cognitive-core 0.2.4 → 0.3.1

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 (1278) hide show
  1. package/LICENSE +21 -0
  2. package/README.md +88 -143
  3. package/dist/agents/mock-provider.d.ts.map +1 -1
  4. package/dist/agents/mock-provider.js.map +1 -1
  5. package/dist/agents/types.d.ts.map +1 -1
  6. package/dist/agents/types.js +5 -4
  7. package/dist/agents/types.js.map +1 -1
  8. package/dist/atlas.d.ts +309 -97
  9. package/dist/atlas.d.ts.map +1 -1
  10. package/dist/atlas.js +743 -277
  11. package/dist/atlas.js.map +1 -1
  12. package/dist/bin/cli-utils.d.ts +25 -2
  13. package/dist/bin/cli-utils.d.ts.map +1 -1
  14. package/dist/bin/cli-utils.js +79 -9
  15. package/dist/bin/cli-utils.js.map +1 -1
  16. package/dist/bin/cognitive-core.js +32 -0
  17. package/dist/bin/cognitive-core.js.map +1 -1
  18. package/dist/bin/commands/dream.d.ts +17 -0
  19. package/dist/bin/commands/dream.d.ts.map +1 -0
  20. package/dist/bin/commands/dream.js +168 -0
  21. package/dist/bin/commands/dream.js.map +1 -0
  22. package/dist/bin/commands/eval.d.ts +181 -0
  23. package/dist/bin/commands/eval.d.ts.map +1 -0
  24. package/dist/bin/commands/eval.js +1582 -0
  25. package/dist/bin/commands/eval.js.map +1 -0
  26. package/dist/bin/commands/kb.js +0 -8
  27. package/dist/bin/commands/kb.js.map +1 -1
  28. package/dist/bin/commands/learn.js +12 -4
  29. package/dist/bin/commands/learn.js.map +1 -1
  30. package/dist/bin/commands/legacy.d.ts.map +1 -1
  31. package/dist/bin/commands/legacy.js +1 -1
  32. package/dist/bin/commands/legacy.js.map +1 -1
  33. package/dist/bin/commands/run.d.ts.map +1 -1
  34. package/dist/bin/commands/run.js +10 -9
  35. package/dist/bin/commands/run.js.map +1 -1
  36. package/dist/bin/commands/sessions.d.ts.map +1 -1
  37. package/dist/bin/commands/sessions.js.map +1 -1
  38. package/dist/bin/commands/skills.d.ts.map +1 -1
  39. package/dist/bin/commands/skills.js +283 -1
  40. package/dist/bin/commands/skills.js.map +1 -1
  41. package/dist/embeddings/bm25.d.ts.map +1 -1
  42. package/dist/embeddings/bm25.js +1 -2
  43. package/dist/embeddings/bm25.js.map +1 -1
  44. package/dist/embeddings/index.d.ts +3 -3
  45. package/dist/embeddings/index.d.ts.map +1 -1
  46. package/dist/embeddings/index.js +3 -3
  47. package/dist/embeddings/index.js.map +1 -1
  48. package/dist/embeddings/inverted-index.js +62 -7
  49. package/dist/embeddings/inverted-index.js.map +1 -1
  50. package/dist/embeddings/manager.d.ts.map +1 -1
  51. package/dist/embeddings/manager.js +6 -5
  52. package/dist/embeddings/manager.js.map +1 -1
  53. package/dist/embeddings/provider.d.ts +96 -5
  54. package/dist/embeddings/provider.d.ts.map +1 -1
  55. package/dist/embeddings/provider.js +171 -48
  56. package/dist/embeddings/provider.js.map +1 -1
  57. package/dist/embeddings/vector-store.d.ts.map +1 -1
  58. package/dist/embeddings/vector-store.js +3 -1
  59. package/dist/embeddings/vector-store.js.map +1 -1
  60. package/dist/eval/ablation-validate.d.ts +12 -0
  61. package/dist/eval/ablation-validate.d.ts.map +1 -0
  62. package/dist/eval/ablation-validate.js +78 -0
  63. package/dist/eval/ablation-validate.js.map +1 -0
  64. package/dist/eval/ablation.d.ts +94 -0
  65. package/dist/eval/ablation.d.ts.map +1 -0
  66. package/dist/eval/ablation.js +161 -0
  67. package/dist/eval/ablation.js.map +1 -0
  68. package/dist/eval/ci-smoke.d.ts +11 -0
  69. package/dist/eval/ci-smoke.d.ts.map +1 -0
  70. package/dist/eval/ci-smoke.js +49 -0
  71. package/dist/eval/ci-smoke.js.map +1 -0
  72. package/dist/eval/diagnostics.d.ts +214 -0
  73. package/dist/eval/diagnostics.d.ts.map +1 -0
  74. package/dist/eval/diagnostics.js +107 -0
  75. package/dist/eval/diagnostics.js.map +1 -0
  76. package/dist/eval/index.d.ts +27 -0
  77. package/dist/eval/index.d.ts.map +1 -0
  78. package/dist/eval/index.js +27 -0
  79. package/dist/eval/index.js.map +1 -0
  80. package/dist/eval/live/alfworld.d.ts +248 -0
  81. package/dist/eval/live/alfworld.d.ts.map +1 -0
  82. package/dist/eval/live/alfworld.js +637 -0
  83. package/dist/eval/live/alfworld.js.map +1 -0
  84. package/dist/eval/live/alfworld_h1.d.ts +16 -0
  85. package/dist/eval/live/alfworld_h1.d.ts.map +1 -0
  86. package/dist/eval/live/alfworld_h1.js +75 -0
  87. package/dist/eval/live/alfworld_h1.js.map +1 -0
  88. package/dist/eval/live/alfworld_memory.d.ts +38 -0
  89. package/dist/eval/live/alfworld_memory.d.ts.map +1 -0
  90. package/dist/eval/live/alfworld_memory.js +184 -0
  91. package/dist/eval/live/alfworld_memory.js.map +1 -0
  92. package/dist/eval/live/alfworld_strategy.d.ts +65 -0
  93. package/dist/eval/live/alfworld_strategy.d.ts.map +1 -0
  94. package/dist/eval/live/alfworld_strategy.js +89 -0
  95. package/dist/eval/live/alfworld_strategy.js.map +1 -0
  96. package/dist/eval/live/azure-openai-llm-provider.d.ts +41 -0
  97. package/dist/eval/live/azure-openai-llm-provider.d.ts.map +1 -0
  98. package/dist/eval/live/azure-openai-llm-provider.js +92 -0
  99. package/dist/eval/live/azure-openai-llm-provider.js.map +1 -0
  100. package/dist/eval/live/bridges/skillflow_e2b_bridge.d.ts +3 -0
  101. package/dist/eval/live/bridges/skillflow_e2b_bridge.d.ts.map +1 -0
  102. package/dist/eval/live/bridges/skillflow_e2b_bridge.js +12 -0
  103. package/dist/eval/live/bridges/skillflow_e2b_bridge.js.map +1 -0
  104. package/dist/eval/live/claude-cli-llm-provider.d.ts +53 -0
  105. package/dist/eval/live/claude-cli-llm-provider.d.ts.map +1 -0
  106. package/dist/eval/live/claude-cli-llm-provider.js +71 -0
  107. package/dist/eval/live/claude-cli-llm-provider.js.map +1 -0
  108. package/dist/eval/live/claude-cli.d.ts +96 -0
  109. package/dist/eval/live/claude-cli.d.ts.map +1 -0
  110. package/dist/eval/live/claude-cli.js +314 -0
  111. package/dist/eval/live/claude-cli.js.map +1 -0
  112. package/dist/eval/live/claude-delegate.d.ts +41 -0
  113. package/dist/eval/live/claude-delegate.d.ts.map +1 -0
  114. package/dist/eval/live/claude-delegate.js +59 -0
  115. package/dist/eval/live/claude-delegate.js.map +1 -0
  116. package/dist/eval/live/claude-workspace-backend.d.ts +24 -0
  117. package/dist/eval/live/claude-workspace-backend.d.ts.map +1 -0
  118. package/dist/eval/live/claude-workspace-backend.js +76 -0
  119. package/dist/eval/live/claude-workspace-backend.js.map +1 -0
  120. package/dist/eval/live/clbench.d.ts +64 -0
  121. package/dist/eval/live/clbench.d.ts.map +1 -0
  122. package/dist/eval/live/clbench.js +291 -0
  123. package/dist/eval/live/clbench.js.map +1 -0
  124. package/dist/eval/live/cold-reuse.d.ts +21 -0
  125. package/dist/eval/live/cold-reuse.d.ts.map +1 -0
  126. package/dist/eval/live/cold-reuse.js +102 -0
  127. package/dist/eval/live/cold-reuse.js.map +1 -0
  128. package/dist/eval/live/difficulty-registry.d.ts +68 -0
  129. package/dist/eval/live/difficulty-registry.d.ts.map +1 -0
  130. package/dist/eval/live/difficulty-registry.js +169 -0
  131. package/dist/eval/live/difficulty-registry.js.map +1 -0
  132. package/dist/eval/live/docker.d.ts +52 -0
  133. package/dist/eval/live/docker.d.ts.map +1 -0
  134. package/dist/eval/live/docker.js +164 -0
  135. package/dist/eval/live/docker.js.map +1 -0
  136. package/dist/eval/live/env-bridge.d.ts +68 -0
  137. package/dist/eval/live/env-bridge.d.ts.map +1 -0
  138. package/dist/eval/live/env-bridge.js +188 -0
  139. package/dist/eval/live/env-bridge.js.map +1 -0
  140. package/dist/eval/live/gsm8k.d.ts +44 -0
  141. package/dist/eval/live/gsm8k.d.ts.map +1 -0
  142. package/dist/eval/live/gsm8k.js +142 -0
  143. package/dist/eval/live/gsm8k.js.map +1 -0
  144. package/dist/eval/live/index.d.ts +38 -0
  145. package/dist/eval/live/index.d.ts.map +1 -0
  146. package/dist/eval/live/index.js +38 -0
  147. package/dist/eval/live/index.js.map +1 -0
  148. package/dist/eval/live/lifelongagent.d.ts +64 -0
  149. package/dist/eval/live/lifelongagent.d.ts.map +1 -0
  150. package/dist/eval/live/lifelongagent.js +216 -0
  151. package/dist/eval/live/lifelongagent.js.map +1 -0
  152. package/dist/eval/live/lift-registry.d.ts +128 -0
  153. package/dist/eval/live/lift-registry.d.ts.map +1 -0
  154. package/dist/eval/live/lift-registry.js +147 -0
  155. package/dist/eval/live/lift-registry.js.map +1 -0
  156. package/dist/eval/live/live-benchmark.d.ts +84 -0
  157. package/dist/eval/live/live-benchmark.d.ts.map +1 -0
  158. package/dist/eval/live/live-benchmark.js +2 -0
  159. package/dist/eval/live/live-benchmark.js.map +1 -0
  160. package/dist/eval/live/live-runner.d.ts +609 -0
  161. package/dist/eval/live/live-runner.d.ts.map +1 -0
  162. package/dist/eval/live/live-runner.js +1409 -0
  163. package/dist/eval/live/live-runner.js.map +1 -0
  164. package/dist/eval/live/refine.d.ts +88 -0
  165. package/dist/eval/live/refine.d.ts.map +1 -0
  166. package/dist/eval/live/refine.js +135 -0
  167. package/dist/eval/live/refine.js.map +1 -0
  168. package/dist/eval/live/registry-store.d.ts +109 -0
  169. package/dist/eval/live/registry-store.d.ts.map +1 -0
  170. package/dist/eval/live/registry-store.js +272 -0
  171. package/dist/eval/live/registry-store.js.map +1 -0
  172. package/dist/eval/live/run-checkpoint.d.ts +55 -0
  173. package/dist/eval/live/run-checkpoint.d.ts.map +1 -0
  174. package/dist/eval/live/run-checkpoint.js +86 -0
  175. package/dist/eval/live/run-checkpoint.js.map +1 -0
  176. package/dist/eval/live/run-resume.d.ts +22 -0
  177. package/dist/eval/live/run-resume.d.ts.map +1 -0
  178. package/dist/eval/live/run-resume.js +70 -0
  179. package/dist/eval/live/run-resume.js.map +1 -0
  180. package/dist/eval/live/skill-injector.d.ts +36 -0
  181. package/dist/eval/live/skill-injector.d.ts.map +1 -0
  182. package/dist/eval/live/skill-injector.js +81 -0
  183. package/dist/eval/live/skill-injector.js.map +1 -0
  184. package/dist/eval/live/skillflow-embedded-workbook-bundle.d.ts +3 -0
  185. package/dist/eval/live/skillflow-embedded-workbook-bundle.d.ts.map +1 -0
  186. package/dist/eval/live/skillflow-embedded-workbook-bundle.js +391 -0
  187. package/dist/eval/live/skillflow-embedded-workbook-bundle.js.map +1 -0
  188. package/dist/eval/live/skillflow.d.ts +154 -0
  189. package/dist/eval/live/skillflow.d.ts.map +1 -0
  190. package/dist/eval/live/skillflow.js +612 -0
  191. package/dist/eval/live/skillflow.js.map +1 -0
  192. package/dist/eval/live/skilllearnbench.d.ts +43 -0
  193. package/dist/eval/live/skilllearnbench.d.ts.map +1 -0
  194. package/dist/eval/live/skilllearnbench.js +154 -0
  195. package/dist/eval/live/skilllearnbench.js.map +1 -0
  196. package/dist/eval/live/skillopt_v1.d.ts +96 -0
  197. package/dist/eval/live/skillopt_v1.d.ts.map +1 -0
  198. package/dist/eval/live/skillopt_v1.js +0 -0
  199. package/dist/eval/live/skillopt_v1.js.map +1 -0
  200. package/dist/eval/live/skillopt_v1_causal.d.ts +26 -0
  201. package/dist/eval/live/skillopt_v1_causal.d.ts.map +1 -0
  202. package/dist/eval/live/skillopt_v1_causal.js +106 -0
  203. package/dist/eval/live/skillopt_v1_causal.js.map +1 -0
  204. package/dist/eval/live/skillopt_v1_memory.d.ts +53 -0
  205. package/dist/eval/live/skillopt_v1_memory.d.ts.map +1 -0
  206. package/dist/eval/live/skillopt_v1_memory.js +147 -0
  207. package/dist/eval/live/skillopt_v1_memory.js.map +1 -0
  208. package/dist/eval/live/skillopt_v1_rules.d.ts +39 -0
  209. package/dist/eval/live/skillopt_v1_rules.d.ts.map +1 -0
  210. package/dist/eval/live/skillopt_v1_rules.js +91 -0
  211. package/dist/eval/live/skillopt_v1_rules.js.map +1 -0
  212. package/dist/eval/live/skillsbench.d.ts +138 -0
  213. package/dist/eval/live/skillsbench.d.ts.map +1 -0
  214. package/dist/eval/live/skillsbench.js +626 -0
  215. package/dist/eval/live/skillsbench.js.map +1 -0
  216. package/dist/eval/live/stopping.d.ts +69 -0
  217. package/dist/eval/live/stopping.d.ts.map +1 -0
  218. package/dist/eval/live/stopping.js +213 -0
  219. package/dist/eval/live/stopping.js.map +1 -0
  220. package/dist/eval/live/trace.d.ts +20 -0
  221. package/dist/eval/live/trace.d.ts.map +1 -0
  222. package/dist/eval/live/trace.js +41 -0
  223. package/dist/eval/live/trace.js.map +1 -0
  224. package/dist/eval/mechanism-trace-adapter.d.ts +8 -0
  225. package/dist/eval/mechanism-trace-adapter.d.ts.map +1 -0
  226. package/dist/eval/mechanism-trace-adapter.js +118 -0
  227. package/dist/eval/mechanism-trace-adapter.js.map +1 -0
  228. package/dist/eval/memory-ab.d.ts +61 -0
  229. package/dist/eval/memory-ab.d.ts.map +1 -0
  230. package/dist/eval/memory-ab.js +161 -0
  231. package/dist/eval/memory-ab.js.map +1 -0
  232. package/dist/eval/memory-mock.d.ts +54 -0
  233. package/dist/eval/memory-mock.d.ts.map +1 -0
  234. package/dist/eval/memory-mock.js +129 -0
  235. package/dist/eval/memory-mock.js.map +1 -0
  236. package/dist/eval/memory-qa/index.d.ts +13 -0
  237. package/dist/eval/memory-qa/index.d.ts.map +1 -0
  238. package/dist/eval/memory-qa/index.js +13 -0
  239. package/dist/eval/memory-qa/index.js.map +1 -0
  240. package/dist/eval/memory-qa/iterative.d.ts +80 -0
  241. package/dist/eval/memory-qa/iterative.d.ts.map +1 -0
  242. package/dist/eval/memory-qa/iterative.js +107 -0
  243. package/dist/eval/memory-qa/iterative.js.map +1 -0
  244. package/dist/eval/memory-qa/loaders.d.ts +11 -0
  245. package/dist/eval/memory-qa/loaders.d.ts.map +1 -0
  246. package/dist/eval/memory-qa/loaders.js +21 -0
  247. package/dist/eval/memory-qa/loaders.js.map +1 -0
  248. package/dist/eval/memory-qa/metrics.d.ts +3 -0
  249. package/dist/eval/memory-qa/metrics.d.ts.map +1 -0
  250. package/dist/eval/memory-qa/metrics.js +2 -0
  251. package/dist/eval/memory-qa/metrics.js.map +1 -0
  252. package/dist/eval/memory-qa/qa.d.ts +186 -0
  253. package/dist/eval/memory-qa/qa.d.ts.map +1 -0
  254. package/dist/eval/memory-qa/qa.js +424 -0
  255. package/dist/eval/memory-qa/qa.js.map +1 -0
  256. package/dist/eval/memory-qa/retrieval.d.ts +33 -0
  257. package/dist/eval/memory-qa/retrieval.d.ts.map +1 -0
  258. package/dist/eval/memory-qa/retrieval.js +44 -0
  259. package/dist/eval/memory-qa/retrieval.js.map +1 -0
  260. package/dist/eval/memory-qa/serving.d.ts +60 -0
  261. package/dist/eval/memory-qa/serving.d.ts.map +1 -0
  262. package/dist/eval/memory-qa/serving.js +267 -0
  263. package/dist/eval/memory-qa/serving.js.map +1 -0
  264. package/dist/eval/memory-qa/types.d.ts +2 -0
  265. package/dist/eval/memory-qa/types.d.ts.map +1 -0
  266. package/dist/eval/memory-qa/types.js +2 -0
  267. package/dist/eval/memory-qa/types.js.map +1 -0
  268. package/dist/eval/metrics.d.ts +120 -0
  269. package/dist/eval/metrics.d.ts.map +1 -0
  270. package/dist/eval/metrics.js +361 -0
  271. package/dist/eval/metrics.js.map +1 -0
  272. package/dist/eval/oracle-smoke.d.ts +75 -0
  273. package/dist/eval/oracle-smoke.d.ts.map +1 -0
  274. package/dist/eval/oracle-smoke.js +143 -0
  275. package/dist/eval/oracle-smoke.js.map +1 -0
  276. package/dist/eval/playbook-calibration-report.d.ts +40 -0
  277. package/dist/eval/playbook-calibration-report.d.ts.map +1 -0
  278. package/dist/eval/playbook-calibration-report.js +246 -0
  279. package/dist/eval/playbook-calibration-report.js.map +1 -0
  280. package/dist/eval/playbook-candidate-decisions-report.d.ts +36 -0
  281. package/dist/eval/playbook-candidate-decisions-report.d.ts.map +1 -0
  282. package/dist/eval/playbook-candidate-decisions-report.js +201 -0
  283. package/dist/eval/playbook-candidate-decisions-report.js.map +1 -0
  284. package/dist/eval/playbook-usage-report.d.ts +7 -0
  285. package/dist/eval/playbook-usage-report.d.ts.map +1 -0
  286. package/dist/eval/playbook-usage-report.js +222 -0
  287. package/dist/eval/playbook-usage-report.js.map +1 -0
  288. package/dist/eval/report.d.ts +28 -0
  289. package/dist/eval/report.d.ts.map +1 -0
  290. package/dist/eval/report.js +567 -0
  291. package/dist/eval/report.js.map +1 -0
  292. package/dist/eval/retrieval/corpus.d.ts +41 -0
  293. package/dist/eval/retrieval/corpus.d.ts.map +1 -0
  294. package/dist/eval/retrieval/corpus.js +158 -0
  295. package/dist/eval/retrieval/corpus.js.map +1 -0
  296. package/dist/eval/retrieval/index.d.ts +12 -0
  297. package/dist/eval/retrieval/index.d.ts.map +1 -0
  298. package/dist/eval/retrieval/index.js +14 -0
  299. package/dist/eval/retrieval/index.js.map +1 -0
  300. package/dist/eval/retrieval/knowledge-corpus.d.ts +91 -0
  301. package/dist/eval/retrieval/knowledge-corpus.d.ts.map +1 -0
  302. package/dist/eval/retrieval/knowledge-corpus.js +268 -0
  303. package/dist/eval/retrieval/knowledge-corpus.js.map +1 -0
  304. package/dist/eval/retrieval/knowledge-quality.d.ts +91 -0
  305. package/dist/eval/retrieval/knowledge-quality.d.ts.map +1 -0
  306. package/dist/eval/retrieval/knowledge-quality.js +247 -0
  307. package/dist/eval/retrieval/knowledge-quality.js.map +1 -0
  308. package/dist/eval/retrieval/ladder.d.ts +69 -0
  309. package/dist/eval/retrieval/ladder.d.ts.map +1 -0
  310. package/dist/eval/retrieval/ladder.js +80 -0
  311. package/dist/eval/retrieval/ladder.js.map +1 -0
  312. package/dist/eval/retrieval/quality.d.ts +39 -0
  313. package/dist/eval/retrieval/quality.d.ts.map +1 -0
  314. package/dist/eval/retrieval/quality.js +75 -0
  315. package/dist/eval/retrieval/quality.js.map +1 -0
  316. package/dist/eval/rng.d.ts +25 -0
  317. package/dist/eval/rng.d.ts.map +1 -0
  318. package/dist/eval/rng.js +43 -0
  319. package/dist/eval/rng.js.map +1 -0
  320. package/dist/eval/runner.d.ts +65 -0
  321. package/dist/eval/runner.d.ts.map +1 -0
  322. package/dist/eval/runner.js +216 -0
  323. package/dist/eval/runner.js.map +1 -0
  324. package/dist/eval/skill-bundle.d.ts +18 -0
  325. package/dist/eval/skill-bundle.d.ts.map +1 -0
  326. package/dist/eval/skill-bundle.js +38 -0
  327. package/dist/eval/skill-bundle.js.map +1 -0
  328. package/dist/eval/swarmkit-report.d.ts +21 -0
  329. package/dist/eval/swarmkit-report.d.ts.map +1 -0
  330. package/dist/eval/swarmkit-report.js +93 -0
  331. package/dist/eval/swarmkit-report.js.map +1 -0
  332. package/dist/eval/task-stream.d.ts +22 -0
  333. package/dist/eval/task-stream.d.ts.map +1 -0
  334. package/dist/eval/task-stream.js +99 -0
  335. package/dist/eval/task-stream.js.map +1 -0
  336. package/dist/eval/types.d.ts +173 -0
  337. package/dist/eval/types.d.ts.map +1 -0
  338. package/dist/eval/types.js +11 -0
  339. package/dist/eval/types.js.map +1 -0
  340. package/dist/experimental/causal-store.d.ts +97 -0
  341. package/dist/experimental/causal-store.d.ts.map +1 -0
  342. package/dist/experimental/causal-store.js +213 -0
  343. package/dist/experimental/causal-store.js.map +1 -0
  344. package/dist/experimental/graph-layers/base.d.ts +29 -0
  345. package/dist/experimental/graph-layers/base.d.ts.map +1 -0
  346. package/dist/experimental/graph-layers/base.js +143 -0
  347. package/dist/experimental/graph-layers/base.js.map +1 -0
  348. package/dist/experimental/graph-layers/causal.d.ts +14 -0
  349. package/dist/experimental/graph-layers/causal.d.ts.map +1 -0
  350. package/dist/experimental/graph-layers/causal.js +14 -0
  351. package/dist/experimental/graph-layers/causal.js.map +1 -0
  352. package/dist/experimental/graph-layers/entity.d.ts +14 -0
  353. package/dist/experimental/graph-layers/entity.d.ts.map +1 -0
  354. package/dist/experimental/graph-layers/entity.js +14 -0
  355. package/dist/experimental/graph-layers/entity.js.map +1 -0
  356. package/dist/experimental/graph-layers/index.d.ts +6 -0
  357. package/dist/experimental/graph-layers/index.d.ts.map +1 -0
  358. package/dist/experimental/graph-layers/index.js +6 -0
  359. package/dist/experimental/graph-layers/index.js.map +1 -0
  360. package/dist/experimental/graph-layers/semantic.d.ts +14 -0
  361. package/dist/experimental/graph-layers/semantic.d.ts.map +1 -0
  362. package/dist/experimental/graph-layers/semantic.js +14 -0
  363. package/dist/experimental/graph-layers/semantic.js.map +1 -0
  364. package/dist/experimental/graph-layers/temporal.d.ts +14 -0
  365. package/dist/experimental/graph-layers/temporal.d.ts.map +1 -0
  366. package/dist/experimental/graph-layers/temporal.js +14 -0
  367. package/dist/experimental/graph-layers/temporal.js.map +1 -0
  368. package/dist/experimental/healing-orchestrator.d.ts +182 -0
  369. package/dist/experimental/healing-orchestrator.d.ts.map +1 -0
  370. package/dist/experimental/healing-orchestrator.js +250 -0
  371. package/dist/experimental/healing-orchestrator.js.map +1 -0
  372. package/dist/experimental/knowledge-graph-types.d.ts +148 -0
  373. package/dist/experimental/knowledge-graph-types.d.ts.map +1 -0
  374. package/dist/experimental/knowledge-graph-types.js +40 -0
  375. package/dist/experimental/knowledge-graph-types.js.map +1 -0
  376. package/dist/experimental/knowledge-graph.d.ts +41 -0
  377. package/dist/experimental/knowledge-graph.d.ts.map +1 -0
  378. package/dist/experimental/knowledge-graph.js +270 -0
  379. package/dist/experimental/knowledge-graph.js.map +1 -0
  380. package/dist/experimental/meta-reflection.d.ts +47 -0
  381. package/dist/experimental/meta-reflection.d.ts.map +1 -0
  382. package/dist/experimental/meta-reflection.js +143 -0
  383. package/dist/experimental/meta-reflection.js.map +1 -0
  384. package/dist/experimental/moe-gate.d.ts +134 -0
  385. package/dist/experimental/moe-gate.d.ts.map +1 -0
  386. package/dist/experimental/moe-gate.js +240 -0
  387. package/dist/experimental/moe-gate.js.map +1 -0
  388. package/dist/experimental/reasoning-bank.d.ts +130 -0
  389. package/dist/experimental/reasoning-bank.d.ts.map +1 -0
  390. package/dist/experimental/reasoning-bank.js +427 -0
  391. package/dist/experimental/reasoning-bank.js.map +1 -0
  392. package/dist/experimental/team/team-config.d.ts +127 -0
  393. package/dist/experimental/team/team-config.d.ts.map +1 -0
  394. package/dist/experimental/team/team-config.js +48 -0
  395. package/dist/experimental/team/team-config.js.map +1 -0
  396. package/dist/experimental/team/team-experience.d.ts +300 -0
  397. package/dist/experimental/team/team-experience.d.ts.map +1 -0
  398. package/dist/experimental/team/team-experience.js +355 -0
  399. package/dist/experimental/team/team-experience.js.map +1 -0
  400. package/dist/experimental/team/team-ingester.d.ts +152 -0
  401. package/dist/experimental/team/team-ingester.d.ts.map +1 -0
  402. package/dist/experimental/team/team-ingester.js +330 -0
  403. package/dist/experimental/team/team-ingester.js.map +1 -0
  404. package/dist/experimental/team/team-meta-learner.d.ts +51 -0
  405. package/dist/experimental/team/team-meta-learner.d.ts.map +1 -0
  406. package/dist/experimental/team/team-meta-learner.js +404 -0
  407. package/dist/experimental/team/team-meta-learner.js.map +1 -0
  408. package/dist/experimental/team/team-meta.d.ts +160 -0
  409. package/dist/experimental/team/team-meta.d.ts.map +1 -0
  410. package/dist/experimental/team/team-meta.js +42 -0
  411. package/dist/experimental/team/team-meta.js.map +1 -0
  412. package/dist/experimental/team/team-pipeline.d.ts +77 -0
  413. package/dist/experimental/team/team-pipeline.d.ts.map +1 -0
  414. package/dist/experimental/team/team-pipeline.js +262 -0
  415. package/dist/experimental/team/team-pipeline.js.map +1 -0
  416. package/dist/experimental/team/team-playbook-extraction.d.ts +45 -0
  417. package/dist/experimental/team/team-playbook-extraction.d.ts.map +1 -0
  418. package/dist/experimental/team/team-playbook-extraction.js +526 -0
  419. package/dist/experimental/team/team-playbook-extraction.js.map +1 -0
  420. package/dist/experimental/team/team-playbook.d.ts +276 -0
  421. package/dist/experimental/team/team-playbook.d.ts.map +1 -0
  422. package/dist/experimental/team/team-playbook.js +85 -0
  423. package/dist/experimental/team/team-playbook.js.map +1 -0
  424. package/dist/experimental/team/team-router.d.ts +98 -0
  425. package/dist/experimental/team/team-router.d.ts.map +1 -0
  426. package/dist/experimental/team/team-router.js +312 -0
  427. package/dist/experimental/team/team-router.js.map +1 -0
  428. package/dist/experimental/team/team-skill-library.d.ts +180 -0
  429. package/dist/experimental/team/team-skill-library.d.ts.map +1 -0
  430. package/dist/experimental/team/team-skill-library.js +385 -0
  431. package/dist/experimental/team/team-skill-library.js.map +1 -0
  432. package/dist/experimental/team/team-trajectory-analysis.d.ts +19 -0
  433. package/dist/experimental/team/team-trajectory-analysis.d.ts.map +1 -0
  434. package/dist/experimental/team/team-trajectory-analysis.js +442 -0
  435. package/dist/experimental/team/team-trajectory-analysis.js.map +1 -0
  436. package/dist/experimental/team/team-trajectory.d.ts +305 -0
  437. package/dist/experimental/team/team-trajectory.d.ts.map +1 -0
  438. package/dist/experimental/team/team-trajectory.js +315 -0
  439. package/dist/experimental/team/team-trajectory.js.map +1 -0
  440. package/dist/factory.d.ts +2 -3
  441. package/dist/factory.d.ts.map +1 -1
  442. package/dist/factory.js +12 -11
  443. package/dist/factory.js.map +1 -1
  444. package/dist/index.d.ts +9 -11
  445. package/dist/index.d.ts.map +1 -1
  446. package/dist/index.js +26 -32
  447. package/dist/index.js.map +1 -1
  448. package/dist/learning/analyzer.d.ts +26 -36
  449. package/dist/learning/analyzer.d.ts.map +1 -1
  450. package/dist/learning/analyzer.js +50 -11
  451. package/dist/learning/analyzer.js.map +1 -1
  452. package/dist/learning/compound-bootstrap/external-acquisition.d.ts +63 -0
  453. package/dist/learning/compound-bootstrap/external-acquisition.d.ts.map +1 -0
  454. package/dist/learning/compound-bootstrap/external-acquisition.js +255 -0
  455. package/dist/learning/compound-bootstrap/external-acquisition.js.map +1 -0
  456. package/dist/learning/compound-bootstrap/index.d.ts +3 -0
  457. package/dist/learning/compound-bootstrap/index.d.ts.map +1 -0
  458. package/dist/learning/compound-bootstrap/index.js +3 -0
  459. package/dist/learning/compound-bootstrap/index.js.map +1 -0
  460. package/dist/learning/compound-bootstrap/workspace-analyzer.d.ts +33 -0
  461. package/dist/learning/compound-bootstrap/workspace-analyzer.d.ts.map +1 -0
  462. package/dist/learning/compound-bootstrap/workspace-analyzer.js +379 -0
  463. package/dist/learning/compound-bootstrap/workspace-analyzer.js.map +1 -0
  464. package/dist/learning/consolidation-coordinator.d.ts +49 -0
  465. package/dist/learning/consolidation-coordinator.d.ts.map +1 -0
  466. package/dist/learning/consolidation-coordinator.js +75 -0
  467. package/dist/learning/consolidation-coordinator.js.map +1 -0
  468. package/dist/learning/convention-distiller.d.ts +84 -0
  469. package/dist/learning/convention-distiller.d.ts.map +1 -0
  470. package/dist/learning/convention-distiller.js +268 -0
  471. package/dist/learning/convention-distiller.js.map +1 -0
  472. package/dist/learning/distiller.d.ts +82 -0
  473. package/dist/learning/distiller.d.ts.map +1 -0
  474. package/dist/learning/distiller.js +353 -0
  475. package/dist/learning/distiller.js.map +1 -0
  476. package/dist/learning/effectiveness.d.ts +12 -2
  477. package/dist/learning/effectiveness.d.ts.map +1 -1
  478. package/dist/learning/effectiveness.js +18 -14
  479. package/dist/learning/effectiveness.js.map +1 -1
  480. package/dist/learning/escalation.d.ts +89 -0
  481. package/dist/learning/escalation.d.ts.map +1 -0
  482. package/dist/learning/escalation.js +168 -0
  483. package/dist/learning/escalation.js.map +1 -0
  484. package/dist/learning/external-exposure.d.ts +57 -0
  485. package/dist/learning/external-exposure.d.ts.map +1 -0
  486. package/dist/learning/external-exposure.js +95 -0
  487. package/dist/learning/external-exposure.js.map +1 -0
  488. package/dist/learning/gepa-proposer.d.ts +116 -0
  489. package/dist/learning/gepa-proposer.d.ts.map +1 -0
  490. package/dist/learning/gepa-proposer.js +258 -0
  491. package/dist/learning/gepa-proposer.js.map +1 -0
  492. package/dist/learning/index.d.ts +21 -8
  493. package/dist/learning/index.d.ts.map +1 -1
  494. package/dist/learning/index.js +28 -10
  495. package/dist/learning/index.js.map +1 -1
  496. package/dist/learning/instant-loop.d.ts +46 -16
  497. package/dist/learning/instant-loop.d.ts.map +1 -1
  498. package/dist/learning/instant-loop.js +114 -111
  499. package/dist/learning/instant-loop.js.map +1 -1
  500. package/dist/learning/knowledge-extractor.d.ts.map +1 -1
  501. package/dist/learning/knowledge-extractor.js +1 -2
  502. package/dist/learning/knowledge-extractor.js.map +1 -1
  503. package/dist/learning/llm-extractor.d.ts +88 -0
  504. package/dist/learning/llm-extractor.d.ts.map +1 -0
  505. package/dist/learning/llm-extractor.js +372 -0
  506. package/dist/learning/llm-extractor.js.map +1 -0
  507. package/dist/learning/loop-coordinator.d.ts +61 -0
  508. package/dist/learning/loop-coordinator.d.ts.map +1 -0
  509. package/dist/learning/loop-coordinator.js +96 -0
  510. package/dist/learning/loop-coordinator.js.map +1 -0
  511. package/dist/learning/maintenance-scheduler.d.ts +15 -4
  512. package/dist/learning/maintenance-scheduler.d.ts.map +1 -1
  513. package/dist/learning/maintenance-scheduler.js +19 -16
  514. package/dist/learning/maintenance-scheduler.js.map +1 -1
  515. package/dist/learning/mechanism-trace.d.ts +196 -0
  516. package/dist/learning/mechanism-trace.d.ts.map +1 -0
  517. package/dist/learning/mechanism-trace.js +81 -0
  518. package/dist/learning/mechanism-trace.js.map +1 -0
  519. package/dist/learning/merge-proposer.d.ts +102 -0
  520. package/dist/learning/merge-proposer.d.ts.map +1 -0
  521. package/dist/learning/merge-proposer.js +253 -0
  522. package/dist/learning/merge-proposer.js.map +1 -0
  523. package/dist/learning/meta-learner.d.ts +9 -7
  524. package/dist/learning/meta-learner.d.ts.map +1 -1
  525. package/dist/learning/meta-learner.js +16 -45
  526. package/dist/learning/meta-learner.js.map +1 -1
  527. package/dist/learning/mutation-gate.d.ts +218 -0
  528. package/dist/learning/mutation-gate.d.ts.map +1 -0
  529. package/dist/learning/mutation-gate.js +390 -0
  530. package/dist/learning/mutation-gate.js.map +1 -0
  531. package/dist/learning/oracle-signal.d.ts +67 -0
  532. package/dist/learning/oracle-signal.d.ts.map +1 -0
  533. package/dist/learning/oracle-signal.js +94 -0
  534. package/dist/learning/oracle-signal.js.map +1 -0
  535. package/dist/learning/pipeline.d.ts +31 -4
  536. package/dist/learning/pipeline.d.ts.map +1 -1
  537. package/dist/learning/pipeline.js +64 -12
  538. package/dist/learning/pipeline.js.map +1 -1
  539. package/dist/learning/playbook-extractor.d.ts +8 -0
  540. package/dist/learning/playbook-extractor.d.ts.map +1 -1
  541. package/dist/learning/playbook-extractor.js +119 -34
  542. package/dist/learning/playbook-extractor.js.map +1 -1
  543. package/dist/learning/playbook-optimizer.d.ts +37 -0
  544. package/dist/learning/playbook-optimizer.d.ts.map +1 -0
  545. package/dist/learning/playbook-optimizer.js +125 -0
  546. package/dist/learning/playbook-optimizer.js.map +1 -0
  547. package/dist/learning/reflective-proposer.d.ts +173 -0
  548. package/dist/learning/reflective-proposer.d.ts.map +1 -0
  549. package/dist/learning/reflective-proposer.js +354 -0
  550. package/dist/learning/reflective-proposer.js.map +1 -0
  551. package/dist/learning/reflexion-generator.d.ts.map +1 -1
  552. package/dist/learning/reflexion-generator.js +4 -3
  553. package/dist/learning/reflexion-generator.js.map +1 -1
  554. package/dist/learning/replay-executor.d.ts +64 -0
  555. package/dist/learning/replay-executor.d.ts.map +1 -0
  556. package/dist/learning/replay-executor.js +92 -0
  557. package/dist/learning/replay-executor.js.map +1 -0
  558. package/dist/learning/replay-types.d.ts +150 -0
  559. package/dist/learning/replay-types.d.ts.map +1 -0
  560. package/dist/learning/replay-types.js +207 -0
  561. package/dist/learning/replay-types.js.map +1 -0
  562. package/dist/learning/replay.d.ts +166 -0
  563. package/dist/learning/replay.d.ts.map +1 -0
  564. package/dist/learning/replay.js +228 -0
  565. package/dist/learning/replay.js.map +1 -0
  566. package/dist/learning/seed-report.d.ts +88 -0
  567. package/dist/learning/seed-report.d.ts.map +1 -0
  568. package/dist/learning/seed-report.js +120 -0
  569. package/dist/learning/seed-report.js.map +1 -0
  570. package/dist/learning/trajectory-sources/dataclaw.d.ts.map +1 -1
  571. package/dist/learning/trajectory-sources/dataclaw.js +6 -18
  572. package/dist/learning/trajectory-sources/dataclaw.js.map +1 -1
  573. package/dist/learning/trajectory-sources/entire.d.ts.map +1 -1
  574. package/dist/learning/trajectory-sources/entire.js +30 -9
  575. package/dist/learning/trajectory-sources/entire.js.map +1 -1
  576. package/dist/learning/trajectory-sources/file.d.ts.map +1 -1
  577. package/dist/learning/trajectory-sources/file.js.map +1 -1
  578. package/dist/learning/trajectory-sources/huggingface.d.ts.map +1 -1
  579. package/dist/learning/trajectory-sources/huggingface.js +7 -7
  580. package/dist/learning/trajectory-sources/huggingface.js.map +1 -1
  581. package/dist/learning/trajectory-sources/in-memory.d.ts.map +1 -1
  582. package/dist/learning/trajectory-sources/in-memory.js.map +1 -1
  583. package/dist/learning/trajectory-sources/index.d.ts +3 -4
  584. package/dist/learning/trajectory-sources/index.d.ts.map +1 -1
  585. package/dist/learning/trajectory-sources/index.js +3 -4
  586. package/dist/learning/trajectory-sources/index.js.map +1 -1
  587. package/dist/learning/trajectory-sources/pipeline.d.ts +4 -3
  588. package/dist/learning/trajectory-sources/pipeline.d.ts.map +1 -1
  589. package/dist/learning/trajectory-sources/pipeline.js +4 -3
  590. package/dist/learning/trajectory-sources/pipeline.js.map +1 -1
  591. package/dist/learning/unified-pipeline.d.ts +486 -61
  592. package/dist/learning/unified-pipeline.d.ts.map +1 -1
  593. package/dist/learning/unified-pipeline.js +1501 -232
  594. package/dist/learning/unified-pipeline.js.map +1 -1
  595. package/dist/learning/usage-inference.d.ts.map +1 -1
  596. package/dist/learning/usage-inference.js +4 -9
  597. package/dist/learning/usage-inference.js.map +1 -1
  598. package/dist/learning/workspace-snapshot.d.ts +146 -0
  599. package/dist/learning/workspace-snapshot.d.ts.map +1 -0
  600. package/dist/learning/workspace-snapshot.js +269 -0
  601. package/dist/learning/workspace-snapshot.js.map +1 -0
  602. package/dist/mcp/playbook-server.d.ts +61 -2
  603. package/dist/mcp/playbook-server.d.ts.map +1 -1
  604. package/dist/mcp/playbook-server.js +179 -9
  605. package/dist/mcp/playbook-server.js.map +1 -1
  606. package/dist/memory/coherence.d.ts.map +1 -1
  607. package/dist/memory/coherence.js +5 -7
  608. package/dist/memory/coherence.js.map +1 -1
  609. package/dist/memory/curated-loader.d.ts +24 -0
  610. package/dist/memory/curated-loader.d.ts.map +1 -1
  611. package/dist/memory/curated-loader.js +22 -5
  612. package/dist/memory/curated-loader.js.map +1 -1
  613. package/dist/memory/evolution/llm-evolver.d.ts +36 -0
  614. package/dist/memory/evolution/llm-evolver.d.ts.map +1 -0
  615. package/dist/memory/evolution/llm-evolver.js +106 -0
  616. package/dist/memory/evolution/llm-evolver.js.map +1 -0
  617. package/dist/memory/experience.d.ts +34 -13
  618. package/dist/memory/experience.d.ts.map +1 -1
  619. package/dist/memory/experience.js +59 -48
  620. package/dist/memory/experience.js.map +1 -1
  621. package/dist/memory/extraction/conversation-extractor.d.ts +62 -0
  622. package/dist/memory/extraction/conversation-extractor.d.ts.map +1 -0
  623. package/dist/memory/extraction/conversation-extractor.js +169 -0
  624. package/dist/memory/extraction/conversation-extractor.js.map +1 -0
  625. package/dist/memory/filesystem-cascade.d.ts +116 -0
  626. package/dist/memory/filesystem-cascade.d.ts.map +1 -0
  627. package/dist/memory/filesystem-cascade.js +221 -0
  628. package/dist/memory/filesystem-cascade.js.map +1 -0
  629. package/dist/memory/index.d.ts +14 -10
  630. package/dist/memory/index.d.ts.map +1 -1
  631. package/dist/memory/index.js +15 -12
  632. package/dist/memory/index.js.map +1 -1
  633. package/dist/memory/knowledge-bank.d.ts +145 -29
  634. package/dist/memory/knowledge-bank.d.ts.map +1 -1
  635. package/dist/memory/knowledge-bank.js +511 -200
  636. package/dist/memory/knowledge-bank.js.map +1 -1
  637. package/dist/memory/knowledge-defrag.d.ts +32 -9
  638. package/dist/memory/knowledge-defrag.d.ts.map +1 -1
  639. package/dist/memory/knowledge-defrag.js +57 -49
  640. package/dist/memory/knowledge-defrag.js.map +1 -1
  641. package/dist/memory/meta-playbook-bank.d.ts +24 -0
  642. package/dist/memory/meta-playbook-bank.d.ts.map +1 -0
  643. package/dist/memory/meta-playbook-bank.js +135 -0
  644. package/dist/memory/meta-playbook-bank.js.map +1 -0
  645. package/dist/memory/meta.d.ts.map +1 -1
  646. package/dist/memory/meta.js +33 -14
  647. package/dist/memory/meta.js.map +1 -1
  648. package/dist/memory/playbook-applicability.d.ts +43 -0
  649. package/dist/memory/playbook-applicability.d.ts.map +1 -0
  650. package/dist/memory/playbook-applicability.js +155 -0
  651. package/dist/memory/playbook-applicability.js.map +1 -0
  652. package/dist/memory/playbook-incorporation.d.ts +24 -0
  653. package/dist/memory/playbook-incorporation.d.ts.map +1 -0
  654. package/dist/memory/playbook-incorporation.js +41 -0
  655. package/dist/memory/playbook-incorporation.js.map +1 -0
  656. package/dist/memory/playbook.d.ts +444 -7
  657. package/dist/memory/playbook.d.ts.map +1 -1
  658. package/dist/memory/playbook.js +1231 -50
  659. package/dist/memory/playbook.js.map +1 -1
  660. package/dist/memory/procedural-skill.d.ts +142 -0
  661. package/dist/memory/procedural-skill.d.ts.map +1 -0
  662. package/dist/memory/procedural-skill.js +846 -0
  663. package/dist/memory/procedural-skill.js.map +1 -0
  664. package/dist/memory/search-provider.d.ts +8 -3
  665. package/dist/memory/search-provider.d.ts.map +1 -1
  666. package/dist/memory/search-providers/embedding.d.ts +64 -0
  667. package/dist/memory/search-providers/embedding.d.ts.map +1 -0
  668. package/dist/memory/search-providers/embedding.js +96 -0
  669. package/dist/memory/search-providers/embedding.js.map +1 -0
  670. package/dist/memory/search-providers/hybrid.d.ts +74 -0
  671. package/dist/memory/search-providers/hybrid.d.ts.map +1 -0
  672. package/dist/memory/search-providers/hybrid.js +94 -0
  673. package/dist/memory/search-providers/hybrid.js.map +1 -0
  674. package/dist/memory/search-providers/index.d.ts +3 -0
  675. package/dist/memory/search-providers/index.d.ts.map +1 -1
  676. package/dist/memory/search-providers/index.js +3 -0
  677. package/dist/memory/search-providers/index.js.map +1 -1
  678. package/dist/memory/search-providers/keyword-expanding.d.ts +22 -0
  679. package/dist/memory/search-providers/keyword-expanding.d.ts.map +1 -0
  680. package/dist/memory/search-providers/keyword-expanding.js +35 -0
  681. package/dist/memory/search-providers/keyword-expanding.js.map +1 -0
  682. package/dist/memory/search-providers/minimem.d.ts +16 -4
  683. package/dist/memory/search-providers/minimem.d.ts.map +1 -1
  684. package/dist/memory/search-providers/minimem.js +14 -3
  685. package/dist/memory/search-providers/minimem.js.map +1 -1
  686. package/dist/memory/search-providers/text-similarity.d.ts +2 -4
  687. package/dist/memory/search-providers/text-similarity.d.ts.map +1 -1
  688. package/dist/memory/search-providers/text-similarity.js.map +1 -1
  689. package/dist/memory/seeds/meta-playbook-seeds.d.ts +27 -0
  690. package/dist/memory/seeds/meta-playbook-seeds.d.ts.map +1 -0
  691. package/dist/memory/seeds/meta-playbook-seeds.js +560 -0
  692. package/dist/memory/seeds/meta-playbook-seeds.js.map +1 -0
  693. package/dist/memory/seeds/meta-strategy-seeds.d.ts +23 -0
  694. package/dist/memory/seeds/meta-strategy-seeds.d.ts.map +1 -0
  695. package/dist/memory/seeds/meta-strategy-seeds.js +82 -0
  696. package/dist/memory/seeds/meta-strategy-seeds.js.map +1 -0
  697. package/dist/memory/seeds/seed-loader.d.ts +34 -0
  698. package/dist/memory/seeds/seed-loader.d.ts.map +1 -0
  699. package/dist/memory/seeds/seed-loader.js +60 -0
  700. package/dist/memory/seeds/seed-loader.js.map +1 -0
  701. package/dist/memory/skill-exporter.d.ts.map +1 -1
  702. package/dist/memory/skill-exporter.js +1 -2
  703. package/dist/memory/skill-exporter.js.map +1 -1
  704. package/dist/memory/source-resolver.d.ts.map +1 -1
  705. package/dist/memory/source-resolver.js +7 -8
  706. package/dist/memory/source-resolver.js.map +1 -1
  707. package/dist/memory/system.d.ts +119 -14
  708. package/dist/memory/system.d.ts.map +1 -1
  709. package/dist/memory/system.js +416 -60
  710. package/dist/memory/system.js.map +1 -1
  711. package/dist/memory/temporal-compressor.d.ts +1 -1
  712. package/dist/memory/temporal-compressor.d.ts.map +1 -1
  713. package/dist/memory/temporal-compressor.js +26 -16
  714. package/dist/memory/temporal-compressor.js.map +1 -1
  715. package/dist/persistence/index.d.ts +2 -2
  716. package/dist/persistence/index.d.ts.map +1 -1
  717. package/dist/persistence/index.js +1 -1
  718. package/dist/persistence/index.js.map +1 -1
  719. package/dist/persistence/migrator.d.ts +34 -1
  720. package/dist/persistence/migrator.d.ts.map +1 -1
  721. package/dist/persistence/migrator.js +163 -33
  722. package/dist/persistence/migrator.js.map +1 -1
  723. package/dist/persistence/sqlite-persistence.d.ts +412 -1
  724. package/dist/persistence/sqlite-persistence.d.ts.map +1 -1
  725. package/dist/persistence/sqlite-persistence.js +1038 -47
  726. package/dist/persistence/sqlite-persistence.js.map +1 -1
  727. package/dist/runtime/backends/acp-protocol.d.ts.map +1 -1
  728. package/dist/runtime/backends/acp-protocol.js +2 -1
  729. package/dist/runtime/backends/acp-protocol.js.map +1 -1
  730. package/dist/runtime/backends/acp.d.ts.map +1 -1
  731. package/dist/runtime/backends/acp.js +5 -4
  732. package/dist/runtime/backends/acp.js.map +1 -1
  733. package/dist/runtime/backends/index.d.ts +2 -2
  734. package/dist/runtime/backends/index.d.ts.map +1 -1
  735. package/dist/runtime/backends/index.js +2 -2
  736. package/dist/runtime/backends/index.js.map +1 -1
  737. package/dist/runtime/backends/macro-agent.d.ts +17 -1
  738. package/dist/runtime/backends/macro-agent.d.ts.map +1 -1
  739. package/dist/runtime/backends/macro-agent.js +1 -1
  740. package/dist/runtime/backends/macro-agent.js.map +1 -1
  741. package/dist/runtime/backends/mock.d.ts.map +1 -1
  742. package/dist/runtime/backends/mock.js.map +1 -1
  743. package/dist/runtime/backends/subprocess.d.ts.map +1 -1
  744. package/dist/runtime/backends/subprocess.js +5 -4
  745. package/dist/runtime/backends/subprocess.js.map +1 -1
  746. package/dist/runtime/compute-provider.d.ts +2 -18
  747. package/dist/runtime/compute-provider.d.ts.map +1 -1
  748. package/dist/runtime/compute-provider.js +8 -8
  749. package/dist/runtime/compute-provider.js.map +1 -1
  750. package/dist/runtime/flows/learning.d.ts +2 -1
  751. package/dist/runtime/flows/learning.d.ts.map +1 -1
  752. package/dist/runtime/flows/learning.js +4 -5
  753. package/dist/runtime/flows/learning.js.map +1 -1
  754. package/dist/runtime/flows/validation.d.ts.map +1 -1
  755. package/dist/runtime/flows/validation.js +2 -5
  756. package/dist/runtime/flows/validation.js.map +1 -1
  757. package/dist/runtime/index.d.ts +1 -1
  758. package/dist/runtime/index.d.ts.map +1 -1
  759. package/dist/runtime/index.js.map +1 -1
  760. package/dist/runtime/manager.d.ts +128 -0
  761. package/dist/runtime/manager.d.ts.map +1 -1
  762. package/dist/runtime/manager.js +375 -64
  763. package/dist/runtime/manager.js.map +1 -1
  764. package/dist/runtime/types.d.ts +92 -1
  765. package/dist/runtime/types.d.ts.map +1 -1
  766. package/dist/search/evaluator.d.ts +71 -0
  767. package/dist/search/evaluator.d.ts.map +1 -1
  768. package/dist/search/evaluator.js +40 -5
  769. package/dist/search/evaluator.js.map +1 -1
  770. package/dist/search/index.d.ts +3 -5
  771. package/dist/search/index.d.ts.map +1 -1
  772. package/dist/search/index.js +4 -7
  773. package/dist/search/index.js.map +1 -1
  774. package/dist/search/moe-gate.d.ts +48 -2
  775. package/dist/search/moe-gate.d.ts.map +1 -1
  776. package/dist/search/moe-gate.js +117 -4
  777. package/dist/search/moe-gate.js.map +1 -1
  778. package/dist/search/refinement-loop.d.ts +63 -3
  779. package/dist/search/refinement-loop.d.ts.map +1 -1
  780. package/dist/search/refinement-loop.js +278 -23
  781. package/dist/search/refinement-loop.js.map +1 -1
  782. package/dist/search/refinement-types.d.ts +4 -2
  783. package/dist/search/refinement-types.d.ts.map +1 -1
  784. package/dist/search/refinement-types.js +2 -11
  785. package/dist/search/refinement-types.js.map +1 -1
  786. package/dist/search/router.d.ts +41 -34
  787. package/dist/search/router.d.ts.map +1 -1
  788. package/dist/search/router.js +156 -90
  789. package/dist/search/router.js.map +1 -1
  790. package/dist/search/solver.js +1 -1
  791. package/dist/search/solver.js.map +1 -1
  792. package/dist/search/task-difficulty.d.ts +54 -0
  793. package/dist/search/task-difficulty.d.ts.map +1 -0
  794. package/dist/search/task-difficulty.js +116 -0
  795. package/dist/search/task-difficulty.js.map +1 -0
  796. package/dist/search/tau-bench-verifier.d.ts +74 -0
  797. package/dist/search/tau-bench-verifier.d.ts.map +1 -0
  798. package/dist/search/tau-bench-verifier.js +96 -0
  799. package/dist/search/tau-bench-verifier.js.map +1 -0
  800. package/dist/search/verification-runner.d.ts.map +1 -1
  801. package/dist/search/verification-runner.js +6 -5
  802. package/dist/search/verification-runner.js.map +1 -1
  803. package/dist/session-bank/git-reader.d.ts.map +1 -1
  804. package/dist/session-bank/git-reader.js +5 -19
  805. package/dist/session-bank/git-reader.js.map +1 -1
  806. package/dist/session-bank/parser.d.ts.map +1 -1
  807. package/dist/session-bank/parser.js +26 -0
  808. package/dist/session-bank/parser.js.map +1 -1
  809. package/dist/session-bank/session-bank.d.ts.map +1 -1
  810. package/dist/session-bank/session-bank.js +2 -5
  811. package/dist/session-bank/session-bank.js.map +1 -1
  812. package/dist/session-bank/types.d.ts +24 -0
  813. package/dist/session-bank/types.d.ts.map +1 -1
  814. package/dist/surfacing/index.d.ts +3 -3
  815. package/dist/surfacing/index.d.ts.map +1 -1
  816. package/dist/surfacing/index.js +2 -2
  817. package/dist/surfacing/index.js.map +1 -1
  818. package/dist/surfacing/shared-bank.d.ts +43 -0
  819. package/dist/surfacing/shared-bank.d.ts.map +1 -0
  820. package/dist/surfacing/shared-bank.js +104 -0
  821. package/dist/surfacing/shared-bank.js.map +1 -0
  822. package/dist/surfacing/skill-library.d.ts +98 -2
  823. package/dist/surfacing/skill-library.d.ts.map +1 -1
  824. package/dist/surfacing/skill-library.js +192 -17
  825. package/dist/surfacing/skill-library.js.map +1 -1
  826. package/dist/surfacing/skill-publisher.d.ts +3 -1
  827. package/dist/surfacing/skill-publisher.d.ts.map +1 -1
  828. package/dist/surfacing/skill-publisher.js +28 -28
  829. package/dist/surfacing/skill-publisher.js.map +1 -1
  830. package/dist/surfacing/sqlite-storage-adapter.d.ts +2 -2
  831. package/dist/surfacing/sqlite-storage-adapter.d.ts.map +1 -1
  832. package/dist/surfacing/sqlite-storage-adapter.js +27 -18
  833. package/dist/surfacing/sqlite-storage-adapter.js.map +1 -1
  834. package/dist/types/analysis.d.ts +43 -0
  835. package/dist/types/analysis.d.ts.map +1 -0
  836. package/dist/types/analysis.js +10 -0
  837. package/dist/types/analysis.js.map +1 -0
  838. package/dist/types/config.d.ts +2178 -425
  839. package/dist/types/config.d.ts.map +1 -1
  840. package/dist/types/config.js +600 -88
  841. package/dist/types/config.js.map +1 -1
  842. package/dist/types/dataclaw.d.ts +3 -3
  843. package/dist/types/index.d.ts +4 -7
  844. package/dist/types/index.d.ts.map +1 -1
  845. package/dist/types/index.js +2 -10
  846. package/dist/types/index.js.map +1 -1
  847. package/dist/types/knowledge-graph.d.ts +1 -1
  848. package/dist/types/knowledge-graph.d.ts.map +1 -1
  849. package/dist/types/knowledge.d.ts +219 -11
  850. package/dist/types/knowledge.d.ts.map +1 -1
  851. package/dist/types/knowledge.js +90 -30
  852. package/dist/types/knowledge.js.map +1 -1
  853. package/dist/types/memory.d.ts +8 -0
  854. package/dist/types/memory.d.ts.map +1 -1
  855. package/dist/types/memory.js +12 -4
  856. package/dist/types/memory.js.map +1 -1
  857. package/dist/types/outcome.d.ts +11 -0
  858. package/dist/types/outcome.d.ts.map +1 -1
  859. package/dist/types/outcome.js +9 -0
  860. package/dist/types/outcome.js.map +1 -1
  861. package/dist/types/playbook.d.ts +63 -0
  862. package/dist/types/playbook.d.ts.map +1 -1
  863. package/dist/types/playbook.js +51 -0
  864. package/dist/types/playbook.js.map +1 -1
  865. package/dist/types/trajectory-source.d.ts +1 -1
  866. package/dist/types/trajectory-source.d.ts.map +1 -1
  867. package/dist/types/trajectory.d.ts +61 -0
  868. package/dist/types/trajectory.d.ts.map +1 -1
  869. package/dist/types/trajectory.js +25 -0
  870. package/dist/types/trajectory.js.map +1 -1
  871. package/dist/utils/clock.d.ts +36 -0
  872. package/dist/utils/clock.d.ts.map +1 -0
  873. package/dist/utils/clock.js +48 -0
  874. package/dist/utils/clock.js.map +1 -0
  875. package/dist/utils/frontmatter.d.ts.map +1 -1
  876. package/dist/utils/frontmatter.js +10 -2
  877. package/dist/utils/frontmatter.js.map +1 -1
  878. package/dist/utils/index.d.ts +4 -1
  879. package/dist/utils/index.d.ts.map +1 -1
  880. package/dist/utils/index.js +4 -1
  881. package/dist/utils/index.js.map +1 -1
  882. package/dist/utils/logger.d.ts +31 -0
  883. package/dist/utils/logger.d.ts.map +1 -0
  884. package/dist/utils/logger.js +50 -0
  885. package/dist/utils/logger.js.map +1 -0
  886. package/dist/utils/objectives.d.ts +45 -0
  887. package/dist/utils/objectives.d.ts.map +1 -0
  888. package/dist/utils/objectives.js +74 -0
  889. package/dist/utils/objectives.js.map +1 -0
  890. package/dist/utils/partitioned-store.d.ts.map +1 -1
  891. package/dist/utils/partitioned-store.js.map +1 -1
  892. package/dist/utils/playbook-hash.d.ts +19 -0
  893. package/dist/utils/playbook-hash.d.ts.map +1 -0
  894. package/dist/utils/playbook-hash.js +51 -0
  895. package/dist/utils/playbook-hash.js.map +1 -0
  896. package/dist/utils/similarity.d.ts.map +1 -1
  897. package/dist/utils/similarity.js.map +1 -1
  898. package/dist/utils/solver-canonicalizer.d.ts +22 -0
  899. package/dist/utils/solver-canonicalizer.d.ts.map +1 -0
  900. package/dist/utils/solver-canonicalizer.js +41 -0
  901. package/dist/utils/solver-canonicalizer.js.map +1 -0
  902. package/dist/utils/storage.d.ts.map +1 -1
  903. package/dist/utils/storage.js +1 -3
  904. package/dist/utils/storage.js.map +1 -1
  905. package/dist/utils/task-signature.d.ts +34 -0
  906. package/dist/utils/task-signature.d.ts.map +1 -0
  907. package/dist/utils/task-signature.js +112 -0
  908. package/dist/utils/task-signature.js.map +1 -0
  909. package/dist/utils/trajectory-helpers.d.ts.map +1 -1
  910. package/dist/utils/trajectory-helpers.js.map +1 -1
  911. package/dist/utils/validation.d.ts.map +1 -1
  912. package/dist/utils/validation.js.map +1 -1
  913. package/dist/workspace/efficacy-toolkit.d.ts.map +1 -1
  914. package/dist/workspace/efficacy-toolkit.js +18 -26
  915. package/dist/workspace/efficacy-toolkit.js.map +1 -1
  916. package/dist/workspace/index.d.ts +4 -3
  917. package/dist/workspace/index.d.ts.map +1 -1
  918. package/dist/workspace/index.js +5 -3
  919. package/dist/workspace/index.js.map +1 -1
  920. package/dist/workspace/runner.d.ts.map +1 -1
  921. package/dist/workspace/runner.js +95 -21
  922. package/dist/workspace/runner.js.map +1 -1
  923. package/dist/workspace/skill-converter.d.ts +66 -1
  924. package/dist/workspace/skill-converter.d.ts.map +1 -1
  925. package/dist/workspace/skill-converter.js +238 -19
  926. package/dist/workspace/skill-converter.js.map +1 -1
  927. package/dist/workspace/skill-import.d.ts +43 -0
  928. package/dist/workspace/skill-import.d.ts.map +1 -0
  929. package/dist/workspace/skill-import.js +40 -0
  930. package/dist/workspace/skill-import.js.map +1 -0
  931. package/dist/workspace/templates/index.d.ts +6 -8
  932. package/dist/workspace/templates/index.d.ts.map +1 -1
  933. package/dist/workspace/templates/index.js +7 -11
  934. package/dist/workspace/templates/index.js.map +1 -1
  935. package/dist/workspace/templates/knowledge-defrag.d.ts.map +1 -1
  936. package/dist/workspace/templates/knowledge-defrag.js +0 -2
  937. package/dist/workspace/templates/knowledge-defrag.js.map +1 -1
  938. package/dist/workspace/templates/knowledge-extraction.d.ts +11 -0
  939. package/dist/workspace/templates/knowledge-extraction.d.ts.map +1 -1
  940. package/dist/workspace/templates/knowledge-extraction.js +11 -6
  941. package/dist/workspace/templates/knowledge-extraction.js.map +1 -1
  942. package/dist/workspace/templates/playbook-conditioning.d.ts +38 -0
  943. package/dist/workspace/templates/playbook-conditioning.d.ts.map +1 -0
  944. package/dist/workspace/templates/playbook-conditioning.js +147 -0
  945. package/dist/workspace/templates/playbook-conditioning.js.map +1 -0
  946. package/dist/workspace/templates/playbook-decay-detection.d.ts.map +1 -1
  947. package/dist/workspace/templates/playbook-decay-detection.js +24 -18
  948. package/dist/workspace/templates/playbook-decay-detection.js.map +1 -1
  949. package/dist/workspace/templates/playbook-efficacy-audit.d.ts.map +1 -1
  950. package/dist/workspace/templates/playbook-efficacy-audit.js +26 -11
  951. package/dist/workspace/templates/playbook-efficacy-audit.js.map +1 -1
  952. package/dist/workspace/templates/playbook-extraction.d.ts.map +1 -1
  953. package/dist/workspace/templates/playbook-extraction.js +75 -15
  954. package/dist/workspace/templates/playbook-extraction.js.map +1 -1
  955. package/dist/workspace/templates/playbook-lifecycle-review.d.ts.map +1 -1
  956. package/dist/workspace/templates/playbook-lifecycle-review.js +12 -8
  957. package/dist/workspace/templates/playbook-lifecycle-review.js.map +1 -1
  958. package/dist/workspace/templates/refinement-analysis.d.ts.map +1 -1
  959. package/dist/workspace/templates/refinement-analysis.js +11 -5
  960. package/dist/workspace/templates/refinement-analysis.js.map +1 -1
  961. package/dist/workspace/templates/skill-enrichment.d.ts.map +1 -1
  962. package/dist/workspace/templates/skill-enrichment.js +7 -11
  963. package/dist/workspace/templates/skill-enrichment.js.map +1 -1
  964. package/dist/workspace/templates/solution-evaluation.d.ts.map +1 -1
  965. package/dist/workspace/templates/solution-evaluation.js +15 -8
  966. package/dist/workspace/templates/solution-evaluation.js.map +1 -1
  967. package/dist/workspace/templates/trajectory-analysis.d.ts.map +1 -1
  968. package/dist/workspace/templates/trajectory-analysis.js +1 -3
  969. package/dist/workspace/templates/trajectory-analysis.js.map +1 -1
  970. package/dist/workspace/templates/usage-inference.d.ts.map +1 -1
  971. package/dist/workspace/templates/usage-inference.js +9 -7
  972. package/dist/workspace/templates/usage-inference.js.map +1 -1
  973. package/dist/workspace/types.d.ts +32 -2
  974. package/dist/workspace/types.d.ts.map +1 -1
  975. package/dist/workspace/types.js +29 -0
  976. package/dist/workspace/types.js.map +1 -1
  977. package/package.json +43 -19
  978. package/playbooks/agent-ops/long-task-checkpointing.json +54 -0
  979. package/playbooks/agent-ops/progressive-context-gathering.json +56 -0
  980. package/playbooks/agent-ops/safe-destructive-operations.json +59 -0
  981. package/playbooks/agent-ops/scope-discipline.json +56 -0
  982. package/playbooks/agent-ops/stall-detection-strategy-switch.json +55 -0
  983. package/playbooks/agent-ops/verification-before-completion.json +54 -0
  984. package/playbooks/alfworld/clean-and-place.json +35 -0
  985. package/playbooks/alfworld/cool-and-place.json +35 -0
  986. package/playbooks/alfworld/examine-in-light.json +32 -0
  987. package/playbooks/alfworld/heat-and-place.json +35 -0
  988. package/playbooks/alfworld/pick-and-place.json +37 -0
  989. package/playbooks/alfworld/pick-two-and-place.json +35 -0
  990. package/.claude/settings.json +0 -197
  991. package/.entire/settings.json +0 -4
  992. package/.sessionlog/settings.json +0 -4
  993. package/CLAUDE.md +0 -233
  994. package/SKILL.md +0 -193
  995. package/docs/DESIGN-agentic-workspace.md +0 -2057
  996. package/docs/DESIGN-semantic-memory-knowledge-bank.md +0 -1789
  997. package/docs/DESIGN-session-bank.md +0 -1134
  998. package/docs/DESIGN-team-extraction-pipeline.md +0 -1369
  999. package/docs/PROPOSAL-team-aware-learning.md +0 -1080
  1000. package/docs/RESEARCH-semantic-memory-knowledge-structures.md +0 -517
  1001. package/docs/RESEARCH-team-trajectory-learning.md +0 -553
  1002. package/gaps.md +0 -204
  1003. package/scripts/migrate-to-playbooks.ts +0 -307
  1004. package/src/agents/index.ts +0 -14
  1005. package/src/agents/mock-provider.ts +0 -93
  1006. package/src/agents/types.ts +0 -137
  1007. package/src/atlas.ts +0 -1198
  1008. package/src/bin/cli-utils.ts +0 -220
  1009. package/src/bin/cognitive-core.ts +0 -162
  1010. package/src/bin/commands/kb.ts +0 -266
  1011. package/src/bin/commands/learn.ts +0 -100
  1012. package/src/bin/commands/legacy.ts +0 -182
  1013. package/src/bin/commands/run.ts +0 -113
  1014. package/src/bin/commands/sessions.ts +0 -221
  1015. package/src/bin/commands/skills.ts +0 -146
  1016. package/src/embeddings/bm25.ts +0 -337
  1017. package/src/embeddings/index.ts +0 -42
  1018. package/src/embeddings/inverted-index.ts +0 -134
  1019. package/src/embeddings/manager.ts +0 -293
  1020. package/src/embeddings/provider.ts +0 -318
  1021. package/src/embeddings/vector-store.ts +0 -365
  1022. package/src/factory.ts +0 -280
  1023. package/src/index.ts +0 -495
  1024. package/src/learning/analyzer.ts +0 -319
  1025. package/src/learning/effectiveness.ts +0 -428
  1026. package/src/learning/energy-evaluator.ts +0 -282
  1027. package/src/learning/healing-orchestrator.ts +0 -383
  1028. package/src/learning/index.ts +0 -150
  1029. package/src/learning/instant-loop.ts +0 -357
  1030. package/src/learning/knowledge-extractor.ts +0 -470
  1031. package/src/learning/maintenance-scheduler.ts +0 -271
  1032. package/src/learning/meta-learner.ts +0 -552
  1033. package/src/learning/playbook-extractor.ts +0 -702
  1034. package/src/learning/reflexion-generator.ts +0 -273
  1035. package/src/learning/team-ingester.ts +0 -499
  1036. package/src/learning/team-meta-learner.ts +0 -558
  1037. package/src/learning/team-pipeline.ts +0 -364
  1038. package/src/learning/trajectory-sources/dataclaw.ts +0 -403
  1039. package/src/learning/trajectory-sources/entire.ts +0 -251
  1040. package/src/learning/trajectory-sources/file.ts +0 -136
  1041. package/src/learning/trajectory-sources/huggingface.ts +0 -248
  1042. package/src/learning/trajectory-sources/in-memory.ts +0 -61
  1043. package/src/learning/trajectory-sources/index.ts +0 -15
  1044. package/src/learning/trajectory-sources/pipeline.ts +0 -69
  1045. package/src/learning/unified-pipeline.ts +0 -1191
  1046. package/src/learning/usage-inference.ts +0 -372
  1047. package/src/mcp/index.ts +0 -12
  1048. package/src/mcp/playbook-server.ts +0 -565
  1049. package/src/memory/candidate-retrieval.ts +0 -72
  1050. package/src/memory/causal-store.ts +0 -273
  1051. package/src/memory/coherence.ts +0 -252
  1052. package/src/memory/curated-loader.ts +0 -213
  1053. package/src/memory/experience.ts +0 -682
  1054. package/src/memory/graph-layers/base.ts +0 -184
  1055. package/src/memory/graph-layers/causal.ts +0 -16
  1056. package/src/memory/graph-layers/entity.ts +0 -16
  1057. package/src/memory/graph-layers/index.ts +0 -5
  1058. package/src/memory/graph-layers/semantic.ts +0 -16
  1059. package/src/memory/graph-layers/temporal.ts +0 -16
  1060. package/src/memory/index.ts +0 -131
  1061. package/src/memory/knowledge-bank.ts +0 -1317
  1062. package/src/memory/knowledge-defrag.ts +0 -329
  1063. package/src/memory/knowledge-graph.ts +0 -361
  1064. package/src/memory/meta.ts +0 -488
  1065. package/src/memory/playbook.ts +0 -676
  1066. package/src/memory/reasoning-bank.ts +0 -458
  1067. package/src/memory/reflexion.ts +0 -122
  1068. package/src/memory/search-provider.ts +0 -35
  1069. package/src/memory/search-providers/index.ts +0 -3
  1070. package/src/memory/search-providers/minimem.ts +0 -84
  1071. package/src/memory/search-providers/text-similarity.ts +0 -35
  1072. package/src/memory/skill-exporter.ts +0 -357
  1073. package/src/memory/source-resolver.ts +0 -422
  1074. package/src/memory/system.ts +0 -620
  1075. package/src/memory/team-experience.ts +0 -604
  1076. package/src/memory/temporal-compressor.ts +0 -409
  1077. package/src/persistence/index.ts +0 -37
  1078. package/src/persistence/migrator.ts +0 -298
  1079. package/src/persistence/serializers.ts +0 -79
  1080. package/src/persistence/sqlite-persistence.ts +0 -925
  1081. package/src/runtime/backends/acp-protocol.ts +0 -231
  1082. package/src/runtime/backends/acp.ts +0 -239
  1083. package/src/runtime/backends/index.ts +0 -38
  1084. package/src/runtime/backends/macro-agent.ts +0 -177
  1085. package/src/runtime/backends/mock.ts +0 -218
  1086. package/src/runtime/backends/subprocess.ts +0 -356
  1087. package/src/runtime/compute-provider.ts +0 -206
  1088. package/src/runtime/flows/learning.ts +0 -166
  1089. package/src/runtime/flows/validation.ts +0 -381
  1090. package/src/runtime/index.ts +0 -75
  1091. package/src/runtime/manager.ts +0 -685
  1092. package/src/runtime/types.ts +0 -198
  1093. package/src/search/evaluator.ts +0 -499
  1094. package/src/search/index.ts +0 -78
  1095. package/src/search/moe-gate.ts +0 -304
  1096. package/src/search/refinement-loop.ts +0 -457
  1097. package/src/search/refinement-types.ts +0 -159
  1098. package/src/search/router.ts +0 -368
  1099. package/src/search/solver.ts +0 -303
  1100. package/src/search/team-router.ts +0 -459
  1101. package/src/search/verification-runner.ts +0 -570
  1102. package/src/session-bank/git-reader.ts +0 -200
  1103. package/src/session-bank/index.ts +0 -26
  1104. package/src/session-bank/parser.ts +0 -548
  1105. package/src/session-bank/session-bank.ts +0 -485
  1106. package/src/session-bank/types.ts +0 -176
  1107. package/src/surfacing/index.ts +0 -29
  1108. package/src/surfacing/publisher.ts +0 -23
  1109. package/src/surfacing/skill-library.ts +0 -625
  1110. package/src/surfacing/skill-publisher.ts +0 -257
  1111. package/src/surfacing/sqlite-storage-adapter.ts +0 -285
  1112. package/src/surfacing/team-skill-library.ts +0 -610
  1113. package/src/types/config.ts +0 -661
  1114. package/src/types/dataclaw.ts +0 -99
  1115. package/src/types/huggingface-transformers.d.ts +0 -12
  1116. package/src/types/index.ts +0 -293
  1117. package/src/types/knowledge-graph.ts +0 -246
  1118. package/src/types/knowledge.ts +0 -388
  1119. package/src/types/memory.ts +0 -300
  1120. package/src/types/meta.ts +0 -218
  1121. package/src/types/outcome.ts +0 -66
  1122. package/src/types/playbook.ts +0 -215
  1123. package/src/types/step.ts +0 -40
  1124. package/src/types/task.ts +0 -52
  1125. package/src/types/team-meta.ts +0 -212
  1126. package/src/types/team-playbook.ts +0 -384
  1127. package/src/types/team-trajectory.ts +0 -673
  1128. package/src/types/trajectory-source.ts +0 -47
  1129. package/src/types/trajectory.ts +0 -80
  1130. package/src/utils/error-classifier.ts +0 -113
  1131. package/src/utils/frontmatter.ts +0 -118
  1132. package/src/utils/index.ts +0 -63
  1133. package/src/utils/partitioned-store.ts +0 -299
  1134. package/src/utils/similarity.ts +0 -139
  1135. package/src/utils/storage.ts +0 -249
  1136. package/src/utils/trajectory-helpers.ts +0 -79
  1137. package/src/utils/validation.ts +0 -286
  1138. package/src/workspace/efficacy-toolkit.ts +0 -496
  1139. package/src/workspace/index.ts +0 -77
  1140. package/src/workspace/runner.ts +0 -281
  1141. package/src/workspace/skill-converter.ts +0 -288
  1142. package/src/workspace/templates/index.ts +0 -95
  1143. package/src/workspace/templates/knowledge-defrag.ts +0 -223
  1144. package/src/workspace/templates/knowledge-extraction.ts +0 -337
  1145. package/src/workspace/templates/meta-reflection.ts +0 -208
  1146. package/src/workspace/templates/playbook-decay-detection.ts +0 -272
  1147. package/src/workspace/templates/playbook-efficacy-audit.ts +0 -246
  1148. package/src/workspace/templates/playbook-extraction.ts +0 -240
  1149. package/src/workspace/templates/playbook-lifecycle-review.ts +0 -274
  1150. package/src/workspace/templates/refinement-analysis.ts +0 -162
  1151. package/src/workspace/templates/skill-enrichment.ts +0 -275
  1152. package/src/workspace/templates/solution-evaluation.ts +0 -199
  1153. package/src/workspace/templates/team-playbook-extraction.ts +0 -631
  1154. package/src/workspace/templates/team-trajectory-analysis.ts +0 -564
  1155. package/src/workspace/templates/trajectory-analysis.ts +0 -234
  1156. package/src/workspace/templates/usage-inference.ts +0 -163
  1157. package/src/workspace/types.ts +0 -185
  1158. package/tests/atlas-knowledge.test.ts +0 -165
  1159. package/tests/embeddings/bm25.test.ts +0 -130
  1160. package/tests/embeddings/inverted-index.test.ts +0 -138
  1161. package/tests/embeddings/manager.test.ts +0 -205
  1162. package/tests/feature-toggles.test.ts +0 -275
  1163. package/tests/fixtures/behavioral-trajectories.ts +0 -210
  1164. package/tests/gap-fixes.test.ts +0 -1116
  1165. package/tests/integration/atlas.test.ts +0 -266
  1166. package/tests/integration/cli-e2e.test.ts +0 -621
  1167. package/tests/integration/curated-sources-e2e.test.ts +0 -502
  1168. package/tests/integration/dataclaw-e2e.test.ts +0 -559
  1169. package/tests/integration/e2e.test.ts +0 -1337
  1170. package/tests/integration/entire-e2e.test.ts +0 -376
  1171. package/tests/integration/huggingface-e2e.test.ts +0 -627
  1172. package/tests/integration/persistence-e2e.test.ts +0 -741
  1173. package/tests/integration/phase-e2e.test.ts +0 -1143
  1174. package/tests/integration/pipeline-data-correctness.test.ts +0 -794
  1175. package/tests/integration/session-bank.test.ts +0 -231
  1176. package/tests/integration/sessionlog-e2e.test.ts +0 -329
  1177. package/tests/integration/skill-tree-wiring.test.ts +0 -152
  1178. package/tests/integration/unified-pipeline-e2e.test.ts +0 -634
  1179. package/tests/learning/analyzer.test.ts +0 -426
  1180. package/tests/learning/dataclaw-trajectory-source.test.ts +0 -642
  1181. package/tests/learning/effectiveness.test.ts +0 -542
  1182. package/tests/learning/energy-evaluator.test.ts +0 -180
  1183. package/tests/learning/entire-trajectory-source.test.ts +0 -380
  1184. package/tests/learning/healing-orchestrator.test.ts +0 -269
  1185. package/tests/learning/huggingface-trajectory-source.test.ts +0 -817
  1186. package/tests/learning/instant-loop.test.ts +0 -243
  1187. package/tests/learning/knowledge-extractor.test.ts +0 -491
  1188. package/tests/learning/maintenance-scheduler.test.ts +0 -191
  1189. package/tests/learning/meta-learner.test.ts +0 -418
  1190. package/tests/learning/pipeline-memory-updates.test.ts +0 -721
  1191. package/tests/learning/playbook-extractor-provenance.test.ts +0 -114
  1192. package/tests/learning/reflexion-generator.test.ts +0 -411
  1193. package/tests/learning/team-ingester.test.ts +0 -349
  1194. package/tests/learning/team-meta-learner.test.ts +0 -618
  1195. package/tests/learning/team-pipeline.test.ts +0 -334
  1196. package/tests/learning/trajectory-sources.test.ts +0 -320
  1197. package/tests/learning/unified-pipeline-efficacy.test.ts +0 -232
  1198. package/tests/learning/unified-pipeline.test.ts +0 -322
  1199. package/tests/learning/usage-inference.test.ts +0 -254
  1200. package/tests/mcp/playbook-server.test.ts +0 -257
  1201. package/tests/memory/candidate-retrieval.test.ts +0 -167
  1202. package/tests/memory/causal-store.test.ts +0 -276
  1203. package/tests/memory/coherence.test.ts +0 -232
  1204. package/tests/memory/compound-engineering-seed.test.ts +0 -338
  1205. package/tests/memory/curated-loader-extended.test.ts +0 -225
  1206. package/tests/memory/experience.test.ts +0 -203
  1207. package/tests/memory/graph-layers/semantic.test.ts +0 -219
  1208. package/tests/memory/knowledge-bank-extraction.test.ts +0 -558
  1209. package/tests/memory/knowledge-bank.test.ts +0 -705
  1210. package/tests/memory/knowledge-defrag.test.ts +0 -366
  1211. package/tests/memory/knowledge-evolution.test.ts +0 -563
  1212. package/tests/memory/knowledge-graph.test.ts +0 -492
  1213. package/tests/memory/knowledge-inbox.test.ts +0 -258
  1214. package/tests/memory/knowledge-minimem.test.ts +0 -251
  1215. package/tests/memory/meta.test.ts +0 -399
  1216. package/tests/memory/playbook-quality-validation.test.ts +0 -430
  1217. package/tests/memory/playbook.test.ts +0 -644
  1218. package/tests/memory/provenance.test.ts +0 -648
  1219. package/tests/memory/reasoning-bank.test.ts +0 -239
  1220. package/tests/memory/reflexion.test.ts +0 -166
  1221. package/tests/memory/skill-exporter.test.ts +0 -475
  1222. package/tests/memory/source-resolver.test.ts +0 -700
  1223. package/tests/memory/system.test.ts +0 -330
  1224. package/tests/memory/team-experience.test.ts +0 -411
  1225. package/tests/memory/temporal-compressor.test.ts +0 -318
  1226. package/tests/persistence/migrator.test.ts +0 -1009
  1227. package/tests/persistence/sqlite-persistence.test.ts +0 -635
  1228. package/tests/runtime/agent-manager.test.ts +0 -517
  1229. package/tests/runtime/compute-provider.test.ts +0 -288
  1230. package/tests/runtime/delegate.test.ts +0 -354
  1231. package/tests/runtime/macro-agent-backend.test.ts +0 -266
  1232. package/tests/runtime/mock-backend.test.ts +0 -248
  1233. package/tests/search/evaluator.test.ts +0 -257
  1234. package/tests/search/moe-gate.test.ts +0 -250
  1235. package/tests/search/refinement-loop.test.ts +0 -477
  1236. package/tests/search/refinement.test.ts +0 -267
  1237. package/tests/search/router.test.ts +0 -506
  1238. package/tests/search/team-router.test.ts +0 -376
  1239. package/tests/search/verification-runner.test.ts +0 -357
  1240. package/tests/session-bank/fixtures/multi-tool-session/full.jsonl +0 -12
  1241. package/tests/session-bank/fixtures/multi-tool-session/metadata.json +0 -28
  1242. package/tests/session-bank/fixtures/root-metadata.json +0 -18
  1243. package/tests/session-bank/fixtures/sessionlog-root-metadata.json +0 -16
  1244. package/tests/session-bank/fixtures/sessionlog-session/full.jsonl +0 -6
  1245. package/tests/session-bank/fixtures/sessionlog-session/metadata.json +0 -55
  1246. package/tests/session-bank/fixtures/simple-session/full.jsonl +0 -6
  1247. package/tests/session-bank/fixtures/simple-session/metadata.json +0 -38
  1248. package/tests/session-bank/git-reader.test.ts +0 -232
  1249. package/tests/session-bank/parser.test.ts +0 -453
  1250. package/tests/session-bank/session-bank.test.ts +0 -546
  1251. package/tests/surfacing/skill-library.test.ts +0 -297
  1252. package/tests/surfacing/skill-publisher.test.ts +0 -437
  1253. package/tests/surfacing/sqlite-storage-adapter.test.ts +0 -206
  1254. package/tests/surfacing/team-skill-library.test.ts +0 -444
  1255. package/tests/types/outcome.test.ts +0 -147
  1256. package/tests/types/step.test.ts +0 -133
  1257. package/tests/types/task.test.ts +0 -158
  1258. package/tests/types/team-meta.test.ts +0 -147
  1259. package/tests/types/team-playbook.test.ts +0 -246
  1260. package/tests/types/team-trajectory.test.ts +0 -557
  1261. package/tests/types/trajectory.test.ts +0 -253
  1262. package/tests/utils/error-classifier.test.ts +0 -149
  1263. package/tests/utils/frontmatter.test.ts +0 -208
  1264. package/tests/utils/partitioned-store.test.ts +0 -230
  1265. package/tests/utils/similarity.test.ts +0 -188
  1266. package/tests/utils/trajectory-helpers.test.ts +0 -163
  1267. package/tests/utils/validation.test.ts +0 -252
  1268. package/tests/workspace/efficacy-toolkit.test.ts +0 -404
  1269. package/tests/workspace/full-flow.test.ts +0 -845
  1270. package/tests/workspace/manager.test.ts +0 -215
  1271. package/tests/workspace/runner.test.ts +0 -336
  1272. package/tests/workspace/skill-converter.test.ts +0 -205
  1273. package/tests/workspace/templates/knowledge-extraction.test.ts +0 -235
  1274. package/tests/workspace/templates/playbook-efficacy.test.ts +0 -377
  1275. package/tests/workspace/templates/team-playbook-extraction.test.ts +0 -341
  1276. package/tests/workspace/templates/team-trajectory-analysis.test.ts +0 -417
  1277. package/tsconfig.json +0 -25
  1278. package/vitest.config.ts +0 -22
@@ -0,0 +1,314 @@
1
+ import { spawn } from 'node:child_process';
2
+ import { existsSync } from 'node:fs';
3
+ import { delimiter, dirname } from 'node:path';
4
+ import { generateResponsesText, isResponsesProvider } from 'swarmkit-eval';
5
+ /** Tools disabled when `disallowTools` is set — forces a single-inference step. */
6
+ const DISALLOWED_TOOLS = [
7
+ 'Bash',
8
+ 'Read',
9
+ 'Write',
10
+ 'Edit',
11
+ 'NotebookEdit',
12
+ 'Glob',
13
+ 'Grep',
14
+ 'WebFetch',
15
+ 'WebSearch',
16
+ 'Task',
17
+ 'TodoWrite',
18
+ ];
19
+ const CLAUDE_COMMAND_FALLBACKS = [
20
+ '/usr/local/bin/claude',
21
+ '/opt/homebrew/bin/claude',
22
+ '/Applications/cmux.app/Contents/Resources/bin/claude',
23
+ ];
24
+ export function resolveClaudeCommand() {
25
+ const configured = process.env.CLAUDE_CLI_PATH;
26
+ if (configured && existsSync(configured))
27
+ return configured;
28
+ for (const candidate of CLAUDE_COMMAND_FALLBACKS) {
29
+ if (existsSync(candidate))
30
+ return candidate;
31
+ }
32
+ return 'claude';
33
+ }
34
+ export function claudePathEntries(command = resolveClaudeCommand()) {
35
+ const entries = new Set();
36
+ if (command !== 'claude')
37
+ entries.add(dirname(command));
38
+ for (const candidate of CLAUDE_COMMAND_FALLBACKS) {
39
+ if (existsSync(candidate))
40
+ entries.add(dirname(candidate));
41
+ }
42
+ return [...entries];
43
+ }
44
+ export function buildClaudeChildEnv(base, command = resolveClaudeCommand()) {
45
+ const pathEntries = claudePathEntries(command);
46
+ const existingPath = base.PATH ? [base.PATH] : [];
47
+ return {
48
+ ...base,
49
+ CLAUDE_CLI_PATH: command !== 'claude' ? command : base.CLAUDE_CLI_PATH,
50
+ PATH: [...pathEntries, ...existingPath].filter(Boolean).join(delimiter),
51
+ };
52
+ }
53
+ /**
54
+ * Friendly model name → Bedrock cross-region inference-profile id. Used for
55
+ * ANTHROPIC_MODEL when `bedrock` is set. Unknown names pass through (so a full
56
+ * profile id can be supplied directly). Ids verified against the account's
57
+ * enabled profiles (us-east-1).
58
+ */
59
+ export const BEDROCK_MODELS = {
60
+ haiku: 'us.anthropic.claude-haiku-4-5-20251001-v1:0',
61
+ sonnet: 'us.anthropic.claude-sonnet-4-6',
62
+ opus: 'us.anthropic.claude-opus-4-1-20250805-v1:0',
63
+ };
64
+ /**
65
+ * Cross-region inference profiles are geo-prefixed (`us.`, `eu.`, `apac.`) and
66
+ * a profile only resolves from a region in its own geo. Daily token quotas are
67
+ * per calling region, so being able to point runs at eu-* regions matters when
68
+ * the us-* quotas are exhausted.
69
+ */
70
+ function bedrockGeoPrefix(region) {
71
+ if (region.startsWith('eu-'))
72
+ return 'eu';
73
+ if (region.startsWith('ap-'))
74
+ return 'apac';
75
+ return 'us';
76
+ }
77
+ export function bedrockModelId(name, region) {
78
+ const id = BEDROCK_MODELS[name] ?? name;
79
+ const prefix = bedrockGeoPrefix(region ?? 'us-east-1');
80
+ return prefix === 'us' ? id : id.replace(/^us\./, `${prefix}.`);
81
+ }
82
+ /**
83
+ * Run the Claude CLI with retry-on-failure. A single hung/throttled `claude -p`
84
+ * call should not abort an entire multi-hour eval; we retry a couple of times
85
+ * with backoff before giving up.
86
+ */
87
+ export async function runClaude(prompt, opts = {}) {
88
+ const retries = opts.retries ?? 2;
89
+ let lastErr;
90
+ let nextCommandOverride = opts.commandOverride;
91
+ for (let attempt = 0; attempt <= retries; attempt++) {
92
+ try {
93
+ const attemptOpts = nextCommandOverride
94
+ ? { ...opts, commandOverride: nextCommandOverride }
95
+ : opts;
96
+ return await (isResponsesProvider(opts.provider)
97
+ ? runResponsesOnce(prompt, attemptOpts)
98
+ : runClaudeOnce(prompt, attemptOpts));
99
+ }
100
+ catch (err) {
101
+ lastErr = err;
102
+ // Surface the failure cause so Bedrock throttling/timeouts are visible
103
+ // instead of being silently retried away.
104
+ console.error(`[claude-cli] attempt ${attempt + 1}/${retries + 1} failed: ${lastErr.message.slice(0, 300)}`);
105
+ if (isClaudePathLookupError(lastErr)) {
106
+ const fallback = alternateClaudeCommand(nextCommandOverride ?? resolveClaudeCommand());
107
+ if (fallback) {
108
+ nextCommandOverride = fallback;
109
+ console.error(`[claude-cli] retrying with alternate Claude binary: ${fallback}`);
110
+ }
111
+ }
112
+ if (attempt < retries) {
113
+ // Linear backoff: 5s, 10s — gives a throttled CLI room to recover.
114
+ await new Promise((r) => setTimeout(r, 5000 * (attempt + 1)));
115
+ }
116
+ }
117
+ }
118
+ throw lastErr ?? new Error('claude CLI failed');
119
+ }
120
+ /**
121
+ * Bridge the Azure credential naming gap for the Responses solver/learning path.
122
+ *
123
+ * swarmkit-eval's `generateResponsesText` only reads the canonical
124
+ * `AZURE_OPENAI_*` env names, but this environment commonly exports the shorter
125
+ * `AZURE_API_*` aliases (endpoint/version/key). Resolve both, canonical names
126
+ * winning, and pass them explicitly so `--provider azure` works with whichever
127
+ * credential set is present. Deployment stays `AZURE_OPENAI_DEPLOYMENT` only;
128
+ * when unset the caller's `--model` becomes the deployment name downstream.
129
+ */
130
+ export function resolveAzureResponsesEnv(env = process.env) {
131
+ const apiKey = env.AZURE_OPENAI_API_KEY ?? env.AZURE_OPENAI_KEY ?? env.AZURE_API_KEY;
132
+ const azureEndpoint = env.AZURE_OPENAI_ENDPOINT ?? env.AZURE_OPENAI_BASE_URL ?? env.AZURE_API_BASE;
133
+ const azureApiVersion = env.AZURE_OPENAI_API_VERSION ?? env.AZURE_API_VERSION;
134
+ const azureDeployment = env.AZURE_OPENAI_DEPLOYMENT;
135
+ const out = {};
136
+ if (apiKey)
137
+ out.apiKey = apiKey;
138
+ if (azureEndpoint)
139
+ out.azureEndpoint = azureEndpoint;
140
+ if (azureApiVersion)
141
+ out.azureApiVersion = azureApiVersion;
142
+ if (azureDeployment)
143
+ out.azureDeployment = azureDeployment;
144
+ return out;
145
+ }
146
+ async function runResponsesOnce(prompt, opts = {}) {
147
+ const azure = opts.provider === 'azure' || opts.provider === 'azure-openai' ? resolveAzureResponsesEnv() : {};
148
+ const result = await generateResponsesText({
149
+ provider: opts.provider,
150
+ model: opts.model ?? 'gpt-5.5',
151
+ input: prompt,
152
+ maxOutputTokens: Number(process.env.OPENAI_MAX_OUTPUT_TOKENS) || undefined,
153
+ reasoningEffort: parseReasoningEffort(process.env.OPENAI_REASONING_EFFORT),
154
+ timeoutMs: opts.timeoutMs,
155
+ ...azure,
156
+ });
157
+ return {
158
+ text: result.text,
159
+ ok: true,
160
+ };
161
+ }
162
+ function parseReasoningEffort(value) {
163
+ if (value === 'minimal' || value === 'low' || value === 'medium' || value === 'high') {
164
+ return value;
165
+ }
166
+ return undefined;
167
+ }
168
+ /**
169
+ * Invoke `claude -p --output-format json` with `prompt` on stdin and parse the
170
+ * single JSON result. Rejects on spawn error, non-zero exit, timeout, or
171
+ * unparseable output.
172
+ */
173
+ function runClaudeOnce(prompt, opts = {}) {
174
+ // Always isolate from host plugins/settings. Otherwise every `claude -p` loads
175
+ // the user's global plugins (omc/swarm), each spawning ~5 MCP subprocesses that
176
+ // leak and hang at MCP-init — the dominant cause of multi-minute eval stalls.
177
+ // --strict-mcp-config (no host MCP) + empty --setting-sources keeps eval calls
178
+ // clean and fast (auth via ~/.claude / Bedrock env is unaffected).
179
+ const args = ['-p', '--output-format', 'json', '--strict-mcp-config', '--setting-sources', ''];
180
+ const command = opts.commandOverride ?? resolveClaudeCommand();
181
+ const env = buildClaudeChildEnv(process.env, command);
182
+ // `provider: 'bedrock'` must take the SAME env-var routing as `bedrock: true`.
183
+ // The old behavior passed `--settings '{"apiProvider":"bedrock"}'` — which is
184
+ // not a Claude Code setting, so the CLI silently routed to the user's
185
+ // Anthropic subscription instead of AWS. Found 2026-07-02 when "bedrock"
186
+ // SkillFlow runs died with a subscription 429 ("session limit") — every
187
+ // adapter that passed provider without bedrock (skillflow, clbench,
188
+ // skillsbench) had been burning the subscription.
189
+ const useBedrock = opts.bedrock || opts.provider === 'bedrock';
190
+ if (useBedrock) {
191
+ // Env-var Bedrock routing; DROP --model (CLI aliases are invalid Bedrock ids,
192
+ // → 400). Route the model via ANTHROPIC_MODEL = inference-profile id.
193
+ env.CLAUDE_CODE_USE_BEDROCK = '1';
194
+ env.AWS_REGION = opts.awsRegion ?? process.env.AWS_REGION ?? 'us-east-1';
195
+ if (opts.model)
196
+ env.ANTHROPIC_MODEL = bedrockModelId(opts.model, env.AWS_REGION);
197
+ }
198
+ else if (opts.model) {
199
+ args.push('--model', opts.model);
200
+ }
201
+ if (opts.resumeSessionId)
202
+ args.push('--resume', opts.resumeSessionId);
203
+ if (opts.permissionMode) {
204
+ args.push('--permission-mode', opts.permissionMode);
205
+ }
206
+ if (opts.allowedTools && opts.allowedTools.length > 0) {
207
+ args.push('--allowedTools', ...opts.allowedTools);
208
+ }
209
+ if (opts.provider && !useBedrock) {
210
+ // Legacy settings pass-through for non-bedrock providers. NOTE: this is
211
+ // dubious — `apiProvider` is not a documented Claude Code setting — but
212
+ // responses-providers (openai/azure) never reach runClaudeOnce, so this
213
+ // only fires for exotic values. Bedrock is env-routed above.
214
+ const settings = { apiProvider: opts.provider };
215
+ args.push('--settings', JSON.stringify(settings));
216
+ }
217
+ if (opts.disallowTools) {
218
+ args.push('--disallowedTools', ...DISALLOWED_TOOLS);
219
+ }
220
+ const timeoutMs = opts.timeoutMs ?? 90_000;
221
+ return new Promise((resolve, reject) => {
222
+ const child = spawn(command, args, {
223
+ stdio: ['pipe', 'pipe', 'pipe'],
224
+ env,
225
+ cwd: opts.cwd,
226
+ });
227
+ let stdout = '';
228
+ let stderr = '';
229
+ let settled = false;
230
+ const timer = setTimeout(() => {
231
+ if (settled)
232
+ return;
233
+ settled = true;
234
+ child.kill('SIGKILL');
235
+ reject(new Error(`claude CLI timed out after ${timeoutMs}ms`));
236
+ }, timeoutMs);
237
+ child.stdout.on('data', (d) => (stdout += d.toString()));
238
+ child.stderr.on('data', (d) => (stderr += d.toString()));
239
+ child.on('error', (err) => {
240
+ if (settled)
241
+ return;
242
+ settled = true;
243
+ clearTimeout(timer);
244
+ reject(new Error(`Failed to spawn claude CLI: ${err.message}`));
245
+ });
246
+ child.on('close', (code) => {
247
+ if (settled)
248
+ return;
249
+ settled = true;
250
+ clearTimeout(timer);
251
+ if (code !== 0) {
252
+ const errMsg = stderr.slice(0, 500);
253
+ // Nonzero exit with no stdout is an infra failure (CLI crash, auth
254
+ // error, rate-limit). Use a message pattern isEnvError() can match.
255
+ if (!stdout.trim()) {
256
+ reject(new Error(`claude CLI exited ${code} with no output: ${errMsg}`));
257
+ return;
258
+ }
259
+ reject(new Error(`claude CLI exited ${code}: ${errMsg}`));
260
+ return;
261
+ }
262
+ try {
263
+ const json = JSON.parse(stdout);
264
+ // Bedrock throttling/errors surface as api_error_status or is_error even
265
+ // on a clean exit. Log them so rate-limiting is diagnosable.
266
+ if (json.api_error_status || json.is_error === true) {
267
+ const status = String(json.api_error_status ?? '');
268
+ const subtype = String(json.subtype ?? '');
269
+ console.error(`[claude-cli] API error (provider=${opts.provider ?? 'default'}): ` +
270
+ `status=${status} subtype=${subtype} ` +
271
+ `result=${String(json.result ?? '').slice(0, 200)}`);
272
+ // 429 / 529 rate-limit or overload — surface as env_error-classifiable.
273
+ if (status === '429' ||
274
+ status === '529' ||
275
+ /rate.?limit|too many requests|overloaded/i.test(subtype)) {
276
+ reject(new Error(`claude CLI API ${status || subtype}: rate-limit/overloaded`));
277
+ return;
278
+ }
279
+ }
280
+ resolve({
281
+ text: typeof json.result === 'string' ? json.result : '',
282
+ sessionId: json.session_id,
283
+ ok: json.is_error !== true && !json.api_error_status,
284
+ costUsd: typeof json.total_cost_usd === 'number' ? json.total_cost_usd : undefined,
285
+ });
286
+ }
287
+ catch (err) {
288
+ reject(new Error(`Could not parse claude CLI JSON output: ${err.message}. ` +
289
+ `First 300 chars: ${stdout.slice(0, 300)}`));
290
+ }
291
+ });
292
+ child.stdin.write(prompt);
293
+ child.stdin.end();
294
+ });
295
+ }
296
+ function isClaudePathLookupError(error) {
297
+ return /claude not found in PATH/i.test(error.message);
298
+ }
299
+ function alternateClaudeCommand(current) {
300
+ for (const candidate of CLAUDE_COMMAND_FALLBACKS) {
301
+ if (candidate !== current && existsSync(candidate))
302
+ return candidate;
303
+ }
304
+ return undefined;
305
+ }
306
+ /** Quick availability probe — true if the `claude` binary is on PATH. */
307
+ export async function claudeAvailable() {
308
+ return new Promise((resolve) => {
309
+ const child = spawn(resolveClaudeCommand(), ['--version'], { stdio: 'ignore' });
310
+ child.on('error', () => resolve(false));
311
+ child.on('close', (code) => resolve(code === 0));
312
+ });
313
+ }
314
+ //# sourceMappingURL=claude-cli.js.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"claude-cli.js","sourceRoot":"","sources":["../../../src/eval/live/claude-cli.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,KAAK,EAAE,MAAM,oBAAoB,CAAC;AAC3C,OAAO,EAAE,UAAU,EAAE,MAAM,SAAS,CAAC;AACrC,OAAO,EAAE,SAAS,EAAE,OAAO,EAAE,MAAM,WAAW,CAAC;AAC/C,OAAO,EAAE,qBAAqB,EAAE,mBAAmB,EAAE,MAAM,eAAe,CAAC;AAoD3E,mFAAmF;AACnF,MAAM,gBAAgB,GAAG;IACvB,MAAM;IACN,MAAM;IACN,OAAO;IACP,MAAM;IACN,cAAc;IACd,MAAM;IACN,MAAM;IACN,UAAU;IACV,WAAW;IACX,MAAM;IACN,WAAW;CACZ,CAAC;AAEF,MAAM,wBAAwB,GAAG;IAC/B,uBAAuB;IACvB,0BAA0B;IAC1B,sDAAsD;CACvD,CAAC;AAEF,MAAM,UAAU,oBAAoB;IAClC,MAAM,UAAU,GAAG,OAAO,CAAC,GAAG,CAAC,eAAe,CAAC;IAC/C,IAAI,UAAU,IAAI,UAAU,CAAC,UAAU,CAAC;QAAE,OAAO,UAAU,CAAC;IAC5D,KAAK,MAAM,SAAS,IAAI,wBAAwB,EAAE,CAAC;QACjD,IAAI,UAAU,CAAC,SAAS,CAAC;YAAE,OAAO,SAAS,CAAC;IAC9C,CAAC;IACD,OAAO,QAAQ,CAAC;AAClB,CAAC;AAED,MAAM,UAAU,iBAAiB,CAAC,OAAO,GAAG,oBAAoB,EAAE;IAChE,MAAM,OAAO,GAAG,IAAI,GAAG,EAAU,CAAC;IAClC,IAAI,OAAO,KAAK,QAAQ;QAAE,OAAO,CAAC,GAAG,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC,CAAC;IACxD,KAAK,MAAM,SAAS,IAAI,wBAAwB,EAAE,CAAC;QACjD,IAAI,UAAU,CAAC,SAAS,CAAC;YAAE,OAAO,CAAC,GAAG,CAAC,OAAO,CAAC,SAAS,CAAC,CAAC,CAAC;IAC7D,CAAC;IACD,OAAO,CAAC,GAAG,OAAO,CAAC,CAAC;AACtB,CAAC;AAED,MAAM,UAAU,mBAAmB,CACjC,IAAuB,EACvB,OAAO,GAAG,oBAAoB,EAAE;IAEhC,MAAM,WAAW,GAAG,iBAAiB,CAAC,OAAO,CAAC,CAAC;IAC/C,MAAM,YAAY,GAAG,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;IAClD,OAAO;QACL,GAAG,IAAI;QACP,eAAe,EAAE,OAAO,KAAK,QAAQ,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,IAAI,CAAC,eAAe;QACtE,IAAI,EAAE,CAAC,GAAG,WAAW,EAAE,GAAG,YAAY,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,IAAI,CAAC,SAAS,CAAC;KACxE,CAAC;AACJ,CAAC;AAED;;;;;GAKG;AACH,MAAM,CAAC,MAAM,cAAc,GAA2B;IACpD,KAAK,EAAE,6CAA6C;IACpD,MAAM,EAAE,gCAAgC;IACxC,IAAI,EAAE,4CAA4C;CACnD,CAAC;AAEF;;;;;GAKG;AACH,SAAS,gBAAgB,CAAC,MAAc;IACtC,IAAI,MAAM,CAAC,UAAU,CAAC,KAAK,CAAC;QAAE,OAAO,IAAI,CAAC;IAC1C,IAAI,MAAM,CAAC,UAAU,CAAC,KAAK,CAAC;QAAE,OAAO,MAAM,CAAC;IAC5C,OAAO,IAAI,CAAC;AACd,CAAC;AAED,MAAM,UAAU,cAAc,CAAC,IAAY,EAAE,MAAe;IAC1D,MAAM,EAAE,GAAG,cAAc,CAAC,IAAI,CAAC,IAAI,IAAI,CAAC;IACxC,MAAM,MAAM,GAAG,gBAAgB,CAAC,MAAM,IAAI,WAAW,CAAC,CAAC;IACvD,OAAO,MAAM,KAAK,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC,OAAO,CAAC,OAAO,EAAE,GAAG,MAAM,GAAG,CAAC,CAAC;AAClE,CAAC;AAED;;;;GAIG;AACH,MAAM,CAAC,KAAK,UAAU,SAAS,CAC7B,MAAc,EACd,OAAyB,EAAE;IAE3B,MAAM,OAAO,GAAG,IAAI,CAAC,OAAO,IAAI,CAAC,CAAC;IAClC,IAAI,OAA0B,CAAC;IAC/B,IAAI,mBAAmB,GAAG,IAAI,CAAC,eAAe,CAAC;IAC/C,KAAK,IAAI,OAAO,GAAG,CAAC,EAAE,OAAO,IAAI,OAAO,EAAE,OAAO,EAAE,EAAE,CAAC;QACpD,IAAI,CAAC;YACH,MAAM,WAAW,GAAG,mBAAmB;gBACrC,CAAC,CAAC,EAAE,GAAG,IAAI,EAAE,eAAe,EAAE,mBAAmB,EAAE;gBACnD,CAAC,CAAC,IAAI,CAAC;YACT,OAAO,MAAM,CAAC,mBAAmB,CAAC,IAAI,CAAC,QAAQ,CAAC;gBAC9C,CAAC,CAAC,gBAAgB,CAAC,MAAM,EAAE,WAAW,CAAC;gBACvC,CAAC,CAAC,aAAa,CAAC,MAAM,EAAE,WAAW,CAAC,CAAC,CAAC;QAC1C,CAAC;QAAC,OAAO,GAAG,EAAE,CAAC;YACb,OAAO,GAAG,GAAY,CAAC;YACvB,uEAAuE;YACvE,0CAA0C;YAC1C,OAAO,CAAC,KAAK,CACX,wBAAwB,OAAO,GAAG,CAAC,IAAI,OAAO,GAAG,CAAC,YAAY,OAAO,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC,EAAE,CAC9F,CAAC;YACF,IAAI,uBAAuB,CAAC,OAAO,CAAC,EAAE,CAAC;gBACrC,MAAM,QAAQ,GAAG,sBAAsB,CAAC,mBAAmB,IAAI,oBAAoB,EAAE,CAAC,CAAC;gBACvF,IAAI,QAAQ,EAAE,CAAC;oBACb,mBAAmB,GAAG,QAAQ,CAAC;oBAC/B,OAAO,CAAC,KAAK,CAAC,uDAAuD,QAAQ,EAAE,CAAC,CAAC;gBACnF,CAAC;YACH,CAAC;YACD,IAAI,OAAO,GAAG,OAAO,EAAE,CAAC;gBACtB,mEAAmE;gBACnE,MAAM,IAAI,OAAO,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,UAAU,CAAC,CAAC,EAAE,IAAI,GAAG,CAAC,OAAO,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC;YAChE,CAAC;QACH,CAAC;IACH,CAAC;IACD,MAAM,OAAO,IAAI,IAAI,KAAK,CAAC,mBAAmB,CAAC,CAAC;AAClD,CAAC;AAUD;;;;;;;;;GASG;AACH,MAAM,UAAU,wBAAwB,CAAC,MAAyB,OAAO,CAAC,GAAG;IAC3E,MAAM,MAAM,GAAG,GAAG,CAAC,oBAAoB,IAAI,GAAG,CAAC,gBAAgB,IAAI,GAAG,CAAC,aAAa,CAAC;IACrF,MAAM,aAAa,GACjB,GAAG,CAAC,qBAAqB,IAAI,GAAG,CAAC,qBAAqB,IAAI,GAAG,CAAC,cAAc,CAAC;IAC/E,MAAM,eAAe,GAAG,GAAG,CAAC,wBAAwB,IAAI,GAAG,CAAC,iBAAiB,CAAC;IAC9E,MAAM,eAAe,GAAG,GAAG,CAAC,uBAAuB,CAAC;IAEpD,MAAM,GAAG,GAAsB,EAAE,CAAC;IAClC,IAAI,MAAM;QAAE,GAAG,CAAC,MAAM,GAAG,MAAM,CAAC;IAChC,IAAI,aAAa;QAAE,GAAG,CAAC,aAAa,GAAG,aAAa,CAAC;IACrD,IAAI,eAAe;QAAE,GAAG,CAAC,eAAe,GAAG,eAAe,CAAC;IAC3D,IAAI,eAAe;QAAE,GAAG,CAAC,eAAe,GAAG,eAAe,CAAC;IAC3D,OAAO,GAAG,CAAC;AACb,CAAC;AAED,KAAK,UAAU,gBAAgB,CAC7B,MAAc,EACd,OAAyB,EAAE;IAE3B,MAAM,KAAK,GACT,IAAI,CAAC,QAAQ,KAAK,OAAO,IAAI,IAAI,CAAC,QAAQ,KAAK,cAAc,CAAC,CAAC,CAAC,wBAAwB,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;IAClG,MAAM,MAAM,GAAG,MAAM,qBAAqB,CAAC;QACzC,QAAQ,EAAE,IAAI,CAAC,QAAQ;QACvB,KAAK,EAAE,IAAI,CAAC,KAAK,IAAI,SAAS;QAC9B,KAAK,EAAE,MAAM;QACb,eAAe,EAAE,MAAM,CAAC,OAAO,CAAC,GAAG,CAAC,wBAAwB,CAAC,IAAI,SAAS;QAC1E,eAAe,EAAE,oBAAoB,CAAC,OAAO,CAAC,GAAG,CAAC,uBAAuB,CAAC;QAC1E,SAAS,EAAE,IAAI,CAAC,SAAS;QACzB,GAAG,KAAK;KACT,CAAC,CAAC;IACH,OAAO;QACL,IAAI,EAAE,MAAM,CAAC,IAAI;QACjB,EAAE,EAAE,IAAI;KACT,CAAC;AACJ,CAAC;AAaD,SAAS,oBAAoB,CAC3B,KAAyB;IAEzB,IAAI,KAAK,KAAK,SAAS,IAAI,KAAK,KAAK,KAAK,IAAI,KAAK,KAAK,QAAQ,IAAI,KAAK,KAAK,MAAM,EAAE,CAAC;QACrF,OAAO,KAAK,CAAC;IACf,CAAC;IACD,OAAO,SAAS,CAAC;AACnB,CAAC;AAED;;;;GAIG;AACH,SAAS,aAAa,CAAC,MAAc,EAAE,OAAyB,EAAE;IAChE,+EAA+E;IAC/E,gFAAgF;IAChF,8EAA8E;IAC9E,+EAA+E;IAC/E,mEAAmE;IACnE,MAAM,IAAI,GAAG,CAAC,IAAI,EAAE,iBAAiB,EAAE,MAAM,EAAE,qBAAqB,EAAE,mBAAmB,EAAE,EAAE,CAAC,CAAC;IAC/F,MAAM,OAAO,GAAG,IAAI,CAAC,eAAe,IAAI,oBAAoB,EAAE,CAAC;IAC/D,MAAM,GAAG,GAAG,mBAAmB,CAAC,OAAO,CAAC,GAAG,EAAE,OAAO,CAAC,CAAC;IAEtD,+EAA+E;IAC/E,8EAA8E;IAC9E,sEAAsE;IACtE,yEAAyE;IACzE,wEAAwE;IACxE,oEAAoE;IACpE,kDAAkD;IAClD,MAAM,UAAU,GAAG,IAAI,CAAC,OAAO,IAAI,IAAI,CAAC,QAAQ,KAAK,SAAS,CAAC;IAC/D,IAAI,UAAU,EAAE,CAAC;QACf,8EAA8E;QAC9E,sEAAsE;QACtE,GAAG,CAAC,uBAAuB,GAAG,GAAG,CAAC;QAClC,GAAG,CAAC,UAAU,GAAG,IAAI,CAAC,SAAS,IAAI,OAAO,CAAC,GAAG,CAAC,UAAU,IAAI,WAAW,CAAC;QACzE,IAAI,IAAI,CAAC,KAAK;YAAE,GAAG,CAAC,eAAe,GAAG,cAAc,CAAC,IAAI,CAAC,KAAK,EAAE,GAAG,CAAC,UAAU,CAAC,CAAC;IACnF,CAAC;SAAM,IAAI,IAAI,CAAC,KAAK,EAAE,CAAC;QACtB,IAAI,CAAC,IAAI,CAAC,SAAS,EAAE,IAAI,CAAC,KAAK,CAAC,CAAC;IACnC,CAAC;IACD,IAAI,IAAI,CAAC,eAAe;QAAE,IAAI,CAAC,IAAI,CAAC,UAAU,EAAE,IAAI,CAAC,eAAe,CAAC,CAAC;IACtE,IAAI,IAAI,CAAC,cAAc,EAAE,CAAC;QACxB,IAAI,CAAC,IAAI,CAAC,mBAAmB,EAAE,IAAI,CAAC,cAAc,CAAC,CAAC;IACtD,CAAC;IACD,IAAI,IAAI,CAAC,YAAY,IAAI,IAAI,CAAC,YAAY,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QACtD,IAAI,CAAC,IAAI,CAAC,gBAAgB,EAAE,GAAG,IAAI,CAAC,YAAY,CAAC,CAAC;IACpD,CAAC;IACD,IAAI,IAAI,CAAC,QAAQ,IAAI,CAAC,UAAU,EAAE,CAAC;QACjC,wEAAwE;QACxE,wEAAwE;QACxE,wEAAwE;QACxE,6DAA6D;QAC7D,MAAM,QAAQ,GAA2B,EAAE,WAAW,EAAE,IAAI,CAAC,QAAQ,EAAE,CAAC;QACxE,IAAI,CAAC,IAAI,CAAC,YAAY,EAAE,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAC,CAAC,CAAC;IACpD,CAAC;IACD,IAAI,IAAI,CAAC,aAAa,EAAE,CAAC;QACvB,IAAI,CAAC,IAAI,CAAC,mBAAmB,EAAE,GAAG,gBAAgB,CAAC,CAAC;IACtD,CAAC;IAED,MAAM,SAAS,GAAG,IAAI,CAAC,SAAS,IAAI,MAAM,CAAC;IAE3C,OAAO,IAAI,OAAO,CAAkB,CAAC,OAAO,EAAE,MAAM,EAAE,EAAE;QACtD,MAAM,KAAK,GAAG,KAAK,CAAC,OAAO,EAAE,IAAI,EAAE;YACjC,KAAK,EAAE,CAAC,MAAM,EAAE,MAAM,EAAE,MAAM,CAAC;YAC/B,GAAG;YACH,GAAG,EAAE,IAAI,CAAC,GAAG;SACd,CAAC,CAAC;QAEH,IAAI,MAAM,GAAG,EAAE,CAAC;QAChB,IAAI,MAAM,GAAG,EAAE,CAAC;QAChB,IAAI,OAAO,GAAG,KAAK,CAAC;QAEpB,MAAM,KAAK,GAAG,UAAU,CAAC,GAAG,EAAE;YAC5B,IAAI,OAAO;gBAAE,OAAO;YACpB,OAAO,GAAG,IAAI,CAAC;YACf,KAAK,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC;YACtB,MAAM,CAAC,IAAI,KAAK,CAAC,8BAA8B,SAAS,IAAI,CAAC,CAAC,CAAC;QACjE,CAAC,EAAE,SAAS,CAAC,CAAC;QAEd,KAAK,CAAC,MAAM,CAAC,EAAE,CAAC,MAAM,EAAE,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,MAAM,IAAI,CAAC,CAAC,QAAQ,EAAE,CAAC,CAAC,CAAC;QACzD,KAAK,CAAC,MAAM,CAAC,EAAE,CAAC,MAAM,EAAE,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,MAAM,IAAI,CAAC,CAAC,QAAQ,EAAE,CAAC,CAAC,CAAC;QAEzD,KAAK,CAAC,EAAE,CAAC,OAAO,EAAE,CAAC,GAAG,EAAE,EAAE;YACxB,IAAI,OAAO;gBAAE,OAAO;YACpB,OAAO,GAAG,IAAI,CAAC;YACf,YAAY,CAAC,KAAK,CAAC,CAAC;YACpB,MAAM,CAAC,IAAI,KAAK,CAAC,+BAA+B,GAAG,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC;QAClE,CAAC,CAAC,CAAC;QAEH,KAAK,CAAC,EAAE,CAAC,OAAO,EAAE,CAAC,IAAI,EAAE,EAAE;YACzB,IAAI,OAAO;gBAAE,OAAO;YACpB,OAAO,GAAG,IAAI,CAAC;YACf,YAAY,CAAC,KAAK,CAAC,CAAC;YACpB,IAAI,IAAI,KAAK,CAAC,EAAE,CAAC;gBACf,MAAM,MAAM,GAAG,MAAM,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC,CAAC;gBACpC,mEAAmE;gBACnE,oEAAoE;gBACpE,IAAI,CAAC,MAAM,CAAC,IAAI,EAAE,EAAE,CAAC;oBACnB,MAAM,CAAC,IAAI,KAAK,CAAC,qBAAqB,IAAI,oBAAoB,MAAM,EAAE,CAAC,CAAC,CAAC;oBACzE,OAAO;gBACT,CAAC;gBACD,MAAM,CAAC,IAAI,KAAK,CAAC,qBAAqB,IAAI,KAAK,MAAM,EAAE,CAAC,CAAC,CAAC;gBAC1D,OAAO;YACT,CAAC;YACD,IAAI,CAAC;gBACH,MAAM,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC;gBAChC,yEAAyE;gBACzE,6DAA6D;gBAC7D,IAAI,IAAI,CAAC,gBAAgB,IAAI,IAAI,CAAC,QAAQ,KAAK,IAAI,EAAE,CAAC;oBACpD,MAAM,MAAM,GAAG,MAAM,CAAC,IAAI,CAAC,gBAAgB,IAAI,EAAE,CAAC,CAAC;oBACnD,MAAM,OAAO,GAAG,MAAM,CAAC,IAAI,CAAC,OAAO,IAAI,EAAE,CAAC,CAAC;oBAC3C,OAAO,CAAC,KAAK,CACX,oCAAoC,IAAI,CAAC,QAAQ,IAAI,SAAS,KAAK;wBACjE,UAAU,MAAM,YAAY,OAAO,GAAG;wBACtC,UAAU,MAAM,CAAC,IAAI,CAAC,MAAM,IAAI,EAAE,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC,EAAE,CACtD,CAAC;oBACF,wEAAwE;oBACxE,IACE,MAAM,KAAK,KAAK;wBAChB,MAAM,KAAK,KAAK;wBAChB,2CAA2C,CAAC,IAAI,CAAC,OAAO,CAAC,EACzD,CAAC;wBACD,MAAM,CAAC,IAAI,KAAK,CAAC,kBAAkB,MAAM,IAAI,OAAO,yBAAyB,CAAC,CAAC,CAAC;wBAChF,OAAO;oBACT,CAAC;gBACH,CAAC;gBACD,OAAO,CAAC;oBACN,IAAI,EAAE,OAAO,IAAI,CAAC,MAAM,KAAK,QAAQ,CAAC,CAAC,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE;oBACxD,SAAS,EAAE,IAAI,CAAC,UAAU;oBAC1B,EAAE,EAAE,IAAI,CAAC,QAAQ,KAAK,IAAI,IAAI,CAAC,IAAI,CAAC,gBAAgB;oBACpD,OAAO,EAAE,OAAO,IAAI,CAAC,cAAc,KAAK,QAAQ,CAAC,CAAC,CAAC,IAAI,CAAC,cAAc,CAAC,CAAC,CAAC,SAAS;iBACnF,CAAC,CAAC;YACL,CAAC;YAAC,OAAO,GAAG,EAAE,CAAC;gBACb,MAAM,CACJ,IAAI,KAAK,CACP,2CAA4C,GAAa,CAAC,OAAO,IAAI;oBACnE,oBAAoB,MAAM,CAAC,KAAK,CAAC,CAAC,EAAE,GAAG,CAAC,EAAE,CAC7C,CACF,CAAC;YACJ,CAAC;QACH,CAAC,CAAC,CAAC;QAEH,KAAK,CAAC,KAAK,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC;QAC1B,KAAK,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC;IACpB,CAAC,CAAC,CAAC;AACL,CAAC;AAED,SAAS,uBAAuB,CAAC,KAAY;IAC3C,OAAO,2BAA2B,CAAC,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC;AACzD,CAAC;AAED,SAAS,sBAAsB,CAAC,OAAe;IAC7C,KAAK,MAAM,SAAS,IAAI,wBAAwB,EAAE,CAAC;QACjD,IAAI,SAAS,KAAK,OAAO,IAAI,UAAU,CAAC,SAAS,CAAC;YAAE,OAAO,SAAS,CAAC;IACvE,CAAC;IACD,OAAO,SAAS,CAAC;AACnB,CAAC;AAED,yEAAyE;AACzE,MAAM,CAAC,KAAK,UAAU,eAAe;IACnC,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,EAAE,EAAE;QAC7B,MAAM,KAAK,GAAG,KAAK,CAAC,oBAAoB,EAAE,EAAE,CAAC,WAAW,CAAC,EAAE,EAAE,KAAK,EAAE,QAAQ,EAAE,CAAC,CAAC;QAChF,KAAK,CAAC,EAAE,CAAC,OAAO,EAAE,GAAG,EAAE,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC,CAAC;QACxC,KAAK,CAAC,EAAE,CAAC,OAAO,EAAE,CAAC,IAAI,EAAE,EAAE,CAAC,OAAO,CAAC,IAAI,KAAK,CAAC,CAAC,CAAC,CAAC;IACnD,CAAC,CAAC,CAAC;AACL,CAAC"}
@@ -0,0 +1,41 @@
1
+ import type { AgentDelegate, AgentDelegateOptions, AgentDelegateResult } from '../../runtime/types.js';
2
+ import type { Task } from '../../types/index.js';
3
+ import type { LiveBenchmark } from './live-benchmark.js';
4
+ /**
5
+ * Single-turn agent delegate backed by the headless Claude CLI.
6
+ *
7
+ * Atlas injects its learned knowledge into `options.systemContext`; this
8
+ * delegate prepends it to the task prompt, runs `claude -p`, then **verifies the
9
+ * answer against ground truth** via the benchmark. The verified correctness
10
+ * becomes the returned `success`, so it flows into `trajectory.outcome.success`
11
+ * — meaning the learning pipeline learns from correctly-labeled outcomes, not
12
+ * the agent's self-report.
13
+ *
14
+ * The runner sets `currentTask` immediately before each `atlas.solve`, which is
15
+ * safe because the eval runs tasks strictly sequentially.
16
+ */
17
+ export declare class ClaudeBenchmarkDelegate implements AgentDelegate {
18
+ private readonly benchmark;
19
+ private readonly opts;
20
+ private currentTask;
21
+ private lastCostUsd;
22
+ private totalCostUsd;
23
+ constructor(benchmark: LiveBenchmark, opts?: {
24
+ model?: string;
25
+ timeoutMs?: number;
26
+ provider?: string;
27
+ awsRegion?: string;
28
+ /** Per-call retry override (failFast → 0 = single attempt). */
29
+ cliRetries?: number;
30
+ });
31
+ /** Set the task being solved next (its gold answer is read during verify). */
32
+ setCurrentTask(task: Task): void;
33
+ /** Cumulative USD cost across all calls (best-effort, from the CLI). */
34
+ getTotalCostUsd(): number;
35
+ execute(prompt: string, options: AgentDelegateOptions): Promise<AgentDelegateResult>;
36
+ }
37
+ export declare function createClaudeBenchmarkDelegate(benchmark: LiveBenchmark, opts?: {
38
+ model?: string;
39
+ timeoutMs?: number;
40
+ }): ClaudeBenchmarkDelegate;
41
+ //# sourceMappingURL=claude-delegate.d.ts.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"claude-delegate.d.ts","sourceRoot":"","sources":["../../../src/eval/live/claude-delegate.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EACV,aAAa,EACb,oBAAoB,EACpB,mBAAmB,EACpB,MAAM,wBAAwB,CAAC;AAChC,OAAO,KAAK,EAAE,IAAI,EAAE,MAAM,sBAAsB,CAAC;AACjD,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,qBAAqB,CAAC;AAGzD;;;;;;;;;;;;GAYG;AACH,qBAAa,uBAAwB,YAAW,aAAa;IAMzD,OAAO,CAAC,QAAQ,CAAC,SAAS;IAC1B,OAAO,CAAC,QAAQ,CAAC,IAAI;IANvB,OAAO,CAAC,WAAW,CAAqB;IACxC,OAAO,CAAC,WAAW,CAAK;IACxB,OAAO,CAAC,YAAY,CAAK;gBAGN,SAAS,EAAE,aAAa,EACxB,IAAI,GAAE;QACrB,KAAK,CAAC,EAAE,MAAM,CAAC;QACf,SAAS,CAAC,EAAE,MAAM,CAAC;QACnB,QAAQ,CAAC,EAAE,MAAM,CAAC;QAClB,SAAS,CAAC,EAAE,MAAM,CAAC;QACnB,+DAA+D;QAC/D,UAAU,CAAC,EAAE,MAAM,CAAC;KAChB;IAGR,8EAA8E;IAC9E,cAAc,CAAC,IAAI,EAAE,IAAI,GAAG,IAAI;IAIhC,wEAAwE;IACxE,eAAe,IAAI,MAAM;IAInB,OAAO,CAAC,MAAM,EAAE,MAAM,EAAE,OAAO,EAAE,oBAAoB,GAAG,OAAO,CAAC,mBAAmB,CAAC;CA0B3F;AAED,wBAAgB,6BAA6B,CAC3C,SAAS,EAAE,aAAa,EACxB,IAAI,CAAC,EAAE;IAAE,KAAK,CAAC,EAAE,MAAM,CAAC;IAAC,SAAS,CAAC,EAAE,MAAM,CAAA;CAAE,GAC5C,uBAAuB,CAEzB"}
@@ -0,0 +1,59 @@
1
+ import { runClaude } from './claude-cli.js';
2
+ /**
3
+ * Single-turn agent delegate backed by the headless Claude CLI.
4
+ *
5
+ * Atlas injects its learned knowledge into `options.systemContext`; this
6
+ * delegate prepends it to the task prompt, runs `claude -p`, then **verifies the
7
+ * answer against ground truth** via the benchmark. The verified correctness
8
+ * becomes the returned `success`, so it flows into `trajectory.outcome.success`
9
+ * — meaning the learning pipeline learns from correctly-labeled outcomes, not
10
+ * the agent's self-report.
11
+ *
12
+ * The runner sets `currentTask` immediately before each `atlas.solve`, which is
13
+ * safe because the eval runs tasks strictly sequentially.
14
+ */
15
+ export class ClaudeBenchmarkDelegate {
16
+ benchmark;
17
+ opts;
18
+ currentTask = null;
19
+ lastCostUsd = 0;
20
+ totalCostUsd = 0;
21
+ constructor(benchmark, opts = {}) {
22
+ this.benchmark = benchmark;
23
+ this.opts = opts;
24
+ }
25
+ /** Set the task being solved next (its gold answer is read during verify). */
26
+ setCurrentTask(task) {
27
+ this.currentTask = task;
28
+ }
29
+ /** Cumulative USD cost across all calls (best-effort, from the CLI). */
30
+ getTotalCostUsd() {
31
+ return this.totalCostUsd;
32
+ }
33
+ async execute(prompt, options) {
34
+ // systemContext carries Atlas's injected playbooks/experiences — keep it
35
+ // first so accumulated learning shapes the agent's approach.
36
+ const fullPrompt = [options.systemContext, prompt].filter(Boolean).join('\n\n');
37
+ const res = await runClaude(fullPrompt, {
38
+ model: this.opts.model,
39
+ timeoutMs: this.opts.timeoutMs ?? options.timeoutMs,
40
+ provider: this.opts.provider,
41
+ awsRegion: this.opts.awsRegion,
42
+ retries: this.opts.cliRetries,
43
+ });
44
+ this.lastCostUsd = res.costUsd ?? 0;
45
+ this.totalCostUsd += this.lastCostUsd;
46
+ const verdict = this.currentTask
47
+ ? this.benchmark.verify(this.currentTask, res.text)
48
+ : { correct: false, extracted: undefined };
49
+ return {
50
+ success: verdict.correct,
51
+ output: res.text,
52
+ structured: { extracted: verdict.extracted, costUsd: this.lastCostUsd },
53
+ };
54
+ }
55
+ }
56
+ export function createClaudeBenchmarkDelegate(benchmark, opts) {
57
+ return new ClaudeBenchmarkDelegate(benchmark, opts);
58
+ }
59
+ //# sourceMappingURL=claude-delegate.js.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"claude-delegate.js","sourceRoot":"","sources":["../../../src/eval/live/claude-delegate.ts"],"names":[],"mappings":"AAOA,OAAO,EAAE,SAAS,EAAE,MAAM,iBAAiB,CAAC;AAE5C;;;;;;;;;;;;GAYG;AACH,MAAM,OAAO,uBAAuB;IAMf;IACA;IANX,WAAW,GAAgB,IAAI,CAAC;IAChC,WAAW,GAAG,CAAC,CAAC;IAChB,YAAY,GAAG,CAAC,CAAC;IAEzB,YACmB,SAAwB,EACxB,OAOb,EAAE;QARW,cAAS,GAAT,SAAS,CAAe;QACxB,SAAI,GAAJ,IAAI,CAOf;IACL,CAAC;IAEJ,8EAA8E;IAC9E,cAAc,CAAC,IAAU;QACvB,IAAI,CAAC,WAAW,GAAG,IAAI,CAAC;IAC1B,CAAC;IAED,wEAAwE;IACxE,eAAe;QACb,OAAO,IAAI,CAAC,YAAY,CAAC;IAC3B,CAAC;IAED,KAAK,CAAC,OAAO,CAAC,MAAc,EAAE,OAA6B;QACzD,yEAAyE;QACzE,6DAA6D;QAC7D,MAAM,UAAU,GAAG,CAAC,OAAO,CAAC,aAAa,EAAE,MAAM,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;QAEhF,MAAM,GAAG,GAAG,MAAM,SAAS,CAAC,UAAU,EAAE;YACtC,KAAK,EAAE,IAAI,CAAC,IAAI,CAAC,KAAK;YACtB,SAAS,EAAE,IAAI,CAAC,IAAI,CAAC,SAAS,IAAI,OAAO,CAAC,SAAS;YACnD,QAAQ,EAAE,IAAI,CAAC,IAAI,CAAC,QAAQ;YAC5B,SAAS,EAAE,IAAI,CAAC,IAAI,CAAC,SAAS;YAC9B,OAAO,EAAE,IAAI,CAAC,IAAI,CAAC,UAAU;SAC9B,CAAC,CAAC;QAEH,IAAI,CAAC,WAAW,GAAG,GAAG,CAAC,OAAO,IAAI,CAAC,CAAC;QACpC,IAAI,CAAC,YAAY,IAAI,IAAI,CAAC,WAAW,CAAC;QAEtC,MAAM,OAAO,GAAG,IAAI,CAAC,WAAW;YAC9B,CAAC,CAAC,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC,IAAI,CAAC,WAAW,EAAE,GAAG,CAAC,IAAI,CAAC;YACnD,CAAC,CAAC,EAAE,OAAO,EAAE,KAAK,EAAE,SAAS,EAAE,SAAS,EAAE,CAAC;QAE7C,OAAO;YACL,OAAO,EAAE,OAAO,CAAC,OAAO;YACxB,MAAM,EAAE,GAAG,CAAC,IAAI;YAChB,UAAU,EAAE,EAAE,SAAS,EAAE,OAAO,CAAC,SAAS,EAAE,OAAO,EAAE,IAAI,CAAC,WAAW,EAAE;SACxE,CAAC;IACJ,CAAC;CACF;AAED,MAAM,UAAU,6BAA6B,CAC3C,SAAwB,EACxB,IAA6C;IAE7C,OAAO,IAAI,uBAAuB,CAAC,SAAS,EAAE,IAAI,CAAC,CAAC;AACtD,CAAC"}
@@ -0,0 +1,24 @@
1
+ import type { AgentBackend, AgentSession, AgentSpawnConfig } from '../../runtime/types.js';
2
+ /**
3
+ * General Claude CLI backend for offline workspace-analysis agents.
4
+ *
5
+ * Unlike benchmark delegates, this backend does not verify task success against
6
+ * a benchmark. It runs Claude in the supplied workspace cwd and lets templates
7
+ * validate the files they asked the agent to write.
8
+ */
9
+ export declare class ClaudeWorkspaceBackend implements AgentBackend {
10
+ private readonly opts;
11
+ readonly name = "claude-workspace";
12
+ readonly supportedTypes: string[];
13
+ private readonly sessions;
14
+ constructor(opts?: {
15
+ model?: string;
16
+ timeoutMs?: number;
17
+ provider?: string;
18
+ awsRegion?: string;
19
+ });
20
+ isAvailable(): Promise<boolean>;
21
+ spawn(config: AgentSpawnConfig): Promise<AgentSession>;
22
+ getSession(sessionId: string): Promise<AgentSession | undefined>;
23
+ }
24
+ //# sourceMappingURL=claude-workspace-backend.d.ts.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"claude-workspace-backend.d.ts","sourceRoot":"","sources":["../../../src/eval/live/claude-workspace-backend.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EACV,YAAY,EAEZ,YAAY,EACZ,gBAAgB,EACjB,MAAM,wBAAwB,CAAC;AAGhC;;;;;;GAMG;AACH,qBAAa,sBAAuB,YAAW,YAAY;IAOvD,OAAO,CAAC,QAAQ,CAAC,IAAI;IANvB,QAAQ,CAAC,IAAI,sBAAsB;IACnC,QAAQ,CAAC,cAAc,WAAmB;IAE1C,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAmC;gBAGzC,IAAI,GAAE;QACrB,KAAK,CAAC,EAAE,MAAM,CAAC;QACf,SAAS,CAAC,EAAE,MAAM,CAAC;QACnB,QAAQ,CAAC,EAAE,MAAM,CAAC;QAClB,SAAS,CAAC,EAAE,MAAM,CAAC;KACf;IAGF,WAAW,IAAI,OAAO,CAAC,OAAO,CAAC;IAI/B,KAAK,CAAC,MAAM,EAAE,gBAAgB,GAAG,OAAO,CAAC,YAAY,CAAC;IAuDtD,UAAU,CAAC,SAAS,EAAE,MAAM,GAAG,OAAO,CAAC,YAAY,GAAG,SAAS,CAAC;CAGvE"}
@@ -0,0 +1,76 @@
1
+ import { randomUUID } from 'node:crypto';
2
+ import { runClaude } from './claude-cli.js';
3
+ /**
4
+ * General Claude CLI backend for offline workspace-analysis agents.
5
+ *
6
+ * Unlike benchmark delegates, this backend does not verify task success against
7
+ * a benchmark. It runs Claude in the supplied workspace cwd and lets templates
8
+ * validate the files they asked the agent to write.
9
+ */
10
+ export class ClaudeWorkspaceBackend {
11
+ opts;
12
+ name = 'claude-workspace';
13
+ supportedTypes = ['claude-code'];
14
+ sessions = new Map();
15
+ constructor(opts = {}) {
16
+ this.opts = opts;
17
+ }
18
+ async isAvailable() {
19
+ return true;
20
+ }
21
+ async spawn(config) {
22
+ const startTime = new Date();
23
+ const session = {
24
+ id: randomUUID(),
25
+ agentType: config.agentType,
26
+ task: config.task,
27
+ state: 'running',
28
+ messages: [],
29
+ toolCalls: [],
30
+ startTime,
31
+ metadata: { backend: this.name },
32
+ };
33
+ this.sessions.set(session.id, session);
34
+ try {
35
+ const prompt = [
36
+ config.systemPromptAdditions,
37
+ 'You are running as a workspace automation agent. Required output files must be created on disk in this working directory. Use file tools to write the requested output files; do not only describe the result in your final answer. The caller validates files on disk and ignores prose unless the files exist.',
38
+ config.task.description,
39
+ ]
40
+ .filter(Boolean)
41
+ .join('\n\n');
42
+ const result = await runClaude(prompt, {
43
+ model: this.opts.model,
44
+ timeoutMs: config.timeout ?? this.opts.timeoutMs,
45
+ provider: this.opts.provider,
46
+ awsRegion: this.opts.awsRegion,
47
+ cwd: config.cwd,
48
+ permissionMode: 'acceptEdits',
49
+ allowedTools: ['Read', 'Write', 'Edit', 'Glob', 'Grep'],
50
+ });
51
+ const message = {
52
+ role: 'assistant',
53
+ content: result.text,
54
+ timestamp: new Date(),
55
+ };
56
+ session.messages.push(message);
57
+ session.result = result.text;
58
+ session.state = result.ok ? 'completed' : 'failed';
59
+ session.endTime = new Date();
60
+ if (!result.ok) {
61
+ session.error = 'Claude CLI returned an error result';
62
+ }
63
+ }
64
+ catch (error) {
65
+ session.state = 'failed';
66
+ session.error = error instanceof Error ? error.message : String(error);
67
+ session.endTime = new Date();
68
+ }
69
+ this.sessions.set(session.id, session);
70
+ return session;
71
+ }
72
+ async getSession(sessionId) {
73
+ return this.sessions.get(sessionId);
74
+ }
75
+ }
76
+ //# sourceMappingURL=claude-workspace-backend.js.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"claude-workspace-backend.js","sourceRoot":"","sources":["../../../src/eval/live/claude-workspace-backend.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,UAAU,EAAE,MAAM,aAAa,CAAC;AAOzC,OAAO,EAAE,SAAS,EAAE,MAAM,iBAAiB,CAAC;AAE5C;;;;;;GAMG;AACH,MAAM,OAAO,sBAAsB;IAOd;IANV,IAAI,GAAG,kBAAkB,CAAC;IAC1B,cAAc,GAAG,CAAC,aAAa,CAAC,CAAC;IAEzB,QAAQ,GAAG,IAAI,GAAG,EAAwB,CAAC;IAE5D,YACmB,OAKb,EAAE;QALW,SAAI,GAAJ,IAAI,CAKf;IACL,CAAC;IAEJ,KAAK,CAAC,WAAW;QACf,OAAO,IAAI,CAAC;IACd,CAAC;IAED,KAAK,CAAC,KAAK,CAAC,MAAwB;QAClC,MAAM,SAAS,GAAG,IAAI,IAAI,EAAE,CAAC;QAC7B,MAAM,OAAO,GAAiB;YAC5B,EAAE,EAAE,UAAU,EAAE;YAChB,SAAS,EAAE,MAAM,CAAC,SAAS;YAC3B,IAAI,EAAE,MAAM,CAAC,IAAI;YACjB,KAAK,EAAE,SAAS;YAChB,QAAQ,EAAE,EAAE;YACZ,SAAS,EAAE,EAAE;YACb,SAAS;YACT,QAAQ,EAAE,EAAE,OAAO,EAAE,IAAI,CAAC,IAAI,EAAE;SACjC,CAAC;QACF,IAAI,CAAC,QAAQ,CAAC,GAAG,CAAC,OAAO,CAAC,EAAE,EAAE,OAAO,CAAC,CAAC;QAEvC,IAAI,CAAC;YACH,MAAM,MAAM,GAAG;gBACb,MAAM,CAAC,qBAAqB;gBAC5B,kTAAkT;gBAClT,MAAM,CAAC,IAAI,CAAC,WAAW;aACxB;iBACE,MAAM,CAAC,OAAO,CAAC;iBACf,IAAI,CAAC,MAAM,CAAC,CAAC;YAEhB,MAAM,MAAM,GAAG,MAAM,SAAS,CAAC,MAAM,EAAE;gBACrC,KAAK,EAAE,IAAI,CAAC,IAAI,CAAC,KAAK;gBACtB,SAAS,EAAE,MAAM,CAAC,OAAO,IAAI,IAAI,CAAC,IAAI,CAAC,SAAS;gBAChD,QAAQ,EAAE,IAAI,CAAC,IAAI,CAAC,QAAQ;gBAC5B,SAAS,EAAE,IAAI,CAAC,IAAI,CAAC,SAAS;gBAC9B,GAAG,EAAE,MAAM,CAAC,GAAG;gBACf,cAAc,EAAE,aAAa;gBAC7B,YAAY,EAAE,CAAC,MAAM,EAAE,OAAO,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,CAAC;aACxD,CAAC,CAAC;YAEH,MAAM,OAAO,GAAiB;gBAC5B,IAAI,EAAE,WAAW;gBACjB,OAAO,EAAE,MAAM,CAAC,IAAI;gBACpB,SAAS,EAAE,IAAI,IAAI,EAAE;aACtB,CAAC;YACF,OAAO,CAAC,QAAQ,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC;YAC/B,OAAO,CAAC,MAAM,GAAG,MAAM,CAAC,IAAI,CAAC;YAC7B,OAAO,CAAC,KAAK,GAAG,MAAM,CAAC,EAAE,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC,CAAC,QAAQ,CAAC;YACnD,OAAO,CAAC,OAAO,GAAG,IAAI,IAAI,EAAE,CAAC;YAC7B,IAAI,CAAC,MAAM,CAAC,EAAE,EAAE,CAAC;gBACf,OAAO,CAAC,KAAK,GAAG,qCAAqC,CAAC;YACxD,CAAC;QACH,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YACf,OAAO,CAAC,KAAK,GAAG,QAAQ,CAAC;YACzB,OAAO,CAAC,KAAK,GAAG,KAAK,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC;YACvE,OAAO,CAAC,OAAO,GAAG,IAAI,IAAI,EAAE,CAAC;QAC/B,CAAC;QAED,IAAI,CAAC,QAAQ,CAAC,GAAG,CAAC,OAAO,CAAC,EAAE,EAAE,OAAO,CAAC,CAAC;QACvC,OAAO,OAAO,CAAC;IACjB,CAAC;IAED,KAAK,CAAC,UAAU,CAAC,SAAiB;QAChC,OAAO,IAAI,CAAC,QAAQ,CAAC,GAAG,CAAC,SAAS,CAAC,CAAC;IACtC,CAAC;CACF"}
@@ -0,0 +1,64 @@
1
+ import { type Task } from '../../types/index.js';
2
+ import type { EvalTask } from '../types.js';
3
+ import type { ExtendedLiveBenchmark, BenchmarkVerdict } from './live-benchmark.js';
4
+ import type { BenchmarkDelegate } from './live-runner.js';
5
+ import { EnvBridge } from './env-bridge.js';
6
+ export declare class ClBenchBenchmark implements ExtendedLiveBenchmark {
7
+ readonly name = "clbench";
8
+ readonly interactive = true;
9
+ readonly preserveOrder = true;
10
+ readonly playbookMinTrajectories = 3;
11
+ private bridge;
12
+ private readonly maxSteps;
13
+ private readonly reinject;
14
+ private readonly historyWindow;
15
+ constructor(opts?: {
16
+ maxSteps?: number;
17
+ reinject?: boolean;
18
+ historyWindow?: number;
19
+ });
20
+ setup(): Promise<void>;
21
+ teardown(): Promise<void>;
22
+ private getBridge;
23
+ freshBridge(): Promise<EnvBridge>;
24
+ /**
25
+ * Load conversations as tasks. Applies a seeded Fisher-Yates shuffle so
26
+ * different seeds produce different orderings (reproducible per seed).
27
+ */
28
+ loadTasks(opts: {
29
+ count: number;
30
+ seed: number;
31
+ }): Promise<EvalTask[]>;
32
+ verify(_task: Task, _output: string): BenchmarkVerdict;
33
+ getBridgeInstance(): EnvBridge | null;
34
+ delegateFactory: (_benchmark: ExtendedLiveBenchmark, opts?: {
35
+ model?: string;
36
+ timeoutMs?: number;
37
+ provider?: string;
38
+ awsRegion?: string;
39
+ cliRetries?: number;
40
+ }) => BenchmarkDelegate;
41
+ }
42
+ export declare function createClBenchBenchmark(opts?: {
43
+ maxSteps?: number;
44
+ reinject?: boolean;
45
+ historyWindow?: number;
46
+ }): ClBenchBenchmark;
47
+ /**
48
+ * Structured per-step prompt for CL-Bench CRM conversations.
49
+ * Surfaces tool catalog, conversation context, and injected knowledge.
50
+ */
51
+ export declare function buildClBenchPrompt(p: {
52
+ knowledge?: string;
53
+ instruction: string;
54
+ category: string;
55
+ complexity: string;
56
+ history: Array<{
57
+ action: string;
58
+ obs: string;
59
+ }>;
60
+ obs: string;
61
+ availableTools?: string[];
62
+ context?: Record<string, unknown>;
63
+ }): string;
64
+ //# sourceMappingURL=clbench.d.ts.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"clbench.d.ts","sourceRoot":"","sources":["../../../src/eval/live/clbench.ts"],"names":[],"mappings":"AAEA,OAAO,EAML,KAAK,IAAI,EACV,MAAM,sBAAsB,CAAC;AAE9B,OAAO,KAAK,EAAE,QAAQ,EAAE,MAAM,aAAa,CAAC;AAC5C,OAAO,KAAK,EAAE,qBAAqB,EAAE,gBAAgB,EAAE,MAAM,qBAAqB,CAAC;AACnF,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,kBAAkB,CAAC;AAC1D,OAAO,EAAE,SAAS,EAAE,MAAM,iBAAiB,CAAC;AAuE5C,qBAAa,gBAAiB,YAAW,qBAAqB;IAC5D,QAAQ,CAAC,IAAI,aAAa;IAC1B,QAAQ,CAAC,WAAW,QAAQ;IAC5B,QAAQ,CAAC,aAAa,QAAQ;IAC9B,QAAQ,CAAC,uBAAuB,KAAK;IAErC,OAAO,CAAC,MAAM,CAA0B;IACxC,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAS;IAClC,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAU;IACnC,OAAO,CAAC,QAAQ,CAAC,aAAa,CAAS;gBAGrC,IAAI,GAAE;QACJ,QAAQ,CAAC,EAAE,MAAM,CAAC;QAClB,QAAQ,CAAC,EAAE,OAAO,CAAC;QACnB,aAAa,CAAC,EAAE,MAAM,CAAC;KACnB;IAOF,KAAK,IAAI,OAAO,CAAC,IAAI,CAAC;IAItB,QAAQ,IAAI,OAAO,CAAC,IAAI,CAAC;YAKjB,SAAS;IAiBjB,WAAW,IAAI,OAAO,CAAC,SAAS,CAAC;IAMvC;;;OAGG;IACG,SAAS,CAAC,IAAI,EAAE;QAAE,KAAK,EAAE,MAAM,CAAC;QAAC,IAAI,EAAE,MAAM,CAAA;KAAE,GAAG,OAAO,CAAC,QAAQ,EAAE,CAAC;IA4C3E,MAAM,CAAC,KAAK,EAAE,IAAI,EAAE,OAAO,EAAE,MAAM,GAAG,gBAAgB;IAItD,iBAAiB,IAAI,SAAS,GAAG,IAAI;IAIrC,eAAe,GACb,YAAY,qBAAqB,EACjC,OAAO;QACL,KAAK,CAAC,EAAE,MAAM,CAAC;QACf,SAAS,CAAC,EAAE,MAAM,CAAC;QACnB,QAAQ,CAAC,EAAE,MAAM,CAAC;QAClB,SAAS,CAAC,EAAE,MAAM,CAAC;QACnB,UAAU,CAAC,EAAE,MAAM,CAAC;KACrB,KACA,iBAAiB,CAWlB;CACH;AAED,wBAAgB,sBAAsB,CAAC,IAAI,CAAC,EAAE;IAC5C,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,QAAQ,CAAC,EAAE,OAAO,CAAC;IACnB,aAAa,CAAC,EAAE,MAAM,CAAC;CACxB,GAAG,gBAAgB,CAEnB;AAwID;;;GAGG;AACH,wBAAgB,kBAAkB,CAAC,CAAC,EAAE;IACpC,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,WAAW,EAAE,MAAM,CAAC;IACpB,QAAQ,EAAE,MAAM,CAAC;IACjB,UAAU,EAAE,MAAM,CAAC;IACnB,OAAO,EAAE,KAAK,CAAC;QAAE,MAAM,EAAE,MAAM,CAAC;QAAC,GAAG,EAAE,MAAM,CAAA;KAAE,CAAC,CAAC;IAChD,GAAG,EAAE,MAAM,CAAC;IACZ,cAAc,CAAC,EAAE,MAAM,EAAE,CAAC;IAC1B,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;CACnC,GAAG,MAAM,CAkCT"}