niceeval 0.17.1 → 0.19.0

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 (692) hide show
  1. package/INDEX.md +5 -2
  2. package/dist/adapter-usage.cjs.map +1 -1
  3. package/dist/adapter-usage.d.cts +1 -1
  4. package/dist/adapter-usage.d.mts +1 -1
  5. package/dist/adapter-usage.d.ts +1 -1
  6. package/dist/adapter.cjs +8 -3
  7. package/dist/adapter.cjs.map +1 -1
  8. package/dist/adapter.d.cts +26 -15
  9. package/dist/adapter.d.mts +26 -15
  10. package/dist/adapter.d.ts +26 -15
  11. package/dist/agents/{effect-runtime.cjs → agent-effects.cjs} +12 -18
  12. package/dist/agents/agent-effects.cjs.map +1 -0
  13. package/dist/agents/agent-effects.d.cts +23 -0
  14. package/dist/agents/agent-effects.d.mts +23 -0
  15. package/dist/agents/agent-effects.d.ts +23 -0
  16. package/dist/agents/agent-effects.js +16 -0
  17. package/dist/agents/agent-effects.js.map +1 -0
  18. package/dist/agents/ai-sdk.cjs +1 -1
  19. package/dist/agents/ai-sdk.cjs.map +1 -1
  20. package/dist/agents/bub.cjs +31 -18
  21. package/dist/agents/bub.cjs.map +1 -1
  22. package/dist/agents/bub.d.cts +9 -12
  23. package/dist/agents/bub.d.mts +9 -12
  24. package/dist/agents/bub.d.ts +9 -12
  25. package/dist/agents/claude-code-native.cjs +2 -2
  26. package/dist/agents/claude-code-native.cjs.map +1 -1
  27. package/dist/agents/claude-code.cjs +17 -11
  28. package/dist/agents/claude-code.cjs.map +1 -1
  29. package/dist/agents/claude-code.d.cts +11 -11
  30. package/dist/agents/claude-code.d.mts +11 -11
  31. package/dist/agents/claude-code.d.ts +11 -11
  32. package/dist/agents/codex-app-server.cjs +2 -2
  33. package/dist/agents/codex-app-server.cjs.map +1 -1
  34. package/dist/agents/codex.cjs +18 -12
  35. package/dist/agents/codex.cjs.map +1 -1
  36. package/dist/agents/codex.d.cts +12 -12
  37. package/dist/agents/codex.d.mts +12 -12
  38. package/dist/agents/codex.d.ts +12 -12
  39. package/dist/agents/deepseek-harness.cjs +2 -2
  40. package/dist/agents/deepseek-harness.cjs.map +1 -1
  41. package/dist/agents/hermes.cjs +3 -3
  42. package/dist/agents/hermes.cjs.map +1 -1
  43. package/dist/agents/index.cjs +3 -3
  44. package/dist/agents/index.cjs.map +1 -1
  45. package/dist/agents/index.d.cts +3 -3
  46. package/dist/agents/index.d.mts +3 -3
  47. package/dist/agents/index.d.ts +3 -3
  48. package/dist/agents/index.js +2 -2
  49. package/dist/agents/index.mjs +2 -2
  50. package/dist/agents/mcp.cjs.map +1 -1
  51. package/dist/agents/mcp.d.cts +2 -2
  52. package/dist/agents/mcp.d.mts +2 -2
  53. package/dist/agents/mcp.d.ts +2 -2
  54. package/dist/agents/omp.cjs +2 -2
  55. package/dist/agents/omp.cjs.map +1 -1
  56. package/dist/agents/openai-compat.cjs +2 -2
  57. package/dist/agents/openai-compat.cjs.map +1 -1
  58. package/dist/agents/openclaw.cjs +3 -3
  59. package/dist/agents/openclaw.cjs.map +1 -1
  60. package/dist/agents/opencode.cjs +16 -4
  61. package/dist/agents/opencode.cjs.map +1 -1
  62. package/dist/agents/plugin-extensions.cjs +71 -0
  63. package/dist/agents/plugin-extensions.cjs.map +1 -0
  64. package/dist/agents/plugin-extensions.d.cts +26 -0
  65. package/dist/agents/plugin-extensions.d.mts +26 -0
  66. package/dist/agents/plugin-extensions.d.ts +26 -0
  67. package/dist/agents/plugin-extensions.js +8 -0
  68. package/dist/agents/plugin-extensions.js.map +1 -0
  69. package/dist/agents/post-setup.cjs +14 -3
  70. package/dist/agents/post-setup.cjs.map +1 -1
  71. package/dist/agents/post-setup.d.cts +4 -2
  72. package/dist/agents/post-setup.d.mts +4 -2
  73. package/dist/agents/post-setup.d.ts +4 -2
  74. package/dist/agents/post-setup.js +8 -6
  75. package/dist/agents/sdk-streams.cjs +3 -3
  76. package/dist/agents/sdk-streams.cjs.map +1 -1
  77. package/dist/agents/{session-slot.cjs → session-key.cjs} +4 -4
  78. package/dist/agents/session-key.cjs.map +1 -0
  79. package/dist/agents/session-key.d.cts +12 -0
  80. package/dist/agents/session-key.d.mts +12 -0
  81. package/dist/agents/session-key.d.ts +12 -0
  82. package/dist/agents/session-key.js +6 -0
  83. package/dist/agents/session-key.js.map +1 -0
  84. package/dist/agents/shared.cjs +2 -2
  85. package/dist/agents/shared.cjs.map +1 -1
  86. package/dist/agents/shared.d.cts +4 -4
  87. package/dist/agents/shared.d.mts +4 -4
  88. package/dist/agents/shared.d.ts +4 -4
  89. package/dist/agents/skills.cjs.map +1 -1
  90. package/dist/agents/skills.d.cts +3 -3
  91. package/dist/agents/skills.d.mts +3 -3
  92. package/dist/agents/skills.d.ts +3 -3
  93. package/dist/agents/types.cjs.map +1 -1
  94. package/dist/agents/types.d.cts +55 -54
  95. package/dist/agents/types.d.mts +55 -54
  96. package/dist/agents/types.d.ts +55 -54
  97. package/dist/agents/ui-message-stream.cjs +4 -4
  98. package/dist/agents/ui-message-stream.cjs.map +1 -1
  99. package/dist/agents/unrecognized-events.cjs +32 -0
  100. package/dist/agents/unrecognized-events.cjs.map +1 -0
  101. package/dist/agents/unrecognized-events.d.cts +8 -0
  102. package/dist/agents/unrecognized-events.d.mts +8 -0
  103. package/dist/agents/unrecognized-events.d.ts +8 -0
  104. package/dist/agents/unrecognized-events.js +6 -0
  105. package/dist/agents/unrecognized-events.js.map +1 -0
  106. package/dist/assertions/api.cjs.map +1 -1
  107. package/dist/assertions/api.d.cts +3 -0
  108. package/dist/assertions/api.d.mts +3 -0
  109. package/dist/assertions/api.d.ts +3 -0
  110. package/dist/assertions/collection.cjs +1 -1
  111. package/dist/assertions/collection.cjs.map +1 -1
  112. package/dist/assertions/judge.cjs +1 -1
  113. package/dist/assertions/judge.cjs.map +1 -1
  114. package/dist/assertions/match.cjs +52 -21
  115. package/dist/assertions/match.cjs.map +1 -1
  116. package/dist/assertions/match.d.cts +16 -5
  117. package/dist/assertions/match.d.mts +16 -5
  118. package/dist/assertions/match.d.ts +16 -5
  119. package/dist/assertions/match.js +96 -94
  120. package/dist/assertions/measurements.cjs +61 -0
  121. package/dist/assertions/measurements.cjs.map +1 -0
  122. package/dist/assertions/measurements.d.cts +24 -0
  123. package/dist/assertions/measurements.d.mts +24 -0
  124. package/dist/assertions/measurements.d.ts +24 -0
  125. package/dist/assertions/measurements.js +8 -0
  126. package/dist/assertions/measurements.js.map +1 -0
  127. package/dist/assertions/numeric.cjs +2 -0
  128. package/dist/assertions/numeric.cjs.map +1 -1
  129. package/dist/assertions/numeric.d.cts +1 -1
  130. package/dist/assertions/numeric.d.mts +1 -1
  131. package/dist/assertions/numeric.d.ts +1 -1
  132. package/dist/assertions/quantity.cjs +42 -0
  133. package/dist/assertions/quantity.cjs.map +1 -0
  134. package/dist/assertions/quantity.d.cts +18 -0
  135. package/dist/assertions/quantity.d.mts +18 -0
  136. package/dist/assertions/quantity.d.ts +18 -0
  137. package/dist/assertions/quantity.js +22 -0
  138. package/dist/assertions/quantity.js.map +1 -0
  139. package/dist/assertions/record/attachment.cjs +2 -0
  140. package/dist/assertions/record/attachment.cjs.map +1 -1
  141. package/dist/assertions/record/codec.cjs +5 -1
  142. package/dist/assertions/record/codec.cjs.map +1 -1
  143. package/dist/assertions/record/model.cjs.map +1 -1
  144. package/dist/assertions/record/model.d.cts +3 -1
  145. package/dist/assertions/record/model.d.mts +3 -1
  146. package/dist/assertions/record/model.d.ts +3 -1
  147. package/dist/assertions/runtime.cjs +11 -6
  148. package/dist/assertions/runtime.cjs.map +1 -1
  149. package/dist/cli/program.cjs +3 -0
  150. package/dist/cli/program.cjs.map +1 -1
  151. package/dist/context/assert-first.cjs +6 -4
  152. package/dist/context/assert-first.cjs.map +1 -1
  153. package/dist/context/assert-first.d.cts +9 -3
  154. package/dist/context/assert-first.d.mts +9 -3
  155. package/dist/context/assert-first.d.ts +9 -3
  156. package/dist/context/notes.cjs +47 -0
  157. package/dist/context/notes.cjs.map +1 -0
  158. package/dist/context/notes.d.cts +10 -0
  159. package/dist/context/notes.d.mts +10 -0
  160. package/dist/context/notes.d.ts +10 -0
  161. package/dist/context/notes.js +6 -0
  162. package/dist/context/notes.js.map +1 -0
  163. package/dist/context/send-failures.cjs +1 -7
  164. package/dist/context/send-failures.cjs.map +1 -1
  165. package/dist/context/send-failures.d.cts +2 -3
  166. package/dist/context/send-failures.d.mts +2 -3
  167. package/dist/context/send-failures.d.ts +2 -3
  168. package/dist/context/send-retry.cjs +0 -1
  169. package/dist/context/send-retry.cjs.map +1 -1
  170. package/dist/context/send-retry.d.cts +1 -2
  171. package/dist/context/send-retry.d.mts +1 -2
  172. package/dist/context/send-retry.d.ts +1 -2
  173. package/dist/context/session.cjs +10 -17
  174. package/dist/context/session.cjs.map +1 -1
  175. package/dist/context/session.d.cts +4 -10
  176. package/dist/context/session.d.mts +4 -10
  177. package/dist/context/session.d.ts +4 -10
  178. package/dist/define.cjs +89 -64
  179. package/dist/define.cjs.map +1 -1
  180. package/dist/define.d.cts +3 -3
  181. package/dist/define.d.mts +3 -3
  182. package/dist/define.d.ts +3 -3
  183. package/dist/eval/cli/contribution.cjs +1 -1
  184. package/dist/eval/cli/contribution.cjs.map +1 -1
  185. package/dist/eval/host/index.cjs +1 -1
  186. package/dist/eval/host/index.cjs.map +1 -1
  187. package/dist/eval/record/evaluation.cjs +1 -1
  188. package/dist/eval/record/evaluation.cjs.map +1 -1
  189. package/dist/eval/record/membership-provenance.d.cts +3 -3
  190. package/dist/eval/record/membership-provenance.d.mts +3 -3
  191. package/dist/eval/record/membership-provenance.d.ts +3 -3
  192. package/dist/eval/record/score.cjs +2 -2
  193. package/dist/eval/record/score.cjs.map +1 -1
  194. package/dist/eval/record/sealed-assertion.cjs +1 -1
  195. package/dist/eval/record/sealed-assertion.cjs.map +1 -1
  196. package/dist/eval/record/verdict.cjs +2 -2
  197. package/dist/eval/record/verdict.cjs.map +1 -1
  198. package/dist/expect/index.cjs +13 -1
  199. package/dist/expect/index.cjs.map +1 -1
  200. package/dist/expect/index.d.cts +5 -1
  201. package/dist/expect/index.d.mts +5 -1
  202. package/dist/expect/index.d.ts +5 -1
  203. package/dist/expect/index.js +92 -72
  204. package/dist/expect/index.mjs +92 -72
  205. package/dist/experiment/host/cli/contribution.cjs +52 -39
  206. package/dist/experiment/host/cli/contribution.cjs.map +1 -1
  207. package/dist/experiment/host/cli/contribution.d.cts +18 -4
  208. package/dist/experiment/host/cli/contribution.d.mts +18 -4
  209. package/dist/experiment/host/cli/contribution.d.ts +18 -4
  210. package/dist/experiment/host/cli/contribution.js +2 -2
  211. package/dist/experiment/host/cli/debug-protocol.cjs +2 -2
  212. package/dist/experiment/host/cli/debug-protocol.cjs.map +1 -1
  213. package/dist/experiment/host/cli/debug-protocol.d.cts +2 -2
  214. package/dist/experiment/host/cli/debug-protocol.d.mts +2 -2
  215. package/dist/experiment/host/cli/debug-protocol.d.ts +2 -2
  216. package/dist/experiment/host/cli/output-protocol.cjs +1 -1
  217. package/dist/experiment/host/cli/output-protocol.cjs.map +1 -1
  218. package/dist/experiment/host/cli/output-protocol.d.cts +2 -2
  219. package/dist/experiment/host/cli/output-protocol.d.mts +2 -2
  220. package/dist/experiment/host/cli/output-protocol.d.ts +2 -2
  221. package/dist/experiment/host/current.cjs +2 -2
  222. package/dist/experiment/host/current.cjs.map +1 -1
  223. package/dist/experiment/host/index.cjs +1 -1
  224. package/dist/experiment/host/index.cjs.map +1 -1
  225. package/dist/experiment/host/operations.cjs +24 -23
  226. package/dist/experiment/host/operations.cjs.map +1 -1
  227. package/dist/experiment/host/runtime.cjs +11 -10
  228. package/dist/experiment/host/runtime.cjs.map +1 -1
  229. package/dist/experiment/host/runtime.d.cts +3 -3
  230. package/dist/experiment/host/runtime.d.mts +3 -3
  231. package/dist/experiment/host/runtime.d.ts +3 -3
  232. package/dist/experiment/host/teardown.cjs +4 -3
  233. package/dist/experiment/host/teardown.cjs.map +1 -1
  234. package/dist/experiment/host/teardown.d.cts +3 -3
  235. package/dist/experiment/host/teardown.d.mts +3 -3
  236. package/dist/experiment/host/teardown.d.ts +3 -3
  237. package/dist/experiment/host/types.cjs.map +1 -1
  238. package/dist/experiment/host/types.d.cts +19 -19
  239. package/dist/experiment/host/types.d.mts +19 -19
  240. package/dist/experiment/host/types.d.ts +19 -19
  241. package/dist/index.cjs +3 -19
  242. package/dist/index.cjs.map +1 -1
  243. package/dist/index.d.cts +4 -8
  244. package/dist/index.d.mts +4 -8
  245. package/dist/index.d.ts +4 -8
  246. package/dist/index.js +12 -34
  247. package/dist/index.mjs +12 -34
  248. package/dist/inspection/assertions.cjs +1 -1
  249. package/dist/inspection/assertions.cjs.map +1 -1
  250. package/dist/inspection/cli/contribution.cjs +8 -7
  251. package/dist/inspection/cli/contribution.cjs.map +1 -1
  252. package/dist/inspection/facts.cjs +4 -4
  253. package/dist/inspection/facts.cjs.map +1 -1
  254. package/dist/inspection/index.cjs +1 -5
  255. package/dist/inspection/index.cjs.map +1 -1
  256. package/dist/inspection/index.d.cts +1 -2
  257. package/dist/inspection/index.d.mts +1 -2
  258. package/dist/inspection/index.d.ts +1 -2
  259. package/dist/inspection/index.js +20 -26
  260. package/dist/inspection/lifecycle.cjs +7 -0
  261. package/dist/inspection/lifecycle.cjs.map +1 -0
  262. package/dist/inspection/lifecycle.d.cts +4 -0
  263. package/dist/inspection/lifecycle.d.mts +4 -0
  264. package/dist/inspection/lifecycle.d.ts +4 -0
  265. package/dist/inspection/lifecycle.js +6 -0
  266. package/dist/inspection/lifecycle.js.map +1 -0
  267. package/dist/inspection/notes.cjs +40 -0
  268. package/dist/inspection/notes.cjs.map +1 -0
  269. package/dist/inspection/notes.d.cts +12 -0
  270. package/dist/inspection/notes.d.mts +12 -0
  271. package/dist/inspection/notes.d.ts +12 -0
  272. package/dist/inspection/notes.js +8 -0
  273. package/dist/inspection/notes.js.map +1 -0
  274. package/dist/inspection/overview.cjs +12 -15
  275. package/dist/inspection/overview.cjs.map +1 -1
  276. package/dist/inspection/overview.d.cts +8 -6
  277. package/dist/inspection/overview.d.mts +8 -6
  278. package/dist/inspection/overview.d.ts +8 -6
  279. package/dist/inspection/planned-attempt.cjs +45 -0
  280. package/dist/inspection/planned-attempt.cjs.map +1 -0
  281. package/dist/inspection/planned-attempt.d.cts +109 -0
  282. package/dist/inspection/planned-attempt.d.mts +109 -0
  283. package/dist/inspection/planned-attempt.d.ts +109 -0
  284. package/dist/inspection/planned-attempt.js +24 -0
  285. package/dist/inspection/planned-attempt.js.map +1 -0
  286. package/dist/inspection/project-result.cjs +5 -5
  287. package/dist/inspection/project-result.cjs.map +1 -1
  288. package/dist/inspection/project-result.d.cts +58 -57
  289. package/dist/inspection/project-result.d.mts +58 -57
  290. package/dist/inspection/project-result.d.ts +58 -57
  291. package/dist/inspection/project-result.js +4 -4
  292. package/dist/inspection/project.cjs +3 -3
  293. package/dist/inspection/project.cjs.map +1 -1
  294. package/dist/inspection/protocol-values.cjs +1 -1
  295. package/dist/inspection/protocol-values.cjs.map +1 -1
  296. package/dist/inspection/protocol-values.d.cts +1 -1
  297. package/dist/inspection/protocol-values.d.mts +1 -1
  298. package/dist/inspection/protocol-values.d.ts +1 -1
  299. package/dist/inspection/protocol.cjs +8 -8
  300. package/dist/inspection/protocol.cjs.map +1 -1
  301. package/dist/inspection/protocol.d.cts +237 -219
  302. package/dist/inspection/protocol.d.mts +237 -219
  303. package/dist/inspection/protocol.d.ts +237 -219
  304. package/dist/inspection/public.cjs +1 -5
  305. package/dist/inspection/public.cjs.map +1 -1
  306. package/dist/inspection/public.d.cts +1 -2
  307. package/dist/inspection/public.d.mts +1 -2
  308. package/dist/inspection/public.d.ts +1 -2
  309. package/dist/inspection/public.js +14 -20
  310. package/dist/inspection/public.mjs +166 -129
  311. package/dist/inspection/query.cjs +1 -1
  312. package/dist/inspection/query.cjs.map +1 -1
  313. package/dist/inspection/results.cjs +61 -52
  314. package/dist/inspection/results.cjs.map +1 -1
  315. package/dist/inspection/results.d.cts +157 -140
  316. package/dist/inspection/results.d.mts +157 -140
  317. package/dist/inspection/results.d.ts +157 -140
  318. package/dist/inspection/select.cjs +33 -28
  319. package/dist/inspection/select.cjs.map +1 -1
  320. package/dist/inspection/source.cjs +1 -1
  321. package/dist/inspection/source.cjs.map +1 -1
  322. package/dist/inspection/sources.cjs +2 -2
  323. package/dist/inspection/sources.cjs.map +1 -1
  324. package/dist/inspection/trace.cjs +9 -8
  325. package/dist/inspection/trace.cjs.map +1 -1
  326. package/dist/inspection/trace.d.cts +4 -4
  327. package/dist/inspection/trace.d.mts +4 -4
  328. package/dist/inspection/trace.d.ts +4 -4
  329. package/dist/loaders/index.cjs +7 -279
  330. package/dist/loaders/index.cjs.map +1 -1
  331. package/dist/loaders/index.d.cts +2 -95
  332. package/dist/loaders/index.d.mts +2 -95
  333. package/dist/loaders/index.d.ts +2 -95
  334. package/dist/loaders/index.js +10 -12
  335. package/dist/loaders/index.mjs +10 -12
  336. package/dist/loaders/runtime.cjs +283 -0
  337. package/dist/loaders/runtime.cjs.map +1 -0
  338. package/dist/loaders/runtime.d.cts +95 -0
  339. package/dist/loaders/runtime.d.mts +95 -0
  340. package/dist/loaders/runtime.d.ts +95 -0
  341. package/dist/loaders/runtime.js +16 -0
  342. package/dist/loaders/runtime.js.map +1 -0
  343. package/dist/migrations/judge-provider.md +6 -6
  344. package/dist/{model-slots.cjs → model-roles.cjs} +26 -7
  345. package/dist/model-roles.cjs.map +1 -0
  346. package/dist/model-roles.d.cts +26 -0
  347. package/dist/model-roles.d.mts +26 -0
  348. package/dist/model-roles.d.ts +26 -0
  349. package/dist/model-roles.js +14 -0
  350. package/dist/model-roles.js.map +1 -0
  351. package/dist/o11y/adapter-usage-projection.cjs +9 -8
  352. package/dist/o11y/adapter-usage-projection.cjs.map +1 -1
  353. package/dist/o11y/adapter-usage-projection.d.cts +23 -15
  354. package/dist/o11y/adapter-usage-projection.d.mts +23 -15
  355. package/dist/o11y/adapter-usage-projection.d.ts +23 -15
  356. package/dist/o11y/cache-read.cjs +14 -14
  357. package/dist/o11y/cache-read.cjs.map +1 -1
  358. package/dist/o11y/cache-read.d.cts +5 -5
  359. package/dist/o11y/cache-read.d.mts +5 -5
  360. package/dist/o11y/cache-read.d.ts +5 -5
  361. package/dist/o11y/derive.cjs.map +1 -1
  362. package/dist/o11y/derive.d.cts +2 -2
  363. package/dist/o11y/derive.d.mts +2 -2
  364. package/dist/o11y/derive.d.ts +2 -2
  365. package/dist/o11y/eval-usage.cjs +9 -9
  366. package/dist/o11y/eval-usage.cjs.map +1 -1
  367. package/dist/o11y/eval-usage.d.cts +7 -7
  368. package/dist/o11y/eval-usage.d.mts +7 -7
  369. package/dist/o11y/eval-usage.d.ts +7 -7
  370. package/dist/o11y/parsers/bub.cjs +57 -81
  371. package/dist/o11y/parsers/bub.cjs.map +1 -1
  372. package/dist/o11y/parsers/index.cjs.map +1 -1
  373. package/dist/o11y/parsers/index.d.cts +2 -0
  374. package/dist/o11y/parsers/index.d.mts +2 -0
  375. package/dist/o11y/parsers/index.d.ts +2 -0
  376. package/dist/o11y/parsers/opencode.cjs +129 -217
  377. package/dist/o11y/parsers/opencode.cjs.map +1 -1
  378. package/dist/o11y/parsers/opencode.d.cts +2 -5
  379. package/dist/o11y/parsers/opencode.d.mts +2 -5
  380. package/dist/o11y/parsers/opencode.d.ts +2 -5
  381. package/dist/o11y/types.cjs.map +1 -1
  382. package/dist/o11y/types.d.cts +1 -1
  383. package/dist/o11y/types.d.mts +1 -1
  384. package/dist/o11y/types.d.ts +1 -1
  385. package/dist/plugin/contracts.cjs +170 -120
  386. package/dist/plugin/contracts.cjs.map +1 -1
  387. package/dist/plugin/contracts.d.cts +102 -90
  388. package/dist/plugin/contracts.d.mts +102 -90
  389. package/dist/plugin/contracts.d.ts +102 -90
  390. package/dist/plugin/contracts.js +16 -12
  391. package/dist/plugin/index.cjs.map +1 -1
  392. package/dist/plugin/index.d.cts +1 -1
  393. package/dist/plugin/index.d.mts +1 -1
  394. package/dist/plugin/index.d.ts +1 -1
  395. package/dist/plugin/link.cjs +69 -64
  396. package/dist/plugin/link.cjs.map +1 -1
  397. package/dist/plugin/link.d.cts +5 -3
  398. package/dist/plugin/link.d.mts +5 -3
  399. package/dist/plugin/link.d.ts +5 -3
  400. package/dist/plugin/link.js +12 -8
  401. package/dist/record/codec/seal-manifest.d.cts +3 -3
  402. package/dist/record/codec/seal-manifest.d.mts +3 -3
  403. package/dist/record/codec/seal-manifest.d.ts +3 -3
  404. package/dist/record/family/attempt-note/definition.cjs +12 -0
  405. package/dist/record/family/attempt-note/definition.cjs.map +1 -0
  406. package/dist/record/family/attempt-note/definition.d.cts +6 -0
  407. package/dist/record/family/attempt-note/definition.d.mts +6 -0
  408. package/dist/record/family/attempt-note/definition.d.ts +6 -0
  409. package/dist/record/family/attempt-note/definition.js +6 -0
  410. package/dist/record/family/attempt-note/definition.js.map +1 -0
  411. package/dist/record/family/attempt-note/persistence.cjs +10 -0
  412. package/dist/record/family/attempt-note/persistence.cjs.map +1 -0
  413. package/dist/record/family/attempt-note/persistence.d.cts +6 -0
  414. package/dist/record/family/attempt-note/persistence.d.mts +6 -0
  415. package/dist/record/family/attempt-note/persistence.d.ts +6 -0
  416. package/dist/record/family/attempt-note/persistence.js +6 -0
  417. package/dist/record/family/attempt-note/persistence.js.map +1 -0
  418. package/dist/record/family/attempt-note/schema.cjs +20 -0
  419. package/dist/record/family/attempt-note/schema.cjs.map +1 -0
  420. package/dist/record/family/attempt-note/schema.d.cts +21 -0
  421. package/dist/record/family/attempt-note/schema.d.mts +21 -0
  422. package/dist/record/family/attempt-note/schema.d.ts +21 -0
  423. package/dist/record/family/attempt-note/schema.js +10 -0
  424. package/dist/record/family/attempt-note/schema.js.map +1 -0
  425. package/dist/record/family/catalog.cjs +32 -29
  426. package/dist/record/family/catalog.cjs.map +1 -1
  427. package/dist/record/family/catalog.d.cts +7 -2
  428. package/dist/record/family/catalog.d.mts +7 -2
  429. package/dist/record/family/catalog.d.ts +7 -2
  430. package/dist/record/family/current.cjs +48 -45
  431. package/dist/record/family/current.cjs.map +1 -1
  432. package/dist/record/family/current.d.cts +15 -0
  433. package/dist/record/family/current.d.mts +15 -0
  434. package/dist/record/family/current.d.ts +15 -0
  435. package/dist/record/model/index.js +66 -64
  436. package/dist/record/model/run-context.cjs +9 -3
  437. package/dist/record/model/run-context.cjs.map +1 -1
  438. package/dist/record/model/run-context.d.cts +9 -2
  439. package/dist/record/model/run-context.d.mts +9 -2
  440. package/dist/record/model/run-context.d.ts +9 -2
  441. package/dist/record/model/run-context.js +8 -6
  442. package/dist/record/model/seal-manifest.d.cts +1 -1
  443. package/dist/record/model/seal-manifest.d.mts +1 -1
  444. package/dist/record/model/seal-manifest.d.ts +1 -1
  445. package/dist/run/cli/contribution.cjs +6 -32
  446. package/dist/run/cli/contribution.cjs.map +1 -1
  447. package/dist/run/host/runtime.cjs +1 -1
  448. package/dist/run/host/runtime.cjs.map +1 -1
  449. package/dist/run/storage/sqlite.cjs +7 -7
  450. package/dist/run/storage/sqlite.cjs.map +1 -1
  451. package/dist/runner/adapter-usage.cjs +5 -3
  452. package/dist/runner/adapter-usage.cjs.map +1 -1
  453. package/dist/runner/adapter-usage.d.cts +2 -2
  454. package/dist/runner/adapter-usage.d.mts +2 -2
  455. package/dist/runner/adapter-usage.d.ts +2 -2
  456. package/dist/runner/adoption.cjs +10 -12
  457. package/dist/runner/adoption.cjs.map +1 -1
  458. package/dist/runner/attempt-notes.cjs +8 -0
  459. package/dist/runner/attempt-notes.cjs.map +1 -0
  460. package/dist/runner/attempt-notes.d.cts +4 -0
  461. package/dist/runner/attempt-notes.d.mts +4 -0
  462. package/dist/runner/attempt-notes.d.ts +4 -0
  463. package/dist/runner/attempt-notes.js +8 -0
  464. package/dist/runner/attempt-notes.js.map +1 -0
  465. package/dist/runner/attempt.cjs +61 -73
  466. package/dist/runner/attempt.cjs.map +1 -1
  467. package/dist/runner/attempt.d.cts +3 -3
  468. package/dist/runner/attempt.d.mts +3 -3
  469. package/dist/runner/attempt.d.ts +3 -3
  470. package/dist/runner/command-plan.cjs +21 -26
  471. package/dist/runner/command-plan.cjs.map +1 -1
  472. package/dist/runner/command-plan.d.cts +4 -4
  473. package/dist/runner/command-plan.d.mts +4 -4
  474. package/dist/runner/command-plan.d.ts +4 -4
  475. package/dist/runner/command-plan.js +4 -4
  476. package/dist/runner/config-identity.cjs.map +1 -1
  477. package/dist/runner/config-identity.d.cts +2 -2
  478. package/dist/runner/config-identity.d.mts +2 -2
  479. package/dist/runner/config-identity.d.ts +2 -2
  480. package/dist/runner/discover.cjs +6 -6
  481. package/dist/runner/discover.cjs.map +1 -1
  482. package/dist/runner/eval-selection.cjs +15 -1
  483. package/dist/runner/eval-selection.cjs.map +1 -1
  484. package/dist/runner/eval-selection.d.cts +4 -0
  485. package/dist/runner/eval-selection.d.mts +4 -0
  486. package/dist/runner/eval-selection.d.ts +4 -0
  487. package/dist/runner/feedback/human.cjs +1 -1
  488. package/dist/runner/feedback/human.cjs.map +1 -1
  489. package/dist/runner/feedback/json.cjs +4 -2
  490. package/dist/runner/feedback/json.cjs.map +1 -1
  491. package/dist/runner/feedback/json.d.cts +1 -1
  492. package/dist/runner/feedback/json.d.mts +1 -1
  493. package/dist/runner/feedback/json.d.ts +1 -1
  494. package/dist/runner/feedback/sink.cjs +1 -1
  495. package/dist/runner/feedback/sink.cjs.map +1 -1
  496. package/dist/runner/feedback/sink.d.cts +1 -1
  497. package/dist/runner/feedback/sink.d.mts +1 -1
  498. package/dist/runner/feedback/sink.d.ts +1 -1
  499. package/dist/runner/fingerprint.cjs +1 -1
  500. package/dist/runner/fingerprint.cjs.map +1 -1
  501. package/dist/runner/lock.cjs +1 -3
  502. package/dist/runner/lock.cjs.map +1 -1
  503. package/dist/runner/lock.d.cts +0 -2
  504. package/dist/runner/lock.d.mts +0 -2
  505. package/dist/runner/lock.d.ts +0 -2
  506. package/dist/runner/lock.js +14 -16
  507. package/dist/runner/project-target.cjs.map +1 -1
  508. package/dist/runner/project-target.d.cts +2 -2
  509. package/dist/runner/project-target.d.mts +2 -2
  510. package/dist/runner/project-target.d.ts +2 -2
  511. package/dist/runner/record/attachments.cjs +2 -2
  512. package/dist/runner/record/attachments.cjs.map +1 -1
  513. package/dist/runner/record/planning.cjs +8 -2
  514. package/dist/runner/record/planning.cjs.map +1 -1
  515. package/dist/runner/record.cjs +2 -0
  516. package/dist/runner/record.cjs.map +1 -1
  517. package/dist/runner/report.cjs +3 -3
  518. package/dist/runner/report.cjs.map +1 -1
  519. package/dist/runner/report.d.cts +1 -1
  520. package/dist/runner/report.d.mts +1 -1
  521. package/dist/runner/report.d.ts +1 -1
  522. package/dist/runner/reuse-plan.cjs +2 -2
  523. package/dist/runner/reuse-plan.cjs.map +1 -1
  524. package/dist/runner/run.cjs +10 -45
  525. package/dist/runner/run.cjs.map +1 -1
  526. package/dist/runner/sandbox-pool.cjs +1 -2
  527. package/dist/runner/sandbox-pool.cjs.map +1 -1
  528. package/dist/runner/setup-prefix-preparation.cjs +1 -1
  529. package/dist/runner/setup-prefix-preparation.cjs.map +1 -1
  530. package/dist/runner/teardown-registry.cjs +2 -2
  531. package/dist/runner/teardown-registry.cjs.map +1 -1
  532. package/dist/runner/teardown-registry.d.cts +1 -1
  533. package/dist/runner/teardown-registry.d.mts +1 -1
  534. package/dist/runner/teardown-registry.d.ts +1 -1
  535. package/dist/runner/types.cjs +4 -4
  536. package/dist/runner/types.cjs.map +1 -1
  537. package/dist/runner/types.d.cts +64 -92
  538. package/dist/runner/types.d.mts +64 -92
  539. package/dist/runner/types.d.ts +64 -92
  540. package/dist/runner/types.js +2 -2
  541. package/dist/sandbox/action.cjs +1 -1
  542. package/dist/sandbox/action.cjs.map +1 -1
  543. package/dist/sandbox/cli/contribution.cjs +17 -14
  544. package/dist/sandbox/cli/contribution.cjs.map +1 -1
  545. package/dist/sandbox/cli/contribution.d.cts +1 -1
  546. package/dist/sandbox/cli/contribution.d.mts +1 -1
  547. package/dist/sandbox/cli/contribution.d.ts +1 -1
  548. package/dist/sandbox/commands.cjs +1 -1
  549. package/dist/sandbox/commands.cjs.map +1 -1
  550. package/dist/sandbox/commands.d.cts +1 -1
  551. package/dist/sandbox/commands.d.mts +1 -1
  552. package/dist/sandbox/commands.d.ts +1 -1
  553. package/dist/sandbox/compose.cjs +2 -2
  554. package/dist/sandbox/compose.cjs.map +1 -1
  555. package/dist/sandbox/docker-profile/registry.cjs +5 -10
  556. package/dist/sandbox/docker-profile/registry.cjs.map +1 -1
  557. package/dist/sandbox/docker-profile/registry.d.cts +0 -4
  558. package/dist/sandbox/docker-profile/registry.d.mts +0 -4
  559. package/dist/sandbox/docker-profile/registry.d.ts +0 -4
  560. package/dist/sandbox/docker-profile/registry.js +6 -14
  561. package/dist/sandbox/docker-profile/runtime.cjs +1 -1
  562. package/dist/sandbox/docker-profile/runtime.cjs.map +1 -1
  563. package/dist/sandbox/docker-profile/schema.cjs +12 -16
  564. package/dist/sandbox/docker-profile/schema.cjs.map +1 -1
  565. package/dist/sandbox/docker-profile/schema.d.cts +2 -6
  566. package/dist/sandbox/docker-profile/schema.d.mts +2 -6
  567. package/dist/sandbox/docker-profile/schema.d.ts +2 -6
  568. package/dist/sandbox/docker-profile/schema.js +24 -32
  569. package/dist/sandbox/incus.cjs +2 -2
  570. package/dist/sandbox/incus.cjs.map +1 -1
  571. package/dist/sandbox/index.cjs +4 -36
  572. package/dist/sandbox/index.cjs.map +1 -1
  573. package/dist/sandbox/index.d.cts +4 -8
  574. package/dist/sandbox/index.d.mts +4 -8
  575. package/dist/sandbox/index.d.ts +4 -8
  576. package/dist/sandbox/index.js +116 -174
  577. package/dist/sandbox/index.mjs +116 -174
  578. package/dist/sandbox/layer.cjs +3 -3
  579. package/dist/sandbox/layer.cjs.map +1 -1
  580. package/dist/sandbox/link.cjs +5 -0
  581. package/dist/sandbox/link.cjs.map +1 -1
  582. package/dist/sandbox/link.d.cts +3 -0
  583. package/dist/sandbox/link.d.mts +3 -0
  584. package/dist/sandbox/link.d.ts +3 -0
  585. package/dist/sandbox/runtime.cjs +9 -4
  586. package/dist/sandbox/runtime.cjs.map +1 -1
  587. package/dist/shared/failure-class.cjs +5 -5
  588. package/dist/shared/failure-class.cjs.map +1 -1
  589. package/dist/shared/failure-class.d.cts +4 -10
  590. package/dist/shared/failure-class.d.mts +4 -10
  591. package/dist/shared/failure-class.d.ts +4 -10
  592. package/dist/shared/types.cjs.map +1 -1
  593. package/dist/shared/types.d.cts +1 -1
  594. package/dist/shared/types.d.mts +1 -1
  595. package/dist/shared/types.d.ts +1 -1
  596. package/dist/show/contribution.cjs +17 -2
  597. package/dist/show/contribution.cjs.map +1 -1
  598. package/dist/show/contribution.d.cts +14 -0
  599. package/dist/show/contribution.d.mts +14 -0
  600. package/dist/show/contribution.d.ts +14 -0
  601. package/dist/show/model.cjs +8 -5
  602. package/dist/show/model.cjs.map +1 -1
  603. package/dist/show/model.d.cts +11 -8
  604. package/dist/show/model.d.mts +11 -8
  605. package/dist/show/model.d.ts +11 -8
  606. package/dist/show/render.cjs +34 -22
  607. package/dist/show/render.cjs.map +1 -1
  608. package/dist/sources/projector.cjs +1 -1
  609. package/dist/sources/projector.cjs.map +1 -1
  610. package/dist/types.cjs +1 -1
  611. package/dist/types.cjs.map +1 -1
  612. package/dist/types.d.cts +1 -1
  613. package/dist/types.d.mts +1 -1
  614. package/dist/types.d.ts +1 -1
  615. package/dist/types.js +2 -2
  616. package/dist/view/app-dist/.vite/manifest.json +2 -2
  617. package/dist/view/app-dist/assets/{index-B3ijZRph.js → index-CxrOUvHC.js} +17 -17
  618. package/dist/view/app-dist/assets/{quality-cost-scatter-B--DfJ9s.js → quality-cost-scatter-B6Aiss7x.js} +1 -1
  619. package/dist/view/app-dist/index.html +1 -1
  620. package/dist/view/cli/contribution.cjs +1 -1
  621. package/dist/view/cli/contribution.cjs.map +1 -1
  622. package/dist/view/function-runtime.mjs +318 -121
  623. package/dist/view/http-protocol.d.cts +1 -1
  624. package/dist/view/http-protocol.d.mts +1 -1
  625. package/dist/view/http-protocol.d.ts +1 -1
  626. package/dist/view/inspection-host.cjs +2 -2
  627. package/dist/view/inspection-host.cjs.map +1 -1
  628. package/dist/view/inspection-host.d.cts +349 -331
  629. package/dist/view/inspection-host.d.mts +349 -331
  630. package/dist/view/inspection-host.d.ts +349 -331
  631. package/docs-site/zh/examples/integrations/ai-sdk-v7.mdx +9 -9
  632. package/docs-site/zh/examples/integrations/claude-sdk.mdx +8 -8
  633. package/docs-site/zh/examples/integrations/codex-sdk.mdx +4 -4
  634. package/docs-site/zh/examples/integrations/langgraph.mdx +6 -6
  635. package/docs-site/zh/examples/integrations/pi-sdk.mdx +8 -8
  636. package/docs-site/zh/explanation/adapter.mdx +6 -6
  637. package/docs-site/zh/explanation/assert.mdx +7 -7
  638. package/docs-site/zh/explanation/drive.mdx +4 -4
  639. package/docs-site/zh/explanation/hitl.mdx +2 -2
  640. package/docs-site/zh/explanation/judge.mdx +3 -3
  641. package/docs-site/zh/index.mdx +18 -0
  642. package/docs-site/zh/introduction.mdx +4 -3
  643. package/docs-site/zh/reference/builtin-agents.mdx +46 -49
  644. package/docs-site/zh/reference/cli.mdx +16 -14
  645. package/docs-site/zh/reference/define-agent.mdx +19 -19
  646. package/docs-site/zh/reference/define-config.mdx +7 -15
  647. package/docs-site/zh/reference/define-eval.mdx +20 -14
  648. package/docs-site/zh/reference/define-experiment.mdx +18 -39
  649. package/docs-site/zh/reference/events.mdx +1 -1
  650. package/docs-site/zh/reference/expect.mdx +21 -0
  651. package/docs-site/zh/reference/official-adapters.mdx +12 -5
  652. package/docs-site/zh/troubleshooting/debugging.mdx +2 -2
  653. package/docs-site/zh/troubleshooting/recover-after-kill.mdx +1 -1
  654. package/docs-site/zh/tutorials/agent-onboarding.mdx +2 -2
  655. package/docs-site/zh/tutorials/authoring.mdx +4 -3
  656. package/docs-site/zh/tutorials/compare-runs.mdx +1 -1
  657. package/docs-site/zh/tutorials/concurrency.mdx +8 -8
  658. package/docs-site/zh/tutorials/configuration.mdx +5 -5
  659. package/docs-site/zh/tutorials/control-run-cost.mdx +49 -9
  660. package/docs-site/zh/tutorials/eval-groups.mdx +1 -1
  661. package/docs-site/zh/tutorials/evaluation-kinds.mdx +5 -5
  662. package/docs-site/zh/tutorials/experiment-services.mdx +32 -32
  663. package/docs-site/zh/tutorials/handle-execution-failures.mdx +2 -28
  664. package/docs-site/zh/tutorials/plugins.mdx +49 -143
  665. package/docs-site/zh/tutorials/sandbox-agent.mdx +1 -1
  666. package/docs-site/zh/tutorials/sandbox-reuse.mdx +7 -7
  667. package/docs-site/zh/tutorials/upgrade-0-18.mdx +64 -0
  668. package/docs-site/zh/tutorials/upgrade-0-19.mdx +121 -0
  669. package/docs-site/zh/tutorials/upgrade-0-20.mdx +12 -0
  670. package/docs-site/zh/tutorials/verify-judge.mdx +4 -3
  671. package/docs-site/zh/tutorials/viewing-results.mdx +2 -2
  672. package/docs-site/zh/tutorials/write-experiment.mdx +1 -1
  673. package/docs-site/zh/tutorials/write-send.mdx +13 -13
  674. package/package.json +1 -1
  675. package/dist/agents/effect-runtime.cjs.map +0 -1
  676. package/dist/agents/effect-runtime.d.cts +0 -29
  677. package/dist/agents/effect-runtime.d.mts +0 -29
  678. package/dist/agents/effect-runtime.d.ts +0 -29
  679. package/dist/agents/effect-runtime.js +0 -16
  680. package/dist/agents/effect-runtime.js.map +0 -1
  681. package/dist/agents/session-slot.cjs.map +0 -1
  682. package/dist/agents/session-slot.d.cts +0 -12
  683. package/dist/agents/session-slot.d.mts +0 -12
  684. package/dist/agents/session-slot.d.ts +0 -12
  685. package/dist/agents/session-slot.js +0 -6
  686. package/dist/agents/session-slot.js.map +0 -1
  687. package/dist/model-slots.cjs.map +0 -1
  688. package/dist/model-slots.d.cts +0 -19
  689. package/dist/model-slots.d.mts +0 -19
  690. package/dist/model-slots.d.ts +0 -19
  691. package/dist/model-slots.js +0 -8
  692. package/dist/model-slots.js.map +0 -1
@@ -1,4 +1,4 @@
1
- import type { ResolvedModelSlots } from "../../model-slots.mjs";
1
+ import type { ResolvedModelRoles } from "../../model-roles.mjs";
2
2
  import { type Effect } from "effect";
3
3
  import type { Config } from "../../types.mjs";
4
4
  import type { FeedbackCoordinator } from "../../runner/feedback/coordinator.mjs";
@@ -38,7 +38,7 @@ export interface ExperimentHostEvalSummary {
38
38
  readonly evaluationKind: "pass" | "score";
39
39
  }
40
40
  export interface ExperimentHostExperimentSummary {
41
- readonly models?: ResolvedModelSlots;
41
+ readonly models?: ResolvedModelRoles;
42
42
  readonly id: string;
43
43
  readonly description?: string;
44
44
  readonly adapter: string;
@@ -46,8 +46,8 @@ export interface ExperimentHostExperimentSummary {
46
46
  readonly attempts: number;
47
47
  readonly evalIds: readonly string[];
48
48
  readonly labels: Readonly<Record<string, string | number>>;
49
- readonly hasSetup: boolean;
50
- readonly hasTeardown: boolean;
49
+ readonly hasBefore: boolean;
50
+ readonly hasAfter: boolean;
51
51
  readonly sharedStateKey?: string;
52
52
  }
53
53
  export interface ExperimentHostCatalog {
@@ -79,7 +79,7 @@ export interface ExperimentHostRunOverrides {
79
79
  readonly keepSandbox?: "failed" | "all";
80
80
  readonly maxConcurrency?: number;
81
81
  readonly maxBuildConcurrency?: number;
82
- readonly maxSetupPrefixConcurrency?: number;
82
+ readonly maxBeforeCacheConcurrency?: number;
83
83
  }
84
84
  export interface ExperimentHostCheckRequest extends ExperimentHostSelectionInput {
85
85
  readonly config: Config;
@@ -108,7 +108,7 @@ export interface ExperimentHostInvocationShape {
108
108
  readonly attempts: number;
109
109
  readonly maxConcurrency: number;
110
110
  readonly maxBuildConcurrency: number;
111
- readonly maxSetupPrefixConcurrency: number;
111
+ readonly maxBeforeCacheConcurrency: number;
112
112
  readonly experimentConcurrency: Readonly<Record<string, number>>;
113
113
  }
114
114
  export interface ExperimentHostDrySlotTarget {
@@ -217,7 +217,7 @@ export interface ExperimentHostInvocationSummary {
217
217
  readonly inputTokens?: number;
218
218
  readonly outputTokens?: number;
219
219
  readonly estimatedCostUSD?: number;
220
- readonly setupPrefixes: SetupPrefixPreparationSummary;
220
+ readonly beforeCaches: SetupPrefixPreparationSummary;
221
221
  }
222
222
  export interface ExperimentHostInvocationResult {
223
223
  readonly status: "finished";
@@ -313,14 +313,14 @@ export interface ExperimentHostSharedStateEvidence {
313
313
  readonly processIdentity: string;
314
314
  readonly heartbeatAt: string;
315
315
  }
316
- export interface ExperimentHostTeardownInspectRequest {
316
+ export interface ExperimentHostAfterInspectRequest {
317
317
  readonly cwd: string;
318
318
  readonly experimentSelector?: string;
319
319
  readonly coordinationRoot?: string;
320
320
  readonly currentHost: string;
321
321
  readonly recoveryKey?: string;
322
322
  }
323
- export type ExperimentHostTeardownInspection = {
323
+ export type ExperimentHostAfterInspection = {
324
324
  readonly status: "no-evidence";
325
325
  readonly key: string;
326
326
  } | ExperimentHostSelectionProblem | {
@@ -332,7 +332,7 @@ export type ExperimentHostTeardownInspection = {
332
332
  readonly selectedExperimentId: string;
333
333
  readonly evidence: ExperimentHostSharedStateEvidence;
334
334
  } | {
335
- readonly status: "teardown-required";
335
+ readonly status: "after-required";
336
336
  readonly experimentId: string;
337
337
  readonly evidence: ExperimentHostSharedStateEvidence;
338
338
  } | {
@@ -341,14 +341,14 @@ export type ExperimentHostTeardownInspection = {
341
341
  readonly orphanedRegistrations: number;
342
342
  readonly evidence?: ExperimentHostSharedStateEvidence;
343
343
  };
344
- export interface ExperimentHostTeardownRequest extends ExperimentHostTeardownInspectRequest {
345
- readonly observer?: ExperimentHostTeardownObserver;
344
+ export interface ExperimentHostAfterRequest extends ExperimentHostAfterInspectRequest {
345
+ readonly observer?: ExperimentHostAfterObserver;
346
346
  readonly signal?: AbortSignal;
347
347
  readonly ownerToken?: string;
348
348
  readonly confirmOwnerTerminated?: boolean;
349
349
  readonly confirmRemoteQuiesced?: boolean;
350
350
  }
351
- export type ExperimentHostTeardownEvent = {
351
+ export type ExperimentHostAfterEvent = {
352
352
  readonly type: "experiment-progress";
353
353
  readonly experimentId: string;
354
354
  readonly detail: string;
@@ -366,10 +366,10 @@ export type ExperimentHostTeardownEvent = {
366
366
  readonly durationMs?: number;
367
367
  readonly recovery?: true;
368
368
  };
369
- export interface ExperimentHostTeardownObserver {
370
- readonly observe: (event: ExperimentHostTeardownEvent) => void;
369
+ export interface ExperimentHostAfterObserver {
370
+ readonly observe: (event: ExperimentHostAfterEvent) => void;
371
371
  }
372
- export type ExperimentHostTeardownResult = ExperimentHostTeardownInspection | {
372
+ export type ExperimentHostAfterResult = ExperimentHostAfterInspection | {
373
373
  readonly status: "completed";
374
374
  readonly experiments: readonly {
375
375
  readonly experimentId: string;
@@ -405,9 +405,9 @@ export interface ExperimentHostHighLevelSDK {
405
405
  readonly plan: (input: ExperimentHostRenameRequest) => Effect.Effect<ExperimentHostRenamePlan, ExperimentHostError>;
406
406
  readonly apply: (input: ExperimentHostRenameRequest) => Effect.Effect<ExperimentHostRenameResult, ExperimentHostError>;
407
407
  };
408
- readonly teardown: {
409
- readonly inspect: (input: ExperimentHostTeardownInspectRequest) => Effect.Effect<ExperimentHostTeardownInspection, ExperimentHostError>;
410
- readonly run: (input: ExperimentHostTeardownRequest) => Effect.Effect<ExperimentHostTeardownResult, ExperimentHostError>;
408
+ readonly after: {
409
+ readonly inspect: (input: ExperimentHostAfterInspectRequest) => Effect.Effect<ExperimentHostAfterInspection, ExperimentHostError>;
410
+ readonly run: (input: ExperimentHostAfterRequest) => Effect.Effect<ExperimentHostAfterResult, ExperimentHostError>;
411
411
  };
412
412
  readonly accept: (input: ExperimentHostAcceptRequest) => Effect.Effect<readonly ExperimentHostAcceptedAttempt[], ExperimentHostError>;
413
413
  readonly acceptRun: {
@@ -1,4 +1,4 @@
1
- import type { ResolvedModelSlots } from "../../model-slots.js";
1
+ import type { ResolvedModelRoles } from "../../model-roles.js";
2
2
  import { type Effect } from "effect";
3
3
  import type { Config } from "../../types.js";
4
4
  import type { FeedbackCoordinator } from "../../runner/feedback/coordinator.js";
@@ -38,7 +38,7 @@ export interface ExperimentHostEvalSummary {
38
38
  readonly evaluationKind: "pass" | "score";
39
39
  }
40
40
  export interface ExperimentHostExperimentSummary {
41
- readonly models?: ResolvedModelSlots;
41
+ readonly models?: ResolvedModelRoles;
42
42
  readonly id: string;
43
43
  readonly description?: string;
44
44
  readonly adapter: string;
@@ -46,8 +46,8 @@ export interface ExperimentHostExperimentSummary {
46
46
  readonly attempts: number;
47
47
  readonly evalIds: readonly string[];
48
48
  readonly labels: Readonly<Record<string, string | number>>;
49
- readonly hasSetup: boolean;
50
- readonly hasTeardown: boolean;
49
+ readonly hasBefore: boolean;
50
+ readonly hasAfter: boolean;
51
51
  readonly sharedStateKey?: string;
52
52
  }
53
53
  export interface ExperimentHostCatalog {
@@ -79,7 +79,7 @@ export interface ExperimentHostRunOverrides {
79
79
  readonly keepSandbox?: "failed" | "all";
80
80
  readonly maxConcurrency?: number;
81
81
  readonly maxBuildConcurrency?: number;
82
- readonly maxSetupPrefixConcurrency?: number;
82
+ readonly maxBeforeCacheConcurrency?: number;
83
83
  }
84
84
  export interface ExperimentHostCheckRequest extends ExperimentHostSelectionInput {
85
85
  readonly config: Config;
@@ -108,7 +108,7 @@ export interface ExperimentHostInvocationShape {
108
108
  readonly attempts: number;
109
109
  readonly maxConcurrency: number;
110
110
  readonly maxBuildConcurrency: number;
111
- readonly maxSetupPrefixConcurrency: number;
111
+ readonly maxBeforeCacheConcurrency: number;
112
112
  readonly experimentConcurrency: Readonly<Record<string, number>>;
113
113
  }
114
114
  export interface ExperimentHostDrySlotTarget {
@@ -217,7 +217,7 @@ export interface ExperimentHostInvocationSummary {
217
217
  readonly inputTokens?: number;
218
218
  readonly outputTokens?: number;
219
219
  readonly estimatedCostUSD?: number;
220
- readonly setupPrefixes: SetupPrefixPreparationSummary;
220
+ readonly beforeCaches: SetupPrefixPreparationSummary;
221
221
  }
222
222
  export interface ExperimentHostInvocationResult {
223
223
  readonly status: "finished";
@@ -313,14 +313,14 @@ export interface ExperimentHostSharedStateEvidence {
313
313
  readonly processIdentity: string;
314
314
  readonly heartbeatAt: string;
315
315
  }
316
- export interface ExperimentHostTeardownInspectRequest {
316
+ export interface ExperimentHostAfterInspectRequest {
317
317
  readonly cwd: string;
318
318
  readonly experimentSelector?: string;
319
319
  readonly coordinationRoot?: string;
320
320
  readonly currentHost: string;
321
321
  readonly recoveryKey?: string;
322
322
  }
323
- export type ExperimentHostTeardownInspection = {
323
+ export type ExperimentHostAfterInspection = {
324
324
  readonly status: "no-evidence";
325
325
  readonly key: string;
326
326
  } | ExperimentHostSelectionProblem | {
@@ -332,7 +332,7 @@ export type ExperimentHostTeardownInspection = {
332
332
  readonly selectedExperimentId: string;
333
333
  readonly evidence: ExperimentHostSharedStateEvidence;
334
334
  } | {
335
- readonly status: "teardown-required";
335
+ readonly status: "after-required";
336
336
  readonly experimentId: string;
337
337
  readonly evidence: ExperimentHostSharedStateEvidence;
338
338
  } | {
@@ -341,14 +341,14 @@ export type ExperimentHostTeardownInspection = {
341
341
  readonly orphanedRegistrations: number;
342
342
  readonly evidence?: ExperimentHostSharedStateEvidence;
343
343
  };
344
- export interface ExperimentHostTeardownRequest extends ExperimentHostTeardownInspectRequest {
345
- readonly observer?: ExperimentHostTeardownObserver;
344
+ export interface ExperimentHostAfterRequest extends ExperimentHostAfterInspectRequest {
345
+ readonly observer?: ExperimentHostAfterObserver;
346
346
  readonly signal?: AbortSignal;
347
347
  readonly ownerToken?: string;
348
348
  readonly confirmOwnerTerminated?: boolean;
349
349
  readonly confirmRemoteQuiesced?: boolean;
350
350
  }
351
- export type ExperimentHostTeardownEvent = {
351
+ export type ExperimentHostAfterEvent = {
352
352
  readonly type: "experiment-progress";
353
353
  readonly experimentId: string;
354
354
  readonly detail: string;
@@ -366,10 +366,10 @@ export type ExperimentHostTeardownEvent = {
366
366
  readonly durationMs?: number;
367
367
  readonly recovery?: true;
368
368
  };
369
- export interface ExperimentHostTeardownObserver {
370
- readonly observe: (event: ExperimentHostTeardownEvent) => void;
369
+ export interface ExperimentHostAfterObserver {
370
+ readonly observe: (event: ExperimentHostAfterEvent) => void;
371
371
  }
372
- export type ExperimentHostTeardownResult = ExperimentHostTeardownInspection | {
372
+ export type ExperimentHostAfterResult = ExperimentHostAfterInspection | {
373
373
  readonly status: "completed";
374
374
  readonly experiments: readonly {
375
375
  readonly experimentId: string;
@@ -405,9 +405,9 @@ export interface ExperimentHostHighLevelSDK {
405
405
  readonly plan: (input: ExperimentHostRenameRequest) => Effect.Effect<ExperimentHostRenamePlan, ExperimentHostError>;
406
406
  readonly apply: (input: ExperimentHostRenameRequest) => Effect.Effect<ExperimentHostRenameResult, ExperimentHostError>;
407
407
  };
408
- readonly teardown: {
409
- readonly inspect: (input: ExperimentHostTeardownInspectRequest) => Effect.Effect<ExperimentHostTeardownInspection, ExperimentHostError>;
410
- readonly run: (input: ExperimentHostTeardownRequest) => Effect.Effect<ExperimentHostTeardownResult, ExperimentHostError>;
408
+ readonly after: {
409
+ readonly inspect: (input: ExperimentHostAfterInspectRequest) => Effect.Effect<ExperimentHostAfterInspection, ExperimentHostError>;
410
+ readonly run: (input: ExperimentHostAfterRequest) => Effect.Effect<ExperimentHostAfterResult, ExperimentHostError>;
411
411
  };
412
412
  readonly accept: (input: ExperimentHostAcceptRequest) => Effect.Effect<readonly ExperimentHostAcceptedAttempt[], ExperimentHostError>;
413
413
  readonly acceptRun: {
package/dist/index.cjs CHANGED
@@ -1,6 +1,6 @@
1
1
  "use strict";
2
2
  Object.defineProperty(exports, "__esModule", { value: true });
3
- exports.defineContextMatch = exports.defineMaterialMatch = exports.MigrationRequiredError = exports.pairwisePreference = exports.instructionFollowing = exports.faithfulness = exports.factuality = exports.failureClassOf = exports.EvalFatalError = exports.ExperimentFatalError = exports.stripComments = exports.getEnv = exports.requireEnv = exports.readScoreMatchAudit = exports.defineScoreMatch = exports.defineJudge = exports.defineAdapterContract = exports.defineAdapter = exports.defineExperiment = exports.defineConfig = exports.defineEvalGroup = exports.defineScoreEval = exports.defineEval = void 0;
3
+ exports.defineModelRoles = exports.MigrationRequiredError = exports.failureClassOf = exports.EvalFatalError = exports.ExperimentFatalError = exports.defineAdapterContract = exports.defineAdapter = exports.defineExperiment = exports.defineConfig = exports.defineEvalGroup = exports.defineScoreEval = exports.defineEval = void 0;
4
4
  // niceeval 公开导出(import { … } from "niceeval")。
5
5
  // Agent/Adapter 相关见 "niceeval/adapter";Sandbox 相关见 "niceeval/sandbox"。
6
6
  var define_ts_1 = require("./define.cjs");
@@ -12,30 +12,14 @@ Object.defineProperty(exports, "defineExperiment", { enumerable: true, get: func
12
12
  var adapter_ts_1 = require("./adapter.cjs");
13
13
  Object.defineProperty(exports, "defineAdapter", { enumerable: true, get: function () { return adapter_ts_1.defineAdapter; } });
14
14
  Object.defineProperty(exports, "defineAdapterContract", { enumerable: true, get: function () { return adapter_ts_1.defineAdapterContract; } });
15
- var judge_ts_1 = require("./assertions/judge.cjs");
16
- Object.defineProperty(exports, "defineJudge", { enumerable: true, get: function () { return judge_ts_1.defineJudge; } });
17
- var match_ts_1 = require("./assertions/match.cjs");
18
- Object.defineProperty(exports, "defineScoreMatch", { enumerable: true, get: function () { return match_ts_1.defineScoreMatch; } });
19
- var score_match_audit_ts_1 = require("./assertions/score-match-audit.cjs");
20
- Object.defineProperty(exports, "readScoreMatchAudit", { enumerable: true, get: function () { return score_match_audit_ts_1.readScoreMatchAudit; } });
21
- var util_ts_1 = require("./util.cjs");
22
- Object.defineProperty(exports, "requireEnv", { enumerable: true, get: function () { return util_ts_1.requireEnv; } });
23
- Object.defineProperty(exports, "getEnv", { enumerable: true, get: function () { return util_ts_1.getEnv; } });
24
- Object.defineProperty(exports, "stripComments", { enumerable: true, get: function () { return util_ts_1.stripComments; } });
25
15
  // 执行失败分类:抛出点糖衣类(声明死因波及多远)+ 结构守卫。判据、分类链与止损语义见
26
16
  // docs/feature/error-classification/README.md;`niceeval/adapter` 复导出同一份词表类型。
27
17
  var failure_class_ts_1 = require("./shared/failure-class.cjs");
28
18
  Object.defineProperty(exports, "ExperimentFatalError", { enumerable: true, get: function () { return failure_class_ts_1.ExperimentFatalError; } });
29
19
  Object.defineProperty(exports, "EvalFatalError", { enumerable: true, get: function () { return failure_class_ts_1.EvalFatalError; } });
30
20
  Object.defineProperty(exports, "failureClassOf", { enumerable: true, get: function () { return failure_class_ts_1.failureClassOf; } });
31
- var judge_presets_ts_1 = require("./assertions/judge-presets.cjs");
32
- Object.defineProperty(exports, "factuality", { enumerable: true, get: function () { return judge_presets_ts_1.factuality; } });
33
- Object.defineProperty(exports, "faithfulness", { enumerable: true, get: function () { return judge_presets_ts_1.faithfulness; } });
34
- Object.defineProperty(exports, "instructionFollowing", { enumerable: true, get: function () { return judge_presets_ts_1.instructionFollowing; } });
35
- Object.defineProperty(exports, "pairwisePreference", { enumerable: true, get: function () { return judge_presets_ts_1.pairwisePreference; } });
36
21
  var index_ts_1 = require("./error-assistance/index.cjs");
37
22
  Object.defineProperty(exports, "MigrationRequiredError", { enumerable: true, get: function () { return index_ts_1.MigrationRequiredError; } });
38
- var context_match_ts_1 = require("./assertions/context-match.cjs");
39
- Object.defineProperty(exports, "defineMaterialMatch", { enumerable: true, get: function () { return context_match_ts_1.defineMaterialMatch; } });
40
- Object.defineProperty(exports, "defineContextMatch", { enumerable: true, get: function () { return context_match_ts_1.defineContextMatch; } });
23
+ var model_roles_ts_1 = require("./model-roles.cjs");
24
+ Object.defineProperty(exports, "defineModelRoles", { enumerable: true, get: function () { return model_roles_ts_1.defineModelRoles; } });
41
25
  //# sourceMappingURL=index.cjs.map
@@ -1 +1 @@
1
- {"version":3,"file":"index.js","sourceRoot":"","sources":["index.ts"],"names":[],"mappings":";;;AAAA,+CAA+C;AAC/C,uEAAuE;AACvE,yCAA2G;AAAlG,uGAAA,UAAU,OAAA;AAAE,4GAAA,eAAe,OAAA;AAAE,4GAAA,eAAe,OAAA;AAAE,yGAAA,YAAY,OAAA;AAAE,6GAAA,gBAAgB,OAAA;AACrF,2CAAoE;AAA3D,2GAAA,aAAa,OAAA;AAAE,mHAAA,qBAAqB,OAAA;AAC7C,kDAAoD;AAA3C,uGAAA,WAAW,OAAA;AAEpB,kDAAyD;AAAhD,4GAAA,gBAAgB,OAAA;AAGzB,0EAAuN;AAA9M,2HAAA,mBAAmB,OAAA;AAG5B,qCAA8D;AAArD,qGAAA,UAAU,OAAA;AAAE,iGAAA,MAAM,OAAA;AAAE,wGAAA,aAAa,OAAA;AAC1C,6CAA6C;AAC7C,6EAA6E;AAC7E,8DAAiG;AAAxF,wHAAA,oBAAoB,OAAA;AAAE,kHAAA,cAAc,OAAA;AAAE,kHAAA,cAAc,OAAA;AAK7D,kEAAmH;AAA1G,8GAAA,UAAU,OAAA;AAAE,gHAAA,YAAY,OAAA;AAAE,wHAAA,oBAAoB,OAAA;AAAE,sHAAA,kBAAkB,OAAA;AAG3E,wDAAiH;AAAxG,kHAAA,sBAAsB,OAAA;AAI/B,kEAAwF;AAA/E,uHAAA,mBAAmB,OAAA;AAAE,sHAAA,kBAAkB,OAAA","sourcesContent":["// niceeval 公开导出(import { … } from \"niceeval\")。\n// Agent/Adapter 相关见 \"niceeval/adapter\";Sandbox 相关见 \"niceeval/sandbox\"。\nexport { defineEval, defineScoreEval, defineEvalGroup, defineConfig, defineExperiment } from \"./define.ts\";\nexport { defineAdapter, defineAdapterContract } from \"./adapter.ts\";\nexport { defineJudge } from \"./assertions/judge.ts\";\nexport type { JudgeAnchor, JudgeDefinition, JudgeOptions } from \"./assertions/judge.ts\";\nexport { defineScoreMatch } from \"./assertions/match.ts\";\nexport type { ManagedScoreMatchOptions, ScoreMatch, ScoreMatchAnchor, ScoreMatchContext, ScoreMatchLlmFailure, ScoreMatchResult } from \"./assertions/match.ts\";\nexport type { AssertionEntryId } from \"./assertions/identity.ts\";\nexport { readScoreMatchAudit, type ScoreMatchAudit, type ScoreMatchAuditV3, type ScoreMatchAuditImageContent, type ScoreMatchAuditEnvelope, type ScoreMatchAuditReadResult, } from \"./assertions/score-match-audit.ts\";\nexport type { AgentSendWindowIdentity, AgentWorkspaceDiffEndpoint, AgentWorkspaceDiffHunks, AgentWorkspaceDiffPolicy, AgentWorkspaceDiff, AgentWorkspaceDiffWindow, AgentWorkspaceDiffWindowChange, } from \"./assertions/workspace-diff.ts\";\nexport type { AssertionCheck, ScoreFunction, PolymorphicBooleanAssertionHandle, PolymorphicMeasurementAssertionHandle, PostRunBooleanAssertionHandle, } from \"./assertions/api.ts\";\nexport { requireEnv, getEnv, stripComments } from \"./util.ts\";\n// 执行失败分类:抛出点糖衣类(声明死因波及多远)+ 结构守卫。判据、分类链与止损语义见\n// docs/feature/error-classification/README.md;`niceeval/adapter` 复导出同一份词表类型。\nexport { ExperimentFatalError, EvalFatalError, failureClassOf } from \"./shared/failure-class.ts\";\n// 类型(eval 作者会用到;跑哪个 agent / 用哪个 sandbox 见对应子路径)\nexport type { StreamEvent, ToolName, JsonValue, FlagValue, ExperimentFlags, JsonMatch, Usage, Turn, EvidenceCoverageStatus, EvidenceCoverageEntry, EvidenceCoverage, TurnEvidenceCoverage, TurnInput, InputFile, AnswerValue, InputResponse, RespondAnswer, TurnHandle, SessionHandle, TestContext, ScoreTestContext, Verdict, EvalAuthorFields, EvalInput, ScoreEvalInput, EvalDefinition, AnyEvalDefinition, EvalGroupMember, EvalGroupInput, EvalGroupDefinition, EvaluationKind, EvalDescriptor, ExperimentAuthorFields, SharedStateConfig, ExperimentInput, ExperimentDefinition, ExperimentHook, ExperimentHookContext, Config, LocalizedText, Reporter, ReporterEvent, EvalResult, AcceptedDifference, AcceptedResult, RetryAttemptRecord, InvocationSummary, InvocationShape, DiagnosticRecord, ExperimentRunInfo, SandboxRunInfo, ScriptResult, CommandResult, InputRequest, InputRequestFilter, O11ySummary, TraceSpan, SpanKind, DerivedFacts, FailureClass, FailureScope, AttemptFailureInfo, AttemptFailureClassifier, ExternalCause, ExternalCauseFact, ExternalCauseLink, ExternalCauseScalar, } from \"./types.ts\";\nexport type { Adapter, AdapterAssertionsFactory, AdapterAssertionsFactoryContext, AdapterCleanupContext, AttemptCancellation, AttemptSignal, AdapterContract, AdapterCreateContext, AdapterDefinition, AdapterEvalDefinition, AdapterEvalInput, AdapterIdentity, AdapterImplementation, AdapterImplementationInput, AdapterRuntimeDefinition, AdapterScoreEvalInput, EvalContext, ReservedAdapterContextKey, } from \"./adapter.ts\";\nexport type { ParsedTranscript } from \"./o11y/parsers/index.ts\";\nexport { factuality, faithfulness, instructionFollowing, pairwisePreference } from \"./assertions/judge-presets.ts\";\nexport type { JudgePresetOptions, FactualityMaterial, FaithfulnessMaterial, InstructionFollowingMaterial, PairwisePreferenceMaterial } from \"./assertions/judge-presets.ts\";\nexport type { JudgeProvider, JudgeSelection } from \"./judge/index.ts\";\nexport { MigrationRequiredError, type MigrationOccurrence, type ErrorSource } from \"./error-assistance/index.ts\";\nexport type { AdapterUsageInput, AdapterUsageSeal } from \"./adapter-usage.ts\";\nexport type { AdapterAttachmentInput, AdapterAttachmentReceipt } from \"./adapter-attachments.ts\";\nexport type { ExecutionDisplayBlock, ExecutionTraceActor, ExecutionTraceEvent, ExecutionTraceEvidence, ExecutionTraceInput, ExecutionTraceLimitation, ExecutionTraceLink, ExecutionTraceReceipt, ExecutionTraceScope, ExecutionTraceScopeMembership, ExecutionTraceSource, ExecutionTraceTime, TraceJson, } from \"./adapter-execution-trace.ts\";\nexport { defineMaterialMatch, defineContextMatch } from \"./assertions/context-match.ts\";\nexport type { MaterialMatch, ContextBooleanMatch, ContextScoreMatch, MatchContext, ReadonlyMaterial, MatchFact, MaterialCollection, MaterialItem, MaterialCaptureBudget, FactCaptureBudget } from \"./assertions/context-match.ts\";\nexport type { AttemptUsageSnapshot, AdapterAttemptUsageSnapshot } from \"./o11y/adapter-usage-projection.ts\";\nexport type { EvalUsage, EvalCacheRead } from \"./o11y/eval-usage.ts\";\nexport type { AgentMatchContext } from \"./context/assert-first.ts\";\nexport type { ModelSlotSelection, ResolvedModelSlot, ResolvedModelSlots } from \"./model-slots.ts\";\nexport type { CacheReadReport, CacheReadSummary } from \"./o11y/cache-read.ts\";\n"]}
1
+ {"version":3,"file":"index.js","sourceRoot":"","sources":["index.ts"],"names":[],"mappings":";;;AAAA,+CAA+C;AAC/C,uEAAuE;AACvE,yCAA2G;AAAlG,uGAAA,UAAU,OAAA;AAAE,4GAAA,eAAe,OAAA;AAAE,4GAAA,eAAe,OAAA;AAAE,yGAAA,YAAY,OAAA;AAAE,6GAAA,gBAAgB,OAAA;AACrF,2CAAoE;AAA3D,2GAAA,aAAa,OAAA;AAAE,mHAAA,qBAAqB,OAAA;AAO7C,6CAA6C;AAC7C,6EAA6E;AAC7E,8DAAiG;AAAxF,wHAAA,oBAAoB,OAAA;AAAE,kHAAA,cAAc,OAAA;AAAE,kHAAA,cAAc,OAAA;AAO7D,wDAAiH;AAAxG,kHAAA,sBAAsB,OAAA;AAS/B,mDAAoD;AAA3C,kHAAA,gBAAgB,OAAA","sourcesContent":["// niceeval 公开导出(import { … } from \"niceeval\")。\n// Agent/Adapter 相关见 \"niceeval/adapter\";Sandbox 相关见 \"niceeval/sandbox\"。\nexport { defineEval, defineScoreEval, defineEvalGroup, defineConfig, defineExperiment } from \"./define.ts\";\nexport { defineAdapter, defineAdapterContract } from \"./adapter.ts\";\nexport type { JudgeAnchor, JudgeDefinition, JudgeOptions } from \"./assertions/judge.ts\";\nexport type { ManagedScoreMatchOptions, ScoreMatch, ScoreMatchAnchor, ScoreMatchContext, ScoreMatchLlmFailure, ScoreMatchResult } from \"./assertions/match.ts\";\nexport type { AssertionEntryId } from \"./assertions/identity.ts\";\nexport { type ScoreMatchAudit, type ScoreMatchAuditV3, type ScoreMatchAuditImageContent, type ScoreMatchAuditEnvelope, type ScoreMatchAuditReadResult, } from \"./assertions/score-match-audit.ts\";\nexport type { AgentSendWindowIdentity, AgentWorkspaceDiffEndpoint, AgentWorkspaceDiffHunks, AgentWorkspaceDiffPolicy, AgentWorkspaceDiff, AgentWorkspaceDiffWindow, AgentWorkspaceDiffWindowChange, } from \"./assertions/workspace-diff.ts\";\nexport type { AssertionCheck, ScoreFunction, PolymorphicBooleanAssertionHandle, PolymorphicMeasurementAssertionHandle, PostRunBooleanAssertionHandle, } from \"./assertions/api.ts\";\n// 执行失败分类:抛出点糖衣类(声明死因波及多远)+ 结构守卫。判据、分类链与止损语义见\n// docs/feature/error-classification/README.md;`niceeval/adapter` 复导出同一份词表类型。\nexport { ExperimentFatalError, EvalFatalError, failureClassOf } from \"./shared/failure-class.ts\";\n// 类型(eval 作者会用到;跑哪个 agent / 用哪个 sandbox 见对应子路径)\nexport type { StreamEvent, ToolName, JsonValue, FlagValue, ExperimentFlags, JsonMatch, Usage, Turn, EvidenceCoverageStatus, EvidenceCoverageEntry, EvidenceCoverage, TurnEvidenceCoverage, TurnInput, InputFile, AnswerValue, InputResponse, RespondAnswer, TurnHandle, SessionHandle, TestContext, ScoreTestContext, Verdict, EvalAuthorFields, EvalInput, ScoreEvalInput, EvalDefinition, AnyEvalDefinition, EvalGroupMember, EvalGroupInput, EvalGroupDefinition, EvaluationKind, EvalDescriptor, ExperimentAuthorFields, SharedStateConfig, SandboxBeforeCache, SandboxCacheConfig, ExperimentInput, ExperimentDefinition, ExperimentHook, ExperimentHookContext, Config, LocalizedText, Reporter, ReporterEvent, EvalResult, AcceptedDifference, AcceptedResult, RetryAttemptRecord, InvocationSummary, InvocationShape, DiagnosticRecord, ExperimentRunInfo, SandboxRunInfo, ScriptResult, CommandResult, InputRequest, InputRequestFilter, ObservabilitySummary, TraceSpan, SpanKind, DerivedFacts, FailureClass, FailureScope, AttemptFailureInfo, ExternalCause, ExternalCauseFact, ExternalCauseLink, ExternalCauseScalar, } from \"./types.ts\";\nexport type { Adapter, AdapterAssertionsFactory, AdapterAssertionsFactoryContext, AdapterCleanupContext, AttemptCancellation, AttemptSignal, AdapterContract, AdapterCreateContext, AdapterDefinition, AdapterEvalDefinition, AdapterEvalInput, AdapterIdentity, AdapterImplementation, AdapterImplementationInput, AdapterRuntimeDefinition, AdapterScoreEvalInput, EvalContext, ReservedAdapterContextKey, } from \"./adapter.ts\";\nexport type { ParsedTranscript } from \"./o11y/parsers/index.ts\";\nexport type { JudgePresetOptions, FactualityMaterial, FaithfulnessMaterial, InstructionFollowingMaterial, PairwisePreferenceMaterial } from \"./assertions/judge-presets.ts\";\nexport type { JudgeProvider, JudgeSelection } from \"./judge/index.ts\";\nexport { MigrationRequiredError, type MigrationOccurrence, type ErrorSource } from \"./error-assistance/index.ts\";\nexport type { AdapterUsageInput, AdapterUsageSeal } from \"./adapter-usage.ts\";\nexport type { AdapterAttachmentInput, AdapterAttachmentReceipt } from \"./adapter-attachments.ts\";\nexport type { ExecutionDisplayBlock, ExecutionTraceActor, ExecutionTraceEvent, ExecutionTraceEvidence, ExecutionTraceInput, ExecutionTraceLimitation, ExecutionTraceLink, ExecutionTraceReceipt, ExecutionTraceScope, ExecutionTraceScopeMembership, ExecutionTraceSource, ExecutionTraceTime, TraceJson, } from \"./adapter-execution-trace.ts\";\nexport type { MaterialMatch, ContextBooleanMatch, ContextScoreMatch, MatchContext, ReadonlyMaterial, MatchFact, MaterialCollection, MaterialItem, MaterialCaptureBudget, FactCaptureBudget } from \"./assertions/context-match.ts\";\nexport type { AttemptUsageSnapshot, AdapterAttemptUsageSnapshot } from \"./o11y/adapter-usage-projection.ts\";\nexport type { EvalUsage, EvalCacheRead } from \"./o11y/eval-usage.ts\";\nexport type { AgentMatchContext } from \"./context/assert-first.ts\";\nexport type { ModelRoleSelection, ResolvedModelRole, ResolvedModelRoles } from \"./model-roles.ts\";\nexport { defineModelRoles } from \"./model-roles.ts\";\nexport type { CacheReadReport, CacheReadSummary } from \"./o11y/cache-read.ts\";\n"]}
package/dist/index.d.cts CHANGED
@@ -1,29 +1,25 @@
1
1
  export { defineEval, defineScoreEval, defineEvalGroup, defineConfig, defineExperiment } from "./define.cjs";
2
2
  export { defineAdapter, defineAdapterContract } from "./adapter.cjs";
3
- export { defineJudge } from "./assertions/judge.cjs";
4
3
  export type { JudgeAnchor, JudgeDefinition, JudgeOptions } from "./assertions/judge.cjs";
5
- export { defineScoreMatch } from "./assertions/match.cjs";
6
4
  export type { ManagedScoreMatchOptions, ScoreMatch, ScoreMatchAnchor, ScoreMatchContext, ScoreMatchLlmFailure, ScoreMatchResult } from "./assertions/match.cjs";
7
5
  export type { AssertionEntryId } from "./assertions/identity.cjs";
8
- export { readScoreMatchAudit, type ScoreMatchAudit, type ScoreMatchAuditV3, type ScoreMatchAuditImageContent, type ScoreMatchAuditEnvelope, type ScoreMatchAuditReadResult, } from "./assertions/score-match-audit.cjs";
6
+ export { type ScoreMatchAudit, type ScoreMatchAuditV3, type ScoreMatchAuditImageContent, type ScoreMatchAuditEnvelope, type ScoreMatchAuditReadResult, } from "./assertions/score-match-audit.cjs";
9
7
  export type { AgentSendWindowIdentity, AgentWorkspaceDiffEndpoint, AgentWorkspaceDiffHunks, AgentWorkspaceDiffPolicy, AgentWorkspaceDiff, AgentWorkspaceDiffWindow, AgentWorkspaceDiffWindowChange, } from "./assertions/workspace-diff.cjs";
10
8
  export type { AssertionCheck, ScoreFunction, PolymorphicBooleanAssertionHandle, PolymorphicMeasurementAssertionHandle, PostRunBooleanAssertionHandle, } from "./assertions/api.cjs";
11
- export { requireEnv, getEnv, stripComments } from "./util.cjs";
12
9
  export { ExperimentFatalError, EvalFatalError, failureClassOf } from "./shared/failure-class.cjs";
13
- export type { StreamEvent, ToolName, JsonValue, FlagValue, ExperimentFlags, JsonMatch, Usage, Turn, EvidenceCoverageStatus, EvidenceCoverageEntry, EvidenceCoverage, TurnEvidenceCoverage, TurnInput, InputFile, AnswerValue, InputResponse, RespondAnswer, TurnHandle, SessionHandle, TestContext, ScoreTestContext, Verdict, EvalAuthorFields, EvalInput, ScoreEvalInput, EvalDefinition, AnyEvalDefinition, EvalGroupMember, EvalGroupInput, EvalGroupDefinition, EvaluationKind, EvalDescriptor, ExperimentAuthorFields, SharedStateConfig, ExperimentInput, ExperimentDefinition, ExperimentHook, ExperimentHookContext, Config, LocalizedText, Reporter, ReporterEvent, EvalResult, AcceptedDifference, AcceptedResult, RetryAttemptRecord, InvocationSummary, InvocationShape, DiagnosticRecord, ExperimentRunInfo, SandboxRunInfo, ScriptResult, CommandResult, InputRequest, InputRequestFilter, O11ySummary, TraceSpan, SpanKind, DerivedFacts, FailureClass, FailureScope, AttemptFailureInfo, AttemptFailureClassifier, ExternalCause, ExternalCauseFact, ExternalCauseLink, ExternalCauseScalar, } from "./types.cjs";
10
+ export type { StreamEvent, ToolName, JsonValue, FlagValue, ExperimentFlags, JsonMatch, Usage, Turn, EvidenceCoverageStatus, EvidenceCoverageEntry, EvidenceCoverage, TurnEvidenceCoverage, TurnInput, InputFile, AnswerValue, InputResponse, RespondAnswer, TurnHandle, SessionHandle, TestContext, ScoreTestContext, Verdict, EvalAuthorFields, EvalInput, ScoreEvalInput, EvalDefinition, AnyEvalDefinition, EvalGroupMember, EvalGroupInput, EvalGroupDefinition, EvaluationKind, EvalDescriptor, ExperimentAuthorFields, SharedStateConfig, SandboxBeforeCache, SandboxCacheConfig, ExperimentInput, ExperimentDefinition, ExperimentHook, ExperimentHookContext, Config, LocalizedText, Reporter, ReporterEvent, EvalResult, AcceptedDifference, AcceptedResult, RetryAttemptRecord, InvocationSummary, InvocationShape, DiagnosticRecord, ExperimentRunInfo, SandboxRunInfo, ScriptResult, CommandResult, InputRequest, InputRequestFilter, ObservabilitySummary, TraceSpan, SpanKind, DerivedFacts, FailureClass, FailureScope, AttemptFailureInfo, ExternalCause, ExternalCauseFact, ExternalCauseLink, ExternalCauseScalar, } from "./types.cjs";
14
11
  export type { Adapter, AdapterAssertionsFactory, AdapterAssertionsFactoryContext, AdapterCleanupContext, AttemptCancellation, AttemptSignal, AdapterContract, AdapterCreateContext, AdapterDefinition, AdapterEvalDefinition, AdapterEvalInput, AdapterIdentity, AdapterImplementation, AdapterImplementationInput, AdapterRuntimeDefinition, AdapterScoreEvalInput, EvalContext, ReservedAdapterContextKey, } from "./adapter.cjs";
15
12
  export type { ParsedTranscript } from "./o11y/parsers/index.cjs";
16
- export { factuality, faithfulness, instructionFollowing, pairwisePreference } from "./assertions/judge-presets.cjs";
17
13
  export type { JudgePresetOptions, FactualityMaterial, FaithfulnessMaterial, InstructionFollowingMaterial, PairwisePreferenceMaterial } from "./assertions/judge-presets.cjs";
18
14
  export type { JudgeProvider, JudgeSelection } from "./judge/index.cjs";
19
15
  export { MigrationRequiredError, type MigrationOccurrence, type ErrorSource } from "./error-assistance/index.cjs";
20
16
  export type { AdapterUsageInput, AdapterUsageSeal } from "./adapter-usage.cjs";
21
17
  export type { AdapterAttachmentInput, AdapterAttachmentReceipt } from "./adapter-attachments.cjs";
22
18
  export type { ExecutionDisplayBlock, ExecutionTraceActor, ExecutionTraceEvent, ExecutionTraceEvidence, ExecutionTraceInput, ExecutionTraceLimitation, ExecutionTraceLink, ExecutionTraceReceipt, ExecutionTraceScope, ExecutionTraceScopeMembership, ExecutionTraceSource, ExecutionTraceTime, TraceJson, } from "./adapter-execution-trace.cjs";
23
- export { defineMaterialMatch, defineContextMatch } from "./assertions/context-match.cjs";
24
19
  export type { MaterialMatch, ContextBooleanMatch, ContextScoreMatch, MatchContext, ReadonlyMaterial, MatchFact, MaterialCollection, MaterialItem, MaterialCaptureBudget, FactCaptureBudget } from "./assertions/context-match.cjs";
25
20
  export type { AttemptUsageSnapshot, AdapterAttemptUsageSnapshot } from "./o11y/adapter-usage-projection.cjs";
26
21
  export type { EvalUsage, EvalCacheRead } from "./o11y/eval-usage.cjs";
27
22
  export type { AgentMatchContext } from "./context/assert-first.cjs";
28
- export type { ModelSlotSelection, ResolvedModelSlot, ResolvedModelSlots } from "./model-slots.cjs";
23
+ export type { ModelRoleSelection, ResolvedModelRole, ResolvedModelRoles } from "./model-roles.cjs";
24
+ export { defineModelRoles } from "./model-roles.cjs";
29
25
  export type { CacheReadReport, CacheReadSummary } from "./o11y/cache-read.cjs";
package/dist/index.d.mts CHANGED
@@ -1,29 +1,25 @@
1
1
  export { defineEval, defineScoreEval, defineEvalGroup, defineConfig, defineExperiment } from "./define.mjs";
2
2
  export { defineAdapter, defineAdapterContract } from "./adapter.mjs";
3
- export { defineJudge } from "./assertions/judge.mjs";
4
3
  export type { JudgeAnchor, JudgeDefinition, JudgeOptions } from "./assertions/judge.mjs";
5
- export { defineScoreMatch } from "./assertions/match.mjs";
6
4
  export type { ManagedScoreMatchOptions, ScoreMatch, ScoreMatchAnchor, ScoreMatchContext, ScoreMatchLlmFailure, ScoreMatchResult } from "./assertions/match.mjs";
7
5
  export type { AssertionEntryId } from "./assertions/identity.mjs";
8
- export { readScoreMatchAudit, type ScoreMatchAudit, type ScoreMatchAuditV3, type ScoreMatchAuditImageContent, type ScoreMatchAuditEnvelope, type ScoreMatchAuditReadResult, } from "./assertions/score-match-audit.mjs";
6
+ export { type ScoreMatchAudit, type ScoreMatchAuditV3, type ScoreMatchAuditImageContent, type ScoreMatchAuditEnvelope, type ScoreMatchAuditReadResult, } from "./assertions/score-match-audit.mjs";
9
7
  export type { AgentSendWindowIdentity, AgentWorkspaceDiffEndpoint, AgentWorkspaceDiffHunks, AgentWorkspaceDiffPolicy, AgentWorkspaceDiff, AgentWorkspaceDiffWindow, AgentWorkspaceDiffWindowChange, } from "./assertions/workspace-diff.mjs";
10
8
  export type { AssertionCheck, ScoreFunction, PolymorphicBooleanAssertionHandle, PolymorphicMeasurementAssertionHandle, PostRunBooleanAssertionHandle, } from "./assertions/api.mjs";
11
- export { requireEnv, getEnv, stripComments } from "./util.mjs";
12
9
  export { ExperimentFatalError, EvalFatalError, failureClassOf } from "./shared/failure-class.mjs";
13
- export type { StreamEvent, ToolName, JsonValue, FlagValue, ExperimentFlags, JsonMatch, Usage, Turn, EvidenceCoverageStatus, EvidenceCoverageEntry, EvidenceCoverage, TurnEvidenceCoverage, TurnInput, InputFile, AnswerValue, InputResponse, RespondAnswer, TurnHandle, SessionHandle, TestContext, ScoreTestContext, Verdict, EvalAuthorFields, EvalInput, ScoreEvalInput, EvalDefinition, AnyEvalDefinition, EvalGroupMember, EvalGroupInput, EvalGroupDefinition, EvaluationKind, EvalDescriptor, ExperimentAuthorFields, SharedStateConfig, ExperimentInput, ExperimentDefinition, ExperimentHook, ExperimentHookContext, Config, LocalizedText, Reporter, ReporterEvent, EvalResult, AcceptedDifference, AcceptedResult, RetryAttemptRecord, InvocationSummary, InvocationShape, DiagnosticRecord, ExperimentRunInfo, SandboxRunInfo, ScriptResult, CommandResult, InputRequest, InputRequestFilter, O11ySummary, TraceSpan, SpanKind, DerivedFacts, FailureClass, FailureScope, AttemptFailureInfo, AttemptFailureClassifier, ExternalCause, ExternalCauseFact, ExternalCauseLink, ExternalCauseScalar, } from "./types.mjs";
10
+ export type { StreamEvent, ToolName, JsonValue, FlagValue, ExperimentFlags, JsonMatch, Usage, Turn, EvidenceCoverageStatus, EvidenceCoverageEntry, EvidenceCoverage, TurnEvidenceCoverage, TurnInput, InputFile, AnswerValue, InputResponse, RespondAnswer, TurnHandle, SessionHandle, TestContext, ScoreTestContext, Verdict, EvalAuthorFields, EvalInput, ScoreEvalInput, EvalDefinition, AnyEvalDefinition, EvalGroupMember, EvalGroupInput, EvalGroupDefinition, EvaluationKind, EvalDescriptor, ExperimentAuthorFields, SharedStateConfig, SandboxBeforeCache, SandboxCacheConfig, ExperimentInput, ExperimentDefinition, ExperimentHook, ExperimentHookContext, Config, LocalizedText, Reporter, ReporterEvent, EvalResult, AcceptedDifference, AcceptedResult, RetryAttemptRecord, InvocationSummary, InvocationShape, DiagnosticRecord, ExperimentRunInfo, SandboxRunInfo, ScriptResult, CommandResult, InputRequest, InputRequestFilter, ObservabilitySummary, TraceSpan, SpanKind, DerivedFacts, FailureClass, FailureScope, AttemptFailureInfo, ExternalCause, ExternalCauseFact, ExternalCauseLink, ExternalCauseScalar, } from "./types.mjs";
14
11
  export type { Adapter, AdapterAssertionsFactory, AdapterAssertionsFactoryContext, AdapterCleanupContext, AttemptCancellation, AttemptSignal, AdapterContract, AdapterCreateContext, AdapterDefinition, AdapterEvalDefinition, AdapterEvalInput, AdapterIdentity, AdapterImplementation, AdapterImplementationInput, AdapterRuntimeDefinition, AdapterScoreEvalInput, EvalContext, ReservedAdapterContextKey, } from "./adapter.mjs";
15
12
  export type { ParsedTranscript } from "./o11y/parsers/index.mjs";
16
- export { factuality, faithfulness, instructionFollowing, pairwisePreference } from "./assertions/judge-presets.mjs";
17
13
  export type { JudgePresetOptions, FactualityMaterial, FaithfulnessMaterial, InstructionFollowingMaterial, PairwisePreferenceMaterial } from "./assertions/judge-presets.mjs";
18
14
  export type { JudgeProvider, JudgeSelection } from "./judge/index.mjs";
19
15
  export { MigrationRequiredError, type MigrationOccurrence, type ErrorSource } from "./error-assistance/index.mjs";
20
16
  export type { AdapterUsageInput, AdapterUsageSeal } from "./adapter-usage.mjs";
21
17
  export type { AdapterAttachmentInput, AdapterAttachmentReceipt } from "./adapter-attachments.mjs";
22
18
  export type { ExecutionDisplayBlock, ExecutionTraceActor, ExecutionTraceEvent, ExecutionTraceEvidence, ExecutionTraceInput, ExecutionTraceLimitation, ExecutionTraceLink, ExecutionTraceReceipt, ExecutionTraceScope, ExecutionTraceScopeMembership, ExecutionTraceSource, ExecutionTraceTime, TraceJson, } from "./adapter-execution-trace.mjs";
23
- export { defineMaterialMatch, defineContextMatch } from "./assertions/context-match.mjs";
24
19
  export type { MaterialMatch, ContextBooleanMatch, ContextScoreMatch, MatchContext, ReadonlyMaterial, MatchFact, MaterialCollection, MaterialItem, MaterialCaptureBudget, FactCaptureBudget } from "./assertions/context-match.mjs";
25
20
  export type { AttemptUsageSnapshot, AdapterAttemptUsageSnapshot } from "./o11y/adapter-usage-projection.mjs";
26
21
  export type { EvalUsage, EvalCacheRead } from "./o11y/eval-usage.mjs";
27
22
  export type { AgentMatchContext } from "./context/assert-first.mjs";
28
- export type { ModelSlotSelection, ResolvedModelSlot, ResolvedModelSlots } from "./model-slots.mjs";
23
+ export type { ModelRoleSelection, ResolvedModelRole, ResolvedModelRoles } from "./model-roles.mjs";
24
+ export { defineModelRoles } from "./model-roles.mjs";
29
25
  export type { CacheReadReport, CacheReadSummary } from "./o11y/cache-read.mjs";
package/dist/index.d.ts CHANGED
@@ -1,29 +1,25 @@
1
1
  export { defineEval, defineScoreEval, defineEvalGroup, defineConfig, defineExperiment } from "./define.js";
2
2
  export { defineAdapter, defineAdapterContract } from "./adapter.js";
3
- export { defineJudge } from "./assertions/judge.js";
4
3
  export type { JudgeAnchor, JudgeDefinition, JudgeOptions } from "./assertions/judge.js";
5
- export { defineScoreMatch } from "./assertions/match.js";
6
4
  export type { ManagedScoreMatchOptions, ScoreMatch, ScoreMatchAnchor, ScoreMatchContext, ScoreMatchLlmFailure, ScoreMatchResult } from "./assertions/match.js";
7
5
  export type { AssertionEntryId } from "./assertions/identity.js";
8
- export { readScoreMatchAudit, type ScoreMatchAudit, type ScoreMatchAuditV3, type ScoreMatchAuditImageContent, type ScoreMatchAuditEnvelope, type ScoreMatchAuditReadResult, } from "./assertions/score-match-audit.js";
6
+ export { type ScoreMatchAudit, type ScoreMatchAuditV3, type ScoreMatchAuditImageContent, type ScoreMatchAuditEnvelope, type ScoreMatchAuditReadResult, } from "./assertions/score-match-audit.js";
9
7
  export type { AgentSendWindowIdentity, AgentWorkspaceDiffEndpoint, AgentWorkspaceDiffHunks, AgentWorkspaceDiffPolicy, AgentWorkspaceDiff, AgentWorkspaceDiffWindow, AgentWorkspaceDiffWindowChange, } from "./assertions/workspace-diff.js";
10
8
  export type { AssertionCheck, ScoreFunction, PolymorphicBooleanAssertionHandle, PolymorphicMeasurementAssertionHandle, PostRunBooleanAssertionHandle, } from "./assertions/api.js";
11
- export { requireEnv, getEnv, stripComments } from "./util.js";
12
9
  export { ExperimentFatalError, EvalFatalError, failureClassOf } from "./shared/failure-class.js";
13
- export type { StreamEvent, ToolName, JsonValue, FlagValue, ExperimentFlags, JsonMatch, Usage, Turn, EvidenceCoverageStatus, EvidenceCoverageEntry, EvidenceCoverage, TurnEvidenceCoverage, TurnInput, InputFile, AnswerValue, InputResponse, RespondAnswer, TurnHandle, SessionHandle, TestContext, ScoreTestContext, Verdict, EvalAuthorFields, EvalInput, ScoreEvalInput, EvalDefinition, AnyEvalDefinition, EvalGroupMember, EvalGroupInput, EvalGroupDefinition, EvaluationKind, EvalDescriptor, ExperimentAuthorFields, SharedStateConfig, ExperimentInput, ExperimentDefinition, ExperimentHook, ExperimentHookContext, Config, LocalizedText, Reporter, ReporterEvent, EvalResult, AcceptedDifference, AcceptedResult, RetryAttemptRecord, InvocationSummary, InvocationShape, DiagnosticRecord, ExperimentRunInfo, SandboxRunInfo, ScriptResult, CommandResult, InputRequest, InputRequestFilter, O11ySummary, TraceSpan, SpanKind, DerivedFacts, FailureClass, FailureScope, AttemptFailureInfo, AttemptFailureClassifier, ExternalCause, ExternalCauseFact, ExternalCauseLink, ExternalCauseScalar, } from "./types.js";
10
+ export type { StreamEvent, ToolName, JsonValue, FlagValue, ExperimentFlags, JsonMatch, Usage, Turn, EvidenceCoverageStatus, EvidenceCoverageEntry, EvidenceCoverage, TurnEvidenceCoverage, TurnInput, InputFile, AnswerValue, InputResponse, RespondAnswer, TurnHandle, SessionHandle, TestContext, ScoreTestContext, Verdict, EvalAuthorFields, EvalInput, ScoreEvalInput, EvalDefinition, AnyEvalDefinition, EvalGroupMember, EvalGroupInput, EvalGroupDefinition, EvaluationKind, EvalDescriptor, ExperimentAuthorFields, SharedStateConfig, SandboxBeforeCache, SandboxCacheConfig, ExperimentInput, ExperimentDefinition, ExperimentHook, ExperimentHookContext, Config, LocalizedText, Reporter, ReporterEvent, EvalResult, AcceptedDifference, AcceptedResult, RetryAttemptRecord, InvocationSummary, InvocationShape, DiagnosticRecord, ExperimentRunInfo, SandboxRunInfo, ScriptResult, CommandResult, InputRequest, InputRequestFilter, ObservabilitySummary, TraceSpan, SpanKind, DerivedFacts, FailureClass, FailureScope, AttemptFailureInfo, ExternalCause, ExternalCauseFact, ExternalCauseLink, ExternalCauseScalar, } from "./types.js";
14
11
  export type { Adapter, AdapterAssertionsFactory, AdapterAssertionsFactoryContext, AdapterCleanupContext, AttemptCancellation, AttemptSignal, AdapterContract, AdapterCreateContext, AdapterDefinition, AdapterEvalDefinition, AdapterEvalInput, AdapterIdentity, AdapterImplementation, AdapterImplementationInput, AdapterRuntimeDefinition, AdapterScoreEvalInput, EvalContext, ReservedAdapterContextKey, } from "./adapter.js";
15
12
  export type { ParsedTranscript } from "./o11y/parsers/index.js";
16
- export { factuality, faithfulness, instructionFollowing, pairwisePreference } from "./assertions/judge-presets.js";
17
13
  export type { JudgePresetOptions, FactualityMaterial, FaithfulnessMaterial, InstructionFollowingMaterial, PairwisePreferenceMaterial } from "./assertions/judge-presets.js";
18
14
  export type { JudgeProvider, JudgeSelection } from "./judge/index.js";
19
15
  export { MigrationRequiredError, type MigrationOccurrence, type ErrorSource } from "./error-assistance/index.js";
20
16
  export type { AdapterUsageInput, AdapterUsageSeal } from "./adapter-usage.js";
21
17
  export type { AdapterAttachmentInput, AdapterAttachmentReceipt } from "./adapter-attachments.js";
22
18
  export type { ExecutionDisplayBlock, ExecutionTraceActor, ExecutionTraceEvent, ExecutionTraceEvidence, ExecutionTraceInput, ExecutionTraceLimitation, ExecutionTraceLink, ExecutionTraceReceipt, ExecutionTraceScope, ExecutionTraceScopeMembership, ExecutionTraceSource, ExecutionTraceTime, TraceJson, } from "./adapter-execution-trace.js";
23
- export { defineMaterialMatch, defineContextMatch } from "./assertions/context-match.js";
24
19
  export type { MaterialMatch, ContextBooleanMatch, ContextScoreMatch, MatchContext, ReadonlyMaterial, MatchFact, MaterialCollection, MaterialItem, MaterialCaptureBudget, FactCaptureBudget } from "./assertions/context-match.js";
25
20
  export type { AttemptUsageSnapshot, AdapterAttemptUsageSnapshot } from "./o11y/adapter-usage-projection.js";
26
21
  export type { EvalUsage, EvalCacheRead } from "./o11y/eval-usage.js";
27
22
  export type { AgentMatchContext } from "./context/assert-first.js";
28
- export type { ModelSlotSelection, ResolvedModelSlot, ResolvedModelSlots } from "./model-slots.js";
23
+ export type { ModelRoleSelection, ResolvedModelRole, ResolvedModelRoles } from "./model-roles.js";
24
+ export { defineModelRoles } from "./model-roles.js";
29
25
  export type { CacheReadReport, CacheReadSummary } from "./o11y/cache-read.js";
package/dist/index.js CHANGED
@@ -13,38 +13,16 @@ const __niceevalExport4 = __niceevalCanonical["defineAdapterContract"];
13
13
  export { __niceevalExport4 as defineAdapterContract };
14
14
  const __niceevalExport5 = __niceevalCanonical["defineConfig"];
15
15
  export { __niceevalExport5 as defineConfig };
16
- const __niceevalExport6 = __niceevalCanonical["defineContextMatch"];
17
- export { __niceevalExport6 as defineContextMatch };
18
- const __niceevalExport7 = __niceevalCanonical["defineEval"];
19
- export { __niceevalExport7 as defineEval };
20
- const __niceevalExport8 = __niceevalCanonical["defineEvalGroup"];
21
- export { __niceevalExport8 as defineEvalGroup };
22
- const __niceevalExport9 = __niceevalCanonical["defineExperiment"];
23
- export { __niceevalExport9 as defineExperiment };
24
- const __niceevalExport10 = __niceevalCanonical["defineJudge"];
25
- export { __niceevalExport10 as defineJudge };
26
- const __niceevalExport11 = __niceevalCanonical["defineMaterialMatch"];
27
- export { __niceevalExport11 as defineMaterialMatch };
28
- const __niceevalExport12 = __niceevalCanonical["defineScoreEval"];
29
- export { __niceevalExport12 as defineScoreEval };
30
- const __niceevalExport13 = __niceevalCanonical["defineScoreMatch"];
31
- export { __niceevalExport13 as defineScoreMatch };
32
- const __niceevalExport14 = __niceevalCanonical["factuality"];
33
- export { __niceevalExport14 as factuality };
34
- const __niceevalExport15 = __niceevalCanonical["failureClassOf"];
35
- export { __niceevalExport15 as failureClassOf };
36
- const __niceevalExport16 = __niceevalCanonical["faithfulness"];
37
- export { __niceevalExport16 as faithfulness };
38
- const __niceevalExport17 = __niceevalCanonical["getEnv"];
39
- export { __niceevalExport17 as getEnv };
40
- const __niceevalExport18 = __niceevalCanonical["instructionFollowing"];
41
- export { __niceevalExport18 as instructionFollowing };
42
- const __niceevalExport19 = __niceevalCanonical["pairwisePreference"];
43
- export { __niceevalExport19 as pairwisePreference };
44
- const __niceevalExport20 = __niceevalCanonical["readScoreMatchAudit"];
45
- export { __niceevalExport20 as readScoreMatchAudit };
46
- const __niceevalExport21 = __niceevalCanonical["requireEnv"];
47
- export { __niceevalExport21 as requireEnv };
48
- const __niceevalExport22 = __niceevalCanonical["stripComments"];
49
- export { __niceevalExport22 as stripComments };
16
+ const __niceevalExport6 = __niceevalCanonical["defineEval"];
17
+ export { __niceevalExport6 as defineEval };
18
+ const __niceevalExport7 = __niceevalCanonical["defineEvalGroup"];
19
+ export { __niceevalExport7 as defineEvalGroup };
20
+ const __niceevalExport8 = __niceevalCanonical["defineExperiment"];
21
+ export { __niceevalExport8 as defineExperiment };
22
+ const __niceevalExport9 = __niceevalCanonical["defineModelRoles"];
23
+ export { __niceevalExport9 as defineModelRoles };
24
+ const __niceevalExport10 = __niceevalCanonical["defineScoreEval"];
25
+ export { __niceevalExport10 as defineScoreEval };
26
+ const __niceevalExport11 = __niceevalCanonical["failureClassOf"];
27
+ export { __niceevalExport11 as failureClassOf };
50
28
  //# sourceMappingURL=index.js.map
package/dist/index.mjs CHANGED
@@ -13,38 +13,16 @@ const __niceevalExport4 = __niceevalCanonical["defineAdapterContract"];
13
13
  export { __niceevalExport4 as defineAdapterContract };
14
14
  const __niceevalExport5 = __niceevalCanonical["defineConfig"];
15
15
  export { __niceevalExport5 as defineConfig };
16
- const __niceevalExport6 = __niceevalCanonical["defineContextMatch"];
17
- export { __niceevalExport6 as defineContextMatch };
18
- const __niceevalExport7 = __niceevalCanonical["defineEval"];
19
- export { __niceevalExport7 as defineEval };
20
- const __niceevalExport8 = __niceevalCanonical["defineEvalGroup"];
21
- export { __niceevalExport8 as defineEvalGroup };
22
- const __niceevalExport9 = __niceevalCanonical["defineExperiment"];
23
- export { __niceevalExport9 as defineExperiment };
24
- const __niceevalExport10 = __niceevalCanonical["defineJudge"];
25
- export { __niceevalExport10 as defineJudge };
26
- const __niceevalExport11 = __niceevalCanonical["defineMaterialMatch"];
27
- export { __niceevalExport11 as defineMaterialMatch };
28
- const __niceevalExport12 = __niceevalCanonical["defineScoreEval"];
29
- export { __niceevalExport12 as defineScoreEval };
30
- const __niceevalExport13 = __niceevalCanonical["defineScoreMatch"];
31
- export { __niceevalExport13 as defineScoreMatch };
32
- const __niceevalExport14 = __niceevalCanonical["factuality"];
33
- export { __niceevalExport14 as factuality };
34
- const __niceevalExport15 = __niceevalCanonical["failureClassOf"];
35
- export { __niceevalExport15 as failureClassOf };
36
- const __niceevalExport16 = __niceevalCanonical["faithfulness"];
37
- export { __niceevalExport16 as faithfulness };
38
- const __niceevalExport17 = __niceevalCanonical["getEnv"];
39
- export { __niceevalExport17 as getEnv };
40
- const __niceevalExport18 = __niceevalCanonical["instructionFollowing"];
41
- export { __niceevalExport18 as instructionFollowing };
42
- const __niceevalExport19 = __niceevalCanonical["pairwisePreference"];
43
- export { __niceevalExport19 as pairwisePreference };
44
- const __niceevalExport20 = __niceevalCanonical["readScoreMatchAudit"];
45
- export { __niceevalExport20 as readScoreMatchAudit };
46
- const __niceevalExport21 = __niceevalCanonical["requireEnv"];
47
- export { __niceevalExport21 as requireEnv };
48
- const __niceevalExport22 = __niceevalCanonical["stripComments"];
49
- export { __niceevalExport22 as stripComments };
16
+ const __niceevalExport6 = __niceevalCanonical["defineEval"];
17
+ export { __niceevalExport6 as defineEval };
18
+ const __niceevalExport7 = __niceevalCanonical["defineEvalGroup"];
19
+ export { __niceevalExport7 as defineEvalGroup };
20
+ const __niceevalExport8 = __niceevalCanonical["defineExperiment"];
21
+ export { __niceevalExport8 as defineExperiment };
22
+ const __niceevalExport9 = __niceevalCanonical["defineModelRoles"];
23
+ export { __niceevalExport9 as defineModelRoles };
24
+ const __niceevalExport10 = __niceevalCanonical["defineScoreEval"];
25
+ export { __niceevalExport10 as defineScoreEval };
26
+ const __niceevalExport11 = __niceevalCanonical["failureClassOf"];
27
+ export { __niceevalExport11 as failureClassOf };
50
28
  //# sourceMappingURL=index.mjs.map
@@ -3,7 +3,7 @@ Object.defineProperty(exports, "__esModule", { value: true });
3
3
  exports.projectAttemptAssertionIndex = projectAttemptAssertionIndex;
4
4
  exports.projectAttemptAssertionDetail = projectAttemptAssertionDetail;
5
5
  exports.projectAttemptAssertionImage = projectAttemptAssertionImage;
6
- // @concord-file ne-eval-inspection-assertion-projection
6
+ // @concord-file
7
7
  // @concord-implements docs/feature/inspection/README.md
8
8
  // @concord-implements docs/feature/assertions/README.md
9
9
  const effect_1 = require("effect");