@zhushanwen/pi-subagent-workflow 8.5.0 → 8.7.0

package/src/orchestration/__tests__/execute-agent-call.test.ts

@@ -1,314 +0,0 @@
-// src/orchestration/__tests__/execute-agent-call.test.ts
-//
-// U2: executeAgentCall 透传 stream 给 runner.run
-// U3: executeAgentCall retry 递归也透传 stream（不丢、不重建）
-
-import { describe, expect, it, vi } from "vitest";
-
-import {
-  executeAgentCall,
-  isStaleContextErrorMsg,
-  STALE_CONTEXT_PATTERNS,
-} from "../execute-agent-call.ts";
-import { AgentCall } from "../models/agent-call.ts";
-import { Budget } from "../models/budget.ts";
-import type { AgentRunner } from "../models/ports.ts";
-import { Trace } from "../models/trace.ts";
-import type { ExecutionTraceNode } from "../models/types.ts";
-import type { AgentCallOpts, AgentResult } from "../models/types.ts";
-
-// ── 测试辅助 ──
-
-function makeMockResult(overrides: Partial<AgentResult> = {}): AgentResult {
-  return {
-    content: "OK",
-    durationMs: 100,
-    error: undefined,
-    toolCalls: [],
-    ...overrides,
-  };
-}
-
-function makeBaseOpts(): AgentCallOpts {
-  return {
-    prompt: "test task",
-    agent: "worker",
-    cwd: "/some/path",
-  } as AgentCallOpts;
-}
-
-/** 构造一个 traceNode（ExecutionTraceNode 最小子集） */
-function makeTraceNode(stepIndex = 0): ExecutionTraceNode {
-  return {
-    stepIndex,
-    agent: "test-agent",
-    task: "test task",
-    model: "default",
-    status: "pending",
-  };
-}
-
-/** 构造 AgentCall + 关联的 Trace（call.id 与 trace 节点 stepIndex 对齐） */
-function makeAgentCallAndTrace(): { call: AgentCall; trace: Trace } {
-  const trace = new Trace();
-  const traceNode = makeTraceNode(0);
-  trace.append(traceNode);
-  const call = new AgentCall(0, makeBaseOpts(), traceNode);
-  return { call, trace };
-}
-
-/** 创建 mock AgentRunner（只实现 run） */
-function createMockRunner(impl?: ReturnType<typeof vi.fn>): AgentRunner & { run: ReturnType<typeof vi.fn> } {
-  const run = impl ?? vi.fn().mockResolvedValue(makeMockResult());
-  return { run } as unknown as AgentRunner & { run: ReturnType<typeof vi.fn> };
-}
-
-// ── U2: executeAgentCall 透传 stream 给 runner.run ──
-
-describe("U2: executeAgentCall 透传 stream", () => {
-  it("executeAgentCall 传 stream → runner.run 第 4 参收到同一 stream", async () => {
-    let capturedStream: unknown;
-    const runner = createMockRunner(
-      vi.fn().mockImplementation((_opts, _sig, _onEvt, stream) => {
-        capturedStream = stream;
-        return Promise.resolve(makeMockResult());
-      }),
-    );
-
-    const fakeStream = { onDelta: vi.fn(), dispose: vi.fn() };
-    const { call, trace } = makeAgentCallAndTrace();
-    const budget = new Budget();
-
-    await executeAgentCall(call, runner, budget, new AbortController().signal, trace, undefined, fakeStream as never);
-
-    expect(capturedStream).toBe(fakeStream);
-  });
-
-  it("executeAgentCall 不传 stream → runner.run 第 4 参为 undefined", async () => {
-    let capturedStream: unknown = "sentinel";
-    const runner = createMockRunner(
-      vi.fn().mockImplementation((_opts, _sig, _onEvt, stream) => {
-        capturedStream = stream;
-        return Promise.resolve(makeMockResult());
-      }),
-    );
-
-    const { call, trace } = makeAgentCallAndTrace();
-    const budget = new Budget();
-
-    await executeAgentCall(call, runner, budget, new AbortController().signal, trace);
-
-    expect(capturedStream).toBeUndefined();
-  });
-});
-
-// ── U1: finalizeCall 透传 sessionFile 到 trace 节点（方案 A）──
-
-describe("U1: finalizeCall sessionFile → trace 节点", () => {
-  it("runner.run 返回带 sessionFile 的 result → trace 节点携带 sessionFile", async () => {
-    const sessionFilePath = "/abs/.pi/agent/subagents/enc/sessions/2026-07-15T_session-abc.jsonl";
-    const runner = createMockRunner(
-      vi.fn().mockResolvedValue(
-        makeMockResult({ sessionId: "session-abc", sessionFile: sessionFilePath }),
-      ),
-    );
-    const { call, trace } = makeAgentCallAndTrace();
-    const budget = new Budget();
-
-    await executeAgentCall(call, runner, budget, new AbortController().signal, trace);
-
-    const node = trace.find(0);
-    expect(node).toBeDefined();
-    expect(node!.sessionFile).toBe(sessionFilePath);
-    expect(node!.sessionId).toBe("session-abc");
-  });
-
-  it("runner.run 返回无 sessionFile 的 result → trace 节点 sessionFile undefined", async () => {
-    const runner = createMockRunner(
-      vi.fn().mockResolvedValue(makeMockResult({ sessionId: "session-xyz" })),
-    );
-    const { call, trace } = makeAgentCallAndTrace();
-    const budget = new Budget();
-
-    await executeAgentCall(call, runner, budget, new AbortController().signal, trace);
-
-    const node = trace.find(0);
-    expect(node).toBeDefined();
-    expect(node!.sessionFile).toBeUndefined();
-  });
-});
-
-// ── U3: retry 递归也透传 stream ──
-
-describe("U3: executeAgentCall retry 透传 stream", () => {
-  it("首次 runner.run 返回 error → retry 时第 4 参仍是同一 stream", async () => {
-    vi.useFakeTimers();
-
-    const capturedStreams: unknown[] = [];
-    const runner = createMockRunner(
-      vi.fn().mockImplementation((_opts, _sig, _onEvt, stream) => {
-        capturedStreams.push(stream);
-        // 首次 error，第二次成功
-        if (capturedStreams.length === 1) {
-          return Promise.resolve(makeMockResult({ error: "transient error" }));
-        }
-        return Promise.resolve(makeMockResult());
-      }),
-    );
-
-    const fakeStream = { onDelta: vi.fn(), dispose: vi.fn() };
-    const { call, trace } = makeAgentCallAndTrace();
-    const budget = new Budget();
-
-    const promise = executeAgentCall(call, runner, budget, new AbortController().signal, trace, undefined, fakeStream as never);
-    // 推进 retry 退避定时器（BACKOFF_BASE_MS = 1000，首退避 1s）
-    await vi.advanceTimersByTimeAsync(2000);
-    await promise;
-
-    expect(runner.run).toHaveBeenCalledTimes(2);
-    // 两次调用的第 4 参都是同一 stream 对象
-    expect(capturedStreams[0]).toBe(fakeStream);
-    expect(capturedStreams[1]).toBe(fakeStream);
-
-    vi.useRealTimers();
-  });
-});
-
-// ── isOrphaned 守卫（OB2：S7 残留瞬时污染根除） ──
-//
-// 谓词为 true 时 finalizeCall 跳过 trace.update，但 markDone 与
-// sessionId/sessionFile 同步保留。U4 锁定不传谓词时的现状回归。
-
-describe("isOrphaned 守卫", () => {
-  it("U1: 谓词 () => true + 终态成功路径 → trace.update 0 次，markDone 保留（status done + result 已设置）", async () => {
-    const { call, trace } = makeAgentCallAndTrace();
-    const updateSpy = vi.spyOn(trace, "update");
-    const runner = createMockRunner(); // 默认成功 result
-    const budget = new Budget();
-
-    await executeAgentCall(call, runner, budget, new AbortController().signal, trace, undefined, undefined, () => true);
-
-    expect(updateSpy).not.toHaveBeenCalled();
-    expect(call.status).toBe("done");
-    expect(call.result).toBeDefined();
-    expect(call.result?.content).toBe("OK");
-  });
-
-  it("U2: 谓词 true + stale-context 失败路径 → trace.update 0 次，markDone 保留（覆盖 stale finalize 调用点）", async () => {
-    const { call, trace } = makeAgentCallAndTrace();
-    const updateSpy = vi.spyOn(trace, "update");
-    const runner = createMockRunner(
-      vi.fn().mockResolvedValue(makeMockResult({ error: "context canceled" })),
-    );
-    const budget = new Budget();
-
-    await executeAgentCall(call, runner, budget, new AbortController().signal, trace, undefined, undefined, () => true);
-
-    expect(updateSpy).not.toHaveBeenCalled();
-    expect(call.status).toBe("done");
-    expect(call.result?.error).toBe("context canceled");
-  });
-
-  it("U3: 谓词 true + 信号 abort 路径 → trace.update 0 次，markDone 保留（覆盖 abort finalize 调用点）", async () => {
-    const { call, trace } = makeAgentCallAndTrace();
-    const updateSpy = vi.spyOn(trace, "update");
-    // 错误文案不含 stale 模式词，确保走 signal.aborted 分支而非 stale 分支
-    const runner = createMockRunner(
-      vi.fn().mockResolvedValue(makeMockResult({ error: "old generation failure" })),
-    );
-    const controller = new AbortController();
-    controller.abort();
-    const budget = new Budget();
-
-    await executeAgentCall(call, runner, budget, controller.signal, trace, undefined, undefined, () => true);
-
-    expect(updateSpy).not.toHaveBeenCalled();
-    expect(call.status).toBe("done");
-    expect(call.result?.error).toBe("old generation failure");
-  });
-
-  it("U4: 谓词 undefined（不传）→ trace.update 恰被调用 1 次（现状回归锁定）", async () => {
-    const { call, trace } = makeAgentCallAndTrace();
-    const updateSpy = vi.spyOn(trace, "update");
-    const runner = createMockRunner();
-    const budget = new Budget();
-
-    await executeAgentCall(call, runner, budget, new AbortController().signal, trace);
-
-    expect(updateSpy).toHaveBeenCalledTimes(1);
-    expect(trace.find(0)?.status).toBe("completed");
-    expect(call.status).toBe("done");
-  });
-
-  it("U5: 递归重试透传——可重试失败后成功，谓词 true → 重试后终态 trace.update 仍 0 次", async () => {
-    vi.useFakeTimers();
-    try {
-      const { call, trace } = makeAgentCallAndTrace();
-      const updateSpy = vi.spyOn(trace, "update");
-      const runner = createMockRunner(
-        vi.fn()
-          .mockResolvedValueOnce(makeMockResult({ error: "transient error" }))
-          .mockResolvedValueOnce(makeMockResult()),
-      );
-      const budget = new Budget();
-
-      const promise = executeAgentCall(call, runner, budget, new AbortController().signal, trace, undefined, undefined, () => true);
-      // 推进首退避 1s（BACKOFF_BASE_MS = 1000）
-      await vi.advanceTimersByTimeAsync(2000);
-      await promise;
-
-      expect(runner.run).toHaveBeenCalledTimes(2);
-      expect(call.attempts).toBe(2);
-      // 谓词穿透递归层：重试后终态也不写 trace
-      expect(updateSpy).not.toHaveBeenCalled();
-      expect(call.status).toBe("done");
-      expect(call.result?.error).toBeUndefined();
-    } finally {
-      vi.useRealTimers();
-    }
-  });
-});
-
-// ── W4b: STALE_CONTEXT_PATTERNS 对齐 pi 真实文案（stale 分诊） ──
-//
-// pi 0.84.x 真实 stale 文案（extensions/runner.ts:531，dist runner.js:567）：
-// "This extension ctx is stale after session replacement or reload. Do not use
-//  a captured pi or command ctx after ctx.newSession(), ..."
-// 旧 patterns（"stale context"/"stalecontext"）与该文案零匹配（词序相反），
-// stale 错误曾退化为普通错误照常重试 3 次——W4b 词序修正 + 对齐 scheduler
-// 已验证 marker 'stale after session replacement'（runtime.ts STALE_CTX_MARKER）。
-
-/** pi 真实 stale 文案前半（含 marker 全文前缀，锚定真实串而非自造缩写）。 */
-const PI_REAL_STALE_MESSAGE =
-  "This extension ctx is stale after session replacement or reload. Do not use a captured pi or command ctx after ctx.newSession(), ctx.fork(), ctx.switchSession(), or ctx.reload().";
-
-describe("W4b: stale 分诊对齐 pi 真实文案", () => {
-  it("真实文案全文 → isStaleContextErrorMsg true（'ctx is stale' 词序 + scheduler marker 双命中）", () => {
-    expect(isStaleContextErrorMsg(PI_REAL_STALE_MESSAGE)).toBe(true);
-  });
-
-  it("patterns 含 scheduler 已验证 marker 'stale after session replacement'", () => {
-    expect(STALE_CONTEXT_PATTERNS).toContain("stale after session replacement");
-    expect(STALE_CONTEXT_PATTERNS).toContain("ctx is stale");
-  });
-
-  it("真实文案 → executeAgentCall 不重试（runner.run 恰 1 次）+ markDone failed", async () => {
-    const { call, trace } = makeAgentCallAndTrace();
-    const runner = createMockRunner(
-      vi.fn().mockResolvedValue(makeMockResult({ error: PI_REAL_STALE_MESSAGE })),
-    );
-    const budget = new Budget();
-
-    await executeAgentCall(call, runner, budget, new AbortController().signal, trace);
-
-    expect(runner.run).toHaveBeenCalledTimes(1);
-    expect(call.status).toBe("done");
-    expect(call.result?.error).toBe(PI_REAL_STALE_MESSAGE);
-    expect(trace.find(0)?.status).toBe("failed");
-  });
-
-  it("普通 transient 错误 → 不命中 stale（照常进入重试路径，现状回归）", () => {
-    expect(isStaleContextErrorMsg("transient network error")).toBe(false);
-    expect(isStaleContextErrorMsg(undefined)).toBe(false);
-  });
-});

package/src/orchestration/__tests__/jsonl-run-store-loadall-sources.test.ts

@@ -1,171 +0,0 @@
-// src/orchestration/__tests__/jsonl-run-store-loadall-sources.test.ts
-//
-// loadAll entry 源扫描（collectEntrySources + loadRunFromStateFile）定向用例（R2-TC S4）。
-//
-// 防的 bug：loadAll 的多 run 重建 / 损坏 state 文件降级。W17 已覆盖单 run 的 entry 重建、
-// link 兼容与读序优先级；本文件补齐三块未定向覆盖的分支：
-// 1. 多 run 并存（多个 workflow-record entry run + link run 同批重建，无丢失）
-// 2. 损坏 state 文件降级（末行非 JSON / 空文件 / link 指向不存在路径 / link data 缺 runId）
-//    ——单文件失败返回 null 跳过，不崩 loadAll、不阻断同批其余 run 重建
-// 3. 同 runId 多条 record entry 末条胜出在多 run 场景下不串扰
-//
-// 不 mock collectEntrySources：走 loadAll 真实通路（ctx.sessionManager.getEntries 返回
-// seed entries，loadRunFromStateFile 读真实临时文件）。record entry 一律经真实 save 产出
-//（deserializeRun 的快照 schema 以真实 serializeRun 为准，手工拼快照会绕开 schema 契约）。
-
-import * as fs from "node:fs";
-import * as os from "node:os";
-import * as path from "node:path";
-
-import { afterEach, beforeEach, describe, expect, it } from "vitest";
-
-import type { CustomEntry } from "@earendil-works/pi-coding-agent";
-
-import { Budget } from "../models/budget.ts";
-import { Trace } from "../models/trace.ts";
-import type { ExecutionTraceNode } from "../models/types.ts";
-import type { RunSpec } from "../models/run-spec.ts";
-import { WorkflowRun } from "../models/workflow-run.ts";
-import { JsonlRunStore } from "../jsonl-run-store.ts";
-import { mkCtx, mkPi } from "./test-mocks.ts";
-
-function makeSpec(): RunSpec {
-  return {
-    scriptSource: "module.exports = async () => {};",
-    args: {},
-    scriptName: "test-script",
-    scriptPath: "/tmp/test.js",
-    description: "test",
-  };
-}
-
-function makeTraceNode(stepIndex: number): ExecutionTraceNode {
-  return { stepIndex, agent: "worker", task: "do thing", model: "default", status: "pending" };
-}
-
-/** running → done 的完整生命周期 save（终态快照可被 deserializeRun 重建）。 */
-async function saveDoneRun(
-  sessionDir: string,
-  entries: CustomEntry[],
-  runId: string,
-): Promise<void> {
-  const store = new JsonlRunStore({ sessionDir, pi: mkPi(entries) });
-  const run = makeRunningRun(runId);
-  await store.save(run);
-  run.transition("done", "completed");
-  await store.save(run);
-}
-
-function makeRunningRun(runId: string): WorkflowRun {
-  const trace = new Trace();
-  trace.append(makeTraceNode(0));
-  return WorkflowRun.reconstruct(runId, makeSpec(), {
-    status: "running",
-    budget: new Budget(),
-    calls: new Map(),
-    trace,
-    errorLogs: [],
-  }, { startedAt: new Date().toISOString() });
-}
-
-/** 手工构造旧 workflow-state-link 指针 entry（loadAll 的 link 兼容输入）。 */
-function linkEntry(runId: string, statePath: string): CustomEntry {
-  return {
-    type: "custom",
-    customType: "workflow-state-link",
-    data: { runId, path: statePath },
-    id: `seed-link-${runId}`,
-    parentId: null,
-    timestamp: new Date().toISOString(),
-  };
-}
-
-describe("loadAll entry 源扫描：多 run 重建与损坏 state 降级（R2-TC S4）", () => {
-  let tmpDir: string;
-
-  beforeEach(() => {
-    tmpDir = fs.mkdtempSync(path.join(os.tmpdir(), "wf-loadall-src-"));
-  });
-
-  afterEach(() => {
-    fs.rmSync(tmpDir, { recursive: true, force: true });
-  });
-
-  it("多 run 并存：2 个 entry run + 1 个 link run 同批全部重建（无丢失）", async () => {
-    // run-a/run-b 经真实 save 通路产出终态 entry（快照形态可信）
-    const entries: CustomEntry[] = [];
-    await saveDoneRun(tmpDir, entries, "run-a");
-    await saveDoneRun(tmpDir, entries, "run-b");
-
-    // run-c：旧 link 形态（state 文件完好，无 record entry）
-    const storeC = new JsonlRunStore({ sessionDir: tmpDir });
-    const runC = makeRunningRun("run-c");
-    await storeC.save(runC);
-    runC.transition("done", "completed");
-    await storeC.save(runC);
-    const linkPath = path.join(tmpDir, "workflow-state", "run-c.jsonl");
-    expect(fs.existsSync(linkPath)).toBe(true);
-    const seedEntries: CustomEntry[] = [...entries, linkEntry("run-c", linkPath)];
-
-    const store = new JsonlRunStore({ sessionDir: tmpDir, ctx: mkCtx(seedEntries) });
-    const loaded = await store.loadAll();
-    expect(loaded.map((r) => r.runId).sort()).toEqual(["run-a", "run-b", "run-c"]);
-  });
-
-  it("损坏 state 文件（末行非 JSON）→ 该 run 跳过不崩，同批其余 run（entry run）正常返回", async () => {
-    const entries: CustomEntry[] = [];
-    await saveDoneRun(tmpDir, entries, "run-good");
-
-    const corruptPath = path.join(tmpDir, "workflow-state", "run-corrupt.jsonl");
-    fs.mkdirSync(path.dirname(corruptPath), { recursive: true });
-    fs.writeFileSync(corruptPath, "{truncated-not-json\n", "utf8");
-
-    const seedEntries: CustomEntry[] = [...entries, linkEntry("run-corrupt", corruptPath)];
-    const store = new JsonlRunStore({ sessionDir: tmpDir, ctx: mkCtx(seedEntries) });
-    const loaded = await store.loadAll();
-    expect(loaded.map((r) => r.runId)).toEqual(["run-good"]);
-  });
-
-  it("损坏形态族：空文件 / link 指向不存在路径 / link data 缺 runId → 均降级跳过，loadAll 不抛", async () => {
-    const emptyPath = path.join(tmpDir, "workflow-state", "run-empty.jsonl");
-    fs.mkdirSync(path.dirname(emptyPath), { recursive: true });
-    fs.writeFileSync(emptyPath, "", "utf8");
-    const missingPath = path.join(tmpDir, "workflow-state", "run-missing.jsonl");
-
-    const malformedLink = linkEntry("run-bad", emptyPath);
-    (malformedLink.data as Record<string, unknown>).runId = undefined; // 缺 runId 的坏指针
-
-    // 对照组：一条合法 record entry run 应正常重建（真实 save 产出）
-    const entries: CustomEntry[] = [];
-    await saveDoneRun(tmpDir, entries, "run-seeded");
-
-    const seedEntries: CustomEntry[] = [
-      ...entries,
-      linkEntry("run-empty", emptyPath),
-      linkEntry("run-missing", missingPath),
-      malformedLink,
-    ];
-    const store = new JsonlRunStore({ sessionDir: tmpDir, ctx: mkCtx(seedEntries) });
-    const loaded = await store.loadAll();
-    expect(loaded.map((r) => r.runId)).toEqual(["run-seeded"]);
-  });
-
-  it("同 runId 多条 record entry 末条胜出（running→done 收敛）+ 另一 run 不受影响（混合批不串扰）", async () => {
-    const entries: CustomEntry[] = [];
-    // run-x：两条 entry（running 中间态 + done 终态），末条胜出
-    const storeX = new JsonlRunStore({ sessionDir: tmpDir, pi: mkPi(entries) });
-    const runX = makeRunningRun("run-x");
-    await storeX.save(runX);
-    runX.transition("done", "completed");
-    await storeX.save(runX);
-    // run-y：done 终态
-    await saveDoneRun(tmpDir, entries, "run-y");
-
-    const store = new JsonlRunStore({ sessionDir: tmpDir, ctx: mkCtx(entries) });
-    const loaded = await store.loadAll();
-    const byId = new Map(loaded.map((r) => [r.runId, r]));
-    expect(loaded).toHaveLength(2);
-    expect(byId.get("run-x")?.state.status).toBe("done");
-    expect(byId.get("run-y")?.state.status).toBe("done");
-  });
-});