Skip to main content

mlua_swarm_server/
worker.rs

1//! HTTP `/v1/worker/*` endpoints (SubAgent self-fetch path).
2//!
3//! # 7-Entry pointer #6 (Output Event design)
4//!
5//! **This endpoint accesses `OutputStore` directly and does NOT go through the engine.**
6//! It is one of the seven entry points enumerated in project `CLAUDE.md` §"Output Event
7//! Design SoT". For the canonical description, see the crate root doc of
8//! `mlua-swarm-output-store` (`cargo doc -p mlua-swarm-output-store`).
9//!
10//! # Path
11//!
12//! A thin-payload path where a SubAgent (= worker process launched by a MainAI) uses
13//! the capability token it received via WS Spawn to self-fetch its prompt and
14//! submit its result — putting the token in `Authorization: Bearer <encoded CapToken>`.
15//!
16//! ## Routes
17//!
18//! - `GET /v1/worker/prompt?task_id=<tid>` — via `engine.fetch_worker_payload`,
19//!   returns `{task_id, attempt, agent, system?, prompt, context?}`.
20//!   `context.steps` (`projection-adapter` ST5, [`assemble_step_pointers`])
21//!   is assembled fresh on every fetch: a `ContextPolicy.steps`-filtered
22//!   pointer list to preceding steps' OUTPUT, resolved through
23//!   `crate::projection::McpQueryAdapter`'s Data-plane + `result_ref`
24//!   enumeration — no separate MCP tool call needed to discover a prior
25//!   step's OUTPUT.
26//! - `POST /v1/worker/result` with body `{task_id, value, ok}` — appends one `Final`
27//!   to the output tail via `engine.submit_output(Final)` (= the canonical path
28//!   through which the dispatch layer decides Pass/Blocked) and updates
29//!   `task.last_result` via `engine.post_result`.
30//! - `POST /v1/worker/artifact?name=<name>` (GH #36 ST1) — stages one named
31//!   part per POST via `engine.stage_worker_artifact_trusted`. Completing the
32//!   attempt is still `POST /v1/worker/submit` / `/v1/worker/result` — this
33//!   route only stages; the dispatch layer's Final-pull folds every staged
34//!   part into `{"out": <final>, "parts": {<name>: <value>, ...}}`. At
35//!   staging time the submit-time projection sink also materializes the part
36//!   raw to `<ctx-dir>/<name>` (the IN file the next Agent step reads;
37//!   fail-open skip when no `work_dir` / `project_root` resolves).
38//! - `GET /v1/worker/prompt/system?task_id=<tid>&attempt=<n>` (GH #31) —
39//!   raw baked `system` bytes for `(task_id, attempt)`, the `Http`-mode
40//!   fetch target for `system_ref.uri`. Same Bearer flow as
41//!   `/v1/worker/prompt`; body is `text/plain`, not JSON.
42//! - `GET /v1/agents/:name/render-size` (GH #31) — no Bearer required, same
43//!   trust tier as `GET /v1/blueprints/:id/head`. Live per-agent most-recently
44//!   observed render size, backing `bp_doctor`'s post-render check.
45//! - `POST /v1/worker/degradation` (GH #32) — structured JSON `{tool, error,
46//!   fallback, note?}`, same Bearer flow as [`worker_submit`]. An
47//!   **independent channel**: entries are appended to `RunRecord.degradations`
48//!   via `RunStore::append_degradation` directly and never touch
49//!   `OutputStore` / the fold path (Crux invariant 2 — a degradation must
50//!   never surface as step OUTPUT). `step_ref` / `attempt` / `at` are
51//!   server-injected, never trusted from the client. Silent `204` (no
52//!   append) when the dispatch task carries no Run linkage — same
53//!   fail-open contract as [`reject_if_run_terminal`]'s own resolution
54//!   steps, since a pre-run-tracking dispatch has nowhere to record a
55//!   degradation and that must not become a client-visible error.
56//!
57//! ## Bearer authentication
58//!
59//! The Bearer value is the string produced by `CapToken::encode()` (= URL-safe
60//! base64 of serde_json). The server decodes it with `CapToken::decode` and then,
61//! inside the engine, verifies HMAC sig + role × verb gate + TTL via
62//! `verify_token_for_task` (= self-contained capability token; no server-side
63//! store lookup required).
64//!
65//! Tokens are minted during the "2) mint outside the lock" phase of
66//! `engine.dispatch_attempt` (`Role::Worker`, 600s TTL, `scopes=["*"]`).
67//! The verb gate covers `FetchPrompt` / `EmitOutput` / `PostResult` — the worker
68//! leaf capability set (`crate::types::WORKER_LEAF_VERBS`).
69
70use axum::{
71    extract::{Query, State},
72    http::{header, header::AUTHORIZATION, HeaderMap, StatusCode},
73    Json,
74};
75use mlua_swarm::core::agent_context::StepPointer;
76use mlua_swarm::core::step_naming::StepNaming;
77use mlua_swarm::store::run::{DegradationEntry, RunStatus, RunStoreError};
78use mlua_swarm::{CapToken, ContentRef, EngineError, OutputEvent, RunId, StepId, WorkerPayload};
79use mlua_swarm_schema::{ContextPolicy, VerdictChannel};
80use serde::Deserialize;
81use serde_json::Value;
82
83use crate::projection::McpQueryAdapter;
84use crate::{ApiError, AppState};
85
86/// Query params for `GET /v1/worker/prompt`.
87#[derive(Debug, Deserialize)]
88pub struct PromptQuery {
89    /// Task the fetched prompt belongs to; cross-checked against the Bearer
90    /// handle/token. Typed [`StepId`] since issue #14 — the wire shape stays
91    /// a plain string; a bad prefix is rejected at deserialize.
92    pub task_id: StepId,
93}
94
95/// `GET /v1/worker/prompt?task_id=<tid>`. Bearer = encoded `CapToken` or short `wh-` handle.
96/// Thin HTTP wrapper over `engine.fetch_worker_payload` / `fetch_worker_payload_trusted`.
97/// Short-handle path (recommended for SubAgents): handle → task_id
98/// cross-check → trusted fetch.
99/// Full-`CapToken` path: token decode → verify → fetch.
100pub async fn worker_prompt(
101    State(state): State<AppState>,
102    headers: HeaderMap,
103    Query(q): Query<PromptQuery>,
104) -> Result<Json<WorkerPayload>, ApiError> {
105    let task_id = q.task_id;
106    let bearer = extract_bearer_raw(&headers)?;
107    let mut payload = if let Some(handle) = parse_worker_handle(&bearer) {
108        // Short-handle path: verify handle → task_id (security: confirm the handle is bound to this task).
109        let resolved = state
110            .engine
111            .task_id_from_handle(handle)
112            .await
113            .map_err(map_handle_lookup_err)?;
114        if resolved != task_id {
115            return Err(ApiError::bad_request(format!(
116                "handle {handle} is bound to task {resolved}, not {task_id}"
117            )));
118        }
119        state
120            .engine
121            .fetch_worker_payload_trusted(&task_id)
122            .await
123            .map_err(|e| ApiError::engine(format!("fetch_worker_payload_trusted: {e}")))?
124    } else {
125        // Full CapToken path (the alternate Bearer form).
126        let token = CapToken::decode(bearer.trim())
127            .map_err(|e| ApiError::bad_request(format!("invalid token: {e}")))?;
128        state
129            .engine
130            .fetch_worker_payload(&token, &task_id)
131            .await
132            .map_err(|e| ApiError::engine(format!("fetch_worker_payload: {e}")))?
133    };
134    assemble_step_pointers(&state, &mut payload).await;
135    Ok(Json(payload))
136}
137
138/// Assembles `payload.context.steps` — the `ContextPolicy.steps`-filtered
139/// pointer list to preceding steps' OUTPUT (`projection-adapter` ST5's
140/// Worker axis; see `mlua_swarm::core::agent_context`'s module doc).
141/// Resolved fresh on every fetch (not baked at spawn time), so a step
142/// submitted after this agent spawned — but before it fetches its prompt
143/// — is still visible.
144///
145/// GH #23 subtask-3: `resolved_steps` (from
146/// `McpQueryAdapter::list_steps_by_run_id`) always reports the CANONICAL
147/// name (see `crate::projection`'s module doc), so both the self-exclusion
148/// check and the `ContextPolicy` match are done against canonical names —
149/// `payload.agent` (the raw `Step.ref` this fetching agent was dispatched
150/// under) is canonicalized via `Engine::step_naming_for(&payload.task_id)`
151/// (the FETCHING agent's own dispatch id — the same `StepNaming` `Arc`
152/// every step of this Blueprint launch shares, see [`StepNaming`]'s module
153/// doc), and `policy.allows_step` itself is left untouched (schema crate
154/// stays name-agnostic) — [`allows_step_canonical`] is the caller-side seam
155/// that resolves each policy-declared name through the table before
156/// comparing.
157///
158/// No-op (`context.steps` stays empty) when: the payload carries no
159/// `context` at all; the context has no `run_id` (a spawn that never
160/// threaded one through — pre-run-tracking callers, or a spawner stack
161/// without the Run-tracking layer); or the addressed Run cannot be
162/// resolved. All three are fail-open, matching this crate's other
163/// best-effort projection hooks (a missing pointer list must never turn a
164/// would-have-succeeded fetch into a failure).
165async fn assemble_step_pointers(state: &AppState, payload: &mut WorkerPayload) {
166    let Some(context) = payload.context.as_mut() else {
167        return;
168    };
169    let Some(run_id_str) = context.run_id.clone() else {
170        return;
171    };
172    let Ok(run_id) = RunId::parse(run_id_str) else {
173        return;
174    };
175
176    let adapter = McpQueryAdapter::new(
177        state.data_store.clone(),
178        state.run_store.clone(),
179        state.engine.clone(),
180    );
181    let Ok((run, resolved_steps)) = adapter.list_steps_by_run_id(&run_id).await else {
182        return;
183    };
184
185    let naming = state.engine.step_naming_for(&payload.task_id).await;
186    let policy = state
187        .engine
188        .context_policy_for(&payload.task_id, payload.attempt)
189        .await;
190    let self_canonical = naming
191        .as_deref()
192        .and_then(|n| n.canonical_of_producer(&payload.agent))
193        .map(str::to_string)
194        .unwrap_or_else(|| payload.agent.clone());
195
196    let mut pointers = Vec::new();
197    for step in &resolved_steps {
198        if step.name == self_canonical
199            || !allows_step_canonical(&policy, naming.as_deref(), &step.name)
200        {
201            continue;
202        }
203        if let Some((size_bytes, file_path, content_url, sha256)) =
204            crate::projection::resolve_step_pointer_fields(state, &run, step).await
205        {
206            pointers.push(StepPointer {
207                name: step.name.clone(),
208                size_bytes,
209                file_path,
210                content_url,
211                sha256,
212            });
213        }
214    }
215    context.steps = pointers;
216}
217
218/// GH #23 subtask-3: caller-side canonical/alias expansion for
219/// `ContextPolicy.allows_step` — same precedence as
220/// `ContextPolicy::allows_step` itself (`steps_exclude` wins; `steps:
221/// None` = pass-all, `Some(list)` = named-only), but each
222/// policy-declared name is resolved through the Blueprint's `StepNaming`
223/// table before comparison, so a Blueprint author's `steps: [...]` entry
224/// naming either the canonical projection name OR any alias (`Step.ref` /
225/// the `out` ctx-path's top-level segment) matches the same step.
226/// `ContextPolicy::allows_step` (schema crate) is untouched — this is the
227/// GH #23 seam, kept out of the name-agnostic schema type. `naming: None`
228/// degrades to a literal string comparison, byte-identical to
229/// `ContextPolicy::allows_step` itself (defensive-only fallback, matching
230/// `crate::projection::McpQueryAdapter::step_naming_for_run`'s own
231/// contract).
232fn allows_step_canonical(
233    policy: &ContextPolicy,
234    naming: Option<&StepNaming>,
235    canonical_name: &str,
236) -> bool {
237    let resolves_to = |raw: &str| -> bool {
238        match naming {
239            Some(n) => n
240                .resolve(raw)
241                .map(|c| c == canonical_name)
242                .unwrap_or(raw == canonical_name),
243            None => raw == canonical_name,
244        }
245    };
246    if policy
247        .steps_exclude
248        .iter()
249        .any(|excluded| resolves_to(excluded))
250    {
251        return false;
252    }
253    match &policy.steps {
254        None => true,
255        Some(list) => list.iter().any(|included| resolves_to(included)),
256    }
257}
258
259/// Body for `POST /v1/worker/result`.
260#[derive(Debug, Deserialize)]
261pub struct WorkerResultReq {
262    /// Task this result belongs to (looked up together with the Bearer
263    /// token). Typed [`StepId`] since issue #14 (see [`PromptQuery`]).
264    pub task_id: StepId,
265    /// `WorkerResult.value` (= the value returned by the Operator: LLM inference result or tool execution result).
266    pub value: Value,
267    /// `WorkerResult.ok`. `false` makes the dispatch path decide Blocked
268    /// (= same semantics as `OutputEvent::Final { ok: false, .. }` from a
269    /// `SpawnerAdapter`). Defaults to `true`.
270    #[serde(default = "default_ok_true")]
271    pub ok: bool,
272    /// Optional explicit attempt. Normally omitted (= the server looks up `task.attempt`).
273    /// A carry for race-condition tests that need to write to a fixed attempt.
274    #[serde(default)]
275    pub attempt: Option<u32>,
276}
277
278fn default_ok_true() -> bool {
279    true
280}
281
282/// `POST /v1/worker/result`. Bearer = encoded `CapToken`.
283/// Fires `engine.submit_output(Final)` + `engine.post_result`.
284pub async fn worker_result(
285    State(state): State<AppState>,
286    headers: HeaderMap,
287    Json(req): Json<WorkerResultReq>,
288) -> Result<StatusCode, ApiError> {
289    let token = decode_worker_bearer(&headers)?;
290    let task_id = req.task_id.clone();
291
292    // Use body-explicit attempt if provided; otherwise the current task.attempt.
293    let attempt = match req.attempt {
294        Some(n) => n,
295        None => state
296            .engine
297            .task_attempt(&task_id)
298            .await
299            .map_err(|e| ApiError::engine(format!("task_attempt: {e}")))?,
300    };
301
302    let event = OutputEvent::Final {
303        content: ContentRef::Inline {
304            value: req.value.clone(),
305        },
306        ok: req.ok,
307    };
308    // GH #51: completion-time verdict-contract enforcement now runs
309    // inside `Engine::submit_output` itself (see
310    // `map_completion_result`'s doc) — this route previously called no
311    // gate at all.
312    map_completion_result(
313        state
314            .engine
315            .submit_output(&token, &task_id, attempt, event)
316            .await,
317        "submit_output",
318    )?;
319    state
320        .engine
321        .post_result(&token, &task_id, req.value)
322        .await
323        .map_err(|e| ApiError::engine(format!("post_result: {e}")))?;
324    Ok(StatusCode::NO_CONTENT)
325}
326
327/// Body-level protocol prefix recognized by [`worker_submit`] and
328/// [`worker_artifact`] (GH #42). When the trimmed request body starts with
329/// this sentinel, the rest is treated as an absolute path; the file is
330/// read and its contents replace the submitted body. Non-sentinel bodies
331/// are unchanged.
332///
333/// See [`resolve_file_sentinel`] for the resolution rules and guards.
334const FILE_SENTINEL_PREFIX: &str = "@file:";
335
336/// Byte ceiling on the resolved-file body, matching the HTTP
337/// `DefaultBodyLimit` applied to inline bodies at the router (2 MiB, see
338/// the `/v1/worker/submit` layer in `crate::app_router`). Sentinel
339/// bodies bypass that axum body layer (the request itself is small), so
340/// the guard is checked in [`resolve_file_sentinel`] instead.
341const FILE_SENTINEL_MAX_BYTES: u64 = 2 * 1024 * 1024;
342
343/// `AgentContextView.extra` key that opts a step into `@file:` sentinel
344/// resolution (GH #43). Declared through the GH #21 meta channels
345/// (`Blueprint.metas` / `AgentMeta.ctx` / step-level `$step_meta`) and
346/// folded into the view at spawn time by `AgentContextMiddleware`.
347///
348/// Default-deny: absent, or any value other than the strict boolean
349/// `true` (a string `"true"` does not count), rejects the sentinel with
350/// `400`. The v0.9.x line has no sentinel at all, so deny-by-default is
351/// the released-behavior-compatible default; a step whose output
352/// contract legitimately needs file submission opts in with one
353/// declaration.
354const FILE_SENTINEL_ALLOW_KEY: &str = "allow_file_submit";
355
356/// Resolves the `@file:<abs-path>` sentinel (GH #42) when present at the
357/// start of `body_str`. When absent, returns `body_str` unchanged — this
358/// is the byte-for-byte compatible path for all pre-#42 workers.
359///
360/// # Sentinel form
361///
362/// The trimmed body is `@file:<abs-path>` on a single line — a worker
363/// materializes the large payload to a file under its task's `work_dir`
364/// with its existing `Write` capability, then submits the sentinel body
365/// instead of streaming the payload back through the LLM.
366///
367/// # Guards
368///
369/// - Empty / multi-line path → `400`.
370/// - Relative path → `400` (the allowlist works only in
371///   canonicalized-absolute form).
372/// - `AgentContextView` not materialized for `(task_id, attempt)` → `400`
373///   (spawn must have run through `AgentContextMiddleware`; without a
374///   view there is no allowlist root to check against).
375/// - `view.extra[`[`FILE_SENTINEL_ALLOW_KEY`]`]` is not boolean `true` →
376///   `400` (GH #43 — file submission is opt-in per step; default-deny).
377/// - `view.work_dir` is `None` → `400`.
378/// - Canonicalized path is not under canonicalized `work_dir` → `400`
379///   (blocks `..`-escapes and symlinks pointing outside the allowlist).
380/// - File does not exist → `404`.
381/// - File size > [`FILE_SENTINEL_MAX_BYTES`] → `413`.
382/// - Any other I/O / canonicalize error → `500`.
383///
384/// The resolved contents are `trim_end()`-ed to match the inline path's
385/// own trailing-whitespace strip, so the downstream `Value::String` is
386/// observationally identical whether the body arrived inline or via
387/// sentinel.
388async fn resolve_file_sentinel(
389    state: &AppState,
390    task_id: &StepId,
391    attempt: u32,
392    body_str: String,
393) -> Result<String, ApiError> {
394    let Some(rest) = body_str.strip_prefix(FILE_SENTINEL_PREFIX) else {
395        return Ok(body_str);
396    };
397    let path_str = rest.trim();
398    if path_str.is_empty() {
399        return Err(ApiError::bad_request(
400            "@file: sentinel: empty path".to_string(),
401        ));
402    }
403    if path_str.contains('\n') || path_str.contains('\r') {
404        return Err(ApiError::bad_request(
405            "@file: sentinel: path must be a single line".to_string(),
406        ));
407    }
408    let path = std::path::Path::new(path_str);
409    if !path.is_absolute() {
410        return Err(ApiError::bad_request(format!(
411            "@file: sentinel: path must be absolute (got {path_str:?})"
412        )));
413    }
414    let view = state
415        .engine
416        .agent_context_for(task_id, attempt)
417        .await
418        .ok_or_else(|| {
419            ApiError::bad_request(
420                "@file: sentinel: no AgentContextView for this task/attempt \
421                 (spawn must run through AgentContextMiddleware to enable \
422                 sentinel resolution)"
423                    .to_string(),
424            )
425        })?;
426    // GH #43: file submission is opt-in per step (default-deny). Strict
427    // boolean `true` only — folded from the Blueprint meta channels by
428    // `AgentContextMiddleware` at spawn time.
429    if view.extra.get(FILE_SENTINEL_ALLOW_KEY) != Some(&Value::Bool(true)) {
430        return Err(ApiError::bad_request(format!(
431            "@file: sentinel: file submission is not allowed for this step \
432             (declare `{FILE_SENTINEL_ALLOW_KEY}: true` via `$step_meta` / \
433             `AgentMeta.ctx` / `Blueprint.metas`; strict boolean `true` \
434             required)"
435        )));
436    }
437    let work_dir = view.work_dir.ok_or_else(|| {
438        ApiError::bad_request("@file: sentinel: task has no resolved work_dir".to_string())
439    })?;
440    let work_dir_canon = tokio::fs::canonicalize(&work_dir).await.map_err(|e| {
441        ApiError::engine(format!(
442            "@file: sentinel: canonicalize work_dir {work_dir:?}: {e}"
443        ))
444    })?;
445    let path_canon = match tokio::fs::canonicalize(path).await {
446        Ok(p) => p,
447        Err(e) if e.kind() == std::io::ErrorKind::NotFound => {
448            return Err(ApiError::not_found(format!(
449                "@file: sentinel: file not found: {path_str}"
450            )));
451        }
452        Err(e) => {
453            return Err(ApiError::engine(format!(
454                "@file: sentinel: canonicalize {path_str:?}: {e}"
455            )));
456        }
457    };
458    if !path_canon.starts_with(&work_dir_canon) {
459        return Err(ApiError::bad_request(format!(
460            "@file: sentinel: path {} is not under work_dir {} (canonicalized: {} vs {})",
461            path_str,
462            work_dir,
463            path_canon.display(),
464            work_dir_canon.display(),
465        )));
466    }
467    let meta = tokio::fs::metadata(&path_canon)
468        .await
469        .map_err(|e| ApiError::engine(format!("@file: sentinel: metadata {path_str:?}: {e}")))?;
470    if meta.len() > FILE_SENTINEL_MAX_BYTES {
471        return Err(ApiError::payload_too_large(format!(
472            "@file: sentinel: file size {} exceeds limit {}",
473            meta.len(),
474            FILE_SENTINEL_MAX_BYTES
475        )));
476    }
477    let bytes = tokio::fs::read(&path_canon)
478        .await
479        .map_err(|e| ApiError::engine(format!("@file: sentinel: read {path_str:?}: {e}")))?;
480    // Match the `trim_end()` the inline path applies (see `worker_submit`).
481    Ok(String::from_utf8_lossy(&bytes).trim_end().to_string())
482}
483
484/// GH #50 (Subtask 2) — submit-time verdict contract gate, shared by
485/// [`worker_submit`] (`channel = Body`) and [`worker_artifact`]
486/// (`channel = Part`, only when `name == "verdict"`). Enforcement Point 2
487/// (the submit-time complement to `Compiler::compile`'s register-time lint
488/// in `mlua_swarm::blueprint::compiler`, Enforcement Point 1) — called
489/// after the final value string is resolved and BEFORE it is handed to
490/// `submit_worker_result_trusted` / `stage_worker_artifact_trusted`, so a
491/// rejected value never reaches the flow ctx.
492///
493/// No-op (`Ok(())`) in every case that must preserve pre-GH-#50 behavior
494/// byte-for-byte:
495/// - the dispatching agent declared no `VerdictContract` at all (opt-in).
496/// - the agent's declared contract addresses the OTHER channel — a
497///   channel/shape mismatch is the compile-time lint's job (Enforcement
498///   Point 1); this gate only validates value membership for the channel
499///   it was called for.
500/// - `value` IS a member of the contract's declared `values`.
501///
502/// `Err(ApiError::unprocessable(..))` (HTTP 422) otherwise, echoing the
503/// expected token set.
504async fn check_verdict_contract(
505    state: &AppState,
506    task_id: &StepId,
507    channel: VerdictChannel,
508    value: &str,
509) -> Result<(), ApiError> {
510    let Some(contract) = state.engine.verdict_contract_for_task(task_id).await else {
511        return Ok(());
512    };
513    if contract.channel != channel {
514        return Ok(());
515    }
516    if contract.values.iter().any(|v| v == value) {
517        return Ok(());
518    }
519    Err(ApiError::unprocessable(format!(
520        "verdict contract violation: {value:?} is not a member of the declared values {:?}",
521        contract.values
522    )))
523}
524
525/// GH #51 — maps the 2 completion-time verdict-contract `EngineError`
526/// variants (raised by the embedded choke point inside
527/// `Engine::submit_worker_result_trusted` / `Engine::submit_output`) to
528/// their `422` HTTP shape; every other `EngineError` variant falls back
529/// to the pre-existing generic `500` `ApiError::engine` wrapping,
530/// unchanged. Shared by [`worker_submit`] and [`worker_result`] — both
531/// routes surface the SAME embedded engine-side check, so their
532/// HTTP-layer error translation is identical too (this is HTTP
533/// status-code translation, not the verdict-contract logic itself, which
534/// stays the single engine-side choke point per GH #51's "not duplicated
535/// into each route handler" constraint).
536///
537/// `context` labels the wrapped `EngineError`'s `Display` text for the
538/// fallback `500` case only, matching the pre-existing
539/// `format!("<call>: {e}")` style each call site used before this
540/// helper.
541fn map_completion_result<T>(result: Result<T, EngineError>, context: &str) -> Result<T, ApiError> {
542    result.map_err(|e| match e {
543        EngineError::VerdictValueRejected { value, allowed } => ApiError::unprocessable(format!(
544            "verdict contract violation: {value:?} is not a member of the declared values {allowed:?}"
545        )),
546        EngineError::VerdictPartMissing { allowed } => ApiError::unprocessable(format!(
547            "verdict contract violation: no staged \"verdict\" part found for this attempt; declared values {allowed:?}"
548        )),
549        other => ApiError::engine(format!("{context}: {other}")),
550    })
551}
552
553/// `POST /v1/worker/submit`. Bearer = encoded `CapToken`. Body = raw text/octet.
554///
555/// Simplification-axis endpoint for SubAgents. Removes the JSON construction,
556/// duplicated `task_id`, and JSON-escape burden of `/v1/worker/result` — the
557/// worker completes a POST with just token + raw body. Origin: the recent clean-up
558/// of the SubAgent contract drift (fewer IDs to pass around, multi-line escape
559/// accidents eliminated).
560///
561/// **GH #42 `@file:` sentinel**: workers whose result body is too large to
562/// re-emit inline (multi-KB structured output) may `Write` the payload to
563/// a file under their task's `work_dir` and submit the body
564/// `@file:<abs-path>` instead — see [`resolve_file_sentinel`].
565/// Non-sentinel bodies pass through unchanged. The step must opt in via
566/// `allow_file_submit: true` (GH #43, default-deny — see
567/// [`FILE_SENTINEL_ALLOW_KEY`]).
568///
569/// Behavior:
570/// - `task_id` is auto-looked-up server-side from the token (already bound to the `CapToken`).
571/// - Body raw bytes go as-is into `Value::String` for `submit_output` + `post_result`.
572/// - `ok=true` fixed (= the submit endpoint is success-path only). For the error
573///   path, use `/v1/worker/result` with an explicit `ok=false`.
574#[derive(Debug, Deserialize, Default)]
575pub struct SubmitQuery {
576    /// Optional. `ok=false` signals failure (= `DispatchOutcome::Blocked`, caught
577    /// by the flow.ir Try path). Unspecified (`None`) is treated as `ok=true`
578    /// (= normal success).
579    #[serde(default)]
580    pub ok: Option<bool>,
581}
582
583/// `POST /v1/worker/submit`. Simplified counterpart of [`worker_result`]:
584/// the caller sends only the raw result body, `task_id` is resolved
585/// server-side from the Bearer handle/token, and `ok` defaults to `true`
586/// unless overridden via [`SubmitQuery::ok`]. See the module doc for the
587/// short-handle vs full-`CapToken` Bearer forms.
588pub async fn worker_submit(
589    State(state): State<AppState>,
590    headers: HeaderMap,
591    Query(q): Query<SubmitQuery>,
592    body: axum::body::Bytes,
593) -> Result<StatusCode, ApiError> {
594    // Bearer accepts either (a) `wh-<8 hex>` short handle (recommended for
595    // SubAgents) or (b) base64-wrapped CapToken JSON (the full-token form).
596    let bearer = extract_bearer_raw(&headers)?;
597    let task_id = if let Some(handle) = parse_worker_handle(&bearer) {
598        state
599            .engine
600            .task_id_from_handle(handle)
601            .await
602            .map_err(map_handle_lookup_err)?
603    } else {
604        let token = CapToken::decode(bearer.trim())
605            .map_err(|e| ApiError::bad_request(format!("invalid token: {e}")))?;
606        state
607            .engine
608            .task_id_from_token(&token)
609            .await
610            .map_err(|e| ApiError::engine(format!("task_id_from_token: {e}")))?
611    };
612    let attempt = state
613        .engine
614        .task_attempt(&task_id)
615        .await
616        .map_err(|e| ApiError::engine(format!("task_attempt: {e}")))?;
617    // GH #37: fail loud (410) instead of silently accepting a submit whose
618    // addressed Run is already terminal — see `reject_if_run_terminal`.
619    reject_if_run_terminal(&state, &task_id, attempt).await?;
620    // Strip trailing whitespace (newlines, etc.) so flow.ir `Eq` string matches
621    // don't drift on `"BLOCKED\n" == "BLOCKED"` false results. Origin: the recent clean-up
622    // verdict_loop smoke — sharp-edge removal. Internal `\n` inside the raw bytes
623    // is preserved (= only trailing).
624    let body_str = String::from_utf8_lossy(&body).trim_end().to_string();
625    // GH #42: `@file:<abs-path>` sentinel — pass through unchanged when
626    // absent (byte-for-byte compat with pre-#42 callers).
627    let body_str = resolve_file_sentinel(&state, &task_id, attempt, body_str).await?;
628    // GH #51: the `channel: "body"` submit-time check formerly performed
629    // here (`check_verdict_contract(&state, &task_id, VerdictChannel::Body,
630    // ..)`) is now performed inside `Engine::submit_worker_result_trusted`
631    // itself — the single completion-time choke point shared by all 3
632    // completion routes (see `map_completion_result`'s doc). No separate
633    // call is needed here; `check_verdict_contract` remains in use by
634    // `worker_artifact`'s staging-time `name == "verdict"` early
635    // validation, unchanged.
636    let value = Value::String(body_str);
637
638    // The handle path = trusted internal API (= the server-minted handle is validated
639    // by the earlier lookup); the full-token path = existing verify-by-token API.
640    // Both are reflected identically into final + last_result.
641    // `?ok=false` in the query signals failure (= `DispatchOutcome::Blocked`,
642    // the flow.ir Try catch path).
643    let ok = q.ok.unwrap_or(true);
644    map_completion_result(
645        state
646            .engine
647            .submit_worker_result_trusted(&task_id, attempt, value, ok)
648            .await,
649        "submit_worker_result_trusted",
650    )?;
651    Ok(StatusCode::NO_CONTENT)
652}
653
654/// Query params for `POST /v1/worker/artifact`.
655#[derive(Debug, Deserialize)]
656pub struct ArtifactQuery {
657    /// Artifact name (GH #36 ST1: named multi-part worker output). Required
658    /// and non-empty (400 otherwise) — becomes the object key
659    /// `Engine::dispatch_attempt_with`'s Final-pull folds this part under
660    /// (`{"out": <final>, "parts": {<name>: <value>, ...}}`, see that
661    /// method's doc). No character restriction is enforced here (a BP
662    /// author references it via bracket notation, e.g. `$.out.parts["a.b"]`).
663    pub name: String,
664}
665
666/// `POST /v1/worker/artifact?name=<name>`. Bearer = same short-handle /
667/// full-`CapToken` forms as [`worker_submit`]. Body = raw text/octet.
668///
669/// Simplification-axis sibling of [`worker_submit`] (GH #36 ST1): lets a
670/// worker with more than one named result POST each part independently —
671/// same 1-part-per-POST simplicity as `/v1/worker/submit`, no Single Big
672/// JSON the worker has to construct/escape itself — then complete the
673/// attempt with an ordinary `/v1/worker/submit` (unchanged). Staging alone
674/// never completes the attempt; `dispatch_attempt_with` only pulls the
675/// tail's `Final` (whichever endpoint submits it) and folds every staged
676/// `Artifact` into `"parts"` at that point.
677///
678/// Behavior:
679/// - `task_id` is auto-looked-up server-side from the token/handle, same as
680///   [`worker_submit`].
681/// - `name` is required and non-empty; missing or blank → 400.
682/// - Body raw bytes go as-is into `Value::String` (same trailing-whitespace
683///   trim as `worker_submit`) and are staged via
684///   [`mlua_swarm::core::engine::Engine::stage_worker_artifact_trusted`].
685/// - Staging the same `name` twice within one attempt: last write wins (the
686///   Final-pull fold walks the tail in event order — see its doc).
687pub async fn worker_artifact(
688    State(state): State<AppState>,
689    headers: HeaderMap,
690    Query(q): Query<ArtifactQuery>,
691    body: axum::body::Bytes,
692) -> Result<StatusCode, ApiError> {
693    let name = q.name.trim();
694    if name.is_empty() {
695        return Err(ApiError::bad_request("name must not be empty".into()));
696    }
697    let name = name.to_string();
698
699    let bearer = extract_bearer_raw(&headers)?;
700    let task_id = if let Some(handle) = parse_worker_handle(&bearer) {
701        state
702            .engine
703            .task_id_from_handle(handle)
704            .await
705            .map_err(map_handle_lookup_err)?
706    } else {
707        let token = CapToken::decode(bearer.trim())
708            .map_err(|e| ApiError::bad_request(format!("invalid token: {e}")))?;
709        state
710            .engine
711            .task_id_from_token(&token)
712            .await
713            .map_err(|e| ApiError::engine(format!("task_id_from_token: {e}")))?
714    };
715    let attempt = state
716        .engine
717        .task_attempt(&task_id)
718        .await
719        .map_err(|e| ApiError::engine(format!("task_attempt: {e}")))?;
720    // GH #37: fail loud (410) instead of silently staging a part whose
721    // addressed Run is already terminal — see `reject_if_run_terminal`.
722    reject_if_run_terminal(&state, &task_id, attempt).await?;
723    let body_str = String::from_utf8_lossy(&body).trim_end().to_string();
724    // GH #42: same `@file:<abs-path>` sentinel as `worker_submit`.
725    let body_str = resolve_file_sentinel(&state, &task_id, attempt, body_str).await?;
726    // GH #50: submit-time verdict contract gate (Enforcement Point 2),
727    // ONLY for the literal `"verdict"` part name (Pattern B's staging
728    // channel — see `blueprint-authoring.md`'s "Returning verdicts to
729    // drive BP flow" section). Every other part name skips the gate
730    // entirely, unchanged from pre-GH-#50 behavior.
731    if name == "verdict" {
732        check_verdict_contract(&state, &task_id, VerdictChannel::Part, &body_str).await?;
733    }
734    let value = Value::String(body_str);
735
736    state
737        .engine
738        .stage_worker_artifact_trusted(&task_id, attempt, name, value)
739        .await
740        .map_err(|e| ApiError::engine(format!("stage_worker_artifact_trusted: {e}")))?;
741    Ok(StatusCode::NO_CONTENT)
742}
743
744/// Body for `POST /v1/worker/degradation` (GH #32).
745#[derive(Debug, Deserialize)]
746pub struct DegradationBody {
747    /// The tool (or capability) the worker attempted to use.
748    pub tool: String,
749    /// The error that triggered the fallback, in the worker's own words.
750    pub error: String,
751    /// What the worker substituted instead of failing.
752    pub fallback: String,
753    /// Optional free-form context from the worker.
754    #[serde(default)]
755    pub note: Option<String>,
756}
757
758/// `POST /v1/worker/degradation` (GH #32). Bearer = same short-handle /
759/// full-`CapToken` forms as [`worker_submit`]. Body = JSON, not raw bytes —
760/// this endpoint carries structured data, unlike its raw-bytes siblings.
761///
762/// Independent channel: appends a [`DegradationEntry`] to
763/// `RunRecord.degradations` via `RunStore::append_degradation` directly.
764/// Never touches `OutputStore` / the fold path (Crux invariant 2 — a
765/// degradation must not surface as step OUTPUT / `$.step.parts`).
766///
767/// Behavior:
768/// - `task_id` is auto-looked-up server-side from the token/handle, same as
769///   [`worker_submit`] / [`worker_artifact`].
770/// - GH #37 terminal-run guard applies first — a degradation addressed at
771///   an already-terminal Run is rejected with `410 Gone`
772///   ([`reject_if_run_terminal`]), same as a submit/artifact would be.
773/// - `step_ref` / `attempt` / `at` are server-injected — `step_ref` is the
774///   fetching agent's resolved name (`AgentContextView.agent`, the best
775///   proxy for `Step.ref` available at this layer), `attempt` is the
776///   task's current attempt, `at` is now (Unix epoch seconds). The client
777///   body never supplies any of the three.
778/// - No Run linkage in `agent_ctx` (a pre-run-tracking dispatch), an
779///   unparseable `run_id`, or an `append_degradation` call against a Run
780///   the store doesn't actually hold (`RunStoreError::NotFound` — the same
781///   condition [`reject_if_run_terminal`] itself fails open on) all take
782///   the same silent `204 No Content` path, logged via `tracing::warn!` —
783///   this is a legitimate no-tracking codepath, not a client error. Any
784///   other `RunStore` failure propagates as `ApiError::engine`.
785pub async fn worker_degradation(
786    State(state): State<AppState>,
787    headers: HeaderMap,
788    Json(body): Json<DegradationBody>,
789) -> Result<StatusCode, ApiError> {
790    let bearer = extract_bearer_raw(&headers)?;
791    let task_id = if let Some(handle) = parse_worker_handle(&bearer) {
792        state
793            .engine
794            .task_id_from_handle(handle)
795            .await
796            .map_err(map_handle_lookup_err)?
797    } else {
798        let token = CapToken::decode(bearer.trim())
799            .map_err(|e| ApiError::bad_request(format!("invalid token: {e}")))?;
800        state
801            .engine
802            .task_id_from_token(&token)
803            .await
804            .map_err(|e| ApiError::engine(format!("task_id_from_token: {e}")))?
805    };
806    let attempt = state
807        .engine
808        .task_attempt(&task_id)
809        .await
810        .map_err(|e| ApiError::engine(format!("task_attempt: {e}")))?;
811    // GH #37: the same terminal-run guard `worker_submit` / `worker_artifact`
812    // apply — a dead Run must not accumulate signals.
813    reject_if_run_terminal(&state, &task_id, attempt).await?;
814
815    // Same `with_state` resolution pattern as `reject_if_run_terminal`: an
816    // engine-level failure here is fail-open too (`_ => ...`), matching
817    // that guard's own "every resolution step is fail-open" contract —
818    // this lookup isn't a second, stricter gate on top of it.
819    let tid = task_id.clone();
820    let (run_id_str, agent) = match state
821        .engine
822        .with_state("worker_degradation_run_lookup", move |s| {
823            s.agent_ctx.get(&(tid, attempt)).and_then(|e| {
824                e.view
825                    .run_id
826                    .clone()
827                    .map(|run_id| (run_id, e.view.agent.clone()))
828            })
829        })
830        .await
831    {
832        Ok(Some(pair)) => pair,
833        _ => {
834            tracing::warn!(%task_id, "worker_degradation: no run linkage for this task; entry dropped");
835            return Ok(StatusCode::NO_CONTENT);
836        }
837    };
838    let Ok(run_id) = RunId::parse(run_id_str) else {
839        tracing::warn!(%task_id, "worker_degradation: run_id failed to parse; entry dropped");
840        return Ok(StatusCode::NO_CONTENT);
841    };
842
843    let entry = DegradationEntry {
844        tool: body.tool,
845        error: body.error,
846        fallback: body.fallback,
847        note: body.note,
848        step_ref: Some(agent),
849        attempt: Some(attempt),
850        at: crate::tasks::now_secs(),
851    };
852    match state.run_store.append_degradation(&run_id, entry).await {
853        Ok(()) => Ok(StatusCode::NO_CONTENT),
854        Err(RunStoreError::NotFound(_)) => {
855            tracing::warn!(%task_id, %run_id, "worker_degradation: run not found in run_store; entry dropped");
856            Ok(StatusCode::NO_CONTENT)
857        }
858        Err(e) => Err(ApiError::engine(format!("append_degradation: {e}"))),
859    }
860}
861
862/// GH #37: terminal-run guard shared by [`worker_submit`] / [`worker_artifact`].
863///
864/// Resolves the dispatch task's `AgentContextView.run_id` (threaded at
865/// spawn time when a `RunContext` accompanied the launch) and rejects the
866/// submit with `410 Gone` when the addressed Run has already reached a
867/// terminal status (`Done` / `Failed` / `Interrupted`) — the flow-eval
868/// driver for that Run is gone, so the staged/final value could never be
869/// folded into a flow context. Before this guard, such a submit was
870/// silently accepted with `204` and the worker's output orphaned — the
871/// exact failure shape observed when a long-running worker outlived the
872/// GH #33 sync launch ceiling.
873///
874/// Every resolution step is fail-open (missing agent-ctx entry / missing
875/// `run_id` / unparseable id / unknown Run → `Ok(())`), matching this
876/// crate's other best-effort projection hooks: a pre-run-tracking dispatch
877/// must keep working exactly as before.
878async fn reject_if_run_terminal(
879    state: &AppState,
880    task_id: &StepId,
881    attempt: u32,
882) -> Result<(), ApiError> {
883    let tid = task_id.clone();
884    let run_id_str = match state
885        .engine
886        .with_state("worker_terminal_run_guard", move |s| {
887            s.agent_ctx
888                .get(&(tid, attempt))
889                .and_then(|e| e.view.run_id.clone())
890        })
891        .await
892    {
893        Ok(Some(rid)) => rid,
894        _ => return Ok(()),
895    };
896    let Ok(run_id) = RunId::parse(run_id_str) else {
897        return Ok(());
898    };
899    let Ok(rec) = state.run_store.get(&run_id).await else {
900        return Ok(());
901    };
902    match rec.status {
903        RunStatus::Done | RunStatus::Failed | RunStatus::Interrupted => {
904            Err(ApiError::gone(format!(
905                "run {run_id} is already terminal ({:?}): this attempt's output cannot be \
906                 delivered to a flow context; re-kick the task (POST /v1/tasks/:id/runs) and \
907                 fetch a fresh prompt",
908                rec.status
909            )))
910        }
911        RunStatus::Pending | RunStatus::Running => Ok(()),
912    }
913}
914
915/// Query params for `GET /v1/worker/prompt/system`. Field names are fixed to
916/// `task_id` / `attempt` — this is the exact shape the engine bakes into
917/// `system_ref.uri`'s query string for `Http` mode (GH #31), so the names
918/// here must match verbatim.
919#[derive(Debug, Deserialize)]
920pub struct PromptSystemQuery {
921    /// Task the fetched raw system prompt belongs to; cross-checked
922    /// against the Bearer handle/token, same as [`PromptQuery::task_id`].
923    pub task_id: StepId,
924    /// Attempt number the baked system prompt was recorded under.
925    pub attempt: u32,
926}
927
928/// `GET /v1/worker/prompt/system?task_id=<tid>&attempt=<n>` (GH #31). The
929/// `Http`-mode fetch target for `system_ref.uri`: serves the exact baked
930/// `system` bytes for `(task_id, attempt)` as a raw `text/plain` body — not
931/// JSON-wrapped, since `mse_worker_fetch` needs the precise byte sequence to
932/// sha256-verify against `system_ref.sha256`.
933///
934/// Same Bearer auth flow as [`worker_prompt`] (short handle or full
935/// `CapToken`); 404 via [`ApiError::not_found`] if no baked system exists for
936/// that `(task_id, attempt)`.
937pub async fn worker_prompt_system(
938    State(state): State<AppState>,
939    headers: HeaderMap,
940    Query(q): Query<PromptSystemQuery>,
941) -> Result<impl axum::response::IntoResponse, ApiError> {
942    let task_id = q.task_id;
943    let attempt = q.attempt;
944    let bearer = extract_bearer_raw(&headers)?;
945    if let Some(handle) = parse_worker_handle(&bearer) {
946        let resolved = state
947            .engine
948            .task_id_from_handle(handle)
949            .await
950            .map_err(map_handle_lookup_err)?;
951        if resolved != task_id {
952            return Err(ApiError::bad_request(format!(
953                "handle {handle} is bound to task {resolved}, not {task_id}"
954            )));
955        }
956    } else {
957        let token = CapToken::decode(bearer.trim())
958            .map_err(|e| ApiError::bad_request(format!("invalid token: {e}")))?;
959        state
960            .engine
961            .verify_token_for_task(&token, mlua_swarm::Verb::FetchPrompt, &task_id)
962            .await
963            .map_err(|e| ApiError::engine(format!("verify_token_for_task: {e}")))?;
964    }
965    let system = state
966        .engine
967        .raw_system_prompt(&task_id, attempt)
968        .await
969        .map_err(|e| ApiError::engine(format!("raw_system_prompt: {e}")))?
970        .ok_or_else(|| {
971            ApiError::not_found(format!(
972                "no baked system prompt for task {task_id} attempt {attempt}"
973            ))
974        })?;
975    Ok((
976        [(header::CONTENT_TYPE, "text/plain; charset=utf-8")],
977        system,
978    ))
979}
980
981/// Response body for `GET /v1/agents/:name/render-size`.
982#[derive(Debug, serde::Serialize)]
983pub struct AgentRenderSizeResponse {
984    /// The agent name looked up (echoed back verbatim from the path param).
985    pub agent: String,
986    /// Most-recently-baked `system_prompt` render size in bytes for this
987    /// agent, or `None` if `bake_worker_system_prompt` has never recorded
988    /// one (a freshly-added agent that has never been dispatched).
989    pub last_rendered_bytes: Option<usize>,
990}
991
992/// `GET /v1/agents/:name/render-size` (GH #31). Live per-agent-name lookup
993/// of the most-recently-baked `system_prompt` render size, backing
994/// `bp_doctor`'s post-render size check. No Bearer required — same
995/// unauthenticated trust tier as `GET /v1/blueprints/:id/head`
996/// (`blueprints::get_head`), an operator-diagnostic route.
997///
998/// `last_rendered_bytes: null` is a normal, expected response (a
999/// freshly-added agent that has never been dispatched yet) — always
1000/// `200 OK`, never a 404.
1001pub async fn agent_render_size(
1002    State(state): State<AppState>,
1003    axum::extract::Path(name): axum::extract::Path<String>,
1004) -> Json<AgentRenderSizeResponse> {
1005    let last_rendered_bytes = state.engine.agent_last_rendered_size(&name).await;
1006    Json(AgentRenderSizeResponse {
1007        agent: name,
1008        last_rendered_bytes,
1009    })
1010}
1011
1012/// Extracts the raw string from the `Authorization` header (= strips the `Bearer `
1013/// prefix). To let `worker_submit` accept both short handles and full tokens, we
1014/// fetch the raw value before any decode.
1015fn extract_bearer_raw(headers: &HeaderMap) -> Result<String, ApiError> {
1016    let v = headers
1017        .get(AUTHORIZATION)
1018        .ok_or_else(|| ApiError::bad_request("missing Authorization header".into()))?
1019        .to_str()
1020        .map_err(|_| ApiError::bad_request("invalid Authorization header encoding".into()))?;
1021    let s = v
1022        .strip_prefix("Bearer ")
1023        .ok_or_else(|| ApiError::bad_request("Authorization must be 'Bearer <token>'".into()))?
1024        .trim();
1025    if s.is_empty() {
1026        return Err(ApiError::bad_request("Bearer is empty".into()));
1027    }
1028    Ok(s.to_string())
1029}
1030
1031/// Maps a `task_id_from_handle` failure on the short-handle Bearer path. A
1032/// `wh-`-prefixed handle that parsed as well-formed (via
1033/// [`parse_worker_handle`]) but is unknown to the engine
1034/// (`EngineError::TokenNotFound`) is a handle that was minted before the
1035/// engine's in-memory state was wiped — typically a server restart. "Once
1036/// valid, now gone" is exactly `410 Gone`: the worker should re-kick its
1037/// task and fetch a fresh handle rather than treat this as a server fault.
1038/// Every other engine error stays a `500` (unchanged), same wrapping style
1039/// as the pre-existing call sites.
1040///
1041/// Only reached on the handle path (`parse_worker_handle` returned `Some`),
1042/// so it never re-labels a full-`CapToken` decode/verify failure.
1043fn map_handle_lookup_err(e: EngineError) -> ApiError {
1044    match e {
1045        EngineError::TokenNotFound(_) => ApiError::gone(
1046            "worker handle is no longer valid (the engine's in-flight state was reset, \
1047             e.g. by a server restart): re-kick the task (POST /v1/tasks/:id/runs) and \
1048             fetch a fresh prompt/handle"
1049                .to_string(),
1050        ),
1051        other => ApiError::engine(format!("task_id_from_handle: {other}")),
1052    }
1053}
1054
1055/// Decides whether the Bearer is a short handle (`wh-XXXXXXXX`). Returns
1056/// `Some(handle)` on a match, `None` otherwise (= caller proceeds to try decoding
1057/// as full `CapToken` JSON).
1058fn parse_worker_handle(s: &str) -> Option<&str> {
1059    let s = s.trim();
1060    if s.starts_with("wh-")
1061        && s.len() >= 5
1062        && s.len() <= 64
1063        && s[3..].chars().all(|c| c.is_ascii_alphanumeric())
1064    {
1065        Some(s)
1066    } else {
1067        None
1068    }
1069}
1070
1071/// Decodes an encoded `CapToken` from `Authorization: Bearer <encoded CapToken>`.
1072/// Kept separate from `extract_bearer` (sid-only) — kept as a distinct fn so
1073/// that sid strings and encoded tokens are not confused, distinguishing them by type.
1074fn decode_worker_bearer(headers: &HeaderMap) -> Result<CapToken, ApiError> {
1075    let v = headers
1076        .get(AUTHORIZATION)
1077        .ok_or_else(|| ApiError::bad_request("missing Authorization header".into()))?
1078        .to_str()
1079        .map_err(|_| ApiError::bad_request("invalid Authorization header encoding".into()))?;
1080    let encoded = v
1081        .strip_prefix("Bearer ")
1082        .ok_or_else(|| ApiError::bad_request("Authorization must be 'Bearer <token>'".into()))?
1083        .trim();
1084    if encoded.is_empty() {
1085        return Err(ApiError::bad_request("Bearer token is empty".into()));
1086    }
1087    CapToken::decode(encoded).map_err(|e| ApiError::bad_request(format!("invalid token: {e}")))
1088}
1089
1090// ──────────────────────────────────────────────────────────────────────────
1091// UT — `assemble_step_pointers` (`projection-adapter` ST5 Worker axis)
1092// ──────────────────────────────────────────────────────────────────────────
1093
1094#[cfg(test)]
1095mod tests {
1096    use super::*;
1097    use axum::response::IntoResponse;
1098    use mlua_swarm::core::agent_context::AgentContextView;
1099    use mlua_swarm::core::config::EngineCfg;
1100    use mlua_swarm::core::engine::Engine;
1101    use mlua_swarm::store::output::{InMemoryOutputStore, OutputStore};
1102    use mlua_swarm::store::run::{InMemoryRunStore, RunRecord, RunStatus, RunStore, StepEntry};
1103    use mlua_swarm::store::task::InMemoryTaskStore;
1104    use mlua_swarm::{RunId, StepId, TaskId};
1105    use serde_json::json;
1106    use std::collections::HashMap;
1107    use std::sync::Arc;
1108    use tokio::sync::Mutex;
1109
1110    /// Per-module test-helper convention (this crate's established
1111    /// pattern — see e.g. `projection::tests::test_state`): a minimal
1112    /// `AppState` wired with the caller-supplied `data_store` / `run_store`
1113    /// so a test can seed both directly rather than driving a real
1114    /// dispatch through them.
1115    fn test_state(data_store: Arc<dyn OutputStore>, run_store: Arc<dyn RunStore>) -> AppState {
1116        let engine = Engine::new(EngineCfg::default());
1117        let compiler = mlua_swarm::Compiler::new(crate::default_registry());
1118        let launch = Arc::new(mlua_swarm::TaskLaunchService::new(engine.clone(), compiler));
1119        AppState {
1120            engine,
1121            sessions: Arc::new(Mutex::new(crate::SessionStore::default())),
1122            task_app: Arc::new(mlua_swarm::TaskApplication::new_inline_only(launch)),
1123            ws_operator_factory: None,
1124            data_store,
1125            operator_sessions: Arc::new(Mutex::new(HashMap::new())),
1126            roles_to_sid: Arc::new(Mutex::new(HashMap::new())),
1127            task_store: Arc::new(InMemoryTaskStore::new()),
1128            run_store,
1129            replay_store: Arc::new(mlua_swarm::store::replay::InMemoryReplayStore::new()),
1130            base_url: None,
1131            sync_timeout_secs: 300,
1132        }
1133    }
1134
1135    async fn append_final(
1136        data_store: &Arc<dyn OutputStore>,
1137        task_id: &str,
1138        producer: &str,
1139        value: Value,
1140    ) {
1141        data_store
1142            .append(
1143                task_id,
1144                1,
1145                producer,
1146                OutputEvent::Final {
1147                    content: ContentRef::Inline { value },
1148                    ok: true,
1149                },
1150                vec![],
1151            )
1152            .await
1153            .expect("append final");
1154    }
1155
1156    fn step_entry(step_id: &StepId, step_ref: &str) -> StepEntry {
1157        StepEntry {
1158            step_id: step_id.clone(),
1159            step_ref: Some(step_ref.to_string()),
1160            status: Some("passed".to_string()),
1161            at: 0,
1162        }
1163    }
1164
1165    fn run_record(task_id: &TaskId, run_id: &RunId, step_entries: Vec<StepEntry>) -> RunRecord {
1166        RunRecord {
1167            id: run_id.clone(),
1168            task_id: task_id.clone(),
1169            status: RunStatus::Running,
1170            step_entries,
1171            degradations: Vec::new(),
1172            operator_sid: None,
1173            result_ref: None,
1174            input_json: None,
1175            created_at: 0,
1176            updated_at: 0,
1177        }
1178    }
1179
1180    fn consumer_payload(consumer_step_id: &StepId, run_id: &RunId) -> WorkerPayload {
1181        WorkerPayload {
1182            task_id: consumer_step_id.clone(),
1183            attempt: 1,
1184            agent: "consumer".to_string(),
1185            system: None,
1186            prompt: String::new(),
1187            context: Some(AgentContextView {
1188                task_id: consumer_step_id.to_string(),
1189                agent: "consumer".to_string(),
1190                attempt: 1,
1191                run_id: Some(run_id.to_string()),
1192                ..Default::default()
1193            }),
1194            system_ref: None,
1195        }
1196    }
1197
1198    /// Test 1: `ContextPolicy.steps` unspecified (no policy seeded at all
1199    /// — `Engine::context_policy_for`'s "no entry" default is `None` /
1200    /// pass-all) → the fetch payload carries every submitted step's
1201    /// `StepPointer`.
1202    #[tokio::test]
1203    async fn context_policy_unspecified_yields_every_submitted_step() {
1204        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
1205        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
1206        let task_id = TaskId::new();
1207        let run_id = RunId::new();
1208        let planner_id = StepId::new();
1209        let coder_id = StepId::new();
1210
1211        append_final(
1212            &data_store,
1213            planner_id.as_str(),
1214            "planner",
1215            json!({"plan": "x"}),
1216        )
1217        .await;
1218        append_final(
1219            &data_store,
1220            coder_id.as_str(),
1221            "coder",
1222            json!({"code": "y"}),
1223        )
1224        .await;
1225        run_store
1226            .create(run_record(
1227                &task_id,
1228                &run_id,
1229                vec![
1230                    step_entry(&planner_id, "planner"),
1231                    step_entry(&coder_id, "coder"),
1232                ],
1233            ))
1234            .await
1235            .expect("create run");
1236
1237        let state = test_state(data_store, run_store);
1238        let consumer_id = StepId::new();
1239        let mut payload = consumer_payload(&consumer_id, &run_id);
1240        assemble_step_pointers(&state, &mut payload).await;
1241
1242        let names: Vec<&str> = payload
1243            .context
1244            .as_ref()
1245            .expect("context")
1246            .steps
1247            .iter()
1248            .map(|p| p.name.as_str())
1249            .collect();
1250        assert!(names.contains(&"planner"), "names: {names:?}");
1251        assert!(names.contains(&"coder"), "names: {names:?}");
1252    }
1253
1254    /// Test 2: `steps: ["planner"]` → only `planner`'s pointer is present.
1255    #[tokio::test]
1256    async fn context_policy_steps_include_list_filters_to_named_steps() {
1257        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
1258        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
1259        let task_id = TaskId::new();
1260        let run_id = RunId::new();
1261        let planner_id = StepId::new();
1262        let coder_id = StepId::new();
1263        append_final(&data_store, planner_id.as_str(), "planner", json!("x")).await;
1264        append_final(&data_store, coder_id.as_str(), "coder", json!("y")).await;
1265        run_store
1266            .create(run_record(
1267                &task_id,
1268                &run_id,
1269                vec![
1270                    step_entry(&planner_id, "planner"),
1271                    step_entry(&coder_id, "coder"),
1272                ],
1273            ))
1274            .await
1275            .expect("create run");
1276
1277        let state = test_state(data_store, run_store);
1278        let consumer_id = StepId::new();
1279        state
1280            .engine
1281            .with_state("test.seed_policy", {
1282                let consumer_id = consumer_id.clone();
1283                move |s| {
1284                    s.agent_ctx.insert(
1285                        (consumer_id, 1),
1286                        mlua_swarm::core::state::AgentCtxEntry {
1287                            policy: mlua_swarm_schema::ContextPolicy {
1288                                steps: Some(vec!["planner".to_string()]),
1289                                ..Default::default()
1290                            },
1291                            ..Default::default()
1292                        },
1293                    );
1294                }
1295            })
1296            .await
1297            .expect("seed policy");
1298
1299        let mut payload = consumer_payload(&consumer_id, &run_id);
1300        assemble_step_pointers(&state, &mut payload).await;
1301
1302        let names: Vec<&str> = payload
1303            .context
1304            .as_ref()
1305            .expect("context")
1306            .steps
1307            .iter()
1308            .map(|p| p.name.as_str())
1309            .collect();
1310        assert_eq!(names, vec!["planner"], "names: {names:?}");
1311    }
1312
1313    /// Test 3: `steps: []` → the pointer list is empty.
1314    #[tokio::test]
1315    async fn context_policy_steps_empty_list_yields_no_pointers() {
1316        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
1317        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
1318        let task_id = TaskId::new();
1319        let run_id = RunId::new();
1320        let planner_id = StepId::new();
1321        append_final(&data_store, planner_id.as_str(), "planner", json!("x")).await;
1322        run_store
1323            .create(run_record(
1324                &task_id,
1325                &run_id,
1326                vec![step_entry(&planner_id, "planner")],
1327            ))
1328            .await
1329            .expect("create run");
1330
1331        let state = test_state(data_store, run_store);
1332        let consumer_id = StepId::new();
1333        state
1334            .engine
1335            .with_state("test.seed_policy", {
1336                let consumer_id = consumer_id.clone();
1337                move |s| {
1338                    s.agent_ctx.insert(
1339                        (consumer_id, 1),
1340                        mlua_swarm::core::state::AgentCtxEntry {
1341                            policy: mlua_swarm_schema::ContextPolicy {
1342                                steps: Some(vec![]),
1343                                ..Default::default()
1344                            },
1345                            ..Default::default()
1346                        },
1347                    );
1348                }
1349            })
1350            .await
1351            .expect("seed policy");
1352
1353        let mut payload = consumer_payload(&consumer_id, &run_id);
1354        assemble_step_pointers(&state, &mut payload).await;
1355
1356        assert!(payload.context.expect("context").steps.is_empty());
1357    }
1358
1359    /// Test 4: `steps_exclude` wins over `steps` for a name in both.
1360    #[tokio::test]
1361    async fn context_policy_steps_exclude_wins_over_steps() {
1362        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
1363        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
1364        let task_id = TaskId::new();
1365        let run_id = RunId::new();
1366        let planner_id = StepId::new();
1367        let coder_id = StepId::new();
1368        append_final(&data_store, planner_id.as_str(), "planner", json!("x")).await;
1369        append_final(&data_store, coder_id.as_str(), "coder", json!("y")).await;
1370        run_store
1371            .create(run_record(
1372                &task_id,
1373                &run_id,
1374                vec![
1375                    step_entry(&planner_id, "planner"),
1376                    step_entry(&coder_id, "coder"),
1377                ],
1378            ))
1379            .await
1380            .expect("create run");
1381
1382        let state = test_state(data_store, run_store);
1383        let consumer_id = StepId::new();
1384        state
1385            .engine
1386            .with_state("test.seed_policy", {
1387                let consumer_id = consumer_id.clone();
1388                move |s| {
1389                    s.agent_ctx.insert(
1390                        (consumer_id, 1),
1391                        mlua_swarm::core::state::AgentCtxEntry {
1392                            policy: mlua_swarm_schema::ContextPolicy {
1393                                steps: Some(vec!["planner".to_string(), "coder".to_string()]),
1394                                steps_exclude: vec!["planner".to_string()],
1395                                ..Default::default()
1396                            },
1397                            ..Default::default()
1398                        },
1399                    );
1400                }
1401            })
1402            .await
1403            .expect("seed policy");
1404
1405        let mut payload = consumer_payload(&consumer_id, &run_id);
1406        assemble_step_pointers(&state, &mut payload).await;
1407
1408        let names: Vec<&str> = payload
1409            .context
1410            .as_ref()
1411            .expect("context")
1412            .steps
1413            .iter()
1414            .map(|p| p.name.as_str())
1415            .collect();
1416        assert_eq!(names, vec!["coder"], "names: {names:?}");
1417    }
1418
1419    /// Test 5 (in-flight window, subtask-4-style invariant): the Run has
1420    /// NOT finalized (`result_ref: None`, mirroring a Run still `Running`)
1421    /// yet the fetch payload still carries a `StepPointer` for a step
1422    /// already visible through the Data-plane store — the same mechanism
1423    /// `crates/mlua-swarm-server/src/projection.rs`'s
1424    /// `steps_list_returns_in_flight_step_output_before_run_completes`
1425    /// proves end-to-end through a real gated 2-step dispatch; this test
1426    /// isolates the same invariant at the `assemble_step_pointers` level.
1427    #[tokio::test]
1428    async fn in_flight_step_output_is_visible_before_run_finalizes() {
1429        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
1430        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
1431        let task_id = TaskId::new();
1432        let run_id = RunId::new();
1433        let step1_id = StepId::new();
1434        append_final(
1435            &data_store,
1436            step1_id.as_str(),
1437            "step1",
1438            json!({"step1_out": "hi"}),
1439        )
1440        .await;
1441        let mut run = run_record(&task_id, &run_id, vec![step_entry(&step1_id, "step1")]);
1442        run.status = RunStatus::Running;
1443        run.result_ref = None; // the in-flight window: not yet finalized.
1444        run_store.create(run).await.expect("create run");
1445
1446        let state = test_state(data_store, run_store);
1447        let consumer_id = StepId::new();
1448        let mut payload = consumer_payload(&consumer_id, &run_id);
1449        assemble_step_pointers(&state, &mut payload).await;
1450
1451        let steps = &payload.context.expect("context").steps;
1452        assert_eq!(steps.len(), 1);
1453        assert_eq!(steps[0].name, "step1");
1454    }
1455
1456    /// Test 6: the fetching agent's own name is always excluded, even if
1457    /// (e.g. a loop re-dispatching the same agent) it also appears in
1458    /// `run.step_entries` with a resolvable Data-plane record.
1459    #[tokio::test]
1460    async fn self_agent_name_is_always_excluded() {
1461        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
1462        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
1463        let task_id = TaskId::new();
1464        let run_id = RunId::new();
1465        let planner_id = StepId::new();
1466        let consumer_prior_id = StepId::new();
1467        append_final(&data_store, planner_id.as_str(), "planner", json!("x")).await;
1468        append_final(
1469            &data_store,
1470            consumer_prior_id.as_str(),
1471            "consumer",
1472            json!("self"),
1473        )
1474        .await;
1475        run_store
1476            .create(run_record(
1477                &task_id,
1478                &run_id,
1479                vec![
1480                    step_entry(&planner_id, "planner"),
1481                    step_entry(&consumer_prior_id, "consumer"),
1482                ],
1483            ))
1484            .await
1485            .expect("create run");
1486
1487        let state = test_state(data_store, run_store);
1488        let consumer_id = StepId::new();
1489        let mut payload = consumer_payload(&consumer_id, &run_id);
1490        assemble_step_pointers(&state, &mut payload).await;
1491
1492        let names: Vec<&str> = payload
1493            .context
1494            .as_ref()
1495            .expect("context")
1496            .steps
1497            .iter()
1498            .map(|p| p.name.as_str())
1499            .collect();
1500        assert!(!names.contains(&"consumer"), "names: {names:?}");
1501        assert!(names.contains(&"planner"), "names: {names:?}");
1502    }
1503
1504    /// Test 7 (pointer-only invariant): a `StepPointer`'s serialized JSON
1505    /// carries no preview / content-bytes field — only `name` /
1506    /// `size_bytes` / `file_path?` / `content_url` / `sha256`.
1507    #[tokio::test]
1508    async fn step_pointer_serializes_with_no_preview_or_content_bytes() {
1509        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
1510        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
1511        let task_id = TaskId::new();
1512        let run_id = RunId::new();
1513        let planner_id = StepId::new();
1514        append_final(
1515            &data_store,
1516            planner_id.as_str(),
1517            "planner",
1518            json!({"plan": "do the thing, at length".repeat(50)}),
1519        )
1520        .await;
1521        run_store
1522            .create(run_record(
1523                &task_id,
1524                &run_id,
1525                vec![step_entry(&planner_id, "planner")],
1526            ))
1527            .await
1528            .expect("create run");
1529
1530        let state = test_state(data_store, run_store);
1531        let consumer_id = StepId::new();
1532        let mut payload = consumer_payload(&consumer_id, &run_id);
1533        assemble_step_pointers(&state, &mut payload).await;
1534
1535        let steps = &payload.context.expect("context").steps;
1536        assert_eq!(steps.len(), 1);
1537        let json_value = serde_json::to_value(&steps[0]).expect("serialize StepPointer");
1538        let obj = json_value.as_object().expect("object");
1539        for forbidden in ["preview", "content", "value", "bytes"] {
1540            assert!(
1541                !obj.contains_key(forbidden),
1542                "StepPointer must not carry a {forbidden:?} field: {obj:?}"
1543            );
1544        }
1545        assert!(obj.contains_key("name"));
1546        assert!(obj.contains_key("size_bytes"));
1547        assert!(obj.contains_key("content_url"));
1548        assert!(obj.contains_key("sha256"));
1549    }
1550
1551    /// A single-step Blueprint whose `planner` agent declares
1552    /// `AgentMeta.projection_name = "plan-out"` — the `StepNaming` fixture
1553    /// for [`declared_projection_name_pointer_name_is_canonical_and_policy_matches_it`],
1554    /// mirroring `crate::projection::tests`' own
1555    /// `declared_projection_name_blueprint` helper (duplicated here rather
1556    /// than shared — this crate's established per-module test-helper
1557    /// convention).
1558    fn declared_name_bp() -> mlua_swarm::blueprint::Blueprint {
1559        use mlua_flow_ir::{Expr, Node};
1560        use mlua_swarm::blueprint::{
1561            current_schema_version, AgentDef, AgentKind, AgentMeta, Blueprint, BlueprintMetadata,
1562            CompilerHints, CompilerStrategy,
1563        };
1564        Blueprint {
1565            schema_version: current_schema_version(),
1566            id: "worker-test-declared-name-bp".into(),
1567            flow: Node::Step {
1568                ref_: "planner".to_string(),
1569                in_: Expr::Path {
1570                    at: "$.in".parse().expect("literal test path: $.in"),
1571                },
1572                out: Expr::Path {
1573                    at: "$.plan".parse().expect("literal test path: $.plan"),
1574                },
1575            },
1576            agents: vec![AgentDef {
1577                name: "planner".to_string(),
1578                kind: AgentKind::RustFn,
1579                spec: json!({"fn_id": "planner"}),
1580                profile: None,
1581                meta: Some(AgentMeta {
1582                    projection_name: Some("plan-out".to_string()),
1583                    ..Default::default()
1584                }),
1585                runner: None,
1586                runner_ref: None,
1587                verdict: None,
1588            }],
1589            operators: vec![],
1590            metas: vec![],
1591            hints: CompilerHints::default(),
1592            strategy: CompilerStrategy::default(),
1593            metadata: BlueprintMetadata::default(),
1594            spawner_hints: Default::default(),
1595            default_agent_kind: AgentKind::Operator,
1596            default_operator_kind: None,
1597            default_init_ctx: None,
1598            default_agent_ctx: None,
1599            default_context_policy: None,
1600            projection_placement: None,
1601            audits: vec![],
1602            degradation_policy: None,
1603            runners: vec![],
1604            default_runner: None,
1605            check_policy: None,
1606            blueprint_ref_includes: Vec::new(),
1607        }
1608    }
1609
1610    /// Test 8 (GH #23 subtask-3, declared-name E2E — Worker axis half): a
1611    /// declared `projection_name` makes `StepPointer.name` the CANONICAL
1612    /// name (not the raw `Step.ref` the Data-plane / `step_entries` still
1613    /// index by), and `ContextPolicy.steps` naming the canonical name
1614    /// matches it.
1615    #[tokio::test]
1616    async fn declared_projection_name_pointer_name_is_canonical_and_policy_matches_it() {
1617        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
1618        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
1619        let task_id = TaskId::new();
1620        let run_id = RunId::new();
1621        let planner_id = StepId::new();
1622
1623        // The Data-plane store is keyed by the CANONICAL name — GH #23
1624        // subtask-2's sink already writes it that way.
1625        append_final(
1626            &data_store,
1627            planner_id.as_str(),
1628            "plan-out",
1629            json!({"plan": "x"}),
1630        )
1631        .await;
1632        run_store
1633            .create(run_record(
1634                &task_id,
1635                &run_id,
1636                vec![step_entry(&planner_id, "planner")],
1637            ))
1638            .await
1639            .expect("create run");
1640
1641        let state = test_state(data_store, run_store);
1642
1643        // Seed the `StepNaming` table the way `Compiler::compile` +
1644        // `EngineDispatcher::dispatch` would have — the same `Arc` stashed
1645        // under every dispatched step's own id, including the FETCHING
1646        // agent's (`consumer_id`), which `assemble_step_pointers` looks up
1647        // via `Engine::step_naming_for(&payload.task_id)`.
1648        let (naming, _warnings) =
1649            mlua_swarm::core::step_naming::StepNaming::from_blueprint(&declared_name_bp())
1650                .expect("no collision");
1651        let naming = Arc::new(naming);
1652        let consumer_id = StepId::new();
1653        state
1654            .engine
1655            .with_state("test.seed_step_naming", {
1656                let naming = naming.clone();
1657                let planner_id = planner_id.clone();
1658                let consumer_id = consumer_id.clone();
1659                move |s| {
1660                    s.step_namings.insert(planner_id, naming.clone());
1661                    s.step_namings.insert(consumer_id, naming);
1662                }
1663            })
1664            .await
1665            .expect("seed step naming");
1666        state
1667            .engine
1668            .with_state("test.seed_policy", {
1669                let consumer_id = consumer_id.clone();
1670                move |s| {
1671                    s.agent_ctx.insert(
1672                        (consumer_id, 1),
1673                        mlua_swarm::core::state::AgentCtxEntry {
1674                            policy: mlua_swarm_schema::ContextPolicy {
1675                                steps: Some(vec!["plan-out".to_string()]),
1676                                ..Default::default()
1677                            },
1678                            ..Default::default()
1679                        },
1680                    );
1681                }
1682            })
1683            .await
1684            .expect("seed policy");
1685
1686        let mut payload = consumer_payload(&consumer_id, &run_id);
1687        assemble_step_pointers(&state, &mut payload).await;
1688
1689        let steps = &payload.context.expect("context").steps;
1690        assert_eq!(steps.len(), 1, "steps: {steps:?}");
1691        assert_eq!(
1692            steps[0].name, "plan-out",
1693            "StepPointer.name must be the canonical name"
1694        );
1695    }
1696
1697    // ──────────────────────────────────────────────────────────────────────
1698    // GH #31 — `/v1/worker/prompt/system` + `/v1/agents/:name/render-size`
1699    // ──────────────────────────────────────────────────────────────────────
1700
1701    /// Seeds a task + baked system prompt + a short worker handle bound to
1702    /// it, mirroring the shape `Engine::dispatch_attempt` would have
1703    /// produced (minus the parts these two routes don't touch: no real
1704    /// HMAC-signed `CapToken`, since `task_id_from_handle`'s handle → fp →
1705    /// task_id chain is what's under test, not signature verification).
1706    async fn seed_task_with_handle(
1707        state: &AppState,
1708        task_id: &StepId,
1709        agent: &str,
1710        attempt: u32,
1711        system: Option<String>,
1712    ) -> String {
1713        let handle = format!("wh-{}", mlua_swarm::types::secure_hex(4));
1714        let task_id = task_id.clone();
1715        let agent = agent.to_string();
1716        let handle_clone = handle.clone();
1717        state
1718            .engine
1719            .with_state("test.seed_task_with_handle", move |s| {
1720                let mut task = mlua_swarm::core::state::TaskState::new(
1721                    task_id.clone(),
1722                    mlua_swarm::core::state::TaskSpec {
1723                        agent: agent.clone(),
1724                        initial_directive: json!("x"),
1725                        step_ctx: None,
1726                        check_policy: None,
1727                    },
1728                );
1729                task.attempt = attempt;
1730                s.tasks.insert(task_id.clone(), task);
1731                s.systems.insert((task_id.clone(), attempt), system);
1732                let token = CapToken {
1733                    agent_id: agent,
1734                    role: mlua_swarm::Role::Worker,
1735                    scopes: vec!["*".to_string()],
1736                    issued_at: 0,
1737                    expire_at: u64::MAX,
1738                    max_uses: None,
1739                    nonce: format!("test-nonce-{task_id}"),
1740                    sig_hex: String::new(),
1741                };
1742                let fp = token.fingerprint();
1743                s.tokens.insert(
1744                    fp.clone(),
1745                    mlua_swarm::core::state::CapTokenRecord {
1746                        token,
1747                        uses_left: None,
1748                        revoked: false,
1749                        task_id: Some(task_id),
1750                    },
1751                );
1752                s.worker_handles.insert(handle_clone, fp);
1753            })
1754            .await
1755            .expect("seed_task_with_handle");
1756        handle
1757    }
1758
1759    fn bearer_headers(handle: &str) -> HeaderMap {
1760        let mut headers = HeaderMap::new();
1761        headers.insert(
1762            AUTHORIZATION,
1763            format!("Bearer {handle}").parse().expect("header value"),
1764        );
1765        headers
1766    }
1767
1768    /// `GET /v1/worker/prompt/system` returns the exact raw baked bytes
1769    /// (not JSON-wrapped) with `Content-Type: text/plain`, for the
1770    /// `(task_id, attempt)` the handle is bound to.
1771    #[tokio::test]
1772    async fn worker_prompt_system_returns_raw_bytes_for_baked_system() {
1773        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
1774        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
1775        let state = test_state(data_store, run_store);
1776        let task_id = StepId::new();
1777        let rendered = "# Hello\n\nThis is the baked system prompt.".to_string();
1778        let handle =
1779            seed_task_with_handle(&state, &task_id, "planner", 1, Some(rendered.clone())).await;
1780
1781        let resp = worker_prompt_system(
1782            State(state.clone()),
1783            bearer_headers(&handle),
1784            Query(PromptSystemQuery {
1785                task_id: task_id.clone(),
1786                attempt: 1,
1787            }),
1788        )
1789        .await
1790        .expect("worker_prompt_system")
1791        .into_response();
1792
1793        assert_eq!(resp.status(), StatusCode::OK);
1794        let content_type = resp
1795            .headers()
1796            .get(header::CONTENT_TYPE)
1797            .expect("content-type header")
1798            .to_str()
1799            .expect("ascii");
1800        assert_eq!(content_type, "text/plain; charset=utf-8");
1801        let body_bytes = axum::body::to_bytes(resp.into_body(), usize::MAX)
1802            .await
1803            .expect("body bytes");
1804        assert_eq!(body_bytes.as_ref(), rendered.as_bytes());
1805    }
1806
1807    /// No baked system for the given `(task_id, attempt)` → 404, not a
1808    /// panic or a 200-with-empty-body.
1809    #[tokio::test]
1810    async fn worker_prompt_system_404s_when_no_baked_system() {
1811        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
1812        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
1813        let state = test_state(data_store, run_store);
1814        let task_id = StepId::new();
1815        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
1816
1817        let result = worker_prompt_system(
1818            State(state.clone()),
1819            bearer_headers(&handle),
1820            Query(PromptSystemQuery {
1821                task_id: task_id.clone(),
1822                attempt: 1,
1823            }),
1824        )
1825        .await;
1826        let err = match result {
1827            Ok(_) => panic!("expected 404 ApiError, got Ok"),
1828            Err(e) => e,
1829        };
1830        assert_eq!(err.into_response().status(), StatusCode::NOT_FOUND);
1831    }
1832
1833    /// A handle bound to a different task than the one requested must be
1834    /// rejected (400) — this is the same cross-check `worker_prompt` does.
1835    #[tokio::test]
1836    async fn worker_prompt_system_rejects_handle_task_mismatch() {
1837        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
1838        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
1839        let state = test_state(data_store, run_store);
1840        let task_id = StepId::new();
1841        let other_task_id = StepId::new();
1842        let handle =
1843            seed_task_with_handle(&state, &task_id, "planner", 1, Some("x".to_string())).await;
1844
1845        let result = worker_prompt_system(
1846            State(state.clone()),
1847            bearer_headers(&handle),
1848            Query(PromptSystemQuery {
1849                task_id: other_task_id,
1850                attempt: 1,
1851            }),
1852        )
1853        .await;
1854        let err = match result {
1855            Ok(_) => panic!("expected 400 ApiError for task mismatch, got Ok"),
1856            Err(e) => e,
1857        };
1858        assert_eq!(err.into_response().status(), StatusCode::BAD_REQUEST);
1859    }
1860
1861    /// `GET /v1/agents/:name/render-size` requires no auth, and reports
1862    /// `last_rendered_bytes: null` for an agent that has never had a
1863    /// `system_prompt` baked — a normal 200, not a 404.
1864    #[tokio::test]
1865    async fn agent_render_size_returns_null_for_unknown_agent() {
1866        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
1867        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
1868        let state = test_state(data_store, run_store);
1869
1870        let Json(body) = agent_render_size(
1871            State(state.clone()),
1872            axum::extract::Path("never-dispatched".to_string()),
1873        )
1874        .await;
1875        assert_eq!(body.agent, "never-dispatched");
1876        assert_eq!(body.last_rendered_bytes, None);
1877    }
1878
1879    /// Once `bake_worker_system_prompt` has recorded a render size for an
1880    /// agent, the route reports the most-recently-observed value.
1881    #[tokio::test]
1882    async fn agent_render_size_reports_last_rendered_bytes() {
1883        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
1884        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
1885        let state = test_state(data_store, run_store);
1886        let task_id = StepId::new();
1887        state
1888            .engine
1889            .with_state("test.seed_agent_ctx_for_bake", {
1890                let task_id = task_id.clone();
1891                move |s| {
1892                    s.tasks.insert(
1893                        task_id.clone(),
1894                        mlua_swarm::core::state::TaskState::new(
1895                            task_id,
1896                            mlua_swarm::core::state::TaskSpec {
1897                                agent: "coder".to_string(),
1898                                initial_directive: json!("x"),
1899                                step_ctx: None,
1900                                check_policy: None,
1901                            },
1902                        ),
1903                    );
1904                }
1905            })
1906            .await
1907            .expect("seed task");
1908        state
1909            .engine
1910            .bake_worker_system_prompt(&task_id, 1, Some("z".repeat(42)))
1911            .await
1912            .expect("bake_worker_system_prompt");
1913
1914        let Json(body) = agent_render_size(
1915            State(state.clone()),
1916            axum::extract::Path("coder".to_string()),
1917        )
1918        .await;
1919        assert_eq!(body.agent, "coder");
1920        assert_eq!(body.last_rendered_bytes, Some(42));
1921    }
1922
1923    // ──────────────────────────────────────────────────────────────────────
1924    // GH #36 ST1 — `POST /v1/worker/artifact`
1925    // ──────────────────────────────────────────────────────────────────────
1926
1927    /// A valid `?name=` + short-handle Bearer stages the raw body (trailing
1928    /// whitespace trimmed, same as `worker_submit`) as an `Artifact` on the
1929    /// task's current-attempt tail, and returns `204 No Content`.
1930    #[tokio::test]
1931    async fn worker_artifact_stages_and_204s_for_valid_request() {
1932        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
1933        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
1934        let state = test_state(data_store, run_store);
1935        let task_id = StepId::new();
1936        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
1937
1938        let status = worker_artifact(
1939            State(state.clone()),
1940            bearer_headers(&handle),
1941            Query(ArtifactQuery {
1942                name: "summary".to_string(),
1943            }),
1944            axum::body::Bytes::from_static(b"hello artifact\n"),
1945        )
1946        .await
1947        .expect("worker_artifact");
1948        assert_eq!(status, StatusCode::NO_CONTENT);
1949
1950        let tail = state.engine.output_tail(&task_id, 1).await;
1951        assert_eq!(tail.len(), 1, "tail: {tail:?}");
1952        match &tail[0] {
1953            OutputEvent::Artifact { name, content } => {
1954                assert_eq!(name, "summary");
1955                match content {
1956                    ContentRef::Inline { value } => {
1957                        assert_eq!(value, &json!("hello artifact"));
1958                    }
1959                    other => panic!("expected Inline content, got {other:?}"),
1960                }
1961            }
1962            other => panic!("expected Artifact event, got {other:?}"),
1963        }
1964    }
1965
1966    /// `?name=` missing entirely → axum's `Query` extractor rejection
1967    /// (400), not a panic. `Query<ArtifactQuery>` is constructed directly
1968    /// in this test (mirroring the other handlers' unit style, which call
1969    /// the handler fn with an already-extracted `Query`) — an empty `name`
1970    /// is exercised separately below since that case is NOT caught by the
1971    /// extractor and must be checked in the handler body.
1972    #[tokio::test]
1973    async fn worker_artifact_rejects_blank_name() {
1974        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
1975        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
1976        let state = test_state(data_store, run_store);
1977        let task_id = StepId::new();
1978        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
1979
1980        let result = worker_artifact(
1981            State(state.clone()),
1982            bearer_headers(&handle),
1983            Query(ArtifactQuery {
1984                name: "   ".to_string(),
1985            }),
1986            axum::body::Bytes::from_static(b"x"),
1987        )
1988        .await;
1989        let err = match result {
1990            Ok(_) => panic!("expected 400 ApiError for blank name, got Ok"),
1991            Err(e) => e,
1992        };
1993        assert_eq!(err.into_response().status(), StatusCode::BAD_REQUEST);
1994
1995        // Nothing was staged.
1996        assert!(state.engine.output_tail(&task_id, 1).await.is_empty());
1997    }
1998
1999    /// Staging the same `name` twice within one attempt is last-write-wins
2000    /// on the folded value (`fold_final_and_parts` in `mlua_swarm::core::
2001    /// engine`) — this test only asserts the raw tail carries both events
2002    /// in order (the fold itself is covered by that crate's own unit
2003    /// tests); `Engine::stage_worker_artifact_trusted`'s doc.
2004    #[tokio::test]
2005    async fn worker_artifact_staging_same_name_twice_appends_both_events_in_order() {
2006        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2007        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2008        let state = test_state(data_store, run_store);
2009        let task_id = StepId::new();
2010        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
2011
2012        for body in [b"first".as_slice(), b"second".as_slice()] {
2013            worker_artifact(
2014                State(state.clone()),
2015                bearer_headers(&handle),
2016                Query(ArtifactQuery {
2017                    name: "a".to_string(),
2018                }),
2019                axum::body::Bytes::copy_from_slice(body),
2020            )
2021            .await
2022            .expect("worker_artifact");
2023        }
2024
2025        let tail = state.engine.output_tail(&task_id, 1).await;
2026        assert_eq!(tail.len(), 2, "tail: {tail:?}");
2027        let values: Vec<&str> = tail
2028            .iter()
2029            .map(|ev| match ev {
2030                OutputEvent::Artifact {
2031                    content: ContentRef::Inline { value },
2032                    ..
2033                } => value.as_str().expect("string value"),
2034                other => panic!("expected Artifact/Inline event, got {other:?}"),
2035            })
2036            .collect();
2037        assert_eq!(values, vec!["first", "second"]);
2038    }
2039
2040    // ──────────────────────────────────────────────────────────────────
2041    // GH #37 — terminal-run guard (`reject_if_run_terminal`)
2042    // ──────────────────────────────────────────────────────────────────
2043
2044    /// Links a seeded dispatch task to a Run the same way
2045    /// `AgentContextMiddleware` does at spawn time: an `agent_ctx` entry
2046    /// whose view carries the `run_id`.
2047    async fn link_task_to_run(state: &AppState, task_id: &StepId, attempt: u32, run_id: &RunId) {
2048        let tid = task_id.clone();
2049        let rid_str = run_id.to_string();
2050        state
2051            .engine
2052            .with_state("test.link_task_to_run", move |s| {
2053                let mut entry = mlua_swarm::core::state::AgentCtxEntry::default();
2054                entry.view.run_id = Some(rid_str);
2055                s.agent_ctx.insert((tid, attempt), entry);
2056            })
2057            .await
2058            .expect("link_task_to_run");
2059    }
2060
2061    /// GH #37: a submit / artifact addressed at a Run that already
2062    /// reached a terminal status must be rejected with `410 Gone` — the
2063    /// flow-eval driver for that Run is gone, so a silent `204` here
2064    /// would orphan the worker's output.
2065    #[tokio::test]
2066    async fn submit_and_artifact_against_terminal_run_return_410() {
2067        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2068        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2069        let state = test_state(data_store, run_store.clone());
2070        let task_id = StepId::new();
2071        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
2072
2073        let owner_task = TaskId::new();
2074        let run_id = RunId::new();
2075        let mut rec = run_record(&owner_task, &run_id, vec![]);
2076        rec.status = RunStatus::Failed;
2077        run_store.create(rec).await.expect("run create");
2078        link_task_to_run(&state, &task_id, 1, &run_id).await;
2079
2080        let err = worker_submit(
2081            State(state.clone()),
2082            bearer_headers(&handle),
2083            Query(SubmitQuery { ok: None }),
2084            axum::body::Bytes::from_static(b"LATE OUTPUT"),
2085        )
2086        .await
2087        .expect_err("a submit against a Failed run must be rejected");
2088        assert_eq!(err.status, StatusCode::GONE);
2089        assert!(
2090            err.message.contains(&run_id.to_string()),
2091            "the 410 must name the terminal run: {}",
2092            err.message
2093        );
2094
2095        let err = worker_artifact(
2096            State(state.clone()),
2097            bearer_headers(&handle),
2098            Query(ArtifactQuery {
2099                name: "part.md".to_string(),
2100            }),
2101            axum::body::Bytes::from_static(b"LATE PART"),
2102        )
2103        .await
2104        .expect_err("an artifact staged against a Failed run must be rejected");
2105        assert_eq!(err.status, StatusCode::GONE);
2106
2107        // The rejected values must not have reached the output tail.
2108        let tail = state.engine.output_tail(&task_id, 1).await;
2109        assert!(tail.is_empty(), "rejected submits must not land: {tail:?}");
2110    }
2111
2112    /// GH #37 fail-open contract: the guard must never turn a
2113    /// would-have-succeeded submit into a failure — no run linkage at
2114    /// all, an unknown Run, and a live (`Running`) Run all pass.
2115    #[tokio::test]
2116    async fn terminal_run_guard_is_fail_open() {
2117        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2118        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2119        let state = test_state(data_store, run_store.clone());
2120        let task_id = StepId::new();
2121        seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
2122
2123        // (a) No agent-ctx linkage at all (pre-run-tracking dispatch).
2124        reject_if_run_terminal(&state, &task_id, 1)
2125            .await
2126            .expect("no linkage must fail open");
2127
2128        // (b) Linked to a Run the store does not know.
2129        let unknown_run = RunId::new();
2130        link_task_to_run(&state, &task_id, 1, &unknown_run).await;
2131        reject_if_run_terminal(&state, &task_id, 1)
2132            .await
2133            .expect("unknown run must fail open");
2134
2135        // (c) Linked to a live Run.
2136        let owner_task = TaskId::new();
2137        let live_run = RunId::new();
2138        run_store
2139            .create(run_record(&owner_task, &live_run, vec![]))
2140            .await
2141            .expect("run create");
2142        link_task_to_run(&state, &task_id, 1, &live_run).await;
2143        reject_if_run_terminal(&state, &task_id, 1)
2144            .await
2145            .expect("a Running run must pass the guard");
2146    }
2147
2148    // ──────────────────────────────────────────────────────────────────
2149    // GH #32 — `POST /v1/worker/degradation`
2150    // ──────────────────────────────────────────────────────────────────
2151
2152    fn degradation_body(tool: &str, note: Option<&str>) -> DegradationBody {
2153        DegradationBody {
2154            tool: tool.to_string(),
2155            error: "boom".to_string(),
2156            fallback: "used cached value".to_string(),
2157            note: note.map(str::to_string),
2158        }
2159    }
2160
2161    /// [`link_task_to_run`] plus the `view.agent` name — production's
2162    /// `AgentContextMiddleware` sets both fields on the same `agent_ctx`
2163    /// entry; the shared GH #37 helper only needed `run_id`, so this
2164    /// sibling fills in `agent` too for tests that assert on the
2165    /// server-injected `step_ref`.
2166    async fn link_task_to_run_with_agent(
2167        state: &AppState,
2168        task_id: &StepId,
2169        attempt: u32,
2170        run_id: &RunId,
2171        agent: &str,
2172    ) {
2173        let tid = task_id.clone();
2174        let rid_str = run_id.to_string();
2175        let agent = agent.to_string();
2176        state
2177            .engine
2178            .with_state("test.link_task_to_run_with_agent", move |s| {
2179                let mut entry = mlua_swarm::core::state::AgentCtxEntry::default();
2180                entry.view.run_id = Some(rid_str);
2181                entry.view.agent = agent;
2182                s.agent_ctx.insert((tid, attempt), entry);
2183            })
2184            .await
2185            .expect("link_task_to_run_with_agent");
2186    }
2187
2188    /// A worker-reported degradation is persisted to the linked Run's
2189    /// `degradations` with the server-injected `step_ref` / `attempt` /
2190    /// `at` fields filled in — the client body never supplies any of the
2191    /// three.
2192    #[tokio::test]
2193    async fn worker_degradation_persists_entry_when_run_tracked() {
2194        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2195        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2196        let state = test_state(data_store, run_store.clone());
2197        let task_id = StepId::new();
2198        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
2199
2200        let owner_task = TaskId::new();
2201        let run_id = RunId::new();
2202        run_store
2203            .create(run_record(&owner_task, &run_id, vec![]))
2204            .await
2205            .expect("run create");
2206        link_task_to_run_with_agent(&state, &task_id, 1, &run_id, "planner").await;
2207
2208        let status = worker_degradation(
2209            State(state.clone()),
2210            bearer_headers(&handle),
2211            Json(degradation_body("web_search", Some("rate limited"))),
2212        )
2213        .await
2214        .expect("worker_degradation");
2215        assert_eq!(status, StatusCode::NO_CONTENT);
2216
2217        let rec = run_store.get(&run_id).await.expect("run get");
2218        assert_eq!(
2219            rec.degradations.len(),
2220            1,
2221            "degradations: {:?}",
2222            rec.degradations
2223        );
2224        let entry = &rec.degradations[0];
2225        assert_eq!(entry.tool, "web_search");
2226        assert_eq!(entry.error, "boom");
2227        assert_eq!(entry.fallback, "used cached value");
2228        assert_eq!(entry.note.as_deref(), Some("rate limited"));
2229        assert_eq!(entry.step_ref.as_deref(), Some("planner"));
2230        assert_eq!(entry.attempt, Some(1));
2231        assert!(entry.at > 0, "at must be a real timestamp: {}", entry.at);
2232    }
2233
2234    /// Two entries POSTed in sequence are appended in order.
2235    #[tokio::test]
2236    async fn worker_degradation_appends_in_order() {
2237        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2238        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2239        let state = test_state(data_store, run_store.clone());
2240        let task_id = StepId::new();
2241        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
2242
2243        let owner_task = TaskId::new();
2244        let run_id = RunId::new();
2245        run_store
2246            .create(run_record(&owner_task, &run_id, vec![]))
2247            .await
2248            .expect("run create");
2249        link_task_to_run(&state, &task_id, 1, &run_id).await;
2250
2251        for tool in ["first_tool", "second_tool"] {
2252            worker_degradation(
2253                State(state.clone()),
2254                bearer_headers(&handle),
2255                Json(degradation_body(tool, None)),
2256            )
2257            .await
2258            .expect("worker_degradation");
2259        }
2260
2261        let rec = run_store.get(&run_id).await.expect("run get");
2262        let tools: Vec<&str> = rec.degradations.iter().map(|e| e.tool.as_str()).collect();
2263        assert_eq!(tools, vec!["first_tool", "second_tool"]);
2264    }
2265
2266    /// A task whose `agent_ctx` carries no Run linkage (pre-run-tracking
2267    /// dispatch) silently 204s — nothing to append to, and this must not
2268    /// surface as a client error.
2269    #[tokio::test]
2270    async fn worker_degradation_silent_ok_when_no_run_tracked() {
2271        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2272        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2273        let state = test_state(data_store, run_store);
2274        let task_id = StepId::new();
2275        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
2276
2277        let status = worker_degradation(
2278            State(state.clone()),
2279            bearer_headers(&handle),
2280            Json(degradation_body("some_tool", None)),
2281        )
2282        .await
2283        .expect("worker_degradation must not error on missing run linkage");
2284        assert_eq!(status, StatusCode::NO_CONTENT);
2285    }
2286
2287    /// GH #37 terminal-run guard applies to the degradation channel too — a
2288    /// dead Run must not accumulate signals.
2289    #[tokio::test]
2290    async fn worker_degradation_rejects_terminal_run() {
2291        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2292        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2293        let state = test_state(data_store, run_store.clone());
2294        let task_id = StepId::new();
2295        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
2296
2297        let owner_task = TaskId::new();
2298        let run_id = RunId::new();
2299        let mut rec = run_record(&owner_task, &run_id, vec![]);
2300        rec.status = RunStatus::Done;
2301        run_store.create(rec).await.expect("run create");
2302        link_task_to_run(&state, &task_id, 1, &run_id).await;
2303
2304        let err = worker_degradation(
2305            State(state.clone()),
2306            bearer_headers(&handle),
2307            Json(degradation_body("some_tool", None)),
2308        )
2309        .await
2310        .expect_err("a degradation against a Done run must be rejected");
2311        assert_eq!(err.status, StatusCode::GONE);
2312
2313        let rec = run_store.get(&run_id).await.expect("run get");
2314        assert!(
2315            rec.degradations.is_empty(),
2316            "rejected degradation must not land: {:?}",
2317            rec.degradations
2318        );
2319    }
2320
2321    // ──────────────────────────────────────────────────────────────────
2322    // GH #42 — `@file:<abs-path>` sentinel resolution in `worker_submit`
2323    // / `worker_artifact`. Guards each verified independently: sentinel
2324    // resolves to the file's trimmed contents; path outside `work_dir`,
2325    // missing file, oversized file, and non-sentinel bodies each get the
2326    // documented behavior.
2327    // ──────────────────────────────────────────────────────────────────
2328
2329    /// Seeds an `agent_ctx` entry whose view carries `work_dir` and, when
2330    /// `allow_file_submit` is `Some`, that value under the GH #43
2331    /// [`FILE_SENTINEL_ALLOW_KEY`] in `view.extra` — matching the shape
2332    /// `AgentContextMiddleware` writes at spawn time. Sentinel resolution
2333    /// requires both the `work_dir` and the strict `Bool(true)` opt-in.
2334    async fn seed_work_dir(
2335        state: &AppState,
2336        task_id: &StepId,
2337        attempt: u32,
2338        work_dir: &str,
2339        allow_file_submit: Option<Value>,
2340    ) {
2341        let tid = task_id.clone();
2342        let work_dir = work_dir.to_string();
2343        state
2344            .engine
2345            .with_state("test.seed_work_dir", move |s| {
2346                let mut entry = mlua_swarm::core::state::AgentCtxEntry::default();
2347                entry.view.work_dir = Some(work_dir);
2348                if let Some(v) = allow_file_submit {
2349                    entry
2350                        .view
2351                        .extra
2352                        .insert(FILE_SENTINEL_ALLOW_KEY.to_string(), v);
2353                }
2354                s.agent_ctx.insert((tid, attempt), entry);
2355            })
2356            .await
2357            .expect("seed_work_dir");
2358    }
2359
2360    /// Sentinel body `@file:<abs-path>` resolves to the file's trimmed
2361    /// contents and reaches the `OutputStore` via the normal Final-append
2362    /// path — same 204 the inline path returns.
2363    #[tokio::test]
2364    async fn worker_submit_resolves_file_sentinel_under_work_dir() {
2365        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2366        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2367        let state = test_state(data_store.clone(), run_store);
2368        let task_id = StepId::new();
2369        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
2370
2371        let tmp = tempfile::tempdir().expect("tempdir");
2372        let work_dir = tmp.path().to_path_buf();
2373        seed_work_dir(
2374            &state,
2375            &task_id,
2376            1,
2377            work_dir.to_str().expect("work_dir utf-8"),
2378            Some(Value::Bool(true)),
2379        )
2380        .await;
2381
2382        let payload_path = work_dir.join("scout.md");
2383        let payload = "## Context Package (broad)\n\nlarge body content\n";
2384        tokio::fs::write(&payload_path, payload)
2385            .await
2386            .expect("write payload");
2387        let body = format!(
2388            "@file:{}",
2389            payload_path.to_str().expect("payload path utf-8")
2390        );
2391
2392        let status = worker_submit(
2393            State(state.clone()),
2394            bearer_headers(&handle),
2395            Query(SubmitQuery { ok: None }),
2396            axum::body::Bytes::from(body),
2397        )
2398        .await
2399        .expect("worker_submit sentinel");
2400        assert_eq!(status, StatusCode::NO_CONTENT);
2401
2402        // Final event lands with the file's trimmed contents on
2403        // `EngineState.output_store` (the in-memory tail
2404        // `submit_worker_result_trusted` writes to).
2405        let tid = task_id.clone();
2406        let value = state
2407            .engine
2408            .with_state("test.inspect_output_store", move |s| {
2409                s.output_store.get(&(tid.clone(), 1)).and_then(|evs| {
2410                    evs.iter().find_map(|ev| match ev {
2411                        OutputEvent::Final {
2412                            content: ContentRef::Inline { value },
2413                            ..
2414                        } => Some(value.clone()),
2415                        _ => None,
2416                    })
2417                })
2418            })
2419            .await
2420            .expect("with_state")
2421            .expect("Final event present");
2422        assert_eq!(value, Value::String(payload.trim_end().to_string()));
2423    }
2424
2425    /// A non-sentinel body is passed through byte-for-byte (pre-#42
2426    /// callers see zero behavior change).
2427    #[tokio::test]
2428    async fn worker_submit_passes_non_sentinel_body_unchanged() {
2429        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2430        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2431        let state = test_state(data_store.clone(), run_store);
2432        let task_id = StepId::new();
2433        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
2434        // No agent_ctx / work_dir seeded — the inline path must not
2435        // require one.
2436
2437        let status = worker_submit(
2438            State(state.clone()),
2439            bearer_headers(&handle),
2440            Query(SubmitQuery { ok: None }),
2441            axum::body::Bytes::from_static(b"DONE yes=1 maybe=0 no=0"),
2442        )
2443        .await
2444        .expect("worker_submit inline");
2445        assert_eq!(status, StatusCode::NO_CONTENT);
2446
2447        let tid = task_id.clone();
2448        let value = state
2449            .engine
2450            .with_state("test.inspect_output_store", move |s| {
2451                s.output_store.get(&(tid.clone(), 1)).and_then(|evs| {
2452                    evs.iter().find_map(|ev| match ev {
2453                        OutputEvent::Final {
2454                            content: ContentRef::Inline { value },
2455                            ..
2456                        } => Some(value.clone()),
2457                        _ => None,
2458                    })
2459                })
2460            })
2461            .await
2462            .expect("with_state")
2463            .expect("Final event present");
2464        assert_eq!(value, Value::String("DONE yes=1 maybe=0 no=0".to_string()));
2465    }
2466
2467    /// Sentinel with a path outside the task's `work_dir` (`..`-escape
2468    /// via a sibling tempdir) → `400`. `canonicalize` collapses the
2469    /// `..`, so a symlink pointing outside the allowlist would be caught
2470    /// by the same check.
2471    #[tokio::test]
2472    async fn worker_submit_rejects_sentinel_path_outside_work_dir() {
2473        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2474        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2475        let state = test_state(data_store, run_store);
2476        let task_id = StepId::new();
2477        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
2478
2479        let allowed = tempfile::tempdir().expect("allowed tempdir");
2480        let outside = tempfile::tempdir().expect("outside tempdir");
2481        seed_work_dir(
2482            &state,
2483            &task_id,
2484            1,
2485            allowed.path().to_str().expect("utf-8"),
2486            Some(Value::Bool(true)),
2487        )
2488        .await;
2489
2490        let outside_file = outside.path().join("leak.md");
2491        tokio::fs::write(&outside_file, b"outside content")
2492            .await
2493            .expect("write outside");
2494        let body = format!(
2495            "@file:{}",
2496            outside_file.to_str().expect("outside path utf-8")
2497        );
2498
2499        let err = worker_submit(
2500            State(state.clone()),
2501            bearer_headers(&handle),
2502            Query(SubmitQuery { ok: None }),
2503            axum::body::Bytes::from(body),
2504        )
2505        .await
2506        .expect_err("outside-work_dir sentinel must be rejected");
2507        assert_eq!(err.status, StatusCode::BAD_REQUEST);
2508    }
2509
2510    /// Sentinel pointing at a non-existent file → `404`.
2511    #[tokio::test]
2512    async fn worker_submit_rejects_sentinel_missing_file() {
2513        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2514        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2515        let state = test_state(data_store, run_store);
2516        let task_id = StepId::new();
2517        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
2518
2519        let tmp = tempfile::tempdir().expect("tempdir");
2520        seed_work_dir(
2521            &state,
2522            &task_id,
2523            1,
2524            tmp.path().to_str().expect("utf-8"),
2525            Some(Value::Bool(true)),
2526        )
2527        .await;
2528        let missing = tmp.path().join("does-not-exist.md");
2529        let body = format!("@file:{}", missing.to_str().expect("utf-8"));
2530
2531        let err = worker_submit(
2532            State(state.clone()),
2533            bearer_headers(&handle),
2534            Query(SubmitQuery { ok: None }),
2535            axum::body::Bytes::from(body),
2536        )
2537        .await
2538        .expect_err("missing-file sentinel must be rejected");
2539        assert_eq!(err.status, StatusCode::NOT_FOUND);
2540    }
2541
2542    /// Sentinel body with a relative path → `400` before any FS lookup.
2543    #[tokio::test]
2544    async fn worker_submit_rejects_sentinel_relative_path() {
2545        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2546        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2547        let state = test_state(data_store, run_store);
2548        let task_id = StepId::new();
2549        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
2550
2551        let err = worker_submit(
2552            State(state.clone()),
2553            bearer_headers(&handle),
2554            Query(SubmitQuery { ok: None }),
2555            axum::body::Bytes::from_static(b"@file:relative/path.md"),
2556        )
2557        .await
2558        .expect_err("relative-path sentinel must be rejected");
2559        assert_eq!(err.status, StatusCode::BAD_REQUEST);
2560    }
2561
2562    /// Sentinel body when the task has no `AgentContextView` (spawn
2563    /// didn't run through `AgentContextMiddleware`) → `400`. This is the
2564    /// documented pre-condition for sentinel use.
2565    #[tokio::test]
2566    async fn worker_submit_rejects_sentinel_without_agent_context_view() {
2567        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2568        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2569        let state = test_state(data_store, run_store);
2570        let task_id = StepId::new();
2571        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
2572        // No seed_work_dir — the agent_ctx map has no entry for this task.
2573
2574        let err = worker_submit(
2575            State(state.clone()),
2576            bearer_headers(&handle),
2577            Query(SubmitQuery { ok: None }),
2578            axum::body::Bytes::from_static(b"@file:/tmp/anywhere.md"),
2579        )
2580        .await
2581        .expect_err("missing AgentContextView must reject sentinel");
2582        assert_eq!(err.status, StatusCode::BAD_REQUEST);
2583    }
2584
2585    /// The same sentinel form works on `POST /v1/worker/artifact` — the
2586    /// artifact endpoint shares the resolver with `worker_submit`, so the
2587    /// resolved file contents land under the artifact's `name` key.
2588    #[tokio::test]
2589    async fn worker_artifact_resolves_file_sentinel_under_work_dir() {
2590        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2591        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2592        let state = test_state(data_store, run_store);
2593        let task_id = StepId::new();
2594        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
2595
2596        let tmp = tempfile::tempdir().expect("tempdir");
2597        seed_work_dir(
2598            &state,
2599            &task_id,
2600            1,
2601            tmp.path().to_str().expect("utf-8"),
2602            Some(Value::Bool(true)),
2603        )
2604        .await;
2605
2606        let payload_path = tmp.path().join("part.md");
2607        let payload = "artifact part body\n";
2608        tokio::fs::write(&payload_path, payload)
2609            .await
2610            .expect("write payload");
2611        let body = format!("@file:{}", payload_path.to_str().expect("utf-8"));
2612
2613        let status = worker_artifact(
2614            State(state.clone()),
2615            bearer_headers(&handle),
2616            Query(ArtifactQuery {
2617                name: "scout".to_string(),
2618            }),
2619            axum::body::Bytes::from(body),
2620        )
2621        .await
2622        .expect("worker_artifact sentinel");
2623        assert_eq!(status, StatusCode::NO_CONTENT);
2624    }
2625
2626    /// GH #43 — sentinel with `work_dir` seeded but no
2627    /// `allow_file_submit` opt-in → `400` (default-deny). The file exists
2628    /// and sits under `work_dir`, so the rejection is attributable to the
2629    /// missing opt-in alone.
2630    #[tokio::test]
2631    async fn worker_submit_rejects_sentinel_without_allow_flag() {
2632        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2633        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2634        let state = test_state(data_store, run_store);
2635        let task_id = StepId::new();
2636        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
2637
2638        let tmp = tempfile::tempdir().expect("tempdir");
2639        seed_work_dir(
2640            &state,
2641            &task_id,
2642            1,
2643            tmp.path().to_str().expect("utf-8"),
2644            None,
2645        )
2646        .await;
2647
2648        let payload_path = tmp.path().join("out.md");
2649        tokio::fs::write(&payload_path, b"resolvable body")
2650            .await
2651            .expect("write payload");
2652        let body = format!("@file:{}", payload_path.to_str().expect("utf-8"));
2653
2654        let err = worker_submit(
2655            State(state.clone()),
2656            bearer_headers(&handle),
2657            Query(SubmitQuery { ok: None }),
2658            axum::body::Bytes::from(body),
2659        )
2660        .await
2661        .expect_err("missing opt-in must reject sentinel");
2662        assert_eq!(err.status, StatusCode::BAD_REQUEST);
2663        assert!(
2664            err.message.contains("not allowed"),
2665            "rejection must name the opt-in guard, got: {}",
2666            err.message
2667        );
2668    }
2669
2670    /// GH #43 — the opt-in is the strict boolean `true`: `Bool(false)`
2671    /// and the string `"true"` are both rejected with `400`.
2672    #[tokio::test]
2673    async fn worker_submit_rejects_sentinel_with_non_true_allow_values() {
2674        for allow in [Value::Bool(false), Value::String("true".to_string())] {
2675            let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2676            let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2677            let state = test_state(data_store, run_store);
2678            let task_id = StepId::new();
2679            let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
2680
2681            let tmp = tempfile::tempdir().expect("tempdir");
2682            seed_work_dir(
2683                &state,
2684                &task_id,
2685                1,
2686                tmp.path().to_str().expect("utf-8"),
2687                Some(allow.clone()),
2688            )
2689            .await;
2690
2691            let payload_path = tmp.path().join("out.md");
2692            tokio::fs::write(&payload_path, b"resolvable body")
2693                .await
2694                .expect("write payload");
2695            let body = format!("@file:{}", payload_path.to_str().expect("utf-8"));
2696
2697            let err = worker_submit(
2698                State(state.clone()),
2699                bearer_headers(&handle),
2700                Query(SubmitQuery { ok: None }),
2701                axum::body::Bytes::from(body),
2702            )
2703            .await
2704            .expect_err("non-true opt-in value must reject sentinel");
2705            assert_eq!(err.status, StatusCode::BAD_REQUEST, "value: {allow:?}");
2706        }
2707    }
2708
2709    // ──────────────────────────────────────────────────────────────────
2710    // GH #50 (Subtask 2) — submit-time verdict contract gate, handler-
2711    // level unit coverage. The full process-boundary HTTP round trip
2712    // (Acceptance Criterion #7) lives in
2713    // `crates/mlua-swarm-server/tests/verdict_contract.rs`; these are the
2714    // fast in-process counterpart exercising `worker_submit` /
2715    // `worker_artifact` directly, same convention as the sentinel tests
2716    // above.
2717    // ──────────────────────────────────────────────────────────────────
2718
2719    fn body_verdict_contract(values: &[&str]) -> mlua_swarm_schema::VerdictContract {
2720        mlua_swarm_schema::VerdictContract {
2721            channel: VerdictChannel::Body,
2722            values: values.iter().map(|v| v.to_string()).collect(),
2723        }
2724    }
2725
2726    fn part_verdict_contract(values: &[&str]) -> mlua_swarm_schema::VerdictContract {
2727        mlua_swarm_schema::VerdictContract {
2728            channel: VerdictChannel::Part,
2729            values: values.iter().map(|v| v.to_string()).collect(),
2730        }
2731    }
2732
2733    /// A `channel: "body"` contract rejects a `worker_submit` body outside
2734    /// its declared `values` with `422`, echoing the expected token set.
2735    #[tokio::test]
2736    async fn worker_submit_rejects_body_outside_contract_values_with_422() {
2737        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2738        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2739        let state = test_state(data_store, run_store);
2740        let task_id = StepId::new();
2741        let handle = seed_task_with_handle(&state, &task_id, "gate", 1, None).await;
2742        state.engine.register_verdict_contracts(HashMap::from([(
2743            "gate".to_string(),
2744            body_verdict_contract(&["PASS", "BLOCKED"]),
2745        )]));
2746
2747        let err = worker_submit(
2748            State(state.clone()),
2749            bearer_headers(&handle),
2750            Query(SubmitQuery { ok: None }),
2751            axum::body::Bytes::from("UNKNOWN"),
2752        )
2753        .await
2754        .expect_err("value outside declared values must reject");
2755        assert_eq!(err.status, StatusCode::UNPROCESSABLE_ENTITY);
2756        assert!(
2757            err.message.contains("PASS") && err.message.contains("BLOCKED"),
2758            "rejection must echo the declared values, got: {}",
2759            err.message
2760        );
2761    }
2762
2763    /// The same contract accepts a body that IS a member of `values` —
2764    /// `204`, unaffected submit.
2765    #[tokio::test]
2766    async fn worker_submit_accepts_body_inside_contract_values() {
2767        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2768        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2769        let state = test_state(data_store, run_store);
2770        let task_id = StepId::new();
2771        let handle = seed_task_with_handle(&state, &task_id, "gate", 1, None).await;
2772        state.engine.register_verdict_contracts(HashMap::from([(
2773            "gate".to_string(),
2774            body_verdict_contract(&["PASS", "BLOCKED"]),
2775        )]));
2776
2777        let status = worker_submit(
2778            State(state.clone()),
2779            bearer_headers(&handle),
2780            Query(SubmitQuery { ok: None }),
2781            axum::body::Bytes::from("PASS"),
2782        )
2783        .await
2784        .expect("value inside declared values must succeed");
2785        assert_eq!(status, StatusCode::NO_CONTENT);
2786    }
2787
2788    /// Opt-in regression guard: an agent with NO declared verdict contract
2789    /// is entirely unaffected — `worker_submit` still returns `204` for an
2790    /// arbitrary body, exactly the pre-GH-#50 behavior.
2791    #[tokio::test]
2792    async fn worker_submit_without_a_declared_contract_is_unaffected() {
2793        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2794        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2795        let state = test_state(data_store, run_store);
2796        let task_id = StepId::new();
2797        // No `register_verdict_contracts` call — the agent declared no contract.
2798        let handle = seed_task_with_handle(&state, &task_id, "undeclared-agent", 1, None).await;
2799
2800        let status = worker_submit(
2801            State(state.clone()),
2802            bearer_headers(&handle),
2803            Query(SubmitQuery { ok: None }),
2804            axum::body::Bytes::from("anything at all, no contract to violate"),
2805        )
2806        .await
2807        .expect("no contract declared must never reject");
2808        assert_eq!(status, StatusCode::NO_CONTENT);
2809    }
2810
2811    /// A `channel: "part"` contract rejects a `worker_artifact?name=verdict`
2812    /// value outside `values` with `422`.
2813    #[tokio::test]
2814    async fn worker_artifact_verdict_part_rejects_value_outside_contract_with_422() {
2815        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2816        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2817        let state = test_state(data_store, run_store);
2818        let task_id = StepId::new();
2819        let handle = seed_task_with_handle(&state, &task_id, "gate", 1, None).await;
2820        state.engine.register_verdict_contracts(HashMap::from([(
2821            "gate".to_string(),
2822            part_verdict_contract(&["PASS", "BLOCKED"]),
2823        )]));
2824
2825        let err = worker_artifact(
2826            State(state.clone()),
2827            bearer_headers(&handle),
2828            Query(ArtifactQuery {
2829                name: "verdict".to_string(),
2830            }),
2831            axum::body::Bytes::from("UNKNOWN"),
2832        )
2833        .await
2834        .expect_err("value outside declared values must reject");
2835        assert_eq!(err.status, StatusCode::UNPROCESSABLE_ENTITY);
2836    }
2837
2838    /// A part named anything OTHER than `"verdict"` skips the gate
2839    /// entirely, even with a `channel: "part"` contract declared — `204`,
2840    /// existing pre-GH-#50 behavior unchanged.
2841    #[tokio::test]
2842    async fn worker_artifact_non_verdict_part_skips_the_gate() {
2843        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2844        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2845        let state = test_state(data_store, run_store);
2846        let task_id = StepId::new();
2847        let handle = seed_task_with_handle(&state, &task_id, "gate", 1, None).await;
2848        state.engine.register_verdict_contracts(HashMap::from([(
2849            "gate".to_string(),
2850            part_verdict_contract(&["PASS", "BLOCKED"]),
2851        )]));
2852
2853        let status = worker_artifact(
2854            State(state.clone()),
2855            bearer_headers(&handle),
2856            Query(ArtifactQuery {
2857                name: "notes".to_string(),
2858            }),
2859            axum::body::Bytes::from("anything at all"),
2860        )
2861        .await
2862        .expect("non-verdict part name must never be gated");
2863        assert_eq!(status, StatusCode::NO_CONTENT);
2864    }
2865}