niceeval 0.15.1-canary.76 → 0.16.1

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (714) hide show
  1. package/INDEX.md +45 -40
  2. package/dist/adapter-execution-trace.cjs.map +1 -1
  3. package/dist/adapter-execution-trace.d.cts +26 -0
  4. package/dist/adapter-execution-trace.d.mts +26 -0
  5. package/dist/adapter-execution-trace.d.ts +26 -0
  6. package/dist/adapter-usage.cjs.map +1 -1
  7. package/dist/adapter-usage.d.cts +9 -0
  8. package/dist/adapter-usage.d.mts +9 -0
  9. package/dist/adapter-usage.d.ts +9 -0
  10. package/dist/adapter.cjs +19 -4
  11. package/dist/adapter.cjs.map +1 -1
  12. package/dist/adapter.d.cts +37 -12
  13. package/dist/adapter.d.mts +37 -12
  14. package/dist/adapter.d.ts +37 -12
  15. package/dist/agents/ai-sdk.cjs +3 -0
  16. package/dist/agents/ai-sdk.cjs.map +1 -1
  17. package/dist/agents/bub.cjs +3 -0
  18. package/dist/agents/bub.cjs.map +1 -1
  19. package/dist/agents/claude-code-native.cjs +28 -2
  20. package/dist/agents/claude-code-native.cjs.map +1 -1
  21. package/dist/agents/claude-code.cjs +3 -0
  22. package/dist/agents/claude-code.cjs.map +1 -1
  23. package/dist/agents/codex.cjs +3 -0
  24. package/dist/agents/codex.cjs.map +1 -1
  25. package/dist/agents/coding-cli-versions.cjs +9 -10
  26. package/dist/agents/coding-cli-versions.cjs.map +1 -1
  27. package/dist/agents/coding-cli-versions.d.cts +1 -1
  28. package/dist/agents/coding-cli-versions.d.mts +1 -1
  29. package/dist/agents/coding-cli-versions.d.ts +1 -1
  30. package/dist/agents/deepseek-harness.cjs +3 -0
  31. package/dist/agents/deepseek-harness.cjs.map +1 -1
  32. package/dist/agents/hermes.cjs +3 -0
  33. package/dist/agents/hermes.cjs.map +1 -1
  34. package/dist/agents/index.cjs.map +1 -1
  35. package/dist/agents/index.d.cts +3 -2
  36. package/dist/agents/index.d.mts +3 -2
  37. package/dist/agents/index.d.ts +3 -2
  38. package/dist/agents/omp.cjs +3 -0
  39. package/dist/agents/omp.cjs.map +1 -1
  40. package/dist/agents/openai-compat.cjs +6 -0
  41. package/dist/agents/openai-compat.cjs.map +1 -1
  42. package/dist/agents/openclaw.cjs +3 -0
  43. package/dist/agents/openclaw.cjs.map +1 -1
  44. package/dist/agents/opencode.cjs +3 -0
  45. package/dist/agents/opencode.cjs.map +1 -1
  46. package/dist/agents/sdk-streams.cjs +9 -0
  47. package/dist/agents/sdk-streams.cjs.map +1 -1
  48. package/dist/agents/ui-message-stream.cjs +3 -0
  49. package/dist/agents/ui-message-stream.cjs.map +1 -1
  50. package/dist/assertions/api.cjs.map +1 -1
  51. package/dist/assertions/api.d.cts +27 -5
  52. package/dist/assertions/api.d.mts +27 -5
  53. package/dist/assertions/api.d.ts +27 -5
  54. package/dist/assertions/collection.cjs +6 -0
  55. package/dist/assertions/collection.cjs.map +1 -1
  56. package/dist/assertions/collection.d.cts +4 -2
  57. package/dist/assertions/collection.d.mts +4 -2
  58. package/dist/assertions/collection.d.ts +4 -2
  59. package/dist/assertions/collection.js +12 -8
  60. package/dist/assertions/context-match.cjs +70 -0
  61. package/dist/assertions/context-match.cjs.map +1 -0
  62. package/dist/assertions/context-match.d.cts +107 -0
  63. package/dist/assertions/context-match.d.mts +107 -0
  64. package/dist/assertions/context-match.d.ts +107 -0
  65. package/dist/assertions/context-match.js +12 -0
  66. package/dist/assertions/context-match.js.map +1 -0
  67. package/dist/assertions/full-content.cjs +109 -0
  68. package/dist/assertions/full-content.cjs.map +1 -0
  69. package/dist/assertions/full-content.d.cts +6 -0
  70. package/dist/assertions/full-content.d.mts +6 -0
  71. package/dist/assertions/full-content.d.ts +6 -0
  72. package/dist/assertions/full-content.js +12 -0
  73. package/dist/assertions/full-content.js.map +1 -0
  74. package/dist/assertions/judge-presets.cjs +2 -34
  75. package/dist/assertions/judge-presets.cjs.map +1 -1
  76. package/dist/assertions/judge-presets.d.cts +6 -9
  77. package/dist/assertions/judge-presets.d.mts +6 -9
  78. package/dist/assertions/judge-presets.d.ts +6 -9
  79. package/dist/assertions/judge-presets.js +8 -8
  80. package/dist/assertions/judge.cjs +107 -29
  81. package/dist/assertions/judge.cjs.map +1 -1
  82. package/dist/assertions/judge.d.cts +2 -0
  83. package/dist/assertions/judge.d.mts +2 -0
  84. package/dist/assertions/judge.d.ts +2 -0
  85. package/dist/assertions/match.cjs +244 -5
  86. package/dist/assertions/match.cjs.map +1 -1
  87. package/dist/assertions/match.d.cts +43 -3
  88. package/dist/assertions/match.d.mts +43 -3
  89. package/dist/assertions/match.d.ts +43 -3
  90. package/dist/assertions/match.js +88 -78
  91. package/dist/assertions/material.cjs +520 -0
  92. package/dist/assertions/material.cjs.map +1 -0
  93. package/dist/assertions/material.d.cts +56 -0
  94. package/dist/assertions/material.d.mts +56 -0
  95. package/dist/assertions/material.d.ts +56 -0
  96. package/dist/assertions/material.js +20 -0
  97. package/dist/assertions/material.js.map +1 -0
  98. package/dist/assertions/record/attachment.cjs +5 -2
  99. package/dist/assertions/record/attachment.cjs.map +1 -1
  100. package/dist/assertions/record/codec.cjs +2 -1
  101. package/dist/assertions/record/codec.cjs.map +1 -1
  102. package/dist/assertions/record/model.cjs.map +1 -1
  103. package/dist/assertions/record/model.d.cts +6 -1
  104. package/dist/assertions/record/model.d.mts +6 -1
  105. package/dist/assertions/record/model.d.ts +6 -1
  106. package/dist/assertions/runtime.cjs +207 -14
  107. package/dist/assertions/runtime.cjs.map +1 -1
  108. package/dist/assertions/runtime.d.cts +9 -0
  109. package/dist/assertions/runtime.d.mts +9 -0
  110. package/dist/assertions/runtime.d.ts +9 -0
  111. package/dist/assertions/runtime.js +14 -8
  112. package/dist/assertions/score-match-audit.cjs +18 -8
  113. package/dist/assertions/score-match-audit.cjs.map +1 -1
  114. package/dist/assertions/score-match-gateway.cjs +83 -25
  115. package/dist/assertions/score-match-gateway.cjs.map +1 -1
  116. package/dist/assertions/score-match-gateway.d.cts +2 -0
  117. package/dist/assertions/score-match-gateway.d.mts +2 -0
  118. package/dist/assertions/score-match-gateway.d.ts +2 -0
  119. package/dist/assertions/value-composition.test-d.cjs +77 -0
  120. package/dist/assertions/value-composition.test-d.cjs.map +1 -0
  121. package/dist/assertions/value-composition.test-d.d.cts +1 -0
  122. package/dist/assertions/value-composition.test-d.d.mts +1 -0
  123. package/dist/assertions/value-composition.test-d.d.ts +1 -0
  124. package/dist/assertions/value-composition.test-d.js +4 -0
  125. package/dist/assertions/value-composition.test-d.js.map +1 -0
  126. package/dist/context/assert-first.cjs +207 -213
  127. package/dist/context/assert-first.cjs.map +1 -1
  128. package/dist/context/assert-first.d.cts +49 -23
  129. package/dist/context/assert-first.d.mts +49 -23
  130. package/dist/context/assert-first.d.ts +49 -23
  131. package/dist/context/session.cjs +111 -9
  132. package/dist/context/session.cjs.map +1 -1
  133. package/dist/context/session.d.cts +18 -0
  134. package/dist/context/session.d.mts +18 -0
  135. package/dist/context/session.d.ts +18 -0
  136. package/dist/define.cjs +17 -1
  137. package/dist/define.cjs.map +1 -1
  138. package/dist/eval/cli/contribution.cjs +2 -0
  139. package/dist/eval/cli/contribution.cjs.map +1 -1
  140. package/dist/eval/host/index.cjs +2 -0
  141. package/dist/eval/host/index.cjs.map +1 -1
  142. package/dist/eval/record/evaluation.cjs +3 -0
  143. package/dist/eval/record/evaluation.cjs.map +1 -1
  144. package/dist/eval/record/membership-provenance.d.cts +15 -15
  145. package/dist/eval/record/membership-provenance.d.mts +15 -15
  146. package/dist/eval/record/membership-provenance.d.ts +15 -15
  147. package/dist/eval/record/score.cjs +7 -0
  148. package/dist/eval/record/score.cjs.map +1 -1
  149. package/dist/eval/record/sealed-assertion.cjs +3 -0
  150. package/dist/eval/record/sealed-assertion.cjs.map +1 -1
  151. package/dist/eval/record/verdict.cjs +7 -0
  152. package/dist/eval/record/verdict.cjs.map +1 -1
  153. package/dist/expect/index.cjs +8 -2
  154. package/dist/expect/index.cjs.map +1 -1
  155. package/dist/expect/index.d.cts +9 -4
  156. package/dist/expect/index.d.mts +9 -4
  157. package/dist/expect/index.d.ts +9 -4
  158. package/dist/expect/index.js +72 -62
  159. package/dist/expect/index.mjs +72 -62
  160. package/dist/experiment/host/cli/contribution.cjs +37 -23
  161. package/dist/experiment/host/cli/contribution.cjs.map +1 -1
  162. package/dist/experiment/host/cli/contribution.d.cts +9 -1
  163. package/dist/experiment/host/cli/contribution.d.mts +9 -1
  164. package/dist/experiment/host/cli/contribution.d.ts +9 -1
  165. package/dist/experiment/host/current.cjs +74 -0
  166. package/dist/experiment/host/current.cjs.map +1 -0
  167. package/dist/experiment/host/current.d.cts +33 -0
  168. package/dist/experiment/host/current.d.mts +33 -0
  169. package/dist/experiment/host/current.d.ts +33 -0
  170. package/dist/experiment/host/current.js +10 -0
  171. package/dist/experiment/host/current.js.map +1 -0
  172. package/dist/experiment/host/index.cjs +17 -174
  173. package/dist/experiment/host/index.cjs.map +1 -1
  174. package/dist/experiment/host/index.d.cts +8 -54
  175. package/dist/experiment/host/index.d.mts +8 -54
  176. package/dist/experiment/host/index.d.ts +8 -54
  177. package/dist/experiment/host/operations.cjs +33 -10
  178. package/dist/experiment/host/operations.cjs.map +1 -1
  179. package/dist/experiment/host/operations.d.cts +29 -1
  180. package/dist/experiment/host/operations.d.mts +29 -1
  181. package/dist/experiment/host/operations.d.ts +29 -1
  182. package/dist/experiment/host/operations.js +6 -4
  183. package/dist/experiment/host/requirements.cjs +3 -0
  184. package/dist/experiment/host/requirements.cjs.map +1 -0
  185. package/dist/experiment/host/requirements.d.cts +3 -0
  186. package/dist/experiment/host/requirements.d.mts +3 -0
  187. package/dist/experiment/host/requirements.d.ts +3 -0
  188. package/dist/experiment/host/requirements.js +4 -0
  189. package/dist/experiment/host/requirements.js.map +1 -0
  190. package/dist/experiment/host/runtime.cjs +179 -0
  191. package/dist/experiment/host/runtime.cjs.map +1 -0
  192. package/dist/experiment/host/runtime.d.cts +79 -0
  193. package/dist/experiment/host/runtime.d.mts +79 -0
  194. package/dist/experiment/host/runtime.d.ts +79 -0
  195. package/dist/experiment/host/runtime.js +6 -0
  196. package/dist/experiment/host/runtime.js.map +1 -0
  197. package/dist/experiment/host/types.cjs.map +1 -1
  198. package/dist/experiment/host/types.d.cts +17 -16
  199. package/dist/experiment/host/types.d.mts +17 -16
  200. package/dist/experiment/host/types.d.ts +17 -16
  201. package/dist/fresh-import.cjs +2 -1
  202. package/dist/fresh-import.cjs.map +1 -1
  203. package/dist/index.cjs +4 -2
  204. package/dist/index.cjs.map +1 -1
  205. package/dist/index.d.cts +12 -6
  206. package/dist/index.d.mts +12 -6
  207. package/dist/index.d.ts +12 -6
  208. package/dist/index.js +32 -30
  209. package/dist/index.mjs +32 -30
  210. package/dist/inspection/artifacts.cjs +83 -54
  211. package/dist/inspection/artifacts.cjs.map +1 -1
  212. package/dist/inspection/artifacts.d.cts +8 -0
  213. package/dist/inspection/artifacts.d.mts +8 -0
  214. package/dist/inspection/artifacts.d.ts +8 -0
  215. package/dist/inspection/artifacts.js +2 -0
  216. package/dist/inspection/assertions.cjs +4 -1
  217. package/dist/inspection/assertions.cjs.map +1 -1
  218. package/dist/inspection/cli/contribution.cjs +27 -4
  219. package/dist/inspection/cli/contribution.cjs.map +1 -1
  220. package/dist/inspection/cli/contribution.d.cts +2 -1
  221. package/dist/inspection/cli/contribution.d.mts +2 -1
  222. package/dist/inspection/cli/contribution.d.ts +2 -1
  223. package/dist/inspection/execution-traces.cjs +114 -8
  224. package/dist/inspection/execution-traces.cjs.map +1 -1
  225. package/dist/inspection/execution-traces.d.cts +38 -0
  226. package/dist/inspection/execution-traces.d.mts +38 -0
  227. package/dist/inspection/execution-traces.d.ts +38 -0
  228. package/dist/inspection/facts.cjs +12 -0
  229. package/dist/inspection/facts.cjs.map +1 -1
  230. package/dist/inspection/limits.cjs +5 -1
  231. package/dist/inspection/limits.cjs.map +1 -1
  232. package/dist/inspection/limits.d.cts +4 -0
  233. package/dist/inspection/limits.d.mts +4 -0
  234. package/dist/inspection/limits.d.ts +4 -0
  235. package/dist/inspection/limits.js +6 -2
  236. package/dist/inspection/overview.cjs +74 -33
  237. package/dist/inspection/overview.cjs.map +1 -1
  238. package/dist/inspection/overview.d.cts +19 -1
  239. package/dist/inspection/overview.d.mts +19 -1
  240. package/dist/inspection/overview.d.ts +19 -1
  241. package/dist/inspection/overview.js +4 -2
  242. package/dist/inspection/project-input.cjs +9 -0
  243. package/dist/inspection/project-input.cjs.map +1 -0
  244. package/dist/inspection/project-input.d.cts +20 -0
  245. package/dist/inspection/project-input.d.mts +20 -0
  246. package/dist/inspection/project-input.d.ts +20 -0
  247. package/dist/inspection/project-input.js +8 -0
  248. package/dist/inspection/project-input.js.map +1 -0
  249. package/dist/inspection/project-result.cjs +60 -0
  250. package/dist/inspection/project-result.cjs.map +1 -0
  251. package/dist/inspection/project-result.d.cts +1264 -0
  252. package/dist/inspection/project-result.d.mts +1264 -0
  253. package/dist/inspection/project-result.d.ts +1264 -0
  254. package/dist/inspection/project-result.js +10 -0
  255. package/dist/inspection/project-result.js.map +1 -0
  256. package/dist/inspection/project-source.cjs +109 -0
  257. package/dist/inspection/project-source.cjs.map +1 -0
  258. package/dist/inspection/project-source.d.cts +10 -0
  259. package/dist/inspection/project-source.d.mts +10 -0
  260. package/dist/inspection/project-source.d.ts +10 -0
  261. package/dist/inspection/project-source.js +6 -0
  262. package/dist/inspection/project-source.js.map +1 -0
  263. package/dist/inspection/project.cjs +100 -0
  264. package/dist/inspection/project.cjs.map +1 -0
  265. package/dist/inspection/project.d.cts +3 -0
  266. package/dist/inspection/project.d.mts +3 -0
  267. package/dist/inspection/project.d.ts +3 -0
  268. package/dist/inspection/project.js +6 -0
  269. package/dist/inspection/project.js.map +1 -0
  270. package/dist/inspection/protocol.cjs +19 -9
  271. package/dist/inspection/protocol.cjs.map +1 -1
  272. package/dist/inspection/protocol.d.cts +5506 -38189
  273. package/dist/inspection/protocol.d.mts +5506 -38189
  274. package/dist/inspection/protocol.d.ts +5506 -38189
  275. package/dist/inspection/public.mjs +896 -84
  276. package/dist/inspection/query.cjs +2 -0
  277. package/dist/inspection/query.cjs.map +1 -1
  278. package/dist/inspection/results.cjs +234 -29
  279. package/dist/inspection/results.cjs.map +1 -1
  280. package/dist/inspection/results.d.cts +870 -13
  281. package/dist/inspection/results.d.mts +870 -13
  282. package/dist/inspection/results.d.ts +870 -13
  283. package/dist/inspection/results.js +36 -34
  284. package/dist/inspection/select.cjs +126 -14
  285. package/dist/inspection/select.cjs.map +1 -1
  286. package/dist/inspection/select.d.cts +1 -1
  287. package/dist/inspection/select.d.mts +1 -1
  288. package/dist/inspection/select.d.ts +1 -1
  289. package/dist/inspection/source.cjs +4 -0
  290. package/dist/inspection/source.cjs.map +1 -1
  291. package/dist/inspection/sources.cjs +5 -0
  292. package/dist/inspection/sources.cjs.map +1 -1
  293. package/dist/inspection/trace.cjs +82 -84
  294. package/dist/inspection/trace.cjs.map +1 -1
  295. package/dist/inspection/trace.d.cts +7 -2
  296. package/dist/inspection/trace.d.mts +7 -2
  297. package/dist/inspection/trace.d.ts +7 -2
  298. package/dist/model-slots.cjs +35 -0
  299. package/dist/model-slots.cjs.map +1 -0
  300. package/dist/model-slots.d.cts +19 -0
  301. package/dist/model-slots.d.mts +19 -0
  302. package/dist/model-slots.d.ts +19 -0
  303. package/dist/model-slots.js +8 -0
  304. package/dist/model-slots.js.map +1 -0
  305. package/dist/o11y/adapter-call-price.cjs +4 -0
  306. package/dist/o11y/adapter-call-price.cjs.map +1 -1
  307. package/dist/o11y/adapter-call-price.d.cts +24 -0
  308. package/dist/o11y/adapter-call-price.d.mts +24 -0
  309. package/dist/o11y/adapter-call-price.d.ts +24 -0
  310. package/dist/o11y/adapter-call-price.js +6 -2
  311. package/dist/o11y/adapter-usage-projection.cjs +158 -0
  312. package/dist/o11y/adapter-usage-projection.cjs.map +1 -0
  313. package/dist/o11y/adapter-usage-projection.d.cts +141 -0
  314. package/dist/o11y/adapter-usage-projection.d.mts +141 -0
  315. package/dist/o11y/adapter-usage-projection.d.ts +141 -0
  316. package/dist/o11y/adapter-usage-projection.js +20 -0
  317. package/dist/o11y/adapter-usage-projection.js.map +1 -0
  318. package/dist/o11y/eval-usage.cjs +149 -0
  319. package/dist/o11y/eval-usage.cjs.map +1 -0
  320. package/dist/o11y/eval-usage.d.cts +46 -0
  321. package/dist/o11y/eval-usage.d.mts +46 -0
  322. package/dist/o11y/eval-usage.d.ts +46 -0
  323. package/dist/o11y/eval-usage.js +8 -0
  324. package/dist/o11y/eval-usage.js.map +1 -0
  325. package/dist/o11y/judge-call-price.cjs +62 -0
  326. package/dist/o11y/judge-call-price.cjs.map +1 -0
  327. package/dist/o11y/judge-call-price.d.cts +3 -0
  328. package/dist/o11y/judge-call-price.d.mts +3 -0
  329. package/dist/o11y/judge-call-price.d.ts +3 -0
  330. package/dist/o11y/judge-call-price.js +6 -0
  331. package/dist/o11y/judge-call-price.js.map +1 -0
  332. package/dist/o11y/judge-usage-projection.cjs +134 -0
  333. package/dist/o11y/judge-usage-projection.cjs.map +1 -0
  334. package/dist/o11y/judge-usage-projection.d.cts +18 -0
  335. package/dist/o11y/judge-usage-projection.d.mts +18 -0
  336. package/dist/o11y/judge-usage-projection.d.ts +18 -0
  337. package/dist/o11y/judge-usage-projection.js +20 -0
  338. package/dist/o11y/judge-usage-projection.js.map +1 -0
  339. package/dist/o11y/judge-usage-receipt.cjs +135 -0
  340. package/dist/o11y/judge-usage-receipt.cjs.map +1 -0
  341. package/dist/o11y/judge-usage-receipt.d.cts +5 -0
  342. package/dist/o11y/judge-usage-receipt.d.mts +5 -0
  343. package/dist/o11y/judge-usage-receipt.d.ts +5 -0
  344. package/dist/o11y/judge-usage-receipt.js +6 -0
  345. package/dist/o11y/judge-usage-receipt.js.map +1 -0
  346. package/dist/o11y/judge-usage.cjs +171 -0
  347. package/dist/o11y/judge-usage.cjs.map +1 -0
  348. package/dist/o11y/judge-usage.d.cts +35 -0
  349. package/dist/o11y/judge-usage.d.mts +35 -0
  350. package/dist/o11y/judge-usage.d.ts +35 -0
  351. package/dist/o11y/judge-usage.js +6 -0
  352. package/dist/o11y/judge-usage.js.map +1 -0
  353. package/dist/plugin/contracts.cjs +2 -0
  354. package/dist/plugin/contracts.cjs.map +1 -1
  355. package/dist/plugin/link.cjs +7 -0
  356. package/dist/plugin/link.cjs.map +1 -1
  357. package/dist/record/attachment/protocol.cjs +8 -6
  358. package/dist/record/attachment/protocol.cjs.map +1 -1
  359. package/dist/record/attachment/protocol.d.cts +10 -0
  360. package/dist/record/attachment/protocol.d.mts +10 -0
  361. package/dist/record/attachment/protocol.d.ts +10 -0
  362. package/dist/record/attachment/protocol.js +38 -36
  363. package/dist/record/codec/seal-manifest.d.cts +3 -3
  364. package/dist/record/codec/seal-manifest.d.mts +3 -3
  365. package/dist/record/codec/seal-manifest.d.ts +3 -3
  366. package/dist/record/family/adapter-usage/definition.d.cts +1 -0
  367. package/dist/record/family/adapter-usage/definition.d.mts +1 -0
  368. package/dist/record/family/adapter-usage/definition.d.ts +1 -0
  369. package/dist/record/family/adapter-usage/persistence.cjs +1 -1
  370. package/dist/record/family/adapter-usage/persistence.cjs.map +1 -1
  371. package/dist/record/family/adapter-usage/persistence.d.cts +2 -1
  372. package/dist/record/family/adapter-usage/persistence.d.mts +2 -1
  373. package/dist/record/family/adapter-usage/persistence.d.ts +2 -1
  374. package/dist/record/family/adapter-usage/schema.cjs +21 -2
  375. package/dist/record/family/adapter-usage/schema.cjs.map +1 -1
  376. package/dist/record/family/adapter-usage/schema.d.cts +85 -0
  377. package/dist/record/family/adapter-usage/schema.d.mts +85 -0
  378. package/dist/record/family/adapter-usage/schema.d.ts +85 -0
  379. package/dist/record/family/adapter-usage/schema.js +18 -12
  380. package/dist/record/family/catalog.cjs +31 -27
  381. package/dist/record/family/catalog.cjs.map +1 -1
  382. package/dist/record/family/catalog.d.cts +94 -3
  383. package/dist/record/family/catalog.d.mts +94 -3
  384. package/dist/record/family/catalog.d.ts +94 -3
  385. package/dist/record/family/current.cjs +45 -42
  386. package/dist/record/family/current.cjs.map +1 -1
  387. package/dist/record/family/current.d.cts +240 -1
  388. package/dist/record/family/current.d.mts +240 -1
  389. package/dist/record/family/current.d.ts +240 -1
  390. package/dist/record/family/execution-traces/definition.d.cts +26 -0
  391. package/dist/record/family/execution-traces/definition.d.mts +26 -0
  392. package/dist/record/family/execution-traces/definition.d.ts +26 -0
  393. package/dist/record/family/execution-traces/definition.js +20 -14
  394. package/dist/record/family/execution-traces/schema.cjs +57 -1
  395. package/dist/record/family/execution-traces/schema.cjs.map +1 -1
  396. package/dist/record/family/execution-traces/schema.d.cts +114 -0
  397. package/dist/record/family/execution-traces/schema.d.mts +114 -0
  398. package/dist/record/family/execution-traces/schema.d.ts +114 -0
  399. package/dist/record/family/execution-traces/schema.js +18 -12
  400. package/dist/record/family/judge-usage/definition.cjs +12 -0
  401. package/dist/record/family/judge-usage/definition.cjs.map +1 -0
  402. package/dist/record/family/judge-usage/definition.d.cts +91 -0
  403. package/dist/record/family/judge-usage/definition.d.mts +91 -0
  404. package/dist/record/family/judge-usage/definition.d.ts +91 -0
  405. package/dist/record/family/judge-usage/definition.js +6 -0
  406. package/dist/record/family/judge-usage/definition.js.map +1 -0
  407. package/dist/record/family/judge-usage/persistence.cjs +10 -0
  408. package/dist/record/family/judge-usage/persistence.cjs.map +1 -0
  409. package/dist/record/family/judge-usage/persistence.d.cts +91 -0
  410. package/dist/record/family/judge-usage/persistence.d.mts +91 -0
  411. package/dist/record/family/judge-usage/persistence.d.ts +91 -0
  412. package/dist/record/family/judge-usage/persistence.js +6 -0
  413. package/dist/record/family/judge-usage/persistence.js.map +1 -0
  414. package/dist/record/family/judge-usage/schema.cjs +307 -0
  415. package/dist/record/family/judge-usage/schema.cjs.map +1 -0
  416. package/dist/record/family/judge-usage/schema.d.cts +351 -0
  417. package/dist/record/family/judge-usage/schema.d.mts +351 -0
  418. package/dist/record/family/judge-usage/schema.d.ts +351 -0
  419. package/dist/record/family/judge-usage/schema.js +18 -0
  420. package/dist/record/family/judge-usage/schema.js.map +1 -0
  421. package/dist/record/host/sqlite-host.cjs +55 -9
  422. package/dist/record/host/sqlite-host.cjs.map +1 -1
  423. package/dist/record/host/types.cjs.map +1 -1
  424. package/dist/record/host/types.d.cts +7 -0
  425. package/dist/record/host/types.d.mts +7 -0
  426. package/dist/record/host/types.d.ts +7 -0
  427. package/dist/record/model/run-context.cjs +7 -1
  428. package/dist/record/model/run-context.cjs.map +1 -1
  429. package/dist/record/model/run-context.d.cts +4 -0
  430. package/dist/record/model/run-context.d.mts +4 -0
  431. package/dist/record/model/run-context.d.ts +4 -0
  432. package/dist/record/model/seal-manifest.d.cts +1 -1
  433. package/dist/record/model/seal-manifest.d.mts +1 -1
  434. package/dist/record/model/seal-manifest.d.ts +1 -1
  435. package/dist/record/sqlite/client.cjs +18 -5
  436. package/dist/record/sqlite/client.cjs.map +1 -1
  437. package/dist/record/sqlite/client.d.cts +1 -1
  438. package/dist/record/sqlite/client.d.mts +1 -1
  439. package/dist/record/sqlite/client.d.ts +1 -1
  440. package/dist/record/sqlite/project-state-database.cjs +29 -8
  441. package/dist/record/sqlite/project-state-database.cjs.map +1 -1
  442. package/dist/run/host/runtime.cjs +4 -0
  443. package/dist/run/host/runtime.cjs.map +1 -1
  444. package/dist/run/storage/sqlite.cjs +20 -0
  445. package/dist/run/storage/sqlite.cjs.map +1 -1
  446. package/dist/runner/adapter-execution-trace.cjs +125 -2
  447. package/dist/runner/adapter-execution-trace.cjs.map +1 -1
  448. package/dist/runner/adapter-execution-trace.d.cts +1 -1
  449. package/dist/runner/adapter-execution-trace.d.mts +1 -1
  450. package/dist/runner/adapter-execution-trace.d.ts +1 -1
  451. package/dist/runner/adapter-usage.cjs +55 -7
  452. package/dist/runner/adapter-usage.cjs.map +1 -1
  453. package/dist/runner/adapter-usage.d.cts +9 -2
  454. package/dist/runner/adapter-usage.d.mts +9 -2
  455. package/dist/runner/adapter-usage.d.ts +9 -2
  456. package/dist/runner/adoption.cjs +9 -0
  457. package/dist/runner/adoption.cjs.map +1 -1
  458. package/dist/runner/attempt-cancellation.cjs +12 -0
  459. package/dist/runner/attempt-cancellation.cjs.map +1 -0
  460. package/dist/runner/attempt-cancellation.d.cts +6 -0
  461. package/dist/runner/attempt-cancellation.d.mts +6 -0
  462. package/dist/runner/attempt-cancellation.d.ts +6 -0
  463. package/dist/runner/attempt-cancellation.js +6 -0
  464. package/dist/runner/attempt-cancellation.js.map +1 -0
  465. package/dist/runner/attempt-resources.cjs +10 -4
  466. package/dist/runner/attempt-resources.cjs.map +1 -1
  467. package/dist/runner/attempt-resources.d.cts +2 -1
  468. package/dist/runner/attempt-resources.d.mts +2 -1
  469. package/dist/runner/attempt-resources.d.ts +2 -1
  470. package/dist/runner/attempt.cjs +1177 -1139
  471. package/dist/runner/attempt.cjs.map +1 -1
  472. package/dist/runner/config-identity.cjs +20 -1
  473. package/dist/runner/config-identity.cjs.map +1 -1
  474. package/dist/runner/config-identity.d.cts +5 -0
  475. package/dist/runner/config-identity.d.mts +5 -0
  476. package/dist/runner/config-identity.d.ts +5 -0
  477. package/dist/runner/discover.cjs +14 -0
  478. package/dist/runner/discover.cjs.map +1 -1
  479. package/dist/runner/discover.d.cts +3 -0
  480. package/dist/runner/discover.d.mts +3 -0
  481. package/dist/runner/discover.d.ts +3 -0
  482. package/dist/runner/feedback/human.cjs +9 -4
  483. package/dist/runner/feedback/human.cjs.map +1 -1
  484. package/dist/runner/fingerprint.cjs +1 -0
  485. package/dist/runner/fingerprint.cjs.map +1 -1
  486. package/dist/runner/judge-usage.cjs +18 -0
  487. package/dist/runner/judge-usage.cjs.map +1 -0
  488. package/dist/runner/judge-usage.d.cts +5 -0
  489. package/dist/runner/judge-usage.d.mts +5 -0
  490. package/dist/runner/judge-usage.d.ts +5 -0
  491. package/dist/runner/judge-usage.js +8 -0
  492. package/dist/runner/judge-usage.js.map +1 -0
  493. package/dist/runner/project-target.cjs.map +1 -1
  494. package/dist/runner/project-target.d.cts +2 -0
  495. package/dist/runner/project-target.d.mts +2 -0
  496. package/dist/runner/project-target.d.ts +2 -0
  497. package/dist/runner/record/planning.cjs +1 -0
  498. package/dist/runner/record/planning.cjs.map +1 -1
  499. package/dist/runner/record.cjs +4 -0
  500. package/dist/runner/record.cjs.map +1 -1
  501. package/dist/runner/reporters/braintrust.cjs +2 -0
  502. package/dist/runner/reporters/braintrust.cjs.map +1 -1
  503. package/dist/runner/reuse-diagnostics.cjs +9 -10
  504. package/dist/runner/reuse-diagnostics.cjs.map +1 -1
  505. package/dist/runner/reuse-diagnostics.d.cts +5 -6
  506. package/dist/runner/reuse-diagnostics.d.mts +5 -6
  507. package/dist/runner/reuse-diagnostics.d.ts +5 -6
  508. package/dist/runner/reuse-diagnostics.js +4 -4
  509. package/dist/runner/reuse-plan.cjs +100 -48
  510. package/dist/runner/reuse-plan.cjs.map +1 -1
  511. package/dist/runner/reuse-plan.d.cts +60 -3
  512. package/dist/runner/reuse-plan.d.mts +60 -3
  513. package/dist/runner/reuse-plan.d.ts +60 -3
  514. package/dist/runner/reuse-plan.js +14 -10
  515. package/dist/runner/run.cjs +8 -6
  516. package/dist/runner/run.cjs.map +1 -1
  517. package/dist/runner/session.cjs +1 -1
  518. package/dist/runner/session.cjs.map +1 -1
  519. package/dist/runner/session.d.cts +1 -1
  520. package/dist/runner/session.d.mts +1 -1
  521. package/dist/runner/session.d.ts +1 -1
  522. package/dist/runner/source-producer.cjs +3 -1
  523. package/dist/runner/source-producer.cjs.map +1 -1
  524. package/dist/runner/types.cjs.map +1 -1
  525. package/dist/runner/types.d.cts +26 -5
  526. package/dist/runner/types.d.mts +26 -5
  527. package/dist/runner/types.d.ts +26 -5
  528. package/dist/sandbox/action.cjs +3 -0
  529. package/dist/sandbox/action.cjs.map +1 -1
  530. package/dist/sandbox/commands.cjs +3 -0
  531. package/dist/sandbox/commands.cjs.map +1 -1
  532. package/dist/sandbox/compose.cjs +6 -0
  533. package/dist/sandbox/compose.cjs.map +1 -1
  534. package/dist/sandbox/docker-profile/registry.cjs +7 -0
  535. package/dist/sandbox/docker-profile/registry.cjs.map +1 -1
  536. package/dist/sandbox/docker-profile/runtime.cjs +3 -0
  537. package/dist/sandbox/docker-profile/runtime.cjs.map +1 -1
  538. package/dist/sandbox/docker-profile/schema.cjs +7 -0
  539. package/dist/sandbox/docker-profile/schema.cjs.map +1 -1
  540. package/dist/sandbox/incus.cjs +7 -0
  541. package/dist/sandbox/incus.cjs.map +1 -1
  542. package/dist/sandbox/layer.cjs +12 -0
  543. package/dist/sandbox/layer.cjs.map +1 -1
  544. package/dist/shared/adapter-cleanup.cjs +18 -0
  545. package/dist/shared/adapter-cleanup.cjs.map +1 -0
  546. package/dist/shared/adapter-cleanup.d.cts +4 -0
  547. package/dist/shared/adapter-cleanup.d.mts +4 -0
  548. package/dist/shared/adapter-cleanup.d.ts +4 -0
  549. package/dist/shared/adapter-cleanup.js +10 -0
  550. package/dist/shared/adapter-cleanup.js.map +1 -0
  551. package/dist/shared/attempt-lifecycle.cjs +3 -0
  552. package/dist/shared/attempt-lifecycle.cjs.map +1 -0
  553. package/dist/shared/attempt-lifecycle.d.cts +13 -0
  554. package/dist/shared/attempt-lifecycle.d.mts +13 -0
  555. package/dist/shared/attempt-lifecycle.d.ts +13 -0
  556. package/dist/shared/attempt-lifecycle.js +4 -0
  557. package/dist/shared/attempt-lifecycle.js.map +1 -0
  558. package/dist/show/contribution.cjs +52 -14
  559. package/dist/show/contribution.cjs.map +1 -1
  560. package/dist/show/contribution.d.cts +23 -1
  561. package/dist/show/contribution.d.mts +23 -1
  562. package/dist/show/contribution.d.ts +23 -1
  563. package/dist/show/model.cjs +8 -1
  564. package/dist/show/model.cjs.map +1 -1
  565. package/dist/show/model.d.cts +10 -1
  566. package/dist/show/model.d.mts +10 -1
  567. package/dist/show/model.d.ts +10 -1
  568. package/dist/show/render.cjs +287 -49
  569. package/dist/show/render.cjs.map +1 -1
  570. package/dist/show/render.d.cts +3 -0
  571. package/dist/show/render.d.mts +3 -0
  572. package/dist/show/render.d.ts +3 -0
  573. package/dist/sources/projector.cjs +3 -0
  574. package/dist/sources/projector.cjs.map +1 -1
  575. package/dist/types.cjs.map +1 -1
  576. package/dist/types.d.cts +1 -1
  577. package/dist/types.d.mts +1 -1
  578. package/dist/types.d.ts +1 -1
  579. package/dist/view/app-dist/.vite/manifest.json +2 -2
  580. package/dist/view/app-dist/assets/index-CmVoXbwG.js +63 -0
  581. package/dist/view/app-dist/assets/{quality-cost-scatter-CFB3wnkN.js → quality-cost-scatter-BuLISaH4.js} +1 -1
  582. package/dist/view/app-dist/index.html +1 -1
  583. package/dist/view/cli/contribution.cjs +54 -21
  584. package/dist/view/cli/contribution.cjs.map +1 -1
  585. package/dist/view/cli/contribution.d.cts +2 -1
  586. package/dist/view/cli/contribution.d.mts +2 -1
  587. package/dist/view/cli/contribution.d.ts +2 -1
  588. package/dist/view/function-runtime.mjs +13231 -11647
  589. package/dist/view/http-protocol.cjs +3 -0
  590. package/dist/view/http-protocol.cjs.map +1 -1
  591. package/dist/view/http-protocol.d.cts +8 -0
  592. package/dist/view/http-protocol.d.mts +8 -0
  593. package/dist/view/http-protocol.d.ts +8 -0
  594. package/dist/view/inspection-host.cjs +17 -1
  595. package/dist/view/inspection-host.cjs.map +1 -1
  596. package/dist/view/render.cjs +1 -0
  597. package/dist/view/render.cjs.map +1 -1
  598. package/dist/view/render.d.cts +2 -1
  599. package/dist/view/render.d.mts +2 -1
  600. package/dist/view/render.d.ts +2 -1
  601. package/dist/view/revision.cjs.map +1 -1
  602. package/dist/view/revision.d.cts +12 -0
  603. package/dist/view/revision.d.mts +12 -0
  604. package/dist/view/revision.d.ts +12 -0
  605. package/dist/view/server.cjs +10 -9
  606. package/dist/view/server.cjs.map +1 -1
  607. package/docs-site/zh/examples/ai-agent-application.mdx +1 -1
  608. package/docs-site/zh/explanation/adapter.mdx +2 -2
  609. package/docs-site/zh/explanation/assert.mdx +60 -24
  610. package/docs-site/zh/explanation/drive.mdx +1 -1
  611. package/docs-site/zh/explanation/evals.mdx +31 -29
  612. package/docs-site/zh/explanation/experiment.mdx +15 -15
  613. package/docs-site/zh/explanation/judge.mdx +23 -15
  614. package/docs-site/zh/explanation/overview.mdx +44 -21
  615. package/docs-site/zh/explanation/record.mdx +28 -13
  616. package/docs-site/zh/explanation/results-flow.mdx +37 -0
  617. package/docs-site/zh/explanation/runner.mdx +27 -6
  618. package/docs-site/zh/index.mdx +22 -115
  619. package/docs-site/zh/introduction.mdx +89 -61
  620. package/docs-site/zh/reference/cli.mdx +50 -26
  621. package/docs-site/zh/reference/define-agent.mdx +37 -5
  622. package/docs-site/zh/reference/define-eval.mdx +130 -29
  623. package/docs-site/zh/reference/define-experiment.mdx +251 -0
  624. package/docs-site/zh/reference/events.mdx +34 -30
  625. package/docs-site/zh/reference/expect.mdx +77 -2
  626. package/docs-site/zh/reference/official-adapters.mdx +2 -2
  627. package/docs-site/zh/troubleshooting/debug-sandbox.mdx +43 -24
  628. package/docs-site/zh/troubleshooting/debugging.mdx +65 -42
  629. package/docs-site/zh/troubleshooting/recover-after-kill.mdx +35 -28
  630. package/docs-site/zh/tutorials/accept-results.mdx +34 -24
  631. package/docs-site/zh/tutorials/agent-feedback-loop.mdx +74 -43
  632. package/docs-site/zh/tutorials/agent-onboarding.mdx +27 -20
  633. package/docs-site/zh/tutorials/authoring.mdx +138 -121
  634. package/docs-site/zh/tutorials/ci-integration.mdx +46 -22
  635. package/docs-site/zh/tutorials/compare-runs.mdx +40 -13
  636. package/docs-site/zh/tutorials/concurrency.mdx +66 -62
  637. package/docs-site/zh/tutorials/configuration.mdx +17 -17
  638. package/docs-site/zh/tutorials/connect-otel.mdx +48 -37
  639. package/docs-site/zh/tutorials/connect-your-agent.mdx +80 -173
  640. package/docs-site/zh/tutorials/control-run-cost.mdx +63 -12
  641. package/docs-site/zh/tutorials/criteria-files.mdx +28 -21
  642. package/docs-site/zh/tutorials/custom-application.mdx +75 -55
  643. package/docs-site/zh/tutorials/dataset-fanout.mdx +32 -22
  644. package/docs-site/zh/tutorials/debug-lifecycle-plan.mdx +38 -31
  645. package/docs-site/zh/tutorials/docker-in-docker.mdx +83 -89
  646. package/docs-site/zh/tutorials/eval-groups.mdx +70 -89
  647. package/docs-site/zh/tutorials/evaluation-kinds.mdx +12 -5
  648. package/docs-site/zh/tutorials/experiment-services.mdx +243 -0
  649. package/docs-site/zh/tutorials/fixtures.mdx +42 -26
  650. package/docs-site/zh/tutorials/handle-execution-failures.mdx +41 -19
  651. package/docs-site/zh/tutorials/install-custom-sandbox-agent.mdx +21 -16
  652. package/docs-site/zh/tutorials/niceeval-taste-application.mdx +394 -0
  653. package/docs-site/zh/tutorials/niceeval-taste.mdx +190 -0
  654. package/docs-site/zh/tutorials/nixos-managed-dind.mdx +51 -39
  655. package/docs-site/zh/tutorials/plugins.mdx +55 -31
  656. package/docs-site/zh/tutorials/prebuilt-sandbox.mdx +204 -0
  657. package/docs-site/zh/tutorials/quickstart.mdx +91 -47
  658. package/docs-site/zh/tutorials/reporters.mdx +45 -30
  659. package/docs-site/zh/tutorials/rerun-and-cache.mdx +39 -40
  660. package/docs-site/zh/tutorials/sandbox-agent.mdx +65 -40
  661. package/docs-site/zh/tutorials/sandbox-providers.mdx +104 -282
  662. package/docs-site/zh/tutorials/sandbox-reuse.mdx +65 -97
  663. package/docs-site/zh/tutorials/select-and-preview.mdx +25 -20
  664. package/docs-site/zh/tutorials/verify-judge.mdx +49 -26
  665. package/docs-site/zh/tutorials/viewing-results.mdx +130 -80
  666. package/docs-site/zh/tutorials/write-experiment.mdx +116 -298
  667. package/docs-site/zh/tutorials/write-send.mdx +53 -52
  668. package/node_modules/undici/docs/docs/api/Agent.md +8 -8
  669. package/node_modules/undici/docs/docs/api/DiagnosticsChannel.md +9 -6
  670. package/node_modules/undici/docs/docs/api/Errors.md +25 -0
  671. package/node_modules/undici/docs/docs/api/Interceptors.md +4 -3
  672. package/node_modules/undici/docs/docs/api/Pool.md +5 -6
  673. package/node_modules/undici/lib/api/readable.js +1 -2
  674. package/node_modules/undici/lib/core/request.js +23 -2
  675. package/node_modules/undici/lib/core/util.js +0 -39
  676. package/node_modules/undici/lib/dispatcher/client-h1.js +11 -4
  677. package/node_modules/undici/lib/dispatcher/client-h2.js +52 -8
  678. package/node_modules/undici/lib/dispatcher/client.js +1 -1
  679. package/node_modules/undici/lib/dispatcher/dispatcher1-wrapper.js +3 -2
  680. package/node_modules/undici/lib/dispatcher/env-http-proxy-agent.js +21 -6
  681. package/node_modules/undici/lib/dispatcher/pool-base.js +63 -2
  682. package/node_modules/undici/lib/dispatcher/pool.js +90 -7
  683. package/node_modules/undici/lib/dispatcher/proxy-agent.js +16 -3
  684. package/node_modules/undici/lib/handler/deduplication-handler.js +24 -11
  685. package/node_modules/undici/lib/handler/redirect-handler.js +5 -0
  686. package/node_modules/undici/lib/handler/retry-handler.js +72 -13
  687. package/node_modules/undici/lib/interceptor/cache.js +4 -1
  688. package/node_modules/undici/lib/interceptor/decompress.js +236 -69
  689. package/node_modules/undici/lib/web/cache/cache.js +17 -8
  690. package/node_modules/undici/lib/web/cache/cachestorage.js +11 -5
  691. package/node_modules/undici/lib/web/cookies/index.js +19 -7
  692. package/node_modules/undici/lib/web/cookies/util.js +3 -1
  693. package/node_modules/undici/lib/web/eventsource/eventsource.js +25 -1
  694. package/node_modules/undici/lib/web/fetch/body.js +22 -15
  695. package/node_modules/undici/lib/web/fetch/data-url.js +6 -6
  696. package/node_modules/undici/lib/web/fetch/formdata.js +30 -36
  697. package/node_modules/undici/lib/web/fetch/headers.js +34 -31
  698. package/node_modules/undici/lib/web/fetch/index.js +38 -14
  699. package/node_modules/undici/lib/web/fetch/request.js +74 -79
  700. package/node_modules/undici/lib/web/fetch/response.js +43 -44
  701. package/node_modules/undici/lib/web/fetch/util.js +7 -7
  702. package/node_modules/undici/lib/web/infra/index.js +16 -1
  703. package/node_modules/undici/lib/web/webidl/index.js +2 -14
  704. package/node_modules/undici/lib/web/websocket/events.js +43 -26
  705. package/node_modules/undici/lib/web/websocket/receiver.js +5 -0
  706. package/node_modules/undici/lib/web/websocket/stream/websocketerror.js +13 -10
  707. package/node_modules/undici/lib/web/websocket/websocket.js +44 -41
  708. package/node_modules/undici/package.json +1 -1
  709. package/node_modules/undici/types/diagnostics-channel.d.ts +2 -1
  710. package/node_modules/undici/types/interceptors.d.ts +1 -1
  711. package/node_modules/undici/types/webidl.d.ts +13 -7
  712. package/package.json +4 -4
  713. package/dist/view/app-dist/assets/index-DFzw1L97.js +0 -63
  714. package/docs-site/zh/tutorials/experiments.mdx +0 -101
@@ -1,3 +1,4 @@
1
+ import type { ModelSlotSelection, ResolvedModelSlots } from "../model-slots.cjs";
1
2
  import type { AdapterFlagsParser, AdapterFlagsOutput, AdapterFlagsValue } from "../adapter-flags.cjs";
2
3
  import type { Effect } from "effect";
3
4
  import type { ExperimentFlags, FlagValue, JsonValue, LocalizedText, ScopedFeedback, SourceArtifact, Verdict } from "../shared/types.cjs";
@@ -22,6 +23,7 @@ import type { PluginInstance, PluginOnUnavailable } from "../plugin/contracts.cj
22
23
  * 这里不复制(见 docs/feature/record/architecture.md「run.json」)。
23
24
  */
24
25
  export interface ExperimentRunInfo {
26
+ readonly models?: ResolvedModelSlots;
25
27
  description?: string;
26
28
  reasoningEffort?: string;
27
29
  flags?: globalThis.Record<string, JsonValue>;
@@ -814,11 +816,13 @@ export interface SandboxCacheConfig {
814
816
  export declare function resolveSandboxSetupCache(cli: SandboxSetupCache | undefined, experiment: SandboxCacheConfig | undefined, config: SandboxCacheConfig | undefined): SandboxSetupCache;
815
817
  /** Experiment 作者自行选择的字段;不包含路径 id 与 factory 品牌。 */
816
818
  export interface ExperimentAuthorFields {
819
+ /** 为自定义 Adapter 的各个具名模型槽选择模型;不能与单模型简写同时使用。 */
820
+ models?: Readonly<Record<string, ModelSlotSelection>>;
817
821
  /** 一句话描述,展示在 view / CLI 里;纯说明,不影响调度或打分。 */
818
822
  description?: string;
819
- /** Conversation Adapter shorthand. Exactly one of `agent` and `adapter` is required. */
823
+ /** 对话 Adapter 的简写;`agent` 与 `adapter` 必须且只能提供一个。 */
820
824
  agent?: Agent;
821
- /** The Adapter implementation selected for this Experiment. */
825
+ /** 本实验使用的 Adapter 实现。 */
822
826
  adapter?: Adapter;
823
827
  /** 单个模型(agent 留空时实验决定);省略=用 agent 原生默认。跨模型对比写多个实验文件,别用数组。 */
824
828
  model?: string;
@@ -863,9 +867,9 @@ export interface ExperimentAuthorFields {
863
867
  * 每个配对恰好一方提供 template-bearing layer。
864
868
  */
865
869
  sandbox?: SandboxLayer;
866
- /** Host execution policy; unlike the Sandbox layer, this does not participate in identity. */
870
+ /** Host 的执行缓存策略;与 Sandbox 声明层不同,此字段不参与身份计算。 */
867
871
  sandboxCache?: SandboxCacheConfig;
868
- /** Explicit Experiment Plugin occurrences, normalized by defineExperiment(). */
872
+ /** 本实验显式声明的 Plugin 实例列表,由 `defineExperiment()` 规范化。 */
869
873
  plugins?: readonly PluginInstance<"experiment">[];
870
874
  /** 同一 Run 内复用沙箱;这种运行与历史携带双向隔离。 */
871
875
  sandboxReuse?: boolean;
@@ -936,11 +940,27 @@ type ExperimentAdapterSelection<A extends Adapter> = {
936
940
  readonly agent: A & Agent;
937
941
  readonly adapter?: never;
938
942
  };
939
- export type ExperimentInput<A extends Adapter = Adapter> = Omit<ExperimentAuthorFields, "agent" | "adapter" | "flags"> & ExperimentAdapterSelection<A> & ExperimentFlagsFields<NoInfer<A>> & {
943
+ type ExperimentModelFields<A extends Adapter> = A extends {
944
+ readonly kind: "custom";
945
+ } ? ({
946
+ readonly models: Readonly<Record<string, ModelSlotSelection>>;
947
+ readonly model?: never;
948
+ readonly reasoningEffort?: never;
949
+ } | {
950
+ readonly models?: never;
951
+ readonly model?: string;
952
+ readonly reasoningEffort?: string;
953
+ }) : {
954
+ readonly models?: never;
955
+ readonly model?: string;
956
+ readonly reasoningEffort?: string;
957
+ };
958
+ export type ExperimentInput<A extends Adapter = Adapter> = Omit<ExperimentAuthorFields, "agent" | "adapter" | "flags" | "models" | "model" | "reasoningEffort"> & ExperimentModelFields<A> & ExperimentAdapterSelection<A> & ExperimentFlagsFields<NoInfer<A>> & {
940
959
  id?: ExperimentIdComesFromFilePath;
941
960
  };
942
961
  /** Factory 完成默认归一后的 Experiment 字段;无默认语义的 Hook 仍保持作者声明。 */
943
962
  export interface ExperimentDefinition<A extends Adapter = Adapter> {
963
+ readonly models?: ResolvedModelSlots;
944
964
  readonly description?: string;
945
965
  readonly adapter: A;
946
966
  /** Present when the selected Adapter is an Agent, regardless of the input shorthand. */
@@ -1095,6 +1115,7 @@ export declare function runWho(run: {
1095
1115
  }): string;
1096
1116
  /** 一个 (agent, model, flags) 的运行配置 —— 由 CLI / 实验展开。 */
1097
1117
  export interface AdapterRun {
1118
+ readonly models?: ResolvedModelSlots;
1098
1119
  readonly adapter: Adapter;
1099
1120
  /** Present exactly when `adapter` is an Agent implementation. */
1100
1121
  readonly agent?: Agent;
@@ -1,3 +1,4 @@
1
+ import type { ModelSlotSelection, ResolvedModelSlots } from "../model-slots.mjs";
1
2
  import type { AdapterFlagsParser, AdapterFlagsOutput, AdapterFlagsValue } from "../adapter-flags.mjs";
2
3
  import type { Effect } from "effect";
3
4
  import type { ExperimentFlags, FlagValue, JsonValue, LocalizedText, ScopedFeedback, SourceArtifact, Verdict } from "../shared/types.mjs";
@@ -22,6 +23,7 @@ import type { PluginInstance, PluginOnUnavailable } from "../plugin/contracts.mj
22
23
  * 这里不复制(见 docs/feature/record/architecture.md「run.json」)。
23
24
  */
24
25
  export interface ExperimentRunInfo {
26
+ readonly models?: ResolvedModelSlots;
25
27
  description?: string;
26
28
  reasoningEffort?: string;
27
29
  flags?: globalThis.Record<string, JsonValue>;
@@ -814,11 +816,13 @@ export interface SandboxCacheConfig {
814
816
  export declare function resolveSandboxSetupCache(cli: SandboxSetupCache | undefined, experiment: SandboxCacheConfig | undefined, config: SandboxCacheConfig | undefined): SandboxSetupCache;
815
817
  /** Experiment 作者自行选择的字段;不包含路径 id 与 factory 品牌。 */
816
818
  export interface ExperimentAuthorFields {
819
+ /** 为自定义 Adapter 的各个具名模型槽选择模型;不能与单模型简写同时使用。 */
820
+ models?: Readonly<Record<string, ModelSlotSelection>>;
817
821
  /** 一句话描述,展示在 view / CLI 里;纯说明,不影响调度或打分。 */
818
822
  description?: string;
819
- /** Conversation Adapter shorthand. Exactly one of `agent` and `adapter` is required. */
823
+ /** 对话 Adapter 的简写;`agent` 与 `adapter` 必须且只能提供一个。 */
820
824
  agent?: Agent;
821
- /** The Adapter implementation selected for this Experiment. */
825
+ /** 本实验使用的 Adapter 实现。 */
822
826
  adapter?: Adapter;
823
827
  /** 单个模型(agent 留空时实验决定);省略=用 agent 原生默认。跨模型对比写多个实验文件,别用数组。 */
824
828
  model?: string;
@@ -863,9 +867,9 @@ export interface ExperimentAuthorFields {
863
867
  * 每个配对恰好一方提供 template-bearing layer。
864
868
  */
865
869
  sandbox?: SandboxLayer;
866
- /** Host execution policy; unlike the Sandbox layer, this does not participate in identity. */
870
+ /** Host 的执行缓存策略;与 Sandbox 声明层不同,此字段不参与身份计算。 */
867
871
  sandboxCache?: SandboxCacheConfig;
868
- /** Explicit Experiment Plugin occurrences, normalized by defineExperiment(). */
872
+ /** 本实验显式声明的 Plugin 实例列表,由 `defineExperiment()` 规范化。 */
869
873
  plugins?: readonly PluginInstance<"experiment">[];
870
874
  /** 同一 Run 内复用沙箱;这种运行与历史携带双向隔离。 */
871
875
  sandboxReuse?: boolean;
@@ -936,11 +940,27 @@ type ExperimentAdapterSelection<A extends Adapter> = {
936
940
  readonly agent: A & Agent;
937
941
  readonly adapter?: never;
938
942
  };
939
- export type ExperimentInput<A extends Adapter = Adapter> = Omit<ExperimentAuthorFields, "agent" | "adapter" | "flags"> & ExperimentAdapterSelection<A> & ExperimentFlagsFields<NoInfer<A>> & {
943
+ type ExperimentModelFields<A extends Adapter> = A extends {
944
+ readonly kind: "custom";
945
+ } ? ({
946
+ readonly models: Readonly<Record<string, ModelSlotSelection>>;
947
+ readonly model?: never;
948
+ readonly reasoningEffort?: never;
949
+ } | {
950
+ readonly models?: never;
951
+ readonly model?: string;
952
+ readonly reasoningEffort?: string;
953
+ }) : {
954
+ readonly models?: never;
955
+ readonly model?: string;
956
+ readonly reasoningEffort?: string;
957
+ };
958
+ export type ExperimentInput<A extends Adapter = Adapter> = Omit<ExperimentAuthorFields, "agent" | "adapter" | "flags" | "models" | "model" | "reasoningEffort"> & ExperimentModelFields<A> & ExperimentAdapterSelection<A> & ExperimentFlagsFields<NoInfer<A>> & {
940
959
  id?: ExperimentIdComesFromFilePath;
941
960
  };
942
961
  /** Factory 完成默认归一后的 Experiment 字段;无默认语义的 Hook 仍保持作者声明。 */
943
962
  export interface ExperimentDefinition<A extends Adapter = Adapter> {
963
+ readonly models?: ResolvedModelSlots;
944
964
  readonly description?: string;
945
965
  readonly adapter: A;
946
966
  /** Present when the selected Adapter is an Agent, regardless of the input shorthand. */
@@ -1095,6 +1115,7 @@ export declare function runWho(run: {
1095
1115
  }): string;
1096
1116
  /** 一个 (agent, model, flags) 的运行配置 —— 由 CLI / 实验展开。 */
1097
1117
  export interface AdapterRun {
1118
+ readonly models?: ResolvedModelSlots;
1098
1119
  readonly adapter: Adapter;
1099
1120
  /** Present exactly when `adapter` is an Agent implementation. */
1100
1121
  readonly agent?: Agent;
@@ -1,3 +1,4 @@
1
+ import type { ModelSlotSelection, ResolvedModelSlots } from "../model-slots.js";
1
2
  import type { AdapterFlagsParser, AdapterFlagsOutput, AdapterFlagsValue } from "../adapter-flags.js";
2
3
  import type { Effect } from "effect";
3
4
  import type { ExperimentFlags, FlagValue, JsonValue, LocalizedText, ScopedFeedback, SourceArtifact, Verdict } from "../shared/types.js";
@@ -22,6 +23,7 @@ import type { PluginInstance, PluginOnUnavailable } from "../plugin/contracts.js
22
23
  * 这里不复制(见 docs/feature/record/architecture.md「run.json」)。
23
24
  */
24
25
  export interface ExperimentRunInfo {
26
+ readonly models?: ResolvedModelSlots;
25
27
  description?: string;
26
28
  reasoningEffort?: string;
27
29
  flags?: globalThis.Record<string, JsonValue>;
@@ -814,11 +816,13 @@ export interface SandboxCacheConfig {
814
816
  export declare function resolveSandboxSetupCache(cli: SandboxSetupCache | undefined, experiment: SandboxCacheConfig | undefined, config: SandboxCacheConfig | undefined): SandboxSetupCache;
815
817
  /** Experiment 作者自行选择的字段;不包含路径 id 与 factory 品牌。 */
816
818
  export interface ExperimentAuthorFields {
819
+ /** 为自定义 Adapter 的各个具名模型槽选择模型;不能与单模型简写同时使用。 */
820
+ models?: Readonly<Record<string, ModelSlotSelection>>;
817
821
  /** 一句话描述,展示在 view / CLI 里;纯说明,不影响调度或打分。 */
818
822
  description?: string;
819
- /** Conversation Adapter shorthand. Exactly one of `agent` and `adapter` is required. */
823
+ /** 对话 Adapter 的简写;`agent` 与 `adapter` 必须且只能提供一个。 */
820
824
  agent?: Agent;
821
- /** The Adapter implementation selected for this Experiment. */
825
+ /** 本实验使用的 Adapter 实现。 */
822
826
  adapter?: Adapter;
823
827
  /** 单个模型(agent 留空时实验决定);省略=用 agent 原生默认。跨模型对比写多个实验文件,别用数组。 */
824
828
  model?: string;
@@ -863,9 +867,9 @@ export interface ExperimentAuthorFields {
863
867
  * 每个配对恰好一方提供 template-bearing layer。
864
868
  */
865
869
  sandbox?: SandboxLayer;
866
- /** Host execution policy; unlike the Sandbox layer, this does not participate in identity. */
870
+ /** Host 的执行缓存策略;与 Sandbox 声明层不同,此字段不参与身份计算。 */
867
871
  sandboxCache?: SandboxCacheConfig;
868
- /** Explicit Experiment Plugin occurrences, normalized by defineExperiment(). */
872
+ /** 本实验显式声明的 Plugin 实例列表,由 `defineExperiment()` 规范化。 */
869
873
  plugins?: readonly PluginInstance<"experiment">[];
870
874
  /** 同一 Run 内复用沙箱;这种运行与历史携带双向隔离。 */
871
875
  sandboxReuse?: boolean;
@@ -936,11 +940,27 @@ type ExperimentAdapterSelection<A extends Adapter> = {
936
940
  readonly agent: A & Agent;
937
941
  readonly adapter?: never;
938
942
  };
939
- export type ExperimentInput<A extends Adapter = Adapter> = Omit<ExperimentAuthorFields, "agent" | "adapter" | "flags"> & ExperimentAdapterSelection<A> & ExperimentFlagsFields<NoInfer<A>> & {
943
+ type ExperimentModelFields<A extends Adapter> = A extends {
944
+ readonly kind: "custom";
945
+ } ? ({
946
+ readonly models: Readonly<Record<string, ModelSlotSelection>>;
947
+ readonly model?: never;
948
+ readonly reasoningEffort?: never;
949
+ } | {
950
+ readonly models?: never;
951
+ readonly model?: string;
952
+ readonly reasoningEffort?: string;
953
+ }) : {
954
+ readonly models?: never;
955
+ readonly model?: string;
956
+ readonly reasoningEffort?: string;
957
+ };
958
+ export type ExperimentInput<A extends Adapter = Adapter> = Omit<ExperimentAuthorFields, "agent" | "adapter" | "flags" | "models" | "model" | "reasoningEffort"> & ExperimentModelFields<A> & ExperimentAdapterSelection<A> & ExperimentFlagsFields<NoInfer<A>> & {
940
959
  id?: ExperimentIdComesFromFilePath;
941
960
  };
942
961
  /** Factory 完成默认归一后的 Experiment 字段;无默认语义的 Hook 仍保持作者声明。 */
943
962
  export interface ExperimentDefinition<A extends Adapter = Adapter> {
963
+ readonly models?: ResolvedModelSlots;
944
964
  readonly description?: string;
945
965
  readonly adapter: A;
946
966
  /** Present when the selected Adapter is an Agent, regardless of the input shorthand. */
@@ -1095,6 +1115,7 @@ export declare function runWho(run: {
1095
1115
  }): string;
1096
1116
  /** 一个 (agent, model, flags) 的运行配置 —— 由 CLI / 实验展开。 */
1097
1117
  export interface AdapterRun {
1118
+ readonly models?: ResolvedModelSlots;
1098
1119
  readonly adapter: Adapter;
1099
1120
  /** Present exactly when `adapter` is an Agent implementation. */
1100
1121
  readonly agent?: Agent;
@@ -547,6 +547,9 @@ function makeAfterAction(data) {
547
547
  SANDBOX_AFTER_ACTION_DATA.set(action, data);
548
548
  return Object.freeze(action);
549
549
  }
550
+ // @concord-code ne-sandbox-define-action
551
+ // @concord-implements docs/feature/sandbox/prepare-commands.md
552
+ // @concord-implements docs/feature/sandbox/architecture.md
550
553
  function defineSandboxAction(definition) {
551
554
  assertRecord(definition, "defineSandboxAction definition");
552
555
  assertOnlyKeys(definition, new Set(["id", "input", "cache", "steps"]), "defineSandboxAction definition");