diff --git a/src/daemon/ingest.ts b/src/daemon/ingest.ts
index 2829b36..5fa3bad 100644
--- a/src/daemon/ingest.ts
+++ b/src/daemon/ingest.ts
@@ -26,7 +26,7 @@ export function ingestRun(store: Store, log: IngestLogger, runId: string): void
   const recs = readOutboxSince(runId, run.lastSeq);
   for (const rec of recs) {
     try {
-      applyRecord(store, log, taskId, runId, rec);
+      applyRecord(store, log, taskId, runId, run.kind, 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}`);
@@ -36,8 +36,8 @@ export function ingestRun(store: Store, log: IngestLogger, runId: string): void
   }
 }
 
-/** 把一条 OutboxRecord 映射为 DB 写。 */
-function applyRecord(store: Store, log: IngestLogger, taskId: string, runId: string, rec: OutboxRecord): void {
+/** 把一条 OutboxRecord 映射为 DB 写。runKind 用于区分 failed 的收尾策略（executor vs planner）。 */
+function applyRecord(store: Store, log: IngestLogger, taskId: string, runId: string, runKind: 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。
@@ -50,8 +50,8 @@ function applyRecord(store: Store, log: IngestLogger, taskId: string, runId: str
       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。
+      // 先把 run 收尾为 failed（带转录/会话），再走失败/重试策略（按 kind 分流）。
+      // failTaskAttempt/failPlanAttempt 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,
@@ -61,8 +61,44 @@ function applyRecord(store: Store, log: IngestLogger, taskId: string, runId: str
       } 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}`);
+      if (runKind === 'planner') store.failPlanAttempt(taskId, runId, rec.error);  // planner：退避留 analyzing/speccing
+      else store.failTaskAttempt(taskId, runId, rec.error);                        // executor：failed→重试/needs_attention
+      log.info(`task=${taskId} run=${runId}(${runKind}) 失败：${rec.error}`);
+      return;
+    }
+
+    case 'spec-result': {
+      // planner-spec 成功：写方案 → 转 spec_review（待你审）→ 收尾 planner run。
+      store.setSpec(taskId, rec.spec);
+      store.transition(taskId, 'spec_review', { by: 'ingest', runId });
+      store.finishRun(runId, 'succeeded', {
+        transcriptRef: rec.transcriptRef ?? undefined,
+        claudeSessionId: rec.sessionId ?? undefined,
+      });
+      log.info(`task=${taskId} run=${runId} 方案完成 → spec_review`);
+      return;
+    }
+
+    case 'decompose-result': {
+      // planner-decompose 成功：写分析 + 建子任务 + 转 plan_review（待你审）→ 收尾 planner run。
+      const task = store.getTask(taskId);
+      if (!task) { log.error(`decompose: 任务不存在 ${taskId}`); return; }
+      store.setPlan(taskId, rec.plan || '（无分析说明）');
+      let created = 0;
+      for (const sub of rec.subtasks) {
+        try {
+          store.createTask({ projectId: task.projectId, parentId: taskId, title: sub.title, complexity: sub.complexity });
+          created++;
+        } catch (e) {
+          log.error(`decompose: 建子任务「${sub.title}」失败：${(e as Error).message}`);
+        }
+      }
+      store.transition(taskId, 'plan_review', { by: 'ingest', runId, subtasks: created });
+      store.finishRun(runId, 'succeeded', {
+        transcriptRef: rec.transcriptRef ?? undefined,
+        claudeSessionId: rec.sessionId ?? undefined,
+      });
+      log.info(`task=${taskId} run=${runId} 拆解完成 → plan_review（${created} 个子任务）`);
       return;
     }
 
@@ -103,12 +139,11 @@ function applyRecord(store: Store, log: IngestLogger, taskId: string, runId: str
 }
 
 /**
- * 摄取所有在途 run：对每个 executing 任务的最近一条 executor run 调 ingestRun。
+ * 摄取所有在途 run（executor + planner）：遍历所有 started 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;
+  for (const { run } of store.liveRunsWithTask()) {
     ingestRun(store, log, run.id);
   }
 }
diff --git a/src/daemon/orchestrator.ts b/src/daemon/orchestrator.ts
index 866b8b6..c11ea2a 100644
--- a/src/daemon/orchestrator.ts
+++ b/src/daemon/orchestrator.ts
@@ -117,17 +117,21 @@ export function createOrchestrator(store: Store, log: OrchestratorLogger, 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。
-   * 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）。
+   * 本项目可领取的任务（三类"可执行"状态）：
+   *   ready/queued → executor（执行改动）；analyzing → planner 拆解 Hard；speccing → planner 写方案 Medium。
+   * 排除：已在途（inflight=有 started run 的任务）、非叶子容器、退避冷却中、deps 未全 done；
+   * auto-easy 只挑 easy（故不做规划——analyzing/speccing 都是 hard/medium 被排除）。按调度分降序返回。
    */
-  function claimable(project: Project): Array<{ task: Task; score: number }> {
+  function claimable(project: Project, inflight: ReadonlySet<string>): Array<{ task: Task; score: number }> {
     const nowMs = d.nowMs();
     const tasks = store.listTasks(project.id);
     const byId = new Map(tasks.map((t) => [t.id, t]));
     const parents = new Set(tasks.filter((t) => t.parentId).map((t) => t.parentId as string));
     const easyOnly = project.autonomy === 'auto-easy';
+    const CLAIMABLE = new Set<string>(['ready', 'queued', 'analyzing', 'speccing']);
     const candidates = tasks.filter((t) => {
-      if (t.status !== 'ready' && t.status !== 'queued') return false;
+      if (!CLAIMABLE.has(t.status)) return false;
+      if (inflight.has(t.id)) return false;                                 // 已有在途 run（executor/planner）
       if (parents.has(t.id)) return false;                                  // 非叶子（容器）跳过
       if (easyOnly && t.complexity !== 'easy') return false;
       if (t.nextEligibleAt && nowMs < Date.parse(t.nextEligibleAt)) return false;  // 退避冷却中
@@ -136,51 +140,66 @@ export function createOrchestrator(store: Store, log: OrchestratorLogger, deps:
     return rankByScore(candidates, tasks);
   }
 
-  /** 领取一个任务：建分支/worktree 路径 → executing → startRun → 写 job.json → spawn worker → 记 pid。 */
+  /**
+   * 领取一个任务并起 worker：
+   * - ready/queued → executor：ready→queued→executing，startRun(executor, worktree)，job.runKind=executor。
+   * - analyzing → planner-decompose；speccing → planner-spec：【不转状态】（留 analyzing/speccing 作"规划中"），
+   *   startRun(planner)，planner 只读跑在 repo（worktreeDir 仅记录用）。失败按 kind 走 failTaskAttempt/failPlanAttempt。
+   */
   function claimOne(project: Project, task: Task, score: number): void {
+    const isPlanner = task.status === 'analyzing' || task.status === 'speccing';
+    const runKind = task.status === 'analyzing' ? 'planner-decompose'
+      : task.status === 'speccing' ? 'planner-spec' : 'executor';
+    const dbKind = isPlanner ? 'planner' : 'executor';
+    const role = isPlanner ? 'planner' : 'executor';
     let runId: string | null = null;
     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 run = store.startRun(task.id, 'executor', { worktree: dir, branch });
+      if (!isPlanner) {
+        if (task.status === 'ready') store.transition(task.id, 'queued', { by: 'orchestrator', score });
+        store.transition(task.id, 'executing', { by: 'orchestrator' });
+      }
+      // planner run 只读跑在主仓，worktree 记 repoPath；executor 记隔离 worktree 路径
+      const run = store.startRun(task.id, dbKind, { worktree: isPlanner ? project.repoPath : dir, branch });
       runId = run.id;
 
       d.writeJobSpec({
         runId: run.id,
-        task: { ...task, status: 'executing' },
-        project,
-        worktreeDir: dir,
-        branch,
+        task: { ...task, status: isPlanner ? task.status : 'executing' },
+        project, worktreeDir: dir, branch, runKind,
       });
 
-      log.info(`领取任务 ${task.id}「${task.title}」score=${score} run=${run.id} model=${pickModel(task, project, 'executor')}`);
+      log.info(`领取 ${runKind} ${task.id}「${task.title}」score=${score} run=${run.id} model=${pickModel(task, project, role)}`);
       const pid = d.spawnWorker(run.id);
       store.setWorkerPid(run.id, pid);
-      log.info(`任务 ${task.id} 已起 worker pid=${pid} worktree=${dir}`);
+      log.info(`任务 ${task.id} 已起 worker pid=${pid}（${runKind}）`);
     } catch (e) {
       const msg = (e as Error).message;
       log.error(`任务 ${task.id} 领取/spawn 失败：${msg}`);
       try {
-        store.failTaskAttempt(task.id, runId, msg);
+        if (isPlanner) store.failPlanAttempt(task.id, runId, msg);
+        else store.failTaskAttempt(task.id, runId, msg);
       } catch (e2) {
         log.error(`任务 ${task.id} 失败收尾出错：${(e2 as Error).message}`);
       }
     }
   }
 
-  /** 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。 */
+  /**
+   * 领取轮：对每个 active 且 autonomy≠manual 的项目，并发闸 = inflightTaskIds（有 started run 的任务，
+   * executor + planner 通用）。在 active<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 = store.countExecuting(p.id);
+        const inflight = store.inflightTaskIds(p.id);
+        let active = inflight.size;
         if (active >= p.concurrency) continue;
 
-        for (const { task, score } of claimable(p)) {
+        for (const { task, score } of claimable(p, inflight)) {
           if (active >= p.concurrency) break;
           claimOne(p, task, score);
           active++;
@@ -200,19 +219,21 @@ export function createOrchestrator(store: Store, log: OrchestratorLogger, deps:
     }
   }
 
-  /** 回收：对每个 executing 任务判活，死 worker → failTaskAttempt（收尾 + 重试/needs_attention）。 */
+  /** 回收：对每个在途 run（executor + planner）判活，死 worker → 按 kind 收尾（executor=failTaskAttempt / planner=failPlanAttempt）。 */
   function reap(): void {
     try {
-      for (const { task, run } of store.executingWithLatestExecutorRun()) {
-        const alive = run !== null && d.isWorkerAlive({
+      for (const { task, run } of store.liveRunsWithTask()) {
+        const alive = 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`);
+        log.error(`任务 ${task.id} worker 异常退出（run=${run.id} ${run.kind}）→ 回收`);
         try {
-          store.failTaskAttempt(task.id, run?.id ?? null, 'worker 异常退出');
+          if (run.kind === 'executor') store.failTaskAttempt(task.id, run.id, 'worker 异常退出');
+          else if (run.kind === 'planner') store.failPlanAttempt(task.id, run.id, 'planner worker 异常退出');
+          else store.finishRun(run.id, 'failed', { error: 'worker 异常退出' });
         } catch (e) {
           log.error(`任务 ${task.id} 回收收尾出错：${(e as Error).message}`);
         }
diff --git a/src/executor/models.ts b/src/executor/models.ts
index 8c30730..0f20c3d 100644
--- a/src/executor/models.ts
+++ b/src/executor/models.ts
@@ -1,8 +1,8 @@
 import type { Project, Task } from '../model/types.js';
 import type { Complexity } from '../model/complexity.js';
 
-/** 角色：executor（执行改动）/ reviewer（执行后复审） */
-export type ModelRole = 'executor' | 'reviewer';
+/** 角色：executor（执行改动）/ reviewer（执行后复审）/ planner（拆解 Hard / 写方案 Medium） */
+export type ModelRole = 'executor' | 'reviewer' | 'planner';
 
 /** 模型不可用时的回退链（按序取第一个 ≠ 失败模型的，同一 run 内只重试一次） */
 export const MODEL_FALLBACK_CHAIN = ['claude-opus-4-8', 'claude-sonnet-4-6'] as const;
@@ -19,6 +19,12 @@ const MODEL_TABLE: Record<ModelRole, Record<Complexity, [env: string, fallback:
     medium: ['MAESTRO_MODEL_REVIEW_MEDIUM', 'claude-opus-4-8'],
     hard: ['MAESTRO_MODEL_REVIEW_HARD', 'claude-opus-4-8'],
   },
+  // planner：拆解/写方案要质量，统一用 opus 档（project.model 设了仍最优先）
+  planner: {
+    easy: ['MAESTRO_MODEL_PLAN_EASY', 'claude-opus-4-8'],
+    medium: ['MAESTRO_MODEL_PLAN_MEDIUM', 'claude-opus-4-8'],
+    hard: ['MAESTRO_MODEL_PLAN_HARD', 'claude-opus-4-8'],
+  },
 };
 
 /**
diff --git a/src/executor/pipeline.ts b/src/executor/pipeline.ts
index 7ff89a8..1e66be5 100644
--- a/src/executor/pipeline.ts
+++ b/src/executor/pipeline.ts
@@ -13,9 +13,9 @@
 //   双复审 + setResult(四字段) + transition(exec_review) → emit result
 //   （worker 退出）                                      → emit done
 
-import type { JobSpec, OutboxPayload, ReviewReport } from './protocol.js';
+import type { JobSpec, OutboxPayload, ReviewReport, DecomposeResult } from './protocol.js';
 import { createWorktree, worktreeDiff, type WorktreeDiff, type WorktreeInfo } from './worktree.js';
-import { runTask, type RunnerFn } from './runner.js';
+import { runTask, runPlanner, type RunnerFn, type PlannerFn } from './runner.js';
 import { runVerify, type VerifyFn } from './verify.js';
 import { reviewCode, reviewSecurity, type ReviewerFn } from './reviewer.js';
 
@@ -27,6 +27,7 @@ export interface PipelineDeps {
   verify: VerifyFn;
   reviewCode: ReviewerFn;     // code review（daemon 落成 kind=reviewer 的 run）
   reviewSecurity: ReviewerFn; // 安全审计（daemon 落成 kind=security 的 run）
+  runPlanner: PlannerFn;      // planner（拆解 Hard / 写方案 Medium，只读跑 CC）
 }
 
 /** 真实依赖（worker 生产用）。 */
@@ -37,6 +38,7 @@ export const realDeps: PipelineDeps = {
   verify: runVerify,
   reviewCode,
   reviewSecurity,
+  runPlanner,
 };
 
 /** emit：把一条 OutboxPayload 交给 outbox（worker 传 appendOutbox 包装，测试传收集器）。 */
@@ -74,6 +76,10 @@ async function runOneReview(
  * 不抛错（除非 emit 自身抛——那由 worker 顶层兜底）。全程总会 emit 一条终态（failed|result）后再 emit done。
  */
 export async function runPipeline(job: JobSpec, deps: PipelineDeps = realDeps, emit: Emit): Promise<void> {
+  // planner-* 走只读的规划管线（不建 worktree、不改文件、不复审）
+  if (job.runKind === 'planner-spec' || job.runKind === 'planner-decompose') {
+    return runPlannerPipeline(job, deps, emit);
+  }
   // 1. 建 worktree（其 dir/branch 应与 job.worktreeDir/branch 一致；以 deps 真建的为准）
   const wt = await deps.createWorktree(job.project.repoPath, job.task.id, job.project.defaultBranch);
 
@@ -118,3 +124,54 @@ export async function runPipeline(job: JobSpec, deps: PipelineDeps = realDeps, e
   });
   emit({ type: 'done' });
 }
+
+/**
+ * planner 管线（只读）：跑一次 CC 写方案 / 拆解，emit spec-result / decompose-result。
+ * - planner-spec：CC 输出即方案正文 → emit spec-result{spec}。
+ * - planner-decompose：解析 CC 输出末尾的 fenced JSON{plan,subtasks} → emit decompose-result；解析失败 → failed。
+ * CC 失败 / 无输出 → failed。全程总 emit 一条终态后再 emit done。
+ */
+async function runPlannerPipeline(job: JobSpec, deps: PipelineDeps, emit: Emit): Promise<void> {
+  const kind = job.runKind === 'planner-spec' ? 'spec' : 'decompose';
+  emit({ type: 'phase', phase: kind === 'spec' ? 'speccing' : 'decomposing' });
+  const r = await deps.runPlanner(job.task, job.project, kind, job.runId);
+  if (!r.ok || !r.finalText) {
+    emit({ type: 'failed', error: r.error ?? 'planner 无输出', transcriptRef: r.transcriptRef, sessionId: r.sessionId });
+    emit({ type: 'done' });
+    return;
+  }
+  if (kind === 'spec') {
+    emit({ type: 'spec-result', spec: r.finalText, transcriptRef: r.transcriptRef, sessionId: r.sessionId });
+    emit({ type: 'done' });
+    return;
+  }
+  const parsed = parseDecompose(r.finalText);
+  if (!parsed) {
+    emit({ type: 'failed', error: 'planner 输出未含合法的子任务 JSON 块', transcriptRef: r.transcriptRef, sessionId: r.sessionId });
+    emit({ type: 'done' });
+    return;
+  }
+  emit({ type: 'decompose-result', plan: parsed.plan, subtasks: parsed.subtasks, transcriptRef: r.transcriptRef, sessionId: r.sessionId });
+  emit({ type: 'done' });
+}
+
+/** 从 CC 文本里抽取拆解结果：优先取最后一个 ```json 块，回退到任意 ``` 块，再回退到整段当 JSON。非法 → null。 */
+export function parseDecompose(text: string): DecomposeResult | null {
+  const blocks = [...text.matchAll(/```(?:json)?\s*([\s\S]*?)```/gi)].map((m) => m[1].trim());
+  const candidates = blocks.length ? [blocks[blocks.length - 1]] : [text.trim()];
+  for (const raw of candidates) {
+    try {
+      const o = JSON.parse(raw) as { plan?: unknown; subtasks?: unknown };
+      if (!Array.isArray(o.subtasks)) continue;
+      const subtasks = o.subtasks
+        .filter((s): s is { title: string; complexity: 'easy' | 'medium' | 'hard' } =>
+          !!s && typeof (s as { title?: unknown }).title === 'string' &&
+          ['easy', 'medium', 'hard'].includes((s as { complexity?: unknown }).complexity as string))
+        .map((s) => ({ title: String(s.title).trim(), complexity: s.complexity }))
+        .filter((s) => s.title.length > 0);
+      if (subtasks.length === 0) continue;
+      return { plan: typeof o.plan === 'string' ? o.plan : '', subtasks };
+    } catch { /* 试下一个候选 */ }
+  }
+  return null;
+}
diff --git a/src/executor/protocol.ts b/src/executor/protocol.ts
index fb836a7..a6c914b 100644
--- a/src/executor/protocol.ts
+++ b/src/executor/protocol.ts
@@ -54,6 +54,9 @@ export function isPlainRunId(runId: string): boolean {
  * worker 的唯一输入。daemon 在 spawn 前写 runs/<runId>/job.json，worker 读它即可执行，
  * 【无需访问 DB】。携带完整 Task / Project（均 JSON 可序列化），worker 自行 pickModel/buildPrompt。
  */
+/** run 类型：executor（在 worktree 执行改动）/ planner-spec（写 Medium 方案）/ planner-decompose（拆解 Hard） */
+export type RunKind = 'executor' | 'planner-spec' | 'planner-decompose';
+
 export interface JobSpec {
   runId: string;
   task: Task;
@@ -61,6 +64,14 @@ export interface JobSpec {
   /** 确定性 worktree 路径与分支（daemon 预填；worker 也能由 worktree.ts 算出，传入避免重复） */
   worktreeDir: string;
   branch: string;
+  /** run 类型，缺省 executor（向后兼容旧 job）。planner-* 在 repo 内只读跑 CC、不建 worktree。 */
+  runKind?: RunKind;
+}
+
+/** planner-decompose 的拆解产出：plan 文本 + 子任务列表 */
+export interface DecomposeResult {
+  plan: string;
+  subtasks: Array<{ title: string; complexity: 'easy' | 'medium' | 'hard' }>;
 }
 
 export function writeJobSpec(job: JobSpec): void {
@@ -86,6 +97,8 @@ export interface ReviewReport {
  * - phase  ：进度阶段（executing|verifying|reviewing…），daemon 落成 event 推看板（可选）
  * - failed ：本次尝试终态失败（executor 报错或 verify 失败）。daemon 据此 finishRun(failed)+重试决策
  * - result ：成功终态。daemon 据此建 reviewer/security run + setResult(四字段) + 转 exec_review
+ * - spec-result    ：planner-spec 成功终态。daemon 据此 setSpec + 转 spec_review
+ * - decompose-result：planner-decompose 成功终态。daemon 据此 setPlan + 建子任务 + 转 plan_review
  * - done   ：worker 即将退出（成功或失败都发，daemon 据此停止 tail）
  */
 export type OutboxRecord =
@@ -99,6 +112,12 @@ export type OutboxRecord =
       code: ReviewReport;
       security: ReviewReport;
     }
+  | { seq: number; at: string; type: 'spec-result'; spec: string; transcriptRef: string | null; sessionId: string | null }
+  | {
+      seq: number; at: string; type: 'decompose-result';
+      plan: string; subtasks: Array<{ title: string; complexity: 'easy' | 'medium' | 'hard' }>;
+      transcriptRef: string | null; sessionId: string | null;
+    }
   | { seq: number; at: string; type: 'done' };
 
 /** OutboxRecord 去掉 seq/at（由 appendOutbox 填） */
@@ -107,6 +126,8 @@ export type OutboxPayload =
   | Omit<Extract<OutboxRecord, { type: 'phase' }>, 'seq' | 'at'>
   | Omit<Extract<OutboxRecord, { type: 'failed' }>, 'seq' | 'at'>
   | Omit<Extract<OutboxRecord, { type: 'result' }>, 'seq' | 'at'>
+  | Omit<Extract<OutboxRecord, { type: 'spec-result' }>, 'seq' | 'at'>
+  | Omit<Extract<OutboxRecord, { type: 'decompose-result' }>, 'seq' | 'at'>
   | Omit<Extract<OutboxRecord, { type: 'done' }>, 'seq' | 'at'>;
 
 /** 追加一条 outbox 记录（worker 侧调用）。seq = 现有行数+1（worker 单线程，无并发写）。返回写入的完整记录。 */
diff --git a/src/executor/runner.ts b/src/executor/runner.ts
index ec0bdb0..dc0db9c 100644
--- a/src/executor/runner.ts
+++ b/src/executor/runner.ts
@@ -94,3 +94,69 @@ export async function runTask(task: Task, project: Project, worktree: WorktreeIn
   }
   return { ok: true, transcriptRef: cc.transcriptRef, sessionId: cc.sessionId, finalText: cc.finalText || null, modelUsed: cc.modelUsed };
 }
+
+// ───────────────────────── planner（拆解 Hard / 写方案 Medium）─────────────────────────
+
+export type PlanKind = 'spec' | 'decompose';
+
+export interface PlannerResult {
+  ok: boolean;
+  transcriptRef: string | null;
+  sessionId: string | null;
+  finalText: string | null;   // CC 的最终文本：spec=方案正文；decompose=分析 + 末尾 fenced JSON
+  error?: string;
+}
+
+/** 测试注入点（pipeline 的 planner 分支用） */
+export type PlannerFn = (task: Task, project: Project, kind: PlanKind, runId: string) => Promise<PlannerResult>;
+
+/** planner 提示词：只读代码库，spec=写改动方案；decompose=拆子任务并在末尾输出 fenced JSON。 */
+export function buildPlannerPrompt(task: Task, kind: PlanKind): string {
+  const head = [`# 任务：${task.title}`, `任务 ID：${task.id}`];
+  if (kind === 'spec') {
+    if (task.spec) head.push('', '## 现有方案草稿（可改进/替换）', task.spec);
+    return [
+      ...head, '',
+      '## 你的角色：方案作者（只读，不改任何文件）',
+      '只读这个代码库，为上述任务写一份「具体改动方案」。不要编辑文件、不要执行有副作用的命令。',
+      '## 输出（markdown 正文）',
+      '## 改动（要改哪些文件/模块、怎么改）',
+      '## 为什么（这么做的理由、取舍）',
+      '## 验收（怎么算完成、怎么验证）',
+      '这份方案将提交给人审；通过后才会有执行 agent 按它改代码。直接输出方案正文即可。',
+    ].join('\n');
+  }
+  if (task.plan) head.push('', '## 现有分析草稿（可改进/替换）', task.plan);
+  return [
+    ...head, '',
+    '## 你的角色：任务拆解者（只读，不改任何文件）',
+    '只读这个代码库，分析上述（复杂）任务，拆成 2–6 个更小、可独立交付的子任务。不要编辑文件、不要执行有副作用的命令。',
+    '每个子任务标注复杂度：easy（单文件机械改动/无设计）、medium（需方案、跨几处）、hard（仍需进一步拆解）。',
+    '## 输出（重要）',
+    '先写一段分析与拆解理由；然后在【最后】输出唯一一个 ```json 代码块，严格格式：',
+    '```json',
+    '{"plan":"一句话分析与拆解理由","subtasks":[{"title":"子任务标题","complexity":"easy"}]}',
+    '```',
+    'subtasks 至少 1 个；complexity 只能是 easy|medium|hard；除这个 JSON 块外不要再写其它 ``` 代码块。',
+  ].join('\n');
+}
+
+/**
+ * 起 headless CC 跑一次 planner：**只读**跑在 project.repoPath（不建 worktree、不改文件）。
+ * 模型用 planner 角色（默认 opus）。不抛错，失败折叠进 { ok:false, error }。
+ */
+export async function runPlanner(task: Task, project: Project, kind: PlanKind, runId: string): Promise<PlannerResult> {
+  const model = pickModel(task, project, 'planner');
+  const timeoutMs = project.timeoutMs ?? DEFAULT_TIMEOUT_MS;
+  const cc = await runClaude({
+    prompt: buildPlannerPrompt(task, kind),
+    cwd: project.repoPath,                  // 只读跑在主仓（planner 不改文件，无需 worktree）
+    model,
+    runId,
+    maxTurns: 60,
+    timeoutMs,
+    permissionMode: 'acceptEdits',          // planner 只读，不会触发编辑；保持工具流不被 prompt 卡住
+    allowedTools: ['Read', 'Glob', 'Grep'], // 纯只读：无 Edit/Write/Bash，零副作用
+  });
+  return { ok: cc.ok, transcriptRef: cc.transcriptRef, sessionId: cc.sessionId, finalText: cc.finalText || null, error: cc.error };
+}
diff --git a/src/executor/worker.ts b/src/executor/worker.ts
index 0d9129d..53159ad 100644
--- a/src/executor/worker.ts
+++ b/src/executor/worker.ts
@@ -32,13 +32,14 @@ async function main(): Promise<void> {
   const emit = (p: OutboxPayload): void => { appendOutbox(runId, p); };
   const job = readJobSpec(runId);
 
-  // 起手汇报：pid / worktree / branch / 实际执行模型
+  // 起手汇报：pid / worktree / branch / 实际执行模型（planner-* 用 planner 角色模型）
+  const role = job.runKind && job.runKind !== 'executor' ? 'planner' : 'executor';
   emit({
     type: 'started',
     pid: process.pid,
     worktree: job.worktreeDir,
     branch: job.branch,
-    model: pickModel(job.task, job.project, 'executor'),
+    model: pickModel(job.task, job.project, role),
   });
 
   // 心跳：立即一次 + 周期刷（结束时清掉）
diff --git a/src/model/status.ts b/src/model/status.ts
index 68411cd..fde0cbd 100644
--- a/src/model/status.ts
+++ b/src/model/status.ts
@@ -57,10 +57,10 @@ export const TERMINAL_STATUS: ReadonlySet<TaskStatus> = new Set<TaskStatus>(['do
 /** 合法状态流转表：from → 允许的 to[] */
 export const TRANSITIONS: Record<TaskStatus, TaskStatus[]> = {
   init:            ['analyzing', 'speccing', 'ready', 'cancelled'],         // 按复杂度分流
-  analyzing:       ['plan_review', 'cancelled', 'paused'],
+  analyzing:       ['plan_review', 'needs_attention', 'cancelled', 'paused'], // planner 拆解 / 失败超限升级
   plan_review:     ['decomposed', 'analyzing', 'cancelled'],               // accept / reject(+意见)
   decomposed:      ['done', 'analyzing', 'cancelled', 'paused'],           // 子全 done / 重拆
-  speccing:        ['spec_review', 'cancelled', 'paused'],
+  speccing:        ['spec_review', 'needs_attention', 'cancelled', 'paused'], // planner 写方案 / 失败超限升级
   spec_review:     ['ready', 'speccing', 'cancelled'],                     // accept / reject(+意见)
   ready:           ['blocked', 'queued', 'speccing', 'analyzing', 'cancelled', 'paused'],
   blocked:         ['ready', 'cancelled', 'paused'],
diff --git a/src/store/store.ts b/src/store/store.ts
index 26cac49..ff4dff8 100644
--- a/src/store/store.ts
+++ b/src/store/store.ts
@@ -660,11 +660,67 @@ export class Store {
     return this.getTask(taskId)!;
   }
 
+  /**
+   * planner（拆解 Hard / 写方案 Medium）失败的落点：收尾 planner run + 退避，任务【留在 analyzing/speccing】
+   * 等退避到期重新被领取；累计 planner 失败超 project.maxRetries → 升级 needs_attention（清退避）。
+   * 与 failTaskAttempt 的区别：planner 失败不把任务转 failed/queued（那是 executor 语义），只退避或升级。
+   */
+  failPlanAttempt(taskId: string, runId: string | null, error: string): Task {
+    const row = this.getTaskRow(taskId);
+    if (!row) throw new StoreError(`任务不存在: ${taskId}`);
+    const project = this.getProject(row.project_id);
+    if (!project) throw new StoreError(`项目不存在: ${row.project_id}`);
+
+    if (runId) {
+      const rr = this.db.prepare(`SELECT status FROM runs WHERE id = ?`).get(runId) as { status: string } | undefined;
+      if (rr && rr.status === 'started') this.finishRun(runId, 'failed', { error });
+    } else {
+      const r = this.startRun(taskId, 'planner');
+      this.finishRun(r.id, 'failed', { error });
+    }
+
+    const allFailed = (this.db.prepare(
+      `SELECT COUNT(*) AS n FROM runs WHERE task_id = ? AND kind = 'planner' AND status = 'failed'`,
+    ).get(taskId) as { n: number }).n;
+    const priorNetFailed = Math.max(0, allFailed - 1);                       // 不含本次
+    if (priorNetFailed < project.maxRetries) {
+      const attempt = priorNetFailed + 1;
+      const backoffMs = Math.min(30_000 * 2 ** (attempt - 1), 10 * 60_000);
+      this.setNextEligibleAt(taskId, new Date(Date.now() + backoffMs).toISOString());
+      // 任务保持 analyzing/speccing（不转状态）——退避到期后重新可领，再起 planner run
+      this.emit(row.project_id, taskId, 'task.updated', { field: 'plan-retry', attempt, backoffMs, error });
+    } else {
+      this.setNextEligibleAt(taskId, null);
+      this.transition(taskId, 'needs_attention', { by: 'failPlanAttempt', allFailed });
+    }
+    return this.getTask(taskId)!;
+  }
+
   /** 持久化退避：早于 next_eligible_at 不被领取（null=即刻可领）。 */
   setNextEligibleAt(taskId: string, iso: string | null): void {
     this.db.prepare(`UPDATE tasks SET next_eligible_at = ?, updated_at = ? WHERE id = ?`).run(iso, now(), taskId);
   }
 
+  /** 项目内"有 started run 的任务" id 集合（多进程并发闸 + claimable 排除在途；executor 与 planner 通用）。 */
+  inflightTaskIds(projectId: string): Set<string> {
+    const rows = this.db.prepare(
+      `SELECT DISTINCT r.task_id AS tid FROM runs r JOIN tasks t ON t.id = r.task_id
+       WHERE r.status = 'started' AND t.project_id = ?`,
+    ).all(projectId) as Array<{ tid: string }>;
+    return new Set(rows.map((r) => r.tid));
+  }
+
+  /** 所有 started run + 其任务（reap / ingest / reconcile 通用，覆盖 executor + planner + 残留复审 run）。 */
+  liveRunsWithTask(): Array<{ task: Task; run: Run }> {
+    const runs = this.db.prepare(`SELECT * FROM runs WHERE status = 'started' ORDER BY started_at`).all() as RunRow[];
+    const out: Array<{ task: Task; run: Run }> = [];
+    for (const r of runs) {
+      const t = this.getTaskRow(r.task_id);
+      if (t) out.push({ task: rowToTask(t), run: rowToRun(r) });
+    }
+    return out;
+  }
+
   /** 项目内 executing 任务数（多进程执行的并发闸：每个 executing 对应一个活 worker）。 */
   countExecuting(projectId: string): number {
     return (this.db.prepare(
@@ -691,9 +747,11 @@ export class Store {
    */
   reconcileInterrupted(isAlive: (run: Run | null) => boolean = () => false): { readopted: number; reclaimed: number } {
     let readopted = 0, reclaimed = 0;
-    for (const { task, run } of this.executingWithLatestExecutorRun()) {
-      if (run && isAlive(run)) { readopted++; continue; }
-      this.failTaskAttempt(task.id, run?.id ?? null, 'daemon 重启时发现 worker 已退出');
+    for (const { task, run } of this.liveRunsWithTask()) {
+      if (isAlive(run)) { readopted++; continue; }           // worker 仍活 → re-adopt，daemon 续 ingest
+      if (run.kind === 'executor') this.failTaskAttempt(task.id, run.id, 'daemon 重启时发现 worker 已退出');
+      else if (run.kind === 'planner') this.failPlanAttempt(task.id, run.id, 'daemon 重启时发现 planner worker 已退出');
+      else this.finishRun(run.id, 'failed', { error: 'daemon 重启中断' });  // 残留复审 run：仅收尾
       reclaimed++;
     }
     return { readopted, reclaimed };
diff --git a/test/ingest.test.ts b/test/ingest.test.ts
index 043851c..265881d 100644
--- a/test/ingest.test.ts
+++ b/test/ingest.test.ts
@@ -203,3 +203,68 @@ test('ingest 续读：先 ingest 半截（started/phase），再追加 result，
     store.close();
   });
 });
+
+// ───────────────────────── planner run 摄取 ─────────────────────────
+
+/** 真 Store + 一个 medium(speccing) 或 hard(analyzing) 任务 + 一条 planner run（模拟已领取规划）。 */
+function setupPlanning(complexity: 'medium' | 'hard'): { store: Store; taskId: string; runId: string } {
+  const store = new Store(':memory:');
+  const p = store.createProject({ name: 'plan', repoPath: '/tmp/plan-' + Math.random(), autonomy: 'auto-approved' });
+  const t = store.createTask({ projectId: p.id, title: '大任务', complexity });
+  // 建任务即落 speccing(medium)/analyzing(hard)，不转状态，直接起 planner run
+  const run = store.startRun(t.id, 'planner', { worktree: p.repoPath, branch: 'n/a' });
+  store.setWorkerPid(run.id, 4243);
+  return { store, taskId: t.id, runId: run.id };
+}
+
+test('ingest planner-spec：setSpec + 转 spec_review + 收尾 planner run', () => {
+  withTmpDataDir(() => {
+    const { store, taskId, runId } = setupPlanning('medium');
+    assert.equal(store.getTask(taskId)!.status, 'speccing');
+    appendOutbox(runId, { type: 'started', pid: 4243, worktree: '/r', branch: 'n/a', model: 'opus' });
+    appendOutbox(runId, { type: 'spec-result', spec: '## 改动\n改 a.ts', transcriptRef: '/p.jsonl', sessionId: 's' });
+    appendOutbox(runId, { type: 'done' });
+    ingestRun(store, noopLog, runId);
+    const t = store.getTask(taskId)!;
+    assert.equal(t.status, 'spec_review');
+    assert.equal(t.spec, '## 改动\n改 a.ts');
+    assert.equal(store.getRun(runId)!.status, 'succeeded');
+    store.close();
+  });
+});
+
+test('ingest planner-decompose：setPlan + 建子任务 + 转 plan_review', () => {
+  withTmpDataDir(() => {
+    const { store, taskId, runId } = setupPlanning('hard');
+    assert.equal(store.getTask(taskId)!.status, 'analyzing');
+    appendOutbox(runId, {
+      type: 'decompose-result', plan: '拆成两步',
+      subtasks: [{ title: '建骨架', complexity: 'easy' }, { title: '实现', complexity: 'medium' }],
+      transcriptRef: null, sessionId: null,
+    });
+    appendOutbox(runId, { type: 'done' });
+    ingestRun(store, noopLog, runId);
+    const t = store.getTask(taskId)!;
+    assert.equal(t.status, 'plan_review');
+    assert.equal(t.plan, '拆成两步');
+    const kids = store.listTasks(t.projectId).filter((x) => x.parentId === taskId);
+    assert.equal(kids.length, 2, '建了 2 个子任务');
+    assert.deepEqual(kids.map((k) => k.complexity).sort(), ['easy', 'medium']);
+    assert.equal(store.getRun(runId)!.status, 'succeeded');
+    store.close();
+  });
+});
+
+test('ingest planner failed：走 failPlanAttempt（退避，任务留 speccing，不进 failed）', () => {
+  withTmpDataDir(() => {
+    const { store, taskId, runId } = setupPlanning('medium');
+    appendOutbox(runId, { type: 'failed', error: 'planner 崩了', transcriptRef: null, sessionId: null });
+    appendOutbox(runId, { type: 'done' });
+    ingestRun(store, noopLog, runId);
+    const t = store.getTask(taskId)!;
+    assert.equal(t.status, 'speccing', '默认 maxRetries=2，第一次失败留 speccing 退避重试');
+    assert.ok(t.nextEligibleAt, '设了退避 nextEligibleAt');
+    assert.equal(store.getRun(runId)!.status, 'failed');
+    store.close();
+  });
+});
diff --git a/test/orchestrator.test.ts b/test/orchestrator.test.ts
index efc04ce..4530090 100644
--- a/test/orchestrator.test.ts
+++ b/test/orchestrator.test.ts
@@ -375,3 +375,68 @@ test(`reaper 反复回收：净失败 ${DEFAULT_MAX_RETRIES} 次后 → needs_at
   assert.equal(failed.length, DEFAULT_MAX_RETRIES + 1);
   store.close();
 });
+
+// ───────────────────────── planner：analyzing / speccing 也被领取 ─────────────────────────
+
+test('claimable 纳入 speccing：medium 任务派 planner-spec（不转状态、startRun=planner、job.runKind=planner-spec）', () => {
+  const { store, projectId } = setup('auto-approved');
+  const t = store.createTask({ projectId, title: 'feat', complexity: 'medium' });  // → speccing
+  assert.equal(store.getTask(t.id)!.status, 'speccing');
+  const state = freshState();
+  const { deps } = mockDeps(state);
+  const orch = createOrchestrator(store, noopLog, deps);
+
+  orch.claimTick();
+  assert.equal(state.spawned.length, 1, '派了一个 planner worker');
+  assert.equal(store.getTask(t.id)!.status, 'speccing', 'planner 不转任务状态（留 speccing 作规划中）');
+  const runs = store.listRuns(t.id);
+  assert.equal(runs.length, 1);
+  assert.equal(runs[0].kind, 'planner', 'run kind=planner');
+  assert.equal(runs[0].status, 'started');
+  assert.equal(state.jobs[0].runKind, 'planner-spec');
+
+  // 已在途（有 started planner run）→ 下一轮不重领
+  orch.claimTick();
+  assert.equal(state.spawned.length, 1, 'inflight 任务不被重复领取');
+  store.close();
+});
+
+test('claimable 纳入 analyzing：hard 任务派 planner-decompose', () => {
+  const { store, projectId } = setup('auto-approved');
+  const t = store.createTask({ projectId, title: 'epic', complexity: 'hard' });   // → analyzing
+  assert.equal(store.getTask(t.id)!.status, 'analyzing');
+  const state = freshState();
+  const { deps } = mockDeps(state);
+  createOrchestrator(store, noopLog, deps).claimTick();
+  assert.equal(state.spawned.length, 1);
+  assert.equal(store.getTask(t.id)!.status, 'analyzing', 'planner 不转状态');
+  assert.equal(store.listRuns(t.id)[0].kind, 'planner');
+  assert.equal(state.jobs[0].runKind, 'planner-decompose');
+  store.close();
+});
+
+test('auto-easy 不做规划：medium(speccing) 不被领取', () => {
+  const { store, projectId } = setup('auto-easy');
+  const t = store.createTask({ projectId, title: 'feat', complexity: 'medium' });  // → speccing
+  const state = freshState();
+  const { deps } = mockDeps(state);
+  createOrchestrator(store, noopLog, deps).claimTick();
+  assert.equal(state.spawned.length, 0, 'auto-easy 只做 easy 执行，不规划 medium/hard');
+  assert.equal(store.getTask(t.id)!.status, 'speccing');
+  store.close();
+});
+
+test('并发闸用 inflight：1 个 executing + 1 个 planner 占满 concurrency=2', () => {
+  const { store, projectId } = setup('auto-approved', 2);
+  const e = store.createTask({ projectId, title: 'easy', complexity: 'easy' });    // → ready
+  const m = store.createTask({ projectId, title: 'med', complexity: 'medium' });   // → speccing
+  store.createTask({ projectId, title: 'easy2', complexity: 'easy' });             // → ready（第三个）
+  const state = freshState();
+  const { deps } = mockDeps(state);
+  createOrchestrator(store, noopLog, deps).claimTick();
+  // concurrency=2：一个 executor(easy) + 一个 planner(medium) 占满，第三个不领
+  assert.equal(state.spawned.length, 2);
+  assert.equal(store.getTask(e.id)!.status, 'executing');
+  assert.equal(store.getTask(m.id)!.status, 'speccing');
+  store.close();
+});
diff --git a/test/pipeline.test.ts b/test/pipeline.test.ts
index 304bd84..6aaeede 100644
--- a/test/pipeline.test.ts
+++ b/test/pipeline.test.ts
@@ -1,6 +1,6 @@
 import { test } from 'node:test';
 import assert from 'node:assert/strict';
-import { runPipeline, type PipelineDeps } from '../src/executor/pipeline.js';
+import { runPipeline, parseDecompose, type PipelineDeps } from '../src/executor/pipeline.js';
 import type { JobSpec, OutboxPayload } from '../src/executor/protocol.js';
 import type { Project, Task } from '../src/model/types.js';
 import type { RunnerResult } from '../src/executor/runner.js';
@@ -55,6 +55,7 @@ function mockDeps(overrides: Partial<PipelineDeps> = {}): PipelineDeps {
     runTask: async () => okRun,
     reviewCode: async () => okReview,
     reviewSecurity: async () => okSecurity,
+    runPlanner: async () => ({ ok: true, transcriptRef: '/tmp/plan.jsonl', sessionId: 'sess-plan', finalText: '方案正文' }),
     ...overrides,
   };
 }
@@ -181,3 +182,62 @@ test('安全审计抛错不挡：security.verdict=null、summary 含「自动复
   assert.equal(result.code.verdict, 'approve');
   assert.equal(emitted.at(-1)?.type, 'done');
 });
+
+// ───────────────────────── planner 分支 ─────────────────────────
+
+const medJob = (over: Partial<JobSpec> = {}): JobSpec => ({ ...fakeJob, runKind: 'planner-spec', task: { ...fakeTask, complexity: 'medium', status: 'speccing' }, ...over });
+const hardJob = (over: Partial<JobSpec> = {}): JobSpec => ({ ...fakeJob, runKind: 'planner-decompose', task: { ...fakeTask, complexity: 'hard', status: 'analyzing' }, ...over });
+
+test('planner-spec：emit spec-result（方案=CC 文本）+ done，不复审/不建 worktree', async () => {
+  const { emitted, emit } = collector();
+  let worktreeCalled = false;
+  await runPipeline(medJob(), mockDeps({
+    createWorktree: async () => { worktreeCalled = true; return { dir: 'x', branch: 'y' }; },
+    runPlanner: async () => ({ ok: true, transcriptRef: '/tmp/p.jsonl', sessionId: 's', finalText: '## 改动\n改 a.ts' }),
+  }), emit);
+  assert.equal(worktreeCalled, false, 'planner 不建 worktree');
+  const sr = find(emitted, 'spec-result');
+  assert.ok(sr, '应 emit spec-result');
+  assert.equal(sr.spec, '## 改动\n改 a.ts');
+  assert.equal(find(emitted, 'result'), undefined);
+  assert.equal(find(emitted, 'failed'), undefined);
+  assert.equal(emitted.at(-1)?.type, 'done');
+});
+
+test('planner-decompose：解析末尾 fenced JSON → emit decompose-result + done', async () => {
+  const { emitted, emit } = collector();
+  const txt = '分析：拆成两步。\n```json\n{"plan":"先骨架再实现","subtasks":[{"title":"建骨架","complexity":"easy"},{"title":"实现逻辑","complexity":"medium"}]}\n```';
+  await runPipeline(hardJob(), mockDeps({ runPlanner: async () => ({ ok: true, transcriptRef: null, sessionId: null, finalText: txt }) }), emit);
+  const dr = find(emitted, 'decompose-result');
+  assert.ok(dr, '应 emit decompose-result');
+  assert.equal(dr.plan, '先骨架再实现');
+  assert.equal(dr.subtasks.length, 2);
+  assert.deepEqual(dr.subtasks.map((s) => s.complexity), ['easy', 'medium']);
+  assert.equal(emitted.at(-1)?.type, 'done');
+});
+
+test('planner-decompose：输出无合法 JSON → emit failed', async () => {
+  const { emitted, emit } = collector();
+  await runPipeline(hardJob(), mockDeps({ runPlanner: async () => ({ ok: true, transcriptRef: null, sessionId: null, finalText: '我忘了输出 JSON' }) }), emit);
+  assert.ok(find(emitted, 'failed'), '解析失败应 emit failed');
+  assert.equal(find(emitted, 'decompose-result'), undefined);
+  assert.equal(emitted.at(-1)?.type, 'done');
+});
+
+test('planner：CC 失败 → emit failed', async () => {
+  const { emitted, emit } = collector();
+  await runPipeline(medJob(), mockDeps({ runPlanner: async () => ({ ok: false, transcriptRef: null, sessionId: null, finalText: null, error: 'CC 崩了' }) }), emit);
+  const f = find(emitted, 'failed');
+  assert.ok(f);
+  assert.match(f.error, /CC 崩了/);
+  assert.equal(emitted.at(-1)?.type, 'done');
+});
+
+test('parseDecompose：取最后一个 json 块；非法/缺失 → null', () => {
+  assert.equal(parseDecompose('no json here'), null);
+  assert.equal(parseDecompose('```json\n{"subtasks":[]}\n```'), null, '空 subtasks → null');
+  const ok = parseDecompose('前文\n```json\n{"subtasks":[{"title":"a","complexity":"hard"},{"title":"","complexity":"easy"},{"title":"b","complexity":"x"}]}\n```');
+  assert.ok(ok);
+  assert.equal(ok.subtasks.length, 1, '过滤空标题与非法复杂度');
+  assert.equal(ok.subtasks[0].title, 'a');
+});