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(|e| ApiError::engine(format!("task_id_from_handle: {e}")))?;
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(|e| ApiError::engine(format!("task_id_from_handle: {e}")))?
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(|e| ApiError::engine(format!("task_id_from_handle: {e}")))?
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(|e| ApiError::engine(format!("task_id_from_handle: {e}")))?
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(|e| ApiError::engine(format!("task_id_from_handle: {e}")))?;
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/// Decides whether the Bearer is a short handle (`wh-XXXXXXXX`). Returns
1032/// `Some(handle)` on a match, `None` otherwise (= caller proceeds to try decoding
1033/// as full `CapToken` JSON).
1034fn parse_worker_handle(s: &str) -> Option<&str> {
1035    let s = s.trim();
1036    if s.starts_with("wh-")
1037        && s.len() >= 5
1038        && s.len() <= 64
1039        && s[3..].chars().all(|c| c.is_ascii_alphanumeric())
1040    {
1041        Some(s)
1042    } else {
1043        None
1044    }
1045}
1046
1047/// Decodes an encoded `CapToken` from `Authorization: Bearer <encoded CapToken>`.
1048/// Kept separate from `extract_bearer` (sid-only) — kept as a distinct fn so
1049/// that sid strings and encoded tokens are not confused, distinguishing them by type.
1050fn decode_worker_bearer(headers: &HeaderMap) -> Result<CapToken, ApiError> {
1051    let v = headers
1052        .get(AUTHORIZATION)
1053        .ok_or_else(|| ApiError::bad_request("missing Authorization header".into()))?
1054        .to_str()
1055        .map_err(|_| ApiError::bad_request("invalid Authorization header encoding".into()))?;
1056    let encoded = v
1057        .strip_prefix("Bearer ")
1058        .ok_or_else(|| ApiError::bad_request("Authorization must be 'Bearer <token>'".into()))?
1059        .trim();
1060    if encoded.is_empty() {
1061        return Err(ApiError::bad_request("Bearer token is empty".into()));
1062    }
1063    CapToken::decode(encoded).map_err(|e| ApiError::bad_request(format!("invalid token: {e}")))
1064}
1065
1066// ──────────────────────────────────────────────────────────────────────────
1067// UT — `assemble_step_pointers` (`projection-adapter` ST5 Worker axis)
1068// ──────────────────────────────────────────────────────────────────────────
1069
1070#[cfg(test)]
1071mod tests {
1072    use super::*;
1073    use axum::response::IntoResponse;
1074    use mlua_swarm::core::agent_context::AgentContextView;
1075    use mlua_swarm::core::config::EngineCfg;
1076    use mlua_swarm::core::engine::Engine;
1077    use mlua_swarm::store::output::{InMemoryOutputStore, OutputStore};
1078    use mlua_swarm::store::run::{InMemoryRunStore, RunRecord, RunStatus, RunStore, StepEntry};
1079    use mlua_swarm::store::task::InMemoryTaskStore;
1080    use mlua_swarm::{RunId, StepId, TaskId};
1081    use serde_json::json;
1082    use std::collections::HashMap;
1083    use std::sync::Arc;
1084    use tokio::sync::Mutex;
1085
1086    /// Per-module test-helper convention (this crate's established
1087    /// pattern — see e.g. `projection::tests::test_state`): a minimal
1088    /// `AppState` wired with the caller-supplied `data_store` / `run_store`
1089    /// so a test can seed both directly rather than driving a real
1090    /// dispatch through them.
1091    fn test_state(data_store: Arc<dyn OutputStore>, run_store: Arc<dyn RunStore>) -> AppState {
1092        let engine = Engine::new(EngineCfg::default());
1093        let compiler = mlua_swarm::Compiler::new(crate::default_registry());
1094        let launch = Arc::new(mlua_swarm::TaskLaunchService::new(engine.clone(), compiler));
1095        AppState {
1096            engine,
1097            sessions: Arc::new(Mutex::new(crate::SessionStore::default())),
1098            task_app: Arc::new(mlua_swarm::TaskApplication::new_inline_only(launch)),
1099            ws_operator_factory: None,
1100            data_store,
1101            operator_sessions: Arc::new(Mutex::new(HashMap::new())),
1102            roles_to_sid: Arc::new(Mutex::new(HashMap::new())),
1103            task_store: Arc::new(InMemoryTaskStore::new()),
1104            run_store,
1105            base_url: None,
1106            sync_timeout_secs: 300,
1107        }
1108    }
1109
1110    async fn append_final(
1111        data_store: &Arc<dyn OutputStore>,
1112        task_id: &str,
1113        producer: &str,
1114        value: Value,
1115    ) {
1116        data_store
1117            .append(
1118                task_id,
1119                1,
1120                producer,
1121                OutputEvent::Final {
1122                    content: ContentRef::Inline { value },
1123                    ok: true,
1124                },
1125                vec![],
1126            )
1127            .await
1128            .expect("append final");
1129    }
1130
1131    fn step_entry(step_id: &StepId, step_ref: &str) -> StepEntry {
1132        StepEntry {
1133            step_id: step_id.clone(),
1134            step_ref: Some(step_ref.to_string()),
1135            status: Some("passed".to_string()),
1136            at: 0,
1137        }
1138    }
1139
1140    fn run_record(task_id: &TaskId, run_id: &RunId, step_entries: Vec<StepEntry>) -> RunRecord {
1141        RunRecord {
1142            id: run_id.clone(),
1143            task_id: task_id.clone(),
1144            status: RunStatus::Running,
1145            step_entries,
1146            degradations: Vec::new(),
1147            operator_sid: None,
1148            result_ref: None,
1149            created_at: 0,
1150            updated_at: 0,
1151        }
1152    }
1153
1154    fn consumer_payload(consumer_step_id: &StepId, run_id: &RunId) -> WorkerPayload {
1155        WorkerPayload {
1156            task_id: consumer_step_id.clone(),
1157            attempt: 1,
1158            agent: "consumer".to_string(),
1159            system: None,
1160            prompt: String::new(),
1161            context: Some(AgentContextView {
1162                task_id: consumer_step_id.to_string(),
1163                agent: "consumer".to_string(),
1164                attempt: 1,
1165                run_id: Some(run_id.to_string()),
1166                ..Default::default()
1167            }),
1168            system_ref: None,
1169        }
1170    }
1171
1172    /// Test 1: `ContextPolicy.steps` unspecified (no policy seeded at all
1173    /// — `Engine::context_policy_for`'s "no entry" default is `None` /
1174    /// pass-all) → the fetch payload carries every submitted step's
1175    /// `StepPointer`.
1176    #[tokio::test]
1177    async fn context_policy_unspecified_yields_every_submitted_step() {
1178        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
1179        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
1180        let task_id = TaskId::new();
1181        let run_id = RunId::new();
1182        let planner_id = StepId::new();
1183        let coder_id = StepId::new();
1184
1185        append_final(
1186            &data_store,
1187            planner_id.as_str(),
1188            "planner",
1189            json!({"plan": "x"}),
1190        )
1191        .await;
1192        append_final(
1193            &data_store,
1194            coder_id.as_str(),
1195            "coder",
1196            json!({"code": "y"}),
1197        )
1198        .await;
1199        run_store
1200            .create(run_record(
1201                &task_id,
1202                &run_id,
1203                vec![
1204                    step_entry(&planner_id, "planner"),
1205                    step_entry(&coder_id, "coder"),
1206                ],
1207            ))
1208            .await
1209            .expect("create run");
1210
1211        let state = test_state(data_store, run_store);
1212        let consumer_id = StepId::new();
1213        let mut payload = consumer_payload(&consumer_id, &run_id);
1214        assemble_step_pointers(&state, &mut payload).await;
1215
1216        let names: Vec<&str> = payload
1217            .context
1218            .as_ref()
1219            .expect("context")
1220            .steps
1221            .iter()
1222            .map(|p| p.name.as_str())
1223            .collect();
1224        assert!(names.contains(&"planner"), "names: {names:?}");
1225        assert!(names.contains(&"coder"), "names: {names:?}");
1226    }
1227
1228    /// Test 2: `steps: ["planner"]` → only `planner`'s pointer is present.
1229    #[tokio::test]
1230    async fn context_policy_steps_include_list_filters_to_named_steps() {
1231        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
1232        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
1233        let task_id = TaskId::new();
1234        let run_id = RunId::new();
1235        let planner_id = StepId::new();
1236        let coder_id = StepId::new();
1237        append_final(&data_store, planner_id.as_str(), "planner", json!("x")).await;
1238        append_final(&data_store, coder_id.as_str(), "coder", json!("y")).await;
1239        run_store
1240            .create(run_record(
1241                &task_id,
1242                &run_id,
1243                vec![
1244                    step_entry(&planner_id, "planner"),
1245                    step_entry(&coder_id, "coder"),
1246                ],
1247            ))
1248            .await
1249            .expect("create run");
1250
1251        let state = test_state(data_store, run_store);
1252        let consumer_id = StepId::new();
1253        state
1254            .engine
1255            .with_state("test.seed_policy", {
1256                let consumer_id = consumer_id.clone();
1257                move |s| {
1258                    s.agent_ctx.insert(
1259                        (consumer_id, 1),
1260                        mlua_swarm::core::state::AgentCtxEntry {
1261                            policy: mlua_swarm_schema::ContextPolicy {
1262                                steps: Some(vec!["planner".to_string()]),
1263                                ..Default::default()
1264                            },
1265                            ..Default::default()
1266                        },
1267                    );
1268                }
1269            })
1270            .await
1271            .expect("seed policy");
1272
1273        let mut payload = consumer_payload(&consumer_id, &run_id);
1274        assemble_step_pointers(&state, &mut payload).await;
1275
1276        let names: Vec<&str> = payload
1277            .context
1278            .as_ref()
1279            .expect("context")
1280            .steps
1281            .iter()
1282            .map(|p| p.name.as_str())
1283            .collect();
1284        assert_eq!(names, vec!["planner"], "names: {names:?}");
1285    }
1286
1287    /// Test 3: `steps: []` → the pointer list is empty.
1288    #[tokio::test]
1289    async fn context_policy_steps_empty_list_yields_no_pointers() {
1290        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
1291        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
1292        let task_id = TaskId::new();
1293        let run_id = RunId::new();
1294        let planner_id = StepId::new();
1295        append_final(&data_store, planner_id.as_str(), "planner", json!("x")).await;
1296        run_store
1297            .create(run_record(
1298                &task_id,
1299                &run_id,
1300                vec![step_entry(&planner_id, "planner")],
1301            ))
1302            .await
1303            .expect("create run");
1304
1305        let state = test_state(data_store, run_store);
1306        let consumer_id = StepId::new();
1307        state
1308            .engine
1309            .with_state("test.seed_policy", {
1310                let consumer_id = consumer_id.clone();
1311                move |s| {
1312                    s.agent_ctx.insert(
1313                        (consumer_id, 1),
1314                        mlua_swarm::core::state::AgentCtxEntry {
1315                            policy: mlua_swarm_schema::ContextPolicy {
1316                                steps: Some(vec![]),
1317                                ..Default::default()
1318                            },
1319                            ..Default::default()
1320                        },
1321                    );
1322                }
1323            })
1324            .await
1325            .expect("seed policy");
1326
1327        let mut payload = consumer_payload(&consumer_id, &run_id);
1328        assemble_step_pointers(&state, &mut payload).await;
1329
1330        assert!(payload.context.expect("context").steps.is_empty());
1331    }
1332
1333    /// Test 4: `steps_exclude` wins over `steps` for a name in both.
1334    #[tokio::test]
1335    async fn context_policy_steps_exclude_wins_over_steps() {
1336        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
1337        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
1338        let task_id = TaskId::new();
1339        let run_id = RunId::new();
1340        let planner_id = StepId::new();
1341        let coder_id = StepId::new();
1342        append_final(&data_store, planner_id.as_str(), "planner", json!("x")).await;
1343        append_final(&data_store, coder_id.as_str(), "coder", json!("y")).await;
1344        run_store
1345            .create(run_record(
1346                &task_id,
1347                &run_id,
1348                vec![
1349                    step_entry(&planner_id, "planner"),
1350                    step_entry(&coder_id, "coder"),
1351                ],
1352            ))
1353            .await
1354            .expect("create run");
1355
1356        let state = test_state(data_store, run_store);
1357        let consumer_id = StepId::new();
1358        state
1359            .engine
1360            .with_state("test.seed_policy", {
1361                let consumer_id = consumer_id.clone();
1362                move |s| {
1363                    s.agent_ctx.insert(
1364                        (consumer_id, 1),
1365                        mlua_swarm::core::state::AgentCtxEntry {
1366                            policy: mlua_swarm_schema::ContextPolicy {
1367                                steps: Some(vec!["planner".to_string(), "coder".to_string()]),
1368                                steps_exclude: vec!["planner".to_string()],
1369                                ..Default::default()
1370                            },
1371                            ..Default::default()
1372                        },
1373                    );
1374                }
1375            })
1376            .await
1377            .expect("seed policy");
1378
1379        let mut payload = consumer_payload(&consumer_id, &run_id);
1380        assemble_step_pointers(&state, &mut payload).await;
1381
1382        let names: Vec<&str> = payload
1383            .context
1384            .as_ref()
1385            .expect("context")
1386            .steps
1387            .iter()
1388            .map(|p| p.name.as_str())
1389            .collect();
1390        assert_eq!(names, vec!["coder"], "names: {names:?}");
1391    }
1392
1393    /// Test 5 (in-flight window, subtask-4-style invariant): the Run has
1394    /// NOT finalized (`result_ref: None`, mirroring a Run still `Running`)
1395    /// yet the fetch payload still carries a `StepPointer` for a step
1396    /// already visible through the Data-plane store — the same mechanism
1397    /// `crates/mlua-swarm-server/src/projection.rs`'s
1398    /// `steps_list_returns_in_flight_step_output_before_run_completes`
1399    /// proves end-to-end through a real gated 2-step dispatch; this test
1400    /// isolates the same invariant at the `assemble_step_pointers` level.
1401    #[tokio::test]
1402    async fn in_flight_step_output_is_visible_before_run_finalizes() {
1403        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
1404        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
1405        let task_id = TaskId::new();
1406        let run_id = RunId::new();
1407        let step1_id = StepId::new();
1408        append_final(
1409            &data_store,
1410            step1_id.as_str(),
1411            "step1",
1412            json!({"step1_out": "hi"}),
1413        )
1414        .await;
1415        let mut run = run_record(&task_id, &run_id, vec![step_entry(&step1_id, "step1")]);
1416        run.status = RunStatus::Running;
1417        run.result_ref = None; // the in-flight window: not yet finalized.
1418        run_store.create(run).await.expect("create run");
1419
1420        let state = test_state(data_store, run_store);
1421        let consumer_id = StepId::new();
1422        let mut payload = consumer_payload(&consumer_id, &run_id);
1423        assemble_step_pointers(&state, &mut payload).await;
1424
1425        let steps = &payload.context.expect("context").steps;
1426        assert_eq!(steps.len(), 1);
1427        assert_eq!(steps[0].name, "step1");
1428    }
1429
1430    /// Test 6: the fetching agent's own name is always excluded, even if
1431    /// (e.g. a loop re-dispatching the same agent) it also appears in
1432    /// `run.step_entries` with a resolvable Data-plane record.
1433    #[tokio::test]
1434    async fn self_agent_name_is_always_excluded() {
1435        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
1436        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
1437        let task_id = TaskId::new();
1438        let run_id = RunId::new();
1439        let planner_id = StepId::new();
1440        let consumer_prior_id = StepId::new();
1441        append_final(&data_store, planner_id.as_str(), "planner", json!("x")).await;
1442        append_final(
1443            &data_store,
1444            consumer_prior_id.as_str(),
1445            "consumer",
1446            json!("self"),
1447        )
1448        .await;
1449        run_store
1450            .create(run_record(
1451                &task_id,
1452                &run_id,
1453                vec![
1454                    step_entry(&planner_id, "planner"),
1455                    step_entry(&consumer_prior_id, "consumer"),
1456                ],
1457            ))
1458            .await
1459            .expect("create run");
1460
1461        let state = test_state(data_store, run_store);
1462        let consumer_id = StepId::new();
1463        let mut payload = consumer_payload(&consumer_id, &run_id);
1464        assemble_step_pointers(&state, &mut payload).await;
1465
1466        let names: Vec<&str> = payload
1467            .context
1468            .as_ref()
1469            .expect("context")
1470            .steps
1471            .iter()
1472            .map(|p| p.name.as_str())
1473            .collect();
1474        assert!(!names.contains(&"consumer"), "names: {names:?}");
1475        assert!(names.contains(&"planner"), "names: {names:?}");
1476    }
1477
1478    /// Test 7 (pointer-only invariant): a `StepPointer`'s serialized JSON
1479    /// carries no preview / content-bytes field — only `name` /
1480    /// `size_bytes` / `file_path?` / `content_url` / `sha256`.
1481    #[tokio::test]
1482    async fn step_pointer_serializes_with_no_preview_or_content_bytes() {
1483        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
1484        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
1485        let task_id = TaskId::new();
1486        let run_id = RunId::new();
1487        let planner_id = StepId::new();
1488        append_final(
1489            &data_store,
1490            planner_id.as_str(),
1491            "planner",
1492            json!({"plan": "do the thing, at length".repeat(50)}),
1493        )
1494        .await;
1495        run_store
1496            .create(run_record(
1497                &task_id,
1498                &run_id,
1499                vec![step_entry(&planner_id, "planner")],
1500            ))
1501            .await
1502            .expect("create run");
1503
1504        let state = test_state(data_store, run_store);
1505        let consumer_id = StepId::new();
1506        let mut payload = consumer_payload(&consumer_id, &run_id);
1507        assemble_step_pointers(&state, &mut payload).await;
1508
1509        let steps = &payload.context.expect("context").steps;
1510        assert_eq!(steps.len(), 1);
1511        let json_value = serde_json::to_value(&steps[0]).expect("serialize StepPointer");
1512        let obj = json_value.as_object().expect("object");
1513        for forbidden in ["preview", "content", "value", "bytes"] {
1514            assert!(
1515                !obj.contains_key(forbidden),
1516                "StepPointer must not carry a {forbidden:?} field: {obj:?}"
1517            );
1518        }
1519        assert!(obj.contains_key("name"));
1520        assert!(obj.contains_key("size_bytes"));
1521        assert!(obj.contains_key("content_url"));
1522        assert!(obj.contains_key("sha256"));
1523    }
1524
1525    /// A single-step Blueprint whose `planner` agent declares
1526    /// `AgentMeta.projection_name = "plan-out"` — the `StepNaming` fixture
1527    /// for [`declared_projection_name_pointer_name_is_canonical_and_policy_matches_it`],
1528    /// mirroring `crate::projection::tests`' own
1529    /// `declared_projection_name_blueprint` helper (duplicated here rather
1530    /// than shared — this crate's established per-module test-helper
1531    /// convention).
1532    fn declared_name_bp() -> mlua_swarm::blueprint::Blueprint {
1533        use mlua_flow_ir::{Expr, Node};
1534        use mlua_swarm::blueprint::{
1535            current_schema_version, AgentDef, AgentKind, AgentMeta, Blueprint, BlueprintMetadata,
1536            CompilerHints, CompilerStrategy,
1537        };
1538        Blueprint {
1539            schema_version: current_schema_version(),
1540            id: "worker-test-declared-name-bp".into(),
1541            flow: Node::Step {
1542                ref_: "planner".to_string(),
1543                in_: Expr::Path {
1544                    at: "$.in".parse().expect("literal test path: $.in"),
1545                },
1546                out: Expr::Path {
1547                    at: "$.plan".parse().expect("literal test path: $.plan"),
1548                },
1549            },
1550            agents: vec![AgentDef {
1551                name: "planner".to_string(),
1552                kind: AgentKind::RustFn,
1553                spec: json!({"fn_id": "planner"}),
1554                profile: None,
1555                meta: Some(AgentMeta {
1556                    projection_name: Some("plan-out".to_string()),
1557                    ..Default::default()
1558                }),
1559                runner: None,
1560                runner_ref: None,
1561                verdict: None,
1562            }],
1563            operators: vec![],
1564            metas: vec![],
1565            hints: CompilerHints::default(),
1566            strategy: CompilerStrategy::default(),
1567            metadata: BlueprintMetadata::default(),
1568            spawner_hints: Default::default(),
1569            default_agent_kind: AgentKind::Operator,
1570            default_operator_kind: None,
1571            default_init_ctx: None,
1572            default_agent_ctx: None,
1573            default_context_policy: None,
1574            projection_placement: None,
1575            audits: vec![],
1576            degradation_policy: None,
1577            runners: vec![],
1578            default_runner: None,
1579            check_policy: None,
1580        }
1581    }
1582
1583    /// Test 8 (GH #23 subtask-3, declared-name E2E — Worker axis half): a
1584    /// declared `projection_name` makes `StepPointer.name` the CANONICAL
1585    /// name (not the raw `Step.ref` the Data-plane / `step_entries` still
1586    /// index by), and `ContextPolicy.steps` naming the canonical name
1587    /// matches it.
1588    #[tokio::test]
1589    async fn declared_projection_name_pointer_name_is_canonical_and_policy_matches_it() {
1590        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
1591        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
1592        let task_id = TaskId::new();
1593        let run_id = RunId::new();
1594        let planner_id = StepId::new();
1595
1596        // The Data-plane store is keyed by the CANONICAL name — GH #23
1597        // subtask-2's sink already writes it that way.
1598        append_final(
1599            &data_store,
1600            planner_id.as_str(),
1601            "plan-out",
1602            json!({"plan": "x"}),
1603        )
1604        .await;
1605        run_store
1606            .create(run_record(
1607                &task_id,
1608                &run_id,
1609                vec![step_entry(&planner_id, "planner")],
1610            ))
1611            .await
1612            .expect("create run");
1613
1614        let state = test_state(data_store, run_store);
1615
1616        // Seed the `StepNaming` table the way `Compiler::compile` +
1617        // `EngineDispatcher::dispatch` would have — the same `Arc` stashed
1618        // under every dispatched step's own id, including the FETCHING
1619        // agent's (`consumer_id`), which `assemble_step_pointers` looks up
1620        // via `Engine::step_naming_for(&payload.task_id)`.
1621        let (naming, _warnings) =
1622            mlua_swarm::core::step_naming::StepNaming::from_blueprint(&declared_name_bp())
1623                .expect("no collision");
1624        let naming = Arc::new(naming);
1625        let consumer_id = StepId::new();
1626        state
1627            .engine
1628            .with_state("test.seed_step_naming", {
1629                let naming = naming.clone();
1630                let planner_id = planner_id.clone();
1631                let consumer_id = consumer_id.clone();
1632                move |s| {
1633                    s.step_namings.insert(planner_id, naming.clone());
1634                    s.step_namings.insert(consumer_id, naming);
1635                }
1636            })
1637            .await
1638            .expect("seed step naming");
1639        state
1640            .engine
1641            .with_state("test.seed_policy", {
1642                let consumer_id = consumer_id.clone();
1643                move |s| {
1644                    s.agent_ctx.insert(
1645                        (consumer_id, 1),
1646                        mlua_swarm::core::state::AgentCtxEntry {
1647                            policy: mlua_swarm_schema::ContextPolicy {
1648                                steps: Some(vec!["plan-out".to_string()]),
1649                                ..Default::default()
1650                            },
1651                            ..Default::default()
1652                        },
1653                    );
1654                }
1655            })
1656            .await
1657            .expect("seed policy");
1658
1659        let mut payload = consumer_payload(&consumer_id, &run_id);
1660        assemble_step_pointers(&state, &mut payload).await;
1661
1662        let steps = &payload.context.expect("context").steps;
1663        assert_eq!(steps.len(), 1, "steps: {steps:?}");
1664        assert_eq!(
1665            steps[0].name, "plan-out",
1666            "StepPointer.name must be the canonical name"
1667        );
1668    }
1669
1670    // ──────────────────────────────────────────────────────────────────────
1671    // GH #31 — `/v1/worker/prompt/system` + `/v1/agents/:name/render-size`
1672    // ──────────────────────────────────────────────────────────────────────
1673
1674    /// Seeds a task + baked system prompt + a short worker handle bound to
1675    /// it, mirroring the shape `Engine::dispatch_attempt` would have
1676    /// produced (minus the parts these two routes don't touch: no real
1677    /// HMAC-signed `CapToken`, since `task_id_from_handle`'s handle → fp →
1678    /// task_id chain is what's under test, not signature verification).
1679    async fn seed_task_with_handle(
1680        state: &AppState,
1681        task_id: &StepId,
1682        agent: &str,
1683        attempt: u32,
1684        system: Option<String>,
1685    ) -> String {
1686        let handle = format!("wh-{}", mlua_swarm::types::secure_hex(4));
1687        let task_id = task_id.clone();
1688        let agent = agent.to_string();
1689        let handle_clone = handle.clone();
1690        state
1691            .engine
1692            .with_state("test.seed_task_with_handle", move |s| {
1693                let mut task = mlua_swarm::core::state::TaskState::new(
1694                    task_id.clone(),
1695                    mlua_swarm::core::state::TaskSpec {
1696                        agent: agent.clone(),
1697                        initial_directive: json!("x"),
1698                        step_ctx: None,
1699                        check_policy: None,
1700                    },
1701                );
1702                task.attempt = attempt;
1703                s.tasks.insert(task_id.clone(), task);
1704                s.systems.insert((task_id.clone(), attempt), system);
1705                let token = CapToken {
1706                    agent_id: agent,
1707                    role: mlua_swarm::Role::Worker,
1708                    scopes: vec!["*".to_string()],
1709                    issued_at: 0,
1710                    expire_at: u64::MAX,
1711                    max_uses: None,
1712                    nonce: format!("test-nonce-{task_id}"),
1713                    sig_hex: String::new(),
1714                };
1715                let fp = token.fingerprint();
1716                s.tokens.insert(
1717                    fp.clone(),
1718                    mlua_swarm::core::state::CapTokenRecord {
1719                        token,
1720                        uses_left: None,
1721                        revoked: false,
1722                        task_id: Some(task_id),
1723                    },
1724                );
1725                s.worker_handles.insert(handle_clone, fp);
1726            })
1727            .await
1728            .expect("seed_task_with_handle");
1729        handle
1730    }
1731
1732    fn bearer_headers(handle: &str) -> HeaderMap {
1733        let mut headers = HeaderMap::new();
1734        headers.insert(
1735            AUTHORIZATION,
1736            format!("Bearer {handle}").parse().expect("header value"),
1737        );
1738        headers
1739    }
1740
1741    /// `GET /v1/worker/prompt/system` returns the exact raw baked bytes
1742    /// (not JSON-wrapped) with `Content-Type: text/plain`, for the
1743    /// `(task_id, attempt)` the handle is bound to.
1744    #[tokio::test]
1745    async fn worker_prompt_system_returns_raw_bytes_for_baked_system() {
1746        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
1747        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
1748        let state = test_state(data_store, run_store);
1749        let task_id = StepId::new();
1750        let rendered = "# Hello\n\nThis is the baked system prompt.".to_string();
1751        let handle =
1752            seed_task_with_handle(&state, &task_id, "planner", 1, Some(rendered.clone())).await;
1753
1754        let resp = worker_prompt_system(
1755            State(state.clone()),
1756            bearer_headers(&handle),
1757            Query(PromptSystemQuery {
1758                task_id: task_id.clone(),
1759                attempt: 1,
1760            }),
1761        )
1762        .await
1763        .expect("worker_prompt_system")
1764        .into_response();
1765
1766        assert_eq!(resp.status(), StatusCode::OK);
1767        let content_type = resp
1768            .headers()
1769            .get(header::CONTENT_TYPE)
1770            .expect("content-type header")
1771            .to_str()
1772            .expect("ascii");
1773        assert_eq!(content_type, "text/plain; charset=utf-8");
1774        let body_bytes = axum::body::to_bytes(resp.into_body(), usize::MAX)
1775            .await
1776            .expect("body bytes");
1777        assert_eq!(body_bytes.as_ref(), rendered.as_bytes());
1778    }
1779
1780    /// No baked system for the given `(task_id, attempt)` → 404, not a
1781    /// panic or a 200-with-empty-body.
1782    #[tokio::test]
1783    async fn worker_prompt_system_404s_when_no_baked_system() {
1784        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
1785        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
1786        let state = test_state(data_store, run_store);
1787        let task_id = StepId::new();
1788        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
1789
1790        let result = worker_prompt_system(
1791            State(state.clone()),
1792            bearer_headers(&handle),
1793            Query(PromptSystemQuery {
1794                task_id: task_id.clone(),
1795                attempt: 1,
1796            }),
1797        )
1798        .await;
1799        let err = match result {
1800            Ok(_) => panic!("expected 404 ApiError, got Ok"),
1801            Err(e) => e,
1802        };
1803        assert_eq!(err.into_response().status(), StatusCode::NOT_FOUND);
1804    }
1805
1806    /// A handle bound to a different task than the one requested must be
1807    /// rejected (400) — this is the same cross-check `worker_prompt` does.
1808    #[tokio::test]
1809    async fn worker_prompt_system_rejects_handle_task_mismatch() {
1810        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
1811        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
1812        let state = test_state(data_store, run_store);
1813        let task_id = StepId::new();
1814        let other_task_id = StepId::new();
1815        let handle =
1816            seed_task_with_handle(&state, &task_id, "planner", 1, Some("x".to_string())).await;
1817
1818        let result = worker_prompt_system(
1819            State(state.clone()),
1820            bearer_headers(&handle),
1821            Query(PromptSystemQuery {
1822                task_id: other_task_id,
1823                attempt: 1,
1824            }),
1825        )
1826        .await;
1827        let err = match result {
1828            Ok(_) => panic!("expected 400 ApiError for task mismatch, got Ok"),
1829            Err(e) => e,
1830        };
1831        assert_eq!(err.into_response().status(), StatusCode::BAD_REQUEST);
1832    }
1833
1834    /// `GET /v1/agents/:name/render-size` requires no auth, and reports
1835    /// `last_rendered_bytes: null` for an agent that has never had a
1836    /// `system_prompt` baked — a normal 200, not a 404.
1837    #[tokio::test]
1838    async fn agent_render_size_returns_null_for_unknown_agent() {
1839        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
1840        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
1841        let state = test_state(data_store, run_store);
1842
1843        let Json(body) = agent_render_size(
1844            State(state.clone()),
1845            axum::extract::Path("never-dispatched".to_string()),
1846        )
1847        .await;
1848        assert_eq!(body.agent, "never-dispatched");
1849        assert_eq!(body.last_rendered_bytes, None);
1850    }
1851
1852    /// Once `bake_worker_system_prompt` has recorded a render size for an
1853    /// agent, the route reports the most-recently-observed value.
1854    #[tokio::test]
1855    async fn agent_render_size_reports_last_rendered_bytes() {
1856        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
1857        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
1858        let state = test_state(data_store, run_store);
1859        let task_id = StepId::new();
1860        state
1861            .engine
1862            .with_state("test.seed_agent_ctx_for_bake", {
1863                let task_id = task_id.clone();
1864                move |s| {
1865                    s.tasks.insert(
1866                        task_id.clone(),
1867                        mlua_swarm::core::state::TaskState::new(
1868                            task_id,
1869                            mlua_swarm::core::state::TaskSpec {
1870                                agent: "coder".to_string(),
1871                                initial_directive: json!("x"),
1872                                step_ctx: None,
1873                                check_policy: None,
1874                            },
1875                        ),
1876                    );
1877                }
1878            })
1879            .await
1880            .expect("seed task");
1881        state
1882            .engine
1883            .bake_worker_system_prompt(&task_id, 1, Some("z".repeat(42)))
1884            .await
1885            .expect("bake_worker_system_prompt");
1886
1887        let Json(body) = agent_render_size(
1888            State(state.clone()),
1889            axum::extract::Path("coder".to_string()),
1890        )
1891        .await;
1892        assert_eq!(body.agent, "coder");
1893        assert_eq!(body.last_rendered_bytes, Some(42));
1894    }
1895
1896    // ──────────────────────────────────────────────────────────────────────
1897    // GH #36 ST1 — `POST /v1/worker/artifact`
1898    // ──────────────────────────────────────────────────────────────────────
1899
1900    /// A valid `?name=` + short-handle Bearer stages the raw body (trailing
1901    /// whitespace trimmed, same as `worker_submit`) as an `Artifact` on the
1902    /// task's current-attempt tail, and returns `204 No Content`.
1903    #[tokio::test]
1904    async fn worker_artifact_stages_and_204s_for_valid_request() {
1905        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
1906        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
1907        let state = test_state(data_store, run_store);
1908        let task_id = StepId::new();
1909        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
1910
1911        let status = worker_artifact(
1912            State(state.clone()),
1913            bearer_headers(&handle),
1914            Query(ArtifactQuery {
1915                name: "summary".to_string(),
1916            }),
1917            axum::body::Bytes::from_static(b"hello artifact\n"),
1918        )
1919        .await
1920        .expect("worker_artifact");
1921        assert_eq!(status, StatusCode::NO_CONTENT);
1922
1923        let tail = state.engine.output_tail(&task_id, 1).await;
1924        assert_eq!(tail.len(), 1, "tail: {tail:?}");
1925        match &tail[0] {
1926            OutputEvent::Artifact { name, content } => {
1927                assert_eq!(name, "summary");
1928                match content {
1929                    ContentRef::Inline { value } => {
1930                        assert_eq!(value, &json!("hello artifact"));
1931                    }
1932                    other => panic!("expected Inline content, got {other:?}"),
1933                }
1934            }
1935            other => panic!("expected Artifact event, got {other:?}"),
1936        }
1937    }
1938
1939    /// `?name=` missing entirely → axum's `Query` extractor rejection
1940    /// (400), not a panic. `Query<ArtifactQuery>` is constructed directly
1941    /// in this test (mirroring the other handlers' unit style, which call
1942    /// the handler fn with an already-extracted `Query`) — an empty `name`
1943    /// is exercised separately below since that case is NOT caught by the
1944    /// extractor and must be checked in the handler body.
1945    #[tokio::test]
1946    async fn worker_artifact_rejects_blank_name() {
1947        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
1948        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
1949        let state = test_state(data_store, run_store);
1950        let task_id = StepId::new();
1951        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
1952
1953        let result = worker_artifact(
1954            State(state.clone()),
1955            bearer_headers(&handle),
1956            Query(ArtifactQuery {
1957                name: "   ".to_string(),
1958            }),
1959            axum::body::Bytes::from_static(b"x"),
1960        )
1961        .await;
1962        let err = match result {
1963            Ok(_) => panic!("expected 400 ApiError for blank name, got Ok"),
1964            Err(e) => e,
1965        };
1966        assert_eq!(err.into_response().status(), StatusCode::BAD_REQUEST);
1967
1968        // Nothing was staged.
1969        assert!(state.engine.output_tail(&task_id, 1).await.is_empty());
1970    }
1971
1972    /// Staging the same `name` twice within one attempt is last-write-wins
1973    /// on the folded value (`fold_final_and_parts` in `mlua_swarm::core::
1974    /// engine`) — this test only asserts the raw tail carries both events
1975    /// in order (the fold itself is covered by that crate's own unit
1976    /// tests); `Engine::stage_worker_artifact_trusted`'s doc.
1977    #[tokio::test]
1978    async fn worker_artifact_staging_same_name_twice_appends_both_events_in_order() {
1979        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
1980        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
1981        let state = test_state(data_store, run_store);
1982        let task_id = StepId::new();
1983        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
1984
1985        for body in [b"first".as_slice(), b"second".as_slice()] {
1986            worker_artifact(
1987                State(state.clone()),
1988                bearer_headers(&handle),
1989                Query(ArtifactQuery {
1990                    name: "a".to_string(),
1991                }),
1992                axum::body::Bytes::copy_from_slice(body),
1993            )
1994            .await
1995            .expect("worker_artifact");
1996        }
1997
1998        let tail = state.engine.output_tail(&task_id, 1).await;
1999        assert_eq!(tail.len(), 2, "tail: {tail:?}");
2000        let values: Vec<&str> = tail
2001            .iter()
2002            .map(|ev| match ev {
2003                OutputEvent::Artifact {
2004                    content: ContentRef::Inline { value },
2005                    ..
2006                } => value.as_str().expect("string value"),
2007                other => panic!("expected Artifact/Inline event, got {other:?}"),
2008            })
2009            .collect();
2010        assert_eq!(values, vec!["first", "second"]);
2011    }
2012
2013    // ──────────────────────────────────────────────────────────────────
2014    // GH #37 — terminal-run guard (`reject_if_run_terminal`)
2015    // ──────────────────────────────────────────────────────────────────
2016
2017    /// Links a seeded dispatch task to a Run the same way
2018    /// `AgentContextMiddleware` does at spawn time: an `agent_ctx` entry
2019    /// whose view carries the `run_id`.
2020    async fn link_task_to_run(state: &AppState, task_id: &StepId, attempt: u32, run_id: &RunId) {
2021        let tid = task_id.clone();
2022        let rid_str = run_id.to_string();
2023        state
2024            .engine
2025            .with_state("test.link_task_to_run", move |s| {
2026                let mut entry = mlua_swarm::core::state::AgentCtxEntry::default();
2027                entry.view.run_id = Some(rid_str);
2028                s.agent_ctx.insert((tid, attempt), entry);
2029            })
2030            .await
2031            .expect("link_task_to_run");
2032    }
2033
2034    /// GH #37: a submit / artifact addressed at a Run that already
2035    /// reached a terminal status must be rejected with `410 Gone` — the
2036    /// flow-eval driver for that Run is gone, so a silent `204` here
2037    /// would orphan the worker's output.
2038    #[tokio::test]
2039    async fn submit_and_artifact_against_terminal_run_return_410() {
2040        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2041        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2042        let state = test_state(data_store, run_store.clone());
2043        let task_id = StepId::new();
2044        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
2045
2046        let owner_task = TaskId::new();
2047        let run_id = RunId::new();
2048        let mut rec = run_record(&owner_task, &run_id, vec![]);
2049        rec.status = RunStatus::Failed;
2050        run_store.create(rec).await.expect("run create");
2051        link_task_to_run(&state, &task_id, 1, &run_id).await;
2052
2053        let err = worker_submit(
2054            State(state.clone()),
2055            bearer_headers(&handle),
2056            Query(SubmitQuery { ok: None }),
2057            axum::body::Bytes::from_static(b"LATE OUTPUT"),
2058        )
2059        .await
2060        .expect_err("a submit against a Failed run must be rejected");
2061        assert_eq!(err.status, StatusCode::GONE);
2062        assert!(
2063            err.message.contains(&run_id.to_string()),
2064            "the 410 must name the terminal run: {}",
2065            err.message
2066        );
2067
2068        let err = worker_artifact(
2069            State(state.clone()),
2070            bearer_headers(&handle),
2071            Query(ArtifactQuery {
2072                name: "part.md".to_string(),
2073            }),
2074            axum::body::Bytes::from_static(b"LATE PART"),
2075        )
2076        .await
2077        .expect_err("an artifact staged against a Failed run must be rejected");
2078        assert_eq!(err.status, StatusCode::GONE);
2079
2080        // The rejected values must not have reached the output tail.
2081        let tail = state.engine.output_tail(&task_id, 1).await;
2082        assert!(tail.is_empty(), "rejected submits must not land: {tail:?}");
2083    }
2084
2085    /// GH #37 fail-open contract: the guard must never turn a
2086    /// would-have-succeeded submit into a failure — no run linkage at
2087    /// all, an unknown Run, and a live (`Running`) Run all pass.
2088    #[tokio::test]
2089    async fn terminal_run_guard_is_fail_open() {
2090        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2091        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2092        let state = test_state(data_store, run_store.clone());
2093        let task_id = StepId::new();
2094        seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
2095
2096        // (a) No agent-ctx linkage at all (pre-run-tracking dispatch).
2097        reject_if_run_terminal(&state, &task_id, 1)
2098            .await
2099            .expect("no linkage must fail open");
2100
2101        // (b) Linked to a Run the store does not know.
2102        let unknown_run = RunId::new();
2103        link_task_to_run(&state, &task_id, 1, &unknown_run).await;
2104        reject_if_run_terminal(&state, &task_id, 1)
2105            .await
2106            .expect("unknown run must fail open");
2107
2108        // (c) Linked to a live Run.
2109        let owner_task = TaskId::new();
2110        let live_run = RunId::new();
2111        run_store
2112            .create(run_record(&owner_task, &live_run, vec![]))
2113            .await
2114            .expect("run create");
2115        link_task_to_run(&state, &task_id, 1, &live_run).await;
2116        reject_if_run_terminal(&state, &task_id, 1)
2117            .await
2118            .expect("a Running run must pass the guard");
2119    }
2120
2121    // ──────────────────────────────────────────────────────────────────
2122    // GH #32 — `POST /v1/worker/degradation`
2123    // ──────────────────────────────────────────────────────────────────
2124
2125    fn degradation_body(tool: &str, note: Option<&str>) -> DegradationBody {
2126        DegradationBody {
2127            tool: tool.to_string(),
2128            error: "boom".to_string(),
2129            fallback: "used cached value".to_string(),
2130            note: note.map(str::to_string),
2131        }
2132    }
2133
2134    /// [`link_task_to_run`] plus the `view.agent` name — production's
2135    /// `AgentContextMiddleware` sets both fields on the same `agent_ctx`
2136    /// entry; the shared GH #37 helper only needed `run_id`, so this
2137    /// sibling fills in `agent` too for tests that assert on the
2138    /// server-injected `step_ref`.
2139    async fn link_task_to_run_with_agent(
2140        state: &AppState,
2141        task_id: &StepId,
2142        attempt: u32,
2143        run_id: &RunId,
2144        agent: &str,
2145    ) {
2146        let tid = task_id.clone();
2147        let rid_str = run_id.to_string();
2148        let agent = agent.to_string();
2149        state
2150            .engine
2151            .with_state("test.link_task_to_run_with_agent", move |s| {
2152                let mut entry = mlua_swarm::core::state::AgentCtxEntry::default();
2153                entry.view.run_id = Some(rid_str);
2154                entry.view.agent = agent;
2155                s.agent_ctx.insert((tid, attempt), entry);
2156            })
2157            .await
2158            .expect("link_task_to_run_with_agent");
2159    }
2160
2161    /// A worker-reported degradation is persisted to the linked Run's
2162    /// `degradations` with the server-injected `step_ref` / `attempt` /
2163    /// `at` fields filled in — the client body never supplies any of the
2164    /// three.
2165    #[tokio::test]
2166    async fn worker_degradation_persists_entry_when_run_tracked() {
2167        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2168        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2169        let state = test_state(data_store, run_store.clone());
2170        let task_id = StepId::new();
2171        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
2172
2173        let owner_task = TaskId::new();
2174        let run_id = RunId::new();
2175        run_store
2176            .create(run_record(&owner_task, &run_id, vec![]))
2177            .await
2178            .expect("run create");
2179        link_task_to_run_with_agent(&state, &task_id, 1, &run_id, "planner").await;
2180
2181        let status = worker_degradation(
2182            State(state.clone()),
2183            bearer_headers(&handle),
2184            Json(degradation_body("web_search", Some("rate limited"))),
2185        )
2186        .await
2187        .expect("worker_degradation");
2188        assert_eq!(status, StatusCode::NO_CONTENT);
2189
2190        let rec = run_store.get(&run_id).await.expect("run get");
2191        assert_eq!(
2192            rec.degradations.len(),
2193            1,
2194            "degradations: {:?}",
2195            rec.degradations
2196        );
2197        let entry = &rec.degradations[0];
2198        assert_eq!(entry.tool, "web_search");
2199        assert_eq!(entry.error, "boom");
2200        assert_eq!(entry.fallback, "used cached value");
2201        assert_eq!(entry.note.as_deref(), Some("rate limited"));
2202        assert_eq!(entry.step_ref.as_deref(), Some("planner"));
2203        assert_eq!(entry.attempt, Some(1));
2204        assert!(entry.at > 0, "at must be a real timestamp: {}", entry.at);
2205    }
2206
2207    /// Two entries POSTed in sequence are appended in order.
2208    #[tokio::test]
2209    async fn worker_degradation_appends_in_order() {
2210        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2211        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2212        let state = test_state(data_store, run_store.clone());
2213        let task_id = StepId::new();
2214        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
2215
2216        let owner_task = TaskId::new();
2217        let run_id = RunId::new();
2218        run_store
2219            .create(run_record(&owner_task, &run_id, vec![]))
2220            .await
2221            .expect("run create");
2222        link_task_to_run(&state, &task_id, 1, &run_id).await;
2223
2224        for tool in ["first_tool", "second_tool"] {
2225            worker_degradation(
2226                State(state.clone()),
2227                bearer_headers(&handle),
2228                Json(degradation_body(tool, None)),
2229            )
2230            .await
2231            .expect("worker_degradation");
2232        }
2233
2234        let rec = run_store.get(&run_id).await.expect("run get");
2235        let tools: Vec<&str> = rec.degradations.iter().map(|e| e.tool.as_str()).collect();
2236        assert_eq!(tools, vec!["first_tool", "second_tool"]);
2237    }
2238
2239    /// A task whose `agent_ctx` carries no Run linkage (pre-run-tracking
2240    /// dispatch) silently 204s — nothing to append to, and this must not
2241    /// surface as a client error.
2242    #[tokio::test]
2243    async fn worker_degradation_silent_ok_when_no_run_tracked() {
2244        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2245        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2246        let state = test_state(data_store, run_store);
2247        let task_id = StepId::new();
2248        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
2249
2250        let status = worker_degradation(
2251            State(state.clone()),
2252            bearer_headers(&handle),
2253            Json(degradation_body("some_tool", None)),
2254        )
2255        .await
2256        .expect("worker_degradation must not error on missing run linkage");
2257        assert_eq!(status, StatusCode::NO_CONTENT);
2258    }
2259
2260    /// GH #37 terminal-run guard applies to the degradation channel too — a
2261    /// dead Run must not accumulate signals.
2262    #[tokio::test]
2263    async fn worker_degradation_rejects_terminal_run() {
2264        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2265        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2266        let state = test_state(data_store, run_store.clone());
2267        let task_id = StepId::new();
2268        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
2269
2270        let owner_task = TaskId::new();
2271        let run_id = RunId::new();
2272        let mut rec = run_record(&owner_task, &run_id, vec![]);
2273        rec.status = RunStatus::Done;
2274        run_store.create(rec).await.expect("run create");
2275        link_task_to_run(&state, &task_id, 1, &run_id).await;
2276
2277        let err = worker_degradation(
2278            State(state.clone()),
2279            bearer_headers(&handle),
2280            Json(degradation_body("some_tool", None)),
2281        )
2282        .await
2283        .expect_err("a degradation against a Done run must be rejected");
2284        assert_eq!(err.status, StatusCode::GONE);
2285
2286        let rec = run_store.get(&run_id).await.expect("run get");
2287        assert!(
2288            rec.degradations.is_empty(),
2289            "rejected degradation must not land: {:?}",
2290            rec.degradations
2291        );
2292    }
2293
2294    // ──────────────────────────────────────────────────────────────────
2295    // GH #42 — `@file:<abs-path>` sentinel resolution in `worker_submit`
2296    // / `worker_artifact`. Guards each verified independently: sentinel
2297    // resolves to the file's trimmed contents; path outside `work_dir`,
2298    // missing file, oversized file, and non-sentinel bodies each get the
2299    // documented behavior.
2300    // ──────────────────────────────────────────────────────────────────
2301
2302    /// Seeds an `agent_ctx` entry whose view carries `work_dir` and, when
2303    /// `allow_file_submit` is `Some`, that value under the GH #43
2304    /// [`FILE_SENTINEL_ALLOW_KEY`] in `view.extra` — matching the shape
2305    /// `AgentContextMiddleware` writes at spawn time. Sentinel resolution
2306    /// requires both the `work_dir` and the strict `Bool(true)` opt-in.
2307    async fn seed_work_dir(
2308        state: &AppState,
2309        task_id: &StepId,
2310        attempt: u32,
2311        work_dir: &str,
2312        allow_file_submit: Option<Value>,
2313    ) {
2314        let tid = task_id.clone();
2315        let work_dir = work_dir.to_string();
2316        state
2317            .engine
2318            .with_state("test.seed_work_dir", move |s| {
2319                let mut entry = mlua_swarm::core::state::AgentCtxEntry::default();
2320                entry.view.work_dir = Some(work_dir);
2321                if let Some(v) = allow_file_submit {
2322                    entry
2323                        .view
2324                        .extra
2325                        .insert(FILE_SENTINEL_ALLOW_KEY.to_string(), v);
2326                }
2327                s.agent_ctx.insert((tid, attempt), entry);
2328            })
2329            .await
2330            .expect("seed_work_dir");
2331    }
2332
2333    /// Sentinel body `@file:<abs-path>` resolves to the file's trimmed
2334    /// contents and reaches the `OutputStore` via the normal Final-append
2335    /// path — same 204 the inline path returns.
2336    #[tokio::test]
2337    async fn worker_submit_resolves_file_sentinel_under_work_dir() {
2338        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2339        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2340        let state = test_state(data_store.clone(), run_store);
2341        let task_id = StepId::new();
2342        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
2343
2344        let tmp = tempfile::tempdir().expect("tempdir");
2345        let work_dir = tmp.path().to_path_buf();
2346        seed_work_dir(
2347            &state,
2348            &task_id,
2349            1,
2350            work_dir.to_str().expect("work_dir utf-8"),
2351            Some(Value::Bool(true)),
2352        )
2353        .await;
2354
2355        let payload_path = work_dir.join("scout.md");
2356        let payload = "## Context Package (broad)\n\nlarge body content\n";
2357        tokio::fs::write(&payload_path, payload)
2358            .await
2359            .expect("write payload");
2360        let body = format!(
2361            "@file:{}",
2362            payload_path.to_str().expect("payload path utf-8")
2363        );
2364
2365        let status = worker_submit(
2366            State(state.clone()),
2367            bearer_headers(&handle),
2368            Query(SubmitQuery { ok: None }),
2369            axum::body::Bytes::from(body),
2370        )
2371        .await
2372        .expect("worker_submit sentinel");
2373        assert_eq!(status, StatusCode::NO_CONTENT);
2374
2375        // Final event lands with the file's trimmed contents on
2376        // `EngineState.output_store` (the in-memory tail
2377        // `submit_worker_result_trusted` writes to).
2378        let tid = task_id.clone();
2379        let value = state
2380            .engine
2381            .with_state("test.inspect_output_store", move |s| {
2382                s.output_store.get(&(tid.clone(), 1)).and_then(|evs| {
2383                    evs.iter().find_map(|ev| match ev {
2384                        OutputEvent::Final {
2385                            content: ContentRef::Inline { value },
2386                            ..
2387                        } => Some(value.clone()),
2388                        _ => None,
2389                    })
2390                })
2391            })
2392            .await
2393            .expect("with_state")
2394            .expect("Final event present");
2395        assert_eq!(value, Value::String(payload.trim_end().to_string()));
2396    }
2397
2398    /// A non-sentinel body is passed through byte-for-byte (pre-#42
2399    /// callers see zero behavior change).
2400    #[tokio::test]
2401    async fn worker_submit_passes_non_sentinel_body_unchanged() {
2402        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2403        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2404        let state = test_state(data_store.clone(), run_store);
2405        let task_id = StepId::new();
2406        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
2407        // No agent_ctx / work_dir seeded — the inline path must not
2408        // require one.
2409
2410        let status = worker_submit(
2411            State(state.clone()),
2412            bearer_headers(&handle),
2413            Query(SubmitQuery { ok: None }),
2414            axum::body::Bytes::from_static(b"DONE yes=1 maybe=0 no=0"),
2415        )
2416        .await
2417        .expect("worker_submit inline");
2418        assert_eq!(status, StatusCode::NO_CONTENT);
2419
2420        let tid = task_id.clone();
2421        let value = state
2422            .engine
2423            .with_state("test.inspect_output_store", move |s| {
2424                s.output_store.get(&(tid.clone(), 1)).and_then(|evs| {
2425                    evs.iter().find_map(|ev| match ev {
2426                        OutputEvent::Final {
2427                            content: ContentRef::Inline { value },
2428                            ..
2429                        } => Some(value.clone()),
2430                        _ => None,
2431                    })
2432                })
2433            })
2434            .await
2435            .expect("with_state")
2436            .expect("Final event present");
2437        assert_eq!(value, Value::String("DONE yes=1 maybe=0 no=0".to_string()));
2438    }
2439
2440    /// Sentinel with a path outside the task's `work_dir` (`..`-escape
2441    /// via a sibling tempdir) → `400`. `canonicalize` collapses the
2442    /// `..`, so a symlink pointing outside the allowlist would be caught
2443    /// by the same check.
2444    #[tokio::test]
2445    async fn worker_submit_rejects_sentinel_path_outside_work_dir() {
2446        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2447        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2448        let state = test_state(data_store, run_store);
2449        let task_id = StepId::new();
2450        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
2451
2452        let allowed = tempfile::tempdir().expect("allowed tempdir");
2453        let outside = tempfile::tempdir().expect("outside tempdir");
2454        seed_work_dir(
2455            &state,
2456            &task_id,
2457            1,
2458            allowed.path().to_str().expect("utf-8"),
2459            Some(Value::Bool(true)),
2460        )
2461        .await;
2462
2463        let outside_file = outside.path().join("leak.md");
2464        tokio::fs::write(&outside_file, b"outside content")
2465            .await
2466            .expect("write outside");
2467        let body = format!(
2468            "@file:{}",
2469            outside_file.to_str().expect("outside path utf-8")
2470        );
2471
2472        let err = worker_submit(
2473            State(state.clone()),
2474            bearer_headers(&handle),
2475            Query(SubmitQuery { ok: None }),
2476            axum::body::Bytes::from(body),
2477        )
2478        .await
2479        .expect_err("outside-work_dir sentinel must be rejected");
2480        assert_eq!(err.status, StatusCode::BAD_REQUEST);
2481    }
2482
2483    /// Sentinel pointing at a non-existent file → `404`.
2484    #[tokio::test]
2485    async fn worker_submit_rejects_sentinel_missing_file() {
2486        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2487        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2488        let state = test_state(data_store, run_store);
2489        let task_id = StepId::new();
2490        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
2491
2492        let tmp = tempfile::tempdir().expect("tempdir");
2493        seed_work_dir(
2494            &state,
2495            &task_id,
2496            1,
2497            tmp.path().to_str().expect("utf-8"),
2498            Some(Value::Bool(true)),
2499        )
2500        .await;
2501        let missing = tmp.path().join("does-not-exist.md");
2502        let body = format!("@file:{}", missing.to_str().expect("utf-8"));
2503
2504        let err = worker_submit(
2505            State(state.clone()),
2506            bearer_headers(&handle),
2507            Query(SubmitQuery { ok: None }),
2508            axum::body::Bytes::from(body),
2509        )
2510        .await
2511        .expect_err("missing-file sentinel must be rejected");
2512        assert_eq!(err.status, StatusCode::NOT_FOUND);
2513    }
2514
2515    /// Sentinel body with a relative path → `400` before any FS lookup.
2516    #[tokio::test]
2517    async fn worker_submit_rejects_sentinel_relative_path() {
2518        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2519        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2520        let state = test_state(data_store, run_store);
2521        let task_id = StepId::new();
2522        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
2523
2524        let err = worker_submit(
2525            State(state.clone()),
2526            bearer_headers(&handle),
2527            Query(SubmitQuery { ok: None }),
2528            axum::body::Bytes::from_static(b"@file:relative/path.md"),
2529        )
2530        .await
2531        .expect_err("relative-path sentinel must be rejected");
2532        assert_eq!(err.status, StatusCode::BAD_REQUEST);
2533    }
2534
2535    /// Sentinel body when the task has no `AgentContextView` (spawn
2536    /// didn't run through `AgentContextMiddleware`) → `400`. This is the
2537    /// documented pre-condition for sentinel use.
2538    #[tokio::test]
2539    async fn worker_submit_rejects_sentinel_without_agent_context_view() {
2540        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2541        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2542        let state = test_state(data_store, run_store);
2543        let task_id = StepId::new();
2544        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
2545        // No seed_work_dir — the agent_ctx map has no entry for this task.
2546
2547        let err = worker_submit(
2548            State(state.clone()),
2549            bearer_headers(&handle),
2550            Query(SubmitQuery { ok: None }),
2551            axum::body::Bytes::from_static(b"@file:/tmp/anywhere.md"),
2552        )
2553        .await
2554        .expect_err("missing AgentContextView must reject sentinel");
2555        assert_eq!(err.status, StatusCode::BAD_REQUEST);
2556    }
2557
2558    /// The same sentinel form works on `POST /v1/worker/artifact` — the
2559    /// artifact endpoint shares the resolver with `worker_submit`, so the
2560    /// resolved file contents land under the artifact's `name` key.
2561    #[tokio::test]
2562    async fn worker_artifact_resolves_file_sentinel_under_work_dir() {
2563        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2564        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2565        let state = test_state(data_store, run_store);
2566        let task_id = StepId::new();
2567        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
2568
2569        let tmp = tempfile::tempdir().expect("tempdir");
2570        seed_work_dir(
2571            &state,
2572            &task_id,
2573            1,
2574            tmp.path().to_str().expect("utf-8"),
2575            Some(Value::Bool(true)),
2576        )
2577        .await;
2578
2579        let payload_path = tmp.path().join("part.md");
2580        let payload = "artifact part body\n";
2581        tokio::fs::write(&payload_path, payload)
2582            .await
2583            .expect("write payload");
2584        let body = format!("@file:{}", payload_path.to_str().expect("utf-8"));
2585
2586        let status = worker_artifact(
2587            State(state.clone()),
2588            bearer_headers(&handle),
2589            Query(ArtifactQuery {
2590                name: "scout".to_string(),
2591            }),
2592            axum::body::Bytes::from(body),
2593        )
2594        .await
2595        .expect("worker_artifact sentinel");
2596        assert_eq!(status, StatusCode::NO_CONTENT);
2597    }
2598
2599    /// GH #43 — sentinel with `work_dir` seeded but no
2600    /// `allow_file_submit` opt-in → `400` (default-deny). The file exists
2601    /// and sits under `work_dir`, so the rejection is attributable to the
2602    /// missing opt-in alone.
2603    #[tokio::test]
2604    async fn worker_submit_rejects_sentinel_without_allow_flag() {
2605        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2606        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2607        let state = test_state(data_store, run_store);
2608        let task_id = StepId::new();
2609        let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
2610
2611        let tmp = tempfile::tempdir().expect("tempdir");
2612        seed_work_dir(
2613            &state,
2614            &task_id,
2615            1,
2616            tmp.path().to_str().expect("utf-8"),
2617            None,
2618        )
2619        .await;
2620
2621        let payload_path = tmp.path().join("out.md");
2622        tokio::fs::write(&payload_path, b"resolvable body")
2623            .await
2624            .expect("write payload");
2625        let body = format!("@file:{}", payload_path.to_str().expect("utf-8"));
2626
2627        let err = worker_submit(
2628            State(state.clone()),
2629            bearer_headers(&handle),
2630            Query(SubmitQuery { ok: None }),
2631            axum::body::Bytes::from(body),
2632        )
2633        .await
2634        .expect_err("missing opt-in must reject sentinel");
2635        assert_eq!(err.status, StatusCode::BAD_REQUEST);
2636        assert!(
2637            err.message.contains("not allowed"),
2638            "rejection must name the opt-in guard, got: {}",
2639            err.message
2640        );
2641    }
2642
2643    /// GH #43 — the opt-in is the strict boolean `true`: `Bool(false)`
2644    /// and the string `"true"` are both rejected with `400`.
2645    #[tokio::test]
2646    async fn worker_submit_rejects_sentinel_with_non_true_allow_values() {
2647        for allow in [Value::Bool(false), Value::String("true".to_string())] {
2648            let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2649            let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2650            let state = test_state(data_store, run_store);
2651            let task_id = StepId::new();
2652            let handle = seed_task_with_handle(&state, &task_id, "planner", 1, None).await;
2653
2654            let tmp = tempfile::tempdir().expect("tempdir");
2655            seed_work_dir(
2656                &state,
2657                &task_id,
2658                1,
2659                tmp.path().to_str().expect("utf-8"),
2660                Some(allow.clone()),
2661            )
2662            .await;
2663
2664            let payload_path = tmp.path().join("out.md");
2665            tokio::fs::write(&payload_path, b"resolvable body")
2666                .await
2667                .expect("write payload");
2668            let body = format!("@file:{}", payload_path.to_str().expect("utf-8"));
2669
2670            let err = worker_submit(
2671                State(state.clone()),
2672                bearer_headers(&handle),
2673                Query(SubmitQuery { ok: None }),
2674                axum::body::Bytes::from(body),
2675            )
2676            .await
2677            .expect_err("non-true opt-in value must reject sentinel");
2678            assert_eq!(err.status, StatusCode::BAD_REQUEST, "value: {allow:?}");
2679        }
2680    }
2681
2682    // ──────────────────────────────────────────────────────────────────
2683    // GH #50 (Subtask 2) — submit-time verdict contract gate, handler-
2684    // level unit coverage. The full process-boundary HTTP round trip
2685    // (Acceptance Criterion #7) lives in
2686    // `crates/mlua-swarm-server/tests/verdict_contract.rs`; these are the
2687    // fast in-process counterpart exercising `worker_submit` /
2688    // `worker_artifact` directly, same convention as the sentinel tests
2689    // above.
2690    // ──────────────────────────────────────────────────────────────────
2691
2692    fn body_verdict_contract(values: &[&str]) -> mlua_swarm_schema::VerdictContract {
2693        mlua_swarm_schema::VerdictContract {
2694            channel: VerdictChannel::Body,
2695            values: values.iter().map(|v| v.to_string()).collect(),
2696        }
2697    }
2698
2699    fn part_verdict_contract(values: &[&str]) -> mlua_swarm_schema::VerdictContract {
2700        mlua_swarm_schema::VerdictContract {
2701            channel: VerdictChannel::Part,
2702            values: values.iter().map(|v| v.to_string()).collect(),
2703        }
2704    }
2705
2706    /// A `channel: "body"` contract rejects a `worker_submit` body outside
2707    /// its declared `values` with `422`, echoing the expected token set.
2708    #[tokio::test]
2709    async fn worker_submit_rejects_body_outside_contract_values_with_422() {
2710        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2711        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2712        let state = test_state(data_store, run_store);
2713        let task_id = StepId::new();
2714        let handle = seed_task_with_handle(&state, &task_id, "gate", 1, None).await;
2715        state.engine.register_verdict_contracts(HashMap::from([(
2716            "gate".to_string(),
2717            body_verdict_contract(&["PASS", "BLOCKED"]),
2718        )]));
2719
2720        let err = worker_submit(
2721            State(state.clone()),
2722            bearer_headers(&handle),
2723            Query(SubmitQuery { ok: None }),
2724            axum::body::Bytes::from("UNKNOWN"),
2725        )
2726        .await
2727        .expect_err("value outside declared values must reject");
2728        assert_eq!(err.status, StatusCode::UNPROCESSABLE_ENTITY);
2729        assert!(
2730            err.message.contains("PASS") && err.message.contains("BLOCKED"),
2731            "rejection must echo the declared values, got: {}",
2732            err.message
2733        );
2734    }
2735
2736    /// The same contract accepts a body that IS a member of `values` —
2737    /// `204`, unaffected submit.
2738    #[tokio::test]
2739    async fn worker_submit_accepts_body_inside_contract_values() {
2740        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2741        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2742        let state = test_state(data_store, run_store);
2743        let task_id = StepId::new();
2744        let handle = seed_task_with_handle(&state, &task_id, "gate", 1, None).await;
2745        state.engine.register_verdict_contracts(HashMap::from([(
2746            "gate".to_string(),
2747            body_verdict_contract(&["PASS", "BLOCKED"]),
2748        )]));
2749
2750        let status = worker_submit(
2751            State(state.clone()),
2752            bearer_headers(&handle),
2753            Query(SubmitQuery { ok: None }),
2754            axum::body::Bytes::from("PASS"),
2755        )
2756        .await
2757        .expect("value inside declared values must succeed");
2758        assert_eq!(status, StatusCode::NO_CONTENT);
2759    }
2760
2761    /// Opt-in regression guard: an agent with NO declared verdict contract
2762    /// is entirely unaffected — `worker_submit` still returns `204` for an
2763    /// arbitrary body, exactly the pre-GH-#50 behavior.
2764    #[tokio::test]
2765    async fn worker_submit_without_a_declared_contract_is_unaffected() {
2766        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2767        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2768        let state = test_state(data_store, run_store);
2769        let task_id = StepId::new();
2770        // No `register_verdict_contracts` call — the agent declared no contract.
2771        let handle = seed_task_with_handle(&state, &task_id, "undeclared-agent", 1, None).await;
2772
2773        let status = worker_submit(
2774            State(state.clone()),
2775            bearer_headers(&handle),
2776            Query(SubmitQuery { ok: None }),
2777            axum::body::Bytes::from("anything at all, no contract to violate"),
2778        )
2779        .await
2780        .expect("no contract declared must never reject");
2781        assert_eq!(status, StatusCode::NO_CONTENT);
2782    }
2783
2784    /// A `channel: "part"` contract rejects a `worker_artifact?name=verdict`
2785    /// value outside `values` with `422`.
2786    #[tokio::test]
2787    async fn worker_artifact_verdict_part_rejects_value_outside_contract_with_422() {
2788        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2789        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2790        let state = test_state(data_store, run_store);
2791        let task_id = StepId::new();
2792        let handle = seed_task_with_handle(&state, &task_id, "gate", 1, None).await;
2793        state.engine.register_verdict_contracts(HashMap::from([(
2794            "gate".to_string(),
2795            part_verdict_contract(&["PASS", "BLOCKED"]),
2796        )]));
2797
2798        let err = worker_artifact(
2799            State(state.clone()),
2800            bearer_headers(&handle),
2801            Query(ArtifactQuery {
2802                name: "verdict".to_string(),
2803            }),
2804            axum::body::Bytes::from("UNKNOWN"),
2805        )
2806        .await
2807        .expect_err("value outside declared values must reject");
2808        assert_eq!(err.status, StatusCode::UNPROCESSABLE_ENTITY);
2809    }
2810
2811    /// A part named anything OTHER than `"verdict"` skips the gate
2812    /// entirely, even with a `channel: "part"` contract declared — `204`,
2813    /// existing pre-GH-#50 behavior unchanged.
2814    #[tokio::test]
2815    async fn worker_artifact_non_verdict_part_skips_the_gate() {
2816        let data_store: Arc<dyn OutputStore> = Arc::new(InMemoryOutputStore::new());
2817        let run_store: Arc<dyn RunStore> = Arc::new(InMemoryRunStore::new());
2818        let state = test_state(data_store, run_store);
2819        let task_id = StepId::new();
2820        let handle = seed_task_with_handle(&state, &task_id, "gate", 1, None).await;
2821        state.engine.register_verdict_contracts(HashMap::from([(
2822            "gate".to_string(),
2823            part_verdict_contract(&["PASS", "BLOCKED"]),
2824        )]));
2825
2826        let status = worker_artifact(
2827            State(state.clone()),
2828            bearer_headers(&handle),
2829            Query(ArtifactQuery {
2830                name: "notes".to_string(),
2831            }),
2832            axum::body::Bytes::from("anything at all"),
2833        )
2834        .await
2835        .expect("non-verdict part name must never be gated");
2836        assert_eq!(status, StatusCode::NO_CONTENT);
2837    }
2838}