).await.is_err());
12436    }
12437
12438    #[tokio::test]
12439    async fn spawn_depth_remote_lineage_accepts_existing_tree_above_default() {
12440        let (_temp, store, locked, inbox, root, _claim) =
12441            actor_inbox_fixture("deep-remote-root").await;
12442        let mut parent = root;
12443        for depth in 1..=6 {
12444            let child =
12445                Session::new_child_of(format!("remote-depth-{depth}"), &parent, "model", "child");
12446            store.save_session(&child).await.unwrap();
12447            parent = child;
12448        }
12449        let binding = actor_binding(store, inbox, locked);
12450        let chain = remote_canonical_lineage(&binding, &parent).await.unwrap();
12451        assert_eq!(chain.len(), 7);
12452        assert_eq!(chain.last().unwrap().spawn_depth, 0);
12453    }
12454
12455    #[tokio::test]
12456    async fn child_creation_requires_canonical_host_lifetime_not_cached_candidate() {
12457        let (_temp, store, locked, inbox, parent, _claim) =
12458            actor_inbox_fixture("creation-parent").await;
12459        let mut child = Session::new_child_of("creation-child", &parent, "model", "child");
12460        child.set_project_id_meta("creation-project");
12461        store.save_session(&child).await.unwrap();
12462        let binding = actor_binding(store.clone(), inbox, locked);
12463        let job = SpawnJob {
12464            parent_session_id: parent.id.clone(),
12465            child_session_id: child.id.clone(),
12466            model: "model".into(),
12467            disabled_tools: None,
12468        };
12469        let creation = canonical_child_creation(&binding, &child, &job)
12470            .await
12471            .unwrap();
12472        assert_eq!(creation.created_at, child.created_at);
12473        assert_eq!(creation.spawn_depth, child.spawn_depth);
12474        for damage in 0..7 {
12475            let mut held = child.clone();
12476            match damage {
12477                0 => held.created_at += chrono::Duration::nanoseconds(1),
12478                1 => held.parent_session_id = Some("foreign-parent".into()),
12479                2 => held.root_session_id = "foreign-root".into(),
12480                3 => held.spawn_depth += 1,
12481                4 => held.set_project_id_meta("foreign-project"),
12482                5 => held.kind = bamboo_domain::SessionKind::Root,
12483                _ => held.id = " creation-child ".into(),
12484            }
12485            assert!(
12486                canonical_child_creation(&binding, &held, &job)
12487                    .await
12488                    .is_err(),
12489                "damage {damage}"
12490            );
12491        }
12492        store.delete_session(&child.id).await.unwrap();
12493        assert!(canonical_child_creation(&binding, &child, &job)
12494            .await
12495            .is_err());
12496        let mut replacement = Session::new_child_of(&child.id, &parent, "model", "replacement");
12497        replacement.set_project_id_meta("creation-project");
12498        assert_ne!(replacement.created_at, child.created_at);
12499        store.save_session(&replacement).await.unwrap();
12500        assert!(
12501            canonical_child_creation(&binding, &child, &job)
12502                .await
12503                .is_err(),
12504            "old held spec cannot adopt replacement birth"
12505        );
12506        canonical_child_creation(&binding, &replacement, &job)
12507            .await
12508            .unwrap();
12509    }
12510
12511    #[tokio::test]
12512    async fn actor_immediate_claims_keep_exact_policy_across_mixed_order_ack_and_restart() {
12513        use bamboo_domain::{SessionActivationPolicy, SessionInboxLimits, SessionMessageEnvelope};
12514        for policies in [
12515            [
12516                SessionActivationPolicy::RespectSpecificWait,
12517                SessionActivationPolicy::InterruptSpecificWait,
12518            ],
12519            [
12520                SessionActivationPolicy::InterruptSpecificWait,
12521                SessionActivationPolicy::RespectSpecificWait,
12522            ],
12523        ] {
12524            let temp = tempfile::tempdir().unwrap();
12525            let store = Arc::new(
12526                bamboo_storage::SessionStoreV2::new(temp.path().into())
12527                    .await
12528                    .unwrap(),
12529            );
12530            let session_id = "actor-mixed-intent-policy";
12531            store
12532                .save_session(&Session::new(session_id, "model"))
12533                .await
12534                .unwrap();
12535            let inbox = bamboo_storage::FileSessionInbox::new(store, SessionInboxLimits::default());
12536            let staged = SessionMessageEnvelope::user_input(session_id, "staged sibling");
12537            let staged_receipt = inbox.deliver(&staged).await.unwrap();
12538            let mut sent = Vec::new();
12539            for policy in policies {
12540                let envelope = SessionMessageEnvelope::user_input(session_id, "immediate");
12541                let receipt = inbox
12542                    .deliver_with_activation_intent(&envelope, policy, None)
12543                    .await
12544                    .unwrap();
12545                sent.push((envelope, policy, receipt));
12546            }
12547            let restarted = Arc::new(
12548                bamboo_storage::SessionStoreV2::new(temp.path().into())
12549                    .await
12550                    .unwrap(),
12551            );
12552            let inbox: Arc<dyn SessionInboxPort> = Arc::new(bamboo_storage::FileSessionInbox::new(
12553                restarted.clone(),
12554                SessionInboxLimits::default(),
12555            ));
12556            let binding = actor_binding(
12557                restarted.clone(),
12558                inbox.clone(),
12559                Arc::new(bamboo_storage::LockedSessionStore::new(restarted.clone())),
12560            );
12561            let mut session = restarted.load_session(session_id).await.unwrap().unwrap();
12562            let pairs =
12563                claim_canonical_deliveries(&binding, &mut session, "run-1", usize::MAX, None)
12564                    .await
12565                    .unwrap();
12566            assert_eq!(pairs.len(), 2, "the staged sibling has no permission");
12567            for ((claim, delivery), (envelope, policy, receipt)) in pairs.iter().zip(&sent) {
12568                assert_eq!(claim.envelope.id, envelope.id);
12569                assert_eq!(claim.generation, receipt.generation);
12570                assert_eq!(claim.activation_policy, *policy);
12571                assert_eq!(delivery.activation_policy, *policy);
12572                checkpoint_and_ack_canonical_claim(&binding, &mut session, claim, None)
12573                    .await
12574                    .unwrap();
12575            }
12576            let restarted = Arc::new(
12577                bamboo_storage::SessionStoreV2::new(temp.path().into())
12578                    .await
12579                    .unwrap(),
12580            );
12581            let inbox: Arc<dyn SessionInboxPort> = Arc::new(bamboo_storage::FileSessionInbox::new(
12582                restarted.clone(),
12583                SessionInboxLimits::default(),
12584            ));
12585            let binding = actor_binding(
12586                restarted.clone(),
12587                inbox.clone(),
12588                Arc::new(bamboo_storage::LockedSessionStore::new(restarted.clone())),
12589            );
12590            session = restarted.load_session(session_id).await.unwrap().unwrap();
12591            for (envelope, policy, receipt) in &sent {
12592                assert_eq!(
12593                    inbox
12594                        .deliver_with_activation_intent(envelope, *policy, None)
12595                        .await
12596                        .unwrap(),
12597                    *receipt
12598                );
12599            }
12600            assert!(
12601                claim_canonical_deliveries(&binding, &mut session, "run-2", usize::MAX, None)
12602                    .await
12603                    .unwrap()
12604                    .is_empty()
12605            );
12606            inbox
12607                .mark_activation_eligible(
12608                    session_id,
12609                    staged_receipt.generation,
12610                    SessionActivationPolicy::RespectSpecificWait,
12611                )
12612                .await
12613                .unwrap();
12614            let pairs =
12615                claim_canonical_deliveries(&binding, &mut session, "run-2", usize::MAX, None)
12616                    .await
12617                    .unwrap();
12618            assert_eq!(pairs.len(), 1);
12619            assert_eq!(pairs[0].0.envelope.id, staged.id);
12620            assert_eq!(
12621                pairs[0].1.activation_policy,
12622                SessionActivationPolicy::RespectSpecificWait
12623            );
12624        }
12625    }
12626
12627    #[tokio::test]
12628    async fn actor_mismatched_typed_marker_id_collision_never_acks() {
12629        let (_temp, store, locked, inbox, mut session, claim) =
12630            actor_inbox_fixture("actor-live-id-collision").await;
12631        let mut forged = claim.envelope.to_provider_message().unwrap();
12632        forged.metadata = Some(serde_json::json!({
12633            "session_message": {
12634                "id": claim.envelope.id,
12635                "target_session_id": "different-session"
12636            }
12637        }));
12638        session.add_message(forged);
12639        let persistence: Arc<dyn RuntimeSessionPersistence> = locked;
12640        let binding = actor_binding(store, inbox.clone(), persistence);
12641
12642        assert!(
12643            checkpoint_and_ack_canonical_claim(&binding, &mut session, &claim, None)
12644                .await
12645                .is_err()
12646        );
12647        assert_eq!(inbox.inspect(&session.id).await.unwrap().claimed, 1);
12648        assert!(!inbox
12649            .was_admitted(&session.id, &claim.envelope.id)
12650            .await
12651            .unwrap());
12652    }
12653
12654    #[tokio::test]
12655    async fn actor_concurrent_durable_id_collision_after_claim_never_acks() {
12656        let (_temp, store, locked, inbox, mut session, claim) =
12657            actor_inbox_fixture("actor-durable-id-collision").await;
12658        let mut concurrent = store.load_session(&session.id).await.unwrap().unwrap();
12659        let mut forged = bamboo_agent_core::Message::user("concurrent actor collision");
12660        forged.id = claim.envelope.id.to_string();
12661        concurrent.add_message(forged);
12662        store.save_session(&concurrent).await.unwrap();
12663        let persistence: Arc<dyn RuntimeSessionPersistence> = locked;
12664        let binding = actor_binding(store.clone(), inbox.clone(), persistence);
12665
12666        assert!(
12667            checkpoint_and_ack_canonical_claim(&binding, &mut session, &claim, None)
12668                .await
12669                .is_err()
12670        );
12671        assert_eq!(inbox.inspect(&session.id).await.unwrap().claimed, 1);
12672        assert!(!inbox
12673            .was_admitted(&session.id, &claim.envelope.id)
12674            .await
12675            .unwrap());
12676        let durable = store.load_session(&session.id).await.unwrap().unwrap();
12677        assert!(!durable.messages.iter().any(|message| {
12678            bamboo_domain::is_matching_session_message(message, &claim.envelope)
12679        }));
12680    }
12681
12682    #[tokio::test]
12683    async fn actor_concurrent_durable_typed_body_mismatch_never_acks() {
12684        let (_temp, store, locked, inbox, mut session, claim) =
12685            actor_inbox_fixture("actor-durable-typed-body-collision").await;
12686        let mut different = claim.envelope.clone();
12687        different.body = bamboo_domain::SessionMessageBody::Content(
12688            bamboo_domain::SessionMessageContent::text("forged actor body"),
12689        );
12690        let mut concurrent = store.load_session(&session.id).await.unwrap().unwrap();
12691        concurrent.add_message(different.to_provider_message().unwrap());
12692        store.save_session(&concurrent).await.unwrap();
12693        let persistence: Arc<dyn RuntimeSessionPersistence> = locked;
12694        let binding = actor_binding(store.clone(), inbox.clone(), persistence);
12695
12696        assert!(
12697            checkpoint_and_ack_canonical_claim(&binding, &mut session, &claim, None)
12698                .await
12699                .is_err()
12700        );
12701        assert_eq!(inbox.inspect(&session.id).await.unwrap().claimed, 1);
12702        assert!(!inbox
12703            .was_admitted(&session.id, &claim.envelope.id)
12704            .await
12705            .unwrap());
12706        let durable = store.load_session(&session.id).await.unwrap().unwrap();
12707        assert!(!durable
12708            .messages
12709            .iter()
12710            .any(|message| bamboo_domain::is_matching_session_message(message, &claim.envelope)));
12711    }
12712
12713    #[tokio::test]
12714    async fn actor_checkpoint_failure_rolls_back_and_restart_admits_once() {
12715        let (_temp, store, locked, inbox, mut session, claim) =
12716            actor_inbox_fixture("actor-checkpoint-failure").await;
12717        let envelope_id = claim.envelope.id.clone();
12718        let fault: Arc<dyn RuntimeSessionPersistence> = Arc::new(ActorFaultingPersistence {
12719            inner: locked.clone(),
12720            fail_checkpoint_once: std::sync::atomic::AtomicBool::new(true),
12721        });
12722        let binding = actor_binding(store.clone(), inbox.clone(), fault);
12723
12724        assert!(
12725            checkpoint_and_ack_canonical_claim(&binding, &mut session, &claim, None)
12726                .await
12727                .is_err()
12728        );
12729        assert!(!session
12730            .messages
12731            .iter()
12732            .any(|message| message.id == envelope_id.as_str()));
12733        assert_eq!(inbox.inspect(&session.id).await.unwrap().claimed, 1);
12734        assert!(!inbox.was_admitted(&session.id, &envelope_id).await.unwrap());
12735
12736        let reopened: Arc<dyn SessionInboxPort> = Arc::new(bamboo_storage::FileSessionInbox::new(
12737            store.clone(),
12738            bamboo_domain::SessionInboxLimits::default(),
12739        ));
12740        let recovered = reopened.claim(&session.id, 1).await.unwrap().remove(0);
12741        let persistence: Arc<dyn RuntimeSessionPersistence> = locked;
12742        let binding = actor_binding(store.clone(), reopened.clone(), persistence);
12743        let mut restarted = store.load_session(&session.id).await.unwrap().unwrap();
12744        checkpoint_and_ack_canonical_claim(&binding, &mut restarted, &recovered, None)
12745            .await
12746            .unwrap();
12747        assert_eq!(
12748            restarted
12749                .messages
12750                .iter()
12751                .filter(|message| message.id == envelope_id.as_str())
12752                .count(),
12753            1
12754        );
12755        assert!(reopened
12756            .was_admitted(&session.id, &envelope_id)
12757            .await
12758            .unwrap());
12759        let backlog = reopened.inspect(&session.id).await.unwrap();
12760        assert_eq!(backlog.pending + backlog.claimed, 0);
12761    }
12762
12763    #[tokio::test]
12764    async fn actor_checkpoint_success_pre_ack_failure_recovers_without_duplicate() {
12765        let (_temp, store, locked, real_inbox, mut session, claim) =
12766            actor_inbox_fixture("actor-pre-ack-failure").await;
12767        let envelope_id = claim.envelope.id.clone();
12768        let faulted: Arc<dyn SessionInboxPort> = Arc::new(ActorFailBeforeAckInbox {
12769            inner: real_inbox.clone(),
12770            fail_once: std::sync::atomic::AtomicBool::new(true),
12771        });
12772        let persistence: Arc<dyn RuntimeSessionPersistence> = locked.clone();
12773        let binding = actor_binding(store.clone(), faulted, persistence);
12774
12775        assert!(
12776            checkpoint_and_ack_canonical_claim(&binding, &mut session, &claim, None)
12777                .await
12778                .is_err()
12779        );
12780        let durable = store.load_session(&session.id).await.unwrap().unwrap();
12781        assert_eq!(
12782            durable
12783                .messages
12784                .iter()
12785                .filter(|message| message.id == envelope_id.as_str())
12786                .count(),
12787            1
12788        );
12789        assert_eq!(real_inbox.inspect(&session.id).await.unwrap().claimed, 1);
12790        assert!(!real_inbox
12791            .was_admitted(&session.id, &envelope_id)
12792            .await
12793            .unwrap());
12794
12795        let reopened: Arc<dyn SessionInboxPort> = Arc::new(bamboo_storage::FileSessionInbox::new(
12796            store.clone(),
12797            bamboo_domain::SessionInboxLimits::default(),
12798        ));
12799        let recovered = reopened.claim(&session.id, 1).await.unwrap().remove(0);
12800        let persistence: Arc<dyn RuntimeSessionPersistence> = locked;
12801        let binding = actor_binding(store.clone(), reopened.clone(), persistence);
12802        let mut restarted = durable;
12803        checkpoint_and_ack_canonical_claim(&binding, &mut restarted, &recovered, None)
12804            .await
12805            .unwrap();
12806        assert_eq!(
12807            restarted
12808                .messages
12809                .iter()
12810                .filter(|message| message.id == envelope_id.as_str())
12811                .count(),
12812            1
12813        );
12814        assert!(reopened
12815            .was_admitted(&session.id, &envelope_id)
12816            .await
12817            .unwrap());
12818        let backlog = reopened.inspect(&session.id).await.unwrap();
12819        assert_eq!(backlog.pending + backlog.claimed, 0);
12820    }
12821
12822    struct ConfirmationSequenceLink {
12823        frames: VecDeque<ChildFrame>,
12824        sent: Vec<ParentFrame>,
12825    }
12826
12827    #[async_trait]
12828    impl bamboo_subagent::ChildLink for ConfirmationSequenceLink {
12829        async fn send(&mut self, frame: ParentFrame) -> bamboo_subagent::TransportResult<()> {
12830            self.sent.push(frame);
12831            Ok(())
12832        }
12833
12834        async fn next_frame(&mut self) -> bamboo_subagent::TransportResult<Option<ChildFrame>> {
12835            match self.frames.pop_front() {
12836                Some(frame) => Ok(Some(frame)),
12837                None => std::future::pending().await,
12838            }
12839        }
12840    }
12841
12842    #[tokio::test]
12843    async fn failed_actor_retry_checkpoints_new_input_and_rejects_old_or_stale_state() {
12844        for case in [
12845            "accepted",
12846            "readonly-refusal",
12847            "prefix-refusal",
12848            "old-claim",
12849            "old-checkpoint",
12850            "succeeded",
12851            "changed-birth",
12852            "oversize",
12853        ] {
12854            let temp = tempfile::tempdir().unwrap();
12855            let store = Arc::new(
12856                bamboo_storage::SessionStoreV2::new(temp.path().into())
12857                    .await
12858                    .unwrap(),
12859            );
12860            let parent = Session::new("retry-parent", "model");
12861            store.save_session(&parent).await.unwrap();
12862            let mut child =
12863                Session::new_child_of("retry-child", &parent, "model", "original assignment");
12864            child.add_message(bamboo_agent_core::Message::user("original assignment"));
12865            store.save_session(&child).await.unwrap();
12866            let policy = Arc::new(bamboo_tools::permission::PermissionConfig::new());
12867            bind_local_control_plane(store.as_ref(), &child.id, policy.as_ref()).await;
12868            let locked = Arc::new(bamboo_storage::LockedSessionStore::new(store.clone()));
12869            let inbox: Arc<dyn SessionInboxPort> = Arc::new(bamboo_storage::FileSessionInbox::new(
12870                store.clone(),
12871                bamboo_domain::SessionInboxLimits::default(),
12872            ));
12873            let binding = actor_binding(store.clone(), inbox, locked);
12874            let registration = binding
12875                .router
12876                .register_run(&child.id, "first-run")
12877                .await
12878                .unwrap();
12879            let first = PlainActorActivation::start(
12880                store.clone(),
12881                &child,
12882                &binding,
12883                Some("first-run"),
12884                Some(policy.clone()),
12885                true,
12886                None,
12887            )
12888            .await
12889            .unwrap();
12890            if matches!(case, "old-claim" | "old-checkpoint") {
12891                let old = bamboo_domain::SessionMessageEnvelope::user_input(
12892                    &child.id,
12893                    "unconfirmed old input",
12894                );
12895                let receipt = first.input_inbox.deliver(&old).await.unwrap();
12896                first
12897                    .input_inbox
12898                    .mark_activation_eligible(
12899                        &child.id,
12900                        receipt.generation,
12901                        bamboo_domain::SessionActivationPolicy::InterruptSpecificWait,
12902                    )
12903                    .await
12904                    .unwrap();
12905                let claim = first
12906                    .claim_input(&binding, &child, "first-run")
12907                    .await
12908                    .unwrap()
12909                    .unwrap();
12910                if case == "old-checkpoint" {
12911                    first
12912                        .checkpoint_input(&mut child, "first-run", &claim)
12913                        .await
12914                        .unwrap();
12915                }
12916            }
12917            first
12918                .finish(if case == "succeeded" {
12919                    ActorActivationFinish::Succeeded
12920                } else {
12921                    ActorActivationFinish::Failed
12922                })
12923                .await
12924                .unwrap();
12925            // Drop schedules cleanup; the retry needs the old owner released first.
12926            registration.abandon().await;
12927            assert!(
12928                !binding.router.owns_run(&child.id, "first-run").await,
12929                "{case}: abandoned first run must release its owner before retry"
12930            );
12931            let envelope = bamboo_domain::SessionMessageEnvelope::user_input(
12932                &child.id,
12933                if case == "oversize" {
12934                    "x".repeat(8193)
12935                } else {
12936                    "NEW_CORRECTION".into()
12937                },
12938            );
12939            let receipt = first.input_inbox.deliver(&envelope).await.unwrap();
12940            first
12941                .input_inbox
12942                .mark_activation_eligible(
12943                    &child.id,
12944                    receipt.generation,
12945                    bamboo_domain::SessionActivationPolicy::InterruptSpecificWait,
12946                )
12947                .await
12948                .unwrap();
12949            let before = store.load_session(&child.id).await.unwrap().unwrap();
12950            let expected_messages = serde_json::to_value(&before.messages).unwrap();
12951            child = before.clone();
12952            if case == "prefix-refusal" {
12953                child.messages.last_mut().unwrap().content = "stale cached prefix".into();
12954            }
12955            if case == "changed-birth" {
12956                child.created_at += chrono::Duration::milliseconds(1);
12957            }
12958            let _retry_registration = binding
12959                .router
12960                .register_run(&child.id, "retry-run")
12961                .await
12962                .unwrap();
12963            let result = PlainActorActivation::start_and_prepare(
12964                store.clone(),
12965                &mut child,
12966                &binding,
12967                Some("retry-run"),
12968                Some(policy.clone()),
12969                case != "readonly-refusal",
12970                None,
12971            )
12972            .await;
12973            if case != "accepted" {
12974                assert!(result.is_err(), "{case}: no retry Run may be dispatched");
12975                let actual = store.load_session(&before.id).await.unwrap().unwrap();
12976                assert_eq!(
12977                    serde_json::to_value(&actual.messages).unwrap(),
12978                    expected_messages,
12979                    "{case}: failed preparation preserves Main history"
12980                );
12981                assert!(
12982                    !first
12983                        .input_inbox
12984                        .was_admitted(&before.id, &envelope.id)
12985                        .await
12986                        .unwrap(),
12987                    "{case}: no ACK"
12988                );
12989                if case == "readonly-refusal" {
12990                    let entry = store.inspect_actor(&before.id).await.unwrap();
12991                    assert_eq!(entry.actor.current_attempt, 1);
12992                    assert_eq!(
12993                        entry.activation.unwrap().activation_id,
12994                        first.fence.activation_id
12995                    );
12996                    let backlog = first.input_inbox.inspect(&before.id).await.unwrap();
12997                    assert_eq!((backlog.pending, backlog.claimed), (1, 0));
12998                }
12999                if matches!(case, "prefix-refusal" | "changed-birth" | "oversize") {
13000                    assert_eq!(
13001                        first.input_inbox.inspect(&before.id).await.unwrap().claimed,
13002                        1
13003                    );
13004                    assert_eq!(
13005                        store
13006                            .inspect_actor(&before.id)
13007                            .await
13008                            .unwrap()
13009                            .activation
13010                            .unwrap()
13011                            .status,
13012                        bamboo_domain::ActorActivationStatus::Failed
13013                    );
13014                }
13015                if case == "oversize" {
13016                    let reopened = Arc::new(
13017                        bamboo_storage::SessionStoreV2::new(temp.path().into())
13018                            .await
13019                            .unwrap(),
13020                    );
13021                    let cold_inbox = bamboo_storage::FileSessionInbox::new(
13022                        reopened,
13023                        bamboo_domain::SessionInboxLimits::default(),
13024                    );
13025                    let now = chrono::Utc::now();
13026                    let leases = cold_inbox
13027                        .inspect_owned_leases(&before.id, 2, now)
13028                        .await
13029                        .unwrap();
13030                    assert_eq!(leases.len(), 1);
13031                    assert_eq!(leases[0].failure_count, 1);
13032                    assert_eq!(
13033                        leases[0].last_error_code.as_deref(),
13034                        Some("unsupported_actor_input")
13035                    );
13036                    let retry_after = leases[0].retry_after.unwrap();
13037                    assert!(retry_after > now);
13038                    let readiness = cold_inbox
13039                        .inspect_wake_readiness(&before.id, now)
13040                        .await
13041                        .unwrap();
13042                    assert!(
13043                        readiness.ready.is_none(),
13044                        "failed correction must not hot-loop"
13045                    );
13046                    assert_eq!(readiness.next_due_at, Some(retry_after));
13047                }
13048                continue;
13049            }
13050            let (retry, prepared) = result.unwrap();
13051            let (prefix, mut deliveries) = prepared.unwrap();
13052            assert_eq!(deliveries.len(), 1);
13053            let delivery = deliveries.pop().unwrap();
13054            assert_eq!(
13055                prefix,
13056                serde_json::to_value(&before.messages)
13057                    .unwrap()
13058                    .as_array()
13059                    .unwrap()
13060                    .clone()
13061            );
13062            assert_eq!(delivery.envelope, envelope);
13063            assert_eq!(delivery.activation_run_id, "retry-run");
13064            assert_eq!(delivery.canonical_claim_generation, receipt.generation);
13065            assert!(store
13066                .validate_fence(&first.fence, chrono::Utc::now())
13067                .await
13068                .is_err());
13069            let running = store.inspect_actor(&child.id).await.unwrap();
13070            assert_eq!(running.actor.current_attempt, 2);
13071            assert_eq!(
13072                running.activation.as_ref().unwrap().status,
13073                bamboo_domain::ActorActivationStatus::Running
13074            );
13075            assert!(running.activation.unwrap().lease_epoch > first.fence.lease_epoch);
13076            let claim = retry.initial_inputs.first().unwrap();
13077            assert!(child
13078                .session_inbox_admission()
13079                .unwrap()
13080                .contains(&envelope.id));
13081            assert!(
13082                !retry
13083                    .input_inbox
13084                    .was_admitted(&child.id, &envelope.id)
13085                    .await
13086                    .unwrap(),
13087                "checkpoint is not startup confirmation"
13088            );
13089            let mut confirmation = admission_confirmation(&child.id, &claim.claim, "retry-run");
13090            confirmation.activation_run_id = "first-run".into();
13091            assert!(retry
13092                .confirm_input(&binding, &child, "retry-run", claim, &confirmation)
13093                .await
13094                .is_err());
13095            assert!(!retry
13096                .input_inbox
13097                .was_admitted(&child.id, &envelope.id)
13098                .await
13099                .unwrap());
13100            confirmation.activation_run_id = "retry-run".into();
13101            retry
13102                .confirm_input(&binding, &child, "retry-run", claim, &confirmation)
13103                .await
13104                .unwrap();
13105            let (tx, mut rx) = mpsc::channel(1);
13106            retry
13107                .append_reply(&mut child, "RETRY_REPLY".into(), &tx)
13108                .await
13109                .unwrap();
13110            rx.recv().await.unwrap();
13111            retry
13112                .finish(ActorActivationFinish::Succeeded)
13113                .await
13114                .unwrap();
13115            let reopened = Arc::new(
13116                bamboo_storage::SessionStoreV2::new(temp.path().into())
13117                    .await
13118                    .unwrap(),
13119            );
13120            let cold = reopened.load_session(&child.id).await.unwrap().unwrap();
13121            assert_eq!(cold.created_at, before.created_at);
13122            assert_eq!(cold.messages.len(), before.messages.len() + 2);
13123            assert_eq!(
13124                cold.messages[before.messages.len()].id,
13125                envelope.id.as_str()
13126            );
13127            assert_eq!(cold.messages.last().unwrap().content, "RETRY_REPLY");
13128            assert_eq!(
13129                cold.messages
13130                    .iter()
13131                    .filter(|m| m.id == envelope.id.as_str())
13132                    .count(),
13133                1
13134            );
13135            let cold_inbox = bamboo_storage::FileSessionInbox::new(
13136                reopened,
13137                bamboo_domain::SessionInboxLimits::default(),
13138            );
13139            assert!(cold_inbox
13140                .was_admitted(&cold.id, &envelope.id)
13141                .await
13142                .unwrap());
13143            assert_eq!(cold_inbox.inspect(&cold.id).await.unwrap().claimed, 0);
13144        }
13145    }
13146
13147    #[tokio::test]
13148    async fn pre_ack_owned_recovery_requires_expiry_provenance_and_verified_already() {
13149        for case in [
13150            "accepted",
13151            "actor-live",
13152            "claim-live",
13153            "legacy",
13154            "ack",
13155            "readonly",
13156            "prefix",
13157            "duplicate",
13158            "missing-marker",
13159            "mixed",
13160            "project",
13161            "birth",
13162        ] {
13163            let temp = tempfile::tempdir().unwrap();
13164            let store = Arc::new(
13165                bamboo_storage::SessionStoreV2::new(temp.path().into())
13166                    .await
13167                    .unwrap(),
13168            );
13169            let parent = Session::new("recovery-parent", "model");
13170            store.save_session(&parent).await.unwrap();
13171            let mut child =
13172                Session::new_child_of("recovery-child", &parent, "model", "original task");
13173            child.add_message(bamboo_agent_core::Message::user("original task"));
13174            store.save_session(&child).await.unwrap();
13175            let policy = Arc::new(bamboo_tools::permission::PermissionConfig::new());
13176            bind_local_control_plane(store.as_ref(), &child.id, policy.as_ref()).await;
13177            let inbox = bamboo_storage::FileSessionInbox::new(
13178                store.clone(),
13179                bamboo_domain::SessionInboxLimits::default(),
13180            );
13181            // Complete unfenced persistence before starting the setup lease window.
13182            let envelope =
13183                bamboo_domain::SessionMessageEnvelope::user_input(&child.id, "RECOVER_SAME_INPUT");
13184            let receipt = inbox.deliver(&envelope).await.unwrap();
13185            inbox
13186                .mark_activation_eligible(
13187                    &child.id,
13188                    receipt.generation,
13189                    bamboo_domain::SessionActivationPolicy::InterruptSpecificWait,
13190                )
13191                .await
13192                .unwrap();
13193            let now = chrono::Utc::now();
13194            let cutoff = now + chrono::Duration::seconds(10);
13195            let consumer = SessionInboxConsumerId::new();
13196            let old = store
13197                .claim_activation(&ActorActivationClaim {
13198                    actor_id: child.id.clone(),
13199                    run_id: "old-run".into(),
13200                    lease_owner: consumer.as_str().into(),
13201                    lease_expires_at: if case == "actor-live" {
13202                        now + chrono::Duration::minutes(5)
13203                    } else {
13204                        cutoff
13205                    },
13206                    inbox_generation: 0,
13207                    placement_ref: if case == "legacy" {
13208                        None
13209                    } else {
13210                        Some(bamboo_domain::ActorPlacementRef {
13211                            class: bamboo_domain::ActorPlacementClass::Local,
13212                            lease_id: "owned-initial-release-v1:actual-old-worker".into(),
13213                            slot_epoch: None,
13214                        })
13215                    },
13216                    now,
13217                })
13218                .await
13219                .unwrap();
13220            store
13221                .start_activation(&old.fence(), chrono::Utc::now())
13222                .await
13223                .unwrap();
13224            let claim_now = chrono::Utc::now();
13225            let claim = inbox
13226                .claim_owned(
13227                    &child.id,
13228                    1,
13229                    Some("old-run"),
13230                    &SessionInboxLeaseRequest {
13231                        consumer,
13232                        now: claim_now,
13233                        duration: if case == "claim-live" {
13234                            chrono::Duration::minutes(5)
13235                        } else {
13236                            cutoff - claim_now
13237                        },
13238                    },
13239                )
13240                .await
13241                .unwrap()
13242                .pop()
13243                .unwrap();
13244            child = inbox
13245                .checkpoint_actor_input(bamboo_storage::ActorInputCheckpoint {
13246                    fence: old.fence(),
13247                    expected_created_at: child.created_at,
13248                    expected_conversation_generation: child.conversation_generation_observation(),
13249                    claim: claim.clone(),
13250                    expected_messages: child.messages.clone(),
13251                    expected_provider_transcript: child.provider_transcript.clone(),
13252                    expected_admission: None,
13253                })
13254                .await
13255                .unwrap()
13256                .session;
13257            // The real checkpoint must succeed under the original live authority.
13258            store
13259                .validate_fence(&old.fence(), chrono::Utc::now())
13260                .await
13261                .unwrap_or_else(|error| panic!("{case}: setup fence: {error:?}"));
13262            let setup_leases = inbox
13263                .inspect_owned_leases(&child.id, 2, chrono::Utc::now())
13264                .await
13265                .unwrap();
13266            assert_eq!(setup_leases.len(), 1, "{case}: original setup claim");
13267            assert_eq!(setup_leases[0].epoch, claim.lease.epoch, "{case}");
13268            assert_eq!(setup_leases[0].expires_at, claim.lease.expires_at, "{case}");
13269            assert!(!setup_leases[0].expired, "{case}: setup claim must be live");
13270            let setup_at = chrono::Utc::now();
13271            assert!(
13272                setup_at < old.lease_expires_at,
13273                "{case}: setup Actor expired"
13274            );
13275            assert!(
13276                setup_at < claim.lease.expires_at,
13277                "{case}: setup claim expired"
13278            );
13279            eprintln!("pre-ACK {case}: setup live at {setup_at}, cutoff {cutoff}");
13280            if case == "ack" {
13281                inbox
13282                    .ack_owned(&child.id, &claim, chrono::Utc::now())
13283                    .await
13284                    .unwrap();
13285            }
13286            if case == "mixed" {
13287                inbox
13288                    .deliver(&bamboo_domain::SessionMessageEnvelope::user_input(
13289                        &child.id,
13290                        "different input",
13291                    ))
13292                    .await
13293                    .unwrap();
13294            }
13295            let directory = store
13296                .bamboo_home_dir()
13297                .join(store.resolve_rel_path(&child.id).await.unwrap());
13298            let main = std::fs::read(directory.join("session.json")).unwrap();
13299            let canonical = child.clone();
13300            match case {
13301                "prefix" => child.messages[0].content.push_str(" stale"),
13302                "duplicate" => child.messages.push(child.messages.last().unwrap().clone()),
13303                "missing-marker" => {
13304                    child
13305                        .messages
13306                        .last_mut()
13307                        .unwrap()
13308                        .metadata
13309                        .as_mut()
13310                        .unwrap()
13311                        .as_object_mut()
13312                        .unwrap()
13313                        .remove("_bamboo_owned_input_checkpoint");
13314                }
13315                "project" => child.set_project_id_meta(bamboo_domain::ProjectId::new().to_string()),
13316                "birth" => child.created_at += chrono::Duration::nanoseconds(1),
13317                _ => {}
13318            }
13319            // One shared short deadline per case; never wait for the live controls.
13320            // No reclaim, renewal or authority-file edit occurs between these stages.
13321            let delay = (cutoff - chrono::Utc::now()).to_std().unwrap_or_default();
13322            tokio::time::sleep(delay + Duration::from_millis(10)).await;
13323            let recovery_at = chrono::Utc::now();
13324            assert!(recovery_at >= cutoff, "{case}: real expiry is required");
13325            let before_recovery = store.inspect_actor(&child.id).await.unwrap();
13326            let original = before_recovery.activation.unwrap();
13327            assert_eq!(original.fence(), old.fence(), "{case}: same authority");
13328            assert_eq!(original.lease_expires_at, old.lease_expires_at, "{case}");
13329            assert_eq!(
13330                store
13331                    .validate_fence(&old.fence(), recovery_at)
13332                    .await
13333                    .is_ok(),
13334                case == "actor-live",
13335                "{case}: Actor expiry control"
13336            );
13337            let recovery_leases = inbox
13338                .inspect_owned_leases(&child.id, 2, recovery_at)
13339                .await
13340                .unwrap();
13341            if case == "ack" {
13342                assert!(recovery_leases.is_empty(), "{case}: terminal ACK");
13343            } else {
13344                assert_eq!(recovery_leases.len(), 1, "{case}: original recovery claim");
13345                assert_eq!(recovery_leases[0].epoch, claim.lease.epoch, "{case}");
13346                assert_eq!(
13347                    recovery_leases[0].expires_at, claim.lease.expires_at,
13348                    "{case}"
13349                );
13350                assert_eq!(recovery_leases[0].expired, case != "claim-live", "{case}");
13351            }
13352            eprintln!("pre-ACK {case}: recovery expiry control at {recovery_at}");
13353            let second = Arc::new(
13354                bamboo_storage::SessionStoreV2::new(temp.path().into())
13355                    .await
13356                    .unwrap(),
13357            );
13358            let binding = actor_binding(
13359                second.clone(),
13360                Arc::new(bamboo_storage::FileSessionInbox::new(
13361                    second.clone(),
13362                    bamboo_domain::SessionInboxLimits::default(),
13363                )),
13364                Arc::new(bamboo_storage::LockedSessionStore::new(second.clone())),
13365            );
13366            let _owner = binding
13367                .router
13368                .register_run(&child.id, "replacement-run")
13369                .await
13370                .unwrap();
13371            let prepared = PlainActorActivation::start_and_prepare(
13372                second.clone(),
13373                &mut child,
13374                &binding,
13375                Some("replacement-run"),
13376                Some(policy.clone()),
13377                case != "readonly",
13378                Some("probed-new-worker"),
13379            )
13380            .await;
13381            assert_eq!(
13382                std::fs::read(directory.join("session.json")).unwrap(),
13383                main,
13384                "{case}: Already or refusal never rewrites Main"
13385            );
13386            if case != "accepted" {
13387                assert!(prepared.is_err(), "{case}");
13388                assert_eq!(
13389                    inbox.was_admitted(&child.id, &envelope.id).await.unwrap(),
13390                    case == "ack"
13391                );
13392                continue;
13393            }
13394            let (replacement, delivery) = prepared.unwrap();
13395            assert!(replacement.recovering_pre_ack);
13396            let (messages, mut deliveries) = delivery.unwrap();
13397            assert_eq!(deliveries.len(), 1);
13398            let delivery = deliveries.pop().unwrap();
13399            let new_claim = replacement.initial_inputs.first().unwrap();
13400            assert_eq!(new_claim.claim.envelope, claim.claim.envelope);
13401            assert_eq!(new_claim.claim.generation, claim.claim.generation);
13402            assert_eq!(
13403                new_claim.claim.activation_policy,
13404                claim.claim.activation_policy
13405            );
13406            assert!(new_claim.lease.epoch > claim.lease.epoch);
13407            assert_ne!(new_claim.lease.incarnation, claim.lease.incarnation);
13408            assert_eq!(
13409                replacement.inbox_generation, 1,
13410                "old Actor generation was zero"
13411            );
13412            assert_eq!(
13413                second
13414                    .inspect_actor(&child.id)
13415                    .await
13416                    .unwrap()
13417                    .actor
13418                    .current_attempt,
13419                2
13420            );
13421            assert!(second
13422                .validate_fence(&old.fence(), chrono::Utc::now())
13423                .await
13424                .is_err());
13425            assert!(inbox
13426                .ack_owned(&child.id, &claim, chrono::Utc::now())
13427                .await
13428                .is_err());
13429            // The original Store/claim survives the handoff. Even with the
13430            // current exact prefix and dedupe cursor, its delayed checkpoint
13431            // must fail on authority before AlreadyCheckpointed can succeed.
13432            let authority = second.inspect_actor(&child.id).await.unwrap();
13433            let leases = inbox
13434                .inspect_owned_leases(&child.id, 2, chrono::Utc::now())
13435                .await
13436                .unwrap();
13437            let delayed = inbox
13438                .checkpoint_actor_input(bamboo_storage::ActorInputCheckpoint {
13439                    fence: old.fence(),
13440                    expected_created_at: child.created_at,
13441                    expected_conversation_generation: child.conversation_generation_observation(),
13442                    claim: claim.clone(),
13443                    expected_messages: child.messages.clone(),
13444                    expected_provider_transcript: child.provider_transcript.clone(),
13445                    expected_admission: child.session_inbox_admission().cloned(),
13446                })
13447                .await;
13448            assert!(matches!(
13449                delayed,
13450                Err(bamboo_storage::ActorInputCheckpointError::Actor(
13451                    bamboo_domain::ActorDirectoryError::StaleFence
13452                ))
13453            ));
13454            assert_eq!(std::fs::read(directory.join("session.json")).unwrap(), main);
13455            assert_eq!(second.inspect_actor(&child.id).await.unwrap(), authority);
13456            assert_eq!(
13457                inbox
13458                    .inspect_owned_leases(&child.id, 2, chrono::Utc::now())
13459                    .await
13460                    .unwrap(),
13461                leases
13462            );
13463            assert!(!inbox.was_admitted(&child.id, &envelope.id).await.unwrap());
13464            assert_eq!(
13465                serde_json::to_value(&child.messages).unwrap(),
13466                serde_json::to_value(&canonical.messages).unwrap()
13467            );
13468            let mut worker: Vec<bamboo_agent_core::Message> = messages
13469                .iter()
13470                .cloned()
13471                .map(|v| serde_json::from_value(v).unwrap())
13472                .collect();
13473            let target = worker.last_mut().unwrap();
13474            assert!(target
13475                .metadata
13476                .as_ref()
13477                .unwrap()
13478                .get("_bamboo_owned_input_checkpoint")
13479                .is_none());
13480            assert!(bamboo_domain::is_matching_session_message(
13481                target, &envelope
13482            ));
13483            let mut wanted = canonical.messages.last().unwrap().clone();
13484            wanted
13485                .metadata
13486                .as_mut()
13487                .unwrap()
13488                .as_object_mut()
13489                .unwrap()
13490                .remove("_bamboo_owned_input_checkpoint");
13491            assert_eq!(
13492                serde_json::to_value(target).unwrap(),
13493                serde_json::to_value(wanted).unwrap()
13494            );
13495            let run = RunSpec {
13496                assignment: "original task".into(),
13497                logical_session: Some(LogicalSessionIdentity {
13498                    session_id: child.id.clone(),
13499                    parent_session_id: child.parent_session_id.clone(),
13500                    root_session_id: child.root_session_id.clone(),
13501                    creation: Some(bamboo_subagent::proto::ChildCreationIdentity {
13502                        created_at: child.created_at,
13503                        spawn_depth: child.spawn_depth,
13504                    }),
13505                }),
13506                project_id: None,
13507                reasoning_effort: None,
13508                permission_policy: Some(PermissionPolicyContext {
13509                    revision: policy.policy_revision(),
13510                    requested_mode: "default".into(),
13511                    effective_mode: "default".into(),
13512                    bypass_permissions: false,
13513                    auto_approve_permissions: false,
13514                    session_id: child.id.clone(),
13515                    workspace_path: child.workspace.clone(),
13516                    environment_lease: None,
13517                    inherit_session_grants: false,
13518                    policy: serde_json::to_value(policy.to_serializable()).unwrap(),
13519                }),
13520                messages,
13521                activation_run_id: Some("replacement-run".into()),
13522                execution_epoch: 2,
13523                initial_session_messages: vec![delivery.clone()],
13524                secrets: Default::default(),
13525            };
13526            let policy_revision = policy.policy_revision();
13527            let mut expected = Some(expected_default_permission_posture(policy_revision));
13528            let ChildFrame::Event { event } = permission_posture_frame(&child.id, policy_revision)
13529            else {
13530                unreachable!()
13531            };
13532            let (events, mut rx) = mpsc::channel(1);
13533            process_actor_event(
13534                event,
13535                true,
13536                &mut PermissionPostureHandshake::Awaiting,
13537                expected.as_ref(),
13538                Some(&binding),
13539                &mut child,
13540                &events,
13541                &mut ActorEventDisplay::default(),
13542            )
13543            .await
13544            .unwrap();
13545            rx.recv().await.unwrap();
13546            let request = bamboo_subagent::proto::InitialInputReleaseRequest::from_run(
13547                &run,
13548                &delivery,
13549                uuid::Uuid::new_v4().to_string(),
13550            )
13551            .unwrap();
13552            replacement
13553                .release_initial_input(
13554                    &binding,
13555                    &child,
13556                    &run,
13557                    2,
13558                    new_claim,
13559                    &request,
13560                    &mut expected,
13561                    None,
13562                )
13563                .await
13564                .unwrap();
13565            assert!(inbox.was_admitted(&child.id, &envelope.id).await.unwrap());
13566            replacement
13567                .append_reply(&mut child, "RECOVERY_REPLY".into(), &events)
13568                .await
13569                .unwrap();
13570            rx.recv().await.unwrap();
13571            replacement
13572                .finish(ActorActivationFinish::Succeeded)
13573                .await
13574                .unwrap();
13575            let cold = second.load_session(&child.id).await.unwrap().unwrap();
13576            assert_eq!(cold.created_at, canonical.created_at);
13577            assert_eq!(
13578                serde_json::to_value(&cold.messages[..canonical.messages.len()]).unwrap(),
13579                serde_json::to_value(&canonical.messages).unwrap()
13580            );
13581            assert_eq!(cold.messages.last().unwrap().content, "RECOVERY_REPLY");
13582            assert_eq!(
13583                cold.messages
13584                    .iter()
13585                    .filter(|m| m.id == envelope.id.as_str())
13586                    .count(),
13587                1
13588            );
13589        }
13590    }
13591
13592    #[cfg(unix)]
13593    #[tokio::test]
13594    async fn owned_initial_release_missing_only_new_capability_rejects_before_spawn_or_run() {
13595        use bamboo_subagent::provision::*;
13596        use std::os::unix::fs::PermissionsExt;
13597        let temp = tempfile::tempdir().unwrap();
13598        let worker = temp.path().join("mixed-worker");
13599        let spawned = temp.path().join("spawned");
13600        let run = temp.path().join("run");
13601        let probes = temp.path().join("probes");
13602        let fabric = temp.path().join("fabric");
13603        let mut report = WorkerCapabilityReport::current();
13604        report
13605            .capabilities
13606            .retain(|name| name != INITIAL_INPUT_RELEASE_WORKER_CAPABILITY);
13607        assert!(!report.supports(INITIAL_INPUT_RELEASE_WORKER_CAPABILITY));
13608        let report = serde_json::to_string(&report).unwrap();
13609        std::fs::write(&worker, format!(
13610            "#!/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"
13611        )).unwrap();
13612        std::fs::set_permissions(&worker, std::fs::Permissions::from_mode(0o700)).unwrap();
13613        let args = vec![
13614            spawned.to_str().unwrap().into(),
13615            run.to_str().unwrap().into(),
13616            probes.to_str().unwrap().into(),
13617        ];
13618        // Real subprocess probes prove the rejection cannot be an earlier missing bit.
13619        for capability in [
13620            REQUIRED_CHILD_CONTEXT_WORKER_CAPABILITY,
13621            CHILD_CREATION_IDENTITY_WORKER_CAPABILITY,
13622            NATIVE_TOOL_CEILING_WORKER_CAPABILITY,
13623            TYPED_READ_ONLY_WORKER_CAPABILITY,
13624        ] {
13625            bamboo_subagent::fleet::require_worker_capability(&worker, &args, capability)
13626                .await
13627                .unwrap();
13628        }
13629        let runner = ActorChildRunner::new(
13630            "mixed".into(),
13631            worker,
13632            args,
13633            fabric.clone(),
13634            ExecutorSpec::BambooRuntime,
13635            vec![],
13636            "openai".into(),
13637            1,
13638        );
13639        // The actual Host method remains mandatory before claim/provision/Run.
13640        let error = runner
13641            .require_initial_input_release_worker()
13642            .await
13643            .unwrap_err();
13644        assert!(error
13645            .to_string()
13646            .contains("owned initial release worker capability unconfirmed"));
13647        assert_eq!(std::fs::read_to_string(probes).unwrap().lines().count(), 5);
13648        assert!(!spawned.exists() && !run.exists() && !fabric.exists());
13649        assert!(runner.pool.lock().await.is_empty());
13650    }
13651
13652    #[tokio::test]
13653    async fn ordinary_glob_only_observes_own_initialization_signals() {
13654        let temp = tempfile::tempdir().unwrap();
13655        let home = temp.path().canonicalize().unwrap();
13656        let store = bamboo_storage::SessionStoreV2::new(home.clone())
13657            .await
13658            .unwrap();
13659        let parent = Session::new("glob-standard-parent", "model");
13660        store.save_session(&parent).await.unwrap();
13661        let child = Session::new_child_of("glob-standard-child", &parent, "model", "one Glob");
13662        store.save_session(&child).await.unwrap();
13663        let directory = home.join(store.resolve_rel_path(&child.id).await.unwrap());
13664        // An unrelated bad child would reject a full tree snapshot, but is not
13665        // part of this own-session no-replay observation.
13666        let unrelated = directory.parent().unwrap().join("unrelated");
13667        std::fs::create_dir(&unrelated).unwrap();
13668        std::fs::write(unrelated.join("session.json"), b"invalid").unwrap();
13669        require_unowned_glob(&store, &child).await.unwrap();
13670        for name in ["actor-authority.json", "actor-authority.initialized.json"] {
13671            let path = directory.join(name);
13672            std::fs::write(&path, b"invalid prior signal").unwrap();
13673            assert!(require_unowned_glob(&store, &child).await.is_err());
13674            std::fs::remove_file(&path).unwrap();
13675        }
13676        #[cfg(unix)]
13677        {
13678            let marker = directory.join("actor-authority.initialized.json");
13679            std::os::unix::fs::symlink(directory.join("missing"), &marker).unwrap();
13680            assert!(require_unowned_glob(&store, &child).await.is_err());
13681        }
13682    }
13683
13684    #[test]
13685    fn owned_glob_collector_requires_actual_ordered_tool_trace_and_typed_tail() {
13686        let call = bamboo_domain::ToolCall {
13687            id: "glob".into(),
13688            tool_type: "function".into(),
13689            function: bamboo_domain::FunctionCall {
13690                name: "Glob".into(),
13691                arguments: r#"{"pattern":"marker.txt"}"#.into(),
13692            },
13693        };
13694        let messages = vec![
13695            bamboo_agent_core::Message::assistant("", Some(vec![call])),
13696            bamboo_agent_core::Message::tool_result("glob", "marker.txt"),
13697            bamboo_agent_core::Message::assistant("done", None),
13698        ];
13699        let start = serde_json::to_value(AgentEvent::ToolStart {
13700            tool_call_id: "glob".into(),
13701            tool_name: "Glob".into(),
13702            arguments: serde_json::json!({"pattern":"marker.txt"}),
13703        })
13704        .unwrap();
13705        let complete = serde_json::to_value(AgentEvent::ToolComplete {
13706            tool_call_id: "glob".into(),
13707            result: bamboo_domain::ToolResult::text(true, "marker.txt"),
13708        })
13709        .unwrap();
13710        let tail =
13711            serde_json::to_value(bamboo_subagent::proto::ReadOnlyActorTranscript::Complete {
13712                messages: messages
13713                    .iter()
13714                    .map(serde_json::to_value)
13715                    .collect::<Result<_, _>>()
13716                    .unwrap(),
13717            })
13718            .unwrap();
13719        let mut ordered = ReadOnlyActorCollector::default();
13720        assert!(ordered.event(&start).unwrap());
13721        assert!(ordered.event(&complete).unwrap());
13722        assert!(!ordered.event(&tail).unwrap());
13723        assert!(ordered.event(&tail).is_err(), "duplicate suffix");
13724        assert_eq!(ordered.finish(Some("done")).unwrap().len(), 3);
13725        assert!(ReadOnlyActorCollector::default()
13726            .finish(Some("done"))
13727            .is_err());
13728        let mut unfinished = ReadOnlyActorCollector::default();
13729        unfinished.event(&start).unwrap();
13730        unfinished.event(&complete).unwrap();
13731        assert!(unfinished.finish(Some("done")).is_err(), "missing suffix");
13732        let mut closed = tail.clone();
13733        closed["grant"] = true.into();
13734        assert!(
13735            serde_json::from_value::<bamboo_subagent::proto::ReadOnlyActorTranscript>(closed)
13736                .is_err()
13737        );
13738        let mut oversized = tail.clone();
13739        oversized["messages"][2]["content"] = "x".repeat(64 * 1024).into();
13740        assert!(
13741            serde_json::from_value::<bamboo_subagent::proto::ReadOnlyActorTranscript>(oversized)
13742                .unwrap()
13743                .validate()
13744                .is_err()
13745        );
13746        assert!(ReadOnlyActorCollector::default().event(&complete).is_err());
13747        assert!(ReadOnlyActorCollector::default().event(&tail).is_err());
13748        let mut bad = start.clone();
13749        bad["tool_name"] = "Bash".into();
13750        assert!(ReadOnlyActorCollector::default().event(&bad).is_err());
13751        let mut duplicate = ReadOnlyActorCollector::default();
13752        duplicate.event(&start).unwrap();
13753        assert!(duplicate.event(&start).is_err());
13754        let hidden = serde_json::to_value(AgentEvent::ReasoningToken {
13755            content: "hidden".into(),
13756        })
13757        .unwrap();
13758        assert!(ReadOnlyActorCollector::default().event(&hidden).is_err());
13759        let mut mismatched = ReadOnlyActorCollector::default();
13760        mismatched.event(&start).unwrap();
13761        let mut bad = complete.clone();
13762        bad["tool_call_id"] = "foreign".into();
13763        assert!(mismatched.event(&bad).is_err());
13764        mismatched.event(&complete).unwrap();
13765        let mut changed = tail.clone();
13766        changed["messages"][1]["content"] = "fabricated".into();
13767        assert!(mismatched.event(&changed).is_err());
13768    }
13769
13770    #[cfg(unix)]
13771    #[tokio::test]
13772    async fn owned_initial_release_requires_current_prefix_ack_and_same_receipt() {
13773        let temp = tempfile::tempdir().unwrap();
13774        let store = Arc::new(
13775            bamboo_storage::SessionStoreV2::new(temp.path().into())
13776                .await
13777                .unwrap(),
13778        );
13779        let parent = Session::new("release-host-parent", "model");
13780        store.save_session(&parent).await.unwrap();
13781        let mut child = Session::new_child_of("release-host-child", &parent, "model", "assignment");
13782        store.save_session(&child).await.unwrap();
13783        let inbox: Arc<dyn SessionInboxPort> = Arc::new(bamboo_storage::FileSessionInbox::new(
13784            store.clone(),
13785            bamboo_domain::SessionInboxLimits::default(),
13786        ));
13787        let binding = actor_binding(
13788            store.clone(),
13789            inbox,
13790            Arc::new(bamboo_storage::LockedSessionStore::new(store.clone())),
13791        );
13792        let _owner = binding
13793            .router
13794            .register_run(&child.id, "release-host-run")
13795            .await
13796            .unwrap();
13797        let policy = Arc::new(bamboo_tools::permission::PermissionConfig::new());
13798        let revision = policy.policy_revision();
13799        bind_local_control_plane(store.as_ref(), &child.id, policy.as_ref()).await;
13800        let activation = PlainActorActivation::start(
13801            store.clone(),
13802            &child,
13803            &binding,
13804            Some("release-host-run"),
13805            Some(policy.clone()),
13806            true,
13807            Some("probed-physical-worker"),
13808        )
13809        .await
13810        .unwrap();
13811        let envelope = bamboo_domain::SessionMessageEnvelope::user_input(&child.id, "RELEASE_ONCE");
13812        let receipt = activation.input_inbox.deliver(&envelope).await.unwrap();
13813        activation
13814            .input_inbox
13815            .mark_activation_eligible(
13816                &child.id,
13817                receipt.generation,
13818                bamboo_domain::SessionActivationPolicy::InterruptSpecificWait,
13819            )
13820            .await
13821            .unwrap();
13822        let claim = activation
13823            .claim_input(&binding, &child, "release-host-run")
13824            .await
13825            .unwrap()
13826            .unwrap();
13827        let (prefix, delivery) = activation
13828            .checkpoint_input(&mut child, "release-host-run", &claim)
13829            .await
13830            .unwrap();
13831        let mut expected = Some(expected_default_permission_posture(revision));
13832        let ChildFrame::Event { event } = permission_posture_frame(&child.id, revision) else {
13833            unreachable!()
13834        };
13835        let (events, mut events_rx) = mpsc::channel(1);
13836        process_actor_event(
13837            event,
13838            true,
13839            &mut PermissionPostureHandshake::Awaiting,
13840            expected.as_ref(),
13841            Some(&binding),
13842            &mut child,
13843            &events,
13844            &mut ActorEventDisplay::default(),
13845        )
13846        .await
13847        .unwrap();
13848        events_rx.recv().await.unwrap();
13849        let run = RunSpec {
13850            assignment: "assignment".into(),
13851            logical_session: Some(LogicalSessionIdentity {
13852                session_id: child.id.clone(),
13853                parent_session_id: child.parent_session_id.clone(),
13854                root_session_id: child.root_session_id.clone(),
13855                creation: Some(bamboo_subagent::proto::ChildCreationIdentity {
13856                    created_at: child.created_at,
13857                    spawn_depth: child.spawn_depth,
13858                }),
13859            }),
13860            project_id: None,
13861            reasoning_effort: None,
13862            permission_policy: Some(bamboo_subagent::proto::PermissionPolicyContext {
13863                revision,
13864                requested_mode: "default".into(),
13865                effective_mode: "default".into(),
13866                bypass_permissions: false,
13867                auto_approve_permissions: false,
13868                session_id: child.id.clone(),
13869                workspace_path: child.workspace.clone(),
13870                environment_lease: None,
13871                inherit_session_grants: false,
13872                policy: serde_json::to_value(policy.to_serializable()).unwrap(),
13873            }),
13874            messages: prefix,
13875            activation_run_id: Some("release-host-run".into()),
13876            execution_epoch: 9,
13877            initial_session_messages: vec![delivery.clone()],
13878            secrets: Default::default(),
13879        };
13880        let request = bamboo_subagent::proto::InitialInputReleaseRequest::from_run(
13881            &run,
13882            &delivery,
13883            uuid::Uuid::new_v4().to_string(),
13884        )
13885        .unwrap();
13886        let mut stale = child.clone();
13887        stale.add_message(bamboo_agent_core::Message::user("not canonical"));
13888        assert!(activation
13889            .release_initial_input(
13890                &binding,
13891                &stale,
13892                &run,
13893                9,
13894                &claim,
13895                &request,
13896                &mut expected,
13897                None
13898            )
13899            .await
13900            .is_err());
13901        assert!(!activation
13902            .input_inbox
13903            .was_admitted(&child.id, &envelope.id)
13904            .await
13905            .unwrap());
13906        // Keep receipt lookup readable so Already reaches the real ACK write.
13907        // Unix root can bypass these modes; this FS fault needs an unprivileged user.
13908        use std::os::unix::fs::PermissionsExt;
13909        struct RestorePermissions(std::path::PathBuf, std::fs::Permissions);
13910        impl Drop for RestorePermissions {
13911            fn drop(&mut self) {
13912                let _ = std::fs::set_permissions(&self.0, self.1.clone());
13913            }
13914        }
13915        let admitted = store
13916            .bamboo_home_dir()
13917            .join(store.resolve_rel_path(&child.id).await.unwrap())
13918            .join("inbox/admitted");
13919        std::fs::create_dir_all(&admitted).unwrap();
13920        let restore = RestorePermissions(
13921            admitted.clone(),
13922            std::fs::metadata(&admitted).unwrap().permissions(),
13923        );
13924        std::fs::set_permissions(&admitted, std::fs::Permissions::from_mode(0o555)).unwrap();
13925        let verified = activation
13926            .input_inbox
13927            .checkpoint_actor_input(bamboo_storage::ActorInputCheckpoint {
13928                fence: activation.fence.clone(),
13929                expected_created_at: child.created_at,
13930                expected_conversation_generation: child.conversation_generation_observation(),
13931                claim: claim.clone(),
13932                expected_messages: child.messages.clone(),
13933                expected_provider_transcript: child.provider_transcript.clone(),
13934                expected_admission: child.session_inbox_admission().cloned(),
13935            })
13936            .await
13937            .unwrap();
13938        assert_eq!(
13939            verified.status,
13940            bamboo_storage::ActorInputCheckpointStatus::AlreadyCheckpointed
13941        );
13942        let error = activation
13943            .release_initial_input(
13944                &binding,
13945                &child,
13946                &run,
13947                9,
13948                &claim,
13949                &request,
13950                &mut expected,
13951                None,
13952            )
13953            .await
13954            .unwrap_err();
13955        std::fs::set_permissions(&admitted, restore.1.clone()).unwrap();
13956        drop(restore);
13957        assert!(
13958            error.to_string().contains("ACK unresolved"),
13959            "actual release error (<=512B): {}",
13960            error.to_string().chars().take(128).collect::<String>()
13961        );
13962        let release = activation
13963            .release_initial_input(
13964                &binding,
13965                &child,
13966                &run,
13967                9,
13968                &claim,
13969                &request,
13970                &mut expected,
13971                None,
13972            )
13973            .await
13974            .unwrap();
13975        assert!(activation
13976            .input_inbox
13977            .was_admitted(&child.id, &envelope.id)
13978            .await
13979            .unwrap());
13980        assert_eq!(
13981            release.expires_at,
13982            activation
13983                .actor_deadline
13984                .min(claim.lease.expires_at)
13985                .min(activation.input_deadline)
13986        );
13987        assert_eq!(
13988            activation
13989                .release_initial_input(
13990                    &binding,
13991                    &child,
13992                    &run,
13993                    9,
13994                    &claim,
13995                    &request,
13996                    &mut expected,
13997                    Some(&release)
13998                )
13999                .await
14000                .unwrap(),
14001            release
14002        );
14003        let mut wrong = request.clone();
14004        wrong.nonce = uuid::Uuid::new_v4().to_string();
14005        assert!(activation
14006            .release_initial_input(
14007                &binding,
14008                &child,
14009                &run,
14010                9,
14011                &claim,
14012                &wrong,
14013                &mut expected,
14014                Some(&release)
14015            )
14016            .await
14017            .is_err());
14018        assert!(activation
14019            .release_initial_input(
14020                &binding,
14021                &child,
14022                &run,
14023                10,
14024                &claim,
14025                &request,
14026                &mut expected,
14027                Some(&release)
14028            )
14029            .await
14030            .is_err());
14031        // A second real checkpoint/release reuses the activation, not its old epoch or nonce.
14032        let deadlines = (
14033            activation.fence.clone(),
14034            activation.actor_deadline,
14035            activation.input_deadline,
14036        );
14037        let second = bamboo_domain::SessionMessageEnvelope::user_input(&child.id, "RELEASE_TWICE");
14038        let receipt = activation.input_inbox.deliver(&second).await.unwrap();
14039        activation
14040            .input_inbox
14041            .mark_activation_eligible(
14042                &child.id,
14043                receipt.generation,
14044                bamboo_domain::SessionActivationPolicy::InterruptSpecificWait,
14045            )
14046            .await
14047            .unwrap();
14048        let mut link = ConfirmationSequenceLink {
14049            frames: VecDeque::new(),
14050            sent: Vec::new(),
14051        };
14052        let epochs = AtomicU64::new(9);
14053        let (claim2, epoch2) = activation
14054            .continue_input(
14055                &mut link,
14056                &binding,
14057                &mut child,
14058                "release-host-run",
14059                Some((&run, &epochs)),
14060                &mut expected,
14061                "SECOND_REPLY".into(),
14062                &events,
14063            )
14064            .await
14065            .unwrap()
14066            .unwrap();
14067        events_rx.recv().await.unwrap();
14068        assert_eq!(epoch2, 10);
14069        let ParentFrame::Run(next) = &link.sent[0] else {
14070            panic!("second correlated Run")
14071        };
14072        assert_eq!(next.logical_session, run.logical_session);
14073        assert_eq!(next.activation_run_id, run.activation_run_id);
14074        assert_eq!(next.initial_session_messages.len(), 1);
14075        assert!(next
14076            .messages
14077            .iter()
14078            .any(|m| m["id"] == envelope.id.as_str()));
14079        assert!(next.messages.iter().any(|m| m["content"] == "SECOND_REPLY"));
14080        assert!(!next.messages.iter().any(|m| m["id"] == second.id.as_str()));
14081        let request2 = bamboo_subagent::proto::InitialInputReleaseRequest::from_run(
14082            next,
14083            &next.initial_session_messages[0],
14084            uuid::Uuid::new_v4().to_string(),
14085        )
14086        .unwrap();
14087        assert_ne!(request2.nonce, request.nonce);
14088        macro_rules! release_second {
14089            ($candidate:expr, $request:expr, $epoch:expr, $previous:expr) => {
14090                activation
14091                    .release_initial_input(
14092                        &binding,
14093                        $candidate,
14094                        next,
14095                        $epoch,
14096                        &claim2,
14097                        $request,
14098                        &mut expected,
14099                        $previous,
14100                    )
14101                    .await
14102            };
14103        }
14104        assert!(release_second!(&child, &request, epoch2, None).is_err());
14105        assert!(release_second!(&child, &request2, 9, None).is_err());
14106        let mut stale2 = child.clone();
14107        stale2.add_message(bamboo_agent_core::Message::user("stale second prefix"));
14108        assert!(release_second!(&stale2, &request2, epoch2, None).is_err());
14109        assert!(!activation
14110            .input_inbox
14111            .was_admitted(&child.id, &second.id)
14112            .await
14113            .unwrap());
14114        let release2 = release_second!(&child, &request2, epoch2, None).unwrap();
14115        assert_eq!(
14116            release_second!(&child, &request2, epoch2, Some(&release2)).unwrap(),
14117            release2
14118        );
14119        let mut wrong2 = request2.clone();
14120        wrong2.nonce = request.nonce.clone();
14121        assert!(release_second!(&child, &wrong2, epoch2, Some(&release2)).is_err());
14122        assert!(activation
14123            .input_inbox
14124            .was_admitted(&child.id, &second.id)
14125            .await
14126            .unwrap());
14127        let expired = bamboo_domain::SessionMessageEnvelope::user_input(&child.id, "EXPIRED_INPUT");
14128        let receipt = activation.input_inbox.deliver(&expired).await.unwrap();
14129        activation
14130            .input_inbox
14131            .mark_activation_eligible(
14132                &child.id,
14133                receipt.generation,
14134                bamboo_domain::SessionActivationPolicy::InterruptSpecificWait,
14135            )
14136            .await
14137            .unwrap();
14138        let now = chrono::Utc::now();
14139        let expired_claim = activation
14140            .input_inbox
14141            .claim_owned(
14142                &child.id,
14143                1,
14144                Some("release-host-run"),
14145                &SessionInboxLeaseRequest {
14146                    consumer: activation.input_consumer.clone(),
14147                    now,
14148                    duration: chrono::Duration::seconds(1),
14149                },
14150            )
14151            .await
14152            .unwrap()
14153            .pop()
14154            .unwrap();
14155        let (prefix, delivery3) = activation
14156            .checkpoint_input(&mut child, "release-host-run", &expired_claim)
14157            .await
14158            .unwrap();
14159        let mut expired_run = next.clone();
14160        expired_run.execution_epoch = 11;
14161        expired_run.messages = prefix;
14162        expired_run.initial_session_messages = vec![delivery3.clone()];
14163        let expired_request = bamboo_subagent::proto::InitialInputReleaseRequest::from_run(
14164            &expired_run,
14165            &delivery3,
14166            uuid::Uuid::new_v4().to_string(),
14167        )
14168        .unwrap();
14169        tokio::time::sleep(
14170            (expired_claim.lease.expires_at - chrono::Utc::now())
14171                .to_std()
14172                .unwrap_or_default(),
14173        )
14174        .await;
14175        assert!(chrono::Utc::now() >= expired_claim.lease.expires_at);
14176        assert!(activation
14177            .release_initial_input(
14178                &binding,
14179                &child,
14180                &expired_run,
14181                11,
14182                &expired_claim,
14183                &expired_request,
14184                &mut expected,
14185                None
14186            )
14187            .await
14188            .is_err());
14189        assert!(!activation
14190            .input_inbox
14191            .was_admitted(&child.id, &expired.id)
14192            .await
14193            .unwrap());
14194        assert_eq!(
14195            (
14196                activation.fence.clone(),
14197                activation.actor_deadline,
14198                activation.input_deadline
14199            ),
14200            deadlines
14201        );
14202        assert_eq!(
14203            link.sent.len(),
14204            1,
14205            "no dispatch from rejected release checks"
14206        );
14207        let cold = store.load_session(&child.id).await.unwrap().unwrap();
14208        for id in [&envelope.id, &second.id] {
14209            assert_eq!(
14210                cold.messages.iter().filter(|m| m.id == id.as_str()).count(),
14211                1
14212            );
14213            assert!(cold.session_inbox_admission().unwrap().contains(id));
14214        }
14215        // Mutate the actual captured policy Arc; cached ACK/release is not a fresh grant.
14216        policy.deny_scoped_session_permission(
14217            &child.id,
14218            bamboo_tools::permission::PermissionType::ExecuteCommand,
14219            "blocked-action",
14220        );
14221        assert!(release_second!(&child, &request2, epoch2, Some(&release2)).is_err());
14222        assert!(activation
14223            .input_inbox
14224            .was_admitted(&child.id, &second.id)
14225            .await
14226            .unwrap());
14227        activation
14228            .finish(ActorActivationFinish::Failed)
14229            .await
14230            .unwrap();
14231        // Terminal ACK can be idempotent; stopped/expired Actor cannot grant.
14232        activation
14233            .input_inbox
14234            .ack_owned(&child.id, &claim, chrono::Utc::now())
14235            .await
14236            .unwrap();
14237        assert!(activation
14238            .release_initial_input(
14239                &binding,
14240                &child,
14241                &run,
14242                9,
14243                &claim,
14244                &request,
14245                &mut expected,
14246                Some(&release)
14247            )
14248            .await
14249            .is_err());
14250        let cold = bamboo_storage::SessionStoreV2::new(temp.path().into())
14251            .await
14252            .unwrap()
14253            .load_session(&child.id)
14254            .await
14255            .unwrap()
14256            .unwrap();
14257        assert_eq!(
14258            cold.messages
14259                .iter()
14260                .filter(|m| m.id == envelope.id.as_str())
14261                .count(),
14262            1
14263        );
14264    }
14265
14266    #[tokio::test]
14267    async fn owned_actor_input_checkpoint_precedes_dispatch_and_exact_ack() {
14268        for reject_checkpoint in [false, true] {
14269            let temp = tempfile::tempdir().unwrap();
14270            let store = Arc::new(
14271                bamboo_storage::SessionStoreV2::new(temp.path().into())
14272                    .await
14273                    .unwrap(),
14274            );
14275            let parent = Session::new("owned-input-parent", "model");
14276            store.save_session(&parent).await.unwrap();
14277            let mut child =
14278                Session::new_child_of("owned-input-child", &parent, "model", "one plain task");
14279            store.save_session(&child).await.unwrap();
14280            let policy = Arc::new(bamboo_tools::permission::PermissionConfig::new());
14281            let raw_storage: Arc<dyn Storage> = store.clone();
14282            let locked = Arc::new(bamboo_storage::LockedSessionStore::new(raw_storage));
14283            let inbox: Arc<dyn SessionInboxPort> = Arc::new(bamboo_storage::FileSessionInbox::new(
14284                store.clone(),
14285                bamboo_domain::SessionInboxLimits::default(),
14286            ));
14287            let binding = actor_binding(store.clone(), inbox, locked);
14288            let run_id = "owned-input-run";
14289            let _registration = binding
14290                .router
14291                .register_run(&child.id, run_id)
14292                .await
14293                .unwrap();
14294            assert!(PlainActorActivation::start(
14295                store.clone(),
14296                &child,
14297                &binding,
14298                Some(run_id),
14299                None,
14300                true,
14301                None,
14302            )
14303            .await
14304            .is_err());
14305            assert!(PlainActorActivation::start(
14306                store.clone(),
14307                &child,
14308                &binding,
14309                Some(run_id),
14310                Some(policy.clone()),
14311                true,
14312                None,
14313            )
14314            .await
14315            .is_err());
14316            assert_eq!(
14317                store
14318                    .inspect_actor(&child.id)
14319                    .await
14320                    .unwrap()
14321                    .actor
14322                    .current_attempt,
14323                0
14324            );
14325            bind_local_control_plane(store.as_ref(), &child.id, policy.as_ref()).await;
14326            let activation = PlainActorActivation::start(
14327                store.clone(),
14328                &child,
14329                &binding,
14330                Some(run_id),
14331                Some(policy.clone()),
14332                true,
14333                None,
14334            )
14335            .await
14336            .unwrap();
14337            let envelope =
14338                bamboo_domain::SessionMessageEnvelope::user_input(&child.id, "CORRECTION_ONCE");
14339            let receipt = activation.input_inbox.deliver(&envelope).await.unwrap();
14340            activation
14341                .input_inbox
14342                .mark_activation_eligible(
14343                    &child.id,
14344                    receipt.generation,
14345                    bamboo_domain::SessionActivationPolicy::InterruptSpecificWait,
14346                )
14347                .await
14348                .unwrap();
14349            let mut link = ConfirmationSequenceLink {
14350                frames: VecDeque::new(),
14351                sent: Vec::new(),
14352            };
14353            let policy_revision = policy.policy_revision();
14354            let run = RunSpec {
14355                assignment: "one plain task".into(),
14356                logical_session: None,
14357                project_id: None,
14358                reasoning_effort: None,
14359                permission_policy: Some(PermissionPolicyContext {
14360                    revision: policy_revision,
14361                    requested_mode: "default".into(),
14362                    effective_mode: "default".into(),
14363                    bypass_permissions: false,
14364                    auto_approve_permissions: false,
14365                    session_id: child.id.clone(),
14366                    workspace_path: child.workspace.clone(),
14367                    environment_lease: None,
14368                    inherit_session_grants: false,
14369                    policy: serde_json::to_value(policy.to_serializable()).unwrap(),
14370                }),
14371                messages: vec![],
14372                activation_run_id: Some(run_id.into()),
14373                execution_epoch: 1,
14374                initial_session_messages: vec![],
14375                secrets: Default::default(),
14376            };
14377            let epochs = AtomicU64::new(1);
14378            let (tx, mut rx) = mpsc::channel(1);
14379            let mut expected = Some(expected_default_permission_posture(policy_revision));
14380            let ChildFrame::Event { event } = permission_posture_frame(&child.id, policy_revision)
14381            else {
14382                unreachable!()
14383            };
14384            process_actor_event(
14385                event,
14386                true,
14387                &mut PermissionPostureHandshake::Awaiting,
14388                expected.as_ref(),
14389                Some(&binding),
14390                &mut child,
14391                &tx,
14392                &mut ActorEventDisplay::default(),
14393            )
14394            .await
14395            .unwrap();
14396            rx.recv().await.unwrap();
14397            let first_audit =
14398                bamboo_domain::PermissionAuditSnapshot::from_metadata(&child.metadata)
14399                    .unwrap()
14400                    .audit_revision;
14401            if reject_checkpoint {
14402                // A real bounded publication channel blocks AFTER first-reply
14403                // commit, so change actual Actor authority before checkpoint.
14404                tx.send(AgentEvent::Complete {
14405                    usage: Default::default(),
14406                })
14407                .await
14408                .unwrap();
14409            }
14410            let mut operation = activation.continue_input(
14411                &mut link,
14412                &binding,
14413                &mut child,
14414                run_id,
14415                Some((&run, &epochs)),
14416                &mut expected,
14417                "FIRST_PLAIN_REPLY".into(),
14418                &tx,
14419            );
14420            if reject_checkpoint {
14421                tokio::select! {
14422                    result = &mut operation => panic!("publication barrier released early: {result:?}"),
14423                    _ = async {
14424                        tokio::time::timeout(Duration::from_secs(5), async {
14425                            loop {
14426                                let actual = store.load_session("owned-input-child").await.unwrap().unwrap();
14427                                if actual.messages.iter().any(|m| m.content == "FIRST_PLAIN_REPLY") { break; }
14428                                tokio::task::yield_now().await;
14429                            }
14430                        }).await.unwrap();
14431                        activation.finish(ActorActivationFinish::Failed).await.unwrap();
14432                        rx.recv().await.unwrap();
14433                    } => {}
14434                }
14435            }
14436            let result = operation.await;
14437            if reject_checkpoint {
14438                assert!(result
14439                    .unwrap_err()
14440                    .to_string()
14441                    .contains("checkpoint unconfirmed"));
14442                assert!(
14443                    link.sent.is_empty(),
14444                    "zero worker/provider correction dispatch"
14445                );
14446                let durable = store.load_session(&child.id).await.unwrap().unwrap();
14447                assert!(durable
14448                    .messages
14449                    .iter()
14450                    .any(|m| m.content == "FIRST_PLAIN_REPLY"));
14451                assert!(!durable
14452                    .messages
14453                    .iter()
14454                    .any(|m| m.id == envelope.id.as_str()));
14455                assert!(!activation
14456                    .input_inbox
14457                    .was_admitted(&child.id, &envelope.id)
14458                    .await
14459                    .unwrap());
14460                assert_eq!(
14461                    activation
14462                        .input_inbox
14463                        .inspect(&child.id)
14464                        .await
14465                        .unwrap()
14466                        .claimed,
14467                    1
14468                );
14469                continue;
14470            }
14471            let (claim, epoch) = result.unwrap().unwrap();
14472            rx.recv().await.unwrap(); // Actual first-reply publication.
14473            assert_eq!(epoch, 2);
14474            assert_eq!(link.sent.len(), 1);
14475            let ParentFrame::Run(next) = &link.sent[0] else {
14476                panic!("correlated continuation Run")
14477            };
14478            assert_eq!(next.execution_epoch, 2);
14479            assert_eq!(next.activation_run_id.as_deref(), Some(run_id));
14480            assert!(next
14481                .messages
14482                .iter()
14483                .any(|m| m["content"] == "FIRST_PLAIN_REPLY"));
14484            assert!(!next
14485                .messages
14486                .iter()
14487                .any(|m| m["id"] == envelope.id.as_str()));
14488            let delivery = &next.initial_session_messages[0];
14489            assert_eq!(delivery.envelope, envelope);
14490            let durable = store.load_session(&child.id).await.unwrap().unwrap();
14491            assert!(durable
14492                .session_inbox_admission()
14493                .unwrap()
14494                .contains(&envelope.id));
14495            assert_eq!(
14496                durable
14497                    .messages
14498                    .iter()
14499                    .filter(|m| m.id == envelope.id.as_str())
14500                    .count(),
14501                1
14502            );
14503            assert!(
14504                !activation
14505                    .input_inbox
14506                    .was_admitted(&child.id, &envelope.id)
14507                    .await
14508                    .unwrap(),
14509                "checkpoint is not a worker admission receipt"
14510            );
14511            let mut confirmation = admission_confirmation(&child.id, &claim.claim, run_id);
14512            confirmation.canonical_claim_generation += 1;
14513            assert!(activation
14514                .confirm_input(&binding, &child, run_id, &claim, &confirmation)
14515                .await
14516                .is_err());
14517            assert!(!activation
14518                .input_inbox
14519                .was_admitted(&child.id, &envelope.id)
14520                .await
14521                .unwrap());
14522            confirmation.canonical_claim_generation = claim.claim.generation;
14523            activation
14524                .confirm_input(&binding, &child, run_id, &claim, &confirmation)
14525                .await
14526                .unwrap();
14527            assert!(activation
14528                .input_inbox
14529                .was_admitted(&child.id, &envelope.id)
14530                .await
14531                .unwrap());
14532            assert_eq!(
14533                expected.as_ref().unwrap().expected_audit_revision,
14534                Some(first_audit)
14535            );
14536            let ChildFrame::Event { event } = permission_posture_frame(&child.id, policy_revision)
14537            else {
14538                unreachable!()
14539            };
14540            process_actor_event(
14541                event,
14542                true,
14543                &mut PermissionPostureHandshake::Awaiting,
14544                expected.as_ref(),
14545                Some(&binding),
14546                &mut child,
14547                &tx,
14548                &mut ActorEventDisplay::default(),
14549            )
14550            .await
14551            .unwrap();
14552            rx.recv().await.unwrap();
14553            assert!(
14554                bamboo_domain::PermissionAuditSnapshot::from_metadata(&child.metadata)
14555                    .unwrap()
14556                    .audit_revision
14557                    > first_audit
14558            );
14559            let reopened = bamboo_storage::SessionStoreV2::new(temp.path().into())
14560                .await
14561                .unwrap();
14562            let cold = reopened.load_session(&child.id).await.unwrap().unwrap();
14563            assert_eq!(
14564                cold.messages
14565                    .iter()
14566                    .filter(|m| m.id == envelope.id.as_str())
14567                    .count(),
14568                1
14569            );
14570            assert!(cold
14571                .session_inbox_admission()
14572                .unwrap()
14573                .contains(&envelope.id));
14574        }
14575    }
14576
14577    fn admission_confirmation(
14578        session_id: &str,
14579        claim: &SessionInboxClaim,
14580        run_id: &str,
14581    ) -> bamboo_subagent::proto::SessionMessageAdmissionConfirmation {
14582        bamboo_subagent::proto::SessionMessageAdmissionConfirmation {
14583            target_session_id: session_id.to_string(),
14584            envelope_id: claim.envelope.id.to_string(),
14585            canonical_claim_generation: claim.generation,
14586            activation_run_id: run_id.to_string(),
14587        }
14588    }
14589
14590    fn expected_default_permission_posture(policy_revision: u64) -> ExpectedPermissionPosture {
14591        ExpectedPermissionPosture {
14592            policy_revision,
14593            resolution: bamboo_domain::PermissionModeResolution {
14594                requested: bamboo_domain::SessionPermissionMode::Default,
14595                effective: bamboo_domain::PermissionMode::Default,
14596            },
14597            expected_audit_revision: None,
14598            executor_mapping: "test_actor:permission_mode=default".to_string(),
14599        }
14600    }
14601
14602    fn permission_posture_frame(session_id: &str, policy_revision: u64) -> ChildFrame {
14603        ChildFrame::Event {
14604            event: serde_json::to_value(AgentEvent::PermissionPostureActivated {
14605                session_id: session_id.to_string(),
14606                policy_revision,
14607                requested_mode: "default".to_string(),
14608                effective_mode: "default".to_string(),
14609                executor_mapping: "test_actor:permission_mode=default".to_string(),
14610            })
14611            .expect("serialize permission posture event"),
14612        }
14613    }
14614
14615    fn actor_event_frame(event: AgentEvent) -> ChildFrame {
14616        ChildFrame::Event {
14617            event: serde_json::to_value(event).expect("serialize actor event"),
14618        }
14619    }
14620
14621    #[test]
14622    fn actor_event_batch_is_fenced_by_session_activation_and_epoch() {
14623        let session = Session::new_child("logical-child", "logical-parent", "model", "child");
14624        let mut batch = ActorEventBatch {
14625            logical_session: Some(LogicalSessionIdentity {
14626                creation: None,
14627                session_id: session.id.clone(),
14628                parent_session_id: session.parent_session_id.clone(),
14629                root_session_id: session.root_session_id.clone(),
14630            }),
14631            activation_id: Some("activation-7".into()),
14632            execution_epoch: 11,
14633            source_node_id: Some("node-a".into()),
14634            source_actor_id: Some("worker-a".into()),
14635            first_seq: 1,
14636            last_seq: 1,
14637            qos: bamboo_subagent::ActorEventQos::Ephemeral,
14638            events: vec![serde_json::json!({"type":"token","content":"ok"})],
14639        };
14640        validate_actor_event_batch(
14641            &batch,
14642            &session,
14643            "logical-parent",
14644            Some("activation-7"),
14645            11,
14646            "worker-a",
14647            None,
14648        )
14649        .unwrap();
14650
14651        let creation = bamboo_subagent::proto::ChildCreationIdentity {
14652            created_at: session.created_at,
14653            spawn_depth: session.spawn_depth,
14654        };
14655        assert!(validate_actor_event_batch(
14656            &batch,
14657            &session,
14658            "logical-parent",
14659            Some("activation-7"),
14660            11,
14661            "worker-a",
14662            Some(&creation)
14663        )
14664        .is_err());
14665        batch.logical_session.as_mut().unwrap().creation = Some(creation.clone());
14666        validate_actor_event_batch(
14667            &batch,
14668            &session,
14669            "logical-parent",
14670            Some("activation-7"),
14671            11,
14672            "worker-a",
14673            Some(&creation),
14674        )
14675        .unwrap();
14676        batch
14677            .logical_session
14678            .as_mut()
14679            .unwrap()
14680            .creation
14681            .as_mut()
14682            .unwrap()
14683            .created_at += chrono::Duration::nanoseconds(1);
14684        assert!(validate_actor_event_batch(
14685            &batch,
14686            &session,
14687            "logical-parent",
14688            Some("activation-7"),
14689            11,
14690            "worker-a",
14691            Some(&creation)
14692        )
14693        .is_err());
14694        batch.logical_session.as_mut().unwrap().creation = None;
14695
14696        batch.execution_epoch = 10;
14697        assert!(validate_actor_event_batch(
14698            &batch,
14699            &session,
14700            "logical-parent",
14701            Some("activation-7"),
14702            11,
14703            "worker-a",
14704            None,
14705        )
14706        .unwrap_err()
14707        .to_string()
14708        .contains("execution epoch"));
14709        batch.execution_epoch = 11;
14710        batch.activation_id = Some("replaced-activation".into());
14711        assert!(validate_actor_event_batch(
14712            &batch,
14713            &session,
14714            "logical-parent",
14715            Some("activation-7"),
14716            11,
14717            "worker-a",
14718            None,
14719        )
14720        .unwrap_err()
14721        .to_string()
14722        .contains("stale activation"));
14723        batch.activation_id = Some("activation-7".into());
14724        batch
14725            .logical_session
14726            .as_mut()
14727            .expect("logical identity")
14728            .session_id = "foreign-child".into();
14729        assert!(validate_actor_event_batch(
14730            &batch,
14731            &session,
14732            "logical-parent",
14733            Some("activation-7"),
14734            11,
14735            "worker-a",
14736            None,
14737        )
14738        .unwrap_err()
14739        .to_string()
14740        .contains("different logical session"));
14741        batch
14742            .logical_session
14743            .as_mut()
14744            .expect("logical identity")
14745            .session_id = session.id.clone();
14746        batch.source_actor_id = Some("worker-b".into());
14747        assert!(validate_actor_event_batch(
14748            &batch,
14749            &session,
14750            "logical-parent",
14751            Some("activation-7"),
14752            11,
14753            "worker-a",
14754            None,
14755        )
14756        .unwrap_err()
14757        .to_string()
14758        .contains("different physical actor"));
14759    }
14760
14761    #[tokio::test]
14762    async fn directory_event_ingress_rejects_old_attempt_and_forged_identity() {
14763        use bamboo_domain::Storage;
14764
14765        let temp = tempfile::tempdir().unwrap();
14766        let store = bamboo_storage::SessionStoreV2::new(temp.path().into())
14767            .await
14768            .unwrap();
14769        let parent = Session::new("event-parent", "model");
14770        let child = Session::new_child_of("event-child", &parent, "model", "task");
14771        store.save_session(&parent).await.unwrap();
14772        store.save_session(&child).await.unwrap();
14773        store.ensure_actor(&child.id).await.unwrap();
14774        let now = chrono::Utc::now();
14775        let claim = |run_id: &str| ActorActivationClaim {
14776            actor_id: child.id.clone(),
14777            run_id: run_id.into(),
14778            lease_owner: "event-owner".into(),
14779            lease_expires_at: now + chrono::Duration::minutes(2),
14780            inbox_generation: 0,
14781            placement_ref: None,
14782            now,
14783        };
14784        let first = store.claim_activation(&claim("run-one")).await.unwrap();
14785        store.start_activation(&first.fence(), now).await.unwrap();
14786        let first_fence = first.fence();
14787        validate_directory_actor_event(&store, &first_fence, &child, Some("run-one"))
14788            .await
14789            .unwrap();
14790        validate_directory_actor_event(&store, &first_fence, &child, None)
14791            .await
14792            .unwrap(); // rolling-upgrade frame uses the same owner
14793        assert!(
14794            validate_directory_actor_event(&store, &first_fence, &child, Some("forged-run"))
14795                .await
14796                .is_err()
14797        );
14798        let mut forged = child.clone();
14799        forged.root_session_id = "foreign-root".into();
14800        assert!(
14801            validate_directory_actor_event(&store, &first_fence, &forged, Some("run-one"))
14802                .await
14803                .is_err()
14804        );
14805
14806        store
14807            .finish_activation(&first_fence, now, ActorActivationFinish::Failed)
14808            .await
14809            .unwrap();
14810        let second = store.claim_activation(&claim("run-two")).await.unwrap();
14811        store.start_activation(&second.fence(), now).await.unwrap();
14812        assert!(
14813            validate_directory_actor_event(&store, &first_fence, &child, Some("run-one"))
14814                .await
14815                .is_err()
14816        );
14817        assert!(
14818            validate_directory_actor_event(&store, &first_fence, &child, None)
14819                .await
14820                .is_err()
14821        );
14822        validate_directory_actor_event(&store, &second.fence(), &child, Some("run-two"))
14823            .await
14824            .unwrap();
14825    }
14826
14827    fn completed_actor_frame() -> ChildFrame {
14828        ChildFrame::Terminal {
14829            status: TerminalStatus::Completed,
14830            result: Some("done".to_string()),
14831            error: None,
14832            final_event_watermark: None,
14833            transcript: Vec::new(),
14834        }
14835    }
14836
14837    async fn drive_permission_handshake_frames(
14838        session_id: &str,
14839        frames: impl IntoIterator<Item = ChildFrame>,
14840        expected: ExpectedPermissionPosture,
14841    ) -> (
14842        crate::runtime::runner::Result<Option<String>>,
14843        Session,
14844        Vec<AgentEvent>,
14845        Vec<ParentFrame>,
14846    ) {
14847        let mut link = ConfirmationSequenceLink {
14848            frames: frames.into_iter().collect(),
14849            sent: Vec::new(),
14850        };
14851        let (event_tx, mut event_rx) = mpsc::channel(128);
14852        let cancel = CancellationToken::new();
14853        let (_live_tx, mut live_rx) = mpsc::unbounded_channel();
14854        let (_delivery_tx, mut delivery_rx) = mpsc::unbounded_channel();
14855        let mut session = Session::new(session_id, "model");
14856        let result = drive(ActorDriveContext {
14857            expected_creation: None,
14858            client: &mut link,
14859            parent_session_id: "permission-parent",
14860            child_session_id: session_id,
14861            child_attempt: 0,
14862            approval_registry: None,
14863            approval_decider: None,
14864            approval_reviewer: None,
14865            escalation_bridge: None,
14866            event_tx: &event_tx,
14867            cancel_token: &cancel,
14868            live_rx: &mut live_rx,
14869            delivery_rx: &mut delivery_rx,
14870            logical_session: &mut session,
14871            expected_permission_posture: Some(expected),
14872            session_inbox_runtime: None,
14873            actor_directory_store: None,
14874            canonical_subagent_tool: None,
14875            ticket_service: None,
14876            activation_run_id: None,
14877            execution_epoch: 0,
14878            expected_source_actor_id: session_id,
14879            initial_inflight_claims: VecDeque::new(),
14880            plain_actor: false,
14881            remote_environment_lease: false,
14882            readonly_output: None,
14883            local_history_tools: None,
14884            local_history_read_only: false,
14885            strict_history_output: None,
14886            plain_input: None,
14887            canonical_activation: None,
14888            canonical_placement_ref: None,
14889            actor_event_observer: None,
14890            plain_run: None,
14891            first_frame_timeout: Some(Duration::from_secs(1)),
14892        })
14893        .await;
14894        let events = std::iter::from_fn(|| event_rx.try_recv().ok()).collect();
14895        (result, session, events, link.sent)
14896    }
14897
14898    #[tokio::test]
14899    async fn actor_permission_handshake_rejects_terminal_without_posture() {
14900        let session_id = "permission-missing";
14901        let (result, session, events, _) = drive_permission_handshake_frames(
14902            session_id,
14903            [completed_actor_frame()],
14904            expected_default_permission_posture(7),
14905        )
14906        .await;
14907
14908        assert!(result
14909            .unwrap_err()
14910            .to_string()
14911            .contains("terminated before permission posture confirmation"));
14912        assert!(events.is_empty());
14913        assert!(bamboo_domain::PermissionAuditSnapshot::from_metadata(&session.metadata).is_none());
14914    }
14915
14916    #[tokio::test]
14917    async fn remote_environment_rejection_accepts_only_allowlisted_durable_terminal() {
14918        struct TerminalProofProbe {
14919            frame: Option<ChildFrame>,
14920            accepted: bool,
14921        }
14922
14923        #[async_trait]
14924        impl bamboo_subagent::ChildLink for TerminalProofProbe {
14925            async fn send(&mut self, _frame: ParentFrame) -> bamboo_subagent::TransportResult<()> {
14926                Ok(())
14927            }
14928
14929            async fn next_frame(&mut self) -> bamboo_subagent::TransportResult<Option<ChildFrame>> {
14930                Ok(self.frame.take())
14931            }
14932
14933            fn accept_durable_terminal(&mut self, status: TerminalStatus) {
14934                self.accepted = status == TerminalStatus::Error;
14935            }
14936
14937            fn has_pending_durable_terminal(&self) -> bool {
14938                self.accepted
14939            }
14940        }
14941
14942        for (status, error, expected_proof) in [
14943            (
14944                TerminalStatus::Error,
14945                Some("remote_environment_checkout_not_clean"),
14946                true,
14947            ),
14948            (TerminalStatus::Error, Some("untrusted worker error"), false),
14949            (
14950                TerminalStatus::Cancelled,
14951                Some("remote_environment_checkout_not_clean"),
14952                false,
14953            ),
14954        ] {
14955            let mut link = TerminalProofProbe {
14956                frame: Some(ChildFrame::Terminal {
14957                    status,
14958                    result: None,
14959                    error: error.map(str::to_owned),
14960                    final_event_watermark: None,
14961                    transcript: Vec::new(),
14962                }),
14963                accepted: false,
14964            };
14965            let (event_tx, _event_rx) = mpsc::channel(8);
14966            let cancel = CancellationToken::new();
14967            let (_live_tx, mut live_rx) = mpsc::unbounded_channel();
14968            let (_delivery_tx, mut delivery_rx) = mpsc::unbounded_channel();
14969            let mut session = Session::new("remote-terminal-proof", "model");
14970            let result = drive(ActorDriveContext {
14971                client: &mut link,
14972                parent_session_id: "proof-parent",
14973                child_session_id: "remote-terminal-proof",
14974                child_attempt: 0,
14975                approval_registry: None,
14976                approval_decider: None,
14977                approval_reviewer: None,
14978                escalation_bridge: None,
14979                event_tx: &event_tx,
14980                cancel_token: &cancel,
14981                live_rx: &mut live_rx,
14982                delivery_rx: &mut delivery_rx,
14983                logical_session: &mut session,
14984                expected_permission_posture: Some(expected_default_permission_posture(7)),
14985                expected_creation: None,
14986                session_inbox_runtime: None,
14987                actor_directory_store: None,
14988                canonical_subagent_tool: None,
14989                ticket_service: None,
14990                activation_run_id: None,
14991                execution_epoch: 0,
14992                expected_source_actor_id: "remote-terminal-proof",
14993                initial_inflight_claims: VecDeque::new(),
14994                plain_actor: false,
14995                remote_environment_lease: true,
14996                readonly_output: None,
14997                local_history_tools: None,
14998                local_history_read_only: false,
14999                strict_history_output: None,
15000                plain_input: None,
15001                canonical_activation: None,
15002                canonical_placement_ref: None,
15003                actor_event_observer: None,
15004                plain_run: None,
15005                first_frame_timeout: Some(Duration::from_secs(1)),
15006            })
15007            .await;
15008            assert!(result.is_err());
15009            assert_eq!(
15010                bamboo_subagent::ChildLink::has_pending_durable_terminal(&link),
15011                expected_proof
15012            );
15013        }
15014    }
15015
15016    #[tokio::test(start_paused = true)]
15017    async fn remote_cancel_requires_selected_worker_terminal_for_release_proof() {
15018        struct CancelOutcomeLink {
15019            cancel_sent: bool,
15020            settle: bool,
15021            accepted: bool,
15022        }
15023
15024        #[async_trait]
15025        impl bamboo_subagent::ChildLink for CancelOutcomeLink {
15026            async fn send(&mut self, frame: ParentFrame) -> bamboo_subagent::TransportResult<()> {
15027                self.cancel_sent = matches!(frame, ParentFrame::Cancel);
15028                Ok(())
15029            }
15030
15031            async fn next_frame(&mut self) -> bamboo_subagent::TransportResult<Option<ChildFrame>> {
15032                if !self.cancel_sent || !self.settle {
15033                    std::future::pending::<bamboo_subagent::TransportResult<Option<ChildFrame>>>()
15034                        .await
15035                } else {
15036                    self.cancel_sent = false;
15037                    Ok(Some(ChildFrame::Terminal {
15038                        status: TerminalStatus::Cancelled,
15039                        result: None,
15040                        error: None,
15041                        final_event_watermark: None,
15042                        transcript: Vec::new(),
15043                    }))
15044                }
15045            }
15046
15047            fn accept_durable_terminal(&mut self, status: TerminalStatus) {
15048                self.accepted = status == TerminalStatus::Cancelled;
15049            }
15050
15051            fn has_pending_durable_terminal(&self) -> bool {
15052                self.accepted
15053            }
15054        }
15055
15056        let temp = tempfile::tempdir().unwrap();
15057        let store = bamboo_storage::SessionStoreV2::new(temp.path().into())
15058            .await
15059            .unwrap();
15060        let fence = ActorActivationFence {
15061            schema_version: 2,
15062            actor_id: "cancelled-remote-child".into(),
15063            activation_id: "cancelled-remote-activation".into(),
15064            attempt: 1,
15065            run_id: "cancelled-remote-run".into(),
15066            lease_owner: "cancelled-remote-host".into(),
15067            lease_epoch: 1,
15068        };
15069        for settle in [true, false] {
15070            let mut link = CancelOutcomeLink {
15071                cancel_sent: false,
15072                settle,
15073                accepted: false,
15074            };
15075            let (event_tx, _event_rx) = mpsc::channel(8);
15076            let cancel = CancellationToken::new();
15077            cancel.cancel();
15078            let (_live_tx, mut live_rx) = mpsc::unbounded_channel();
15079            let (_delivery_tx, mut delivery_rx) = mpsc::unbounded_channel();
15080            let mut session = Session::new("cancelled-remote-child", "model");
15081            let result = drive(ActorDriveContext {
15082                expected_creation: None,
15083                client: &mut link,
15084                parent_session_id: "cancelled-remote-parent",
15085                child_session_id: "cancelled-remote-child",
15086                child_attempt: 1,
15087                approval_registry: None,
15088                approval_decider: None,
15089                approval_reviewer: None,
15090                escalation_bridge: None,
15091                event_tx: &event_tx,
15092                cancel_token: &cancel,
15093                live_rx: &mut live_rx,
15094                delivery_rx: &mut delivery_rx,
15095                logical_session: &mut session,
15096                expected_permission_posture: Some(expected_default_permission_posture(7)),
15097                session_inbox_runtime: None,
15098                actor_directory_store: Some(&store),
15099                canonical_subagent_tool: None,
15100                ticket_service: None,
15101                activation_run_id: Some("cancelled-remote-run"),
15102                execution_epoch: 0,
15103                expected_source_actor_id: "cancelled-remote-worker",
15104                initial_inflight_claims: VecDeque::new(),
15105                plain_actor: false,
15106                remote_environment_lease: true,
15107                readonly_output: None,
15108                local_history_tools: None,
15109                local_history_read_only: false,
15110                strict_history_output: None,
15111                plain_input: None,
15112                canonical_activation: Some(&fence),
15113                canonical_placement_ref: None,
15114                actor_event_observer: None,
15115                plain_run: None,
15116                first_frame_timeout: None,
15117            })
15118            .await;
15119            assert!(matches!(result, Err(AgentError::Cancelled)));
15120            assert_eq!(
15121                link.accepted, settle,
15122                "only a consumed exact terminal can release the remote fence"
15123            );
15124            assert_eq!(
15125                bamboo_subagent::ChildLink::has_pending_durable_terminal(&link),
15126                settle
15127            );
15128        }
15129    }
15130
15131    #[tokio::test]
15132    async fn actor_permission_handshake_rejects_malformed_agent_event() {
15133        let session_id = "permission-malformed";
15134        let (result, session, events, _) = drive_permission_handshake_frames(
15135            session_id,
15136            [ChildFrame::Event {
15137                event: serde_json::json!({"type": "token", "content": 42}),
15138            }],
15139            expected_default_permission_posture(7),
15140        )
15141        .await;
15142
15143        assert!(result
15144            .unwrap_err()
15145            .to_string()
15146            .contains("malformed AgentEvent"));
15147        assert!(events.is_empty());
15148        assert!(bamboo_domain::PermissionAuditSnapshot::from_metadata(&session.metadata).is_none());
15149
15150        let (result, _, events, _) = drive_permission_handshake_frames(
15151            session_id,
15152            [
15153                permission_posture_frame(session_id, 7),
15154                ChildFrame::Event {
15155                    event: serde_json::json!({"type":"private-download-secret"}),
15156                },
15157            ],
15158            expected_default_permission_posture(7),
15159        )
15160        .await;
15161        assert!(!result
15162            .unwrap_err()
15163            .to_string()
15164            .contains("private-download-secret"));
15165        assert_eq!(
15166            events.len(),
15167            1,
15168            "only the confirmed posture may be forwarded"
15169        );
15170    }
15171
15172    #[tokio::test]
15173    async fn actor_permission_handshake_rejects_execution_event_before_posture() {
15174        let session_id = "permission-early-event";
15175        let early_events = [
15176            (
15177                "progress",
15178                AgentEvent::RunnerProgress {
15179                    session_id: session_id.to_string(),
15180                    round_count: 1,
15181                },
15182            ),
15183            (
15184                "token",
15185                AgentEvent::Token {
15186                    content: "must-not-forward".to_string(),
15187                },
15188            ),
15189            (
15190                "tool",
15191                AgentEvent::ToolStart {
15192                    tool_call_id: "early-tool".to_string(),
15193                    tool_name: "Read".to_string(),
15194                    arguments: serde_json::json!({"file_path": "README.md"}),
15195                },
15196            ),
15197        ];
15198        for (kind, event) in early_events {
15199            let (result, session, events, _) = drive_permission_handshake_frames(
15200                session_id,
15201                [
15202                    actor_event_frame(event),
15203                    permission_posture_frame(session_id, 7),
15204                    completed_actor_frame(),
15205                ],
15206                expected_default_permission_posture(7),
15207            )
15208            .await;
15209
15210            assert!(
15211                result
15212                    .unwrap_err()
15213                    .to_string()
15214                    .contains("execution event before permission posture confirmation"),
15215                "{kind} must fail closed before posture"
15216            );
15217            assert!(events.is_empty(), "{kind} must not be forwarded");
15218            assert!(
15219                bamboo_domain::PermissionAuditSnapshot::from_metadata(&session.metadata).is_none(),
15220                "{kind} must not advance the permission audit"
15221            );
15222        }
15223    }
15224
15225    #[tokio::test]
15226    async fn actor_permission_handshake_rejects_approval_before_posture() {
15227        let session_id = "permission-early-approval";
15228        let (result, session, events, sent) = drive_permission_handshake_frames(
15229            session_id,
15230            [ChildFrame::ApprovalRequest {
15231                id: "approval-before-posture".to_string(),
15232                body: serde_json::json!({
15233                    "tool_name": "Bash",
15234                    "permission": "execute",
15235                    "resource": "echo must-not-run"
15236                }),
15237            }],
15238            expected_default_permission_posture(7),
15239        )
15240        .await;
15241
15242        assert!(result
15243            .unwrap_err()
15244            .to_string()
15245            .contains("requested approval before permission posture confirmation"));
15246        assert!(events.is_empty());
15247        assert!(
15248            sent.is_empty(),
15249            "an unconfirmed actor must receive no approval reply"
15250        );
15251        assert!(bamboo_domain::PermissionAuditSnapshot::from_metadata(&session.metadata).is_none());
15252    }
15253
15254    #[tokio::test]
15255    async fn actor_permission_handshake_rejects_mismatched_posture() {
15256        let session_id = "permission-mismatch";
15257        let (result, session, events, _) = drive_permission_handshake_frames(
15258            session_id,
15259            [permission_posture_frame(session_id, 8)],
15260            expected_default_permission_posture(7),
15261        )
15262        .await;
15263
15264        assert!(result
15265            .unwrap_err()
15266            .to_string()
15267            .contains("does not match the host-dispatched policy"));
15268        assert!(events.is_empty());
15269        assert!(bamboo_domain::PermissionAuditSnapshot::from_metadata(&session.metadata).is_none());
15270    }
15271
15272    #[tokio::test]
15273    async fn actor_permission_handshake_rejects_untrusted_executor_mapping() {
15274        let session_id = "permission-hostile-mapping";
15275        for hostile_mapping in [
15276            "wrong_executor:permission_mode=default",
15277            "test_actor:permission_mode=default;credential=must-not-persist",
15278        ] {
15279            let frame = ChildFrame::Event {
15280                event: serde_json::to_value(AgentEvent::PermissionPostureActivated {
15281                    session_id: session_id.to_string(),
15282                    policy_revision: 7,
15283                    requested_mode: "default".to_string(),
15284                    effective_mode: "default".to_string(),
15285                    executor_mapping: hostile_mapping.to_string(),
15286                })
15287                .expect("serialize hostile posture fixture"),
15288            };
15289            let (result, session, events, _) = drive_permission_handshake_frames(
15290                session_id,
15291                [frame],
15292                expected_default_permission_posture(7),
15293            )
15294            .await;
15295
15296            let error = result.unwrap_err().to_string();
15297            assert!(error.contains("host-dispatched executor mapping"));
15298            assert!(
15299                !error.contains(hostile_mapping),
15300                "untrusted mapping must not be reflected in host errors"
15301            );
15302            assert!(events.is_empty());
15303            assert!(
15304                bamboo_domain::PermissionAuditSnapshot::from_metadata(&session.metadata).is_none(),
15305                "untrusted mapping must not reach durable or in-memory audit state"
15306            );
15307        }
15308    }
15309
15310    #[tokio::test]
15311    async fn actor_permission_handshake_rejects_duplicate_posture() {
15312        let session_id = "permission-duplicate";
15313        let (result, session, events, _) = drive_permission_handshake_frames(
15314            session_id,
15315            [
15316                permission_posture_frame(session_id, 7),
15317                permission_posture_frame(session_id, 7),
15318                completed_actor_frame(),
15319            ],
15320            expected_default_permission_posture(7),
15321        )
15322        .await;
15323
15324        assert!(result
15325            .unwrap_err()
15326            .to_string()
15327            .contains("duplicate permission posture activation"));
15328        assert_eq!(
15329            events.len(),
15330            1,
15331            "only the confirmed posture may be forwarded"
15332        );
15333        assert!(matches!(
15334            events[0],
15335            AgentEvent::PermissionPostureActivated { .. }
15336        ));
15337        let audit = bamboo_domain::PermissionAuditSnapshot::from_metadata(&session.metadata)
15338            .expect("the first matching posture must be recorded");
15339        assert_eq!(audit.policy_revision, 7);
15340    }
15341
15342    #[tokio::test]
15343    async fn actor_permission_handshake_happy_path_persists_before_forwarding_execution() {
15344        let session_id = "permission-happy";
15345        let (result, session, events, _) = drive_permission_handshake_frames(
15346            session_id,
15347            [
15348                permission_posture_frame(session_id, 7),
15349                actor_event_frame(AgentEvent::RunnerProgress {
15350                    session_id: session_id.to_string(),
15351                    round_count: 1,
15352                }),
15353                actor_event_frame(AgentEvent::Token {
15354                    content: "working".to_string(),
15355                }),
15356                actor_event_frame(AgentEvent::ToolStart {
15357                    tool_call_id: "tool-1".to_string(),
15358                    tool_name: "Read".to_string(),
15359                    arguments: serde_json::json!({"file_path": "README.md"}),
15360                }),
15361                completed_actor_frame(),
15362            ],
15363            expected_default_permission_posture(7),
15364        )
15365        .await;
15366
15367        assert_eq!(result.unwrap().as_deref(), Some("done"));
15368        assert!(matches!(
15369            events.as_slice(),
15370            [
15371                AgentEvent::PermissionPostureActivated { .. },
15372                AgentEvent::RunnerProgress { .. },
15373                AgentEvent::Token { .. },
15374                AgentEvent::ToolStart { .. }
15375            ]
15376        ));
15377        let audit = bamboo_domain::PermissionAuditSnapshot::from_metadata(&session.metadata)
15378            .expect("matching posture must be recorded before execution events are accepted");
15379        assert_eq!(audit.policy_revision, 7);
15380        assert_eq!(audit.executor_mapping, "test_actor:permission_mode=default");
15381    }
15382
15383    #[tokio::test]
15384    async fn actor_download_tool_start_is_projected_before_parent_forwarding() {
15385        for (index, tool_name) in ["browser", "default::browser", "private-namespace::browser"]
15386            .into_iter()
15387            .enumerate()
15388        {
15389            let session_id = format!("actor-download-{index}");
15390            let args = serde_json::json!({
15391                "action":"download",
15392                "selector":"a[data-secret='private-selector']",
15393                "expected_epoch":17,
15394                "data_base64":"private-extra-bytes",
15395                "extra":{"url":"https://private.test/file"},
15396            });
15397            let original = args.clone();
15398            let read_args = serde_json::json!({"file_path":"README.md"});
15399            let (result, _session, events, _) = drive_permission_handshake_frames(
15400                &session_id,
15401                [
15402                    permission_posture_frame(&session_id, 7),
15403                    actor_event_frame(AgentEvent::ToolStart {
15404                        tool_call_id: "download-call".to_string(),
15405                        tool_name: tool_name.to_string(),
15406                        arguments: args.clone(),
15407                    }),
15408                    actor_event_frame(AgentEvent::ToolStart {
15409                        tool_call_id: "read-call".to_string(),
15410                        tool_name: "Read".to_string(),
15411                        arguments: read_args.clone(),
15412                    }),
15413                    completed_actor_frame(),
15414                ],
15415                expected_default_permission_posture(7),
15416            )
15417            .await;
15418            assert_eq!(result.expect("actor completed").as_deref(), Some("done"));
15419            assert_eq!(args, original);
15420            let starts: Vec<_> = events
15421                .iter()
15422                .filter_map(|event| match event {
15423                    AgentEvent::ToolStart {
15424                        tool_name,
15425                        arguments,
15426                        ..
15427                    } => Some((tool_name.as_str(), arguments)),
15428                    _ => None,
15429                })
15430                .collect();
15431            assert_eq!(starts.len(), 2);
15432            assert_eq!(starts[0].0, "browser");
15433            assert_eq!(
15434                starts[0].1,
15435                &serde_json::json!({"action":"download","expected_epoch":17})
15436            );
15437            assert_eq!(starts[1], ("Read", &read_args));
15438            assert!(!format!("{events:?}").contains("private"));
15439        }
15440    }
15441
15442    #[tokio::test]
15443    async fn actor_hook_lifecycle_hides_every_decision_text_across_event_batch_and_nested_stream() {
15444        let session_id = "actor-hook-display";
15445        let private = "private-selector-url-filename-response";
15446        let hook_name = format!("audit-{private}");
15447        let hook = |duration_ms, phase: &str, decision| AgentEvent::HookLifecycle {
15448            hook_name: hook_name.clone(),
15449            point: bamboo_domain::AgentHookPoint::BeforeToolExecution,
15450            phase: phase.into(),
15451            duration_ms,
15452            decision,
15453        };
15454        let decisions = [
15455            HookResult::Deny {
15456                reason: private.into(),
15457            },
15458            HookResult::InjectContext {
15459                text: private.into(),
15460            },
15461            HookResult::WithContext {
15462                result: Box::new(HookResult::WithContext {
15463                    result: Box::new(HookResult::Deny {
15464                        reason: private.into(),
15465                    }),
15466                    text: private.into(),
15467                }),
15468                text: private.into(),
15469            },
15470            HookResult::Suspend {
15471                reason: private.into(),
15472            },
15473            HookResult::Abort {
15474                reason: private.into(),
15475            },
15476        ];
15477        let originals = decisions.to_vec();
15478        let batch_events = [
15479            hook(2, "completed", decisions[1].clone()),
15480            hook(3, "completed", decisions[2].clone()),
15481            hook(4, "completed", decisions[3].clone()),
15482        ]
15483        .into_iter()
15484        .map(|event| serde_json::to_value(event).unwrap())
15485        .collect::<Vec<_>>();
15486        let batch = ActorEventBatch {
15487            logical_session: Some(LogicalSessionIdentity {
15488                creation: None,
15489                session_id: session_id.into(),
15490                parent_session_id: Some("permission-parent".into()),
15491                root_session_id: session_id.into(),
15492            }),
15493            activation_id: None,
15494            execution_epoch: 0,
15495            source_node_id: None,
15496            source_actor_id: Some(session_id.into()),
15497            first_seq: 1,
15498            last_seq: batch_events.len() as u64,
15499            qos: bamboo_subagent::ActorEventQos::classify(&batch_events[0]),
15500            events: batch_events,
15501        };
15502        let nested = AgentEvent::SubAgentEvent {
15503            parent_session_id: "parent".into(),
15504            child_session_id: "child".into(),
15505            event: Box::new(hook(5, "completed", decisions[4].clone())),
15506        };
15507        let raw = serde_json::to_string(&originals).unwrap();
15508        assert!(raw.contains(private));
15509
15510        let (outcome, _session, forwarded, _) = drive_permission_handshake_frames(
15511            session_id,
15512            [
15513                permission_posture_frame(session_id, 7),
15514                actor_event_frame(hook(1, private, decisions[0].clone())),
15515                ChildFrame::EventBatch { batch },
15516                actor_event_frame(nested),
15517                completed_actor_frame(),
15518            ],
15519            expected_default_permission_posture(7),
15520        )
15521        .await;
15522        assert_eq!(outcome.unwrap().as_deref(), Some("done"));
15523        let wire = serde_json::to_string(&forwarded).unwrap();
15524        assert!(
15525            !wire.contains(private),
15526            "private hook text reached SSE: {wire}"
15527        );
15528
15529        let hooks = forwarded
15530            .iter()
15531            .filter_map(|event| match event {
15532                AgentEvent::HookLifecycle {
15533                    hook_name,
15534                    point,
15535                    phase,
15536                    duration_ms,
15537                    decision,
15538                } => Some((hook_name, point, phase, duration_ms, decision)),
15539                _ => None,
15540            })
15541            .collect::<Vec<_>>();
15542        assert_eq!(hooks.len(), 4);
15543        assert!(hooks.iter().all(|(name, point, _, _, _)| {
15544            name == &&"actor-hook-1".to_string()
15545                && **point == bamboo_domain::AgentHookPoint::BeforeToolExecution
15546        }));
15547        assert_eq!(hooks[0].2, "event");
15548        assert_eq!(hooks[1].2, "completed");
15549        assert_eq!(*hooks[0].3, 1);
15550        assert_eq!(*hooks[3].3, 4);
15551        assert!(matches!(hooks[0].4, HookResult::Deny { reason } if reason == "Hook denied"));
15552        assert!(
15553            matches!(hooks[1].4, HookResult::InjectContext { text } if text == "Hook context hidden")
15554        );
15555        assert!(
15556            matches!(hooks[2].4, HookResult::WithContext { result, text }
15557            if text == "Hook context hidden" && matches!(result.as_ref(),
15558                HookResult::WithContext { result, text }
15559                    if text == "Hook context hidden" && matches!(result.as_ref(),
15560                        HookResult::Deny { reason } if reason == "Hook denied")))
15561        );
15562        assert!(matches!(hooks[3].4, HookResult::Suspend { reason } if reason == "Hook suspended"));
15563        assert!(forwarded.iter().any(|event| matches!(event,
15564            AgentEvent::SubAgentEvent { event, .. }
15565                if matches!(event.as_ref(), AgentEvent::HookLifecycle {
15566                    hook_name, duration_ms: 5,
15567                    decision: HookResult::Abort { reason }, ..
15568                } if hook_name == "actor-hook-1" && reason == "Hook aborted")
15569        )));
15570    }
15571
15572    #[test]
15573    fn actor_hook_display_depth_and_identity_limits_fail_closed() {
15574        let mut display = ActorEventDisplay::default();
15575        let mut deep = HookResult::Allow;
15576        for _ in 0..=MAX_DISPLAY_HOOK_RESULT_DEPTH {
15577            deep = HookResult::WithContext {
15578                result: Box::new(deep),
15579                text: "private-download-bytes".into(),
15580            };
15581        }
15582        let event = |hook_name: String, decision| AgentEvent::HookLifecycle {
15583            hook_name,
15584            point: bamboo_domain::AgentHookPoint::BeforeToolExecution,
15585            phase: "completed".into(),
15586            duration_ms: 7,
15587            decision,
15588        };
15589        assert!(display.project(event("deep".into(), deep), 0).is_none());
15590        assert!(display.overflowed);
15591        assert!(display
15592            .project(
15593                event(
15594                    "private-download-bytes".into(),
15595                    HookResult::Deny {
15596                        reason: "private-download-bytes".into(),
15597                    },
15598                ),
15599                0,
15600            )
15601            .is_none());
15602
15603        let mut oversized = ActorEventDisplay::default();
15604        assert!(oversized
15605            .project(
15606                event("x".repeat(MAX_DISPLAY_ID_BYTES + 1), HookResult::Allow),
15607                0,
15608            )
15609            .is_none());
15610        assert!(oversized.overflowed);
15611
15612        // A new actor connection gets a fresh opaque alias, never the raw
15613        // hook name. Repeated names keep the same alias until the shared
15614        // identity budget is exhausted, then the stream fails closed.
15615        let mut reconnected = ActorEventDisplay::default();
15616        let private = "private-download-bytes";
15617        for _ in 0..2 {
15618            let projected = reconnected
15619                .project(event(private.into(), HookResult::Allow), 0)
15620                .unwrap();
15621            assert!(
15622                matches!(projected, AgentEvent::HookLifecycle { hook_name, .. }
15623                if hook_name == "actor-hook-1")
15624            );
15625        }
15626        assert_eq!(reconnected.identities.load(Ordering::Relaxed), 1);
15627        for index in 1..MAX_DISPLAY_CALLS {
15628            assert!(reconnected
15629                .project(event(format!("hook-{index}"), HookResult::Allow), 0)
15630                .is_some());
15631        }
15632        assert!(reconnected
15633            .project(event("overflow-private".into(), HookResult::Allow), 0)
15634            .is_none());
15635        assert!(reconnected.overflowed);
15636        assert!(reconnected.hook_names.is_empty());
15637    }
15638
15639    #[test]
15640    fn actor_hook_display_preserves_nontext_decision_kinds() {
15641        for result in [
15642            HookResult::Continue,
15643            HookResult::Mutated,
15644            HookResult::Allow,
15645            HookResult::Ask,
15646        ] {
15647            assert_eq!(hook_result_for_display(result.clone(), 0), Some(result));
15648        }
15649    }
15650
15651    #[tokio::test]
15652    async fn malformed_actor_hook_events_never_fall_back_to_raw_display() {
15653        let session_id = "actor-malformed-hook";
15654        let private = "private-download-selector";
15655        let original = AgentEvent::HookLifecycle {
15656            hook_name: "audit".into(),
15657            point: bamboo_domain::AgentHookPoint::BeforeToolExecution,
15658            phase: "completed".into(),
15659            duration_ms: 9,
15660            decision: HookResult::Deny {
15661                reason: private.into(),
15662            },
15663        };
15664        let mut invalid = serde_json::to_value(&original).unwrap();
15665        invalid["decision"]["type"] = serde_json::json!("unknown_decision");
15666        let nested = AgentEvent::SubAgentEvent {
15667            parent_session_id: "parent".into(),
15668            child_session_id: "child".into(),
15669            event: Box::new(original),
15670        };
15671        let mut invalid_nested = serde_json::to_value(nested).unwrap();
15672        invalid_nested["event"]["decision"]["type"] = serde_json::json!("unknown_decision");
15673        let batch = ActorEventBatch {
15674            logical_session: Some(LogicalSessionIdentity {
15675                creation: None,
15676                session_id: session_id.into(),
15677                parent_session_id: Some("permission-parent".into()),
15678                root_session_id: session_id.into(),
15679            }),
15680            activation_id: None,
15681            execution_epoch: 0,
15682            source_node_id: None,
15683            source_actor_id: Some(session_id.into()),
15684            first_seq: 1,
15685            last_seq: 1,
15686            qos: bamboo_subagent::ActorEventQos::classify(&invalid),
15687            events: vec![invalid.clone()],
15688        };
15689        for frame in [
15690            ChildFrame::Event {
15691                event: invalid.clone(),
15692            },
15693            ChildFrame::EventBatch { batch },
15694            ChildFrame::Event {
15695                event: invalid_nested,
15696            },
15697        ] {
15698            let (outcome, _session, forwarded, _) = drive_permission_handshake_frames(
15699                session_id,
15700                [permission_posture_frame(session_id, 7), frame],
15701                expected_default_permission_posture(7),
15702            )
15703            .await;
15704            let error = outcome.unwrap_err().to_string();
15705            assert!(error.contains("malformed AgentEvent"), "{error}");
15706            assert!(!error.contains(private));
15707            assert!(!serde_json::to_string(&forwarded).unwrap().contains(private));
15708        }
15709
15710        let (event_tx, mut event_rx) = mpsc::channel(8);
15711        let mut session = Session::new("legacy-hook", "model");
15712        let mut handshake = PermissionPostureHandshake::NotRequired;
15713        let mut display = ActorEventDisplay::default();
15714        process_actor_event(
15715            invalid,
15716            false,
15717            &mut handshake,
15718            None,
15719            None,
15720            &mut session,
15721            &event_tx,
15722            &mut display,
15723        )
15724        .await
15725        .unwrap();
15726        assert!(display.overflowed);
15727        assert!(event_rx.try_recv().is_err());
15728    }
15729
15730    #[tokio::test]
15731    async fn actor_download_event_batch_hides_output_and_preserves_read_output() {
15732        let session_id = "actor-download-batch";
15733        let private = "private-download-bytes-and-url";
15734        let events = vec![
15735            AgentEvent::ToolStart {
15736                tool_call_id: "download-call".into(),
15737                tool_name: "default::browser".into(),
15738                arguments: serde_json::json!({
15739                    "action":"download", "selector": private, "expected_epoch": 17,
15740                }),
15741            },
15742            AgentEvent::ToolApprovalRequested {
15743                tool_call_id: "download-call".into(),
15744                tool_name: format!("{private}::browser"),
15745                parameters: serde_json::json!({
15746                    "action":"download", "selector": private, "expected_epoch": 17,
15747                }),
15748            },
15749            AgentEvent::ToolStart {
15750                tool_call_id: "read-call".into(),
15751                tool_name: "Read".into(),
15752                arguments: serde_json::json!({"file_path":"README.md"}),
15753            },
15754        ];
15755        let batch = ActorEventBatch {
15756            logical_session: Some(LogicalSessionIdentity {
15757                creation: None,
15758                session_id: session_id.into(),
15759                parent_session_id: Some("permission-parent".into()),
15760                root_session_id: session_id.into(),
15761            }),
15762            activation_id: None,
15763            execution_epoch: 0,
15764            source_node_id: None,
15765            source_actor_id: Some(session_id.into()),
15766            first_seq: 1,
15767            last_seq: events.len() as u64,
15768            qos: bamboo_subagent::ActorEventQos::Durable,
15769            events: events
15770                .into_iter()
15771                .map(|event| serde_json::to_value(event).unwrap())
15772                .collect(),
15773        };
15774        let result = bamboo_agent_core::tools::ToolResult {
15775            success: true,
15776            result: format!("{{\"data_base64\":\"{private}\"}}"),
15777            display_preference: Some(private.into()),
15778            images: vec![bamboo_agent_core::tools::ToolResultImage {
15779                mime_type: "image/png".into(),
15780                data: private.into(),
15781            }],
15782        };
15783        let (outcome, _, forwarded, _) = drive_permission_handshake_frames(
15784            session_id,
15785            [
15786                permission_posture_frame(session_id, 7),
15787                ChildFrame::EventBatch { batch },
15788                actor_event_frame(AgentEvent::ToolToken {
15789                    tool_call_id: "download-call".into(),
15790                    content: private.into(),
15791                }),
15792                actor_event_frame(AgentEvent::ToolToken {
15793                    tool_call_id: "read-call".into(),
15794                    content: "public live output".into(),
15795                }),
15796                actor_event_frame(AgentEvent::ToolComplete {
15797                    tool_call_id: "download-call".into(),
15798                    result,
15799                }),
15800                actor_event_frame(AgentEvent::ToolLifecycle {
15801                    tool_call_id: "download-call".into(),
15802                    tool_name: "browser".into(),
15803                    phase: "finished".into(),
15804                    elapsed_ms: Some(1),
15805                    is_mutating: false,
15806                    auto_approved: false,
15807                    summary: Some(private.into()),
15808                    error: Some(private.into()),
15809                }),
15810                actor_event_frame(AgentEvent::ToolError {
15811                    tool_call_id: "download-call".into(),
15812                    error: private.into(),
15813                }),
15814                actor_event_frame(AgentEvent::ToolComplete {
15815                    tool_call_id: "read-call".into(),
15816                    result: bamboo_agent_core::tools::ToolResult::text(true, "public README text"),
15817                }),
15818                completed_actor_frame(),
15819            ],
15820            expected_default_permission_posture(7),
15821        )
15822        .await;
15823        assert_eq!(outcome.unwrap().as_deref(), Some("done"));
15824        let wire = serde_json::to_string(&forwarded).unwrap();
15825        assert!(
15826            !wire.contains(private),
15827            "private content reached actor SSE: {wire}"
15828        );
15829        assert!(wire.contains("public README text"));
15830        assert!(wire.contains("public live output"));
15831        assert!(forwarded.iter().any(|event| matches!(event,
15832            AgentEvent::ToolComplete { tool_call_id, result }
15833                if tool_call_id == "download-call"
15834                && result.result == "Browser tool result hidden"
15835                && result.images.is_empty() && result.display_preference.is_none()
15836        )));
15837    }
15838
15839    #[tokio::test]
15840    async fn actor_missing_reused_and_nested_tool_identity_fails_closed() {
15841        let private = "private-download-content";
15842        let nested = |event| AgentEvent::SubAgentEvent {
15843            parent_session_id: "parent".into(),
15844            child_session_id: "nested-child".into(),
15845            event: Box::new(event),
15846        };
15847        let (outcome, _, forwarded, _) = drive_permission_handshake_frames(
15848            "actor-reorder",
15849            [
15850                permission_posture_frame("actor-reorder", 7),
15851                actor_event_frame(AgentEvent::ToolComplete {
15852                    tool_call_id: "missing".into(),
15853                    result: bamboo_agent_core::tools::ToolResult::text(true, private),
15854                }),
15855                actor_event_frame(AgentEvent::ToolLifecycle {
15856                    tool_call_id: "missing".into(),
15857                    tool_name: "browser".into(),
15858                    phase: private.into(),
15859                    elapsed_ms: None,
15860                    is_mutating: false,
15861                    auto_approved: false,
15862                    summary: Some(private.into()),
15863                    error: Some(private.into()),
15864                }),
15865                actor_event_frame(AgentEvent::ToolApprovalRequested {
15866                    tool_call_id: "approval-only".into(),
15867                    tool_name: "browser".into(),
15868                    parameters: serde_json::json!({"action":"download","selector":private}),
15869                }),
15870                actor_event_frame(AgentEvent::ToolToken {
15871                    tool_call_id: "approval-only".into(),
15872                    content: private.into(),
15873                }),
15874                actor_event_frame(AgentEvent::ToolStart {
15875                    tool_call_id: "malformed-browser".into(),
15876                    tool_name: format!("{private}::browser"),
15877                    arguments: serde_json::json!({"selector":private}),
15878                }),
15879                actor_event_frame(AgentEvent::ToolComplete {
15880                    tool_call_id: "malformed-browser".into(),
15881                    result: bamboo_agent_core::tools::ToolResult::text(true, private),
15882                }),
15883                actor_event_frame(AgentEvent::ToolStart {
15884                    tool_call_id: "reused".into(),
15885                    tool_name: "Read".into(),
15886                    arguments: serde_json::json!({"file_path":"README.md"}),
15887                }),
15888                actor_event_frame(AgentEvent::ToolStart {
15889                    tool_call_id: "reused".into(),
15890                    tool_name: "browser".into(),
15891                    arguments: serde_json::json!({"action":"download","selector":private}),
15892                }),
15893                actor_event_frame(AgentEvent::ToolToken {
15894                    tool_call_id: "reused".into(),
15895                    content: private.into(),
15896                }),
15897                actor_event_frame(nested(AgentEvent::ToolStart {
15898                    tool_call_id: "nested".into(),
15899                    tool_name: "browser".into(),
15900                    arguments: serde_json::json!({"action":"download","selector":private}),
15901                })),
15902                actor_event_frame(nested(AgentEvent::ToolComplete {
15903                    tool_call_id: "nested".into(),
15904                    result: bamboo_agent_core::tools::ToolResult::text(true, private),
15905                })),
15906                actor_event_frame(AgentEvent::ChildApprovalRequested {
15907                    child_session_id: "nested-child".into(),
15908                    request_id: "approval".into(),
15909                    tool_name: format!("{private}::browser"),
15910                    permission: private.into(),
15911                    resource: private.into(),
15912                }),
15913                actor_event_frame(AgentEvent::ChildApprovalChanged {
15914                    parent_session_id: "parent".into(),
15915                    child_session_id: "nested-child".into(),
15916                    child_attempt: 1,
15917                    request_id: "approval".into(),
15918                    version: 1,
15919                    status: "pending".into(),
15920                    reason: Some(private.into()),
15921                    tool_name: "browser".into(),
15922                    permission: private.into(),
15923                    resource: private.into(),
15924                    created_at: "2026-01-01T00:00:00Z".into(),
15925                    resolved_at: None,
15926                }),
15927                completed_actor_frame(),
15928            ],
15929            expected_default_permission_posture(7),
15930        )
15931        .await;
15932        assert_eq!(outcome.unwrap().as_deref(), Some("done"));
15933        let wire = serde_json::to_string(&forwarded).unwrap();
15934        assert!(
15935            !wire.contains(private),
15936            "private content reached actor SSE: {wire}"
15937        );
15938        assert!(wire.contains("Tool result hidden"));
15939        assert!(wire.contains("Browser tool result hidden"));
15940        assert!(wire.contains("Browser interaction approval"));
15941        assert!(!forwarded
15942            .iter()
15943            .any(|event| matches!(event, AgentEvent::ChildApprovalChanged { .. })));
15944    }
15945
15946    #[test]
15947    fn actor_display_tracking_overflow_never_reclassifies_old_private_id() {
15948        let mut display = ActorEventDisplay::default();
15949        display
15950            .project(
15951                AgentEvent::ToolStart {
15952                    tool_call_id: "private-call".into(),
15953                    tool_name: "browser".into(),
15954                    arguments: serde_json::json!({"action":"download","selector":"private"}),
15955                },
15956                0,
15957            )
15958            .unwrap();
15959        for index in 0..MAX_DISPLAY_CALLS {
15960            display
15961                .project(
15962                    AgentEvent::ToolStart {
15963                        tool_call_id: format!("ordinary-{index}"),
15964                        tool_name: "Read".into(),
15965                        arguments: serde_json::json!({"file_path":"README.md"}),
15966                    },
15967                    0,
15968                )
15969                .unwrap();
15970        }
15971        assert!(display.overflowed);
15972        let forwarded = display
15973            .project(
15974                AgentEvent::ToolComplete {
15975                    tool_call_id: "private-call".into(),
15976                    result: bamboo_agent_core::tools::ToolResult::text(true, "private-bytes"),
15977                },
15978                0,
15979            )
15980            .unwrap();
15981        assert!(!serde_json::to_string(&forwarded)
15982            .unwrap()
15983            .contains("private-bytes"));
15984    }
15985
15986    #[test]
15987    fn nested_actor_display_uses_one_identity_budget_and_bounded_keys() {
15988        let mut display = ActorEventDisplay::default();
15989        let nested = |index: usize| AgentEvent::SubAgentEvent {
15990            parent_session_id: "parent".into(),
15991            child_session_id: format!("child-{index}"),
15992            event: Box::new(AgentEvent::ToolStart {
15993                tool_call_id: format!("read-{index}"),
15994                tool_name: "Read".into(),
15995                arguments: serde_json::json!({"file_path":"README.md"}),
15996            }),
15997        };
15998        for index in 0..(MAX_DISPLAY_CALLS / 2) {
15999            assert!(display.project(nested(index), 0).is_some());
16000        }
16001        assert_eq!(
16002            display.identities.load(Ordering::Relaxed),
16003            MAX_DISPLAY_CALLS
16004        );
16005        assert!(display.project(nested(MAX_DISPLAY_CALLS / 2), 0).is_none());
16006        assert!(display.overflowed);
16007        assert!(display.children.is_empty());
16008
16009        let mut oversized = ActorEventDisplay::default();
16010        assert!(oversized
16011            .project(
16012                AgentEvent::SubAgentEvent {
16013                    parent_session_id: "parent".into(),
16014                    child_session_id: "x".repeat(MAX_DISPLAY_ID_BYTES + 1),
16015                    event: Box::new(AgentEvent::ToolComplete {
16016                        tool_call_id: "call".into(),
16017                        result: bamboo_agent_core::tools::ToolResult::text(
16018                            true,
16019                            "private-download-bytes",
16020                        ),
16021                    }),
16022                },
16023                0,
16024            )
16025            .is_none());
16026        assert!(oversized.overflowed);
16027        assert!(oversized
16028            .project(
16029                AgentEvent::ToolStart {
16030                    tool_call_id: "x".repeat(MAX_DISPLAY_ID_BYTES + 1),
16031                    tool_name: "Read".into(),
16032                    arguments: serde_json::json!({"file_path":"README.md"}),
16033                },
16034                0,
16035            )
16036            .is_none());
16037    }
16038
16039    #[tokio::test]
16040    async fn actor_batch_sequence_gap_invalidates_prior_tool_identity() {
16041        let session_id = "actor-gap";
16042        let batch = |seq, event: AgentEvent| {
16043            let event = serde_json::to_value(event).unwrap();
16044            ChildFrame::EventBatch {
16045                batch: ActorEventBatch {
16046                    logical_session: Some(LogicalSessionIdentity {
16047                        creation: None,
16048                        session_id: session_id.into(),
16049                        parent_session_id: Some("permission-parent".into()),
16050                        root_session_id: session_id.into(),
16051                    }),
16052                    activation_id: None,
16053                    execution_epoch: 0,
16054                    source_node_id: None,
16055                    source_actor_id: Some(session_id.into()),
16056                    first_seq: seq,
16057                    last_seq: seq,
16058                    qos: bamboo_subagent::ActorEventQos::classify(&event),
16059                    events: vec![event],
16060                },
16061            }
16062        };
16063        let (outcome, _, forwarded, _) = drive_permission_handshake_frames(
16064            session_id,
16065            [
16066                permission_posture_frame(session_id, 7),
16067                batch(
16068                    1,
16069                    AgentEvent::ToolStart {
16070                        tool_call_id: "reused-call".into(),
16071                        tool_name: "Read".into(),
16072                        arguments: serde_json::json!({"file_path":"README.md"}),
16073                    },
16074                ),
16075                // Sequence 2 may have replaced this ID with a private browser
16076                // call. The actor's next result must not inherit Read display.
16077                batch(
16078                    3,
16079                    AgentEvent::ToolComplete {
16080                        tool_call_id: "reused-call".into(),
16081                        result: bamboo_agent_core::tools::ToolResult::text(
16082                            true,
16083                            "private-download-bytes",
16084                        ),
16085                    },
16086                ),
16087                completed_actor_frame(),
16088            ],
16089            expected_default_permission_posture(7),
16090        )
16091        .await;
16092        assert_eq!(outcome.unwrap().as_deref(), Some("done"));
16093        let wire = serde_json::to_string(&forwarded).unwrap();
16094        assert!(!wire.contains("private-download-bytes"), "{wire}");
16095        assert!(wire.contains("Tool result hidden"));
16096    }
16097
16098    #[tokio::test]
16099    async fn legacy_malformed_actor_event_invalidates_prior_tool_identity() {
16100        let (event_tx, mut event_rx) = mpsc::channel(8);
16101        let mut session = Session::new("legacy-malformed-private", "model");
16102        let mut handshake = PermissionPostureHandshake::NotRequired;
16103        let mut display = ActorEventDisplay::default();
16104        let read_start = serde_json::to_value(AgentEvent::ToolStart {
16105            tool_call_id: "reused-call".into(),
16106            tool_name: "Read".into(),
16107            arguments: serde_json::json!({"file_path":"README.md"}),
16108        })
16109        .unwrap();
16110        process_actor_event(
16111            read_start,
16112            false,
16113            &mut handshake,
16114            None,
16115            None,
16116            &mut session,
16117            &event_tx,
16118            &mut display,
16119        )
16120        .await
16121        .unwrap();
16122
16123        // A mixed-version actor may send a browser ToolStart whose name has
16124        // the wrong wire type. Serde drops it; the reused ID is no longer safe.
16125        let mut malformed = serde_json::to_value(AgentEvent::ToolStart {
16126            tool_call_id: "reused-call".into(),
16127            tool_name: "browser".into(),
16128            arguments: serde_json::json!({
16129                "action":"download", "selector":"private-download-selector"
16130            }),
16131        })
16132        .unwrap();
16133        malformed["tool_name"] = serde_json::json!(7);
16134        process_actor_event(
16135            malformed,
16136            false,
16137            &mut handshake,
16138            None,
16139            None,
16140            &mut session,
16141            &event_tx,
16142            &mut display,
16143        )
16144        .await
16145        .unwrap();
16146        process_actor_event(
16147            serde_json::to_value(AgentEvent::ToolComplete {
16148                tool_call_id: "reused-call".into(),
16149                result: bamboo_agent_core::tools::ToolResult::text(true, "private-download-bytes"),
16150            })
16151            .unwrap(),
16152            false,
16153            &mut handshake,
16154            None,
16155            None,
16156            &mut session,
16157            &event_tx,
16158            &mut display,
16159        )
16160        .await
16161        .unwrap();
16162        let forwarded = std::iter::from_fn(|| event_rx.try_recv().ok()).collect::<Vec<_>>();
16163        assert_eq!(forwarded.len(), 2);
16164        let wire = serde_json::to_string(&forwarded).unwrap();
16165        assert!(!wire.contains("private-download"), "{wire}");
16166        assert!(wire.contains("Tool result hidden"));
16167    }
16168
16169    #[tokio::test]
16170    async fn actor_initial_batch_acks_in_order_and_rejects_stale_confirmation() {
16171        let temp = tempfile::tempdir().unwrap();
16172        let store = Arc::new(
16173            bamboo_storage::SessionStoreV2::new(temp.path().to_path_buf())
16174                .await
16175                .unwrap(),
16176        );
16177        let storage: Arc<dyn Storage> = store.clone();
16178        let locked = Arc::new(bamboo_storage::LockedSessionStore::new(storage.clone()));
16179        let inbox: Arc<dyn SessionInboxPort> = Arc::new(bamboo_storage::FileSessionInbox::new(
16180            store.clone(),
16181            bamboo_domain::SessionInboxLimits::default(),
16182        ));
16183        let session_id = "actor-confirmation-order";
16184        let run_id = "actor-run-current";
16185        let mut session = Session::new(session_id, "model");
16186        store.save_session(&session).await.unwrap();
16187        for (id, text) in [("actor-first", "first"), ("actor-second", "second")] {
16188            let mut envelope = bamboo_domain::SessionMessageEnvelope::user_input(session_id, text);
16189            envelope.id = bamboo_domain::SessionMessageId::parse(id).unwrap();
16190            inbox.deliver(&envelope).await.unwrap();
16191        }
16192        inbox
16193            .mark_activation_eligible(
16194                session_id,
16195                2,
16196                bamboo_domain::SessionActivationPolicy::InterruptSpecificWait,
16197            )
16198            .await
16199            .unwrap();
16200        let router = SessionActivationRouter::new();
16201        let mut owner_registration = router.register_run(session_id, run_id).await.unwrap();
16202        let binding = SessionInboxRuntimeBinding {
16203            router,
16204            inbox: inbox.clone(),
16205            storage,
16206            persistence: locked,
16207            parent_question_lock: None,
16208        };
16209        let pairs = claim_canonical_deliveries(&binding, &mut session, run_id, usize::MAX, None)
16210            .await
16211            .unwrap();
16212        assert_eq!(
16213            pairs
16214                .iter()
16215                .map(|(claim, _)| claim.envelope.id.as_str())
16216                .collect::<Vec<_>>(),
16217            vec!["actor-first", "actor-second"]
16218        );
16219        let seeded = store.load_session(session_id).await.unwrap().unwrap();
16220        assert_eq!(
16221            seeded
16222                .messages
16223                .iter()
16224                .filter(|message| matches!(message.id.as_str(), "actor-first" | "actor-second"))
16225                .map(|message| message.id.as_str())
16226                .collect::<Vec<_>>(),
16227            vec!["actor-first", "actor-second"],
16228            "host context must be durable before actor dispatch"
16229        );
16230        for (claim, _) in &pairs {
16231            assert_eq!(
16232                seeded
16233                    .messages
16234                    .iter()
16235                    .filter(|message| bamboo_domain::is_matching_session_message(
16236                        message,
16237                        &claim.envelope
16238                    ))
16239                    .count(),
16240                1,
16241                "pre-dispatch host checkpoint must contain exactly one canonical marker for {}",
16242                claim.envelope.id
16243            );
16244        }
16245        assert!(
16246            seeded.session_inbox_admission().is_none_or(|cursor| {
16247                !cursor.contains(&pairs[0].0.envelope.id)
16248                    && !cursor.contains(&pairs[1].0.envelope.id)
16249            }),
16250            "pre-dispatch transcript seeding must not forge worker confirmation"
16251        );
16252        assert_eq!(inbox.inspect(session_id).await.unwrap().claimed, 2);
16253        let claims = pairs
16254            .into_iter()
16255            .map(|(claim, _)| claim)
16256            .collect::<VecDeque<_>>();
16257        let first = claims[0].clone();
16258        let second = claims[1].clone();
16259        let mut stale = admission_confirmation(session_id, &first, "stale-run");
16260        stale.canonical_claim_generation = second.generation;
16261        let mut link = ConfirmationSequenceLink {
16262            frames: VecDeque::from([
16263                ChildFrame::SessionMessageAdmitted {
16264                    confirmation: stale,
16265                },
16266                ChildFrame::SessionMessageAdmitted {
16267                    confirmation: admission_confirmation(session_id, &first, run_id),
16268                },
16269                ChildFrame::SessionMessageAdmitted {
16270                    confirmation: admission_confirmation(session_id, &second, run_id),
16271                },
16272                ChildFrame::Terminal {
16273                    status: TerminalStatus::Completed,
16274                    result: Some("done".to_string()),
16275                    error: None,
16276                    final_event_watermark: None,
16277                    transcript: Vec::new(),
16278                },
16279            ]),
16280            sent: Vec::new(),
16281        };
16282        let (event_tx, _event_rx) = mpsc::channel(8);
16283        let cancel = CancellationToken::new();
16284        let (_live_tx, mut live_rx) = mpsc::unbounded_channel();
16285        let (_delivery_tx, mut delivery_rx) = mpsc::unbounded_channel();
16286        let result = drive(ActorDriveContext {
16287            expected_creation: None,
16288            client: &mut link,
16289            parent_session_id: "parent",
16290            child_session_id: session_id,
16291            child_attempt: 0,
16292            approval_registry: None,
16293            approval_decider: None,
16294            approval_reviewer: None,
16295            escalation_bridge: None,
16296            event_tx: &event_tx,
16297            cancel_token: &cancel,
16298            live_rx: &mut live_rx,
16299            delivery_rx: &mut delivery_rx,
16300            logical_session: &mut session,
16301            expected_permission_posture: None,
16302            session_inbox_runtime: Some(&binding),
16303            actor_directory_store: None,
16304            canonical_subagent_tool: None,
16305            ticket_service: None,
16306            activation_run_id: Some(run_id),
16307            execution_epoch: 0,
16308            expected_source_actor_id: session_id,
16309            initial_inflight_claims: claims,
16310            plain_actor: false,
16311            remote_environment_lease: false,
16312            readonly_output: None,
16313            local_history_tools: None,
16314            local_history_read_only: false,
16315            strict_history_output: None,
16316            plain_input: None,
16317            canonical_activation: None,
16318            canonical_placement_ref: None,
16319            actor_event_observer: None,
16320            plain_run: None,
16321            first_frame_timeout: Some(Duration::from_secs(1)),
16322        })
16323        .await
16324        .unwrap();
16325        assert_eq!(result.as_deref(), Some("done"));
16326        assert_eq!(
16327            session
16328                .messages
16329                .iter()
16330                .filter(|message| matches!(message.id.as_str(), "actor-first" | "actor-second"))
16331                .map(|message| message.id.as_str())
16332                .collect::<Vec<_>>(),
16333            vec!["actor-first", "actor-second"]
16334        );
16335        let backlog = inbox.inspect(session_id).await.unwrap();
16336        assert_eq!(backlog.pending + backlog.claimed, 0);
16337        let confirmed = store.load_session(session_id).await.unwrap().unwrap();
16338        let cursor = confirmed
16339            .session_inbox_admission()
16340            .expect("exact worker confirmation must checkpoint the admission cursor");
16341        assert!(cursor.contains(&first.envelope.id));
16342        assert!(cursor.contains(&second.envelope.id));
16343        for claim in [&first, &second] {
16344            assert_eq!(
16345                confirmed
16346                    .messages
16347                    .iter()
16348                    .filter(|message| bamboo_domain::is_matching_session_message(
16349                        message,
16350                        &claim.envelope
16351                    ))
16352                    .count(),
16353                1,
16354                "confirmation must retain one exact canonical marker for {}",
16355                claim.envelope.id
16356            );
16357        }
16358        assert!(inbox
16359            .was_admitted(session_id, &first.envelope.id)
16360            .await
16361            .unwrap());
16362        assert!(inbox
16363            .was_admitted(session_id, &second.envelope.id)
16364            .await
16365            .unwrap());
16366        owner_registration.begin_finalization().await;
16367        owner_registration.finish(2).await.unwrap();
16368    }
16369
16370    #[derive(Default)]
16371    struct RecordingCodexTokenAuthority {
16372        issued_for: std::sync::Mutex<Vec<String>>,
16373        revoked: std::sync::Mutex<Vec<String>>,
16374    }
16375
16376    impl CodexRunTokenAuthority for RecordingCodexTokenAuthority {
16377        fn issue(&self, session_id: &str) -> Result<IssuedCodexRunToken, String> {
16378            self.issued_for
16379                .lock()
16380                .expect("issued fixture lock")
16381                .push(session_id.to_string());
16382            Ok(IssuedCodexRunToken {
16383                token_id: format!("id-{session_id}"),
16384                token: format!("bcx1_secret-{session_id}"),
16385            })
16386        }
16387
16388        fn revoke(&self, token_id: &str) {
16389            self.revoked
16390                .lock()
16391                .expect("revoked fixture lock")
16392                .push(token_id.to_string());
16393        }
16394    }
16395
16396    fn codex_executor(auth_mode: Option<&str>, inherit_user_config: Option<bool>) -> ExecutorSpec {
16397        let bamboo_mode = auth_mode == Some("bamboo")
16398            || (auth_mode.is_none() && !inherit_user_config.unwrap_or(false));
16399        ExecutorSpec::Codex {
16400            binary: None,
16401            model: None,
16402            mode: None,
16403            sandbox: None,
16404            inherit_user_config,
16405            auth_mode: auth_mode.map(str::to_string),
16406            base_url: bamboo_mode.then(|| "http://127.0.0.1:9562/openai/v1".to_string()),
16407            wire_api: Some("responses".to_string()),
16408            provider_key_ref: None,
16409            forward_env: None,
16410            approval_policy: None,
16411            network_access: None,
16412            allow_danger_bypass: None,
16413            permission_profile: None,
16414            workspace_owned: None,
16415        }
16416    }
16417
16418    fn permission_resolution(
16419        requested: bamboo_domain::SessionPermissionMode,
16420        effective: bamboo_domain::PermissionMode,
16421    ) -> bamboo_domain::PermissionModeResolution {
16422        bamboo_domain::PermissionModeResolution {
16423            requested,
16424            effective,
16425        }
16426    }
16427
16428    #[test]
16429    fn permission_posture_mapping_contract_is_exact_for_supported_executors() {
16430        use bamboo_domain::{PermissionMode, SessionPermissionMode};
16431
16432        let default =
16433            permission_resolution(SessionPermissionMode::Default, PermissionMode::Default);
16434        assert_eq!(
16435            expected_permission_executor_mapping(&ExecutorSpec::BambooRuntime, default, false)
16436                .unwrap()
16437                .as_deref(),
16438            Some("bamboo_runtime:default")
16439        );
16440        assert_eq!(
16441            expected_permission_executor_mapping(&ExecutorSpec::Echo, default, false).unwrap(),
16442            None,
16443            "transport-only Echo must not claim the typed permission contract"
16444        );
16445        assert_eq!(
16446            expected_permission_executor_mapping(
16447                &ExecutorSpec::CliAdapter {
16448                    command: "must-not-appear-in-contract".to_string(),
16449                    args: vec!["credential-like-argument".to_string()],
16450                },
16451                default,
16452                false,
16453            )
16454            .unwrap(),
16455            None,
16456            "unimplemented CliAdapter must not leak command data into a contract"
16457        );
16458
16459        let claude = ExecutorSpec::ClaudeCode {
16460            binary: None,
16461            model: None,
16462            permission_mode: Some("default".to_string()),
16463            inherit_user_config: None,
16464            forward_env: None,
16465        };
16466        for (resolution, mapping) in [
16467            (
16468                permission_resolution(SessionPermissionMode::Default, PermissionMode::Plan),
16469                "claude_code:permission_mode=plan",
16470            ),
16471            (
16472                permission_resolution(SessionPermissionMode::Auto, PermissionMode::Auto),
16473                "claude_code:permission_mode=bypassPermissions",
16474            ),
16475            (
16476                permission_resolution(SessionPermissionMode::Default, PermissionMode::AcceptEdits),
16477                "claude_code:permission_mode=acceptEdits",
16478            ),
16479            (
16480                permission_resolution(SessionPermissionMode::Default, PermissionMode::DontAsk),
16481                "claude_code:permission_mode=dontAsk",
16482            ),
16483            (
16484                permission_resolution(
16485                    SessionPermissionMode::Bypass,
16486                    PermissionMode::BypassPermissions,
16487                ),
16488                "claude_code:permission_mode=default",
16489            ),
16490        ] {
16491            assert_eq!(
16492                expected_permission_executor_mapping(&claude, resolution, false)
16493                    .unwrap()
16494                    .as_deref(),
16495                Some(mapping)
16496            );
16497        }
16498        assert_eq!(
16499            expected_permission_executor_mapping(&claude, default, true)
16500                .unwrap()
16501                .as_deref(),
16502            Some("claude_code:blocked_explicit_deny")
16503        );
16504
16505        let mut codex_exec = codex_executor(Some("inherit"), Some(true));
16506        if let ExecutorSpec::Codex {
16507            approval_policy, ..
16508        } = &mut codex_exec
16509        {
16510            *approval_policy = Some("on-failure".to_string());
16511        }
16512        assert_eq!(
16513            expected_permission_executor_mapping(&codex_exec, default, false)
16514                .unwrap()
16515                .as_deref(),
16516            Some("codex_exec:approval_policy=on-failure")
16517        );
16518        assert_eq!(
16519            expected_permission_executor_mapping(
16520                &codex_exec,
16521                permission_resolution(SessionPermissionMode::Auto, PermissionMode::Auto),
16522                false,
16523            )
16524            .unwrap()
16525            .as_deref(),
16526            Some("codex_exec:approval_policy=never")
16527        );
16528        assert_eq!(
16529            expected_permission_executor_mapping(&codex_exec, default, true)
16530                .unwrap()
16531                .as_deref(),
16532            Some("codex_exec:blocked_explicit_deny")
16533        );
16534
16535        let mut codex_app_server = codex_executor(Some("inherit"), Some(true));
16536        if let ExecutorSpec::Codex {
16537            mode,
16538            approval_policy,
16539            ..
16540        } = &mut codex_app_server
16541        {
16542            *mode = Some("app_server".to_string());
16543            *approval_policy = Some("on-request".to_string());
16544        }
16545        assert_eq!(
16546            expected_permission_executor_mapping(&codex_app_server, default, false)
16547                .unwrap()
16548                .as_deref(),
16549            Some("codex_app_server:approvalPolicy=on-request")
16550        );
16551        assert_eq!(
16552            expected_permission_executor_mapping(
16553                &codex_app_server,
16554                permission_resolution(SessionPermissionMode::Auto, PermissionMode::Auto),
16555                false,
16556            )
16557            .unwrap()
16558            .as_deref(),
16559            Some("codex_app_server:approvalPolicy=never")
16560        );
16561        assert_eq!(
16562            expected_permission_executor_mapping(&codex_app_server, default, true)
16563                .unwrap()
16564                .as_deref(),
16565            Some("codex_app_server:blocked_explicit_deny")
16566        );
16567    }
16568
16569    #[test]
16570    fn only_bamboo_managed_non_git_workspaces_are_marked_owned() {
16571        let project = tempfile::tempdir().unwrap();
16572        let managed = project.path().join(".bamboo/worktree/child-571");
16573        std::fs::create_dir_all(&managed).unwrap();
16574        assert!(!workspace_is_bamboo_owned(managed.to_str().unwrap()));
16575        let marker = project
16576            .path()
16577            .join(".bamboo/worktree/.bamboo-owned/child-571");
16578        std::fs::create_dir_all(marker.parent().unwrap()).unwrap();
16579        std::fs::write(&marker, "bamboo/child-571").unwrap();
16580        assert!(workspace_is_bamboo_owned(managed.to_str().unwrap()));
16581        let nested = managed.join("nested/path");
16582        std::fs::create_dir_all(&nested).unwrap();
16583        assert!(workspace_is_bamboo_owned(nested.to_str().unwrap()));
16584
16585        let arbitrary = tempfile::tempdir().unwrap();
16586        assert!(!workspace_is_bamboo_owned(
16587            arbitrary.path().to_str().unwrap()
16588        ));
16589    }
16590
16591    #[test]
16592    fn bamboo_codex_token_is_per_run_redacted_and_revoked_on_guard_drop() {
16593        let authority = Arc::new(RecordingCodexTokenAuthority::default());
16594        let authority_dyn: Arc<dyn CodexRunTokenAuthority> = authority.clone();
16595
16596        let (secrets, guard) = build_codex_run_secrets(
16597            &codex_executor(Some("bamboo"), None),
16598            Some(authority_dyn),
16599            "child-570",
16600        )
16601        .unwrap();
16602
16603        let token = secrets
16604            .codex_provider_token
16605            .as_ref()
16606            .expect("bamboo mode mints a token");
16607        assert_eq!(token.expose(), "bcx1_secret-child-570");
16608        assert!(!format!("{token:?}").contains("secret-child-570"));
16609        assert_eq!(
16610            authority.issued_for.lock().unwrap().as_slice(),
16611            ["child-570"]
16612        );
16613        assert!(authority.revoked.lock().unwrap().is_empty());
16614
16615        drop(guard);
16616        assert_eq!(
16617            authority.revoked.lock().unwrap().as_slice(),
16618            ["id-child-570"]
16619        );
16620    }
16621
16622    #[test]
16623    fn non_bamboo_codex_never_mints_and_bamboo_fails_closed_without_authority() {
16624        let authority = Arc::new(RecordingCodexTokenAuthority::default());
16625        let authority_dyn: Arc<dyn CodexRunTokenAuthority> = authority.clone();
16626        let (secrets, guard) = build_codex_run_secrets(
16627            &codex_executor(Some("custom"), None),
16628            Some(authority_dyn),
16629            "child-custom",
16630        )
16631        .unwrap();
16632        assert!(secrets.codex_provider_token.is_none());
16633        assert!(guard.is_none());
16634        assert!(authority.issued_for.lock().unwrap().is_empty());
16635
16636        let error = build_codex_run_secrets(
16637            &codex_executor(Some("bamboo"), None),
16638            None,
16639            "child-no-authority",
16640        )
16641        .err()
16642        .expect("bamboo mode without an authority must fail closed");
16643        assert!(error.to_string().contains("per-run token authority"));
16644    }
16645
16646    #[test]
16647    fn codex_provisioning_never_leaks_the_session_provider_credential() {
16648        let credentials = vec![ScopedCredential {
16649            provider: "openai".to_string(),
16650            api_key: "upstream-secret-must-not-cross".to_string(),
16651            base_url: None,
16652            provider_type: Some("openai".to_string()),
16653            credential_ref: Some("provider.openai.api_key".to_string()),
16654        }];
16655
16656        for (mode, label) in [
16657            (Some("inherit"), "inherit"),
16658            (Some("api_key"), "api_key"),
16659            (Some("bamboo"), "bamboo"),
16660            (None, "default-bamboo"),
16661        ] {
16662            let runner = ActorChildRunner::new(
16663                format!("codex-{label}-test"),
16664                PathBuf::from("/bin/false"),
16665                Vec::new(),
16666                std::env::temp_dir().join(format!("bamboo-codex-{label}-570")),
16667                codex_executor(mode, None),
16668                credentials.clone(),
16669                "openai".to_string(),
16670                1,
16671            );
16672            let mut session = Session::new(format!("child-{label}"), "model");
16673            session.add_message(bamboo_agent_core::Message::user("test"));
16674            let spec = runner.build_spec(
16675                &session,
16676                &crate::runtime::execution::SpawnJob {
16677                    parent_session_id: "parent".to_string(),
16678                    child_session_id: format!("child-{label}"),
16679                    model: "gpt-5.4".to_string(),
16680                    disabled_tools: None,
16681                },
16682            );
16683            assert!(
16684                spec.secrets.provider_credentials.is_empty(),
16685                "{label} Codex must not receive the session provider key"
16686            );
16687        }
16688    }
16689
16690    #[test]
16691    fn non_codex_provisioning_still_receives_only_its_selected_provider_credential() {
16692        let credentials = vec![
16693            ScopedCredential {
16694                provider: "openai".to_string(),
16695                api_key: "selected-openai-secret".to_string(),
16696                base_url: None,
16697                provider_type: Some("openai".to_string()),
16698                credential_ref: Some("provider.openai.api_key".to_string()),
16699            },
16700            ScopedCredential {
16701                provider: "other".to_string(),
16702                api_key: "unrelated-secret".to_string(),
16703                base_url: None,
16704                provider_type: Some("openai".to_string()),
16705                credential_ref: Some("provider.other.api_key".to_string()),
16706            },
16707        ];
16708        let runner = ActorChildRunner::new(
16709            "echo-test".to_string(),
16710            PathBuf::from("/bin/false"),
16711            Vec::new(),
16712            std::env::temp_dir().join("bamboo-echo-provider-570"),
16713            ExecutorSpec::Echo,
16714            credentials,
16715            "openai".to_string(),
16716            1,
16717        );
16718        let spec = runner.build_spec(
16719            &Session::new("child-echo", "model"),
16720            &crate::runtime::execution::SpawnJob {
16721                parent_session_id: "parent".to_string(),
16722                child_session_id: "child-echo".to_string(),
16723                model: "gpt-5.4".to_string(),
16724                disabled_tools: None,
16725            },
16726        );
16727
16728        assert_eq!(spec.secrets.provider_credentials.len(), 1);
16729        assert_eq!(
16730            spec.secrets.provider_credentials[0].api_key,
16731            "selected-openai-secret"
16732        );
16733    }
16734
16735    #[tokio::test]
16736    async fn live_provider_config_refreshes_credentials_and_default_per_activation() {
16737        fn provider_instance(
16738            provider_type: &str,
16739            api_key: &str,
16740        ) -> bamboo_config::ProviderInstanceConfig {
16741            bamboo_config::ProviderInstanceConfig {
16742                provider_type: provider_type.to_string(),
16743                label: None,
16744                api_key: api_key.to_string(),
16745                api_key_encrypted: None,
16746                credential_ref: None,
16747                base_url: None,
16748                model: None,
16749                fast_model: None,
16750                vision_model: None,
16751                reasoning_effort: None,
16752                responses_only_models: Vec::new(),
16753                request_overrides: None,
16754                enabled: true,
16755                extra: Default::default(),
16756            }
16757        }
16758
16759        let mut initial = bamboo_llm::Config::default();
16760        initial.provider_instances.insert(
16761            "old-provider".to_string(),
16762            provider_instance("openai", "old-key"),
16763        );
16764        initial.default_provider_instance = Some("old-provider".to_string());
16765        let live = Arc::new(tokio::sync::RwLock::new(initial));
16766
16767        let runner = ActorChildRunner::new(
16768            "echo-live-config-test".to_string(),
16769            PathBuf::from("/bin/false"),
16770            Vec::new(),
16771            std::env::temp_dir().join("bamboo-echo-live-config"),
16772            ExecutorSpec::Echo,
16773            Vec::new(),
16774            "stale-provider".to_string(),
16775            1,
16776        )
16777        .with_live_provider_config(live.clone());
16778        let job = crate::runtime::execution::SpawnJob {
16779            parent_session_id: "parent".to_string(),
16780            child_session_id: "child".to_string(),
16781            model: "model".to_string(),
16782            disabled_tools: None,
16783        };
16784
16785        let first = runner
16786            .build_live_spec(&Session::new("child", "model"), &job)
16787            .await;
16788        assert_eq!(
16789            first.model.as_ref().map(|model| model.provider.as_str()),
16790            Some("old-provider")
16791        );
16792        assert_eq!(first.secrets.provider_credentials.len(), 1);
16793        assert_eq!(first.secrets.provider_credentials[0].api_key, "old-key");
16794
16795        let mut reloaded = bamboo_llm::Config::default();
16796        reloaded.provider_instances.insert(
16797            "new-provider".to_string(),
16798            provider_instance("openai", "new-key"),
16799        );
16800        reloaded.default_provider_instance = Some("new-provider".to_string());
16801        *live.write().await = reloaded;
16802
16803        let second = runner
16804            .build_live_spec(&Session::new("child", "model"), &job)
16805            .await;
16806        assert_eq!(
16807            second.model.as_ref().map(|model| model.provider.as_str()),
16808            Some("new-provider")
16809        );
16810        assert_eq!(second.secrets.provider_credentials.len(), 1);
16811        assert_eq!(second.secrets.provider_credentials[0].api_key, "new-key");
16812    }
16813
16814    #[test]
16815    fn custom_codex_provisioning_scopes_only_the_referenced_credential() {
16816        let mut executor = codex_executor(Some("custom"), None);
16817        if let ExecutorSpec::Codex {
16818            base_url,
16819            provider_key_ref,
16820            ..
16821        } = &mut executor
16822        {
16823            *base_url = Some("https://provider.example/v1".to_string());
16824            *provider_key_ref = Some("provider.custom.api_key".to_string());
16825        }
16826        let credentials = vec![
16827            ScopedCredential {
16828                provider: "openai".to_string(),
16829                api_key: "session-provider-secret".to_string(),
16830                base_url: None,
16831                provider_type: Some("openai".to_string()),
16832                credential_ref: Some("provider.openai.api_key".to_string()),
16833            },
16834            ScopedCredential {
16835                provider: "custom".to_string(),
16836                api_key: "selected-secret".to_string(),
16837                base_url: None,
16838                provider_type: Some("openai".to_string()),
16839                credential_ref: Some("provider.custom.api_key".to_string()),
16840            },
16841            ScopedCredential {
16842                provider: "other".to_string(),
16843                api_key: "unrelated-secret".to_string(),
16844                base_url: None,
16845                provider_type: Some("openai".to_string()),
16846                credential_ref: Some("provider.other.api_key".to_string()),
16847            },
16848        ];
16849        let runner = ActorChildRunner::new(
16850            "codex-test".to_string(),
16851            PathBuf::from("/bin/false"),
16852            Vec::new(),
16853            std::env::temp_dir().join("bamboo-codex-570"),
16854            executor,
16855            credentials,
16856            "openai".to_string(),
16857            1,
16858        );
16859        let mut session = Session::new("child-custom", "model");
16860        session.add_message(bamboo_agent_core::Message::user("test"));
16861        let spec = runner.build_spec(
16862            &session,
16863            &crate::runtime::execution::SpawnJob {
16864                parent_session_id: "parent".to_string(),
16865                child_session_id: "child-custom".to_string(),
16866                model: "gpt-5.4".to_string(),
16867                disabled_tools: None,
16868            },
16869        );
16870
16871        assert_eq!(spec.secrets.provider_credentials.len(), 1);
16872        assert_eq!(
16873            spec.secrets.provider_credentials[0]
16874                .credential_ref
16875                .as_deref(),
16876            Some("provider.custom.api_key")
16877        );
16878        assert_eq!(
16879            spec.secrets.provider_credentials[0].api_key,
16880            "selected-secret"
16881        );
16882    }
16883
16884    fn spec_with(
16885        role: &str,
16886        provider: &str,
16887        model: &str,
16888        workspace: Option<&str>,
16889        disabled: Option<Vec<&str>>,
16890    ) -> ProvisionSpec {
16891        let mut spec = ProvisionSpec::new(
16892            ChildIdentity {
16893                child_id: "c".into(),
16894                parent_id: None,
16895                project_key: None,
16896                role: role.into(),
16897                depth: 0,
16898            },
16899            ExecutorSpec::Echo,
16900            "/tmp/fab".into(),
16901        );
16902        spec.workspace = workspace.map(|w| w.to_string());
16903        spec.model = Some(ModelRefSpec {
16904            provider: provider.into(),
16905            model: model.into(),
16906        });
16907        spec.disabled_tools = disabled.map(|d| d.into_iter().map(String::from).collect());
16908        spec
16909    }
16910
16911    #[test]
16912    fn fingerprint_matches_interchangeable_children() {
16913        // Same role/provider/model/workspace and equal tool sets (order-insensitive)
16914        // are interchangeable on one warm worker — and differ only in child_id.
16915        let a = spec_with(
16916            "explorer",
16917            "p",
16918            "m",
16919            Some("/ws"),
16920            Some(vec!["Bash", "Edit"]),
16921        );
16922        let mut b = spec_with(
16923            "explorer",
16924            "p",
16925            "m",
16926            Some("/ws"),
16927            Some(vec!["Edit", "Bash"]),
16928        );
16929        b.identity.child_id = "other".into();
16930        assert_eq!(
16931            ActorChildRunner::fingerprint(&a),
16932            ActorChildRunner::fingerprint(&b)
16933        );
16934    }
16935
16936    #[test]
16937    fn logical_identity_is_invariant_across_local_remote_scheduled_and_warm_reuse() {
16938        let mut session =
16939            Session::new_child("logical-child-681", "logical-parent-681", "model", "child");
16940        session.root_session_id = "logical-root-681".to_string();
16941        let job = SpawnJob {
16942            parent_session_id: "logical-parent-681".to_string(),
16943            child_session_id: "logical-child-681".to_string(),
16944            model: "model".to_string(),
16945            disabled_tools: None,
16946        };
16947        let expected = LogicalSessionIdentity {
16948            creation: None,
16949            session_id: "logical-child-681".to_string(),
16950            parent_session_id: Some("logical-parent-681".to_string()),
16951            root_session_id: "logical-root-681".to_string(),
16952        };
16953
16954        let placements_and_transport_ids = [
16955            (Placement::Local, "local-mailbox-first"),
16956            (
16957                Placement::Remote {
16958                    endpoint: "wss://remote.example/actor".to_string(),
16959                },
16960                "remote-process-44",
16961            ),
16962            (
16963                Placement::Schedulable {
16964                    pool: "gpu-pool".to_string(),
16965                },
16966                "scheduled-mailbox-9",
16967            ),
16968            // Same logical child reactivated on a different pooled mailbox.
16969            (Placement::Local, "warm-mailbox-reused-77"),
16970        ];
16971        for (placement, transport_id) in placements_and_transport_ids {
16972            let mut provision = spec_with("worker", "provider", "model", None, None);
16973            provision.placement = placement;
16974            provision.identity.child_id = transport_id.to_string();
16975            assert_eq!(logical_identity_for_actor_run(&session, &job), expected);
16976            assert_ne!(
16977                provision.identity.child_id, expected.session_id,
16978                "test fixture must prove transport identity is independent"
16979            );
16980        }
16981        let mut original = spec_with("worker", "provider", "model", Some("/ws"), None);
16982        original.executor = ExecutorSpec::BambooRuntime;
16983        original.identity.depth = session.spawn_depth;
16984        original.capabilities.required_child_context = true;
16985        original.capabilities.child_creation_identity = true;
16986        original.capabilities.enforce_permissions = true;
16987        original.capabilities.native_tool_ceiling_required = true;
16988        original.capabilities.native_tool_ceiling =
16989            Some(bamboo_subagent::proto::NativeToolCeiling {
16990                version: 1,
16991                child_session_id: session.id.clone(),
16992                parent_session_id: job.parent_session_id.clone(),
16993                root_session_id: session.root_session_id.clone(),
16994                created_at: session.created_at,
16995                spawn_depth: session.spawn_depth,
16996                project_id: None,
16997                tools: vec![],
16998            });
16999        original.validate().unwrap();
17000        let first = ActorChildRunner::fresh_required_worker_spec(&original);
17001        let second = ActorChildRunner::fresh_required_worker_spec(&original);
17002        assert_ne!(first.identity.child_id, second.identity.child_id);
17003        assert_ne!(first.identity.child_id, original.identity.child_id);
17004        for physical in [first, second] {
17005            physical.validate().unwrap();
17006            let mut actual = serde_json::to_value(&physical).unwrap();
17007            actual["identity"]["child_id"] = serde_json::json!(original.identity.child_id);
17008            assert_eq!(actual, serde_json::to_value(&original).unwrap());
17009            assert_eq!(logical_identity_for_actor_run(&session, &job), expected);
17010        }
17011    }
17012
17013    #[test]
17014    fn fingerprint_separates_distinct_runtimes() {
17015        let base = spec_with("explorer", "p", "m", Some("/ws"), None);
17016        let base_fp = ActorChildRunner::fingerprint(&base);
17017        // Each axis that is baked into the worker must split the pool bucket.
17018        assert_ne!(
17019            base_fp,
17020            ActorChildRunner::fingerprint(&spec_with("writer", "p", "m", Some("/ws"), None))
17021        );
17022        assert_ne!(
17023            base_fp,
17024            ActorChildRunner::fingerprint(&spec_with("explorer", "p2", "m", Some("/ws"), None))
17025        );
17026        assert_ne!(
17027            base_fp,
17028            ActorChildRunner::fingerprint(&spec_with("explorer", "p", "m2", Some("/ws"), None))
17029        );
17030        assert_ne!(
17031            base_fp,
17032            ActorChildRunner::fingerprint(&spec_with("explorer", "p", "m", Some("/ws2"), None))
17033        );
17034        assert_ne!(
17035            base_fp,
17036            ActorChildRunner::fingerprint(&spec_with(
17037                "explorer",
17038                "p",
17039                "m",
17040                Some("/ws"),
17041                Some(vec!["Bash"])
17042            ))
17043        );
17044    }
17045
17046    #[test]
17047    fn fingerprint_splits_rotated_provider_credentials_without_exposing_them() {
17048        let mut before = spec_with("explorer", "provider", "model", Some("/ws"), None);
17049        before.secrets.provider_credentials.push(ScopedCredential {
17050            provider: "provider".to_string(),
17051            api_key: "old-secret-key".to_string(),
17052            base_url: Some("https://provider.example/v1".to_string()),
17053            provider_type: Some("openai".to_string()),
17054            credential_ref: None,
17055        });
17056        let mut after = before.clone();
17057        after.secrets.provider_credentials[0].api_key = "new-secret-key".to_string();
17058
17059        let before_fingerprint = ActorChildRunner::fingerprint(&before);
17060        let after_fingerprint = ActorChildRunner::fingerprint(&after);
17061        assert_ne!(before_fingerprint, after_fingerprint);
17062        assert!(!before_fingerprint.contains("old-secret-key"));
17063        assert!(!after_fingerprint.contains("new-secret-key"));
17064    }
17065
17066    #[test]
17067    fn fingerprint_splits_on_baked_capabilities() {
17068        // Every capability baked once at provision time must split the pool
17069        // bucket, else a worker baked for one posture gets reused for another
17070        // (e.g. a depth-1 worker re-stamping spawn_depth onto a depth-4 child,
17071        // breaking the depth cap; or a bypass worker reused for a non-bypass one).
17072        let base_fp =
17073            ActorChildRunner::fingerprint(&spec_with("explorer", "p", "m", Some("/ws"), None));
17074
17075        let mut creation = spec_with("explorer", "p", "m", Some("/ws"), None);
17076        creation.capabilities.child_creation_identity = true;
17077        let typed = ActorChildRunner::fingerprint(&creation);
17078        assert_ne!(
17079            base_fp, typed,
17080            "legacy worker pool cannot satisfy birth-required Run"
17081        );
17082        creation.identity.child_id = "different-logical-child".into();
17083        assert_eq!(
17084            typed,
17085            ActorChildRunner::fingerprint(&creation),
17086            "birth-required mode retains cross-Child pool reuse"
17087        );
17088
17089        let mut depth = spec_with("explorer", "p", "m", Some("/ws"), None);
17090        depth.identity.depth = 2;
17091        assert_ne!(
17092            base_fp,
17093            ActorChildRunner::fingerprint(&depth),
17094            "depth must split"
17095        );
17096
17097        let mut nested = spec_with("explorer", "p", "m", Some("/ws"), None);
17098        nested.capabilities.nested_spawn = true;
17099        assert_ne!(
17100            base_fp,
17101            ActorChildRunner::fingerprint(&nested),
17102            "nested_spawn must split"
17103        );
17104
17105        let mut bypass = spec_with("explorer", "p", "m", Some("/ws"), None);
17106        bypass.capabilities.bypass = true;
17107        assert_ne!(
17108            base_fp,
17109            ActorChildRunner::fingerprint(&bypass),
17110            "bypass must split"
17111        );
17112
17113        let mut auto = spec_with("explorer", "p", "m", Some("/ws"), None);
17114        auto.capabilities.auto_approve_permissions = true;
17115        assert_ne!(
17116            base_fp,
17117            ActorChildRunner::fingerprint(&auto),
17118            "auto_approve_permissions must split"
17119        );
17120
17121        let mut global_auto = spec_with("explorer", "p", "m", Some("/ws"), None);
17122        global_auto.capabilities.permission_requested_mode = "default".to_string();
17123        global_auto.capabilities.permission_effective_mode = "auto".to_string();
17124        global_auto.capabilities.auto_approve_permissions = true;
17125        let mut explicit_auto = global_auto.clone();
17126        explicit_auto.capabilities.permission_requested_mode = "auto".to_string();
17127        assert_ne!(
17128            ActorChildRunner::fingerprint(&global_auto),
17129            ActorChildRunner::fingerprint(&explicit_auto),
17130            "permission_requested_mode must split global and explicit Auto"
17131        );
17132
17133        let mut plan_overlay = explicit_auto.clone();
17134        plan_overlay.capabilities.permission_effective_mode = "plan".to_string();
17135        assert_ne!(
17136            ActorChildRunner::fingerprint(&explicit_auto),
17137            ActorChildRunner::fingerprint(&plan_overlay),
17138            "permission_effective_mode must split Plan overlay from Auto"
17139        );
17140
17141        let mut enforce = spec_with("explorer", "p", "m", Some("/ws"), None);
17142        enforce.capabilities.enforce_permissions = true;
17143        assert_ne!(
17144            base_fp,
17145            ActorChildRunner::fingerprint(&enforce),
17146            "enforce_permissions must split"
17147        );
17148
17149        let mut cap = spec_with("explorer", "p", "m", Some("/ws"), None);
17150        cap.capabilities.max_spawn_depth = Some(8);
17151        assert_ne!(
17152            base_fp,
17153            ActorChildRunner::fingerprint(&cap),
17154            "max_spawn_depth must split"
17155        );
17156
17157        // #73 (P1): the worker bakes `no_human_review` from this flag once at
17158        // build(), so it MUST split the pool or a worker baked for one approval
17159        // posture is reused for the opposite one.
17160        let mut nha = spec_with("explorer", "p", "m", Some("/ws"), None);
17161        nha.capabilities.no_human_approver = true;
17162        assert_ne!(
17163            base_fp,
17164            ActorChildRunner::fingerprint(&nha),
17165            "no_human_approver must split"
17166        );
17167
17168        // The generic read-only no-shell checker is baked once at build(), so a
17169        // planner/Guardian worker must not be reused for an ordinary child.
17170        let mut read_only = spec_with("explorer", "p", "m", Some("/ws"), None);
17171        read_only.capabilities.read_only = true;
17172        assert_ne!(
17173            base_fp,
17174            ActorChildRunner::fingerprint(&read_only),
17175            "read_only must split"
17176        );
17177
17178        // Preserve the legacy Guardian bit as its own rolling-upgrade axis.
17179        let mut gro = spec_with("explorer", "p", "m", Some("/ws"), None);
17180        gro.capabilities.guardian_read_only = true;
17181        assert_ne!(
17182            base_fp,
17183            ActorChildRunner::fingerprint(&gro),
17184            "guardian_read_only must split"
17185        );
17186    }
17187
17188    #[test]
17189    fn fingerprint_splits_codex_exec_and_app_server_workers() {
17190        let mut exec = spec_with("explorer", "p", "m", Some("/ws"), None);
17191        exec.executor = codex_executor(Some("inherit"), None);
17192        let mut app_server = exec.clone();
17193        if let ExecutorSpec::Codex { mode, .. } = &mut app_server.executor {
17194            *mode = Some("app_server".to_string());
17195        }
17196        assert_ne!(
17197            ActorChildRunner::fingerprint(&exec),
17198            ActorChildRunner::fingerprint(&app_server)
17199        );
17200    }
17201
17202    struct StaticDecider(bool);
17203
17204    #[async_trait]
17205    impl ChildApprovalDecider for StaticDecider {
17206        async fn decide(&self, _child: &str, _req: &serde_json::Value) -> bool {
17207            self.0
17208        }
17209    }
17210
17211    struct RecordingReviewer {
17212        reviewed: mpsc::UnboundedSender<(String, String, serde_json::Value)>,
17213    }
17214
17215    #[async_trait]
17216    impl ChildApprovalReviewer for RecordingReviewer {
17217        async fn review(&self, parent: &str, child: &str, request: &serde_json::Value) -> bool {
17218            let _ = self
17219                .reviewed
17220                .send((parent.to_string(), child.to_string(), request.clone()));
17221            true
17222        }
17223    }
17224
17225    // ---- first-frame watchdog (dead-pooled-worker recovery) -----------------
17226
17227    /// A link that never yields a frame — models a worker that died (or never
17228    /// subscribed) so its Run sits queued with no server.
17229    struct SilentLink;
17230    #[async_trait]
17231    impl bamboo_subagent::ChildLink for SilentLink {
17232        async fn send(&mut self, _: ParentFrame) -> bamboo_subagent::TransportResult<()> {
17233            Ok(())
17234        }
17235        async fn next_frame(&mut self) -> bamboo_subagent::TransportResult<Option<ChildFrame>> {
17236            std::future::pending().await
17237        }
17238    }
17239
17240    /// A link that immediately yields one terminal frame (a healthy fast worker).
17241    struct InstantTerminalLink {
17242        done: bool,
17243    }
17244
17245    struct ApprovalRoundTripLink {
17246        step: u8,
17247        approval_reply: Option<(String, bool)>,
17248    }
17249
17250    #[async_trait]
17251    impl bamboo_subagent::ChildLink for ApprovalRoundTripLink {
17252        async fn send(&mut self, frame: ParentFrame) -> bamboo_subagent::TransportResult<()> {
17253            if let ParentFrame::ApprovalReply { id, approved } = frame {
17254                self.approval_reply = Some((id, approved));
17255                self.step = 2;
17256            }
17257            Ok(())
17258        }
17259
17260        async fn next_frame(&mut self) -> bamboo_subagent::TransportResult<Option<ChildFrame>> {
17261            match self.step {
17262                0 => {
17263                    self.step = 1;
17264                    Ok(Some(ChildFrame::ApprovalRequest {
17265                        id: "approval-1".into(),
17266                        body: serde_json::json!({
17267                            "tool_name": "Bash",
17268                            "permission": "execute",
17269                            "resource": "rm -rf target",
17270                            "permission_request": {"reason_code": "hard_dangerous"}
17271                        }),
17272                    }))
17273                }
17274                1 => std::future::pending().await,
17275                2 => {
17276                    self.step = 3;
17277                    Ok(Some(ChildFrame::Terminal {
17278                        status: TerminalStatus::Completed,
17279                        result: Some("done".into()),
17280                        error: None,
17281                        final_event_watermark: None,
17282                        transcript: vec![],
17283                    }))
17284                }
17285                _ => std::future::pending().await,
17286            }
17287        }
17288    }
17289
17290    #[tokio::test]
17291    async fn drive_routes_forced_ask_to_parent_reviewer_without_human_event() {
17292        let (event_tx, mut event_rx) = mpsc::channel::<AgentEvent>(8);
17293        let (review_tx, mut review_rx) = mpsc::unbounded_channel();
17294        let reviewer: Arc<dyn ChildApprovalReviewer> = Arc::new(RecordingReviewer {
17295            reviewed: review_tx,
17296        });
17297        let cancel = CancellationToken::new();
17298        let (live_tx, mut live_rx) = mpsc::unbounded_channel::<ParentFrame>();
17299        let (_delivery_tx, mut delivery_rx) = mpsc::unbounded_channel();
17300        let mut logical_session = Session::new("child-reviewer", "model");
17301        let live_guard = crate::external_agents::live::register("child-reviewer", live_tx, 0, None);
17302        let mut link = ApprovalRoundTripLink {
17303            step: 0,
17304            approval_reply: None,
17305        };
17306
17307        let result = tokio::time::timeout(
17308            Duration::from_secs(1),
17309            drive(ActorDriveContext {
17310                expected_creation: None,
17311                client: &mut link,
17312                parent_session_id: "parent-reviewer",
17313                child_session_id: "child-reviewer",
17314                child_attempt: 0,
17315                approval_registry: None,
17316                approval_decider: None,
17317                approval_reviewer: Some(&reviewer),
17318                escalation_bridge: None,
17319                event_tx: &event_tx,
17320                cancel_token: &cancel,
17321                live_rx: &mut live_rx,
17322                delivery_rx: &mut delivery_rx,
17323                logical_session: &mut logical_session,
17324                expected_permission_posture: None,
17325                session_inbox_runtime: None,
17326                actor_directory_store: None,
17327                canonical_subagent_tool: None,
17328                ticket_service: None,
17329                activation_run_id: None,
17330                execution_epoch: 0,
17331                expected_source_actor_id: "child-reviewer",
17332                initial_inflight_claims: VecDeque::new(),
17333                plain_actor: false,
17334                remote_environment_lease: false,
17335                readonly_output: None,
17336                local_history_tools: None,
17337                local_history_read_only: false,
17338                strict_history_output: None,
17339                plain_input: None,
17340                canonical_activation: None,
17341                canonical_placement_ref: None,
17342                actor_event_observer: None,
17343                plain_run: None,
17344                first_frame_timeout: None,
17345            }),
17346        )
17347        .await
17348        .expect("worker must receive the reviewer verdict before terminating");
17349
17350        assert_eq!(result.ok().flatten().as_deref(), Some("done"));
17351        assert_eq!(
17352            link.approval_reply,
17353            Some(("approval-1".to_string(), true)),
17354            "reviewer verdict must traverse the live route back to the worker"
17355        );
17356        let (parent, child, body) = tokio::time::timeout(Duration::from_secs(1), review_rx.recv())
17357            .await
17358            .expect("reviewer should be invoked off-loop")
17359            .expect("review channel should remain open");
17360        assert_eq!(parent, "parent-reviewer");
17361        assert_eq!(child, "child-reviewer");
17362        assert_eq!(
17363            body.pointer("/permission_request/reason_code")
17364                .and_then(serde_json::Value::as_str),
17365            Some("hard_dangerous")
17366        );
17367        assert!(
17368            event_rx.try_recv().is_err(),
17369            "must not emit a human-review event"
17370        );
17371        drop(live_guard);
17372    }
17373
17374    #[tokio::test]
17375    async fn drive_denies_forced_ask_without_parent_reviewer_or_manual_event() {
17376        let (event_tx, mut event_rx) = mpsc::channel::<AgentEvent>(8);
17377        let cancel = CancellationToken::new();
17378        let (_live_tx, mut live_rx) = mpsc::unbounded_channel::<ParentFrame>();
17379        let (_delivery_tx, mut delivery_rx) = mpsc::unbounded_channel();
17380        let mut logical_session = Session::new("child-no-reviewer", "model");
17381        let mut link = ApprovalRoundTripLink {
17382            step: 0,
17383            approval_reply: None,
17384        };
17385
17386        let result = tokio::time::timeout(
17387            Duration::from_secs(1),
17388            drive(ActorDriveContext {
17389                expected_creation: None,
17390                client: &mut link,
17391                parent_session_id: "parent-no-reviewer",
17392                child_session_id: "child-no-reviewer",
17393                child_attempt: 0,
17394                approval_registry: None,
17395                approval_decider: None,
17396                approval_reviewer: None,
17397                escalation_bridge: None,
17398                event_tx: &event_tx,
17399                cancel_token: &cancel,
17400                live_rx: &mut live_rx,
17401                delivery_rx: &mut delivery_rx,
17402                logical_session: &mut logical_session,
17403                expected_permission_posture: None,
17404                session_inbox_runtime: None,
17405                actor_directory_store: None,
17406                canonical_subagent_tool: None,
17407                ticket_service: None,
17408                activation_run_id: None,
17409                execution_epoch: 0,
17410                expected_source_actor_id: "child-no-reviewer",
17411                initial_inflight_claims: VecDeque::new(),
17412                plain_actor: false,
17413                remote_environment_lease: false,
17414                readonly_output: None,
17415                local_history_tools: None,
17416                local_history_read_only: false,
17417                strict_history_output: None,
17418                plain_input: None,
17419                canonical_activation: None,
17420                canonical_placement_ref: None,
17421                actor_event_observer: None,
17422                plain_run: None,
17423                first_frame_timeout: None,
17424            }),
17425        )
17426        .await
17427        .expect("fail-closed reply must unblock the child immediately");
17428
17429        assert_eq!(result.ok().flatten().as_deref(), Some("done"));
17430        assert_eq!(link.approval_reply, Some(("approval-1".to_string(), false)));
17431        assert!(
17432            event_rx.try_recv().is_err(),
17433            "missing parent review must not surface a manual approval event"
17434        );
17435    }
17436    #[async_trait]
17437    impl bamboo_subagent::ChildLink for InstantTerminalLink {
17438        async fn send(&mut self, _: ParentFrame) -> bamboo_subagent::TransportResult<()> {
17439            Ok(())
17440        }
17441        async fn next_frame(&mut self) -> bamboo_subagent::TransportResult<Option<ChildFrame>> {
17442            if self.done {
17443                std::future::pending().await
17444            } else {
17445                self.done = true;
17446                Ok(Some(ChildFrame::Terminal {
17447                    status: TerminalStatus::Completed,
17448                    result: Some("done".into()),
17449                    error: None,
17450                    final_event_watermark: None,
17451                    transcript: vec![],
17452                }))
17453            }
17454        }
17455    }
17456
17457    #[tokio::test]
17458    async fn drive_trips_first_frame_watchdog_on_a_silent_worker() {
17459        let (event_tx, _rx) = mpsc::channel::<AgentEvent>(8);
17460        let cancel = CancellationToken::new();
17461        let (_live_tx, mut live_rx) = mpsc::unbounded_channel::<ParentFrame>();
17462        let (_delivery_tx, mut delivery_rx) = mpsc::unbounded_channel();
17463        let mut logical_session = Session::new("child-x", "model");
17464        let mut link = SilentLink;
17465        let r = drive(ActorDriveContext {
17466            expected_creation: None,
17467            client: &mut link,
17468            parent_session_id: "parent-x",
17469            child_session_id: "child-x",
17470            child_attempt: 0,
17471            approval_registry: None,
17472            approval_decider: None,
17473            approval_reviewer: None,
17474            escalation_bridge: None,
17475            event_tx: &event_tx,
17476            cancel_token: &cancel,
17477            live_rx: &mut live_rx,
17478            delivery_rx: &mut delivery_rx,
17479            logical_session: &mut logical_session,
17480            expected_permission_posture: None,
17481            session_inbox_runtime: None,
17482            actor_directory_store: None,
17483            canonical_subagent_tool: None,
17484            ticket_service: None,
17485            activation_run_id: None,
17486            execution_epoch: 0,
17487            expected_source_actor_id: "child-x",
17488            initial_inflight_claims: VecDeque::new(),
17489            plain_actor: false,
17490            remote_environment_lease: false,
17491            readonly_output: None,
17492            local_history_tools: None,
17493            local_history_read_only: false,
17494            strict_history_output: None,
17495            plain_input: None,
17496            canonical_activation: None,
17497            canonical_placement_ref: None,
17498            actor_event_observer: None,
17499            plain_run: None,
17500            first_frame_timeout: Some(Duration::from_millis(100)),
17501        })
17502        .await;
17503        assert!(
17504            matches!(r, Err(AgentError::WorkerUnresponsive(_))),
17505            "a silent worker must trip the first-frame watchdog, got {r:?}"
17506        );
17507    }
17508
17509    #[tokio::test]
17510    async fn drive_does_not_trip_when_a_frame_arrives() {
17511        let (event_tx, _rx) = mpsc::channel::<AgentEvent>(8);
17512        let cancel = CancellationToken::new();
17513        let (_live_tx, mut live_rx) = mpsc::unbounded_channel::<ParentFrame>();
17514        let (_delivery_tx, mut delivery_rx) = mpsc::unbounded_channel();
17515        let mut logical_session = Session::new("child-y", "model");
17516        let mut link = InstantTerminalLink { done: false };
17517        // Even a tiny timeout must NOT trip: the terminal frame arrives first and
17518        // disarms the watchdog.
17519        let r = drive(ActorDriveContext {
17520            expected_creation: None,
17521            client: &mut link,
17522            parent_session_id: "parent-y",
17523            child_session_id: "child-y",
17524            child_attempt: 0,
17525            approval_registry: None,
17526            approval_decider: None,
17527            approval_reviewer: None,
17528            escalation_bridge: None,
17529            event_tx: &event_tx,
17530            cancel_token: &cancel,
17531            live_rx: &mut live_rx,
17532            delivery_rx: &mut delivery_rx,
17533            logical_session: &mut logical_session,
17534            expected_permission_posture: None,
17535            session_inbox_runtime: None,
17536            actor_directory_store: None,
17537            canonical_subagent_tool: None,
17538            ticket_service: None,
17539            activation_run_id: None,
17540            execution_epoch: 0,
17541            expected_source_actor_id: "child-y",
17542            initial_inflight_claims: VecDeque::new(),
17543            plain_actor: false,
17544            remote_environment_lease: false,
17545            readonly_output: None,
17546            local_history_tools: None,
17547            local_history_read_only: false,
17548            strict_history_output: None,
17549            plain_input: None,
17550            canonical_activation: None,
17551            canonical_placement_ref: None,
17552            actor_event_observer: None,
17553            plain_run: None,
17554            first_frame_timeout: Some(Duration::from_millis(50)),
17555        })
17556        .await;
17557        assert_eq!(r.ok().flatten().as_deref(), Some("done"));
17558    }
17559
17560    #[tokio::test]
17561    async fn child_approval_fails_closed_without_decider() {
17562        // No decider wired ⇒ the host denies (safe default), unchanged behavior.
17563        let body = serde_json::json!({"tool_name":"Bash","permission":"run","resource":"rm -rf /"});
17564        assert!(!decide_child_approval(None, "child-1", &body).await);
17565    }
17566
17567    #[tokio::test]
17568    async fn child_approval_honors_wired_decider() {
17569        let body =
17570            serde_json::json!({"tool_name":"Write","permission":"write","resource":"/tmp/x"});
17571        let approve: Arc<dyn ChildApprovalDecider> = Arc::new(StaticDecider(true));
17572        let deny: Arc<dyn ChildApprovalDecider> = Arc::new(StaticDecider(false));
17573        assert!(decide_child_approval(Some(&approve), "child-1", &body).await);
17574        assert!(!decide_child_approval(Some(&deny), "child-1", &body).await);
17575    }
17576
17577    // ---- #193: remote placement routing -------------------------------------
17578
17579    use crate::runtime::execution::SpawnJob;
17580    use bamboo_agent_core::Session;
17581
17582    /// A runner with a BOGUS worker_bin (`/bin/false`): a local spawn here would
17583    /// FAIL, so a passing remote test proves the remote path never spawns.
17584    fn bogus_runner(placements: HashMap<String, ResolvedRemotePlacement>) -> ActorChildRunner {
17585        ActorChildRunner::new(
17586            "test-actor".into(),
17587            PathBuf::from("/bin/false"),
17588            vec![],
17589            std::env::temp_dir().join("bamboo-test-fab-193"),
17590            ExecutorSpec::Echo,
17591            vec![],
17592            "anthropic".into(),
17593            4,
17594        )
17595        .with_remote_placements(placements)
17596    }
17597
17598    /// A child session of the given role (the role rides `subagent_type`, the
17599    /// path build_spec + the remote lookup both read).
17600    fn session_of_role(role: &str, assignment: &str) -> Session {
17601        let mut s = Session::new("child-1", "test-model");
17602        s.metadata
17603            .insert("subagent_type".to_string(), role.to_string());
17604        s.add_message(bamboo_agent_core::Message::user(assignment));
17605        s
17606    }
17607
17608    fn job_for(child: &str) -> SpawnJob {
17609        SpawnJob {
17610            parent_session_id: "parent-1".into(),
17611            child_session_id: child.into(),
17612            model: String::new(),
17613            disabled_tools: None,
17614        }
17615    }
17616
17617    #[tokio::test]
17618    async fn spawn_depth_provisioning_uses_live_policy_for_local_and_remote_workers() {
17619        let live = Arc::new(tokio::sync::RwLock::new(bamboo_llm::Config::default()));
17620        live.write().await.subagents_mut().max_spawn_depth = Some(1);
17621        let mut remote = HashMap::new();
17622        remote.insert(
17623            "remote-role".into(),
17624            ResolvedRemotePlacement {
17625                endpoint: "wss://worker.example:8443".into(),
17626                token: None,
17627                ca_cert_file: None,
17628                host_label: None,
17629                broker_peer: None,
17630                requirements: None,
17631            },
17632        );
17633        let runner = bogus_runner(remote).with_live_provider_config(live.clone());
17634        let root = Session::new("provision-root", "model");
17635        for role in ["local-role", "remote-role"] {
17636            let mut child =
17637                Session::new_child_of(format!("provision-{role}"), &root, "model", "child");
17638            child.metadata.insert("subagent_type".into(), role.into());
17639            let first = runner.build_live_spec(&child, &job_for(&child.id)).await;
17640            assert_eq!(first.capabilities.max_spawn_depth, Some(1));
17641            assert!(!first.capabilities.nested_spawn);
17642            assert_eq!(
17643                matches!(first.placement, Placement::Remote { .. }),
17644                role == "remote-role"
17645            );
17646            live.write().await.subagents_mut().max_spawn_depth = Some(7);
17647            let second = runner.build_live_spec(&child, &job_for(&child.id)).await;
17648            assert_eq!(second.capabilities.max_spawn_depth, Some(7));
17649            assert!(second.capabilities.nested_spawn);
17650            assert_ne!(
17651                ActorChildRunner::fingerprint(&first),
17652                ActorChildRunner::fingerprint(&second)
17653            );
17654            assert_eq!(second.identity.depth, 1);
17655            live.write().await.subagents_mut().max_spawn_depth = Some(1);
17656        }
17657        let fallback = bogus_runner(HashMap::new()).with_max_spawn_depth(0);
17658        let spec = fallback.build_spec(&root, &job_for(&root.id));
17659        assert_eq!(spec.capabilities.max_spawn_depth, Some(0));
17660        assert!(!spec.capabilities.nested_spawn);
17661    }
17662
17663    struct RecordingChildSessionPort {
17664        parent: Session,
17665        saved: std::sync::Mutex<Option<Session>>,
17666    }
17667
17668    impl RecordingChildSessionPort {
17669        fn saved_child(&self) -> Session {
17670            self.saved
17671                .lock()
17672                .expect("saved-child fixture lock")
17673                .clone()
17674                .expect("create_child_action must save the child")
17675        }
17676    }
17677
17678    #[async_trait]
17679    impl crate::session_app::child_session::ChildSessionPort for RecordingChildSessionPort {
17680        async fn load_root_session(
17681            &self,
17682            root_id: &str,
17683        ) -> Result<Session, crate::session_app::child_session::ChildSessionError> {
17684            assert_eq!(root_id, self.parent.id);
17685            Ok(self.parent.clone())
17686        }
17687
17688        async fn load_child_for_parent(
17689            &self,
17690            _parent_id: &str,
17691            _child_id: &str,
17692        ) -> Result<Session, crate::session_app::child_session::ChildSessionError> {
17693            unreachable!("create_child_action does not reload the child")
17694        }
17695
17696        async fn save_child_session(
17697            &self,
17698            child: &mut Session,
17699        ) -> Result<(), crate::session_app::child_session::ChildSessionError> {
17700            *self.saved.lock().expect("saved-child fixture lock") = Some(child.clone());
17701            Ok(())
17702        }
17703
17704        async fn save_child_session_authoritative_flags(
17705            &self,
17706            _child: &mut Session,
17707        ) -> Result<(), crate::session_app::child_session::ChildSessionError> {
17708            unreachable!("new-child creation uses the ordinary save")
17709        }
17710
17711        async fn is_child_running(&self, _child_id: &str) -> bool {
17712            false
17713        }
17714
17715        async fn list_children(
17716            &self,
17717            _parent_id: &str,
17718        ) -> Vec<crate::session_app::child_session::ChildSessionEntry> {
17719            Vec::new()
17720        }
17721
17722        async fn enqueue_child_run(
17723            &self,
17724            _parent: &Session,
17725            _child: &Session,
17726        ) -> Result<(), crate::session_app::child_session::ChildSessionError> {
17727            unreachable!("fixture creates the child with auto_run=false")
17728        }
17729
17730        async fn cancel_child_run_and_wait(
17731            &self,
17732            _child_id: &str,
17733        ) -> Result<(), crate::session_app::child_session::ChildSessionError> {
17734            unreachable!("create_child_action does not cancel")
17735        }
17736
17737        async fn delete_child_session(
17738            &self,
17739            _parent_id: &str,
17740            _child_id: &str,
17741        ) -> Result<
17742            crate::session_app::child_session::DeleteChildResult,
17743            crate::session_app::child_session::ChildSessionError,
17744        > {
17745            unreachable!("create_child_action does not delete")
17746        }
17747
17748        async fn get_child_runner_info(
17749            &self,
17750            _child_id: &str,
17751        ) -> Option<crate::session_app::child_session::ChildRunnerInfo> {
17752            None
17753        }
17754
17755        async fn register_parent_wait_for_child(
17756            &self,
17757            _parent_session_id: &str,
17758            _child_session_id: &str,
17759            _tool_call_id: Option<&str>,
17760        ) -> Result<(), crate::session_app::child_session::ChildSessionError> {
17761            unreachable!("create_child_action does not register a wait")
17762        }
17763
17764        async fn register_parent_wait_for_children(
17765            &self,
17766            _parent_session_id: &str,
17767            _child_session_ids: &[String],
17768            _policy: bamboo_domain::session::runtime_state::ChildWaitPolicy,
17769        ) -> Result<usize, crate::session_app::child_session::ChildSessionError> {
17770            unreachable!("create_child_action does not register a wait")
17771        }
17772
17773        async fn active_child_ids(&self, _parent_session_id: &str) -> Vec<String> {
17774            Vec::new()
17775        }
17776
17777        async fn find_resident_child(
17778            &self,
17779            _root_session_id: &str,
17780            _resident_name: &str,
17781        ) -> Option<String> {
17782            None
17783        }
17784
17785        async fn ensure_child_indexed(&self, _child_session_id: &str) {}
17786    }
17787
17788    #[test]
17789    fn build_spec_sets_remote_placement_for_matching_role() {
17790        let mut placements = HashMap::new();
17791        placements.insert(
17792            "explorer".to_string(),
17793            ResolvedRemotePlacement {
17794                endpoint: "wss://gpu-host:8443".into(),
17795                token: Some("T-secret".into()),
17796                ca_cert_file: None,
17797                host_label: None,
17798                broker_peer: None,
17799                requirements: None,
17800            },
17801        );
17802        let runner = bogus_runner(placements);
17803
17804        // The broker credential never rides the child RunSpec secrets envelope.
17805        let s = session_of_role("explorer", "do the thing");
17806        let spec = runner.build_spec(&s, &job_for("child-1"));
17807        match &spec.placement {
17808            Placement::Remote { endpoint } => assert_eq!(endpoint, "wss://gpu-host:8443"),
17809            other => panic!("expected Remote, got {other:?}"),
17810        }
17811        assert!(spec.secrets.worker_auth_token.is_none());
17812    }
17813
17814    #[test]
17815    fn build_spec_leaves_local_for_unmatched_role() {
17816        let mut placements = HashMap::new();
17817        placements.insert(
17818            "explorer".to_string(),
17819            ResolvedRemotePlacement {
17820                endpoint: "wss://gpu-host:8443".into(),
17821                token: Some("T".into()),
17822                ca_cert_file: None,
17823                host_label: None,
17824                broker_peer: None,
17825                requirements: None,
17826            },
17827        );
17828        let runner = bogus_runner(placements);
17829
17830        // A DIFFERENT role keeps the default Local placement + no bearer.
17831        let s = session_of_role("writer", "do the thing");
17832        let spec = runner.build_spec(&s, &job_for("child-1"));
17833        assert_eq!(spec.placement, Placement::Local);
17834        assert!(spec.secrets.worker_auth_token.is_none());
17835    }
17836
17837    #[test]
17838    fn build_spec_local_when_no_placements() {
17839        let runner = bogus_runner(HashMap::new());
17840        let s = session_of_role("explorer", "do the thing");
17841        let spec = runner.build_spec(&s, &job_for("child-1"));
17842        assert_eq!(spec.placement, Placement::Local);
17843        assert!(spec.secrets.worker_auth_token.is_none());
17844    }
17845
17846    #[test]
17847    fn typed_child_state_is_the_read_only_authority_under_auto_and_bypass() {
17848        for mode in [
17849            bamboo_domain::SessionPermissionMode::Auto,
17850            bamboo_domain::SessionPermissionMode::Bypass,
17851        ] {
17852            let runner = bogus_runner(HashMap::new());
17853            let mut child = session_of_role("planner", "inspect and plan");
17854            let runtime = child
17855                .agent_runtime_state
17856                .get_or_insert_with(bamboo_domain::AgentRuntimeState::default);
17857            runtime.set_permission_mode(mode);
17858            runtime.read_only = true;
17859
17860            let spec = runner.build_spec(&child, &job_for("read-only-child"));
17861
17862            assert!(spec.capabilities.read_only);
17863            assert!(spec.capabilities.read_only_enforced());
17864            assert!(
17865                spec.capabilities.guardian_read_only,
17866                "the compatibility alias must protect planners on older workers"
17867            );
17868            assert_eq!(spec.capabilities.permission_requested_mode, mode.as_str());
17869            assert_eq!(spec.capabilities.permission_effective_mode, "plan");
17870            assert!(
17871                !spec.capabilities.bypass,
17872                "read-only must win over requested {mode:?}"
17873            );
17874            assert_eq!(
17875                spec.capabilities.auto_approve_permissions,
17876                mode == bamboo_domain::SessionPermissionMode::Auto,
17877                "Auto may suppress prompts but must not widen Plan authorization"
17878            );
17879            let disabled = spec
17880                .disabled_tools
17881                .as_ref()
17882                .expect("read-only state must provision a host-owned denylist");
17883            for tool_name in [
17884                "Edit",
17885                "Write",
17886                "Plan",
17887                "SubAgent",
17888                "session_control",
17889                "request_permissions",
17890            ] {
17891                assert!(
17892                    disabled.iter().any(|name| name == tool_name),
17893                    "read-only provisioning must disable {tool_name}"
17894                );
17895            }
17896        }
17897
17898        let runner = bogus_runner(HashMap::new());
17899        let cosmetic_planner = runner.build_spec(
17900            &session_of_role("planner", "label only"),
17901            &job_for("cosmetic"),
17902        );
17903        assert!(
17904            !cosmetic_planner.capabilities.read_only_enforced(),
17905            "a cosmetic role label must not change authority"
17906        );
17907        assert!(cosmetic_planner.disabled_tools.is_none());
17908
17909        let cosmetic_guardian = runner.build_spec(
17910            &session_of_role("guardian", "label only"),
17911            &job_for("cosmetic-guardian"),
17912        );
17913        assert!(
17914            !cosmetic_guardian.capabilities.read_only_enforced(),
17915            "a Guardian label without typed state or the legacy denylist is cosmetic"
17916        );
17917
17918        let legacy_guardian = session_of_role("guardian", "pre-upgrade reviewer");
17919        let mut legacy_job = job_for("legacy-guardian");
17920        legacy_job.disabled_tools = Some(vec![
17921            "Edit".to_string(),
17922            "Write".to_string(),
17923            "SubAgent".to_string(),
17924            "WebFetch".to_string(),
17925        ]);
17926        let legacy_spec = runner.build_spec(&legacy_guardian, &legacy_job);
17927        assert!(
17928            legacy_spec.capabilities.read_only_enforced(),
17929            "an in-flight pre-upgrade Guardian keeps its read-only authority"
17930        );
17931
17932        let mut differently_named = session_of_role("explorer", "typed authority");
17933        differently_named
17934            .agent_runtime_state
17935            .get_or_insert_with(bamboo_domain::AgentRuntimeState::default)
17936            .read_only = true;
17937        let typed = runner.build_spec(&differently_named, &job_for("typed"));
17938        assert!(
17939            typed.capabilities.read_only_enforced(),
17940            "typed state must enforce read-only independently of role"
17941        );
17942    }
17943
17944    #[test]
17945    fn typed_read_only_forces_native_external_executor_profiles() {
17946        let mut planner = session_of_role("planner", "inspect and plan");
17947        planner
17948            .agent_runtime_state
17949            .get_or_insert_with(bamboo_domain::AgentRuntimeState::default)
17950            .read_only = true;
17951
17952        let claude_runner = ActorChildRunner::new(
17953            "test-actor".into(),
17954            PathBuf::from("/bin/false"),
17955            vec![],
17956            std::env::temp_dir().join("bamboo-test-plan-claude"),
17957            ExecutorSpec::ClaudeCode {
17958                binary: None,
17959                model: None,
17960                permission_mode: Some("bypassPermissions".to_string()),
17961                inherit_user_config: Some(true),
17962                forward_env: None,
17963            },
17964            vec![],
17965            "anthropic".into(),
17966            4,
17967        );
17968        let claude_spec = claude_runner.build_spec(&planner, &job_for("planner-claude"));
17969        let disabled = claude_spec
17970            .disabled_tools
17971            .as_ref()
17972            .expect("typed read-only child must receive a host-owned denylist");
17973        for tool in ["Bash", "BashInput", "BashOutput", "KillShell"] {
17974            assert!(
17975                disabled.iter().any(|name| name == tool),
17976                "typed read-only child must not advertise {tool}"
17977            );
17978        }
17979        assert!(matches!(
17980            claude_spec.executor,
17981            ExecutorSpec::ClaudeCode {
17982                permission_mode: Some(ref mode),
17983                inherit_user_config: Some(false),
17984                ..
17985            } if mode == "plan"
17986        ));
17987        assert_eq!(claude_spec.capabilities.permission_effective_mode, "plan");
17988
17989        let mut codex_exec = codex_executor(Some("inherit"), Some(true));
17990        if let ExecutorSpec::Codex {
17991            sandbox,
17992            approval_policy,
17993            network_access,
17994            allow_danger_bypass,
17995            permission_profile,
17996            ..
17997        } = &mut codex_exec
17998        {
17999            *sandbox = Some("danger-full-access".to_string());
18000            *approval_policy = Some("on-failure".to_string());
18001            *network_access = Some(true);
18002            *allow_danger_bypass = Some(true);
18003            *permission_profile = Some("full-access".to_string());
18004        }
18005        let codex_runner = ActorChildRunner::new(
18006            "test-actor".into(),
18007            PathBuf::from("/bin/false"),
18008            vec![],
18009            std::env::temp_dir().join("bamboo-test-plan-codex"),
18010            codex_exec,
18011            vec![],
18012            "openai".into(),
18013            4,
18014        );
18015        let codex_spec = codex_runner.build_spec(&planner, &job_for("planner-codex"));
18016        assert_eq!(
18017            expected_permission_executor_mapping(
18018                &codex_spec.executor,
18019                codex_spec.capabilities.permission_resolution().unwrap(),
18020                false,
18021            )
18022            .unwrap()
18023            .as_deref(),
18024            Some("codex_exec:approval_policy=never")
18025        );
18026        assert!(matches!(
18027            codex_spec.executor,
18028            ExecutorSpec::Codex {
18029                sandbox: Some(ref sandbox),
18030                inherit_user_config: Some(false),
18031                approval_policy: Some(ref approval_policy),
18032                network_access: Some(false),
18033                allow_danger_bypass: Some(false),
18034                permission_profile: Some(ref profile),
18035                ..
18036            } if sandbox == "read-only"
18037                && approval_policy == "never"
18038                && profile == "read-only"
18039        ));
18040
18041        let mut app_server = codex_executor(Some("inherit"), Some(true));
18042        if let ExecutorSpec::Codex {
18043            mode,
18044            sandbox,
18045            approval_policy,
18046            network_access,
18047            allow_danger_bypass,
18048            permission_profile,
18049            ..
18050        } = &mut app_server
18051        {
18052            *mode = Some("app_server".to_string());
18053            *sandbox = Some("danger-full-access".to_string());
18054            *approval_policy = Some("on-request".to_string());
18055            *network_access = Some(true);
18056            *allow_danger_bypass = Some(true);
18057            *permission_profile = Some("full-access".to_string());
18058        }
18059        let app_server_runner = ActorChildRunner::new(
18060            "test-actor".into(),
18061            PathBuf::from("/bin/false"),
18062            vec![],
18063            std::env::temp_dir().join("bamboo-test-plan-codex-app-server"),
18064            app_server,
18065            vec![],
18066            "openai".into(),
18067            4,
18068        );
18069        let app_server_spec =
18070            app_server_runner.build_spec(&planner, &job_for("planner-codex-app-server"));
18071        assert_eq!(
18072            expected_permission_executor_mapping(
18073                &app_server_spec.executor,
18074                app_server_spec
18075                    .capabilities
18076                    .permission_resolution()
18077                    .unwrap(),
18078                false,
18079            )
18080            .unwrap()
18081            .as_deref(),
18082            Some("codex_app_server:approvalPolicy=never")
18083        );
18084        assert!(matches!(
18085            app_server_spec.executor,
18086            ExecutorSpec::Codex {
18087                mode: Some(ref mode),
18088                sandbox: Some(ref sandbox),
18089                inherit_user_config: Some(false),
18090                approval_policy: Some(ref approval_policy),
18091                network_access: Some(false),
18092                allow_danger_bypass: Some(false),
18093                permission_profile: Some(ref profile),
18094                ..
18095            } if mode == "app_server"
18096                && sandbox == "read-only"
18097                && approval_policy == "on-request"
18098                && profile == "read-only"
18099        ));
18100    }
18101
18102    #[test]
18103    fn typed_read_only_rejects_both_codex_executor_modes() {
18104        let mut planner = session_of_role("planner", "inspect and plan");
18105        planner
18106            .agent_runtime_state
18107            .get_or_insert_with(bamboo_domain::AgentRuntimeState::default)
18108            .read_only = true;
18109
18110        for mode in ["exec", "app_server"] {
18111            let mut executor = codex_executor(Some("inherit"), Some(true));
18112            if let ExecutorSpec::Codex {
18113                mode: configured_mode,
18114                ..
18115            } = &mut executor
18116            {
18117                *configured_mode = Some(mode.to_string());
18118            }
18119            let runner = ActorChildRunner::new(
18120                "test-actor".into(),
18121                PathBuf::from("/bin/false"),
18122                vec![],
18123                std::env::temp_dir().join(format!("bamboo-test-plan-codex-{mode}")),
18124                executor,
18125                vec![],
18126                "openai".into(),
18127                4,
18128            );
18129            let spec = runner.build_spec(&planner, &job_for(&format!("planner-codex-{mode}")));
18130            let error = ensure_read_only_activation_is_enforceable(&spec)
18131                .expect_err("Codex command execution must fail closed for read-only activations");
18132            let message = error.to_string();
18133            assert!(message.contains("Codex executor"), "{message}");
18134            assert!(message.contains("command execution"), "{message}");
18135        }
18136    }
18137
18138    #[tokio::test]
18139    async fn typed_read_only_codex_fails_before_worker_spawn_or_dispatch() {
18140        let mut planner = session_of_role("planner", "inspect and plan");
18141        planner
18142            .agent_runtime_state
18143            .get_or_insert_with(bamboo_domain::AgentRuntimeState::default)
18144            .read_only = true;
18145
18146        let runner = ActorChildRunner::new(
18147            "test-actor".into(),
18148            PathBuf::from("/bin/false"),
18149            vec![],
18150            std::env::temp_dir().join("bamboo-test-plan-codex-no-spawn"),
18151            codex_executor(Some("inherit"), Some(true)),
18152            vec![],
18153            "openai".into(),
18154            4,
18155        );
18156        let (event_tx, _event_rx) = mpsc::channel::<AgentEvent>(4);
18157        let error = runner
18158            .execute_external_child(
18159                &mut planner,
18160                &job_for("planner-codex-no-spawn"),
18161                event_tx,
18162                CancellationToken::new(),
18163            )
18164            .await
18165            .expect_err("typed read-only Codex must fail before invoking /bin/false");
18166        let message = error.to_string();
18167        assert!(message.contains("Codex executor"), "{message}");
18168        assert!(message.contains("command execution"), "{message}");
18169    }
18170
18171    #[test]
18172    fn ticket_placement_keeps_local_and_ordinary_remote_execution_available() {
18173        let runner = bogus_runner(HashMap::new());
18174        let ordinary = session_of_role("worker", "ordinary work");
18175        for placement in [
18176            Placement::Local,
18177            Placement::Remote {
18178                endpoint: "wss://fixture.invalid/actor".into(),
18179            },
18180            Placement::Schedulable {
18181                pool: "fixture-pool".into(),
18182            },
18183        ] {
18184            let mut spec = runner.build_spec(&ordinary, &job_for("placement"));
18185            spec.placement = placement;
18186            ensure_ticket_placement_is_enforceable(&ordinary, &spec).unwrap();
18187            for marker in [
18188                crate::ticket_runtime::TICKET_DISPATCH_KEY,
18189                crate::ticket_worker_plan::TICKET_LOCAL_PLAN_KEY,
18190            ] {
18191                let mut ticket = ordinary.clone();
18192                ticket.metadata.insert(marker.into(), "fixture".into());
18193                assert_eq!(
18194                    ensure_ticket_placement_is_enforceable(&ticket, &spec).is_ok(),
18195                    matches!(spec.placement, Placement::Local)
18196                );
18197            }
18198        }
18199    }
18200
18201    #[tokio::test]
18202    async fn ticket_placement_rejected_before_remote_route_claim_or_worker_spawn() {
18203        for marker in [
18204            crate::ticket_runtime::TICKET_DISPATCH_KEY,
18205            crate::ticket_worker_plan::TICKET_LOCAL_PLAN_KEY,
18206        ] {
18207            for schedulable in [false, true] {
18208                let mut session = session_of_role("worker", "Ticket work");
18209                session.metadata.insert(marker.into(), "fixture".into());
18210                let before = serde_json::to_value(&session).unwrap();
18211                let remote = if schedulable {
18212                    HashMap::new()
18213                } else {
18214                    HashMap::from([(
18215                        "worker".into(),
18216                        ResolvedRemotePlacement {
18217                            endpoint: "wss://fixture.invalid/actor".into(),
18218                            ..Default::default()
18219                        },
18220                    )])
18221                };
18222                let sched = if schedulable {
18223                    HashMap::from([(
18224                        "worker".into(),
18225                        sched_placement("fixture-pool", "wss://fixture.invalid"),
18226                    )])
18227                } else {
18228                    HashMap::new()
18229                };
18230                let runner = bogus_sched_runner(remote, sched);
18231                let (events, mut observations) = mpsc::channel::<AgentEvent>(4);
18232                let error = runner
18233                    .execute_external_child(
18234                        &mut session,
18235                        &job_for("ticket-no-remote"),
18236                        events,
18237                        CancellationToken::new(),
18238                    )
18239                    .await
18240                    .unwrap_err();
18241                assert!(
18242                    error
18243                        .to_string()
18244                        .contains("Ticket execution requires local placement"),
18245                    "{error}"
18246                );
18247                assert_eq!(serde_json::to_value(&session).unwrap(), before);
18248                assert!(observations.try_recv().is_err());
18249            }
18250        }
18251    }
18252
18253    #[test]
18254    fn typed_read_only_rejects_resident_actor_placements() {
18255        let runner = bogus_runner(HashMap::new());
18256        let mut planner = session_of_role("planner", "inspect and plan");
18257        planner
18258            .agent_runtime_state
18259            .get_or_insert_with(bamboo_domain::AgentRuntimeState::default)
18260            .read_only = true;
18261
18262        for placement in [
18263            Placement::Remote {
18264                endpoint: "wss://resident.example/actor".to_string(),
18265            },
18266            Placement::Schedulable {
18267                pool: "resident-pool".to_string(),
18268            },
18269        ] {
18270            let mut spec = runner.build_spec(&planner, &job_for("read-only-resident"));
18271            spec.placement = placement;
18272
18273            let error = ensure_read_only_activation_is_enforceable(&spec)
18274                .expect_err("resident placement must fail closed for read-only activation");
18275            let message = error.to_string();
18276            assert!(message.contains("requires local placement"), "{message}");
18277            assert!(message.contains("read-only tool boundary"), "{message}");
18278        }
18279
18280        let local = runner.build_spec(&planner, &job_for("read-only-local"));
18281        ensure_read_only_activation_is_enforceable(&local)
18282            .expect("a locally provisioned read-only worker is enforceable");
18283    }
18284
18285    #[tokio::test]
18286    async fn build_spec_preserves_exact_inherited_permission_mode_for_child_worker() {
18287        // Exercise the real creation path instead of pre-seeding the child by
18288        // hand: both legacy Bypass and zero-prompt Auto must survive child
18289        // creation and the actor provisioning boundary without collapsing.
18290        for (label, mode) in [
18291            ("bypass", bamboo_domain::SessionPermissionMode::Bypass),
18292            ("auto", bamboo_domain::SessionPermissionMode::Auto),
18293        ] {
18294            let runner = bogus_runner(HashMap::new());
18295            let mut parent = Session::new(format!("parent-{label}"), "test-model");
18296            parent
18297                .agent_runtime_state
18298                .get_or_insert_with(bamboo_domain::AgentRuntimeState::default)
18299                .set_permission_mode(mode);
18300            let workspace = tempfile::tempdir().expect("workspace fixture");
18301            let port = RecordingChildSessionPort {
18302                parent: parent.clone(),
18303                saved: Default::default(),
18304            };
18305            let child_id = format!("child-{label}-{}", uuid::Uuid::new_v4());
18306            crate::session_app::child_session::create_child_action(
18307                &port,
18308                crate::session_app::child_session::CreateChildInput {
18309                    parent_session: parent,
18310                    child_id: child_id.clone(),
18311                    title: format!("{label} child"),
18312                    responsibility: "Run ordinary commands".to_string(),
18313                    assignment_prompt: "run an ordinary command".to_string(),
18314                    subagent_type: "explorer".to_string(),
18315                    workspace: workspace.path().to_string_lossy().into_owned(),
18316                    workspace_source: crate::project_context::WorkspaceSource::Explicit,
18317                    model_override: None,
18318                    model_ref_override: None,
18319                    runtime_metadata: HashMap::new(),
18320                    read_only: false,
18321                    auto_run: false,
18322                    reasoning_effort: None,
18323                    lifecycle: None,
18324                    resident_name: None,
18325                    resident_context: None,
18326                    disabled_tools: None,
18327                    context_fork: None,
18328                },
18329            )
18330            .await
18331            .expect("create inherited-permission child");
18332            let child = port.saved_child();
18333
18334            assert_eq!(
18335                child
18336                    .agent_runtime_state
18337                    .as_ref()
18338                    .map(bamboo_domain::AgentRuntimeState::effective_permission_mode),
18339                Some(mode),
18340                "create_child_action must inherit {label} from the parent"
18341            );
18342
18343            let spec = runner.build_spec(&child, &job_for(&child_id));
18344
18345            assert_eq!(
18346                spec.capabilities.bypass,
18347                mode == bamboo_domain::SessionPermissionMode::Bypass
18348            );
18349            assert_eq!(
18350                spec.capabilities.auto_approve_permissions,
18351                mode == bamboo_domain::SessionPermissionMode::Auto
18352            );
18353            assert!(
18354                spec.capabilities.enforce_permissions,
18355                "policy evaluation must remain active under {label}"
18356            );
18357        }
18358    }
18359
18360    #[tokio::test]
18361    async fn child_resident_and_guardian_inherit_project_through_actor_run_spec() {
18362        let project_id = bamboo_domain::ProjectId::parse("project-inherited").expect("Project id");
18363        let workspace = tempfile::tempdir().expect("workspace fixture");
18364
18365        for (role, lifecycle, resident_name) in [
18366            ("explorer", None, None),
18367            ("resident", Some("resident"), Some("stable-reviewer")),
18368            ("guardian", None, None),
18369        ] {
18370            let mut parent = Session::new(format!("parent-{role}"), "test-model");
18371            parent.set_project_id_meta(project_id.to_string());
18372            let port = RecordingChildSessionPort {
18373                parent: parent.clone(),
18374                saved: Default::default(),
18375            };
18376            let child_id = format!("child-{role}-{}", uuid::Uuid::new_v4());
18377            crate::session_app::child_session::create_child_action(
18378                &port,
18379                crate::session_app::child_session::CreateChildInput {
18380                    parent_session: parent,
18381                    child_id: child_id.clone(),
18382                    title: format!("{role} child"),
18383                    responsibility: "Review the assigned work".to_string(),
18384                    assignment_prompt: "inspect the change".to_string(),
18385                    subagent_type: role.to_string(),
18386                    workspace: workspace.path().to_string_lossy().into_owned(),
18387                    workspace_source: crate::project_context::WorkspaceSource::Explicit,
18388                    model_override: None,
18389                    model_ref_override: None,
18390                    runtime_metadata: HashMap::new(),
18391                    read_only: false,
18392                    auto_run: false,
18393                    reasoning_effort: None,
18394                    lifecycle: lifecycle.map(str::to_string),
18395                    resident_name: resident_name.map(str::to_string),
18396                    resident_context: None,
18397                    disabled_tools: None,
18398                    context_fork: None,
18399                },
18400            )
18401            .await
18402            .expect("create Project-inheriting child");
18403            let child = port.saved_child();
18404            assert_eq!(
18405                crate::project_context::ProjectContextResolver::project_id_from_session(&child),
18406                Some(project_id.clone()),
18407                "{role} child must inherit its parent's Project"
18408            );
18409
18410            assert_eq!(
18411                project_id_for_actor_run(&child).expect("valid actor Project identity"),
18412                Some(project_id.clone()),
18413                "{role} actor RunSpec must preserve inherited Project identity"
18414            );
18415        }
18416    }
18417
18418    #[test]
18419    fn placement_metadata_stamps_remote_and_schedulable_not_local() {
18420        // Local children carry no stamp — the DTO defaults them to the backend host.
18421        assert_eq!(placement_metadata(&Placement::Local, None), None);
18422
18423        // Remote, no node label → host derived from the endpoint.
18424        let r = placement_metadata(
18425            &Placement::Remote {
18426                endpoint: "wss://10.0.0.5:8443/stream".into(),
18427            },
18428            None,
18429        )
18430        .unwrap();
18431        assert!(r.contains(r#""kind":"remote""#), "{r}");
18432        assert!(r.contains(r#""host":"10.0.0.5""#), "{r}");
18433
18434        // A cluster node's label (its metadata) OVERRIDES the raw endpoint host.
18435        let labeled = placement_metadata(
18436            &Placement::Remote {
18437                endpoint: "ws://169.254.230.101:8899".into(),
18438            },
18439            Some("mini"),
18440        )
18441        .unwrap();
18442        assert!(labeled.contains(r#""host":"mini""#), "{labeled}");
18443
18444        // Schedulable → {kind:"remote", host:<node label, else pool>}.
18445        let s = placement_metadata(
18446            &Placement::Schedulable {
18447                pool: "explorers".into(),
18448            },
18449            Some("mini"),
18450        )
18451        .unwrap();
18452        assert!(s.contains(r#""kind":"remote""#), "{s}");
18453        assert!(s.contains(r#""host":"mini""#), "{s}");
18454
18455        // The stamp round-trips through the storage placement type.
18456        let p: bamboo_storage::SessionPlacement = serde_json::from_str(&labeled).unwrap();
18457        assert_eq!(p.kind, "remote");
18458        assert_eq!(p.host, "mini");
18459    }
18460
18461    /// A legacy direct worker endpoint cannot bypass the broker activation
18462    /// contract, even when the endpoint is reachable and has a bearer token.
18463    #[tokio::test]
18464    async fn execute_external_child_rejects_direct_remote_worker() {
18465        let mut placements = HashMap::new();
18466        placements.insert(
18467            "explorer".to_string(),
18468            ResolvedRemotePlacement {
18469                endpoint: "ws://127.0.0.1:1".into(),
18470                token: Some("remote-test-token".into()),
18471                ca_cert_file: None,
18472                host_label: Some("mini-e2e".into()),
18473                broker_peer: None,
18474                requirements: None,
18475            },
18476        );
18477        let runner = bogus_runner(placements);
18478        let mut session = session_of_role("explorer", "hello remote");
18479        let job = job_for("child-1");
18480        let (event_tx, _event_rx) = mpsc::channel::<AgentEvent>(64);
18481        let error = runner
18482            .execute_external_child(&mut session, &job, event_tx, CancellationToken::new())
18483            .await
18484            .err()
18485            .expect("direct remote transport must fail closed");
18486        assert!(error
18487            .to_string()
18488            .contains("remote_broker_activation_unavailable"));
18489        assert!(session
18490            .messages
18491            .iter()
18492            .all(|message| message.role != Role::Assistant));
18493    }
18494
18495    // ---- #181 (P2b): schedulable placement routing --------------------------
18496
18497    /// A bogus-worker_bin runner carrying SCHEDULABLE placements (and optionally
18498    /// remote ones, to test precedence). A local spawn here would fail on
18499    /// `/bin/false`, so a passing schedulable test proves no subprocess spawned.
18500    fn bogus_sched_runner(
18501        remote: HashMap<String, ResolvedRemotePlacement>,
18502        sched: HashMap<String, ResolvedSchedulablePlacement>,
18503    ) -> ActorChildRunner {
18504        ActorChildRunner::new(
18505            "test-actor".into(),
18506            PathBuf::from("/bin/false"),
18507            vec![],
18508            std::env::temp_dir().join("bamboo-test-fab-181"),
18509            ExecutorSpec::Echo,
18510            vec![],
18511            "anthropic".into(),
18512            4,
18513        )
18514        .with_remote_placements(remote)
18515        .with_schedulable_placements(sched)
18516    }
18517
18518    fn sched_placement(
18519        pool: &str,
18520        _registry_url: impl Into<String>,
18521    ) -> ResolvedSchedulablePlacement {
18522        ResolvedSchedulablePlacement {
18523            pool: pool.into(),
18524            host_label: None,
18525            requirements: None,
18526            ..Default::default()
18527        }
18528    }
18529
18530    #[test]
18531    fn build_spec_sets_schedulable_placement_for_matching_role() {
18532        let mut sched = HashMap::new();
18533        sched.insert(
18534            "explorer".to_string(),
18535            sched_placement("gpu-pool", "unused"),
18536        );
18537        let runner = bogus_sched_runner(HashMap::new(), sched);
18538
18539        let s = session_of_role("explorer", "do the thing");
18540        let spec = runner.build_spec(&s, &job_for("child-1"));
18541        match &spec.placement {
18542            Placement::Schedulable { pool } => assert_eq!(pool, "gpu-pool"),
18543            other => panic!("expected Schedulable, got {other:?}"),
18544        }
18545        // No per-placement bearer now — the bus connection carries the bus token.
18546        assert!(spec.secrets.worker_auth_token.is_none());
18547    }
18548
18549    #[test]
18550    fn build_spec_remote_wins_when_role_in_both_maps() {
18551        // A role present in BOTH remote_placements and schedulable_placements must
18552        // resolve to the FIXED remote placement (documented precedence).
18553        let mut remote = HashMap::new();
18554        remote.insert(
18555            "explorer".to_string(),
18556            ResolvedRemotePlacement {
18557                endpoint: "wss://fixed-host:8443".into(),
18558                token: Some("T-remote".into()),
18559                ca_cert_file: None,
18560                host_label: None,
18561                broker_peer: None,
18562                requirements: None,
18563            },
18564        );
18565        let mut sched = HashMap::new();
18566        sched.insert(
18567            "explorer".to_string(),
18568            sched_placement("gpu-pool", "https://control-plane:9562"),
18569        );
18570        let runner = bogus_sched_runner(remote, sched);
18571
18572        let s = session_of_role("explorer", "do the thing");
18573        let spec = runner.build_spec(&s, &job_for("child-1"));
18574        match &spec.placement {
18575            Placement::Remote { endpoint } => assert_eq!(endpoint, "wss://fixed-host:8443"),
18576            other => panic!("expected Remote (precedence), got {other:?}"),
18577        }
18578        assert!(spec.secrets.worker_auth_token.is_none());
18579    }
18580
18581    #[test]
18582    fn build_spec_local_for_unmatched_schedulable_role() {
18583        let mut sched = HashMap::new();
18584        sched.insert(
18585            "explorer".to_string(),
18586            sched_placement("gpu-pool", "https://control-plane:9562"),
18587        );
18588        let runner = bogus_sched_runner(HashMap::new(), sched);
18589        let s = session_of_role("writer", "do the thing");
18590        let spec = runner.build_spec(&s, &job_for("child-1"));
18591        assert_eq!(spec.placement, Placement::Local);
18592        assert!(spec.secrets.worker_auth_token.is_none());
18593    }
18594
18595    /// The full role → resolved-placement → badge-host chain: a child routed to a
18596    /// remote/schedulable placement carrying a cluster node's `host_label` stamps
18597    /// that label; without a label it falls back to the endpoint host / pool; a
18598    /// Local child gets no stamp (the DTO defaults it to the backend host).
18599    #[test]
18600    fn placement_stamp_uses_node_label_for_remote_and_schedulable() {
18601        // Remote WITH a node label → {remote, <label>}, overriding the raw IP.
18602        let mut remote = HashMap::new();
18603        remote.insert(
18604            "explorer".to_string(),
18605            ResolvedRemotePlacement {
18606                endpoint: "ws://169.254.230.101:8899".into(),
18607                token: None,
18608                ca_cert_file: None,
18609                host_label: Some("mini".into()),
18610                broker_peer: None,
18611                requirements: None,
18612            },
18613        );
18614        let runner = bogus_runner(remote);
18615        let spec = runner.build_spec(&session_of_role("explorer", "go"), &job_for("c1"));
18616        let stamp = runner
18617            .placement_stamp_for(&spec)
18618            .expect("remote child is stamped");
18619        assert!(stamp.contains(r#""kind":"remote""#), "{stamp}");
18620        assert!(stamp.contains(r#""host":"mini""#), "{stamp}");
18621
18622        // Remote WITHOUT a node label → falls back to the endpoint host.
18623        let mut remote_nolabel = HashMap::new();
18624        remote_nolabel.insert(
18625            "explorer".to_string(),
18626            ResolvedRemotePlacement {
18627                endpoint: "ws://169.254.230.101:8899".into(),
18628                token: None,
18629                ca_cert_file: None,
18630                host_label: None,
18631                broker_peer: None,
18632                requirements: None,
18633            },
18634        );
18635        let r2 = bogus_runner(remote_nolabel);
18636        let spec2 = r2.build_spec(&session_of_role("explorer", "go"), &job_for("c1"));
18637        assert!(r2
18638            .placement_stamp_for(&spec2)
18639            .unwrap()
18640            .contains(r#""host":"169.254.230.101""#));
18641
18642        // Schedulable WITH a node label → {remote, <label>} (a node, not a pool name).
18643        let mut sched = HashMap::new();
18644        sched.insert(
18645            "mac-mini-monitor".to_string(),
18646            ResolvedSchedulablePlacement {
18647                pool: "mac-mini-monitor".into(),
18648                host_label: Some("mini".into()),
18649                requirements: None,
18650                ..Default::default()
18651            },
18652        );
18653        let sr = bogus_sched_runner(HashMap::new(), sched);
18654        let spec3 = sr.build_spec(&session_of_role("mac-mini-monitor", "go"), &job_for("c1"));
18655        let stamp3 = sr
18656            .placement_stamp_for(&spec3)
18657            .expect("scheduled child is stamped");
18658        assert!(stamp3.contains(r#""kind":"remote""#), "{stamp3}");
18659        assert!(stamp3.contains(r#""host":"mini""#), "{stamp3}");
18660
18661        // A Local (unmatched) child gets NO stamp.
18662        let local = bogus_runner(HashMap::new());
18663        let spec4 = local.build_spec(&session_of_role("writer", "go"), &job_for("c1"));
18664        assert_eq!(local.placement_stamp_for(&spec4), None);
18665    }
18666
18667    #[tokio::test]
18668    async fn legacy_schedulable_route_fails_closed_without_local_fallback() {
18669        let mut sched = HashMap::new();
18670        sched.insert(
18671            "explorer".to_string(),
18672            sched_placement("gpu-pool", "unused"),
18673        );
18674        let runner = bogus_sched_runner(HashMap::new(), sched);
18675        let mut session = session_of_role("explorer", "do the thing");
18676        let (event_tx, _rx) = mpsc::channel::<AgentEvent>(8);
18677        let error = runner
18678            .execute_external_child(
18679                &mut session,
18680                &job_for("child-1"),
18681                event_tx,
18682                CancellationToken::new(),
18683            )
18684            .await
18685            .expect_err("legacy bus presence cannot grant placement");
18686        assert!(error
18687            .to_string()
18688            .contains("remote_broker_activation_unavailable"));
18689    }
18690}