principles-disciple 1.72.0 → 1.73.0

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 (309) hide show
  1. package/openclaw.plugin.json +10 -5
  2. package/package.json +17 -19
  3. package/scripts/acceptance-test.mjs +16 -73
  4. package/scripts/sync-plugin.mjs +382 -77
  5. package/src/commands/archive-impl.ts +2 -1
  6. package/src/commands/capabilities.ts +2 -2
  7. package/src/commands/context.ts +2 -2
  8. package/src/commands/disable-impl.ts +2 -1
  9. package/src/commands/evolution-status.ts +16 -16
  10. package/src/commands/export.ts +12 -67
  11. package/src/commands/pain.ts +91 -1
  12. package/src/commands/principle-rollback.ts +2 -1
  13. package/src/commands/promote-impl.ts +7 -43
  14. package/src/commands/rollback-impl.ts +2 -1
  15. package/src/commands/rollback.ts +2 -1
  16. package/src/commands/samples.ts +2 -1
  17. package/src/commands/thinking-os.ts +2 -1
  18. package/src/config/errors.ts +18 -2
  19. package/src/constants/diagnostician.ts +2 -2
  20. package/src/constants/tools.ts +2 -1
  21. package/src/core/__tests__/focus-history.test.ts +210 -0
  22. package/src/core/config.ts +1 -1
  23. package/src/core/confirm-first-gate.ts +255 -0
  24. package/src/core/correction-cue-learner.ts +2 -136
  25. package/src/core/correction-types.ts +16 -88
  26. package/src/core/dictionary.ts +19 -20
  27. package/src/core/empathy-keyword-matcher.ts +17 -289
  28. package/src/core/empathy-types.ts +18 -229
  29. package/src/core/event-log.ts +38 -132
  30. package/src/core/evolution-reducer.ts +21 -2
  31. package/src/core/evolution-types.ts +76 -464
  32. package/src/core/file-store.ts +80 -0
  33. package/src/core/focus-history.ts +228 -955
  34. package/src/core/local-worker-routing.ts +34 -314
  35. package/src/core/merge-gate-audit.ts +0 -195
  36. package/src/core/pain-diagnostic-gate.ts +154 -0
  37. package/src/core/pain-signal.ts +21 -138
  38. package/src/core/pain.ts +15 -88
  39. package/src/core/pd-task-reconciler.ts +26 -115
  40. package/src/core/pd-task-service.ts +9 -9
  41. package/src/core/pd-task-types.ts +23 -127
  42. package/src/core/principle-compiler/__tests__/compiler-replay-gate.test.ts +174 -0
  43. package/src/core/principle-compiler/code-validator.ts +15 -42
  44. package/src/core/principle-compiler/compiler.ts +100 -15
  45. package/src/core/principle-compiler/index.ts +5 -2
  46. package/src/core/principle-compiler/template-generator.ts +4 -104
  47. package/src/core/principle-injection.ts +10 -202
  48. package/src/core/principle-internalization/filesystem-lifecycle-datasource.ts +42 -0
  49. package/src/core/principle-internalization/lifecycle-read-model.ts +39 -242
  50. package/src/core/principle-internalization/principle-lifecycle-service.ts +12 -10
  51. package/src/core/principle-tree-ledger-adapter.ts +145 -0
  52. package/src/core/principle-tree-ledger.ts +8 -6
  53. package/src/core/reflection/reflection-context.ts +14 -109
  54. package/src/core/replay-engine.ts +8 -500
  55. package/src/core/rule-host-helpers.ts +5 -35
  56. package/src/core/rule-host-types.ts +10 -82
  57. package/src/core/rule-host.ts +6 -63
  58. package/src/core/runtime-v2-prompt-activation-reader.ts +231 -0
  59. package/src/core/session-tracker.ts +87 -101
  60. package/src/core/shadow-observation-registry.ts +19 -48
  61. package/src/core/trajectory.ts +3 -1
  62. package/src/core/workflow-funnel-loader.ts +62 -68
  63. package/src/core/workspace-context.ts +46 -0
  64. package/src/core/workspace-dir-service.ts +1 -1
  65. package/src/core/workspace-dir-validation.ts +18 -9
  66. package/src/hooks/AGENTS.md +1 -1
  67. package/src/hooks/gate-block-helper.ts +46 -44
  68. package/src/hooks/gate.ts +207 -7
  69. package/src/hooks/lifecycle.ts +30 -32
  70. package/src/hooks/llm.ts +60 -32
  71. package/src/hooks/pain.ts +297 -103
  72. package/src/hooks/prompt.ts +459 -439
  73. package/src/hooks/subagent.ts +2 -29
  74. package/src/i18n/commands.ts +2 -10
  75. package/src/index.ts +95 -85
  76. package/src/openclaw-sdk.ts +311 -0
  77. package/src/service/central-database.ts +8 -4
  78. package/src/service/evolution-queue-migration.ts +2 -1
  79. package/src/service/evolution-worker.ts +163 -1786
  80. package/src/service/internalization-trigger-adapter.ts +302 -0
  81. package/src/service/keyword-optimization-service.ts +4 -4
  82. package/src/service/monitoring-query-service.ts +1 -215
  83. package/src/service/queue-io.ts +60 -331
  84. package/src/service/runtime-summary-service.ts +59 -16
  85. package/src/service/subagent-workflow/index.ts +0 -41
  86. package/src/service/subagent-workflow/types.ts +9 -120
  87. package/src/service/subagent-workflow/workflow-store.ts +2 -119
  88. package/src/service/workflow-watchdog.ts +0 -43
  89. package/src/types/event-payload.ts +16 -74
  90. package/src/types/event-types.ts +39 -547
  91. package/src/types/hygiene-types.ts +7 -30
  92. package/src/types/principle-tree-schema.ts +20 -222
  93. package/src/types/queue.ts +15 -70
  94. package/src/types/runtime-summary.ts +5 -49
  95. package/src/utils/io.ts +10 -0
  96. package/src/utils/retry.ts +1 -1
  97. package/src/utils/shadow-fingerprint.ts +2 -2
  98. package/src/utils/workspace-resolver.ts +50 -0
  99. package/templates/langs/en/core/AGENTS.md +2 -2
  100. package/templates/langs/en/core/BOOT.md +1 -1
  101. package/templates/langs/en/core/HEARTBEAT.md +2 -2
  102. package/templates/langs/en/skills/ai-sprint-orchestration/references/agent-registry.json +1 -72
  103. package/templates/langs/en/skills/ai-sprint-orchestration/references/specs/bugfix-complex-template.json +6 -6
  104. package/templates/langs/en/skills/ai-sprint-orchestration/references/specs/feature-complex-template.json +6 -6
  105. package/templates/langs/en/skills/ai-sprint-orchestration/references/specs/workflow-validation-minimal-verify.json +2 -12
  106. package/templates/langs/en/skills/ai-sprint-orchestration/references/specs/workflow-validation-minimal.json +2 -12
  107. package/templates/langs/en/skills/ai-sprint-orchestration/runtime/.gitignore +2 -2
  108. package/templates/langs/en/skills/ai-sprint-orchestration/scripts/run.mjs +51 -15
  109. package/templates/langs/en/skills/evolve-task/SKILL.md +1 -1
  110. package/templates/langs/en/skills/pd-cli-operator/SKILL.md +67 -0
  111. package/templates/langs/en/skills/pd-diagnostician/SKILL.md +1 -1
  112. package/templates/langs/en/skills/pd-mentor/SKILL.md +1 -1
  113. package/templates/langs/en/skills/pd-pain-signal/SKILL.md +17 -39
  114. package/templates/langs/en/skills/pd-runtime-v2/SKILL.md +61 -0
  115. package/templates/langs/zh/core/AGENTS.md +2 -2
  116. package/templates/langs/zh/core/BOOT.md +1 -1
  117. package/templates/langs/zh/core/HEARTBEAT.md +2 -2
  118. package/templates/langs/zh/skills/ai-sprint-orchestration/references/agent-registry.json +1 -72
  119. package/templates/langs/zh/skills/ai-sprint-orchestration/references/specs/bugfix-complex-template.json +6 -6
  120. package/templates/langs/zh/skills/ai-sprint-orchestration/references/specs/feature-complex-template.json +6 -6
  121. package/templates/langs/zh/skills/ai-sprint-orchestration/references/specs/nocturnal-trinity-quality-enhancement.json +8 -8
  122. package/templates/langs/zh/skills/ai-sprint-orchestration/references/specs/workflow-validation-minimal-verify.json +2 -12
  123. package/templates/langs/zh/skills/ai-sprint-orchestration/references/specs/workflow-validation-minimal.json +2 -12
  124. package/templates/langs/zh/skills/ai-sprint-orchestration/runtime/.gitignore +2 -2
  125. package/templates/langs/zh/skills/ai-sprint-orchestration/scripts/run.mjs +51 -15
  126. package/templates/langs/zh/skills/ai-sprint-orchestration/test/run.test.mjs +21 -5
  127. package/templates/langs/zh/skills/evolve-task/SKILL.md +2 -2
  128. package/templates/langs/zh/skills/pd-cli-operator/SKILL.md +67 -0
  129. package/templates/langs/zh/skills/pd-diagnostician/SKILL.md +1 -1
  130. package/templates/langs/zh/skills/pd-mentor/SKILL.md +1 -1
  131. package/templates/langs/zh/skills/pd-pain-signal/SKILL.md +17 -38
  132. package/templates/langs/zh/skills/pd-runtime-v2/SKILL.md +61 -0
  133. package/tests/build-artifacts.test.ts +1 -3
  134. package/tests/commands/evolution-status.test.ts +0 -118
  135. package/tests/core/bootstrap-rules.test.ts +1 -1
  136. package/tests/core/config.test.ts +1 -1
  137. package/tests/core/event-log.test.ts +35 -0
  138. package/tests/core/evolution-engine.test.ts +610 -0
  139. package/tests/core/file-store.test.ts +102 -0
  140. package/tests/core/focus-history.test.ts +203 -11
  141. package/tests/core/merge-gate-audit.test.ts +2 -169
  142. package/tests/core/model-deployment-registry.test.ts +7 -1
  143. package/tests/core/model-training-registry.test.ts +19 -0
  144. package/tests/core/observability.test.ts +0 -1
  145. package/tests/core/pain-diagnostic-gate.test.ts +498 -0
  146. package/tests/core/pain.test.ts +0 -1
  147. package/tests/core/principle-internalization/deprecated-readiness.test.ts +2 -2
  148. package/tests/core/principle-internalization/lifecycle-metrics.test.ts +2 -2
  149. package/tests/core/principle-internalization/{internalization-routing-policy.test.ts → lifecycle-routing-policy.test.ts} +6 -6
  150. package/tests/core/principle-internalization/lineage-source-retired.test.ts +56 -0
  151. package/tests/core/principle-internalization/principle-lifecycle-service.test.ts +1 -23
  152. package/tests/core/principle-tree-ledger-adapter.test.ts +253 -0
  153. package/tests/core/reflection-context.test.ts +0 -14
  154. package/tests/core/replay-engine.test.ts +127 -215
  155. package/tests/core/rule-host-helpers.test.ts +2 -2
  156. package/tests/core/rule-implementation-runtime.test.ts +0 -27
  157. package/tests/core/workflow-funnel-loader.test.ts +162 -0
  158. package/tests/core/workspace-dir-validation.test.ts +8 -1
  159. package/tests/core-anti-growth.test.ts +192 -0
  160. package/tests/hook-workspace-nextaction-contract.test.ts +42 -0
  161. package/tests/hooks/confirm-first-gate.test.ts +333 -0
  162. package/tests/hooks/gate-auto-correct-shadow.test.ts +310 -0
  163. package/tests/hooks/gate-auto-correct.test.ts +665 -0
  164. package/tests/hooks/gate-rule-host-pipeline.test.ts +2 -1
  165. package/tests/hooks/pain.test.ts +269 -12
  166. package/tests/hooks/prompt-characterization.test.ts +500 -0
  167. package/tests/hooks/prompt-size-guard.test.ts +32 -17
  168. package/tests/hooks/runtime-v2-prompt-activation.test.ts +869 -0
  169. package/tests/index.test.ts +94 -1
  170. package/tests/integration/auto-entry-gate.test.ts +248 -0
  171. package/tests/integration/internalization-trigger-guard.test.ts +69 -0
  172. package/tests/integration/m8-legacy-paths.test.ts +63 -0
  173. package/tests/integration/runtime-v2-pain-guard.test.ts +125 -0
  174. package/tests/plugin-config-resolution-cutover.test.ts +359 -0
  175. package/tests/runtime-v2-discovery-guard.test.ts +154 -0
  176. package/tests/service/central-database.test.ts +457 -0
  177. package/tests/service/evolution-worker.correction-observer.test.ts +173 -0
  178. package/tests/service/evolution-worker.timeout.test.ts +11 -129
  179. package/tests/service/internalization-trigger-adapter.test.ts +251 -0
  180. package/tests/service/monitoring-query-service.test.ts +1 -47
  181. package/tests/service/queue-io.test.ts +1 -62
  182. package/tests/service/runtime-summary-service.test.ts +3 -1
  183. package/tests/service/workflow-watchdog.test.ts +0 -91
  184. package/tests/utils/file-lock.test.ts +5 -3
  185. package/tests/utils/session-key.test.ts +52 -0
  186. package/tests/utils/subagent-probe.test.ts +48 -1
  187. package/vitest.config.ts +4 -11
  188. package/.planning/codebase/ARCHITECTURE.md +0 -157
  189. package/.planning/codebase/CONCERNS.md +0 -145
  190. package/.planning/codebase/CONVENTIONS.md +0 -148
  191. package/.planning/codebase/INTEGRATIONS.md +0 -81
  192. package/.planning/codebase/STACK.md +0 -87
  193. package/.planning/codebase/STRUCTURE.md +0 -193
  194. package/.planning/codebase/TESTING.md +0 -243
  195. package/.planning/phases/01-basic-visualization/01-GAP-CLOSURE-VERIFICATION.md +0 -113
  196. package/docs/COMMAND_REFERENCE.md +0 -76
  197. package/docs/COMMAND_REFERENCE_EN.md +0 -79
  198. package/scripts/build-web.mjs +0 -46
  199. package/scripts/diagnose-nocturnal.mjs +0 -537
  200. package/scripts/seed-nocturnal-scenarios.mjs +0 -384
  201. package/src/commands/nocturnal-review.ts +0 -322
  202. package/src/commands/nocturnal-rollout.ts +0 -790
  203. package/src/commands/nocturnal-train.ts +0 -986
  204. package/src/commands/pd-reflect.ts +0 -88
  205. package/src/core/adaptive-thresholds.ts +0 -478
  206. package/src/core/diagnostician-task-store.ts +0 -192
  207. package/src/core/nocturnal-arbiter.ts +0 -715
  208. package/src/core/nocturnal-artifact-lineage.ts +0 -116
  209. package/src/core/nocturnal-artificer.ts +0 -257
  210. package/src/core/nocturnal-candidate-scoring.ts +0 -530
  211. package/src/core/nocturnal-compliance.ts +0 -1146
  212. package/src/core/nocturnal-dataset.ts +0 -763
  213. package/src/core/nocturnal-executability.ts +0 -428
  214. package/src/core/nocturnal-export.ts +0 -499
  215. package/src/core/nocturnal-paths.ts +0 -240
  216. package/src/core/nocturnal-reasoning-deriver.ts +0 -343
  217. package/src/core/nocturnal-rule-implementation-validator.ts +0 -246
  218. package/src/core/nocturnal-snapshot-contract.ts +0 -99
  219. package/src/core/nocturnal-trajectory-extractor.ts +0 -512
  220. package/src/core/nocturnal-trinity-types.ts +0 -218
  221. package/src/core/nocturnal-trinity.ts +0 -2680
  222. package/src/core/principle-internalization/deprecated-readiness.ts +0 -93
  223. package/src/core/principle-internalization/internalization-routing-policy.ts +0 -208
  224. package/src/core/principle-internalization/lifecycle-metrics.ts +0 -152
  225. package/src/http/principles-console-route.ts +0 -709
  226. package/src/service/central-health-service.ts +0 -49
  227. package/src/service/central-overview-service.ts +0 -138
  228. package/src/service/control-ui-query-service.ts +0 -900
  229. package/src/service/cooldown-strategy.ts +0 -97
  230. package/src/service/evolution-pain-context.ts +0 -79
  231. package/src/service/evolution-query-service.ts +0 -407
  232. package/src/service/health-query-service.ts +0 -1038
  233. package/src/service/nocturnal-config.ts +0 -214
  234. package/src/service/nocturnal-runtime.ts +0 -734
  235. package/src/service/nocturnal-service.ts +0 -1605
  236. package/src/service/nocturnal-target-selector.ts +0 -545
  237. package/src/service/sleep-cycle.ts +0 -157
  238. package/src/service/startup-reconciler.ts +0 -112
  239. package/src/service/subagent-workflow/correction-observer-types.ts +0 -82
  240. package/src/service/subagent-workflow/correction-observer-workflow-manager.ts +0 -250
  241. package/src/service/subagent-workflow/deep-reflect-workflow-manager.ts +0 -1
  242. package/src/service/subagent-workflow/dynamic-timeout.ts +0 -30
  243. package/src/service/subagent-workflow/empathy-observer-workflow-manager.ts +0 -268
  244. package/src/service/subagent-workflow/nocturnal-workflow-manager.ts +0 -795
  245. package/src/service/subagent-workflow/runtime-direct-driver.ts +0 -268
  246. package/src/service/subagent-workflow/workflow-manager-base.ts +0 -580
  247. package/src/tools/write-pain-flag.ts +0 -215
  248. package/tests/commands/nocturnal-review.test.ts +0 -448
  249. package/tests/commands/nocturnal-train.test.ts +0 -97
  250. package/tests/commands/pd-reflect.test.ts +0 -49
  251. package/tests/core/adaptive-thresholds.test.ts +0 -261
  252. package/tests/core/nocturnal-arbiter.test.ts +0 -559
  253. package/tests/core/nocturnal-artifact-lineage.test.ts +0 -53
  254. package/tests/core/nocturnal-artificer.test.ts +0 -241
  255. package/tests/core/nocturnal-candidate-scoring.test.ts +0 -532
  256. package/tests/core/nocturnal-compliance-p-principles.test.ts +0 -133
  257. package/tests/core/nocturnal-compliance.test.ts +0 -646
  258. package/tests/core/nocturnal-dataset.test.ts +0 -892
  259. package/tests/core/nocturnal-e2e.test.ts +0 -234
  260. package/tests/core/nocturnal-executability.test.ts +0 -357
  261. package/tests/core/nocturnal-export.test.ts +0 -517
  262. package/tests/core/nocturnal-reasoning-deriver.test.ts +0 -372
  263. package/tests/core/nocturnal-reviewed-subset-comparison.test.ts +0 -428
  264. package/tests/core/nocturnal-rule-implementation-validator.test.ts +0 -127
  265. package/tests/core/nocturnal-snapshot-contract.test.ts +0 -121
  266. package/tests/core/nocturnal-trajectory-extractor.test.ts +0 -634
  267. package/tests/core/nocturnal-trinity.test.ts +0 -2053
  268. package/tests/core/pain-auto-repair.test.ts +0 -96
  269. package/tests/core/pain-integration.test.ts +0 -510
  270. package/tests/fixtures/nocturnal-reviewed-subset.json +0 -183
  271. package/tests/http/principles-console-route.test.ts +0 -162
  272. package/tests/integration/chaos-resilience.test.ts +0 -348
  273. package/tests/integration/empathy-workflow-integration.test.ts +0 -626
  274. package/tests/integration/pain-diagnostician-loop.e2e.test.ts +0 -380
  275. package/tests/service/control-ui-query-service.test.ts +0 -121
  276. package/tests/service/cooldown-strategy.test.ts +0 -164
  277. package/tests/service/data-endpoints-regression.test.ts +0 -834
  278. package/tests/service/empathy-observer-workflow-manager.test.ts +0 -175
  279. package/tests/service/evolution-worker.nocturnal.test.ts +0 -601
  280. package/tests/service/nocturnal-runtime-hardening.test.ts +0 -118
  281. package/tests/service/nocturnal-runtime.test.ts +0 -473
  282. package/tests/service/nocturnal-service-code-candidate.test.ts +0 -330
  283. package/tests/service/nocturnal-target-selector.test.ts +0 -615
  284. package/tests/service/startup-reconciler.test.ts +0 -148
  285. package/tests/tools/write-pain-flag.test.ts +0 -358
  286. package/ui/src/App.tsx +0 -45
  287. package/ui/src/api.ts +0 -220
  288. package/ui/src/charts.tsx +0 -955
  289. package/ui/src/components/ErrorState.tsx +0 -6
  290. package/ui/src/components/Loading.tsx +0 -13
  291. package/ui/src/components/ProtectedRoute.tsx +0 -12
  292. package/ui/src/components/Shell.tsx +0 -91
  293. package/ui/src/components/WorkspaceConfig.tsx +0 -178
  294. package/ui/src/components/index.ts +0 -5
  295. package/ui/src/context/auth.tsx +0 -80
  296. package/ui/src/context/theme.tsx +0 -66
  297. package/ui/src/hooks/useAutoRefresh.ts +0 -39
  298. package/ui/src/i18n/ui.ts +0 -473
  299. package/ui/src/main.tsx +0 -16
  300. package/ui/src/pages/EvolutionPage.tsx +0 -333
  301. package/ui/src/pages/FeedbackPage.tsx +0 -138
  302. package/ui/src/pages/GateMonitorPage.tsx +0 -136
  303. package/ui/src/pages/LoginPage.tsx +0 -89
  304. package/ui/src/pages/OverviewPage.tsx +0 -599
  305. package/ui/src/pages/SamplesPage.tsx +0 -174
  306. package/ui/src/pages/ThinkingModelsPage.tsx +0 -702
  307. package/ui/src/styles.css +0 -2020
  308. package/ui/src/types.ts +0 -384
  309. package/ui/src/utils/format.ts +0 -15
@@ -1,499 +0,0 @@
1
- /**
2
- * Nocturnal ORPO Export — Approved Dataset to Decision-Point JSONL
3
- * =================================================================
4
- *
5
- * PURPOSE: Export approved nocturnal samples as ORPO-formatted decision-point
6
- * training JSONL, strictly separated from legacy correction export.
7
- *
8
- * ARCHITECTURE:
9
- * - Export output: .state/exports/orpo/{exportId}.jsonl
10
- * - Export manifest: .state/exports/orpo/{exportId}-manifest.json
11
- * - Legacy corrections: untouched, separate path
12
- *
13
- * ORPO FORMAT (each line):
14
- * {
15
- * sampleFingerprint: string,
16
- * artifactId: string,
17
- * sessionId: string,
18
- * principleId: string,
19
- * targetModelFamily: string,
20
- * prompt: string, // badDecision (the wrong choice)
21
- * chosen: string, // betterDecision (the right choice)
22
- * rejected: string, // badDecision (for ORPO)
23
- * rationale: string,
24
- * datasetMetadata: {
25
- * sampleFingerprint: string,
26
- * artifactPath: string,
27
- * createdAt: string,
28
- * exportedAt: string,
29
- * exportId: string,
30
- * datasetFingerprint: string
31
- * }
32
- * }
33
- *
34
- * EXPORT GATING (fail-closed):
35
- * - reviewStatus === 'approved_for_training'
36
- * - targetModelFamily matches requested target (or any if not specified)
37
- * - Lineage fields complete (sampleFingerprint, artifactId, sessionId, principleId)
38
- * - Source artifact file exists and is approved
39
- *
40
- * DESIGN CONSTRAINTS:
41
- * - No trainer invocation
42
- * - No automatic training
43
- * - No checkpoint deploy
44
- * - Export is read-only from dataset perspective
45
- */
46
-
47
- import * as fs from 'fs';
48
- import * as path from 'path';
49
- import * as crypto from 'crypto';
50
- import { atomicWriteFileSync } from '../utils/io.js';
51
- import {
52
- listDatasetRecords,
53
- readDatasetArtifact,
54
- type NocturnalDatasetRecord,
55
- } from './nocturnal-dataset.js';
56
- import {
57
- listArtifactLineageRecords,
58
- type ArtifactLineageRecord,
59
- } from './nocturnal-artifact-lineage.js';
60
- import { NocturnalPathResolver } from './nocturnal-paths.js';
61
-
62
- // ---------------------------------------------------------------------------
63
- // Types
64
- // ---------------------------------------------------------------------------
65
-
66
- /**
67
- * A single ORPO training sample in JSONL format.
68
- */
69
- export interface ORPOSample {
70
- sampleFingerprint: string;
71
- artifactId: string;
72
- sessionId: string;
73
- principleId: string;
74
- targetModelFamily: string;
75
- /** The suboptimal decision (what the agent did wrong) */
76
- prompt: string;
77
- /** The correct decision (what should have been done) */
78
- chosen: string;
79
- /** The suboptimal decision (same as prompt, for ORPO structure) */
80
- rejected: string;
81
- rationale: string;
82
- datasetMetadata: {
83
- sampleFingerprint: string;
84
- artifactPath: string;
85
- createdAt: string;
86
- exportedAt: string;
87
- exportId: string;
88
- datasetFingerprint: string;
89
- evidenceSummary: ORPOEvidenceSummary;
90
- };
91
- }
92
-
93
- export type EvidenceState = 'observed' | 'not_observed' | 'unknown';
94
-
95
- export interface ORPOEvidenceSummary {
96
- lineageStatus: 'observed' | 'unknown';
97
- painSignals: {
98
- status: EvidenceState;
99
- count: number | null;
100
- ids: string[];
101
- };
102
- gateBlocks: {
103
- status: EvidenceState;
104
- count: number | null;
105
- ids: string[];
106
- };
107
- }
108
-
109
- /**
110
- * Export manifest containing metadata about the entire export.
111
- */
112
- export interface ORPOExportManifest {
113
- exportId: string;
114
- createdAt: string;
115
- sampleCount: number;
116
- targetModelFamily: string;
117
- /** SHA-256 of all sample fingerprints, sorted — for reproducibility */
118
- datasetFingerprint: string;
119
- exportPath: string;
120
- manifestPath: string;
121
- samples: {
122
- sampleFingerprint: string;
123
- artifactId: string;
124
- sessionId: string;
125
- principleId: string;
126
- }[];
127
- }
128
-
129
- /**
130
- * Result of an export operation.
131
- */
132
- export interface ExportResult {
133
- success: boolean;
134
- manifest?: ORPOExportManifest;
135
- error?: string;
136
- emptyReason?: 'no_approved_samples' | 'family_mismatch' | 'all_samples_missing_artifacts';
137
- }
138
-
139
- // ---------------------------------------------------------------------------
140
- // Dataset Fingerprint (for reproducibility)
141
- // ---------------------------------------------------------------------------
142
-
143
- /**
144
- * Compute a deterministic dataset fingerprint from a sorted list of sample fingerprints.
145
- * This allows reproducible exports — same dataset always produces same fingerprint.
146
- */
147
- function computeDatasetFingerprint(sampleFingerprints: string[]): string {
148
- const sorted = [...sampleFingerprints].sort();
149
- const combined = sorted.join('|');
150
- return crypto.createHash('sha256').update(combined, 'utf8').digest('hex');
151
- }
152
-
153
- // ---------------------------------------------------------------------------
154
- // Individual Sample Serialization
155
- // ---------------------------------------------------------------------------
156
-
157
- /**
158
- * Serialize a single dataset record + artifact to ORPO JSONL line.
159
- * Caller guarantees record.targetModelFamily is non-null.
160
- */
161
-
162
-
163
- function serializeORPOSample(
164
- record: NocturnalDatasetRecord,
165
- artifact: ReturnType<typeof readDatasetArtifact>,
166
- evidenceSummary: ORPOEvidenceSummary,
167
- exportId: string,
168
- datasetFingerprint: string
169
- ): ORPOSample {
170
- const now = new Date().toISOString();
171
-
172
- const rejected = buildEvidenceBoundedRejected(artifact, evidenceSummary);
173
-
174
- return {
175
- sampleFingerprint: record.sampleFingerprint,
176
- artifactId: record.artifactId,
177
- sessionId: record.sessionId,
178
- principleId: record.principleId,
179
- targetModelFamily: record.targetModelFamily as string, // validated non-null by caller
180
- // Export only evidence-bounded narratives. Free-form artifact text can overstate what was observed.
181
- prompt: rejected,
182
- chosen: artifact.betterDecision,
183
- rejected,
184
-
185
- rationale: buildEvidenceBoundedRationale(evidenceSummary),
186
- datasetMetadata: {
187
- sampleFingerprint: record.sampleFingerprint,
188
- artifactPath: record.artifactPath,
189
- createdAt: record.createdAt,
190
- exportedAt: now,
191
- exportId,
192
- datasetFingerprint,
193
- evidenceSummary,
194
- },
195
- };
196
- }
197
-
198
- function buildEvidenceSummary(
199
- lineageRecord: ArtifactLineageRecord | null
200
- ): ORPOEvidenceSummary {
201
- if (!lineageRecord) {
202
- return {
203
- lineageStatus: 'unknown',
204
- painSignals: { status: 'unknown', count: null, ids: [] },
205
- gateBlocks: { status: 'unknown', count: null, ids: [] },
206
- };
207
- }
208
-
209
- // Defensive: old lineage files on disk may lack these fields
210
- const painIds = lineageRecord.sourcePainIds ?? [];
211
- const gateBlockIds = lineageRecord.sourceGateBlockIds ?? [];
212
- const painCount = painIds.length;
213
- const gateCount = gateBlockIds.length;
214
-
215
- return {
216
- lineageStatus: 'observed',
217
- painSignals: {
218
- status: painCount > 0 ? 'observed' : 'not_observed',
219
- count: painCount,
220
- ids: [...painIds],
221
- },
222
- gateBlocks: {
223
- status: gateCount > 0 ? 'observed' : 'not_observed',
224
- count: gateCount,
225
- ids: [...gateBlockIds],
226
- },
227
- };
228
- }
229
-
230
- function buildEvidenceBoundedRejected(
231
- artifact: ReturnType<typeof readDatasetArtifact>,
232
- evidenceSummary: ORPOEvidenceSummary
233
- ): string {
234
- if (evidenceSummary.lineageStatus === 'unknown') {
235
- return 'Take the next action without verified source evidence.';
236
- }
237
-
238
- const clauses: string[] = [];
239
- if (evidenceSummary.painSignals.status === 'observed' && evidenceSummary.painSignals.count) {
240
- clauses.push(`continue despite ${evidenceSummary.painSignals.count} observed pain signals`);
241
- }
242
- if (evidenceSummary.gateBlocks.status === 'observed' && evidenceSummary.gateBlocks.count) {
243
- clauses.push(`ignore ${evidenceSummary.gateBlocks.count} observed gate blocks`);
244
- }
245
-
246
- if (clauses.length === 0) {
247
- return 'Proceed without first verifying the relevant state from the source session.';
248
- }
249
-
250
- const prefix = artifact.badDecision.trim().length > 0
251
- ? 'Proceed with the rejected action and '
252
- : 'Take the rejected action and ';
253
- return `${prefix}${clauses.join(' and ')}.`;
254
- }
255
-
256
- function buildEvidenceBoundedRationale(evidenceSummary: ORPOEvidenceSummary): string {
257
- if (evidenceSummary.lineageStatus === 'unknown') {
258
- return 'Source evidence is unknown. Export uses a neutral rationale instead of narrating unverified failures or violations.';
259
- }
260
-
261
- const painCount = evidenceSummary.painSignals.count ?? 0;
262
- const gateCount = evidenceSummary.gateBlocks.count ?? 0;
263
- if (painCount === 0 && gateCount === 0) {
264
- return 'Source lineage is present but records no pain signals or gate blocks. Export keeps the corrective preference while avoiding invented failure narratives.';
265
- }
266
-
267
- return `Observed source evidence: ${painCount} pain signals and ${gateCount} gate blocks. Prefer the bounded corrective action over repeating the rejected choice.`;
268
- }
269
-
270
- // ---------------------------------------------------------------------------
271
- // Core Export Function
272
- // ---------------------------------------------------------------------------
273
-
274
- /**
275
- * Export approved nocturnal samples as ORPO decision-point JSONL.
276
- *
277
- * @param workspaceDir - Workspace directory
278
- * @param targetModelFamily - Specific model family to export, or undefined for all
279
- * @param options - Additional export options
280
- * @returns ExportResult
281
- */
282
- export function exportORPOSamples(
283
- workspaceDir: string,
284
- targetModelFamily?: string | null,
285
- _options: Record<string, never> = {}
286
- ): ExportResult {
287
- const exportId = crypto.randomUUID();
288
- const now = new Date().toISOString();
289
- const lineageRecords = listArtifactLineageRecords(workspaceDir, 'behavioral-sample');
290
-
291
- // Step 1: Collect eligible records
292
- // Use listDatasetRecords directly to have full control over the family filter
293
- // (listExportReadyRecords uses ?? which maps null→undefined, losing the null distinction)
294
- const allApprovedRecords = listDatasetRecords(workspaceDir, {
295
- reviewStatus: 'approved_for_training',
296
- });
297
-
298
-
299
-
300
- let eligibleRecords: typeof allApprovedRecords;
301
-
302
- if (targetModelFamily !== undefined && targetModelFamily !== null) {
303
- // Specific family: check if ANY records (regardless of status) have this family
304
- const allRecords = listDatasetRecords(workspaceDir);
305
- const hasAnyWithFamily = allRecords.some((r) => r.targetModelFamily === targetModelFamily);
306
- if (!hasAnyWithFamily) {
307
- // Family doesn't exist in any record
308
- return {
309
- success: false,
310
- error: 'No samples found for the requested target model family',
311
- emptyReason: 'family_mismatch',
312
- };
313
- }
314
- // Family exists but none are approved
315
- eligibleRecords = allApprovedRecords.filter((r) => r.targetModelFamily === targetModelFamily);
316
- } else {
317
- // All families
318
- eligibleRecords = allApprovedRecords;
319
- }
320
-
321
- // Step 2: Validate we have records
322
- if (eligibleRecords.length === 0) {
323
- return {
324
- success: false,
325
- error: 'No approved samples found for export',
326
- emptyReason: 'no_approved_samples',
327
- };
328
- }
329
-
330
- // Step 3: Verify lineage completeness and read artifacts
331
- const orpoSamples: ORPOSample[] = [];
332
- const failedFingerprints: string[] = [];
333
-
334
- for (const record of eligibleRecords) {
335
- // Enforce targetModelFamily binding — samples without a family cannot enter training
336
- if (record.targetModelFamily === null) {
337
- failedFingerprints.push(record.sampleFingerprint);
338
- continue;
339
- }
340
-
341
- // Verify lineage completeness
342
- if (!record.sampleFingerprint || !record.artifactId || !record.sessionId || !record.principleId) {
343
- failedFingerprints.push(record.sampleFingerprint);
344
- continue;
345
- }
346
-
347
- // Read artifact (throws on error — distinguishes read failure from missing artifact)
348
-
349
-
350
- let artifact;
351
- try {
352
- artifact = readDatasetArtifact(workspaceDir, record.sampleFingerprint);
353
- } catch {
354
- failedFingerprints.push(record.sampleFingerprint);
355
- continue;
356
- }
357
-
358
- const lineageRecord =
359
- lineageRecords.find((candidate) => candidate.artifactId === record.artifactId) ?? null;
360
- const evidenceSummary = buildEvidenceSummary(lineageRecord);
361
-
362
- // Serialize
363
- orpoSamples.push(serializeORPOSample(record, artifact, evidenceSummary, exportId, ''));
364
- }
365
-
366
- // Step 4: Fail if all samples failed validation
367
- if (orpoSamples.length === 0) {
368
- return {
369
- success: false,
370
- error: `All ${eligibleRecords.length} eligible samples failed validation (missing artifacts or lineage)`,
371
- emptyReason: 'all_samples_missing_artifacts',
372
- };
373
- }
374
-
375
- // Step 5: Compute dataset fingerprint for manifest
376
- const datasetFingerprint = computeDatasetFingerprint(
377
- orpoSamples.map((s) => s.sampleFingerprint)
378
- );
379
-
380
- // Step 6: Fill in dataset fingerprint in all samples
381
- for (const sample of orpoSamples) {
382
- sample.datasetMetadata.datasetFingerprint = datasetFingerprint;
383
- }
384
-
385
- // Step 7: Write JSONL file
386
- const exportsDir = NocturnalPathResolver.exportsDir(workspaceDir);
387
- const jsonlPath = path.join(exportsDir, `${exportId}.jsonl`);
388
- const lines = orpoSamples.map((s) => JSON.stringify(s)).join('\n') + '\n';
389
- atomicWriteFileSync(jsonlPath, lines);
390
-
391
- // Step 8: Write manifest
392
- const manifest: ORPOExportManifest = {
393
- exportId,
394
- createdAt: now,
395
- sampleCount: orpoSamples.length,
396
- targetModelFamily: targetModelFamily ?? 'all',
397
- datasetFingerprint,
398
- exportPath: jsonlPath,
399
- manifestPath: path.join(exportsDir, `${exportId}-manifest.json`),
400
- samples: orpoSamples.map((s) => ({
401
- sampleFingerprint: s.sampleFingerprint,
402
- artifactId: s.artifactId,
403
- sessionId: s.sessionId,
404
- principleId: s.principleId,
405
- })),
406
- };
407
-
408
- atomicWriteFileSync(manifest.manifestPath, JSON.stringify(manifest, null, 2));
409
-
410
- return {
411
- success: true,
412
- manifest,
413
- };
414
- }
415
-
416
- /**
417
- * Verify an existing export by re-computing its dataset fingerprint.
418
- * Returns true if the export is intact and reproducible.
419
- */
420
- export function verifyExportIntegrity(
421
- workspaceDir: string,
422
- exportId: string
423
- ): { valid: boolean; computedFingerprint: string; manifestFingerprint: string } | null {
424
- const exportsDir = NocturnalPathResolver.exportsDir(workspaceDir);
425
- const manifestPath = path.join(exportsDir, `${exportId}-manifest.json`);
426
-
427
- if (!fs.existsSync(manifestPath)) {
428
- return null;
429
- }
430
-
431
- try {
432
- const manifest = JSON.parse(fs.readFileSync(manifestPath, 'utf-8')) as ORPOExportManifest;
433
- const computedFingerprint = computeDatasetFingerprint(
434
- manifest.samples.map((s) => s.sampleFingerprint)
435
- );
436
-
437
- return {
438
- valid: computedFingerprint === manifest.datasetFingerprint,
439
- computedFingerprint,
440
- manifestFingerprint: manifest.datasetFingerprint,
441
- };
442
- } catch {
443
- return null;
444
- }
445
- }
446
-
447
- /**
448
- * List all exports in the exports directory.
449
- */
450
- export function listExports(workspaceDir: string): ORPOExportManifest[] {
451
- const exportsDir = NocturnalPathResolver.exportsDir(workspaceDir);
452
- if (!fs.existsSync(exportsDir)) {
453
- return [];
454
- }
455
-
456
- try {
457
- const files = fs.readdirSync(exportsDir);
458
- const manifests: ORPOExportManifest[] = [];
459
-
460
- for (const file of files) {
461
- if (!file.endsWith('-manifest.json')) continue;
462
- try {
463
- const manifest = JSON.parse(
464
- fs.readFileSync(path.join(exportsDir, file), 'utf-8')
465
- ) as ORPOExportManifest;
466
- manifests.push(manifest);
467
- } catch {
468
- // Skip malformed manifest
469
- }
470
- }
471
-
472
- return manifests.sort(
473
- (a, b) => new Date(b.createdAt).getTime() - new Date(a.createdAt).getTime()
474
- );
475
- } catch {
476
- return [];
477
- }
478
- }
479
-
480
- /**
481
- * Read an export manifest by ID.
482
- */
483
- export function getExportManifest(
484
- workspaceDir: string,
485
- exportId: string
486
- ): ORPOExportManifest | null {
487
- const exportsDir = NocturnalPathResolver.exportsDir(workspaceDir);
488
- const manifestPath = path.join(exportsDir, `${exportId}-manifest.json`);
489
-
490
- if (!fs.existsSync(manifestPath)) {
491
- return null;
492
- }
493
-
494
- try {
495
- return JSON.parse(fs.readFileSync(manifestPath, 'utf-8')) as ORPOExportManifest;
496
- } catch {
497
- return null;
498
- }
499
- }