@zhushanwen/subagent-core 0.4.0 → 0.5.1

package/src/execution/__tests__/delivery-methods.test.ts

@@ -1,14 +1,14 @@
 // src/execution/__tests__/delivery-methods.test.ts
 //
-// resumeRound + deliverMessage 单元测试（M2-B1 投递基础设施）。
+// 冷路径续轮 + chatMode 统一投递单元测试（M2-B1 投递基础设施；D2 单轨后口径）。
 //
 // mock session-runner（runSpawn 受控 + killAllSpawnedChildren 空实现 + getChildByRecord/spawnedChildren
-// 真实 Map 语义），走 SubagentService 真实的 resumeRound/deliverMessage 逻辑：
-//   - resumeRound：idle→running→detached kickOff runSpawn 收到 resume 参数；chatMode+done 回 idle/round+1；
-//     非 idle / 无 sessionFile / 无 controller throw
-//   - deliverMessage：按进程死活分流（热路径 prompt+streamingBehavior / 冷路径 resume）
+// 真实 Map 语义），走 SubagentService.deliverChatMessage → PiEngine.deliverPrompt 真实逻辑：
+//   - 冷路径（进程死）：续轮 resume spawn（runSpawn 收到 resume 参数）；chatMode+done 回 running/round+1；
+//     终态 / 无 sessionFile / 无 controller throw 行动语言
+//   - 热路径（进程活）：prompt + streamingBehavior（interrupt → steer/followUp）
 //
-// stdin-writer 不 mock（端到端验证 deliverMessage→sendPromptCommand→child.stdin 字节）。
+// stdin-writer 不 mock（端到端验证 PiEngine.deliverPrompt→sendPromptCommand→child.stdin 字节）。
 // [review 修复] 已删除 deliverToRunning describe（busy follow_up/steer 投递）——随
 // deliverToRunning 方法一并移除（无生产调用方，pendingMessages 消费确认制死机制）。
 
@@ -29,7 +29,7 @@ vi.mock("../../core/logger.ts", () => ({ getLogger: () => loggerMock }));
 
 // mock session-runner：runSpawn 受控 + killAllSpawnedChildren 空实现；
 // getChildByRecord/spawnedChildren 提供真实 Map 语义（deliverToRunning 注册 mock child 用）。
-vi.mock("../session-runner.ts", () => {
+vi.mock("../engine/engines/pi/session-runner.ts", () => {
   const spawnedChildren = new Map<string, unknown>();
   return {
     runSpawn: vi.fn(),
@@ -41,7 +41,7 @@ vi.mock("../session-runner.ts", () => {
   };
 });
 
-import { runSpawn, spawnedChildren, type SessionRunnerContext } from "../session-runner.ts";
+import { runSpawn, spawnedChildren, type SessionRunnerContext } from "../engine/engines/pi/session-runner.ts";
 import * as lifecycle from "../lifecycle-manager.ts";
 import { createRecord } from "../execution-record.ts";
 import { ModelConfigService } from "../model-config-service.ts";
@@ -59,11 +59,15 @@ function makeTmpAgentDir(): string {
 }
 
 function makePi(): PiLike & {
-  appendEntry: ReturnType<typeof vi.fn>;
-  events: { emit: ReturnType<typeof vi.fn> };
-  sendMessage: ReturnType<typeof vi.fn>;
+  appendEntry: ReturnType<typeof vi.fn<(customType: string, data?: unknown) => void>>;
+  events: { emit: ReturnType<typeof vi.fn<(channel: string, data: unknown) => void>> };
+  sendMessage: ReturnType<typeof vi.fn<(message: Parameters<PiLike["sendMessage"]>[0], options?: Parameters<PiLike["sendMessage"]>[1]) => void>>;
 } {
-  return { appendEntry: vi.fn(), events: { emit: vi.fn() }, sendMessage: vi.fn() };
+  return {
+    appendEntry: vi.fn((customType: string, data?: unknown) => {}),
+    events: { emit: vi.fn((channel: string, data: unknown) => {}) },
+    sendMessage: vi.fn(() => {}),
+  };
 }
 
 function makeResult(success: boolean): AgentResult {
@@ -100,7 +104,7 @@ function makeIdleRecord(id = "sa-chat"): ExecutionRecord {
 
 /** PassThrough child（可读出 stdin 字节验证 deliverToRunning 写入）。 */
 function makeStreamChild(): ChildProcess {
-  // exitCode/signalCode：真 ChildProcess 未退出时均为 null（deliverMessage [race-F5]
+  // exitCode/signalCode：真 ChildProcess 未退出时均为 null（热路径投递 [race-F5]
   // 写后死进程检测读这两个字段，缺省 undefined 会被误判为已死触发 warn）。
   return { stdin: new PassThrough(), exitCode: null, signalCode: null } as unknown as ChildProcess;
 }
@@ -108,7 +112,7 @@ function makeStreamChild(): ChildProcess {
 function readStdinLines(child: ChildProcess): unknown[] {
   const stream = child.stdin as unknown as PassThrough;
   stream.pause();
-  const text = stream.read()?.toString() ?? "";
+  const text: string = stream.read()?.toString() ?? "";
   return text
     .split("\n")
     .map((l) => l.trim())
@@ -116,14 +120,14 @@ function readStdinLines(child: ChildProcess): unknown[] {
     .map((l) => JSON.parse(l));
 }
 
-describe("resumeRound (M2-B1 idle 投递)", () => {
+describe("冷路径续轮（M2-B1 idle 投递；D2 后经 deliverChatMessage 无活进程到达）", () => {
   let agentDir: string;
   let service: SubagentService;
   let record: ExecutionRecord;
 
   beforeEach(() => {
     agentDir = makeTmpAgentDir();
-    const modelService = new ModelConfigService({ agentDir });
+    const modelService = new ModelConfigService({ agentDir, cwd: agentDir });
     service = new SubagentService({ cwd: agentDir, modelService });
     service.initSession({ pi: makePi(), sessionId: "root-session" });
     record = makeIdleRecord();
@@ -134,16 +138,16 @@ describe("resumeRound (M2-B1 idle 投递)", () => {
 
   afterEach(() => {
     service.dispose();
-    fs.rmSync(agentDir, { recursive: true, force: true });
+    fs.rmSync(agentDir, { recursive: true, force: true, maxRetries: 5, retryDelay: 20 });
   });
 
-  it("resumeRound(running) → kickOff runSpawn 收到 resume 参数；chatMode+done 回 running/round+1", async () => {
+  it("冷路径续轮(running) → kickOff runSpawn 收到 resume 参数；chatMode+done 回 running/round+1", async () => {
     mockRunSpawn.mockResolvedValueOnce(makeResult(true));
     const beforeRound = record.round;
 
-    service.resumeRound(record, "next round msg");
+    await service.chatActions.deliverChatMessage(record, "next round msg", false);
 
-    // 同步：status 已手动设回 running（M2-A 边界，绕过 tryTransition）
+    // 冷路径守卫通过后：status 已手动设回 running（M2-A 边界，绕过 tryTransition）
     expect(record.status).toBe("running");
 
     // detached：等 runSpawn 被调
@@ -163,42 +167,42 @@ describe("resumeRound (M2-B1 idle 投递)", () => {
     expect(record.round).toBe(beforeRound! + 1);
   });
 
-  it("终态 closed record → throw 行动语言（MF-4，仅 running 可续聊），不触发 kickOff", () => {
+  it("终态 closed record → throw 行动语言（MF-4，仅 running 可续聊），不触发 kickOff", async () => {
     record.status = "closed";
     // MF-4：行动语言（spec §3.1），不暴露 resume/controller 内部词汇
-    expect(() => service.resumeRound(record, "msg")).toThrow(/not ready for a new message/);
+    await expect(service.chatActions.deliverChatMessage(record, "msg", false)).rejects.toThrow(/not ready for a new message/);
     expect(mockRunSpawn).not.toHaveBeenCalled();
   });
 
-  it("record 无 sessionFile → throw 行动语言（MF-4 canonical session unavailable），不触发 kickOff", () => {
+  it("record 无 sessionFile → throw 行动语言（MF-4 canonical session unavailable），不触发 kickOff", async () => {
     record.sessionFile = undefined;
-    expect(() => service.resumeRound(record, "msg")).toThrow(/session unavailable/);
+    await expect(service.chatActions.deliverChatMessage(record, "msg", false)).rejects.toThrow(/session unavailable/);
     expect(mockRunSpawn).not.toHaveBeenCalled();
   });
 
-  it("record 无 controller → throw 行动语言（MF-4），不触发 kickOff", () => {
+  it("record 无 controller → throw 行动语言（MF-4），不触发 kickOff", async () => {
     record.controller = undefined;
-    expect(() => service.resumeRound(record, "msg")).toThrow(/not ready for a new message/);
+    await expect(service.chatActions.deliverChatMessage(record, "msg", false)).rejects.toThrow(/not ready for a new message/);
     expect(mockRunSpawn).not.toHaveBeenCalled();
   });
 });
 
 // ============================================================
-// deliverMessage（V2 决策 3 chatMode 统一投递：按进程死活分流）
+// deliverChatMessage（V2 决策 3 chatMode 统一投递：按进程死活分流；D2 后经 engine.interactRecord）
 // ============================================================
 
-describe("deliverMessage (V2 决策 3 chatMode 统一投递)", () => {
+describe("deliverChatMessage (V2 决策 3 chatMode 统一投递)", () => {
   let agentDir: string;
   let service: SubagentService;
   let record: ExecutionRecord;
 
   beforeEach(() => {
     agentDir = makeTmpAgentDir();
-    const modelService = new ModelConfigService({ agentDir });
+    const modelService = new ModelConfigService({ agentDir, cwd: agentDir });
     service = new SubagentService({ cwd: agentDir, modelService });
     service.initSession({ pi: makePi(), sessionId: "root-session" });
     record = makeIdleRecord(); // chatMode:true, idle, round=1
-    // sessionFile：冷路径 resumeRound 需要（热路径不用，设了无害）
+    // sessionFile：冷路径续轮需要（热路径不用，设了无害）
     record.sessionFile = path.join(agentDir, "fake-session.jsonl");
     spawnedChildren.clear();
     mockRunSpawn.mockReset();
@@ -209,15 +213,15 @@ describe("deliverMessage (V2 决策 3 chatMode 统一投递)", () => {
     service.dispose();
     spawnedChildren.clear();
     lifecycle._resetLifecycleState();
-    fs.rmSync(agentDir, { recursive: true, force: true });
+    fs.rmSync(agentDir, { recursive: true, force: true, maxRetries: 5, retryDelay: 20 });
   });
 
-  it("热路径 interrupt=false：进程活 → prompt streamingBehavior:followUp + status=running + pid 记录", () => {
+  it("热路径 interrupt=false：进程活 → prompt streamingBehavior:followUp + status=running + pid 记录", async () => {
     const child = makeStreamChild();
     Object.assign(child, { pid: 12345 });
     spawnedChildren.set(record.id, child);
 
-    service.deliverMessage(record, "after you finish", false);
+    await service.chatActions.deliverChatMessage(record, "after you finish", false);
 
     const lines = readStdinLines(child);
     expect(lines).toHaveLength(1);
@@ -227,11 +231,11 @@ describe("deliverMessage (V2 决策 3 chatMode 统一投递)", () => {
     expect(record.pid).toBe(12345);
   });
 
-  it("热路径 interrupt=true：进程活 → prompt streamingBehavior:steer", () => {
+  it("热路径 interrupt=true：进程活 → prompt streamingBehavior:steer", async () => {
     const child = makeStreamChild();
     spawnedChildren.set(record.id, child);
 
-    service.deliverMessage(record, "stop now", true);
+    await service.chatActions.deliverChatMessage(record, "stop now", true);
 
     const lines = readStdinLines(child);
     expect(lines[0]).toMatchObject({ type: "prompt", message: "stop now", streamingBehavior: "steer" });
@@ -241,13 +245,13 @@ describe("deliverMessage (V2 决策 3 chatMode 统一投递)", () => {
   // [race-F5] 写后死进程检测：热路径 write 同步成功（数据进内核缓冲）但子进程在读取前
   // 已死（gate/idle kill 竞速）→ 消息将随缓冲静默丢弃。修复：写后检查 exitCode/signalCode，
   // 已死则 warn 留证（含 runId 与消息类型），不抛错不重试（终态已由 kill 路径保证）。
-  it("[race-F5] 热路径写后子进程已死 → logger.warn 被记录（含 runId 与消息类型），不抛错", () => {
+  it("[race-F5] 热路径写后子进程已死 → logger.warn 被记录（含 runId 与消息类型），不抛错", async () => {
     const child = makeStreamChild();
     // 模拟 gate/idle kill 竞速：写 stdin 成功但子进程已死（SIGTERM 终止形态）
     Object.assign(child, { signalCode: "SIGTERM" });
     spawnedChildren.set(record.id, child);
 
-    expect(() => service.deliverMessage(record, "lost msg", true)).not.toThrow();
+    await expect(service.chatActions.deliverChatMessage(record, "lost msg", true)).resolves.toBeUndefined();
 
     // 写入照常发生（热路径语义不变，不做二次分发）
     const lines = readStdinLines(child);
@@ -259,13 +263,13 @@ describe("deliverMessage (V2 决策 3 chatMode 统一投递)", () => {
     );
   });
 
-  it("[race-F5] 活进程正常投递不触发死进程 warn（守卫不误报）", () => {
+  it("[race-F5] 活进程正常投递不触发死进程 warn（守卫不误报）", async () => {
     // loggerMock 是模块级共享 mock：清历史调用后再断言（防前一用例的 warn 干扰）
     loggerMock.warn.mockClear();
     const child = makeStreamChild();
     spawnedChildren.set(record.id, child);
 
-    service.deliverMessage(record, "normal", false);
+    await service.chatActions.deliverChatMessage(record, "normal", false);
 
     expect(loggerMock.warn).not.toHaveBeenCalledWith(
       expect.stringContaining("died around stdin write"),
@@ -273,24 +277,24 @@ describe("deliverMessage (V2 决策 3 chatMode 统一投递)", () => {
     );
   });
 
-  it("热路径 disarm idle timer：arm 后 deliverMessage → timer 清除（防 turn 期间误杀）", () => {
+  it("热路径 disarm idle timer：arm 后投递 → timer 清除（防 turn 期间误杀）", async () => {
     const child = makeStreamChild();
     spawnedChildren.set(record.id, child);
     // 先 arm idle timer（模拟 Step 4a agent_settled 后 armed）
     lifecycle.armIdleTimer(record.id, () => {}, 10000);
     expect(lifecycle.hasIdleTimer(record.id)).toBe(true);
 
-    service.deliverMessage(record, "msg", false);
+    await service.chatActions.deliverChatMessage(record, "msg", false);
 
     // disarmIdleTimer 被调 → timer 清除（新 turn 不被 idle timer 误杀）
     expect(lifecycle.hasIdleTimer(record.id)).toBe(false);
   });
 
-  it("冷路径：进程死（无 child）→ resumeRound spawn（runSpawn 收到 resume 参数）", async () => {
+  it("冷路径：进程死（无 child）→ 续轮 resume spawn（runSpawn 收到 resume 参数）", async () => {
     mockRunSpawn.mockResolvedValueOnce(makeResult(true));
     // spawnedChildren 无该 record → getChildByRecord 返回 undefined → 冷路径
 
-    service.deliverMessage(record, "resume msg", false);
+    await service.chatActions.deliverChatMessage(record, "resume msg", false);
 
     await vi.waitFor(() => expect(mockRunSpawn).toHaveBeenCalledTimes(1));
     const call = mockRunSpawn.mock.calls[0]!;
@@ -308,7 +312,7 @@ describe("deliverMessage (V2 决策 3 chatMode 统一投递)", () => {
     Object.assign(child, { killed: true }); // 进程已 kill
     spawnedChildren.set(record.id, child);
 
-    service.deliverMessage(record, "after kill", false);
+    await service.chatActions.deliverChatMessage(record, "after kill", false);
 
     await vi.waitFor(() => expect(mockRunSpawn).toHaveBeenCalledTimes(1));
   });
@@ -316,7 +320,7 @@ describe("deliverMessage (V2 决策 3 chatMode 统一投递)", () => {
   it("续聊后 agent_settled → onRoundSettled 设 running + round+1（v4 B-1 idle 折入 running，复用 Step 4a 链路）", () => {
     // buildSessionRunnerContext 是 private，类型断言访问（测试专用）
     const ctx = (service as unknown as { buildSessionRunnerContext(): SessionRunnerContext }).buildSessionRunnerContext();
-    record.status = "running"; // 模拟 deliverMessage 热路径设的新 turn 状态
+    record.status = "running"; // 模拟热路径投递设的新 turn 状态
     const beforeRound = record.round;
 
     // 模拟 session-runner 在 agent_settled 时调本回调（Step 4a 接入点）
@@ -331,21 +335,21 @@ describe("deliverMessage (V2 决策 3 chatMode 统一投递)", () => {
 // 冷路径并发守卫（review round2 MF1：同 turn 批量两条 message 双冷路径双 spawn）
 // ============================================================
 // 复现链（reviewer 探针实证）：pi 对同一条 assistant message 的 tool calls 顺序执行
-// （subagent tool sequential），tool1 的 deliverMessage 在冷路径 resumeRound 返回即
+// （subagent tool sequential），tool1 的投递在冷路径续轮返回即
 // resolve——早于 runSpawn 完成 spawn 注册（session-runner spawnedChildren.set 前有
 // pool.acquire await / writePromptToTempFile 等多个异步点）；tool2 立即执行 →
-// getChildByRecord 仍 undefined → 再次冷路径。v4 两态收敛后 resumeRound 的
+// getChildByRecord 仍 undefined → 再次冷路径。v4 两态收敛后续轮的
 // `status !== "running"` 守卫对 idle-resumable record 恒放行（idle 本来就是 running）、
 // `status = "running"` 是幂等写 → 两次 kickOff → runSpawn 被调 2 次 → 两个 pi 子进程
 // 以 --session 同一 JSONL 双写 + 第一个进程脱离 kill 记账成孤儿。
-describe("deliverMessage 冷路径并发守卫（review round2 MF1）", () => {
+describe("deliverChatMessage 冷路径并发守卫（review round2 MF1）", () => {
   let agentDir: string;
   let service: SubagentService;
   let record: ExecutionRecord;
 
   beforeEach(() => {
     agentDir = makeTmpAgentDir();
-    const modelService = new ModelConfigService({ agentDir });
+    const modelService = new ModelConfigService({ agentDir, cwd: agentDir });
     service = new SubagentService({ cwd: agentDir, modelService });
     service.initSession({ pi: makePi(), sessionId: "root-session" });
     record = makeIdleRecord();
@@ -359,7 +363,7 @@ describe("deliverMessage 冷路径并发守卫（review round2 MF1）", () => {
     service.dispose();
     spawnedChildren.clear();
     lifecycle._resetLifecycleState();
-    fs.rmSync(agentDir, { recursive: true, force: true });
+    fs.rmSync(agentDir, { recursive: true, force: true, maxRetries: 5, retryDelay: 20 });
   });
 
   it("同 record 连续两条 message 冷路径 → 第二条 throw 行动语言，runSpawn 仅 1 次", async () => {
@@ -371,13 +375,13 @@ describe("deliverMessage 冷路径并发守卫（review round2 MF1）", () => {
     );
 
     // 第一条：正常冷路径 resume
-    await expect(service.deliverMessage(record, "first msg", false)).resolves.toBeUndefined();
+    await expect(service.chatActions.deliverChatMessage(record, "first msg", false)).resolves.toBeUndefined();
     await vi.waitFor(() => expect(mockRunSpawn).toHaveBeenCalledTimes(1));
 
     // 第二条：spawn 仍在途（getChildByRecord undefined）→ 再走冷路径。
-    // 修复前：resumeRound 守卫恒放行 → 第二次 kickOff → runSpawn 2 次（双 spawn 双写 session）。
+    // 修复前：续轮守卫恒放行 → 第二次 kickOff → runSpawn 2 次（双 spawn 双写 session）。
     // 修复后：in-flight 守卫 throw 行动语言（MF-4）。
-    await expect(service.deliverMessage(record, "second msg", false)).rejects.toThrow(
+    await expect(service.chatActions.deliverChatMessage(record, "second msg", false)).rejects.toThrow(
       /already starting a new round/,
     );
     expect(mockRunSpawn).toHaveBeenCalledTimes(1);
@@ -388,7 +392,7 @@ describe("deliverMessage 冷路径并发守卫（review round2 MF1）", () => {
       expect((service as unknown as { resumesInFlight: Set<string> }).resumesInFlight.has(record.id)).toBe(false),
     );
     mockRunSpawn.mockResolvedValueOnce(makeResult(true));
-    await expect(service.deliverMessage(record, "third msg", false)).resolves.toBeUndefined();
+    await expect(service.chatActions.deliverChatMessage(record, "third msg", false)).resolves.toBeUndefined();
     await vi.waitFor(() => expect(mockRunSpawn).toHaveBeenCalledTimes(2));
     expect(mockRunSpawn.mock.calls[1]![1]).toBe("third msg");
   });
@@ -401,25 +405,34 @@ describe("deliverMessage 冷路径并发守卫（review round2 MF1）", () => {
       () => new Promise<AgentResult>((res) => { releaseA = res; }),
     );
 
-    await service.deliverMessage(record, "A msg", false);
+    await service.chatActions.deliverChatMessage(record, "A msg", false);
     await vi.waitFor(() => expect(mockRunSpawn).toHaveBeenCalledTimes(1));
 
     // B 的冷路径不受 A 在途影响
-    await expect(service.deliverMessage(recordB, "B msg", false)).resolves.toBeUndefined();
+    await expect(service.chatActions.deliverChatMessage(recordB, "B msg", false)).resolves.toBeUndefined();
     await vi.waitFor(() => expect(mockRunSpawn).toHaveBeenCalledTimes(2));
     expect(mockRunSpawn.mock.calls[1]![0]).toBe(recordB);
 
     releaseA(makeResult(true));
   });
 
-  it("resume 在途时直接调 resumeRound（不经 deliverMessage）同样被守卫拦截", () => {
-    // promise 同步创建：release 立即可用（runSpawn 在 pool.acquire 微任务后才被调）
+  it("冷路径续轮在途时 EPIPE 兜底重入同样被守卫拦截（不持锁路径，review round2 MF1）", async () => {
+    // D2 后续轮本体是编排层私有，不持锁的重入形态只剩 EPIPE 兜底（PiEngine.deliverPrompt
+    // catch 分支直调续轮）——此处经热路径 EPIPE 复现：活 child 的 stdin 已销毁 → 写入
+    // EPIPE → 兜底转冷路径。首条消息已占在途守卫时，第二条的 EPIPE 兜底被拒。
     let release!: (r: AgentResult) => void;
-    const gate = new Promise<AgentResult>((res) => { release = res; });
-    mockRunSpawn.mockImplementationOnce(() => gate);
+    mockRunSpawn.mockImplementationOnce(() => new Promise<AgentResult>((res) => { release = res; }));
+    // 第一条：冷路径（无 child）→ 续轮在途
+    await expect(service.chatActions.deliverChatMessage(record, "first", false)).resolves.toBeUndefined();
+    await vi.waitFor(() => expect(mockRunSpawn).toHaveBeenCalledTimes(1));
 
-    service.resumeRound(record, "first");
-    expect(() => service.resumeRound(record, "second")).toThrow(/already starting a new round/);
+    // 第二条：EPIPE 兜底（child 活但 stdin 断）→ 兜底转冷路径 → 在途守卫 throw 行动语言
+    const epipe = new Error("write EPIPE") as NodeJS.ErrnoException;
+    epipe.code = "EPIPE"; // writeStdinLine 的 R3 判据：err.code === 'EPIPE' 才转 throw
+    const child = { stdin: { destroyed: false, write: () => { throw epipe; } }, exitCode: null, signalCode: null, killed: false } as unknown as ChildProcess;
+    spawnedChildren.set(record.id, child);
+    await expect(service.chatActions.deliverChatMessage(record, "second", false)).rejects.toThrow(/already starting a new round/);
+    expect(mockRunSpawn).toHaveBeenCalledTimes(1);
 
     release(makeResult(true));
   });

package/src/execution/__tests__/descendant-sweep-guards.test.ts

@@ -0,0 +1,220 @@
+// src/execution/__tests__/descendant-sweep-guards.test.ts
+//
+// [S-17c / T2-②] sweepDescendantsOfSession 守卫分支补充（与 descendant-sweep.test.ts
+// 互补，不改动其既有用例）。
+//
+// 覆盖该文件未锁定的分支：
+//   - register entry 缺 sessionId（TODO T5 marker fallback 锚点）→ skipped 留痕不杀；
+//   - readProcessCmdline 两条失败边：ps 非零退出（status≠0/error）与 spawnSync 同步
+//     抛错 → cmdline undefined → "cmdline probe failed" 保守跳过；
+//   - killPidWithEscalation 的 SIGTERM 发送失败（EPERM 等权限/垂死窗口）→ debug
+//     留痕不武装升级（30s 后无 SIGKILL）；
+//   - SIGKILL 升级窗口内目标自行退出（ESRCH）→ 吞掉不崩（目标已达成）。
+//
+// 真实 fs（临时目录）+ mock spawnSync / process.kill / logger——同 descendant-sweep
+// 布局，新增 logger mock 用于 SIGTERM 失败留痕断言。
+
+import { spawnSync } from "node:child_process";
+import * as fs from "node:fs";
+import * as os from "node:os";
+import * as path from "node:path";
+
+import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
+
+const { loggerMock } = vi.hoisted(() => ({
+  loggerMock: { debug: vi.fn(), warn: vi.fn(), error: vi.fn() },
+}));
+vi.mock("../../core/logger.ts", () => ({ getLogger: () => loggerMock }));
+
+// 只 mock spawnSync（ps 探测受控）；spawn 等其余实现保留 actual（本文件不经 runSpawn）。
+vi.mock("node:child_process", async () => {
+  const actual = await import("node:child_process");
+  return {
+    ...actual,
+    spawnSync: vi.fn(),
+  };
+});
+
+import { sweepDescendantsOfSession } from "../engine/engines/pi/session-runner.ts";
+import { clearPendingCursors } from "../session-pending.ts";
+import { writeAliveMarker } from "../alive-store.ts";
+
+const mockSpawnSync = vi.mocked(spawnSync);
+
+const ALIVE_PID = 424_242;
+
+/** process.kill 受控：signal=0 探活与真实信号分别处置，信号调用记录到 kills。 */
+const kills: Array<{ pid: number; signal: string | number }> = [];
+
+function entryLine(customType: string, data: Record<string, unknown>): string {
+  return `${JSON.stringify({
+    type: "custom",
+    customType,
+    data,
+    timestamp: "2026-09-01T00:00:00.000Z",
+    id: `e-${Math.random().toString(36).slice(2)}`,
+  })}\n`;
+}
+
+function registerData(fields: Record<string, unknown>): Record<string, unknown> {
+  return { id: "bg-1", type: "session", name: "desc", status: "active", registeredAt: 1, ...fields };
+}
+
+/** ps -p <pid> -o command= 的完整 SpawnSyncReturns 形态（字段补齐省 as 链）。 */
+function psResult(stdout: string, status = 0): ReturnType<typeof spawnSync> {
+  return {
+    pid: 0,
+    output: [null, stdout, ""],
+    stdout,
+    stderr: "",
+    status,
+    signal: null,
+  } as ReturnType<typeof spawnSync>;
+}
+
+let sessionDir = "";
+
+beforeEach(() => {
+  vi.clearAllMocks();
+  clearPendingCursors();
+  kills.length = 0;
+  sessionDir = fs.mkdtempSync(path.join(os.tmpdir(), "desc-sweep-guard-"));
+  vi.spyOn(process, "kill").mockImplementation(((pid: number, signal?: string | number) => {
+    if (signal === undefined || signal === 0) return true; // 探活恒真（目标在窗口内存活）
+    kills.push({ pid, signal });
+    return true;
+  }) as typeof process.kill);
+});
+
+afterEach(() => {
+  vi.restoreAllMocks();
+  clearPendingCursors();
+  fs.rmSync(sessionDir, { recursive: true, force: true, maxRetries: 5, retryDelay: 20 });
+});
+
+/** 层主 + 一个带 .alive marker 的直接后代（标准单层树）。 */
+function writeSingleLevelTree(): { rootFile: string; childFile: string } {
+  const rootFile = path.join(sessionDir, "20260901T000000-000_root-sess.jsonl");
+  const childFile = path.join(sessionDir, "20260901T000001-000_child-sess.jsonl");
+  fs.writeFileSync(rootFile, entryLine("pending:register", registerData({ sessionId: "child-sess" })));
+  fs.writeFileSync(childFile, "other entry line\n");
+  writeAliveMarker(childFile, { pid: ALIVE_PID, id: "child-sess", startedAt: Date.now() });
+  return { rootFile, childFile };
+}
+
+/** ps 探测默认桩：存活 pid 返回 pi rpc 形态。 */
+function stubCmdlineByPid(): void {
+  mockSpawnSync.mockImplementation((_cmd: string, args?: readonly string[]) => {
+    const pid = Number(args?.[1]);
+    return psResult(`node /usr/local/bin/pi --mode rpc --session-dir /tmp/sess-${pid}\n`);
+  });
+}
+
+describe("[S-17c] sweep 守卫：pending 项缺 sessionId（T5 marker fallback 锚点）", () => {
+  it("register entry 无 sessionId → skipped 留痕（reason 指向 T5 兜底），不杀任何进程", () => {
+    stubCmdlineByPid();
+    const rootFile = path.join(sessionDir, "20260901T000000-000_root2.jsonl");
+    fs.writeFileSync(rootFile, entryLine("pending:register", registerData({ sessionId: undefined })));
+
+    const r = sweepDescendantsOfSession(rootFile, sessionDir, "test-nosessid");
+
+    expect(r.killed).toHaveLength(0);
+    expect(r.skipped).toHaveLength(1);
+    // sessionId 字段回退为 pending 操作 id（诊断线索），reason 说明 marker 兜底锚点
+    expect(r.skipped[0]).toMatchObject({
+      sessionId: "bg-1",
+      reason: expect.stringContaining("no sessionId (marker-based fallback pending T5)"),
+    });
+    expect(kills).toHaveLength(0);
+  });
+});
+
+describe("[S-17c] sweep 守卫：readProcessCmdline 失败边（保守跳过不动手）", () => {
+  it.each([
+    [
+      "ps 非零退出（进程刚死，ps -p 无此进程）",
+      () =>
+        mockSpawnSync.mockImplementation(() => psResult("", 1)),
+    ],
+    [
+      "spawnSync 同步抛错（ps 不可用）",
+      () =>
+        mockSpawnSync.mockImplementation(() => {
+          throw new Error("ps binary missing");
+        }),
+    ],
+  ])("%s → cmdline probe failed，跳过不杀", (_label, setupProbe) => {
+    const { rootFile } = writeSingleLevelTree();
+    setupProbe();
+
+    const r = sweepDescendantsOfSession(rootFile, sessionDir, "test-ps-fail");
+
+    // [防误杀] 探测失败 = 无法证明是 pi rpc 形态 → 一律不动手
+    expect(r.killed).toHaveLength(0);
+    expect(r.skipped).toHaveLength(1);
+    expect(r.skipped[0]).toMatchObject({
+      pid: ALIVE_PID,
+      reason: expect.stringContaining("cmdline probe failed (ps unavailable)"),
+    });
+    expect(kills).toHaveLength(0);
+  });
+});
+
+describe("[S-17c] killPidWithEscalation：SIGTERM/ SIGKILL 失败边", () => {
+  it("SIGTERM 发送即失败（权限/垂死窗口 EPERM）→ debug 留痕不武装升级，30s 后无 SIGKILL", async () => {
+    vi.useFakeTimers({ toFake: ["setTimeout", "clearTimeout"] });
+    stubCmdlineByPid();
+    vi.spyOn(process, "kill").mockImplementation(((pid: number, signal?: string | number) => {
+      if (signal === undefined || signal === 0) return true;
+      if (signal === "SIGTERM") {
+        const err = new Error("operation not permitted") as NodeJS.ErrnoException;
+        err.code = "EPERM";
+        throw err;
+      }
+      kills.push({ pid, signal });
+      return true;
+    }) as typeof process.kill);
+
+    const { rootFile } = writeSingleLevelTree();
+    const r = sweepDescendantsOfSession(rootFile, sessionDir, "test-sigterm-fail");
+
+    // killed 列表记录的是处置意图；实际 SIGTERM 未送达（失败留痕，best-effort continue）
+    expect(r.killed).toEqual([ALIVE_PID]);
+    expect(loggerMock.debug).toHaveBeenCalledWith(
+      expect.stringContaining(`SIGTERM to pid ${ALIVE_PID} failed (best-effort continue)`),
+    );
+
+    // 无升级 timer：SIGTERM 失败后直接 return，越过 30s 窗口也不补发 SIGKILL
+    await vi.advanceTimersByTimeAsync(30_000);
+    expect(kills).toHaveLength(0);
+  });
+
+  it("SIGKILL 升级窗口内目标自行退出（ESRCH）→ 吞掉不崩，SIGTERM 已达即目标达成", async () => {
+    vi.useFakeTimers({ toFake: ["setTimeout", "clearTimeout"] });
+    stubCmdlineByPid();
+    // 探活恒真（fire 时窗口内仍活）；SIGTERM 送达成功；SIGKILL 时进程已退出（ESRCH）
+    vi.spyOn(process, "kill").mockImplementation(((pid: number, signal?: string | number) => {
+      if (signal === undefined || signal === 0) return true;
+      if (signal === "SIGKILL") {
+        const err = new Error("No such process") as NodeJS.ErrnoException;
+        err.code = "ESRCH";
+        throw err;
+      }
+      kills.push({ pid, signal });
+      return true;
+    }) as typeof process.kill);
+
+    const { rootFile } = writeSingleLevelTree();
+    const r = sweepDescendantsOfSession(rootFile, sessionDir, "test-sigkill-esrch");
+
+    expect(r.killed).toEqual([ALIVE_PID]);
+    expect(kills).toEqual([{ pid: ALIVE_PID, signal: "SIGTERM" }]);
+
+    // 升级 fire：ESRCH 是预期终态（进程在窗口内自行退出），吞掉不崩不重试
+    await vi.advanceTimersByTimeAsync(30_000);
+    expect(kills).toEqual([{ pid: ALIVE_PID, signal: "SIGTERM" }]);
+    expect(loggerMock.debug).not.toHaveBeenCalledWith(
+      expect.stringContaining("SIGTERM to pid"),
+    );
+  });
+});

package/src/execution/__tests__/descendant-sweep.test.ts

@@ -34,7 +34,7 @@ vi.mock("node:child_process", async () => {
 import {
   looksLikePiRpcProcess,
   sweepDescendantsOfSession,
-} from "../session-runner.ts";
+} from "../engine/engines/pi/session-runner.ts";
 import {
   clearPendingCursors,
   listActivePendingFromSessionFile,
@@ -89,7 +89,7 @@ beforeEach(() => {
 afterEach(() => {
   vi.restoreAllMocks();
   clearPendingCursors();
-  fs.rmSync(sessionDir, { recursive: true, force: true });
+  fs.rmSync(sessionDir, { recursive: true, force: true, maxRetries: 5, retryDelay: 20 });
 });
 
 /** 标准三层树：root（层主）→ child-sess → grand-sess（叶）。 */
@@ -123,8 +123,8 @@ function psResult(stdout: string): ReturnType<typeof spawnSync> {
 
 /** ps 探测默认桩：全部存活 pid 返回 pi rpc 形态。 */
 function stubCmdlineByPid(): void {
-  mockSpawnSync.mockImplementation((_cmd: string, args: readonly string[]) => {
-    const pid = Number(args[1]);
+  mockSpawnSync.mockImplementation((_cmd: string, args?: readonly string[]) => {
+    const pid = Number(args?.[1]);
     return psResult(`node /usr/local/bin/pi --mode rpc --no-extensions --session-dir /tmp/sess-${pid}\n`);
   });
 }
@@ -189,8 +189,8 @@ describe("[T2-②] sweepDescendantsOfSession（迭代补杀）", () => {
   it("cmdline 非 pi --mode rpc → pid 复用守卫拦截，不杀（但迭代继续展开其后代）", () => {
     const { rootFile, childFile, grandFile } = writeThreeLevelTree();
     // 伪造无关进程占住 child 的 pid；grand 正常
-    mockSpawnSync.mockImplementation((_cmd: string, args: readonly string[]) => {
-      const pid = Number(args[1]);
+    mockSpawnSync.mockImplementation((_cmd: string, args?: readonly string[]) => {
+      const pid = Number(args?.[1]);
       return psResult(
         pid === ALIVE_PID_CHILD ? "vim notes.txt\n" : "pi --mode rpc\n",
       );

package/src/execution/__tests__/dialog-queue.test.ts

@@ -343,7 +343,7 @@ describe("DialogGlobalQueue — dialog 超时上界（LC-3/T2⑦）", () => {
 
     // 完整错误消息承载在父进程日志：含已等待时长 + 「重新发起提问」恢复指引
     expect(loggerMock.warn).toHaveBeenCalledTimes(1);
-    const msg = loggerMock.warn.mock.calls[0][0] as string;
+    const msg = loggerMock.warn.mock.calls[0]![0] as string;
     expect(msg).toContain("1000ms");
     expect(msg).toContain("重新发起提问");
     // 队列状态已释放
@@ -364,7 +364,7 @@ describe("DialogGlobalQueue — dialog 超时上界（LC-3/T2⑦）", () => {
     await vi.advanceTimersByTimeAsync(1);
     await expect(p).resolves.toEqual({ cancelled: true });
     expect(loggerMock.warn).toHaveBeenCalledTimes(1);
-    const msg = loggerMock.warn.mock.calls[0][0] as string;
+    const msg = loggerMock.warn.mock.calls[0]![0] as string;
     expect(msg).toContain(`${DEFAULT_DIALOG_TIMEOUT_MS}ms`);
     expect(msg).toContain("重新发起提问");
   });