@claude-flow/cli 3.27.4 → 3.29.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (464) hide show
  1. package/.claude/agents/analysis/analyze-code-quality.md +178 -178
  2. package/.claude/agents/analysis/code-analyzer.md +209 -209
  3. package/.claude/agents/analysis/code-review/analyze-code-quality.md +178 -178
  4. package/.claude/agents/architecture/arch-system-design.md +156 -156
  5. package/.claude/agents/architecture/system-design/arch-system-design.md +154 -154
  6. package/.claude/agents/browser/browser-agent.yaml +182 -182
  7. package/.claude/agents/consensus/byzantine-coordinator.md +62 -62
  8. package/.claude/agents/consensus/crdt-synchronizer.md +996 -996
  9. package/.claude/agents/consensus/gossip-coordinator.md +62 -62
  10. package/.claude/agents/consensus/performance-benchmarker.md +850 -850
  11. package/.claude/agents/consensus/quorum-manager.md +822 -822
  12. package/.claude/agents/consensus/raft-manager.md +62 -62
  13. package/.claude/agents/consensus/security-manager.md +621 -621
  14. package/.claude/agents/core/planner.md +374 -374
  15. package/.claude/agents/custom/test-long-runner.md +44 -44
  16. package/.claude/agents/data/data-ml-model.md +444 -444
  17. package/.claude/agents/data/ml/data-ml-model.md +192 -192
  18. package/.claude/agents/development/backend/dev-backend-api.md +141 -141
  19. package/.claude/agents/development/dev-backend-api.md +344 -344
  20. package/.claude/agents/devops/ci-cd/ops-cicd-github.md +163 -163
  21. package/.claude/agents/devops/ops-cicd-github.md +164 -164
  22. package/.claude/agents/documentation/api-docs/docs-api-openapi.md +173 -173
  23. package/.claude/agents/documentation/docs-api-openapi.md +354 -354
  24. package/.claude/agents/flow-nexus/app-store.md +87 -87
  25. package/.claude/agents/flow-nexus/authentication.md +68 -68
  26. package/.claude/agents/flow-nexus/challenges.md +80 -80
  27. package/.claude/agents/flow-nexus/neural-network.md +87 -87
  28. package/.claude/agents/flow-nexus/payments.md +82 -82
  29. package/.claude/agents/flow-nexus/sandbox.md +75 -75
  30. package/.claude/agents/flow-nexus/swarm.md +75 -75
  31. package/.claude/agents/flow-nexus/user-tools.md +95 -95
  32. package/.claude/agents/flow-nexus/workflow.md +83 -83
  33. package/.claude/agents/github/code-review-swarm.md +377 -377
  34. package/.claude/agents/github/github-modes.md +172 -172
  35. package/.claude/agents/github/issue-tracker.md +575 -575
  36. package/.claude/agents/github/multi-repo-swarm.md +552 -552
  37. package/.claude/agents/github/pr-manager.md +437 -437
  38. package/.claude/agents/github/project-board-sync.md +508 -508
  39. package/.claude/agents/github/release-manager.md +604 -604
  40. package/.claude/agents/github/release-swarm.md +582 -582
  41. package/.claude/agents/github/repo-architect.md +397 -397
  42. package/.claude/agents/github/swarm-issue.md +572 -572
  43. package/.claude/agents/github/swarm-pr.md +427 -427
  44. package/.claude/agents/github/sync-coordinator.md +451 -451
  45. package/.claude/agents/github/workflow-automation.md +902 -902
  46. package/.claude/agents/goal/agent.md +815 -815
  47. package/.claude/agents/optimization/benchmark-suite.md +664 -664
  48. package/.claude/agents/optimization/load-balancer.md +430 -430
  49. package/.claude/agents/optimization/performance-monitor.md +671 -671
  50. package/.claude/agents/optimization/resource-allocator.md +673 -673
  51. package/.claude/agents/optimization/topology-optimizer.md +807 -807
  52. package/.claude/agents/payments/agentic-payments.md +126 -126
  53. package/.claude/agents/sona/sona-learning-optimizer.md +74 -74
  54. package/.claude/agents/sparc/architecture.md +698 -698
  55. package/.claude/agents/sparc/pseudocode.md +519 -519
  56. package/.claude/agents/sparc/refinement.md +801 -801
  57. package/.claude/agents/sparc/specification.md +477 -477
  58. package/.claude/agents/specialized/mobile/spec-mobile-react-native.md +224 -224
  59. package/.claude/agents/specialized/spec-mobile-react-native.md +226 -226
  60. package/.claude/agents/sublinear/consensus-coordinator.md +337 -337
  61. package/.claude/agents/sublinear/matrix-optimizer.md +184 -184
  62. package/.claude/agents/sublinear/pagerank-analyzer.md +298 -298
  63. package/.claude/agents/sublinear/performance-optimizer.md +367 -367
  64. package/.claude/agents/sublinear/trading-predictor.md +245 -245
  65. package/.claude/agents/swarm/adaptive-coordinator.md +1126 -1126
  66. package/.claude/agents/swarm/hierarchical-coordinator.md +709 -709
  67. package/.claude/agents/swarm/mesh-coordinator.md +962 -962
  68. package/.claude/agents/templates/automation-smart-agent.md +204 -204
  69. package/.claude/agents/templates/base-template-generator.md +289 -289
  70. package/.claude/agents/templates/coordinator-swarm-init.md +89 -89
  71. package/.claude/agents/templates/github-pr-manager.md +176 -176
  72. package/.claude/agents/templates/implementer-sparc-coder.md +258 -258
  73. package/.claude/agents/templates/memory-coordinator.md +186 -186
  74. package/.claude/agents/templates/orchestrator-task.md +138 -138
  75. package/.claude/agents/templates/performance-analyzer.md +198 -198
  76. package/.claude/agents/templates/sparc-coordinator.md +513 -513
  77. package/.claude/agents/testing/production-validator.md +394 -394
  78. package/.claude/agents/testing/tdd-london-swarm.md +243 -243
  79. package/.claude/agents/v3/aidefence-guardian.md +282 -282
  80. package/.claude/agents/v3/claims-authorizer.md +208 -208
  81. package/.claude/agents/v3/collective-intelligence-coordinator.md +993 -993
  82. package/.claude/agents/v3/ddd-domain-expert.md +220 -220
  83. package/.claude/agents/v3/injection-analyst.md +236 -236
  84. package/.claude/agents/v3/performance-engineer.md +1233 -1233
  85. package/.claude/agents/v3/pii-detector.md +151 -151
  86. package/.claude/agents/v3/reasoningbank-learner.md +213 -213
  87. package/.claude/agents/v3/security-architect-aidefence.md +410 -410
  88. package/.claude/agents/v3/security-architect.md +867 -867
  89. package/.claude/agents/v3/swarm-memory-manager.md +157 -157
  90. package/.claude/agents/v3/v3-integration-architect.md +205 -205
  91. package/.claude/commands/agents/README.md +50 -50
  92. package/.claude/commands/agents/agent-capabilities.md +140 -140
  93. package/.claude/commands/agents/agent-coordination.md +28 -28
  94. package/.claude/commands/agents/agent-spawning.md +28 -28
  95. package/.claude/commands/agents/agent-types.md +216 -216
  96. package/.claude/commands/agents/health.md +139 -139
  97. package/.claude/commands/agents/list.md +100 -100
  98. package/.claude/commands/agents/logs.md +130 -130
  99. package/.claude/commands/agents/metrics.md +122 -122
  100. package/.claude/commands/agents/pool.md +127 -127
  101. package/.claude/commands/agents/spawn.md +140 -140
  102. package/.claude/commands/agents/status.md +115 -115
  103. package/.claude/commands/agents/stop.md +102 -102
  104. package/.claude/commands/analysis/COMMAND_COMPLIANCE_REPORT.md +53 -53
  105. package/.claude/commands/analysis/README.md +9 -9
  106. package/.claude/commands/analysis/bottleneck-detect.md +162 -162
  107. package/.claude/commands/analysis/performance-bottlenecks.md +58 -58
  108. package/.claude/commands/analysis/performance-report.md +25 -25
  109. package/.claude/commands/analysis/token-efficiency.md +44 -44
  110. package/.claude/commands/analysis/token-usage.md +25 -25
  111. package/.claude/commands/automation/README.md +9 -9
  112. package/.claude/commands/automation/auto-agent.md +122 -122
  113. package/.claude/commands/automation/self-healing.md +105 -105
  114. package/.claude/commands/automation/session-memory.md +89 -89
  115. package/.claude/commands/automation/smart-agents.md +72 -72
  116. package/.claude/commands/automation/smart-spawn.md +25 -25
  117. package/.claude/commands/automation/workflow-select.md +25 -25
  118. package/.claude/commands/claude-flow-help.md +103 -103
  119. package/.claude/commands/claude-flow-memory.md +107 -107
  120. package/.claude/commands/claude-flow-swarm.md +205 -205
  121. package/.claude/commands/coordination/README.md +9 -9
  122. package/.claude/commands/coordination/agent-spawn.md +25 -25
  123. package/.claude/commands/coordination/init.md +44 -44
  124. package/.claude/commands/coordination/orchestrate.md +43 -43
  125. package/.claude/commands/coordination/spawn.md +45 -45
  126. package/.claude/commands/coordination/swarm-init.md +85 -85
  127. package/.claude/commands/coordination/task-orchestrate.md +25 -25
  128. package/.claude/commands/github/README.md +11 -11
  129. package/.claude/commands/github/code-review-swarm.md +513 -513
  130. package/.claude/commands/github/code-review.md +25 -25
  131. package/.claude/commands/github/github-modes.md +146 -146
  132. package/.claude/commands/github/github-swarm.md +121 -121
  133. package/.claude/commands/github/issue-tracker.md +291 -291
  134. package/.claude/commands/github/issue-triage.md +25 -25
  135. package/.claude/commands/github/multi-repo-swarm.md +518 -518
  136. package/.claude/commands/github/pr-enhance.md +26 -26
  137. package/.claude/commands/github/pr-manager.md +169 -169
  138. package/.claude/commands/github/project-board-sync.md +470 -470
  139. package/.claude/commands/github/release-manager.md +339 -339
  140. package/.claude/commands/github/release-swarm.md +543 -543
  141. package/.claude/commands/github/repo-analyze.md +25 -25
  142. package/.claude/commands/github/repo-architect.md +366 -366
  143. package/.claude/commands/github/swarm-issue.md +484 -484
  144. package/.claude/commands/github/swarm-pr.md +287 -287
  145. package/.claude/commands/github/sync-coordinator.md +302 -302
  146. package/.claude/commands/github/workflow-automation.md +441 -441
  147. package/.claude/commands/hive-mind/README.md +17 -17
  148. package/.claude/commands/hive-mind/hive-mind-consensus.md +8 -8
  149. package/.claude/commands/hive-mind/hive-mind-init.md +18 -18
  150. package/.claude/commands/hive-mind/hive-mind-memory.md +8 -8
  151. package/.claude/commands/hive-mind/hive-mind-metrics.md +8 -8
  152. package/.claude/commands/hive-mind/hive-mind-resume.md +8 -8
  153. package/.claude/commands/hive-mind/hive-mind-sessions.md +8 -8
  154. package/.claude/commands/hive-mind/hive-mind-spawn.md +21 -21
  155. package/.claude/commands/hive-mind/hive-mind-status.md +8 -8
  156. package/.claude/commands/hive-mind/hive-mind-stop.md +8 -8
  157. package/.claude/commands/hive-mind/hive-mind-wizard.md +8 -8
  158. package/.claude/commands/hive-mind/hive-mind.md +27 -27
  159. package/.claude/commands/hooks/README.md +11 -11
  160. package/.claude/commands/hooks/overview.md +57 -57
  161. package/.claude/commands/hooks/post-edit.md +117 -117
  162. package/.claude/commands/hooks/post-task.md +112 -112
  163. package/.claude/commands/hooks/pre-edit.md +113 -113
  164. package/.claude/commands/hooks/pre-task.md +111 -111
  165. package/.claude/commands/hooks/session-end.md +118 -118
  166. package/.claude/commands/hooks/setup.md +102 -102
  167. package/.claude/commands/memory/README.md +9 -9
  168. package/.claude/commands/memory/memory-persist.md +25 -25
  169. package/.claude/commands/memory/memory-search.md +25 -25
  170. package/.claude/commands/memory/memory-usage.md +25 -25
  171. package/.claude/commands/memory/neural.md +47 -47
  172. package/.claude/commands/monitoring/README.md +9 -9
  173. package/.claude/commands/monitoring/agent-metrics.md +25 -25
  174. package/.claude/commands/monitoring/agents.md +44 -44
  175. package/.claude/commands/monitoring/real-time-view.md +25 -25
  176. package/.claude/commands/monitoring/status.md +46 -46
  177. package/.claude/commands/monitoring/swarm-monitor.md +25 -25
  178. package/.claude/commands/optimization/README.md +9 -9
  179. package/.claude/commands/optimization/auto-topology.md +61 -61
  180. package/.claude/commands/optimization/cache-manage.md +25 -25
  181. package/.claude/commands/optimization/parallel-execute.md +25 -25
  182. package/.claude/commands/optimization/parallel-execution.md +49 -49
  183. package/.claude/commands/optimization/topology-optimize.md +25 -25
  184. package/.claude/commands/pair/README.md +260 -260
  185. package/.claude/commands/pair/commands.md +545 -545
  186. package/.claude/commands/pair/config.md +509 -509
  187. package/.claude/commands/pair/examples.md +511 -511
  188. package/.claude/commands/pair/modes.md +347 -347
  189. package/.claude/commands/pair/session.md +406 -406
  190. package/.claude/commands/pair/start.md +208 -208
  191. package/.claude/commands/sparc/analyzer.md +51 -51
  192. package/.claude/commands/sparc/architect.md +53 -53
  193. package/.claude/commands/sparc/ask.md +97 -97
  194. package/.claude/commands/sparc/batch-executor.md +54 -54
  195. package/.claude/commands/sparc/code.md +89 -89
  196. package/.claude/commands/sparc/coder.md +54 -54
  197. package/.claude/commands/sparc/debug.md +83 -83
  198. package/.claude/commands/sparc/debugger.md +54 -54
  199. package/.claude/commands/sparc/designer.md +53 -53
  200. package/.claude/commands/sparc/devops.md +109 -109
  201. package/.claude/commands/sparc/docs-writer.md +80 -80
  202. package/.claude/commands/sparc/documenter.md +54 -54
  203. package/.claude/commands/sparc/innovator.md +54 -54
  204. package/.claude/commands/sparc/integration.md +83 -83
  205. package/.claude/commands/sparc/mcp.md +117 -117
  206. package/.claude/commands/sparc/memory-manager.md +54 -54
  207. package/.claude/commands/sparc/optimizer.md +54 -54
  208. package/.claude/commands/sparc/orchestrator.md +131 -131
  209. package/.claude/commands/sparc/post-deployment-monitoring-mode.md +83 -83
  210. package/.claude/commands/sparc/refinement-optimization-mode.md +83 -83
  211. package/.claude/commands/sparc/researcher.md +54 -54
  212. package/.claude/commands/sparc/reviewer.md +54 -54
  213. package/.claude/commands/sparc/security-review.md +80 -80
  214. package/.claude/commands/sparc/sparc-modes.md +174 -174
  215. package/.claude/commands/sparc/sparc.md +111 -111
  216. package/.claude/commands/sparc/spec-pseudocode.md +80 -80
  217. package/.claude/commands/sparc/supabase-admin.md +348 -348
  218. package/.claude/commands/sparc/swarm-coordinator.md +54 -54
  219. package/.claude/commands/sparc/tdd.md +54 -54
  220. package/.claude/commands/sparc/tester.md +54 -54
  221. package/.claude/commands/sparc/tutorial.md +79 -79
  222. package/.claude/commands/sparc/workflow-manager.md +54 -54
  223. package/.claude/commands/sparc.md +166 -166
  224. package/.claude/commands/stream-chain/pipeline.md +120 -120
  225. package/.claude/commands/stream-chain/run.md +69 -69
  226. package/.claude/commands/swarm/README.md +15 -15
  227. package/.claude/commands/swarm/analysis.md +95 -95
  228. package/.claude/commands/swarm/development.md +96 -96
  229. package/.claude/commands/swarm/examples.md +168 -168
  230. package/.claude/commands/swarm/maintenance.md +102 -102
  231. package/.claude/commands/swarm/optimization.md +117 -117
  232. package/.claude/commands/swarm/research.md +136 -136
  233. package/.claude/commands/swarm/swarm-analysis.md +8 -8
  234. package/.claude/commands/swarm/swarm-background.md +8 -8
  235. package/.claude/commands/swarm/swarm-init.md +19 -19
  236. package/.claude/commands/swarm/swarm-modes.md +8 -8
  237. package/.claude/commands/swarm/swarm-monitor.md +8 -8
  238. package/.claude/commands/swarm/swarm-spawn.md +19 -19
  239. package/.claude/commands/swarm/swarm-status.md +8 -8
  240. package/.claude/commands/swarm/swarm-strategies.md +8 -8
  241. package/.claude/commands/swarm/swarm.md +87 -87
  242. package/.claude/commands/swarm/testing.md +131 -131
  243. package/.claude/commands/training/README.md +9 -9
  244. package/.claude/commands/training/model-update.md +25 -25
  245. package/.claude/commands/training/neural-patterns.md +107 -107
  246. package/.claude/commands/training/neural-train.md +75 -75
  247. package/.claude/commands/training/pattern-learn.md +25 -25
  248. package/.claude/commands/training/specialization.md +62 -62
  249. package/.claude/commands/truth/start.md +142 -142
  250. package/.claude/commands/verify/check.md +49 -49
  251. package/.claude/commands/verify/start.md +127 -127
  252. package/.claude/commands/workflows/README.md +9 -9
  253. package/.claude/commands/workflows/development.md +77 -77
  254. package/.claude/commands/workflows/research.md +62 -62
  255. package/.claude/commands/workflows/workflow-create.md +25 -25
  256. package/.claude/commands/workflows/workflow-execute.md +25 -25
  257. package/.claude/commands/workflows/workflow-export.md +25 -25
  258. package/.claude/eval/human-relevance-frozen-v1.json +17 -17
  259. package/.claude/evolve-proof/generation-0.json +211 -211
  260. package/.claude/evolve-proof/real-generation-0.json +406 -406
  261. package/.claude/evolve-proof/real-generation-1.json +406 -406
  262. package/.claude/helpers/README.md +96 -96
  263. package/.claude/helpers/adr-compliance.sh +186 -186
  264. package/.claude/helpers/auto-commit.sh +178 -178
  265. package/.claude/helpers/auto-memory-hook.mjs +430 -430
  266. package/.claude/helpers/checkpoint-manager.sh +251 -251
  267. package/.claude/helpers/daemon-manager.sh +252 -252
  268. package/.claude/helpers/ddd-tracker.sh +144 -144
  269. package/.claude/helpers/github-safe.js +156 -156
  270. package/.claude/helpers/github-setup.sh +45 -45
  271. package/.claude/helpers/guidance-hook.sh +13 -13
  272. package/.claude/helpers/guidance-hooks.sh +102 -102
  273. package/.claude/helpers/health-monitor.sh +108 -108
  274. package/.claude/helpers/helpers.manifest.json +6 -6
  275. package/.claude/helpers/hook-handler.cjs +460 -460
  276. package/.claude/helpers/intelligence.cjs +1058 -1058
  277. package/.claude/helpers/learning-hooks.sh +329 -329
  278. package/.claude/helpers/learning-optimizer.sh +127 -127
  279. package/.claude/helpers/learning-service.mjs +1144 -1144
  280. package/.claude/helpers/memory.js +83 -83
  281. package/.claude/helpers/metrics-db.mjs +503 -503
  282. package/.claude/helpers/pattern-consolidator.sh +86 -86
  283. package/.claude/helpers/perf-worker.sh +160 -160
  284. package/.claude/helpers/post-commit +16 -16
  285. package/.claude/helpers/pre-commit +26 -26
  286. package/.claude/helpers/quick-start.sh +19 -19
  287. package/.claude/helpers/router.js +105 -105
  288. package/.claude/helpers/security-scanner.sh +127 -127
  289. package/.claude/helpers/session.js +157 -157
  290. package/.claude/helpers/setup-mcp.sh +18 -18
  291. package/.claude/helpers/standard-checkpoint-hooks.sh +189 -189
  292. package/.claude/helpers/statusline-hook.sh +21 -21
  293. package/.claude/helpers/statusline.cjs +925 -925
  294. package/.claude/helpers/statusline.js +352 -352
  295. package/.claude/helpers/swarm-comms.sh +353 -353
  296. package/.claude/helpers/swarm-hooks.sh +761 -761
  297. package/.claude/helpers/swarm-monitor.sh +210 -210
  298. package/.claude/helpers/sync-v3-metrics.sh +245 -245
  299. package/.claude/helpers/update-v3-progress.sh +165 -165
  300. package/.claude/helpers/v3-quick-status.sh +57 -57
  301. package/.claude/helpers/v3.sh +110 -110
  302. package/.claude/helpers/validate-v3-config.sh +215 -215
  303. package/.claude/helpers/worker-manager.sh +170 -170
  304. package/.claude/proven-config.json +1 -1
  305. package/.claude/proven-config.manifest.json +37 -37
  306. package/.claude/proven-config.signed.json +41 -41
  307. package/.claude/settings.json +182 -182
  308. package/.claude/skills/agentdb-advanced/SKILL.md +550 -550
  309. package/.claude/skills/agentdb-learning/SKILL.md +545 -545
  310. package/.claude/skills/agentdb-memory-patterns/SKILL.md +339 -339
  311. package/.claude/skills/agentdb-optimization/SKILL.md +509 -509
  312. package/.claude/skills/agentdb-vector-search/SKILL.md +339 -339
  313. package/.claude/skills/browser/SKILL.md +204 -204
  314. package/.claude/skills/dual-mode/README.md +71 -71
  315. package/.claude/skills/dual-mode/dual-collect.md +103 -103
  316. package/.claude/skills/dual-mode/dual-coordinate.md +85 -85
  317. package/.claude/skills/dual-mode/dual-spawn.md +81 -81
  318. package/.claude/skills/flow-nexus-neural/SKILL.md +727 -727
  319. package/.claude/skills/flow-nexus-platform/SKILL.md +1154 -1154
  320. package/.claude/skills/flow-nexus-swarm/SKILL.md +604 -604
  321. package/.claude/skills/github-code-review/SKILL.md +1125 -1125
  322. package/.claude/skills/github-multi-repo/SKILL.md +862 -862
  323. package/.claude/skills/github-project-management/SKILL.md +1262 -1262
  324. package/.claude/skills/github-release-management/SKILL.md +1064 -1064
  325. package/.claude/skills/github-workflow-automation/SKILL.md +1047 -1047
  326. package/.claude/skills/hooks-automation/SKILL.md +1201 -1201
  327. package/.claude/skills/pair-programming/SKILL.md +1202 -1202
  328. package/.claude/skills/reasoningbank-agentdb/SKILL.md +446 -446
  329. package/.claude/skills/reasoningbank-intelligence/SKILL.md +201 -201
  330. package/.claude/skills/skill-builder/SKILL.md +910 -910
  331. package/.claude/skills/sparc-methodology/SKILL.md +1106 -1106
  332. package/.claude/skills/stream-chain/SKILL.md +560 -560
  333. package/.claude/skills/swarm-advanced/SKILL.md +970 -970
  334. package/.claude/skills/swarm-orchestration/SKILL.md +179 -179
  335. package/.claude/skills/v3-cli-modernization/SKILL.md +871 -871
  336. package/.claude/skills/v3-core-implementation/SKILL.md +796 -796
  337. package/.claude/skills/v3-ddd-architecture/SKILL.md +441 -441
  338. package/.claude/skills/v3-integration-deep/SKILL.md +240 -240
  339. package/.claude/skills/v3-mcp-optimization/SKILL.md +776 -776
  340. package/.claude/skills/v3-memory-unification/SKILL.md +173 -173
  341. package/.claude/skills/v3-performance-optimization/SKILL.md +389 -389
  342. package/.claude/skills/v3-security-overhaul/SKILL.md +81 -81
  343. package/.claude/skills/v3-swarm-coordination/SKILL.md +339 -339
  344. package/.claude/skills/verification-quality/SKILL.md +691 -691
  345. package/README.md +419 -419
  346. package/bin/cli.js +314 -314
  347. package/bin/mcp-server.js +224 -224
  348. package/bin/preinstall.cjs +2 -2
  349. package/catalog-manifest.json +2 -2
  350. package/dist/src/benchmarks/gaia-critic.js +24 -24
  351. package/dist/src/business-pods/bbs-budget-tracker.js +53 -53
  352. package/dist/src/commands/completions.js +409 -409
  353. package/dist/src/commands/daemon.d.ts +23 -0
  354. package/dist/src/commands/daemon.js +216 -44
  355. package/dist/src/commands/embeddings.js +26 -26
  356. package/dist/src/commands/funnel.d.ts +2 -0
  357. package/dist/src/commands/funnel.js +110 -2
  358. package/dist/src/commands/hive-mind.js +97 -97
  359. package/dist/src/commands/hooks.js +9 -9
  360. package/dist/src/commands/init.js +94 -38
  361. package/dist/src/commands/memory.js +85 -1
  362. package/dist/src/commands/ruvector/backup.js +23 -23
  363. package/dist/src/commands/ruvector/benchmark.js +31 -31
  364. package/dist/src/commands/ruvector/import.js +14 -14
  365. package/dist/src/commands/ruvector/init.js +115 -115
  366. package/dist/src/commands/ruvector/migrate.js +99 -99
  367. package/dist/src/commands/ruvector/optimize.js +51 -51
  368. package/dist/src/commands/ruvector/setup.js +624 -624
  369. package/dist/src/commands/ruvector/status.js +38 -38
  370. package/dist/src/config/proven-config.js +2 -2
  371. package/dist/src/funnel/disclosure.d.ts +1 -0
  372. package/dist/src/funnel/disclosure.js +12 -0
  373. package/dist/src/funnel/index.d.ts +1 -1
  374. package/dist/src/funnel/index.js +1 -1
  375. package/dist/src/init/claudemd-generator.js +231 -231
  376. package/dist/src/init/executor.js +453 -453
  377. package/dist/src/init/helper-signing.d.ts +8 -1
  378. package/dist/src/init/helper-signing.js +9 -2
  379. package/dist/src/init/helpers-generator.js +751 -751
  380. package/dist/src/init/statusline-generator.js +994 -951
  381. package/dist/src/mcp-tools/agentdb-tools.js +15 -15
  382. package/dist/src/mcp-tools/browser-intent-tools.js +19 -19
  383. package/dist/src/memory/graph-edge-writer.js +22 -22
  384. package/dist/src/memory/memory-bridge.d.ts +10 -0
  385. package/dist/src/memory/memory-bridge.js +139 -88
  386. package/dist/src/memory/memory-initializer.d.ts +18 -0
  387. package/dist/src/memory/memory-initializer.js +539 -407
  388. package/dist/src/memory/rabitq-index.js +5 -5
  389. package/dist/src/runtime/headless.js +28 -28
  390. package/dist/src/ruvector/flash-attention.d.ts +195 -0
  391. package/dist/src/ruvector/flash-attention.js +643 -0
  392. package/dist/src/ruvector/moe-router.d.ts +206 -0
  393. package/dist/src/ruvector/moe-router.js +626 -0
  394. package/dist/src/services/distill-tuning.js +7 -7
  395. package/dist/src/services/event-stream.d.ts +25 -0
  396. package/dist/src/services/event-stream.js +27 -0
  397. package/dist/src/services/global-ai-budget.d.ts +25 -0
  398. package/dist/src/services/global-ai-budget.js +56 -0
  399. package/dist/src/services/headless-worker-executor.d.ts +21 -0
  400. package/dist/src/services/headless-worker-executor.js +152 -86
  401. package/dist/src/services/loop-worker-runner.d.ts +16 -0
  402. package/dist/src/services/loop-worker-runner.js +34 -0
  403. package/dist/src/services/memory-distillation.js +4 -4
  404. package/dist/src/services/repo-supervisor.d.ts +70 -0
  405. package/dist/src/services/repo-supervisor.js +228 -0
  406. package/dist/src/services/runtime-capabilities.d.ts +22 -0
  407. package/dist/src/services/runtime-capabilities.js +45 -0
  408. package/dist/src/services/worker-daemon.d.ts +26 -0
  409. package/dist/src/services/worker-daemon.js +95 -5
  410. package/dist/src/services/workspace-lease.d.ts +55 -0
  411. package/dist/src/services/workspace-lease.js +191 -0
  412. package/dist/src/transfer/deploy-seraphine.js +23 -23
  413. package/package.json +134 -133
  414. package/plugins/ruflo-metaharness/.claude-plugin/plugin.json +32 -32
  415. package/plugins/ruflo-metaharness/README.md +72 -72
  416. package/plugins/ruflo-metaharness/agents/metaharness-architect.md +58 -58
  417. package/plugins/ruflo-metaharness/commands/ruflo-metaharness.md +48 -48
  418. package/plugins/ruflo-metaharness/scripts/_darwin.mjs +210 -210
  419. package/plugins/ruflo-metaharness/scripts/_harness.mjs +330 -330
  420. package/plugins/ruflo-metaharness/scripts/_invoke.mjs +231 -231
  421. package/plugins/ruflo-metaharness/scripts/_redblue.mjs +143 -143
  422. package/plugins/ruflo-metaharness/scripts/_similarity.mjs +161 -161
  423. package/plugins/ruflo-metaharness/scripts/_spike-similarity.mjs +223 -223
  424. package/plugins/ruflo-metaharness/scripts/audit-list.mjs +158 -158
  425. package/plugins/ruflo-metaharness/scripts/audit-trend.mjs +272 -272
  426. package/plugins/ruflo-metaharness/scripts/bench-parse-mcp-scan.mjs +146 -146
  427. package/plugins/ruflo-metaharness/scripts/bench-recordpair-overhead.mjs +186 -186
  428. package/plugins/ruflo-metaharness/scripts/bench-similarity.mjs +177 -177
  429. package/plugins/ruflo-metaharness/scripts/bench.mjs +95 -95
  430. package/plugins/ruflo-metaharness/scripts/drift-from-history.mjs +363 -363
  431. package/plugins/ruflo-metaharness/scripts/evolve.mjs +404 -404
  432. package/plugins/ruflo-metaharness/scripts/genome.mjs +80 -80
  433. package/plugins/ruflo-metaharness/scripts/gepa.mjs +153 -153
  434. package/plugins/ruflo-metaharness/scripts/learn.mjs +127 -127
  435. package/plugins/ruflo-metaharness/scripts/mcp-scan.mjs +111 -111
  436. package/plugins/ruflo-metaharness/scripts/mint.mjs +126 -126
  437. package/plugins/ruflo-metaharness/scripts/oia-audit.mjs +228 -228
  438. package/plugins/ruflo-metaharness/scripts/redblue.mjs +286 -286
  439. package/plugins/ruflo-metaharness/scripts/router-parallel-analyze.mjs +250 -250
  440. package/plugins/ruflo-metaharness/scripts/score.mjs +92 -92
  441. package/plugins/ruflo-metaharness/scripts/security-bench.mjs +174 -174
  442. package/plugins/ruflo-metaharness/scripts/similarity.mjs +158 -158
  443. package/plugins/ruflo-metaharness/scripts/smoke.sh +2353 -2353
  444. package/plugins/ruflo-metaharness/scripts/test-graceful-degradation.mjs +165 -165
  445. package/plugins/ruflo-metaharness/scripts/test-mcp-tools.mjs +472 -472
  446. package/plugins/ruflo-metaharness/scripts/test-parallel-pipeline.mjs +204 -204
  447. package/plugins/ruflo-metaharness/scripts/test-pipeline-roundtrip.mjs +586 -586
  448. package/plugins/ruflo-metaharness/scripts/test-similarity.mjs +334 -334
  449. package/plugins/ruflo-metaharness/scripts/test-with-openrouter.mjs +229 -229
  450. package/plugins/ruflo-metaharness/scripts/threat-model.mjs +59 -59
  451. package/plugins/ruflo-metaharness/skills/harness-bench/SKILL.md +64 -64
  452. package/plugins/ruflo-metaharness/skills/harness-drift-from-history/SKILL.md +65 -65
  453. package/plugins/ruflo-metaharness/skills/harness-evolve/SKILL.md +131 -131
  454. package/plugins/ruflo-metaharness/skills/harness-genome/SKILL.md +54 -54
  455. package/plugins/ruflo-metaharness/skills/harness-gepa/SKILL.md +65 -65
  456. package/plugins/ruflo-metaharness/skills/harness-learn/SKILL.md +65 -65
  457. package/plugins/ruflo-metaharness/skills/harness-mcp-scan/SKILL.md +49 -49
  458. package/plugins/ruflo-metaharness/skills/harness-mint/SKILL.md +72 -72
  459. package/plugins/ruflo-metaharness/skills/harness-oia-audit/SKILL.md +79 -79
  460. package/plugins/ruflo-metaharness/skills/harness-score/SKILL.md +66 -66
  461. package/plugins/ruflo-metaharness/skills/harness-security-bench/SKILL.md +101 -101
  462. package/plugins/ruflo-metaharness/skills/harness-similarity/SKILL.md +67 -67
  463. package/plugins/ruflo-metaharness/skills/harness-threat-model/SKILL.md +41 -41
  464. package/scripts/postinstall.cjs +153 -153
package/bin/mcp-server.js CHANGED
@@ -1,224 +1,224 @@
1
- #!/usr/bin/env node
2
- /**
3
- * @claude-flow/cli - MCP Server Entry Point
4
- *
5
- * Direct stdio MCP server for Claude Code integration.
6
- * This entry point handles stdin/stdout directly for MCP protocol
7
- * without any CLI formatting output that would corrupt the protocol.
8
- */
9
-
10
- import { randomUUID } from 'crypto';
11
-
12
- // Suppress the SPECIFIC cosmetic "[AgentDB Patch] Controller index not found"
13
- // noise. Tight match (both prefix AND "Controller index not found") so other
14
- // [AgentDB Patch] warnings about real issues still flow through. Also patch
15
- // console.log because the underlying call site uses it. See bin/cli.js for
16
- // the same rationale.
17
- const _origWarn = console.warn;
18
- const _origLog = console.log;
19
- const _isCosmeticAgentdbPatchNoise = (msg) =>
20
- msg.includes('[AgentDB Patch]') && msg.includes('Controller index not found');
21
- console.warn = (...args) => {
22
- if (_isCosmeticAgentdbPatchNoise(String(args[0] ?? ''))) return;
23
- _origWarn.apply(console, args);
24
- };
25
- console.log = (...args) => {
26
- if (_isCosmeticAgentdbPatchNoise(String(args[0] ?? ''))) return;
27
- _origLog.apply(console, args);
28
- };
29
-
30
- import { listMCPTools, callMCPTool, hasTool } from '../dist/src/mcp-client.js';
31
-
32
- const VERSION = '3.0.0';
33
- const sessionId = `mcp-${Date.now()}-${randomUUID().slice(0, 8)}`;
34
-
35
- // Log to stderr (doesn't corrupt stdout for MCP protocol)
36
- console.error(
37
- `[${new Date().toISOString()}] INFO [claude-flow-mcp] (${sessionId}) Starting in stdio mode`
38
- );
39
- console.error(JSON.stringify({
40
- arch: process.arch,
41
- mode: 'mcp-stdio',
42
- nodeVersion: process.version,
43
- pid: process.pid,
44
- platform: process.platform,
45
- protocol: 'stdio',
46
- sessionId,
47
- version: VERSION,
48
- }));
49
-
50
- // Handle stdin messages
51
- // Audit-flagged DoS protection (audit_1776483149979): cap stdin buffer
52
- // to 10MB. See bin/cli.js for the same protection on the auto-detect path.
53
- const MCP_MAX_BUFFER_BYTES = 10 * 1024 * 1024;
54
- let buffer = '';
55
-
56
- process.stdin.setEncoding('utf8');
57
- process.stdin.on('data', async (chunk) => {
58
- buffer += chunk;
59
-
60
- if (buffer.length > MCP_MAX_BUFFER_BYTES) {
61
- console.log(JSON.stringify({
62
- jsonrpc: '2.0',
63
- id: null,
64
- error: {
65
- code: -32700,
66
- message: `Buffered stdin exceeds ${MCP_MAX_BUFFER_BYTES} bytes without newline; resetting`,
67
- },
68
- }));
69
- buffer = '';
70
- return;
71
- }
72
-
73
- // Process complete JSON messages (newline-delimited)
74
- let lines = buffer.split('\n');
75
- buffer = lines.pop() || ''; // Keep incomplete line in buffer
76
-
77
- for (const line of lines) {
78
- if (line.trim()) {
79
- try {
80
- const message = JSON.parse(line);
81
- const response = await handleMessage(message);
82
- if (response) {
83
- console.log(JSON.stringify(response));
84
- }
85
- } catch (error) {
86
- console.error(
87
- `[${new Date().toISOString()}] ERROR [claude-flow-mcp] Failed to parse:`,
88
- error instanceof Error ? error.message : String(error)
89
- );
90
- // Send parse error response
91
- console.log(JSON.stringify({
92
- jsonrpc: '2.0',
93
- id: null,
94
- error: { code: -32700, message: 'Parse error' },
95
- }));
96
- }
97
- }
98
- }
99
- });
100
-
101
- process.stdin.on('end', () => {
102
- console.error(
103
- `[${new Date().toISOString()}] INFO [claude-flow-mcp] (${sessionId}) stdin closed, shutting down...`
104
- );
105
- process.exit(0);
106
- });
107
-
108
- // Handle process termination
109
- process.on('SIGINT', () => {
110
- console.error(`[${new Date().toISOString()}] INFO [claude-flow-mcp] Received SIGINT`);
111
- process.exit(0);
112
- });
113
-
114
- process.on('SIGTERM', () => {
115
- console.error(`[${new Date().toISOString()}] INFO [claude-flow-mcp] Received SIGTERM`);
116
- process.exit(0);
117
- });
118
-
119
- /**
120
- * Handle MCP message
121
- */
122
- async function handleMessage(message) {
123
- if (!message.method) {
124
- return {
125
- jsonrpc: '2.0',
126
- id: message.id,
127
- error: { code: -32600, message: 'Invalid Request: missing method' },
128
- };
129
- }
130
-
131
- const params = message.params || {};
132
-
133
- try {
134
- switch (message.method) {
135
- case 'initialize':
136
- return {
137
- jsonrpc: '2.0',
138
- id: message.id,
139
- result: {
140
- protocolVersion: '2024-11-05',
141
- serverInfo: { name: 'ruflo', version: VERSION },
142
- capabilities: {
143
- tools: { listChanged: true },
144
- resources: { subscribe: true, listChanged: true },
145
- },
146
- },
147
- };
148
-
149
- case 'tools/list': {
150
- const tools = listMCPTools();
151
- return {
152
- jsonrpc: '2.0',
153
- id: message.id,
154
- result: {
155
- tools: tools.map(tool => ({
156
- name: tool.name,
157
- description: tool.description,
158
- inputSchema: tool.inputSchema,
159
- })),
160
- },
161
- };
162
- }
163
-
164
- case 'tools/call': {
165
- const toolName = params.name;
166
- const toolParams = params.arguments || {};
167
-
168
- if (!hasTool(toolName)) {
169
- return {
170
- jsonrpc: '2.0',
171
- id: message.id,
172
- error: { code: -32601, message: `Tool not found: ${toolName}` },
173
- };
174
- }
175
-
176
- try {
177
- const result = await callMCPTool(toolName, toolParams, { sessionId });
178
- return {
179
- jsonrpc: '2.0',
180
- id: message.id,
181
- result: { content: [{ type: 'text', text: JSON.stringify(result, null, 2) }] },
182
- };
183
- } catch (error) {
184
- return {
185
- jsonrpc: '2.0',
186
- id: message.id,
187
- error: {
188
- code: -32603,
189
- message: error instanceof Error ? error.message : 'Tool execution failed',
190
- },
191
- };
192
- }
193
- }
194
-
195
- case 'notifications/initialized':
196
- console.error(`[${new Date().toISOString()}] INFO [claude-flow-mcp] Client initialized`);
197
- return null; // No response for notifications
198
-
199
- case 'ping':
200
- return {
201
- jsonrpc: '2.0',
202
- id: message.id,
203
- result: {},
204
- };
205
-
206
- default:
207
- return {
208
- jsonrpc: '2.0',
209
- id: message.id,
210
- error: { code: -32601, message: `Method not found: ${message.method}` },
211
- };
212
- }
213
- } catch (error) {
214
- console.error(`[${new Date().toISOString()}] ERROR [claude-flow-mcp] ${message.method}:`, error);
215
- return {
216
- jsonrpc: '2.0',
217
- id: message.id,
218
- error: {
219
- code: -32603,
220
- message: error instanceof Error ? error.message : 'Internal error',
221
- },
222
- };
223
- }
224
- }
1
+ #!/usr/bin/env node
2
+ /**
3
+ * @claude-flow/cli - MCP Server Entry Point
4
+ *
5
+ * Direct stdio MCP server for Claude Code integration.
6
+ * This entry point handles stdin/stdout directly for MCP protocol
7
+ * without any CLI formatting output that would corrupt the protocol.
8
+ */
9
+
10
+ import { randomUUID } from 'crypto';
11
+
12
+ // Suppress the SPECIFIC cosmetic "[AgentDB Patch] Controller index not found"
13
+ // noise. Tight match (both prefix AND "Controller index not found") so other
14
+ // [AgentDB Patch] warnings about real issues still flow through. Also patch
15
+ // console.log because the underlying call site uses it. See bin/cli.js for
16
+ // the same rationale.
17
+ const _origWarn = console.warn;
18
+ const _origLog = console.log;
19
+ const _isCosmeticAgentdbPatchNoise = (msg) =>
20
+ msg.includes('[AgentDB Patch]') && msg.includes('Controller index not found');
21
+ console.warn = (...args) => {
22
+ if (_isCosmeticAgentdbPatchNoise(String(args[0] ?? ''))) return;
23
+ _origWarn.apply(console, args);
24
+ };
25
+ console.log = (...args) => {
26
+ if (_isCosmeticAgentdbPatchNoise(String(args[0] ?? ''))) return;
27
+ _origLog.apply(console, args);
28
+ };
29
+
30
+ import { listMCPTools, callMCPTool, hasTool } from '../dist/src/mcp-client.js';
31
+
32
+ const VERSION = '3.0.0';
33
+ const sessionId = `mcp-${Date.now()}-${randomUUID().slice(0, 8)}`;
34
+
35
+ // Log to stderr (doesn't corrupt stdout for MCP protocol)
36
+ console.error(
37
+ `[${new Date().toISOString()}] INFO [claude-flow-mcp] (${sessionId}) Starting in stdio mode`
38
+ );
39
+ console.error(JSON.stringify({
40
+ arch: process.arch,
41
+ mode: 'mcp-stdio',
42
+ nodeVersion: process.version,
43
+ pid: process.pid,
44
+ platform: process.platform,
45
+ protocol: 'stdio',
46
+ sessionId,
47
+ version: VERSION,
48
+ }));
49
+
50
+ // Handle stdin messages
51
+ // Audit-flagged DoS protection (audit_1776483149979): cap stdin buffer
52
+ // to 10MB. See bin/cli.js for the same protection on the auto-detect path.
53
+ const MCP_MAX_BUFFER_BYTES = 10 * 1024 * 1024;
54
+ let buffer = '';
55
+
56
+ process.stdin.setEncoding('utf8');
57
+ process.stdin.on('data', async (chunk) => {
58
+ buffer += chunk;
59
+
60
+ if (buffer.length > MCP_MAX_BUFFER_BYTES) {
61
+ console.log(JSON.stringify({
62
+ jsonrpc: '2.0',
63
+ id: null,
64
+ error: {
65
+ code: -32700,
66
+ message: `Buffered stdin exceeds ${MCP_MAX_BUFFER_BYTES} bytes without newline; resetting`,
67
+ },
68
+ }));
69
+ buffer = '';
70
+ return;
71
+ }
72
+
73
+ // Process complete JSON messages (newline-delimited)
74
+ let lines = buffer.split('\n');
75
+ buffer = lines.pop() || ''; // Keep incomplete line in buffer
76
+
77
+ for (const line of lines) {
78
+ if (line.trim()) {
79
+ try {
80
+ const message = JSON.parse(line);
81
+ const response = await handleMessage(message);
82
+ if (response) {
83
+ console.log(JSON.stringify(response));
84
+ }
85
+ } catch (error) {
86
+ console.error(
87
+ `[${new Date().toISOString()}] ERROR [claude-flow-mcp] Failed to parse:`,
88
+ error instanceof Error ? error.message : String(error)
89
+ );
90
+ // Send parse error response
91
+ console.log(JSON.stringify({
92
+ jsonrpc: '2.0',
93
+ id: null,
94
+ error: { code: -32700, message: 'Parse error' },
95
+ }));
96
+ }
97
+ }
98
+ }
99
+ });
100
+
101
+ process.stdin.on('end', () => {
102
+ console.error(
103
+ `[${new Date().toISOString()}] INFO [claude-flow-mcp] (${sessionId}) stdin closed, shutting down...`
104
+ );
105
+ process.exit(0);
106
+ });
107
+
108
+ // Handle process termination
109
+ process.on('SIGINT', () => {
110
+ console.error(`[${new Date().toISOString()}] INFO [claude-flow-mcp] Received SIGINT`);
111
+ process.exit(0);
112
+ });
113
+
114
+ process.on('SIGTERM', () => {
115
+ console.error(`[${new Date().toISOString()}] INFO [claude-flow-mcp] Received SIGTERM`);
116
+ process.exit(0);
117
+ });
118
+
119
+ /**
120
+ * Handle MCP message
121
+ */
122
+ async function handleMessage(message) {
123
+ if (!message.method) {
124
+ return {
125
+ jsonrpc: '2.0',
126
+ id: message.id,
127
+ error: { code: -32600, message: 'Invalid Request: missing method' },
128
+ };
129
+ }
130
+
131
+ const params = message.params || {};
132
+
133
+ try {
134
+ switch (message.method) {
135
+ case 'initialize':
136
+ return {
137
+ jsonrpc: '2.0',
138
+ id: message.id,
139
+ result: {
140
+ protocolVersion: '2024-11-05',
141
+ serverInfo: { name: 'ruflo', version: VERSION },
142
+ capabilities: {
143
+ tools: { listChanged: true },
144
+ resources: { subscribe: true, listChanged: true },
145
+ },
146
+ },
147
+ };
148
+
149
+ case 'tools/list': {
150
+ const tools = listMCPTools();
151
+ return {
152
+ jsonrpc: '2.0',
153
+ id: message.id,
154
+ result: {
155
+ tools: tools.map(tool => ({
156
+ name: tool.name,
157
+ description: tool.description,
158
+ inputSchema: tool.inputSchema,
159
+ })),
160
+ },
161
+ };
162
+ }
163
+
164
+ case 'tools/call': {
165
+ const toolName = params.name;
166
+ const toolParams = params.arguments || {};
167
+
168
+ if (!hasTool(toolName)) {
169
+ return {
170
+ jsonrpc: '2.0',
171
+ id: message.id,
172
+ error: { code: -32601, message: `Tool not found: ${toolName}` },
173
+ };
174
+ }
175
+
176
+ try {
177
+ const result = await callMCPTool(toolName, toolParams, { sessionId });
178
+ return {
179
+ jsonrpc: '2.0',
180
+ id: message.id,
181
+ result: { content: [{ type: 'text', text: JSON.stringify(result, null, 2) }] },
182
+ };
183
+ } catch (error) {
184
+ return {
185
+ jsonrpc: '2.0',
186
+ id: message.id,
187
+ error: {
188
+ code: -32603,
189
+ message: error instanceof Error ? error.message : 'Tool execution failed',
190
+ },
191
+ };
192
+ }
193
+ }
194
+
195
+ case 'notifications/initialized':
196
+ console.error(`[${new Date().toISOString()}] INFO [claude-flow-mcp] Client initialized`);
197
+ return null; // No response for notifications
198
+
199
+ case 'ping':
200
+ return {
201
+ jsonrpc: '2.0',
202
+ id: message.id,
203
+ result: {},
204
+ };
205
+
206
+ default:
207
+ return {
208
+ jsonrpc: '2.0',
209
+ id: message.id,
210
+ error: { code: -32601, message: `Method not found: ${message.method}` },
211
+ };
212
+ }
213
+ } catch (error) {
214
+ console.error(`[${new Date().toISOString()}] ERROR [claude-flow-mcp] ${message.method}:`, error);
215
+ return {
216
+ jsonrpc: '2.0',
217
+ id: message.id,
218
+ error: {
219
+ code: -32603,
220
+ message: error instanceof Error ? error.message : 'Internal error',
221
+ },
222
+ };
223
+ }
224
+ }
@@ -1,2 +1,2 @@
1
- #!/usr/bin/env node
2
- // Preinstall hook — no-op (cache repair removed in 3.1.0-alpha.53)
1
+ #!/usr/bin/env node
2
+ // Preinstall hook — no-op (cache repair removed in 3.1.0-alpha.53)
@@ -1,8 +1,8 @@
1
1
  {
2
2
  "schemaVersion": 1,
3
3
  "generation": 1,
4
- "generatedAt": "2026-07-14T04:41:45.824Z",
5
- "gitSha": "b254e321",
4
+ "generatedAt": "2026-07-14T16:52:03.409Z",
5
+ "gitSha": "0052b1b0",
6
6
  "catalog": {
7
7
  "agents": 164,
8
8
  "tools": 387,
@@ -70,25 +70,25 @@ function resolveApiKey(override) {
70
70
  * Instructs Sonnet to act as an adversarial reviewer and respond in JSON.
71
71
  */
72
72
  function buildCriticSystemPrompt() {
73
- return `You are an adversarial reviewer of agent answers on the GAIA benchmark. \
74
- Your job is to find flaws in candidate answers before they are submitted.
75
-
76
- Given a question, a candidate answer, and a summary of the agent's reasoning \
77
- trajectory, you must decide whether the candidate answer is correct.
78
-
79
- Evaluation criteria:
80
- 1. Does the answer directly address what the question asks?
81
- 2. Is there evidence in the trajectory that supports the answer?
82
- 3. Are there obvious flaws (wrong unit, wrong format, missed constraint, \
83
- fabricated source, off-by-one error, truncation)?
84
- 4. For numeric answers: is the magnitude and unit plausible?
85
- 5. For list answers: are all required items present and correctly ordered?
86
-
87
- Respond ONLY with a JSON object — no markdown, no prose outside the JSON:
88
- {"verdict":"pass"|"fail"|"uncertain","reasoning":"<1-3 sentences>","suggestedRevision":"<corrected answer or empty string>"}
89
-
90
- Use "uncertain" only when you genuinely cannot determine correctness from the \
91
- available information. Never use "uncertain" to avoid making a decision when \
73
+ return `You are an adversarial reviewer of agent answers on the GAIA benchmark. \
74
+ Your job is to find flaws in candidate answers before they are submitted.
75
+
76
+ Given a question, a candidate answer, and a summary of the agent's reasoning \
77
+ trajectory, you must decide whether the candidate answer is correct.
78
+
79
+ Evaluation criteria:
80
+ 1. Does the answer directly address what the question asks?
81
+ 2. Is there evidence in the trajectory that supports the answer?
82
+ 3. Are there obvious flaws (wrong unit, wrong format, missed constraint, \
83
+ fabricated source, off-by-one error, truncation)?
84
+ 4. For numeric answers: is the magnitude and unit plausible?
85
+ 5. For list answers: are all required items present and correctly ordered?
86
+
87
+ Respond ONLY with a JSON object — no markdown, no prose outside the JSON:
88
+ {"verdict":"pass"|"fail"|"uncertain","reasoning":"<1-3 sentences>","suggestedRevision":"<corrected answer or empty string>"}
89
+
90
+ Use "uncertain" only when you genuinely cannot determine correctness from the \
91
+ available information. Never use "uncertain" to avoid making a decision when \
92
92
  evidence is available.`;
93
93
  }
94
94
  /**
@@ -108,11 +108,11 @@ function buildCriticUserMessage(question, candidateAnswer, trajectory) {
108
108
  const result = s.result ? ` → ${s.result.slice(0, 200)}` : '';
109
109
  return `${label}${result}`;
110
110
  }).join('\n');
111
- return `QUESTION: ${question.question}
112
-
113
- CANDIDATE ANSWER: ${candidateAnswer}
114
-
115
- AGENT TRAJECTORY (${trajectory.turns} turns, summarised):
111
+ return `QUESTION: ${question.question}
112
+
113
+ CANDIDATE ANSWER: ${candidateAnswer}
114
+
115
+ AGENT TRAJECTORY (${trajectory.turns} turns, summarised):
116
116
  ${trajectoryText}`;
117
117
  }
118
118
  /**