feat(planner): 编排器把 analyzing/speccing 也当可执行——派 planner run 自动拆解/写方案
修复"hard/medium 任务建后永远停在 analyzing/speccing"的缺口：编排器现在对 analyzing(hard) 派 planner-decompose、对 speccing(medium) 派 planner-spec（auto-approved 下；auto-easy 仍只执行 easy）， planner 只读跑 CC（无 worktree/无写）产出 → daemon 落库 + 自动提交到 plan_review/spec_review 闸（仍人审）。 - protocol: JobSpec.runKind + spec-result/decompose-result 两型 outbox + DecomposeResult - runner: runPlanner（只读 Read/Glob/Grep 跑在 repo，opus 档）+ buildPlannerPrompt - pipeline: runKind 分支 + parseDecompose（取末尾 fenced JSON，非法→failed） - orchestrator: claimable 纳入 analyzing/speccing + 排除在途；并发/in-flight 从 countExecuting 泛化为 inflightTaskIds（有 started run 的任务，executor+planner 通用）；claimOne 按状态分 executor/planner （planner 不转状态）；reap 泛化（死 planner→failPlanAttempt） - ingest: ingestAll 覆盖所有 started run；spec-result→setSpec+spec_review；decompose-result→setPlan+建子任务+plan_review； failed 按 kind 分流（planner→failPlanAttempt 退避留态、超限→needs_attention） - store: inflightTaskIds / liveRunsWithTask / failPlanAttempt；reconcile 泛化到所有 started run - status: analyzing/speccing 加 →needs_attention（planner 失败超限升级） - models: planner 角色（opus）；status: RunKind 已含 planner - 测试 +12（pipeline 5 / ingest 3 / orchestrator 4），194 全绿 Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
wangjia
+45
-10
@@ -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);
|
||||
}
|
||||
}
|
||||
|
||||
+45
-24
@@ -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}`);
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user