/** * Unit Test — stopRun 不能给「正在 executeSubgraph 里 await planWorkGraph」的节点留孤儿子 run * （T-mtgxcnfz 复现） * * 真实事故时序：run_e4ba218e9446 于 2026-08-31T07:30:50.644Z 落 stopped，它的子 run * run_12aa199b07f2 created_at 是 07:30:50.607Z（早 37ms，同一次事件循环附近），但从未被置 * stop_requested，一直跑到人工手动 /stop 才收掉。 * * 根因：`stopRun`（graph-engine.js）在函数开头一次性拍 `descendantRuns(run)` 快照决定要停 * 哪些 run；`executeSubgraph`（同文件）建子 run 之前有一次真实 `await planWorkGraph(ctx)`—— * 如果这次 await 跨越了 stopRun 执行的时间点，子 run 在 stopRun 拍完快照**之后**才落库， * 永远不在 targets 里，`stopRun` 本身又是全同步函数、没有内部 await，事后补扫赶不上。 * * 修法钉在 executeSubgraph：`await planWorkGraph` 之后、真正 `createRun` 之前重新读一次 * 父 run，`stop_requested` 已经变真就放弃创建子 run，就地按节点失败结算——不产生孤儿。 * * 用真实的 org run + subgraph 节点复现（不是打桩内部实现细节）：`setPlanWorkGraph` 注入一个 * 可以从测试里手动控制何时 resolve 的假规划，制造出与生产一致的 await 窗口，在窗口内调用 * `engine.stopRun`，再放行规划，断言子 run 没有被建出来、父节点按 fail 收敛、run 落 stopped。 */ import { describe, it, expect, beforeAll, afterAll, beforeEach } from 'vitest'; import fs from 'fs'; import os from 'os'; import path from 'path'; describe('stopRun 与 executeSubgraph 的建子 run 竞态（T-mtgxcnfz）', () => { let orgSvc; let wdb; let engine; let runsDir; const TEST_DB_PATH = '/tmp/test-graph-engine-stop-run-orphan-race.sqlite'; beforeAll(() => { for (const f of [TEST_DB_PATH, `${TEST_DB_PATH}-shm`, `${TEST_DB_PATH}-wal`]) { try { fs.unlinkSync(f); } catch {} } process.env.GRAPH_DB_PATH = TEST_DB_PATH; runsDir = fs.mkdtempSync(path.join(os.tmpdir(), 'graph-engine-stop-run-orphan-race-')); process.env.GRAPH_RUNS_DIR = runsDir; for (const m of [ '../../services/graph-db', '../../services/org-graph-service', '../../services/work-graph-db', '../../services/graph-run-sse', '../../services/graph-engine', ]) { delete require.cache[require.resolve(m)]; } orgSvc = require('../../services/org-graph-service'); wdb = require('../../services/work-graph-db'); engine = require('../../services/graph-engine'); }); afterAll(() => { for (const f of [TEST_DB_PATH, `${TEST_DB_PATH}-shm`, `${TEST_DB_PATH}-wal`]) { try { fs.unlinkSync(f); } catch {} } delete process.env.GRAPH_DB_PATH; try { fs.rmSync(runsDir, { recursive: true, force: true }); } catch {} delete process.env.GRAPH_RUNS_DIR; }); beforeEach(() => { const db = wdb.getDb(); db.pragma('foreign_keys = OFF'); db.exec(` DELETE FROM graph_run_loop_records; DELETE FROM graph_run_nodes; DELETE FROM graph_runs; DELETE FROM work_graph_edges; DELETE FROM work_graph_nodes; DELETE FROM work_graphs; DELETE FROM org_graph_edges; DELETE FROM org_graph_nodes; DELETE FROM system_agents; `); db.pragma('foreign_keys = ON'); engine.setAgentExecutor(null); engine.setPlanWorkGraph(null); }); let seq = 0; const nextCompany = () => `co-stop-race-${++seq}`; it('planWorkGraph 还没 resolve 时被 stopRun 停掉：不建子 run，节点按 fail 收敛，run 落 stopped', async () => { const companyId = nextCompany(); expect(orgSvc.writeGraph(companyId, { nodes: ['org-coder'], edges: [] }).ok).toBe(true); let releasePlanning; const planningStarted = new Promise((resolveStarted) => { engine.setPlanWorkGraph((ctx) => { resolveStarted(); return new Promise((resolve) => { releasePlanning = () => { // 造一张真实可用的子图定义——workGraphId 不能是 null，否则会在 // executeSubgraph 里被「未返回 workGraphId」那道**无关**的校验拦下， // 那不是这条测试要盯的路径，会把「没建成子 run」的原因张冠李戴。 const worker = `${ctx.nodeName}-worker`; const db = wdb.getDb(); const now = new Date().toISOString(); db.prepare(` INSERT OR IGNORE INTO system_agents (id, name, scope, owner_company_id, legacy_path, status, created_at, updated_at) VALUES (?, ?, 'public', NULL, ?, 'active', ?, ?) `).run(worker, worker, `system/agents/${worker}`, now, now); const g = wdb.createWorkGraph({ companyId, nodes: [worker], edges: [] }); resolve({ workGraphId: g.id }); }; }); }); }); const run = wdb.createRun({ graphKind: 'org', graphId: companyId, companyId, executionMode: 'manual' }); engine.markRunReady(run.id); engine.startRun(run.id, { mode: 'manual' }); // 不 await：advanceOne 内部会在 executeSubgraph 里卡在 planWorkGraph 那次 await 上， // 这条 promise 要等我们手动 releasePlanning 之后才会继续往下走。 const stepPromise = engine.nextStep(run.id); await planningStarted; // 确认已经真的进了 await 窗口，不是还没排到 // 窗口期间人工调 /stop——此时子 run 还没建出来，descendantRuns(run) 快照里只有 run 自己。 const stopResult = engine.stopRun(run.id); expect(stopResult.ok).toBe(true); expect(wdb.getRun(run.id).run_status).toBe('stopped'); // 这一刻整棵树只有 root 自己——旧代码的 bug 正是「这个断言此刻成立，但之后会被打破」 expect(wdb.listRunTree(run.id)).toHaveLength(1); // 放行规划，让 executeSubgraph 走到「要不要建子 run」那一步 releasePlanning(); const stepResult = await stepPromise; expect(stepResult.ok).toBe(true); // 核心断言：没有孤儿子 run——树上除了 root 一个都没多出来 const tree = wdb.listRunTree(run.id); expect(tree).toHaveLength(1); // 节点没有 pending 的 child_run_id，且不再停在 running 干等一个不存在的子 run—— // 没有匹配的 fail 边时 routeFailure 落 blocked（需要人工介入），这里不咬死具体是 // fail 还是 blocked，只钉住「不是孤儿」这条：没有子 run、没有继续假装在跑。 const node = wdb.getRunState(run.id).nodes.find((n) => n.name === 'org-coder'); expect(node.childRunId).toBeNull(); expect(node.status).not.toBe('running'); expect(['fail', 'blocked']).toContain(node.status); }); it('对照组：没有并发 stop 时，同样的 await 窗口结束后子 run 正常建出来', async () => { const companyId = nextCompany(); expect(orgSvc.writeGraph(companyId, { nodes: ['org-coder'], edges: [] }).ok).toBe(true); engine.setPlanWorkGraph(async (ctx) => { const worker = `${ctx.nodeName}-worker`; const db = wdb.getDb(); const now = new Date().toISOString(); db.prepare(` INSERT OR IGNORE INTO system_agents (id, name, scope, owner_company_id, legacy_path, status, created_at, updated_at) VALUES (?, ?, 'public', NULL, ?, 'active', ?, ?) `).run(worker, worker, `system/agents/${worker}`, now, now); const g = wdb.createWorkGraph({ companyId, nodes: [worker], edges: [] }); expect(g.errors || []).toEqual([]); return { workGraphId: g.id }; }); const run = wdb.createRun({ graphKind: 'org', graphId: companyId, companyId, executionMode: 'manual' }); engine.markRunReady(run.id); engine.startRun(run.id, { mode: 'manual' }); const stepResult = await engine.nextStep(run.id); expect(stepResult.ok).toBe(true); expect(wdb.listRunTree(run.id)).toHaveLength(2); const node = wdb.getRunState(run.id).nodes.find((n) => n.name === 'org-coder'); expect(node.childRunId).toBeTruthy(); expect(node.status).toBe('running'); // 等子 run 收敛，父节点停在 running }); });