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
@@ -1,634 +0,0 @@
1
- import { describe, it, expect, beforeEach, afterEach } from 'vitest';
2
- import * as fs from 'fs';
3
- import * as path from 'path';
4
- import * as os from 'os';
5
- import {
6
- NocturnalTrajectoryExtractor,
7
- listNocturnalCandidateSessions,
8
- getNocturnalSessionSnapshot,
9
- computeThinkingModelActivation,
10
- computePlanningRatio,
11
- computeThinkingModelDelta,
12
- computePlanningRatioGain,
13
- type NocturnalSessionSnapshot,
14
- type NocturnalSessionSummary,
15
- } from '../../src/core/nocturnal-trajectory-extractor.js';
16
- import { TrajectoryDatabase } from '../../src/core/trajectory.js';
17
-
18
- describe('NocturnalTrajectoryExtractor', () => {
19
- let tmpDir: string;
20
- let trajectory: TrajectoryDatabase;
21
- let extractor: NocturnalTrajectoryExtractor;
22
-
23
- beforeEach(() => {
24
- tmpDir = fs.mkdtempSync(path.join(os.tmpdir(), 'pd-nocturnal-extractor-test-'));
25
- trajectory = new TrajectoryDatabase({ workspaceDir: tmpDir });
26
- extractor = new NocturnalTrajectoryExtractor(trajectory);
27
- });
28
-
29
- afterEach(() => {
30
- trajectory.dispose();
31
- fs.rmSync(tmpDir, { recursive: true, force: true });
32
- });
33
-
34
- // -------------------------------------------------------------------------
35
- // Helper: Seed trajectory data
36
- // -------------------------------------------------------------------------
37
-
38
- function seedSession(sessionId: string, startedAt?: string): void {
39
- trajectory.recordSession({ sessionId, startedAt: startedAt ?? new Date().toISOString() });
40
- }
41
-
42
- function seedAssistantTurn(
43
- sessionId: string,
44
- sanitizedText: string,
45
- rawText: string,
46
- model = 'gpt-4'
47
- ): void {
48
- trajectory.recordAssistantTurn({
49
- sessionId,
50
- runId: 'run-1',
51
- provider: 'openai',
52
- model,
53
- rawText,
54
- sanitizedText,
55
- usageJson: {},
56
- empathySignalJson: {},
57
- });
58
- }
59
-
60
- function seedToolCall(
61
- sessionId: string,
62
- toolName: string,
63
- outcome: 'success' | 'failure' | 'blocked',
64
- errorMessage?: string
65
- ): void {
66
- trajectory.recordToolCall({
67
- sessionId,
68
- toolName,
69
- outcome,
70
- errorMessage: errorMessage ?? null,
71
- });
72
- }
73
-
74
- function seedPainEvent(
75
- sessionId: string,
76
- score: number,
77
- source: string,
78
- reason?: string
79
- ): void {
80
- trajectory.recordPainEvent({
81
- sessionId,
82
- source,
83
- score,
84
- reason: reason ?? null,
85
- });
86
- }
87
-
88
- function seedGateBlock(
89
- sessionId: string,
90
- toolName: string,
91
- reason: string
92
- ): void {
93
- trajectory.recordGateBlock({
94
- sessionId,
95
- toolName,
96
- reason,
97
- });
98
- }
99
-
100
- // -------------------------------------------------------------------------
101
- // listRecentNocturnalCandidateSessions
102
- // -------------------------------------------------------------------------
103
-
104
- describe('listRecentNocturnalCandidateSessions', () => {
105
- it('returns empty array when no sessions exist', () => {
106
- const result = extractor.listRecentNocturnalCandidateSessions();
107
- expect(result).toEqual([]);
108
- });
109
-
110
- it('returns sessions with tool calls above threshold', () => {
111
- seedSession('session-1');
112
- seedSession('session-2');
113
- seedToolCall('session-1', 'read_file', 'success');
114
- seedToolCall('session-2', 'read_file', 'success');
115
-
116
- const result = extractor.listRecentNocturnalCandidateSessions();
117
- expect(result).toHaveLength(2);
118
- expect(result.map((s) => s.sessionId).sort()).toEqual(['session-1', 'session-2']);
119
- });
120
-
121
- it('filters out sessions below minToolCalls threshold', () => {
122
- seedSession('session-1');
123
- seedSession('session-2');
124
- seedToolCall('session-1', 'read_file', 'success');
125
- // session-2 has no tool calls
126
-
127
- const result = extractor.listRecentNocturnalCandidateSessions({ minToolCalls: 1 });
128
- expect(result).toHaveLength(1);
129
- expect(result[0].sessionId).toBe('session-1');
130
- });
131
-
132
- it('counts failures correctly', () => {
133
- seedSession('session-1');
134
- seedToolCall('session-1', 'bash', 'failure', 'command failed');
135
- seedToolCall('session-1', 'read_file', 'success');
136
-
137
- const result = extractor.listRecentNocturnalCandidateSessions();
138
- expect(result[0].failureCount).toBe(1);
139
- expect(result[0].toolCallCount).toBe(2);
140
- });
141
-
142
- it('counts pain events and gate blocks', () => {
143
- seedSession('session-1');
144
- seedToolCall('session-1', 'bash', 'failure', 'command failed');
145
- seedPainEvent('session-1', 50, 'tool_failure', 'bash failed');
146
- seedGateBlock('session-1', 'delete_file', 'risky operation');
147
-
148
- const result = extractor.listRecentNocturnalCandidateSessions();
149
- expect(result[0].painEventCount).toBe(1);
150
- expect(result[0].gateBlockCount).toBe(1);
151
- });
152
-
153
- it('respects limit option', () => {
154
- for (let i = 0; i < 10; i++) {
155
- seedSession(`session-${i}`);
156
- seedToolCall(`session-${i}`, 'read_file', 'success');
157
- }
158
-
159
- const result = extractor.listRecentNocturnalCandidateSessions({ limit: 5 });
160
- expect(result).toHaveLength(5);
161
- });
162
-
163
- it('returns sessions ordered by most recently updated', () => {
164
- const old = new Date(Date.now() - 86400000).toISOString(); // 1 day ago
165
- const recent = new Date().toISOString();
166
- seedSession('session-old', old);
167
- seedSession('session-recent', recent);
168
- seedToolCall('session-old', 'read_file', 'success');
169
- seedToolCall('session-recent', 'read_file', 'success');
170
-
171
- const result = extractor.listRecentNocturnalCandidateSessions();
172
- expect(result[0].sessionId).toBe('session-recent');
173
- });
174
- });
175
-
176
- // -------------------------------------------------------------------------
177
- // getNocturnalSessionSnapshot
178
- // -------------------------------------------------------------------------
179
-
180
- describe('getNocturnalSessionSnapshot', () => {
181
- it('returns null for non-existent session', () => {
182
- const result = extractor.getNocturnalSessionSnapshot('does-not-exist');
183
- expect(result).toBeNull();
184
- });
185
-
186
- it('returns full snapshot with all turn types', () => {
187
- seedSession('session-1');
188
- seedAssistantTurn('session-1', 'I will read the file first.', 'User asked me to read the file.');
189
- seedAssistantTurn(
190
- 'session-1',
191
- 'I am editing the file now.',
192
- 'User asked me to edit the file.'
193
- );
194
- trajectory.recordUserTurn({
195
- sessionId: 'session-1',
196
- turnIndex: 0,
197
- rawText: 'Good, keep going',
198
- correctionDetected: false,
199
- });
200
- trajectory.recordUserTurn({
201
- sessionId: 'session-1',
202
- turnIndex: 1,
203
- rawText: 'No, that is wrong!',
204
- correctionDetected: true,
205
- correctionCue: 'explicit correction',
206
- });
207
- seedToolCall('session-1', 'read_file', 'success');
208
- seedToolCall('session-1', 'edit_file', 'success');
209
- seedPainEvent('session-1', 40, 'tool_failure', 'minor issue');
210
- seedGateBlock('session-1', 'delete_file', 'risky');
211
-
212
- const snapshot = extractor.getNocturnalSessionSnapshot('session-1');
213
-
214
- expect(snapshot).not.toBeNull();
215
- snapshot!;
216
-
217
- expect(snapshot!.sessionId).toBe('session-1');
218
- expect(snapshot!.assistantTurns).toHaveLength(2);
219
- expect(snapshot!.userTurns).toHaveLength(2);
220
- expect(snapshot!.toolCalls).toHaveLength(2);
221
- expect(snapshot!.painEvents).toHaveLength(1);
222
- expect(snapshot!.gateBlocks).toHaveLength(1);
223
- expect(snapshot!.stats.totalAssistantTurns).toBe(2);
224
- expect(snapshot!.stats.totalToolCalls).toBe(2);
225
- expect(snapshot!.stats.failureCount).toBe(0);
226
- });
227
-
228
- it('sanitizedText only — never raw_text', () => {
229
- seedSession('session-1');
230
- seedAssistantTurn(
231
- 'session-1',
232
- 'I will fix the bug now.',
233
- 'User private API key was exposed in the file content. Fix the bug now please!'
234
- );
235
-
236
- const snapshot = extractor.getNocturnalSessionSnapshot('session-1');
237
-
238
- expect(snapshot).not.toBeNull();
239
- // Sanitized text should not contain the private content
240
- expect(snapshot!.assistantTurns[0].sanitizedText).toBe('I will fix the bug now.');
241
- // Ensure raw text is NOT in the snapshot
242
- expect(JSON.stringify(snapshot)).not.toContain('API key');
243
- expect(JSON.stringify(snapshot)).not.toContain('exposed');
244
- });
245
-
246
- it('user turns expose only correctionCue — never raw user text', () => {
247
- seedSession('session-1');
248
- trajectory.recordUserTurn({
249
- sessionId: 'session-1',
250
- turnIndex: 0,
251
- rawText: 'DELETE ALL THE FILES immediately!',
252
- correctionDetected: true,
253
- correctionCue: 'User said "No, that is wrong!"',
254
- });
255
-
256
- const snapshot = extractor.getNocturnalSessionSnapshot('session-1');
257
-
258
- expect(snapshot).not.toBeNull();
259
- // Only correctionCue should be present
260
- expect(snapshot!.userTurns[0].correctionCue).toBe('User said "No, that is wrong!"');
261
- // rawText should NOT be in the snapshot
262
- expect(JSON.stringify(snapshot)).not.toContain('DELETE ALL THE FILES');
263
- });
264
-
265
- it('tool calls include outcome and error info but not params', () => {
266
- seedSession('session-1');
267
- seedToolCall('session-1', 'bash', 'failure', 'rm: cannot delete /protected: Permission denied');
268
-
269
- const snapshot = extractor.getNocturnalSessionSnapshot('session-1');
270
-
271
- expect(snapshot).not.toBeNull();
272
- expect(snapshot!.toolCalls[0].toolName).toBe('bash');
273
- expect(snapshot!.toolCalls[0].outcome).toBe('failure');
274
- expect(snapshot!.toolCalls[0].errorMessage).toBe('rm: cannot delete /protected: Permission denied');
275
- // params should not be included (we don't seed params in seedToolCall)
276
- expect(snapshot!.toolCalls[0]).not.toHaveProperty('paramsJson');
277
- });
278
-
279
- it('pain events include score and reason only', () => {
280
- seedSession('session-1');
281
- seedPainEvent('session-1', 65, 'gate_block', 'Agent attempted risky delete without plan');
282
-
283
- const snapshot = extractor.getNocturnalSessionSnapshot('session-1');
284
-
285
- expect(snapshot).not.toBeNull();
286
- expect(snapshot!.painEvents[0].score).toBe(65);
287
- expect(snapshot!.painEvents[0].reason).toBe('Agent attempted risky delete without plan');
288
- expect(snapshot!.painEvents[0].source).toBe('gate_block');
289
- });
290
-
291
- it('gate blocks include tool and reason only', () => {
292
- seedSession('session-1');
293
- seedGateBlock('session-1', 'delete_file', 'No PLAN.md found for risky path');
294
-
295
- const snapshot = extractor.getNocturnalSessionSnapshot('session-1');
296
-
297
- expect(snapshot).not.toBeNull();
298
- expect(snapshot!.gateBlocks[0].toolName).toBe('delete_file');
299
- expect(snapshot!.gateBlocks[0].reason).toBe('No PLAN.md found for risky path');
300
- });
301
-
302
- it('snapshot stats compute failureCount correctly', () => {
303
- seedSession('session-1');
304
- seedToolCall('session-1', 'bash', 'failure', 'error 1');
305
- seedToolCall('session-1', 'read_file', 'success');
306
- seedToolCall('session-1', 'bash', 'failure', 'error 2');
307
- seedToolCall('session-1', 'edit_file', 'success');
308
-
309
- const snapshot = extractor.getNocturnalSessionSnapshot('session-1');
310
-
311
- expect(snapshot).not.toBeNull();
312
- expect(snapshot!.stats.failureCount).toBe(2);
313
- expect(snapshot!.stats.totalToolCalls).toBe(4);
314
- });
315
- });
316
-
317
- // -------------------------------------------------------------------------
318
- // Convenience wrappers
319
- // -------------------------------------------------------------------------
320
-
321
- describe('module-level convenience functions', () => {
322
- it('listNocturnalCandidateSessions works as a standalone function', () => {
323
- seedSession('session-1');
324
- seedToolCall('session-1', 'read_file', 'success');
325
-
326
- const result = listNocturnalCandidateSessions(trajectory);
327
- expect(result).toHaveLength(1);
328
- });
329
-
330
- it('getNocturnalSessionSnapshot works as a standalone function', () => {
331
- seedSession('session-1');
332
- seedAssistantTurn('session-1', 'Hello', 'Hello');
333
-
334
- const result = getNocturnalSessionSnapshot(trajectory, 'session-1');
335
- expect(result).not.toBeNull();
336
- expect(result!.sessionId).toBe('session-1');
337
- });
338
-
339
- it('getNocturnalSessionSnapshot returns null for non-existent session', () => {
340
- const result = getNocturnalSessionSnapshot(trajectory, 'non-existent');
341
- expect(result).toBeNull();
342
- });
343
- });
344
-
345
- // -------------------------------------------------------------------------
346
- // Security / sanitization guarantees
347
- // -------------------------------------------------------------------------
348
-
349
- describe('sanitization guarantees', () => {
350
- it('snapshot JSON stringified output does not contain raw_text field names', () => {
351
- seedSession('session-1');
352
- seedAssistantTurn(
353
- 'session-1',
354
- 'Fixed the issue.',
355
- 'SECRET_API_KEY=sk-12345678 user private data here'
356
- );
357
-
358
- const snapshot = extractor.getNocturnalSessionSnapshot('session-1');
359
- const jsonStr = JSON.stringify(snapshot);
360
-
361
- // Verify sanitized text is present
362
- expect(jsonStr).toContain('Fixed the issue.');
363
- // Verify raw text content is NOT present
364
- expect(jsonStr).not.toContain('SECRET_API_KEY');
365
- expect(jsonStr).not.toContain('sk-12345678');
366
- expect(jsonStr).not.toContain('user private data');
367
- // Verify field name 'rawText' is NOT in the output
368
- expect(jsonStr).not.toContain('"rawText"');
369
- });
370
-
371
- it('session with very long raw text only stores sanitized version', () => {
372
- const longContent = 'A'.repeat(10000);
373
- seedSession('session-1');
374
- seedAssistantTurn('session-1', 'Summary: large file processed', longContent);
375
-
376
- const snapshot = extractor.getNocturnalSessionSnapshot('session-1');
377
- expect(snapshot).not.toBeNull();
378
- // The sanitized text should be short
379
- expect(snapshot!.assistantTurns[0].sanitizedText).toBe('Summary: large file processed');
380
- // The JSON should not contain the long raw content
381
- expect(JSON.stringify(snapshot)).not.toContain(longContent);
382
- });
383
- });
384
- });
385
-
386
- describe('Reflection Quality Metrics', () => {
387
- // -------------------------------------------------------------------------
388
- // computeThinkingModelActivation
389
- // -------------------------------------------------------------------------
390
-
391
- describe('computeThinkingModelActivation', () => {
392
- it('returns 0 for empty text', () => {
393
- expect(computeThinkingModelActivation('')).toBe(0);
394
- expect(computeThinkingModelActivation(' ')).toBe(0);
395
- });
396
-
397
- it('returns 0 for text with no thinking model patterns', () => {
398
- const text = 'Just do it now without any planning';
399
- const activation = computeThinkingModelActivation(text);
400
- expect(activation).toBeGreaterThanOrEqual(0);
401
- expect(activation).toBeLessThanOrEqual(1);
402
- });
403
-
404
- it('returns positive value for text with thinking model patterns', () => {
405
- // T-01 pattern: "let me first understand the structure"
406
- const text = 'Let me first understand the structure before editing anything';
407
- const activation = computeThinkingModelActivation(text);
408
- expect(activation).toBeGreaterThan(0);
409
- expect(activation).toBeLessThanOrEqual(1);
410
- });
411
-
412
- it('returns value rounded to 2 decimal places', () => {
413
- const text = 'Based on the evidence and logs, let me check the actual source code';
414
- const activation = computeThinkingModelActivation(text);
415
- // Should be rounded to 2 decimal places
416
- expect(activation * 100).toBe(Math.round(activation * 100));
417
- });
418
- });
419
-
420
- // -------------------------------------------------------------------------
421
- // computePlanningRatio
422
- // -------------------------------------------------------------------------
423
-
424
- describe('computePlanningRatio', () => {
425
- it('returns 0 for snapshot with no tool calls', () => {
426
- const snapshot: NocturnalSessionSnapshot = {
427
- sessionId: 'test',
428
- startedAt: '',
429
- principleId: '',
430
- assistantTurns: [],
431
- userTurns: [],
432
- toolCalls: [],
433
- painEvents: [],
434
- gateBlocks: [],
435
- stats: { failureCount: 0, totalPainEvents: 0, totalGateBlocks: 0 },
436
- };
437
- expect(computePlanningRatio(snapshot)).toBe(0);
438
- });
439
-
440
- it('returns 0 for snapshot with writes but no preceding reads', () => {
441
- const snapshot: NocturnalSessionSnapshot = {
442
- sessionId: 'test',
443
- startedAt: '',
444
- principleId: '',
445
- assistantTurns: [],
446
- userTurns: [],
447
- toolCalls: [
448
- { toolName: 'edit', outcome: 'success', filePath: null, durationMs: 100, exitCode: null, errorType: null, errorMessage: null, createdAt: '' },
449
- { toolName: 'write', outcome: 'success', filePath: null, durationMs: 100, exitCode: null, errorType: null, errorMessage: null, createdAt: '' },
450
- ],
451
- painEvents: [],
452
- gateBlocks: [],
453
- stats: { failureCount: 0, totalPainEvents: 0, totalGateBlocks: 0 },
454
- };
455
- expect(computePlanningRatio(snapshot)).toBe(0);
456
- });
457
-
458
- it('returns 1 for snapshot where all writes are preceded by reads', () => {
459
- const snapshot: NocturnalSessionSnapshot = {
460
- sessionId: 'test',
461
- startedAt: '',
462
- principleId: '',
463
- assistantTurns: [],
464
- userTurns: [],
465
- toolCalls: [
466
- { toolName: 'read', outcome: 'success', filePath: null, durationMs: 100, exitCode: null, errorType: null, errorMessage: null, createdAt: '' },
467
- { toolName: 'edit', outcome: 'success', filePath: null, durationMs: 100, exitCode: null, errorType: null, errorMessage: null, createdAt: '' },
468
- ],
469
- painEvents: [],
470
- gateBlocks: [],
471
- stats: { failureCount: 0, totalPainEvents: 0, totalGateBlocks: 0 },
472
- };
473
- expect(computePlanningRatio(snapshot)).toBe(1);
474
- });
475
-
476
- it('returns 0.5 when half of writes are preceded by reads', () => {
477
- const snapshot: NocturnalSessionSnapshot = {
478
- sessionId: 'test',
479
- startedAt: '',
480
- principleId: '',
481
- assistantTurns: [],
482
- userTurns: [],
483
- toolCalls: [
484
- { toolName: 'read', outcome: 'success', filePath: null, durationMs: 100, exitCode: null, errorType: null, errorMessage: null, createdAt: '' },
485
- { toolName: 'edit', outcome: 'success', filePath: null, durationMs: 100, exitCode: null, errorType: null, errorMessage: null, createdAt: '' },
486
- { toolName: 'edit', outcome: 'success', filePath: null, durationMs: 100, exitCode: null, errorType: null, errorMessage: null, createdAt: '' },
487
- ],
488
- painEvents: [],
489
- gateBlocks: [],
490
- stats: { failureCount: 0, totalPainEvents: 0, totalGateBlocks: 0 },
491
- };
492
- expect(computePlanningRatio(snapshot)).toBe(0.5);
493
- });
494
-
495
- it('returns value rounded to 2 decimal places', () => {
496
- const snapshot: NocturnalSessionSnapshot = {
497
- sessionId: 'test',
498
- startedAt: '',
499
- principleId: '',
500
- assistantTurns: [],
501
- userTurns: [],
502
- toolCalls: [
503
- { toolName: 'read', outcome: 'success', filePath: null, durationMs: 100, exitCode: null, errorType: null, errorMessage: null, createdAt: '' },
504
- { toolName: 'read', outcome: 'success', filePath: null, durationMs: 100, exitCode: null, errorType: null, errorMessage: null, createdAt: '' },
505
- { toolName: 'edit', outcome: 'success', filePath: null, durationMs: 100, exitCode: null, errorType: null, errorMessage: null, createdAt: '' },
506
- ],
507
- painEvents: [],
508
- gateBlocks: [],
509
- stats: { failureCount: 0, totalPainEvents: 0, totalGateBlocks: 0 },
510
- };
511
- expect(computePlanningRatio(snapshot)).toBe(1);
512
- });
513
- });
514
-
515
- // -------------------------------------------------------------------------
516
- // computeThinkingModelDelta
517
- // -------------------------------------------------------------------------
518
-
519
- describe('computeThinkingModelDelta', () => {
520
- it('returns 0 for identical texts', () => {
521
- const text = 'Just do it now';
522
- expect(computeThinkingModelDelta(text, text)).toBe(0);
523
- });
524
-
525
- it('returns positive delta when improved has more thinking models', () => {
526
- const original = 'Edit the file now';
527
- const improved = 'Let me first understand the structure before editing anything';
528
- const delta = computeThinkingModelDelta(original, improved);
529
- expect(delta).toBeGreaterThan(0);
530
- });
531
-
532
- it('returns negative delta when improved has fewer thinking models', () => {
533
- const original = 'Let me first understand the structure before editing anything';
534
- const improved = 'Edit the file now';
535
- const delta = computeThinkingModelDelta(original, improved);
536
- expect(delta).toBeLessThan(0);
537
- });
538
-
539
- it('returns delta rounded to 2 decimal places', () => {
540
- const original = 'Edit the file';
541
- const improved = 'Based on the evidence, let me check the actual source and verify before editing';
542
- const delta = computeThinkingModelDelta(original, improved);
543
- expect(delta * 100).toBe(Math.round(delta * 100));
544
- });
545
- });
546
-
547
- // -------------------------------------------------------------------------
548
- // computePlanningRatioGain
549
- // -------------------------------------------------------------------------
550
-
551
- describe('computePlanningRatioGain', () => {
552
- it('returns 0 for identical snapshots', () => {
553
- const snapshot: NocturnalSessionSnapshot = {
554
- sessionId: 'test',
555
- startedAt: '',
556
- principleId: '',
557
- assistantTurns: [],
558
- userTurns: [],
559
- toolCalls: [
560
- { toolName: 'read', outcome: 'success', filePath: null, durationMs: 100, exitCode: null, errorType: null, errorMessage: null, createdAt: '' },
561
- { toolName: 'edit', outcome: 'success', filePath: null, durationMs: 100, exitCode: null, errorType: null, errorMessage: null, createdAt: '' },
562
- ],
563
- painEvents: [],
564
- gateBlocks: [],
565
- stats: { failureCount: 0, totalPainEvents: 0, totalGateBlocks: 0 },
566
- };
567
- expect(computePlanningRatioGain(snapshot, snapshot)).toBe(0);
568
- });
569
-
570
- it('returns positive gain when improved has better planning ratio', () => {
571
- const original: NocturnalSessionSnapshot = {
572
- sessionId: 'test',
573
- startedAt: '',
574
- principleId: '',
575
- assistantTurns: [],
576
- userTurns: [],
577
- toolCalls: [
578
- { toolName: 'edit', outcome: 'success', filePath: null, durationMs: 100, exitCode: null, errorType: null, errorMessage: null, createdAt: '' },
579
- ],
580
- painEvents: [],
581
- gateBlocks: [],
582
- stats: { failureCount: 0, totalPainEvents: 0, totalGateBlocks: 0 },
583
- };
584
- const improved: NocturnalSessionSnapshot = {
585
- sessionId: 'test',
586
- startedAt: '',
587
- principleId: '',
588
- assistantTurns: [],
589
- userTurns: [],
590
- toolCalls: [
591
- { toolName: 'read', outcome: 'success', filePath: null, durationMs: 100, exitCode: null, errorType: null, errorMessage: null, createdAt: '' },
592
- { toolName: 'edit', outcome: 'success', filePath: null, durationMs: 100, exitCode: null, errorType: null, errorMessage: null, createdAt: '' },
593
- ],
594
- painEvents: [],
595
- gateBlocks: [],
596
- stats: { failureCount: 0, totalPainEvents: 0, totalGateBlocks: 0 },
597
- };
598
- const gain = computePlanningRatioGain(original, improved);
599
- expect(gain).toBeGreaterThan(0);
600
- });
601
-
602
- it('returns negative gain when improved has worse planning ratio', () => {
603
- const original: NocturnalSessionSnapshot = {
604
- sessionId: 'test',
605
- startedAt: '',
606
- principleId: '',
607
- assistantTurns: [],
608
- userTurns: [],
609
- toolCalls: [
610
- { toolName: 'read', outcome: 'success', filePath: null, durationMs: 100, exitCode: null, errorType: null, errorMessage: null, createdAt: '' },
611
- { toolName: 'edit', outcome: 'success', filePath: null, durationMs: 100, exitCode: null, errorType: null, errorMessage: null, createdAt: '' },
612
- ],
613
- painEvents: [],
614
- gateBlocks: [],
615
- stats: { failureCount: 0, totalPainEvents: 0, totalGateBlocks: 0 },
616
- };
617
- const improved: NocturnalSessionSnapshot = {
618
- sessionId: 'test',
619
- startedAt: '',
620
- principleId: '',
621
- assistantTurns: [],
622
- userTurns: [],
623
- toolCalls: [
624
- { toolName: 'edit', outcome: 'success', filePath: null, durationMs: 100, exitCode: null, errorType: null, errorMessage: null, createdAt: '' },
625
- ],
626
- painEvents: [],
627
- gateBlocks: [],
628
- stats: { failureCount: 0, totalPainEvents: 0, totalGateBlocks: 0 },
629
- };
630
- const gain = computePlanningRatioGain(original, improved);
631
- expect(gain).toBeLessThan(0);
632
- });
633
- });
634
- });