agent-hitch 0.2.6 → 0.2.8

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 (562) hide show
  1. package/README.md +30 -3
  2. package/README.zh-CN.md +14 -2
  3. package/dist/bin/hitch.js +25 -3
  4. package/dist/bin/hitch.js.map +1 -1
  5. package/dist/scripts/canary-benchmark.js +54 -0
  6. package/dist/scripts/canary-benchmark.js.map +1 -0
  7. package/dist/scripts/canary-buildkit-secrets.js +166 -0
  8. package/dist/scripts/canary-buildkit-secrets.js.map +1 -0
  9. package/dist/scripts/canary-collision-domains.js +77 -0
  10. package/dist/scripts/canary-collision-domains.js.map +1 -0
  11. package/dist/scripts/canary-docker-image-observation.js +181 -0
  12. package/dist/scripts/canary-docker-image-observation.js.map +1 -0
  13. package/dist/scripts/canary-eval-scheduler-throughput.js +47 -0
  14. package/dist/scripts/canary-eval-scheduler-throughput.js.map +1 -0
  15. package/dist/scripts/canary-gpu-hardware.js +95 -0
  16. package/dist/scripts/canary-gpu-hardware.js.map +1 -0
  17. package/dist/scripts/canary-harbor-load.js +365 -0
  18. package/dist/scripts/canary-harbor-load.js.map +1 -0
  19. package/dist/scripts/canary-packaged-harness.js +75 -0
  20. package/dist/scripts/canary-packaged-harness.js.map +1 -0
  21. package/dist/scripts/canary-resource-load.js +188 -0
  22. package/dist/scripts/canary-resource-load.js.map +1 -0
  23. package/dist/scripts/check-architecture.js +10 -5
  24. package/dist/scripts/check-architecture.js.map +1 -1
  25. package/dist/scripts/docker-reaper-crash-child.js +21 -0
  26. package/dist/scripts/docker-reaper-crash-child.js.map +1 -0
  27. package/dist/src/adapters/catalog.js +3 -1
  28. package/dist/src/adapters/catalog.js.map +1 -1
  29. package/dist/src/adapters/discovery.js +1 -1
  30. package/dist/src/adapters/discovery.js.map +1 -1
  31. package/dist/src/adapters/providers/claude.js +6 -0
  32. package/dist/src/adapters/providers/claude.js.map +1 -1
  33. package/dist/src/adapters/providers/codex-auth.js +29 -0
  34. package/dist/src/adapters/providers/codex-auth.js.map +1 -0
  35. package/dist/src/adapters/providers/codex.js +29 -6
  36. package/dist/src/adapters/providers/codex.js.map +1 -1
  37. package/dist/src/adapters/providers/deepseek.js +6 -0
  38. package/dist/src/adapters/providers/deepseek.js.map +1 -1
  39. package/dist/src/adapters/providers/model-call.js +26 -0
  40. package/dist/src/adapters/providers/model-call.js.map +1 -0
  41. package/dist/src/adapters/providers/opencode.js +6 -0
  42. package/dist/src/adapters/providers/opencode.js.map +1 -1
  43. package/dist/src/adapters/providers/pi.js +6 -0
  44. package/dist/src/adapters/providers/pi.js.map +1 -1
  45. package/dist/src/artifacts/index.js +1 -0
  46. package/dist/src/artifacts/index.js.map +1 -1
  47. package/dist/src/artifacts/preparer.js +7 -2
  48. package/dist/src/artifacts/preparer.js.map +1 -1
  49. package/dist/src/backends/harbor/agent-budget.js +31 -0
  50. package/dist/src/backends/harbor/agent-budget.js.map +1 -0
  51. package/dist/src/backends/harbor/backend.js +59 -107
  52. package/dist/src/backends/harbor/backend.js.map +1 -1
  53. package/dist/src/backends/harbor/environment-config.js +92 -0
  54. package/dist/src/backends/harbor/environment-config.js.map +1 -0
  55. package/dist/src/backends/harbor/index.js +6 -1
  56. package/dist/src/backends/harbor/index.js.map +1 -1
  57. package/dist/src/backends/harbor/model-proxy-config.js +23 -0
  58. package/dist/src/backends/harbor/model-proxy-config.js.map +1 -0
  59. package/dist/src/backends/harbor/process.js +177 -0
  60. package/dist/src/backends/harbor/process.js.map +1 -0
  61. package/dist/src/backends/harbor/regrade.js +77 -0
  62. package/dist/src/backends/harbor/regrade.js.map +1 -0
  63. package/dist/src/backends/harbor/runtime-toolchain.js +30 -0
  64. package/dist/src/backends/harbor/runtime-toolchain.js.map +1 -0
  65. package/dist/src/backends/harbor/supervisor.js +74 -0
  66. package/dist/src/backends/harbor/supervisor.js.map +1 -0
  67. package/dist/src/backends/harbor/task-resources.js +206 -0
  68. package/dist/src/backends/harbor/task-resources.js.map +1 -0
  69. package/dist/src/backends/harbor/tools.js +6 -29
  70. package/dist/src/backends/harbor/tools.js.map +1 -1
  71. package/dist/src/backends/index.js +5 -1
  72. package/dist/src/backends/index.js.map +1 -1
  73. package/dist/src/benchmarks/index.js +3 -0
  74. package/dist/src/benchmarks/index.js.map +1 -0
  75. package/dist/src/benchmarks/loader.js +271 -0
  76. package/dist/src/benchmarks/loader.js.map +1 -0
  77. package/dist/src/benchmarks/metrics.js +16 -0
  78. package/dist/src/benchmarks/metrics.js.map +1 -0
  79. package/dist/src/benchmarks/toml.js +13 -0
  80. package/dist/src/benchmarks/toml.js.map +1 -0
  81. package/dist/src/benchmarks/validation.js +242 -0
  82. package/dist/src/benchmarks/validation.js.map +1 -0
  83. package/dist/src/cli/arguments.js +6 -4
  84. package/dist/src/cli/arguments.js.map +1 -1
  85. package/dist/src/cli/commands/benchmark.js +22 -0
  86. package/dist/src/cli/commands/benchmark.js.map +1 -0
  87. package/dist/src/cli/commands/capabilities.js +18 -0
  88. package/dist/src/cli/commands/capabilities.js.map +1 -0
  89. package/dist/src/cli/commands/daemon.js +135 -8
  90. package/dist/src/cli/commands/daemon.js.map +1 -1
  91. package/dist/src/cli/commands/eval.js +228 -6
  92. package/dist/src/cli/commands/eval.js.map +1 -1
  93. package/dist/src/cli/commands/images.js +40 -0
  94. package/dist/src/cli/commands/images.js.map +1 -0
  95. package/dist/src/cli/commands/run.js +42 -17
  96. package/dist/src/cli/commands/run.js.map +1 -1
  97. package/dist/src/cli/commands/runs.js +25 -3
  98. package/dist/src/cli/commands/runs.js.map +1 -1
  99. package/dist/src/cli/commands/trajectory.js +96 -5
  100. package/dist/src/cli/commands/trajectory.js.map +1 -1
  101. package/dist/src/cli/commands/verifier.js +36 -0
  102. package/dist/src/cli/commands/verifier.js.map +1 -0
  103. package/dist/src/cli/commands/worker.js +79 -0
  104. package/dist/src/cli/commands/worker.js.map +1 -0
  105. package/dist/src/cli/main.js +10 -0
  106. package/dist/src/cli/main.js.map +1 -1
  107. package/dist/src/cli/output.js +84 -4
  108. package/dist/src/cli/output.js.map +1 -1
  109. package/dist/src/control-plane/build-admission.js +69 -0
  110. package/dist/src/control-plane/build-admission.js.map +1 -0
  111. package/dist/src/control-plane/build-records.js +6 -0
  112. package/dist/src/control-plane/build-records.js.map +1 -0
  113. package/dist/src/control-plane/collisions.js +40 -0
  114. package/dist/src/control-plane/collisions.js.map +1 -0
  115. package/dist/src/control-plane/eval-control-state.js +22 -0
  116. package/dist/src/control-plane/eval-control-state.js.map +1 -0
  117. package/dist/src/control-plane/eval-control-work.js +44 -0
  118. package/dist/src/control-plane/eval-control-work.js.map +1 -0
  119. package/dist/src/control-plane/eval-image-resolution.js +52 -0
  120. package/dist/src/control-plane/eval-image-resolution.js.map +1 -0
  121. package/dist/src/control-plane/eval-image-services.js +30 -0
  122. package/dist/src/control-plane/eval-image-services.js.map +1 -0
  123. package/dist/src/control-plane/eval-records.js +252 -0
  124. package/dist/src/control-plane/eval-records.js.map +1 -0
  125. package/dist/src/control-plane/eval-recovery.js +181 -0
  126. package/dist/src/control-plane/eval-recovery.js.map +1 -0
  127. package/dist/src/control-plane/eval-scheduler.js +450 -0
  128. package/dist/src/control-plane/eval-scheduler.js.map +1 -0
  129. package/dist/src/control-plane/index.js +20 -0
  130. package/dist/src/control-plane/index.js.map +1 -0
  131. package/dist/src/control-plane/local-worker.js +49 -0
  132. package/dist/src/control-plane/local-worker.js.map +1 -0
  133. package/dist/src/control-plane/model-capture-planning.js +12 -0
  134. package/dist/src/control-plane/model-capture-planning.js.map +1 -0
  135. package/dist/src/control-plane/remote-result-transport.js +175 -0
  136. package/dist/src/control-plane/remote-result-transport.js.map +1 -0
  137. package/dist/src/control-plane/remote-work-coordinator.js +333 -0
  138. package/dist/src/control-plane/remote-work-coordinator.js.map +1 -0
  139. package/dist/src/control-plane/remote-work-inputs.js +266 -0
  140. package/dist/src/control-plane/remote-work-inputs.js.map +1 -0
  141. package/dist/src/control-plane/remote-work-item.js +62 -0
  142. package/dist/src/control-plane/remote-work-item.js.map +1 -0
  143. package/dist/src/control-plane/remote-work-recovery.js +220 -0
  144. package/dist/src/control-plane/remote-work-recovery.js.map +1 -0
  145. package/dist/src/control-plane/remote-worker-artifacts.js +135 -0
  146. package/dist/src/control-plane/remote-worker-artifacts.js.map +1 -0
  147. package/dist/src/control-plane/remote-worker-client.js +202 -0
  148. package/dist/src/control-plane/remote-worker-client.js.map +1 -0
  149. package/dist/src/control-plane/remote-worker-configuration.js +34 -0
  150. package/dist/src/control-plane/remote-worker-configuration.js.map +1 -0
  151. package/dist/src/control-plane/remote-worker-credentials.js +58 -0
  152. package/dist/src/control-plane/remote-worker-credentials.js.map +1 -0
  153. package/dist/src/control-plane/remote-worker-protocol.js +489 -0
  154. package/dist/src/control-plane/remote-worker-protocol.js.map +1 -0
  155. package/dist/src/control-plane/remote-worker-runner.js +214 -0
  156. package/dist/src/control-plane/remote-worker-runner.js.map +1 -0
  157. package/dist/src/control-plane/remote-workers.js +377 -0
  158. package/dist/src/control-plane/remote-workers.js.map +1 -0
  159. package/dist/src/control-plane/rerun-scheduler.js +448 -0
  160. package/dist/src/control-plane/rerun-scheduler.js.map +1 -0
  161. package/dist/src/control-plane/rerun-submission.js +62 -0
  162. package/dist/src/control-plane/rerun-submission.js.map +1 -0
  163. package/dist/src/control-plane/resources.js +182 -0
  164. package/dist/src/control-plane/resources.js.map +1 -0
  165. package/dist/src/control-plane/scheduler-eval-entry.js +32 -0
  166. package/dist/src/control-plane/scheduler-eval-entry.js.map +1 -0
  167. package/dist/src/control-plane/synthetic-result.js +21 -0
  168. package/dist/src/control-plane/synthetic-result.js.map +1 -0
  169. package/dist/src/control-plane/work-admission.js +23 -0
  170. package/dist/src/control-plane/work-admission.js.map +1 -0
  171. package/dist/src/control-plane/work-dispatcher.js +237 -0
  172. package/dist/src/control-plane/work-dispatcher.js.map +1 -0
  173. package/dist/src/controller-runtime/hash.js +19 -3
  174. package/dist/src/controller-runtime/hash.js.map +1 -1
  175. package/dist/src/controller-runtime/index.js +1 -1
  176. package/dist/src/controller-runtime/index.js.map +1 -1
  177. package/dist/src/controller-runtime/store.js +15 -6
  178. package/dist/src/controller-runtime/store.js.map +1 -1
  179. package/dist/src/daemon/auth.js +5 -0
  180. package/dist/src/daemon/auth.js.map +1 -1
  181. package/dist/src/daemon/client.js.map +1 -1
  182. package/dist/src/daemon/event-stream.js +46 -0
  183. package/dist/src/daemon/event-stream.js.map +1 -0
  184. package/dist/src/daemon/health.js +66 -0
  185. package/dist/src/daemon/health.js.map +1 -0
  186. package/dist/src/daemon/launcher.js +28 -1
  187. package/dist/src/daemon/launcher.js.map +1 -1
  188. package/dist/src/daemon/logging.js +23 -0
  189. package/dist/src/daemon/logging.js.map +1 -0
  190. package/dist/src/daemon/scheduler.js +26 -5
  191. package/dist/src/daemon/scheduler.js.map +1 -1
  192. package/dist/src/daemon/server.js +255 -98
  193. package/dist/src/daemon/server.js.map +1 -1
  194. package/dist/src/daemon/telemetry.js +163 -0
  195. package/dist/src/daemon/telemetry.js.map +1 -0
  196. package/dist/src/daemon/worker-routes.js +219 -0
  197. package/dist/src/daemon/worker-routes.js.map +1 -0
  198. package/dist/src/domain/benchmarks.js +2 -0
  199. package/dist/src/domain/benchmarks.js.map +1 -0
  200. package/dist/src/domain/bundles.js +2 -0
  201. package/dist/src/domain/bundles.js.map +1 -0
  202. package/dist/src/domain/control-plane.js +2 -0
  203. package/dist/src/domain/control-plane.js.map +1 -0
  204. package/dist/src/domain/execution-evidence.js +2 -0
  205. package/dist/src/domain/execution-evidence.js.map +1 -0
  206. package/dist/src/domain/execution-plan.js +2 -0
  207. package/dist/src/domain/execution-plan.js.map +1 -0
  208. package/dist/src/domain/images.js +2 -0
  209. package/dist/src/domain/images.js.map +1 -0
  210. package/dist/src/domain/index.js +13 -0
  211. package/dist/src/domain/index.js.map +1 -1
  212. package/dist/src/domain/interactions.js +2 -0
  213. package/dist/src/domain/interactions.js.map +1 -0
  214. package/dist/src/domain/providers.js +2 -0
  215. package/dist/src/domain/providers.js.map +1 -0
  216. package/dist/src/domain/resources.js +2 -0
  217. package/dist/src/domain/resources.js.map +1 -0
  218. package/dist/src/domain/runs.js.map +1 -1
  219. package/dist/src/domain/validation.js +11 -4
  220. package/dist/src/domain/validation.js.map +1 -1
  221. package/dist/src/domain/verifier-evidence-validation.js +214 -0
  222. package/dist/src/domain/verifier-evidence-validation.js.map +1 -0
  223. package/dist/src/domain/verifier-evidence.js +3 -0
  224. package/dist/src/domain/verifier-evidence.js.map +1 -0
  225. package/dist/src/domain/verifier-score-contract.js +232 -0
  226. package/dist/src/domain/verifier-score-contract.js.map +1 -0
  227. package/dist/src/domain/workers.js +2 -0
  228. package/dist/src/domain/workers.js.map +1 -0
  229. package/dist/src/evals/benchmark-adapter-manifest.js +219 -0
  230. package/dist/src/evals/benchmark-adapter-manifest.js.map +1 -0
  231. package/dist/src/evals/benchmark-candidate.js +36 -0
  232. package/dist/src/evals/benchmark-candidate.js.map +1 -0
  233. package/dist/src/evals/benchmark-run.js +166 -0
  234. package/dist/src/evals/benchmark-run.js.map +1 -0
  235. package/dist/src/evals/collect-only-rerun.js +183 -0
  236. package/dist/src/evals/collect-only-rerun.js.map +1 -0
  237. package/dist/src/evals/directory.js +30 -0
  238. package/dist/src/evals/directory.js.map +1 -0
  239. package/dist/src/evals/docker-engine-stats.js +89 -0
  240. package/dist/src/evals/docker-engine-stats.js.map +1 -0
  241. package/dist/src/evals/docker-ownership.js +51 -0
  242. package/dist/src/evals/docker-ownership.js.map +1 -0
  243. package/dist/src/evals/docker-reaper.js +144 -0
  244. package/dist/src/evals/docker-reaper.js.map +1 -0
  245. package/dist/src/evals/docker-resource-observer.js +207 -0
  246. package/dist/src/evals/docker-resource-observer.js.map +1 -0
  247. package/dist/src/evals/duration-estimator.js +126 -0
  248. package/dist/src/evals/duration-estimator.js.map +1 -0
  249. package/dist/src/evals/environment-image-planning.js +164 -0
  250. package/dist/src/evals/environment-image-planning.js.map +1 -0
  251. package/dist/src/evals/eval-artifact-planning.js +56 -0
  252. package/dist/src/evals/eval-artifact-planning.js.map +1 -0
  253. package/dist/src/evals/eval-finalization.js +10 -0
  254. package/dist/src/evals/eval-finalization.js.map +1 -0
  255. package/dist/src/evals/eval-lifecycle-events.js +29 -0
  256. package/dist/src/evals/eval-lifecycle-events.js.map +1 -0
  257. package/dist/src/evals/eval-logical-plan.js +51 -0
  258. package/dist/src/evals/eval-logical-plan.js.map +1 -0
  259. package/dist/src/evals/eval-result-builder.js +81 -0
  260. package/dist/src/evals/eval-result-builder.js.map +1 -0
  261. package/dist/src/evals/events.js +53 -5
  262. package/dist/src/evals/events.js.map +1 -1
  263. package/dist/src/evals/evolution-baseline.js +191 -0
  264. package/dist/src/evals/evolution-baseline.js.map +1 -0
  265. package/dist/src/evals/execution-evidence.js +116 -0
  266. package/dist/src/evals/execution-evidence.js.map +1 -0
  267. package/dist/src/evals/execution-leases.js +342 -0
  268. package/dist/src/evals/execution-leases.js.map +1 -0
  269. package/dist/src/evals/execution-plan-artifacts.js +72 -0
  270. package/dist/src/evals/execution-plan-artifacts.js.map +1 -0
  271. package/dist/src/evals/execution-plan-images.js +69 -0
  272. package/dist/src/evals/execution-plan-images.js.map +1 -0
  273. package/dist/src/evals/execution-plan-resources.js +174 -0
  274. package/dist/src/evals/execution-plan-resources.js.map +1 -0
  275. package/dist/src/evals/execution-plan.js +451 -0
  276. package/dist/src/evals/execution-plan.js.map +1 -0
  277. package/dist/src/evals/failure-classifier.js +61 -0
  278. package/dist/src/evals/failure-classifier.js.map +1 -0
  279. package/dist/src/evals/harbor-artifact-builder-image.js +99 -0
  280. package/dist/src/evals/harbor-artifact-builder-image.js.map +1 -0
  281. package/dist/src/evals/harbor-artifact-builder.js +358 -0
  282. package/dist/src/evals/harbor-artifact-builder.js.map +1 -0
  283. package/dist/src/evals/harbor-bridge-error.js +58 -3
  284. package/dist/src/evals/harbor-bridge-error.js.map +1 -1
  285. package/dist/src/evals/harbor-node-runtime.js +146 -0
  286. package/dist/src/evals/harbor-node-runtime.js.map +1 -0
  287. package/dist/src/evals/harbor-phase-timings.js +60 -0
  288. package/dist/src/evals/harbor-phase-timings.js.map +1 -0
  289. package/dist/src/evals/index.js +40 -2
  290. package/dist/src/evals/index.js.map +1 -1
  291. package/dist/src/evals/infrastructure-retry.js +208 -84
  292. package/dist/src/evals/infrastructure-retry.js.map +1 -1
  293. package/dist/src/evals/interaction-capture-import.js +49 -0
  294. package/dist/src/evals/interaction-capture-import.js.map +1 -0
  295. package/dist/src/evals/local-docker-provider.js +372 -0
  296. package/dist/src/evals/local-docker-provider.js.map +1 -0
  297. package/dist/src/evals/local-eval-planning.js +80 -0
  298. package/dist/src/evals/local-eval-planning.js.map +1 -0
  299. package/dist/src/evals/model-capture-plan.js +22 -0
  300. package/dist/src/evals/model-capture-plan.js.map +1 -0
  301. package/dist/src/evals/model-capture-runtime.js +114 -0
  302. package/dist/src/evals/model-capture-runtime.js.map +1 -0
  303. package/dist/src/evals/model-proxy-runtime-state.js +53 -0
  304. package/dist/src/evals/model-proxy-runtime-state.js.map +1 -0
  305. package/dist/src/evals/native-phase-evidence.js +440 -0
  306. package/dist/src/evals/native-phase-evidence.js.map +1 -0
  307. package/dist/src/evals/physical-retry-work.js +31 -0
  308. package/dist/src/evals/physical-retry-work.js.map +1 -0
  309. package/dist/src/evals/planned-execution-support.js +66 -0
  310. package/dist/src/evals/planned-execution-support.js.map +1 -0
  311. package/dist/src/evals/planned-execution.js +469 -0
  312. package/dist/src/evals/planned-execution.js.map +1 -0
  313. package/dist/src/evals/planned-progress-publisher.js +62 -0
  314. package/dist/src/evals/planned-progress-publisher.js.map +1 -0
  315. package/dist/src/evals/planned-remote-execution.js +40 -0
  316. package/dist/src/evals/planned-remote-execution.js.map +1 -0
  317. package/dist/src/evals/planned-retry-execution.js +174 -0
  318. package/dist/src/evals/planned-retry-execution.js.map +1 -0
  319. package/dist/src/evals/planned-retry-lifecycle.js +188 -0
  320. package/dist/src/evals/planned-retry-lifecycle.js.map +1 -0
  321. package/dist/src/evals/preparation-rerun.js +225 -0
  322. package/dist/src/evals/preparation-rerun.js.map +1 -0
  323. package/dist/src/evals/prepared-harness.js +76 -0
  324. package/dist/src/evals/prepared-harness.js.map +1 -0
  325. package/dist/src/evals/progress.js +52 -12
  326. package/dist/src/evals/progress.js.map +1 -1
  327. package/dist/src/evals/records.js +20 -3
  328. package/dist/src/evals/records.js.map +1 -1
  329. package/dist/src/evals/recovery.js +242 -0
  330. package/dist/src/evals/recovery.js.map +1 -0
  331. package/dist/src/evals/regrade-evidence.js +77 -0
  332. package/dist/src/evals/regrade-evidence.js.map +1 -0
  333. package/dist/src/evals/remote-infrastructure-retry.js +165 -0
  334. package/dist/src/evals/remote-infrastructure-retry.js.map +1 -0
  335. package/dist/src/evals/request.js +21 -3
  336. package/dist/src/evals/request.js.map +1 -1
  337. package/dist/src/evals/rerun-inputs.js +108 -0
  338. package/dist/src/evals/rerun-inputs.js.map +1 -0
  339. package/dist/src/evals/rerun-slots.js +4 -3
  340. package/dist/src/evals/rerun-slots.js.map +1 -1
  341. package/dist/src/evals/rerun-types.js +45 -0
  342. package/dist/src/evals/rerun-types.js.map +1 -0
  343. package/dist/src/evals/rerun.js +133 -88
  344. package/dist/src/evals/rerun.js.map +1 -1
  345. package/dist/src/evals/result-helpers.js +78 -0
  346. package/dist/src/evals/result-helpers.js.map +1 -0
  347. package/dist/src/evals/resume-state.js +36 -0
  348. package/dist/src/evals/resume-state.js.map +1 -0
  349. package/dist/src/evals/retry-backoff.js +20 -0
  350. package/dist/src/evals/retry-backoff.js.map +1 -0
  351. package/dist/src/evals/retry-state.js +195 -0
  352. package/dist/src/evals/retry-state.js.map +1 -0
  353. package/dist/src/evals/scheduler-metrics.js +95 -0
  354. package/dist/src/evals/scheduler-metrics.js.map +1 -0
  355. package/dist/src/evals/scheduler-trace-replay.js +107 -0
  356. package/dist/src/evals/scheduler-trace-replay.js.map +1 -0
  357. package/dist/src/evals/service-types.js +2 -0
  358. package/dist/src/evals/service-types.js.map +1 -0
  359. package/dist/src/evals/service.js +331 -347
  360. package/dist/src/evals/service.js.map +1 -1
  361. package/dist/src/evals/task-resources.js +209 -0
  362. package/dist/src/evals/task-resources.js.map +1 -0
  363. package/dist/src/evals/trial-environment-evidence.js +87 -0
  364. package/dist/src/evals/trial-environment-evidence.js.map +1 -0
  365. package/dist/src/evals/trial-execution-evidence.js +13 -0
  366. package/dist/src/evals/trial-execution-evidence.js.map +1 -0
  367. package/dist/src/evals/trial-import-identity.js +31 -0
  368. package/dist/src/evals/trial-import-identity.js.map +1 -0
  369. package/dist/src/evals/trial-import.js +183 -188
  370. package/dist/src/evals/trial-import.js.map +1 -1
  371. package/dist/src/evals/trial-publication-recovery.js +144 -0
  372. package/dist/src/evals/trial-publication-recovery.js.map +1 -0
  373. package/dist/src/evals/trial-publication.js +25 -0
  374. package/dist/src/evals/trial-publication.js.map +1 -0
  375. package/dist/src/evals/trial-reference-validation.js +49 -0
  376. package/dist/src/evals/trial-reference-validation.js.map +1 -0
  377. package/dist/src/evals/verifier-artifacts.js +317 -0
  378. package/dist/src/evals/verifier-artifacts.js.map +1 -0
  379. package/dist/src/evals/verifier-eligibility.js +25 -0
  380. package/dist/src/evals/verifier-eligibility.js.map +1 -0
  381. package/dist/src/evals/verifier-only-rerun.js +210 -0
  382. package/dist/src/evals/verifier-only-rerun.js.map +1 -0
  383. package/dist/src/evals/verifier-runtime.js +61 -0
  384. package/dist/src/evals/verifier-runtime.js.map +1 -0
  385. package/dist/src/evals/verifier-score-artifacts.js +119 -0
  386. package/dist/src/evals/verifier-score-artifacts.js.map +1 -0
  387. package/dist/src/evals/work-item-artifacts.js +23 -0
  388. package/dist/src/evals/work-item-artifacts.js.map +1 -0
  389. package/dist/src/foundation/config.js +17 -0
  390. package/dist/src/foundation/config.js.map +1 -1
  391. package/dist/src/foundation/contained-file.js +85 -0
  392. package/dist/src/foundation/contained-file.js.map +1 -0
  393. package/dist/src/foundation/credential-redaction.js +162 -0
  394. package/dist/src/foundation/credential-redaction.js.map +1 -0
  395. package/dist/src/foundation/environment-image-references.js +59 -0
  396. package/dist/src/foundation/environment-image-references.js.map +1 -0
  397. package/dist/src/foundation/executable.js +45 -5
  398. package/dist/src/foundation/executable.js.map +1 -1
  399. package/dist/src/foundation/fs.js +2 -1
  400. package/dist/src/foundation/fs.js.map +1 -1
  401. package/dist/src/foundation/index.js +6 -2
  402. package/dist/src/foundation/index.js.map +1 -1
  403. package/dist/src/foundation/process-identity.js +69 -0
  404. package/dist/src/foundation/process-identity.js.map +1 -0
  405. package/dist/src/images/context.js +47 -0
  406. package/dist/src/images/context.js.map +1 -0
  407. package/dist/src/images/docker-buildkit.js +137 -0
  408. package/dist/src/images/docker-buildkit.js.map +1 -0
  409. package/dist/src/images/dockerfile.js +67 -0
  410. package/dist/src/images/dockerfile.js.map +1 -0
  411. package/dist/src/images/gc.js +261 -0
  412. package/dist/src/images/gc.js.map +1 -0
  413. package/dist/src/images/index.js +10 -0
  414. package/dist/src/images/index.js.map +1 -0
  415. package/dist/src/images/manifest.js +109 -0
  416. package/dist/src/images/manifest.js.map +1 -0
  417. package/dist/src/images/ownership.js +5 -0
  418. package/dist/src/images/ownership.js.map +1 -0
  419. package/dist/src/images/records.js +32 -0
  420. package/dist/src/images/records.js.map +1 -0
  421. package/dist/src/images/registry.js +154 -0
  422. package/dist/src/images/registry.js.map +1 -0
  423. package/dist/src/images/service.js +224 -0
  424. package/dist/src/images/service.js.map +1 -0
  425. package/dist/src/model-access/capture.js +255 -0
  426. package/dist/src/model-access/capture.js.map +1 -0
  427. package/dist/src/model-access/index.js +5 -0
  428. package/dist/src/model-access/index.js.map +1 -0
  429. package/dist/src/model-access/policy.js +49 -0
  430. package/dist/src/model-access/policy.js.map +1 -0
  431. package/dist/src/model-access/proxy.js +336 -0
  432. package/dist/src/model-access/proxy.js.map +1 -0
  433. package/dist/src/model-access/records.js +167 -0
  434. package/dist/src/model-access/records.js.map +1 -0
  435. package/dist/src/runs/adapter-process.js +12 -0
  436. package/dist/src/runs/adapter-process.js.map +1 -0
  437. package/dist/src/runs/bundle.js +450 -0
  438. package/dist/src/runs/bundle.js.map +1 -0
  439. package/dist/src/runs/events.js +12 -1
  440. package/dist/src/runs/events.js.map +1 -1
  441. package/dist/src/runs/executor.js +46 -45
  442. package/dist/src/runs/executor.js.map +1 -1
  443. package/dist/src/runs/finalizer.js +17 -0
  444. package/dist/src/runs/finalizer.js.map +1 -1
  445. package/dist/src/runs/index.js +8 -1
  446. package/dist/src/runs/index.js.map +1 -1
  447. package/dist/src/runs/manifest.js +5 -5
  448. package/dist/src/runs/manifest.js.map +1 -1
  449. package/dist/src/runs/phase-bundle.js +86 -0
  450. package/dist/src/runs/phase-bundle.js.map +1 -0
  451. package/dist/src/runs/phase-cancellation.js +107 -0
  452. package/dist/src/runs/phase-cancellation.js.map +1 -0
  453. package/dist/src/runs/phase-group.js +113 -0
  454. package/dist/src/runs/phase-group.js.map +1 -0
  455. package/dist/src/runs/query.js +1 -1
  456. package/dist/src/runs/query.js.map +1 -1
  457. package/dist/src/runs/records.js +3 -0
  458. package/dist/src/runs/records.js.map +1 -1
  459. package/dist/src/runs/request.js +10 -3
  460. package/dist/src/runs/request.js.map +1 -1
  461. package/dist/src/runs/training-candidate.js +161 -0
  462. package/dist/src/runs/training-candidate.js.map +1 -0
  463. package/dist/src/runs/verifier-evidence-redaction.js +61 -0
  464. package/dist/src/runs/verifier-evidence-redaction.js.map +1 -0
  465. package/dist/src/runs/verifier-evidence.js +459 -0
  466. package/dist/src/runs/verifier-evidence.js.map +1 -0
  467. package/dist/src/runs/verifier-structured-evidence.js +83 -0
  468. package/dist/src/runs/verifier-structured-evidence.js.map +1 -0
  469. package/dist/src/trajectories/analysis.js +288 -0
  470. package/dist/src/trajectories/analysis.js.map +1 -0
  471. package/dist/src/trajectories/chunk-projection.js +148 -0
  472. package/dist/src/trajectories/chunk-projection.js.map +1 -0
  473. package/dist/src/trajectories/content-projection.js +389 -0
  474. package/dist/src/trajectories/content-projection.js.map +1 -0
  475. package/dist/src/trajectories/dsh-chunk-contract.js +176 -0
  476. package/dist/src/trajectories/dsh-chunk-contract.js.map +1 -0
  477. package/dist/src/trajectories/dsh-contract.js +284 -0
  478. package/dist/src/trajectories/dsh-contract.js.map +1 -0
  479. package/dist/src/trajectories/events-chunk-drill.js +95 -0
  480. package/dist/src/trajectories/events-chunk-drill.js.map +1 -0
  481. package/dist/src/trajectories/events-page.js +367 -0
  482. package/dist/src/trajectories/events-page.js.map +1 -0
  483. package/dist/src/trajectories/index.js +3 -0
  484. package/dist/src/trajectories/index.js.map +1 -1
  485. package/dist/src/trajectories/provider-capture.js +15 -24
  486. package/dist/src/trajectories/provider-capture.js.map +1 -1
  487. package/dist/src/trajectories/providers/deepseek.js +3 -3
  488. package/dist/src/trajectories/providers/deepseek.js.map +1 -1
  489. package/dist/src/trajectories/request-attempt.js +47 -0
  490. package/dist/src/trajectories/request-attempt.js.map +1 -0
  491. package/dist/src/trajectories/stream-reader.js +311 -0
  492. package/dist/src/trajectories/stream-reader.js.map +1 -0
  493. package/dist/src/trajectories/surface-fold.js +126 -0
  494. package/dist/src/trajectories/surface-fold.js.map +1 -0
  495. package/dist/src/workers/index.js +3 -0
  496. package/dist/src/workers/index.js.map +1 -0
  497. package/dist/src/workers/remote-harbor-work-spec.js +97 -0
  498. package/dist/src/workers/remote-harbor-work-spec.js.map +1 -0
  499. package/dist/src/workers/remote-harbor-worker.js +223 -0
  500. package/dist/src/workers/remote-harbor-worker.js.map +1 -0
  501. package/docs/schemas/benchmark-hook-request-v1.schema.json +65 -0
  502. package/docs/schemas/benchmark-hook-response-v1.schema.json +75 -0
  503. package/docs/schemas/benchmark-hook-v1.schema.json +33 -0
  504. package/docs/schemas/benchmark-lock-v1.schema.json +351 -0
  505. package/docs/schemas/benchmark-metric-v1.schema.json +37 -0
  506. package/docs/schemas/benchmark-package-v1.schema.json +167 -0
  507. package/docs/schemas/benchmark-phase-group.schema.json +41 -0
  508. package/docs/schemas/benchmark-profile-v1.schema.json +141 -0
  509. package/docs/schemas/benchmark-task-v1.schema.json +384 -0
  510. package/docs/schemas/benchmark-tool-result-v1.schema.json +30 -0
  511. package/docs/schemas/environment-build-index.schema.json +13 -0
  512. package/docs/schemas/environment-build-record.schema.json +25 -0
  513. package/docs/schemas/environment-image-manifest.schema.json +62 -0
  514. package/docs/schemas/error.schema.json +22 -0
  515. package/docs/schemas/eval-control.schema.json +36 -0
  516. package/docs/schemas/eval-execution-plan.schema.json +216 -0
  517. package/docs/schemas/eval-progress.schema.json +1 -15
  518. package/docs/schemas/eval-rerun-result.schema.json +58 -0
  519. package/docs/schemas/eval-rerun-state.schema.json +61 -0
  520. package/docs/schemas/eval-rerun-submission.schema.json +60 -0
  521. package/docs/schemas/eval-result.schema.json +32 -13
  522. package/docs/schemas/eval-submission-input.schema.json +14 -0
  523. package/docs/schemas/eval-submission.schema.json +68 -0
  524. package/docs/schemas/eval-trial-publication.schema.json +15 -0
  525. package/docs/schemas/eval-trial-reference.schema.json +129 -0
  526. package/docs/schemas/execution-evidence.schema.json +96 -0
  527. package/docs/schemas/execution-lease.schema.json +50 -0
  528. package/docs/schemas/execution-provider-status.schema.json +65 -0
  529. package/docs/schemas/execution-worker.schema.json +53 -0
  530. package/docs/schemas/harbor-process-exit.schema.json +14 -0
  531. package/docs/schemas/interaction-capture-ref.schema.json +45 -0
  532. package/docs/schemas/local-provider-execution.schema.json +33 -0
  533. package/docs/schemas/model-interaction.schema.json +41 -0
  534. package/docs/schemas/regrade-assessment-reference.schema.json +11 -0
  535. package/docs/schemas/remote-result-envelope.schema.json +32 -0
  536. package/docs/schemas/remote-tree-envelope.schema.json +25 -0
  537. package/docs/schemas/remote-work-offer.schema.json +58 -0
  538. package/docs/schemas/remote-work-receipt.schema.json +50 -0
  539. package/docs/schemas/remote-worker-artifact.schema.json +17 -0
  540. package/docs/schemas/remote-worker-event.schema.json +18 -0
  541. package/docs/schemas/remote-worker-heartbeat.schema.json +37 -0
  542. package/docs/schemas/remote-worker-registration.schema.json +52 -0
  543. package/docs/schemas/result-bundle-index.schema.json +133 -0
  544. package/docs/schemas/run-context.schema.json +15 -0
  545. package/docs/schemas/run-environment-images.schema.json +39 -0
  546. package/docs/schemas/run-manifest.schema.json +4 -0
  547. package/docs/schemas/run-request.schema.json +4 -0
  548. package/docs/schemas/training-data-candidate.schema.json +46 -0
  549. package/docs/schemas/trajectory-analysis.schema.json +258 -0
  550. package/docs/schemas/trajectory-events-page.schema.json +40 -0
  551. package/docs/schemas/verifier-assessment.schema.json +77 -0
  552. package/docs/schemas/verifier-evidence.schema.json +276 -0
  553. package/integrations/harbor/hitch_benchmark.py +252 -0
  554. package/integrations/harbor/hitch_candidate_recycle.py +262 -0
  555. package/integrations/harbor/hitch_harbor_agent.py +698 -568
  556. package/integrations/harbor/hitch_harbor_environment.py +469 -0
  557. package/integrations/harbor/hitch_harbor_task_resources.py +275 -0
  558. package/integrations/harbor/hitch_harbor_verifier.py +83 -2
  559. package/integrations/harbor/hitch_phase_supervisor.py +452 -0
  560. package/integrations/harbor/hitch_tool_client.mjs +91 -0
  561. package/integrations/model-call/cli.js +57 -0
  562. package/package.json +30 -6
@@ -4,47 +4,85 @@ import { runHarborBackend } from "../backends/index.js";
4
4
  import { HitchError } from "../foundation/index.js";
5
5
  import { importEvalTrialRun, importEvalTrialRuns, TrialBundlePendingError, validateEvalTrialReferences } from "./trial-import.js";
6
6
  import { replaceInvalidEvalProgressTrial, writeEvalProgress } from "./progress.js";
7
+ import { ensurePhysicalRetryDecision, transitionRetryDecision } from "./retry-state.js";
8
+ import { physicalRetryWorkItem } from "./physical-retry-work.js";
9
+ import { classifyTrialFailure, physicalRetryAllowed } from "./failure-classifier.js";
10
+ import { retryBackoffMs } from "./retry-backoff.js";
11
+ import { harborPhaseTimingEvents } from "./harbor-phase-timings.js";
7
12
  const INFRASTRUCTURE_REASONS = new Set([
8
13
  "infrastructure_failure",
9
14
  "verifier_infrastructure_failure",
10
15
  "verifier_result_missing",
11
- ]);
12
- // Verifier failures are retried inside the original live Harbor trial by the
13
- // custom verifier wrapper. Starting another Harbor trial would execute the
14
- // candidate agent again, so only non-verifier trial infrastructure failures
15
- // are eligible for this outer retry path.
16
- const RETRYABLE_TRIAL_INFRASTRUCTURE_REASONS = new Set([
17
- "infrastructure_failure",
16
+ "provider_quota_exhausted",
17
+ "provider_auth_failed",
18
+ "provider_configuration_invalid",
19
+ "provider_rate_limited",
20
+ "provider_transport_transient",
21
+ "worker_lost_before_candidate",
22
+ "sandbox_setup_failed",
23
+ "execution_state_ambiguous",
18
24
  ]);
19
25
  export async function runInfrastructureRetries(options) {
20
26
  let progress = options.progress;
21
27
  let retryCandidates = retryableInfrastructureTrials(options.initialRefs);
22
28
  const runs = [];
23
- for (let retry = 1; retry <= options.request.infrastructure_retries && retryCandidates.length > 0; retry += 1) {
29
+ const firstRetryIndex = options.firstRetryIndex ?? 1;
30
+ if (!Number.isSafeInteger(firstRetryIndex) || firstRetryIndex < 1)
31
+ throw new TypeError("first infrastructure retry index is invalid");
32
+ for (let retry = firstRetryIndex; retry <= options.request.infrastructure_retries && retryCandidates.length > 0; retry += 1) {
24
33
  if (options.signal?.aborted)
25
34
  break;
26
35
  const retryTriggers = [...retryCandidates];
27
36
  const taskNames = [...new Set(retryCandidates.map((ref) => ref.task_id))].sort();
28
- const backoffMs = options.request.infrastructure_retry_backoff_ms * retry;
37
+ const retryWork = options.originWorkItem ? physicalRetryWorkItem(options.originWorkItem, retry, retryTriggers) : undefined;
38
+ const backoffMs = retryBackoffMs(options.request.infrastructure_retry_backoff_ms, retry, retryWork?.work_id ?? retryTriggers.map((trial) => trial.trial_id).join("\0"));
39
+ const notBefore = new Date(Date.now() + backoffMs).toISOString();
40
+ const retryDecisions = options.originWorkItem
41
+ ? await Promise.all(retryTriggers.map((trigger) => ensurePhysicalRetryDecision({
42
+ evalDirectory: options.evalDirectory, evalId: options.evalId, item: options.originWorkItem,
43
+ retryIndex: retry, trigger, notBefore,
44
+ })))
45
+ : [];
46
+ if (retryWork && retryDecisions.some((decision) => decision.retry_work_id !== retryWork.work_id)) {
47
+ throw new Error(`persisted retry work identity conflicts with execution: ${retryWork.work_id}`);
48
+ }
49
+ if (retryWork)
50
+ await options.onRetryWorkPlanned?.(retryWork.work_id);
29
51
  options.sink.emit({
30
- type: "eval.infrastructure-retry.scheduled",
52
+ type: "eval.retry.decision",
53
+ execution_kind: "physical-infrastructure-retry",
54
+ candidate_executes: true,
55
+ ...(retryWork ? { work_id: retryWork.work_id } : {}),
56
+ decision_ids: retryDecisions.map((decision) => decision.decision_id),
31
57
  attempt: options.logicalAttempt,
32
58
  retry,
33
59
  tasks: taskNames,
34
60
  backoff_ms: backoffMs,
35
61
  });
36
- if (backoffMs > 0) {
62
+ const remainingBackoffMs = Math.max(0, Date.parse(retryDecisions[0]?.not_before ?? notBefore) - Date.now());
63
+ options.onRetryBackoff?.(remainingBackoffMs);
64
+ if (remainingBackoffMs > 0) {
37
65
  if (options.signal)
38
- await delay(backoffMs, undefined, { signal: options.signal });
66
+ await delay(remainingBackoffMs, undefined, { signal: options.signal });
39
67
  else
40
- await delay(backoffMs);
68
+ await delay(remainingBackoffMs);
41
69
  }
42
70
  if (options.signal?.aborted)
43
71
  break;
44
- const backendDirectory = path.join(options.evalDirectory, "infrastructure-retries", `retry-${String(retry).padStart(4, "0")}`, options.request.attempts === 1 ? "harbor" : attemptDirectoryName(options.logicalAttempt));
72
+ if (retryWork)
73
+ options.sink.emit({
74
+ type: "eval.retry.ready", work_id: retryWork.work_id, decision_ids: retryDecisions.map((decision) => decision.decision_id),
75
+ retry, tasks: taskNames, priority: options.originWorkItem?.scheduling?.remaining_path_ms ?? 0,
76
+ });
77
+ const backendDirectory = path.join(options.backendBaseDirectory || path.join(options.evalDirectory, "infrastructure-retries"), `retry-${String(retry).padStart(4, "0")}`, options.request.attempts === 1 ? "harbor" : attemptDirectoryName(options.logicalAttempt));
45
78
  const harborJobDirectory = path.join(backendDirectory, "job");
46
79
  const retryRefs = [];
47
80
  const selectedTasks = new Set(taskNames);
81
+ let lifecycle;
82
+ const retryEnvironmentImages = () => lifecycle?.environmentImages ?? options.environmentImages;
83
+ const executionEvidence = async () => lifecycle?.captureExecutionEvidence
84
+ ? { executionEvidence: await lifecycle.captureExecutionEvidence() }
85
+ : {};
48
86
  const publish = async (ref) => {
49
87
  if (ref.attempt !== options.logicalAttempt || !selectedTasks.has(ref.task_id)) {
50
88
  throw new HitchError(`Harbor infrastructure retry returned an unselected trial: ${ref.task_id}#${ref.attempt}`, {
@@ -59,13 +97,26 @@ export async function runInfrastructureRetries(options) {
59
97
  return;
60
98
  }
61
99
  retryRefs.push(ref);
100
+ if (ref.invalid_reason === "candidate_evidence_unavailable")
101
+ options.sink.emit({
102
+ type: "eval.verifier.skipped", work_id: retryWork?.work_id, trial_id: ref.trial_id, task_id: ref.task_id,
103
+ reason: "candidate_evidence_unavailable", candidate_executes: true, verifier_executes: false,
104
+ });
62
105
  await validateEvalTrialReferences(options.root, options.evalId, [ref], {
63
106
  benchmarkId: options.request.benchmark_id,
64
107
  benchmarkRevision: options.request.benchmark_revision,
65
108
  });
66
109
  if (ref.observation_status === "valid") {
67
- progress = replaceInvalidEvalProgressTrial(progress, ref);
68
- await writeEvalProgress(options.evalDirectory, progress);
110
+ if (options.replaceProgressTrial) {
111
+ if (!lifecycle)
112
+ throw new Error("planned infrastructure retry is missing its lifecycle identity");
113
+ await options.replaceProgressTrial(ref, lifecycle.workId);
114
+ progress = options.currentProgress?.() ?? progress;
115
+ }
116
+ else {
117
+ progress = replaceInvalidEvalProgressTrial(progress, ref);
118
+ await writeEvalProgress(options.evalDirectory, progress);
119
+ }
69
120
  }
70
121
  options.sink.emit({
71
122
  type: ref.observation_status === "valid"
@@ -78,73 +129,148 @@ export async function runInfrastructureRetries(options) {
78
129
  run_id: ref.run_id,
79
130
  observation_status: ref.observation_status,
80
131
  ...(ref.invalid_reason ? { invalid_reason: ref.invalid_reason } : {}),
81
- generation: progress.generation,
132
+ generation: (options.currentProgress?.() ?? progress).generation,
82
133
  });
83
134
  };
84
- const run = await runHarborBackend({
85
- evalId: options.evalId,
86
- evalDirectory: options.evalDirectory,
87
- backendDirectory,
88
- logicalAttempt: options.logicalAttempt,
89
- taskNames,
90
- request: { ...options.request, attempts: 1 },
91
- root: options.root,
92
- resolvedRevision: options.resolvedRevision,
93
- runtimeDirectory: options.controllerRuntime.directory,
94
- runtimeId: options.controllerRuntime.runtime_id,
95
- preparedArtifact: options.preparedArtifact,
96
- ...(options.localTransport ? { localTransport: options.localTransport } : {}),
97
- env: options.env,
98
- ...(options.harborExecutable !== undefined ? { harborExecutable: options.harborExecutable } : {}),
99
- ...(options.signal ? { signal: options.signal } : {}),
100
- ...(options.trialBundleGraceMs === undefined ? {} : { trialBundleGraceMs: options.trialBundleGraceMs }),
101
- emit: (event) => options.sink.emit({ ...event, infrastructure_retry: retry, logical_attempt: options.logicalAttempt }),
102
- onTrialSettled: async (trial, context) => {
103
- try {
104
- const ref = await importEvalTrialRun({
105
- root: options.root,
106
- evalId: options.evalId,
107
- evalDirectory: options.evalDirectory,
108
- harborJobDirectory,
109
- expectedAttempt: options.logicalAttempt,
110
- request: options.request,
111
- resolvedRevision: options.resolvedRevision,
112
- benchmarkId: options.request.benchmark_id,
113
- benchmarkRevision: options.request.benchmark_revision,
114
- runtimeId: options.controllerRuntime.runtime_id,
115
- requireCompleteMarker: true,
116
- allowMissingBundleDiagnostic: context.bundleWaitExpired,
117
- }, trial, retryRefs.length, retryRefs);
118
- await publish(ref);
119
- return true;
120
- }
121
- catch (error) {
122
- if (error instanceof TrialBundlePendingError)
123
- return false;
124
- throw error;
125
- }
126
- },
127
- });
128
- const terminalRefs = await importEvalTrialRuns({
129
- root: options.root,
130
- evalId: options.evalId,
131
- evalDirectory: options.evalDirectory,
132
- harborJobDirectory,
133
- expectedAttempt: options.logicalAttempt,
134
- request: options.request,
135
- resolvedRevision: options.resolvedRevision,
136
- benchmarkId: options.request.benchmark_id,
137
- benchmarkRevision: options.request.benchmark_revision,
138
- runtimeId: options.controllerRuntime.runtime_id,
139
- rawResult: run.rawResult,
140
- }, retryRefs);
141
- for (const ref of terminalRefs)
142
- await publish(ref);
143
- if (run.rawResult !== null)
144
- assertBackendTrialSet(run.rawResult, retryRefs);
145
- runs.push({ attempt: options.logicalAttempt, retry, tasks: taskNames, triggers: retryTriggers, refs: retryRefs, run });
135
+ let run;
136
+ let nextCandidates = [];
137
+ let runRecord;
138
+ let executionStarted = false;
139
+ try {
140
+ lifecycle = await options.beginRetry?.({ retry, logicalAttempt: options.logicalAttempt, taskNames, triggers: retryTriggers, backendDirectory });
141
+ if (retryWork && lifecycle?.workId !== retryWork.work_id)
142
+ throw new Error(`retry lifecycle work identity conflicts with persisted decision: ${retryWork.work_id}`);
143
+ for (const decision of retryDecisions)
144
+ await transitionRetryDecision({
145
+ evalDirectory: options.evalDirectory, evalId: options.evalId, decisionId: decision.decision_id, state: "running",
146
+ });
147
+ if (retryWork)
148
+ options.sink.emit({ type: "eval.retry.admitted", work_id: retryWork.work_id, decision_ids: retryDecisions.map((decision) => decision.decision_id) });
149
+ if (retryWork) {
150
+ options.onRetryExecutionStarted?.(retryWork.work_id);
151
+ executionStarted = true;
152
+ }
153
+ run = await runHarborBackend({
154
+ evalId: options.evalId,
155
+ evalDirectory: options.evalDirectory,
156
+ backendDirectory,
157
+ logicalAttempt: options.logicalAttempt,
158
+ taskNames,
159
+ request: { ...options.request, attempts: 1 },
160
+ root: options.root,
161
+ resolvedRevision: options.resolvedRevision,
162
+ runtimeDirectory: options.controllerRuntime.directory,
163
+ runtimeId: options.controllerRuntime.runtime_id,
164
+ preparedArtifact: options.preparedArtifact,
165
+ ...(options.executionResources ? { executionResources: options.executionResources } : {}),
166
+ ...(options.resolvedImages ? { resolvedImages: options.resolvedImages } : {}),
167
+ ...(options.interactionCaptureExporter ? { modelProxy: options.interactionCaptureExporter.route } : {}),
168
+ ...lifecycle?.backend,
169
+ env: options.env,
170
+ ...(options.harborExecutable !== undefined ? { harborExecutable: options.harborExecutable } : {}),
171
+ ...(options.signal ? { signal: options.signal } : {}),
172
+ ...(options.trialBundleGraceMs === undefined ? {} : { trialBundleGraceMs: options.trialBundleGraceMs }),
173
+ emit: (event) => options.sink.emit({ ...event, execution_kind: "physical-infrastructure-retry", infrastructure_retry: retry, logical_attempt: options.logicalAttempt, ...(lifecycle ? { work_id: lifecycle.workId, lease_id: lifecycle.leaseId } : {}) }),
174
+ onTrialSettled: async (trial, context) => {
175
+ try {
176
+ const environmentImages = retryEnvironmentImages();
177
+ const ref = await importEvalTrialRun({
178
+ root: options.root,
179
+ evalId: options.evalId,
180
+ evalDirectory: options.evalDirectory,
181
+ harborJobDirectory,
182
+ expectedAttempt: options.logicalAttempt,
183
+ request: options.request,
184
+ resolvedRevision: options.resolvedRevision,
185
+ benchmarkId: options.request.benchmark_id,
186
+ benchmarkRevision: options.request.benchmark_revision,
187
+ publicationMode: "replace-invalid",
188
+ runtimeId: options.controllerRuntime.runtime_id,
189
+ env: options.env,
190
+ ...(options.signal ? { signal: options.signal } : {}),
191
+ ...(options.modelCapturePlan ? { modelCapturePlan: options.modelCapturePlan } : {}),
192
+ ...(options.interactionCaptureExporter ? {
193
+ interactionCaptureExporter: options.interactionCaptureExporter,
194
+ } : {}),
195
+ ...await executionEvidence(),
196
+ ...(environmentImages ? { environmentImages } : {}),
197
+ requireCompleteMarker: true,
198
+ allowMissingBundleDiagnostic: context.bundleWaitExpired,
199
+ }, trial, retryRefs.length, retryRefs);
200
+ await publish(ref);
201
+ return true;
202
+ }
203
+ catch (error) {
204
+ if (error instanceof TrialBundlePendingError)
205
+ return false;
206
+ throw error;
207
+ }
208
+ },
209
+ });
210
+ for (const event of await harborPhaseTimingEvents(harborJobDirectory, run.rawResult)) {
211
+ options.sink.emit(event);
212
+ if (event.type === "eval.verifier.completed" && typeof event.duration_ms === "number")
213
+ options.onVerifierDuration?.(event.duration_ms);
214
+ }
215
+ const environmentImages = retryEnvironmentImages();
216
+ const terminalRefs = await importEvalTrialRuns({
217
+ root: options.root,
218
+ evalId: options.evalId,
219
+ evalDirectory: options.evalDirectory,
220
+ harborJobDirectory,
221
+ expectedAttempt: options.logicalAttempt,
222
+ request: options.request,
223
+ resolvedRevision: options.resolvedRevision,
224
+ benchmarkId: options.request.benchmark_id,
225
+ benchmarkRevision: options.request.benchmark_revision,
226
+ publicationMode: "replace-invalid",
227
+ runtimeId: options.controllerRuntime.runtime_id,
228
+ env: options.env,
229
+ ...(options.signal ? { signal: options.signal } : {}),
230
+ ...(options.modelCapturePlan ? { modelCapturePlan: options.modelCapturePlan } : {}),
231
+ ...(options.interactionCaptureExporter ? {
232
+ interactionCaptureExporter: options.interactionCaptureExporter,
233
+ } : {}),
234
+ ...await executionEvidence(),
235
+ ...(environmentImages ? { environmentImages } : {}),
236
+ rawResult: run.rawResult,
237
+ }, retryRefs);
238
+ for (const ref of terminalRefs)
239
+ await publish(ref);
240
+ if (run.rawResult !== null)
241
+ assertBackendTrialSet(run.rawResult, retryRefs);
242
+ runRecord = { attempt: options.logicalAttempt, retry, tasks: taskNames, triggers: retryTriggers, refs: retryRefs, run, ...(lifecycle ? { leaseId: lifecycle.leaseId, workId: lifecycle.workId } : {}) };
243
+ runs.push(runRecord);
244
+ nextCandidates = retryableInfrastructureTrials(retryRefs);
245
+ if (options.originWorkItem && retry < options.request.infrastructure_retries && nextCandidates.length > 0) {
246
+ const nextWork = physicalRetryWorkItem(options.originWorkItem, retry + 1, nextCandidates);
247
+ const nextNotBefore = new Date(Date.now() + retryBackoffMs(options.request.infrastructure_retry_backoff_ms, retry + 1, nextWork.work_id)).toISOString();
248
+ await Promise.all(nextCandidates.map((trigger) => ensurePhysicalRetryDecision({
249
+ evalDirectory: options.evalDirectory, evalId: options.evalId, item: options.originWorkItem,
250
+ retryIndex: retry + 1, trigger, notBefore: nextNotBefore,
251
+ })));
252
+ await options.onRetryWorkPlanned?.(nextWork.work_id);
253
+ }
254
+ const retryState = retryRefs.some((ref) => ref.observation_status === "valid")
255
+ ? "repaired"
256
+ : retry < options.request.infrastructure_retries && nextCandidates.length > 0 ? "invalid" : "exhausted";
257
+ for (const decision of retryDecisions)
258
+ await transitionRetryDecision({
259
+ evalDirectory: options.evalDirectory, evalId: options.evalId, decisionId: decision.decision_id, state: retryState,
260
+ });
261
+ }
262
+ finally {
263
+ await lifecycle?.close();
264
+ if (executionStarted && retryWork) {
265
+ const durationMs = options.onRetryExecutionFinished?.(retryWork.work_id);
266
+ if (runRecord && durationMs !== undefined)
267
+ runRecord.durationMs = durationMs;
268
+ }
269
+ }
146
270
  options.sink.emit({
147
271
  type: "eval.infrastructure-retry.completed",
272
+ execution_kind: "physical-infrastructure-retry",
273
+ candidate_executes: true,
148
274
  attempt: options.logicalAttempt,
149
275
  retry,
150
276
  tasks: taskNames,
@@ -153,14 +279,12 @@ export async function runInfrastructureRetries(options) {
153
279
  });
154
280
  if (run.backend.process_exit_code !== 0 || run.rawResult === null || options.stopAfterResult?.(run.rawResult))
155
281
  break;
156
- retryCandidates = retryableInfrastructureTrials(retryRefs);
282
+ retryCandidates = nextCandidates;
157
283
  }
158
- return { progress, runs };
284
+ return { progress: options.currentProgress?.() ?? progress, runs };
159
285
  }
160
286
  export function retryableInfrastructureTrials(trials) {
161
- return trials.filter((trial) => trial.observation_status === "invalid"
162
- && trial.invalid_reason !== undefined
163
- && RETRYABLE_TRIAL_INFRASTRUCTURE_REASONS.has(trial.invalid_reason));
287
+ return trials.filter((trial) => physicalRetryAllowed(classifyTrialFailure(trial)));
164
288
  }
165
289
  export function infrastructureFailureTrials(trials) {
166
290
  return trials.filter((trial) => trial.observation_status === "invalid"
@@ -1 +1 @@
1
- {"version":3,"file":"infrastructure-retry.js","sourceRoot":"","sources":["../../../src/evals/infrastructure-retry.ts"],"names":[],"mappings":"AAAA,OAAO,IAAI,MAAM,WAAW,CAAC;AAC7B,OAAO,EAAE,UAAU,IAAI,KAAK,EAAE,MAAM,sBAAsB,CAAC;AAE3D,OAAO,EAAE,gBAAgB,EAAE,MAAM,sBAAsB,CAAC;AAIxD,OAAO,EAAE,UAAU,EAAE,MAAM,wBAAwB,CAAC;AAEpD,OAAO,EAAE,kBAAkB,EAAE,mBAAmB,EAAE,uBAAuB,EAAE,2BAA2B,EAAE,MAAM,mBAAmB,CAAC;AAClI,OAAO,EAAE,+BAA+B,EAAE,iBAAiB,EAAE,MAAM,eAAe,CAAC;AAEnF,MAAM,sBAAsB,GAAG,IAAI,GAAG,CAAC;IACrC,wBAAwB;IACxB,iCAAiC;IACjC,yBAAyB;CAC1B,CAAC,CAAC;AAEH,6EAA6E;AAC7E,2EAA2E;AAC3E,4EAA4E;AAC5E,0CAA0C;AAC1C,MAAM,sCAAsC,GAAG,IAAI,GAAG,CAAC;IACrD,wBAAwB;CACzB,CAAC,CAAC;AA+BH,MAAM,CAAC,KAAK,UAAU,wBAAwB,CAC5C,OAAwC;IAExC,IAAI,QAAQ,GAAG,OAAO,CAAC,QAAQ,CAAC;IAChC,IAAI,eAAe,GAAG,6BAA6B,CAAC,OAAO,CAAC,WAAW,CAAC,CAAC;IACzE,MAAM,IAAI,GAA6B,EAAE,CAAC;IAC1C,KAAK,IAAI,KAAK,GAAG,CAAC,EAAE,KAAK,IAAI,OAAO,CAAC,OAAO,CAAC,sBAAsB,IAAI,eAAe,CAAC,MAAM,GAAG,CAAC,EAAE,KAAK,IAAI,CAAC,EAAE,CAAC;QAC9G,IAAI,OAAO,CAAC,MAAM,EAAE,OAAO;YAAE,MAAM;QACnC,MAAM,aAAa,GAAG,CAAC,GAAG,eAAe,CAAC,CAAC;QAC3C,MAAM,SAAS,GAAG,CAAC,GAAG,IAAI,GAAG,CAAC,eAAe,CAAC,GAAG,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC;QACjF,MAAM,SAAS,GAAG,OAAO,CAAC,OAAO,CAAC,+BAA+B,GAAG,KAAK,CAAC;QAC1E,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC;YAChB,IAAI,EAAE,qCAAqC;YAC3C,OAAO,EAAE,OAAO,CAAC,cAAc;YAC/B,KAAK;YACL,KAAK,EAAE,SAAS;YAChB,UAAU,EAAE,SAAS;SACtB,CAAC,CAAC;QACH,IAAI,SAAS,GAAG,CAAC,EAAE,CAAC;YAClB,IAAI,OAAO,CAAC,MAAM;gBAAE,MAAM,KAAK,CAAC,SAAS,EAAE,SAAS,EAAE,EAAE,MAAM,EAAE,OAAO,CAAC,MAAM,EAAE,CAAC,CAAC;;gBAC7E,MAAM,KAAK,CAAC,SAAS,CAAC,CAAC;QAC9B,CAAC;QACD,IAAI,OAAO,CAAC,MAAM,EAAE,OAAO;YAAE,MAAM;QAEnC,MAAM,gBAAgB,GAAG,IAAI,CAAC,IAAI,CAChC,OAAO,CAAC,aAAa,EACrB,wBAAwB,EACxB,SAAS,MAAM,CAAC,KAAK,CAAC,CAAC,QAAQ,CAAC,CAAC,EAAE,GAAG,CAAC,EAAE,EACzC,OAAO,CAAC,OAAO,CAAC,QAAQ,KAAK,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,oBAAoB,CAAC,OAAO,CAAC,cAAc,CAAC,CACzF,CAAC;QACF,MAAM,kBAAkB,GAAG,IAAI,CAAC,IAAI,CAAC,gBAAgB,EAAE,KAAK,CAAC,CAAC;QAC9D,MAAM,SAAS,GAAqB,EAAE,CAAC;QACvC,MAAM,aAAa,GAAG,IAAI,GAAG,CAAC,SAAS,CAAC,CAAC;QACzC,MAAM,OAAO,GAAG,KAAK,EAAE,GAAmB,EAAiB,EAAE;YAC3D,IAAI,GAAG,CAAC,OAAO,KAAK,OAAO,CAAC,cAAc,IAAI,CAAC,aAAa,CAAC,GAAG,CAAC,GAAG,CAAC,OAAO,CAAC,EAAE,CAAC;gBAC9E,MAAM,IAAI,UAAU,CAAC,6DAA6D,GAAG,CAAC,OAAO,IAAI,GAAG,CAAC,OAAO,EAAE,EAAE;oBAC9G,IAAI,EAAE,0CAA0C;oBAChD,QAAQ,EAAE,EAAE;iBACb,CAAC,CAAC;YACL,CAAC;YACD,MAAM,QAAQ,GAAG,SAAS,CAAC,IAAI,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,OAAO,CAAC,QAAQ,KAAK,GAAG,CAAC,QAAQ,CAAC,CAAC;YAChF,IAAI,QAAQ,EAAE,CAAC;gBACb,IAAI,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAC,KAAK,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC;oBAAE,MAAM,IAAI,KAAK,CAAC,gDAAgD,GAAG,CAAC,QAAQ,EAAE,CAAC,CAAC;gBACtI,OAAO;YACT,CAAC;YACD,SAAS,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;YACpB,MAAM,2BAA2B,CAAC,OAAO,CAAC,IAAI,EAAE,OAAO,CAAC,MAAM,EAAE,CAAC,GAAG,CAAC,EAAE;gBACrE,WAAW,EAAE,OAAO,CAAC,OAAO,CAAC,YAAY;gBACzC,iBAAiB,EAAE,OAAO,CAAC,OAAO,CAAC,kBAAkB;aACtD,CAAC,CAAC;YACH,IAAI,GAAG,CAAC,kBAAkB,KAAK,OAAO,EAAE,CAAC;gBACvC,QAAQ,GAAG,+BAA+B,CAAC,QAAQ,EAAE,GAAG,CAAC,CAAC;gBAC1D,MAAM,iBAAiB,CAAC,OAAO,CAAC,aAAa,EAAE,QAAQ,CAAC,CAAC;YAC3D,CAAC;YACD,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC;gBAChB,IAAI,EAAE,GAAG,CAAC,kBAAkB,KAAK,OAAO;oBACtC,CAAC,CAAC,oCAAoC;oBACtC,CAAC,CAAC,kCAAkC;gBACtC,OAAO,EAAE,OAAO,CAAC,cAAc;gBAC/B,KAAK;gBACL,OAAO,EAAE,GAAG,CAAC,OAAO;gBACpB,QAAQ,EAAE,GAAG,CAAC,QAAQ;gBACtB,MAAM,EAAE,GAAG,CAAC,MAAM;gBAClB,kBAAkB,EAAE,GAAG,CAAC,kBAAkB;gBAC1C,GAAG,CAAC,GAAG,CAAC,cAAc,CAAC,CAAC,CAAC,EAAE,cAAc,EAAE,GAAG,CAAC,cAAc,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;gBACrE,UAAU,EAAE,QAAQ,CAAC,UAAU;aAChC,CAAC,CAAC;QACL,CAAC,CAAC;QACF,MAAM,GAAG,GAAG,MAAM,gBAAgB,CAAC;YACjC,MAAM,EAAE,OAAO,CAAC,MAAM;YACtB,aAAa,EAAE,OAAO,CAAC,aAAa;YACpC,gBAAgB;YAChB,cAAc,EAAE,OAAO,CAAC,cAAc;YACtC,SAAS;YACT,OAAO,EAAE,EAAE,GAAG,OAAO,CAAC,OAAO,EAAE,QAAQ,EAAE,CAAC,EAAE;YAC5C,IAAI,EAAE,OAAO,CAAC,IAAI;YAClB,gBAAgB,EAAE,OAAO,CAAC,gBAAgB;YAC1C,gBAAgB,EAAE,OAAO,CAAC,iBAAiB,CAAC,SAAS;YACrD,SAAS,EAAE,OAAO,CAAC,iBAAiB,CAAC,UAAU;YAC/C,gBAAgB,EAAE,OAAO,CAAC,gBAAgB;YAC1C,GAAG,CAAC,OAAO,CAAC,cAAc,CAAC,CAAC,CAAC,EAAE,cAAc,EAAE,OAAO,CAAC,cAAc,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;YAC7E,GAAG,EAAE,OAAO,CAAC,GAAG;YAChB,GAAG,CAAC,OAAO,CAAC,gBAAgB,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,gBAAgB,EAAE,OAAO,CAAC,gBAAgB,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;YACjG,GAAG,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,OAAO,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;YACrD,GAAG,CAAC,OAAO,CAAC,kBAAkB,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,kBAAkB,EAAE,OAAO,CAAC,kBAAkB,EAAE,CAAC;YACvG,IAAI,EAAE,CAAC,KAAK,EAAE,EAAE,CAAC,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,GAAG,KAAK,EAAE,oBAAoB,EAAE,KAAK,EAAE,eAAe,EAAE,OAAO,CAAC,cAAc,EAAE,CAAC;YACtH,cAAc,EAAE,KAAK,EAAE,KAAK,EAAE,OAAO,EAAoB,EAAE;gBACzD,IAAI,CAAC;oBACH,MAAM,GAAG,GAAG,MAAM,kBAAkB,CAAC;wBACnC,IAAI,EAAE,OAAO,CAAC,IAAI;wBAClB,MAAM,EAAE,OAAO,CAAC,MAAM;wBACtB,aAAa,EAAE,OAAO,CAAC,aAAa;wBACpC,kBAAkB;wBAClB,eAAe,EAAE,OAAO,CAAC,cAAc;wBACvC,OAAO,EAAE,OAAO,CAAC,OAAO;wBACxB,gBAAgB,EAAE,OAAO,CAAC,gBAAgB;wBAC1C,WAAW,EAAE,OAAO,CAAC,OAAO,CAAC,YAAY;wBACzC,iBAAiB,EAAE,OAAO,CAAC,OAAO,CAAC,kBAAkB;wBACrD,SAAS,EAAE,OAAO,CAAC,iBAAiB,CAAC,UAAU;wBAC/C,qBAAqB,EAAE,IAAI;wBAC3B,4BAA4B,EAAE,OAAO,CAAC,iBAAiB;qBACxD,EAAE,KAAK,EAAE,SAAS,CAAC,MAAM,EAAE,SAAS,CAAC,CAAC;oBACvC,MAAM,OAAO,CAAC,GAAG,CAAC,CAAC;oBACnB,OAAO,IAAI,CAAC;gBACd,CAAC;gBAAC,OAAO,KAAK,EAAE,CAAC;oBACf,IAAI,KAAK,YAAY,uBAAuB;wBAAE,OAAO,KAAK,CAAC;oBAC3D,MAAM,KAAK,CAAC;gBACd,CAAC;YACH,CAAC;SACF,CAAC,CAAC;QACH,MAAM,YAAY,GAAG,MAAM,mBAAmB,CAAC;YAC7C,IAAI,EAAE,OAAO,CAAC,IAAI;YAClB,MAAM,EAAE,OAAO,CAAC,MAAM;YACtB,aAAa,EAAE,OAAO,CAAC,aAAa;YACpC,kBAAkB;YAClB,eAAe,EAAE,OAAO,CAAC,cAAc;YACvC,OAAO,EAAE,OAAO,CAAC,OAAO;YACxB,gBAAgB,EAAE,OAAO,CAAC,gBAAgB;YAC1C,WAAW,EAAE,OAAO,CAAC,OAAO,CAAC,YAAY;YACzC,iBAAiB,EAAE,OAAO,CAAC,OAAO,CAAC,kBAAkB;YACrD,SAAS,EAAE,OAAO,CAAC,iBAAiB,CAAC,UAAU;YAC/C,SAAS,EAAE,GAAG,CAAC,SAAS;SACzB,EAAE,SAAS,CAAC,CAAC;QACd,KAAK,MAAM,GAAG,IAAI,YAAY;YAAE,MAAM,OAAO,CAAC,GAAG,CAAC,CAAC;QACnD,IAAI,GAAG,CAAC,SAAS,KAAK,IAAI;YAAE,qBAAqB,CAAC,GAAG,CAAC,SAAS,EAAE,SAAS,CAAC,CAAC;QAC5E,IAAI,CAAC,IAAI,CAAC,EAAE,OAAO,EAAE,OAAO,CAAC,cAAc,EAAE,KAAK,EAAE,KAAK,EAAE,SAAS,EAAE,QAAQ,EAAE,aAAa,EAAE,IAAI,EAAE,SAAS,EAAE,GAAG,EAAE,CAAC,CAAC;QACvH,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC;YAChB,IAAI,EAAE,qCAAqC;YAC3C,OAAO,EAAE,OAAO,CAAC,cAAc;YAC/B,KAAK;YACL,KAAK,EAAE,SAAS;YAChB,cAAc,EAAE,SAAS,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,GAAG,CAAC,kBAAkB,KAAK,OAAO,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,IAAI,EAAE;YAC9G,eAAe,EAAE,6BAA6B,CAAC,SAAS,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,IAAI,EAAE;SAC3F,CAAC,CAAC;QACH,IAAI,GAAG,CAAC,OAAO,CAAC,iBAAiB,KAAK,CAAC,IAAI,GAAG,CAAC,SAAS,KAAK,IAAI,IAAI,OAAO,CAAC,eAAe,EAAE,CAAC,GAAG,CAAC,SAAS,CAAC;YAAE,MAAM;QACrH,eAAe,GAAG,6BAA6B,CAAC,SAAS,CAAC,CAAC;IAC7D,CAAC;IACD,OAAO,EAAE,QAAQ,EAAE,IAAI,EAAE,CAAC;AAC5B,CAAC;AAED,MAAM,UAAU,6BAA6B,CAAC,MAAiC;IAC7E,OAAO,MAAM,CAAC,MAAM,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,kBAAkB,KAAK,SAAS;WACjE,KAAK,CAAC,cAAc,KAAK,SAAS;WAClC,sCAAsC,CAAC,GAAG,CAAC,KAAK,CAAC,cAAc,CAAC,CAAC,CAAC;AACzE,CAAC;AAED,MAAM,UAAU,2BAA2B,CAAC,MAAiC;IAC3E,OAAO,MAAM,CAAC,MAAM,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,kBAAkB,KAAK,SAAS;WACjE,KAAK,CAAC,cAAc,KAAK,SAAS;WAClC,sBAAsB,CAAC,GAAG,CAAC,KAAK,CAAC,cAAc,CAAC,CAAC,CAAC;AACzD,CAAC;AAED,SAAS,qBAAqB,CAAC,SAAkC,EAAE,IAA+B;IAChG,MAAM,MAAM,GAAG,KAAK,CAAC,OAAO,CAAC,SAAS,CAAC,aAAa,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC,aAA0C,CAAC,CAAC,CAAC,EAAE,CAAC;IAClH,MAAM,UAAU,GAAG,IAAI,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,OAAO,KAAK,CAAC,UAAU,KAAK,QAAQ,CAAC,CAAC,CAAC,KAAK,CAAC,UAAU,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,MAAM,CAAC,CAAC,KAAK,EAAmB,EAAE,CAAC,KAAK,KAAK,IAAI,CAAC,CAAC,CAAC;IACrK,MAAM,MAAM,GAAG,IAAI,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,GAAG,CAAC,QAAQ,CAAC,CAAC,CAAC;IACxD,IAAI,UAAU,CAAC,IAAI,KAAK,MAAM,CAAC,MAAM,IAAI,UAAU,CAAC,IAAI,KAAK,MAAM,CAAC,IAAI,IAAI,CAAC,GAAG,UAAU,CAAC,CAAC,IAAI,CAAC,CAAC,EAAE,EAAE,EAAE,CAAC,CAAC,MAAM,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC,EAAE,CAAC;QAC1H,MAAM,IAAI,KAAK,CAAC,8EAA8E,CAAC,CAAC;IAClG,CAAC;AACH,CAAC;AAED,SAAS,oBAAoB,CAAC,OAAe;IAC3C,OAAO,WAAW,MAAM,CAAC,OAAO,CAAC,CAAC,QAAQ,CAAC,CAAC,EAAE,GAAG,CAAC,EAAE,CAAC;AACvD,CAAC"}
1
+ {"version":3,"file":"infrastructure-retry.js","sourceRoot":"","sources":["../../../src/evals/infrastructure-retry.ts"],"names":[],"mappings":"AAAA,OAAO,IAAI,MAAM,WAAW,CAAC;AAC7B,OAAO,EAAE,UAAU,IAAI,KAAK,EAAE,MAAM,sBAAsB,CAAC;AAE3D,OAAO,EAAE,gBAAgB,EAAE,MAAM,sBAAsB,CAAC;AAIxD,OAAO,EAAE,UAAU,EAAE,MAAM,wBAAwB,CAAC;AAEpD,OAAO,EAAE,kBAAkB,EAAE,mBAAmB,EAAE,uBAAuB,EAAE,2BAA2B,EAAE,MAAM,mBAAmB,CAAC;AAClI,OAAO,EAAE,+BAA+B,EAAE,iBAAiB,EAAE,MAAM,eAAe,CAAC;AAGnF,OAAO,EAAE,2BAA2B,EAAE,uBAAuB,EAAE,MAAM,kBAAkB,CAAC;AACxF,OAAO,EAAE,qBAAqB,EAAE,MAAM,0BAA0B,CAAC;AACjE,OAAO,EAAE,oBAAoB,EAAE,oBAAoB,EAAE,MAAM,yBAAyB,CAAC;AACrF,OAAO,EAAE,cAAc,EAAE,MAAM,oBAAoB,CAAC;AACpD,OAAO,EAAE,uBAAuB,EAAE,MAAM,2BAA2B,CAAC;AAEpE,MAAM,sBAAsB,GAAG,IAAI,GAAG,CAAC;IACrC,wBAAwB;IACxB,iCAAiC;IACjC,yBAAyB;IACzB,0BAA0B;IAC1B,sBAAsB;IACtB,gCAAgC;IAChC,uBAAuB;IACvB,8BAA8B;IAC9B,8BAA8B;IAC9B,sBAAsB;IACtB,2BAA2B;CAC5B,CAAC,CAAC;AAuEH,MAAM,CAAC,KAAK,UAAU,wBAAwB,CAC5C,OAAwC;IAExC,IAAI,QAAQ,GAAG,OAAO,CAAC,QAAQ,CAAC;IAChC,IAAI,eAAe,GAAG,6BAA6B,CAAC,OAAO,CAAC,WAAW,CAAC,CAAC;IACzE,MAAM,IAAI,GAA6B,EAAE,CAAC;IAC1C,MAAM,eAAe,GAAG,OAAO,CAAC,eAAe,IAAI,CAAC,CAAC;IACrD,IAAI,CAAC,MAAM,CAAC,aAAa,CAAC,eAAe,CAAC,IAAI,eAAe,GAAG,CAAC;QAAE,MAAM,IAAI,SAAS,CAAC,6CAA6C,CAAC,CAAC;IACtI,KAAK,IAAI,KAAK,GAAG,eAAe,EAAE,KAAK,IAAI,OAAO,CAAC,OAAO,CAAC,sBAAsB,IAAI,eAAe,CAAC,MAAM,GAAG,CAAC,EAAE,KAAK,IAAI,CAAC,EAAE,CAAC;QAC5H,IAAI,OAAO,CAAC,MAAM,EAAE,OAAO;YAAE,MAAM;QACnC,MAAM,aAAa,GAAG,CAAC,GAAG,eAAe,CAAC,CAAC;QAC3C,MAAM,SAAS,GAAG,CAAC,GAAG,IAAI,GAAG,CAAC,eAAe,CAAC,GAAG,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC;QACjF,MAAM,SAAS,GAAG,OAAO,CAAC,cAAc,CAAC,CAAC,CAAC,qBAAqB,CAAC,OAAO,CAAC,cAAc,EAAE,KAAK,EAAE,aAAa,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;QAC3H,MAAM,SAAS,GAAG,cAAc,CAAC,OAAO,CAAC,OAAO,CAAC,+BAA+B,EAAE,KAAK,EAAE,SAAS,EAAE,OAAO,IAAI,aAAa,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,QAAQ,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC;QACxK,MAAM,SAAS,GAAG,IAAI,IAAI,CAAC,IAAI,CAAC,GAAG,EAAE,GAAG,SAAS,CAAC,CAAC,WAAW,EAAE,CAAC;QACjE,MAAM,cAAc,GAAG,OAAO,CAAC,cAAc;YAC3C,CAAC,CAAC,MAAM,OAAO,CAAC,GAAG,CAAC,aAAa,CAAC,GAAG,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,2BAA2B,CAAC;gBAC7E,aAAa,EAAE,OAAO,CAAC,aAAa,EAAE,MAAM,EAAE,OAAO,CAAC,MAAM,EAAE,IAAI,EAAE,OAAO,CAAC,cAAmC;gBAC/G,UAAU,EAAE,KAAK,EAAE,OAAO,EAAE,SAAS;aACtC,CAAC,CAAC,CAAC;YACJ,CAAC,CAAC,EAAE,CAAC;QACP,IAAI,SAAS,IAAI,cAAc,CAAC,IAAI,CAAC,CAAC,QAAQ,EAAE,EAAE,CAAC,QAAQ,CAAC,aAAa,KAAK,SAAS,CAAC,OAAO,CAAC,EAAE,CAAC;YACjG,MAAM,IAAI,KAAK,CAAC,2DAA2D,SAAS,CAAC,OAAO,EAAE,CAAC,CAAC;QAClG,CAAC;QACD,IAAI,SAAS;YAAE,MAAM,OAAO,CAAC,kBAAkB,EAAE,CAAC,SAAS,CAAC,OAAO,CAAC,CAAC;QACrE,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC;YAChB,IAAI,EAAE,qBAAqB;YAC3B,cAAc,EAAE,+BAA+B;YAC/C,kBAAkB,EAAE,IAAI;YACxB,GAAG,CAAC,SAAS,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,SAAS,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;YACpD,YAAY,EAAE,cAAc,CAAC,GAAG,CAAC,CAAC,QAAQ,EAAE,EAAE,CAAC,QAAQ,CAAC,WAAW,CAAC;YACpE,OAAO,EAAE,OAAO,CAAC,cAAc;YAC/B,KAAK;YACL,KAAK,EAAE,SAAS;YAChB,UAAU,EAAE,SAAS;SACtB,CAAC,CAAC;QACH,MAAM,kBAAkB,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,KAAK,CAAC,cAAc,CAAC,CAAC,CAAC,EAAE,UAAU,IAAI,SAAS,CAAC,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC,CAAC;QAC5G,OAAO,CAAC,cAAc,EAAE,CAAC,kBAAkB,CAAC,CAAC;QAC7C,IAAI,kBAAkB,GAAG,CAAC,EAAE,CAAC;YAC3B,IAAI,OAAO,CAAC,MAAM;gBAAE,MAAM,KAAK,CAAC,kBAAkB,EAAE,SAAS,EAAE,EAAE,MAAM,EAAE,OAAO,CAAC,MAAM,EAAE,CAAC,CAAC;;gBACtF,MAAM,KAAK,CAAC,kBAAkB,CAAC,CAAC;QACvC,CAAC;QACD,IAAI,OAAO,CAAC,MAAM,EAAE,OAAO;YAAE,MAAM;QACnC,IAAI,SAAS;YAAE,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC;gBAC/B,IAAI,EAAE,kBAAkB,EAAE,OAAO,EAAE,SAAS,CAAC,OAAO,EAAE,YAAY,EAAE,cAAc,CAAC,GAAG,CAAC,CAAC,QAAQ,EAAE,EAAE,CAAC,QAAQ,CAAC,WAAW,CAAC;gBAC1H,KAAK,EAAE,KAAK,EAAE,SAAS,EAAE,QAAQ,EAAE,OAAO,CAAC,cAAc,EAAE,UAAU,EAAE,iBAAiB,IAAI,CAAC;aAC9F,CAAC,CAAC;QAEH,MAAM,gBAAgB,GAAG,IAAI,CAAC,IAAI,CAChC,OAAO,CAAC,oBAAoB,IAAI,IAAI,CAAC,IAAI,CAAC,OAAO,CAAC,aAAa,EAAE,wBAAwB,CAAC,EAC1F,SAAS,MAAM,CAAC,KAAK,CAAC,CAAC,QAAQ,CAAC,CAAC,EAAE,GAAG,CAAC,EAAE,EACzC,OAAO,CAAC,OAAO,CAAC,QAAQ,KAAK,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,oBAAoB,CAAC,OAAO,CAAC,cAAc,CAAC,CACzF,CAAC;QACF,MAAM,kBAAkB,GAAG,IAAI,CAAC,IAAI,CAAC,gBAAgB,EAAE,KAAK,CAAC,CAAC;QAC9D,MAAM,SAAS,GAAqB,EAAE,CAAC;QACvC,MAAM,aAAa,GAAG,IAAI,GAAG,CAAC,SAAS,CAAC,CAAC;QACzC,IAAI,SAAmD,CAAC;QACxD,MAAM,sBAAsB,GAAG,GAAyC,EAAE,CAAC,SAAS,EAAE,iBAAiB,IAAI,OAAO,CAAC,iBAAiB,CAAC;QACrI,MAAM,iBAAiB,GAAG,KAAK,IAA0D,EAAE,CAAC,SAAS,EAAE,wBAAwB;YAC7H,CAAC,CAAC,EAAE,iBAAiB,EAAE,MAAM,SAAS,CAAC,wBAAwB,EAAE,EAAE;YACnE,CAAC,CAAC,EAAE,CAAC;QACP,MAAM,OAAO,GAAG,KAAK,EAAE,GAAmB,EAAiB,EAAE;YAC3D,IAAI,GAAG,CAAC,OAAO,KAAK,OAAO,CAAC,cAAc,IAAI,CAAC,aAAa,CAAC,GAAG,CAAC,GAAG,CAAC,OAAO,CAAC,EAAE,CAAC;gBAC9E,MAAM,IAAI,UAAU,CAAC,6DAA6D,GAAG,CAAC,OAAO,IAAI,GAAG,CAAC,OAAO,EAAE,EAAE;oBAC9G,IAAI,EAAE,0CAA0C;oBAChD,QAAQ,EAAE,EAAE;iBACb,CAAC,CAAC;YACL,CAAC;YACD,MAAM,QAAQ,GAAG,SAAS,CAAC,IAAI,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,OAAO,CAAC,QAAQ,KAAK,GAAG,CAAC,QAAQ,CAAC,CAAC;YAChF,IAAI,QAAQ,EAAE,CAAC;gBACb,IAAI,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAC,KAAK,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC;oBAAE,MAAM,IAAI,KAAK,CAAC,gDAAgD,GAAG,CAAC,QAAQ,EAAE,CAAC,CAAC;gBACtI,OAAO;YACT,CAAC;YACD,SAAS,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;YACpB,IAAI,GAAG,CAAC,cAAc,KAAK,gCAAgC;gBAAE,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC;oBAC7E,IAAI,EAAE,uBAAuB,EAAE,OAAO,EAAE,SAAS,EAAE,OAAO,EAAE,QAAQ,EAAE,GAAG,CAAC,QAAQ,EAAE,OAAO,EAAE,GAAG,CAAC,OAAO;oBACxG,MAAM,EAAE,gCAAgC,EAAE,kBAAkB,EAAE,IAAI,EAAE,iBAAiB,EAAE,KAAK;iBAC7F,CAAC,CAAC;YACH,MAAM,2BAA2B,CAAC,OAAO,CAAC,IAAI,EAAE,OAAO,CAAC,MAAM,EAAE,CAAC,GAAG,CAAC,EAAE;gBACrE,WAAW,EAAE,OAAO,CAAC,OAAO,CAAC,YAAY;gBACzC,iBAAiB,EAAE,OAAO,CAAC,OAAO,CAAC,kBAAkB;aACtD,CAAC,CAAC;YACH,IAAI,GAAG,CAAC,kBAAkB,KAAK,OAAO,EAAE,CAAC;gBACvC,IAAI,OAAO,CAAC,oBAAoB,EAAE,CAAC;oBACjC,IAAI,CAAC,SAAS;wBAAE,MAAM,IAAI,KAAK,CAAC,gEAAgE,CAAC,CAAC;oBAClG,MAAM,OAAO,CAAC,oBAAoB,CAAC,GAAG,EAAE,SAAS,CAAC,MAAM,CAAC,CAAC;oBAC1D,QAAQ,GAAG,OAAO,CAAC,eAAe,EAAE,EAAE,IAAI,QAAQ,CAAC;gBACrD,CAAC;qBAAM,CAAC;oBACN,QAAQ,GAAG,+BAA+B,CAAC,QAAQ,EAAE,GAAG,CAAC,CAAC;oBAC1D,MAAM,iBAAiB,CAAC,OAAO,CAAC,aAAa,EAAE,QAAQ,CAAC,CAAC;gBAC3D,CAAC;YACH,CAAC;YACD,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC;gBAChB,IAAI,EAAE,GAAG,CAAC,kBAAkB,KAAK,OAAO;oBACtC,CAAC,CAAC,oCAAoC;oBACtC,CAAC,CAAC,kCAAkC;gBACtC,OAAO,EAAE,OAAO,CAAC,cAAc;gBAC/B,KAAK;gBACL,OAAO,EAAE,GAAG,CAAC,OAAO;gBACpB,QAAQ,EAAE,GAAG,CAAC,QAAQ;gBACtB,MAAM,EAAE,GAAG,CAAC,MAAM;gBAClB,kBAAkB,EAAE,GAAG,CAAC,kBAAkB;gBAC1C,GAAG,CAAC,GAAG,CAAC,cAAc,CAAC,CAAC,CAAC,EAAE,cAAc,EAAE,GAAG,CAAC,cAAc,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;gBACrE,UAAU,EAAE,CAAC,OAAO,CAAC,eAAe,EAAE,EAAE,IAAI,QAAQ,CAAC,CAAC,UAAU;aACjE,CAAC,CAAC;QACL,CAAC,CAAC;QACF,IAAI,GAAwB,CAAC;QAC7B,IAAI,cAAc,GAAqB,EAAE,CAAC;QAC1C,IAAI,SAA6C,CAAC;QAClD,IAAI,gBAAgB,GAAG,KAAK,CAAC;QAC7B,IAAI,CAAC;YACH,SAAS,GAAG,MAAM,OAAO,CAAC,UAAU,EAAE,CAAC,EAAE,KAAK,EAAE,cAAc,EAAE,OAAO,CAAC,cAAc,EAAE,SAAS,EAAE,QAAQ,EAAE,aAAa,EAAE,gBAAgB,EAAE,CAAC,CAAC;YAChJ,IAAI,SAAS,IAAI,SAAS,EAAE,MAAM,KAAK,SAAS,CAAC,OAAO;gBAAE,MAAM,IAAI,KAAK,CAAC,oEAAoE,SAAS,CAAC,OAAO,EAAE,CAAC,CAAC;YACnK,KAAK,MAAM,QAAQ,IAAI,cAAc;gBAAE,MAAM,uBAAuB,CAAC;oBACnE,aAAa,EAAE,OAAO,CAAC,aAAa,EAAE,MAAM,EAAE,OAAO,CAAC,MAAM,EAAE,UAAU,EAAE,QAAQ,CAAC,WAAW,EAAE,KAAK,EAAE,SAAS;iBACjH,CAAC,CAAC;YACH,IAAI,SAAS;gBAAE,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,qBAAqB,EAAE,OAAO,EAAE,SAAS,CAAC,OAAO,EAAE,YAAY,EAAE,cAAc,CAAC,GAAG,CAAC,CAAC,QAAQ,EAAE,EAAE,CAAC,QAAQ,CAAC,WAAW,CAAC,EAAE,CAAC,CAAC;YACpK,IAAI,SAAS,EAAE,CAAC;gBACd,OAAO,CAAC,uBAAuB,EAAE,CAAC,SAAS,CAAC,OAAO,CAAC,CAAC;gBACrD,gBAAgB,GAAG,IAAI,CAAC;YAC1B,CAAC;YACD,GAAG,GAAG,MAAM,gBAAgB,CAAC;gBAC3B,MAAM,EAAE,OAAO,CAAC,MAAM;gBACtB,aAAa,EAAE,OAAO,CAAC,aAAa;gBACpC,gBAAgB;gBAChB,cAAc,EAAE,OAAO,CAAC,cAAc;gBACtC,SAAS;gBACT,OAAO,EAAE,EAAE,GAAG,OAAO,CAAC,OAAO,EAAE,QAAQ,EAAE,CAAC,EAAE;gBAC5C,IAAI,EAAE,OAAO,CAAC,IAAI;gBAClB,gBAAgB,EAAE,OAAO,CAAC,gBAAgB;gBAC1C,gBAAgB,EAAE,OAAO,CAAC,iBAAiB,CAAC,SAAS;gBACrD,SAAS,EAAE,OAAO,CAAC,iBAAiB,CAAC,UAAU;gBAC/C,gBAAgB,EAAE,OAAO,CAAC,gBAAgB;gBAC1C,GAAG,CAAC,OAAO,CAAC,kBAAkB,CAAC,CAAC,CAAC,EAAE,kBAAkB,EAAE,OAAO,CAAC,kBAAkB,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;gBACzF,GAAG,CAAC,OAAO,CAAC,cAAc,CAAC,CAAC,CAAC,EAAE,cAAc,EAAE,OAAO,CAAC,cAAc,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;gBAC7E,GAAG,CAAC,OAAO,CAAC,0BAA0B,CAAC,CAAC,CAAC,EAAE,UAAU,EAAE,OAAO,CAAC,0BAA0B,CAAC,KAAK,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;gBACvG,GAAG,SAAS,EAAE,OAAO;gBACrB,GAAG,EAAE,OAAO,CAAC,GAAG;gBAChB,GAAG,CAAC,OAAO,CAAC,gBAAgB,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,gBAAgB,EAAE,OAAO,CAAC,gBAAgB,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;gBACjG,GAAG,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,OAAO,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;gBACrD,GAAG,CAAC,OAAO,CAAC,kBAAkB,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,kBAAkB,EAAE,OAAO,CAAC,kBAAkB,EAAE,CAAC;gBACvG,IAAI,EAAE,CAAC,KAAK,EAAE,EAAE,CAAC,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,GAAG,KAAK,EAAE,cAAc,EAAE,+BAA+B,EAAE,oBAAoB,EAAE,KAAK,EAAE,eAAe,EAAE,OAAO,CAAC,cAAc,EAAE,GAAG,CAAC,SAAS,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,SAAS,CAAC,MAAM,EAAE,QAAQ,EAAE,SAAS,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,CAAC;gBACzP,cAAc,EAAE,KAAK,EAAE,KAAK,EAAE,OAAO,EAAoB,EAAE;oBACzD,IAAI,CAAC;wBACH,MAAM,iBAAiB,GAAG,sBAAsB,EAAE,CAAC;wBACnD,MAAM,GAAG,GAAG,MAAM,kBAAkB,CAAC;4BACnC,IAAI,EAAE,OAAO,CAAC,IAAI;4BAClB,MAAM,EAAE,OAAO,CAAC,MAAM;4BACtB,aAAa,EAAE,OAAO,CAAC,aAAa;4BACpC,kBAAkB;4BAClB,eAAe,EAAE,OAAO,CAAC,cAAc;4BACvC,OAAO,EAAE,OAAO,CAAC,OAAO;4BACxB,gBAAgB,EAAE,OAAO,CAAC,gBAAgB;4BAC1C,WAAW,EAAE,OAAO,CAAC,OAAO,CAAC,YAAY;4BACzC,iBAAiB,EAAE,OAAO,CAAC,OAAO,CAAC,kBAAkB;4BACrD,eAAe,EAAE,iBAAiB;4BAClC,SAAS,EAAE,OAAO,CAAC,iBAAiB,CAAC,UAAU;4BAC/C,GAAG,EAAE,OAAO,CAAC,GAAG;4BAChB,GAAG,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,OAAO,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;4BACrD,GAAG,CAAC,OAAO,CAAC,gBAAgB,CAAC,CAAC,CAAC,EAAE,gBAAgB,EAAE,OAAO,CAAC,gBAAgB,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;4BACnF,GAAG,CAAC,OAAO,CAAC,0BAA0B,CAAC,CAAC,CAAC;gCACvC,0BAA0B,EAAE,OAAO,CAAC,0BAA0B;6BAC/D,CAAC,CAAC,CAAC,EAAE,CAAC;4BACP,GAAG,MAAM,iBAAiB,EAAE;4BAC5B,GAAG,CAAC,iBAAiB,CAAC,CAAC,CAAC,EAAE,iBAAiB,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;4BACnD,qBAAqB,EAAE,IAAI;4BAC3B,4BAA4B,EAAE,OAAO,CAAC,iBAAiB;yBACxD,EAAE,KAAK,EAAE,SAAS,CAAC,MAAM,EAAE,SAAS,CAAC,CAAC;wBACvC,MAAM,OAAO,CAAC,GAAG,CAAC,CAAC;wBACnB,OAAO,IAAI,CAAC;oBACd,CAAC;oBAAC,OAAO,KAAK,EAAE,CAAC;wBACf,IAAI,KAAK,YAAY,uBAAuB;4BAAE,OAAO,KAAK,CAAC;wBAC3D,MAAM,KAAK,CAAC;oBACd,CAAC;gBACH,CAAC;aACF,CAAC,CAAC;YACH,KAAK,MAAM,KAAK,IAAI,MAAM,uBAAuB,CAAC,kBAAkB,EAAE,GAAG,CAAC,SAAS,CAAC,EAAE,CAAC;gBACrF,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;gBACzB,IAAI,KAAK,CAAC,IAAI,KAAK,yBAAyB,IAAI,OAAO,KAAK,CAAC,WAAW,KAAK,QAAQ;oBAAE,OAAO,CAAC,kBAAkB,EAAE,CAAC,KAAK,CAAC,WAAW,CAAC,CAAC;YACzI,CAAC;YACD,MAAM,iBAAiB,GAAG,sBAAsB,EAAE,CAAC;YACnD,MAAM,YAAY,GAAG,MAAM,mBAAmB,CAAC;gBAC7C,IAAI,EAAE,OAAO,CAAC,IAAI;gBAClB,MAAM,EAAE,OAAO,CAAC,MAAM;gBACtB,aAAa,EAAE,OAAO,CAAC,aAAa;gBACpC,kBAAkB;gBAClB,eAAe,EAAE,OAAO,CAAC,cAAc;gBACvC,OAAO,EAAE,OAAO,CAAC,OAAO;gBACxB,gBAAgB,EAAE,OAAO,CAAC,gBAAgB;gBAC1C,WAAW,EAAE,OAAO,CAAC,OAAO,CAAC,YAAY;gBACzC,iBAAiB,EAAE,OAAO,CAAC,OAAO,CAAC,kBAAkB;gBACrD,eAAe,EAAE,iBAAiB;gBAClC,SAAS,EAAE,OAAO,CAAC,iBAAiB,CAAC,UAAU;gBAC/C,GAAG,EAAE,OAAO,CAAC,GAAG;gBAChB,GAAG,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,OAAO,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;gBACrD,GAAG,CAAC,OAAO,CAAC,gBAAgB,CAAC,CAAC,CAAC,EAAE,gBAAgB,EAAE,OAAO,CAAC,gBAAgB,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;gBACnF,GAAG,CAAC,OAAO,CAAC,0BAA0B,CAAC,CAAC,CAAC;oBACvC,0BAA0B,EAAE,OAAO,CAAC,0BAA0B;iBAC/D,CAAC,CAAC,CAAC,EAAE,CAAC;gBACP,GAAG,MAAM,iBAAiB,EAAE;gBAC5B,GAAG,CAAC,iBAAiB,CAAC,CAAC,CAAC,EAAE,iBAAiB,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;gBACnD,SAAS,EAAE,GAAG,CAAC,SAAS;aACzB,EAAE,SAAS,CAAC,CAAC;YACd,KAAK,MAAM,GAAG,IAAI,YAAY;gBAAE,MAAM,OAAO,CAAC,GAAG,CAAC,CAAC;YACnD,IAAI,GAAG,CAAC,SAAS,KAAK,IAAI;gBAAE,qBAAqB,CAAC,GAAG,CAAC,SAAS,EAAE,SAAS,CAAC,CAAC;YAC5E,SAAS,GAAG,EAAE,OAAO,EAAE,OAAO,CAAC,cAAc,EAAE,KAAK,EAAE,KAAK,EAAE,SAAS,EAAE,QAAQ,EAAE,aAAa,EAAE,IAAI,EAAE,SAAS,EAAE,GAAG,EAAE,GAAG,CAAC,SAAS,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,SAAS,CAAC,OAAO,EAAE,MAAM,EAAE,SAAS,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,CAAC;YACxM,IAAI,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC;YACrB,cAAc,GAAG,6BAA6B,CAAC,SAAS,CAAC,CAAC;YAC1D,IAAI,OAAO,CAAC,cAAc,IAAI,KAAK,GAAG,OAAO,CAAC,OAAO,CAAC,sBAAsB,IAAI,cAAc,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;gBAC1G,MAAM,QAAQ,GAAG,qBAAqB,CAAC,OAAO,CAAC,cAAc,EAAE,KAAK,GAAG,CAAC,EAAE,cAAc,CAAC,CAAC;gBAC1F,MAAM,aAAa,GAAG,IAAI,IAAI,CAAC,IAAI,CAAC,GAAG,EAAE,GAAG,cAAc,CAAC,OAAO,CAAC,OAAO,CAAC,+BAA+B,EAAE,KAAK,GAAG,CAAC,EAAE,QAAQ,CAAC,OAAO,CAAC,CAAC,CAAC,WAAW,EAAE,CAAC;gBACxJ,MAAM,OAAO,CAAC,GAAG,CAAC,cAAc,CAAC,GAAG,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,2BAA2B,CAAC;oBAC5E,aAAa,EAAE,OAAO,CAAC,aAAa,EAAE,MAAM,EAAE,OAAO,CAAC,MAAM,EAAE,IAAI,EAAE,OAAO,CAAC,cAAmC;oBAC/G,UAAU,EAAE,KAAK,GAAG,CAAC,EAAE,OAAO,EAAE,SAAS,EAAE,aAAa;iBACzD,CAAC,CAAC,CAAC,CAAC;gBACL,MAAM,OAAO,CAAC,kBAAkB,EAAE,CAAC,QAAQ,CAAC,OAAO,CAAC,CAAC;YACvD,CAAC;YACD,MAAM,UAAU,GAAG,SAAS,CAAC,IAAI,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,GAAG,CAAC,kBAAkB,KAAK,OAAO,CAAC;gBAC5E,CAAC,CAAC,UAAU;gBACZ,CAAC,CAAC,KAAK,GAAG,OAAO,CAAC,OAAO,CAAC,sBAAsB,IAAI,cAAc,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,WAAW,CAAC;YAC1G,KAAK,MAAM,QAAQ,IAAI,cAAc;gBAAE,MAAM,uBAAuB,CAAC;oBACnE,aAAa,EAAE,OAAO,CAAC,aAAa,EAAE,MAAM,EAAE,OAAO,CAAC,MAAM,EAAE,UAAU,EAAE,QAAQ,CAAC,WAAW,EAAE,KAAK,EAAE,UAAU;iBAClH,CAAC,CAAC;QACL,CAAC;gBAAS,CAAC;YACT,MAAM,SAAS,EAAE,KAAK,EAAE,CAAC;YACzB,IAAI,gBAAgB,IAAI,SAAS,EAAE,CAAC;gBAClC,MAAM,UAAU,GAAG,OAAO,CAAC,wBAAwB,EAAE,CAAC,SAAS,CAAC,OAAO,CAAC,CAAC;gBACzE,IAAI,SAAS,IAAI,UAAU,KAAK,SAAS;oBAAE,SAAS,CAAC,UAAU,GAAG,UAAU,CAAC;YAC/E,CAAC;QACH,CAAC;QACD,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC;YAChB,IAAI,EAAE,qCAAqC;YAC3C,cAAc,EAAE,+BAA+B;YAC/C,kBAAkB,EAAE,IAAI;YACxB,OAAO,EAAE,OAAO,CAAC,cAAc;YAC/B,KAAK;YACL,KAAK,EAAE,SAAS;YAChB,cAAc,EAAE,SAAS,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,GAAG,CAAC,kBAAkB,KAAK,OAAO,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,IAAI,EAAE;YAC9G,eAAe,EAAE,6BAA6B,CAAC,SAAS,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,IAAI,EAAE;SAC3F,CAAC,CAAC;QACH,IAAI,GAAG,CAAC,OAAO,CAAC,iBAAiB,KAAK,CAAC,IAAI,GAAG,CAAC,SAAS,KAAK,IAAI,IAAI,OAAO,CAAC,eAAe,EAAE,CAAC,GAAG,CAAC,SAAS,CAAC;YAAE,MAAM;QACrH,eAAe,GAAG,cAAc,CAAC;IACnC,CAAC;IACD,OAAO,EAAE,QAAQ,EAAE,OAAO,CAAC,eAAe,EAAE,EAAE,IAAI,QAAQ,EAAE,IAAI,EAAE,CAAC;AACrE,CAAC;AAED,MAAM,UAAU,6BAA6B,CAAC,MAAiC;IAC7E,OAAO,MAAM,CAAC,MAAM,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,oBAAoB,CAAC,oBAAoB,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;AACrF,CAAC;AAED,MAAM,UAAU,2BAA2B,CAAC,MAAiC;IAC3E,OAAO,MAAM,CAAC,MAAM,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,kBAAkB,KAAK,SAAS;WACjE,KAAK,CAAC,cAAc,KAAK,SAAS;WAClC,sBAAsB,CAAC,GAAG,CAAC,KAAK,CAAC,cAAc,CAAC,CAAC,CAAC;AACzD,CAAC;AAED,SAAS,qBAAqB,CAAC,SAAkC,EAAE,IAA+B;IAChG,MAAM,MAAM,GAAG,KAAK,CAAC,OAAO,CAAC,SAAS,CAAC,aAAa,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC,aAA0C,CAAC,CAAC,CAAC,EAAE,CAAC;IAClH,MAAM,UAAU,GAAG,IAAI,GAAG,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,OAAO,KAAK,CAAC,UAAU,KAAK,QAAQ,CAAC,CAAC,CAAC,KAAK,CAAC,UAAU,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,MAAM,CAAC,CAAC,KAAK,EAAmB,EAAE,CAAC,KAAK,KAAK,IAAI,CAAC,CAAC,CAAC;IACrK,MAAM,MAAM,GAAG,IAAI,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,GAAG,EAAE,EAAE,CAAC,GAAG,CAAC,QAAQ,CAAC,CAAC,CAAC;IACxD,IAAI,UAAU,CAAC,IAAI,KAAK,MAAM,CAAC,MAAM,IAAI,UAAU,CAAC,IAAI,KAAK,MAAM,CAAC,IAAI,IAAI,CAAC,GAAG,UAAU,CAAC,CAAC,IAAI,CAAC,CAAC,EAAE,EAAE,EAAE,CAAC,CAAC,MAAM,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC,EAAE,CAAC;QAC1H,MAAM,IAAI,KAAK,CAAC,8EAA8E,CAAC,CAAC;IAClG,CAAC;AACH,CAAC;AAED,SAAS,oBAAoB,CAAC,OAAe;IAC3C,OAAO,WAAW,MAAM,CAAC,OAAO,CAAC,CAAC,QAAQ,CAAC,CAAC,EAAE,GAAG,CAAC,EAAE,CAAC;AACvD,CAAC"}
@@ -0,0 +1,49 @@
1
+ import { HitchError, atomicWriteJSON, ensureDir } from "../foundation/index.js";
2
+ import { loadInteractionCapture } from "../model-access/index.js";
3
+ export async function importTrialInteractionCapture(input, runId, staging) {
4
+ const plan = input.modelCapturePlan;
5
+ await writeTrialCapturePolicy(staging, plan);
6
+ const expectsProxy = plan?.effective_mode === "proxy" || plan?.effective_mode === "hybrid";
7
+ const existing = expectsProxy
8
+ ? await loadInteractionCapture(staging).catch((error) => {
9
+ if (error.code === "ENOENT")
10
+ return null;
11
+ throw error;
12
+ })
13
+ : null;
14
+ if (existing) {
15
+ const ref = existing.ref;
16
+ if (ref.run_id !== runId || ref.mode !== plan?.effective_mode || ref.required !== plan.required
17
+ || ref.topology !== plan.topology || plan.required && ref.completeness !== "complete") {
18
+ throw new HitchError("remote model interaction capture does not match its sealed plan", { code: "model_capture_incomplete", exitCode: 12 });
19
+ }
20
+ return;
21
+ }
22
+ if (!input.interactionCaptureExporter) {
23
+ if (expectsProxy && plan.required) {
24
+ throw new HitchError("required model interaction capture is unavailable", { code: "model_capture_incomplete", exitCode: 12 });
25
+ }
26
+ return;
27
+ }
28
+ const ref = await input.interactionCaptureExporter.finalizeRun(runId, staging);
29
+ if (ref.run_id !== runId || ref.mode !== input.interactionCaptureExporter.route.mode) {
30
+ throw new HitchError("model interaction capture identity does not match the run", { code: "model_capture_incomplete", exitCode: 12 });
31
+ }
32
+ if (input.interactionCaptureExporter.plan.required && ref.completeness !== "complete") {
33
+ throw new HitchError("required model interaction capture is incomplete", { code: "model_capture_incomplete", exitCode: 12 });
34
+ }
35
+ }
36
+ export async function writeTrialCapturePolicy(staging, plan) {
37
+ if (!plan)
38
+ return;
39
+ await ensureDir(`${staging}/interactions`);
40
+ await atomicWriteJSON(`${staging}/interactions/capture.policy.json`, {
41
+ schema_version: "1",
42
+ requested_mode: plan.requested_mode,
43
+ effective_mode: plan.effective_mode,
44
+ required: plan.required,
45
+ ...(plan.topology ? { topology: plan.topology } : {}),
46
+ ...(plan.degraded_reason ? { degraded_reason: plan.degraded_reason } : {}),
47
+ });
48
+ }
49
+ //# sourceMappingURL=interaction-capture-import.js.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"interaction-capture-import.js","sourceRoot":"","sources":["../../../src/evals/interaction-capture-import.ts"],"names":[],"mappings":"AACA,OAAO,EAAE,UAAU,EAAE,eAAe,EAAE,SAAS,EAAE,MAAM,wBAAwB,CAAC;AAChF,OAAO,EAAE,sBAAsB,EAAE,MAAM,0BAA0B,CAAC;AAGlE,MAAM,CAAC,KAAK,UAAU,6BAA6B,CAAC,KAGnD,EAAE,KAAa,EAAE,OAAe;IAC/B,MAAM,IAAI,GAAG,KAAK,CAAC,gBAAgB,CAAC;IACpC,MAAM,uBAAuB,CAAC,OAAO,EAAE,IAAI,CAAC,CAAC;IAC7C,MAAM,YAAY,GAAG,IAAI,EAAE,cAAc,KAAK,OAAO,IAAI,IAAI,EAAE,cAAc,KAAK,QAAQ,CAAC;IAC3F,MAAM,QAAQ,GAAG,YAAY;QAC3B,CAAC,CAAC,MAAM,sBAAsB,CAAC,OAAO,CAAC,CAAC,KAAK,CAAC,CAAC,KAAK,EAAE,EAAE;YACtD,IAAK,KAA+B,CAAC,IAAI,KAAK,QAAQ;gBAAE,OAAO,IAAI,CAAC;YACpE,MAAM,KAAK,CAAC;QACd,CAAC,CAAC;QACF,CAAC,CAAC,IAAI,CAAC;IACT,IAAI,QAAQ,EAAE,CAAC;QACb,MAAM,GAAG,GAAG,QAAQ,CAAC,GAAG,CAAC;QACzB,IAAI,GAAG,CAAC,MAAM,KAAK,KAAK,IAAI,GAAG,CAAC,IAAI,KAAK,IAAI,EAAE,cAAc,IAAI,GAAG,CAAC,QAAQ,KAAK,IAAI,CAAC,QAAQ;eAC1F,GAAG,CAAC,QAAQ,KAAK,IAAI,CAAC,QAAQ,IAAI,IAAI,CAAC,QAAQ,IAAI,GAAG,CAAC,YAAY,KAAK,UAAU,EAAE,CAAC;YACxF,MAAM,IAAI,UAAU,CAAC,iEAAiE,EAAE,EAAE,IAAI,EAAE,0BAA0B,EAAE,QAAQ,EAAE,EAAE,EAAE,CAAC,CAAC;QAC9I,CAAC;QACD,OAAO;IACT,CAAC;IACD,IAAI,CAAC,KAAK,CAAC,0BAA0B,EAAE,CAAC;QACtC,IAAI,YAAY,IAAI,IAAI,CAAC,QAAQ,EAAE,CAAC;YAClC,MAAM,IAAI,UAAU,CAAC,mDAAmD,EAAE,EAAE,IAAI,EAAE,0BAA0B,EAAE,QAAQ,EAAE,EAAE,EAAE,CAAC,CAAC;QAChI,CAAC;QACD,OAAO;IACT,CAAC;IACD,MAAM,GAAG,GAAG,MAAM,KAAK,CAAC,0BAA0B,CAAC,WAAW,CAAC,KAAK,EAAE,OAAO,CAAC,CAAC;IAC/E,IAAI,GAAG,CAAC,MAAM,KAAK,KAAK,IAAI,GAAG,CAAC,IAAI,KAAK,KAAK,CAAC,0BAA0B,CAAC,KAAK,CAAC,IAAI,EAAE,CAAC;QACrF,MAAM,IAAI,UAAU,CAAC,2DAA2D,EAAE,EAAE,IAAI,EAAE,0BAA0B,EAAE,QAAQ,EAAE,EAAE,EAAE,CAAC,CAAC;IACxI,CAAC;IACD,IAAI,KAAK,CAAC,0BAA0B,CAAC,IAAI,CAAC,QAAQ,IAAI,GAAG,CAAC,YAAY,KAAK,UAAU,EAAE,CAAC;QACtF,MAAM,IAAI,UAAU,CAAC,kDAAkD,EAAE,EAAE,IAAI,EAAE,0BAA0B,EAAE,QAAQ,EAAE,EAAE,EAAE,CAAC,CAAC;IAC/H,CAAC;AACH,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,uBAAuB,CAAC,OAAe,EAAE,IAAyB;IACtF,IAAI,CAAC,IAAI;QAAE,OAAO;IAClB,MAAM,SAAS,CAAC,GAAG,OAAO,eAAe,CAAC,CAAC;IAC3C,MAAM,eAAe,CAAC,GAAG,OAAO,mCAAmC,EAAE;QACnE,cAAc,EAAE,GAAG;QACnB,cAAc,EAAE,IAAI,CAAC,cAAc;QACnC,cAAc,EAAE,IAAI,CAAC,cAAc;QACnC,QAAQ,EAAE,IAAI,CAAC,QAAQ;QACvB,GAAG,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC,CAAC,EAAE,QAAQ,EAAE,IAAI,CAAC,QAAQ,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QACrD,GAAG,CAAC,IAAI,CAAC,eAAe,CAAC,CAAC,CAAC,EAAE,eAAe,EAAE,IAAI,CAAC,eAAe,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;KAC3E,CAAC,CAAC;AACL,CAAC"}