phase3(daemon): 监工 tick(spawn worker)+ ingest(outbox→DB)+ reaper + reconcile 接真判活

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

src/daemon/index.ts

@@ -6,6 +6,7 @@ import { registerStatic } from '../api/static.js';
 import { loadConfig } from './config.js';
 import { startOrchestrator } from './orchestrator.js';
 import { startNotifier } from './notify.js';
+import { isWorkerAlive, heartbeatAgeMs } from '../executor/protocol.js';
 import { syncProject, todoJsonPath } from '../sync/todo-sync.js';
 
 /**
@@ -53,7 +54,15 @@ async function main(): Promise<void> {
   const cfg = loadConfig();
   const store = new Store(cfg.dbFile);
   const rec = store.reconcileDeps();   // 启动对账：ready↔blocked 按依赖纠正存量数据
-  const itr = store.reconcileInterrupted();   // 中断恢复：上次退出时在跑的任务重新入队
+  // 中断恢复（多进程执行）：用 worker_pid + 心跳真判活——worker 仍活则 re-adopt（daemon 续 ingest 其 outbox），
+  // 死则回收（failTaskAttempt 收尾 + 重试/needs_attention）。
+  const itr = store.reconcileInterrupted((run) =>
+    run !== null && isWorkerAlive({
+      pid: run.workerPid,
+      heartbeatAgeMs: heartbeatAgeMs(run.id),
+      startedAgeMs: Date.now() - Date.parse(run.startedAt),
+    }),
+  );
   const app = buildServer({ store, logger: true });
 
   registerStatic(app);          // Web 看板（web/ 静态文件）

src/daemon/ingest.ts

@@ -0,0 +1,114 @@
+import { Store } from '../store/index.js';
+import { readOutboxSince, type OutboxRecord } from '../executor/protocol.js';
+
+/**
+ * Outbox → DB 摄取（daemon 是唯一 DB 写者）。
+ *
+ * worker 全程不碰 DB，只把进度/结果追加进 runs/<runId>/outbox.ndjson。daemon 在每轮 tick 里把
+ * 这些记录读出来，按 seq 升序逐条映射成 DB 写（finishRun / setResult / transition / failTaskAttempt …）。
+ *
+ * 幂等靠 run.lastSeq 游标：只处理 seq>lastSeq 的记录，每处理一条就 setLastSeq(seq)。崩溃重启续读时
+ * 已落库的 seq 不会重做——故 result/failed/done 这类终态记录即便文件还在，也只会被消费一次。
+ */
+export interface IngestLogger {
+  info(msg: string): void;
+  error(msg: string): void;
+}
+
+/**
+ * 摄取单个 run 的新 outbox 记录（seq>run.lastSeq），按序落库。
+ * run 不存在则忽略（已被删/异常）。每条处理后推进 lastSeq 游标。
+ */
+export function ingestRun(store: Store, log: IngestLogger, runId: string): void {
+  const run = store.getRun(runId);
+  if (!run) return;
+  const taskId = run.taskId;
+  const recs = readOutboxSince(runId, run.lastSeq);
+  for (const rec of recs) {
+    try {
+      applyRecord(store, log, taskId, runId, rec);
+    } catch (e) {
+      // 单条映射失败不阻断后续 run 的摄取；但本条不推进游标，下轮重试。
+      log.error(`ingest run=${runId} seq=${rec.seq} 失败：${(e as Error).message}`);
+      return;
+    }
+    store.setLastSeq(runId, rec.seq);
+  }
+}
+
+/** 把一条 OutboxRecord 映射为 DB 写。 */
+function applyRecord(store: Store, log: IngestLogger, taskId: string, runId: string, rec: OutboxRecord): void {
+  switch (rec.type) {
+    case 'started':
+      // worker 启动自报；pid 已由 daemon spawn 时 setWorkerPid 写过，这里仅记日志。
+      log.info(`worker 启动 task=${taskId} run=${runId} pid=${rec.pid} model=${rec.model}`);
+      return;
+
+    case 'phase':
+      // 进度阶段，仅日志（看板事件可后续接）。
+      log.info(`task=${taskId} run=${runId} 阶段=${rec.phase}`);
+      return;
+
+    case 'failed': {
+      // 先把 executor run 收尾为 failed（带转录/会话），再走失败/重试策略。
+      // failTaskAttempt 见 run 已 ended（非 started）不会重复收尾，只做重试决策。
+      try {
+        store.finishRun(runId, 'failed', {
+          error: rec.error,
+          transcriptRef: rec.transcriptRef ?? undefined,
+          claudeSessionId: rec.sessionId ?? undefined,
+        });
+      } catch (e) {
+        log.error(`task=${taskId} run=${runId} 收尾 failed 出错（继续重试决策）：${(e as Error).message}`);
+      }
+      store.failTaskAttempt(taskId, runId, rec.error);
+      log.info(`task=${taskId} run=${runId} 执行失败：${rec.error}`);
+      return;
+    }
+
+    case 'result': {
+      // 成功终态：建 reviewer/security 两条复审 run + 收尾，写 result 四字段，转 exec_review，收尾 executor run。
+      const cr = store.startRun(taskId, 'reviewer', { worktree: rec.worktree, branch: rec.branch });
+      store.finishRun(cr.id, 'succeeded', { transcriptRef: rec.code.transcriptRef ?? undefined });
+
+      const sr = store.startRun(taskId, 'security', { worktree: rec.worktree, branch: rec.branch });
+      store.finishRun(sr.id, 'succeeded', { transcriptRef: rec.security.transcriptRef ?? undefined });
+
+      store.setResult(taskId, {
+        branch: rec.branch,
+        worktree: rec.worktree,
+        diffSummary: rec.diffSummary,
+        commits: rec.commits,
+        prUrl: null,
+        summary: rec.code.summary,
+        verdict: rec.code.verdict,
+        securitySummary: rec.security.summary,
+        securityVerdict: rec.security.verdict,
+        mergeTaskId: null,
+      });
+      store.transition(taskId, 'exec_review', { by: 'ingest', runId });
+      store.finishRun(runId, 'succeeded', {
+        transcriptRef: rec.executor.transcriptRef ?? undefined,
+        claudeSessionId: rec.executor.sessionId ?? undefined,
+      });
+      log.info(`task=${taskId} run=${runId} 执行完成 → exec_review（${rec.commits.length} commits）`);
+      return;
+    }
+
+    case 'done':
+      // worker 即将退出：终态标记。run 已由 result/failed 收尾，这里无须额外 DB 写。
+      log.info(`task=${taskId} run=${runId} worker 退出`);
+      return;
+  }
+}
+
+/**
+ * 摄取所有在途 run：对每个 executing 任务的最近一条 executor run 调 ingestRun。
+ * daemon 每轮 tick 调一次，把 worker 期间累积的 outbox 落库。
+ */
+export function ingestAll(store: Store, log: IngestLogger): void {
+  for (const { run } of store.executingWithLatestExecutorRun()) {
+    if (!run) continue;
+    ingestRun(store, log, run.id);
+  }
+}

src/daemon/orchestrator.ts

@@ -1,66 +1,107 @@
+import { spawn } from 'node:child_process';
+import { fileURLToPath } from 'node:url';
+import { dirname, join } from 'node:path';
 import { Store } from '../store/index.js';
-import type { Project, Task, ReviewVerdict } from '../model/types.js';
+import type { Project, Task } from '../model/types.js';
 import { rankByScore } from '../model/scoring.js';
-import { createWorktree, worktreeDiff, type WorktreeDiff, type WorktreeInfo } from '../executor/worktree.js';
+import { branchFor, worktreeDirFor } from '../executor/worktree.js';
-import { runTask, type RunnerFn } from '../executor/runner.js';
+import { writeJobSpec, isWorkerAlive, heartbeatAgeMs, type JobSpec } from '../executor/protocol.js';
-import { runVerify, type VerifyFn } from '../executor/verify.js';
-import { reviewCode, reviewSecurity, type ReviewerFn } from '../executor/reviewer.js';
 import { pickModel } from '../executor/models.js';
+import { ingestAll, type IngestLogger } from './ingest.js';
 
-/** 失败后默认最多自动重试次数（项目级 maxRetries 未设置时回退此值） */
+/** 失败后默认最多自动重试次数（项目级 maxRetries 未设置时回退此值）。失败/退避策略本体在 store.failTaskAttempt。 */
 export const DEFAULT_MAX_RETRIES = 2;
 
-/** 指数退避基础等待（第 n 次重试等待 BASE * 2^(n-1)，上限 MAX） */
-const BACKOFF_BASE_MS = 30_000;   // 30s
-const BACKOFF_MAX_MS = 10 * 60_000; // 10min
-
-/** 第 attempt 次重试（1-indexed）的等待毫秒 */
-export function computeBackoffMs(attempt: number): number {
-  return Math.min(BACKOFF_BASE_MS * (2 ** (attempt - 1)), BACKOFF_MAX_MS);
-}
-
 export interface OrchestratorLogger {
   info(msg: string): void;
   error(msg: string): void;
 }
 
-/** 依赖注入点：测试传 mock，生产用真实现 */
+/**
+ * worker 入口解析：返回 spawn 用的 [cmd, ...args]（不含 runId，由调用方追加）。
+ * - 默认：node <本文件同级 ../executor/worker.js>（编译后 dist 布局：daemon/ 与 executor/ 同级）。
+ * - env MAESTRO_WORKER_CMD 覆盖整条命令（空格分隔），供测试 / tsx 跑 .ts 入口用，
+ *   例：MAESTRO_WORKER_CMD="npx tsx src/executor/worker.ts"。
+ */
+export function workerEntry(): string[] {
+  const override = process.env.MAESTRO_WORKER_CMD?.trim();
+  if (override) return override.split(/\s+/);
+  const here = dirname(fileURLToPath(import.meta.url));
+  return ['node', join(here, '..', 'executor', 'worker.js')];
+}
+
+/**
+ * spawn worker 时传给子进程的最小 env：只保留 PATH/HOME/LANG 与 ANTHROPIC_ / CLAUDE_ 前缀变量，
+ * 不把 daemon 全量 env（含 MAESTRO_ 控制变量、无关密钥）漏给独立 worker 进程。
+ * MAESTRO_DATA_DIR 例外透传——worker 必须和 daemon 看同一个 runs/ 目录（job.json/outbox 路径一致）。
+ */
+export function workerEnv(src: NodeJS.ProcessEnv = process.env): NodeJS.ProcessEnv {
+  const out: NodeJS.ProcessEnv = {};
+  for (const k of ['PATH', 'HOME', 'LANG', 'MAESTRO_DATA_DIR', 'MAESTRO_WORKER_CMD'] as const) {
+    if (src[k] !== undefined) out[k] = src[k];
+  }
+  for (const [k, v] of Object.entries(src)) {
+    if (v !== undefined && (k.startsWith('ANTHROPIC_') || k.startsWith('CLAUDE_'))) out[k] = v;
+  }
+  return out;
+}
+
+/** 默认 spawnWorker：detached + 脱离 stdio + 收敛 env，unref 后自存活（daemon 退出不杀 worker）。返回子进程 pid。 */
+function defaultSpawnWorker(runId: string): number {
+  const [cmd, ...args] = workerEntry();
+  const child = spawn(cmd, [...args, runId], {
+    detached: true,
+    stdio: 'ignore',
+    env: workerEnv(),
+  });
+  child.unref();
+  if (child.pid === undefined) throw new Error('spawn worker 未返回 pid');
+  return child.pid;
+}
+
+/** 依赖注入点：生产用默认实现，测试传 mock（不真 spawn / 不真判活 / ingest no-op）。 */
 export interface OrchestratorDeps {
-  runner: RunnerFn;
+  /** spawn 一个独立 worker 进程跑该 run，返回其 pid。 */
-  reviewCode: ReviewerFn;        // code review（kind=reviewer 的 run）
+  spawnWorker: (runId: string) => number;
-  reviewSecurity: ReviewerFn;    // 安全审计（kind=security 的 run）
+  /** worker 是否存活（reaper 判死用），默认 protocol.isWorkerAlive。 */
-  verify: VerifyFn;
+  isWorkerAlive: typeof isWorkerAlive;
-  createWorktree: (repoPath: string, taskId: string, baseBranch: string) => Promise<WorktreeInfo>;
+  /** 把所有在途 run 的 outbox 摄取入库（daemon 唯一写者），默认 ingest.ingestAll。 */
-  worktreeDiff: (repoPath: string, dir: string, branch: string, baseBranch: string) => Promise<WorktreeDiff>;
+  ingestAll: (store: Store, log: IngestLogger) => void;
-  /** 当前时间戳（ms）；测试注入可控时钟（默认 Date.now） */
+  /** 写 worker 输入 job.json，默认 protocol.writeJobSpec（测试可 mock，避免落盘）。 */
+  writeJobSpec: (job: JobSpec) => void;
+  /** 当前时间戳（ms）；测试注入可控时钟（默认 Date.now）。 */
   nowMs: () => number;
 }
 
 export interface Orchestrator {
-  /** 跑一轮领取（同步领取 + 异步执行，不阻塞）。错误只记日志。 */
+  /** 跑一轮：先 ingest（outbox→DB）+ reaper（回收死 worker），再领新任务 spawn worker。错误只记日志、不抛。 */
   tick(): void;
-  /** 等待所有在途执行收尾（测试用） */
+  /** 仅领取/spawn 那一步（测试细粒度断言用，不含 ingest/reaper）。 */
-  drain(): Promise<void>;
+  claimTick(): void;
-  /** 在途任务 id（防同任务重复领取） */
+  /** ingest 所有在途 run（测试用）。 */
-  readonly inflight: ReadonlyMap<string, string>;
+  ingest(): void;
+  /** 回收死 worker（测试用）。 */
+  reap(): void;
 }
 
 /**
- * 编排器核心（与定时器解耦，便于测试）。
+ * 编排器（监工）：自身不跑 agent，只领任务 + spawn 独立 worker + 把 worker 的 outbox 摄取入库 + 回收死 worker。
- * 每轮对 status=active 且 autonomy≠manual 的项目：在途数 < concurrency 时领任务——
+ * 全部 DB 写都在 daemon 进程内（含 ingest 落的 run/result/transition），WS 推送照常经 store.subscribe 触发。
- * ready 叶子（deps 全 done；auto-easy 只领 easy）或 queued（重试/重启遗留）。
+ * 无内存在途态——并发闸读 store.countExecuting，退避读 task.nextEligibleAt，故跨 daemon 重启天然持久。
- * 成功 → setResult + exec_review；失败 → failed → 重试 ≤MAX_RETRIES 次 → needs_attention。
  */
 export function createOrchestrator(store: Store, log: OrchestratorLogger, deps: Partial<OrchestratorDeps> = {}): Orchestrator {
-  const d: OrchestratorDeps = { runner: runTask, reviewCode, reviewSecurity, verify: runVerify, createWorktree, worktreeDiff, nowMs: Date.now, ...deps };
+  const d: OrchestratorDeps = {
-  const inflight = new Map<string, string>();       // taskId → projectId
+    spawnWorker: defaultSpawnWorker,
-  const pending = new Set<Promise<void>>();
+    isWorkerAlive,
-  /** 退避表：taskId → 下次可领取时间戳（ms）；daemon 重启后清空，已有 queued 任务即刻可领 */
+    ingestAll,
-  const backoffUntil = new Map<string, number>();
+    writeJobSpec,
+    nowMs: Date.now,
+    ...deps,
+  };
 
   /**
    * 本项目可领取的任务：queued（重试/孤儿）+ ready 叶子且 deps 全 done；auto-easy 只挑 easy。
-   * 按调度分降序返回（score = 自身分 + 已完成依赖分 + 等待解锁的 blocked 任务分，见 model/scoring.ts）。
+   * 退避读持久化的 task.nextEligibleAt（早于它不领）。按调度分降序返回（见 model/scoring.ts）。
    */
   function claimable(project: Project): Array<{ task: Task; score: number }> {
     const nowMs = d.nowMs();
@@ -69,178 +110,113 @@ export function createOrchestrator(store: Store, log: OrchestratorLogger, deps:
     const parents = new Set(tasks.filter((t) => t.parentId).map((t) => t.parentId as string));
     const easyOnly = project.autonomy === 'auto-easy';
     const candidates = tasks.filter((t) => {
-      if (inflight.has(t.id)) return false;
       if (t.status !== 'ready' && t.status !== 'queued') return false;
       if (parents.has(t.id)) return false;                                  // 非叶子（容器）跳过
       if (easyOnly && t.complexity !== 'easy') return false;
-      const until = backoffUntil.get(t.id);
+      if (t.nextEligibleAt && nowMs < Date.parse(t.nextEligibleAt)) return false;  // 退避冷却中
-      if (until !== undefined && nowMs < until) return false;               // 退避冷却中
       return t.deps.every((dep) => byId.get(dep)?.status === 'done');
     });
     return rankByScore(candidates, tasks);
   }
 
-  /** 跑一个独立复审 run（kind=reviewer/security）。任何失败折叠为 summary=失败原因、verdict=null，不抛错、不挡任务。 */
+  /** 领取一个任务：建分支/worktree 路径 → executing → startRun → 写 job.json → spawn worker → 记 pid。 */
-  async function runOneReview(
+  function claimOne(project: Project, task: Task, score: number): void {
-    kind: 'reviewer' | 'security',
-    fn: ReviewerFn,
-    task: Task,
-    project: Project,
-    wt: WorktreeInfo,
-    executorReport: string,
-  ): Promise<{ summary: string | null; verdict: ReviewVerdict | null }> {
-    const label = kind === 'reviewer' ? 'code review' : '安全审计';
-    let reviewRunId: string | null = null;
-    try {
-      const review = store.startRun(task.id, kind, { worktree: wt.dir, branch: wt.branch });
-      reviewRunId = review.id;
-      log.info(`${label} 任务 ${task.id} run=${review.id} model=${pickModel(task, project, 'reviewer')}`);
-      const rv = await fn(task, project, wt, review.id, executorReport);
-      store.finishRun(review.id, 'succeeded', {
-        transcriptRef: rv.transcriptRef ?? undefined,
-        claudeSessionId: rv.sessionId ?? undefined,
-      });
-      log.info(`任务 ${task.id} ${label} 完成 verdict=${rv.verdict ?? '（未解析到）'}`);
-      return { summary: rv.summary, verdict: rv.verdict };
-    } catch (e) {
-      const reMsg = (e as Error).message;
-      if (reviewRunId) {
-        try { store.finishRun(reviewRunId, 'failed', { error: reMsg }); } catch { /* 收尾失败不影响主流程 */ }
-      }
-      log.error(`任务 ${task.id} ${label} 失败（不挡任务，照常进 exec_review）：${reMsg}`);
-      return { summary: `自动复审失败：${reMsg}`, verdict: null };
-    }
-  }
-
-  /** 单任务全流程：executing → worktree → run → verify → code review run → 安全审计 run（均失败不挡）→ setResult(四字段) → exec_review / failed(重试) */
-  async function executeTask(project: Project, task: Task): Promise<void> {
     let runId: string | null = null;
-    let runClosed = false;
     try {
+      if (task.status === 'ready') store.transition(task.id, 'queued', { by: 'orchestrator', score });
+
+      const branch = branchFor(task.id);
+      const dir = worktreeDirFor(project.repoPath, task.id);
+
       store.transition(task.id, 'executing', { by: 'orchestrator' });
-      const wt = await d.createWorktree(project.repoPath, task.id, project.defaultBranch);
+      const run = store.startRun(task.id, 'executor', { worktree: dir, branch });
-      const run = store.startRun(task.id, 'executor', { worktree: wt.dir, branch: wt.branch });
       runId = run.id;
-      log.info(`执行任务 ${task.id}「${task.title}」 run=${run.id} worktree=${wt.dir}`);
 
-      const rr = await d.runner(task, project, wt, run.id);
+      d.writeJobSpec({
-      if (!rr.ok) {
+        runId: run.id,
-        store.finishRun(run.id, 'failed', {
+        task: { ...task, status: 'executing' },
-          error: rr.error ?? '执行失败',
+        project,
-          transcriptRef: rr.transcriptRef ?? undefined,
+        worktreeDir: dir,
-          claudeSessionId: rr.sessionId ?? undefined,
+        branch,
-        });
-        runClosed = true;
-        throw new Error(rr.error ?? '执行失败');
-      }
-
-      const vr = await d.verify(project, wt.dir, run.id);
-      if (!vr.ok) {
-        store.finishRun(run.id, 'failed', {
-          error: vr.error ?? 'verify 失败',
-          transcriptRef: rr.transcriptRef ?? undefined,
-          claudeSessionId: rr.sessionId ?? undefined,
-        });
-        runClosed = true;
-        throw new Error(vr.error ?? 'verify 失败');
-      }
-
-      const diff = await d.worktreeDiff(project.repoPath, wt.dir, wt.branch, project.defaultBranch);
-
-      // 双复审（顺序）：code review run（kind=reviewer）→ 安全审计 run（kind=security）。任一失败不挡任务。
-      const code = await runOneReview('reviewer', d.reviewCode, task, project, wt, rr.finalText ?? '');
-      const sec = await runOneReview('security', d.reviewSecurity, task, project, wt, rr.finalText ?? '');
-
-      store.setResult(task.id, {
-        branch: wt.branch, worktree: wt.dir,
-        diffSummary: diff.diffSummary, commits: diff.commits, prUrl: null,
-        summary: code.summary, verdict: code.verdict,
-        securitySummary: sec.summary, securityVerdict: sec.verdict,
-        mergeTaskId: null,
       });
-      store.transition(task.id, 'exec_review', { by: 'orchestrator', runId: run.id });
+
-      store.finishRun(run.id, 'succeeded', {
+      log.info(`领取任务 ${task.id}「${task.title}」score=${score} run=${run.id} model=${pickModel(task, project, 'executor')}`);
-        transcriptRef: rr.transcriptRef ?? undefined,
+      const pid = d.spawnWorker(run.id);
-        claudeSessionId: rr.sessionId ?? undefined,
+      store.setWorkerPid(run.id, pid);
-      });
+      log.info(`任务 ${task.id} 已起 worker pid=${pid} worktree=${dir}`);
-      log.info(`任务 ${task.id} 执行完成 → exec_review（${diff.commits.length} commits）`);
     } catch (e) {
       const msg = (e as Error).message;
-      log.error(`任务 ${task.id} 执行失败：${msg}`);
+      log.error(`任务 ${task.id} 领取/spawn 失败：${msg}`);
       try {
-        if (runId && !runClosed) {
+        store.failTaskAttempt(task.id, runId, msg);
-          store.finishRun(runId, 'failed', { error: msg });                 // worktree 创建后抛错（diff 等）时收尾
-        } else if (!runId) {
-          // run 还没建（如 createWorktree 失败）：补记一条 failed run，保证重试计数不漏
-          const r = store.startRun(task.id, 'executor');
-          store.finishRun(r.id, 'failed', { error: msg });
-        }
-        store.transition(task.id, 'failed', { by: 'orchestrator', error: msg });
-        const allFailed = store.listRuns(task.id).filter((r) => r.kind === 'executor' && r.status === 'failed').length;
-        // 相对于上次手动重投基线的净失败次数（retryBaseline 在 requeueTask 时设置）
-        const netFailed = Math.max(0, allFailed - task.retryBaseline);
-        const priorNetFailed = Math.max(0, netFailed - 1);                  // 不含本次
-        const maxRetries = project.maxRetries;
-        if (priorNetFailed < maxRetries) {
-          const attempt = priorNetFailed + 1;                               // 本次是第几次重试
-          const backoffMs = computeBackoffMs(attempt);
-          backoffUntil.set(task.id, d.nowMs() + backoffMs);
-          store.transition(task.id, 'queued', { by: 'orchestrator', retry: attempt, backoffMs });
-          log.info(`任务 ${task.id} 重新入队（第 ${attempt} 次重试，退避 ${backoffMs}ms 后可领）`);
-        } else {
-          backoffUntil.delete(task.id);
-          store.transition(task.id, 'needs_attention', { by: 'orchestrator', allFailed, netFailed });
-          log.error(`任务 ${task.id} 净失败 ${netFailed} 次（上限 ${maxRetries}）→ needs_attention`);
-        }
       } catch (e2) {
         log.error(`任务 ${task.id} 失败收尾出错：${(e2 as Error).message}`);
       }
     }
   }
 
-  function tick(): void {
+  /** 领取轮：对每个 active 且 autonomy≠manual 的项目，在 countExecuting<concurrency 时按 claimable 领新任务并 spawn worker。 */
+  function claimTick(): void {
     try {
       for (const p of store.listProjects()) {
         if (p.status !== 'active' || p.autonomy === 'manual') continue;
-        let active = 0;
+        let active = store.countExecuting(p.id);
-        for (const pid of inflight.values()) if (pid === p.id) active++;
         if (active >= p.concurrency) continue;
 
         for (const { task, score } of claimable(p)) {
           if (active >= p.concurrency) break;
-          try {
+          claimOne(p, task, score);
-            if (task.status === 'ready') store.transition(task.id, 'queued', { by: 'orchestrator', score });
-          } catch (e) {
-            log.error(`任务 ${task.id} 入队失败：${(e as Error).message}`);
-            continue;
-          }
-          log.info(`领取任务 ${task.id}「${task.title}」score=${score} model=${pickModel(task, p, 'executor')}`);
-          inflight.set(task.id, p.id);
           active++;
-          const job: Promise<void> = executeTask(p, { ...task, status: 'queued' })
-            .catch((e) => log.error(`任务 ${task.id} 执行异常：${(e as Error).message}`))
-            .finally(() => {
-              inflight.delete(task.id);
-              pending.delete(job);
-            });
-          pending.add(job);
         }
       }
     } catch (e) {
-      log.error(`编排器轮询失败：${(e as Error).message}`);
+      log.error(`编排器领取轮失败：${(e as Error).message}`);
     }
   }
 
-  async function drain(): Promise<void> {
+  /** ingest 所有在途 run 的 outbox（daemon 唯一 DB 写者把 worker 进度/结果落库）。 */
-    while (pending.size > 0) await Promise.allSettled([...pending]);
+  function ingest(): void {
+    try {
+      d.ingestAll(store, log);
+    } catch (e) {
+      log.error(`ingest 轮失败：${(e as Error).message}`);
+    }
   }
 
-  return { tick, drain, inflight };
+  /** 回收：对每个 executing 任务判活，死 worker → failTaskAttempt（收尾 + 重试/needs_attention）。 */
+  function reap(): void {
+    try {
+      for (const { task, run } of store.executingWithLatestExecutorRun()) {
+        const alive = run !== null && d.isWorkerAlive({
+          pid: run.workerPid,
+          heartbeatAgeMs: heartbeatAgeMs(run.id, d.nowMs()),
+          startedAgeMs: d.nowMs() - Date.parse(run.startedAt),
+        });
+        if (alive) continue;
+        log.error(`任务 ${task.id} worker 异常退出（run=${run?.id ?? '无'}）→ 回收`);
+        try {
+          store.failTaskAttempt(task.id, run?.id ?? null, 'worker 异常退出');
+        } catch (e) {
+          log.error(`任务 ${task.id} 回收收尾出错：${(e as Error).message}`);
+        }
+      }
+    } catch (e) {
+      log.error(`编排器回收轮失败：${(e as Error).message}`);
+    }
+  }
+
+  function tick(): void {
+    ingest();      // 先把 worker 进度/结果落库（可能把 executing→exec_review，腾出并发槽）
+    reap();        // 再回收死 worker（可能把 executing→queued/needs_attention，腾出并发槽）
+    claimTick();   // 最后领新任务
+  }
+
+  return { tick, claimTick, ingest, reap };
 }
 
 /**
  * 接线入口：MAESTRO_ORCH_INTERVAL（秒）控制轮询间隔，默认 15，0=关闭。
- * 返回 timer 供 shutdown 时 clearInterval。
+ * 每轮依次 ingest → reaper → 领取（见 tick）。返回 timer 供 shutdown 时 clearInterval。
  */
 export function startOrchestrator(
   store: Store,
@@ -255,6 +231,6 @@ export function startOrchestrator(
   const orch = createOrchestrator(store, app.log, deps);
   const timer = setInterval(() => orch.tick(), intervalSec * 1000);
   timer.unref();
-  app.log.info(`编排器已启用：每 ${intervalSec}s 一轮领取（autonomy≠manual 的 active 项目）`);
+  app.log.info(`编排器已启用：每 ${intervalSec}s 一轮（ingest→回收→领取，autonomy≠manual 的 active 项目）`);
   return timer;
 }

test/ingest.test.ts

@@ -0,0 +1,205 @@
+import { test } from 'node:test';
+import assert from 'node:assert/strict';
+import { mkdtempSync, rmSync } from 'node:fs';
+import { tmpdir } from 'node:os';
+import { join } from 'node:path';
+import { Store } from '../src/store/index.js';
+import { ingestRun } from '../src/daemon/ingest.js';
+import { appendOutbox, type OutboxPayload } from '../src/executor/protocol.js';
+import { branchFor, worktreeDirFor } from '../src/executor/worktree.js';
+
+const noopLog = { info: (): void => undefined, error: (): void => undefined };
+
+/**
+ * 隔离的 MAESTRO_DATA_DIR（outbox 落在临时目录），跑回调，结束清理。
+ * protocol.ts 的 runDir/outboxPath 在调用时读 env，故进 cb 前设好、cb 内的 append/ingest 都命中临时目录。
+ */
+function withTmpDataDir(cb: () => void): void {
+  const dir = mkdtempSync(join(tmpdir(), 'maestro-ingest-'));
+  const prev = process.env.MAESTRO_DATA_DIR;
+  process.env.MAESTRO_DATA_DIR = dir;
+  try {
+    cb();
+  } finally {
+    if (prev === undefined) delete process.env.MAESTRO_DATA_DIR;
+    else process.env.MAESTRO_DATA_DIR = prev;
+    rmSync(dir, { recursive: true, force: true });
+  }
+}
+
+/** 真 Store + 项目 + easy 任务推进到 executing，并建一条 executor run（模拟 daemon 已领取、worker 已 spawn）。 */
+function setupExecuting(opts: { autonomy?: 'auto-easy' | 'auto-approved'; maxRetries?: number } = {}): {
+  store: Store; projectId: string; taskId: string; runId: string; dir: string; branch: string;
+} {
+  const store = new Store(':memory:');
+  const p = store.createProject({
+    name: 'ingest', repoPath: '/tmp/ingest-repo-' + Math.random(),
+    autonomy: opts.autonomy ?? 'auto-easy', maxRetries: opts.maxRetries,
+  });
+  const t = store.createTask({ projectId: p.id, title: 'tweak', complexity: 'easy' });
+  store.setOperations(t.id, '在 README.md 追加一行');
+  const branch = branchFor(t.id);
+  const dir = worktreeDirFor(p.repoPath, t.id);
+  store.transition(t.id, 'queued', { by: 'test' });
+  store.transition(t.id, 'executing', { by: 'test' });
+  const run = store.startRun(t.id, 'executor', { worktree: dir, branch });
+  store.setWorkerPid(run.id, 4242);
+  return { store, projectId: p.id, taskId: t.id, runId: run.id, dir, branch };
+}
+
+function resultPayload(branch: string, worktree: string): OutboxPayload {
+  return {
+    type: 'result',
+    branch, worktree,
+    diffSummary: ' README.md | 1 +',
+    commits: ['abc1234 hello maestro'],
+    executor: { transcriptRef: '/tmp/exec.jsonl', sessionId: 'sess-exec-1' },
+    code: { summary: '## 做了什么\nmock 复审通过', verdict: 'approve', transcriptRef: '/tmp/code.jsonl' },
+    security: { summary: '## 安全审计\nmock 审计通过', verdict: 'approve', transcriptRef: '/tmp/sec.jsonl' },
+  };
+}
+
+test('ingest result：task→exec_review，setResult 四字段正确，reviewer/security/executor 各 1 条 succeeded', () => {
+  withTmpDataDir(() => {
+    const { store, taskId, runId, dir, branch } = setupExecuting();
+
+    appendOutbox(runId, { type: 'started', pid: 4242, worktree: dir, branch, model: 'claude-sonnet-4-6' });
+    appendOutbox(runId, { type: 'phase', phase: 'executing' });
+    appendOutbox(runId, resultPayload(branch, dir));
+    appendOutbox(runId, { type: 'done' });
+
+    ingestRun(store, noopLog, runId);
+
+    const done = store.getTask(taskId)!;
+    assert.equal(done.status, 'exec_review');
+    assert.deepEqual(done.result, {
+      branch,
+      worktree: dir,
+      diffSummary: ' README.md | 1 +',
+      commits: ['abc1234 hello maestro'],
+      prUrl: null,
+      summary: '## 做了什么\nmock 复审通过',
+      verdict: 'approve',
+      securitySummary: '## 安全审计\nmock 审计通过',
+      securityVerdict: 'approve',
+      mergeTaskId: null,
+    });
+
+    const runs = store.listRuns(taskId);
+    const executor = runs.find((r) => r.kind === 'executor')!;
+    assert.equal(executor.status, 'succeeded');
+    assert.equal(executor.transcriptRef, '/tmp/exec.jsonl');
+    assert.equal(executor.claudeSessionId, 'sess-exec-1');
+    const reviewer = runs.filter((r) => r.kind === 'reviewer');
+    assert.equal(reviewer.length, 1);
+    assert.equal(reviewer[0].status, 'succeeded');
+    assert.equal(reviewer[0].transcriptRef, '/tmp/code.jsonl');
+    const security = runs.filter((r) => r.kind === 'security');
+    assert.equal(security.length, 1);
+    assert.equal(security[0].status, 'succeeded');
+    assert.equal(security[0].transcriptRef, '/tmp/sec.jsonl');
+
+    // 游标推进到末条（done.seq=4）
+    assert.equal(store.getRun(runId)!.lastSeq, 4);
+    store.close();
+  });
+});
+
+test('ingest result：任一 verdict=reject 也照常落进 result（裁决归用户）', () => {
+  withTmpDataDir(() => {
+    const { store, taskId, runId, dir, branch } = setupExecuting();
+    appendOutbox(runId, {
+      ...resultPayload(branch, dir),
+      code: { summary: '发现问题', verdict: 'reject', transcriptRef: '/tmp/code.jsonl' },
+      security: { summary: '发现密钥泄露', verdict: 'reject', transcriptRef: '/tmp/sec.jsonl' },
+    } as OutboxPayload);
+    ingestRun(store, noopLog, runId);
+
+    const done = store.getTask(taskId)!;
+    assert.equal(done.status, 'exec_review');
+    assert.equal(done.result!.verdict, 'reject');
+    assert.equal(done.result!.summary, '发现问题');
+    assert.equal(done.result!.securityVerdict, 'reject');
+    assert.equal(done.result!.securitySummary, '发现密钥泄露');
+    store.close();
+  });
+});
+
+test('ingest failed：首次失败 → task 转 queued + 持久化退避（nextEligibleAt 有值）+ executor run failed', () => {
+  withTmpDataDir(() => {
+    const { store, taskId, runId } = setupExecuting();
+
+    appendOutbox(runId, { type: 'started', pid: 4242, worktree: '/x', branch: 'b', model: 'm' });
+    appendOutbox(runId, { type: 'failed', error: 'boom #1', transcriptRef: '/tmp/f.jsonl', sessionId: 'sess-f-1' });
+    appendOutbox(runId, { type: 'done' });
+
+    ingestRun(store, noopLog, runId);
+
+    const t = store.getTask(taskId)!;
+    assert.equal(t.status, 'queued');                       // 第一次失败 → 重入队（默认 maxRetries=2）
+    assert.ok(t.nextEligibleAt, 'nextEligibleAt 应被持久化（退避）');
+    assert.ok(Date.parse(t.nextEligibleAt!) > Date.now() - 1000, 'nextEligibleAt 在未来');
+
+    const executor = store.listRuns(taskId).find((r) => r.kind === 'executor')!;
+    assert.equal(executor.status, 'failed');
+    assert.match(executor.error ?? '', /boom #1/);
+    assert.equal(executor.transcriptRef, '/tmp/f.jsonl');
+    assert.equal(executor.claudeSessionId, 'sess-f-1');
+    store.close();
+  });
+});
+
+test('ingest failed：maxRetries=0 → 首次失败直接 needs_attention', () => {
+  withTmpDataDir(() => {
+    const { store, taskId, runId } = setupExecuting({ maxRetries: 0 });
+    appendOutbox(runId, { type: 'failed', error: 'boom', transcriptRef: null, sessionId: null });
+    ingestRun(store, noopLog, runId);
+    assert.equal(store.getTask(taskId)!.status, 'needs_attention');
+    store.close();
+  });
+});
+
+test('幂等：同一批 outbox ingest 两次，DB 不重复变更（lastSeq 生效）', () => {
+  withTmpDataDir(() => {
+    const { store, taskId, runId, dir, branch } = setupExecuting();
+    appendOutbox(runId, { type: 'started', pid: 4242, worktree: dir, branch, model: 'm' });
+    appendOutbox(runId, resultPayload(branch, dir));
+    appendOutbox(runId, { type: 'done' });
+
+    ingestRun(store, noopLog, runId);
+    const after1 = store.getTask(taskId)!;
+    const runs1 = store.listRuns(taskId);
+    assert.equal(after1.status, 'exec_review');
+    assert.equal(runs1.filter((r) => r.kind === 'reviewer').length, 1);
+    assert.equal(runs1.filter((r) => r.kind === 'security').length, 1);
+    const lastSeq1 = store.getRun(runId)!.lastSeq;
+
+    // 第二次 ingest：seq 全部 ≤ lastSeq → 不处理任何记录
+    ingestRun(store, noopLog, runId);
+    const runs2 = store.listRuns(taskId);
+    assert.equal(runs2.length, runs1.length, '复审 run 不应重复创建');
+    assert.equal(runs2.filter((r) => r.kind === 'reviewer').length, 1);
+    assert.equal(runs2.filter((r) => r.kind === 'security').length, 1);
+    assert.equal(store.getRun(runId)!.lastSeq, lastSeq1, 'lastSeq 不变');
+    assert.equal(store.getTask(taskId)!.status, 'exec_review');
+    store.close();
+  });
+});
+
+test('ingest 续读：先 ingest 半截（started/phase），再追加 result，第二次 ingest 完成 exec_review', () => {
+  withTmpDataDir(() => {
+    const { store, taskId, runId, dir, branch } = setupExecuting();
+    appendOutbox(runId, { type: 'started', pid: 4242, worktree: dir, branch, model: 'm' });
+    appendOutbox(runId, { type: 'phase', phase: 'executing' });
+    ingestRun(store, noopLog, runId);
+    assert.equal(store.getTask(taskId)!.status, 'executing');   // 还没到 result，仍 executing
+    assert.equal(store.getRun(runId)!.lastSeq, 2);
+
+    appendOutbox(runId, resultPayload(branch, dir));
+    appendOutbox(runId, { type: 'done' });
+    ingestRun(store, noopLog, runId);
+    assert.equal(store.getTask(taskId)!.status, 'exec_review');
+    assert.equal(store.getRun(runId)!.lastSeq, 4);
+    store.close();
+  });
+});

test/orchestrator.test.ts

@@ -1,53 +1,44 @@
 import { test } from 'node:test';
 import assert from 'node:assert/strict';
 import { Store } from '../src/store/index.js';
-import { createOrchestrator, DEFAULT_MAX_RETRIES, computeBackoffMs, type OrchestratorDeps } from '../src/daemon/orchestrator.js';
+import {
-import type { RunnerResult } from '../src/executor/runner.js';
+  createOrchestrator, workerEntry, workerEnv, DEFAULT_MAX_RETRIES,
-import type { ReviewResult } from '../src/executor/reviewer.js';
+  type OrchestratorDeps,
+} from '../src/daemon/orchestrator.js';
+import type { JobSpec } from '../src/executor/protocol.js';
 import type { Autonomy } from '../src/model/types.js';
 
 const noopLog = { info: (): void => undefined, error: (): void => undefined };
 
-function deferred<T>(): { promise: Promise<T>; resolve: (v: T) => void } {
+/** 可前进的测试时钟。 */
-  let resolve!: (v: T) => void;
+function makeClock(initialMs = Date.now()): { nowMs: () => number; advance: (ms: number) => void } {
-  const promise = new Promise<T>((r) => { resolve = r; });
+  let t = initialMs;
-  return { promise, resolve };
+  return { nowMs: () => t, advance: (ms) => { t += ms; } };
 }
 
-const okRun: RunnerResult = { ok: true, transcriptRef: '/tmp/fake.jsonl', sessionId: 'sess-mock-1', finalText: '执行自述：改了 README' };
+interface MockState {
+  spawned: string[];          // 被 spawn 的 runId 列表
+  jobs: JobSpec[];            // 被 writeJobSpec 的 job 列表
+  ingestCalls: number;
+  alive: boolean;             // isWorkerAlive 的返回（reaper 用）
+  nextPid: number;
+}
 
-const okReview: ReviewResult = {
+/** 全 mock 依赖：不真 spawn / 不真判活 / ingest no-op / job 不落盘。可按用例覆盖。 */
-  summary: '## 做了什么\nmock 复审通过',
+function mockDeps(state: MockState, overrides: Partial<OrchestratorDeps> = {}): { deps: OrchestratorDeps } {
-  verdict: 'approve',
+  const deps: OrchestratorDeps = {
-  transcriptRef: '/tmp/fake-review.jsonl',
+    spawnWorker: (runId: string): number => { state.spawned.push(runId); return state.nextPid++; },
-  sessionId: 'sess-review-1',
+    isWorkerAlive: () => state.alive,
-};
+    ingestAll: () => { state.ingestCalls++; },
-
+    writeJobSpec: (job: JobSpec) => { state.jobs.push(job); },
-const okSecurity: ReviewResult = {
-  summary: '## 安全审计\nmock 审计通过',
-  verdict: 'approve',
-  transcriptRef: '/tmp/fake-security.jsonl',
-  sessionId: 'sess-security-1',
-};
-
-/** 全 mock 依赖（不真起 CC、不动 git）：可按用例覆盖 */
-function mockDeps(overrides: Partial<OrchestratorDeps> = {}): OrchestratorDeps {
-  return {
-    createWorktree: async (_repo, taskId) => ({ dir: `/tmp/fake-wt/${taskId}`, branch: `maestro/${taskId}` }),
-    worktreeDiff: async () => ({ diffSummary: ' README.md | 1 +', commits: ['abc1234 hello maestro'] }),
-    verify: async () => ({ ok: true, exitCode: 0, logRef: null }),
-    runner: async () => okRun,
-    reviewCode: async () => okReview,
-    reviewSecurity: async () => okSecurity,
     nowMs: Date.now,
     ...overrides,
   };
+  return { deps };
 }
 
-/** 可前进的测试时钟（用于退避相关测试） */
+function freshState(): MockState {
-function makeClock(initialMs = 0): { nowMs: () => number; advance: (ms: number) => void } {
+  return { spawned: [], jobs: [], ingestCalls: 0, alive: true, nextPid: 1000 };
-  let t = initialMs;
-  return { nowMs: () => t, advance: (ms) => { t += ms; } };
 }
 
 function setup(autonomy: Autonomy, concurrency = 1): { store: Store; projectId: string } {
@@ -58,25 +49,139 @@ function setup(autonomy: Autonomy, concurrency = 1): { store: Store; projectId:
   return { store, projectId: p.id };
 }
 
-async function settle(): Promise<void> {
+// ───────────────────────── workerEntry / workerEnv（可读测的小函数）─────────────────────────
-  await new Promise((r) => setImmediate(r));
-}
 
-test('autonomy=manual：编排器不领取任何任务', async () => {
+test('workerEntry：默认指向 ../executor/worker.js（node 运行）', () => {
-  const { store, projectId } = setup('manual');
+  const prev = process.env.MAESTRO_WORKER_CMD;
-  const t = store.createTask({ projectId, title: 'easy task', complexity: 'easy' });
+  delete process.env.MAESTRO_WORKER_CMD;
-  let calls = 0;
+  const [cmd, scriptPath] = workerEntry();
-  const orch = createOrchestrator(store, noopLog, mockDeps({
+  assert.equal(cmd, 'node');
-    runner: async () => { calls++; return okRun; },
+  assert.match(scriptPath, /executor[/\\]worker\.js$/);
-  }));
+  if (prev !== undefined) process.env.MAESTRO_WORKER_CMD = prev;
+});
+
+test('workerEntry：MAESTRO_WORKER_CMD 覆盖整条命令（空格分隔）', () => {
+  const prev = process.env.MAESTRO_WORKER_CMD;
+  process.env.MAESTRO_WORKER_CMD = 'npx tsx src/executor/worker.ts';
+  assert.deepEqual(workerEntry(), ['npx', 'tsx', 'src/executor/worker.ts']);
+  if (prev === undefined) delete process.env.MAESTRO_WORKER_CMD;
+  else process.env.MAESTRO_WORKER_CMD = prev;
+});
+
+test('workerEnv：只透传 PATH/HOME/LANG + ANTHROPIC_/CLAUDE_ 前缀；丢弃无关与 MAESTRO_ 控制变量', () => {
+  const env = workerEnv({
+    PATH: '/usr/bin', HOME: '/home/u', LANG: 'en_US.UTF-8',
+    ANTHROPIC_API_KEY: 'sk-x', CLAUDE_CODE_FOO: 'y',
+    MAESTRO_ORCH_INTERVAL: '15', MAESTRO_DATA_DIR: '/data',
+    SOME_SECRET: 'leak', AWS_SECRET_ACCESS_KEY: 'nope',
+  });
+  assert.equal(env.PATH, '/usr/bin');
+  assert.equal(env.HOME, '/home/u');
+  assert.equal(env.LANG, 'en_US.UTF-8');
+  assert.equal(env.ANTHROPIC_API_KEY, 'sk-x');
+  assert.equal(env.CLAUDE_CODE_FOO, 'y');
+  assert.equal(env.MAESTRO_DATA_DIR, '/data');     // 例外：worker 须看同一 runs/ 目录
+  assert.equal(env.MAESTRO_ORCH_INTERVAL, undefined, 'daemon 控制变量不下传');
+  assert.equal(env.SOME_SECRET, undefined, '无关变量不下传');
+  assert.equal(env.AWS_SECRET_ACCESS_KEY, undefined, '无关密钥不下传');
+});
+
+// ───────────────────────── tick 顺序：ingest + reaper 先于领取 ─────────────────────────
+
+test('tick：每轮先 ingest 再 reaper 再领取', () => {
+  const { store, projectId } = setup('auto-easy');
+  store.createTask({ projectId, title: 'a', complexity: 'easy' });
+  const state = freshState();
+  const { deps } = mockDeps(state);
+  const orch = createOrchestrator(store, noopLog, deps);
   orch.tick();
-  await orch.drain();
+  assert.equal(state.ingestCalls, 1, '每轮 ingest 一次');
-  assert.equal(calls, 0);
+  assert.equal(state.spawned.length, 1, '领取并 spawn 一个');
+});
+
+// ───────────────────────── 领取（监工）─────────────────────────
+
+test('autonomy=manual：不领取', () => {
+  const { store, projectId } = setup('manual');
+  const t = store.createTask({ projectId, title: 'easy', complexity: 'easy' });
+  const state = freshState();
+  const { deps } = mockDeps(state);
+  createOrchestrator(store, noopLog, deps).tick();
+  assert.equal(state.spawned.length, 0);
   assert.equal(store.getTask(t.id)!.status, 'ready');
   store.close();
 });
 
-test('autonomy=auto-easy：只领 easy，medium ready 不动', async () => {
+test('project paused：不领取', () => {
+  const { store, projectId } = setup('auto-approved');
+  const t = store.createTask({ projectId, title: 'x', complexity: 'easy' });
+  store.patchProject(projectId, { status: 'paused' });
+  const state = freshState();
+  const { deps } = mockDeps(state);
+  createOrchestrator(store, noopLog, deps).tick();
+  assert.equal(state.spawned.length, 0);
+  assert.equal(store.getTask(t.id)!.status, 'ready');
+  store.close();
+});
+
+test('并发闸 concurrency=1：一轮只领一个；领取后 executing + 建 executor run + setWorkerPid + writeJobSpec', () => {
+  const { store, projectId } = setup('auto-approved', 1);
+  const t1 = store.createTask({ projectId, title: 'a', complexity: 'easy' });
+  const t2 = store.createTask({ projectId, title: 'b', complexity: 'easy' });
+  const state = freshState();
+  const { deps } = mockDeps(state);
+  const orch = createOrchestrator(store, noopLog, deps);
+
+  orch.claimTick();
+  // 并发=1：只领一个
+  assert.equal(state.spawned.length, 1);
+  assert.equal(state.jobs.length, 1);
+
+  // 被领的那个进 executing，另一个仍 ready
+  const a = store.getTask(t1.id)!;
+  const b = store.getTask(t2.id)!;
+  const executing = a.status === 'executing' ? a : b;
+  const stillReady = a.status === 'executing' ? b : a;
+  assert.equal(executing.status, 'executing');
+  assert.equal(stillReady.status, 'ready');
+
+  // 建了 executor run + setWorkerPid + writeJobSpec 内容正确
+  const runs = store.listRuns(executing.id);
+  assert.equal(runs.length, 1);
+  const run = runs[0];
+  assert.equal(run.kind, 'executor');
+  assert.equal(run.status, 'started');
+  assert.equal(run.branch, `maestro/${executing.id}`);
+  assert.ok(run.workerPid, 'setWorkerPid 已写 pid');
+  assert.equal(state.spawned[0], run.id, 'spawn 的 runId == 新建 executor run');
+
+  const job = state.jobs[0];
+  assert.equal(job.runId, run.id);
+  assert.equal(job.task.id, executing.id);
+  assert.equal(job.task.status, 'executing');
+  assert.equal(job.branch, `maestro/${executing.id}`);
+  assert.equal(job.worktreeDir, run.worktree);
+
+  // 槽位占满 → 下一轮不再领（仍只 spawn 过一个）
+  orch.claimTick();
+  assert.equal(state.spawned.length, 1, '并发=1，executing 占满 → 不再领');
+  store.close();
+});
+
+test('并发 concurrency=2：一轮领满两个', () => {
+  const { store, projectId } = setup('auto-approved', 2);
+  store.createTask({ projectId, title: 'a', complexity: 'easy' });
+  store.createTask({ projectId, title: 'b', complexity: 'easy' });
+  store.createTask({ projectId, title: 'c', complexity: 'easy' });
+  const state = freshState();
+  const { deps } = mockDeps(state);
+  createOrchestrator(store, noopLog, deps).claimTick();
+  assert.equal(state.spawned.length, 2, '并发=2 一轮领两个');
+  assert.equal(store.listTasks(projectId).filter((t) => t.status === 'executing').length, 2);
+  store.close();
+});
+
+test('auto-easy：只领 easy，medium ready 不动', () => {
   const { store, projectId } = setup('auto-easy', 5);
   const easy = store.createTask({ projectId, title: 'small', complexity: 'easy' });
   const medium = store.createTask({ projectId, title: 'mid', complexity: 'medium' });
@@ -85,392 +190,181 @@ test('autonomy=auto-easy：只领 easy，medium ready 不动', async () => {
   store.decide(medium.id, 'accept', 'user');                 // medium → ready
   assert.equal(store.getTask(medium.id)!.status, 'ready');
 
-  const ran: string[] = [];
+  const state = freshState();
-  const orch = createOrchestrator(store, noopLog, mockDeps({
+  const { deps } = mockDeps(state);
-    runner: async (task) => { ran.push(task.id); return okRun; },
+  createOrchestrator(store, noopLog, deps).claimTick();
-  }));
-  orch.tick();
-  await orch.drain();
 
-  assert.deepEqual(ran, [easy.id]);
+  assert.equal(state.spawned.length, 1);
-  assert.equal(store.getTask(easy.id)!.status, 'exec_review');
+  assert.equal(store.getTask(easy.id)!.status, 'executing');
-  assert.equal(store.getTask(medium.id)!.status, 'ready');   // 不自动跑
+  assert.equal(store.getTask(medium.id)!.status, 'ready');   // 不领 medium
   store.close();
 });
 
-test('autonomy=auto-approved：领全部 ready（含 medium），依赖未满足/非叶子不领', async () => {
+test('auto-approved：领 ready（含 medium）；依赖未满足/非叶子不领', () => {
   const { store, projectId } = setup('auto-approved', 5);
   const medium = store.createTask({ projectId, title: 'mid', complexity: 'medium' });
   store.setSpec(medium.id, '方案');
   store.transition(medium.id, 'spec_review');
   store.decide(medium.id, 'accept', 'user');
-  // 依赖未 done 的任务：建在 medium 上 → blocked，不可领
   const dep = store.createTask({ projectId, title: 'after-mid', complexity: 'easy', deps: [medium.id] });
   assert.equal(store.getTask(dep.id)!.status, 'blocked');
 
-  const ran: string[] = [];
+  const state = freshState();
-  const orch = createOrchestrator(store, noopLog, mockDeps({
+  const { deps } = mockDeps(state);
-    runner: async (task) => { ran.push(task.id); return okRun; },
+  createOrchestrator(store, noopLog, deps).claimTick();
-  }));
-  orch.tick();
-  await orch.drain();
 
-  assert.deepEqual(ran, [medium.id]);
+  assert.equal(state.spawned.length, 1);
-  assert.equal(store.getTask(medium.id)!.status, 'exec_review');
+  assert.equal(store.getTask(medium.id)!.status, 'executing');
   assert.equal(store.getTask(dep.id)!.status, 'blocked');
   store.close();
 });
 
-test('concurrency=1：同项目同轮只领 1 个，跑完下一轮再领', async () => {
+test('退避：nextEligibleAt 在未来 → 不领；过期后 → 领', () => {
-  const { store, projectId } = setup('auto-approved', 1);
-  const t1 = store.createTask({ projectId, title: 'a', complexity: 'easy' });
-  const t2 = store.createTask({ projectId, title: 'b', complexity: 'easy' });
-
-  const gate = deferred<void>();
-  const started: string[] = [];
-  const orch = createOrchestrator(store, noopLog, mockDeps({
-    runner: async (task) => { started.push(task.id); await gate.promise; return okRun; },
-  }));
-
-  orch.tick();
-  await settle();
-  assert.deepEqual(started, [t1.id], '并发=1 只应启动第一个任务');
-  assert.equal(store.getTask(t1.id)!.status, 'executing');
-  assert.equal(store.getTask(t2.id)!.status, 'ready');
-  assert.equal(orch.inflight.size, 1);
-
-  orch.tick();                              // 在途占满 → 本轮不领
-  await settle();
-  assert.deepEqual(started, [t1.id]);
-
-  gate.resolve();
-  await orch.drain();
-  assert.equal(store.getTask(t1.id)!.status, 'exec_review');
-
-  orch.tick();                              // 槽位释放 → 领第二个
-  gate.resolve();
-  await orch.drain();
-  assert.deepEqual(started, [t1.id, t2.id]);
-  assert.equal(store.getTask(t2.id)!.status, 'exec_review');
-  store.close();
-});
-
-test('成功路径：状态流转 + setResult(四字段) + executor/reviewer/security 三 run succeeded', async () => {
-  const { store, projectId } = setup('auto-easy');
-  const t = store.createTask({ projectId, title: 'tweak', complexity: 'easy' });
-  store.setOperations(t.id, '在 README.md 追加一行');
-
-  const seen: string[] = [];
-  store.subscribe((e) => { if (e.type === 'status.changed' && e.taskId === t.id) seen.push(String(e.payload.to)); });
-
-  let codeReportSeen: string | null = null;
-  let secReportSeen: string | null = null;
-  const orch = createOrchestrator(store, noopLog, mockDeps({
-    reviewCode: async (_task, _project, _wt, _runId, executorReport) => {
-      codeReportSeen = executorReport;
-      return okReview;
-    },
-    reviewSecurity: async (_task, _project, _wt, _runId, executorReport) => {
-      secReportSeen = executorReport;
-      return okSecurity;
-    },
-  }));
-  orch.tick();
-  await orch.drain();
-
-  const done = store.getTask(t.id)!;
-  assert.equal(done.status, 'exec_review');
-  assert.deepEqual(seen, ['queued', 'executing', 'exec_review']);
-  assert.deepEqual(done.result, {
-    branch: `maestro/${t.id}`,
-    worktree: `/tmp/fake-wt/${t.id}`,
-    diffSummary: ' README.md | 1 +',
-    commits: ['abc1234 hello maestro'],
-    prUrl: null,
-    summary: '## 做了什么\nmock 复审通过',
-    verdict: 'approve',
-    securitySummary: '## 安全审计\nmock 审计通过',
-    securityVerdict: 'approve',
-    mergeTaskId: null,
-  });
-  assert.equal(codeReportSeen, '执行自述：改了 README');   // runner finalText 传给两个复审作执行者自述
-  assert.equal(secReportSeen, '执行自述：改了 README');
-
-  const runs = store.listRuns(t.id);
-  assert.equal(runs.length, 3);
-  const executor = runs.find((r) => r.kind === 'executor')!;
-  assert.equal(executor.status, 'succeeded');
-  assert.equal(executor.branch, `maestro/${t.id}`);
-  assert.equal(executor.transcriptRef, '/tmp/fake.jsonl');
-  assert.equal(executor.claudeSessionId, 'sess-mock-1');
-  const reviewer = runs.find((r) => r.kind === 'reviewer')!;
-  assert.equal(reviewer.status, 'succeeded');
-  assert.equal(reviewer.transcriptRef, '/tmp/fake-review.jsonl');
-  assert.equal(reviewer.claudeSessionId, 'sess-review-1');
-  const security = runs.find((r) => r.kind === 'security')!;
-  assert.equal(security.status, 'succeeded');
-  assert.equal(security.transcriptRef, '/tmp/fake-security.jsonl');
-  assert.equal(security.claudeSessionId, 'sess-security-1');
-  store.close();
-});
-
-test('任一 verdict=reject 也照常落进 result（最终裁决仍归用户）', async () => {
-  const { store, projectId } = setup('auto-easy');
-  const t = store.createTask({ projectId, title: 'risky', complexity: 'easy' });
-  const orch = createOrchestrator(store, noopLog, mockDeps({
-    reviewCode: async () => ({ ...okReview, summary: '发现问题', verdict: 'reject' as const }),
-    reviewSecurity: async () => ({ ...okSecurity, summary: '发现密钥泄露', verdict: 'reject' as const }),
-  }));
-  orch.tick();
-  await orch.drain();
-  const done = store.getTask(t.id)!;
-  assert.equal(done.status, 'exec_review');
-  assert.equal(done.result!.verdict, 'reject');
-  assert.equal(done.result!.summary, '发现问题');
-  assert.equal(done.result!.securityVerdict, 'reject');
-  assert.equal(done.result!.securitySummary, '发现密钥泄露');
-  store.close();
-});
-
-test('code review 失败不挡任务：summary 记失败原因、verdict=null，安全审计照常跑', async () => {
-  const { store, projectId } = setup('auto-easy');
-  const t = store.createTask({ projectId, title: 'review-broken', complexity: 'easy' });
-  const orch = createOrchestrator(store, noopLog, mockDeps({
-    reviewCode: async () => { throw new Error('复审 CC 崩了'); },
-  }));
-  orch.tick();
-  await orch.drain();
-
-  const done = store.getTask(t.id)!;
-  assert.equal(done.status, 'exec_review');                       // 不挡结果闸
-  assert.equal(done.result!.verdict, null);
-  assert.equal(done.result!.summary, '自动复审失败：复审 CC 崩了');
-  assert.equal(done.result!.securityVerdict, 'approve');          // 另一个复审不受影响
-  assert.equal(done.result!.securitySummary, '## 安全审计\nmock 审计通过');
-
-  const runs = store.listRuns(t.id);
-  assert.equal(runs.find((r) => r.kind === 'executor')!.status, 'succeeded');
-  const reviewer = runs.find((r) => r.kind === 'reviewer')!;
-  assert.equal(reviewer.status, 'failed');
-  assert.match(reviewer.error ?? '', /复审 CC 崩了/);
-  assert.equal(runs.find((r) => r.kind === 'security')!.status, 'succeeded');
-  store.close();
-});
-
-test('安全审计失败不挡任务：securitySummary 记失败原因、securityVerdict=null，code review 不受影响', async () => {
-  const { store, projectId } = setup('auto-easy');
-  const t = store.createTask({ projectId, title: 'security-broken', complexity: 'easy' });
-  const orch = createOrchestrator(store, noopLog, mockDeps({
-    reviewSecurity: async () => { throw new Error('审计 CC 崩了'); },
-  }));
-  orch.tick();
-  await orch.drain();
-
-  const done = store.getTask(t.id)!;
-  assert.equal(done.status, 'exec_review');
-  assert.equal(done.result!.verdict, 'approve');
-  assert.equal(done.result!.summary, '## 做了什么\nmock 复审通过');
-  assert.equal(done.result!.securityVerdict, null);
-  assert.equal(done.result!.securitySummary, '自动复审失败：审计 CC 崩了');
-
-  const runs = store.listRuns(t.id);
-  assert.equal(runs.find((r) => r.kind === 'executor')!.status, 'succeeded');
-  assert.equal(runs.find((r) => r.kind === 'reviewer')!.status, 'succeeded');
-  const security = runs.find((r) => r.kind === 'security')!;
-  assert.equal(security.status, 'failed');
-  assert.match(security.error ?? '', /审计 CC 崩了/);
-  store.close();
-});
-
-test('verify 不过：按失败处理（run failed + 重新入队）', async () => {
-  const { store, projectId } = setup('auto-easy');
-  const t = store.createTask({ projectId, title: 'v', complexity: 'easy' });
-  const orch = createOrchestrator(store, noopLog, mockDeps({
-    verify: async () => ({ ok: false, exitCode: 1, logRef: '/tmp/v.log', error: 'verify 失败（exit 1）' }),
-  }));
-  orch.tick();
-  await orch.drain();
-  assert.equal(store.getTask(t.id)!.status, 'queued');       // 第一次失败 → 重新入队
-  const runs = store.listRuns(t.id);
-  assert.equal(runs.length, 1);
-  assert.equal(runs[0].status, 'failed');
-  assert.match(runs[0].error ?? '', /verify 失败/);
-  store.close();
-});
-
-test(`失败重试：重试 ${DEFAULT_MAX_RETRIES} 次后 → needs_attention（共 ${DEFAULT_MAX_RETRIES + 1} 次失败 run）`, async () => {
-  const { store, projectId } = setup('auto-easy');
-  const t = store.createTask({ projectId, title: 'flaky', complexity: 'easy' });
-
-  let attempts = 0;
-  // 使用快进时钟跳过退避冷却（每次 tick 前将时钟推进 10min）
-  const clock = makeClock();
-  const orch = createOrchestrator(store, noopLog, mockDeps({
-    runner: async () => { attempts++; return { ok: false, transcriptRef: null, sessionId: null, error: `boom #${attempts}` }; },
-    nowMs: clock.nowMs,
-  }));
-
-  for (let i = 1; i <= DEFAULT_MAX_RETRIES; i++) {
-    clock.advance(10 * 60_000);            // 跳过退避冷却
-    orch.tick();
-    await orch.drain();
-    assert.equal(store.getTask(t.id)!.status, 'queued', `第 ${i} 次失败后应重新入队`);
-  }
-
-  clock.advance(10 * 60_000);             // 跳过最后一次退避
-  orch.tick();                              // 最后一次重试也失败
-  await orch.drain();
-  assert.equal(store.getTask(t.id)!.status, 'needs_attention');
-  assert.equal(attempts, DEFAULT_MAX_RETRIES + 1);
-
-  const failed = store.listRuns(t.id).filter((r) => r.status === 'failed');
-  assert.equal(failed.length, DEFAULT_MAX_RETRIES + 1);
-
-  clock.advance(10 * 60_000);
-  orch.tick();                              // needs_attention 不会再被领取
-  await orch.drain();
-  assert.equal(attempts, DEFAULT_MAX_RETRIES + 1);
-  store.close();
-});
-
-test('项目级 maxRetries=1：只重试 1 次就 → needs_attention', async () => {
-  const store = new Store(':memory:');
-  const p = store.createProject({
-    name: 'orch', repoPath: '/tmp/orch-repo-retries-' + Math.random(),
-    autonomy: 'auto-easy', maxRetries: 1,
-  });
-  const projectId = p.id;
-  const t = store.createTask({ projectId, title: 'fail1', complexity: 'easy' });
-
-  let attempts = 0;
-  const clock = makeClock();
-  const orch = createOrchestrator(store, noopLog, mockDeps({
-    runner: async () => { attempts++; return { ok: false, transcriptRef: null, sessionId: null, error: 'boom' }; },
-    nowMs: clock.nowMs,
-  }));
-
-  // 第 1 次执行失败 → 重新入队（还有 1 次重试机会）
-  orch.tick();
-  await orch.drain();
-  assert.equal(store.getTask(t.id)!.status, 'queued');
-
-  // 第 2 次失败 → 超过 maxRetries=1 → needs_attention（快进时钟跳过退避）
-  clock.advance(10 * 60_000);
-  orch.tick();
-  await orch.drain();
-  assert.equal(store.getTask(t.id)!.status, 'needs_attention');
-  assert.equal(attempts, 2);
-  store.close();
-});
-
-test('项目级 maxRetries=0：首次失败直接 needs_attention（不重试）', async () => {
-  const store = new Store(':memory:');
-  const p = store.createProject({
-    name: 'orch', repoPath: '/tmp/orch-repo-zero-' + Math.random(),
-    autonomy: 'auto-easy', maxRetries: 0,
-  });
-  const t = store.createTask({ projectId: p.id, title: 'fail0', complexity: 'easy' });
-
-  let attempts = 0;
-  const orch = createOrchestrator(store, noopLog, mockDeps({
-    runner: async () => { attempts++; return { ok: false, transcriptRef: null, sessionId: null, error: 'boom' }; },
-  }));
-
-  orch.tick();
-  await orch.drain();
-  assert.equal(store.getTask(t.id)!.status, 'needs_attention');
-  assert.equal(attempts, 1);
-  store.close();
-});
-
-test('computeBackoffMs：指数退避公式，30s 基数，上限 10min', () => {
-  assert.equal(computeBackoffMs(1), 30_000);
-  assert.equal(computeBackoffMs(2), 60_000);
-  assert.equal(computeBackoffMs(3), 120_000);
-  assert.equal(computeBackoffMs(4), 240_000);
-  assert.equal(computeBackoffMs(10), 10 * 60_000); // capped at 10min
-  assert.equal(computeBackoffMs(100), 10 * 60_000);
-});
-
-test('退避冷却：首次失败重入队后，退避期内不被领取；退避过期后正常领取', async () => {
   const { store, projectId } = setup('auto-easy');
   const t = store.createTask({ projectId, title: 'backoff', complexity: 'easy' });
+  // 任务已在 queued 且退避到未来
+  store.transition(t.id, 'queued', { by: 'test' });
+  const future = new Date(Date.now() + 60_000).toISOString();
+  store.setNextEligibleAt(t.id, future);
 
-  let attempts = 0;
+  const clock = makeClock(Date.now());
-  // 使用可控时钟：初始时间 0，退避 30s（attempt=1 → BACKOFF_BASE=30000ms）
+  const state = freshState();
-  const clock = makeClock(0);
+  const { deps } = mockDeps(state, { nowMs: clock.nowMs });
-  const orch = createOrchestrator(store, noopLog, mockDeps({
+  const orch = createOrchestrator(store, noopLog, deps);
-    runner: async () => { attempts++; return { ok: false, transcriptRef: null, sessionId: null, error: 'boom' }; },
-    nowMs: clock.nowMs,
-  }));
 
-  // 第 1 次失败（t=0）→ 重入队 + 退避（backoffUntil = 0 + 30000 = 30000ms）
+  orch.claimTick();
-  orch.tick();
+  assert.equal(state.spawned.length, 0, '退避期内不领');
-  await orch.drain();
-  assert.equal(store.getTask(t.id)!.status, 'queued');
-  assert.equal(attempts, 1);
-
-  // t=29999：退避期内，不被领取
-  clock.advance(29_999);
-  orch.tick();
-  await orch.drain();
-  assert.equal(attempts, 1, '退避期内不应再次执行');
   assert.equal(store.getTask(t.id)!.status, 'queued');
 
-  // t=30001：退避过期，任务应被重新领取
+  clock.advance(61_000);                                     // 退避过期
-  clock.advance(2);
+  orch.claimTick();
-  orch.tick();
+  assert.equal(state.spawned.length, 1, '退避过期后领取');
-  await orch.drain();
+  assert.equal(store.getTask(t.id)!.status, 'executing');
-  assert.equal(attempts, 2, '退避过期后应再次执行');
   store.close();
 });
 
-test('requeueTask：needs_attention → queued，重置重试基线，再次允许重试', async () => {
+test('spawnWorker 抛错：兜底 failTaskAttempt（task 转 queued，executor run failed）', () => {
   const { store, projectId } = setup('auto-easy');
-  const t = store.createTask({ projectId, title: 'requeue', complexity: 'easy' });
+  const t = store.createTask({ projectId, title: 'boom-spawn', complexity: 'easy' });
+  const state = freshState();
+  const { deps } = mockDeps(state, {
+    spawnWorker: () => { throw new Error('spawn 失败：ENOENT'); },
+  });
+  createOrchestrator(store, noopLog, deps).claimTick();
 
-  let attempts = 0;
+  const after = store.getTask(t.id)!;
-  const clock = makeClock();
+  assert.equal(after.status, 'queued', '默认 maxRetries=2，首次失败 → 重入队');
-  const orch = createOrchestrator(store, noopLog, mockDeps({
+  assert.ok(after.nextEligibleAt, '退避已持久化');
-    runner: async () => { attempts++; return { ok: false, transcriptRef: null, sessionId: null, error: 'boom' }; },
+  const executor = store.listRuns(t.id).find((r) => r.kind === 'executor')!;
-    nowMs: clock.nowMs,
+  assert.equal(executor.status, 'failed');
-  }));
+  assert.match(executor.error ?? '', /spawn 失败/);
+  store.close();
+});
 
-  // 耗尽默认重试次数 → needs_attention（每次需快进时钟跳过退避）
+// ───────────────────────── reaper（回收死 worker）─────────────────────────
-  for (let i = 0; i <= DEFAULT_MAX_RETRIES; i++) {
+
-    clock.advance(10 * 60_000);
+test('reaper：isWorkerAlive=false → failTaskAttempt（executing→queued，executor run failed）', () => {
-    orch.tick();
+  const { store, projectId } = setup('auto-easy');
-    await orch.drain();
+  const t = store.createTask({ projectId, title: 'dead', complexity: 'easy' });
-  }
+  // 先让它进 executing + 建 run（用一次正常领取，alive=true 不会被回收）
+  const state = freshState();
+  const { deps } = mockDeps(state);
+  const orch = createOrchestrator(store, noopLog, deps);
+  orch.claimTick();
+  assert.equal(store.getTask(t.id)!.status, 'executing');
+  const runId = store.listRuns(t.id)[0].id;
+
+  // 标记 worker 已死 → reaper 回收
+  state.alive = false;
+  orch.reap();
+
+  const after = store.getTask(t.id)!;
+  assert.equal(after.status, 'queued', '默认 maxRetries=2，首次回收 → 重入队');
+  assert.ok(after.nextEligibleAt, '退避已持久化');
+  const executor = store.getRun(runId)!;
+  assert.equal(executor.status, 'failed');
+  assert.match(executor.error ?? '', /worker 异常退出/);
+  store.close();
+});
+
+test('reaper：isWorkerAlive=true → 不回收（保持 executing）', () => {
+  const { store, projectId } = setup('auto-easy');
+  const t = store.createTask({ projectId, title: 'alive', complexity: 'easy' });
+  const state = freshState();
+  const { deps } = mockDeps(state);
+  const orch = createOrchestrator(store, noopLog, deps);
+  orch.claimTick();
+  assert.equal(store.getTask(t.id)!.status, 'executing');
+
+  state.alive = true;
+  orch.reap();
+  assert.equal(store.getTask(t.id)!.status, 'executing', 'worker 活 → 不回收');
+  store.close();
+});
+
+test('reaper：maxRetries=0 → 死 worker 直接 needs_attention', () => {
+  const store = new Store(':memory:');
+  const p = store.createProject({
+    name: 'orch', repoPath: '/tmp/orch-reap0-' + Math.random(), autonomy: 'auto-easy', maxRetries: 0,
+  });
+  const t = store.createTask({ projectId: p.id, title: 'x', complexity: 'easy' });
+  const state = freshState();
+  const { deps } = mockDeps(state);
+  const orch = createOrchestrator(store, noopLog, deps);
+  orch.claimTick();
+  state.alive = false;
+  orch.reap();
   assert.equal(store.getTask(t.id)!.status, 'needs_attention');
-  assert.equal(attempts, DEFAULT_MAX_RETRIES + 1);
+  store.close();
+});
 
-  // 手动重投：重置基线，转 queued
+// ───────────────────────── 多轮：reaper 腾槽后重领（退避到期）─────────────────────────
-  const requeued = store.requeueTask(t.id);
-  assert.equal(requeued.status, 'queued');
-  assert.equal(requeued.retryBaseline, DEFAULT_MAX_RETRIES + 1);
 
-  // 重投后编排器应能再次执行（快进时钟确保无退避阻拦）
+test('死 worker 回收 + 退避到期后下一轮重新领取（监工闭环）', () => {
+  const { store, projectId } = setup('auto-easy', 1);
+  const t = store.createTask({ projectId, title: 'recycle', complexity: 'easy' });
+  const clock = makeClock(Date.now());
+  const state = freshState();
+  const { deps } = mockDeps(state, { nowMs: clock.nowMs });
+  const orch = createOrchestrator(store, noopLog, deps);
+
+  // 第一轮：领取 + spawn（alive=true）
+  orch.tick();
+  assert.equal(state.spawned.length, 1);
+  assert.equal(store.getTask(t.id)!.status, 'executing');
+
+  // worker 死 → 下一轮 reaper 回收（→ queued + 退避），退避期内不重领
+  state.alive = false;
+  orch.tick();
+  assert.equal(store.getTask(t.id)!.status, 'queued');
+  assert.equal(state.spawned.length, 1, '退避期内不重领');
+
+  // 退避到期 → 再下一轮重新领取
   clock.advance(10 * 60_000);
   orch.tick();
-  await orch.drain();
+  assert.equal(state.spawned.length, 2, '退避到期 → 重新领取 spawn');
-  assert.equal(attempts, DEFAULT_MAX_RETRIES + 2, '重投后应再次执行');
+  assert.equal(store.getTask(t.id)!.status, 'executing');
-  // 第 1 次净失败后应重入队（基线已重置，净失败=1 < maxRetries=2）
-  assert.equal(store.getTask(t.id)!.status, 'queued');
   store.close();
 });
 
-test('project paused：不领取', async () => {
+test(`reaper 反复回收：净失败 ${DEFAULT_MAX_RETRIES} 次后 → needs_attention`, () => {
-  const { store, projectId } = setup('auto-approved');
+  const { store, projectId } = setup('auto-easy', 1);
-  const t = store.createTask({ projectId, title: 'x', complexity: 'easy' });
+  const t = store.createTask({ projectId, title: 'flaky', complexity: 'easy' });
-  store.patchProject(projectId, { status: 'paused' });
+  const clock = makeClock(Date.now());
-  let calls = 0;
+  const state = freshState();
-  const orch = createOrchestrator(store, noopLog, mockDeps({
+  const { deps } = mockDeps(state, { nowMs: clock.nowMs });
-    runner: async () => { calls++; return okRun; },
+  const orch = createOrchestrator(store, noopLog, deps);
-  }));
+
-  orch.tick();
+  // 反复：领取（alive=true 领取那刻）→ 标死 → reaper 回收
-  await orch.drain();
+  for (let i = 0; i <= DEFAULT_MAX_RETRIES; i++) {
-  assert.equal(calls, 0);
+    clock.advance(10 * 60_000);          // 跳过退避
-  assert.equal(store.getTask(t.id)!.status, 'ready');
+    state.alive = true;
+    orch.claimTick();
+    state.alive = false;
+    orch.reap();
+  }
+  assert.equal(store.getTask(t.id)!.status, 'needs_attention');
+  const failed = store.listRuns(t.id).filter((r) => r.kind === 'executor' && r.status === 'failed');
+  assert.equal(failed.length, DEFAULT_MAX_RETRIES + 1);
   store.close();
 });