Consider it done. The open-source AI agent that works out of the box · 想到,就能做到。开源、开箱即用的 AI Agent。
TypeScript agentai-agentai-assistantandroidclaude-codecodexelectroniosllmmacosreact-nativetypescriptwindows
fix(orca): 取消未落地的 worker 唤醒,避免归档后被重新拉起
Zhang Yunjina816525b6448
diff --git a/apps/desktop/src/main/__tests__/makerSendToSessionOrdering.test.ts b/apps/desktop/src/main/__tests__/makerSendToSessionOrdering.test.tsindex d44d953537..133e21abfd 100644--- a/apps/desktop/src/main/__tests__/makerSendToSessionOrdering.test.ts+++ b/apps/desktop/src/main/__tests__/makerSendToSessionOrdering.test.ts@@ -789,6 +789,23 @@ describe('sendToSession ordering', () => {'orcaWorkerResumeScheduler.requestInBackground(target, (err) => {',);expect(switchFocusMcpBlock).not.toContain('await resumeOrcaWorkerSessionIfMissing(target)');+ // 关闭协同 / 归档 / 显式 idle 与冷启动交错时,唤醒必须可取消,并在 bootstrap+ // 返回后关掉刚拉起的 session(review P1:不得把已归档 worker 留在运行态)。+ expect(resumeBranch).toContain('if (resumeOpts?.isCancelled?.()) return false;');+ expect(resumeBranch).toContain('await maker.closeSession(resumedSession.id).catch((err) => {');+ expect(serviceDepsBlock).toContain(+ 'cancelWorkerResume: (sessionId) => orcaWorkerResumeScheduler.cancel(sessionId),',+ );+ });++ it('cancels pending worker resumes before releasing runtimes', () => {+ const disableOrcaBlock = extractBetween(+ source,+ 'async function disableOrcaInternal',+ 'ipcMain.handle(MAKER_INVOKE.SESSION_DISABLE_ORCA',+ );+ expect(disableOrcaBlock).toContain('orcaWorkerResumeScheduler.cancel(w.sessionId);');+ expect(orcaTeamServiceSource).toContain('deps.cancelWorkerResume?.(worker.sessionId);');});it('keeps IPC and MCP createWorker delegated to the shared lifecycle service', () => {diff --git a/apps/desktop/src/main/maker-ipc/__tests__/orcaWorkerResumeScheduler.test.ts b/apps/desktop/src/main/maker-ipc/__tests__/orcaWorkerResumeScheduler.test.tsindex ea65c45cd2..b023ef2add 100644--- a/apps/desktop/src/main/maker-ipc/__tests__/orcaWorkerResumeScheduler.test.ts+++ b/apps/desktop/src/main/maker-ipc/__tests__/orcaWorkerResumeScheduler.test.ts@@ -106,4 +106,93 @@ describe('createOrcaWorkerResumeScheduler', () => {await expect(scheduler.request({ sessionId: 'worker-1' })).resolves.toBe(true);expect(resume).toHaveBeenCalledTimes(2);});++ it('aborts a cancelled resume before it starts booting', async () => {+ let releaseLock!: () => void;+ const lockGate = new Promise<void>((resolve) => {+ releaseLock = resolve;+ });+ const resume = vi.fn(async (_target: unknown, isCancelled: () => boolean) => {+ return !isCancelled();+ });+ const scheduler = createOrcaWorkerResumeScheduler({+ resume,+ withSessionLock: async (_sessionId, task) => {+ await lockGate;+ return task();+ },+ });++ const pending = scheduler.request({ sessionId: 'worker-1' });+ await Promise.resolve();+ scheduler.cancel('worker-1');+ releaseLock();++ await expect(pending).resolves.toBe(false);+ expect(resume).toHaveBeenCalledTimes(1);+ expect(resume.mock.calls[0]?.[1]?.()).toBe(true);+ expect(scheduler.pendingCount()).toBe(0);+ });++ it('lets the resume implementation observe cancellation after bootstrap', async () => {+ let releaseBoot!: () => void;+ const bootGate = new Promise<void>((resolve) => {+ releaseBoot = resolve;+ });+ const resume = vi.fn(async (_target: unknown, isCancelled: () => boolean) => {+ await bootGate;+ return !isCancelled();+ });+ const scheduler = createOrcaWorkerResumeScheduler({+ resume,+ withSessionLock: (_sessionId, task) => task(),+ });++ const pending = scheduler.request({ sessionId: 'worker-1' });+ await Promise.resolve();+ scheduler.cancel('worker-1');+ releaseBoot();++ await expect(pending).resolves.toBe(false);+ expect(scheduler.pendingCount()).toBe(0);+ });++ it('keeps a cancelled in-flight entry deduplicated until it settles', async () => {+ let releaseLock!: () => void;+ const lockGate = new Promise<void>((resolve) => {+ releaseLock = resolve;+ });+ const resume = vi.fn(async (_target: unknown, isCancelled: () => boolean) => !isCancelled());+ const scheduler = createOrcaWorkerResumeScheduler({+ resume,+ withSessionLock: async (_sessionId, task) => {+ await lockGate;+ return task();+ },+ });++ const first = scheduler.request({ sessionId: 'worker-1' });+ await Promise.resolve();+ scheduler.cancel('worker-1');+ const second = scheduler.request({ sessionId: 'worker-1' });+ expect(second).toBe(first);+ expect(resume).toHaveBeenCalledTimes(0);++ releaseLock();+ await expect(first).resolves.toBe(false);+ expect(scheduler.pendingCount()).toBe(0);++ // A settled cancelled entry must not poison the next request.+ await expect(scheduler.request({ sessionId: 'worker-1' })).resolves.toBe(true);+ expect(resume).toHaveBeenCalledTimes(2);+ });++ it('ignores cancel for sessions without an in-flight resume', () => {+ const scheduler = createOrcaWorkerResumeScheduler({+ resume: async () => true,+ withSessionLock: (_sessionId, task) => task(),+ });+ expect(() => scheduler.cancel('worker-1')).not.toThrow();+ expect(scheduler.pendingCount()).toBe(0);+ });});diff --git a/apps/desktop/src/main/maker-ipc/orcaTeamService.ts b/apps/desktop/src/main/maker-ipc/orcaTeamService.tsindex 381f304328..235a2040f2 100644--- a/apps/desktop/src/main/maker-ipc/orcaTeamService.ts+++ b/apps/desktop/src/main/maker-ipc/orcaTeamService.ts@@ -230,6 +230,8 @@ export interface OrcaTeamServiceDeps {worker: OrcaWorkerRecordSnapshot,link: OrcaWorkerLinkSnapshot,): Promise<void>;+ /** 取消该 worker session 尚未落地的后台唤醒(focus 预热);释放 runtime 前调用。 */+ cancelWorkerResume?(sessionId: string): void;updateWorkerStatus(workerId: string, status: OrcaWorkerStatus): Promise<void>;markWorkerIdle(workerId: string): Promise<void>;markWorkerIdleIfStatus(workerId: string, expectedStatus: 'done'): Promise<boolean>;@@ -1195,6 +1197,8 @@ export function createOrcaTeamService(deps: OrcaTeamServiceDeps): OrcaTeamServic};}}+ // 显式 idle 会释放 runtime;先取消未落地的后台预热,避免释放后被重新拉起。+ deps.cancelWorkerResume?.(worker.sessionId);clearRuntimeState(worker.sessionId);if (!params.expectedStatus) {await closeWorkerSessionBestEffort(worker.sessionId, 'idleWorker');@@ -1231,6 +1235,8 @@ export function createOrcaTeamService(deps: OrcaTeamServiceDeps): OrcaTeamServicclearRuntimeState(worker.sessionId);deps.forgetWorkerSession?.(worker.sessionId);deferredDoneAcknowledgements.delete(worker.id);+ // 归档会释放 runtime;先取消未落地的后台预热,避免归档后被重新拉起(review P1)。+ deps.cancelWorkerResume?.(worker.sessionId);await deps.cancelWorkerSessionOperations(worker.sessionId);await closeWorkerSessionBestEffort(worker.sessionId, 'archiveWorker');await deps.archiveWorkerSession(worker.sessionId);diff --git a/apps/desktop/src/main/maker-ipc/orcaWorkerResumeScheduler.ts b/apps/desktop/src/main/maker-ipc/orcaWorkerResumeScheduler.tsindex 2498b9e9b1..b03e90ef5e 100644--- a/apps/desktop/src/main/maker-ipc/orcaWorkerResumeScheduler.ts+++ b/apps/desktop/src/main/maker-ipc/orcaWorkerResumeScheduler.ts@@ -7,12 +7,15 @@* gateway)才返回,面板切换被冷启动阻塞。唤醒本身仍然需要(发送、派活都依赖 live* runtime),但它可以后台化:worker 历史来自 DB,不需要 runtime。*- * 本调度器给出两个保证:+ * 本调度器给出三个保证:* 1. per-session 去重:并发的 focus 切换 / 派活共享同一次 resume,不会把同一个会话* bootstrap 两次(两个调用方都会看到同一结果);* 2. 与发送路径共用同一把 per-session 锁(由调用方注入 `withSessionLock`,main 侧注入* `withSendToSessionLock`):resume 期间到达的发送排在锁后;发送先持锁完成 lazy- * bootstrap 时,resume 在锁内重查 live 并直接跳过。+ * bootstrap 时,resume 在锁内重查 live 并直接跳过;+ * 3. 可取消:关闭协同 / 归档 / 显式 idle 等释放路径通过 `cancel` 取消尚未落地的唤醒。+ * 等锁中的唤醒直接放弃;已经在冷启动的由 `resume` 实现按 `isCancelled` 在+ * bootstrap 返回后关掉刚起的 session,避免把已归档 worker 重新拉活。** 唤醒实现与锁都由调用方注入,逻辑本身可单测。*/@@ -23,8 +26,12 @@ export interface OrcaWorkerResumeTarget {}export interface OrcaWorkerResumeSchedulerDeps<Target extends OrcaWorkerResumeTarget> {- /** 真正的唤醒实现;返回是否真的启动了 runtime(已 live 时 false)。 */- resume(target: Target): Promise<boolean>;+ /**+ * 真正的唤醒实现;返回是否真的启动了 runtime(已 live 时 false)。+ * `isCancelled` 在释放路径取消本次唤醒后变为 true:实现必须在冷启动前检查一次,+ * 并在 bootstrap 返回后再检查一次——已启动的 session 要自己关掉。+ */+ resume(target: Target, isCancelled: () => boolean): Promise<boolean>;/** per-session 串行锁;必须与发送路径共用同一把。 */withSessionLock<T>(sessionId: string, task: () => Promise<T>): Promise<T>;}@@ -37,32 +44,51 @@ export interface OrcaWorkerResumeScheduler<Target extends OrcaWorkerResumeTarget* focus 切换用这个入口,resume 结果不构成切换成功与否的一部分。*/requestInBackground(target: Target, onError: (error: unknown) => void): void;+ /**+ * 取消该 session 尚未落地的唤醒。取消不删除 in-flight 条目——条目要等真正 settle,+ * 期间新的 request 仍复用它,避免同一会话被 boot 两次;settle 之后的新请求照常重试。+ */+ cancel(sessionId: string): void;/** 诊断/测试用:当前处于 in-flight 的 session 数。 */pendingCount(): number;}+interface InFlightResume {+ cancelled: boolean;+ promise: Promise<boolean>;+}+export function createOrcaWorkerResumeScheduler<Target extends OrcaWorkerResumeTarget>(deps: OrcaWorkerResumeSchedulerDeps<Target>,): OrcaWorkerResumeScheduler<Target> {- const inFlight = new Map<string, Promise<boolean>>();+ const inFlight = new Map<string, InFlightResume>();function request(target: Target): Promise<boolean> {const existing = inFlight.get(target.sessionId);- if (existing) return existing;+ if (existing) return existing.promise;+ const entry: InFlightResume = {+ cancelled: false,+ promise: Promise.resolve(false),+ };// 锁在 resume 内部持有整个冷启动窗口:与发送、idle 释放等同一 session 的 critical// section 串行;失败也要清掉 in-flight,让下一次 focus / 派活可以重试。- const promise = deps- .withSessionLock(target.sessionId, () => deps.resume(target))+ entry.promise = deps+ .withSessionLock(target.sessionId, () => deps.resume(target, () => entry.cancelled)).finally(() => {- if (inFlight.get(target.sessionId) === promise) inFlight.delete(target.sessionId);+ if (inFlight.get(target.sessionId) === entry) inFlight.delete(target.sessionId);});- inFlight.set(target.sessionId, promise);- return promise;+ inFlight.set(target.sessionId, entry);+ return entry.promise;}function requestInBackground(target: Target, onError: (error: unknown) => void): void {void request(target).catch(onError);}- return { request, requestInBackground, pendingCount: () => inFlight.size };+ function cancel(sessionId: string): void {+ const entry = inFlight.get(sessionId);+ if (entry) entry.cancelled = true;+ }++ return { request, requestInBackground, cancel, pendingCount: () => inFlight.size };}diff --git a/apps/desktop/src/main/maker-ipc/register.ts b/apps/desktop/src/main/maker-ipc/register.tsindex 27779da878..81807a90db 100644--- a/apps/desktop/src/main/maker-ipc/register.ts+++ b/apps/desktop/src/main/maker-ipc/register.ts@@ -7373,14 +7373,20 @@ export function registerMakerIpc(maker: Maker, options: RegisterMakerIpcOptions)}// switchFocus 和 sendToWorker 都可能唤醒 idle worker;统一走这里才能保留 extraDirs。- async function resumeOrcaWorkerSessionIfMissing(target: {- id: string;- teamId: string;- leadSessionId: string;- sessionId: string;- }): Promise<boolean> {+ // resumeOpts.isCancelled:关闭协同 / 归档 / 显式 idle 会取消尚未落地的唤醒;冷启动前与+ // bootstrap 返回后都要检查,已经拉起的 runtime 要自己关掉(review P1)。+ async function resumeOrcaWorkerSessionIfMissing(+ target: {+ id: string;+ teamId: string;+ leadSessionId: string;+ sessionId: string;+ },+ resumeOpts?: { isCancelled?: () => boolean },+ ): Promise<boolean> {const live = maker.getSession(target.sessionId);if (live) return false;+ if (resumeOpts?.isCancelled?.()) return false;const db = getDbClient().drizzle;const [row] = await db@@ -7389,6 +7395,9 @@ export function registerMakerIpc(maker: Maker, options: RegisterMakerIpcOptions).where(eq(sessions.id, target.sessionId)).limit(1);if (!row) return false;+ // 归档 / 删除后的 worker 不允许被唤醒:关闭协同与 archive 可能和后台预热交错。+ if (row.status !== 'active') return false;+ if (resumeOpts?.isCancelled?.()) return false;const workerVendorOptions = {orcaRole: 'worker' as const,@@ -7419,22 +7428,55 @@ export function registerMakerIpc(maker: Maker, options: RegisterMakerIpcOptions)...(writableDirs.length > 0 ? { writableDirs } : {}),});await ensureRemoteReadyForSessionStart({ createOpts: opts });+ if (resumeOpts?.isCancelled?.()) return false;const { session: resumedSession } = await bootstrapSession(opts);+ // 关闭协同 / 归档 / 显式 idle 可能和冷启动交错:bootstrap 返回后必须重新确认这次+ // 唤醒仍然有效,否则关掉刚拉起的 runtime,不能把已归档 worker 留在运行态。+ if (resumeOpts?.isCancelled?.() || !(await isOrcaWorkerSessionResumable(target.sessionId))) {+ await maker.closeSession(resumedSession.id).catch((err) => {+ log.warn('orcaWorkerResume: cancelled after bootstrap, close failed', {+ sessionId: resumedSession.id,+ err: err instanceof Error ? err.message : String(err),+ });+ });+ return false;+ }await markOrcaRoleIfNeeded(resumedSession.id, 'worker');return true;}+ /** 唤醒后的会话仍必须是 active;读不到状态按不可唤醒处理(fail-closed)。 */+ async function isOrcaWorkerSessionResumable(sessionId: string): Promise<boolean> {+ try {+ const [row] = await getDbClient()+ .drizzle.select({ status: sessions.status })+ .from(sessions)+ .where(eq(sessions.id, sessionId))+ .limit(1);+ return row?.status === 'active';+ } catch (err) {+ log.warn('orcaWorkerResume: read session status failed', {+ sessionId,+ err: err instanceof Error ? err.message : String(err),+ });+ return false;+ }+ }+// focus 切换是纯 UI 操作,不再同步等待冷启动(2026-09 实报: 首次切到 dormant worker// 要等 ~5s)。resume 仍要发生,但转后台并走共享调度器: per-session 去重 + 与发送路径// 共用 sendToSession 锁,避免并发双 bootstrap(实现与不变量见 orcaWorkerResumeScheduler.ts)。const orcaWorkerResumeScheduler = createOrcaWorkerResumeScheduler({- resume: async (target: {- id: string;- teamId: string;- leadSessionId: string;- sessionId: string;- }) => {- const didResume = await resumeOrcaWorkerSessionIfMissing(target);+ resume: async (+ target: {+ id: string;+ teamId: string;+ leadSessionId: string;+ sessionId: string;+ },+ isCancelled,+ ) => {+ const didResume = await resumeOrcaWorkerSessionIfMissing(target, { isCancelled });if (didResume) {log.info('orcaWorkerResume: resumed idle worker (session only, no status change)', {workerId: target.id,@@ -10973,6 +11015,11 @@ export function registerMakerIpc(maker: Maker, options: RegisterMakerIpcOptions)const workers = await listWorkersByLead(leadSessionId);const activeWorkers = workers.filter((w) => w.teamId === team.id);+ // 先取消尚未落地的后台预热:关闭流程不等待 sendToSession 锁,若不取消,等锁中或+ // 冷启动中的 resume 会在归档之后把 worker 重新拉活(review P1)。+ for (const w of workers) {+ orcaWorkerResumeScheduler.cancel(w.sessionId);+ }for (const w of activeWorkers) {orcaTeamService.clearAutoBridgeState(w.sessionId);await cancelIOSSimulatorSessionOperations(w.sessionId);@@ -11154,6 +11201,8 @@ export function registerMakerIpc(maker: Maker, options: RegisterMakerIpcOptions)// 与 focus 后台预热共享同一套去重/锁:派活不会和切换触发两次 bootstrap。await orcaWorkerResumeScheduler.request(target);},+ // 归档 / 显式 idle 会释放 runtime;先取消未落地的后台预热,避免释放后被重新拉起。+ cancelWorkerResume: (sessionId) => orcaWorkerResumeScheduler.cancel(sessionId),updateWorkerStatus,markWorkerIdle: async (workerId) => {const now = Date.now();diff --git a/docs/dev-rules/orca-team-architecture.md b/docs/dev-rules/orca-team-architecture.mdindex 50e5a8c67e..b4816ffe2f 100644--- a/docs/dev-rules/orca-team-architecture.md+++ b/docs/dev-rules/orca-team-architecture.md@@ -320,7 +320,7 @@ Worker turn 被 vendor 报终止型 error,但 interrupted-turn auto-resume 仍idle worker 被 `switch_focus` 或 `send_to_worker` 唤醒时,要从 DB 读取 `extra_dirs` 并带回 `bootstrapSession`,否则恢复后的 worker 会丢附加目录上下文。实现指针:`register.ts` 的 idle worker resume helper。3a. **focus 切换不等待 worker 冷启动(状态:不变量)**<br>- `switch_focus` 是纯 UI focus 操作:`setWorkerFocus` 成功并广播 `ORCA_WORKER_CHANGED` 后必须立即返回,不得同步 `await` 冷会话的 resume / `bootstrapSession`(Pi 冷启实测 2~3s,会阻塞面板切换)。idle worker 的 runtime 预热转后台,并与发送/派活路径共享同一套 per-session 去重 + `sendToSession` 锁:并发 focus / 派活只 bootstrap 一次;发送先取得锁完成懒恢复时,后台 resume 在锁内重查 live 并跳过;预热失败只告警,不改变 `switch_focus` 结果。worker 历史来自 DB,展示不需要 live runtime。实现指针:`register.ts` 的 `resumeOrcaWorkerSessionIfMissing` / `orcaWorkerResumeScheduler.ts`。+ `switch_focus` 是纯 UI focus 操作:`setWorkerFocus` 成功并广播 `ORCA_WORKER_CHANGED` 后必须立即返回,不得同步 `await` 冷会话的 resume / `bootstrapSession`(Pi 冷启实测 2~3s,会阻塞面板切换)。idle worker 的 runtime 预热转后台,并与发送/派活路径共享同一套 per-session 去重 + `sendToSession` 锁:并发 focus / 派活只 bootstrap 一次;发送先取得锁完成懒恢复时,后台 resume 在锁内重查 live 并跳过;预热失败只告警,不改变 `switch_focus` 结果。worker 历史来自 DB,展示不需要 live runtime。后台预热必须可取消:关闭协同 / 归档 / 显式 idle 在释放 runtime 前先取消未落地的唤醒;冷启动前与 `bootstrapSession` 返回后都要复核会话仍为 active,失效则关掉刚拉起的 runtime,不得把已归档 worker 留在运行态。实现指针:`register.ts` 的 `resumeOrcaWorkerSessionIfMissing` / `disableOrcaInternal` / `orcaWorkerResumeScheduler.ts`、`orcaTeamService.ts` 的 `cancelWorkerResume`。4. **worker 状态变更必须广播给 renderer(状态:不变量)**创建 worker、`enableOrca` 创建首个 worker、任意真实 worker turn 开始后的 running、idle、archive、terminal done/error 都必须广播 `ORCA_WORKER_CHANGED`。worker DB `status` 跟随真实 turn 生命周期:Lead 派活或用户直接对话 worker 时,只有真实 turn 开始才置 `running`,**产品** terminal 才置 `done/error`(auto-resume 仍 pending/deferred 的 vendor 终态不是产品终态,见消息派发与 auto-bridge 5a);`switch_focus` / resume / restore 只能恢复可访问性,不能凭空置 `running`。实现指针:`orcaLifecycleService.ts` 的 `createWorker` / `enableTeam`、`orcaWorkerCreationService.ts` 的 `createWorkerInTeam`、`orcaTeamService.ts` 的 `dispatchWorkerTask` / `handleWorkerTurnStarted` / `handleWorkerTerminalTurn`、`register.ts` 的 status event adapter、`useWorkers.ts` 的 `useWorkers`。