@cohortapp/agent-sdk 2.17.0 → 2.18.4

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 (526) hide show
  1. package/.claude/settings.json +18 -0
  2. package/.env.example +18 -5
  3. package/README.md +1 -0
  4. package/bin/maestro.mjs +62 -0
  5. package/docs/guides/billing-console-keys.md +60 -0
  6. package/docs/guides/front-door-session.md +54 -9
  7. package/docs/guides/mac-mini.md +20 -25
  8. package/docs/guides/setup-wizard.md +1 -1
  9. package/docs/runbooks/fleet-rollout.md +156 -0
  10. package/docs/runbooks/mac-mini-bootstrap.md +12 -14
  11. package/lib/action-executor.js +19 -3
  12. package/lib/budget-guard.mjs +279 -3
  13. package/lib/channels/base-adapter.mjs +3 -1
  14. package/lib/channels/contract.mjs +2 -1
  15. package/lib/channels/inbox-item.mjs +8 -0
  16. package/lib/claude-bin.mjs +5 -6
  17. package/lib/cli/doctor-checks.mjs +141 -10
  18. package/lib/cli/global-setup-extras.mjs +5 -1
  19. package/lib/cli/inbox.mjs +100 -15
  20. package/lib/cli/seat-auth.mjs +463 -0
  21. package/lib/cli/session.mjs +80 -12
  22. package/lib/collective/capture.mjs +8 -6
  23. package/lib/collective/global-config.mjs +63 -1
  24. package/lib/collective/presence.mjs +142 -5
  25. package/lib/comms/send-gate.mjs +559 -1
  26. package/lib/diagnostics/alerts.mjs +49 -0
  27. package/lib/diagnostics/cadence-output-freshness.mjs +288 -0
  28. package/lib/engine/agents/definitions.mjs +343 -0
  29. package/lib/engine/agents/persist.mjs +275 -0
  30. package/lib/engine/agents/runtime.mjs +748 -0
  31. package/lib/engine/agents/usage.mjs +95 -0
  32. package/lib/engine/auth-status.mjs +139 -0
  33. package/lib/engine/budget.mjs +194 -0
  34. package/lib/engine/cli.mjs +1204 -0
  35. package/lib/engine/commands/index.mjs +269 -0
  36. package/lib/engine/context/budget.mjs +219 -0
  37. package/lib/engine/context/cache.mjs +125 -0
  38. package/lib/engine/context/child-env.mjs +215 -0
  39. package/lib/engine/context/compaction.mjs +342 -0
  40. package/lib/engine/context/images.mjs +90 -0
  41. package/lib/engine/context/instructions.mjs +327 -0
  42. package/lib/engine/context/lazy-instructions.mjs +169 -0
  43. package/lib/engine/context/manager.mjs +182 -0
  44. package/lib/engine/context/real-path.mjs +91 -0
  45. package/lib/engine/context/secret-values.mjs +163 -0
  46. package/lib/engine/context/settings.mjs +274 -0
  47. package/lib/engine/context/stream-input.mjs +159 -0
  48. package/lib/engine/guard.mjs +152 -0
  49. package/lib/engine/hooks.mjs +713 -0
  50. package/lib/engine/loop.mjs +560 -0
  51. package/lib/engine/mcp/client.mjs +254 -0
  52. package/lib/engine/mcp/config.mjs +301 -0
  53. package/lib/engine/mcp/http.mjs +201 -0
  54. package/lib/engine/mcp/index.mjs +146 -0
  55. package/lib/engine/mcp/jsonrpc.mjs +147 -0
  56. package/lib/engine/mcp/naming.mjs +66 -0
  57. package/lib/engine/mcp/resources.mjs +89 -0
  58. package/lib/engine/mcp/results.mjs +133 -0
  59. package/lib/engine/mcp/stdio.mjs +137 -0
  60. package/lib/engine/mcp/supervisor.mjs +116 -0
  61. package/lib/engine/messages.mjs +104 -0
  62. package/lib/engine/output/json.mjs +164 -0
  63. package/lib/engine/output/stream-json.mjs +266 -0
  64. package/lib/engine/permissions.mjs +845 -0
  65. package/lib/engine/process-identity.mjs +164 -0
  66. package/lib/engine/process-tree.mjs +551 -0
  67. package/lib/engine/prompt.mjs +60 -0
  68. package/lib/engine/session/store.mjs +299 -0
  69. package/lib/engine/session-runtime/args.mjs +97 -0
  70. package/lib/engine/session-runtime/host.mjs +143 -0
  71. package/lib/engine/session-runtime/inbox.mjs +122 -0
  72. package/lib/engine/session-runtime/notifications.mjs +129 -0
  73. package/lib/engine/session-runtime/registry.mjs +328 -0
  74. package/lib/engine/session-runtime/runner.mjs +344 -0
  75. package/lib/engine/session-runtime/socket.mjs +212 -0
  76. package/lib/engine/session-runtime/wakeup.mjs +115 -0
  77. package/lib/engine/skills/index.mjs +321 -0
  78. package/lib/engine/tools/bash-background.mjs +533 -0
  79. package/lib/engine/tools/bash.mjs +216 -0
  80. package/lib/engine/tools/edit.mjs +97 -0
  81. package/lib/engine/tools/glob.mjs +81 -0
  82. package/lib/engine/tools/grep.mjs +224 -0
  83. package/lib/engine/tools/index.mjs +84 -0
  84. package/lib/engine/tools/list-agents.mjs +32 -0
  85. package/lib/engine/tools/ls.mjs +127 -0
  86. package/lib/engine/tools/monitor.mjs +82 -0
  87. package/lib/engine/tools/notebook-edit.mjs +218 -0
  88. package/lib/engine/tools/read.mjs +103 -0
  89. package/lib/engine/tools/schedule-wakeup.mjs +45 -0
  90. package/lib/engine/tools/schema.mjs +144 -0
  91. package/lib/engine/tools/send-message.mjs +77 -0
  92. package/lib/engine/tools/session.mjs +70 -0
  93. package/lib/engine/tools/todo.mjs +144 -0
  94. package/lib/engine/tools/toolsearch.mjs +217 -0
  95. package/lib/engine/tools/walk.mjs +193 -0
  96. package/lib/engine/tools/web-switch.mjs +31 -0
  97. package/lib/engine/tools/webfetch-html.mjs +387 -0
  98. package/lib/engine/tools/webfetch-net.mjs +340 -0
  99. package/lib/engine/tools/webfetch.mjs +198 -0
  100. package/lib/engine/tools/websearch.mjs +91 -0
  101. package/lib/engine/tools/workflow.mjs +95 -0
  102. package/lib/engine/tools/write.mjs +76 -0
  103. package/lib/engine/tui/line-editor.mjs +137 -0
  104. package/lib/engine/tui/render.mjs +86 -0
  105. package/lib/engine/tui/tui.mjs +274 -0
  106. package/lib/engine/wire/anthropic-messages.mjs +263 -0
  107. package/lib/engine/wire/effort.mjs +36 -0
  108. package/lib/engine/wire/errors.mjs +496 -0
  109. package/lib/engine/wire/http.mjs +441 -0
  110. package/lib/engine/wire/index.mjs +76 -0
  111. package/lib/engine/wire/openai-chat.mjs +332 -0
  112. package/lib/engine/wire/prompt-cache.mjs +79 -0
  113. package/lib/engine/wire/search.mjs +140 -0
  114. package/lib/engine/wire/sse.mjs +114 -0
  115. package/lib/engine/wire/stall.mjs +349 -0
  116. package/lib/engine/wire/token-provider.mjs +175 -0
  117. package/lib/engine/wire/usage.mjs +192 -0
  118. package/lib/engine/workflow/host.mjs +524 -0
  119. package/lib/engine/workflow/journal.mjs +188 -0
  120. package/lib/engine/workflow/json-schema.mjs +171 -0
  121. package/lib/engine/workflow/meta.mjs +329 -0
  122. package/lib/engine/workflow/notifications.mjs +52 -0
  123. package/lib/engine/workflow/runtime.mjs +447 -0
  124. package/lib/engine/workflow/sandbox.mjs +534 -0
  125. package/lib/engine/workflow/worker.mjs +141 -0
  126. package/lib/engine/workflow/worktree.mjs +74 -0
  127. package/lib/execution/disposition.mjs +1 -1
  128. package/lib/execution/intake.mjs +10 -0
  129. package/lib/execution/surface-policy.mjs +15 -0
  130. package/lib/learning/curator.mjs +8 -6
  131. package/lib/learning/reflect.mjs +8 -6
  132. package/lib/model-router/catalog/cohort.yaml +137 -0
  133. package/lib/model-router/catalog.mjs +118 -1
  134. package/lib/model-router/failover.mjs +67 -16
  135. package/lib/model-router/llm-task.mjs +39 -3
  136. package/lib/model-router/resolve.mjs +89 -3
  137. package/lib/model-router/spawn.mjs +46 -47
  138. package/lib/model-router/taxonomy.mjs +126 -4
  139. package/lib/org/cost-sync.mjs +141 -11
  140. package/lib/org/inbound/broadcast.mjs +289 -0
  141. package/lib/org/inbound/collective.mjs +375 -0
  142. package/lib/org/inbound/directedness.mjs +96 -8
  143. package/lib/org/inbound/facts.mjs +78 -2
  144. package/lib/org/inbound/project.mjs +22 -0
  145. package/lib/org/inbound/surfaces.mjs +14 -0
  146. package/lib/org/llm-token.mjs +879 -0
  147. package/lib/org/mesh.mjs +61 -0
  148. package/lib/org/messaging.mjs +3 -1
  149. package/lib/org/protocol.checksum +1 -1
  150. package/lib/org/protocol.mjs +15 -0
  151. package/lib/org/quota.mjs +520 -0
  152. package/lib/org/tool-surface.mjs +104 -16
  153. package/lib/org/ui-parity.mjs +16 -1
  154. package/lib/org/work-ledger.mjs +37 -6
  155. package/lib/rate-guard.mjs +114 -1
  156. package/lib/resource-governor.mjs +41 -6
  157. package/lib/runtime/adapter.mjs +823 -0
  158. package/lib/runtime/child-env.mjs +191 -0
  159. package/lib/runtime/legacy-shell-guard.mjs +97 -0
  160. package/lib/runtime/seat-engine.mjs +162 -0
  161. package/lib/session/ask-ledger.mjs +271 -0
  162. package/lib/session/current-work.mjs +676 -0
  163. package/lib/session/feed-core.mjs +40 -3
  164. package/lib/session/launch-args.mjs +56 -4
  165. package/lib/session/status-summary.mjs +26 -9
  166. package/lib/session/upgrade-notice.mjs +42 -0
  167. package/lib/setup/claude-probe.mjs +117 -13
  168. package/lib/setup/enrich.mjs +13 -10
  169. package/lib/setup/sections/model.mjs +39 -13
  170. package/lib/telemetry/collect.mjs +208 -9
  171. package/lib/upgrade/ignored-drift.mjs +105 -0
  172. package/lib/voice/post-call-brief.mjs +30 -17
  173. package/package.json +13 -3
  174. package/plugins/maestro-skills/skills/board-work.md +5 -0
  175. package/plugins/maestro-skills/skills/inbound-triage.md +56 -15
  176. package/plugins/maestro-skills/skills/main-session.md +18 -7
  177. package/scripts/ci/check-tarball-fidelity.mjs +126 -2
  178. package/scripts/ci/run-tests.mjs +47 -19
  179. package/scripts/cohort-llm/api-key-helper.mjs +92 -0
  180. package/scripts/collective/hook-runner.mjs +29 -2
  181. package/scripts/continuous-monitor.sh +13 -0
  182. package/scripts/cost/track-claude-usage.mjs +15 -0
  183. package/scripts/daemon/agent-daemon.mjs +408 -20
  184. package/scripts/daemon/assurance.mjs +48 -12
  185. package/scripts/daemon/cadence-consumer.mjs +218 -68
  186. package/scripts/daemon/cadence-handlers.mjs +73 -4
  187. package/scripts/daemon/classifier.mjs +75 -26
  188. package/scripts/daemon/context-compiler.mjs +51 -37
  189. package/scripts/daemon/deliver.mjs +30 -1
  190. package/scripts/daemon/dispatcher.mjs +595 -149
  191. package/scripts/daemon/health.mjs +14 -1
  192. package/scripts/daemon/maestro-daemon.mjs +11 -0
  193. package/scripts/daemon/prompt-builder.mjs +24 -0
  194. package/scripts/daemon/responder.mjs +246 -79
  195. package/scripts/daemon/sdk-version.mjs +98 -16
  196. package/scripts/eval/probe-gateway.mjs +635 -0
  197. package/scripts/eval/replay/extract.mjs +270 -0
  198. package/scripts/eval/replay/grade.mjs +260 -0
  199. package/scripts/eval/replay/lib/config.mjs +50 -0
  200. package/scripts/eval/replay/lib/effects.mjs +65 -0
  201. package/scripts/eval/replay/lib/fixture.mjs +188 -0
  202. package/scripts/eval/replay/lib/judge.mjs +72 -0
  203. package/scripts/eval/replay/lib/redact.mjs +136 -0
  204. package/scripts/eval/replay/lib/sandbox.mjs +170 -0
  205. package/scripts/eval/replay/lib/schema-check.mjs +63 -0
  206. package/scripts/eval/replay/lib/transcript.mjs +76 -0
  207. package/scripts/eval/replay/mcp-replay-stub.mjs +101 -0
  208. package/scripts/eval/replay/report.mjs +185 -0
  209. package/scripts/eval/replay/run.mjs +404 -0
  210. package/scripts/fleet/rollout.mjs +1094 -0
  211. package/scripts/hooks/pre-send-audit.sh +36 -245
  212. package/scripts/hooks/pre-write-yaml-validate.mjs +275 -0
  213. package/scripts/hooks/validate-state-yaml.sh +190 -0
  214. package/scripts/huddle/huddle-llm.mjs +361 -0
  215. package/scripts/huddle/huddle-server.mjs +46 -121
  216. package/scripts/local-triggers/autoupdate.sh +448 -78
  217. package/scripts/local-triggers/run-trigger.sh +13 -0
  218. package/scripts/maintenance/pin-integrity.mjs +364 -0
  219. package/scripts/poll-slack-events.sh +41 -9
  220. package/scripts/poller/slack-socket-mode.mjs +28 -3
  221. package/scripts/session/supervisor.mjs +80 -13
  222. package/scripts/spawn-session.sh +13 -0
  223. package/bin/maestro.test.mjs +0 -1574
  224. package/lib/action-executor.test.mjs +0 -871
  225. package/lib/archetype.test.mjs +0 -132
  226. package/lib/assurance/plan-note.test.mjs +0 -234
  227. package/lib/assurance/room-budget.test.mjs +0 -486
  228. package/lib/assurance/tier.test.mjs +0 -174
  229. package/lib/autonomy.test.mjs +0 -66
  230. package/lib/backlog.test.mjs +0 -302
  231. package/lib/backup/policy.test.mjs +0 -305
  232. package/lib/budget-escalate.test.mjs +0 -232
  233. package/lib/budget-guard.envelope.test.mjs +0 -476
  234. package/lib/budget-guard.test.mjs +0 -427
  235. package/lib/cadence-bus-requeue.test.mjs +0 -83
  236. package/lib/cadence-bus-schedule.test.mjs +0 -194
  237. package/lib/cadence-bus.test.mjs +0 -720
  238. package/lib/cadences.test.mjs +0 -230
  239. package/lib/capability/inventory.test.mjs +0 -232
  240. package/lib/capability.test.mjs +0 -78
  241. package/lib/channels/base-adapter.test.mjs +0 -590
  242. package/lib/channels/channels.test.mjs +0 -371
  243. package/lib/channels/contract.test.mjs +0 -162
  244. package/lib/channels/inbox-item.test.mjs +0 -368
  245. package/lib/channels/orgmail/adapter.test.mjs +0 -448
  246. package/lib/channels/pairing.test.mjs +0 -270
  247. package/lib/channels/repeat-suppressor.test.mjs +0 -134
  248. package/lib/channels/slack-adapter.test.mjs +0 -212
  249. package/lib/channels/telegram-adapter.test.mjs +0 -306
  250. package/lib/channels/voice/adapter.test.mjs +0 -278
  251. package/lib/channels/whatsapp/adapter-baileys.test.mjs +0 -359
  252. package/lib/channels/whatsapp/baileys-typing.test.mjs +0 -154
  253. package/lib/charter.test.mjs +0 -89
  254. package/lib/claude-bin.test.mjs +0 -131
  255. package/lib/cli/board.test.mjs +0 -227
  256. package/lib/cli/design.test.mjs +0 -270
  257. package/lib/cli/doctor-checks.test.mjs +0 -336
  258. package/lib/cli/global-setup-extras.test.mjs +0 -462
  259. package/lib/cli/inbox.test.mjs +0 -230
  260. package/lib/cli/session-ack.test.mjs +0 -63
  261. package/lib/cli/session.test.mjs +0 -613
  262. package/lib/collective/capture.test.mjs +0 -121
  263. package/lib/collective/cards.test.mjs +0 -114
  264. package/lib/collective/config.test.mjs +0 -123
  265. package/lib/collective/global-config.test.mjs +0 -220
  266. package/lib/collective/global-skills.test.mjs +0 -126
  267. package/lib/collective/presence.test.mjs +0 -95
  268. package/lib/collective/recall.test.mjs +0 -116
  269. package/lib/collective/vendor-skills.test.mjs +0 -306
  270. package/lib/comms/send-gate.test.mjs +0 -770
  271. package/lib/comms.test.mjs +0 -41
  272. package/lib/context/budget.test.mjs +0 -252
  273. package/lib/context/history-scope.test.mjs +0 -79
  274. package/lib/cost/ledger-row.test.mjs +0 -183
  275. package/lib/design/design-md.test.mjs +0 -318
  276. package/lib/design/fixtures/DESIGN.golden.md +0 -238
  277. package/lib/design/fixtures/PRODUCT.golden.md +0 -67
  278. package/lib/design/fixtures/foundation.json +0 -133
  279. package/lib/design/refresh-gate.test.mjs +0 -144
  280. package/lib/design/write.test.mjs +0 -241
  281. package/lib/diagnostics/alerts.test.mjs +0 -318
  282. package/lib/diagnostics/backup-freshness.test.mjs +0 -185
  283. package/lib/diagnostics/counters.test.mjs +0 -206
  284. package/lib/diagnostics/events.test.mjs +0 -290
  285. package/lib/diagnostics/otel.test.mjs +0 -196
  286. package/lib/diagnostics/trace.test.mjs +0 -251
  287. package/lib/env-compat.test.mjs +0 -104
  288. package/lib/execution/disposition.test.mjs +0 -553
  289. package/lib/execution/drive.test.mjs +0 -270
  290. package/lib/execution/effects.test.mjs +0 -344
  291. package/lib/execution/intake.test.mjs +0 -389
  292. package/lib/execution/journal.test.mjs +0 -261
  293. package/lib/execution/match.test.mjs +0 -235
  294. package/lib/execution/pipeline.test.mjs +0 -392
  295. package/lib/execution/route.test.mjs +0 -186
  296. package/lib/execution/surface-policy.test.mjs +0 -162
  297. package/lib/fs-atomic.test.mjs +0 -72
  298. package/lib/fs-ownership.test.mjs +0 -158
  299. package/lib/goals/admission.test.mjs +0 -164
  300. package/lib/goals/classify.test.mjs +0 -167
  301. package/lib/goals/collaborate.test.mjs +0 -336
  302. package/lib/goals/gaps.test.mjs +0 -284
  303. package/lib/goals/loop.test.mjs +0 -845
  304. package/lib/hooks/bus.test.mjs +0 -387
  305. package/lib/identity/persona.test.mjs +0 -142
  306. package/lib/kpi-sensors.test.mjs +0 -278
  307. package/lib/kpi.test.mjs +0 -244
  308. package/lib/learning/config.test.mjs +0 -75
  309. package/lib/learning/counters.test.mjs +0 -69
  310. package/lib/learning/curator-consolidate.test.mjs +0 -238
  311. package/lib/learning/curator.test.mjs +0 -106
  312. package/lib/learning/reflect.test.mjs +0 -0
  313. package/lib/learning/session-index.test.mjs +0 -125
  314. package/lib/learning/skill-writer.test.mjs +0 -210
  315. package/lib/mandate/audit.test.mjs +0 -195
  316. package/lib/mandate/contract.test.mjs +0 -185
  317. package/lib/mandate/derive.test.mjs +0 -274
  318. package/lib/mandate/model.test.mjs +0 -164
  319. package/lib/mandate/refresh.test.mjs +0 -389
  320. package/lib/mcp/server.test.mjs +0 -426
  321. package/lib/model-router/auth-profiles.test.mjs +0 -580
  322. package/lib/model-router/catalog.test.mjs +0 -385
  323. package/lib/model-router/economics.test.mjs +0 -438
  324. package/lib/model-router/failover.test.mjs +0 -439
  325. package/lib/model-router/health.test.mjs +0 -338
  326. package/lib/model-router/integration-coverage.test.mjs +0 -831
  327. package/lib/model-router/integration.test.mjs +0 -564
  328. package/lib/model-router/ledger.test.mjs +0 -415
  329. package/lib/model-router/llm-task.test.mjs +0 -392
  330. package/lib/model-router/org-credentials.test.mjs +0 -265
  331. package/lib/model-router/pricing-refresh.test.mjs +0 -286
  332. package/lib/model-router/reconcile.test.mjs +0 -316
  333. package/lib/model-router/repair.test.mjs +0 -180
  334. package/lib/model-router/spawn.test.mjs +0 -446
  335. package/lib/model-router/taxonomy.test.mjs +0 -410
  336. package/lib/model-router.test.mjs +0 -1207
  337. package/lib/org/activity.test.mjs +0 -134
  338. package/lib/org/approvals.test.mjs +0 -216
  339. package/lib/org/awareness.test.mjs +0 -159
  340. package/lib/org/board-mine-cache.test.mjs +0 -53
  341. package/lib/org/board.test.mjs +0 -187
  342. package/lib/org/bootstrap-context.test.mjs +0 -153
  343. package/lib/org/client.test.mjs +0 -1206
  344. package/lib/org/cohort-client.test.mjs +0 -126
  345. package/lib/org/cost-sync.test.mjs +0 -153
  346. package/lib/org/doctor.test.mjs +0 -346
  347. package/lib/org/engagement-ledger.test.mjs +0 -112
  348. package/lib/org/engagement.test.mjs +0 -739
  349. package/lib/org/handoff.test.mjs +0 -269
  350. package/lib/org/inbound/directedness.test.mjs +0 -668
  351. package/lib/org/inbound/facts.test.mjs +0 -471
  352. package/lib/org/inbound/hydrate.test.mjs +0 -908
  353. package/lib/org/inbound/index.test.mjs +0 -429
  354. package/lib/org/inbound/project.test.mjs +0 -287
  355. package/lib/org/integration-tools.test.mjs +0 -160
  356. package/lib/org/keys.test.mjs +0 -92
  357. package/lib/org/knowledge.test.mjs +0 -326
  358. package/lib/org/leases.test.mjs +0 -235
  359. package/lib/org/mesh-directives.test.mjs +0 -110
  360. package/lib/org/mesh-integration.test.mjs +0 -127
  361. package/lib/org/mesh.test.mjs +0 -400
  362. package/lib/org/messaging.test.mjs +0 -471
  363. package/lib/org/param-contract.test.mjs +0 -477
  364. package/lib/org/policy.test.mjs +0 -237
  365. package/lib/org/protocol.checksum.test.mjs +0 -90
  366. package/lib/org/protocol.test.mjs +0 -323
  367. package/lib/org/push.test.mjs +0 -792
  368. package/lib/org/registry.test.mjs +0 -100
  369. package/lib/org/resource-tools.test.mjs +0 -361
  370. package/lib/org/tool-access.test.mjs +0 -144
  371. package/lib/org/tool-surface-integration.test.mjs +0 -120
  372. package/lib/org/tool-surface.test.mjs +0 -1268
  373. package/lib/org/typing.test.mjs +0 -291
  374. package/lib/org/ui-parity.test.mjs +0 -560
  375. package/lib/org/verify.test.mjs +0 -194
  376. package/lib/org/work-ledger.test.mjs +0 -273
  377. package/lib/plan/adoption-e2e.test.mjs +0 -366
  378. package/lib/plan/budget-enforcement.test.mjs +0 -400
  379. package/lib/plan/compile.test.mjs +0 -382
  380. package/lib/plan/emit.test.mjs +0 -269
  381. package/lib/plan/explain.test.mjs +0 -188
  382. package/lib/prompts/parallelism.test.mjs +0 -177
  383. package/lib/rag/rag.test.mjs +0 -505
  384. package/lib/rate-guard.test.mjs +0 -272
  385. package/lib/reactive-gate.test.mjs +0 -57
  386. package/lib/render.test.mjs +0 -68
  387. package/lib/resource-governor.test.mjs +0 -488
  388. package/lib/scheduling/dynamic-jobs.test.mjs +0 -344
  389. package/lib/scheduling/jitter.test.mjs +0 -140
  390. package/lib/secrets/broker.test.mjs +0 -280
  391. package/lib/secrets/providers.test.mjs +0 -274
  392. package/lib/security/audit-engine.test.mjs +0 -424
  393. package/lib/security/coerce-args.test.mjs +0 -281
  394. package/lib/security/dangerous-tools.test.mjs +0 -68
  395. package/lib/security/external-content.test.mjs +0 -84
  396. package/lib/security/redact.test.mjs +0 -441
  397. package/lib/security/secret-equal.test.mjs +0 -55
  398. package/lib/session/config.test.mjs +0 -92
  399. package/lib/session/feed-core.test.mjs +0 -198
  400. package/lib/session/first-run.test.mjs +0 -121
  401. package/lib/session/frontdoor.test.mjs +0 -205
  402. package/lib/session/handoffs.test.mjs +0 -183
  403. package/lib/session/identity.test.mjs +0 -180
  404. package/lib/session/inbox-claims.test.mjs +0 -286
  405. package/lib/session/launch-args.test.mjs +0 -157
  406. package/lib/session/liveness.test.mjs +0 -100
  407. package/lib/session/status-summary.test.mjs +0 -118
  408. package/lib/session-permissions.test.mjs +0 -120
  409. package/lib/setup/claude-probe.test.mjs +0 -187
  410. package/lib/setup/completeness.test.mjs +0 -110
  411. package/lib/setup/context-pack.test.mjs +0 -89
  412. package/lib/setup/enrich.test.mjs +0 -115
  413. package/lib/setup/enroll-from-cohort.test.mjs +0 -300
  414. package/lib/setup/integration.test.mjs +0 -162
  415. package/lib/setup/io.test.mjs +0 -77
  416. package/lib/setup/runner.test.mjs +0 -132
  417. package/lib/setup/sections/identity.test.mjs +0 -234
  418. package/lib/setup/sections/inventory.test.mjs +0 -198
  419. package/lib/setup/sections/learning.test.mjs +0 -81
  420. package/lib/setup/sections/mandate.test.mjs +0 -388
  421. package/lib/setup/sections/messaging.test.mjs +0 -127
  422. package/lib/setup/sections/model.test.mjs +0 -240
  423. package/lib/setup/sections/org.test.mjs +0 -346
  424. package/lib/setup/sections/orgmail.test.mjs +0 -118
  425. package/lib/setup/sections/recovery.test.mjs +0 -98
  426. package/lib/setup/sections/subagents.test.mjs +0 -429
  427. package/lib/setup/sections/verify.test.mjs +0 -175
  428. package/lib/setup/sot.test.mjs +0 -81
  429. package/lib/setup/state.test.mjs +0 -115
  430. package/lib/singleton.test.mjs +0 -151
  431. package/lib/subagents/cli.test.mjs +0 -389
  432. package/lib/subagents/client.test.mjs +0 -309
  433. package/lib/subagents/gap.test.mjs +0 -234
  434. package/lib/subagents/lock.test.mjs +0 -248
  435. package/lib/subagents/manifest.test.mjs +0 -175
  436. package/lib/subagents/refs.test.mjs +0 -204
  437. package/lib/subagents/resolve.test.mjs +0 -422
  438. package/lib/subagents/schema.test.mjs +0 -328
  439. package/lib/telemetry/alerts.test.mjs +0 -109
  440. package/lib/telemetry/collect.test.mjs +0 -1274
  441. package/lib/tool-definitions-integration.test.mjs +0 -83
  442. package/lib/tool-definitions.test.mjs +0 -437
  443. package/lib/upgrade/global-refresh.test.mjs +0 -65
  444. package/lib/upgrade/launchd-reconcile.test.mjs +0 -272
  445. package/lib/upgrade/post-steps.test.mjs +0 -200
  446. package/lib/upgrade/verify.test.mjs +0 -164
  447. package/lib/util/fetch-timeout.test.mjs +0 -202
  448. package/lib/util/reconnect.test.mjs +0 -369
  449. package/lib/util/unhandled.test.mjs +0 -216
  450. package/lib/voice/outbound.test.mjs +0 -69
  451. package/lib/voice/session-rotation.test.mjs +0 -114
  452. package/lib/voice/stt.test.mjs +0 -226
  453. package/lib/voice/voice.test.mjs +0 -990
  454. package/scripts/cadence/enqueue-cadence-tick.test.mjs +0 -187
  455. package/scripts/ci/check-docs-accuracy.test.mjs +0 -409
  456. package/scripts/ci/check-durable-write-seam.test.mjs +0 -90
  457. package/scripts/ci/check-no-build-artifacts.test.mjs +0 -71
  458. package/scripts/ci/check-no-residual-identity.test.mjs +0 -202
  459. package/scripts/ci/check-skill-packs.test.mjs +0 -495
  460. package/scripts/ci/check-subagent-frontmatter.test.mjs +0 -124
  461. package/scripts/ci/check.test.mjs +0 -194
  462. package/scripts/ci/conformance-org-api.test.mjs +0 -425
  463. package/scripts/cloud-relay/voice/relay-identity.test.mjs +0 -96
  464. package/scripts/collective/hook-runner.test.mjs +0 -173
  465. package/scripts/cost/fleet-digest.test.mjs +0 -207
  466. package/scripts/cost/track-claude-usage-pricing.test.mjs +0 -183
  467. package/scripts/cost/track-claude-usage.test.mjs +0 -148
  468. package/scripts/daemon/agent-daemon-board-mine.test.mjs +0 -96
  469. package/scripts/daemon/agent-daemon-design.test.mjs +0 -238
  470. package/scripts/daemon/agent-daemon-frontdoor.test.mjs +0 -60
  471. package/scripts/daemon/agent-daemon.test.mjs +0 -995
  472. package/scripts/daemon/assurance-e2e.test.mjs +0 -613
  473. package/scripts/daemon/assurance.test.mjs +0 -1791
  474. package/scripts/daemon/board-mirror.test.mjs +0 -165
  475. package/scripts/daemon/cadence-consumer-frontdoor.test.mjs +0 -393
  476. package/scripts/daemon/cadence-consumer-governance.test.mjs +0 -276
  477. package/scripts/daemon/cadence-consumer.test.mjs +0 -776
  478. package/scripts/daemon/cadence-handlers.test.mjs +0 -837
  479. package/scripts/daemon/classifier-identity.test.mjs +0 -137
  480. package/scripts/daemon/classifier.test.mjs +0 -266
  481. package/scripts/daemon/classify-kind.test.mjs +0 -40
  482. package/scripts/daemon/context-compiler.test.mjs +0 -406
  483. package/scripts/daemon/deliver.test.mjs +0 -564
  484. package/scripts/daemon/dispatcher-cooldown.test.mjs +0 -122
  485. package/scripts/daemon/dispatcher-governance.test.mjs +0 -1013
  486. package/scripts/daemon/dispatcher-resume.test.mjs +0 -166
  487. package/scripts/daemon/dispatcher-session-continuity.test.mjs +0 -365
  488. package/scripts/daemon/execution-ladder.test.mjs +0 -470
  489. package/scripts/daemon/goal-steward-cadence.test.mjs +0 -312
  490. package/scripts/daemon/inbox-deferral-session.test.mjs +0 -49
  491. package/scripts/daemon/inbox-deferral.test.mjs +0 -336
  492. package/scripts/daemon/inbox-wake.test.mjs +0 -199
  493. package/scripts/daemon/integration.test.mjs +0 -149
  494. package/scripts/daemon/lib/self-echo.test.mjs +0 -153
  495. package/scripts/daemon/lib/session-router.test.mjs +0 -554
  496. package/scripts/daemon/prompt-builder-preamble.test.mjs +0 -210
  497. package/scripts/daemon/prompt-builder.test.mjs +0 -556
  498. package/scripts/daemon/responder-cost.test.mjs +0 -68
  499. package/scripts/daemon/responder-history.test.mjs +0 -221
  500. package/scripts/daemon/sdk-version.test.mjs +0 -31
  501. package/scripts/daemon/session-lock.test.mjs +0 -252
  502. package/scripts/daemon/session-outcomes.test.mjs +0 -533
  503. package/scripts/daemon/typing-registry.test.mjs +0 -102
  504. package/scripts/hooks/pre-send-audit.test.mjs +0 -354
  505. package/scripts/huddle/huddle-prompt.test.mjs +0 -176
  506. package/scripts/local-triggers/autoupdate.test.mjs +0 -518
  507. package/scripts/local-triggers/generate-plists.test.mjs +0 -456
  508. package/scripts/media-generation/brand-clause.test.mjs +0 -135
  509. package/scripts/org/send-orgmail.first-contact.test.mjs +0 -102
  510. package/scripts/poller/inbox-privilege-injection.test.mjs +0 -167
  511. package/scripts/poller/inbox-scan-poller.test.mjs +0 -295
  512. package/scripts/poller/lib/cloud-relay-dedup.test.mjs +0 -133
  513. package/scripts/poller/slack-socket-mode.test.mjs +0 -805
  514. package/scripts/poller-launchd/install.test.mjs +0 -243
  515. package/scripts/restore-from-backup.test.mjs +0 -181
  516. package/scripts/session/feed.test.mjs +0 -196
  517. package/scripts/session/supervisor-sh.test.mjs +0 -218
  518. package/scripts/session/supervisor.test.mjs +0 -482
  519. package/scripts/setup/configure-macos.test.mjs +0 -306
  520. package/scripts/setup/gen-subagent-manifest.test.mjs +0 -124
  521. package/scripts/setup/generate-agent-package-json.test.mjs +0 -143
  522. package/scripts/setup/generate-capability.test.mjs +0 -134
  523. package/scripts/setup/init-agent.test.mjs +0 -370
  524. package/scripts/setup/init-skill-marketplace.test.mjs +0 -193
  525. package/scripts/vendor/sync-skill-packs.test.mjs +0 -103
  526. package/scripts/watchdog/memory-watchdog.test.mjs +0 -64
@@ -1,278 +0,0 @@
1
- /**
2
- * lib/kpi-sensors.test.mjs — the sensor registry.
3
- *
4
- * The load-bearing tests are the NEGATIVES: a sensor that cannot measure must
5
- * produce an honest `{ok:false, reason}` that `measureKpi` refuses to turn into
6
- * a sample, and that `admit`'s honesty gate therefore refuses to turn into work.
7
- * A silent zero here would let an agent invent its own justification for
8
- * anything, which is the whole reason the honesty gate exists.
9
- */
10
-
11
- "use strict";
12
-
13
- import { test } from "node:test";
14
- import assert from "node:assert/strict";
15
- import { mkdtempSync, writeFileSync, mkdirSync, rmSync } from "node:fs";
16
- import { tmpdir } from "node:os";
17
- import { join } from "node:path";
18
-
19
- import {
20
- SENSOR_SPECS,
21
- buildSensorRegistry,
22
- reachableCapabilities,
23
- refusalFor,
24
- hintedCapabilities,
25
- reduceGeneric,
26
- firstArray,
27
- pluck,
28
- } from "./kpi-sensors.mjs";
29
- import { measureKpi, recordMeasurement, readSeries } from "./kpi.mjs";
30
- import { admit } from "./goals/admission.mjs";
31
-
32
- const NOW = Date.parse("2026-08-11T12:00:00.000Z");
33
- const now = () => NOW;
34
-
35
- function root() {
36
- const d = mkdtempSync(join(tmpdir(), "kpi-sensors-"));
37
- mkdirSync(join(d, "config"), { recursive: true });
38
- return d;
39
- }
40
-
41
- /** A fake org: capability → error-or-result frame. */
42
- function fakeExec(table, calls = []) {
43
- return async (name, input) => {
44
- calls.push({ name, input });
45
- const hit = table[name];
46
- if (hit === undefined) return { ok: false, error: { code: "NOT_FOUND", message: `no fake for ${name}` } };
47
- if (typeof hit === "function") return hit(input);
48
- return { ok: true, result: hit };
49
- };
50
- }
51
-
52
- const OBJ = (capability, params = {}) => ({
53
- key: "k", kind: "GOAL", state: "active", metric: "m", direction: "up",
54
- target: 90, tolerance: 2, cadence: "weekly",
55
- // `adoptedById` is part of the law the admission gate enforces, not decoration:
56
- // an active objective with nobody on record as having adopted it admits no work.
57
- memberId: "M-OWNER", adoptedById: "M-SUPERVISOR",
58
- sensor: { capability, params, source: "method" },
59
- });
60
-
61
- // ---------------------------------------------------------------------------
62
- // Reductions
63
- // ---------------------------------------------------------------------------
64
-
65
- test("board_ready measures an on-time rate and shows what it walked", async () => {
66
- const calls = [];
67
- const sensors = buildSensorRegistry({
68
- agentRoot: root(), now,
69
- executeImpl: fakeExec({ board_ready: { items: [
70
- { id: "a", dueAt: "2026-06-01T00:00:00Z" }, // overdue
71
- { id: "b", dueAt: "2026-12-01T00:00:00Z" }, // not yet
72
- { id: "c", dueAt: null }, // cannot be late
73
- { id: "d", dueAt: null },
74
- ] } }, calls),
75
- });
76
- const m = await measureKpi(OBJ("board_ready", { metric: "on_time_rate" }), { sensors, now });
77
- assert.equal(m.ok, true);
78
- assert.equal(m.value, 75);
79
- assert.equal(m.source, "method");
80
- assert.equal(m.evidence.rowCount, 4);
81
- assert.equal(m.evidence.detail.overdue, 1);
82
- assert.equal(m.evidence.detail.withDueDate, 2);
83
- assert.equal(calls[0].name, "board_ready");
84
- });
85
-
86
- test("decision_list distinguishes an honest zero from an absence", async () => {
87
- const sensors = (table) => buildSensorRegistry({ agentRoot: root(), now, executeImpl: fakeExec(table) });
88
-
89
- // Rows exist, none undecided → a REAL zero-day backlog.
90
- const zero = await measureKpi(OBJ("decision_list"), {
91
- sensors: sensors({ decision_list: { decisions: [{ id: "1", status: "adopted", createdAt: "2026-08-01T00:00:00Z", decidedAt: "2026-08-02T00:00:00Z" }] } }),
92
- now,
93
- });
94
- assert.equal(zero.ok, true);
95
- assert.equal(zero.value, 0);
96
- assert.match(zero.evidence.detail.note, /honest zero/);
97
-
98
- // No rows at all → NOT a zero.
99
- const none = await measureKpi(OBJ("decision_list"), { sensors: sensors({ decision_list: { decisions: [] } }), now });
100
- assert.equal(none.ok, false);
101
- assert.equal(none.value, null);
102
- assert.equal(none.reason, "no-rows");
103
- });
104
-
105
- test("email_inbox pairs a reply to its inbound message and never averages silence in as speed", async () => {
106
- const sensors = buildSensorRegistry({
107
- agentRoot: root(), now,
108
- executeImpl: fakeExec({ email_inbox: { messages: [
109
- { id: "1", threadId: "t1", direction: "INBOUND", receivedAt: "2026-08-11T00:00:00Z" },
110
- { id: "2", threadId: "t1", direction: "OUTBOUND", sentAt: "2026-08-11T02:00:00Z" },
111
- { id: "3", threadId: "t2", direction: "INBOUND", receivedAt: "2026-08-10T00:00:00Z" }, // never answered
112
- ] } }),
113
- });
114
- const m = await measureKpi(OBJ("email_inbox"), { sensors, now });
115
- assert.equal(m.ok, true);
116
- assert.equal(m.value, 2); // 2h, not 1h — the unanswered one is excluded
117
- assert.equal(m.evidence.detail.answered, 1);
118
- assert.equal(m.evidence.detail.unanswered, 1);
119
- });
120
-
121
- test("crm_list_deals never sends hq an illegal stage filter", async () => {
122
- const calls = [];
123
- const sensors = buildSensorRegistry({
124
- agentRoot: root(), now,
125
- executeImpl: fakeExec({ crm_list_deals: { deals: [
126
- { id: "1", stage: "INBOUND", valueCents: 100_000_000 },
127
- { id: "2", stage: "CLOSED", valueCents: 900_000_000 },
128
- ] } }, calls),
129
- });
130
- // "open" is not in hq's enum — it must not reach the wire.
131
- const m = await measureKpi(OBJ("crm_list_deals", { stage: "open" }), { sensors, now });
132
- assert.deepEqual(calls[0].input, {});
133
- assert.equal(m.value, 1_000_000); // the CLOSED deal is excluded
134
- assert.equal(m.evidence.detail.open, 1);
135
- });
136
-
137
- test("reduceGeneric is the escape hatch: filter an array, sum a field", () => {
138
- const r = reduceGeneric(
139
- { byAgent: [{ agent: "A016", usd: 4384.87 }, { agent: "A017", usd: 3455.27 }] },
140
- { arrayPath: "byAgent", where: { key: "agent", equals: "A016" }, sum: "usd" },
141
- );
142
- assert.equal(r.value, 4384.87);
143
- assert.equal(r.rowCount, 1);
144
- assert.equal(pluck({ a: { b: [1, 2] } }, "a.b.1"), 2);
145
- assert.equal(firstArray({ x: 1, deals: [1, 2, 3] }).key, "deals");
146
- });
147
-
148
- // ---------------------------------------------------------------------------
149
- // THE NEGATIVES
150
- // ---------------------------------------------------------------------------
151
-
152
- test("NEGATIVE: an unregistered capability yields no-implementation and NO sample", async () => {
153
- const dir = root();
154
- const sensors = buildSensorRegistry({ agentRoot: dir, now, executeImpl: fakeExec({}) });
155
- const o = OBJ("messaging_send");
156
- const m = await measureKpi(o, { sensors, now });
157
-
158
- assert.equal(m.ok, false);
159
- assert.equal(m.reason, "no-implementation");
160
- assert.equal(m.value, null);
161
-
162
- // The loop only records when m.ok — assert the ledger stays empty either way.
163
- assert.equal(readSeries(dir, o.key).length, 0);
164
- });
165
-
166
- test("NEGATIVE: a WRITE capability is refused by name, not silently missing", async () => {
167
- const sensors = buildSensorRegistry({ agentRoot: root(), now, executeImpl: fakeExec({}) });
168
- assert.equal(typeof sensors.task_update, "function", "the derivation can pick it, so it must have an answer");
169
- const m = await measureKpi(OBJ("task_update"), { sensors, now });
170
- assert.equal(m.ok, false);
171
- assert.equal(m.reason, "capability-is-not-a-sensor");
172
- assert.match(m.evidence.detail.why, /may not mutate the org/);
173
- assert.match(refusalFor("escalation_raise"), /may not mutate the org/);
174
- assert.match(refusalFor("approval_wait"), /long-poll/);
175
- assert.equal(refusalFor("board_ready"), null);
176
- });
177
-
178
- test("NEGATIVE: a failed org read is a read-failure, never a zero", async () => {
179
- const dir = root();
180
- const sensors = buildSensorRegistry({
181
- agentRoot: dir, now,
182
- executeImpl: fakeExec({ board_ready: () => ({ ok: false, error: { code: "UNAUTHORIZED", message: "no org credential" } }) }),
183
- });
184
- const m = await measureKpi(OBJ("board_ready"), { sensors, now });
185
- assert.equal(m.ok, false);
186
- assert.equal(m.reason, "read-failed");
187
- assert.equal(m.value, null);
188
- assert.equal(m.evidence.detail.code, "UNAUTHORIZED");
189
- });
190
-
191
- test("NEGATIVE: a sensor that throws is caught and reported, not propagated", async () => {
192
- const sensors = buildSensorRegistry({
193
- agentRoot: root(), now,
194
- executeImpl: async () => { throw new Error("socket hang up"); },
195
- });
196
- const m = await measureKpi(OBJ("board_ready"), { sensors, now });
197
- assert.equal(m.ok, false);
198
- assert.equal(m.reason, "sensor-threw");
199
- });
200
-
201
- test("NEGATIVE: an unmeasurable objective cannot reach the admission honesty gate", async () => {
202
- const dir = root();
203
- const sensors = buildSensorRegistry({ agentRoot: dir, now, executeImpl: fakeExec({ board_ready: { items: [] } }) });
204
- const o = OBJ("board_ready");
205
-
206
- const m = await measureKpi(o, { sensors, now });
207
- assert.equal(m.ok, false, "empty board → no measurement");
208
-
209
- // Exactly what the loop does: record ONLY when ok.
210
- if (m.ok) recordMeasurement(dir, { objectiveKey: o.key, value: m.value, window: "2026-W33", source: m.source }, { now });
211
- const series = readSeries(dir, o.key);
212
- assert.equal(series.length, 0);
213
-
214
- const verdict = admit(
215
- { title: "do something", advances: [o.key], why: { reason: "kpi-gap" } },
216
- { objective: o, gap: { gap: 15, withinTolerance: false, trend: "unknown" }, latestSample: series[series.length - 1] || null, openItems: [], staleness: { outcome: "full" } },
217
- );
218
- assert.equal(verdict.admitted, false);
219
- assert.equal(verdict.gate, "honesty");
220
- rmSync(dir, { recursive: true, force: true });
221
- });
222
-
223
- test("a MEASURED sample satisfies the honesty gate — the joint, end to end", async () => {
224
- const dir = root();
225
- const sensors = buildSensorRegistry({
226
- agentRoot: dir, now,
227
- executeImpl: fakeExec({ board_ready: { items: [{ id: "a", dueAt: "2026-06-01T00:00:00Z" }, { id: "b", dueAt: null }, { id: "c", dueAt: null }, { id: "d", dueAt: null }] } }),
228
- });
229
- const o = OBJ("board_ready");
230
- const m = await measureKpi(o, { sensors, now });
231
- assert.equal(m.ok, true);
232
- recordMeasurement(dir, { objectiveKey: o.key, value: m.value, window: "2026-W33", source: m.source, evidence: m.evidence }, { now });
233
-
234
- const series = readSeries(dir, o.key);
235
- const verdict = admit(
236
- { title: "raise the on-time rate", advances: [o.key], why: { reason: "kpi-gap" }, estCents: 50 },
237
- { objective: o, gap: { gap: 15, withinTolerance: false, trend: "unknown" }, latestSample: series[series.length - 1], openItems: [], remainingCents: 500, staleness: { outcome: "full" } },
238
- );
239
- assert.equal(verdict.admitted, true, verdict.reason);
240
- assert.equal(series[0].evidence.method, "board_ready");
241
- assert.equal(series[0].evidence.rowCount, 4);
242
- rmSync(dir, { recursive: true, force: true });
243
- });
244
-
245
- // ---------------------------------------------------------------------------
246
- // Reachability — the fail-open that must not become fail-closed
247
- // ---------------------------------------------------------------------------
248
-
249
- test("a missing or empty manifest returns undefined, not an empty Set", () => {
250
- const dir = root();
251
- const logs = [];
252
- const log = (l, m) => logs.push(`${l}:${m}`);
253
- assert.equal(reachableCapabilities(dir, { log }), undefined);
254
- assert.ok(logs.some((l) => /no capability manifest/.test(l)), "the degradation is logged");
255
-
256
- writeFileSync(join(dir, "config", "capability-manifest.json"), JSON.stringify({ entries: [] }));
257
- assert.equal(reachableCapabilities(dir, { log }), undefined, "zero reachable entries must not refuse every objective");
258
-
259
- writeFileSync(join(dir, "config", "capability-manifest.json"), "{not json");
260
- assert.equal(reachableCapabilities(dir, { log }), undefined);
261
-
262
- writeFileSync(join(dir, "config", "capability-manifest.json"), JSON.stringify({ entries: [{ id: "board_ready", reachable: true }, { id: "task_update", reachable: false }] }));
263
- const set = reachableCapabilities(dir, { log });
264
- assert.ok(set.has("board_ready"));
265
- assert.ok(!set.has("task_update"));
266
- rmSync(dir, { recursive: true, force: true });
267
- });
268
-
269
- test("every capability the mandate derivation can bind has an answer in the registry", () => {
270
- const sensors = buildSensorRegistry({ agentRoot: root(), now, executeImpl: fakeExec({}) });
271
- for (const cap of hintedCapabilities()) {
272
- assert.equal(typeof sensors[cap], "function", `${cap} is hinted by derive.SENSOR_HINTS but has no registry entry`);
273
- }
274
- // …and every spec is a READ.
275
- for (const cap of Object.keys(SENSOR_SPECS)) {
276
- assert.equal(refusalFor(cap), null, `${cap} has a spec so it must not be refused`);
277
- }
278
- });
package/lib/kpi.test.mjs DELETED
@@ -1,244 +0,0 @@
1
- /**
2
- * kpi.test.mjs — measurement, the series, and the gap.
3
- * Run: `node --test lib/kpi.test.mjs`
4
- */
5
-
6
- "use strict";
7
-
8
- import { test } from "node:test";
9
- import assert from "node:assert/strict";
10
- import { mkdtempSync, rmSync } from "node:fs";
11
- import { tmpdir } from "node:os";
12
- import { join } from "node:path";
13
-
14
- import {
15
- windowFor,
16
- isoWeek,
17
- isDue,
18
- readSeries,
19
- recordMeasurement,
20
- recordIntervention,
21
- readInterventions,
22
- measureKpi,
23
- gapFor,
24
- periodSeconds,
25
- } from "./kpi.mjs";
26
-
27
- function root() {
28
- return mkdtempSync(join(tmpdir(), "kpi-test-"));
29
- }
30
- function cleanup(p) {
31
- try { rmSync(p, { recursive: true, force: true }); } catch { /* */ }
32
- }
33
-
34
- const T = Date.parse("2026-08-11T09:00:00Z"); // a Tuesday, ISO week 33
35
- const now = () => T;
36
-
37
- const OBJ = {
38
- key: "pipeline",
39
- metric: "pipeline_coverage_x",
40
- target: 3,
41
- tolerance: 0.2,
42
- direction: "up",
43
- cadence: "weekly",
44
- sensor: { capability: "crm_list_deals", params: { stage: "open" }, source: "method" },
45
- };
46
-
47
- // ---------------------------------------------------------------------------
48
- // windows
49
- // ---------------------------------------------------------------------------
50
-
51
- test("windowFor produces the right idempotency key per cadence", () => {
52
- assert.equal(windowFor("weekly", T), "2026-W33");
53
- assert.equal(windowFor("monthly", T), "2026-08");
54
- assert.equal(windowFor("quarterly", T), "2026-Q3");
55
- assert.equal(windowFor("daily", T), "2026-08-11");
56
- assert.equal(isoWeek(new Date(T)).week, 33);
57
- });
58
-
59
- test("periodSeconds covers every cadence and defaults to a day", () => {
60
- assert.equal(periodSeconds("weekly"), 604800);
61
- assert.equal(periodSeconds("monthly"), 2592000);
62
- assert.equal(periodSeconds("nonsense"), 86400);
63
- });
64
-
65
- // ---------------------------------------------------------------------------
66
- // the ledger
67
- // ---------------------------------------------------------------------------
68
-
69
- test("recordMeasurement is idempotent on (objective, window) and refuses a weaker duplicate", () => {
70
- const r = root();
71
- try {
72
- const a = recordMeasurement(r, { objectiveKey: "pipeline", value: 2.1, window: "2026-W33", source: "method" }, { now });
73
- assert.equal(a.recorded, true);
74
-
75
- const dup = recordMeasurement(r, { objectiveKey: "pipeline", value: 9.9, window: "2026-W33", source: "llm" }, { now });
76
- assert.equal(dup.recorded, false);
77
- assert.equal(dup.reason, "duplicate-window");
78
- assert.equal(readSeries(r, "pipeline")[0].value, 2.1, "the llm guess must not overwrite the measured number");
79
-
80
- // A STRONGER source may supersede a weaker one in the same window.
81
- const r2 = root();
82
- try {
83
- recordMeasurement(r2, { objectiveKey: "p", value: 1, window: "w", source: "llm" }, { now });
84
- const better = recordMeasurement(r2, { objectiveKey: "p", value: 5, window: "w", source: "method" }, { now });
85
- assert.equal(better.recorded, true);
86
- assert.equal(readSeries(r2, "p")[0].value, 5);
87
- assert.equal(readSeries(r2, "p").length, 1, "the window collapses to the strongest source");
88
- } finally { cleanup(r2); }
89
- } finally { cleanup(r); }
90
- });
91
-
92
- test("recordMeasurement refuses a non-numeric value and logs it", () => {
93
- const r = root();
94
- const logs = [];
95
- try {
96
- const res = recordMeasurement(r, { objectiveKey: "p", value: "banana", window: "w", source: "method" }, { now, log: (l, m) => logs.push(`${l}:${m}`) });
97
- assert.equal(res.recorded, false);
98
- assert.equal(res.reason, "non-numeric value");
99
- assert.equal(logs.length, 1, "a refusal is never silent");
100
- } finally { cleanup(r); }
101
- });
102
-
103
- test("interventions are the learning signal and read back newest-last", () => {
104
- const r = root();
105
- try {
106
- for (const d of [-1, 0.5, 2]) {
107
- recordIntervention(r, { objectiveKey: "pipeline", taskId: `t${d}`, expectedDelta: 1, realizedDelta: d }, { now });
108
- }
109
- const last2 = readInterventions(r, "pipeline", 2);
110
- assert.deepEqual(last2.map((i) => i.realizedDelta), [0.5, 2]);
111
- } finally { cleanup(r); }
112
- });
113
-
114
- test("isDue opens once per window and closes after a sample; no cadence is never due", () => {
115
- const r = root();
116
- try {
117
- assert.equal(isDue(OBJ, [], { now }).due, true);
118
- recordMeasurement(r, { objectiveKey: "pipeline", value: 2, window: "2026-W33", source: "method" }, { now });
119
- assert.equal(isDue(OBJ, readSeries(r, "pipeline"), { now }).due, false);
120
- const none = isDue({ ...OBJ, cadence: null }, [], { now });
121
- assert.equal(none.due, false);
122
- assert.match(none.reason, /no measure cadence/);
123
- } finally { cleanup(r); }
124
- });
125
-
126
- // ---------------------------------------------------------------------------
127
- // measurement — the negative cases matter most
128
- // ---------------------------------------------------------------------------
129
-
130
- test("measureKpi executes a registered sensor and stamps method-sourced evidence", async () => {
131
- const seen = [];
132
- const m = await measureKpi(OBJ, {
133
- sensors: { crm_list_deals: async ({ params }) => { seen.push(params); return { value: 2.4, rowCount: 17 }; } },
134
- reachable: new Set(["crm_list_deals"]),
135
- });
136
- assert.equal(m.ok, true);
137
- assert.equal(m.value, 2.4);
138
- assert.equal(m.source, "method");
139
- assert.equal(m.evidence.method, "crm_list_deals");
140
- assert.equal(m.evidence.rowCount, 17);
141
- assert.deepEqual(seen, [{ stage: "open" }]);
142
- });
143
-
144
- test("NEGATIVE — an unreachable capability refuses to measure rather than recording a zero", async () => {
145
- const logs = [];
146
- const m = await measureKpi(OBJ, {
147
- sensors: { crm_list_deals: async () => 2.4 },
148
- reachable: new Set(["something_else"]),
149
- log: (l, msg) => logs.push(msg),
150
- });
151
- assert.equal(m.ok, false);
152
- assert.equal(m.reason, "unreachable-capability");
153
- assert.equal(m.value, null);
154
- assert.match(logs.join(" "), /NOT reachable/);
155
- });
156
-
157
- test("NEGATIVE — a missing sensor implementation, a throwing sensor and a non-numeric result all fail loudly", async () => {
158
- const logs = [];
159
- const log = (l, msg) => logs.push(msg);
160
-
161
- const missing = await measureKpi(OBJ, { sensors: {}, reachable: new Set(["crm_list_deals"]), log });
162
- assert.equal(missing.reason, "no-implementation");
163
-
164
- const threw = await measureKpi(OBJ, {
165
- sensors: { crm_list_deals: async () => { throw new Error("server unreachable"); } },
166
- reachable: new Set(["crm_list_deals"]),
167
- log,
168
- });
169
- assert.equal(threw.reason, "sensor-threw");
170
-
171
- const junk = await measureKpi(OBJ, {
172
- sensors: { crm_list_deals: async () => ({ value: "n/a" }) },
173
- reachable: new Set(["crm_list_deals"]),
174
- log,
175
- });
176
- assert.equal(junk.reason, "non-numeric");
177
- assert.equal(logs.length, 3, "every degradation is logged");
178
- });
179
-
180
- test("an objective with a human sensor is not an error — it is a human obligation", async () => {
181
- const m = await measureKpi({ ...OBJ, sensor: { capability: null, source: "human" } }, { sensors: {} });
182
- assert.equal(m.ok, false);
183
- assert.equal(m.reason, "no-sensor");
184
- assert.equal(m.source, "human");
185
- });
186
-
187
- // ---------------------------------------------------------------------------
188
- // the gap
189
- // ---------------------------------------------------------------------------
190
-
191
- function series(values, cadence = "weekly", at = "2026-08-11T09:00:00Z") {
192
- return values.map((v, i) => ({ objectiveKey: "pipeline", value: v, window: `w${i}`, at, source: "method" }));
193
- }
194
-
195
- test("gapFor honours direction: up is a shortfall, down is an excess", () => {
196
- const up = gapFor(OBJ, series([2.0]), { now });
197
- assert.equal(up.gap, 1);
198
- assert.equal(up.withinTolerance, false);
199
- assert.equal(up.normalizedGap, +(1 / 3).toFixed(6));
200
-
201
- const down = gapFor({ ...OBJ, direction: "down", target: 5 }, series([7]), { now });
202
- assert.equal(down.gap, 2, "for a down metric, above target is the gap");
203
-
204
- const under = gapFor({ ...OBJ, direction: "down", target: 5 }, series([3]), { now });
205
- assert.equal(under.gap, -2);
206
- assert.equal(under.withinTolerance, true, "beating a down-target is not a gap");
207
- });
208
-
209
- test("gapFor honours tolerance and band direction", () => {
210
- const inTol = gapFor(OBJ, series([2.85]), { now });
211
- assert.ok(inTol.gap <= 0.2 && inTol.withinTolerance, "0.15 short of 3 is inside a 0.2 tolerance");
212
-
213
- const band = gapFor({ ...OBJ, direction: "band", target: 10, tolerance: 2 }, series([13]), { now });
214
- assert.equal(band.gap, 1, "a band gap is the distance outside the band");
215
- const inBand = gapFor({ ...OBJ, direction: "band", target: 10, tolerance: 2 }, series([11]), { now });
216
- assert.equal(inBand.gap, 0);
217
- });
218
-
219
- test("gapFor computes a real trend over the gap, not the raw value", () => {
220
- assert.equal(gapFor(OBJ, series([1.0, 1.8, 2.5]), { now }).trend, "improving");
221
- assert.equal(gapFor(OBJ, series([2.5, 1.8, 1.0]), { now }).trend, "worsening");
222
- assert.equal(gapFor(OBJ, series([2.0, 2.0, 2.0]), { now }).trend, "flat");
223
- assert.equal(gapFor(OBJ, series([2.0]), { now }).trend, "unknown");
224
- });
225
-
226
- test("NEGATIVE — no samples, and no target, both report rather than compute", () => {
227
- const none = gapFor(OBJ, [], { now });
228
- assert.equal(none.gap, null);
229
- assert.equal(none.reason, "no samples");
230
-
231
- const noTarget = gapFor({ ...OBJ, target: null }, series([2]), { now });
232
- assert.equal(noTarget.gap, null);
233
- assert.match(noTarget.reason, /no numeric target/);
234
- assert.equal(noTarget.target, null, "a null target must never coerce to 0");
235
- });
236
-
237
- test("a sensor silent for 2× its cadence is itself a gap", () => {
238
- const stale = gapFor(OBJ, series([2], "weekly", "2026-07-01T09:00:00Z"), { now });
239
- assert.equal(stale.stale, true);
240
- assert.ok(stale.stalenessPeriods > 2);
241
-
242
- const fresh = gapFor(OBJ, series([2], "weekly", "2026-08-10T09:00:00Z"), { now });
243
- assert.equal(fresh.stale, false);
244
- });
@@ -1,75 +0,0 @@
1
- /**
2
- * config.test.mjs — learning config loader tests.
3
- * Run: node --test lib/learning/config.test.mjs
4
- */
5
- "use strict";
6
-
7
- import { test } from "node:test";
8
- import assert from "node:assert/strict";
9
- import { mkdtempSync, mkdirSync, writeFileSync } from "node:fs";
10
- import { tmpdir } from "node:os";
11
- import { join } from "node:path";
12
-
13
- import { loadLearningConfig, isLearningEnabled, DEFAULTS } from "./config.mjs";
14
-
15
- function tmpRoot() {
16
- const root = mkdtempSync(join(tmpdir(), "maestro-learncfg-"));
17
- mkdirSync(join(root, "config"), { recursive: true });
18
- return root;
19
- }
20
-
21
- test("loadLearningConfig: defaults when no file / no root", async () => {
22
- assert.deepEqual(await loadLearningConfig(null), DEFAULTS);
23
- const root = tmpRoot();
24
- const cfg = await loadLearningConfig(root);
25
- assert.equal(cfg.minTurns, 4);
26
- assert.equal(cfg.reviewModel, "sonnet");
27
- });
28
-
29
- test("loadLearningConfig: file overrides merge over defaults", async () => {
30
- const root = tmpRoot();
31
- writeFileSync(join(root, "config", "learning.yaml"), "minTurns: 6\nreviewModel: sonnet\nenabled: false\n");
32
- const cfg = await loadLearningConfig(root);
33
- assert.equal(cfg.minTurns, 6);
34
- assert.equal(cfg.enabled, false);
35
- assert.equal(isLearningEnabled(cfg), false);
36
- // Untouched keys keep their defaults.
37
- assert.equal(cfg.archiveDays, 90);
38
- });
39
-
40
- test("loadLearningConfig: inherits + unions the collective denylist", async () => {
41
- const root = tmpRoot();
42
- writeFileSync(join(root, "config", "learning.yaml"), "repoDenylist: [extra-repo]\n");
43
- const cfg = await loadLearningConfig(root, { repoDenylist: ["client-work", "secrets"] });
44
- assert.ok(cfg.repoDenylist.includes("client-work"));
45
- assert.ok(cfg.repoDenylist.includes("secrets"));
46
- assert.ok(cfg.repoDenylist.includes("extra-repo"), "file extends the inherited floor");
47
- });
48
-
49
- test("loadLearningConfig: consolidation defaults present (enabled / 5 / 14 / sonnet)", async () => {
50
- const root = tmpRoot();
51
- const cfg = await loadLearningConfig(root);
52
- assert.equal(cfg.consolidation.enabled, true);
53
- assert.equal(cfg.consolidation.minSkills, 5);
54
- assert.equal(cfg.consolidation.intervalDays, 14);
55
- assert.equal(cfg.consolidation.model, "sonnet");
56
- });
57
-
58
- test("loadLearningConfig: consolidation overrides deep-merge (keep untouched sub-keys)", async () => {
59
- const root = tmpRoot();
60
- writeFileSync(join(root, "config", "learning.yaml"), "consolidation:\n enabled: false\n minSkills: 3\n");
61
- const cfg = await loadLearningConfig(root);
62
- assert.equal(cfg.consolidation.enabled, false);
63
- assert.equal(cfg.consolidation.minSkills, 3);
64
- // intervalDays / model untouched → defaults.
65
- assert.equal(cfg.consolidation.intervalDays, 14);
66
- assert.equal(cfg.consolidation.model, "sonnet");
67
- });
68
-
69
- test("loadLearningConfig: corrupt yaml fails open to defaults+inheritance", async () => {
70
- const root = tmpRoot();
71
- writeFileSync(join(root, "config", "learning.yaml"), "minTurns: [unterminated");
72
- const cfg = await loadLearningConfig(root, { repoDenylist: ["ring-fenced"] });
73
- assert.equal(cfg.minTurns, 4);
74
- assert.deepEqual(cfg.repoDenylist, ["ring-fenced"]);
75
- });
@@ -1,69 +0,0 @@
1
- /**
2
- * counters.test.mjs — nudge-counter bump/reset/read tests.
3
- * Run: node --test lib/learning/counters.test.mjs
4
- */
5
- "use strict";
6
-
7
- import { test } from "node:test";
8
- import assert from "node:assert/strict";
9
- import { mkdtempSync, existsSync, readFileSync, writeFileSync, mkdirSync } from "node:fs";
10
- import { tmpdir } from "node:os";
11
- import { join } from "node:path";
12
-
13
- import {
14
- countersPath, emptyCounters, readCounters,
15
- bumpToolUses, bumpSessions, resetSkillCounters,
16
- } from "./counters.mjs";
17
-
18
- function tmpRoot() {
19
- return mkdtempSync(join(tmpdir(), "maestro-counters-"));
20
- }
21
-
22
- test("readCounters: fresh root reads as empty, never throws", () => {
23
- const root = tmpRoot();
24
- assert.deepEqual(readCounters(root), emptyCounters());
25
- // No agentRoot → still empty, no throw.
26
- assert.deepEqual(readCounters(null), emptyCounters());
27
- });
28
-
29
- test("readCounters: corrupt file fails open to empty", () => {
30
- const root = tmpRoot();
31
- const p = countersPath(root);
32
- mkdirSync(join(root, "state", "learning"), { recursive: true });
33
- writeFileSync(p, "{ not json");
34
- assert.deepEqual(readCounters(root), emptyCounters());
35
- });
36
-
37
- test("bumpToolUses: increments and persists atomically", () => {
38
- const root = tmpRoot();
39
- bumpToolUses(root);
40
- bumpToolUses(root, 4);
41
- const c = readCounters(root);
42
- assert.equal(c.tool_uses_since_skill, 5);
43
- assert.ok(existsSync(countersPath(root)), "counters.json written");
44
- // Round-trips through disk (not just in-memory).
45
- const onDisk = JSON.parse(readFileSync(countersPath(root), "utf8"));
46
- assert.equal(onDisk.tool_uses_since_skill, 5);
47
- });
48
-
49
- test("bumpSessions: increments + stamps last_reflect_at", () => {
50
- const root = tmpRoot();
51
- const nowMs = Date.parse("2026-06-09T10:00:00Z");
52
- bumpSessions(root, { nowMs });
53
- bumpSessions(root, { nowMs });
54
- const c = readCounters(root);
55
- assert.equal(c.sessions_since_skill, 2);
56
- assert.equal(c.last_reflect_at, new Date(nowMs).toISOString());
57
- });
58
-
59
- test("resetSkillCounters: zeroes both counters + stamps last_skill_at", () => {
60
- const root = tmpRoot();
61
- bumpToolUses(root, 50);
62
- bumpSessions(root);
63
- const nowMs = Date.parse("2026-06-09T12:00:00Z");
64
- resetSkillCounters(root, { nowMs });
65
- const c = readCounters(root);
66
- assert.equal(c.tool_uses_since_skill, 0);
67
- assert.equal(c.sessions_since_skill, 0);
68
- assert.equal(c.last_skill_at, new Date(nowMs).toISOString());
69
- });