pi-crew 0.2.2 → 0.2.4

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.
Files changed (354) hide show
  1. package/AGENTS.md +57 -32
  2. package/CHANGELOG.md +466 -413
  3. package/LICENSE +21 -21
  4. package/NOTICE.md +16 -16
  5. package/README.md +323 -323
  6. package/docs/FEATURE_INTAKE.md +126 -0
  7. package/docs/HARNESS.md +86 -0
  8. package/docs/HARNESS_BACKLOG.md +41 -0
  9. package/docs/TEST_MATRIX.md +49 -0
  10. package/docs/actions-reference.md +595 -595
  11. package/docs/architecture.md +180 -180
  12. package/docs/code-review-2026-05-11.md +592 -0
  13. package/docs/commands-reference.md +347 -347
  14. package/docs/comparison-pi-subagents-vs-pi-crew.md +303 -0
  15. package/docs/decisions/0001-durable-state.md +41 -0
  16. package/docs/decisions/0002-child-process-for-async.md +42 -0
  17. package/docs/decisions/0003-depth-guard.md +36 -0
  18. package/docs/decisions/0004-execfile-over-exec.md +34 -0
  19. package/docs/decisions/0005-no-parameter-properties.md +49 -0
  20. package/docs/decisions/0006-publish-bundled-esm.md +63 -0
  21. package/docs/decisions/0007-active-run-binary-index.md +54 -0
  22. package/docs/decisions/0008-child-pi-warm-pool.md +61 -0
  23. package/docs/decisions/README.md +23 -0
  24. package/docs/followup-plan-2026-05-12.md +463 -0
  25. package/docs/followup-review-2026-05-12.md +297 -0
  26. package/docs/followup-review-round3-2026-05-12.md +342 -0
  27. package/docs/followup-review-round4-2026-05-13.md +107 -0
  28. package/docs/implementation-plan-top3.md +333 -0
  29. package/docs/live-mailbox-runtime.md +36 -36
  30. package/docs/next-upgrade-roadmap.md +808 -808
  31. package/docs/oh-my-pi-research.md +509 -0
  32. package/docs/perf/baseline-2026-05.md +113 -0
  33. package/docs/perf/final-report-2026-05.md +206 -0
  34. package/docs/perf/sprint-1-report.md +71 -0
  35. package/docs/perf/sprint-2-report.md +81 -0
  36. package/docs/perf/sprint-2.5-report.md +53 -0
  37. package/docs/perf/sprint-3-report.md +36 -0
  38. package/docs/perf/sprint-4-report.md +47 -0
  39. package/docs/perf/sprint-5-report.md +51 -0
  40. package/docs/perf/sprint-6-report.md +94 -0
  41. package/docs/perf/sprint-7-report.md +74 -0
  42. package/docs/perf/upgrade-plan-2026-05.md +147 -0
  43. package/docs/pi-subagents3-deep-analysis.md +508 -0
  44. package/docs/product/README.md +31 -0
  45. package/docs/product/platform.md +27 -0
  46. package/docs/product/runtime-safety.md +37 -0
  47. package/docs/product/team-run.md +39 -0
  48. package/docs/product/team-tool.md +37 -0
  49. package/docs/publishing.md +65 -65
  50. package/docs/resource-formats.md +134 -134
  51. package/docs/runtime-analysis-child-vs-live.md +171 -0
  52. package/docs/runtime-flow.md +148 -148
  53. package/docs/runtime-migration-in-process-analysis.md +250 -0
  54. package/docs/stories/README.md +30 -0
  55. package/docs/stories/backlog.md +36 -0
  56. package/docs/templates/decision.md +27 -0
  57. package/docs/templates/story.md +44 -0
  58. package/docs/templates/validation-report.md +32 -0
  59. package/docs/usage.md +238 -238
  60. package/index.ts +7 -6
  61. package/install.mjs +65 -65
  62. package/package.json +107 -99
  63. package/schema.json +222 -222
  64. package/skills/child-pi-spawning/SKILL.md +213 -0
  65. package/skills/context-artifact-hygiene/SKILL.md +32 -0
  66. package/skills/event-log-tracing/SKILL.md +299 -0
  67. package/skills/git-master/SKILL.md +225 -24
  68. package/skills/live-agent-lifecycle/SKILL.md +192 -0
  69. package/skills/mailbox-interactive/SKILL.md +300 -19
  70. package/skills/model-routing-context/SKILL.md +94 -0
  71. package/skills/multi-perspective-review/SKILL.md +88 -0
  72. package/skills/read-only-explorer/SKILL.md +250 -26
  73. package/skills/safe-bash/SKILL.md +307 -21
  74. package/skills/verification-before-done/SKILL.md +11 -2
  75. package/skills/widget-rendering/SKILL.md +258 -0
  76. package/skills/workspace-isolation/SKILL.md +202 -0
  77. package/skills/worktree-isolation/SKILL.md +202 -18
  78. package/src/adapters/claude-adapter.ts +25 -25
  79. package/src/adapters/codex-adapter.ts +21 -21
  80. package/src/adapters/cursor-adapter.ts +17 -17
  81. package/src/adapters/export-util.ts +137 -137
  82. package/src/adapters/index.ts +15 -15
  83. package/src/adapters/registry.ts +18 -18
  84. package/src/adapters/types.ts +23 -23
  85. package/src/agents/agent-config.ts +38 -38
  86. package/src/agents/agent-serializer.ts +38 -38
  87. package/src/agents/discover-agents.ts +121 -118
  88. package/src/config/config.ts +740 -858
  89. package/src/config/defaults.ts +96 -96
  90. package/src/config/drift-detector.ts +211 -211
  91. package/src/config/markers.ts +327 -327
  92. package/src/config/resilient-parser.ts +109 -108
  93. package/src/config/suggestions.ts +74 -74
  94. package/src/config/types.ts +199 -0
  95. package/src/extension/async-notifier.ts +123 -89
  96. package/src/extension/autonomous-policy.ts +169 -169
  97. package/src/extension/cross-extension-rpc.ts +104 -103
  98. package/src/extension/help.ts +47 -47
  99. package/src/extension/import-index.ts +69 -69
  100. package/src/extension/management.ts +395 -382
  101. package/src/extension/notification-router.ts +116 -116
  102. package/src/extension/notification-sink.ts +51 -51
  103. package/src/extension/project-init.ts +168 -168
  104. package/src/extension/register.ts +859 -668
  105. package/src/extension/registration/artifact-cleanup.ts +15 -15
  106. package/src/extension/registration/command-utils.ts +54 -54
  107. package/src/extension/registration/commands.ts +559 -452
  108. package/src/extension/registration/compaction-guard.ts +125 -125
  109. package/src/extension/registration/subagent-helpers.ts +102 -102
  110. package/src/extension/registration/subagent-tools.ts +220 -158
  111. package/src/extension/registration/team-tool.ts +159 -98
  112. package/src/extension/registration/viewers.ts +29 -0
  113. package/src/extension/result-watcher.ts +128 -128
  114. package/src/extension/run-bundle-schema.ts +89 -89
  115. package/src/extension/run-export.ts +73 -73
  116. package/src/extension/run-import.ts +84 -84
  117. package/src/extension/run-index.ts +94 -94
  118. package/src/extension/run-maintenance.ts +142 -142
  119. package/src/extension/session-summary.ts +8 -8
  120. package/src/extension/team-manager-command.ts +96 -95
  121. package/src/extension/team-recommendation.ts +188 -188
  122. package/src/extension/team-tool/api.ts +5 -2
  123. package/src/extension/team-tool/cancel.ts +224 -209
  124. package/src/extension/team-tool/config-patch.ts +36 -36
  125. package/src/extension/team-tool/context.ts +60 -60
  126. package/src/extension/team-tool/doctor.ts +242 -242
  127. package/src/extension/team-tool/handle-settings.ts +421 -195
  128. package/src/extension/team-tool/inspect.ts +41 -41
  129. package/src/extension/team-tool/lifecycle-actions.ts +139 -139
  130. package/src/extension/team-tool/parallel-dispatch.ts +156 -156
  131. package/src/extension/team-tool/plan.ts +19 -19
  132. package/src/extension/team-tool/respond.ts +112 -111
  133. package/src/extension/team-tool/run.ts +246 -228
  134. package/src/extension/team-tool/status.ts +110 -110
  135. package/src/extension/team-tool-types.ts +13 -13
  136. package/src/extension/team-tool.ts +16 -4
  137. package/src/extension/tool-result.ts +16 -16
  138. package/src/extension/validate-resources.ts +77 -77
  139. package/src/hooks/registry.ts +61 -61
  140. package/src/hooks/types.ts +40 -40
  141. package/src/i18n.ts +184 -184
  142. package/src/observability/correlation.ts +35 -35
  143. package/src/observability/event-to-metric.ts +68 -68
  144. package/src/observability/exporters/adapter.ts +30 -30
  145. package/src/observability/exporters/otlp-exporter.ts +106 -92
  146. package/src/observability/exporters/prometheus-exporter.ts +54 -54
  147. package/src/observability/metric-registry.ts +87 -87
  148. package/src/observability/metric-retention.ts +54 -54
  149. package/src/observability/metric-sink.ts +81 -56
  150. package/src/observability/metrics-primitives.ts +167 -167
  151. package/src/prompt/prompt-runtime.ts +72 -72
  152. package/src/runtime/adaptive-plan.ts +338 -0
  153. package/src/runtime/agent-control.ts +169 -169
  154. package/src/runtime/agent-memory.ts +72 -72
  155. package/src/runtime/agent-observability.ts +114 -114
  156. package/src/runtime/async-marker.ts +26 -26
  157. package/src/runtime/async-runner.ts +153 -79
  158. package/src/runtime/attention-events.ts +28 -28
  159. package/src/runtime/auto-resume.ts +100 -100
  160. package/src/runtime/background-runner.ts +122 -88
  161. package/src/runtime/cancellation.ts +61 -61
  162. package/src/runtime/capability-inventory.ts +116 -116
  163. package/src/runtime/child-pi-pool.ts +68 -0
  164. package/src/runtime/child-pi.ts +541 -463
  165. package/src/runtime/code-summary.ts +247 -247
  166. package/src/runtime/compaction-summary.ts +271 -271
  167. package/src/runtime/concurrency.ts +58 -58
  168. package/src/runtime/crash-recovery.ts +317 -301
  169. package/src/runtime/crew-agent-records.ts +379 -281
  170. package/src/runtime/crew-agent-runtime.ts +60 -60
  171. package/src/runtime/cross-extension-rpc.ts +72 -0
  172. package/src/runtime/custom-tools/irc-tool.ts +201 -201
  173. package/src/runtime/custom-tools/submit-result-tool.ts +90 -90
  174. package/src/runtime/deadletter.ts +47 -47
  175. package/src/runtime/delivery-coordinator.ts +176 -176
  176. package/src/runtime/delta-conflict.ts +360 -360
  177. package/src/runtime/diagnostic-export.ts +102 -102
  178. package/src/runtime/direct-run.ts +35 -35
  179. package/src/runtime/effectiveness.ts +82 -81
  180. package/src/runtime/errors/crew-errors.ts +166 -0
  181. package/src/runtime/event-stream-bridge.ts +92 -92
  182. package/src/runtime/foreground-control.ts +82 -82
  183. package/src/runtime/green-contract.ts +46 -46
  184. package/src/runtime/group-join.ts +234 -106
  185. package/src/runtime/heartbeat-watcher.ts +145 -124
  186. package/src/runtime/iteration-hooks.ts +267 -264
  187. package/src/runtime/live-agent-control.ts +88 -88
  188. package/src/runtime/live-agent-manager.ts +377 -179
  189. package/src/runtime/live-control-realtime.ts +36 -36
  190. package/src/runtime/live-session-runtime.ts +676 -599
  191. package/src/runtime/loop-gates.ts +129 -129
  192. package/src/runtime/manifest-cache.ts +263 -263
  193. package/src/runtime/mcp-proxy.ts +113 -113
  194. package/src/runtime/metric-parser.ts +40 -40
  195. package/src/runtime/model-fallback.ts +282 -274
  196. package/src/runtime/model-resolver.ts +118 -0
  197. package/src/runtime/output-validator.ts +187 -187
  198. package/src/runtime/overflow-recovery.ts +175 -175
  199. package/src/runtime/parallel-research.ts +44 -44
  200. package/src/runtime/parallel-utils.ts +156 -156
  201. package/src/runtime/parent-guard.ts +80 -80
  202. package/src/runtime/phase-progress.ts +217 -217
  203. package/src/runtime/pi-args.ts +165 -165
  204. package/src/runtime/pi-json-output.ts +111 -111
  205. package/src/runtime/pi-spawn.ts +167 -167
  206. package/src/runtime/policy-engine.ts +79 -79
  207. package/src/runtime/post-checks.ts +125 -122
  208. package/src/runtime/post-exit-stdio-guard.ts +86 -86
  209. package/src/runtime/process-status.ts +97 -73
  210. package/src/runtime/progress-event-coalescer.ts +43 -43
  211. package/src/runtime/recovery-recipes.ts +74 -74
  212. package/src/runtime/retry-executor.ts +81 -81
  213. package/src/runtime/role-permission.ts +39 -39
  214. package/src/runtime/run-tracker.ts +99 -0
  215. package/src/runtime/runtime-policy.ts +21 -0
  216. package/src/runtime/runtime-resolver.ts +94 -90
  217. package/src/runtime/scheduler.ts +294 -0
  218. package/src/runtime/semaphore.ts +131 -131
  219. package/src/runtime/sensitive-paths.ts +92 -92
  220. package/src/runtime/session-usage.ts +79 -79
  221. package/src/runtime/settings-store.ts +103 -0
  222. package/src/runtime/sidechain-output.ts +29 -29
  223. package/src/runtime/skill-instructions.ts +222 -222
  224. package/src/runtime/stale-reconciler.ts +198 -189
  225. package/src/runtime/streaming-output.ts +47 -0
  226. package/src/runtime/subagent-manager.ts +404 -395
  227. package/src/runtime/subprocess-tool-registry.ts +67 -67
  228. package/src/runtime/task-display.ts +38 -38
  229. package/src/runtime/task-graph-scheduler.ts +122 -122
  230. package/src/runtime/task-graph.ts +207 -207
  231. package/src/runtime/task-output-context.ts +177 -177
  232. package/src/runtime/task-packet.ts +93 -93
  233. package/src/runtime/task-quality.ts +207 -207
  234. package/src/runtime/task-runner/capabilities.ts +78 -78
  235. package/src/runtime/task-runner/live-executor.ts +131 -113
  236. package/src/runtime/task-runner/progress.ts +119 -119
  237. package/src/runtime/task-runner/prompt-builder.ts +139 -139
  238. package/src/runtime/task-runner/prompt-pipeline.ts +64 -64
  239. package/src/runtime/task-runner/result-utils.ts +14 -14
  240. package/src/runtime/task-runner/run-projection.ts +103 -103
  241. package/src/runtime/task-runner/state-helpers.ts +22 -22
  242. package/src/runtime/task-runner.ts +469 -458
  243. package/src/runtime/team-runner.ts +693 -945
  244. package/src/runtime/usage-tracker.ts +71 -0
  245. package/src/runtime/worker-heartbeat.ts +21 -21
  246. package/src/runtime/worker-startup.ts +57 -57
  247. package/src/runtime/workflow-state.ts +187 -187
  248. package/src/runtime/yield-handler.ts +190 -189
  249. package/src/schema/config-schema.ts +172 -168
  250. package/src/schema/team-tool-schema.ts +126 -125
  251. package/src/schema/validation-types.ts +151 -148
  252. package/src/skills/discover-skills.ts +67 -67
  253. package/src/skills/skill-templates.ts +374 -374
  254. package/src/state/active-run-registry.ts +227 -191
  255. package/src/state/artifact-store.ts +130 -129
  256. package/src/state/atomic-write.ts +262 -178
  257. package/src/state/blob-store.ts +116 -116
  258. package/src/state/contracts.ts +111 -111
  259. package/src/state/event-log-rotation.ts +161 -158
  260. package/src/state/event-log.ts +383 -240
  261. package/src/state/event-reconstructor.ts +217 -217
  262. package/src/state/jsonl-writer.ts +82 -82
  263. package/src/state/locks.ts +146 -148
  264. package/src/state/mailbox.ts +446 -405
  265. package/src/state/state-store.ts +364 -351
  266. package/src/state/task-claims.ts +44 -44
  267. package/src/state/types.ts +285 -285
  268. package/src/state/usage.ts +29 -29
  269. package/src/subagents/async-entry.ts +1 -1
  270. package/src/subagents/index.ts +3 -3
  271. package/src/subagents/live/control.ts +1 -1
  272. package/src/subagents/live/manager.ts +1 -1
  273. package/src/subagents/live/realtime.ts +1 -1
  274. package/src/subagents/live/session-runtime.ts +1 -1
  275. package/src/subagents/manager.ts +1 -1
  276. package/src/subagents/spawn.ts +1 -1
  277. package/src/teams/discover-teams.ts +116 -116
  278. package/src/teams/team-config.ts +27 -27
  279. package/src/teams/team-serializer.ts +38 -38
  280. package/src/types/diff.d.ts +18 -18
  281. package/src/ui/agent-management-overlay.ts +144 -144
  282. package/src/ui/crew-widget.ts +487 -370
  283. package/src/ui/dashboard-panes/agents-pane.ts +109 -28
  284. package/src/ui/dashboard-panes/cancellation-pane.ts +42 -42
  285. package/src/ui/dashboard-panes/capability-pane.ts +59 -59
  286. package/src/ui/dashboard-panes/health-pane.ts +30 -30
  287. package/src/ui/dashboard-panes/mailbox-pane.ts +35 -35
  288. package/src/ui/dashboard-panes/progress-pane.ts +30 -30
  289. package/src/ui/dashboard-panes/transcript-pane.ts +10 -10
  290. package/src/ui/heartbeat-aggregator.ts +63 -63
  291. package/src/ui/keybinding-map.ts +97 -94
  292. package/src/ui/live-conversation-overlay.ts +152 -0
  293. package/src/ui/live-run-sidebar.ts +180 -180
  294. package/src/ui/mascot.ts +442 -442
  295. package/src/ui/overlays/agent-picker-overlay.ts +57 -57
  296. package/src/ui/overlays/confirm-overlay.ts +58 -58
  297. package/src/ui/overlays/mailbox-compose-overlay.ts +144 -144
  298. package/src/ui/overlays/mailbox-compose-preview.ts +63 -63
  299. package/src/ui/overlays/mailbox-detail-overlay.ts +122 -122
  300. package/src/ui/pi-ui-compat.ts +57 -57
  301. package/src/ui/powerbar-publisher.ts +221 -197
  302. package/src/ui/render-scheduler.ts +216 -143
  303. package/src/ui/run-action-dispatcher.ts +118 -117
  304. package/src/ui/run-dashboard.ts +526 -464
  305. package/src/ui/run-event-bus.ts +208 -208
  306. package/src/ui/run-snapshot-cache.ts +826 -777
  307. package/src/ui/settings-overlay.ts +721 -0
  308. package/src/ui/snapshot-types.ts +86 -70
  309. package/src/ui/theme-adapter.ts +190 -190
  310. package/src/ui/tool-progress-formatter.ts +89 -0
  311. package/src/ui/transcript-cache.ts +94 -94
  312. package/src/ui/transcript-viewer.ts +335 -335
  313. package/src/utils/conflict-detect.ts +662 -0
  314. package/src/utils/env-filter.ts +30 -0
  315. package/src/utils/file-coalescer.ts +86 -86
  316. package/src/utils/frontmatter.ts +68 -68
  317. package/src/utils/fs-watch.ts +88 -31
  318. package/src/utils/gh-protocol.ts +479 -0
  319. package/src/utils/ids.ts +17 -17
  320. package/src/utils/incremental-reader.ts +104 -104
  321. package/src/utils/internal-error.ts +6 -6
  322. package/src/utils/names.ts +27 -27
  323. package/src/utils/paths.ts +102 -63
  324. package/src/utils/redaction.ts +44 -44
  325. package/src/utils/resolve-shell.ts +34 -0
  326. package/src/utils/safe-paths.ts +47 -47
  327. package/src/utils/scan-cache.ts +136 -136
  328. package/src/utils/sleep.ts +2 -1
  329. package/src/utils/sse-parser.ts +134 -134
  330. package/src/utils/task-name-generator.ts +337 -337
  331. package/src/utils/timings.ts +33 -33
  332. package/src/utils/visual.ts +243 -198
  333. package/src/workflows/discover-workflows.ts +139 -139
  334. package/src/workflows/validate-workflow.ts +40 -40
  335. package/src/workflows/workflow-config.ts +26 -26
  336. package/src/workflows/workflow-serializer.ts +32 -32
  337. package/src/worktree/branch-freshness.ts +45 -45
  338. package/src/worktree/cleanup.ts +75 -72
  339. package/src/worktree/worktree-manager.ts +188 -146
  340. package/teams/default.team.md +12 -12
  341. package/teams/fast-fix.team.md +11 -11
  342. package/teams/implementation.team.md +18 -18
  343. package/teams/parallel-research.team.md +14 -14
  344. package/teams/research.team.md +11 -11
  345. package/teams/review.team.md +12 -12
  346. package/tsconfig.json +19 -19
  347. package/workflows/default.workflow.md +30 -30
  348. package/workflows/fast-fix.workflow.md +23 -23
  349. package/workflows/implementation.workflow.md +43 -43
  350. package/workflows/parallel-research.workflow.md +46 -46
  351. package/workflows/research.workflow.md +22 -22
  352. package/workflows/review.workflow.md +30 -30
  353. package/skills/task-packet/SKILL.md +0 -28
  354. package/skills/verify-evidence/SKILL.md +0 -27
@@ -1,458 +1,469 @@
1
- import * as fs from "node:fs";
2
- import type { AgentConfig } from "../agents/agent-config.ts";
3
- import type { CrewLimitsConfig, CrewRuntimeConfig } from "../config/config.ts";
4
- import type { ArtifactDescriptor, OperationTerminalEvidence, TeamRunManifest, TeamTaskState, UsageState } from "../state/types.ts";
5
- import { writeArtifact } from "../state/artifact-store.ts";
6
- import { appendEvent } from "../state/event-log.ts";
7
- import { saveRunManifest } from "../state/state-store.ts";
8
- import { createTaskClaim } from "../state/task-claims.ts";
9
- import { createWorkerHeartbeat, touchWorkerHeartbeat } from "./worker-heartbeat.ts";
10
- import type { WorkflowStep } from "../workflows/workflow-config.ts";
11
- import { captureWorktreeDiff, captureWorktreeDiffStat, prepareTaskWorkspace } from "../worktree/worktree-manager.ts";
12
- import { buildConfiguredModelRouting, formatModelAttemptNote, isRetryableModelFailure, type ModelAttemptSummary } from "./model-fallback.ts";
13
- import { parsePiJsonOutput, type ParsedPiJsonOutput } from "./pi-json-output.ts";
14
- import { runChildPi } from "./child-pi.ts";
15
- import { buildTaskPacket } from "./task-packet.ts";
16
- import { executeHook, appendHookEvent } from "../hooks/registry.ts";
17
- import { createVerificationEvidence } from "./green-contract.ts";
18
- import { createStartupEvidence } from "./worker-startup.ts";
19
- import { permissionForRole } from "./role-permission.ts";
20
- import { collectDependencyOutputContext, renderDependencyOutputContext, writeTaskInputsArtifact, writeTaskSharedOutput } from "./task-output-context.ts";
21
- import { appendCrewAgentEvent, appendCrewAgentOutput, emptyCrewAgentProgress, recordFromTask, upsertCrewAgent } from "./crew-agent-records.ts";
22
- import { reserveControlChannel } from "./agent-control.ts";
23
- import { parseSessionUsage } from "./session-usage.ts";
24
- import type { CrewAgentProgress, CrewRuntimeKind } from "./crew-agent-runtime.ts";
25
- import { shouldAppendProgressEventUpdate, type ProgressEventSummary } from "./progress-event-coalescer.ts";
26
- import { coordinationBridgeInstructions, renderTaskPrompt } from "./task-runner/prompt-builder.ts";
27
- import { buildWorkerPromptPipeline } from "./task-runner/prompt-pipeline.ts";
28
- import { buildWorkerCapabilityInventory } from "./task-runner/capabilities.ts";
29
- import { applyAgentProgressEvent, applyUsageToProgress, progressEventSummary, shouldFlushProgressEvent } from "./task-runner/progress.ts";
30
- import { checkpointTask, persistSingleTaskUpdate, updateTask } from "./task-runner/state-helpers.ts";
31
- import { cleanResultText, isFinalChildEvent } from "./task-runner/result-utils.ts";
32
- import { evaluateCompletionMutationGuard } from "./completion-guard.ts";
33
- import { cancellationReasonFromSignal, buildSyntheticTerminalEvidence } from "./cancellation.ts";
34
- import { appendTaskAttentionEvent } from "./attention-events.ts";
35
- import { parseSupervisorContactFromLine, recordSupervisorContact } from "./supervisor-contact.ts";
36
- import { registerStreamBridge, bridgeEventFromJsonEvent } from "./event-stream-bridge.ts";
37
- import { renderSkillInstructions } from "./skill-instructions.ts";
38
- import { DEFAULT_YIELD_CONFIG, extractYieldResult, hasYieldInOutput, isYieldEvent, registerYieldTool, type YieldResult } from "./yield-handler.ts";
39
- import { validateWorkerOutput, type OutputValidationResult } from "./output-validator.ts";
40
-
41
- // Register the submit_result tool handler so subprocess events can extract yield data.
42
- registerYieldTool();
43
-
44
- export interface TaskRunnerInput {
45
- manifest: TeamRunManifest;
46
- tasks: TeamTaskState[];
47
- task: TeamTaskState;
48
- step: WorkflowStep;
49
- agent: AgentConfig;
50
- signal?: AbortSignal;
51
- executeWorkers: boolean;
52
- runtimeKind?: CrewRuntimeKind;
53
- runtimeConfig?: CrewRuntimeConfig;
54
- parentContext?: string;
55
- parentModel?: unknown;
56
- modelRegistry?: unknown;
57
- modelOverride?: string;
58
- teamRoleModel?: string;
59
- teamRoleSkills?: string[] | false;
60
- skillOverride?: string[] | false;
61
- limits?: CrewLimitsConfig;
62
- dependencyContextText?: string;
63
- skillBlock?: string;
64
- skillNames?: string[];
65
- skillPaths?: string[];
66
- /** Optional callback for JSON events from child Pi. Used for overflow recovery tracking. */
67
- onJsonEvent?: (taskId: string, runId: string, event: unknown) => void;
68
- }
69
-
70
- export async function runTeamTask(input: TaskRunnerInput): Promise<{ manifest: TeamRunManifest; tasks: TeamTaskState[] }> {
71
- let manifest = input.manifest;
72
- // H4: registerStreamBridge inside try so dispose() in finally is safe
73
- let streamBridge: ReturnType<typeof registerStreamBridge> | undefined;
74
- try {
75
- streamBridge = registerStreamBridge(manifest.runId);
76
- const workspace = prepareTaskWorkspace(manifest, input.task);
77
- const worktree = workspace.worktreePath && workspace.branch ? { path: workspace.worktreePath, branch: workspace.branch, reused: workspace.reused ?? false } : input.task.worktree;
78
- const taskPacket = buildTaskPacket({ manifest, step: input.step, taskId: input.task.id, cwd: workspace.cwd, worktreePath: worktree?.path });
79
- const dependencyContext = collectDependencyOutputContext(manifest, input.tasks, input.task, input.step);
80
- const dependencyContextText = input.dependencyContextText ?? renderDependencyOutputContext(dependencyContext);
81
- let task: TeamTaskState = {
82
- ...input.task,
83
- cwd: workspace.cwd,
84
- worktree,
85
- taskPacket,
86
- status: "running",
87
- startedAt: new Date().toISOString(),
88
- claim: createTaskClaim(`task-runner:${input.task.id}`),
89
- heartbeat: createWorkerHeartbeat(input.task.id),
90
- agentProgress: input.task.agentProgress ?? emptyCrewAgentProgress(),
91
- ...(dependencyContextText ? { dependencyContextText } : {}),
92
- // Reserve control channel before spawn so cancel/steer can target this task immediately
93
- controlReservation: reserveControlChannel(input.task.id, manifest.runId),
94
- } as TeamTaskState;
95
- let tasks = updateTask(input.tasks, task);
96
- const runtimeKind = input.runtimeKind ?? (input.executeWorkers ? "child-process" : "scaffold");
97
- tasks = persistSingleTaskUpdate(manifest, tasks, task);
98
- if (runtimeKind === "child-process") ({ task, tasks } = checkpointTask(manifest, tasks, task, "started"));
99
- upsertCrewAgent(manifest, recordFromTask(manifest, task, runtimeKind));
100
- appendEvent(manifest.eventsPath, { type: "task.started", runId: manifest.runId, taskId: task.id, data: { role: task.role, agent: task.agent, runtime: runtimeKind, cwd: task.cwd, worktreePath: workspace.worktreePath, worktreeBranch: workspace.branch, worktreeReused: workspace.reused } });
101
- // Emit immediate UI notification so widget shows agent as "running" within ~100ms
102
- // instead of waiting for child process first JSON event (2-5s delay).
103
- streamBridge?.handler({ runId: manifest.runId, taskId: task.id, eventType: "task.started", timestamp: Date.now() });
104
- const permissionMode = permissionForRole(task.role);
105
- const renderedSkills = input.skillBlock === undefined ? renderSkillInstructions({ cwd: task.cwd, role: task.role, agent: input.agent, teamRole: { skills: input.teamRoleSkills }, step: input.step, override: input.skillOverride }) : undefined;
106
- const skillBlock = input.skillBlock ?? renderedSkills?.block;
107
- const skillNames = input.skillNames ?? renderedSkills?.names;
108
- const skillPaths = input.skillPaths ?? renderedSkills?.paths;
109
-
110
- const promptResult = await renderTaskPrompt(manifest, input.step, task, input.agent, skillBlock);
111
- const prompt = promptResult.full;
112
- const promptArtifact = writeArtifact(manifest.artifactsRoot, {
113
- kind: "prompt",
114
- relativePath: `prompts/${task.id}.md`,
115
- content: `${prompt}\n`,
116
- producer: task.id,
117
- });
118
-
119
- let resultArtifact: ArtifactDescriptor;
120
- let logArtifact: ArtifactDescriptor | undefined;
121
- let transcriptArtifact: ArtifactDescriptor | undefined;
122
- let exitCode: number | null = 0;
123
- let error: string | undefined;
124
- let modelAttempts: ModelAttemptSummary[] | undefined;
125
- let parsedOutput: ParsedPiJsonOutput | undefined;
126
- let finalStdout = "";
127
- let transcriptPath: string | undefined;
128
- let terminalEvidence: OperationTerminalEvidence[] = [];
129
- const collectedJsonEvents: Record<string, unknown>[] = [];
130
-
131
- let startupEvidence = createStartupEvidence({ command: runtimeKind === "child-process" ? "pi" : runtimeKind === "live-session" ? "live-session" : "safe-scaffold", startedAt: new Date(task.startedAt ?? new Date().toISOString()), finishedAt: new Date(), promptSentAt: new Date(task.startedAt ?? new Date().toISOString()), promptAccepted: true, exitCode: 0 });
132
- const inputsArtifact = writeTaskInputsArtifact(manifest, task, dependencyContext);
133
- const skillArtifact = skillBlock ? writeArtifact(manifest.artifactsRoot, {
134
- kind: "metadata",
135
- relativePath: `metadata/${task.id}.skills.md`,
136
- content: [`Selected skills: ${skillNames?.join(", ") ?? "(none)"}`, `Skill paths passed to child Pi: ${(skillPaths ?? []).length}`, "", skillBlock, ""].join("\n"),
137
- producer: task.id,
138
- }) : undefined;
139
- const coordinationArtifact = writeArtifact(manifest.artifactsRoot, {
140
- kind: "metadata",
141
- relativePath: `metadata/${task.id}.coordination-bridge.md`,
142
- content: `${coordinationBridgeInstructions(task)}\n`,
143
- producer: task.id,
144
- });
145
- if (runtimeKind === "child-process") {
146
- const modelRoutingPlan = buildConfiguredModelRouting({ overrideModel: input.modelOverride, stepModel: input.step.model, teamRoleModel: input.teamRoleModel, agentModel: input.agent.model, fallbackModels: input.agent.fallbackModels, parentModel: input.parentModel, modelRegistry: input.modelRegistry, cwd: task.cwd });
147
- const candidates = modelRoutingPlan.candidates;
148
- const attemptModels = candidates.length > 0 ? candidates : [undefined];
149
- const logs: string[] = [];
150
- let finalStderr = "";
151
- modelAttempts = [];
152
- transcriptPath = `${manifest.artifactsRoot}/transcripts/${task.id}.jsonl`;
153
- let finalCheckpointWritten = false;
154
- let lastAgentRecordPersistedAt = 0;
155
- let lastHeartbeatPersistedAt = 0;
156
- let lastRunProgressPersistedAt = 0;
157
- let lastRunProgressSummary: ProgressEventSummary | undefined;
158
- const persistHeartbeat = (force = false): void => {
159
- const now = Date.now();
160
- if (!force && now - lastHeartbeatPersistedAt < 1000) return;
161
- lastHeartbeatPersistedAt = now;
162
- task = { ...task, heartbeat: touchWorkerHeartbeat(task.heartbeat ?? createWorkerHeartbeat(task.id)) };
163
- tasks = persistSingleTaskUpdate(manifest, tasks, task);
164
- };
165
- const persistChildProgress = (event: unknown, force = false): void => {
166
- const now = Date.now();
167
- if (force || shouldFlushProgressEvent(event) || now - lastAgentRecordPersistedAt >= 500) {
168
- upsertCrewAgent(manifest, recordFromTask(manifest, task, "child-process"));
169
- lastAgentRecordPersistedAt = now;
170
- }
171
- const summary = progressEventSummary(task, event);
172
- const decision = shouldAppendProgressEventUpdate({ previous: lastRunProgressSummary, next: summary, nowMs: now, lastAppendMs: lastRunProgressPersistedAt || undefined, minIntervalMs: 1000, force });
173
- if (decision.shouldAppend) {
174
- appendEvent(manifest.eventsPath, { type: "task.progress", runId: manifest.runId, taskId: task.id, data: { ...summary, coalesceReason: decision.reason } });
175
- lastRunProgressSummary = summary;
176
- lastRunProgressPersistedAt = now;
177
- }
178
- };
179
- for (let i = 0; i < attemptModels.length; i++) {
180
- const model = attemptModels[i];
181
- const attemptStartedAt = new Date();
182
- const pendingAttempt: ModelAttemptSummary = { model: model ?? "default", success: false };
183
- task = { ...task, modelAttempts: [...modelAttempts, pendingAttempt] };
184
- tasks = updateTask(tasks, task);
185
- upsertCrewAgent(manifest, recordFromTask(manifest, task, "child-process"));
186
- const childResult = await runChildPi({
187
- cwd: task.cwd,
188
- task: prompt,
189
- agent: input.agent,
190
- model,
191
- signal: input.signal,
192
- transcriptPath,
193
- maxDepth: input.limits?.maxTaskDepth,
194
- skillPaths,
195
- onSpawn: (pid) => {
196
- ({ task, tasks } = checkpointTask(manifest, tasks, task, "child-spawned", pid));
197
- },
198
- onStdoutLine: (line) => {
199
- appendCrewAgentOutput(manifest, task.id, line);
200
- persistHeartbeat();
201
- // Check for supervisor contact requests from child Pi
202
- const contact = parseSupervisorContactFromLine(line);
203
- if (contact) {
204
- recordSupervisorContact(manifest, { runId: manifest.runId, ...contact });
205
- }
206
- },
207
- onJsonEvent: (event) => {
208
- appendCrewAgentEvent(manifest, task.id, event);
209
- if (event && typeof event === "object" && !Array.isArray(event)) collectedJsonEvents.push(event as Record<string, unknown>);
210
- persistHeartbeat();
211
- task = { ...task, agentProgress: applyAgentProgressEvent(task.agentProgress ?? emptyCrewAgentProgress(), event, task.startedAt) };
212
- tasks = updateTask(tasks, task);
213
- // Bridge event to UI event bus for near-instant updates
214
- try {
215
- const bridgeEvent = bridgeEventFromJsonEvent(manifest.runId, task.id, event);
216
- if (bridgeEvent) streamBridge?.handler(bridgeEvent);
217
- } catch { /* bridge errors should not affect task */ }
218
- // Feed overflow recovery tracker
219
- if (input.onJsonEvent) {
220
- try {
221
- input.onJsonEvent(task.id, manifest.runId, event);
222
- } catch { /* overflow tracking errors should not affect task */ }
223
- }
224
- if (!finalCheckpointWritten && isFinalChildEvent(event)) {
225
- finalCheckpointWritten = true;
226
- ({ task, tasks } = checkpointTask(manifest, tasks, task, "child-stdout-final"));
227
- }
228
- persistChildProgress(event);
229
- },
230
- });
231
- const evidenceStatus = childResult.exitStatus?.cancelled ? "cancelled" : childResult.error || (childResult.exitCode && childResult.exitCode !== 0) ? "failed" : "completed";
232
- terminalEvidence = [...terminalEvidence, { operation: "worker", status: evidenceStatus, startedAt: attemptStartedAt.toISOString(), finishedAt: new Date().toISOString(), ...(input.signal?.aborted ? { reason: cancellationReasonFromSignal(input.signal) } : {}), ...(childResult.exitStatus ? { exitStatus: childResult.exitStatus } : {}) }];
233
- if (evidenceStatus === "cancelled") {
234
- const cancelReason = input.signal?.aborted ? cancellationReasonFromSignal(input.signal) : { code: "caller_cancelled" as const, message: "Worker cancelled." };
235
- terminalEvidence.push(buildSyntheticTerminalEvidence("tool", cancelReason, attemptStartedAt.toISOString()));
236
- appendEvent(manifest.eventsPath, { type: "worker.cancelled", runId: manifest.runId, taskId: task.id, message: cancelReason.message, data: { terminalEvidence: terminalEvidence.at(-1) } });
237
- }
238
- startupEvidence = createStartupEvidence({ command: "pi", startedAt: attemptStartedAt, finishedAt: new Date(), promptSentAt: attemptStartedAt, promptAccepted: childResult.exitCode === 0 && !childResult.error, stderr: childResult.stderr, error: childResult.error, exitCode: childResult.exitCode });
239
- exitCode = childResult.exitCode;
240
- finalStdout = childResult.stdout;
241
- finalStderr = childResult.stderr;
242
- parsedOutput = parsePiJsonOutput(fs.existsSync(transcriptPath) ? fs.readFileSync(transcriptPath, "utf-8") : childResult.stdout);
243
- error = childResult.error || (childResult.exitCode && childResult.exitCode !== 0 ? childResult.stderr || `Child Pi exited with ${childResult.exitCode}` : undefined);
244
- persistHeartbeat(true);
245
- persistChildProgress({ type: "attempt_finished" }, true);
246
- const attempt: ModelAttemptSummary = { model: model ?? "default", success: !error, exitCode, error };
247
- modelAttempts.push(attempt);
248
- task = { ...task, modelAttempts: [...modelAttempts] };
249
- tasks = updateTask(tasks, task);
250
- logs.push(`MODEL ATTEMPT ${i + 1}: ${attempt.model}`, `success=${attempt.success}`, `exitCode=${attempt.exitCode ?? "null"}`, attempt.error ? `error=${attempt.error}` : "", "");
251
- if (!error) break;
252
- const nextModel = attemptModels[i + 1];
253
- if (!nextModel || !isRetryableModelFailure(error)) break;
254
- logs.push(formatModelAttemptNote(attempt, nextModel), "");
255
- }
256
- resultArtifact = writeArtifact(manifest.artifactsRoot, {
257
- kind: "result",
258
- relativePath: `results/${task.id}.txt`,
259
- content: cleanResultText(parsedOutput?.finalText) ?? cleanResultText(finalStdout) ?? cleanResultText(finalStderr) ?? "(no output)",
260
- producer: task.id,
261
- });
262
- logArtifact = writeArtifact(manifest.artifactsRoot, {
263
- kind: "log",
264
- relativePath: `logs/${task.id}.log`,
265
- content: [...logs, `finalExitCode=${exitCode ?? "null"}`, `jsonEvents=${parsedOutput?.jsonEvents ?? 0}`, parsedOutput?.usage ? `usage=${JSON.stringify(parsedOutput.usage)}` : "", "", "STDOUT:", finalStdout, "", "STDERR:", finalStderr].join("\n"),
266
- producer: task.id,
267
- });
268
- const successfulAttemptIndex = modelAttempts.findIndex((attempt) => attempt.success);
269
- const usedAttempt = successfulAttemptIndex === -1 ? Math.max(0, modelAttempts.length - 1) : successfulAttemptIndex;
270
- const resolvedModel = modelAttempts[usedAttempt]?.model ?? candidates[0] ?? "default";
271
- const fallbackReason = usedAttempt > 0 ? modelAttempts[usedAttempt - 1]?.error : undefined;
272
- task = { ...task, modelRouting: { requested: modelRoutingPlan.requested, resolved: resolvedModel, fallbackChain: candidates, reason: fallbackReason ?? modelRoutingPlan.reason, usedAttempt } };
273
- tasks = updateTask(tasks, task);
274
- const sessionUsage = parseSessionUsage(transcriptPath);
275
- const effectiveUsage = parsedOutput?.usage ?? sessionUsage;
276
- if (effectiveUsage) {
277
- parsedOutput = { ...(parsedOutput ?? { jsonEvents: 0, textEvents: [] }), usage: effectiveUsage };
278
- task = { ...task, usage: effectiveUsage, agentProgress: applyUsageToProgress(task.agentProgress, effectiveUsage) };
279
- tasks = updateTask(tasks, task);
280
- upsertCrewAgent(manifest, recordFromTask(manifest, task, "child-process"));
281
- }
282
- if (fs.existsSync(transcriptPath)) {
283
- transcriptArtifact = writeArtifact(manifest.artifactsRoot, {
284
- kind: "log",
285
- relativePath: `transcripts/${task.id}.jsonl`,
286
- content: fs.readFileSync(transcriptPath, "utf-8"),
287
- producer: task.id,
288
- });
289
- }
290
- task = { ...task, resultArtifact, ...(logArtifact ? { logArtifact } : {}), ...(transcriptArtifact ? { transcriptArtifact } : {}) };
291
- tasks = updateTask(tasks, task);
292
- ({ task, tasks } = checkpointTask(manifest, tasks, task, "artifact-written"));
293
- } else if (runtimeKind === "live-session") {
294
- const { runLiveTask } = await import("./task-runner/live-executor.ts");
295
- const live = await runLiveTask({ manifest, tasks, task, step: input.step, agent: input.agent, prompt, signal: input.signal, runtimeConfig: input.runtimeConfig, parentContext: input.parentContext, parentModel: input.parentModel, modelRegistry: input.modelRegistry, modelOverride: input.modelOverride, teamRoleModel: input.teamRoleModel });
296
- task = live.task;
297
- tasks = live.tasks;
298
- startupEvidence = live.startupEvidence;
299
- exitCode = live.exitCode;
300
- error = live.error;
301
- parsedOutput = live.parsedOutput;
302
- resultArtifact = live.resultArtifact;
303
- logArtifact = live.logArtifact;
304
- transcriptArtifact = live.transcriptArtifact;
305
- } else {
306
- resultArtifact = writeArtifact(manifest.artifactsRoot, {
307
- kind: "result",
308
- relativePath: `results/${task.id}.md`,
309
- content: [
310
- `# ${task.id}`,
311
- "",
312
- "Worker execution is disabled in this scaffold-safe run.",
313
- "The prompt artifact contains the exact task that will be sent to a child Pi worker when execution is enabled.",
314
- ].join("\n"),
315
- producer: task.id,
316
- });
317
- }
318
-
319
- // --- Yield-based completion contract ---
320
- let yieldResult: YieldResult | undefined;
321
- const yieldEnabled = input.runtimeConfig?.yield?.enabled ?? DEFAULT_YIELD_CONFIG.enabled;
322
- if (yieldEnabled && collectedJsonEvents.length > 0) {
323
- if (hasYieldInOutput(collectedJsonEvents)) {
324
- const yieldEvent = collectedJsonEvents.find((e) => isYieldEvent(e));
325
- if (yieldEvent) {
326
- yieldResult = extractYieldResult(yieldEvent);
327
- }
328
- } else if (!error) {
329
- appendEvent(manifest.eventsPath, { type: "task.attention", runId: manifest.runId, taskId: task.id, message: "Worker completed without calling submit_result tool.", data: { activityState: "needs_attention", reason: "no_yield" } });
330
- }
331
- }
332
-
333
- const diffArtifact = workspace.worktreePath ? writeArtifact(manifest.artifactsRoot, {
334
- kind: "diff",
335
- relativePath: `diffs/${task.id}.diff`,
336
- content: captureWorktreeDiff(workspace.worktreePath),
337
- producer: task.id,
338
- }) : undefined;
339
- const diffStatArtifact = workspace.worktreePath ? writeArtifact(manifest.artifactsRoot, {
340
- kind: "metadata",
341
- relativePath: `metadata/${task.id}.diff-stat.json`,
342
- content: `${JSON.stringify({ ...captureWorktreeDiffStat(workspace.worktreePath), syntheticPaths: workspace.syntheticPaths ?? [], nodeModulesLinked: workspace.nodeModulesLinked ?? false }, null, 2)}\n`,
343
- producer: task.id,
344
- }) : undefined;
345
-
346
- const mutationGuardMode = input.runtimeConfig?.completionMutationGuard ?? "warn";
347
- const mutationGuard = !error && mutationGuardMode !== "off" ? evaluateCompletionMutationGuard({ role: task.role, taskText: `${task.title}\n${input.step.task}`, transcriptPath: runtimeKind === "child-process" ? transcriptPath : transcriptArtifact?.path, stdout: finalStdout }) : undefined;
348
- if (mutationGuard?.reason === "no_mutation_observed") {
349
- appendTaskAttentionEvent({
350
- manifest,
351
- taskId: task.id,
352
- message: "Implementation-style task completed without an observed mutation tool call.",
353
- data: { activityState: "needs_attention", reason: "completion_guard", taskId: task.id, agentName: task.agent, observedTools: mutationGuard.observedTools, suggestedAction: mutationGuardMode === "fail" ? "Review the worker output and rerun with a concrete implementation task." : "Review the worker output; set runtime.completionMutationGuard='fail' to enforce this." },
354
- });
355
- task = { ...task, agentProgress: { ...(task.agentProgress ?? emptyCrewAgentProgress()), activityState: "needs_attention" } };
356
- if (mutationGuardMode === "fail") {
357
- error = "Completion mutation guard failed: implementation-style task completed without an observed mutation tool call.";
358
- exitCode = exitCode === 0 ? 1 : exitCode;
359
- if (modelAttempts?.length) {
360
- modelAttempts = modelAttempts.map((attempt, index) => index === modelAttempts!.length - 1 ? { ...attempt, success: false, exitCode, error } : attempt);
361
- }
362
- }
363
- tasks = updateTask(tasks, task);
364
- }
365
-
366
- // --- Output format validation (caveman Phase 4) ---
367
- // Validate worker output against the role's output contract.
368
- // On failure: emit attention event but don't fail the task.
369
- let outputValidation: OutputValidationResult | undefined;
370
- if (!error) {
371
- const outputText = parsedOutput?.finalText ?? finalStdout;
372
- if (outputText) {
373
- outputValidation = validateWorkerOutput(task.role, outputText);
374
- if (!outputValidation.valid) {
375
- appendEvent(manifest.eventsPath, { type: "task.output_validation", runId: manifest.runId, taskId: task.id, data: { valid: false, formatMatch: outputValidation.formatMatch, structurePreserved: outputValidation.structurePreserved, issues: outputValidation.issues } });
376
- task = { ...task, agentProgress: { ...(task.agentProgress ?? emptyCrewAgentProgress()), activityState: "needs_attention" } };
377
- tasks = updateTask(tasks, task);
378
- }
379
- }
380
- }
381
-
382
- task = {
383
- ...task,
384
- status: error ? "failed" : "completed",
385
- finishedAt: new Date().toISOString(),
386
- exitCode,
387
- modelAttempts,
388
- usage: parsedOutput?.usage,
389
- jsonEvents: parsedOutput?.jsonEvents,
390
- agentProgress: error && task.agentProgress?.currentTool ? { ...task.agentProgress, failedTool: task.agentProgress.currentTool } : task.agentProgress,
391
- error,
392
- verification: createVerificationEvidence(taskPacket.verification, !error, error ? `Task failed: ${error}` : runtimeKind === "scaffold" ? "Safe scaffold mode; verification commands were not executed." : `${runtimeKind} worker finished without reporting a verification failure.`),
393
- promptArtifact,
394
- resultArtifact,
395
- claim: undefined,
396
- heartbeat: touchWorkerHeartbeat(task.heartbeat ?? createWorkerHeartbeat(task.id), { alive: false }),
397
- workerExitStatus: terminalEvidence.at(-1)?.exitStatus,
398
- terminalEvidence: terminalEvidence.length ? [...(task.terminalEvidence ?? []), ...terminalEvidence] : task.terminalEvidence,
399
- ...(logArtifact ? { logArtifact } : {}),
400
- ...(transcriptArtifact ? { transcriptArtifact } : {}),
401
- };
402
- tasks = updateTask(tasks, task);
403
- const packetArtifact = writeArtifact(manifest.artifactsRoot, {
404
- kind: "metadata",
405
- relativePath: `metadata/${task.id}.task-packet.json`,
406
- content: `${JSON.stringify(task.taskPacket, null, 2)}\n`,
407
- producer: task.id,
408
- });
409
- const verificationArtifact = writeArtifact(manifest.artifactsRoot, {
410
- kind: "metadata",
411
- relativePath: `metadata/${task.id}.verification.json`,
412
- content: `${JSON.stringify(task.verification, null, 2)}\n`,
413
- producer: task.id,
414
- });
415
- const sharedOutputArtifact = writeTaskSharedOutput(manifest, input.step, task);
416
- const startupArtifact = writeArtifact(manifest.artifactsRoot, {
417
- kind: "metadata",
418
- relativePath: `metadata/${task.id}.startup-evidence.json`,
419
- content: `${JSON.stringify(startupEvidence, null, 2)}\n`,
420
- producer: task.id,
421
- });
422
- const permissionArtifact = writeArtifact(manifest.artifactsRoot, {
423
- kind: "metadata",
424
- relativePath: `metadata/${task.id}.permission.json`,
425
- content: `${JSON.stringify({ role: task.role, permissionMode }, null, 2)}\n`,
426
- producer: task.id,
427
- });
428
- const capabilityArtifact = writeArtifact(manifest.artifactsRoot, {
429
- kind: "metadata",
430
- relativePath: `metadata/${task.id}.capabilities.json`,
431
- content: `${JSON.stringify(buildWorkerCapabilityInventory({ taskId: task.id, role: task.role, agent: input.agent, runtime: runtimeKind, permissionMode, skillNames, skillPaths, skillsDisabled: input.skillOverride === false || input.teamRoleSkills === false, modelOverride: input.modelOverride, teamRoleModel: input.teamRoleModel, stepModel: input.step.model }), null, 2)}\n`,
432
- producer: task.id,
433
- });
434
- const promptPipelineArtifact = writeArtifact(manifest.artifactsRoot, {
435
- kind: "metadata",
436
- relativePath: `metadata/${task.id}.prompt-pipeline.json`,
437
- content: `${JSON.stringify(buildWorkerPromptPipeline({ artifactsRoot: manifest.artifactsRoot, taskId: task.id, promptArtifact, inputsArtifact, skillArtifact, capabilityArtifact, coordinationArtifact, skillInstructionCount: skillNames?.length ?? 0, skillsDisabled: input.skillOverride === false || input.teamRoleSkills === false }), null, 2)}\n`,
438
- producer: task.id,
439
- });
440
- const outputValidationArtifact = outputValidation ? writeArtifact(manifest.artifactsRoot, {
441
- kind: "metadata",
442
- relativePath: `metadata/${task.id}.output-validation.json`,
443
- content: `${JSON.stringify(outputValidation, null, 2)}\n`,
444
- producer: task.id,
445
- }) : undefined;
446
- manifest = { ...manifest, updatedAt: new Date().toISOString(), artifacts: [...manifest.artifacts, promptArtifact, resultArtifact, inputsArtifact, coordinationArtifact, ...(skillArtifact ? [skillArtifact] : []), packetArtifact, verificationArtifact, startupArtifact, permissionArtifact, capabilityArtifact, promptPipelineArtifact, ...(outputValidationArtifact ? [outputValidationArtifact] : []), ...(sharedOutputArtifact ? [sharedOutputArtifact] : []), ...(logArtifact ? [logArtifact] : []), ...(transcriptArtifact ? [transcriptArtifact] : []), ...(diffArtifact ? [diffArtifact] : []), ...(diffStatArtifact ? [diffStatArtifact] : [])] };
447
- saveRunManifest(manifest);
448
- tasks = persistSingleTaskUpdate(manifest, tasks, task);
449
- upsertCrewAgent(manifest, recordFromTask(manifest, task, runtimeKind));
450
- // Execute task_result hook before emitting terminal event
451
- const hookReport = await executeHook("task_result", { runId: manifest.runId, taskId: task.id, cwd: manifest.cwd });
452
- appendHookEvent(manifest, hookReport);
453
- appendEvent(manifest.eventsPath, { type: error ? "task.failed" : "task.completed", runId: manifest.runId, taskId: task.id, message: error });
454
- return { manifest, tasks };
455
- } finally {
456
- streamBridge?.dispose();
457
- }
458
- }
1
+ import * as fs from "node:fs";
2
+ import type { AgentConfig } from "../agents/agent-config.ts";
3
+ import type { CrewLimitsConfig, CrewRuntimeConfig } from "../config/config.ts";
4
+ import type { ArtifactDescriptor, OperationTerminalEvidence, TeamRunManifest, TeamTaskState, UsageState } from "../state/types.ts";
5
+ import { writeArtifact } from "../state/artifact-store.ts";
6
+ import { appendEvent, appendEventFireAndForget } from "../state/event-log.ts";
7
+ import { saveRunManifest } from "../state/state-store.ts";
8
+ import { createTaskClaim } from "../state/task-claims.ts";
9
+ import { createWorkerHeartbeat, touchWorkerHeartbeat } from "./worker-heartbeat.ts";
10
+ import type { WorkflowStep } from "../workflows/workflow-config.ts";
11
+ import { captureWorktreeDiff, captureWorktreeDiffStat, prepareTaskWorkspace } from "../worktree/worktree-manager.ts";
12
+ import { buildConfiguredModelRouting, formatModelAttemptNote, isRetryableModelFailure, type ModelAttemptSummary } from "./model-fallback.ts";
13
+ import { parsePiJsonOutput, type ParsedPiJsonOutput } from "./pi-json-output.ts";
14
+ import { runChildPi, type ChildPiLifecycleEvent } from "./child-pi.ts";
15
+ import { buildTaskPacket } from "./task-packet.ts";
16
+ import { executeHook, appendHookEvent } from "../hooks/registry.ts";
17
+ import { createVerificationEvidence } from "./green-contract.ts";
18
+ import { createStartupEvidence } from "./worker-startup.ts";
19
+ import { permissionForRole } from "./role-permission.ts";
20
+ import { collectDependencyOutputContext, renderDependencyOutputContext, writeTaskInputsArtifact, writeTaskSharedOutput } from "./task-output-context.ts";
21
+ import { appendCrewAgentEvent, appendCrewAgentOutput, emptyCrewAgentProgress, recordFromTask, upsertCrewAgent } from "./crew-agent-records.ts";
22
+ import { reserveControlChannel } from "./agent-control.ts";
23
+ import { parseSessionUsage } from "./session-usage.ts";
24
+ import type { CrewAgentProgress, CrewRuntimeKind } from "./crew-agent-runtime.ts";
25
+ import { shouldAppendProgressEventUpdate, type ProgressEventSummary } from "./progress-event-coalescer.ts";
26
+ import { coordinationBridgeInstructions, renderTaskPrompt } from "./task-runner/prompt-builder.ts";
27
+ import { buildWorkerPromptPipeline } from "./task-runner/prompt-pipeline.ts";
28
+ import { buildWorkerCapabilityInventory } from "./task-runner/capabilities.ts";
29
+ import { applyAgentProgressEvent, applyUsageToProgress, progressEventSummary, shouldFlushProgressEvent } from "./task-runner/progress.ts";
30
+ import { checkpointTask, persistSingleTaskUpdate, updateTask } from "./task-runner/state-helpers.ts";
31
+ import { cleanResultText, isFinalChildEvent } from "./task-runner/result-utils.ts";
32
+ import { evaluateCompletionMutationGuard } from "./completion-guard.ts";
33
+ import { cancellationReasonFromSignal, buildSyntheticTerminalEvidence } from "./cancellation.ts";
34
+ import { appendTaskAttentionEvent } from "./attention-events.ts";
35
+ import { parseSupervisorContactFromLine, recordSupervisorContact } from "./supervisor-contact.ts";
36
+ import { registerStreamBridge, bridgeEventFromJsonEvent } from "./event-stream-bridge.ts";
37
+ import { renderSkillInstructions } from "./skill-instructions.ts";
38
+ import { DEFAULT_YIELD_CONFIG, extractYieldResult, hasYieldInOutput, isYieldEvent, registerYieldTool, type YieldResult } from "./yield-handler.ts";
39
+ import { validateWorkerOutput, type OutputValidationResult } from "./output-validator.ts";
40
+
41
+ // Register the submit_result tool handler so subprocess events can extract yield data.
42
+ registerYieldTool();
43
+
44
+ export interface TaskRunnerInput {
45
+ manifest: TeamRunManifest;
46
+ tasks: TeamTaskState[];
47
+ task: TeamTaskState;
48
+ step: WorkflowStep;
49
+ agent: AgentConfig;
50
+ signal?: AbortSignal;
51
+ executeWorkers: boolean;
52
+ runtimeKind?: CrewRuntimeKind;
53
+ /** Per-role runtime override resolved from isolation policy. Takes precedence over runtimeKind. */
54
+ taskRuntimeOverride?: CrewRuntimeKind;
55
+ runtimeConfig?: CrewRuntimeConfig;
56
+ parentContext?: string;
57
+ parentModel?: unknown;
58
+ modelRegistry?: unknown;
59
+ modelOverride?: string;
60
+ teamRoleModel?: string;
61
+ teamRoleSkills?: string[] | false;
62
+ skillOverride?: string[] | false;
63
+ limits?: CrewLimitsConfig;
64
+ dependencyContextText?: string;
65
+ skillBlock?: string;
66
+ skillNames?: string[];
67
+ skillPaths?: string[];
68
+ /** Workspace where this task run was initiated — used for session-scoped live-agent visibility. */
69
+ workspaceId: string;
70
+ /** Optional callback for JSON events from child Pi. Used for overflow recovery tracking. */
71
+ onJsonEvent?: (taskId: string, runId: string, event: unknown) => void;
72
+ }
73
+
74
+ export async function runTeamTask(input: TaskRunnerInput): Promise<{ manifest: TeamRunManifest; tasks: TeamTaskState[] }> {
75
+ let manifest = input.manifest;
76
+ // H4: registerStreamBridge inside try so dispose() in finally is safe
77
+ let streamBridge: ReturnType<typeof registerStreamBridge> | undefined;
78
+ try {
79
+ streamBridge = registerStreamBridge(manifest.runId);
80
+ const workspace = prepareTaskWorkspace(manifest, input.task);
81
+ const worktree = workspace.worktreePath && workspace.branch ? { path: workspace.worktreePath, branch: workspace.branch, reused: workspace.reused ?? false } : input.task.worktree;
82
+ const taskPacket = buildTaskPacket({ manifest, step: input.step, taskId: input.task.id, cwd: workspace.cwd, worktreePath: worktree?.path });
83
+ const dependencyContext = collectDependencyOutputContext(manifest, input.tasks, input.task, input.step);
84
+ const dependencyContextText = input.dependencyContextText ?? renderDependencyOutputContext(dependencyContext);
85
+ let task: TeamTaskState = {
86
+ ...input.task,
87
+ cwd: workspace.cwd,
88
+ worktree,
89
+ taskPacket,
90
+ status: "running",
91
+ startedAt: new Date().toISOString(),
92
+ claim: createTaskClaim(`task-runner:${input.task.id}`),
93
+ heartbeat: createWorkerHeartbeat(input.task.id),
94
+ agentProgress: input.task.agentProgress ?? emptyCrewAgentProgress(),
95
+ ...(dependencyContextText ? { dependencyContextText } : {}),
96
+ // Reserve control channel before spawn so cancel/steer can target this task immediately
97
+ controlReservation: reserveControlChannel(input.task.id, manifest.runId),
98
+ } as TeamTaskState;
99
+ let tasks = updateTask(input.tasks, task);
100
+ const runtimeKind = input.taskRuntimeOverride ?? input.runtimeKind ?? (input.executeWorkers ? "child-process" : "scaffold");
101
+ tasks = persistSingleTaskUpdate(manifest, tasks, task);
102
+ if (runtimeKind === "child-process") ({ task, tasks } = checkpointTask(manifest, tasks, task, "started"));
103
+ upsertCrewAgent(manifest, recordFromTask(manifest, task, runtimeKind));
104
+ appendEvent(manifest.eventsPath, { type: "task.started", runId: manifest.runId, taskId: task.id, data: { role: task.role, agent: task.agent, runtime: runtimeKind, cwd: task.cwd, worktreePath: workspace.worktreePath, worktreeBranch: workspace.branch, worktreeReused: workspace.reused } });
105
+ // Emit immediate UI notification so widget shows agent as "running" within ~100ms
106
+ // instead of waiting for child process first JSON event (2-5s delay).
107
+ streamBridge?.handler({ runId: manifest.runId, taskId: task.id, eventType: "task.started", timestamp: Date.now() });
108
+ const permissionMode = permissionForRole(task.role);
109
+ const renderedSkills = input.skillBlock === undefined ? renderSkillInstructions({ cwd: task.cwd, role: task.role, agent: input.agent, teamRole: { skills: input.teamRoleSkills }, step: input.step, override: input.skillOverride }) : undefined;
110
+ const skillBlock = input.skillBlock ?? renderedSkills?.block;
111
+ const skillNames = input.skillNames ?? renderedSkills?.names;
112
+ const skillPaths = input.skillPaths ?? renderedSkills?.paths;
113
+
114
+ const promptResult = await renderTaskPrompt(manifest, input.step, task, input.agent, skillBlock);
115
+ const prompt = promptResult.full;
116
+ const promptArtifact = writeArtifact(manifest.artifactsRoot, {
117
+ kind: "prompt",
118
+ relativePath: `prompts/${task.id}.md`,
119
+ content: `${prompt}\n`,
120
+ producer: task.id,
121
+ });
122
+
123
+ let resultArtifact: ArtifactDescriptor;
124
+ let logArtifact: ArtifactDescriptor | undefined;
125
+ let transcriptArtifact: ArtifactDescriptor | undefined;
126
+ let exitCode: number | null = 0;
127
+ let error: string | undefined;
128
+ let modelAttempts: ModelAttemptSummary[] | undefined;
129
+ let parsedOutput: ParsedPiJsonOutput | undefined;
130
+ let finalStdout = "";
131
+ let transcriptPath: string | undefined;
132
+ let terminalEvidence: OperationTerminalEvidence[] = [];
133
+ const collectedJsonEvents: Record<string, unknown>[] = [];
134
+
135
+ let startupEvidence = createStartupEvidence({ command: runtimeKind === "child-process" ? "pi" : runtimeKind === "live-session" ? "live-session" : "safe-scaffold", startedAt: new Date(task.startedAt ?? new Date().toISOString()), finishedAt: new Date(), promptSentAt: new Date(task.startedAt ?? new Date().toISOString()), promptAccepted: true, exitCode: 0 });
136
+ const inputsArtifact = writeTaskInputsArtifact(manifest, task, dependencyContext);
137
+ const skillArtifact = skillBlock ? writeArtifact(manifest.artifactsRoot, {
138
+ kind: "metadata",
139
+ relativePath: `metadata/${task.id}.skills.md`,
140
+ content: [`Selected skills: ${skillNames?.join(", ") ?? "(none)"}`, `Skill paths passed to child Pi: ${(skillPaths ?? []).length}`, "", skillBlock, ""].join("\n"),
141
+ producer: task.id,
142
+ }) : undefined;
143
+ const coordinationArtifact = writeArtifact(manifest.artifactsRoot, {
144
+ kind: "metadata",
145
+ relativePath: `metadata/${task.id}.coordination-bridge.md`,
146
+ content: `${coordinationBridgeInstructions(task)}\n`,
147
+ producer: task.id,
148
+ });
149
+ if (runtimeKind === "child-process") {
150
+ const modelRoutingPlan = buildConfiguredModelRouting({ overrideModel: input.modelOverride, stepModel: input.step.model, teamRoleModel: input.teamRoleModel, agentModel: input.agent.model, fallbackModels: input.agent.fallbackModels, parentModel: input.parentModel, modelRegistry: input.modelRegistry, cwd: task.cwd });
151
+ const candidates = modelRoutingPlan.candidates;
152
+ const attemptModels = candidates.length > 0 ? candidates : [undefined];
153
+ const logs: string[] = [];
154
+ let finalStderr = "";
155
+ modelAttempts = [];
156
+ transcriptPath = `${manifest.artifactsRoot}/transcripts/${task.id}.jsonl`;
157
+ let finalCheckpointWritten = false;
158
+ let lastAgentRecordPersistedAt = 0;
159
+ let lastHeartbeatPersistedAt = 0;
160
+ let lastRunProgressPersistedAt = 0;
161
+ let lastRunProgressSummary: ProgressEventSummary | undefined;
162
+ const persistHeartbeat = (force = false): void => {
163
+ const now = Date.now();
164
+ if (!force && now - lastHeartbeatPersistedAt < 1000) return;
165
+ lastHeartbeatPersistedAt = now;
166
+ task = { ...task, heartbeat: touchWorkerHeartbeat(task.heartbeat ?? createWorkerHeartbeat(task.id)) };
167
+ tasks = persistSingleTaskUpdate(manifest, tasks, task);
168
+ };
169
+ const persistChildProgress = (event: unknown, force = false): void => {
170
+ const now = Date.now();
171
+ if (force || shouldFlushProgressEvent(event) || now - lastAgentRecordPersistedAt >= 500) {
172
+ upsertCrewAgent(manifest, recordFromTask(manifest, task, "child-process"));
173
+ lastAgentRecordPersistedAt = now;
174
+ }
175
+ const summary = progressEventSummary(task, event);
176
+ const decision = shouldAppendProgressEventUpdate({ previous: lastRunProgressSummary, next: summary, nowMs: now, lastAppendMs: lastRunProgressPersistedAt || undefined, minIntervalMs: 1000, force });
177
+ if (decision.shouldAppend) {
178
+ // 2.2 caller migration: high-frequency task.progress goes through
179
+ // the buffered path; loss-on-kill is acceptable because progress
180
+ // is informational and re-derivable from per-agent records.
181
+ appendEventFireAndForget(manifest.eventsPath, { type: "task.progress", runId: manifest.runId, taskId: task.id, data: { ...summary, coalesceReason: decision.reason } });
182
+ lastRunProgressSummary = summary;
183
+ lastRunProgressPersistedAt = now;
184
+ }
185
+ };
186
+ for (let i = 0; i < attemptModels.length; i++) {
187
+ const model = attemptModels[i];
188
+ const attemptStartedAt = new Date();
189
+ const pendingAttempt: ModelAttemptSummary = { model: model ?? "default", success: false };
190
+ task = { ...task, modelAttempts: [...modelAttempts, pendingAttempt] };
191
+ tasks = updateTask(tasks, task);
192
+ upsertCrewAgent(manifest, recordFromTask(manifest, task, "child-process"));
193
+ const childResult = await runChildPi({
194
+ cwd: task.cwd,
195
+ task: prompt,
196
+ agent: input.agent,
197
+ model,
198
+ signal: input.signal,
199
+ transcriptPath,
200
+ maxDepth: input.limits?.maxTaskDepth,
201
+ skillPaths,
202
+ onSpawn: (pid) => {
203
+ ({ task, tasks } = checkpointTask(manifest, tasks, task, "child-spawned", pid));
204
+ },
205
+ onLifecycleEvent: (event: ChildPiLifecycleEvent) => {
206
+ appendEvent(manifest.eventsPath, { type: `worker.${event.type}` as const, runId: manifest.runId, taskId: task.id, message: `Worker lifecycle: ${event.type}${event.error ? ` error=${event.error}` : ""}${event.exitCode != null ? ` exit=${event.exitCode}` : ""}`, data: { ...event } });
207
+ },
208
+ onStdoutLine: (line) => {
209
+ appendCrewAgentOutput(manifest, task.id, line);
210
+ persistHeartbeat();
211
+ // Check for supervisor contact requests from child Pi
212
+ const contact = parseSupervisorContactFromLine(line);
213
+ if (contact) {
214
+ recordSupervisorContact(manifest, { runId: manifest.runId, ...contact });
215
+ }
216
+ },
217
+ onJsonEvent: (event) => {
218
+ appendCrewAgentEvent(manifest, task.id, event);
219
+ if (event && typeof event === "object" && !Array.isArray(event)) collectedJsonEvents.push(event as Record<string, unknown>);
220
+ persistHeartbeat();
221
+ task = { ...task, agentProgress: applyAgentProgressEvent(task.agentProgress ?? emptyCrewAgentProgress(), event, task.startedAt) };
222
+ tasks = updateTask(tasks, task);
223
+ // Bridge event to UI event bus for near-instant updates
224
+ try {
225
+ const bridgeEvent = bridgeEventFromJsonEvent(manifest.runId, task.id, event);
226
+ if (bridgeEvent) streamBridge?.handler(bridgeEvent);
227
+ } catch { /* bridge errors should not affect task */ }
228
+ // Feed overflow recovery tracker
229
+ if (input.onJsonEvent) {
230
+ try {
231
+ input.onJsonEvent(task.id, manifest.runId, event);
232
+ } catch { /* overflow tracking errors should not affect task */ }
233
+ }
234
+ if (!finalCheckpointWritten && isFinalChildEvent(event)) {
235
+ finalCheckpointWritten = true;
236
+ ({ task, tasks } = checkpointTask(manifest, tasks, task, "child-stdout-final"));
237
+ }
238
+ persistChildProgress(event);
239
+ },
240
+ });
241
+ const evidenceStatus = childResult.exitStatus?.cancelled ? "cancelled" : childResult.error || (childResult.exitCode && childResult.exitCode !== 0) ? "failed" : "completed";
242
+ terminalEvidence = [...terminalEvidence, { operation: "worker", status: evidenceStatus, startedAt: attemptStartedAt.toISOString(), finishedAt: new Date().toISOString(), ...(input.signal?.aborted ? { reason: cancellationReasonFromSignal(input.signal) } : {}), ...(childResult.exitStatus ? { exitStatus: childResult.exitStatus } : {}) }];
243
+ if (evidenceStatus === "cancelled") {
244
+ const cancelReason = input.signal?.aborted ? cancellationReasonFromSignal(input.signal) : { code: "caller_cancelled" as const, message: "Worker cancelled." };
245
+ terminalEvidence.push(buildSyntheticTerminalEvidence("tool", cancelReason, attemptStartedAt.toISOString()));
246
+ appendEvent(manifest.eventsPath, { type: "worker.cancelled", runId: manifest.runId, taskId: task.id, message: cancelReason.message, data: { terminalEvidence: terminalEvidence.at(-1) } });
247
+ }
248
+ startupEvidence = createStartupEvidence({ command: "pi", startedAt: attemptStartedAt, finishedAt: new Date(), promptSentAt: attemptStartedAt, promptAccepted: childResult.exitCode === 0 && !childResult.error, stderr: childResult.stderr, error: childResult.error, exitCode: childResult.exitCode });
249
+ exitCode = childResult.exitCode;
250
+ finalStdout = childResult.stdout;
251
+ finalStderr = childResult.stderr;
252
+ parsedOutput = parsePiJsonOutput(fs.existsSync(transcriptPath) ? fs.readFileSync(transcriptPath, "utf-8") : childResult.stdout);
253
+ error = childResult.error || (childResult.exitCode && childResult.exitCode !== 0 ? childResult.stderr || `Child Pi exited with ${childResult.exitCode}` : undefined);
254
+ persistHeartbeat(true);
255
+ persistChildProgress({ type: "attempt_finished" }, true);
256
+ const attempt: ModelAttemptSummary = { model: model ?? "default", success: !error, exitCode, error };
257
+ modelAttempts.push(attempt);
258
+ task = { ...task, modelAttempts: [...modelAttempts] };
259
+ tasks = updateTask(tasks, task);
260
+ logs.push(`MODEL ATTEMPT ${i + 1}: ${attempt.model}`, `success=${attempt.success}`, `exitCode=${attempt.exitCode ?? "null"}`, attempt.error ? `error=${attempt.error}` : "", "");
261
+ if (!error) break;
262
+ const nextModel = attemptModels[i + 1];
263
+ if (!nextModel || !isRetryableModelFailure(error)) break;
264
+ logs.push(formatModelAttemptNote(attempt, nextModel), "");
265
+ }
266
+ resultArtifact = writeArtifact(manifest.artifactsRoot, {
267
+ kind: "result",
268
+ relativePath: `results/${task.id}.txt`,
269
+ content: cleanResultText(parsedOutput?.finalText) ?? cleanResultText(finalStdout) ?? cleanResultText(finalStderr) ?? "(no output)",
270
+ producer: task.id,
271
+ });
272
+ logArtifact = writeArtifact(manifest.artifactsRoot, {
273
+ kind: "log",
274
+ relativePath: `logs/${task.id}.log`,
275
+ content: [...logs, `finalExitCode=${exitCode ?? "null"}`, `jsonEvents=${parsedOutput?.jsonEvents ?? 0}`, parsedOutput?.usage ? `usage=${JSON.stringify(parsedOutput.usage)}` : "", "", "STDOUT:", finalStdout, "", "STDERR:", finalStderr].join("\n"),
276
+ producer: task.id,
277
+ });
278
+ const successfulAttemptIndex = modelAttempts.findIndex((attempt) => attempt.success);
279
+ const usedAttempt = successfulAttemptIndex === -1 ? Math.max(0, modelAttempts.length - 1) : successfulAttemptIndex;
280
+ const resolvedModel = modelAttempts[usedAttempt]?.model ?? candidates[0] ?? "default";
281
+ const fallbackReason = usedAttempt > 0 ? modelAttempts[usedAttempt - 1]?.error : undefined;
282
+ task = { ...task, modelRouting: { requested: modelRoutingPlan.requested, resolved: resolvedModel, fallbackChain: candidates, reason: fallbackReason ?? modelRoutingPlan.reason, usedAttempt } };
283
+ tasks = updateTask(tasks, task);
284
+ const sessionUsage = parseSessionUsage(transcriptPath);
285
+ const effectiveUsage = parsedOutput?.usage ?? sessionUsage;
286
+ if (effectiveUsage) {
287
+ parsedOutput = { ...(parsedOutput ?? { jsonEvents: 0, textEvents: [] }), usage: effectiveUsage };
288
+ task = { ...task, usage: effectiveUsage, agentProgress: applyUsageToProgress(task.agentProgress, effectiveUsage) };
289
+ tasks = updateTask(tasks, task);
290
+ upsertCrewAgent(manifest, recordFromTask(manifest, task, "child-process"));
291
+ }
292
+ if (fs.existsSync(transcriptPath)) {
293
+ transcriptArtifact = writeArtifact(manifest.artifactsRoot, {
294
+ kind: "log",
295
+ relativePath: `transcripts/${task.id}.jsonl`,
296
+ content: fs.readFileSync(transcriptPath, "utf-8"),
297
+ producer: task.id,
298
+ });
299
+ }
300
+ task = { ...task, resultArtifact, ...(logArtifact ? { logArtifact } : {}), ...(transcriptArtifact ? { transcriptArtifact } : {}) };
301
+ tasks = updateTask(tasks, task);
302
+ ({ task, tasks } = checkpointTask(manifest, tasks, task, "artifact-written"));
303
+ } else if (runtimeKind === "live-session") {
304
+ // LAZY: live-executor is only needed for live-session runtime branches.
305
+ const { runLiveTask } = await import("./task-runner/live-executor.ts");
306
+ const live = await runLiveTask({ manifest, tasks, task, step: input.step, agent: input.agent, prompt, signal: input.signal, runtimeConfig: input.runtimeConfig, parentContext: input.parentContext, parentModel: input.parentModel, modelRegistry: input.modelRegistry, modelOverride: input.modelOverride, teamRoleModel: input.teamRoleModel, workspaceId: input.workspaceId });
307
+ task = live.task;
308
+ tasks = live.tasks;
309
+ startupEvidence = live.startupEvidence;
310
+ exitCode = live.exitCode;
311
+ error = live.error;
312
+ parsedOutput = live.parsedOutput;
313
+ resultArtifact = live.resultArtifact;
314
+ logArtifact = live.logArtifact;
315
+ transcriptArtifact = live.transcriptArtifact;
316
+ } else {
317
+ resultArtifact = writeArtifact(manifest.artifactsRoot, {
318
+ kind: "result",
319
+ relativePath: `results/${task.id}.md`,
320
+ content: [
321
+ `# ${task.id}`,
322
+ "",
323
+ "Worker execution is disabled in this scaffold-safe run.",
324
+ "The prompt artifact contains the exact task that will be sent to a child Pi worker when execution is enabled.",
325
+ ].join("\n"),
326
+ producer: task.id,
327
+ });
328
+ }
329
+
330
+ // --- Yield-based completion contract ---
331
+ let yieldResult: YieldResult | undefined;
332
+ const yieldEnabled = input.runtimeConfig?.yield?.enabled ?? DEFAULT_YIELD_CONFIG.enabled;
333
+ if (yieldEnabled && collectedJsonEvents.length > 0) {
334
+ if (hasYieldInOutput(collectedJsonEvents)) {
335
+ const yieldEvent = collectedJsonEvents.find((e) => isYieldEvent(e));
336
+ if (yieldEvent) {
337
+ yieldResult = extractYieldResult(yieldEvent);
338
+ }
339
+ } else if (!error) {
340
+ appendEvent(manifest.eventsPath, { type: "task.attention", runId: manifest.runId, taskId: task.id, message: "Worker completed without calling submit_result tool.", data: { activityState: "needs_attention", reason: "no_yield" } });
341
+ }
342
+ }
343
+
344
+ const diffArtifact = workspace.worktreePath ? writeArtifact(manifest.artifactsRoot, {
345
+ kind: "diff",
346
+ relativePath: `diffs/${task.id}.diff`,
347
+ content: captureWorktreeDiff(workspace.worktreePath),
348
+ producer: task.id,
349
+ }) : undefined;
350
+ const diffStatArtifact = workspace.worktreePath ? writeArtifact(manifest.artifactsRoot, {
351
+ kind: "metadata",
352
+ relativePath: `metadata/${task.id}.diff-stat.json`,
353
+ content: `${JSON.stringify({ ...captureWorktreeDiffStat(workspace.worktreePath), syntheticPaths: workspace.syntheticPaths ?? [], nodeModulesLinked: workspace.nodeModulesLinked ?? false }, null, 2)}\n`,
354
+ producer: task.id,
355
+ }) : undefined;
356
+
357
+ const mutationGuardMode = input.runtimeConfig?.completionMutationGuard ?? "warn";
358
+ const mutationGuard = !error && mutationGuardMode !== "off" ? evaluateCompletionMutationGuard({ role: task.role, taskText: `${task.title}\n${input.step.task}`, transcriptPath: runtimeKind === "child-process" ? transcriptPath : transcriptArtifact?.path, stdout: finalStdout }) : undefined;
359
+ if (mutationGuard?.reason === "no_mutation_observed") {
360
+ appendTaskAttentionEvent({
361
+ manifest,
362
+ taskId: task.id,
363
+ message: "Implementation-style task completed without an observed mutation tool call.",
364
+ data: { activityState: "needs_attention", reason: "completion_guard", taskId: task.id, agentName: task.agent, observedTools: mutationGuard.observedTools, suggestedAction: mutationGuardMode === "fail" ? "Review the worker output and rerun with a concrete implementation task." : "Review the worker output; set runtime.completionMutationGuard='fail' to enforce this." },
365
+ });
366
+ task = { ...task, agentProgress: { ...(task.agentProgress ?? emptyCrewAgentProgress()), activityState: "needs_attention" } };
367
+ if (mutationGuardMode === "fail") {
368
+ error = "Completion mutation guard failed: implementation-style task completed without an observed mutation tool call.";
369
+ exitCode = exitCode === 0 ? 1 : exitCode;
370
+ if (modelAttempts?.length) {
371
+ modelAttempts = modelAttempts.map((attempt, index) => index === modelAttempts!.length - 1 ? { ...attempt, success: false, exitCode, error } : attempt);
372
+ }
373
+ }
374
+ tasks = updateTask(tasks, task);
375
+ }
376
+
377
+ // --- Output format validation (caveman Phase 4) ---
378
+ // Validate worker output against the role's output contract.
379
+ // On failure: emit attention event but don't fail the task.
380
+ let outputValidation: OutputValidationResult | undefined;
381
+ if (!error) {
382
+ const outputText = parsedOutput?.finalText ?? finalStdout;
383
+ if (outputText) {
384
+ outputValidation = validateWorkerOutput(task.role, outputText);
385
+ if (!outputValidation.valid) {
386
+ appendEvent(manifest.eventsPath, { type: "task.output_validation", runId: manifest.runId, taskId: task.id, data: { valid: false, formatMatch: outputValidation.formatMatch, structurePreserved: outputValidation.structurePreserved, issues: outputValidation.issues } });
387
+ task = { ...task, agentProgress: { ...(task.agentProgress ?? emptyCrewAgentProgress()), activityState: "needs_attention" } };
388
+ tasks = updateTask(tasks, task);
389
+ }
390
+ }
391
+ }
392
+
393
+ task = {
394
+ ...task,
395
+ status: error ? "failed" : "completed",
396
+ finishedAt: new Date().toISOString(),
397
+ exitCode,
398
+ modelAttempts,
399
+ usage: parsedOutput?.usage,
400
+ jsonEvents: parsedOutput?.jsonEvents,
401
+ agentProgress: error && task.agentProgress?.currentTool ? { ...task.agentProgress, failedTool: task.agentProgress.currentTool } : task.agentProgress,
402
+ error,
403
+ verification: createVerificationEvidence(taskPacket.verification, !error, error ? `Task failed: ${error}` : runtimeKind === "scaffold" ? "Safe scaffold mode; verification commands were not executed." : `${runtimeKind} worker finished without reporting a verification failure.`),
404
+ promptArtifact,
405
+ resultArtifact,
406
+ claim: undefined,
407
+ heartbeat: touchWorkerHeartbeat(task.heartbeat ?? createWorkerHeartbeat(task.id), { alive: false }),
408
+ workerExitStatus: terminalEvidence.at(-1)?.exitStatus,
409
+ terminalEvidence: terminalEvidence.length ? [...(task.terminalEvidence ?? []), ...terminalEvidence] : task.terminalEvidence,
410
+ ...(logArtifact ? { logArtifact } : {}),
411
+ ...(transcriptArtifact ? { transcriptArtifact } : {}),
412
+ };
413
+ tasks = updateTask(tasks, task);
414
+ const packetArtifact = writeArtifact(manifest.artifactsRoot, {
415
+ kind: "metadata",
416
+ relativePath: `metadata/${task.id}.task-packet.json`,
417
+ content: `${JSON.stringify(task.taskPacket, null, 2)}\n`,
418
+ producer: task.id,
419
+ });
420
+ const verificationArtifact = writeArtifact(manifest.artifactsRoot, {
421
+ kind: "metadata",
422
+ relativePath: `metadata/${task.id}.verification.json`,
423
+ content: `${JSON.stringify(task.verification, null, 2)}\n`,
424
+ producer: task.id,
425
+ });
426
+ const sharedOutputArtifact = writeTaskSharedOutput(manifest, input.step, task);
427
+ const startupArtifact = writeArtifact(manifest.artifactsRoot, {
428
+ kind: "metadata",
429
+ relativePath: `metadata/${task.id}.startup-evidence.json`,
430
+ content: `${JSON.stringify(startupEvidence, null, 2)}\n`,
431
+ producer: task.id,
432
+ });
433
+ const permissionArtifact = writeArtifact(manifest.artifactsRoot, {
434
+ kind: "metadata",
435
+ relativePath: `metadata/${task.id}.permission.json`,
436
+ content: `${JSON.stringify({ role: task.role, permissionMode }, null, 2)}\n`,
437
+ producer: task.id,
438
+ });
439
+ const capabilityArtifact = writeArtifact(manifest.artifactsRoot, {
440
+ kind: "metadata",
441
+ relativePath: `metadata/${task.id}.capabilities.json`,
442
+ content: `${JSON.stringify(buildWorkerCapabilityInventory({ taskId: task.id, role: task.role, agent: input.agent, runtime: runtimeKind, permissionMode, skillNames, skillPaths, skillsDisabled: input.skillOverride === false || input.teamRoleSkills === false, modelOverride: input.modelOverride, teamRoleModel: input.teamRoleModel, stepModel: input.step.model }), null, 2)}\n`,
443
+ producer: task.id,
444
+ });
445
+ const promptPipelineArtifact = writeArtifact(manifest.artifactsRoot, {
446
+ kind: "metadata",
447
+ relativePath: `metadata/${task.id}.prompt-pipeline.json`,
448
+ content: `${JSON.stringify(buildWorkerPromptPipeline({ artifactsRoot: manifest.artifactsRoot, taskId: task.id, promptArtifact, inputsArtifact, skillArtifact, capabilityArtifact, coordinationArtifact, skillInstructionCount: skillNames?.length ?? 0, skillsDisabled: input.skillOverride === false || input.teamRoleSkills === false }), null, 2)}\n`,
449
+ producer: task.id,
450
+ });
451
+ const outputValidationArtifact = outputValidation ? writeArtifact(manifest.artifactsRoot, {
452
+ kind: "metadata",
453
+ relativePath: `metadata/${task.id}.output-validation.json`,
454
+ content: `${JSON.stringify(outputValidation, null, 2)}\n`,
455
+ producer: task.id,
456
+ }) : undefined;
457
+ manifest = { ...manifest, updatedAt: new Date().toISOString(), artifacts: [...manifest.artifacts, promptArtifact, resultArtifact, inputsArtifact, coordinationArtifact, ...(skillArtifact ? [skillArtifact] : []), packetArtifact, verificationArtifact, startupArtifact, permissionArtifact, capabilityArtifact, promptPipelineArtifact, ...(outputValidationArtifact ? [outputValidationArtifact] : []), ...(sharedOutputArtifact ? [sharedOutputArtifact] : []), ...(logArtifact ? [logArtifact] : []), ...(transcriptArtifact ? [transcriptArtifact] : []), ...(diffArtifact ? [diffArtifact] : []), ...(diffStatArtifact ? [diffStatArtifact] : [])] };
458
+ saveRunManifest(manifest);
459
+ tasks = persistSingleTaskUpdate(manifest, tasks, task);
460
+ upsertCrewAgent(manifest, recordFromTask(manifest, task, runtimeKind));
461
+ // Execute task_result hook before emitting terminal event
462
+ const hookReport = await executeHook("task_result", { runId: manifest.runId, taskId: task.id, cwd: manifest.cwd });
463
+ appendHookEvent(manifest, hookReport);
464
+ appendEvent(manifest.eventsPath, { type: error ? "task.failed" : "task.completed", runId: manifest.runId, taskId: task.id, message: error });
465
+ return { manifest, tasks };
466
+ } finally {
467
+ streamBridge?.dispose();
468
+ }
469
+ }