pi-crew 0.9.12 → 0.9.14
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +86 -0
- package/README.md +1 -135
- package/docs/bugs/bug-021-notification-badge-counter-misleading.md +293 -0
- package/docs/bugs/bug-023-chain-windows-path-resolution.md +106 -0
- package/package.json +1 -1
- package/src/config/defaults.ts +3 -1
- package/src/extension/knowledge-injection.ts +236 -15
- package/src/extension/pi-api.ts +1 -1
- package/src/extension/team-tool/chain-dispatch.ts +103 -0
- package/src/extension/team-tool/chain-executor.ts +400 -0
- package/src/extension/team-tool/run.ts +9 -0
- package/src/runtime/agent-memory.ts +5 -2
- package/src/runtime/background-runner.ts +49 -0
- package/src/runtime/chain-runner.ts +47 -8
- package/src/runtime/child-pi.ts +119 -0
- package/src/runtime/goal-achievement.ts +131 -0
- package/src/runtime/handoff-manager.ts +7 -0
- package/src/runtime/live-session-runtime.ts +10 -4
- package/src/runtime/pi-json-output.ts +5 -1
- package/src/runtime/recovery-recipes.ts +35 -0
- package/src/runtime/task-output-context.ts +36 -7
- package/src/runtime/task-runner/prompt-builder.ts +5 -1
- package/src/runtime/task-runner.ts +13 -0
- package/src/runtime/team-runner.ts +47 -2
- package/src/schema/team-tool-schema.ts +9 -0
- package/src/state/types.ts +3 -0
- package/src/ui/widget/index.ts +1 -1
- package/src/ui/widget/widget-formatters.ts +14 -1
- package/src/ui/widget/widget-renderer.ts +13 -2
- package/src/utils/redaction.ts +20 -0
|
@@ -0,0 +1,400 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Concrete `ChainTaskRunner` that executes each chain step as a REAL team run.
|
|
3
|
+
*
|
|
4
|
+
* This is the production wiring that makes `ChainRunner` (and Fix C's
|
|
5
|
+
* `__chainHistoryNotes`) reachable: previously `ChainRunner` had zero production
|
|
6
|
+
* callers and only existed with unit-test mock runners. Each step delegates to
|
|
7
|
+
* the existing `handleRun` (which owns config load, team/workflow discovery,
|
|
8
|
+
* worktree preconditions, manifest creation, and `executeTeamRun`), so the
|
|
9
|
+
* ~200 lines of run machinery are NOT duplicated.
|
|
10
|
+
*
|
|
11
|
+
* Context passage (the whole point of the chain feature): `enrichContextFromHandoffs`
|
|
12
|
+
* places `__chainHistory` + `__chainHistoryNotes` into `packet.context`; this
|
|
13
|
+
* executor serializes them into a `# Previous Steps in This Chain` block that is
|
|
14
|
+
* prepended to the step's goal. `handleRun` → `executeTeamRun` → worker prompt,
|
|
15
|
+
* so step N+1's worker sees step N's handoff summary and the Fix C markers.
|
|
16
|
+
*
|
|
17
|
+
* Circular-import avoidance: `handleRun` is received as an INJECTED function
|
|
18
|
+
* reference (dependency injection) rather than a static import. `run.ts` lazy-
|
|
19
|
+
* imports `chain-dispatch.ts`, which imports this file. A static import of
|
|
20
|
+
* `run.ts` here would create a `run.ts ↔ chain-executor.ts` cycle that races
|
|
21
|
+
* module-record instantiation under jiti. The DI pattern breaks the cycle.
|
|
22
|
+
*
|
|
23
|
+
* @see src/runtime/chain-runner.ts (runChain, enrichContextFromHandoffs — Fix C)
|
|
24
|
+
* @see src/extension/team-tool/run.ts (handleRun — the reused machinery)
|
|
25
|
+
*/
|
|
26
|
+
|
|
27
|
+
import type { ChainTaskRunner } from "../../runtime/chain-runner.ts";
|
|
28
|
+
import type { TaskPacket, TaskResult, Decision } from "../../runtime/handoff-manager.ts";
|
|
29
|
+
import { loadRunManifestById, createRunPaths } from "../../state/state-store.ts";
|
|
30
|
+
import { readIfSmallWithTee } from "../../runtime/task-output-context.ts";
|
|
31
|
+
import type { TeamRunManifest, TeamTaskState } from "../../state/types.ts";
|
|
32
|
+
import type { PiTeamsToolResult } from "../tool-result.ts";
|
|
33
|
+
import { textFromToolResult } from "../tool-result.ts";
|
|
34
|
+
import type { TeamToolParamsValue } from "../../schema/team-tool-schema.ts";
|
|
35
|
+
import type { TeamContext } from "./context.ts";
|
|
36
|
+
import * as fs from "node:fs";
|
|
37
|
+
import * as path from "node:path";
|
|
38
|
+
|
|
39
|
+
/**
|
|
40
|
+
* Signature of `handleRun`, declared locally (type-only) to break the
|
|
41
|
+
* import cycle with run.ts. Structurally identical to
|
|
42
|
+
* `(params, ctx) => Promise<PiTeamsToolResult>`.
|
|
43
|
+
*/
|
|
44
|
+
export type HandleRunFn = (
|
|
45
|
+
params: TeamToolParamsValue,
|
|
46
|
+
ctx: TeamContext,
|
|
47
|
+
) => Promise<PiTeamsToolResult>;
|
|
48
|
+
|
|
49
|
+
/** Per-step team/workflow/model overrides forwarded from the chain invocation. */
|
|
50
|
+
export interface ChainExecutorOverrides {
|
|
51
|
+
team?: string;
|
|
52
|
+
workflow?: string;
|
|
53
|
+
model?: string;
|
|
54
|
+
}
|
|
55
|
+
|
|
56
|
+
/** Shape of one entry in `context.__chainHistory` produced by enrichContextFromHandoffs. */
|
|
57
|
+
interface ChainHistoryEntry {
|
|
58
|
+
step?: string;
|
|
59
|
+
outcome?: string;
|
|
60
|
+
filesCreated?: string[];
|
|
61
|
+
filesModified?: string[];
|
|
62
|
+
decisions?: Decision[];
|
|
63
|
+
nextSteps?: string[];
|
|
64
|
+
outputText?: string;
|
|
65
|
+
}
|
|
66
|
+
|
|
67
|
+
/**
|
|
68
|
+
* Serialize `packet.context.__chainHistory` (+ `__chainHistoryNotes`) into a
|
|
69
|
+
* human-readable block prefixed with `# Previous Steps in This Chain`.
|
|
70
|
+
*
|
|
71
|
+
* This is the CRITICAL coupling that makes Fix C's `__chainHistoryNotes`
|
|
72
|
+
* markers visible to step N+1's worker: those markers travel as a sibling
|
|
73
|
+
* array on the context object, and only become worker-visible once formatted
|
|
74
|
+
* into the goal text here. Returns `""` when there is no history.
|
|
75
|
+
*
|
|
76
|
+
* Extracted as a pure function for unit testability (acceptance criterion 6c).
|
|
77
|
+
*/
|
|
78
|
+
export function formatChainHistory(context: Record<string, unknown>): string {
|
|
79
|
+
const history = context.__chainHistory;
|
|
80
|
+
if (!Array.isArray(history) || history.length === 0) {
|
|
81
|
+
return "";
|
|
82
|
+
}
|
|
83
|
+
|
|
84
|
+
const lines: string[] = ["# Previous Steps in This Chain", ""];
|
|
85
|
+
|
|
86
|
+
for (const raw of history) {
|
|
87
|
+
if (!raw || typeof raw !== "object") continue;
|
|
88
|
+
const entry = raw as ChainHistoryEntry;
|
|
89
|
+
const stepName = entry.step ?? "unknown";
|
|
90
|
+
const outcome = entry.outcome ?? "unknown";
|
|
91
|
+
lines.push(`## Step ${stepName}: ${outcome}`);
|
|
92
|
+
lines.push(`Files created: ${entry.filesCreated?.length ? entry.filesCreated.join(", ") : "none"}`);
|
|
93
|
+
lines.push(`Files modified: ${entry.filesModified?.length ? entry.filesModified.join(", ") : "none"}`);
|
|
94
|
+
if (entry.decisions?.length) {
|
|
95
|
+
lines.push("Decisions:");
|
|
96
|
+
for (const d of entry.decisions) {
|
|
97
|
+
const rationale = (d && typeof d === "object" && "rationale" in d) ? String(d.rationale) : "(undocumented)";
|
|
98
|
+
const decisionOutcome = (d && typeof d === "object" && "outcome" in d) ? String(d.outcome) : "";
|
|
99
|
+
lines.push(`- ${rationale}${decisionOutcome ? ` → ${decisionOutcome}` : ""}`);
|
|
100
|
+
}
|
|
101
|
+
}
|
|
102
|
+
if (entry.nextSteps?.length) {
|
|
103
|
+
lines.push("Next steps:");
|
|
104
|
+
for (const ns of entry.nextSteps) lines.push(`- ${ns}`);
|
|
105
|
+
}
|
|
106
|
+
if (entry.outputText) {
|
|
107
|
+
lines.push("Output:");
|
|
108
|
+
lines.push(entry.outputText);
|
|
109
|
+
}
|
|
110
|
+
lines.push("");
|
|
111
|
+
}
|
|
112
|
+
|
|
113
|
+
// Append Fix C honesty markers verbatim — these note elided/oversized history.
|
|
114
|
+
const notes = context.__chainHistoryNotes;
|
|
115
|
+
if (Array.isArray(notes)) {
|
|
116
|
+
for (const note of notes) {
|
|
117
|
+
if (typeof note === "string" && note.length > 0) lines.push(note);
|
|
118
|
+
}
|
|
119
|
+
}
|
|
120
|
+
|
|
121
|
+
return lines.join("\n").trimEnd();
|
|
122
|
+
}
|
|
123
|
+
|
|
124
|
+
/**
|
|
125
|
+
* Map a completed team run's manifest+tasks → a chain `TaskResult`.
|
|
126
|
+
*
|
|
127
|
+
* Outcome is derived from the manifest status, cross-checked against task
|
|
128
|
+
* statuses (a run that completed but with a failed task reads as "partial").
|
|
129
|
+
* `usage.totalTokens` sums each task's input+output tokens; `duration` from
|
|
130
|
+
* manifest timestamps; `error` is best-effort from the first failed task or
|
|
131
|
+
* the manifest summary. `filesCreated`/`filesModified` are intentionally left
|
|
132
|
+
* unset — `TeamTaskState` does not carry them, so we never fabricate them.
|
|
133
|
+
*
|
|
134
|
+
* Extracted as a pure function for unit testability (acceptance criterion 6b).
|
|
135
|
+
*/
|
|
136
|
+
export function mapRunToTaskResult(
|
|
137
|
+
manifest: TeamRunManifest,
|
|
138
|
+
tasks: TeamTaskState[],
|
|
139
|
+
): TaskResult {
|
|
140
|
+
// Determine outcome from manifest status.
|
|
141
|
+
let outcome: TaskResult["outcome"];
|
|
142
|
+
if (manifest.status === "completed") {
|
|
143
|
+
outcome = "success";
|
|
144
|
+
} else if (
|
|
145
|
+
manifest.status === "failed" ||
|
|
146
|
+
manifest.status === "blocked" ||
|
|
147
|
+
manifest.status === "cancelled"
|
|
148
|
+
) {
|
|
149
|
+
outcome = "failure";
|
|
150
|
+
} else {
|
|
151
|
+
// queued / planning / running → not terminal yet
|
|
152
|
+
outcome = "partial";
|
|
153
|
+
}
|
|
154
|
+
|
|
155
|
+
// Cross-check tasks: if a manifest claims completed but a task failed, it is partial.
|
|
156
|
+
const hasFailed = tasks.some((t) => t.status === "failed" || t.status === "needs_attention");
|
|
157
|
+
const hasCompleted = tasks.some((t) => t.status === "completed");
|
|
158
|
+
if (hasFailed && (hasCompleted || outcome === "success")) {
|
|
159
|
+
outcome = "partial";
|
|
160
|
+
}
|
|
161
|
+
|
|
162
|
+
// Token usage: sum task input + output.
|
|
163
|
+
let totalTokens = 0;
|
|
164
|
+
for (const t of tasks) {
|
|
165
|
+
totalTokens += (t.usage?.input ?? 0) + (t.usage?.output ?? 0);
|
|
166
|
+
}
|
|
167
|
+
|
|
168
|
+
// Duration from manifest timestamps.
|
|
169
|
+
let duration = 0;
|
|
170
|
+
const created = Date.parse(manifest.createdAt);
|
|
171
|
+
const updated = Date.parse(manifest.updatedAt);
|
|
172
|
+
if (!Number.isNaN(created) && !Number.isNaN(updated) && updated >= created) {
|
|
173
|
+
duration = updated - created;
|
|
174
|
+
}
|
|
175
|
+
|
|
176
|
+
// Error: best-effort from first failed task, else manifest summary.
|
|
177
|
+
let error: string | undefined;
|
|
178
|
+
if (outcome !== "success") {
|
|
179
|
+
const failedTask = tasks.find((t) => t.status === "failed");
|
|
180
|
+
error = failedTask?.error ?? manifest.summary;
|
|
181
|
+
}
|
|
182
|
+
|
|
183
|
+
const taskResult: TaskResult = {
|
|
184
|
+
outcome,
|
|
185
|
+
usage: { totalTokens },
|
|
186
|
+
duration,
|
|
187
|
+
};
|
|
188
|
+
if (error) taskResult.error = error;
|
|
189
|
+
return taskResult;
|
|
190
|
+
}
|
|
191
|
+
|
|
192
|
+
/**
|
|
193
|
+
* Read completed tasks' result artifacts and concatenate their output text.
|
|
194
|
+
* Mirrors the pattern in task-output-context.ts:collectDependencyOutputContext
|
|
195
|
+
* (readIfSmallWithTee with baseDir = artifactsRoot).
|
|
196
|
+
*
|
|
197
|
+
* Returns the concatenated output text, or undefined if no artifacts were readable.
|
|
198
|
+
*/
|
|
199
|
+
export function readChainStepOutput(
|
|
200
|
+
manifest: TeamRunManifest,
|
|
201
|
+
tasks: TeamTaskState[],
|
|
202
|
+
): string | undefined {
|
|
203
|
+
const parts: string[] = [];
|
|
204
|
+
for (const t of tasks) {
|
|
205
|
+
if (t.status !== "completed" || !t.resultArtifact?.path) continue;
|
|
206
|
+
const read = readIfSmallWithTee(t.resultArtifact.path, {
|
|
207
|
+
baseDir: manifest.artifactsRoot,
|
|
208
|
+
});
|
|
209
|
+
if (read?.content && read.content.trim().length > 0) {
|
|
210
|
+
parts.push(read.content.trim());
|
|
211
|
+
}
|
|
212
|
+
}
|
|
213
|
+
return parts.length > 0 ? parts.join("\n\n") : undefined;
|
|
214
|
+
}
|
|
215
|
+
|
|
216
|
+
/**
|
|
217
|
+
* Concrete `ChainTaskRunner`: each step is a full team run via the injected
|
|
218
|
+
* `handleRun`. Step runIds are captured on `stepRunIds` so the dispatch can
|
|
219
|
+
* surface them in the summary and acceptance verification.
|
|
220
|
+
*
|
|
221
|
+
* Errors are intentionally NOT caught here — `runChain` wraps each step's
|
|
222
|
+
* `runTask` call in try/catch, records `outcome: "failure"`, and respects
|
|
223
|
+
* `continueOnError`. Letting exceptions propagate keeps the chain-runner's
|
|
224
|
+
* existing failure semantics intact.
|
|
225
|
+
*/
|
|
226
|
+
export class ChainTeamRunExecutor implements ChainTaskRunner {
|
|
227
|
+
/** RunId captured per executed step, in execution order. */
|
|
228
|
+
readonly stepRunIds: string[] = [];
|
|
229
|
+
|
|
230
|
+
private readonly handleRunRef: HandleRunFn;
|
|
231
|
+
private readonly ctx: TeamContext;
|
|
232
|
+
private readonly overrides: ChainExecutorOverrides;
|
|
233
|
+
|
|
234
|
+
constructor(opts: {
|
|
235
|
+
handleRun: HandleRunFn;
|
|
236
|
+
ctx: TeamContext;
|
|
237
|
+
overrides?: ChainExecutorOverrides;
|
|
238
|
+
}) {
|
|
239
|
+
this.handleRunRef = opts.handleRun;
|
|
240
|
+
this.ctx = opts.ctx;
|
|
241
|
+
this.overrides = opts.overrides ?? {};
|
|
242
|
+
}
|
|
243
|
+
|
|
244
|
+
async runTask(packet: TaskPacket): Promise<TaskResult> {
|
|
245
|
+
const context = packet.context ?? {};
|
|
246
|
+
|
|
247
|
+
// 1. Format chain history into the goal — this is how context reaches the worker.
|
|
248
|
+
// Mirrors the parentContext pattern in child-pi.ts (parent context + task delimiter).
|
|
249
|
+
const historyPrefix = formatChainHistory(context);
|
|
250
|
+
const enrichedGoal = historyPrefix
|
|
251
|
+
? `${historyPrefix}\n\n---\n# Current Chain Step\n${packet.goal}`
|
|
252
|
+
: packet.goal;
|
|
253
|
+
|
|
254
|
+
// 2. Resolve team/workflow/model: step config (set by executeStep) → overrides → default.
|
|
255
|
+
const stepTeam =
|
|
256
|
+
(context.__chainStepTeam as string | undefined) ??
|
|
257
|
+
this.overrides.team ??
|
|
258
|
+
"default";
|
|
259
|
+
const stepWorkflow =
|
|
260
|
+
(context.__chainStepWorkflow as string | undefined) ??
|
|
261
|
+
this.overrides.workflow;
|
|
262
|
+
const stepModel =
|
|
263
|
+
(context.__chainStepModel as string | undefined) ??
|
|
264
|
+
this.overrides.model;
|
|
265
|
+
|
|
266
|
+
// 3. Call handleRun for the heavy lifting. async:false forces each step to
|
|
267
|
+
// complete synchronously (overrides asyncByDefault) so the chain is sequential.
|
|
268
|
+
// NO `chain` key here — so the run.ts guard does not re-enter chain dispatch.
|
|
269
|
+
const runParams: TeamToolParamsValue = {
|
|
270
|
+
action: "run",
|
|
271
|
+
goal: enrichedGoal,
|
|
272
|
+
team: stepTeam,
|
|
273
|
+
async: false,
|
|
274
|
+
...(stepWorkflow ? { workflow: stepWorkflow } : {}),
|
|
275
|
+
...(stepModel ? { model: stepModel } : {}),
|
|
276
|
+
};
|
|
277
|
+
const runRes = await this.handleRunRef(runParams, this.ctx);
|
|
278
|
+
|
|
279
|
+
// 4. Extract runId + map to TaskResult. If no runId, the run was blocked before
|
|
280
|
+
// a manifest existed — map as a failure.
|
|
281
|
+
const runId = runRes.details?.runId;
|
|
282
|
+
if (!runId) {
|
|
283
|
+
return {
|
|
284
|
+
outcome: "failure",
|
|
285
|
+
error: textFromToolResult(runRes) || "Chain step produced no runId",
|
|
286
|
+
};
|
|
287
|
+
}
|
|
288
|
+
this.stepRunIds.push(runId);
|
|
289
|
+
|
|
290
|
+
const cwd = this.ctx.cwd ?? process.cwd();
|
|
291
|
+
const loaded = loadRunManifestById(cwd, runId);
|
|
292
|
+
if (!loaded) {
|
|
293
|
+
// Manifest not loadable — fall back to the tool result's own status.
|
|
294
|
+
return {
|
|
295
|
+
outcome: runRes.details?.status === "error" ? "failure" : "partial",
|
|
296
|
+
};
|
|
297
|
+
}
|
|
298
|
+
const result = mapRunToTaskResult(loaded.manifest, loaded.tasks);
|
|
299
|
+
const outputText = readChainStepOutput(loaded.manifest, loaded.tasks);
|
|
300
|
+
if (outputText) result.outputText = outputText;
|
|
301
|
+
return result;
|
|
302
|
+
}
|
|
303
|
+
}
|
|
304
|
+
|
|
305
|
+
// ─── test helpers (exported for unit/integration tests) ──────────────────
|
|
306
|
+
|
|
307
|
+
/**
|
|
308
|
+
* Write a minimal valid run fixture (manifest + tasks + events) to `cwd`'s
|
|
309
|
+
* state root so `loadRunManifestById(cwd, runId)` validates and returns it.
|
|
310
|
+
* Used by integration tests to exercise `ChainTeamRunExecutor` with a mocked
|
|
311
|
+
* `handleRun` while still going through the REAL manifest loader + result mapper.
|
|
312
|
+
*/
|
|
313
|
+
export function writeRunFixture(
|
|
314
|
+
cwd: string,
|
|
315
|
+
runId: string,
|
|
316
|
+
opts: {
|
|
317
|
+
status?: TeamRunManifest["status"];
|
|
318
|
+
tasks?: TeamTaskState[];
|
|
319
|
+
summary?: string;
|
|
320
|
+
/** Worker result text to write to results/<taskId>.txt (sets resultArtifact on tasks). */
|
|
321
|
+
resultText?: string;
|
|
322
|
+
/** Task ID for the result artifact (default: "01_task"). */
|
|
323
|
+
taskId?: string;
|
|
324
|
+
},
|
|
325
|
+
): TeamRunManifest {
|
|
326
|
+
const paths = createRunPaths(cwd, runId);
|
|
327
|
+
fs.mkdirSync(paths.stateRoot, { recursive: true });
|
|
328
|
+
const now = new Date().toISOString();
|
|
329
|
+
const start = new Date(Date.now() - 10_000).toISOString();
|
|
330
|
+
const taskId = opts.taskId ?? "01_task";
|
|
331
|
+
|
|
332
|
+
// Build tasks: if resultText is provided without explicit tasks, create a
|
|
333
|
+
// default completed task carrying a resultArtifact pointing at the result file.
|
|
334
|
+
let tasks = opts.tasks;
|
|
335
|
+
if (opts.resultText) {
|
|
336
|
+
const resultArtifact = {
|
|
337
|
+
kind: "result" as const,
|
|
338
|
+
path: `results/${taskId}.txt`,
|
|
339
|
+
createdAt: now,
|
|
340
|
+
producer: "worker",
|
|
341
|
+
retention: "run" as const,
|
|
342
|
+
};
|
|
343
|
+
if (!tasks || tasks.length === 0) {
|
|
344
|
+
tasks = [{
|
|
345
|
+
id: taskId,
|
|
346
|
+
runId,
|
|
347
|
+
role: "executor",
|
|
348
|
+
agent: "executor",
|
|
349
|
+
title: "(chain step fixture)",
|
|
350
|
+
status: "completed",
|
|
351
|
+
dependsOn: [],
|
|
352
|
+
cwd,
|
|
353
|
+
resultArtifact,
|
|
354
|
+
}];
|
|
355
|
+
} else {
|
|
356
|
+
// Attach resultArtifact to the first task if it doesn't have one.
|
|
357
|
+
tasks = tasks.map((t, i) => i === 0 && !t.resultArtifact
|
|
358
|
+
? { ...t, resultArtifact }
|
|
359
|
+
: t);
|
|
360
|
+
}
|
|
361
|
+
}
|
|
362
|
+
|
|
363
|
+
const manifest: TeamRunManifest = {
|
|
364
|
+
schemaVersion: 1,
|
|
365
|
+
runId,
|
|
366
|
+
team: "default",
|
|
367
|
+
workflow: "default",
|
|
368
|
+
goal: "(chain step fixture)",
|
|
369
|
+
status: opts.status ?? "completed",
|
|
370
|
+
workspaceMode: "single",
|
|
371
|
+
createdAt: start,
|
|
372
|
+
updatedAt: now,
|
|
373
|
+
cwd,
|
|
374
|
+
stateRoot: paths.stateRoot,
|
|
375
|
+
artifactsRoot: paths.artifactsRoot,
|
|
376
|
+
tasksPath: paths.tasksPath,
|
|
377
|
+
eventsPath: paths.eventsPath,
|
|
378
|
+
artifacts: [],
|
|
379
|
+
...(opts.summary ? { summary: opts.summary } : {}),
|
|
380
|
+
};
|
|
381
|
+
|
|
382
|
+
// Write the result artifact file + register on manifest if resultText provided.
|
|
383
|
+
if (opts.resultText) {
|
|
384
|
+
const resultDir = path.join(paths.artifactsRoot, "results");
|
|
385
|
+
fs.mkdirSync(resultDir, { recursive: true });
|
|
386
|
+
fs.writeFileSync(path.join(resultDir, `${taskId}.txt`), opts.resultText, "utf-8");
|
|
387
|
+
manifest.artifacts.push({
|
|
388
|
+
kind: "result",
|
|
389
|
+
path: `results/${taskId}.txt`,
|
|
390
|
+
createdAt: now,
|
|
391
|
+
producer: "worker",
|
|
392
|
+
retention: "run",
|
|
393
|
+
});
|
|
394
|
+
}
|
|
395
|
+
|
|
396
|
+
fs.writeFileSync(paths.manifestPath, JSON.stringify(manifest), "utf-8");
|
|
397
|
+
fs.writeFileSync(paths.tasksPath, JSON.stringify(tasks ?? []), "utf-8");
|
|
398
|
+
fs.writeFileSync(paths.eventsPath, "", "utf-8");
|
|
399
|
+
return manifest;
|
|
400
|
+
}
|
|
@@ -113,6 +113,15 @@ function scheduleBackgroundEarlyExitGuard(cwd: string, runId: string, pid: numbe
|
|
|
113
113
|
}
|
|
114
114
|
|
|
115
115
|
export async function handleRun(params: TeamToolParamsValue, ctx: TeamContext): Promise<PiTeamsToolResult> {
|
|
116
|
+
// CHAIN DISPATCH: runs before goal validation since a chain has no top-level
|
|
117
|
+
// goal. The injected handleRun reference breaks the run.ts ↔ chain-dispatch.ts
|
|
118
|
+
// import cycle; the lazy import defers the chain-executor cost until a chain is
|
|
119
|
+
// actually requested. Existing run/workflow paths below are unchanged.
|
|
120
|
+
if (params.chain) {
|
|
121
|
+
// LAZY: defer chain-dispatch import until a chain is actually requested.
|
|
122
|
+
const { handleChainRun } = await import("./chain-dispatch.ts");
|
|
123
|
+
return handleChainRun(params, ctx, handleRun);
|
|
124
|
+
}
|
|
116
125
|
const goal = params.goal ?? params.task;
|
|
117
126
|
if (!goal) return result("Run requires goal or task.", { action: "run", status: "error" }, true);
|
|
118
127
|
const intentPrefix = goal.length > 60 ? `${goal.slice(0, 57)}...` : goal;
|
|
@@ -43,10 +43,13 @@ export function ensureMemoryDir(memoryDir: string): void {
|
|
|
43
43
|
|
|
44
44
|
export function readMemoryIndex(memoryDir: string): string | undefined {
|
|
45
45
|
if (isSymlink(memoryDir)) return undefined;
|
|
46
|
-
const
|
|
46
|
+
const memPath = path.join(memoryDir, "MEMORY.md");
|
|
47
|
+
const content = safeReadMemoryFile(memPath);
|
|
47
48
|
if (content === undefined) return undefined;
|
|
48
49
|
const lines = content.split(/\r?\n/);
|
|
49
|
-
return lines.length > MAX_MEMORY_LINES
|
|
50
|
+
return lines.length > MAX_MEMORY_LINES
|
|
51
|
+
? `${lines.slice(0, MAX_MEMORY_LINES).join("\n")}\n... (truncated at 200 lines). Full file: ${memPath} — use the \`read\` tool if you need entries beyond the head.`
|
|
52
|
+
: content;
|
|
50
53
|
}
|
|
51
54
|
|
|
52
55
|
export function buildMemoryBlock(agentName: string, scope: AgentMemoryScope, cwd: string, writable: boolean): string {
|
|
@@ -25,6 +25,19 @@ import type { executeTeamRun as ExecuteTeamRunFn } from "./team-runner.ts";
|
|
|
25
25
|
import type { TeamRunManifest, TeamTaskState } from "../state/types.ts";
|
|
26
26
|
|
|
27
27
|
let _cachedExecuteTeamRun: typeof ExecuteTeamRunFn | undefined;
|
|
28
|
+
|
|
29
|
+
/** Maximum runtime for a single background run before the watchdog force-aborts
|
|
30
|
+
* it. Prevents zombie background-runner processes when a team run hangs forever
|
|
31
|
+
* (e.g. a hung child Pi process, a stuck lock, or a test that spawns a run
|
|
32
|
+
* without cleanup). Default 2h — generous for legitimate long runs (research
|
|
33
|
+
* workflows, goal-loops) but catches true zombies (observed: 10h+ stale test
|
|
34
|
+
* runs). Override via PI_CREW_MAX_RUN_MS env (milliseconds). The watchdog
|
|
35
|
+
* aborts via the shared AbortController, then force-exits after a grace
|
|
36
|
+
* period in case the abort signal does not propagate to all execution paths. */
|
|
37
|
+
const MAX_BACKGROUND_RUN_MS = (() => {
|
|
38
|
+
const env = Number.parseInt(process.env.PI_CREW_MAX_RUN_MS ?? "", 10);
|
|
39
|
+
return Number.isFinite(env) && env > 0 ? env : 2 * 60 * 60 * 1000;
|
|
40
|
+
})();
|
|
28
41
|
async function executeTeamRun(
|
|
29
42
|
...args: Parameters<typeof ExecuteTeamRunFn>
|
|
30
43
|
): Promise<Awaited<ReturnType<typeof ExecuteTeamRunFn>>> {
|
|
@@ -234,6 +247,7 @@ function runCleanup(
|
|
|
234
247
|
stopParentGuard: () => void,
|
|
235
248
|
stopHeartbeat: () => void,
|
|
236
249
|
keepAlive: NodeJS.Timeout,
|
|
250
|
+
watchdogTimer: NodeJS.Timeout,
|
|
237
251
|
exitDueToRejection: boolean,
|
|
238
252
|
eventsPath?: string,
|
|
239
253
|
): void {
|
|
@@ -245,6 +259,9 @@ function runCleanup(
|
|
|
245
259
|
// FIX: clearInterval FIRST, then kill children. This ensures the heartbeat
|
|
246
260
|
// interval is always cleaned up even if terminateActiveChildPiProcesses throws.
|
|
247
261
|
clearInterval(keepAlive);
|
|
262
|
+
// Clear the anti-zombie watchdog so a normally-completing run does not
|
|
263
|
+
// carry a pending force-exit timer into the exit path.
|
|
264
|
+
clearTimeout(watchdogTimer);
|
|
248
265
|
// FIX Issues #1, #2, #4: Wrap child process termination in try/catch so errors
|
|
249
266
|
// don't prevent the cleanup from completing. We log but don't re-throw since
|
|
250
267
|
// we're already exiting.
|
|
@@ -563,6 +580,37 @@ async function main(): Promise<void> {
|
|
|
563
580
|
// bounded by the 5s interval. The event loop exit is deferred at most 5s.
|
|
564
581
|
const keepAlive = setInterval(() => {}, 5000);
|
|
565
582
|
|
|
583
|
+
// WATCHDOG (anti-zombie): if the run exceeds MAX_BACKGROUND_RUN_MS without
|
|
584
|
+
// completing, abort it and force-exit. Without this, a hung team run
|
|
585
|
+
// (stuck child Pi, deadlocked lock, test that never cleans up) leaves the
|
|
586
|
+
// background-runner alive forever because keepAlive holds the event loop.
|
|
587
|
+
// The watchdog fires once; it is cleared in the finally block via runCleanup.
|
|
588
|
+
const watchdogTimer = setTimeout(() => {
|
|
589
|
+
console.error(`[background-runner] WATCHDOG: run ${runId} exceeded ${MAX_BACKGROUND_RUN_MS}ms — aborting (zombie prevention)`);
|
|
590
|
+
try {
|
|
591
|
+
appendEvent(manifest.eventsPath, {
|
|
592
|
+
type: "async.watchdog_fired",
|
|
593
|
+
runId,
|
|
594
|
+
message: `Run exceeded ${MAX_BACKGROUND_RUN_MS}ms and was force-aborted to prevent a zombie background-runner process.`,
|
|
595
|
+
data: { maxRunMs: MAX_BACKGROUND_RUN_MS },
|
|
596
|
+
});
|
|
597
|
+
} catch { /* best-effort event log */ }
|
|
598
|
+
// Signal the finally block to exit(1) after cleanup.
|
|
599
|
+
exitDueToRejection = true;
|
|
600
|
+
// Abort the in-flight team run via the shared signal (propagates to
|
|
601
|
+
// executeTeamRun → child-pi → kills child processes).
|
|
602
|
+
abortController.abort();
|
|
603
|
+
// Hard-exit safety net: if the abort does not propagate within 15s
|
|
604
|
+
// (e.g. a hung native call), force-kill so the process cannot linger.
|
|
605
|
+
const forceExit = setTimeout(() => {
|
|
606
|
+
console.error(`[background-runner] WATCHDOG: abort did not propagate within grace period — force-exiting`);
|
|
607
|
+
stopParentGuard();
|
|
608
|
+
try { terminateActiveChildPiProcesses(); } catch { /* best-effort */ }
|
|
609
|
+
process.exit(1);
|
|
610
|
+
}, 15_000);
|
|
611
|
+
forceExit.unref();
|
|
612
|
+
}, MAX_BACKGROUND_RUN_MS);
|
|
613
|
+
|
|
566
614
|
try {
|
|
567
615
|
debugLog(`[background-runner] about to call discoverAgents`);
|
|
568
616
|
const agents = allAgents(discoverAgents(cwd));
|
|
@@ -798,6 +846,7 @@ async function main(): Promise<void> {
|
|
|
798
846
|
stopParentGuard,
|
|
799
847
|
stopHeartbeat,
|
|
800
848
|
keepAlive,
|
|
849
|
+
watchdogTimer,
|
|
801
850
|
exitDueToRejection,
|
|
802
851
|
manifest.eventsPath,
|
|
803
852
|
);
|
|
@@ -496,25 +496,54 @@ export class ChainRunner {
|
|
|
496
496
|
return context;
|
|
497
497
|
}
|
|
498
498
|
|
|
499
|
-
|
|
500
|
-
|
|
499
|
+
const notes: string[] = [];
|
|
500
|
+
|
|
501
|
+
// Limit history size to prevent memory leak (H2). Emit a marker so a dropped
|
|
502
|
+
// entry does not vanish silently — this was the only silent-loss path in the
|
|
503
|
+
// output-handling machinery (see research-findings/output-handling-deep-dive.md §F).
|
|
504
|
+
let limitedHandoffs = handoffs;
|
|
505
|
+
if (handoffs.length > ChainRunner.MAX_CHAIN_HISTORY_SIZE) {
|
|
506
|
+
const dropped = handoffs.length - ChainRunner.MAX_CHAIN_HISTORY_SIZE;
|
|
507
|
+
notes.push(`[chain history limited to last ${ChainRunner.MAX_CHAIN_HISTORY_SIZE} entries; ${dropped} older entr${dropped === 1 ? "y" : "ies"} omitted]`);
|
|
508
|
+
limitedHandoffs = handoffs.slice(-ChainRunner.MAX_CHAIN_HISTORY_SIZE);
|
|
509
|
+
}
|
|
501
510
|
|
|
502
|
-
// Limit per-entry size to prevent memory issues from large artifacts
|
|
511
|
+
// Limit per-entry size to prevent memory issues from large artifacts.
|
|
503
512
|
const filteredHandoffs = limitedHandoffs.filter(h => {
|
|
504
513
|
const size = JSON.stringify(h).length;
|
|
505
514
|
return size <= ChainRunner.MAX_HANDOFF_ENTRY_SIZE;
|
|
506
515
|
});
|
|
516
|
+
const oversizedDropped = limitedHandoffs.length - filteredHandoffs.length;
|
|
517
|
+
if (oversizedDropped > 0) {
|
|
518
|
+
notes.push(`[${oversizedDropped} handoff entr${oversizedDropped === 1 ? "y" : "ies"} omitted (> ${ChainRunner.MAX_HANDOFF_ENTRY_SIZE} bytes each)]`);
|
|
519
|
+
}
|
|
507
520
|
|
|
508
|
-
|
|
509
|
-
|
|
510
|
-
|
|
521
|
+
// Track array-cap drops so capped lists do not vanish silently. Each field
|
|
522
|
+
// over its cap records the full count so a reader knows data was elided.
|
|
523
|
+
const arrayCapNotes: string[] = [];
|
|
524
|
+
const mapped = filteredHandoffs.map(h => {
|
|
525
|
+
const over: string[] = [];
|
|
526
|
+
if (h.filesCreated && h.filesCreated.length > 50) over.push(`filesCreated=${h.filesCreated.length}`);
|
|
527
|
+
if (h.filesModified && h.filesModified.length > 50) over.push(`filesModified=${h.filesModified.length}`);
|
|
528
|
+
if (h.decisions && h.decisions.length > 20) over.push(`decisions=${h.decisions.length}`);
|
|
529
|
+
if (h.nextSteps && h.nextSteps.length > 20) over.push(`nextSteps=${h.nextSteps.length}`);
|
|
530
|
+
if (over.length > 0) arrayCapNotes.push(`step ${h.taskId}: ${over.join(", ")} (capped)`);
|
|
531
|
+
return {
|
|
511
532
|
step: h.taskId,
|
|
512
533
|
outcome: h.outcome,
|
|
513
534
|
filesCreated: h.filesCreated?.slice(0, 50), // Limit array size
|
|
514
535
|
filesModified: h.filesModified?.slice(0, 50), // Limit array size
|
|
515
536
|
decisions: h.decisions?.slice(0, 20), // Limit array size
|
|
516
537
|
nextSteps: h.nextSteps?.slice(0, 20), // Limit array size
|
|
517
|
-
|
|
538
|
+
...(h.outputText ? { outputText: h.outputText.slice(0, 5000) } : {}), // Size-capped worker output
|
|
539
|
+
};
|
|
540
|
+
});
|
|
541
|
+
if (arrayCapNotes.length > 0) notes.push(`[list caps applied: ${arrayCapNotes.join("; ")}]`);
|
|
542
|
+
|
|
543
|
+
return {
|
|
544
|
+
...context,
|
|
545
|
+
__chainHistory: mapped,
|
|
546
|
+
...(notes.length > 0 ? { __chainHistoryNotes: notes } : {}),
|
|
518
547
|
};
|
|
519
548
|
}
|
|
520
549
|
|
|
@@ -525,13 +554,23 @@ export class ChainRunner {
|
|
|
525
554
|
config: ChainStep,
|
|
526
555
|
context: Record<string, unknown>
|
|
527
556
|
): Promise<TaskResult> {
|
|
557
|
+
// Pass step config (team/workflow/model) into the packet context so the
|
|
558
|
+
// concrete ChainTaskRunner can resolve which team/workflow to run. Without
|
|
559
|
+
// this, a step parsed to a @teamName reference would lose its team because
|
|
560
|
+
// the only channel from executeStep to runTask is the TaskPacket. Purely
|
|
561
|
+
// additive — adds sibling keys alongside __chainHistory/Notes.
|
|
528
562
|
const packet: TaskPacket = {
|
|
529
563
|
taskId: `chain-${Date.now()}-${config.name}`,
|
|
530
564
|
runId: "chain",
|
|
531
565
|
goal: config.inlineGoal ?? config.name,
|
|
532
566
|
summarizeThreshold: 3000,
|
|
533
567
|
collapseContext: true,
|
|
534
|
-
context
|
|
568
|
+
context: {
|
|
569
|
+
...context,
|
|
570
|
+
__chainStepTeam: config.team,
|
|
571
|
+
__chainStepWorkflow: config.workflow,
|
|
572
|
+
__chainStepModel: config.model,
|
|
573
|
+
},
|
|
535
574
|
};
|
|
536
575
|
|
|
537
576
|
return this.taskRunner.runTask(packet);
|