@tangle-network/agent-eval 0.94.0 → 0.95.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (1056) hide show
  1. package/CHANGELOG.md +15 -0
  2. package/README.md +44 -30
  3. package/dist/action-policy.d.ts +24 -0
  4. package/dist/action-policy.d.ts.map +1 -0
  5. package/dist/action-policy.test.d.ts +2 -0
  6. package/dist/action-policy.test.d.ts.map +1 -0
  7. package/dist/active-learning.d.ts +41 -0
  8. package/dist/active-learning.d.ts.map +1 -0
  9. package/dist/adapters/http.d.ts +15 -21
  10. package/dist/adapters/http.d.ts.map +1 -0
  11. package/dist/adapters/http.js.map +1 -1
  12. package/dist/adapters/langchain.d.ts +7 -13
  13. package/dist/adapters/langchain.d.ts.map +1 -0
  14. package/dist/adapters/otel.d.ts +13 -24
  15. package/dist/adapters/otel.d.ts.map +1 -0
  16. package/dist/agent-profile-cell.d.ts +101 -0
  17. package/dist/agent-profile-cell.d.ts.map +1 -0
  18. package/dist/agent-profile.d.ts +27 -0
  19. package/dist/agent-profile.d.ts.map +1 -0
  20. package/dist/agent-profile.test.d.ts +2 -0
  21. package/dist/agent-profile.test.d.ts.map +1 -0
  22. package/dist/analyst/adapters.d.ts +62 -0
  23. package/dist/analyst/adapters.d.ts.map +1 -0
  24. package/dist/analyst/analyst.test.d.ts +2 -0
  25. package/dist/analyst/analyst.test.d.ts.map +1 -0
  26. package/dist/analyst/ax-service.d.ts +27 -0
  27. package/dist/analyst/ax-service.d.ts.map +1 -0
  28. package/dist/analyst/behavioral-analyst.d.ts +28 -0
  29. package/dist/analyst/behavioral-analyst.d.ts.map +1 -0
  30. package/dist/analyst/chat-client.d.ts +91 -0
  31. package/dist/analyst/chat-client.d.ts.map +1 -0
  32. package/dist/analyst/default-registry.d.ts +27 -0
  33. package/dist/analyst/default-registry.d.ts.map +1 -0
  34. package/dist/analyst/default-registry.test.d.ts +2 -0
  35. package/dist/analyst/default-registry.test.d.ts.map +1 -0
  36. package/dist/analyst/finding-signature.d.ts +48 -0
  37. package/dist/analyst/finding-signature.d.ts.map +1 -0
  38. package/dist/analyst/finding-subject.d.ts +146 -0
  39. package/dist/analyst/finding-subject.d.ts.map +1 -0
  40. package/dist/analyst/finding-subject.test.d.ts +2 -0
  41. package/dist/analyst/finding-subject.test.d.ts.map +1 -0
  42. package/dist/analyst/findings-store.d.ts +75 -0
  43. package/dist/analyst/findings-store.d.ts.map +1 -0
  44. package/dist/analyst/index.d.ts +28 -256
  45. package/dist/analyst/index.d.ts.map +1 -0
  46. package/dist/analyst/index.js +5 -65
  47. package/dist/analyst/index.js.map +1 -1
  48. package/dist/{kind-factory-0BhLSI27.d.ts → analyst/kind-factory.d.ts} +11 -63
  49. package/dist/analyst/kind-factory.d.ts.map +1 -0
  50. package/dist/analyst/kinds/failure-mode.d.ts +19 -0
  51. package/dist/analyst/kinds/failure-mode.d.ts.map +1 -0
  52. package/dist/analyst/kinds/improvement.d.ts +23 -0
  53. package/dist/analyst/kinds/improvement.d.ts.map +1 -0
  54. package/dist/analyst/kinds/index.d.ts +22 -0
  55. package/dist/analyst/kinds/index.d.ts.map +1 -0
  56. package/dist/analyst/kinds/kinds.test.d.ts +2 -0
  57. package/dist/analyst/kinds/kinds.test.d.ts.map +1 -0
  58. package/dist/analyst/kinds/knowledge-gap.d.ts +28 -0
  59. package/dist/analyst/kinds/knowledge-gap.d.ts.map +1 -0
  60. package/dist/analyst/kinds/knowledge-poisoning.d.ts +22 -0
  61. package/dist/analyst/kinds/knowledge-poisoning.d.ts.map +1 -0
  62. package/dist/analyst/kinds/skill-usage.d.ts +84 -0
  63. package/dist/analyst/kinds/skill-usage.d.ts.map +1 -0
  64. package/dist/analyst/kinds/skill-usage.test.d.ts +2 -0
  65. package/dist/analyst/kinds/skill-usage.test.d.ts.map +1 -0
  66. package/dist/analyst/parse-tolerant.d.ts +26 -0
  67. package/dist/analyst/parse-tolerant.d.ts.map +1 -0
  68. package/dist/analyst/parse-tolerant.test.d.ts +2 -0
  69. package/dist/analyst/parse-tolerant.test.d.ts.map +1 -0
  70. package/dist/analyst/registry.budget.test.d.ts +2 -0
  71. package/dist/analyst/registry.budget.test.d.ts.map +1 -0
  72. package/dist/{default-registry-6dhErQbs.d.ts → analyst/registry.d.ts} +8 -37
  73. package/dist/analyst/registry.d.ts.map +1 -0
  74. package/dist/analyst/steer-firewall.d.ts +35 -0
  75. package/dist/analyst/steer-firewall.d.ts.map +1 -0
  76. package/dist/analyst/steer-firewall.test.d.ts +2 -0
  77. package/dist/analyst/steer-firewall.test.d.ts.map +1 -0
  78. package/dist/analyst/structure-findings.d.ts +37 -0
  79. package/dist/analyst/structure-findings.d.ts.map +1 -0
  80. package/dist/analyst/structure-findings.test.d.ts +2 -0
  81. package/dist/analyst/structure-findings.test.d.ts.map +1 -0
  82. package/dist/analyst/tool-groups.d.ts +34 -0
  83. package/dist/analyst/tool-groups.d.ts.map +1 -0
  84. package/dist/{types-Ce17tDlG.d.ts → analyst/types.d.ts} +19 -112
  85. package/dist/analyst/types.d.ts.map +1 -0
  86. package/dist/anti-slop.d.ts +59 -0
  87. package/dist/anti-slop.d.ts.map +1 -0
  88. package/dist/artifact-validator.d.ts +74 -0
  89. package/dist/artifact-validator.d.ts.map +1 -0
  90. package/dist/attestation.d.ts +63 -0
  91. package/dist/attestation.d.ts.map +1 -0
  92. package/dist/attestation.test.d.ts +2 -0
  93. package/dist/attestation.test.d.ts.map +1 -0
  94. package/dist/authenticity/index.d.ts +15 -16
  95. package/dist/authenticity/index.d.ts.map +1 -0
  96. package/dist/authenticity/index.test.d.ts +2 -0
  97. package/dist/authenticity/index.test.d.ts.map +1 -0
  98. package/dist/auto-pr.d.ts +120 -0
  99. package/dist/auto-pr.d.ts.map +1 -0
  100. package/dist/{baseline-Bbid3WoO.d.ts → baseline.d.ts} +8 -44
  101. package/dist/baseline.d.ts.map +1 -0
  102. package/dist/behavior-dsl.d.ts +73 -0
  103. package/dist/behavior-dsl.d.ts.map +1 -0
  104. package/dist/belief-state/calibration.d.ts +10 -0
  105. package/dist/belief-state/calibration.d.ts.map +1 -0
  106. package/dist/belief-state/calibration.test.d.ts +2 -0
  107. package/dist/belief-state/calibration.test.d.ts.map +1 -0
  108. package/dist/belief-state/code-agent-corpus.d.ts +66 -0
  109. package/dist/belief-state/code-agent-corpus.d.ts.map +1 -0
  110. package/dist/belief-state/code-agent-corpus.test.d.ts +2 -0
  111. package/dist/belief-state/code-agent-corpus.test.d.ts.map +1 -0
  112. package/dist/belief-state/code-agent-evidence.d.ts +22 -0
  113. package/dist/belief-state/code-agent-evidence.d.ts.map +1 -0
  114. package/dist/belief-state/code-agent-evidence.test.d.ts +2 -0
  115. package/dist/belief-state/code-agent-evidence.test.d.ts.map +1 -0
  116. package/dist/belief-state/extract.d.ts +7 -0
  117. package/dist/belief-state/extract.d.ts.map +1 -0
  118. package/dist/belief-state/extract.test.d.ts +2 -0
  119. package/dist/belief-state/extract.test.d.ts.map +1 -0
  120. package/dist/belief-state/index.d.ts +14 -604
  121. package/dist/belief-state/index.d.ts.map +1 -0
  122. package/dist/belief-state/ope.d.ts +17 -0
  123. package/dist/belief-state/ope.d.ts.map +1 -0
  124. package/dist/belief-state/ope.test.d.ts +2 -0
  125. package/dist/belief-state/ope.test.d.ts.map +1 -0
  126. package/dist/belief-state/phase0-measurement.d.ts +55 -0
  127. package/dist/belief-state/phase0-measurement.d.ts.map +1 -0
  128. package/dist/belief-state/report.d.ts +17 -0
  129. package/dist/belief-state/report.d.ts.map +1 -0
  130. package/dist/belief-state/report.test.d.ts +2 -0
  131. package/dist/belief-state/report.test.d.ts.map +1 -0
  132. package/dist/belief-state/research-evidence.d.ts +23 -0
  133. package/dist/belief-state/research-evidence.d.ts.map +1 -0
  134. package/dist/belief-state/research-evidence.test.d.ts +2 -0
  135. package/dist/belief-state/research-evidence.test.d.ts.map +1 -0
  136. package/dist/belief-state/runtime-benchmark-corpus.d.ts +32 -0
  137. package/dist/belief-state/runtime-benchmark-corpus.d.ts.map +1 -0
  138. package/dist/belief-state/runtime-hooks.d.ts +87 -0
  139. package/dist/belief-state/runtime-hooks.d.ts.map +1 -0
  140. package/dist/belief-state/runtime-hooks.test.d.ts +2 -0
  141. package/dist/belief-state/runtime-hooks.test.d.ts.map +1 -0
  142. package/dist/belief-state/selective.d.ts +15 -0
  143. package/dist/belief-state/selective.d.ts.map +1 -0
  144. package/dist/belief-state/selective.test.d.ts +2 -0
  145. package/dist/belief-state/selective.test.d.ts.map +1 -0
  146. package/dist/belief-state/shadow-probe.d.ts +80 -0
  147. package/dist/belief-state/shadow-probe.d.ts.map +1 -0
  148. package/dist/belief-state/shadow-probe.test.d.ts +2 -0
  149. package/dist/belief-state/shadow-probe.test.d.ts.map +1 -0
  150. package/dist/belief-state/types.d.ts +195 -0
  151. package/dist/belief-state/types.d.ts.map +1 -0
  152. package/dist/belief-state/types.test.d.ts +2 -0
  153. package/dist/belief-state/types.test.d.ts.map +1 -0
  154. package/dist/benchmark.d.ts +14 -0
  155. package/dist/benchmark.d.ts.map +1 -0
  156. package/dist/benchmarks/index.d.ts +23 -4
  157. package/dist/benchmarks/index.d.ts.map +1 -0
  158. package/dist/benchmarks/routing/dataset.d.ts +34 -0
  159. package/dist/benchmarks/routing/dataset.d.ts.map +1 -0
  160. package/dist/benchmarks/routing/index.d.ts +34 -0
  161. package/dist/benchmarks/routing/index.d.ts.map +1 -0
  162. package/dist/benchmarks/types.d.ts +49 -0
  163. package/dist/benchmarks/types.d.ts.map +1 -0
  164. package/dist/bisector.d.ts +81 -0
  165. package/dist/bisector.d.ts.map +1 -0
  166. package/dist/budget-guard.d.ts +31 -0
  167. package/dist/budget-guard.d.ts.map +1 -0
  168. package/dist/builder-eval/builder-session.d.ts +111 -0
  169. package/dist/builder-eval/builder-session.d.ts.map +1 -0
  170. package/dist/builder-eval/correlation.d.ts +32 -0
  171. package/dist/builder-eval/correlation.d.ts.map +1 -0
  172. package/dist/builder-eval/index.d.ts +5 -250
  173. package/dist/builder-eval/index.d.ts.map +1 -0
  174. package/dist/builder-eval/project-registry.d.ts +51 -0
  175. package/dist/builder-eval/project-registry.d.ts.map +1 -0
  176. package/dist/builder-eval/three-layer-eval.d.ts +55 -0
  177. package/dist/builder-eval/three-layer-eval.d.ts.map +1 -0
  178. package/dist/campaign/analyst-surface.d.ts +108 -0
  179. package/dist/campaign/analyst-surface.d.ts.map +1 -0
  180. package/dist/campaign/analyst-surface.test.d.ts +2 -0
  181. package/dist/campaign/analyst-surface.test.d.ts.map +1 -0
  182. package/dist/campaign/auto-pr.d.ts +46 -0
  183. package/dist/campaign/auto-pr.d.ts.map +1 -0
  184. package/dist/campaign/distillation/agreement-judge.d.ts +69 -0
  185. package/dist/campaign/distillation/agreement-judge.d.ts.map +1 -0
  186. package/dist/campaign/distillation/cli.d.ts +35 -0
  187. package/dist/campaign/distillation/cli.d.ts.map +1 -0
  188. package/dist/campaign/distillation/distillation.test.d.ts +2 -0
  189. package/dist/campaign/distillation/distillation.test.d.ts.map +1 -0
  190. package/dist/campaign/distillation/gold-scenarios.d.ts +54 -0
  191. package/dist/campaign/distillation/gold-scenarios.d.ts.map +1 -0
  192. package/dist/campaign/distillation/run-distillation.d.ts +119 -0
  193. package/dist/campaign/distillation/run-distillation.d.ts.map +1 -0
  194. package/dist/campaign/gates/compose.d.ts +12 -0
  195. package/dist/campaign/gates/compose.d.ts.map +1 -0
  196. package/dist/campaign/gates/default-production-gate.d.ts +58 -0
  197. package/dist/campaign/gates/default-production-gate.d.ts.map +1 -0
  198. package/dist/campaign/gates/heldout-gate.d.ts +12 -0
  199. package/dist/campaign/gates/heldout-gate.d.ts.map +1 -0
  200. package/dist/campaign/gates/promotion-policy.d.ts +125 -0
  201. package/dist/campaign/gates/promotion-policy.d.ts.map +1 -0
  202. package/dist/campaign/gates/promotion-policy.test.d.ts +2 -0
  203. package/dist/campaign/gates/promotion-policy.test.d.ts.map +1 -0
  204. package/dist/campaign/gates/sequential.d.ts +146 -0
  205. package/dist/campaign/gates/sequential.d.ts.map +1 -0
  206. package/dist/campaign/gates/sequential.test.d.ts +2 -0
  207. package/dist/campaign/gates/sequential.test.d.ts.map +1 -0
  208. package/dist/campaign/gates/statistical-heldout.d.ts +99 -0
  209. package/dist/campaign/gates/statistical-heldout.d.ts.map +1 -0
  210. package/dist/campaign/gates/statistical-heldout.test.d.ts +2 -0
  211. package/dist/campaign/gates/statistical-heldout.test.d.ts.map +1 -0
  212. package/dist/campaign/index.d.ts +38 -1341
  213. package/dist/campaign/index.d.ts.map +1 -0
  214. package/dist/campaign/index.js +1863 -1316
  215. package/dist/campaign/index.js.map +1 -1
  216. package/dist/campaign/labeled-store/fs-adapter.d.ts +59 -0
  217. package/dist/campaign/labeled-store/fs-adapter.d.ts.map +1 -0
  218. package/dist/campaign/presets/compare-proposers.d.ts +146 -0
  219. package/dist/campaign/presets/compare-proposers.d.ts.map +1 -0
  220. package/dist/campaign/presets/playback.d.ts +120 -0
  221. package/dist/campaign/presets/playback.d.ts.map +1 -0
  222. package/dist/campaign/presets/playback.test.d.ts +2 -0
  223. package/dist/campaign/presets/playback.test.d.ts.map +1 -0
  224. package/dist/campaign/presets/run-eval.d.ts +14 -0
  225. package/dist/campaign/presets/run-eval.d.ts.map +1 -0
  226. package/dist/campaign/presets/run-improvement-loop.d.ts +63 -0
  227. package/dist/campaign/presets/run-improvement-loop.d.ts.map +1 -0
  228. package/dist/campaign/presets/run-improvement-loop.test.d.ts +2 -0
  229. package/dist/campaign/presets/run-improvement-loop.test.d.ts.map +1 -0
  230. package/dist/campaign/presets/run-optimization.d.ts +92 -0
  231. package/dist/campaign/presets/run-optimization.d.ts.map +1 -0
  232. package/dist/campaign/presets/run-profile-matrix.d.ts +151 -0
  233. package/dist/campaign/presets/run-profile-matrix.d.ts.map +1 -0
  234. package/dist/campaign/presets/run-skill-opt.d.ts +96 -0
  235. package/dist/campaign/presets/run-skill-opt.d.ts.map +1 -0
  236. package/dist/campaign/proposers/_findings-text.d.ts +22 -0
  237. package/dist/campaign/proposers/_findings-text.d.ts.map +1 -0
  238. package/dist/campaign/proposers/ace.d.ts +33 -0
  239. package/dist/campaign/proposers/ace.d.ts.map +1 -0
  240. package/dist/campaign/proposers/ace.test.d.ts +2 -0
  241. package/dist/campaign/proposers/ace.test.d.ts.map +1 -0
  242. package/dist/campaign/proposers/analysis-edit.d.ts +32 -0
  243. package/dist/campaign/proposers/analysis-edit.d.ts.map +1 -0
  244. package/dist/campaign/proposers/evolutionary.d.ts +20 -0
  245. package/dist/campaign/proposers/evolutionary.d.ts.map +1 -0
  246. package/dist/campaign/proposers/fapo.d.ts +120 -0
  247. package/dist/campaign/proposers/fapo.d.ts.map +1 -0
  248. package/dist/campaign/proposers/gepa.d.ts +86 -0
  249. package/dist/campaign/proposers/gepa.d.ts.map +1 -0
  250. package/dist/campaign/proposers/halo.d.ts +44 -0
  251. package/dist/campaign/proposers/halo.d.ts.map +1 -0
  252. package/dist/campaign/proposers/halo.test.d.ts +2 -0
  253. package/dist/campaign/proposers/halo.test.d.ts.map +1 -0
  254. package/dist/campaign/proposers/memory.d.ts +47 -0
  255. package/dist/campaign/proposers/memory.d.ts.map +1 -0
  256. package/dist/campaign/proposers/memory.test.d.ts +2 -0
  257. package/dist/campaign/proposers/memory.test.d.ts.map +1 -0
  258. package/dist/campaign/proposers/skill-opt.d.ts +88 -0
  259. package/dist/campaign/proposers/skill-opt.d.ts.map +1 -0
  260. package/dist/campaign/proposers/trace-analyst.d.ts +48 -0
  261. package/dist/campaign/proposers/trace-analyst.d.ts.map +1 -0
  262. package/dist/campaign/proposers/trace-analyst.test.d.ts +2 -0
  263. package/dist/campaign/proposers/trace-analyst.test.d.ts.map +1 -0
  264. package/dist/campaign/provenance.d.ts +185 -0
  265. package/dist/campaign/provenance.d.ts.map +1 -0
  266. package/dist/campaign/run-campaign.d.ts +90 -0
  267. package/dist/campaign/run-campaign.d.ts.map +1 -0
  268. package/dist/campaign/score-utils.d.ts +26 -0
  269. package/dist/campaign/score-utils.d.ts.map +1 -0
  270. package/dist/campaign/skill-patch.d.ts +62 -0
  271. package/dist/campaign/skill-patch.d.ts.map +1 -0
  272. package/dist/campaign/storage.d.ts +38 -0
  273. package/dist/campaign/storage.d.ts.map +1 -0
  274. package/dist/{types-BU-7W85F.d.ts → campaign/types.d.ts} +98 -99
  275. package/dist/campaign/types.d.ts.map +1 -0
  276. package/dist/campaign/worktree/index.d.ts +53 -0
  277. package/dist/campaign/worktree/index.d.ts.map +1 -0
  278. package/dist/canary.d.ts +101 -0
  279. package/dist/canary.d.ts.map +1 -0
  280. package/dist/causal-attribution.d.ts +45 -0
  281. package/dist/causal-attribution.d.ts.map +1 -0
  282. package/dist/{chunk-2K6UUZ7P.js → chunk-2T4EZACH.js} +1 -1
  283. package/dist/chunk-2T4EZACH.js.map +1 -0
  284. package/dist/{chunk-CTBHKLEU.js → chunk-77T4STFI.js} +59 -86
  285. package/dist/chunk-77T4STFI.js.map +1 -0
  286. package/dist/{chunk-EGPMSBEZ.js → chunk-7QTQKIDD.js} +178 -177
  287. package/dist/chunk-7QTQKIDD.js.map +1 -0
  288. package/dist/{chunk-MIFZUPEK.js → chunk-AQ5WQAIV.js} +21 -6
  289. package/dist/chunk-AQ5WQAIV.js.map +1 -0
  290. package/dist/chunk-DJWX3GVS.js +81 -0
  291. package/dist/chunk-DJWX3GVS.js.map +1 -0
  292. package/dist/{chunk-S6OZEZQK.js → chunk-HMA63UEO.js} +37 -9
  293. package/dist/{chunk-S6OZEZQK.js.map → chunk-HMA63UEO.js.map} +1 -1
  294. package/dist/{chunk-TBDR6PAI.js → chunk-IZCEK2HR.js} +2 -2
  295. package/dist/{chunk-SD2YFWQQ.js → chunk-KKWJD5E6.js} +20 -20
  296. package/dist/chunk-KKWJD5E6.js.map +1 -0
  297. package/dist/{chunk-KWRRMR3J.js → chunk-LO6IOIJ2.js} +10 -10
  298. package/dist/chunk-LO6IOIJ2.js.map +1 -0
  299. package/dist/{chunk-E4GH6USR.js → chunk-NZEQVRH5.js} +2 -2
  300. package/dist/chunk-NZEQVRH5.js.map +1 -0
  301. package/dist/{chunk-MPQWFX6Y.js → chunk-PSWWQXHF.js} +13 -88
  302. package/dist/chunk-PSWWQXHF.js.map +1 -0
  303. package/dist/{chunk-Q5LIB7BC.js → chunk-S4SYLDFX.js} +2 -2
  304. package/dist/chunk-S4SYLDFX.js.map +1 -0
  305. package/dist/{chunk-KW53MSA5.js → chunk-X74V6ESX.js} +2 -2
  306. package/dist/{chunk-QMUEXQJS.js → chunk-YBIGNSCZ.js} +81 -4
  307. package/dist/chunk-YBIGNSCZ.js.map +1 -0
  308. package/dist/{chunk-2KNZHH3P.js → chunk-Z6L6YSU6.js} +2 -2
  309. package/dist/ci-gate.d.ts +44 -0
  310. package/dist/ci-gate.d.ts.map +1 -0
  311. package/dist/cli.d.ts +2 -0
  312. package/dist/cli.d.ts.map +1 -0
  313. package/dist/client.d.ts +77 -0
  314. package/dist/client.d.ts.map +1 -0
  315. package/dist/client.test.d.ts +2 -0
  316. package/dist/client.test.d.ts.map +1 -0
  317. package/dist/command-runner.d.ts +74 -0
  318. package/dist/command-runner.d.ts.map +1 -0
  319. package/dist/command-runner.test.d.ts +2 -0
  320. package/dist/command-runner.test.d.ts.map +1 -0
  321. package/dist/completion-verifier.d.ts +147 -0
  322. package/dist/completion-verifier.d.ts.map +1 -0
  323. package/dist/completion-verifier.test.d.ts +9 -0
  324. package/dist/completion-verifier.test.d.ts.map +1 -0
  325. package/dist/concurrency.d.ts +23 -0
  326. package/dist/concurrency.d.ts.map +1 -0
  327. package/dist/contamination-guard.d.ts +81 -0
  328. package/dist/contamination-guard.d.ts.map +1 -0
  329. package/dist/{analyze-runs-B6Ljo_dI.d.ts → contract/analyze-runs.d.ts} +9 -10
  330. package/dist/contract/analyze-runs.d.ts.map +1 -0
  331. package/dist/contract/define-agent-eval.d.ts +52 -0
  332. package/dist/contract/define-agent-eval.d.ts.map +1 -0
  333. package/dist/contract/diff.d.ts +114 -0
  334. package/dist/contract/diff.d.ts.map +1 -0
  335. package/dist/contract/index.d.ts +106 -640
  336. package/dist/contract/index.d.ts.map +1 -0
  337. package/dist/contract/index.js +127 -17
  338. package/dist/contract/index.js.map +1 -1
  339. package/dist/{insight-report-DWl3z9tl.d.ts → contract/insight-report.d.ts} +16 -19
  340. package/dist/contract/insight-report.d.ts.map +1 -0
  341. package/dist/contract/insight-types-fwd.d.ts +7 -0
  342. package/dist/contract/insight-types-fwd.d.ts.map +1 -0
  343. package/dist/contract/intake/agent-trace.d.ts +97 -0
  344. package/dist/contract/intake/agent-trace.d.ts.map +1 -0
  345. package/dist/{code-agent-session-BO8nCnv3.d.ts → contract/intake/code-agent-session.d.ts} +15 -17
  346. package/dist/contract/intake/code-agent-session.d.ts.map +1 -0
  347. package/dist/contract/intake/feedback-table.d.ts +87 -0
  348. package/dist/contract/intake/feedback-table.d.ts.map +1 -0
  349. package/dist/contract/intake/index.d.ts +22 -0
  350. package/dist/contract/intake/index.d.ts.map +1 -0
  351. package/dist/contract/intake/otel-spans.d.ts +33 -0
  352. package/dist/contract/intake/otel-spans.d.ts.map +1 -0
  353. package/dist/contract/self-improve.d.ts +284 -0
  354. package/dist/contract/self-improve.d.ts.map +1 -0
  355. package/dist/{control-runtime-Acf9CGhw.d.ts → control-runtime.d.ts} +23 -26
  356. package/dist/control-runtime.d.ts.map +1 -0
  357. package/dist/control-runtime.test.d.ts +2 -0
  358. package/dist/control-runtime.test.d.ts.map +1 -0
  359. package/dist/control.d.ts +11 -9
  360. package/dist/control.d.ts.map +1 -0
  361. package/dist/control.js +2 -2
  362. package/dist/convergence.d.ts +29 -0
  363. package/dist/convergence.d.ts.map +1 -0
  364. package/dist/{cost-ledger-DuSqlw5B.d.ts → cost-ledger.d.ts} +10 -11
  365. package/dist/cost-ledger.d.ts.map +1 -0
  366. package/dist/cost-ledger.test.d.ts +2 -0
  367. package/dist/cost-ledger.test.d.ts.map +1 -0
  368. package/dist/cost-report.d.ts +43 -0
  369. package/dist/cost-report.d.ts.map +1 -0
  370. package/dist/cost-report.test.d.ts +2 -0
  371. package/dist/cost-report.test.d.ts.map +1 -0
  372. package/dist/cost-tracker.d.ts +76 -0
  373. package/dist/cost-tracker.d.ts.map +1 -0
  374. package/dist/{counterfactual-DlOz8PBx.d.ts → counterfactual.d.ts} +12 -13
  375. package/dist/counterfactual.d.ts.map +1 -0
  376. package/dist/cross-trace-diff.d.ts +56 -0
  377. package/dist/cross-trace-diff.d.ts.map +1 -0
  378. package/dist/{dataset-BbGkaN2I.d.ts → dataset.d.ts} +11 -14
  379. package/dist/dataset.d.ts.map +1 -0
  380. package/dist/deploy-gate-layer.d.ts +125 -0
  381. package/dist/deploy-gate-layer.d.ts.map +1 -0
  382. package/dist/deploy-gate-layer.test.d.ts +2 -0
  383. package/dist/deploy-gate-layer.test.d.ts.map +1 -0
  384. package/dist/description-length-gate.d.ts +119 -0
  385. package/dist/description-length-gate.d.ts.map +1 -0
  386. package/dist/detectors/edge.test.d.ts +2 -0
  387. package/dist/detectors/edge.test.d.ts.map +1 -0
  388. package/dist/detectors/index.d.ts +81 -0
  389. package/dist/detectors/index.d.ts.map +1 -0
  390. package/dist/detectors/index.test.d.ts +2 -0
  391. package/dist/detectors/index.test.d.ts.map +1 -0
  392. package/dist/diagnose/causal-sweep.d.ts +100 -0
  393. package/dist/diagnose/causal-sweep.d.ts.map +1 -0
  394. package/dist/diagnose/index.d.ts +36 -0
  395. package/dist/diagnose/index.d.ts.map +1 -0
  396. package/dist/diagnose/remediation.d.ts +68 -0
  397. package/dist/diagnose/remediation.d.ts.map +1 -0
  398. package/dist/diagnose/repair.d.ts +77 -0
  399. package/dist/diagnose/repair.d.ts.map +1 -0
  400. package/dist/diagnose.d.ts +1 -251
  401. package/dist/diagnose.js +1 -1
  402. package/dist/discover-personas.d.ts +35 -0
  403. package/dist/discover-personas.d.ts.map +1 -0
  404. package/dist/driver.d.ts +95 -0
  405. package/dist/driver.d.ts.map +1 -0
  406. package/dist/driver.test.d.ts +8 -0
  407. package/dist/driver.test.d.ts.map +1 -0
  408. package/dist/dual-agent-bench.d.ts +81 -0
  409. package/dist/dual-agent-bench.d.ts.map +1 -0
  410. package/dist/error-count-extractor.d.ts +47 -0
  411. package/dist/error-count-extractor.d.ts.map +1 -0
  412. package/dist/error-count-extractor.test.d.ts +2 -0
  413. package/dist/error-count-extractor.test.d.ts.map +1 -0
  414. package/dist/{errors-CzMUYo7b.d.ts → errors.d.ts} +10 -11
  415. package/dist/errors.d.ts.map +1 -0
  416. package/dist/{researcher-B0C2_fVO.d.ts → eval-campaign.d.ts} +34 -156
  417. package/dist/eval-campaign.d.ts.map +1 -0
  418. package/dist/eval-campaign.test.d.ts +2 -0
  419. package/dist/eval-campaign.test.d.ts.map +1 -0
  420. package/dist/eval-tools.d.ts +55 -0
  421. package/dist/eval-tools.d.ts.map +1 -0
  422. package/dist/eval-trace-store.d.ts +107 -0
  423. package/dist/eval-trace-store.d.ts.map +1 -0
  424. package/dist/eval-trace-store.test.d.ts +2 -0
  425. package/dist/eval-trace-store.test.d.ts.map +1 -0
  426. package/dist/executor.d.ts +38 -0
  427. package/dist/executor.d.ts.map +1 -0
  428. package/dist/executor.test.d.ts +10 -0
  429. package/dist/executor.test.d.ts.map +1 -0
  430. package/dist/experiment-tracker.d.ts +178 -0
  431. package/dist/experiment-tracker.d.ts.map +1 -0
  432. package/dist/experiment-tracker.test.d.ts +2 -0
  433. package/dist/experiment-tracker.test.d.ts.map +1 -0
  434. package/dist/failure-taxonomy.d.ts +38 -0
  435. package/dist/failure-taxonomy.d.ts.map +1 -0
  436. package/dist/{feedback-trajectory-BxY0cKfs.d.ts → feedback-trajectory.d.ts} +36 -38
  437. package/dist/feedback-trajectory.d.ts.map +1 -0
  438. package/dist/feedback-trajectory.test.d.ts +2 -0
  439. package/dist/feedback-trajectory.test.d.ts.map +1 -0
  440. package/dist/flow-layer.d.ts +90 -0
  441. package/dist/flow-layer.d.ts.map +1 -0
  442. package/dist/flow-layer.test.d.ts +2 -0
  443. package/dist/flow-layer.test.d.ts.map +1 -0
  444. package/dist/fuzz/capsule.d.ts +46 -0
  445. package/dist/fuzz/capsule.d.ts.map +1 -0
  446. package/dist/fuzz/cube.d.ts +36 -0
  447. package/dist/fuzz/cube.d.ts.map +1 -0
  448. package/dist/fuzz/explorer-cost.test.d.ts +2 -0
  449. package/dist/fuzz/explorer-cost.test.d.ts.map +1 -0
  450. package/dist/fuzz/explorer.d.ts +64 -0
  451. package/dist/fuzz/explorer.d.ts.map +1 -0
  452. package/dist/fuzz/fuzz-agent.d.ts +16 -0
  453. package/dist/fuzz/fuzz-agent.d.ts.map +1 -0
  454. package/dist/fuzz/fuzz-agent.test.d.ts +2 -0
  455. package/dist/fuzz/fuzz-agent.test.d.ts.map +1 -0
  456. package/dist/fuzz/gates.d.ts +33 -0
  457. package/dist/fuzz/gates.d.ts.map +1 -0
  458. package/dist/fuzz/index.d.ts +26 -0
  459. package/dist/fuzz/index.d.ts.map +1 -0
  460. package/dist/fuzz/policies.d.ts +28 -0
  461. package/dist/fuzz/policies.d.ts.map +1 -0
  462. package/dist/fuzz/tools.d.ts +20 -0
  463. package/dist/fuzz/tools.d.ts.map +1 -0
  464. package/dist/fuzz/types.d.ts +307 -0
  465. package/dist/fuzz/types.d.ts.map +1 -0
  466. package/dist/fuzz.d.ts +1 -547
  467. package/dist/golden-matcher.d.ts +71 -0
  468. package/dist/golden-matcher.d.ts.map +1 -0
  469. package/dist/governance/eu-ai-act.d.ts +37 -0
  470. package/dist/governance/eu-ai-act.d.ts.map +1 -0
  471. package/dist/governance/index.d.ts +5 -135
  472. package/dist/governance/index.d.ts.map +1 -0
  473. package/dist/governance/nist-ai-rmf.d.ts +15 -0
  474. package/dist/governance/nist-ai-rmf.d.ts.map +1 -0
  475. package/dist/governance/soc2.d.ts +12 -0
  476. package/dist/governance/soc2.d.ts.map +1 -0
  477. package/dist/governance/types.d.ts +66 -0
  478. package/dist/governance/types.d.ts.map +1 -0
  479. package/dist/harness-optimizer.d.ts +82 -0
  480. package/dist/harness-optimizer.d.ts.map +1 -0
  481. package/dist/held-out-gate.d.ts +135 -0
  482. package/dist/held-out-gate.d.ts.map +1 -0
  483. package/dist/hosted/client.d.ts +73 -0
  484. package/dist/hosted/client.d.ts.map +1 -0
  485. package/dist/hosted/from-env.test.d.ts +8 -0
  486. package/dist/hosted/from-env.test.d.ts.map +1 -0
  487. package/dist/hosted/index.d.ts +10 -238
  488. package/dist/hosted/index.d.ts.map +1 -0
  489. package/dist/hosted/types.d.ts +159 -0
  490. package/dist/hosted/types.d.ts.map +1 -0
  491. package/dist/index.d.ts +277 -5665
  492. package/dist/index.d.ts.map +1 -0
  493. package/dist/index.js +66 -31
  494. package/dist/index.js.map +1 -1
  495. package/dist/integrity/backend-integrity.d.ts +71 -0
  496. package/dist/integrity/backend-integrity.d.ts.map +1 -0
  497. package/dist/integrity/preflight.d.ts +72 -0
  498. package/dist/integrity/preflight.d.ts.map +1 -0
  499. package/dist/integrity/preflight.test.d.ts +2 -0
  500. package/dist/integrity/preflight.test.d.ts.map +1 -0
  501. package/dist/integrity/single-backend.d.ts +67 -0
  502. package/dist/integrity/single-backend.d.ts.map +1 -0
  503. package/dist/intent-match-judge.d.ts +69 -0
  504. package/dist/intent-match-judge.d.ts.map +1 -0
  505. package/dist/intent-match-judge.test.d.ts +2 -0
  506. package/dist/intent-match-judge.test.d.ts.map +1 -0
  507. package/dist/{judge-calibration-0p2QcWNE.d.ts → judge-calibration.d.ts} +16 -17
  508. package/dist/judge-calibration.d.ts.map +1 -0
  509. package/dist/judge-ensemble.d.ts +66 -0
  510. package/dist/judge-ensemble.d.ts.map +1 -0
  511. package/dist/judge-ensemble.test.d.ts +8 -0
  512. package/dist/judge-ensemble.test.d.ts.map +1 -0
  513. package/dist/judge-families.d.ts +38 -0
  514. package/dist/judge-families.d.ts.map +1 -0
  515. package/dist/judge-panel.d.ts +65 -0
  516. package/dist/judge-panel.d.ts.map +1 -0
  517. package/dist/judge-retry.d.ts +70 -0
  518. package/dist/judge-retry.d.ts.map +1 -0
  519. package/dist/judge-runner.d.ts +36 -0
  520. package/dist/judge-runner.d.ts.map +1 -0
  521. package/dist/judge-runner.test.d.ts +2 -0
  522. package/dist/judge-runner.test.d.ts.map +1 -0
  523. package/dist/judges.d.ts +74 -0
  524. package/dist/judges.d.ts.map +1 -0
  525. package/dist/keyword-coverage-judge.d.ts +89 -0
  526. package/dist/keyword-coverage-judge.d.ts.map +1 -0
  527. package/dist/keyword-coverage-judge.test.d.ts +2 -0
  528. package/dist/keyword-coverage-judge.test.d.ts.map +1 -0
  529. package/dist/knowledge/index.d.ts +3 -103
  530. package/dist/knowledge/index.d.ts.map +1 -0
  531. package/dist/knowledge/readiness.d.ts +26 -0
  532. package/dist/knowledge/readiness.d.ts.map +1 -0
  533. package/dist/knowledge/types.d.ts +75 -0
  534. package/dist/knowledge/types.d.ts.map +1 -0
  535. package/dist/live-proof.d.ts +62 -0
  536. package/dist/live-proof.d.ts.map +1 -0
  537. package/dist/{llm-client-Bj7g0rqu.d.ts → llm-client.d.ts} +21 -23
  538. package/dist/llm-client.d.ts.map +1 -0
  539. package/dist/llm-client.test.d.ts +2 -0
  540. package/dist/llm-client.test.d.ts.map +1 -0
  541. package/dist/locked-jsonl-appender.d.ts +19 -0
  542. package/dist/locked-jsonl-appender.d.ts.map +1 -0
  543. package/dist/matrix/aggregation.d.ts +16 -0
  544. package/dist/matrix/aggregation.d.ts.map +1 -0
  545. package/dist/matrix/index.d.ts +12 -30
  546. package/dist/matrix/index.d.ts.map +1 -0
  547. package/dist/matrix/runner.d.ts +15 -0
  548. package/dist/matrix/runner.d.ts.map +1 -0
  549. package/dist/{types-mn5Aqk7x.d.ts → matrix/types.d.ts} +11 -15
  550. package/dist/matrix/types.d.ts.map +1 -0
  551. package/dist/meta-eval/calibration.d.ts +47 -0
  552. package/dist/meta-eval/calibration.d.ts.map +1 -0
  553. package/dist/meta-eval/correlation-study.d.ts +53 -0
  554. package/dist/meta-eval/correlation-study.d.ts.map +1 -0
  555. package/dist/meta-eval/index.d.ts +6 -181
  556. package/dist/meta-eval/index.d.ts.map +1 -0
  557. package/dist/{outcome-store-rnXLEqSn.d.ts → meta-eval/outcome-store.d.ts} +7 -8
  558. package/dist/meta-eval/outcome-store.d.ts.map +1 -0
  559. package/dist/{rubric-predictive-validity-Cy_W-hWZ.d.ts → meta-eval/rubric-predictive-validity.d.ts} +8 -11
  560. package/dist/meta-eval/rubric-predictive-validity.d.ts.map +1 -0
  561. package/dist/meta-eval/sentinel.d.ts +169 -0
  562. package/dist/meta-eval/sentinel.d.ts.map +1 -0
  563. package/dist/metrics.d.ts +63 -0
  564. package/dist/metrics.d.ts.map +1 -0
  565. package/dist/model-seats.d.ts +71 -0
  566. package/dist/model-seats.d.ts.map +1 -0
  567. package/dist/model-seats.test.d.ts +2 -0
  568. package/dist/model-seats.test.d.ts.map +1 -0
  569. package/dist/muffled-gate-scanner.d.ts +102 -0
  570. package/dist/muffled-gate-scanner.d.ts.map +1 -0
  571. package/dist/{multi-layer-verifier-DUZXrPDA.d.ts → multi-layer-verifier.d.ts} +12 -15
  572. package/dist/multi-layer-verifier.d.ts.map +1 -0
  573. package/dist/multi-layer-verifier.test.d.ts +2 -0
  574. package/dist/multi-layer-verifier.test.d.ts.map +1 -0
  575. package/dist/multi-toolchain-layer.d.ts +80 -0
  576. package/dist/multi-toolchain-layer.d.ts.map +1 -0
  577. package/dist/multi-toolchain-layer.test.d.ts +2 -0
  578. package/dist/multi-toolchain-layer.test.d.ts.map +1 -0
  579. package/dist/multishot/default-tools.d.ts +34 -0
  580. package/dist/multishot/default-tools.d.ts.map +1 -0
  581. package/dist/multishot/index.d.ts +7 -290
  582. package/dist/multishot/index.d.ts.map +1 -0
  583. package/dist/multishot/index.js.map +1 -1
  584. package/dist/multishot/judges.d.ts +32 -0
  585. package/dist/multishot/judges.d.ts.map +1 -0
  586. package/dist/multishot/matrix.d.ts +107 -0
  587. package/dist/multishot/matrix.d.ts.map +1 -0
  588. package/dist/multishot/multishot.d.ts +23 -0
  589. package/dist/multishot/multishot.d.ts.map +1 -0
  590. package/dist/multishot/router.d.ts +37 -0
  591. package/dist/multishot/router.d.ts.map +1 -0
  592. package/dist/multishot/types.d.ts +60 -0
  593. package/dist/multishot/types.d.ts.map +1 -0
  594. package/dist/observability.d.ts +71 -0
  595. package/dist/observability.d.ts.map +1 -0
  596. package/dist/openapi.json +1 -1
  597. package/dist/oracle.d.ts +55 -0
  598. package/dist/oracle.d.ts.map +1 -0
  599. package/dist/orthogonality.d.ts +35 -0
  600. package/dist/orthogonality.d.ts.map +1 -0
  601. package/dist/otel-pipeline.d.ts +31 -0
  602. package/dist/otel-pipeline.d.ts.map +1 -0
  603. package/dist/paraphrase.d.ts +107 -0
  604. package/dist/paraphrase.d.ts.map +1 -0
  605. package/dist/{pareto-E-pembql.d.ts → pareto.d.ts} +9 -10
  606. package/dist/pareto.d.ts.map +1 -0
  607. package/dist/partition-held-out.d.ts +70 -0
  608. package/dist/partition-held-out.d.ts.map +1 -0
  609. package/dist/partition-held-out.test.d.ts +2 -0
  610. package/dist/partition-held-out.test.d.ts.map +1 -0
  611. package/dist/perf/index.d.ts +13 -119
  612. package/dist/perf/index.d.ts.map +1 -0
  613. package/dist/perf/integrity.d.ts +30 -0
  614. package/dist/perf/integrity.d.ts.map +1 -0
  615. package/dist/perf/journey.d.ts +45 -0
  616. package/dist/perf/journey.d.ts.map +1 -0
  617. package/dist/perf/ratchet.d.ts +47 -0
  618. package/dist/perf/ratchet.d.ts.map +1 -0
  619. package/dist/pipelines/budget-breach.d.ts +31 -0
  620. package/dist/pipelines/budget-breach.d.ts.map +1 -0
  621. package/dist/pipelines/budget-breach.test.d.ts +2 -0
  622. package/dist/pipelines/budget-breach.test.d.ts.map +1 -0
  623. package/dist/pipelines/failure-cluster.d.ts +38 -0
  624. package/dist/pipelines/failure-cluster.d.ts.map +1 -0
  625. package/dist/pipelines/failure-cluster.test.d.ts +2 -0
  626. package/dist/pipelines/failure-cluster.test.d.ts.map +1 -0
  627. package/dist/pipelines/first-divergence.d.ts +26 -0
  628. package/dist/pipelines/first-divergence.d.ts.map +1 -0
  629. package/dist/pipelines/first-divergence.test.d.ts +2 -0
  630. package/dist/pipelines/first-divergence.test.d.ts.map +1 -0
  631. package/dist/pipelines/index.d.ts +8 -173
  632. package/dist/pipelines/index.d.ts.map +1 -0
  633. package/dist/pipelines/judge-agreement.d.ts +26 -0
  634. package/dist/pipelines/judge-agreement.d.ts.map +1 -0
  635. package/dist/pipelines/judge-agreement.test.d.ts +2 -0
  636. package/dist/pipelines/judge-agreement.test.d.ts.map +1 -0
  637. package/dist/pipelines/regression.d.ts +23 -0
  638. package/dist/pipelines/regression.d.ts.map +1 -0
  639. package/dist/pipelines/regression.test.d.ts +2 -0
  640. package/dist/pipelines/regression.test.d.ts.map +1 -0
  641. package/dist/pipelines/stuck-loop.d.ts +32 -0
  642. package/dist/pipelines/stuck-loop.d.ts.map +1 -0
  643. package/dist/pipelines/stuck-loop.test.d.ts +2 -0
  644. package/dist/pipelines/stuck-loop.test.d.ts.map +1 -0
  645. package/dist/pipelines/tool-waste.d.ts +34 -0
  646. package/dist/pipelines/tool-waste.d.ts.map +1 -0
  647. package/dist/pipelines/tool-waste.test.d.ts +2 -0
  648. package/dist/pipelines/tool-waste.test.d.ts.map +1 -0
  649. package/dist/playbook.d.ts +16 -0
  650. package/dist/playbook.d.ts.map +1 -0
  651. package/dist/pr-review-benchmark.d.ts +88 -0
  652. package/dist/pr-review-benchmark.d.ts.map +1 -0
  653. package/dist/pr-review-benchmark.test.d.ts +2 -0
  654. package/dist/pr-review-benchmark.test.d.ts.map +1 -0
  655. package/dist/pre-registration.d.ts +125 -0
  656. package/dist/pre-registration.d.ts.map +1 -0
  657. package/dist/prm/builtin-rubrics.d.ts +33 -0
  658. package/dist/prm/builtin-rubrics.d.ts.map +1 -0
  659. package/dist/prm/index.d.ts +5 -104
  660. package/dist/prm/index.d.ts.map +1 -0
  661. package/dist/prm/inference.d.ts +29 -0
  662. package/dist/prm/inference.d.ts.map +1 -0
  663. package/dist/prm/inference.test.d.ts +2 -0
  664. package/dist/prm/inference.test.d.ts.map +1 -0
  665. package/dist/{rubric-Cc6UHvUb.d.ts → prm/rubric.d.ts} +10 -13
  666. package/dist/prm/rubric.d.ts.map +1 -0
  667. package/dist/prm/training-export.d.ts +38 -0
  668. package/dist/prm/training-export.d.ts.map +1 -0
  669. package/dist/produced-state.d.ts +63 -0
  670. package/dist/produced-state.d.ts.map +1 -0
  671. package/dist/produced-state.test.d.ts +8 -0
  672. package/dist/produced-state.test.d.ts.map +1 -0
  673. package/dist/profile/baselines.d.ts +37 -0
  674. package/dist/profile/baselines.d.ts.map +1 -0
  675. package/dist/profile/index.d.ts +105 -0
  676. package/dist/profile/index.d.ts.map +1 -0
  677. package/dist/promotion-gate.d.ts +93 -0
  678. package/dist/promotion-gate.d.ts.map +1 -0
  679. package/dist/prompt-registry.d.ts +41 -0
  680. package/dist/prompt-registry.d.ts.map +1 -0
  681. package/dist/propose-review-control.d.ts +49 -0
  682. package/dist/propose-review-control.d.ts.map +1 -0
  683. package/dist/propose-review-control.test.d.ts +2 -0
  684. package/dist/propose-review-control.test.d.ts.map +1 -0
  685. package/dist/propose-review.d.ts +155 -0
  686. package/dist/propose-review.d.ts.map +1 -0
  687. package/dist/{red-team-BWdoyleI.d.ts → red-team.d.ts} +14 -16
  688. package/dist/red-team.d.ts.map +1 -0
  689. package/dist/reference-replay-steering.d.ts +11 -0
  690. package/dist/reference-replay-steering.d.ts.map +1 -0
  691. package/dist/reference-replay.d.ts +176 -0
  692. package/dist/reference-replay.d.ts.map +1 -0
  693. package/dist/reflective-mutation.d.ts +79 -0
  694. package/dist/reflective-mutation.d.ts.map +1 -0
  695. package/dist/registry.d.ts +31 -0
  696. package/dist/registry.d.ts.map +1 -0
  697. package/dist/release-confidence.d.ts +128 -0
  698. package/dist/release-confidence.d.ts.map +1 -0
  699. package/dist/release-report.d.ts +11 -0
  700. package/dist/release-report.d.ts.map +1 -0
  701. package/dist/replay.d.ts +120 -0
  702. package/dist/replay.d.ts.map +1 -0
  703. package/dist/reporter.d.ts +14 -0
  704. package/dist/reporter.d.ts.map +1 -0
  705. package/dist/reporting.d.ts +15 -15
  706. package/dist/reporting.d.ts.map +1 -0
  707. package/dist/researcher.d.ts +140 -0
  708. package/dist/researcher.d.ts.map +1 -0
  709. package/dist/reviewer.d.ts +118 -0
  710. package/dist/reviewer.d.ts.map +1 -0
  711. package/dist/reviewer.test.d.ts +2 -0
  712. package/dist/reviewer.test.d.ts.map +1 -0
  713. package/dist/reward-model-export.d.ts +60 -0
  714. package/dist/reward-model-export.d.ts.map +1 -0
  715. package/dist/rl/active-curriculum.d.ts +110 -0
  716. package/dist/rl/active-curriculum.d.ts.map +1 -0
  717. package/dist/rl/adaptation-eval.d.ts +109 -0
  718. package/dist/rl/adaptation-eval.d.ts.map +1 -0
  719. package/dist/{adversarial-DIVcDoI_.d.ts → rl/adversarial.d.ts} +6 -7
  720. package/dist/rl/adversarial.d.ts.map +1 -0
  721. package/dist/rl/compute-curves.d.ts +127 -0
  722. package/dist/rl/compute-curves.d.ts.map +1 -0
  723. package/dist/rl/contamination.d.ts +117 -0
  724. package/dist/rl/contamination.d.ts.map +1 -0
  725. package/dist/rl/corpus.d.ts +55 -0
  726. package/dist/rl/corpus.d.ts.map +1 -0
  727. package/dist/rl/corpus.test.d.ts +2 -0
  728. package/dist/rl/corpus.test.d.ts.map +1 -0
  729. package/dist/rl/dataset.d.ts +102 -0
  730. package/dist/rl/dataset.d.ts.map +1 -0
  731. package/dist/rl/dataset.test.d.ts +2 -0
  732. package/dist/rl/dataset.test.d.ts.map +1 -0
  733. package/dist/rl/exporters.d.ts +141 -0
  734. package/dist/rl/exporters.d.ts.map +1 -0
  735. package/dist/rl/index.d.ts +49 -0
  736. package/dist/rl/index.d.ts.map +1 -0
  737. package/dist/{off-policy-DiwuKKg7.d.ts → rl/off-policy.d.ts} +8 -9
  738. package/dist/rl/off-policy.d.ts.map +1 -0
  739. package/dist/rl/predictive-validity-researcher.d.ts +69 -0
  740. package/dist/rl/predictive-validity-researcher.d.ts.map +1 -0
  741. package/dist/rl/preferences.d.ts +141 -0
  742. package/dist/rl/preferences.d.ts.map +1 -0
  743. package/dist/rl/process-reward.d.ts +122 -0
  744. package/dist/rl/process-reward.d.ts.map +1 -0
  745. package/dist/rl/reward-hacking.d.ts +104 -0
  746. package/dist/rl/reward-hacking.d.ts.map +1 -0
  747. package/dist/rl/rl-campaign.d.ts +85 -0
  748. package/dist/rl/rl-campaign.d.ts.map +1 -0
  749. package/dist/rl/run-record-adapters.d.ts +56 -0
  750. package/dist/rl/run-record-adapters.d.ts.map +1 -0
  751. package/dist/rl/sim-fidelity.d.ts +166 -0
  752. package/dist/rl/sim-fidelity.d.ts.map +1 -0
  753. package/dist/rl/sim-fidelity.test.d.ts +2 -0
  754. package/dist/rl/sim-fidelity.test.d.ts.map +1 -0
  755. package/dist/rl/tournament.d.ts +115 -0
  756. package/dist/rl/tournament.d.ts.map +1 -0
  757. package/dist/rl/verifiable-reward.d.ts +124 -0
  758. package/dist/rl/verifiable-reward.d.ts.map +1 -0
  759. package/dist/rl.d.ts +1 -1192
  760. package/dist/rl.js +612 -612
  761. package/dist/rl.js.map +1 -1
  762. package/dist/{run-campaign-7WNXMDSN.js → run-campaign-WXY7KI67.js} +2 -2
  763. package/dist/run-critic.d.ts +23 -0
  764. package/dist/run-critic.d.ts.map +1 -0
  765. package/dist/run-evidence.d.ts +32 -0
  766. package/dist/run-evidence.d.ts.map +1 -0
  767. package/dist/{run-record-e7vj1uZQ.d.ts → run-record.d.ts} +16 -122
  768. package/dist/run-record.d.ts.map +1 -0
  769. package/dist/run-record.test.d.ts +2 -0
  770. package/dist/run-record.test.d.ts.map +1 -0
  771. package/dist/run-score.d.ts +31 -0
  772. package/dist/run-score.d.ts.map +1 -0
  773. package/dist/runtime-trajectory.d.ts +47 -0
  774. package/dist/runtime-trajectory.d.ts.map +1 -0
  775. package/dist/{test-graded-scenario-DeODGLra.d.ts → sandbox-harness.d.ts} +15 -60
  776. package/dist/sandbox-harness.d.ts.map +1 -0
  777. package/dist/sandbox-harness.test.d.ts +2 -0
  778. package/dist/sandbox-harness.test.d.ts.map +1 -0
  779. package/dist/sandbox-pool.d.ts +74 -0
  780. package/dist/sandbox-pool.d.ts.map +1 -0
  781. package/dist/sandbox-pool.test.d.ts +2 -0
  782. package/dist/sandbox-pool.test.d.ts.map +1 -0
  783. package/dist/scorecard.d.ts +133 -0
  784. package/dist/scorecard.d.ts.map +1 -0
  785. package/dist/scorecard.test.d.ts +2 -0
  786. package/dist/scorecard.test.d.ts.map +1 -0
  787. package/dist/self-play.d.ts +69 -0
  788. package/dist/self-play.d.ts.map +1 -0
  789. package/dist/semantic-concept-judge.d.ts +135 -0
  790. package/dist/semantic-concept-judge.d.ts.map +1 -0
  791. package/dist/semantic-concept-judge.test.d.ts +2 -0
  792. package/dist/semantic-concept-judge.test.d.ts.map +1 -0
  793. package/dist/{sequential-5iSVfzl2.d.ts → sequential.d.ts} +9 -10
  794. package/dist/sequential.d.ts.map +1 -0
  795. package/dist/{series-convergence-D5OWMBg6.d.ts → series-convergence.d.ts} +4 -5
  796. package/dist/series-convergence.d.ts.map +1 -0
  797. package/dist/slo.d.ts +48 -0
  798. package/dist/slo.d.ts.map +1 -0
  799. package/dist/state-continuity.d.ts +47 -0
  800. package/dist/state-continuity.d.ts.map +1 -0
  801. package/dist/{statistics-CCJpTGOS.d.ts → statistics.d.ts} +46 -48
  802. package/dist/statistics.d.ts.map +1 -0
  803. package/dist/statistics.test.d.ts +2 -0
  804. package/dist/statistics.test.d.ts.map +1 -0
  805. package/dist/steering-optimizer.d.ts +58 -0
  806. package/dist/steering-optimizer.d.ts.map +1 -0
  807. package/dist/steering.d.ts +24 -0
  808. package/dist/steering.d.ts.map +1 -0
  809. package/dist/storyboard/code-edit.d.ts +64 -0
  810. package/dist/storyboard/code-edit.d.ts.map +1 -0
  811. package/dist/storyboard/code-edit.test.d.ts +2 -0
  812. package/dist/storyboard/code-edit.test.d.ts.map +1 -0
  813. package/dist/storyboard/index.d.ts +16 -81
  814. package/dist/storyboard/index.d.ts.map +1 -0
  815. package/dist/storyboard/index.test.d.ts +2 -0
  816. package/dist/storyboard/index.test.d.ts.map +1 -0
  817. package/dist/{summary-report-BDOFevaT.d.ts → summary-report.d.ts} +23 -160
  818. package/dist/summary-report.d.ts.map +1 -0
  819. package/dist/telemetry/client.d.ts +35 -0
  820. package/dist/telemetry/client.d.ts.map +1 -0
  821. package/dist/telemetry/index.d.ts +17 -35
  822. package/dist/telemetry/index.d.ts.map +1 -0
  823. package/dist/telemetry/schema.d.ts +61 -0
  824. package/dist/telemetry/schema.d.ts.map +1 -0
  825. package/dist/telemetry/sink-fetch.d.ts +39 -0
  826. package/dist/telemetry/sink-fetch.d.ts.map +1 -0
  827. package/dist/telemetry/{file.d.ts → sink-file.d.ts} +5 -7
  828. package/dist/telemetry/sink-file.d.ts.map +1 -0
  829. package/dist/test-graded-scenario.d.ts +42 -0
  830. package/dist/test-graded-scenario.d.ts.map +1 -0
  831. package/dist/testing.d.ts +5 -0
  832. package/dist/testing.d.ts.map +1 -0
  833. package/dist/testing.js +8 -0
  834. package/dist/testing.js.map +1 -0
  835. package/dist/tool-use-metrics.d.ts +35 -0
  836. package/dist/tool-use-metrics.d.ts.map +1 -0
  837. package/dist/trace/capture-fetch.d.ts +48 -0
  838. package/dist/trace/capture-fetch.d.ts.map +1 -0
  839. package/dist/trace/capture-fetch.test.d.ts +2 -0
  840. package/dist/trace/capture-fetch.test.d.ts.map +1 -0
  841. package/dist/{emitter-C2rqGH_l.d.ts → trace/emitter.d.ts} +9 -12
  842. package/dist/trace/emitter.d.ts.map +1 -0
  843. package/dist/trace/extract-usage.d.ts +46 -0
  844. package/dist/trace/extract-usage.d.ts.map +1 -0
  845. package/dist/trace/extract-usage.test.d.ts +2 -0
  846. package/dist/trace/extract-usage.test.d.ts.map +1 -0
  847. package/dist/trace/index.d.ts +15 -0
  848. package/dist/trace/index.d.ts.map +1 -0
  849. package/dist/{integrity-D2t12mMw.d.ts → trace/integrity.d.ts} +11 -14
  850. package/dist/trace/integrity.d.ts.map +1 -0
  851. package/dist/trace/otel-bridge.d.ts +29 -0
  852. package/dist/trace/otel-bridge.d.ts.map +1 -0
  853. package/dist/trace/otel-export.d.ts +52 -0
  854. package/dist/trace/otel-export.d.ts.map +1 -0
  855. package/dist/trace/otel.d.ts +57 -0
  856. package/dist/trace/otel.d.ts.map +1 -0
  857. package/dist/trace/otlp-attributes.d.ts +17 -0
  858. package/dist/trace/otlp-attributes.d.ts.map +1 -0
  859. package/dist/trace/query.d.ts +29 -0
  860. package/dist/trace/query.d.ts.map +1 -0
  861. package/dist/trace/query.test.d.ts +2 -0
  862. package/dist/trace/query.test.d.ts.map +1 -0
  863. package/dist/{raw-provider-sink-C46HDghv.d.ts → trace/raw-provider-sink.d.ts} +13 -14
  864. package/dist/trace/raw-provider-sink.d.ts.map +1 -0
  865. package/dist/{redact-B40YG2M_.d.ts → trace/redact.d.ts} +7 -8
  866. package/dist/trace/redact.d.ts.map +1 -0
  867. package/dist/{schema-m0gsnbt3.d.ts → trace/schema.d.ts} +29 -30
  868. package/dist/trace/schema.d.ts.map +1 -0
  869. package/dist/trace/store-to-otlp.d.ts +72 -0
  870. package/dist/trace/store-to-otlp.d.ts.map +1 -0
  871. package/dist/trace/store-to-otlp.test.d.ts +2 -0
  872. package/dist/trace/store-to-otlp.test.d.ts.map +1 -0
  873. package/dist/{store-BcFXE6LG.d.ts → trace/store.d.ts} +21 -12
  874. package/dist/trace/store.d.ts.map +1 -0
  875. package/dist/trace/store.test.d.ts +2 -0
  876. package/dist/trace/store.test.d.ts.map +1 -0
  877. package/dist/{analyst-C8HHvfJp.d.ts → trace-analyst/analyst.d.ts} +12 -14
  878. package/dist/trace-analyst/analyst.d.ts.map +1 -0
  879. package/dist/trace-analyst/analyst.test.d.ts +2 -0
  880. package/dist/trace-analyst/analyst.test.d.ts.map +1 -0
  881. package/dist/trace-analyst/behavioral-metrics.d.ts +40 -0
  882. package/dist/trace-analyst/behavioral-metrics.d.ts.map +1 -0
  883. package/dist/trace-analyst/behavioral-metrics.test.d.ts +2 -0
  884. package/dist/trace-analyst/behavioral-metrics.test.d.ts.map +1 -0
  885. package/dist/trace-analyst/hook.d.ts +55 -0
  886. package/dist/trace-analyst/hook.d.ts.map +1 -0
  887. package/dist/trace-analyst/index.d.ts +18 -0
  888. package/dist/trace-analyst/index.d.ts.map +1 -0
  889. package/dist/trace-analyst/insights.d.ts +71 -0
  890. package/dist/trace-analyst/insights.d.ts.map +1 -0
  891. package/dist/trace-analyst/insights.test.d.ts +2 -0
  892. package/dist/trace-analyst/insights.test.d.ts.map +1 -0
  893. package/dist/trace-analyst/otlp-flatten.d.ts +42 -0
  894. package/dist/trace-analyst/otlp-flatten.d.ts.map +1 -0
  895. package/dist/trace-analyst/otlp-span.d.ts +85 -0
  896. package/dist/trace-analyst/otlp-span.d.ts.map +1 -0
  897. package/dist/trace-analyst/otlp-span.test.d.ts +2 -0
  898. package/dist/trace-analyst/otlp-span.test.d.ts.map +1 -0
  899. package/dist/trace-analyst/otlp-to-run-records.d.ts +115 -0
  900. package/dist/trace-analyst/otlp-to-run-records.d.ts.map +1 -0
  901. package/dist/trace-analyst/otlp-to-run-records.test.d.ts +2 -0
  902. package/dist/trace-analyst/otlp-to-run-records.test.d.ts.map +1 -0
  903. package/dist/trace-analyst/otlp-to-run-records.timestamps.test.d.ts +2 -0
  904. package/dist/trace-analyst/otlp-to-run-records.timestamps.test.d.ts.map +1 -0
  905. package/dist/trace-analyst/prompts.d.ts +6 -0
  906. package/dist/trace-analyst/prompts.d.ts.map +1 -0
  907. package/dist/trace-analyst/store-otlp.d.ts +126 -0
  908. package/dist/trace-analyst/store-otlp.d.ts.map +1 -0
  909. package/dist/trace-analyst/store-otlp.test.d.ts +8 -0
  910. package/dist/trace-analyst/store-otlp.test.d.ts.map +1 -0
  911. package/dist/trace-analyst/store-otlp.timestamps.test.d.ts +2 -0
  912. package/dist/trace-analyst/store-otlp.timestamps.test.d.ts.map +1 -0
  913. package/dist/trace-analyst/store.d.ts +63 -0
  914. package/dist/trace-analyst/store.d.ts.map +1 -0
  915. package/dist/trace-analyst/tools.d.ts +44 -0
  916. package/dist/trace-analyst/tools.d.ts.map +1 -0
  917. package/dist/trace-analyst/tools.test.d.ts +10 -0
  918. package/dist/trace-analyst/tools.test.d.ts.map +1 -0
  919. package/dist/{store-C1YxJDEK.d.ts → trace-analyst/types.d.ts} +18 -74
  920. package/dist/trace-analyst/types.d.ts.map +1 -0
  921. package/dist/trace-contracts.d.ts +180 -0
  922. package/dist/trace-contracts.d.ts.map +1 -0
  923. package/dist/traced-analyst.d.ts +26 -0
  924. package/dist/traced-analyst.d.ts.map +1 -0
  925. package/dist/traced-judges.d.ts +27 -0
  926. package/dist/traced-judges.d.ts.map +1 -0
  927. package/dist/traces.d.ts +4 -960
  928. package/dist/traces.d.ts.map +1 -0
  929. package/dist/traces.js +41 -11
  930. package/dist/{trajectory-2TkpSEVh.d.ts → trajectory.d.ts} +8 -9
  931. package/dist/trajectory.d.ts.map +1 -0
  932. package/dist/{types-C7DGg5ex.d.ts → types.d.ts} +31 -33
  933. package/dist/types.d.ts.map +1 -0
  934. package/dist/ui-finding.d.ts +104 -0
  935. package/dist/ui-finding.d.ts.map +1 -0
  936. package/dist/verdict-cache.d.ts +78 -0
  937. package/dist/verdict-cache.d.ts.map +1 -0
  938. package/dist/verdict-cache.test.d.ts +2 -0
  939. package/dist/verdict-cache.test.d.ts.map +1 -0
  940. package/dist/{verdict-C9MlYujm.d.ts → verdict.d.ts} +2 -3
  941. package/dist/verdict.d.ts.map +1 -0
  942. package/dist/visual-diff.d.ts +32 -0
  943. package/dist/visual-diff.d.ts.map +1 -0
  944. package/dist/wire/handlers.d.ts +54 -0
  945. package/dist/wire/handlers.d.ts.map +1 -0
  946. package/dist/wire/index.d.ts +13 -570
  947. package/dist/wire/index.d.ts.map +1 -0
  948. package/dist/wire/openapi.d.ts +3 -0
  949. package/dist/wire/openapi.d.ts.map +1 -0
  950. package/dist/wire/rpc.d.ts +21 -0
  951. package/dist/wire/rpc.d.ts.map +1 -0
  952. package/dist/wire/rubrics.d.ts +34 -0
  953. package/dist/wire/rubrics.d.ts.map +1 -0
  954. package/dist/wire/schemas.d.ts +410 -0
  955. package/dist/wire/schemas.d.ts.map +1 -0
  956. package/dist/wire/server.d.ts +60 -0
  957. package/dist/wire/server.d.ts.map +1 -0
  958. package/dist/workflow/event-schema.d.ts +5 -0
  959. package/dist/workflow/event-schema.d.ts.map +1 -0
  960. package/dist/workflow/feedback-pack.d.ts +99 -0
  961. package/dist/workflow/feedback-pack.d.ts.map +1 -0
  962. package/dist/workflow/index.d.ts +22 -495
  963. package/dist/workflow/index.d.ts.map +1 -0
  964. package/dist/workflow/index.js +1 -1
  965. package/dist/workflow/intelligence-export.d.ts +62 -0
  966. package/dist/workflow/intelligence-export.d.ts.map +1 -0
  967. package/dist/workflow/partner-report.d.ts +49 -0
  968. package/dist/workflow/partner-report.d.ts.map +1 -0
  969. package/dist/workflow/phase-graph.d.ts +43 -0
  970. package/dist/workflow/phase-graph.d.ts.map +1 -0
  971. package/dist/workflow/promotion-gate.d.ts +61 -0
  972. package/dist/workflow/promotion-gate.d.ts.map +1 -0
  973. package/dist/workflow/run-record.d.ts +12 -0
  974. package/dist/workflow/run-record.d.ts.map +1 -0
  975. package/dist/workflow/runtime-adapter.d.ts +20 -0
  976. package/dist/workflow/runtime-adapter.d.ts.map +1 -0
  977. package/dist/workflow/sanitize.d.ts +21 -0
  978. package/dist/workflow/sanitize.d.ts.map +1 -0
  979. package/dist/workflow/schema.d.ts +5 -0
  980. package/dist/workflow/schema.d.ts.map +1 -0
  981. package/dist/workflow/summary.d.ts +43 -0
  982. package/dist/workflow/summary.d.ts.map +1 -0
  983. package/dist/workflow/trace-event-fields.d.ts +6 -0
  984. package/dist/workflow/trace-event-fields.d.ts.map +1 -0
  985. package/dist/workflow/trajectory.d.ts +15 -0
  986. package/dist/workflow/trajectory.d.ts.map +1 -0
  987. package/dist/workflow/types.d.ts +68 -0
  988. package/dist/workflow/types.d.ts.map +1 -0
  989. package/dist/workspace-inspector.d.ts +67 -0
  990. package/dist/workspace-inspector.d.ts.map +1 -0
  991. package/dist/wrangler-deploy-runner.test.d.ts +2 -0
  992. package/dist/wrangler-deploy-runner.test.d.ts.map +1 -0
  993. package/docs/campaign-proposers.md +170 -0
  994. package/docs/concepts.md +8 -4
  995. package/docs/customer-journeys.md +15 -13
  996. package/docs/design/loop-taxonomy.md +34 -66
  997. package/docs/distributed-driver.md +14 -14
  998. package/docs/feature-guide.md +1 -1
  999. package/docs/hosted-ingest-spec.md +2 -3
  1000. package/docs/multi-shot-optimization.md +8 -8
  1001. package/docs/product-eval-adoption.md +1 -1
  1002. package/docs/self-improvement-map.md +33 -29
  1003. package/package.json +9 -15
  1004. package/dist/calibration-BPmzuVPk.d.ts +0 -101
  1005. package/dist/chunk-2K6UUZ7P.js.map +0 -1
  1006. package/dist/chunk-CTBHKLEU.js.map +0 -1
  1007. package/dist/chunk-E4GH6USR.js.map +0 -1
  1008. package/dist/chunk-EGPMSBEZ.js.map +0 -1
  1009. package/dist/chunk-KWRRMR3J.js.map +0 -1
  1010. package/dist/chunk-MIFZUPEK.js.map +0 -1
  1011. package/dist/chunk-MPQWFX6Y.js.map +0 -1
  1012. package/dist/chunk-Q5LIB7BC.js.map +0 -1
  1013. package/dist/chunk-QMUEXQJS.js.map +0 -1
  1014. package/dist/chunk-SD2YFWQQ.js.map +0 -1
  1015. package/dist/control-D6qwHXIR.d.ts +0 -259
  1016. package/dist/corpus-B8A4BDR3.d.ts +0 -560
  1017. package/dist/failure-cluster-DH9Flgcf.d.ts +0 -76
  1018. package/dist/harness-optimizer-mOl9XX_O.d.ts +0 -106
  1019. package/dist/index-Bx3gZ8xl.d.ts +0 -159
  1020. package/dist/pre-registration-mAnCugl9.d.ts +0 -523
  1021. package/dist/provenance-P-bCL2Fo.d.ts +0 -441
  1022. package/dist/query-B7GGjRox.d.ts +0 -32
  1023. package/dist/release-report-BEbWmVYj.d.ts +0 -233
  1024. package/dist/run-critic-CmMf05uV.d.ts +0 -56
  1025. package/dist/run-improvement-loop-DBahB8Ax.d.ts +0 -427
  1026. package/dist/runtime-trajectory-BDgfGZSr.d.ts +0 -49
  1027. package/dist/semantic-concept-judge-B9MgmBnM.d.ts +0 -624
  1028. package/dist/sink-fetch-B1Yg4Til.d.ts +0 -101
  1029. package/docs/design/external-agent-wedge.md +0 -89
  1030. package/docs/design/phase-d-rfc.md +0 -125
  1031. package/docs/design/phase4-consumer-migration.md +0 -70
  1032. package/docs/design/primitives-integration-spec.md +0 -393
  1033. package/docs/design/product-self-improvement-loop.md +0 -146
  1034. package/docs/design/self-improvement-engine.md +0 -140
  1035. package/docs/design/self-improvement-protocol.md +0 -223
  1036. package/docs/design/self-improvement-roadmap.md +0 -106
  1037. package/docs/design/substrate-gaps.md +0 -118
  1038. package/docs/phase-b-pairing-kit.md +0 -188
  1039. package/docs/phase-b-runbook.md +0 -176
  1040. package/docs/pilot/README.md +0 -62
  1041. package/docs/pilot/customer-checklist.md +0 -90
  1042. package/docs/pilot/integration-foreign-stack.md +0 -296
  1043. package/docs/pilot/integration-tangle-stack.md +0 -248
  1044. package/docs/pilot/one-pager.md +0 -161
  1045. package/docs/pilot/sample-insight-report.json +0 -172
  1046. package/docs/quickstart-external.md +0 -229
  1047. package/docs/research/belief-state-agent-eval-roadmap.md +0 -593
  1048. package/docs/research/research-roadmap.md +0 -205
  1049. package/docs/specs/driver-honest-spec.md +0 -251
  1050. package/docs/specs/hermes-self-improvement-audit.md +0 -93
  1051. package/docs/specs/profile-versioning.md +0 -291
  1052. package/docs/three-package-architecture.md +0 -168
  1053. /package/dist/{chunk-TBDR6PAI.js.map → chunk-IZCEK2HR.js.map} +0 -0
  1054. /package/dist/{chunk-KW53MSA5.js.map → chunk-X74V6ESX.js.map} +0 -0
  1055. /package/dist/{chunk-2KNZHH3P.js.map → chunk-Z6L6YSU6.js.map} +0 -0
  1056. /package/dist/{run-campaign-7WNXMDSN.js.map → run-campaign-WXY7KI67.js.map} +0 -0
@@ -1,7 +1,3 @@
1
- import { T as TraceEmitter } from './emitter-C2rqGH_l.js';
2
- import { F as FailureClass } from './schema-m0gsnbt3.js';
3
- import { T as TraceStore } from './store-BcFXE6LG.js';
4
-
5
1
  /**
6
2
  * Policy-based agent control runtime.
7
3
  *
@@ -13,10 +9,12 @@ import { T as TraceStore } from './store-BcFXE6LG.js';
13
9
  * critic/revise, driver intervention, specialist calls, and human escalation
14
10
  * are all just actions chosen by the control policy.
15
11
  */
16
-
17
- type ControlSeverity = 'info' | 'warning' | 'error' | 'critical';
18
- type ControlActionFailureMode = 'continue' | 'stop';
19
- interface ControlEvalResult {
12
+ import { TraceEmitter } from './trace/emitter';
13
+ import type { FailureClass } from './trace/schema';
14
+ import type { TraceStore } from './trace/store';
15
+ export type ControlSeverity = 'info' | 'warning' | 'error' | 'critical';
16
+ export type ControlActionFailureMode = 'continue' | 'stop';
17
+ export interface ControlEvalResult {
20
18
  /** Stable validator or judge id. */
21
19
  id: string;
22
20
  /** Whether this check passed. */
@@ -34,12 +32,12 @@ interface ControlEvalResult {
34
32
  /** Structured details for downstream control policies and reports. */
35
33
  metadata?: Record<string, unknown>;
36
34
  }
37
- interface ControlBudget {
35
+ export interface ControlBudget {
38
36
  maxSteps: number;
39
37
  maxWallMs?: number;
40
38
  maxCostUsd?: number;
41
39
  }
42
- interface ControlStopPolicies<TState, TAction> {
40
+ export interface ControlStopPolicies<TState, TAction> {
43
41
  /**
44
42
  * Stop after N consecutive steps with no state fingerprint change and
45
43
  * less than `minScoreDelta` score movement. Disabled when omitted.
@@ -57,7 +55,7 @@ interface ControlStopPolicies<TState, TAction> {
57
55
  /** Override the default JSON/string fingerprint for repeated-action checks. */
58
56
  actionFingerprint?: (action: TAction) => string;
59
57
  }
60
- interface ControlContext<TState, TAction, TActionResult, TEval extends ControlEvalResult = ControlEvalResult> {
58
+ export interface ControlContext<TState, TAction, TActionResult, TEval extends ControlEvalResult = ControlEvalResult> {
61
59
  intent: string;
62
60
  state: TState;
63
61
  evals: TEval[];
@@ -70,7 +68,7 @@ interface ControlContext<TState, TAction, TActionResult, TEval extends ControlEv
70
68
  abortSignal: AbortSignal;
71
69
  emitter?: TraceEmitter;
72
70
  }
73
- type ControlDecision<TAction> = {
71
+ export type ControlDecision<TAction> = {
74
72
  type: 'continue';
75
73
  action: TAction;
76
74
  reason?: string;
@@ -80,26 +78,26 @@ type ControlDecision<TAction> = {
80
78
  pass?: boolean;
81
79
  score?: number;
82
80
  };
83
- interface StopDecision {
81
+ export interface StopDecision {
84
82
  stop: boolean;
85
83
  pass: boolean;
86
84
  reason: string;
87
85
  score?: number;
88
86
  failureClass?: FailureClass;
89
87
  }
90
- interface ControlActionOutcome<TActionResult> {
88
+ export interface ControlActionOutcome<TActionResult> {
91
89
  ok: boolean;
92
90
  result?: TActionResult;
93
91
  error?: string;
94
92
  costUsd?: number;
95
93
  durationMs: number;
96
94
  }
97
- interface ControlRuntimeError {
95
+ export interface ControlRuntimeError {
98
96
  phase: 'observe' | 'validate' | 'decide' | 'act' | 'stop-policy' | 'on-step' | 'trace';
99
97
  stepIndex: number;
100
98
  message: string;
101
99
  }
102
- interface ControlStep<TState, TAction, TActionResult, TEval extends ControlEvalResult = ControlEvalResult> {
100
+ export interface ControlStep<TState, TAction, TActionResult, TEval extends ControlEvalResult = ControlEvalResult> {
103
101
  index: number;
104
102
  decision: ControlDecision<TAction>;
105
103
  beforeState: TState;
@@ -110,7 +108,7 @@ interface ControlStep<TState, TAction, TActionResult, TEval extends ControlEvalR
110
108
  startedAt: string;
111
109
  endedAt: string;
112
110
  }
113
- interface ControlRunResult<TState, TAction, TActionResult, TEval extends ControlEvalResult = ControlEvalResult> {
111
+ export interface ControlRunResult<TState, TAction, TActionResult, TEval extends ControlEvalResult = ControlEvalResult> {
114
112
  intent: string;
115
113
  pass: boolean;
116
114
  completed: boolean;
@@ -127,7 +125,7 @@ interface ControlRunResult<TState, TAction, TActionResult, TEval extends Control
127
125
  runtimeErrors: ControlRuntimeError[];
128
126
  stoppedBy: 'policy' | 'stop-policy' | 'budget' | 'abort' | 'runtime-error';
129
127
  }
130
- interface ControlRuntimeConfig<TState, TAction, TActionResult, TEval extends ControlEvalResult = ControlEvalResult> {
128
+ export interface ControlRuntimeConfig<TState, TAction, TActionResult, TEval extends ControlEvalResult = ControlEvalResult> {
131
129
  intent: string;
132
130
  budget?: Partial<ControlBudget>;
133
131
  signal?: AbortSignal;
@@ -172,11 +170,10 @@ interface ControlRuntimeConfig<TState, TAction, TActionResult, TEval extends Con
172
170
  projectId?: string;
173
171
  variantId?: string;
174
172
  }
175
- declare function runAgentControlLoop<TState, TAction, TActionResult, TEval extends ControlEvalResult = ControlEvalResult>(config: ControlRuntimeConfig<TState, TAction, TActionResult, TEval>): Promise<ControlRunResult<TState, TAction, TActionResult, TEval>>;
176
- declare function stopOnNoProgress<TState, TAction>(maxNoProgressSteps: number, options?: Omit<ControlStopPolicies<TState, TAction>, 'maxNoProgressSteps'>): ControlStopPolicies<TState, TAction>;
177
- declare function stopOnRepeatedAction<TState, TAction>(maxRepeatedActions: number, options?: Omit<ControlStopPolicies<TState, TAction>, 'maxRepeatedActions'>): ControlStopPolicies<TState, TAction>;
178
- declare function objectiveEval(input: Omit<ControlEvalResult, 'objective'>): ControlEvalResult;
179
- declare function subjectiveEval(input: Omit<ControlEvalResult, 'objective'>): ControlEvalResult;
180
- declare function allCriticalPassed(evals: ControlEvalResult[]): boolean;
181
-
182
- export { type ControlEvalResult as C, type StopDecision as S, type ControlRunResult as a, type ControlStep as b, type ControlActionFailureMode as c, type ControlActionOutcome as d, type ControlBudget as e, type ControlContext as f, type ControlDecision as g, type ControlRuntimeConfig as h, type ControlRuntimeError as i, type ControlSeverity as j, type ControlStopPolicies as k, allCriticalPassed as l, stopOnRepeatedAction as m, subjectiveEval as n, objectiveEval as o, runAgentControlLoop as r, stopOnNoProgress as s };
173
+ export declare function runAgentControlLoop<TState, TAction, TActionResult, TEval extends ControlEvalResult = ControlEvalResult>(config: ControlRuntimeConfig<TState, TAction, TActionResult, TEval>): Promise<ControlRunResult<TState, TAction, TActionResult, TEval>>;
174
+ export declare function stopOnNoProgress<TState, TAction>(maxNoProgressSteps: number, options?: Omit<ControlStopPolicies<TState, TAction>, 'maxNoProgressSteps'>): ControlStopPolicies<TState, TAction>;
175
+ export declare function stopOnRepeatedAction<TState, TAction>(maxRepeatedActions: number, options?: Omit<ControlStopPolicies<TState, TAction>, 'maxRepeatedActions'>): ControlStopPolicies<TState, TAction>;
176
+ export declare function objectiveEval(input: Omit<ControlEvalResult, 'objective'>): ControlEvalResult;
177
+ export declare function subjectiveEval(input: Omit<ControlEvalResult, 'objective'>): ControlEvalResult;
178
+ export declare function allCriticalPassed(evals: ControlEvalResult[]): boolean;
179
+ //# sourceMappingURL=control-runtime.d.ts.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"control-runtime.d.ts","sourceRoot":"","sources":["../src/control-runtime.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAGH,OAAO,EAAmB,YAAY,EAAE,MAAM,iBAAiB,CAAA;AAC/D,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,gBAAgB,CAAA;AAClD,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,eAAe,CAAA;AAE/C,MAAM,MAAM,eAAe,GAAG,MAAM,GAAG,SAAS,GAAG,OAAO,GAAG,UAAU,CAAA;AACvE,MAAM,MAAM,wBAAwB,GAAG,UAAU,GAAG,MAAM,CAAA;AAE1D,MAAM,WAAW,iBAAiB;IAChC,oCAAoC;IACpC,EAAE,EAAE,MAAM,CAAA;IACV,iCAAiC;IACjC,MAAM,EAAE,OAAO,CAAA;IACf,sDAAsD;IACtD,KAAK,CAAC,EAAE,MAAM,CAAA;IACd,gFAAgF;IAChF,QAAQ,CAAC,EAAE,eAAe,CAAA;IAC1B,6BAA6B;IAC7B,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,8DAA8D;IAC9D,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB,4EAA4E;IAC5E,SAAS,CAAC,EAAE,OAAO,CAAA;IACnB,sEAAsE;IACtE,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CACnC;AAED,MAAM,WAAW,aAAa;IAC5B,QAAQ,EAAE,MAAM,CAAA;IAChB,SAAS,CAAC,EAAE,MAAM,CAAA;IAClB,UAAU,CAAC,EAAE,MAAM,CAAA;CACpB;AAED,MAAM,WAAW,mBAAmB,CAAC,MAAM,EAAE,OAAO;IAClD;;;OAGG;IACH,kBAAkB,CAAC,EAAE,MAAM,CAAA;IAC3B;;;OAGG;IACH,kBAAkB,CAAC,EAAE,MAAM,CAAA;IAC3B,qEAAqE;IACrE,aAAa,CAAC,EAAE,MAAM,CAAA;IACtB,0EAA0E;IAC1E,gBAAgB,CAAC,EAAE,CAAC,KAAK,EAAE,MAAM,KAAK,MAAM,CAAA;IAC5C,+EAA+E;IAC/E,iBAAiB,CAAC,EAAE,CAAC,MAAM,EAAE,OAAO,KAAK,MAAM,CAAA;CAChD;AAED,MAAM,WAAW,cAAc,CAC7B,MAAM,EACN,OAAO,EACP,aAAa,EACb,KAAK,SAAS,iBAAiB,GAAG,iBAAiB;IAEnD,MAAM,EAAE,MAAM,CAAA;IACd,KAAK,EAAE,MAAM,CAAA;IACb,KAAK,EAAE,KAAK,EAAE,CAAA;IACd,OAAO,EAAE,WAAW,CAAC,MAAM,EAAE,OAAO,EAAE,aAAa,EAAE,KAAK,CAAC,EAAE,CAAA;IAC7D,MAAM,EAAE,aAAa,CAAA;IACrB,SAAS,EAAE,MAAM,CAAA;IACjB,MAAM,EAAE,MAAM,CAAA;IACd,YAAY,EAAE,MAAM,CAAA;IACpB,gBAAgB,CAAC,EAAE,MAAM,CAAA;IACzB,WAAW,EAAE,WAAW,CAAA;IACxB,OAAO,CAAC,EAAE,YAAY,CAAA;CACvB;AAED,MAAM,MAAM,eAAe,CAAC,OAAO,IAC/B;IACE,IAAI,EAAE,UAAU,CAAA;IAChB,MAAM,EAAE,OAAO,CAAA;IACf,MAAM,CAAC,EAAE,MAAM,CAAA;CAChB,GACD;IACE,IAAI,EAAE,MAAM,CAAA;IACZ,MAAM,EAAE,MAAM,CAAA;IACd,IAAI,CAAC,EAAE,OAAO,CAAA;IACd,KAAK,CAAC,EAAE,MAAM,CAAA;CACf,CAAA;AAEL,MAAM,WAAW,YAAY;IAC3B,IAAI,EAAE,OAAO,CAAA;IACb,IAAI,EAAE,OAAO,CAAA;IACb,MAAM,EAAE,MAAM,CAAA;IACd,KAAK,CAAC,EAAE,MAAM,CAAA;IACd,YAAY,CAAC,EAAE,YAAY,CAAA;CAC5B;AAED,MAAM,WAAW,oBAAoB,CAAC,aAAa;IACjD,EAAE,EAAE,OAAO,CAAA;IACX,MAAM,CAAC,EAAE,aAAa,CAAA;IACtB,KAAK,CAAC,EAAE,MAAM,CAAA;IACd,OAAO,CAAC,EAAE,MAAM,CAAA;IAChB,UAAU,EAAE,MAAM,CAAA;CACnB;AAED,MAAM,WAAW,mBAAmB;IAClC,KAAK,EAAE,SAAS,GAAG,UAAU,GAAG,QAAQ,GAAG,KAAK,GAAG,aAAa,GAAG,SAAS,GAAG,OAAO,CAAA;IACtF,SAAS,EAAE,MAAM,CAAA;IACjB,OAAO,EAAE,MAAM,CAAA;CAChB;AAED,MAAM,WAAW,WAAW,CAC1B,MAAM,EACN,OAAO,EACP,aAAa,EACb,KAAK,SAAS,iBAAiB,GAAG,iBAAiB;IAEnD,KAAK,EAAE,MAAM,CAAA;IACb,QAAQ,EAAE,eAAe,CAAC,OAAO,CAAC,CAAA;IAClC,WAAW,EAAE,MAAM,CAAA;IACnB,UAAU,EAAE,MAAM,CAAA;IAClB,WAAW,EAAE,KAAK,EAAE,CAAA;IACpB,UAAU,EAAE,KAAK,EAAE,CAAA;IACnB,aAAa,CAAC,EAAE,oBAAoB,CAAC,aAAa,CAAC,CAAA;IACnD,SAAS,EAAE,MAAM,CAAA;IACjB,OAAO,EAAE,MAAM,CAAA;CAChB;AAED,MAAM,WAAW,gBAAgB,CAC/B,MAAM,EACN,OAAO,EACP,aAAa,EACb,KAAK,SAAS,iBAAiB,GAAG,iBAAiB;IAEnD,MAAM,EAAE,MAAM,CAAA;IACd,IAAI,EAAE,OAAO,CAAA;IACb,SAAS,EAAE,OAAO,CAAA;IAClB,MAAM,EAAE,MAAM,CAAA;IACd,KAAK,CAAC,EAAE,MAAM,CAAA;IACd,KAAK,EAAE,WAAW,CAAC,MAAM,EAAE,OAAO,EAAE,aAAa,EAAE,KAAK,CAAC,EAAE,CAAA;IAC3D,UAAU,EAAE,MAAM,GAAG,SAAS,CAAA;IAC9B,UAAU,EAAE,KAAK,EAAE,CAAA;IACnB,MAAM,EAAE,MAAM,CAAA;IACd,YAAY,EAAE,MAAM,CAAA;IACpB,6FAA6F;IAC7F,KAAK,EAAE,MAAM,GAAG,IAAI,CAAA;IACpB,YAAY,CAAC,EAAE,YAAY,CAAA;IAC3B,aAAa,EAAE,mBAAmB,EAAE,CAAA;IACpC,SAAS,EAAE,QAAQ,GAAG,aAAa,GAAG,QAAQ,GAAG,OAAO,GAAG,eAAe,CAAA;CAC3E;AAED,MAAM,WAAW,oBAAoB,CACnC,MAAM,EACN,OAAO,EACP,aAAa,EACb,KAAK,SAAS,iBAAiB,GAAG,iBAAiB;IAEnD,MAAM,EAAE,MAAM,CAAA;IACd,MAAM,CAAC,EAAE,OAAO,CAAC,aAAa,CAAC,CAAA;IAC/B,MAAM,CAAC,EAAE,WAAW,CAAA;IACpB,iGAAiG;IACjG,aAAa,CAAC,EAAE,wBAAwB,CAAA;IACxC;;;OAGG;IACH,gBAAgB,CAAC,EAAE,CAAC,GAAG,EAAE;QACvB,MAAM,EAAE,OAAO,CAAA;QACf,MAAM,EAAE,aAAa,CAAA;QACrB,KAAK,EAAE,MAAM,CAAA;QACb,KAAK,EAAE,KAAK,EAAE,CAAA;QACd,OAAO,EAAE,WAAW,CAAC,MAAM,EAAE,OAAO,EAAE,aAAa,EAAE,KAAK,CAAC,EAAE,CAAA;KAC9D,KAAK,MAAM,GAAG,SAAS,CAAA;IAExB,2FAA2F;IAC3F,OAAO,EAAE,CAAC,GAAG,EAAE;QACb,OAAO,EAAE,WAAW,CAAC,MAAM,EAAE,OAAO,EAAE,aAAa,EAAE,KAAK,CAAC,EAAE,CAAA;QAC7D,WAAW,EAAE,WAAW,CAAA;KACzB,KAAK,OAAO,CAAC,MAAM,CAAC,GAAG,MAAM,CAAA;IAE9B,gGAAgG;IAChG,QAAQ,EAAE,CAAC,GAAG,EAAE;QACd,MAAM,EAAE,MAAM,CAAA;QACd,KAAK,EAAE,MAAM,CAAA;QACb,OAAO,EAAE,WAAW,CAAC,MAAM,EAAE,OAAO,EAAE,aAAa,EAAE,KAAK,CAAC,EAAE,CAAA;QAC7D,WAAW,EAAE,WAAW,CAAA;KACzB,KAAK,OAAO,CAAC,KAAK,EAAE,CAAC,GAAG,KAAK,EAAE,CAAA;IAEhC,uGAAuG;IACvG,MAAM,EAAE,CACN,GAAG,EAAE,cAAc,CAAC,MAAM,EAAE,OAAO,EAAE,aAAa,EAAE,KAAK,CAAC,KACvD,OAAO,CAAC,eAAe,CAAC,OAAO,CAAC,CAAC,GAAG,eAAe,CAAC,OAAO,CAAC,CAAA;IAEjE,iDAAiD;IACjD,GAAG,EAAE,CACH,MAAM,EAAE,OAAO,EACf,GAAG,EAAE,cAAc,CAAC,MAAM,EAAE,OAAO,EAAE,aAAa,EAAE,KAAK,CAAC,KACvD,OAAO,CAAC,aAAa,CAAC,GAAG,aAAa,CAAA;IAE3C,yEAAyE;IACzE,UAAU,CAAC,EAAE,CACX,GAAG,EAAE,cAAc,CAAC,MAAM,EAAE,OAAO,EAAE,aAAa,EAAE,KAAK,CAAC,KACvD,OAAO,CAAC,YAAY,CAAC,GAAG,YAAY,CAAA;IAEzC,0DAA0D;IAC1D,MAAM,CAAC,EAAE,CAAC,IAAI,EAAE,WAAW,CAAC,MAAM,EAAE,OAAO,EAAE,aAAa,EAAE,KAAK,CAAC,KAAK,OAAO,CAAC,IAAI,CAAC,GAAG,IAAI,CAAA;IAE3F,kFAAkF;IAClF,YAAY,CAAC,EAAE,mBAAmB,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;IAEnD,yEAAyE;IACzE,KAAK,CAAC,EAAE,UAAU,CAAA;IAClB,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,SAAS,CAAC,EAAE,MAAM,CAAA;IAClB,SAAS,CAAC,EAAE,MAAM,CAAA;CACnB;AAOD,wBAAsB,mBAAmB,CACvC,MAAM,EACN,OAAO,EACP,aAAa,EACb,KAAK,SAAS,iBAAiB,GAAG,iBAAiB,EAEnD,MAAM,EAAE,oBAAoB,CAAC,MAAM,EAAE,OAAO,EAAE,aAAa,EAAE,KAAK,CAAC,GAClE,OAAO,CAAC,gBAAgB,CAAC,MAAM,EAAE,OAAO,EAAE,aAAa,EAAE,KAAK,CAAC,CAAC,CAgmBlE;AAED,wBAAgB,gBAAgB,CAAC,MAAM,EAAE,OAAO,EAC9C,kBAAkB,EAAE,MAAM,EAC1B,OAAO,GAAE,IAAI,CAAC,mBAAmB,CAAC,MAAM,EAAE,OAAO,CAAC,EAAE,oBAAoB,CAAM,GAC7E,mBAAmB,CAAC,MAAM,EAAE,OAAO,CAAC,CAEtC;AAED,wBAAgB,oBAAoB,CAAC,MAAM,EAAE,OAAO,EAClD,kBAAkB,EAAE,MAAM,EAC1B,OAAO,GAAE,IAAI,CAAC,mBAAmB,CAAC,MAAM,EAAE,OAAO,CAAC,EAAE,oBAAoB,CAAM,GAC7E,mBAAmB,CAAC,MAAM,EAAE,OAAO,CAAC,CAEtC;AAED,wBAAgB,aAAa,CAAC,KAAK,EAAE,IAAI,CAAC,iBAAiB,EAAE,WAAW,CAAC,GAAG,iBAAiB,CAE5F;AAED,wBAAgB,cAAc,CAAC,KAAK,EAAE,IAAI,CAAC,iBAAiB,EAAE,WAAW,CAAC,GAAG,iBAAiB,CAE7F;AAoDD,wBAAgB,iBAAiB,CAAC,KAAK,EAAE,iBAAiB,EAAE,GAAG,OAAO,CAIrE"}
@@ -0,0 +1,2 @@
1
+ export {};
2
+ //# sourceMappingURL=control-runtime.test.d.ts.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"control-runtime.test.d.ts","sourceRoot":"","sources":["../src/control-runtime.test.ts"],"names":[],"mappings":""}
package/dist/control.d.ts CHANGED
@@ -1,9 +1,11 @@
1
- export { A as ActionExecutionPolicy, a as ActionPolicyDecision, C as ControlRunToRunRecordOptions, e as ProposeReviewConfig, f as ProposeReviewControlAction, g as ProposeReviewControlConfig, h as ProposeReviewControlResult, i as ProposeReviewControlState, j as ProposeReviewReport, p as RunEvidenceMetadata, s as controlRunToRunRecord, u as evaluateActionPolicy, x as runProposeReview, y as runProposeReviewAsControlLoop, z as scoreFromEvals } from './control-D6qwHXIR.js';
2
- export { c as ControlActionFailureMode, d as ControlActionOutcome, e as ControlBudget, f as ControlContext, g as ControlDecision, C as ControlEvalResult, a as ControlRunResult, h as ControlRuntimeConfig, i as ControlRuntimeError, j as ControlSeverity, b as ControlStep, k as ControlStopPolicies, S as StopDecision, l as allCriticalPassed, o as objectiveEval, r as runAgentControlLoop, s as stopOnNoProgress, m as stopOnRepeatedAction, n as subjectiveEval } from './control-runtime-Acf9CGhw.js';
3
- import './feedback-trajectory-BxY0cKfs.js';
4
- import './dataset-BbGkaN2I.js';
5
- import './errors-CzMUYo7b.js';
6
- import './emitter-C2rqGH_l.js';
7
- import './schema-m0gsnbt3.js';
8
- import './store-BcFXE6LG.js';
9
- import './run-record-e7vj1uZQ.js';
1
+ export type { ActionExecutionPolicy, ActionPolicyDecision, } from './action-policy';
2
+ export { evaluateActionPolicy } from './action-policy';
3
+ export type { ControlActionFailureMode, ControlActionOutcome, ControlBudget, ControlContext, ControlDecision, ControlEvalResult, ControlRunResult, ControlRuntimeConfig, ControlRuntimeError, ControlSeverity, ControlStep, ControlStopPolicies, StopDecision, } from './control-runtime';
4
+ export { allCriticalPassed, objectiveEval, runAgentControlLoop, stopOnNoProgress, stopOnRepeatedAction, subjectiveEval, } from './control-runtime';
5
+ export type { ProposeReviewConfig, ProposeReviewReport, } from './propose-review';
6
+ export { runProposeReview } from './propose-review';
7
+ export type { ProposeReviewControlAction, ProposeReviewControlConfig, ProposeReviewControlResult, ProposeReviewControlState, } from './propose-review-control';
8
+ export { runProposeReviewAsControlLoop } from './propose-review-control';
9
+ export type { ControlRunToRunRecordOptions, RunEvidenceMetadata, } from './run-evidence';
10
+ export { controlRunToRunRecord, scoreFromEvals, } from './run-evidence';
11
+ //# sourceMappingURL=control.d.ts.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"control.d.ts","sourceRoot":"","sources":["../src/control.ts"],"names":[],"mappings":"AAAA,YAAY,EACV,qBAAqB,EACrB,oBAAoB,GACrB,MAAM,iBAAiB,CAAA;AACxB,OAAO,EAAE,oBAAoB,EAAE,MAAM,iBAAiB,CAAA;AACtD,YAAY,EACV,wBAAwB,EACxB,oBAAoB,EACpB,aAAa,EACb,cAAc,EACd,eAAe,EACf,iBAAiB,EACjB,gBAAgB,EAChB,oBAAoB,EACpB,mBAAmB,EACnB,eAAe,EACf,WAAW,EACX,mBAAmB,EACnB,YAAY,GACb,MAAM,mBAAmB,CAAA;AAC1B,OAAO,EACL,iBAAiB,EACjB,aAAa,EACb,mBAAmB,EACnB,gBAAgB,EAChB,oBAAoB,EACpB,cAAc,GACf,MAAM,mBAAmB,CAAA;AAC1B,YAAY,EACV,mBAAmB,EACnB,mBAAmB,GACpB,MAAM,kBAAkB,CAAA;AACzB,OAAO,EAAE,gBAAgB,EAAE,MAAM,kBAAkB,CAAA;AACnD,YAAY,EACV,0BAA0B,EAC1B,0BAA0B,EAC1B,0BAA0B,EAC1B,yBAAyB,GAC1B,MAAM,0BAA0B,CAAA;AACjC,OAAO,EAAE,6BAA6B,EAAE,MAAM,0BAA0B,CAAA;AACxE,YAAY,EACV,4BAA4B,EAC5B,mBAAmB,GACpB,MAAM,gBAAgB,CAAA;AACvB,OAAO,EACL,qBAAqB,EACrB,cAAc,GACf,MAAM,gBAAgB,CAAA"}
package/dist/control.js CHANGED
@@ -4,7 +4,7 @@ import {
4
4
  runProposeReview,
5
5
  runProposeReviewAsControlLoop,
6
6
  scoreFromEvals
7
- } from "./chunk-E4GH6USR.js";
7
+ } from "./chunk-NZEQVRH5.js";
8
8
  import {
9
9
  allCriticalPassed,
10
10
  objectiveEval,
@@ -13,7 +13,7 @@ import {
13
13
  stopOnRepeatedAction,
14
14
  subjectiveEval
15
15
  } from "./chunk-YEHAEDUD.js";
16
- import "./chunk-KWRRMR3J.js";
16
+ import "./chunk-LO6IOIJ2.js";
17
17
  import "./chunk-TVVP3ZZQ.js";
18
18
  import "./chunk-VSMTAMNK.js";
19
19
  import "./chunk-3BFEG2F6.js";
@@ -0,0 +1,29 @@
1
+ import type { CompletionCriterion, DriverState } from './types';
2
+ /**
3
+ * ConvergenceTracker — tracks completion percentage over turns.
4
+ *
5
+ * Produces convergence curves showing how quickly the agent reaches
6
+ * completion criteria.
7
+ */
8
+ export declare class ConvergenceTracker {
9
+ private criteria;
10
+ private history;
11
+ constructor(criteria: CompletionCriterion[]);
12
+ /** Evaluate criteria against current state, record result */
13
+ record(turn: number, state: DriverState): {
14
+ completionPercent: number;
15
+ complete: boolean;
16
+ criteriaStatus: Record<string, boolean | number>;
17
+ };
18
+ /** Get convergence curve */
19
+ getCurve(): number[];
20
+ /** Get full history with per-criterion status */
21
+ getHistory(): {
22
+ turn: number;
23
+ completionPercent: number;
24
+ criteriaStatus: Record<string, boolean | number>;
25
+ }[];
26
+ /** Find the turn where completion first reached 100% (or null) */
27
+ getTurnToCompletion(): number | null;
28
+ }
29
+ //# sourceMappingURL=convergence.d.ts.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"convergence.d.ts","sourceRoot":"","sources":["../src/convergence.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,mBAAmB,EAAE,WAAW,EAAE,MAAM,SAAS,CAAA;AAE/D;;;;;GAKG;AACH,qBAAa,kBAAkB;IAC7B,OAAO,CAAC,QAAQ,CAAuB;IACvC,OAAO,CAAC,OAAO,CAIP;gBAEI,QAAQ,EAAE,mBAAmB,EAAE;IAI3C,6DAA6D;IAC7D,MAAM,CACJ,IAAI,EAAE,MAAM,EACZ,KAAK,EAAE,WAAW,GACjB;QACD,iBAAiB,EAAE,MAAM,CAAA;QACzB,QAAQ,EAAE,OAAO,CAAA;QACjB,cAAc,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,GAAG,MAAM,CAAC,CAAA;KACjD;IA4BD,4BAA4B;IAC5B,QAAQ,IAAI,MAAM,EAAE;IAIpB,iDAAiD;IACjD,UAAU;cAnDF,MAAM;2BACO,MAAM;wBACT,MAAM,CAAC,MAAM,EAAE,OAAO,GAAG,MAAM,CAAC;;IAqDlD,kEAAkE;IAClE,mBAAmB,IAAI,MAAM,GAAG,IAAI;CAIrC"}
@@ -17,8 +17,8 @@
17
17
  */
18
18
  /** Logical role of an LLM call. Free-form union — consumers add their own
19
19
  * channels; the rollup keys on whatever string is supplied. */
20
- type CostChannel = 'agent' | 'judge' | 'verifier' | 'analyst' | 'driver' | (string & {});
21
- interface CostUsage {
20
+ export type CostChannel = 'agent' | 'judge' | 'verifier' | 'analyst' | 'driver' | (string & {});
21
+ export interface CostUsage {
22
22
  inputTokens: number;
23
23
  outputTokens: number;
24
24
  cachedTokens?: number;
@@ -30,8 +30,8 @@ interface CostUsage {
30
30
  * means any cost computed is `costUnknown` and the 0 must not aggregate as a
31
31
  * measured cost.
32
32
  */
33
- declare function modelPriceKey(model: string): string | null;
34
- interface CostResult {
33
+ export declare function modelPriceKey(model: string): string | null;
34
+ export interface CostResult {
35
35
  costUsd: number;
36
36
  /** True when `model` has no pricing — the 0 is "not priced", NOT "free". */
37
37
  costUnknown: boolean;
@@ -43,8 +43,8 @@ interface CostResult {
43
43
  * model's input rate when present (no separate cache-discount table — callers
44
44
  * that need provider-specific cache pricing supply `actualCostUsd` upstream).
45
45
  */
46
- declare function costForUsage(model: string, usage: CostUsage): CostResult;
47
- interface CostLedgerEntry extends CostUsage {
46
+ export declare function costForUsage(model: string, usage: CostUsage): CostResult;
47
+ export interface CostLedgerEntry extends CostUsage {
48
48
  model: string;
49
49
  channel: CostChannel;
50
50
  costUsd: number;
@@ -55,7 +55,7 @@ interface CostLedgerEntry extends CostUsage {
55
55
  tags?: Record<string, string>;
56
56
  timestamp: number;
57
57
  }
58
- interface ChannelRollup {
58
+ export interface ChannelRollup {
59
59
  channel: CostChannel;
60
60
  calls: number;
61
61
  inputTokens: number;
@@ -65,7 +65,7 @@ interface ChannelRollup {
65
65
  /** Calls whose model was unpriced (their costUsd is 0-but-unknown). */
66
66
  unpricedCalls: number;
67
67
  }
68
- interface CostLedgerSummary {
68
+ export interface CostLedgerSummary {
69
69
  totalCalls: number;
70
70
  inputTokens: number;
71
71
  outputTokens: number;
@@ -86,7 +86,7 @@ interface CostLedgerSummary {
86
86
  * answers "dollars per finished task", the metric every optimizer's
87
87
  * quality-vs-cost tradeoff needs.
88
88
  */
89
- declare class CostLedger {
89
+ export declare class CostLedger {
90
90
  private readonly entries;
91
91
  private completedTasks;
92
92
  /**
@@ -109,5 +109,4 @@ declare class CostLedger {
109
109
  /** Total spend divided by completed tasks; null when nothing completed. */
110
110
  costPerCompletedTask(): number | null;
111
111
  }
112
-
113
- export { CostLedger as C, type CostChannel as a, type ChannelRollup as b, type CostLedgerEntry as c, type CostLedgerSummary as d, type CostResult as e, type CostUsage as f, costForUsage as g, modelPriceKey as m };
112
+ //# sourceMappingURL=cost-ledger.d.ts.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"cost-ledger.d.ts","sourceRoot":"","sources":["../src/cost-ledger.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAKH;gEACgE;AAChE,MAAM,MAAM,WAAW,GAAG,OAAO,GAAG,OAAO,GAAG,UAAU,GAAG,SAAS,GAAG,QAAQ,GAAG,CAAC,MAAM,GAAG,EAAE,CAAC,CAAA;AAE/F,MAAM,WAAW,SAAS;IACxB,WAAW,EAAE,MAAM,CAAA;IACnB,YAAY,EAAE,MAAM,CAAA;IACpB,YAAY,CAAC,EAAE,MAAM,CAAA;CACtB;AAED;;;;;;GAMG;AACH,wBAAgB,aAAa,CAAC,KAAK,EAAE,MAAM,GAAG,MAAM,GAAG,IAAI,CAE1D;AAED,MAAM,WAAW,UAAU;IACzB,OAAO,EAAE,MAAM,CAAA;IACf,4EAA4E;IAC5E,WAAW,EAAE,OAAO,CAAA;CACrB;AAED;;;;;;GAMG;AACH,wBAAgB,YAAY,CAAC,KAAK,EAAE,MAAM,EAAE,KAAK,EAAE,SAAS,GAAG,UAAU,CAQxE;AAED,MAAM,WAAW,eAAgB,SAAQ,SAAS;IAChD,KAAK,EAAE,MAAM,CAAA;IACb,OAAO,EAAE,WAAW,CAAA;IACpB,OAAO,EAAE,MAAM,CAAA;IACf,WAAW,EAAE,OAAO,CAAA;IACpB,4DAA4D;IAC5D,aAAa,CAAC,EAAE,MAAM,CAAA;IACtB,0DAA0D;IAC1D,IAAI,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;IAC7B,SAAS,EAAE,MAAM,CAAA;CAClB;AAED,MAAM,WAAW,aAAa;IAC5B,OAAO,EAAE,WAAW,CAAA;IACpB,KAAK,EAAE,MAAM,CAAA;IACb,WAAW,EAAE,MAAM,CAAA;IACnB,YAAY,EAAE,MAAM,CAAA;IACpB,YAAY,EAAE,MAAM,CAAA;IACpB,OAAO,EAAE,MAAM,CAAA;IACf,uEAAuE;IACvE,aAAa,EAAE,MAAM,CAAA;CACtB;AAED,MAAM,WAAW,iBAAiB;IAChC,UAAU,EAAE,MAAM,CAAA;IAClB,WAAW,EAAE,MAAM,CAAA;IACnB,YAAY,EAAE,MAAM,CAAA;IACpB,YAAY,EAAE,MAAM,CAAA;IACpB,YAAY,EAAE,MAAM,CAAA;IACpB,qDAAqD;IACrD,SAAS,EAAE,aAAa,EAAE,CAAA;IAC1B;2DACuD;IACvD,cAAc,EAAE,MAAM,EAAE,CAAA;IACxB,4EAA4E;IAC5E,WAAW,EAAE,OAAO,CAAA;CACrB;AAED;;;;;;GAMG;AACH,qBAAa,UAAU;IACrB,OAAO,CAAC,QAAQ,CAAC,OAAO,CAAwB;IAChD,OAAO,CAAC,cAAc,CAAI;IAE1B;;;;OAIG;IACH,MAAM,CAAC,KAAK,EAAE;QACZ,KAAK,EAAE,MAAM,CAAA;QACb,OAAO,EAAE,WAAW,CAAA;QACpB,KAAK,EAAE,SAAS,CAAA;QAChB,aAAa,CAAC,EAAE,MAAM,CAAA;QACtB,IAAI,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;QAC7B,SAAS,CAAC,EAAE,MAAM,CAAA;KACnB,GAAG,eAAe;IAqBnB,+EAA+E;IAC/E,aAAa,CAAC,KAAK,SAAI,GAAG,IAAI;IAS9B,IAAI,IAAI,eAAe,EAAE;IAIzB,OAAO,IAAI,iBAAiB;IA0C5B,2EAA2E;IAC3E,oBAAoB,IAAI,MAAM,GAAG,IAAI;CAItC"}
@@ -0,0 +1,2 @@
1
+ export {};
2
+ //# sourceMappingURL=cost-ledger.test.d.ts.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"cost-ledger.test.d.ts","sourceRoot":"","sources":["../src/cost-ledger.test.ts"],"names":[],"mappings":""}
@@ -0,0 +1,43 @@
1
+ /**
2
+ * Program cost report — a thin projection over `CostLedger.summary()` that
3
+ * adds the per-model rollup the summary lacks, plus `attachCostToReport`, the
4
+ * one way every artifact (capsules, campaign results, diagnose reports) gets
5
+ * its cost stamp.
6
+ *
7
+ * Honesty contract carried through from the ledger: `total.unknownEntries`
8
+ * and `perModel[].unpriced` surface the costUnknown axis — a $0 from an
9
+ * unpriced model is a lower bound, never a measured zero.
10
+ */
11
+ import type { ChannelRollup, CostLedger } from './cost-ledger';
12
+ export interface ModelCostRollup {
13
+ model: string;
14
+ usd: number;
15
+ entries: number;
16
+ /** ≥1 entry for this model was costUnknown — `usd` is a lower bound. An
17
+ * `actualCostUsd` override clears the flag for that entry (the dollars are
18
+ * observed, even when the model has no pricing). */
19
+ unpriced: boolean;
20
+ }
21
+ export interface CostReport {
22
+ /** Per-channel breakdown — `CostLedgerSummary.byChannel` verbatim. */
23
+ perChannel: ChannelRollup[];
24
+ total: {
25
+ usd: number;
26
+ /** Entries whose cost was unknown — non-zero means `usd` is a lower bound. */
27
+ unknownEntries: number;
28
+ };
29
+ /** Per-model spend, sorted by model id. */
30
+ perModel: ModelCostRollup[];
31
+ }
32
+ /** Project a ledger into the program cost report. Pure — no I/O, no clock. */
33
+ export declare function costReport(ledger: CostLedger): CostReport;
34
+ /**
35
+ * Stamp a report-shaped object with its cost projection under the `cost` key.
36
+ * Generic so capsules, campaign results, and diagnose reports all stamp the
37
+ * same way. Throws when the report already carries a `cost` key — silently
38
+ * overwriting an existing stamp would corrupt the artifact's provenance.
39
+ */
40
+ export declare function attachCostToReport<R extends object>(report: R, ledger: CostLedger): R & {
41
+ cost: CostReport;
42
+ };
43
+ //# sourceMappingURL=cost-report.d.ts.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"cost-report.d.ts","sourceRoot":"","sources":["../src/cost-report.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AAEH,OAAO,KAAK,EAAE,aAAa,EAAE,UAAU,EAAE,MAAM,eAAe,CAAA;AAG9D,MAAM,WAAW,eAAe;IAC9B,KAAK,EAAE,MAAM,CAAA;IACb,GAAG,EAAE,MAAM,CAAA;IACX,OAAO,EAAE,MAAM,CAAA;IACf;;yDAEqD;IACrD,QAAQ,EAAE,OAAO,CAAA;CAClB;AAED,MAAM,WAAW,UAAU;IACzB,sEAAsE;IACtE,UAAU,EAAE,aAAa,EAAE,CAAA;IAC3B,KAAK,EAAE;QACL,GAAG,EAAE,MAAM,CAAA;QACX,8EAA8E;QAC9E,cAAc,EAAE,MAAM,CAAA;KACvB,CAAA;IACD,2CAA2C;IAC3C,QAAQ,EAAE,eAAe,EAAE,CAAA;CAC5B;AAED,8EAA8E;AAC9E,wBAAgB,UAAU,CAAC,MAAM,EAAE,UAAU,GAAG,UAAU,CAuBzD;AAED;;;;;GAKG;AACH,wBAAgB,kBAAkB,CAAC,CAAC,SAAS,MAAM,EACjD,MAAM,EAAE,CAAC,EACT,MAAM,EAAE,UAAU,GACjB,CAAC,GAAG;IAAE,IAAI,EAAE,UAAU,CAAA;CAAE,CAO1B"}
@@ -0,0 +1,2 @@
1
+ export {};
2
+ //# sourceMappingURL=cost-report.test.d.ts.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"cost-report.test.d.ts","sourceRoot":"","sources":["../src/cost-report.test.ts"],"names":[],"mappings":""}
@@ -0,0 +1,76 @@
1
+ /**
2
+ * Cost tracker — token + USD accounting per scenario and per run.
3
+ *
4
+ * Adapted from generic usage-event accounting. Every
5
+ * optimizer needs to know "is the quality gain worth the cost delta?",
6
+ * and every dashboard needs dollars-per-completed-task. MODEL_PRICING
7
+ * from metrics.ts stays authoritative for estimate math; this module
8
+ * adds the aggregation + per-scenario roll-up that was duplicated
9
+ * across 4 verticals.
10
+ */
11
+ export interface TokenSpec {
12
+ inputTokens: number;
13
+ outputTokens: number;
14
+ cachedTokens?: number;
15
+ reasoningTokens?: number;
16
+ }
17
+ export interface CostEntry extends TokenSpec {
18
+ scenarioId: string;
19
+ model: string;
20
+ /** Override estimate with an observed cost (e.g. from provider response). */
21
+ actualCostUsd?: number;
22
+ timestamp: number;
23
+ /** Free-form tags (variant id, round #, etc.). */
24
+ tags?: Record<string, string>;
25
+ }
26
+ export interface ScenarioCost {
27
+ scenarioId: string;
28
+ entries: CostEntry[];
29
+ totalInputTokens: number;
30
+ totalOutputTokens: number;
31
+ totalCachedTokens: number;
32
+ totalCostUsd: number;
33
+ /** Pass flag — set by consumer via markOutcome; used for cost-per-completed-task. */
34
+ completed?: boolean;
35
+ }
36
+ export declare class CostTracker {
37
+ private byScenario;
38
+ record(entry: Omit<CostEntry, 'timestamp'> & {
39
+ timestamp?: number;
40
+ }): CostEntry;
41
+ markOutcome(scenarioId: string, completed: boolean): void;
42
+ /**
43
+ * Convenience: record + markOutcome in one call from a
44
+ * `{ usage, verdict }`-shaped response (starter-foundry's
45
+ * `invokeMetaJudge` returns this shape; consumers that wrap any
46
+ * judge/critic can follow the same convention).
47
+ *
48
+ * `usage.model` must be present in `MODEL_PRICING` for cost math to
49
+ * populate; otherwise totalCostUsd stays at 0 for the entry but
50
+ * tokens still aggregate.
51
+ */
52
+ recordVerdict(verdict: {
53
+ usage?: {
54
+ inputTokens: number;
55
+ outputTokens: number;
56
+ model: string;
57
+ cachedTokens?: number;
58
+ reasoningTokens?: number;
59
+ };
60
+ verdict?: 'pass' | 'fail' | 'borderline' | string;
61
+ }, scenarioId: string, tags?: Record<string, string>): CostEntry | null;
62
+ get(scenarioId: string): ScenarioCost | undefined;
63
+ list(): ScenarioCost[];
64
+ summary(): CostSummary;
65
+ }
66
+ export interface CostSummary {
67
+ scenarioCount: number;
68
+ completedCount: number;
69
+ totalInputTokens: number;
70
+ totalOutputTokens: number;
71
+ totalCostUsd: number;
72
+ avgCostPerScenarioUsd: number;
73
+ /** Total USD / completed scenarios — null when nothing completed. */
74
+ costPerCompletedTaskUsd: number | null;
75
+ }
76
+ //# sourceMappingURL=cost-tracker.d.ts.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"cost-tracker.d.ts","sourceRoot":"","sources":["../src/cost-tracker.ts"],"names":[],"mappings":"AAAA;;;;;;;;;GASG;AAIH,MAAM,WAAW,SAAS;IACxB,WAAW,EAAE,MAAM,CAAA;IACnB,YAAY,EAAE,MAAM,CAAA;IACpB,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB,eAAe,CAAC,EAAE,MAAM,CAAA;CACzB;AAED,MAAM,WAAW,SAAU,SAAQ,SAAS;IAC1C,UAAU,EAAE,MAAM,CAAA;IAClB,KAAK,EAAE,MAAM,CAAA;IACb,6EAA6E;IAC7E,aAAa,CAAC,EAAE,MAAM,CAAA;IACtB,SAAS,EAAE,MAAM,CAAA;IACjB,kDAAkD;IAClD,IAAI,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;CAC9B;AAED,MAAM,WAAW,YAAY;IAC3B,UAAU,EAAE,MAAM,CAAA;IAClB,OAAO,EAAE,SAAS,EAAE,CAAA;IACpB,gBAAgB,EAAE,MAAM,CAAA;IACxB,iBAAiB,EAAE,MAAM,CAAA;IACzB,iBAAiB,EAAE,MAAM,CAAA;IACzB,YAAY,EAAE,MAAM,CAAA;IACpB,qFAAqF;IACrF,SAAS,CAAC,EAAE,OAAO,CAAA;CACpB;AAED,qBAAa,WAAW;IACtB,OAAO,CAAC,UAAU,CAAkC;IAEpD,MAAM,CAAC,KAAK,EAAE,IAAI,CAAC,SAAS,EAAE,WAAW,CAAC,GAAG;QAAE,SAAS,CAAC,EAAE,MAAM,CAAA;KAAE,GAAG,SAAS;IAwB/E,WAAW,CAAC,UAAU,EAAE,MAAM,EAAE,SAAS,EAAE,OAAO,GAAG,IAAI;IAMzD;;;;;;;;;OASG;IACH,aAAa,CACX,OAAO,EAAE;QACP,KAAK,CAAC,EAAE;YACN,WAAW,EAAE,MAAM,CAAA;YACnB,YAAY,EAAE,MAAM,CAAA;YACpB,KAAK,EAAE,MAAM,CAAA;YACb,YAAY,CAAC,EAAE,MAAM,CAAA;YACrB,eAAe,CAAC,EAAE,MAAM,CAAA;SACzB,CAAA;QACD,OAAO,CAAC,EAAE,MAAM,GAAG,MAAM,GAAG,YAAY,GAAG,MAAM,CAAA;KAClD,EACD,UAAU,EAAE,MAAM,EAClB,IAAI,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,GAC5B,SAAS,GAAG,IAAI;IAenB,GAAG,CAAC,UAAU,EAAE,MAAM,GAAG,YAAY,GAAG,SAAS;IAIjD,IAAI,IAAI,YAAY,EAAE;IAItB,OAAO,IAAI,WAAW;CAiBvB;AAED,MAAM,WAAW,WAAW;IAC1B,aAAa,EAAE,MAAM,CAAA;IACrB,cAAc,EAAE,MAAM,CAAA;IACtB,gBAAgB,EAAE,MAAM,CAAA;IACxB,iBAAiB,EAAE,MAAM,CAAA;IACzB,YAAY,EAAE,MAAM,CAAA;IACpB,qBAAqB,EAAE,MAAM,CAAA;IAC7B,qEAAqE;IACrE,uBAAuB,EAAE,MAAM,GAAG,IAAI,CAAA;CACvC"}
@@ -1,7 +1,3 @@
1
- import { T as TraceEmitter } from './emitter-C2rqGH_l.js';
2
- import { T as TraceStore } from './store-BcFXE6LG.js';
3
- import { T as Trajectory, a as TrajectoryStep } from './trajectory-2TkpSEVh.js';
4
-
5
1
  /**
6
2
  * Counterfactual replay — "what would have happened if we'd changed
7
3
  * exactly one thing at turn N?"
@@ -15,8 +11,11 @@ import { T as Trajectory, a as TrajectoryStep } from './trajectory-2TkpSEVh.js';
15
11
  * `parentRunId = originalRunId`, so downstream diff + correlation
16
12
  * pipelines see them natively.
17
13
  */
18
-
19
- type CounterfactualMutation = {
14
+ import { TraceEmitter } from './trace/emitter';
15
+ import type { Span } from './trace/schema';
16
+ import type { TraceStore } from './trace/store';
17
+ import { type Trajectory, type TrajectoryStep } from './trajectory';
18
+ export type CounterfactualMutation = {
20
19
  kind: 'swap-model';
21
20
  at: number;
22
21
  newModel: string;
@@ -37,7 +36,7 @@ type CounterfactualMutation = {
37
36
  describe: string;
38
37
  apply: (step: TrajectoryStep) => TrajectoryStep;
39
38
  };
40
- interface CounterfactualContext {
39
+ export interface CounterfactualContext {
41
40
  originalRunId: string;
42
41
  originalTrajectory: Trajectory;
43
42
  /** Steps up to (but not including) the mutation point — the prefix the
@@ -49,7 +48,7 @@ interface CounterfactualContext {
49
48
  * re-emit it or continue from there). */
50
49
  mutatedStep: TrajectoryStep;
51
50
  }
52
- interface CounterfactualResult {
51
+ export interface CounterfactualResult {
53
52
  counterfactualRunId: string;
54
53
  originalRunId: string;
55
54
  mutation: CounterfactualMutation;
@@ -60,7 +59,7 @@ interface CounterfactualResult {
60
59
  deltaScore: number | null;
61
60
  };
62
61
  }
63
- interface CounterfactualRunner {
62
+ export interface CounterfactualRunner {
64
63
  /**
65
64
  * Execute the agent from `ctx.prefix` with the mutation applied.
66
65
  * MUST emit spans into the provided emitter so they become part of
@@ -68,18 +67,18 @@ interface CounterfactualRunner {
68
67
  */
69
68
  executeFrom: (ctx: CounterfactualContext, emitter: TraceEmitter) => Promise<void>;
70
69
  }
71
- declare function runCounterfactual(store: TraceStore, originalRunId: string, mutation: CounterfactualMutation, runner: CounterfactualRunner): Promise<CounterfactualResult>;
70
+ export declare function runCounterfactual(store: TraceStore, originalRunId: string, mutation: CounterfactualMutation, runner: CounterfactualRunner): Promise<CounterfactualResult>;
72
71
  /**
73
72
  * Aggregate a batch of counterfactuals into a simple attribution table:
74
73
  * which mutation kinds move outcomes most? (Useful when you run a grid
75
74
  * over the same trajectory — swap-model at every llm span, swap-tool
76
75
  * at every tool span — and want a ranked summary.)
77
76
  */
78
- declare function attributeCounterfactuals(results: CounterfactualResult[]): Array<{
77
+ export declare function attributeCounterfactuals(results: CounterfactualResult[]): Array<{
79
78
  mutationKind: CounterfactualMutation['kind'];
80
79
  n: number;
81
80
  meanAbsDelta: number;
82
81
  meanSignedDelta: number;
83
82
  }>;
84
-
85
- export { type CounterfactualMutation as C, attributeCounterfactuals as a, type CounterfactualRunner as b, type CounterfactualContext as c, type CounterfactualResult as d, runCounterfactual as r };
83
+ export type { Span };
84
+ //# sourceMappingURL=counterfactual.d.ts.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"counterfactual.d.ts","sourceRoot":"","sources":["../src/counterfactual.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;GAYG;AAGH,OAAO,EAAE,YAAY,EAAE,MAAM,iBAAiB,CAAA;AAC9C,OAAO,KAAK,EAAW,IAAI,EAAY,MAAM,gBAAgB,CAAA;AAC7D,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,eAAe,CAAA;AAC/C,OAAO,EAAmB,KAAK,UAAU,EAAE,KAAK,cAAc,EAAE,MAAM,cAAc,CAAA;AAEpF,MAAM,MAAM,sBAAsB,GAC9B;IAAE,IAAI,EAAE,YAAY,CAAC;IAAC,EAAE,EAAE,MAAM,CAAC;IAAC,QAAQ,EAAE,MAAM,CAAA;CAAE,GACpD;IAAE,IAAI,EAAE,kBAAkB,CAAC;IAAC,EAAE,EAAE,MAAM,CAAC;IAAC,SAAS,EAAE,OAAO,CAAA;CAAE,GAC5D;IAAE,IAAI,EAAE,gBAAgB,CAAC;IAAC,EAAE,EAAE,MAAM,CAAA;CAAE,GACtC;IAAE,IAAI,EAAE,uBAAuB,CAAC;IAAC,EAAE,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,MAAM,CAAA;CAAE,GAC9D;IACE,IAAI,EAAE,QAAQ,CAAA;IACd,EAAE,EAAE,MAAM,CAAA;IACV,QAAQ,EAAE,MAAM,CAAA;IAChB,KAAK,EAAE,CAAC,IAAI,EAAE,cAAc,KAAK,cAAc,CAAA;CAChD,CAAA;AAEL,MAAM,WAAW,qBAAqB;IACpC,aAAa,EAAE,MAAM,CAAA;IACrB,kBAAkB,EAAE,UAAU,CAAA;IAC9B;0EACsE;IACtE,MAAM,EAAE,cAAc,EAAE,CAAA;IACxB,QAAQ,EAAE,sBAAsB,CAAA;IAChC;;8CAE0C;IAC1C,WAAW,EAAE,cAAc,CAAA;CAC5B;AAED,MAAM,WAAW,oBAAoB;IACnC,mBAAmB,EAAE,MAAM,CAAA;IAC3B,aAAa,EAAE,MAAM,CAAA;IACrB,QAAQ,EAAE,sBAAsB,CAAA;IAChC,gEAAgE;IAChE,KAAK,EAAE;QACL,oBAAoB,EAAE,MAAM,GAAG,IAAI,CAAA;QACnC,0BAA0B,EAAE,MAAM,GAAG,IAAI,CAAA;QACzC,UAAU,EAAE,MAAM,GAAG,IAAI,CAAA;KAC1B,CAAA;CACF;AAED,MAAM,WAAW,oBAAoB;IACnC;;;;OAIG;IACH,WAAW,EAAE,CAAC,GAAG,EAAE,qBAAqB,EAAE,OAAO,EAAE,YAAY,KAAK,OAAO,CAAC,IAAI,CAAC,CAAA;CAClF;AAED,wBAAsB,iBAAiB,CACrC,KAAK,EAAE,UAAU,EACjB,aAAa,EAAE,MAAM,EACrB,QAAQ,EAAE,sBAAsB,EAChC,MAAM,EAAE,oBAAoB,GAC3B,OAAO,CAAC,oBAAoB,CAAC,CA6C/B;AA0BD;;;;;GAKG;AACH,wBAAgB,wBAAwB,CAAC,OAAO,EAAE,oBAAoB,EAAE,GAAG,KAAK,CAAC;IAC/E,YAAY,EAAE,sBAAsB,CAAC,MAAM,CAAC,CAAA;IAC5C,CAAC,EAAE,MAAM,CAAA;IACT,YAAY,EAAE,MAAM,CAAA;IACpB,eAAe,EAAE,MAAM,CAAA;CACxB,CAAC,CA4BD;AAGD,YAAY,EAAE,IAAI,EAAE,CAAA"}
@@ -0,0 +1,56 @@
1
+ /**
2
+ * Full cross-trace diff — align two trajectories step-by-step, report
3
+ * per-step score deltas, attribute a variant's total outcome lead to
4
+ * specific turns.
5
+ *
6
+ * 0.5 shipped `firstDivergenceView` (finds the first differing step).
7
+ * This does the heavier work: full alignment via LCS, per-step
8
+ * contribution to score delta using PRM verdicts when available,
9
+ * fallback to structural heuristics (latency, token count, tool
10
+ * outcome) otherwise.
11
+ */
12
+ import type { TraceStore } from './trace/store';
13
+ import { type TrajectoryStep } from './trajectory';
14
+ export type AlignmentOp = {
15
+ op: 'match';
16
+ a: TrajectoryStep;
17
+ b: TrajectoryStep;
18
+ } | {
19
+ op: 'insert';
20
+ b: TrajectoryStep;
21
+ } | {
22
+ op: 'delete';
23
+ a: TrajectoryStep;
24
+ } | {
25
+ op: 'replace';
26
+ a: TrajectoryStep;
27
+ b: TrajectoryStep;
28
+ };
29
+ export interface StepAttribution {
30
+ op: AlignmentOp;
31
+ /** Difference in PRM score (or null when not scored by a matching judge). */
32
+ prmDelta: number | null;
33
+ /** Difference in latency (endedAt - startedAt). */
34
+ latencyDeltaMs: number | null;
35
+ /** Difference in token count (LLM spans). */
36
+ tokenDelta: number | null;
37
+ /** Reason this step is / isn't considered a contributor to the outcome delta. */
38
+ note: string;
39
+ }
40
+ export interface CrossTraceDiff {
41
+ runA: string;
42
+ runB: string;
43
+ alignment: AlignmentOp[];
44
+ attributions: StepAttribution[];
45
+ /** Total score delta (B - A). */
46
+ totalScoreDelta: number | null;
47
+ /** Sum of PRM deltas across matched/replaced steps. Close to
48
+ * `totalScoreDelta` when PRM covers the trajectory; gap indicates
49
+ * unmodeled variance. */
50
+ prmDeltaSum: number;
51
+ }
52
+ export interface CrossTraceDiffOptions {
53
+ stepEquals?: (a: TrajectoryStep, b: TrajectoryStep) => boolean;
54
+ }
55
+ export declare function crossTraceDiff(store: TraceStore, runA: string, runB: string, options?: CrossTraceDiffOptions): Promise<CrossTraceDiff>;
56
+ //# sourceMappingURL=cross-trace-diff.d.ts.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"cross-trace-diff.d.ts","sourceRoot":"","sources":["../src/cross-trace-diff.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAIH,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,eAAe,CAAA;AAC/C,OAAO,EAAmB,KAAK,cAAc,EAAE,MAAM,cAAc,CAAA;AAEnE,MAAM,MAAM,WAAW,GACnB;IAAE,EAAE,EAAE,OAAO,CAAC;IAAC,CAAC,EAAE,cAAc,CAAC;IAAC,CAAC,EAAE,cAAc,CAAA;CAAE,GACrD;IAAE,EAAE,EAAE,QAAQ,CAAC;IAAC,CAAC,EAAE,cAAc,CAAA;CAAE,GACnC;IAAE,EAAE,EAAE,QAAQ,CAAC;IAAC,CAAC,EAAE,cAAc,CAAA;CAAE,GACnC;IAAE,EAAE,EAAE,SAAS,CAAC;IAAC,CAAC,EAAE,cAAc,CAAC;IAAC,CAAC,EAAE,cAAc,CAAA;CAAE,CAAA;AAE3D,MAAM,WAAW,eAAe;IAC9B,EAAE,EAAE,WAAW,CAAA;IACf,6EAA6E;IAC7E,QAAQ,EAAE,MAAM,GAAG,IAAI,CAAA;IACvB,mDAAmD;IACnD,cAAc,EAAE,MAAM,GAAG,IAAI,CAAA;IAC7B,6CAA6C;IAC7C,UAAU,EAAE,MAAM,GAAG,IAAI,CAAA;IACzB,iFAAiF;IACjF,IAAI,EAAE,MAAM,CAAA;CACb;AAED,MAAM,WAAW,cAAc;IAC7B,IAAI,EAAE,MAAM,CAAA;IACZ,IAAI,EAAE,MAAM,CAAA;IACZ,SAAS,EAAE,WAAW,EAAE,CAAA;IACxB,YAAY,EAAE,eAAe,EAAE,CAAA;IAC/B,iCAAiC;IACjC,eAAe,EAAE,MAAM,GAAG,IAAI,CAAA;IAC9B;;8BAE0B;IAC1B,WAAW,EAAE,MAAM,CAAA;CACpB;AAED,MAAM,WAAW,qBAAqB;IACpC,UAAU,CAAC,EAAE,CAAC,CAAC,EAAE,cAAc,EAAE,CAAC,EAAE,cAAc,KAAK,OAAO,CAAA;CAC/D;AAED,wBAAsB,cAAc,CAClC,KAAK,EAAE,UAAU,EACjB,IAAI,EAAE,MAAM,EACZ,IAAI,EAAE,MAAM,EACZ,OAAO,GAAE,qBAA0B,GAClC,OAAO,CAAC,cAAc,CAAC,CAwBzB"}