@cohortapp/agent-sdk 2.17.0 → 2.18.5

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 (531) 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-slots.mjs +234 -0
  23. package/lib/collective/capture.mjs +8 -6
  24. package/lib/collective/config.mjs +2 -0
  25. package/lib/collective/global-config.mjs +63 -1
  26. package/lib/collective/loop-guard.mjs +155 -0
  27. package/lib/collective/presence.mjs +142 -5
  28. package/lib/comms/send-gate.mjs +559 -1
  29. package/lib/diagnostics/alerts.mjs +49 -0
  30. package/lib/diagnostics/cadence-output-freshness.mjs +288 -0
  31. package/lib/engine/agents/definitions.mjs +343 -0
  32. package/lib/engine/agents/persist.mjs +275 -0
  33. package/lib/engine/agents/runtime.mjs +748 -0
  34. package/lib/engine/agents/usage.mjs +95 -0
  35. package/lib/engine/auth-status.mjs +139 -0
  36. package/lib/engine/budget.mjs +194 -0
  37. package/lib/engine/cli.mjs +1204 -0
  38. package/lib/engine/commands/index.mjs +269 -0
  39. package/lib/engine/context/budget.mjs +219 -0
  40. package/lib/engine/context/cache.mjs +125 -0
  41. package/lib/engine/context/child-env.mjs +215 -0
  42. package/lib/engine/context/compaction.mjs +342 -0
  43. package/lib/engine/context/images.mjs +90 -0
  44. package/lib/engine/context/instructions.mjs +327 -0
  45. package/lib/engine/context/lazy-instructions.mjs +169 -0
  46. package/lib/engine/context/manager.mjs +182 -0
  47. package/lib/engine/context/real-path.mjs +91 -0
  48. package/lib/engine/context/secret-values.mjs +163 -0
  49. package/lib/engine/context/settings.mjs +274 -0
  50. package/lib/engine/context/stream-input.mjs +159 -0
  51. package/lib/engine/guard.mjs +152 -0
  52. package/lib/engine/hooks.mjs +713 -0
  53. package/lib/engine/loop.mjs +560 -0
  54. package/lib/engine/mcp/client.mjs +254 -0
  55. package/lib/engine/mcp/config.mjs +301 -0
  56. package/lib/engine/mcp/http.mjs +201 -0
  57. package/lib/engine/mcp/index.mjs +146 -0
  58. package/lib/engine/mcp/jsonrpc.mjs +147 -0
  59. package/lib/engine/mcp/naming.mjs +66 -0
  60. package/lib/engine/mcp/resources.mjs +89 -0
  61. package/lib/engine/mcp/results.mjs +133 -0
  62. package/lib/engine/mcp/stdio.mjs +137 -0
  63. package/lib/engine/mcp/supervisor.mjs +116 -0
  64. package/lib/engine/messages.mjs +104 -0
  65. package/lib/engine/output/json.mjs +164 -0
  66. package/lib/engine/output/stream-json.mjs +266 -0
  67. package/lib/engine/permissions.mjs +845 -0
  68. package/lib/engine/process-identity.mjs +164 -0
  69. package/lib/engine/process-tree.mjs +551 -0
  70. package/lib/engine/prompt.mjs +60 -0
  71. package/lib/engine/session/store.mjs +299 -0
  72. package/lib/engine/session-runtime/args.mjs +97 -0
  73. package/lib/engine/session-runtime/host.mjs +143 -0
  74. package/lib/engine/session-runtime/inbox.mjs +122 -0
  75. package/lib/engine/session-runtime/notifications.mjs +129 -0
  76. package/lib/engine/session-runtime/registry.mjs +328 -0
  77. package/lib/engine/session-runtime/runner.mjs +344 -0
  78. package/lib/engine/session-runtime/socket.mjs +212 -0
  79. package/lib/engine/session-runtime/wakeup.mjs +115 -0
  80. package/lib/engine/skills/index.mjs +321 -0
  81. package/lib/engine/tools/bash-background.mjs +533 -0
  82. package/lib/engine/tools/bash.mjs +216 -0
  83. package/lib/engine/tools/edit.mjs +97 -0
  84. package/lib/engine/tools/glob.mjs +81 -0
  85. package/lib/engine/tools/grep.mjs +224 -0
  86. package/lib/engine/tools/index.mjs +84 -0
  87. package/lib/engine/tools/list-agents.mjs +32 -0
  88. package/lib/engine/tools/ls.mjs +127 -0
  89. package/lib/engine/tools/monitor.mjs +82 -0
  90. package/lib/engine/tools/notebook-edit.mjs +218 -0
  91. package/lib/engine/tools/read.mjs +103 -0
  92. package/lib/engine/tools/schedule-wakeup.mjs +45 -0
  93. package/lib/engine/tools/schema.mjs +144 -0
  94. package/lib/engine/tools/send-message.mjs +77 -0
  95. package/lib/engine/tools/session.mjs +70 -0
  96. package/lib/engine/tools/todo.mjs +144 -0
  97. package/lib/engine/tools/toolsearch.mjs +217 -0
  98. package/lib/engine/tools/walk.mjs +193 -0
  99. package/lib/engine/tools/web-switch.mjs +31 -0
  100. package/lib/engine/tools/webfetch-html.mjs +387 -0
  101. package/lib/engine/tools/webfetch-net.mjs +340 -0
  102. package/lib/engine/tools/webfetch.mjs +198 -0
  103. package/lib/engine/tools/websearch.mjs +91 -0
  104. package/lib/engine/tools/workflow.mjs +95 -0
  105. package/lib/engine/tools/write.mjs +76 -0
  106. package/lib/engine/tui/line-editor.mjs +137 -0
  107. package/lib/engine/tui/render.mjs +86 -0
  108. package/lib/engine/tui/tui.mjs +274 -0
  109. package/lib/engine/wire/anthropic-messages.mjs +263 -0
  110. package/lib/engine/wire/effort.mjs +36 -0
  111. package/lib/engine/wire/errors.mjs +496 -0
  112. package/lib/engine/wire/http.mjs +441 -0
  113. package/lib/engine/wire/index.mjs +76 -0
  114. package/lib/engine/wire/openai-chat.mjs +332 -0
  115. package/lib/engine/wire/prompt-cache.mjs +79 -0
  116. package/lib/engine/wire/search.mjs +140 -0
  117. package/lib/engine/wire/sse.mjs +114 -0
  118. package/lib/engine/wire/stall.mjs +349 -0
  119. package/lib/engine/wire/token-provider.mjs +175 -0
  120. package/lib/engine/wire/usage.mjs +192 -0
  121. package/lib/engine/workflow/host.mjs +524 -0
  122. package/lib/engine/workflow/journal.mjs +188 -0
  123. package/lib/engine/workflow/json-schema.mjs +171 -0
  124. package/lib/engine/workflow/meta.mjs +329 -0
  125. package/lib/engine/workflow/notifications.mjs +52 -0
  126. package/lib/engine/workflow/runtime.mjs +447 -0
  127. package/lib/engine/workflow/sandbox.mjs +534 -0
  128. package/lib/engine/workflow/worker.mjs +141 -0
  129. package/lib/engine/workflow/worktree.mjs +74 -0
  130. package/lib/execution/disposition.mjs +1 -1
  131. package/lib/execution/intake.mjs +10 -0
  132. package/lib/execution/surface-policy.mjs +15 -0
  133. package/lib/learning/curator.mjs +8 -6
  134. package/lib/learning/reflect.mjs +8 -6
  135. package/lib/model-router/catalog/cohort.yaml +137 -0
  136. package/lib/model-router/catalog.mjs +118 -1
  137. package/lib/model-router/failover.mjs +67 -16
  138. package/lib/model-router/llm-task.mjs +39 -3
  139. package/lib/model-router/resolve.mjs +89 -3
  140. package/lib/model-router/spawn.mjs +46 -47
  141. package/lib/model-router/taxonomy.mjs +126 -4
  142. package/lib/org/cost-sync.mjs +141 -11
  143. package/lib/org/inbound/broadcast.mjs +289 -0
  144. package/lib/org/inbound/collective.mjs +375 -0
  145. package/lib/org/inbound/directedness.mjs +96 -8
  146. package/lib/org/inbound/facts.mjs +78 -2
  147. package/lib/org/inbound/project.mjs +22 -0
  148. package/lib/org/inbound/surfaces.mjs +14 -0
  149. package/lib/org/llm-token.mjs +879 -0
  150. package/lib/org/mesh.mjs +61 -0
  151. package/lib/org/messaging.mjs +3 -1
  152. package/lib/org/protocol.checksum +1 -1
  153. package/lib/org/protocol.mjs +15 -0
  154. package/lib/org/quota.mjs +520 -0
  155. package/lib/org/tool-surface.mjs +104 -16
  156. package/lib/org/ui-parity.mjs +16 -1
  157. package/lib/org/work-ledger.mjs +37 -6
  158. package/lib/rate-guard.mjs +114 -1
  159. package/lib/resource-governor.mjs +41 -6
  160. package/lib/runtime/adapter.mjs +833 -0
  161. package/lib/runtime/child-env.mjs +191 -0
  162. package/lib/runtime/legacy-shell-guard.mjs +97 -0
  163. package/lib/runtime/seat-engine.mjs +162 -0
  164. package/lib/session/ask-ledger.mjs +271 -0
  165. package/lib/session/current-work.mjs +676 -0
  166. package/lib/session/feed-core.mjs +40 -3
  167. package/lib/session/launch-args.mjs +56 -4
  168. package/lib/session/status-summary.mjs +26 -9
  169. package/lib/session/upgrade-notice.mjs +42 -0
  170. package/lib/setup/claude-probe.mjs +117 -13
  171. package/lib/setup/enrich.mjs +13 -10
  172. package/lib/setup/sections/model.mjs +39 -13
  173. package/lib/telemetry/collect.mjs +229 -11
  174. package/lib/upgrade/ignored-drift.mjs +105 -0
  175. package/lib/voice/post-call-brief.mjs +30 -17
  176. package/package.json +13 -3
  177. package/plugins/maestro-skills/skills/board-work.md +5 -0
  178. package/plugins/maestro-skills/skills/inbound-triage.md +56 -15
  179. package/plugins/maestro-skills/skills/main-session.md +18 -7
  180. package/scaffold/config/collective.yaml +7 -0
  181. package/scripts/ci/check-durable-write-seam.mjs +3 -1
  182. package/scripts/ci/check-tarball-fidelity.mjs +126 -2
  183. package/scripts/ci/run-tests.mjs +47 -19
  184. package/scripts/cohort-llm/api-key-helper.mjs +92 -0
  185. package/scripts/collective/hook-runner.mjs +142 -19
  186. package/scripts/continuous-monitor.sh +13 -0
  187. package/scripts/cost/track-claude-usage.mjs +15 -0
  188. package/scripts/daemon/agent-daemon.mjs +408 -20
  189. package/scripts/daemon/assurance.mjs +48 -12
  190. package/scripts/daemon/cadence-consumer.mjs +218 -68
  191. package/scripts/daemon/cadence-handlers.mjs +73 -4
  192. package/scripts/daemon/classifier.mjs +75 -26
  193. package/scripts/daemon/context-compiler.mjs +51 -37
  194. package/scripts/daemon/deliver.mjs +30 -1
  195. package/scripts/daemon/dispatcher.mjs +595 -149
  196. package/scripts/daemon/health.mjs +14 -1
  197. package/scripts/daemon/maestro-daemon.mjs +11 -0
  198. package/scripts/daemon/prompt-builder.mjs +24 -0
  199. package/scripts/daemon/responder.mjs +246 -79
  200. package/scripts/daemon/sdk-version.mjs +98 -16
  201. package/scripts/eval/probe-gateway.mjs +635 -0
  202. package/scripts/eval/replay/extract.mjs +270 -0
  203. package/scripts/eval/replay/grade.mjs +260 -0
  204. package/scripts/eval/replay/lib/config.mjs +50 -0
  205. package/scripts/eval/replay/lib/effects.mjs +65 -0
  206. package/scripts/eval/replay/lib/fixture.mjs +188 -0
  207. package/scripts/eval/replay/lib/judge.mjs +72 -0
  208. package/scripts/eval/replay/lib/redact.mjs +136 -0
  209. package/scripts/eval/replay/lib/sandbox.mjs +170 -0
  210. package/scripts/eval/replay/lib/schema-check.mjs +63 -0
  211. package/scripts/eval/replay/lib/transcript.mjs +76 -0
  212. package/scripts/eval/replay/mcp-replay-stub.mjs +101 -0
  213. package/scripts/eval/replay/report.mjs +185 -0
  214. package/scripts/eval/replay/run.mjs +404 -0
  215. package/scripts/fleet/rollout.mjs +1151 -0
  216. package/scripts/hooks/pre-send-audit.sh +36 -245
  217. package/scripts/hooks/pre-write-yaml-validate.mjs +275 -0
  218. package/scripts/hooks/validate-state-yaml.sh +190 -0
  219. package/scripts/huddle/huddle-llm.mjs +361 -0
  220. package/scripts/huddle/huddle-server.mjs +46 -121
  221. package/scripts/local-triggers/autoupdate.sh +465 -81
  222. package/scripts/local-triggers/run-trigger.sh +13 -0
  223. package/scripts/maintenance/pin-integrity.mjs +364 -0
  224. package/scripts/poll-slack-events.sh +41 -9
  225. package/scripts/poller/slack-socket-mode.mjs +28 -3
  226. package/scripts/session/supervisor.mjs +80 -13
  227. package/scripts/spawn-session.sh +13 -0
  228. package/bin/maestro.test.mjs +0 -1574
  229. package/lib/action-executor.test.mjs +0 -871
  230. package/lib/archetype.test.mjs +0 -132
  231. package/lib/assurance/plan-note.test.mjs +0 -234
  232. package/lib/assurance/room-budget.test.mjs +0 -486
  233. package/lib/assurance/tier.test.mjs +0 -174
  234. package/lib/autonomy.test.mjs +0 -66
  235. package/lib/backlog.test.mjs +0 -302
  236. package/lib/backup/policy.test.mjs +0 -305
  237. package/lib/budget-escalate.test.mjs +0 -232
  238. package/lib/budget-guard.envelope.test.mjs +0 -476
  239. package/lib/budget-guard.test.mjs +0 -427
  240. package/lib/cadence-bus-requeue.test.mjs +0 -83
  241. package/lib/cadence-bus-schedule.test.mjs +0 -194
  242. package/lib/cadence-bus.test.mjs +0 -720
  243. package/lib/cadences.test.mjs +0 -230
  244. package/lib/capability/inventory.test.mjs +0 -232
  245. package/lib/capability.test.mjs +0 -78
  246. package/lib/channels/base-adapter.test.mjs +0 -590
  247. package/lib/channels/channels.test.mjs +0 -371
  248. package/lib/channels/contract.test.mjs +0 -162
  249. package/lib/channels/inbox-item.test.mjs +0 -368
  250. package/lib/channels/orgmail/adapter.test.mjs +0 -448
  251. package/lib/channels/pairing.test.mjs +0 -270
  252. package/lib/channels/repeat-suppressor.test.mjs +0 -134
  253. package/lib/channels/slack-adapter.test.mjs +0 -212
  254. package/lib/channels/telegram-adapter.test.mjs +0 -306
  255. package/lib/channels/voice/adapter.test.mjs +0 -278
  256. package/lib/channels/whatsapp/adapter-baileys.test.mjs +0 -359
  257. package/lib/channels/whatsapp/baileys-typing.test.mjs +0 -154
  258. package/lib/charter.test.mjs +0 -89
  259. package/lib/claude-bin.test.mjs +0 -131
  260. package/lib/cli/board.test.mjs +0 -227
  261. package/lib/cli/design.test.mjs +0 -270
  262. package/lib/cli/doctor-checks.test.mjs +0 -336
  263. package/lib/cli/global-setup-extras.test.mjs +0 -462
  264. package/lib/cli/inbox.test.mjs +0 -230
  265. package/lib/cli/session-ack.test.mjs +0 -63
  266. package/lib/cli/session.test.mjs +0 -613
  267. package/lib/collective/capture.test.mjs +0 -121
  268. package/lib/collective/cards.test.mjs +0 -114
  269. package/lib/collective/config.test.mjs +0 -123
  270. package/lib/collective/global-config.test.mjs +0 -220
  271. package/lib/collective/global-skills.test.mjs +0 -126
  272. package/lib/collective/presence.test.mjs +0 -95
  273. package/lib/collective/recall.test.mjs +0 -116
  274. package/lib/collective/vendor-skills.test.mjs +0 -306
  275. package/lib/comms/send-gate.test.mjs +0 -770
  276. package/lib/comms.test.mjs +0 -41
  277. package/lib/context/budget.test.mjs +0 -252
  278. package/lib/context/history-scope.test.mjs +0 -79
  279. package/lib/cost/ledger-row.test.mjs +0 -183
  280. package/lib/design/design-md.test.mjs +0 -318
  281. package/lib/design/fixtures/DESIGN.golden.md +0 -238
  282. package/lib/design/fixtures/PRODUCT.golden.md +0 -67
  283. package/lib/design/fixtures/foundation.json +0 -133
  284. package/lib/design/refresh-gate.test.mjs +0 -144
  285. package/lib/design/write.test.mjs +0 -241
  286. package/lib/diagnostics/alerts.test.mjs +0 -318
  287. package/lib/diagnostics/backup-freshness.test.mjs +0 -185
  288. package/lib/diagnostics/counters.test.mjs +0 -206
  289. package/lib/diagnostics/events.test.mjs +0 -290
  290. package/lib/diagnostics/otel.test.mjs +0 -196
  291. package/lib/diagnostics/trace.test.mjs +0 -251
  292. package/lib/env-compat.test.mjs +0 -104
  293. package/lib/execution/disposition.test.mjs +0 -553
  294. package/lib/execution/drive.test.mjs +0 -270
  295. package/lib/execution/effects.test.mjs +0 -344
  296. package/lib/execution/intake.test.mjs +0 -389
  297. package/lib/execution/journal.test.mjs +0 -261
  298. package/lib/execution/match.test.mjs +0 -235
  299. package/lib/execution/pipeline.test.mjs +0 -392
  300. package/lib/execution/route.test.mjs +0 -186
  301. package/lib/execution/surface-policy.test.mjs +0 -162
  302. package/lib/fs-atomic.test.mjs +0 -72
  303. package/lib/fs-ownership.test.mjs +0 -158
  304. package/lib/goals/admission.test.mjs +0 -164
  305. package/lib/goals/classify.test.mjs +0 -167
  306. package/lib/goals/collaborate.test.mjs +0 -336
  307. package/lib/goals/gaps.test.mjs +0 -284
  308. package/lib/goals/loop.test.mjs +0 -845
  309. package/lib/hooks/bus.test.mjs +0 -387
  310. package/lib/identity/persona.test.mjs +0 -142
  311. package/lib/kpi-sensors.test.mjs +0 -278
  312. package/lib/kpi.test.mjs +0 -244
  313. package/lib/learning/config.test.mjs +0 -75
  314. package/lib/learning/counters.test.mjs +0 -69
  315. package/lib/learning/curator-consolidate.test.mjs +0 -238
  316. package/lib/learning/curator.test.mjs +0 -106
  317. package/lib/learning/reflect.test.mjs +0 -0
  318. package/lib/learning/session-index.test.mjs +0 -125
  319. package/lib/learning/skill-writer.test.mjs +0 -210
  320. package/lib/mandate/audit.test.mjs +0 -195
  321. package/lib/mandate/contract.test.mjs +0 -185
  322. package/lib/mandate/derive.test.mjs +0 -274
  323. package/lib/mandate/model.test.mjs +0 -164
  324. package/lib/mandate/refresh.test.mjs +0 -389
  325. package/lib/mcp/server.test.mjs +0 -426
  326. package/lib/model-router/auth-profiles.test.mjs +0 -580
  327. package/lib/model-router/catalog.test.mjs +0 -385
  328. package/lib/model-router/economics.test.mjs +0 -438
  329. package/lib/model-router/failover.test.mjs +0 -439
  330. package/lib/model-router/health.test.mjs +0 -338
  331. package/lib/model-router/integration-coverage.test.mjs +0 -831
  332. package/lib/model-router/integration.test.mjs +0 -564
  333. package/lib/model-router/ledger.test.mjs +0 -415
  334. package/lib/model-router/llm-task.test.mjs +0 -392
  335. package/lib/model-router/org-credentials.test.mjs +0 -265
  336. package/lib/model-router/pricing-refresh.test.mjs +0 -286
  337. package/lib/model-router/reconcile.test.mjs +0 -316
  338. package/lib/model-router/repair.test.mjs +0 -180
  339. package/lib/model-router/spawn.test.mjs +0 -446
  340. package/lib/model-router/taxonomy.test.mjs +0 -410
  341. package/lib/model-router.test.mjs +0 -1207
  342. package/lib/org/activity.test.mjs +0 -134
  343. package/lib/org/approvals.test.mjs +0 -216
  344. package/lib/org/awareness.test.mjs +0 -159
  345. package/lib/org/board-mine-cache.test.mjs +0 -53
  346. package/lib/org/board.test.mjs +0 -187
  347. package/lib/org/bootstrap-context.test.mjs +0 -153
  348. package/lib/org/client.test.mjs +0 -1206
  349. package/lib/org/cohort-client.test.mjs +0 -126
  350. package/lib/org/cost-sync.test.mjs +0 -153
  351. package/lib/org/doctor.test.mjs +0 -346
  352. package/lib/org/engagement-ledger.test.mjs +0 -112
  353. package/lib/org/engagement.test.mjs +0 -739
  354. package/lib/org/handoff.test.mjs +0 -269
  355. package/lib/org/inbound/directedness.test.mjs +0 -668
  356. package/lib/org/inbound/facts.test.mjs +0 -471
  357. package/lib/org/inbound/hydrate.test.mjs +0 -908
  358. package/lib/org/inbound/index.test.mjs +0 -429
  359. package/lib/org/inbound/project.test.mjs +0 -287
  360. package/lib/org/integration-tools.test.mjs +0 -160
  361. package/lib/org/keys.test.mjs +0 -92
  362. package/lib/org/knowledge.test.mjs +0 -326
  363. package/lib/org/leases.test.mjs +0 -235
  364. package/lib/org/mesh-directives.test.mjs +0 -110
  365. package/lib/org/mesh-integration.test.mjs +0 -127
  366. package/lib/org/mesh.test.mjs +0 -400
  367. package/lib/org/messaging.test.mjs +0 -471
  368. package/lib/org/param-contract.test.mjs +0 -477
  369. package/lib/org/policy.test.mjs +0 -237
  370. package/lib/org/protocol.checksum.test.mjs +0 -90
  371. package/lib/org/protocol.test.mjs +0 -323
  372. package/lib/org/push.test.mjs +0 -792
  373. package/lib/org/registry.test.mjs +0 -100
  374. package/lib/org/resource-tools.test.mjs +0 -361
  375. package/lib/org/tool-access.test.mjs +0 -144
  376. package/lib/org/tool-surface-integration.test.mjs +0 -120
  377. package/lib/org/tool-surface.test.mjs +0 -1268
  378. package/lib/org/typing.test.mjs +0 -291
  379. package/lib/org/ui-parity.test.mjs +0 -560
  380. package/lib/org/verify.test.mjs +0 -194
  381. package/lib/org/work-ledger.test.mjs +0 -273
  382. package/lib/plan/adoption-e2e.test.mjs +0 -366
  383. package/lib/plan/budget-enforcement.test.mjs +0 -400
  384. package/lib/plan/compile.test.mjs +0 -382
  385. package/lib/plan/emit.test.mjs +0 -269
  386. package/lib/plan/explain.test.mjs +0 -188
  387. package/lib/prompts/parallelism.test.mjs +0 -177
  388. package/lib/rag/rag.test.mjs +0 -505
  389. package/lib/rate-guard.test.mjs +0 -272
  390. package/lib/reactive-gate.test.mjs +0 -57
  391. package/lib/render.test.mjs +0 -68
  392. package/lib/resource-governor.test.mjs +0 -488
  393. package/lib/scheduling/dynamic-jobs.test.mjs +0 -344
  394. package/lib/scheduling/jitter.test.mjs +0 -140
  395. package/lib/secrets/broker.test.mjs +0 -280
  396. package/lib/secrets/providers.test.mjs +0 -274
  397. package/lib/security/audit-engine.test.mjs +0 -424
  398. package/lib/security/coerce-args.test.mjs +0 -281
  399. package/lib/security/dangerous-tools.test.mjs +0 -68
  400. package/lib/security/external-content.test.mjs +0 -84
  401. package/lib/security/redact.test.mjs +0 -441
  402. package/lib/security/secret-equal.test.mjs +0 -55
  403. package/lib/session/config.test.mjs +0 -92
  404. package/lib/session/feed-core.test.mjs +0 -198
  405. package/lib/session/first-run.test.mjs +0 -121
  406. package/lib/session/frontdoor.test.mjs +0 -205
  407. package/lib/session/handoffs.test.mjs +0 -183
  408. package/lib/session/identity.test.mjs +0 -180
  409. package/lib/session/inbox-claims.test.mjs +0 -286
  410. package/lib/session/launch-args.test.mjs +0 -157
  411. package/lib/session/liveness.test.mjs +0 -100
  412. package/lib/session/status-summary.test.mjs +0 -118
  413. package/lib/session-permissions.test.mjs +0 -120
  414. package/lib/setup/claude-probe.test.mjs +0 -187
  415. package/lib/setup/completeness.test.mjs +0 -110
  416. package/lib/setup/context-pack.test.mjs +0 -89
  417. package/lib/setup/enrich.test.mjs +0 -115
  418. package/lib/setup/enroll-from-cohort.test.mjs +0 -300
  419. package/lib/setup/integration.test.mjs +0 -162
  420. package/lib/setup/io.test.mjs +0 -77
  421. package/lib/setup/runner.test.mjs +0 -132
  422. package/lib/setup/sections/identity.test.mjs +0 -234
  423. package/lib/setup/sections/inventory.test.mjs +0 -198
  424. package/lib/setup/sections/learning.test.mjs +0 -81
  425. package/lib/setup/sections/mandate.test.mjs +0 -388
  426. package/lib/setup/sections/messaging.test.mjs +0 -127
  427. package/lib/setup/sections/model.test.mjs +0 -240
  428. package/lib/setup/sections/org.test.mjs +0 -346
  429. package/lib/setup/sections/orgmail.test.mjs +0 -118
  430. package/lib/setup/sections/recovery.test.mjs +0 -98
  431. package/lib/setup/sections/subagents.test.mjs +0 -429
  432. package/lib/setup/sections/verify.test.mjs +0 -175
  433. package/lib/setup/sot.test.mjs +0 -81
  434. package/lib/setup/state.test.mjs +0 -115
  435. package/lib/singleton.test.mjs +0 -151
  436. package/lib/subagents/cli.test.mjs +0 -389
  437. package/lib/subagents/client.test.mjs +0 -309
  438. package/lib/subagents/gap.test.mjs +0 -234
  439. package/lib/subagents/lock.test.mjs +0 -248
  440. package/lib/subagents/manifest.test.mjs +0 -175
  441. package/lib/subagents/refs.test.mjs +0 -204
  442. package/lib/subagents/resolve.test.mjs +0 -422
  443. package/lib/subagents/schema.test.mjs +0 -328
  444. package/lib/telemetry/alerts.test.mjs +0 -109
  445. package/lib/telemetry/collect.test.mjs +0 -1274
  446. package/lib/tool-definitions-integration.test.mjs +0 -83
  447. package/lib/tool-definitions.test.mjs +0 -437
  448. package/lib/upgrade/global-refresh.test.mjs +0 -65
  449. package/lib/upgrade/launchd-reconcile.test.mjs +0 -272
  450. package/lib/upgrade/post-steps.test.mjs +0 -200
  451. package/lib/upgrade/verify.test.mjs +0 -164
  452. package/lib/util/fetch-timeout.test.mjs +0 -202
  453. package/lib/util/reconnect.test.mjs +0 -369
  454. package/lib/util/unhandled.test.mjs +0 -216
  455. package/lib/voice/outbound.test.mjs +0 -69
  456. package/lib/voice/session-rotation.test.mjs +0 -114
  457. package/lib/voice/stt.test.mjs +0 -226
  458. package/lib/voice/voice.test.mjs +0 -990
  459. package/scripts/cadence/enqueue-cadence-tick.test.mjs +0 -187
  460. package/scripts/ci/check-docs-accuracy.test.mjs +0 -409
  461. package/scripts/ci/check-durable-write-seam.test.mjs +0 -90
  462. package/scripts/ci/check-no-build-artifacts.test.mjs +0 -71
  463. package/scripts/ci/check-no-residual-identity.test.mjs +0 -202
  464. package/scripts/ci/check-skill-packs.test.mjs +0 -495
  465. package/scripts/ci/check-subagent-frontmatter.test.mjs +0 -124
  466. package/scripts/ci/check.test.mjs +0 -194
  467. package/scripts/ci/conformance-org-api.test.mjs +0 -425
  468. package/scripts/cloud-relay/voice/relay-identity.test.mjs +0 -96
  469. package/scripts/collective/hook-runner.test.mjs +0 -173
  470. package/scripts/cost/fleet-digest.test.mjs +0 -207
  471. package/scripts/cost/track-claude-usage-pricing.test.mjs +0 -183
  472. package/scripts/cost/track-claude-usage.test.mjs +0 -148
  473. package/scripts/daemon/agent-daemon-board-mine.test.mjs +0 -96
  474. package/scripts/daemon/agent-daemon-design.test.mjs +0 -238
  475. package/scripts/daemon/agent-daemon-frontdoor.test.mjs +0 -60
  476. package/scripts/daemon/agent-daemon.test.mjs +0 -995
  477. package/scripts/daemon/assurance-e2e.test.mjs +0 -613
  478. package/scripts/daemon/assurance.test.mjs +0 -1791
  479. package/scripts/daemon/board-mirror.test.mjs +0 -165
  480. package/scripts/daemon/cadence-consumer-frontdoor.test.mjs +0 -393
  481. package/scripts/daemon/cadence-consumer-governance.test.mjs +0 -276
  482. package/scripts/daemon/cadence-consumer.test.mjs +0 -776
  483. package/scripts/daemon/cadence-handlers.test.mjs +0 -837
  484. package/scripts/daemon/classifier-identity.test.mjs +0 -137
  485. package/scripts/daemon/classifier.test.mjs +0 -266
  486. package/scripts/daemon/classify-kind.test.mjs +0 -40
  487. package/scripts/daemon/context-compiler.test.mjs +0 -406
  488. package/scripts/daemon/deliver.test.mjs +0 -564
  489. package/scripts/daemon/dispatcher-cooldown.test.mjs +0 -122
  490. package/scripts/daemon/dispatcher-governance.test.mjs +0 -1013
  491. package/scripts/daemon/dispatcher-resume.test.mjs +0 -166
  492. package/scripts/daemon/dispatcher-session-continuity.test.mjs +0 -365
  493. package/scripts/daemon/execution-ladder.test.mjs +0 -470
  494. package/scripts/daemon/goal-steward-cadence.test.mjs +0 -312
  495. package/scripts/daemon/inbox-deferral-session.test.mjs +0 -49
  496. package/scripts/daemon/inbox-deferral.test.mjs +0 -336
  497. package/scripts/daemon/inbox-wake.test.mjs +0 -199
  498. package/scripts/daemon/integration.test.mjs +0 -149
  499. package/scripts/daemon/lib/self-echo.test.mjs +0 -153
  500. package/scripts/daemon/lib/session-router.test.mjs +0 -554
  501. package/scripts/daemon/prompt-builder-preamble.test.mjs +0 -210
  502. package/scripts/daemon/prompt-builder.test.mjs +0 -556
  503. package/scripts/daemon/responder-cost.test.mjs +0 -68
  504. package/scripts/daemon/responder-history.test.mjs +0 -221
  505. package/scripts/daemon/sdk-version.test.mjs +0 -31
  506. package/scripts/daemon/session-lock.test.mjs +0 -252
  507. package/scripts/daemon/session-outcomes.test.mjs +0 -533
  508. package/scripts/daemon/typing-registry.test.mjs +0 -102
  509. package/scripts/hooks/pre-send-audit.test.mjs +0 -354
  510. package/scripts/huddle/huddle-prompt.test.mjs +0 -176
  511. package/scripts/local-triggers/autoupdate.test.mjs +0 -518
  512. package/scripts/local-triggers/generate-plists.test.mjs +0 -456
  513. package/scripts/media-generation/brand-clause.test.mjs +0 -135
  514. package/scripts/org/send-orgmail.first-contact.test.mjs +0 -102
  515. package/scripts/poller/inbox-privilege-injection.test.mjs +0 -167
  516. package/scripts/poller/inbox-scan-poller.test.mjs +0 -295
  517. package/scripts/poller/lib/cloud-relay-dedup.test.mjs +0 -133
  518. package/scripts/poller/slack-socket-mode.test.mjs +0 -805
  519. package/scripts/poller-launchd/install.test.mjs +0 -243
  520. package/scripts/restore-from-backup.test.mjs +0 -181
  521. package/scripts/session/feed.test.mjs +0 -196
  522. package/scripts/session/supervisor-sh.test.mjs +0 -218
  523. package/scripts/session/supervisor.test.mjs +0 -482
  524. package/scripts/setup/configure-macos.test.mjs +0 -306
  525. package/scripts/setup/gen-subagent-manifest.test.mjs +0 -124
  526. package/scripts/setup/generate-agent-package-json.test.mjs +0 -143
  527. package/scripts/setup/generate-capability.test.mjs +0 -134
  528. package/scripts/setup/init-agent.test.mjs +0 -370
  529. package/scripts/setup/init-skill-marketplace.test.mjs +0 -193
  530. package/scripts/vendor/sync-skill-packs.test.mjs +0 -103
  531. package/scripts/watchdog/memory-watchdog.test.mjs +0 -64
@@ -1,286 +0,0 @@
1
- /**
2
- * pricing-refresh.test.mjs — the OpenClaw external pricing oracle.
3
- *
4
- * Verifies: refreshPricing merges an injected LiteLLM + OpenRouter response into
5
- * the cost-only cache ($/MTok, provenance-stamped); a partial failure backfills
6
- * from the existing cache; a fully-offline refresh fails open (cache untouched);
7
- * the written cache is the catalog-consumable flat shape; refresh writes COST
8
- * fields only; an optional refs allowlist narrows the write.
9
- *
10
- * Injected fetchImpl — no network. Run: node --test lib/model-router/pricing-refresh.test.mjs
11
- */
12
- "use strict";
13
-
14
- import { test } from "node:test";
15
- import assert from "node:assert/strict";
16
- import { mkdtempSync, rmSync, existsSync, readFileSync } from "node:fs";
17
- import { tmpdir } from "node:os";
18
- import { join } from "node:path";
19
-
20
- import {
21
- refreshPricing,
22
- parseLiteLLM,
23
- parseOpenRouter,
24
- _internals,
25
- } from "./pricing-refresh.mjs";
26
- import { loadCatalog, _internals as catInternals } from "./catalog.mjs";
27
-
28
- function tmpStateDir() {
29
- return mkdtempSync(join(tmpdir(), "pricing-"));
30
- }
31
-
32
- // A LiteLLM JSON ($/token). deepseek-chat: input 0.14/MTok, output 0.28/MTok.
33
- const LITELLM_JSON = {
34
- sample_spec: { note: "ignored" },
35
- "deepseek/deepseek-chat": {
36
- litellm_provider: "deepseek",
37
- input_cost_per_token: 0.00000014,
38
- output_cost_per_token: 0.00000028,
39
- cache_read_input_token_cost: 0.000000028,
40
- },
41
- "claude-sonnet-4-6": {
42
- litellm_provider: "anthropic",
43
- input_cost_per_token: 0.000003,
44
- output_cost_per_token: 0.000015,
45
- },
46
- };
47
-
48
- // An OpenRouter /models response ($/token strings).
49
- const OPENROUTER_JSON = {
50
- data: [
51
- {
52
- id: "deepseek/deepseek-chat",
53
- pricing: { prompt: "0.00000020", completion: "0.00000030" },
54
- },
55
- {
56
- id: "moonshotai/kimi-k2.6",
57
- pricing: { prompt: "0.0000006", completion: "0.0000025" },
58
- },
59
- ],
60
- };
61
-
62
- /** A fetch that serves LiteLLM/OpenRouter bodies by URL; `fail` set forces errors. */
63
- function fakeFetch({ litellm = LITELLM_JSON, openrouter = OPENROUTER_JSON, fail = new Set() } = {}) {
64
- const calls = [];
65
- const impl = async (url) => {
66
- calls.push(url);
67
- const isLiteLLM = url.includes("litellm") || url.includes("model_prices");
68
- const which = isLiteLLM ? "litellm" : "openrouter";
69
- if (fail.has(which)) {
70
- return { ok: false, status: 503, text: async () => "" };
71
- }
72
- const body = isLiteLLM ? litellm : openrouter;
73
- return { ok: true, status: 200, text: async () => JSON.stringify(body) };
74
- };
75
- impl.calls = calls;
76
- return impl;
77
- }
78
-
79
- // ---------------------------------------------------------------------------
80
- // Parsers
81
- // ---------------------------------------------------------------------------
82
-
83
- test("parseLiteLLM normalizes $/token → $/MTok and resolves provider/model refs", () => {
84
- const m = parseLiteLLM(LITELLM_JSON);
85
- const ds = m.get("deepseek/deepseek-chat");
86
- assert.ok(ds);
87
- assert.equal(ds.input, 0.14);
88
- assert.equal(ds.output, 0.28);
89
- assert.equal(ds.cache_read, 0.028);
90
- // provider-less key resolves via litellm_provider.
91
- assert.ok(m.get("anthropic/claude-sonnet-4-6"));
92
- // sample_spec is skipped.
93
- assert.equal(m.has("sample_spec"), false);
94
- });
95
-
96
- test("parseOpenRouter normalizes pricing strings and canonicalizes the provider", () => {
97
- const m = parseOpenRouter(OPENROUTER_JSON);
98
- const ds = m.get("deepseek/deepseek-chat");
99
- assert.equal(ds.input, 0.2);
100
- assert.equal(ds.output, 0.3);
101
- // moonshotai → moonshot (catalog provider key).
102
- assert.ok(m.get("moonshot/kimi-k2.6"));
103
- });
104
-
105
- // ---------------------------------------------------------------------------
106
- // refreshPricing — merge
107
- // ---------------------------------------------------------------------------
108
-
109
- test("refreshPricing merges both sources into the cost-only cache (LiteLLM wins)", async () => {
110
- const dir = tmpStateDir();
111
- try {
112
- const res = await refreshPricing({
113
- fetchImpl: fakeFetch(),
114
- stateDir: dir,
115
- now: () => Date.parse("2026-06-12T00:00:00Z"),
116
- });
117
- assert.equal(res.ok, true);
118
- assert.ok(res.sources.includes("litellm"));
119
- assert.ok(res.sources.includes("openrouter"));
120
-
121
- // deepseek-chat is in BOTH → LiteLLM wins field-wise, provenance "litellm+openrouter".
122
- const ds = res.cache["deepseek/deepseek-chat"];
123
- assert.equal(ds.input, 0.14); // LiteLLM, not OpenRouter's 0.2
124
- assert.equal(ds.output, 0.28);
125
- assert.equal(ds.cache_read, 0.028);
126
- assert.equal(ds.provenance.source, "litellm+openrouter");
127
- assert.equal(ds.provenance.volatile, true);
128
- assert.equal(ds.provenance.fetched, "2026-06-12T00:00:00.000Z");
129
-
130
- // moonshot only from OpenRouter → provenance "openrouter".
131
- assert.equal(res.cache["moonshot/kimi-k2.6"].provenance.source, "openrouter");
132
-
133
- // The cache file is written in the catalog-consumable { models: {...} } shape.
134
- const p = join(dir, "pricing-cache.json");
135
- assert.equal(existsSync(p), true);
136
- const onDisk = JSON.parse(readFileSync(p, "utf-8"));
137
- assert.ok(onDisk.models);
138
- assert.ok(onDisk.models["deepseek/deepseek-chat"]);
139
- } finally {
140
- rmSync(dir, { recursive: true, force: true });
141
- }
142
- });
143
-
144
- test("refreshPricing writes COST fields only (no capability/context fields)", async () => {
145
- const dir = tmpStateDir();
146
- try {
147
- const res = await refreshPricing({ fetchImpl: fakeFetch(), stateDir: dir });
148
- const ds = res.cache["deepseek/deepseek-chat"];
149
- const allowed = new Set(["input", "output", "cache_read", "cache_write", "tiered", "provenance"]);
150
- for (const k of Object.keys(ds)) {
151
- assert.ok(allowed.has(k), `unexpected field "${k}" in pricing entry`);
152
- }
153
- } finally {
154
- rmSync(dir, { recursive: true, force: true });
155
- }
156
- });
157
-
158
- // ---------------------------------------------------------------------------
159
- // refreshPricing — fail-open + backfill
160
- // ---------------------------------------------------------------------------
161
-
162
- test("refreshPricing fails open when BOTH sources are offline (cache untouched)", async () => {
163
- const dir = tmpStateDir();
164
- try {
165
- const existing = {
166
- "deepseek/deepseek-chat": {
167
- input: 0.99,
168
- output: 1.99,
169
- provenance: { source: "litellm", fetched: "2026-06-01T00:00:00.000Z", volatile: true },
170
- },
171
- };
172
- const res = await refreshPricing({
173
- fetchImpl: fakeFetch({ fail: new Set(["litellm", "openrouter"]) }),
174
- stateDir: dir,
175
- existingCache: { models: existing },
176
- });
177
- assert.equal(res.ok, false);
178
- assert.equal(res.written, 0);
179
- assert.equal(res.path, null); // nothing written
180
- // The returned cache is the untouched existing one.
181
- assert.equal(res.cache["deepseek/deepseek-chat"].input, 0.99);
182
- // No file written either.
183
- assert.equal(existsSync(join(dir, "pricing-cache.json")), false);
184
- } finally {
185
- rmSync(dir, { recursive: true, force: true });
186
- }
187
- });
188
-
189
- test("refreshPricing backfills from the existing cache on a partial failure", async () => {
190
- const dir = tmpStateDir();
191
- try {
192
- // OpenRouter offline; only LiteLLM serves. An existing OpenRouter-only entry
193
- // (moonshot) must survive the merge (not be deleted).
194
- const existing = {
195
- "moonshot/kimi-k2.6": {
196
- input: 0.6,
197
- output: 2.5,
198
- provenance: { source: "openrouter", fetched: "2026-06-01T00:00:00.000Z", volatile: true },
199
- },
200
- };
201
- const res = await refreshPricing({
202
- fetchImpl: fakeFetch({ fail: new Set(["openrouter"]) }),
203
- stateDir: dir,
204
- existingCache: { models: existing },
205
- });
206
- assert.equal(res.ok, true);
207
- assert.deepEqual(res.sources, ["litellm"]);
208
- // Fresh LiteLLM price merged.
209
- assert.equal(res.cache["deepseek/deepseek-chat"].input, 0.14);
210
- assert.equal(res.cache["deepseek/deepseek-chat"].provenance.source, "litellm");
211
- // Backfilled OpenRouter entry survived.
212
- assert.equal(res.cache["moonshot/kimi-k2.6"].input, 0.6);
213
- // One error recorded for the offline source.
214
- assert.equal(res.errors.some((e) => e.source === "openrouter"), true);
215
- } finally {
216
- rmSync(dir, { recursive: true, force: true });
217
- }
218
- });
219
-
220
- test("refreshPricing with no fetchImpl fails open and never throws", async () => {
221
- const dir = tmpStateDir();
222
- try {
223
- const res = await refreshPricing({ stateDir: dir });
224
- assert.equal(res.ok, false);
225
- assert.equal(existsSync(join(dir, "pricing-cache.json")), false);
226
- } finally {
227
- rmSync(dir, { recursive: true, force: true });
228
- }
229
- });
230
-
231
- test("refreshPricing respects an optional refs allowlist", async () => {
232
- const dir = tmpStateDir();
233
- try {
234
- const res = await refreshPricing({
235
- fetchImpl: fakeFetch(),
236
- stateDir: dir,
237
- refs: ["deepseek/deepseek-chat"],
238
- });
239
- assert.equal(res.ok, true);
240
- assert.ok(res.cache["deepseek/deepseek-chat"]);
241
- assert.equal(res.cache["moonshot/kimi-k2.6"], undefined);
242
- assert.equal(res.cache["anthropic/claude-sonnet-4-6"], undefined);
243
- } finally {
244
- rmSync(dir, { recursive: true, force: true });
245
- }
246
- });
247
-
248
- test("a refreshed cache feeds the catalog's cost-only authority layer", () => {
249
- // The cache shape the refresh writes is exactly what loadCatalog's pricing-cache
250
- // layer consumes — patching cost fields on a bundled row without touching caps.
251
- const catalogText = {
252
- deepseek: [
253
- "provider: deepseek",
254
- "auth_env: DEEPSEEK_API_KEY",
255
- "endpoints: { anthropic: https://api.deepseek.com/anthropic }",
256
- "models:",
257
- " - id: deepseek-chat",
258
- " status: available",
259
- " tool_reliability: B",
260
- " context_window: 64000",
261
- " context_tokens: 60000",
262
- " cost: { input: 9.99, output: 9.99 }",
263
- ].join("\n"),
264
- };
265
- const pricingCache = {
266
- models: {
267
- "deepseek/deepseek-chat": {
268
- input: 0.14,
269
- output: 0.28,
270
- cache_read: 0.028,
271
- provenance: { source: "litellm+openrouter", fetched: "2026-06-12T00:00:00.000Z", volatile: true },
272
- },
273
- },
274
- };
275
- const result = loadCatalog(undefined, { catalogText, pricingCache, agentRoot: tmpdir() });
276
- const row = result.byRef.get("deepseek/deepseek-chat");
277
- assert.ok(row);
278
- // Cost overwritten by the cache; capability fields (context) untouched.
279
- assert.equal(row.cost.input, 0.14);
280
- assert.equal(row.cost.output, 0.28);
281
- assert.equal(row.cost.cache_read, 0.028);
282
- assert.equal(row.context_window, 64000);
283
- assert.equal(row.tool_reliability, "B");
284
- assert.equal(row.cost_provenance.source, "litellm+openrouter");
285
- assert.equal(row.cost_provenance.volatile, true);
286
- });
@@ -1,316 +0,0 @@
1
- /**
2
- * Tests for lib/model-router/reconcile.mjs — the nightly three-way spend
3
- * reconcile. The Anthropic Admin Cost API is NEVER called live; every test
4
- * injects a `fetchImpl`. Covers: local-ledger sum, provider-payload parsing
5
- * (both nested + flat shapes, string/object amounts), drift computation, the
6
- * >10% flag, the three-way (rollup) leg, and fail-open behavior.
7
- */
8
-
9
- import { test } from "node:test";
10
- import assert from "node:assert/strict";
11
- import { mkdtempSync, mkdirSync, writeFileSync } from "node:fs";
12
- import { tmpdir } from "node:os";
13
- import { join } from "node:path";
14
-
15
- import {
16
- reconcileDaily,
17
- sumLocalLedger,
18
- fetchProviderCost,
19
- computeDrift,
20
- DRIFT_THRESHOLD_PCT,
21
- _internals,
22
- } from "./reconcile.mjs";
23
-
24
- const DAY = "2026-06-10";
25
-
26
- // ---------------------------------------------------------------------------
27
- // Helpers: write a tmp ledger; build a fake Admin Cost API fetch
28
- // ---------------------------------------------------------------------------
29
-
30
- function tmpLedger(rows, day = DAY) {
31
- const dir = mkdtempSync(join(tmpdir(), "reconcile-"));
32
- const ledgerDir = join(dir, "state", "cost-tracking");
33
- mkdirSync(ledgerDir, { recursive: true });
34
- const file = join(ledgerDir, `${day}.jsonl`);
35
- writeFileSync(file, rows.map((r) => JSON.stringify(r)).join("\n") + "\n");
36
- return { ledgerDir, dir };
37
- }
38
-
39
- /** A fetchImpl that returns a Response-like object wrapping a Cost API payload. */
40
- function fakeFetch(payload, { ok = true, status = 200 } = {}) {
41
- let lastUrl = null;
42
- const f = async (url) => {
43
- lastUrl = url;
44
- return {
45
- ok,
46
- status,
47
- json: async () => payload,
48
- };
49
- };
50
- f.lastUrl = () => lastUrl;
51
- return f;
52
- }
53
-
54
- // Documented nested Cost API shape: data[].results[] with object amounts.
55
- function nestedPayload(items) {
56
- return {
57
- data: [
58
- {
59
- starting_at: `${DAY}T00:00:00Z`,
60
- results: items.map((it) => ({
61
- amount: { amount: String(it.usd), currency: "USD" },
62
- model: it.model,
63
- api_key_id: it.agent || null,
64
- })),
65
- },
66
- ],
67
- };
68
- }
69
-
70
- // ---------------------------------------------------------------------------
71
- // sumLocalLedger
72
- // ---------------------------------------------------------------------------
73
-
74
- test("sumLocalLedger: sums estimated_usd by model", () => {
75
- const { ledgerDir } = tmpLedger([
76
- { agent: "alex", model: "claude-sonnet-4-6", estimated_usd: 1.5 },
77
- { agent: "alex", model: "claude-sonnet-4-6", estimated_usd: 0.5 },
78
- { agent: "alex", model: "claude-haiku-4-5", estimated_usd: 0.25 },
79
- ]);
80
- const r = sumLocalLedger({ ledgerDir, day: DAY });
81
- assert.equal(r.totalUSD, 2.25);
82
- assert.equal(r.rows, 3);
83
- assert.equal(r.byModel["claude-sonnet-4-6"], 2);
84
- assert.equal(r.byModel["claude-haiku-4-5"], 0.25);
85
- });
86
-
87
- test("sumLocalLedger: falls back to total_cost_usd when estimated_usd absent", () => {
88
- const { ledgerDir } = tmpLedger([{ agent: "a", model: "m", total_cost_usd: 3.0 }]);
89
- const r = sumLocalLedger({ ledgerDir, day: DAY });
90
- assert.equal(r.totalUSD, 3.0);
91
- });
92
-
93
- test("sumLocalLedger: missing file → zeroes, never throws", () => {
94
- const r = sumLocalLedger({ ledgerDir: "/nonexistent/path", day: DAY });
95
- assert.equal(r.totalUSD, 0);
96
- assert.equal(r.rows, 0);
97
- });
98
-
99
- test("sumLocalLedger: skips corrupt lines", () => {
100
- const dir = mkdtempSync(join(tmpdir(), "reconcile-"));
101
- const ledgerDir = join(dir, "cost");
102
- mkdirSync(ledgerDir, { recursive: true });
103
- writeFileSync(
104
- join(ledgerDir, `${DAY}.jsonl`),
105
- `{"model":"m","estimated_usd":1.0}\nNOT JSON\n{"model":"m","estimated_usd":2.0}\n`
106
- );
107
- const r = sumLocalLedger({ ledgerDir, day: DAY });
108
- assert.equal(r.totalUSD, 3.0);
109
- assert.equal(r.rows, 2);
110
- });
111
-
112
- // ---------------------------------------------------------------------------
113
- // fetchProviderCost — injected fetch, payload parsing
114
- // ---------------------------------------------------------------------------
115
-
116
- test("fetchProviderCost: parses the nested Cost API shape", async () => {
117
- const fetchImpl = fakeFetch(
118
- nestedPayload([
119
- { usd: 5.0, model: "claude-sonnet-4-6", agent: "key_alex" },
120
- { usd: 1.0, model: "claude-haiku-4-5", agent: "key_alex" },
121
- ])
122
- );
123
- const r = await fetchProviderCost({ fetchImpl, day: DAY });
124
- assert.equal(r.ok, true);
125
- assert.equal(r.totalUSD, 6.0);
126
- assert.equal(r.byModel["claude-sonnet-4-6"], 5.0);
127
- assert.equal(r.byAgent["key_alex"], 6.0);
128
- });
129
-
130
- test("fetchProviderCost: parses a flat shape with bare-number amounts", async () => {
131
- const fetchImpl = fakeFetch({
132
- data: [
133
- { amount: 2.5, model: "claude-opus-4-8" },
134
- { amount: "1.25", model: "claude-haiku-4-5" },
135
- ],
136
- });
137
- const r = await fetchProviderCost({ fetchImpl, day: DAY });
138
- assert.equal(r.totalUSD, 3.75);
139
- });
140
-
141
- test("fetchProviderCost: builds a one-day [starting_at, ending_at) window URL", async () => {
142
- const fetchImpl = fakeFetch(nestedPayload([]));
143
- await fetchProviderCost({ fetchImpl, day: DAY, url: "https://example/cost" });
144
- const url = fetchImpl.lastUrl();
145
- assert.match(url, /starting_at=2026-06-10T00%3A00%3A00Z/);
146
- assert.match(url, /ending_at=2026-06-11T00%3A00%3A00Z/);
147
- assert.match(url, /group_by%5B%5D=model/);
148
- });
149
-
150
- test("fetchProviderCost: no fetchImpl → ok:false, never throws", async () => {
151
- const r = await fetchProviderCost({ day: DAY });
152
- assert.equal(r.ok, false);
153
- assert.equal(r.totalUSD, null);
154
- });
155
-
156
- test("fetchProviderCost: http error → ok:false with note", async () => {
157
- const fetchImpl = fakeFetch({}, { ok: false, status: 403 });
158
- const r = await fetchProviderCost({ fetchImpl, day: DAY });
159
- assert.equal(r.ok, false);
160
- assert.match(r.note, /403/);
161
- });
162
-
163
- test("fetchProviderCost: a throwing fetch is caught", async () => {
164
- const r = await fetchProviderCost({ fetchImpl: async () => { throw new Error("net"); }, day: DAY });
165
- assert.equal(r.ok, false);
166
- assert.match(r.note, /threw/);
167
- });
168
-
169
- test("parseAmount: handles units/nanos protobuf-style money", () => {
170
- assert.equal(_internals.parseAmount({ units: 1, nanos: 500_000_000 }), 1.5);
171
- });
172
-
173
- // ---------------------------------------------------------------------------
174
- // computeDrift
175
- // ---------------------------------------------------------------------------
176
-
177
- test("computeDrift: provider-relative signed percentage", () => {
178
- // provider billed 110, we recorded 100 → +9.09% (under-counted, dangerous dir).
179
- assert.equal(computeDrift(100, 110), 9.09);
180
- // provider billed 90, we recorded 100 → -11.11% (over-counted).
181
- assert.equal(computeDrift(100, 90), -11.11);
182
- assert.equal(computeDrift(100, 100), 0);
183
- });
184
-
185
- test("computeDrift: absent reference → null", () => {
186
- assert.equal(computeDrift(100, null), null);
187
- assert.equal(computeDrift(100, undefined), null);
188
- });
189
-
190
- test("computeDrift: zero reference → 0 when local 0, 100 when local nonzero", () => {
191
- assert.equal(computeDrift(0, 0), 0);
192
- assert.equal(computeDrift(5, 0), 100);
193
- });
194
-
195
- // ---------------------------------------------------------------------------
196
- // reconcileDaily — the three-way reconcile
197
- // ---------------------------------------------------------------------------
198
-
199
- test("reconcileDaily: computes drift and does NOT flag within 10%", async () => {
200
- const { ledgerDir } = tmpLedger([
201
- { agent: "alex", model: "claude-sonnet-4-6", estimated_usd: 100 },
202
- ]);
203
- const fetchImpl = fakeFetch(nestedPayload([{ usd: 105, model: "claude-sonnet-4-6", agent: "key_alex" }]));
204
-
205
- const r = await reconcileDaily({ ledgerDir, day: DAY, fetchImpl });
206
- assert.equal(r.localUSD, 100);
207
- assert.equal(r.providerUSD, 105);
208
- assert.equal(r.driftPct, 4.76); // (105-100)/105
209
- assert.equal(r.flagged, false);
210
- });
211
-
212
- test("reconcileDaily: flags drift > 10%", async () => {
213
- const { ledgerDir } = tmpLedger([
214
- { agent: "alex", model: "claude-sonnet-4-6", estimated_usd: 80 },
215
- ]);
216
- // provider billed 100, we recorded 80 → +20% (we under-counted).
217
- const fetchImpl = fakeFetch(nestedPayload([{ usd: 100, model: "claude-sonnet-4-6", agent: "key_alex" }]));
218
-
219
- const r = await reconcileDaily({ ledgerDir, day: DAY, fetchImpl });
220
- assert.equal(r.driftPct, 20);
221
- assert.equal(r.flagged, true);
222
- assert.ok(r.notes.some((n) => /provider drift/.test(n)));
223
- });
224
-
225
- test("reconcileDaily: three-way per-model view (local vs provider)", async () => {
226
- const { ledgerDir } = tmpLedger([
227
- { agent: "alex", model: "claude-sonnet-4-6", estimated_usd: 50 },
228
- { agent: "alex", model: "claude-haiku-4-5", estimated_usd: 5 },
229
- ]);
230
- const fetchImpl = fakeFetch(
231
- nestedPayload([
232
- { usd: 52, model: "claude-sonnet-4-6", agent: "key_alex" },
233
- { usd: 6, model: "claude-haiku-4-5", agent: "key_alex" },
234
- ])
235
- );
236
- const r = await reconcileDaily({ ledgerDir, day: DAY, fetchImpl });
237
- assert.equal(r.byModel["claude-sonnet-4-6"].local, 50);
238
- assert.equal(r.byModel["claude-sonnet-4-6"].provider, 52);
239
- assert.equal(r.byModel["claude-haiku-4-5"].local, 5);
240
- assert.equal(r.byModel["claude-haiku-4-5"].provider, 6);
241
- assert.equal(r.byAgent["key_alex"], 58);
242
- });
243
-
244
- test("reconcileDaily: surfaces a model the provider billed but the local ledger missed", async () => {
245
- const { ledgerDir } = tmpLedger([
246
- { agent: "alex", model: "claude-sonnet-4-6", estimated_usd: 50 },
247
- ]);
248
- // provider also billed a Haiku flood the ledger never recorded.
249
- const fetchImpl = fakeFetch(
250
- nestedPayload([
251
- { usd: 50, model: "claude-sonnet-4-6", agent: "key_alex" },
252
- { usd: 30, model: "claude-haiku-4-5", agent: "key_alex" },
253
- ])
254
- );
255
- const r = await reconcileDaily({ ledgerDir, day: DAY, fetchImpl });
256
- assert.equal(r.byModel["claude-haiku-4-5"].local, 0);
257
- assert.equal(r.byModel["claude-haiku-4-5"].provider, 30);
258
- assert.equal(r.flagged, true); // local 50 vs provider 80 → 37.5%
259
- });
260
-
261
- test("reconcileDaily: optional rollup leg (third way) + its own drift", async () => {
262
- const { ledgerDir } = tmpLedger([
263
- { agent: "alex", model: "m", estimated_usd: 100 },
264
- ]);
265
- const fetchImpl = fakeFetch(nestedPayload([{ usd: 100, model: "m", agent: "k" }]));
266
- const rollupImpl = async () => ({ totalUSD: 130 }); // rollup says 130 → +23% drift
267
-
268
- const r = await reconcileDaily({ ledgerDir, day: DAY, fetchImpl, rollupImpl });
269
- assert.equal(r.providerUSD, 100);
270
- assert.equal(r.rollupUSD, 130);
271
- assert.equal(r.driftPct, 0); // local==provider
272
- assert.equal(r.rollupDriftPct, 23.08);
273
- assert.equal(r.flagged, true); // rollup leg trips the flag
274
- });
275
-
276
- test("reconcileDaily: provider leg absent (no fetchImpl) → fail-open, providerUSD null", async () => {
277
- const { ledgerDir } = tmpLedger([{ agent: "a", model: "m", estimated_usd: 10 }]);
278
- const r = await reconcileDaily({ ledgerDir, day: DAY });
279
- assert.equal(r.localUSD, 10);
280
- assert.equal(r.providerUSD, null);
281
- assert.equal(r.driftPct, null);
282
- assert.equal(r.flagged, false);
283
- assert.ok(r.notes.some((n) => /provider leg skipped/.test(n)));
284
- });
285
-
286
- test("reconcileDaily: a down provider API does not throw and is reported null", async () => {
287
- const { ledgerDir } = tmpLedger([{ agent: "a", model: "m", estimated_usd: 10 }]);
288
- const fetchImpl = fakeFetch({}, { ok: false, status: 500 });
289
- const r = await reconcileDaily({ ledgerDir, day: DAY, fetchImpl });
290
- assert.equal(r.providerUSD, null);
291
- assert.equal(r.flagged, false);
292
- assert.ok(r.notes.some((n) => /provider leg unavailable/.test(n)));
293
- });
294
-
295
- test("reconcileDaily: a throwing rollupImpl is caught", async () => {
296
- const { ledgerDir } = tmpLedger([{ agent: "a", model: "m", estimated_usd: 10 }]);
297
- const fetchImpl = fakeFetch(nestedPayload([{ usd: 10, model: "m" }]));
298
- const r = await reconcileDaily({
299
- ledgerDir, day: DAY, fetchImpl, rollupImpl: async () => { throw new Error("x"); },
300
- });
301
- assert.equal(r.rollupUSD, null);
302
- assert.ok(r.notes.some((n) => /rollup leg threw/.test(n)));
303
- });
304
-
305
- test("reconcileDaily: custom threshold", async () => {
306
- const { ledgerDir } = tmpLedger([{ agent: "a", model: "m", estimated_usd: 95 }]);
307
- const fetchImpl = fakeFetch(nestedPayload([{ usd: 100, model: "m" }]));
308
- // 5% drift, threshold 3% → flagged.
309
- const r = await reconcileDaily({ ledgerDir, day: DAY, fetchImpl, thresholdPct: 3 });
310
- assert.equal(r.flagged, true);
311
- assert.equal(r.thresholdPct, 3);
312
- });
313
-
314
- test("DRIFT_THRESHOLD_PCT default is 10", () => {
315
- assert.equal(DRIFT_THRESHOLD_PCT, 10);
316
- });