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 +1 @@
1
- {"version":3,"file":"skills.js","sourceRoot":"","sources":["skills.ts"],"names":[],"mappings":";;AA8BA,sCAUC;AA6BD,gDAGC;AA6FD,4CAeC;AAOD,8BAYC;AAED,4BAIC;AAaD,kDASC;AAMD,8DAcC;AAOD,4DAMC;AAQD,0CAKC;AAjRD,uEAAuE;AACvE,yDAAyD;AACzD,iEAAiE;AACjE,qEAAqE;AACrE,EAAE;AACF,mEAAmE;AACnE,0DAA0D;AAC1D,6DAA6D;AAC7D,kEAAkE;AAClE,wDAAwD;AACxD,6CAAyC;AACzC,+CAA2D;AAC3D,yCAA4F;AAC5F,kDAAsD;AAGtD,mDAAmD;AACnD,MAAM,UAAU,GAAG,mBAAmB,CAAC;AAOvC;;;;;GAKG;AACI,KAAK,UAAU,aAAa,CAAC,OAAgB,EAAE,KAA2B,EAAE,IAA0B;IACzG,MAAM,GAAG,GAAsB,EAAE,CAAC;IAClC,KAAK,MAAM,IAAI,IAAI,KAAK,EAAE,CAAC;QACvB,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,KAAK,OAAO;YAC1B,CAAC,CAAC,MAAM,iBAAiB,CAAC,OAAO,EAAE,IAAI,EAAE,IAAI,CAAC;YAC9C,CAAC,CAAC,MAAM,gBAAgB,CAAC,OAAO,EAAE,IAAI,EAAE,IAAI,CAAC,CAAC,CAAC;IACvD,CAAC;IACD,IAAI,GAAG,CAAC,MAAM;QACV,MAAM,eAAe,CAAC,OAAO,EAAE,CAAC,GAAG,IAAI,CAAC,GAAG,GAAG,CAAC,CAAC,CAAC;IACrD,OAAO,GAAG,CAAC;AACf,CAAC;AACD,+DAA+D;AAC/D,KAAK,UAAU,iBAAiB,CAAC,OAAgB,EAAE,IAEjD,EAAE,IAA0B;IAC1B,MAAM,IAAI,GAAG,IAAI,CAAC,WAAW,IAAI,OAAO,CAAC,GAAG,EAAE,CAAC;IAC/C,MAAM,GAAG,GAAG,IAAA,mBAAO,EAAC,IAAI,EAAE,IAAI,CAAC,IAAI,CAAC,CAAC;IACrC,MAAM,IAAI,GAAG,MAAM,IAAA,eAAI,EAAC,GAAG,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,CAAC;IACpD,IAAI,CAAC,IAAI;QACL,MAAM,IAAI,KAAK,CAAC,qBAAqB,IAAI,CAAC,IAAI,iCAAiC,GAAG,kEAAkE,CAAC,CAAC;IAC1J,IAAI,IAAI,CAAC,WAAW,EAAE,EAAE,CAAC;QACrB,MAAM,KAAK,GAAG,MAAM,YAAY,CAAC,GAAG,CAAC,CAAC;QACtC,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,UAAU,CAAC,EAAE,CAAC;YAC5C,MAAM,IAAI,KAAK,CAAC,0BAA0B,IAAI,CAAC,IAAI,gFAAgF,CAAC,CAAC;QACzI,CAAC;QACD,MAAM,IAAI,GAAG,IAAI,CAAC,IAAI,IAAI,IAAA,oBAAQ,EAAC,GAAG,CAAC,CAAC;QACxC,MAAM,MAAM,GAAG,SAAS,CAAC,KAAK,CAAC,CAAC;QAChC,MAAM,OAAO,CAAC,eAAe,CAAC,GAAG,EAAE,iBAAK,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,EAAE,IAAI,CAAC,CAAC,CAAC;QAC/D,OAAO,EAAE,IAAI,EAAE,OAAO,EAAE,IAAI,EAAE,IAAI,EAAE,IAAI,CAAC,IAAI,EAAE,MAAM,EAAE,CAAC;IAC5D,CAAC;IACD,IAAI,CAAC,IAAI,CAAC,MAAM,EAAE,IAAI,IAAA,mBAAO,EAAC,GAAG,CAAC,CAAC,WAAW,EAAE,KAAK,KAAK,EAAE,CAAC;QACzD,MAAM,IAAI,KAAK,CAAC,qBAAqB,IAAI,CAAC,IAAI,8FAA8F,CAAC,CAAC;IAClJ,CAAC;IACD,MAAM,OAAO,GAAG,MAAM,IAAA,mBAAQ,EAAC,GAAG,EAAE,OAAO,CAAC,CAAC;IAC7C,MAAM,IAAI,GAAG,IAAI,CAAC,IAAI,IAAI,kBAAkB,CAAC,GAAG,CAAC,CAAC;IAClD,MAAM,OAAO,CAAC,SAAS,CAAC,iBAAK,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,EAAE,IAAI,EAAE,UAAU,CAAC,EAAE,OAAO,CAAC,CAAC;IACzE,OAAO,EAAE,IAAI,EAAE,OAAO,EAAE,IAAI,EAAE,IAAI,EAAE,IAAI,CAAC,IAAI,EAAE,MAAM,EAAE,SAAS,CAAC,CAAC,EAAE,IAAI,EAAE,UAAU,EAAE,OAAO,EAAE,CAAC,CAAC,EAAE,CAAC;AACxG,CAAC;AACD,qGAAqG;AACrG,SAAgB,kBAAkB,CAAC,OAAe;IAC9C,MAAM,IAAI,GAAG,IAAA,oBAAQ,EAAC,OAAO,CAAC,CAAC;IAC/B,OAAO,IAAI,CAAC,WAAW,EAAE,KAAK,UAAU,CAAC,CAAC,CAAC,IAAA,oBAAQ,EAAC,IAAA,mBAAO,EAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,IAAA,oBAAQ,EAAC,IAAI,EAAE,IAAA,mBAAO,EAAC,IAAI,CAAC,CAAC,CAAC;AAC1G,CAAC;AACD,2DAA2D;AAC3D,SAAS,SAAS,CAAC,KAGhB;IACC,MAAM,CAAC,GAAG,IAAA,wBAAU,EAAC,QAAQ,CAAC,CAAC;IAC/B,KAAK,MAAM,CAAC,IAAI,CAAC,GAAG,KAAK,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,aAAa,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,EAAE,CAAC;QACtE,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC;IACjE,CAAC;IACD,OAAO,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC;AAC3B,CAAC;AACD,iDAAiD;AACjD,KAAK,UAAU,YAAY,CAAC,GAAW;IAInC,MAAM,GAAG,GAGH,EAAE,CAAC;IACT,MAAM,IAAI,GAAG,KAAK,EAAE,OAAe,EAAiB,EAAE;QAClD,KAAK,MAAM,KAAK,IAAI,MAAM,IAAA,kBAAO,EAAC,OAAO,EAAE,EAAE,aAAa,EAAE,IAAI,EAAE,CAAC,EAAE,CAAC;YAClE,MAAM,IAAI,GAAG,IAAA,gBAAI,EAAC,OAAO,EAAE,KAAK,CAAC,IAAI,CAAC,CAAC;YACvC,IAAI,KAAK,CAAC,WAAW,EAAE,EAAE,CAAC;gBACtB,IAAI,KAAK,CAAC,IAAI,KAAK,MAAM;oBACrB,SAAS;gBACb,MAAM,IAAI,CAAC,IAAI,CAAC,CAAC;YACrB,CAAC;iBACI,IAAI,KAAK,CAAC,MAAM,EAAE,EAAE,CAAC;gBACtB,GAAG,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,IAAA,oBAAQ,EAAC,GAAG,EAAE,IAAI,CAAC,CAAC,KAAK,CAAC,eAAG,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,EAAE,OAAO,EAAE,MAAM,IAAA,mBAAQ,EAAC,IAAI,CAAC,EAAE,CAAC,CAAC;YAChG,CAAC;QACL,CAAC;IACL,CAAC,CAAC;IACF,MAAM,IAAI,CAAC,GAAG,CAAC,CAAC;IAChB,OAAO,GAAG,CAAC;AACf,CAAC;AACD,iEAAiE;AACjE,KAAK,UAAU,gBAAgB,CAAC,OAAgB,EAAE,IAEhD,EAAE,IAA0B;IAC1B,MAAM,QAAQ,GAAG,MAAM,SAAS,CAAC,OAAO,EAAE,IAAI,CAAC,MAAM,EAAE,IAAI,CAAC,GAAG,CAAC,CAAC;IACjE,IAAI,CAAC;QACD,MAAM,SAAS,GAAG,MAAM,cAAc,CAAC,OAAO,EAAE,QAAQ,EAAE,IAAI,CAAC,MAAM,CAAC,CAAC;QACvE,MAAM,QAAQ,GAAG,gBAAgB,CAAC,SAAS,EAAE,IAAI,CAAC,CAAC;QACnD,KAAK,MAAM,IAAI,IAAI,QAAQ,EAAE,CAAC;YAC1B,MAAM,GAAG,GAAG,SAAS,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,IAAI,CAAE,CAAC,GAAG,CAAC;YACxD,MAAM,IAAI,GAAG,iBAAK,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,EAAE,IAAI,CAAC,CAAC;YACxC,MAAM,GAAG,GAAG,MAAM,OAAO,CAAC,QAAQ,CAAC,YAAY,IAAA,qBAAC,EAAC,IAAI,CAAC,GAAG,CAAC,cAAc,IAAA,qBAAC,EAAC,IAAI,CAAC,aAAa,IAAA,qBAAC,EAAC,GAAG,CAAC,IAAI,IAAA,qBAAC,EAAC,IAAI,CAAC,EAAE,CAAC,CAAC;YACjH,IAAI,GAAG,CAAC,QAAQ,KAAK,CAAC,EAAE,CAAC;gBACrB,MAAM,IAAI,KAAK,CAAC,4BAA4B,IAAI,UAAU,IAAI;EAC5E,IAAI,CAAC,GAAG,CAAC,MAAM,GAAG,GAAG,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC;YACvB,CAAC;QACL,CAAC;QACD,OAAO;YACH,IAAI,EAAE,MAAM;YACZ,MAAM,EAAE,IAAI,CAAC,MAAM;YACnB,GAAG,CAAC,IAAI,CAAC,GAAG,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,GAAG,EAAE,IAAI,CAAC,GAAG,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;YACpD,MAAM,EAAE,QAAQ;SACnB,CAAC;IACN,CAAC;YACO,CAAC;QACL,MAAM,OAAO,CAAC,QAAQ,CAAC,UAAU,IAAA,qBAAC,EAAC,QAAQ,CAAC,EAAE,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,CAAC;IAC3E,CAAC;AACL,CAAC;AAMD;;;GAGG;AACH,KAAK,UAAU,cAAc,CAAC,OAAgB,EAAE,QAAgB,EAAE,MAAc;IAC5E,MAAM,GAAG,GAAG,MAAM,OAAO,CAAC,QAAQ,CAAC,QAAQ,IAAA,qBAAC,EAAC,QAAQ,CAAC,uCAAuC,CAAC,CAAC;IAC/F,MAAM,KAAK,GAAG,GAAG,CAAC,MAAM,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,IAAI,EAAE,CAAC;IACjF,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC;QAClB,MAAM,IAAI,KAAK,CAAC,cAAc,MAAM,wBAAwB,CAAC,CAAC;IAClE,MAAM,QAAQ,GAAG,GAAG,QAAQ,WAAW,CAAC;IACxC,IAAI,KAAK,CAAC,QAAQ,CAAC,QAAQ,CAAC;QACxB,OAAO,CAAC,EAAE,IAAI,EAAE,QAAQ,CAAC,MAAM,CAAC,EAAE,GAAG,EAAE,QAAQ,EAAE,CAAC,CAAC;IACvD,OAAO,KAAK;SACP,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE;QACX,MAAM,GAAG,GAAG,iBAAK,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC;QAC7B,OAAO,EAAE,IAAI,EAAE,iBAAK,CAAC,QAAQ,CAAC,GAAG,CAAC,EAAE,GAAG,EAAE,CAAC;IAC9C,CAAC,CAAC;SACG,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,aAAa,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,sBAAsB;AAC7E,CAAC;AACD;;;GAGG;AACH,SAAgB,gBAAgB,CAAC,SAAqC,EAAE,IAEtE;IACE,MAAM,KAAK,GAAG,SAAS,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC;IAC3C,IAAI,CAAC,IAAI,CAAC,MAAM,EAAE,MAAM,EAAE,CAAC;QACvB,IAAI,SAAS,CAAC,MAAM,KAAK,CAAC;YACtB,OAAO,CAAC,KAAK,CAAC,CAAC,CAAE,CAAC,CAAC;QACvB,MAAM,IAAI,KAAK,CAAC,cAAc,IAAI,CAAC,MAAM,6FAA6F,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;IAC/J,CAAC;IACD,KAAK,MAAM,IAAI,IAAI,IAAI,CAAC,MAAM,EAAE,CAAC;QAC7B,IAAI,CAAC,KAAK,CAAC,QAAQ,CAAC,IAAI,CAAC,EAAE,CAAC;YACxB,MAAM,IAAI,KAAK,CAAC,cAAc,IAAI,CAAC,MAAM,UAAU,IAAI,CAAC,GAAG,IAAI,WAAW,yBAAyB,IAAI,iBAAiB,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;QACjJ,CAAC;IACL,CAAC;IACD,OAAO,CAAC,GAAG,IAAI,CAAC,MAAM,CAAC,CAAC;AAC5B,CAAC;AACD,0DAA0D;AAC1D;;;;GAIG;AACI,KAAK,UAAU,SAAS,CAAC,OAAgB,EAAE,MAAc,EAAE,GAAY;IAC1E,MAAM,GAAG,GAAG,GAAG,UAAU,IAAI,IAAI,CAAC,MAAM,CAAC,IAAI,IAAI,CAAC,MAAM,EAAE,CAAC,QAAQ,CAAC,EAAE,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,EAAE,CAAC;IACtF,MAAM,GAAG,GAAG,QAAQ,CAAC,MAAM,CAAC,CAAC;IAC7B,MAAM,MAAM,GAAG,UAAU,IAAA,qBAAC,EAAC,GAAG,CAAC,gBAAgB,IAAA,qBAAC,EAAC,UAAU,CAAC,MAAM;QAC9D,qBAAqB,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,YAAY,GAAG,IAAA,qBAAC,EAAC,GAAG,CAAC,IAAI,IAAA,qBAAC,EAAC,GAAG,CAAC,EAAE;QACjE,CAAC,GAAG,CAAC,CAAC,CAAC,cAAc,IAAA,qBAAC,EAAC,GAAG,CAAC,qBAAqB,IAAA,qBAAC,EAAC,GAAG,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;IACnE,MAAM,GAAG,GAAG,MAAM,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAC,CAAC;IAC3C,IAAI,GAAG,CAAC,QAAQ,KAAK,CAAC,EAAE,CAAC;QACrB,MAAM,IAAI,KAAK,CAAC,8BAA8B,MAAM,UAAU,GAAG,IAAI,WAAW;EACtF,IAAI,CAAC,GAAG,CAAC,MAAM,GAAG,GAAG,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC;IAC/B,CAAC;IACD,OAAO,GAAG,CAAC;AACf,CAAC;AACD,qEAAqE;AACrE,SAAgB,QAAQ,CAAC,MAAc;IACnC,IAAI,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,MAAM,CAAC,UAAU,CAAC,MAAM,CAAC;QACnD,OAAO,MAAM,CAAC;IAClB,OAAO,sBAAsB,MAAM,MAAM,CAAC;AAC9C,CAAC;AACD,SAAS,QAAQ,CAAC,MAAc;IAC5B,MAAM,IAAI,GAAG,QAAQ,CAAC,MAAM,CAAC,CAAC,OAAO,CAAC,MAAM,EAAE,EAAE,CAAC,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,GAAG,EAAE,IAAI,MAAM,CAAC;IAC7E,OAAO,IAAI,CAAC,OAAO,CAAC,QAAQ,EAAE,EAAE,CAAC,CAAC;AACtC,CAAC;AACD,SAAS,IAAI,CAAC,MAAc;IACxB,OAAO,MAAM,CAAC,OAAO,CAAC,UAAU,EAAE,GAAG,CAAC,CAAC;AAC3C,CAAC;AACD,SAAS,IAAI,CAAC,MAAc,EAAE,CAAC,GAAG,EAAE;IAChC,OAAO,MAAM,CAAC,IAAI,EAAE,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AAC1D,CAAC;AACD,2DAA2D;AAC3D,mDAAmD;AACnD,SAAgB,mBAAmB,CAAC,MAAkC;IAClE,MAAM,KAAK,GAAa,EAAE,CAAC;IAC3B,KAAK,MAAM,KAAK,IAAI,MAAM,EAAE,CAAC;QACzB,IAAI,KAAK,CAAC,IAAI,KAAK,OAAO;YACtB,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;;YAEvB,KAAK,CAAC,IAAI,CAAC,GAAG,KAAK,CAAC,MAAM,CAAC,CAAC;IACpC,CAAC;IACD,OAAO,CAAC,GAAG,IAAI,GAAG,CAAC,KAAK,CAAC,CAAC,CAAC;AAC/B,CAAC;AACD;;;;GAIG;AACH,SAAgB,yBAAyB,CAAC,GAAW,EAAE,MAAyB;IAC5E,MAAM,IAAI,GAAG,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,KAAK,IAAI,KAAK,iBAAK,CAAC,IAAI,CAAC,GAAG,EAAE,IAAI,CAAC,WAAW,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;IAC7F,OAAO;QACH,EAAE;QACF,WAAW;QACX,EAAE;QACF,kCAAkC,GAAG,MAAM;QAC3C,EAAE;QACF,IAAI;QACJ,EAAE;QACF,8FAA8F;QAC9F,+EAA+E;QAC/E,EAAE;KACL,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AACjB,CAAC;AACD;;;;;GAKG;AACI,KAAK,UAAU,wBAAwB,CAAC,OAA2D,EAAE,IAAY,EAAE,IAAI,GAAG,WAAW;IACxI,MAAM,OAAO,GAAG,MAAM,OAAO,CAAC,UAAU,CAAC,IAAI,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,KAAK,CAAC,CAAC;IAClE,MAAM,KAAK,GAAG,gBAAgB,IAAI,CAAC,MAAM,EAAE,CAAC,QAAQ,CAAC,EAAE,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,EAAE,CAAC;IACvE,MAAM,OAAO,CAAC,QAAQ,CAAC,UAAU,IAAA,qBAAC,EAAC,IAAI,CAAC,OAAO,KAAK,MAAM,IAAI,KAAK,KAAK,IAAI,CAAC,CAAC;IAC9E,IAAI,CAAC,OAAO;QACR,MAAM,eAAe,CAAC,OAAO,EAAE,CAAC,IAAI,CAAC,CAAC,CAAC;AAC/C,CAAC;AACD;;;;;;GAMG;AACI,KAAK,UAAU,eAAe,CAAC,OAA4C,EAAE,KAAwB;IACxG,IAAI,CAAC,KAAK,CAAC,MAAM;QACb,OAAO;IACX,MAAM,KAAK,GAAG,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,kBAAkB,IAAA,qBAAC,EAAC,CAAC,CAAC,uBAAuB,CAAC,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;IAC3F,MAAM,OAAO,CAAC,QAAQ,CAAC,yCAAyC,KAAK,EAAE,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,CAAC;AACpG,CAAC","sourcesContent":["// SkillSpec 的安装实现 —— 三个内置沙箱型 adapter(claude-code / codex / bub)共用的那一半:\n// 「从哪里取得哪份 Skill」(本地目录/文件、repo + ref + 选择集)。另一半——装到哪个目录、\n// 要不要额外写发现指引——留在各 adapter 里,由它们传 `dir` 并自己写 project instruction。\n// 定稿见 docs/feature/adapters/architecture/coding-agent-extensions.md。\n//\n// 为什么不调 `npx skills add`:那个 installer 没有钉 ref 的入口,也没有机器可读的「这个 repo\n// 里有哪些 skill」输出(-l 只打人看的带 ANSI 的清单),而契约要求「来源必须可复现」(钉 ref)\n// 与「多 Skill Repo 必须明确选择启用集合」(选不中要报出可选集)。所以 repo skill 走 git:\n// clone → 枚举 SKILL.md → 按规则选 → 拷进 agent 的 skill 目录。顺带绕开了 memory 里\n// npx-skills-add-headless-hang 那个交互式选择框(无 tty 卡死)的整类问题。\nimport { createHash } from \"node:crypto\";\nimport { readFile, readdir, stat } from \"node:fs/promises\";\nimport { basename, dirname, extname, join, posix, relative, resolve, sep } from \"node:path\";\nimport { shellQuote as q } from \"../sandbox/shell.ts\";\nimport type { SandboxOperations } from \"../sandbox/types.ts\";\nimport type { AgentSetupSkill, Sandbox, SkillSpec } from \"../types.ts\";\n/** 沙箱里放临时 clone 的地方(装完即删;不在 workdir 下,不进 diff)。 */\nconst CLONE_ROOT = \"/tmp/niceeval-git\";\nexport interface InstallSkillsOptions {\n /** 沙箱里的 Skill 安装根目录(相对 workdir):claude-code 用 `.claude/skills`,codex/bub 用 `.agents/skills`。 */\n dir: string;\n /** 本地 Skill 的解析根;省略用跑 niceeval 的项目根(process.cwd())。 */\n projectRoot?: string;\n}\n/**\n * 按配置顺序把 SkillSpec 装进沙箱,返回 manifest 里的 Skill 记录(与配置顺序一一对应)。\n * 同名 Skill 来自多个来源时后装的覆盖磁盘、manifest 保留每一条(不静默合并)。\n * 任何一条装不上(路径不存在 / 形状不支持 / clone 或 ref 解析失败 / 选择集不合法)直接抛错 ——\n * setup 失败即 attempt errored,不伪装成断言失败。\n */\nexport async function installSkills(sandbox: Sandbox, specs: readonly SkillSpec[], opts: InstallSkillsOptions): Promise<AgentSetupSkill[]> {\n const out: AgentSetupSkill[] = [];\n for (const spec of specs) {\n out.push(spec.kind === \"local\"\n ? await installLocalSkill(sandbox, spec, opts)\n : await installRepoSkill(sandbox, spec, opts));\n }\n if (out.length)\n await excludeFromDiff(sandbox, [`${opts.dir}/`]);\n return out;\n}\n// ───────────────────────── 本地 Skill ─────────────────────────\nasync function installLocalSkill(sandbox: Sandbox, spec: Extract<SkillSpec, {\n kind: \"local\";\n}>, opts: InstallSkillsOptions): Promise<AgentSetupSkill> {\n const root = opts.projectRoot ?? process.cwd();\n const abs = resolve(root, spec.path);\n const info = await stat(abs).catch(() => undefined);\n if (!info)\n throw new Error(`Local skill path \"${spec.path}\" does not exist (resolved to ${abs}). Paths are resolved from the project root you run niceeval in.`);\n if (info.isDirectory()) {\n const files = await readDirFiles(abs);\n if (!files.some((f) => f.path === \"SKILL.md\")) {\n throw new Error(`Local skill directory \"${spec.path}\" has no SKILL.md. A directory-shaped skill must contain SKILL.md at its root.`);\n }\n const name = spec.name ?? basename(abs);\n const sha256 = hashFiles(files);\n await sandbox.uploadDirectory(abs, posix.join(opts.dir, name));\n return { kind: \"local\", name, path: spec.path, sha256 };\n }\n if (!info.isFile() || extname(abs).toLowerCase() !== \".md\") {\n throw new Error(`Local skill path \"${spec.path}\" has an unsupported shape. Accepted: a directory containing SKILL.md, or a single .md file.`);\n }\n const content = await readFile(abs, \"utf-8\");\n const name = spec.name ?? localFileSkillName(abs);\n await sandbox.writeText(posix.join(opts.dir, name, \"SKILL.md\"), content);\n return { kind: \"local\", name, path: spec.path, sha256: hashFiles([{ path: \"SKILL.md\", content }]) };\n}\n/** `skills/effect-ts/SKILL.md` → `effect-ts`(SKILL.md 的名字不含信息,取所在目录);`skills/guide.md` → `guide`。 */\nexport function localFileSkillName(absPath: string): string {\n const file = basename(absPath);\n return file.toLowerCase() === \"skill.md\" ? basename(dirname(absPath)) : basename(file, extname(file));\n}\n/** 目录内容的确定性哈希:按相对路径排序,逐条喂 `path\\0content\\0`(二进制附件按原字节)。 */\nfunction hashFiles(files: {\n path: string;\n content: string | Buffer;\n}[]): string {\n const h = createHash(\"sha256\");\n for (const f of [...files].sort((a, b) => a.path.localeCompare(b.path))) {\n h.update(f.path).update(\"\\0\").update(f.content).update(\"\\0\");\n }\n return h.digest(\"hex\");\n}\n/** Skill 目录可以带脚本 / 图片等附件,一律按原字节读 + 上传,不做文本假设。 */\nasync function readDirFiles(dir: string): Promise<{\n path: string;\n content: Buffer;\n}[]> {\n const out: {\n path: string;\n content: Buffer;\n }[] = [];\n const walk = async (current: string): Promise<void> => {\n for (const entry of await readdir(current, { withFileTypes: true })) {\n const full = join(current, entry.name);\n if (entry.isDirectory()) {\n if (entry.name === \".git\")\n continue;\n await walk(full);\n }\n else if (entry.isFile()) {\n out.push({ path: relative(dir, full).split(sep).join(\"/\"), content: await readFile(full) });\n }\n }\n };\n await walk(dir);\n return out;\n}\n// ───────────────────────── Repo Skill ─────────────────────────\nasync function installRepoSkill(sandbox: Sandbox, spec: Extract<SkillSpec, {\n kind: \"repo\";\n}>, opts: InstallSkillsOptions): Promise<AgentSetupSkill> {\n const cloneDir = await cloneRepo(sandbox, spec.source, spec.ref);\n try {\n const available = await discoverSkills(sandbox, cloneDir, spec.source);\n const selected = selectRepoSkills(available, spec);\n for (const name of selected) {\n const src = available.find((s) => s.name === name)!.dir;\n const dest = posix.join(opts.dir, name);\n const res = await sandbox.runShell(`mkdir -p ${q(opts.dir)} && rm -rf ${q(dest)} && cp -R ${q(src)} ${q(dest)}`);\n if (res.exitCode !== 0) {\n throw new Error(`Could not install skill \"${name}\" into ${dest}:\n${tail(res.stdout + res.stderr)}`);\n }\n }\n return {\n kind: \"repo\",\n source: spec.source,\n ...(spec.ref !== undefined ? { ref: spec.ref } : {}),\n skills: selected,\n };\n }\n finally {\n await sandbox.runShell(`rm -rf ${q(cloneDir)}`).catch(() => undefined);\n }\n}\n/** repo 里可用的一个 Skill:名字 + 沙箱内 SKILL.md 所在目录。 */\nexport interface DiscoveredSkill {\n name: string;\n dir: string;\n}\n/**\n * 枚举 clone 里的 Skill:根 `SKILL.md` 存在 → 整个 repo 就是一个 Skill(名字取 repo 名);\n * 否则每个含 `SKILL.md` 的目录是一个 Skill(名字取该目录名),与 `skills` CLI 的默认口径一致。\n */\nasync function discoverSkills(sandbox: Sandbox, cloneDir: string, source: string): Promise<DiscoveredSkill[]> {\n const res = await sandbox.runShell(`find ${q(cloneDir)} -name SKILL.md -not -path '*/.git/*'`);\n const paths = res.stdout.split(\"\\n\").map((l) => l.trim()).filter(Boolean).sort();\n if (paths.length === 0)\n throw new Error(`Repo skill ${source} contains no SKILL.md.`);\n const rootFile = `${cloneDir}/SKILL.md`;\n if (paths.includes(rootFile))\n return [{ name: repoName(source), dir: cloneDir }];\n return paths\n .map((p) => {\n const dir = posix.dirname(p);\n return { name: posix.basename(dir), dir };\n })\n .sort((a, b) => a.name.localeCompare(b.name)); // 报错里的「可选集」按名字排,不按路径排\n}\n/**\n * 选择规则(docs 的表):repo 只有一个 Skill 且省略 `skills` → 装唯一那个;多个且省略 →\n * 失败并列出可选集;指定了不存在的 Skill → 失败并报 source / ref / 名字。\n */\nexport function selectRepoSkills(available: readonly DiscoveredSkill[], spec: Extract<SkillSpec, {\n kind: \"repo\";\n}>): string[] {\n const names = available.map((s) => s.name);\n if (!spec.skills?.length) {\n if (available.length === 1)\n return [names[0]!];\n throw new Error(`Repo skill ${spec.source} contains multiple skills; select which ones to enable with \\`skills: [...]\\`. Available: ${names.join(\", \")}.`);\n }\n for (const want of spec.skills) {\n if (!names.includes(want)) {\n throw new Error(`Repo skill ${spec.source} (ref: ${spec.ref ?? \"(default)\"}) has no skill named \"${want}\". Available: ${names.join(\", \")}.`);\n }\n }\n return [...spec.skills];\n}\n// ───────────────────────── git ─────────────────────────\n/**\n * 把一个 repo clone 进沙箱临时目录,给了 ref 就钉到那个 tag/commit/branch,返回 clone 目录。\n * ref 可能是任意 commit,所以钉 ref 时不能 `--depth 1`(浅克隆 checkout 不到历史 commit)。\n * Skill 与 native plugin 的 marketplace(claude 侧 CLI 没有钉 ref 的入口)共用这一条。\n */\nexport async function cloneRepo(sandbox: Sandbox, source: string, ref?: string): Promise<string> {\n const dir = `${CLONE_ROOT}/${slug(source)}-${Math.random().toString(36).slice(2, 8)}`;\n const url = gitUrlOf(source);\n const script = `rm -rf ${q(dir)} && mkdir -p ${q(CLONE_ROOT)} && ` +\n `git clone --quiet ${ref ? \"\" : \"--depth 1 \"}${q(url)} ${q(dir)}` +\n (ref ? ` && git -C ${q(dir)} checkout --quiet ${q(ref)}` : \"\");\n const res = await sandbox.runShell(script);\n if (res.exitCode !== 0) {\n throw new Error(`Could not fetch repo skill ${source} (ref: ${ref ?? \"(default)\"}):\n${tail(res.stdout + res.stderr)}`);\n }\n return dir;\n}\n/** `owner/repo` → GitHub HTTPS URL;已经是 URL(含 `://` 或 `git@`)则原样用。 */\nexport function gitUrlOf(source: string): string {\n if (source.includes(\"://\") || source.startsWith(\"git@\"))\n return source;\n return `https://github.com/${source}.git`;\n}\nfunction repoName(source: string): string {\n const last = gitUrlOf(source).replace(/\\/+$/, \"\").split(\"/\").pop() ?? source;\n return last.replace(/\\.git$/, \"\");\n}\nfunction slug(source: string): string {\n return source.replace(/[^\\w.-]/g, \"-\");\n}\nfunction tail(output: string, n = 12): string {\n return output.trim().split(\"\\n\").slice(-n).join(\"\\n\");\n}\n// ───────────────────────── 发现指引 ─────────────────────────\n/** manifest 里的 Skill 记录 → 沙箱里实际的目录名(发现指引要逐条点名)。 */\nexport function installedSkillNames(skills: readonly AgentSetupSkill[]): string[] {\n const names: string[] = [];\n for (const skill of skills) {\n if (skill.kind === \"local\")\n names.push(skill.name);\n else\n names.push(...skill.skills);\n }\n return [...new Set(names)];\n}\n/**\n * 给没有原生 Skill 加载机制的 agent(codex / bub)写的 project instruction 段落:\n * 只把文件装到 skill 目录不足以让它们读到(见 memory/codex-no-native-skill-tool.md ——\n * 不提示的话 codex 连一次 shell 读取都不会发生),必须在 AGENTS.md 里给出稳定的发现指引。\n */\nexport function skillDiscoveryInstruction(dir: string, skills: readonly string[]): string {\n const list = skills.map((name) => `- ${name}: ${posix.join(dir, name)}/SKILL.md`).join(\"\\n\");\n return [\n \"\",\n \"## Skills\",\n \"\",\n `Skill files are installed in \\`${dir}/\\`:`,\n \"\",\n list,\n \"\",\n \"Before answering a task, check whether one of these skills covers it and read its `SKILL.md`\",\n \"(e.g. with a shell command). Follow the skill's instructions when it applies.\",\n \"\",\n ].join(\"\\n\");\n}\n/**\n * 把发现指引追加进沙箱里的一个 instruction 文件(不存在则创建)。落在 workspace 里时,\n * 新建的文件顺手排除出 diff —— setup 装的东西不是 agent 的产出(见 excludeFromDiff)。\n * 文件本来就存在(starter 自带 AGENTS.md)时不排除:那是 agent 也可能改的文件,\n * 排掉会把真实改动一起藏起来;此时追加的这一段会如实出现在 diff 里。\n */\nexport async function appendProjectInstruction(sandbox: Pick<SandboxOperations, \"pathExists\" | \"runShell\">, text: string, file = \"AGENTS.md\"): Promise<void> {\n const existed = await sandbox.pathExists(file).catch(() => false);\n const delim = `NICEEVAL_EOF_${Math.random().toString(36).slice(2, 8)}`;\n await sandbox.runShell(`cat >> ${q(file)} <<'${delim}'\\n${text}\\n${delim}\\n`);\n if (!existed)\n await excludeFromDiff(sandbox, [file]);\n}\n/**\n * 把 setup 装进 workspace 的东西(skill 目录、adapter 新建的 instruction 文件)写进\n * `.git/info/exclude`:git 基线是在 agent.setup **之前**打的,不排除的话这些文件会被\n * `captureGeneratedFiles`(git add -A && git diff HEAD)当成「agent 生成的文件」记进\n * diff.json —— 装了 skill 的实验,diff 里凭空多出几十个文件,基于 diff 的断言与展示全被污染。\n * 排除只对未跟踪文件有效,这正是 adapter 自己创建的那一类。\n */\nexport async function excludeFromDiff(sandbox: Pick<SandboxOperations, \"runShell\">, paths: readonly string[]): Promise<void> {\n if (!paths.length)\n return;\n const lines = paths.map((p) => `printf '%s\\\\n' ${q(p)} >> .git/info/exclude`).join(\" && \");\n await sandbox.runShell(`test -d .git && mkdir -p .git/info && ${lines}`).catch(() => undefined);\n}\n"]}
1
+ {"version":3,"file":"skills.js","sourceRoot":"","sources":["skills.ts"],"names":[],"mappings":";;AA8BA,sCAUC;AA6BD,gDAGC;AA6FD,4CAeC;AAOD,8BAYC;AAED,4BAIC;AAaD,kDASC;AAMD,8DAcC;AAOD,4DAMC;AAQD,0CAKC;AAjRD,uEAAuE;AACvE,yDAAyD;AACzD,iEAAiE;AACjE,qEAAqE;AACrE,EAAE;AACF,mEAAmE;AACnE,0DAA0D;AAC1D,6DAA6D;AAC7D,kEAAkE;AAClE,wDAAwD;AACxD,6CAAyC;AACzC,+CAA2D;AAC3D,yCAA4F;AAC5F,kDAAsD;AAGtD,mDAAmD;AACnD,MAAM,UAAU,GAAG,mBAAmB,CAAC;AAOvC;;;;;GAKG;AACI,KAAK,UAAU,aAAa,CAAC,OAAgB,EAAE,KAA2B,EAAE,IAA0B;IACzG,MAAM,GAAG,GAA6B,EAAE,CAAC;IACzC,KAAK,MAAM,IAAI,IAAI,KAAK,EAAE,CAAC;QACvB,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,KAAK,OAAO;YAC1B,CAAC,CAAC,MAAM,iBAAiB,CAAC,OAAO,EAAE,IAAI,EAAE,IAAI,CAAC;YAC9C,CAAC,CAAC,MAAM,gBAAgB,CAAC,OAAO,EAAE,IAAI,EAAE,IAAI,CAAC,CAAC,CAAC;IACvD,CAAC;IACD,IAAI,GAAG,CAAC,MAAM;QACV,MAAM,eAAe,CAAC,OAAO,EAAE,CAAC,GAAG,IAAI,CAAC,GAAG,GAAG,CAAC,CAAC,CAAC;IACrD,OAAO,GAAG,CAAC;AACf,CAAC;AACD,+DAA+D;AAC/D,KAAK,UAAU,iBAAiB,CAAC,OAAgB,EAAE,IAEjD,EAAE,IAA0B;IAC1B,MAAM,IAAI,GAAG,IAAI,CAAC,WAAW,IAAI,OAAO,CAAC,GAAG,EAAE,CAAC;IAC/C,MAAM,GAAG,GAAG,IAAA,mBAAO,EAAC,IAAI,EAAE,IAAI,CAAC,IAAI,CAAC,CAAC;IACrC,MAAM,IAAI,GAAG,MAAM,IAAA,eAAI,EAAC,GAAG,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,CAAC;IACpD,IAAI,CAAC,IAAI;QACL,MAAM,IAAI,KAAK,CAAC,qBAAqB,IAAI,CAAC,IAAI,iCAAiC,GAAG,kEAAkE,CAAC,CAAC;IAC1J,IAAI,IAAI,CAAC,WAAW,EAAE,EAAE,CAAC;QACrB,MAAM,KAAK,GAAG,MAAM,YAAY,CAAC,GAAG,CAAC,CAAC;QACtC,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,UAAU,CAAC,EAAE,CAAC;YAC5C,MAAM,IAAI,KAAK,CAAC,0BAA0B,IAAI,CAAC,IAAI,gFAAgF,CAAC,CAAC;QACzI,CAAC;QACD,MAAM,IAAI,GAAG,IAAI,CAAC,IAAI,IAAI,IAAA,oBAAQ,EAAC,GAAG,CAAC,CAAC;QACxC,MAAM,MAAM,GAAG,SAAS,CAAC,KAAK,CAAC,CAAC;QAChC,MAAM,OAAO,CAAC,eAAe,CAAC,GAAG,EAAE,iBAAK,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,EAAE,IAAI,CAAC,CAAC,CAAC;QAC/D,OAAO,EAAE,IAAI,EAAE,OAAO,EAAE,IAAI,EAAE,IAAI,EAAE,IAAI,CAAC,IAAI,EAAE,MAAM,EAAE,CAAC;IAC5D,CAAC;IACD,IAAI,CAAC,IAAI,CAAC,MAAM,EAAE,IAAI,IAAA,mBAAO,EAAC,GAAG,CAAC,CAAC,WAAW,EAAE,KAAK,KAAK,EAAE,CAAC;QACzD,MAAM,IAAI,KAAK,CAAC,qBAAqB,IAAI,CAAC,IAAI,8FAA8F,CAAC,CAAC;IAClJ,CAAC;IACD,MAAM,OAAO,GAAG,MAAM,IAAA,mBAAQ,EAAC,GAAG,EAAE,OAAO,CAAC,CAAC;IAC7C,MAAM,IAAI,GAAG,IAAI,CAAC,IAAI,IAAI,kBAAkB,CAAC,GAAG,CAAC,CAAC;IAClD,MAAM,OAAO,CAAC,SAAS,CAAC,iBAAK,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,EAAE,IAAI,EAAE,UAAU,CAAC,EAAE,OAAO,CAAC,CAAC;IACzE,OAAO,EAAE,IAAI,EAAE,OAAO,EAAE,IAAI,EAAE,IAAI,EAAE,IAAI,CAAC,IAAI,EAAE,MAAM,EAAE,SAAS,CAAC,CAAC,EAAE,IAAI,EAAE,UAAU,EAAE,OAAO,EAAE,CAAC,CAAC,EAAE,CAAC;AACxG,CAAC;AACD,qGAAqG;AACrG,SAAgB,kBAAkB,CAAC,OAAe;IAC9C,MAAM,IAAI,GAAG,IAAA,oBAAQ,EAAC,OAAO,CAAC,CAAC;IAC/B,OAAO,IAAI,CAAC,WAAW,EAAE,KAAK,UAAU,CAAC,CAAC,CAAC,IAAA,oBAAQ,EAAC,IAAA,mBAAO,EAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,IAAA,oBAAQ,EAAC,IAAI,EAAE,IAAA,mBAAO,EAAC,IAAI,CAAC,CAAC,CAAC;AAC1G,CAAC;AACD,2DAA2D;AAC3D,SAAS,SAAS,CAAC,KAGhB;IACC,MAAM,CAAC,GAAG,IAAA,wBAAU,EAAC,QAAQ,CAAC,CAAC;IAC/B,KAAK,MAAM,CAAC,IAAI,CAAC,GAAG,KAAK,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,aAAa,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,EAAE,CAAC;QACtE,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC;IACjE,CAAC;IACD,OAAO,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC;AAC3B,CAAC;AACD,iDAAiD;AACjD,KAAK,UAAU,YAAY,CAAC,GAAW;IAInC,MAAM,GAAG,GAGH,EAAE,CAAC;IACT,MAAM,IAAI,GAAG,KAAK,EAAE,OAAe,EAAiB,EAAE;QAClD,KAAK,MAAM,KAAK,IAAI,MAAM,IAAA,kBAAO,EAAC,OAAO,EAAE,EAAE,aAAa,EAAE,IAAI,EAAE,CAAC,EAAE,CAAC;YAClE,MAAM,IAAI,GAAG,IAAA,gBAAI,EAAC,OAAO,EAAE,KAAK,CAAC,IAAI,CAAC,CAAC;YACvC,IAAI,KAAK,CAAC,WAAW,EAAE,EAAE,CAAC;gBACtB,IAAI,KAAK,CAAC,IAAI,KAAK,MAAM;oBACrB,SAAS;gBACb,MAAM,IAAI,CAAC,IAAI,CAAC,CAAC;YACrB,CAAC;iBACI,IAAI,KAAK,CAAC,MAAM,EAAE,EAAE,CAAC;gBACtB,GAAG,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,IAAA,oBAAQ,EAAC,GAAG,EAAE,IAAI,CAAC,CAAC,KAAK,CAAC,eAAG,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,EAAE,OAAO,EAAE,MAAM,IAAA,mBAAQ,EAAC,IAAI,CAAC,EAAE,CAAC,CAAC;YAChG,CAAC;QACL,CAAC;IACL,CAAC,CAAC;IACF,MAAM,IAAI,CAAC,GAAG,CAAC,CAAC;IAChB,OAAO,GAAG,CAAC;AACf,CAAC;AACD,iEAAiE;AACjE,KAAK,UAAU,gBAAgB,CAAC,OAAgB,EAAE,IAEhD,EAAE,IAA0B;IAC1B,MAAM,QAAQ,GAAG,MAAM,SAAS,CAAC,OAAO,EAAE,IAAI,CAAC,MAAM,EAAE,IAAI,CAAC,GAAG,CAAC,CAAC;IACjE,IAAI,CAAC;QACD,MAAM,SAAS,GAAG,MAAM,cAAc,CAAC,OAAO,EAAE,QAAQ,EAAE,IAAI,CAAC,MAAM,CAAC,CAAC;QACvE,MAAM,QAAQ,GAAG,gBAAgB,CAAC,SAAS,EAAE,IAAI,CAAC,CAAC;QACnD,KAAK,MAAM,IAAI,IAAI,QAAQ,EAAE,CAAC;YAC1B,MAAM,GAAG,GAAG,SAAS,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,IAAI,CAAE,CAAC,GAAG,CAAC;YACxD,MAAM,IAAI,GAAG,iBAAK,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,EAAE,IAAI,CAAC,CAAC;YACxC,MAAM,GAAG,GAAG,MAAM,OAAO,CAAC,QAAQ,CAAC,YAAY,IAAA,qBAAC,EAAC,IAAI,CAAC,GAAG,CAAC,cAAc,IAAA,qBAAC,EAAC,IAAI,CAAC,aAAa,IAAA,qBAAC,EAAC,GAAG,CAAC,IAAI,IAAA,qBAAC,EAAC,IAAI,CAAC,EAAE,CAAC,CAAC;YACjH,IAAI,GAAG,CAAC,QAAQ,KAAK,CAAC,EAAE,CAAC;gBACrB,MAAM,IAAI,KAAK,CAAC,4BAA4B,IAAI,UAAU,IAAI;EAC5E,IAAI,CAAC,GAAG,CAAC,MAAM,GAAG,GAAG,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC;YACvB,CAAC;QACL,CAAC;QACD,OAAO;YACH,IAAI,EAAE,MAAM;YACZ,MAAM,EAAE,IAAI,CAAC,MAAM;YACnB,GAAG,CAAC,IAAI,CAAC,GAAG,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,GAAG,EAAE,IAAI,CAAC,GAAG,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;YACpD,MAAM,EAAE,QAAQ;SACnB,CAAC;IACN,CAAC;YACO,CAAC;QACL,MAAM,OAAO,CAAC,QAAQ,CAAC,UAAU,IAAA,qBAAC,EAAC,QAAQ,CAAC,EAAE,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,CAAC;IAC3E,CAAC;AACL,CAAC;AAMD;;;GAGG;AACH,KAAK,UAAU,cAAc,CAAC,OAAgB,EAAE,QAAgB,EAAE,MAAc;IAC5E,MAAM,GAAG,GAAG,MAAM,OAAO,CAAC,QAAQ,CAAC,QAAQ,IAAA,qBAAC,EAAC,QAAQ,CAAC,uCAAuC,CAAC,CAAC;IAC/F,MAAM,KAAK,GAAG,GAAG,CAAC,MAAM,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,IAAI,EAAE,CAAC;IACjF,IAAI,KAAK,CAAC,MAAM,KAAK,CAAC;QAClB,MAAM,IAAI,KAAK,CAAC,cAAc,MAAM,wBAAwB,CAAC,CAAC;IAClE,MAAM,QAAQ,GAAG,GAAG,QAAQ,WAAW,CAAC;IACxC,IAAI,KAAK,CAAC,QAAQ,CAAC,QAAQ,CAAC;QACxB,OAAO,CAAC,EAAE,IAAI,EAAE,QAAQ,CAAC,MAAM,CAAC,EAAE,GAAG,EAAE,QAAQ,EAAE,CAAC,CAAC;IACvD,OAAO,KAAK;SACP,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE;QACX,MAAM,GAAG,GAAG,iBAAK,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC;QAC7B,OAAO,EAAE,IAAI,EAAE,iBAAK,CAAC,QAAQ,CAAC,GAAG,CAAC,EAAE,GAAG,EAAE,CAAC;IAC9C,CAAC,CAAC;SACG,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,aAAa,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,sBAAsB;AAC7E,CAAC;AACD;;;GAGG;AACH,SAAgB,gBAAgB,CAAC,SAAqC,EAAE,IAEtE;IACE,MAAM,KAAK,GAAG,SAAS,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC;IAC3C,IAAI,CAAC,IAAI,CAAC,MAAM,EAAE,MAAM,EAAE,CAAC;QACvB,IAAI,SAAS,CAAC,MAAM,KAAK,CAAC;YACtB,OAAO,CAAC,KAAK,CAAC,CAAC,CAAE,CAAC,CAAC;QACvB,MAAM,IAAI,KAAK,CAAC,cAAc,IAAI,CAAC,MAAM,6FAA6F,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;IAC/J,CAAC;IACD,KAAK,MAAM,IAAI,IAAI,IAAI,CAAC,MAAM,EAAE,CAAC;QAC7B,IAAI,CAAC,KAAK,CAAC,QAAQ,CAAC,IAAI,CAAC,EAAE,CAAC;YACxB,MAAM,IAAI,KAAK,CAAC,cAAc,IAAI,CAAC,MAAM,UAAU,IAAI,CAAC,GAAG,IAAI,WAAW,yBAAyB,IAAI,iBAAiB,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;QACjJ,CAAC;IACL,CAAC;IACD,OAAO,CAAC,GAAG,IAAI,CAAC,MAAM,CAAC,CAAC;AAC5B,CAAC;AACD,0DAA0D;AAC1D;;;;GAIG;AACI,KAAK,UAAU,SAAS,CAAC,OAAgB,EAAE,MAAc,EAAE,GAAY;IAC1E,MAAM,GAAG,GAAG,GAAG,UAAU,IAAI,IAAI,CAAC,MAAM,CAAC,IAAI,IAAI,CAAC,MAAM,EAAE,CAAC,QAAQ,CAAC,EAAE,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,EAAE,CAAC;IACtF,MAAM,GAAG,GAAG,QAAQ,CAAC,MAAM,CAAC,CAAC;IAC7B,MAAM,MAAM,GAAG,UAAU,IAAA,qBAAC,EAAC,GAAG,CAAC,gBAAgB,IAAA,qBAAC,EAAC,UAAU,CAAC,MAAM;QAC9D,qBAAqB,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,YAAY,GAAG,IAAA,qBAAC,EAAC,GAAG,CAAC,IAAI,IAAA,qBAAC,EAAC,GAAG,CAAC,EAAE;QACjE,CAAC,GAAG,CAAC,CAAC,CAAC,cAAc,IAAA,qBAAC,EAAC,GAAG,CAAC,qBAAqB,IAAA,qBAAC,EAAC,GAAG,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;IACnE,MAAM,GAAG,GAAG,MAAM,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAC,CAAC;IAC3C,IAAI,GAAG,CAAC,QAAQ,KAAK,CAAC,EAAE,CAAC;QACrB,MAAM,IAAI,KAAK,CAAC,8BAA8B,MAAM,UAAU,GAAG,IAAI,WAAW;EACtF,IAAI,CAAC,GAAG,CAAC,MAAM,GAAG,GAAG,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC;IAC/B,CAAC;IACD,OAAO,GAAG,CAAC;AACf,CAAC;AACD,qEAAqE;AACrE,SAAgB,QAAQ,CAAC,MAAc;IACnC,IAAI,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,MAAM,CAAC,UAAU,CAAC,MAAM,CAAC;QACnD,OAAO,MAAM,CAAC;IAClB,OAAO,sBAAsB,MAAM,MAAM,CAAC;AAC9C,CAAC;AACD,SAAS,QAAQ,CAAC,MAAc;IAC5B,MAAM,IAAI,GAAG,QAAQ,CAAC,MAAM,CAAC,CAAC,OAAO,CAAC,MAAM,EAAE,EAAE,CAAC,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,GAAG,EAAE,IAAI,MAAM,CAAC;IAC7E,OAAO,IAAI,CAAC,OAAO,CAAC,QAAQ,EAAE,EAAE,CAAC,CAAC;AACtC,CAAC;AACD,SAAS,IAAI,CAAC,MAAc;IACxB,OAAO,MAAM,CAAC,OAAO,CAAC,UAAU,EAAE,GAAG,CAAC,CAAC;AAC3C,CAAC;AACD,SAAS,IAAI,CAAC,MAAc,EAAE,CAAC,GAAG,EAAE;IAChC,OAAO,MAAM,CAAC,IAAI,EAAE,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AAC1D,CAAC;AACD,2DAA2D;AAC3D,mDAAmD;AACnD,SAAgB,mBAAmB,CAAC,MAAyC;IACzE,MAAM,KAAK,GAAa,EAAE,CAAC;IAC3B,KAAK,MAAM,KAAK,IAAI,MAAM,EAAE,CAAC;QACzB,IAAI,KAAK,CAAC,IAAI,KAAK,OAAO;YACtB,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;;YAEvB,KAAK,CAAC,IAAI,CAAC,GAAG,KAAK,CAAC,MAAM,CAAC,CAAC;IACpC,CAAC;IACD,OAAO,CAAC,GAAG,IAAI,GAAG,CAAC,KAAK,CAAC,CAAC,CAAC;AAC/B,CAAC;AACD;;;;GAIG;AACH,SAAgB,yBAAyB,CAAC,GAAW,EAAE,MAAyB;IAC5E,MAAM,IAAI,GAAG,MAAM,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,KAAK,IAAI,KAAK,iBAAK,CAAC,IAAI,CAAC,GAAG,EAAE,IAAI,CAAC,WAAW,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;IAC7F,OAAO;QACH,EAAE;QACF,WAAW;QACX,EAAE;QACF,kCAAkC,GAAG,MAAM;QAC3C,EAAE;QACF,IAAI;QACJ,EAAE;QACF,8FAA8F;QAC9F,+EAA+E;QAC/E,EAAE;KACL,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AACjB,CAAC;AACD;;;;;GAKG;AACI,KAAK,UAAU,wBAAwB,CAAC,OAA2D,EAAE,IAAY,EAAE,IAAI,GAAG,WAAW;IACxI,MAAM,OAAO,GAAG,MAAM,OAAO,CAAC,UAAU,CAAC,IAAI,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,KAAK,CAAC,CAAC;IAClE,MAAM,KAAK,GAAG,gBAAgB,IAAI,CAAC,MAAM,EAAE,CAAC,QAAQ,CAAC,EAAE,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,EAAE,CAAC;IACvE,MAAM,OAAO,CAAC,QAAQ,CAAC,UAAU,IAAA,qBAAC,EAAC,IAAI,CAAC,OAAO,KAAK,MAAM,IAAI,KAAK,KAAK,IAAI,CAAC,CAAC;IAC9E,IAAI,CAAC,OAAO;QACR,MAAM,eAAe,CAAC,OAAO,EAAE,CAAC,IAAI,CAAC,CAAC,CAAC;AAC/C,CAAC;AACD;;;;;;GAMG;AACI,KAAK,UAAU,eAAe,CAAC,OAA4C,EAAE,KAAwB;IACxG,IAAI,CAAC,KAAK,CAAC,MAAM;QACb,OAAO;IACX,MAAM,KAAK,GAAG,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,kBAAkB,IAAA,qBAAC,EAAC,CAAC,CAAC,uBAAuB,CAAC,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;IAC3F,MAAM,OAAO,CAAC,QAAQ,CAAC,yCAAyC,KAAK,EAAE,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,SAAS,CAAC,CAAC;AACpG,CAAC","sourcesContent":["// SkillSpec 的安装实现 —— 三个内置沙箱型 adapter(claude-code / codex / bub)共用的那一半:\n// 「从哪里取得哪份 Skill」(本地目录/文件、repo + ref + 选择集)。另一半——装到哪个目录、\n// 要不要额外写发现指引——留在各 adapter 里,由它们传 `dir` 并自己写 project instruction。\n// 定稿见 docs/feature/adapters/architecture/coding-agent-extensions.md。\n//\n// 为什么不调 `npx skills add`:那个 installer 没有钉 ref 的入口,也没有机器可读的「这个 repo\n// 里有哪些 skill」输出(-l 只打人看的带 ANSI 的清单),而契约要求「来源必须可复现」(钉 ref)\n// 与「多 Skill Repo 必须明确选择启用集合」(选不中要报出可选集)。所以 repo skill 走 git:\n// clone → 枚举 SKILL.md → 按规则选 → 拷进 agent 的 skill 目录。顺带绕开了 memory 里\n// npx-skills-add-headless-hang 那个交互式选择框(无 tty 卡死)的整类问题。\nimport { createHash } from \"node:crypto\";\nimport { readFile, readdir, stat } from \"node:fs/promises\";\nimport { basename, dirname, extname, join, posix, relative, resolve, sep } from \"node:path\";\nimport { shellQuote as q } from \"../sandbox/shell.ts\";\nimport type { SandboxOperations } from \"../sandbox/types.ts\";\nimport type { AgentInstallationSkill, Sandbox, SkillSpec } from \"../types.ts\";\n/** 沙箱里放临时 clone 的地方(装完即删;不在 workdir 下,不进 diff)。 */\nconst CLONE_ROOT = \"/tmp/niceeval-git\";\nexport interface InstallSkillsOptions {\n /** 沙箱里的 Skill 安装根目录(相对 workdir):claude-code 用 `.claude/skills`,codex/bub 用 `.agents/skills`。 */\n dir: string;\n /** 本地 Skill 的解析根;省略用跑 niceeval 的项目根(process.cwd())。 */\n projectRoot?: string;\n}\n/**\n * 按配置顺序把 SkillSpec 装进沙箱,返回 manifest 里的 Skill 记录(与配置顺序一一对应)。\n * 同名 Skill 来自多个来源时后装的覆盖磁盘、manifest 保留每一条(不静默合并)。\n * 任何一条装不上(路径不存在 / 形状不支持 / clone 或 ref 解析失败 / 选择集不合法)直接抛错 ——\n * setup 失败即 attempt errored,不伪装成断言失败。\n */\nexport async function installSkills(sandbox: Sandbox, specs: readonly SkillSpec[], opts: InstallSkillsOptions): Promise<AgentInstallationSkill[]> {\n const out: AgentInstallationSkill[] = [];\n for (const spec of specs) {\n out.push(spec.kind === \"local\"\n ? await installLocalSkill(sandbox, spec, opts)\n : await installRepoSkill(sandbox, spec, opts));\n }\n if (out.length)\n await excludeFromDiff(sandbox, [`${opts.dir}/`]);\n return out;\n}\n// ───────────────────────── 本地 Skill ─────────────────────────\nasync function installLocalSkill(sandbox: Sandbox, spec: Extract<SkillSpec, {\n kind: \"local\";\n}>, opts: InstallSkillsOptions): Promise<AgentInstallationSkill> {\n const root = opts.projectRoot ?? process.cwd();\n const abs = resolve(root, spec.path);\n const info = await stat(abs).catch(() => undefined);\n if (!info)\n throw new Error(`Local skill path \"${spec.path}\" does not exist (resolved to ${abs}). Paths are resolved from the project root you run niceeval in.`);\n if (info.isDirectory()) {\n const files = await readDirFiles(abs);\n if (!files.some((f) => f.path === \"SKILL.md\")) {\n throw new Error(`Local skill directory \"${spec.path}\" has no SKILL.md. A directory-shaped skill must contain SKILL.md at its root.`);\n }\n const name = spec.name ?? basename(abs);\n const sha256 = hashFiles(files);\n await sandbox.uploadDirectory(abs, posix.join(opts.dir, name));\n return { kind: \"local\", name, path: spec.path, sha256 };\n }\n if (!info.isFile() || extname(abs).toLowerCase() !== \".md\") {\n throw new Error(`Local skill path \"${spec.path}\" has an unsupported shape. Accepted: a directory containing SKILL.md, or a single .md file.`);\n }\n const content = await readFile(abs, \"utf-8\");\n const name = spec.name ?? localFileSkillName(abs);\n await sandbox.writeText(posix.join(opts.dir, name, \"SKILL.md\"), content);\n return { kind: \"local\", name, path: spec.path, sha256: hashFiles([{ path: \"SKILL.md\", content }]) };\n}\n/** `skills/effect-ts/SKILL.md` → `effect-ts`(SKILL.md 的名字不含信息,取所在目录);`skills/guide.md` → `guide`。 */\nexport function localFileSkillName(absPath: string): string {\n const file = basename(absPath);\n return file.toLowerCase() === \"skill.md\" ? basename(dirname(absPath)) : basename(file, extname(file));\n}\n/** 目录内容的确定性哈希:按相对路径排序,逐条喂 `path\\0content\\0`(二进制附件按原字节)。 */\nfunction hashFiles(files: {\n path: string;\n content: string | Buffer;\n}[]): string {\n const h = createHash(\"sha256\");\n for (const f of [...files].sort((a, b) => a.path.localeCompare(b.path))) {\n h.update(f.path).update(\"\\0\").update(f.content).update(\"\\0\");\n }\n return h.digest(\"hex\");\n}\n/** Skill 目录可以带脚本 / 图片等附件,一律按原字节读 + 上传,不做文本假设。 */\nasync function readDirFiles(dir: string): Promise<{\n path: string;\n content: Buffer;\n}[]> {\n const out: {\n path: string;\n content: Buffer;\n }[] = [];\n const walk = async (current: string): Promise<void> => {\n for (const entry of await readdir(current, { withFileTypes: true })) {\n const full = join(current, entry.name);\n if (entry.isDirectory()) {\n if (entry.name === \".git\")\n continue;\n await walk(full);\n }\n else if (entry.isFile()) {\n out.push({ path: relative(dir, full).split(sep).join(\"/\"), content: await readFile(full) });\n }\n }\n };\n await walk(dir);\n return out;\n}\n// ───────────────────────── Repo Skill ─────────────────────────\nasync function installRepoSkill(sandbox: Sandbox, spec: Extract<SkillSpec, {\n kind: \"repo\";\n}>, opts: InstallSkillsOptions): Promise<AgentInstallationSkill> {\n const cloneDir = await cloneRepo(sandbox, spec.source, spec.ref);\n try {\n const available = await discoverSkills(sandbox, cloneDir, spec.source);\n const selected = selectRepoSkills(available, spec);\n for (const name of selected) {\n const src = available.find((s) => s.name === name)!.dir;\n const dest = posix.join(opts.dir, name);\n const res = await sandbox.runShell(`mkdir -p ${q(opts.dir)} && rm -rf ${q(dest)} && cp -R ${q(src)} ${q(dest)}`);\n if (res.exitCode !== 0) {\n throw new Error(`Could not install skill \"${name}\" into ${dest}:\n${tail(res.stdout + res.stderr)}`);\n }\n }\n return {\n kind: \"repo\",\n source: spec.source,\n ...(spec.ref !== undefined ? { ref: spec.ref } : {}),\n skills: selected,\n };\n }\n finally {\n await sandbox.runShell(`rm -rf ${q(cloneDir)}`).catch(() => undefined);\n }\n}\n/** repo 里可用的一个 Skill:名字 + 沙箱内 SKILL.md 所在目录。 */\nexport interface DiscoveredSkill {\n name: string;\n dir: string;\n}\n/**\n * 枚举 clone 里的 Skill:根 `SKILL.md` 存在 → 整个 repo 就是一个 Skill(名字取 repo 名);\n * 否则每个含 `SKILL.md` 的目录是一个 Skill(名字取该目录名),与 `skills` CLI 的默认口径一致。\n */\nasync function discoverSkills(sandbox: Sandbox, cloneDir: string, source: string): Promise<DiscoveredSkill[]> {\n const res = await sandbox.runShell(`find ${q(cloneDir)} -name SKILL.md -not -path '*/.git/*'`);\n const paths = res.stdout.split(\"\\n\").map((l) => l.trim()).filter(Boolean).sort();\n if (paths.length === 0)\n throw new Error(`Repo skill ${source} contains no SKILL.md.`);\n const rootFile = `${cloneDir}/SKILL.md`;\n if (paths.includes(rootFile))\n return [{ name: repoName(source), dir: cloneDir }];\n return paths\n .map((p) => {\n const dir = posix.dirname(p);\n return { name: posix.basename(dir), dir };\n })\n .sort((a, b) => a.name.localeCompare(b.name)); // 报错里的「可选集」按名字排,不按路径排\n}\n/**\n * 选择规则(docs 的表):repo 只有一个 Skill 且省略 `skills` → 装唯一那个;多个且省略 →\n * 失败并列出可选集;指定了不存在的 Skill → 失败并报 source / ref / 名字。\n */\nexport function selectRepoSkills(available: readonly DiscoveredSkill[], spec: Extract<SkillSpec, {\n kind: \"repo\";\n}>): string[] {\n const names = available.map((s) => s.name);\n if (!spec.skills?.length) {\n if (available.length === 1)\n return [names[0]!];\n throw new Error(`Repo skill ${spec.source} contains multiple skills; select which ones to enable with \\`skills: [...]\\`. Available: ${names.join(\", \")}.`);\n }\n for (const want of spec.skills) {\n if (!names.includes(want)) {\n throw new Error(`Repo skill ${spec.source} (ref: ${spec.ref ?? \"(default)\"}) has no skill named \"${want}\". Available: ${names.join(\", \")}.`);\n }\n }\n return [...spec.skills];\n}\n// ───────────────────────── git ─────────────────────────\n/**\n * 把一个 repo clone 进沙箱临时目录,给了 ref 就钉到那个 tag/commit/branch,返回 clone 目录。\n * ref 可能是任意 commit,所以钉 ref 时不能 `--depth 1`(浅克隆 checkout 不到历史 commit)。\n * Skill 与 native plugin 的 marketplace(claude 侧 CLI 没有钉 ref 的入口)共用这一条。\n */\nexport async function cloneRepo(sandbox: Sandbox, source: string, ref?: string): Promise<string> {\n const dir = `${CLONE_ROOT}/${slug(source)}-${Math.random().toString(36).slice(2, 8)}`;\n const url = gitUrlOf(source);\n const script = `rm -rf ${q(dir)} && mkdir -p ${q(CLONE_ROOT)} && ` +\n `git clone --quiet ${ref ? \"\" : \"--depth 1 \"}${q(url)} ${q(dir)}` +\n (ref ? ` && git -C ${q(dir)} checkout --quiet ${q(ref)}` : \"\");\n const res = await sandbox.runShell(script);\n if (res.exitCode !== 0) {\n throw new Error(`Could not fetch repo skill ${source} (ref: ${ref ?? \"(default)\"}):\n${tail(res.stdout + res.stderr)}`);\n }\n return dir;\n}\n/** `owner/repo` → GitHub HTTPS URL;已经是 URL(含 `://` 或 `git@`)则原样用。 */\nexport function gitUrlOf(source: string): string {\n if (source.includes(\"://\") || source.startsWith(\"git@\"))\n return source;\n return `https://github.com/${source}.git`;\n}\nfunction repoName(source: string): string {\n const last = gitUrlOf(source).replace(/\\/+$/, \"\").split(\"/\").pop() ?? source;\n return last.replace(/\\.git$/, \"\");\n}\nfunction slug(source: string): string {\n return source.replace(/[^\\w.-]/g, \"-\");\n}\nfunction tail(output: string, n = 12): string {\n return output.trim().split(\"\\n\").slice(-n).join(\"\\n\");\n}\n// ───────────────────────── 发现指引 ─────────────────────────\n/** manifest 里的 Skill 记录 → 沙箱里实际的目录名(发现指引要逐条点名)。 */\nexport function installedSkillNames(skills: readonly AgentInstallationSkill[]): string[] {\n const names: string[] = [];\n for (const skill of skills) {\n if (skill.kind === \"local\")\n names.push(skill.name);\n else\n names.push(...skill.skills);\n }\n return [...new Set(names)];\n}\n/**\n * 给没有原生 Skill 加载机制的 agent(codex / bub)写的 project instruction 段落:\n * 只把文件装到 skill 目录不足以让它们读到(见 memory/codex-no-native-skill-tool.md ——\n * 不提示的话 codex 连一次 shell 读取都不会发生),必须在 AGENTS.md 里给出稳定的发现指引。\n */\nexport function skillDiscoveryInstruction(dir: string, skills: readonly string[]): string {\n const list = skills.map((name) => `- ${name}: ${posix.join(dir, name)}/SKILL.md`).join(\"\\n\");\n return [\n \"\",\n \"## Skills\",\n \"\",\n `Skill files are installed in \\`${dir}/\\`:`,\n \"\",\n list,\n \"\",\n \"Before answering a task, check whether one of these skills covers it and read its `SKILL.md`\",\n \"(e.g. with a shell command). Follow the skill's instructions when it applies.\",\n \"\",\n ].join(\"\\n\");\n}\n/**\n * 把发现指引追加进沙箱里的一个 instruction 文件(不存在则创建)。落在 workspace 里时,\n * 新建的文件顺手排除出 diff —— setup 装的东西不是 agent 的产出(见 excludeFromDiff)。\n * 文件本来就存在(starter 自带 AGENTS.md)时不排除:那是 agent 也可能改的文件,\n * 排掉会把真实改动一起藏起来;此时追加的这一段会如实出现在 diff 里。\n */\nexport async function appendProjectInstruction(sandbox: Pick<SandboxOperations, \"pathExists\" | \"runShell\">, text: string, file = \"AGENTS.md\"): Promise<void> {\n const existed = await sandbox.pathExists(file).catch(() => false);\n const delim = `NICEEVAL_EOF_${Math.random().toString(36).slice(2, 8)}`;\n await sandbox.runShell(`cat >> ${q(file)} <<'${delim}'\\n${text}\\n${delim}\\n`);\n if (!existed)\n await excludeFromDiff(sandbox, [file]);\n}\n/**\n * 把 setup 装进 workspace 的东西(skill 目录、adapter 新建的 instruction 文件)写进\n * `.git/info/exclude`:git 基线是在 agent.setup **之前**打的,不排除的话这些文件会被\n * `captureGeneratedFiles`(git add -A && git diff HEAD)当成「agent 生成的文件」记进\n * diff.json —— 装了 skill 的实验,diff 里凭空多出几十个文件,基于 diff 的断言与展示全被污染。\n * 排除只对未跟踪文件有效,这正是 adapter 自己创建的那一类。\n */\nexport async function excludeFromDiff(sandbox: Pick<SandboxOperations, \"runShell\">, paths: readonly string[]): Promise<void> {\n if (!paths.length)\n return;\n const lines = paths.map((p) => `printf '%s\\\\n' ${q(p)} >> .git/info/exclude`).join(\" && \");\n await sandbox.runShell(`test -d .git && mkdir -p .git/info && ${lines}`).catch(() => undefined);\n}\n"]}
@@ -1,5 +1,5 @@
1
1
  import type { SandboxOperations } from "../sandbox/types.cjs";
2
- import type { AgentSetupSkill, Sandbox, SkillSpec } from "../types.cjs";
2
+ import type { AgentInstallationSkill, Sandbox, SkillSpec } from "../types.cjs";
3
3
  export interface InstallSkillsOptions {
4
4
  /** 沙箱里的 Skill 安装根目录(相对 workdir):claude-code 用 `.claude/skills`,codex/bub 用 `.agents/skills`。 */
5
5
  dir: string;
@@ -12,7 +12,7 @@ export interface InstallSkillsOptions {
12
12
  * 任何一条装不上(路径不存在 / 形状不支持 / clone 或 ref 解析失败 / 选择集不合法)直接抛错 ——
13
13
  * setup 失败即 attempt errored,不伪装成断言失败。
14
14
  */
15
- export declare function installSkills(sandbox: Sandbox, specs: readonly SkillSpec[], opts: InstallSkillsOptions): Promise<AgentSetupSkill[]>;
15
+ export declare function installSkills(sandbox: Sandbox, specs: readonly SkillSpec[], opts: InstallSkillsOptions): Promise<AgentInstallationSkill[]>;
16
16
  /** `skills/effect-ts/SKILL.md` → `effect-ts`(SKILL.md 的名字不含信息,取所在目录);`skills/guide.md` → `guide`。 */
17
17
  export declare function localFileSkillName(absPath: string): string;
18
18
  /** repo 里可用的一个 Skill:名字 + 沙箱内 SKILL.md 所在目录。 */
@@ -36,7 +36,7 @@ export declare function cloneRepo(sandbox: Sandbox, source: string, ref?: string
36
36
  /** `owner/repo` → GitHub HTTPS URL;已经是 URL(含 `://` 或 `git@`)则原样用。 */
37
37
  export declare function gitUrlOf(source: string): string;
38
38
  /** manifest 里的 Skill 记录 → 沙箱里实际的目录名(发现指引要逐条点名)。 */
39
- export declare function installedSkillNames(skills: readonly AgentSetupSkill[]): string[];
39
+ export declare function installedSkillNames(skills: readonly AgentInstallationSkill[]): string[];
40
40
  /**
41
41
  * 给没有原生 Skill 加载机制的 agent(codex / bub)写的 project instruction 段落:
42
42
  * 只把文件装到 skill 目录不足以让它们读到(见 memory/codex-no-native-skill-tool.md ——
@@ -1,5 +1,5 @@
1
1
  import type { SandboxOperations } from "../sandbox/types.mjs";
2
- import type { AgentSetupSkill, Sandbox, SkillSpec } from "../types.mjs";
2
+ import type { AgentInstallationSkill, Sandbox, SkillSpec } from "../types.mjs";
3
3
  export interface InstallSkillsOptions {
4
4
  /** 沙箱里的 Skill 安装根目录(相对 workdir):claude-code 用 `.claude/skills`,codex/bub 用 `.agents/skills`。 */
5
5
  dir: string;
@@ -12,7 +12,7 @@ export interface InstallSkillsOptions {
12
12
  * 任何一条装不上(路径不存在 / 形状不支持 / clone 或 ref 解析失败 / 选择集不合法)直接抛错 ——
13
13
  * setup 失败即 attempt errored,不伪装成断言失败。
14
14
  */
15
- export declare function installSkills(sandbox: Sandbox, specs: readonly SkillSpec[], opts: InstallSkillsOptions): Promise<AgentSetupSkill[]>;
15
+ export declare function installSkills(sandbox: Sandbox, specs: readonly SkillSpec[], opts: InstallSkillsOptions): Promise<AgentInstallationSkill[]>;
16
16
  /** `skills/effect-ts/SKILL.md` → `effect-ts`(SKILL.md 的名字不含信息,取所在目录);`skills/guide.md` → `guide`。 */
17
17
  export declare function localFileSkillName(absPath: string): string;
18
18
  /** repo 里可用的一个 Skill:名字 + 沙箱内 SKILL.md 所在目录。 */
@@ -36,7 +36,7 @@ export declare function cloneRepo(sandbox: Sandbox, source: string, ref?: string
36
36
  /** `owner/repo` → GitHub HTTPS URL;已经是 URL(含 `://` 或 `git@`)则原样用。 */
37
37
  export declare function gitUrlOf(source: string): string;
38
38
  /** manifest 里的 Skill 记录 → 沙箱里实际的目录名(发现指引要逐条点名)。 */
39
- export declare function installedSkillNames(skills: readonly AgentSetupSkill[]): string[];
39
+ export declare function installedSkillNames(skills: readonly AgentInstallationSkill[]): string[];
40
40
  /**
41
41
  * 给没有原生 Skill 加载机制的 agent(codex / bub)写的 project instruction 段落:
42
42
  * 只把文件装到 skill 目录不足以让它们读到(见 memory/codex-no-native-skill-tool.md ——
@@ -1,5 +1,5 @@
1
1
  import type { SandboxOperations } from "../sandbox/types.js";
2
- import type { AgentSetupSkill, Sandbox, SkillSpec } from "../types.js";
2
+ import type { AgentInstallationSkill, Sandbox, SkillSpec } from "../types.js";
3
3
  export interface InstallSkillsOptions {
4
4
  /** 沙箱里的 Skill 安装根目录(相对 workdir):claude-code 用 `.claude/skills`,codex/bub 用 `.agents/skills`。 */
5
5
  dir: string;
@@ -12,7 +12,7 @@ export interface InstallSkillsOptions {
12
12
  * 任何一条装不上(路径不存在 / 形状不支持 / clone 或 ref 解析失败 / 选择集不合法)直接抛错 ——
13
13
  * setup 失败即 attempt errored,不伪装成断言失败。
14
14
  */
15
- export declare function installSkills(sandbox: Sandbox, specs: readonly SkillSpec[], opts: InstallSkillsOptions): Promise<AgentSetupSkill[]>;
15
+ export declare function installSkills(sandbox: Sandbox, specs: readonly SkillSpec[], opts: InstallSkillsOptions): Promise<AgentInstallationSkill[]>;
16
16
  /** `skills/effect-ts/SKILL.md` → `effect-ts`(SKILL.md 的名字不含信息,取所在目录);`skills/guide.md` → `guide`。 */
17
17
  export declare function localFileSkillName(absPath: string): string;
18
18
  /** repo 里可用的一个 Skill:名字 + 沙箱内 SKILL.md 所在目录。 */
@@ -36,7 +36,7 @@ export declare function cloneRepo(sandbox: Sandbox, source: string, ref?: string
36
36
  /** `owner/repo` → GitHub HTTPS URL;已经是 URL(含 `://` 或 `git@`)则原样用。 */
37
37
  export declare function gitUrlOf(source: string): string;
38
38
  /** manifest 里的 Skill 记录 → 沙箱里实际的目录名(发现指引要逐条点名)。 */
39
- export declare function installedSkillNames(skills: readonly AgentSetupSkill[]): string[];
39
+ export declare function installedSkillNames(skills: readonly AgentInstallationSkill[]): string[];
40
40
  /**
41
41
  * 给没有原生 Skill 加载机制的 agent(codex / bub)写的 project instruction 段落:
42
42
  * 只把文件装到 skill 目录不足以让它们读到(见 memory/codex-no-native-skill-tool.md ——
@@ -1 +1 @@
1
- {"version":3,"file":"types.js","sourceRoot":"","sources":["types.ts"],"names":[],"mappings":";;;AACA,iDAAiD;AACjD,uDAAgE;AAoIhE,6DAA6D;AAC7D,oCAAoC;AACvB,QAAA,0BAA0B,GAAG,CAAC,UAAU,EAAE,SAAS,EAAE,aAAa,CAAU,CAAC","sourcesContent":["import type { ExperimentFlags } from \"../shared/types.ts\";\n// agent 域类型:Agent / Adapter 契约、会话与 tracing 导出配置。\nimport { AGENT_DOCKERFILE_CACHE_SAFE } from \"./cache-marker.ts\";\n// 「连到哪个被测对象、协议怎么说」的全部契约在这里(见 apps/docs-site/zh/explanation/adapter.mdx)。\n// 能力不再是问卷式声明:t 上解锁什么完全由构造证据决定(见 docs-site 「能力从哪来」一节)。\nimport type { DiagnosticInput, JsonValue, ProgressUpdate } from \"../shared/types.ts\";\nimport type { Effect } from \"effect\";\nimport type { StreamEvent, TraceSpan, Usage } from \"../o11y/types.ts\";\nimport type { Sandbox } from \"../sandbox/types.ts\";\nimport type { SandboxLayer } from \"../sandbox/layer.ts\";\nimport type { AttemptRef, MaybePromise, SandboxCommandTarget, StableSandboxCommand } from \"../sandbox/commands.ts\";\nimport type { RegisteredSandboxContent } from \"../sandbox/content.ts\";\nimport type { SendFailureClassifier } from \"../context/send-failures.ts\";\nimport type { SessionSlot } from \"./session-slot.ts\";\nexport type { SessionSlot } from \"./session-slot.ts\";\n/**\n * 本地 stdio 形态的 MCP server:沙箱内起子进程,按 stdio 说 MCP 协议。\n * 与 {@link McpHttpServer} 按形状判别(有 `command` 的是 stdio,有 `url` 的是 HTTP)。\n */\nexport interface McpStdioServer {\n /** 服务器唯一名(config key)。 */\n name: string;\n /** 启动命令(如 \"npx\"、\"node\"、\"uvx\")。 */\n command: string;\n /** 传给命令的参数。 */\n args?: readonly string[];\n /** 注入服务器进程的环境变量(可能含 secret,不进 manifest)。 */\n env?: Readonly<globalThis.Record<string, string>>;\n /** stdio 与 HTTP 是互斥完成态,不能把另一分支的字段混进来。 */\n url?: never;\n headers?: never;\n}\n/**\n * 远程 Streamable HTTP 形态的 MCP server:沙箱直接连一个 HTTP 端点。\n * `url` 必须沙箱内可达——宿主机上的服务先经隧道(cloudflared / tailscale 等)暴露。\n */\nexport interface McpHttpServer {\n /** 服务器唯一名(config key)。 */\n name: string;\n /** Streamable HTTP 端点(如 https://mem.example.com/mcp/)。 */\n url: string;\n /** 逐字写进每个请求的 HTTP 头(常用于 Authorization;可能含 secret,不进 manifest)。 */\n headers?: Readonly<globalThis.Record<string, string>>;\n /** HTTP 与 stdio 是互斥完成态,不能把另一分支的字段混进来。 */\n command?: never;\n args?: never;\n env?: never;\n}\n/**\n * MCP server 描述符 —— 支持 MCP 的 adapter(Claude Code / Codex)共用的工具服务单元,\n * 不是 native plugin 的一种。stdio 与 Streamable HTTP 两种形态按形状判别,不设 kind 标签\n * (两种形态各有唯一必填判别字段);同时给出 `command` 与 `url` 属配置错误,setup 报错点名。\n * 在 agent factory config 里声明,setup 阶段写进各自的配置文件。\n * 见 docs/feature/adapters/architecture/coding-agent-extensions.md「类型边界」。\n */\nexport type McpServer = McpStdioServer | McpHttpServer;\n/**\n * Skill 的来源描述 —— Claude Code / Codex / Bub 共用的**数据类型**:只统一「从哪里取得\n * 哪份 Skill」,安装位置、发现机制、要不要额外写 project instruction 由各 Adapter 决定。\n * 见 docs/feature/adapters/architecture/coding-agent-extensions.md「类型边界」。\n */\nexport type SkillSpec = {\n kind: \"local\";\n /** 相对项目根(跑 niceeval 的目录)的 Skill 文件或目录:`SKILL.md`、含 `SKILL.md` 的目录,或单个 `.md`。 */\n path: string;\n /** 展示名;省略时由文件或目录名推导(`<dir>/SKILL.md` → `<dir>`,`foo.md` → `foo`)。 */\n name?: string;\n} | {\n kind: \"repo\";\n /** GitHub `owner/repo` 或 Git URL。 */\n source: string;\n /** 多 Skill Repo 中要启用的 Skill;repo 只有一个 Skill 时可省略,多个时省略即 setup 失败。 */\n skills?: readonly string[];\n /** Tag、Commit 或 Branch;省略表示 repo 默认 ref。 */\n ref?: string;\n};\n/** Manifest 里的一条 Skill 安装记录:本地(带内容哈希)或 repo(带来源与 ref)。 */\nexport type AgentSetupSkill = {\n kind: \"local\";\n name: string;\n path: string;\n sha256: string;\n} | {\n kind: \"repo\";\n source: string;\n ref?: string;\n skills: string[];\n};\n/**\n * 一次 Agent setup 实际装了什么 —— 沙箱型 Coding Agent Adapter 在 setup 收尾经\n * `ctx.reportSetup()` 交回的安装清单。它是宿主侧内存对象,运行器把它作为 attempt artifact 存成\n * `agent-setup.json`(见 docs/feature/record/architecture.md);沙箱磁盘上不落它。不参与评分,\n * 只回答「这次实际装了什么」;**环境变量值与 secret 不写进来**(所以 mcpServers 只记\n * name/command/args)。\n */\nexport interface AgentSetupManifest {\n /** 装进去的 Skill(按配置顺序;同名来自多个来源时逐条保留,不静默合并)。 */\n skills: AgentSetupSkill[];\n /** Agent 原生 Plugin(Claude Code / Codex 各自的 Marketplace 协议)。 */\n nativePlugins?: Array<{\n agent: \"claude-code\" | \"codex\";\n marketplace: {\n name: string;\n source: string;\n ref?: string;\n };\n name: string;\n /** 安装后 CLI 报告的版本;取不到时省略。 */\n resolvedVersion?: string;\n }>;\n /** 挂上的 MCP server(只记非 secret 字段:stdio 不含 env,HTTP 不含 headers)。 */\n mcpServers?: Array<{\n name: string;\n command: string;\n args?: string[];\n } | {\n name: string;\n url: string;\n }>;\n /**\n * 官方原生配置文件(Claude Code `settings.json` / Codex `config.toml`):只记 Agent 名、\n * 项目相对来源路径与原始字节的 SHA-256,不落正文 —— 任意官方配置都可能携带敏感字符串,\n * 不能靠字段白名单证明适合原样落盘。\n */\n nativeConfigFile?: {\n agent: \"claude-code\" | \"codex\";\n path: string;\n sha256: string;\n };\n /** Bub 的 Python Plugin(规范化后的 package 串)。 */\n pythonPlugins?: Array<{\n package: string;\n }>;\n}\n// ───────────────────────── 证据覆盖声明 ─────────────────────────\n/** 证据通道的覆盖状态;不存在隐含的 unknown 第四态。 */\nexport const EVIDENCE_COVERAGE_STATUSES = [\"complete\", \"partial\", \"unavailable\"] as const;\nexport type EvidenceCoverageStatus = (typeof EVIDENCE_COVERAGE_STATUSES)[number];\n/** 非完整通道必须解释缺口;完整通道不能携带一个自相矛盾的 reason。 */\nexport type EvidenceCoverageEntry = {\n readonly status: \"complete\";\n readonly reason?: never;\n} | {\n readonly status: Exclude<EvidenceCoverageStatus, \"complete\">;\n readonly reason: string;\n};\n/**\n * 覆盖声明(EvidenceCoverage):完整性不是口头承诺,是随数据走的声明\n * (见 docs/feature/adapters/architecture/evidence.md)。两层:\n * - Agent 级默认(`defineAgent` / `defineSandboxAgent` 的 `evidenceCoverage`)必须逐一声明\n * 六个通道;官方 SDK 适配器可使用 `completeEvidenceCoverage`。\n * - Turn 级降级(`Turn.evidenceCoverage`)只用于相对 Agent 默认值降级。\n */\nexport interface EvidenceCoverage {\n /** 完整事件流(event / notEvent / order 的依据)。 */\n readonly events: EvidenceCoverageEntry;\n /** action 生命周期(工具正负断言、顺序、失败的依据)。 */\n readonly actions: EvidenceCoverageEntry;\n /** assistant / user message(reply 与显式 message 值断言的依据)。 */\n readonly messages: EvidenceCoverageEntry;\n /** usage(token / cost 上限断言的依据)。 */\n readonly usage: EvidenceCoverageEntry;\n /** Turn status 的真实性(succeeded / parked 的依据)——恒 completed 的映射必须声明非 complete。 */\n readonly status: EvidenceCoverageEntry;\n /** Turn.data(供 `t.check(turn.data, match)` 检查结构化输出)。 */\n readonly data: EvidenceCoverageEntry;\n}\n/** 单轮只列相对 Agent 默认值的降级通道;省略的通道沿用 Agent 声明。 */\nexport type TurnEvidenceCoverage = Partial<EvidenceCoverage>;\n/** 随一轮消息附带的文件(图片等多模态输入)。 */\nexport interface InputFile {\n /** 文件名(可选,供 adapter / 模型参考)。 */\n readonly filename?: string;\n /** MIME 类型,如 `image/png`、`image/jpeg`。 */\n readonly mimeType: string;\n /** base64 编码的文件内容(JSON 友好,remote adapter 可直接放进请求体)。 */\n readonly dataBase64: string;\n}\n/**\n * HITL 回答轮里,人的裁决以结构化形式随 `input.responses` 到达——adapter 不需要解析\n * `text` 去猜哪句回答对应哪个请求、算不算批准。见 apps/docs-site/zh/explanation/adapter.mdx\n * 「不同回答的入参」一节的四种典型形态。\n */\nexport type AnswerValue = {\n readonly optionId: string;\n readonly text?: never;\n} | {\n readonly text: string;\n readonly optionId?: never;\n};\nexport type InputResponse = {\n /** 对应哪条 input.requested 请求;多个请求并停时靠它对位。 */\n readonly requestId: string;\n} & AnswerValue;\nexport interface TurnInput {\n readonly text: string;\n /** 本轮附带的文件(图片等)。adapter 自行决定怎么投递;不支持多模态的 adapter 忽略它。 */\n readonly files?: readonly InputFile[];\n /** 仅回答轮(t.respond / t.respondAll):逐请求的结构化回答,按 requestId 对位。 */\n readonly responses?: readonly InputResponse[];\n}\n/** adapter 的 send 返回值(事件流为核心)。 */\nexport interface Turn {\n readonly events: StreamEvent[];\n readonly data?: JsonValue;\n readonly status: \"completed\" | \"failed\" | \"waiting\";\n readonly usage?: Usage;\n /**\n * 本轮相对 Agent 默认覆盖的降级声明(这一轮流断了、拿不到 usage 等);\n * 只能降级,不能把 Agent 未声明的通道升格成 complete(见 EvidenceCoverage)。\n */\n readonly evidenceCoverage?: TurnEvidenceCoverage;\n}\n/**\n * 本次运行的 OTLP traces 接收信息(仅当配置了 OTel 接入时有——agent 的 `tracing` 块 /\n * config 的 telemetry 存在)。经 ctx.telemetry 交给 agent。span 只进瀑布图,不喂断言。\n * 远程 HTTP 接入的 send 只需要 spread `headers`(每轮一个新 traceparent);接收端点是\n * 启动期配置(`defineConfig({ telemetry: { port } } )`)固定的,不从这里传。\n */\nexport interface Telemetry {\n /** 接收端点(完整路径,形如 http://127.0.0.1:PORT/v1/traces)。 */\n readonly endpoint: string;\n /**\n * env-based 导出的 env(= AgentTracing.env(endpoint) 的结果),ready-to-spread。\n * adapter 的 send 直接 `{ ...ctx.telemetry?.env }` 注入,不必手搓 OTEL_* 拼装。\n */\n readonly env?: Readonly<globalThis.Record<string, string>>;\n /**\n * 本轮的 W3C trace context(traceparent),每轮一个新值。send 把它 spread 进 HTTP 请求头\n * (或注入子进程 env)——应用埋点支持 context 传播时,本轮 span 挂到我们给的 trace 下,\n * 并发跑 eval 的 span 归属才精确;不带则回退时间窗口归属(该 agent 自动降为串行)。\n */\n readonly headers?: Readonly<globalThis.Record<string, string>>;\n}\n/**\n * agent 的 OTLP 导出配置 —— 「沙箱里怎么让这个 CLI 把 trace 发到 endpoint」。\n * 刻意从 setup 里拆出来:setup 管运行配置,这里只管 otel 导出。CLI 安装归 `agent.ensure`。两种投递方式\n * (互不排斥,按 CLI 而定):\n * · env-based(标准 OTEL_* 环境变量,如 bub/Python OTel SDK)—— 用 `env`;\n * · file-based(CLI 自有配置文件,如 codex 的 config.toml [otel] 块)—— 用 `configure`。\n */\nexport interface AgentTracing {\n /**\n * 线协议(codex 发 OTLP/JSON、bub 发 OTLP/protobuf)。接收器按 content-type 自动解码,\n * 此字段仅作声明/日志用,也为将来按协议分流留口。\n */\n protocol?: \"http/json\" | \"http/protobuf\";\n /**\n * 接收器粒度(仅非沙箱 agent;沙箱型每沙箱一个,与此无关):\n * · \"attempt\"(默认)—— 每个 attempt 一个接收器。适合每轮能切换导出端点的被测对象\n * (进程内 adapter,如内置 aiSdkAgent 的可切换 exporter),attempt 间天然隔离、可全并发。\n * · \"run\" —— 整个 run 共享一个接收器(每 agent 一个)。**长驻服务必选**:应用的 OTEL_*\n * env 进程启动时读一次,per-attempt 端口会在第一个 attempt 结束时失效。span 逐轮归属\n * (traceparent / 时间窗口 + 串行守卫),见 ctx.telemetry.headers。\n * config 配了 `telemetry`(固定端口,无侵入接入的长驻服务场景)的非沙箱 agent 自动按 \"run\" 处理。\n */\n scope?: \"attempt\" | \"run\";\n /**\n * env-based 导出:给 endpoint → 返回要注入每轮 send 的 env(纯函数)。运行器把结果\n * 放进 ctx.telemetry.env,send 直接 spread。\n */\n env?(endpoint: string): globalThis.Record<string, string>;\n /**\n * file-based 导出:给 sandbox + ctx(ctx.telemetry.endpoint 必有),自己写 / 追加配置文件。\n * 运行器在 agent.setup 之后、首次 send 之前调一次(仅当 tracing 开 + 有 endpoint)。\n * 注:codex 的 [otel.trace_exporter.otlp-http] 是子表,configure 在 setup 写完主配置后\n * 追加到 config.toml 末尾,天然满足「子表在所有上层表之后」。\n */\n configure?(sandbox: Sandbox, ctx: SandboxAgentContext): Promise<void> | void;\n}\n/**\n * 一条会话线。核心承诺只有一句:**同一条会话线的每次 send 拿到同一个 `ctx.session`,\n * 新会话线(eval 第一轮 / t.newSession() 之后)拿到一个全新的。**\n * 服务端历史用 `id` / `capture`,客户端历史与 HITL 停轮现场用 Adapter\n * 私有的 typed slot。\"第一轮\"是新会话线的自然形态,没有要判断的分支。\n */\nexport interface AgentSession {\n /** 会话续接:服务端记历史。本线记过的会话 id;新会话线是 undefined。 */\n readonly id?: string;\n /** 记回传的会话 id;只在还没记过时落地(first-writer-wins),空值忽略。 */\n capture(id: string | undefined): void;\n /** 读 Adapter 私有 slot;新会话线或未写过的 slot 为空。 */\n get<T>(slot: SessionSlot<T>): T | undefined;\n /** 写 Adapter 私有 slot;同一 slot 再写覆盖旧值。 */\n set<T>(slot: SessionSlot<T>, value: T): void;\n /** 读并删除 Adapter 私有 slot,只消费一次。 */\n take<T>(slot: SessionSlot<T>): T | undefined;\n}\n/** Runner-owned Attempt resource registry; adapters use it for cross-send drivers. */\nexport interface AttemptResourceRegistry {\n acquire<T>(acquire: () => Promise<T>, lifecycle: {\n shutdown: (resource: T, signal: AbortSignal) => Promise<void>;\n release: (resource: T, signal: AbortSignal) => Promise<void>;\n }): Promise<T>;\n shutdownAll(signal: AbortSignal): Promise<void>;\n}\nexport interface AgentContext {\n /**\n * 软取消信号:合并了 attempt 超时、run 级中断(用户 Ctrl+C)与评估用例自身的中断请求\n * (见 src/runner/attempt.ts)。adapter 可以选择性检查它(或直接传给 `fetch`)以提前\n * 优雅退出,但这不是唯一的硬边界——即便 adapter 完全忽略它,运行器也会用\n * `Effect.timeoutTo` 兜底强制收尾(停 Sandbox 容器)。\n */\n readonly signal: AbortSignal;\n /** 本次 attempt 用的模型名,透传自 experiment 的 `model` 字段。sandbox 型 agent 通常在 setup 里用它写配置,direct 型通常在 send 里用它选模型。 */\n readonly model?: string;\n /** 模型推理努力程度;归属同 model——实验决定,省略时不覆盖 agent 原生默认。 */\n readonly reasoningEffort?: string;\n /**\n * experiment 的 `flags` 字段原样透传,内容和结构完全由 experiment 作者自定义\n * (如 `{ webResearch: true }`、`{ systemPrompt: \"...\" }`)。adapter 按自己的约定\n * 读取其中的字段;框架本身不解释、不校验它的内容。命名特意避开 CLI 解析出的\n * `flag`(跑法层面的 --timeout/--budget 等),两者是不相关的概念。\n */\n readonly flags: ExperimentFlags;\n /**\n * 路径推导出的实验 id(与结果归属 `runWho` / `AgentRun.experimentId` 同源);不经\n * experiment 跑(如脱离 CLI、直接构造 `AgentRun` 的场景)时为 undefined。典型用途:\n * SandboxLayer command 按实验隔离跨 attempt 的基础设施内容,或 adapter 按实验切换鉴权 / 路由。\n * 与 `flags`(实验条件的\n * 具体取值)是两个维度——这里只是「跑的是哪个实验」的稳定标识,不携带条件内容。\n */\n readonly experimentId?: string;\n /**\n * 当前 Attempt 对应的 eval ID。NiceEval Runner 始终从 discovery 后的评估用例身份填入;\n * 第三方直接构造 AgentContext 时可省略。Adapter 可用它定位与题目同目录的只读宿主资产,\n * 但不能据此绕过 Sandbox 的隐藏判据隔离。\n */\n readonly evalId?: string;\n /** 当前 Attempt 的评估组;未分组评估用例省略。 */\n readonly evalGroup?: {\n readonly id: string;\n readonly definitionHash: string;\n };\n /** Runner 填入的当前 Attempt 引用;第三方直接构造上下文时可省略。 */\n readonly attempt?: AttemptRef;\n readonly session: AgentSession;\n /**\n * 仅当配置了 OTel 接入时有(agent 的 `tracing` 块 / config 的 telemetry 存在):\n * 本次运行的 OTLP traces 接收信息(endpoint + env-based 导出 env)。\n * 怎么把它交给 CLI 由 agent 的 `tracing` 块声明:env-based 的把 ctx.telemetry.env\n * spread 进 send;file-based 的在 tracing.configure 里写配置。远程 HTTP 接入的 send\n * 只需要把 headers spread 进请求头(每轮一个新 traceparent);端点是启动期配置\n * (defineConfig({ telemetry: { port } }))固定的,不从这里传。\n */\n readonly telemetry?: Telemetry;\n /**\n * 作用域反馈:报告此刻正在做什么(turn / tool / 安装进度)。短命状态——Human profile\n * 更新 active 行,`agent`/`ci` 不逐条打印,也不进最终结果;不要每个 token/delta 都调用。\n * runner 按当前回调所处的生命周期阶段(agent.setup / agent.run / agent.teardown)归因,\n * 调用方不能冒充其它阶段(见 docs/feature/experiments/library.md)。\n */\n progress(update: ProgressUpdate): void;\n /**\n * 作用域反馈:报告运行结束后仍应保留的问题(协议降级、数据不完整、cleanup 问题)。\n * 永久事件,落进 attempt 的 diagnostics 并进各 profile 的永久输出;`dedupeKey` 去重。\n * 即使 level 为 \"error\" 也不改变 Turn.status / verdict——无法继续时抛异常。\n */\n diagnostic(input: DiagnosticInput): void;\n /**\n * 显式 timeout breadcrumb。它也更新 Human active 行,但最近若干条会在超时失败时\n * 并入结果的 error 信息;不得用它承载 user message、tool input 或其它只应短命显示的文本。\n */\n log(msg: string): void;\n}\n/** 仅 Sandbox Agent 可见的上下文;Direct Agent 不接收也不伪造 Sandbox。 */\nexport interface SandboxAgentContext extends AgentContext {\n readonly sandbox: Sandbox;\n}\n/** Per-send context; setup and teardown deliberately never receive a turn identity. */\nexport interface AgentSendContext extends AgentContext {\n readonly turnId: import(\"../record/family/source-receipt/model.ts\").TurnId;\n}\nexport interface SandboxAgentSendContext extends SandboxAgentContext {\n readonly turnId: import(\"../record/family/source-receipt/model.ts\").TurnId;\n}\n/**\n * Agent `setup` 的上下文:沙箱上下文再加一条宿主侧安装清单回执通道。只有 `setup` 收得到它——\n * 安装事实在这一步产生,`send` / `teardown` 不再补写。\n */\nexport interface SandboxAgentSetupContext extends SandboxAgentContext {\n /**\n * 交回这次实际安装了什么(Skill / 原生 Plugin / MCP / 原生配置文件)。清单是宿主侧内存对象,\n * 运行器直接存成 attempt artifact `agent-setup.json`;沙箱里不落任何框架文件\n * (见 docs/observability.md「宿主侧行为断言:t.o11y」)。什么都没装就别调——空清单不生成\n * artifact。同一次 `setup` 内多次调用后写覆盖先写。\n */\n reportSetup(manifest: AgentSetupManifest): void;\n}\n/**\n * agent 自己的沙箱运行时生命周期(每个 Attempt 一次,与「每轮 send」分开):\n * CLI 安装属于 runner 的 agent.ensure;`setup` 只写 model/base/auth 等本 Attempt 运行配置,`send` 只管把一轮 prompt\n * 跑起来(第一次 fresh / 后续 resume)+ 解析 transcript,`teardown` 清理。\n * 运行器在 layer prepare → agent.ensure → baseline 后、第一次 send 前调一次 `setup`,不返回值;\n * `teardown` 当且仅当本 attempt 走到过 `setup` 时点才执行(`setup` 抛错不豁免——半初始化\n * 的现场同样要扫尾),在 finally 里跑。要把 `setup` 里创建的句柄传给 `teardown`,以\n * `sandbox` 实例为键存取(同一个 Agent 实例服务并发 attempt,不要用实例字段或模块变量)。\n */\nexport type AgentSetup = (sandbox: Sandbox, ctx: SandboxAgentSetupContext) => MaybePromise<void>;\nexport type AgentTeardown = (sandbox: Sandbox, ctx: SandboxAgentContext) => MaybePromise<void>;\nexport type DirectAgentSetup = (ctx: AgentContext) => MaybePromise<void>;\nexport type DirectAgentTeardown = (ctx: AgentContext) => MaybePromise<void>;\n/**\n * 本 agent 的原生 OTLP span → canonical GenAI semconv 的薄 mapper。\n * 由 adapter 声明(和 tracing.env 一样属于「连到谁」的特殊性),core 只调接口——\n * 省略时 core 走通用 heuristic 兜底,不按 agent 名字分支。\n */\nexport type SpanMapper = (spans: TraceSpan[]) => TraceSpan[];\ninterface AgentBase {\n readonly name: string;\n /** Agent-specialized Adapter Eval factories bound to the session context contract. */\n readonly defineEval: typeof import(\"../define.ts\").defineEval;\n readonly defineScoreEval: typeof import(\"../define.ts\").defineScoreEval;\n /** 该 Adapter 的常态证据覆盖声明;六通道必填。 */\n readonly evidenceCoverage: EvidenceCoverage;\n /** 原生 span → canonical 的薄 mapper;省略走通用 heuristic。只影响瀑布图。 */\n spanMapper?: SpanMapper;\n /**\n * 可选 send 执行失败分类器:归类 adapter reject 的 `SendFailure`;可信的 failed Turn\n * 是领域结果,不进入这条链。\n * 返回 `undefined` 表示不认识、回落保守兜底。链上排在实验的 `classifyFailure` 之后,\n * 实验作者认领过的失败问不到这里。分类器只声明决策轴与诊断词,不影响重试策略(次数、\n * 退避对所有 agent 一致);抛错按 `undefined` 回落并被吞掉,不掩盖原始失败。形状与分类链、\n * 执行体时序见 docs/feature/error-classification/architecture.md。\n */\n classifySendFailure?: SendFailureClassifier;\n}\n// ───────────────────────── Agent Ensure / Installer ─────────────────────────\n/**\n * Agent 安装身份:纯数据,进 configHash 与 `run.json`。\n * 与 sandbox case 身份正交——改 Agent 版本不重建任务 BuildKey。\n * 见 docs/feature/adapters/architecture/agent-ensure.md。\n */\nexport interface AgentIdentity {\n /** Agent 名(与 adapter 显示名对齐,如 `codex`)。 */\n readonly agent: string;\n /** 精确版本;禁止 `latest` 等无钉死取值——无稳定身份不参与可携带结果。 */\n readonly version: string;\n /** 配方修订(内置用数字字符串,自定义可用任意稳定串)。 */\n readonly revision: string;\n}\n/**\n * 三种安装模式,全部显式;失败后不允许在模式之间静默降级。\n * - `staged`:制品在题面网络外准备,经文件 API 送入(内置默认)。\n * - `sandbox-network`:显式声明用沙箱内网络安装。\n * - `verify-only`:只接受预装命中;探测失败立即 errored。\n */\nexport type AgentInstallMode = \"staged\" | \"sandbox-network\" | \"verify-only\";\n/** staged payload 的完成态目标平台;Linux 的 libc 是身份的一部分,不能留成缺值。 */\nexport type AgentArtifactPlatform = {\n readonly _tag: \"Linux\";\n readonly os: \"linux\";\n readonly arch: string;\n readonly libc: \"gnu\" | \"musl\";\n} | {\n readonly _tag: \"Darwin\";\n readonly os: \"darwin\";\n readonly arch: string;\n} | {\n readonly _tag: \"Windows\";\n readonly os: \"windows\";\n readonly arch: string;\n};\n/**\n * Run 级准备好的锁定制品。digest + 实际 platform 只属于运行 provenance,不进入\n * configHash 或 Eval fingerprint;静态 installer identity/revision 与 ProviderPlan target 才是配置身份。\n * 安装时经主 Sandbox 文件 API 上传,不依赖题面网络。\n */\nexport interface AgentStagedArtifact {\n readonly platform: AgentArtifactPlatform;\n /** digest-backed 不可变内容句柄;不向 Attempt 泄露宿主路径。 */\n readonly content: RegisteredSandboxContent;\n /** 沙箱内落点,位于 workdir 外的 Agent 自有目录。 */\n readonly targetPath: string;\n /** 安装协议必须显式,不从缺失字段推断历史默认。 */\n readonly install: AgentArtifactInstallShape;\n}\nexport type AgentArtifactInstallShape = {\n readonly kind: \"npm-tarball\";\n} | {\n readonly kind: \"self-contained\";\n readonly binPath: string;\n};\n/** `agent.ensure` 的取值:区分 probe 命中与本次安装。 */\nexport type AgentEnsureOutcome = \"hit\" | \"installed\";\n/**\n * Adapter 声明的纯 Ensure 义务:身份与只读 probe。probe 的非零退出是未命中,不是\n * Agent 任务失败;Runner 决定是否交给 identity 精确匹配的 installer。\n */\nexport interface AgentEnsure {\n readonly identity: AgentIdentity;\n readonly probe: StableSandboxCommand;\n /** Human-only transient labels; omitted adapters do not add an ensure detail. */\n readonly progress?: {\n readonly checking?: string;\n readonly ready?: string;\n };\n}\nexport interface AgentInstallContext {\n readonly identity: AgentIdentity;\n readonly targetPlatform: AgentArtifactPlatform;\n readonly signal: AbortSignal;\n progress(update: ProgressUpdate): void;\n}\nexport interface AgentArtifactContext {\n readonly targetPlatform: AgentArtifactPlatform;\n readonly signal: AbortSignal;\n}\nexport interface StagedAgentInstallContext extends AgentInstallContext {\n readonly artifact: AgentStagedArtifact;\n}\ninterface AgentInstallerBase {\n /** 必须与某条 ensure identity 完全相同;Runner 不做近似匹配。 */\n readonly identity: AgentIdentity;\n /** 支持的目标平台键(`linux-x64-gnu` 等);省略表示 installer 自己支持所有平台。 */\n readonly platforms?: readonly string[];\n /** Human-only transient label; omitted installers do not add an install detail. */\n readonly progress?: {\n readonly installing?: string;\n };\n /** @internal 只由内置 cache-safe staged installer 写入。 */\n readonly [AGENT_DOCKERFILE_CACHE_SAFE]?: true;\n}\nexport type AgentInstaller = (AgentInstallerBase & {\n readonly installMode: \"staged\";\n prepareArtifact(context: AgentArtifactContext): MaybePromise<AgentStagedArtifact>;\n install(sandbox: SandboxCommandTarget, context: StagedAgentInstallContext): MaybePromise<void>;\n}) | (AgentInstallerBase & {\n readonly installMode: \"sandbox-network\";\n readonly prepareArtifact?: never;\n install(sandbox: SandboxCommandTarget, context: AgentInstallContext): MaybePromise<void>;\n}) | (AgentInstallerBase & {\n readonly installMode: \"verify-only\";\n readonly prepareArtifact?: never;\n readonly install?: never;\n});\n/** 在 NiceEval 管理的 Sandbox 内驱动 CLI 的 Agent。 */\nexport interface SandboxAgent extends AgentBase {\n readonly kind: \"sandbox\";\n /** Adapter-owned command-only contribution to the unified Sandbox action DAG. */\n readonly sandbox?: SandboxLayer<\"command-only\">;\n /** Runner 在 author layers 后、setup 前按声明顺序执行。 */\n readonly ensure: readonly AgentEnsure[];\n /** 官方或第三方随 adapter 提供的安装层;仅 identity 精确匹配时可接手。 */\n readonly installers: readonly AgentInstaller[];\n setup?: AgentSetup;\n tracing?: AgentTracing;\n send(input: TurnInput, ctx: SandboxAgentSendContext): Promise<Turn>;\n teardown?: AgentTeardown;\n}\n/** 由 runner 直接调用函数、SDK 或服务端点的 Agent;目标进程可以在本地或远端。 */\nexport interface DirectAgent extends AgentBase {\n readonly kind: \"direct\";\n setup?: DirectAgentSetup;\n tracing?: Omit<AgentTracing, \"configure\">;\n send(input: TurnInput, ctx: AgentSendContext): Promise<Turn>;\n teardown?: DirectAgentTeardown;\n}\n/** 注册表里的 Agent;判别值描述驱动拓扑,不描述目标所在位置。 */\nexport type Agent = SandboxAgent | DirectAgent;\n/** `defineSandboxAgent()` 的入参形状(见 src/define.ts)——`kind: \"sandbox\"` 由 define 固定填入,不由用户声明。 */\nexport interface SandboxAgentDef {\n /** agent 的显示名/标识,原样进入 `Agent.name`——不是注册表查找 key,只用于展示、结果归属与去重指纹。 */\n name: string;\n /** 该 Adapter 的常态证据覆盖声明;完整采集可用 `completeEvidenceCoverage`。 */\n evidenceCoverage: EvidenceCoverage;\n /** Adapter 自有的 Sandbox action;Agent 不能选择或贡献 template。 */\n sandbox?: SandboxLayer<\"command-only\">;\n /** 单条或数组都按声明顺序规范化为 Agent layer。 */\n ensure: AgentEnsure | readonly AgentEnsure[];\n /**\n * 配对安装层;省略表示这个 adapter 只提供 probe 协议,未命中时 Runner 在\n * `agent.ensure` 点名缺失的精确 identity。工厂会把省略规范化成空数组。\n */\n installers?: readonly AgentInstaller[];\n /**\n * 每条 Attempt 一次(不是每轮 send 一次):写 config.toml / 鉴权配置(model/base/auth 等\n * 本 Attempt 内不变的运行配置)。CLI 的 probe、安装与复检属于 `agent.ensure`。运行器在 Sandbox\n * 备好(layer prepare/ensure/baseline 之后)、第一次 send 前调用一次,不返回值。\n */\n setup?: AgentSetup;\n /** OTLP 导出配置:Sandbox 里怎么让 CLI 把 trace 发到 endpoint(env / 配置文件),从 setup 拆出。 */\n tracing?: AgentTracing;\n /** 原生 span → canonical 的薄 mapper;省略走通用 heuristic。只影响瀑布图。 */\n spanMapper?: SpanMapper;\n /** 每轮一次:跑 prompt(fresh / resume)+ 解析成 events。作者 callback 保持 Promise ABI。 */\n send(input: TurnInput, ctx: SandboxAgentSendContext): Promise<Turn>;\n /** 可选 send 执行失败分类器:见 `Agent.classifySendFailure`。 */\n classifySendFailure?: SendFailureClassifier;\n /** Sandbox 销毁前的清理,当且仅当本 attempt 走到过 `setup` 时点才执行(`setup` 抛错不豁免),\n * 在 finally 里跑一次。 */\n teardown?: AgentTeardown;\n}\n/** `defineAgent()` 的入参形状(见 src/define.ts)——`kind: \"direct\"` 由 define 固定填入,不由用户声明。 */\nexport interface DirectAgentDef {\n /** agent 的显示名/标识,原样进入 `Agent.name`——不是注册表查找 key,只用于展示、结果归属与去重指纹。 */\n name: string;\n /** 该 Adapter 的常态证据覆盖声明;完整采集可用 `completeEvidenceCoverage`。 */\n evidenceCoverage: EvidenceCoverage;\n /**\n * 每个 attempt 一次。Direct Agent 不接收 Sandbox;常用于建立连接、鉴权等一次性准备。\n */\n setup?: DirectAgentSetup;\n /** OTLP 导出配置:被测对象怎么把 trace 发到 endpoint(env-based 注入)。 */\n tracing?: Omit<AgentTracing, \"configure\">;\n /** 原生 span → canonical 的薄 mapper;省略走通用 heuristic。只影响瀑布图。 */\n spanMapper?: SpanMapper;\n /** 每轮一次:把一轮 prompt 直接发给函数、SDK 或服务端点,解析响应成 events。作者 callback 保持 Promise ABI。 */\n send(input: TurnInput, ctx: AgentSendContext): Promise<Turn>;\n /** 可选 send 执行失败分类器:见 `Agent.classifySendFailure`。 */\n classifySendFailure?: SendFailureClassifier;\n /** 运行结束前的清理,当且仅当本 attempt 走到过 `setup` 时点才执行(`setup` 抛错不豁免),\n * 在 finally 里跑一次。 */\n teardown?: DirectAgentTeardown;\n}\n"]}
1
+ {"version":3,"file":"types.js","sourceRoot":"","sources":["types.ts"],"names":[],"mappings":";;;AACA,iDAAiD;AACjD,uDAAgE;AAoIhE,6DAA6D;AAC7D,oCAAoC;AACvB,QAAA,0BAA0B,GAAG,CAAC,UAAU,EAAE,SAAS,EAAE,aAAa,CAAU,CAAC","sourcesContent":["import type { ExperimentFlags } from \"../shared/types.ts\";\n// agent 域类型:Agent / Adapter 契约、会话与 tracing 导出配置。\nimport { AGENT_DOCKERFILE_CACHE_SAFE } from \"./cache-marker.ts\";\n// 「连到哪个被测对象、协议怎么说」的全部契约在这里(见 apps/docs-site/zh/explanation/adapter.mdx)。\n// 能力不再是问卷式声明:t 上解锁什么完全由构造证据决定(见 docs-site 「能力从哪来」一节)。\nimport type { DiagnosticInput, JsonValue, ProgressUpdate } from \"../shared/types.ts\";\nimport type { Effect } from \"effect\";\nimport type { StreamEvent, TraceSpan, Usage } from \"../o11y/types.ts\";\nimport type { Sandbox } from \"../sandbox/types.ts\";\nimport type { SandboxLayer } from \"../sandbox/layer.ts\";\nimport type { AttemptRef, MaybePromise, SandboxCommandTarget, StableSandboxCommand } from \"../sandbox/commands.ts\";\nimport type { RegisteredSandboxContent } from \"../sandbox/content.ts\";\nimport type { SendFailureClassifier } from \"../context/send-failures.ts\";\nimport type { SessionKey } from \"./session-key.ts\";\nexport type { SessionKey } from \"./session-key.ts\";\n/**\n * 本地 stdio 形态的 MCP server:沙箱内起子进程,按 stdio 说 MCP 协议。\n * 与 {@link McpHttpServer} 按形状判别(有 `command` 的是 stdio,有 `url` 的是 HTTP)。\n */\nexport interface McpStdioServer {\n /** 服务器唯一名(config key)。 */\n name: string;\n /** 启动命令(如 \"npx\"、\"node\"、\"uvx\")。 */\n command: string;\n /** 传给命令的参数。 */\n args?: readonly string[];\n /** 注入服务器进程的环境变量(可能含 secret,不进 manifest)。 */\n env?: Readonly<globalThis.Record<string, string>>;\n /** stdio 与 HTTP 是互斥完成态,不能把另一分支的字段混进来。 */\n url?: never;\n headers?: never;\n}\n/**\n * 远程 Streamable HTTP 形态的 MCP server:沙箱直接连一个 HTTP 端点。\n * `url` 必须沙箱内可达——宿主机上的服务先经隧道(cloudflared / tailscale 等)暴露。\n */\nexport interface McpHttpServer {\n /** 服务器唯一名(config key)。 */\n name: string;\n /** Streamable HTTP 端点(如 https://mem.example.com/mcp/)。 */\n url: string;\n /** 逐字写进每个请求的 HTTP 头(常用于 Authorization;可能含 secret,不进 manifest)。 */\n headers?: Readonly<globalThis.Record<string, string>>;\n /** HTTP 与 stdio 是互斥完成态,不能把另一分支的字段混进来。 */\n command?: never;\n args?: never;\n env?: never;\n}\n/**\n * MCP server 描述符 —— 支持 MCP 的 adapter(Claude Code / Codex)共用的工具服务单元,\n * 不是 native plugin 的一种。stdio 与 Streamable HTTP 两种形态按形状判别,不设 kind 标签\n * (两种形态各有唯一必填判别字段);同时给出 `command` 与 `url` 属配置错误,before 报错点名。\n * 在 agent factory config 里声明,before 阶段写进各自的配置文件。\n * 见 docs/feature/adapters/architecture/coding-agent-extensions.md「类型边界」。\n */\nexport type McpServer = McpStdioServer | McpHttpServer;\n/**\n * Skill 的来源描述 —— Claude Code / Codex / Bub 共用的**数据类型**:只统一「从哪里取得\n * 哪份 Skill」,安装位置、发现机制、要不要额外写 project instruction 由各 Adapter 决定。\n * 见 docs/feature/adapters/architecture/coding-agent-extensions.md「类型边界」。\n */\nexport type SkillSpec = {\n kind: \"local\";\n /** 相对项目根(跑 niceeval 的目录)的 Skill 文件或目录:`SKILL.md`、含 `SKILL.md` 的目录,或单个 `.md`。 */\n path: string;\n /** 展示名;省略时由文件或目录名推导(`<dir>/SKILL.md` → `<dir>`,`foo.md` → `foo`)。 */\n name?: string;\n} | {\n kind: \"repo\";\n /** GitHub `owner/repo` 或 Git URL。 */\n source: string;\n /** 多 Skill Repo 中要启用的 Skill;repo 只有一个 Skill 时可省略,多个时省略即 before 失败。 */\n skills?: readonly string[];\n /** Tag、Commit 或 Branch;省略表示 repo 默认 ref。 */\n ref?: string;\n};\n/** Manifest 里的一条 Skill 安装记录:本地(带内容哈希)或 repo(带来源与 ref)。 */\nexport type AgentInstallationSkill = {\n kind: \"local\";\n name: string;\n path: string;\n sha256: string;\n} | {\n kind: \"repo\";\n source: string;\n ref?: string;\n skills: string[];\n};\n/**\n * 一次 Agent before 实际装了什么 —— 沙箱型 Coding Agent Adapter 在 before 收尾经\n * `ctx.reportInstallation()` 交回的安装清单。它是宿主侧内存对象,运行器把它作为 attempt artifact 存成\n * `agent-setup.json`(见 docs/feature/record/architecture.md);沙箱磁盘上不落它。不参与评分,\n * 只回答「这次实际装了什么」;**环境变量值与 secret 不写进来**(所以 mcpServers 只记\n * name/command/args)。\n */\nexport interface AgentInstallationManifest {\n /** 装进去的 Skill(按配置顺序;同名来自多个来源时逐条保留,不静默合并)。 */\n skills: AgentInstallationSkill[];\n /** Agent 原生 Plugin(Claude Code / Codex 各自的 Marketplace 协议)。 */\n nativePlugins?: Array<{\n agent: \"claude-code\" | \"codex\";\n marketplace: {\n name: string;\n source: string;\n ref?: string;\n };\n name: string;\n /** 安装后 CLI 报告的版本;取不到时省略。 */\n resolvedVersion?: string;\n }>;\n /** 挂上的 MCP server(只记非 secret 字段:stdio 不含 env,HTTP 不含 headers)。 */\n mcpServers?: Array<{\n name: string;\n command: string;\n args?: string[];\n } | {\n name: string;\n url: string;\n }>;\n /**\n * 官方原生配置文件(Claude Code `settings.json` / Codex `config.toml`):只记 Agent 名、\n * 项目相对来源路径与原始字节的 SHA-256,不落正文 —— 任意官方配置都可能携带敏感字符串,\n * 不能靠字段白名单证明适合原样落盘。\n */\n nativeConfigFile?: {\n agent: \"claude-code\" | \"codex\";\n path: string;\n sha256: string;\n };\n /** Bub 的 Python Plugin(规范化后的 package 串)。 */\n pythonPlugins?: Array<{\n package: string;\n }>;\n}\n// ───────────────────────── 证据覆盖声明 ─────────────────────────\n/** 证据通道的覆盖状态;不存在隐含的 unknown 第四态。 */\nexport const EVIDENCE_COVERAGE_STATUSES = [\"complete\", \"partial\", \"unavailable\"] as const;\nexport type EvidenceCoverageStatus = (typeof EVIDENCE_COVERAGE_STATUSES)[number];\n/** 非完整通道必须解释缺口;完整通道不能携带一个自相矛盾的 reason。 */\nexport type EvidenceCoverageEntry = {\n readonly status: \"complete\";\n readonly reason?: never;\n} | {\n readonly status: Exclude<EvidenceCoverageStatus, \"complete\">;\n readonly reason: string;\n};\n/**\n * 覆盖声明(EvidenceCoverage):完整性不是口头承诺,是随数据走的声明\n * (见 docs/feature/adapters/architecture/evidence.md)。两层:\n * - Agent 级默认(`defineAgent` / `defineSandboxAgent` 的 `evidenceCoverage`)必须逐一声明\n * 六个通道;官方 SDK 适配器可使用 `completeEvidenceCoverage`。\n * - Turn 级降级(`Turn.evidenceCoverage`)只用于相对 Agent 默认值降级。\n */\nexport interface EvidenceCoverage {\n /** 完整事件流(event / notEvent / order 的依据)。 */\n readonly events: EvidenceCoverageEntry;\n /** action 生命周期(工具正负断言、顺序、失败的依据)。 */\n readonly actions: EvidenceCoverageEntry;\n /** assistant / user message(reply 与显式 message 值断言的依据)。 */\n readonly messages: EvidenceCoverageEntry;\n /** usage(token / cost 上限断言的依据)。 */\n readonly usage: EvidenceCoverageEntry;\n /** Turn status 的真实性(succeeded / parked 的依据)——恒 completed 的映射必须声明非 complete。 */\n readonly status: EvidenceCoverageEntry;\n /** Turn.data(供 `t.check(turn.data, match)` 检查结构化输出)。 */\n readonly data: EvidenceCoverageEntry;\n}\n/** 单轮只列相对 Agent 默认值的降级通道;省略的通道沿用 Agent 声明。 */\nexport type TurnEvidenceCoverage = Partial<EvidenceCoverage>;\n/** 随一轮消息附带的文件(图片等多模态输入)。 */\nexport interface InputFile {\n /** 文件名(可选,供 adapter / 模型参考)。 */\n readonly filename?: string;\n /** MIME 类型,如 `image/png`、`image/jpeg`。 */\n readonly mimeType: string;\n /** base64 编码的文件内容(JSON 友好,remote adapter 可直接放进请求体)。 */\n readonly dataBase64: string;\n}\n/**\n * HITL 回答轮里,人的裁决以结构化形式随 `input.responses` 到达——adapter 不需要解析\n * `text` 去猜哪句回答对应哪个请求、算不算批准。见 apps/docs-site/zh/explanation/adapter.mdx\n * 「不同回答的入参」一节的四种典型形态。\n */\nexport type AnswerValue = {\n readonly optionId: string;\n readonly text?: never;\n} | {\n readonly text: string;\n readonly optionId?: never;\n};\nexport type InputResponse = {\n /** 对应哪条 input.requested 请求;多个请求并停时靠它对位。 */\n readonly requestId: string;\n} & AnswerValue;\nexport interface TurnInput {\n readonly text: string;\n /** 本轮附带的文件(图片等)。adapter 自行决定怎么投递;不支持多模态的 adapter 忽略它。 */\n readonly files?: readonly InputFile[];\n /** 仅回答轮(t.respond / t.respondAll):逐请求的结构化回答,按 requestId 对位。 */\n readonly responses?: readonly InputResponse[];\n}\n/** adapter 的 send 返回值(事件流为核心)。 */\nexport interface Turn {\n readonly events: StreamEvent[];\n readonly data?: JsonValue;\n readonly status: \"completed\" | \"failed\" | \"waiting\";\n readonly usage?: Usage;\n /**\n * 本轮相对 Agent 默认覆盖的降级声明(这一轮流断了、拿不到 usage 等);\n * 只能降级,不能把 Agent 未声明的通道升格成 complete(见 EvidenceCoverage)。\n */\n readonly evidenceCoverage?: TurnEvidenceCoverage;\n}\n/**\n * 本次运行的 OTLP traces 接收信息(仅当配置了 OTel 接入时有——agent 的 `tracing` 块 /\n * config 的 telemetry 存在)。经 ctx.telemetry 交给 agent。span 只进瀑布图,不喂断言。\n * 远程 HTTP 接入的 send 只需要 spread `headers`(每轮一个新 traceparent);接收端点是\n * 启动期配置(`defineConfig({ telemetry: { port } } )`)固定的,不从这里传。\n */\nexport interface Telemetry {\n /** 接收端点(完整路径,形如 http://127.0.0.1:PORT/v1/traces)。 */\n readonly endpoint: string;\n /**\n * env-based 导出的 env(= AgentTracing.env(endpoint) 的结果),ready-to-spread。\n * adapter 的 send 直接 `{ ...ctx.telemetry?.env }` 注入,不必手搓 OTEL_* 拼装。\n */\n readonly env?: Readonly<globalThis.Record<string, string>>;\n /**\n * 本轮的 W3C trace context(traceparent),每轮一个新值。send 把它 spread 进 HTTP 请求头\n * (或注入子进程 env)——应用埋点支持 context 传播时,本轮 span 挂到我们给的 trace 下,\n * 并发跑 eval 的 span 归属才精确;不带则回退时间窗口归属(该 agent 自动降为串行)。\n */\n readonly headers?: Readonly<globalThis.Record<string, string>>;\n}\n/**\n * agent 的 OTLP 导出配置 —— 「沙箱里怎么让这个 CLI 把 trace 发到 endpoint」。\n * 刻意从 before 里拆出来:before 管运行配置,这里只管 otel 导出。CLI 安装归 `agent.ensure`。两种投递方式\n * (互不排斥,按 CLI 而定):\n * · env-based(标准 OTEL_* 环境变量,如 bub/Python OTel SDK)—— 用 `env`;\n * · file-based(CLI 自有配置文件,如 codex 的 config.toml [otel] 块)—— 用 `configure`。\n */\nexport interface AgentTracing {\n /**\n * 线协议(codex 发 OTLP/JSON、bub 发 OTLP/protobuf)。接收器按 content-type 自动解码,\n * 此字段仅作声明/日志用,也为将来按协议分流留口。\n */\n protocol?: \"http/json\" | \"http/protobuf\";\n /**\n * 接收器粒度(仅非沙箱 agent;沙箱型每沙箱一个,与此无关):\n * · \"attempt\"(默认)—— 每个 attempt 一个接收器。适合每轮能切换导出端点的被测对象\n * (进程内 adapter,如内置 aiSdkAgent 的可切换 exporter),attempt 间天然隔离、可全并发。\n * · \"run\" —— 整个 run 共享一个接收器(每 agent 一个)。**长驻服务必选**:应用的 OTEL_*\n * env 进程启动时读一次,per-attempt 端口会在第一个 attempt 结束时失效。span 逐轮归属\n * (traceparent / 时间窗口 + 串行守卫),见 ctx.telemetry.headers。\n * config 配了 `telemetry`(固定端口,无侵入接入的长驻服务场景)的非沙箱 agent 自动按 \"run\" 处理。\n */\n scope?: \"attempt\" | \"run\";\n /**\n * env-based 导出:给 endpoint → 返回要注入每轮 send 的 env(纯函数)。运行器把结果\n * 放进 ctx.telemetry.env,send 直接 spread。\n */\n env?(endpoint: string): globalThis.Record<string, string>;\n /**\n * file-based 导出:给 sandbox + ctx(ctx.telemetry.endpoint 必有),自己写 / 追加配置文件。\n * 运行器在 agent.before 之后、首次 send 之前调一次(仅当 tracing 开 + 有 endpoint)。\n * 注:codex 的 [otel.trace_exporter.otlp-http] 是子表,configure 在 before 写完主配置后\n * 追加到 config.toml 末尾,天然满足「子表在所有上层表之后」。\n */\n configure?(sandbox: Sandbox, ctx: SandboxAgentContext): Promise<void> | void;\n}\n/**\n * 一条会话线。核心承诺只有一句:**同一条会话线的每次 send 拿到同一个 `ctx.session`,\n * 新会话线(eval 第一轮 / t.newSession() 之后)拿到一个全新的。**\n * 服务端历史用 `id` / `capture`,客户端历史与 HITL 停轮现场用 Adapter\n * 私有的 typed key。\"第一轮\"是新会话线的自然形态,没有要判断的分支。\n */\nexport interface AgentSession {\n /** 会话续接:服务端记历史。本线记过的会话 id;新会话线是 undefined。 */\n readonly id?: string;\n /** 记回传的会话 id;只在还没记过时落地(first-writer-wins),空值忽略。 */\n capture(id: string | undefined): void;\n /** 读 Adapter 私有 key;新会话线或未写过的 key 为空。 */\n get<T>(key: SessionKey<T>): T | undefined;\n /** 写 Adapter 私有 key;同一 key 再写覆盖旧值。 */\n set<T>(key: SessionKey<T>, value: T): void;\n /** 读并删除 Adapter 私有 key,只消费一次。 */\n take<T>(key: SessionKey<T>): T | undefined;\n}\n/** Runner-owned Attempt resource registry; adapters use it for cross-send drivers. */\nexport interface AttemptResourceRegistry {\n acquire<T>(acquire: () => Promise<T>, lifecycle: {\n shutdown: (resource: T, signal: AbortSignal) => Promise<void>;\n release: (resource: T, signal: AbortSignal) => Promise<void>;\n }): Promise<T>;\n shutdownAll(signal: AbortSignal): Promise<void>;\n}\nexport interface AgentContext {\n /**\n * 软取消信号:合并了 attempt 超时、run 级中断(用户 Ctrl+C)与评估用例自身的中断请求\n * (见 src/runner/attempt.ts)。adapter 可以选择性检查它(或直接传给 `fetch`)以提前\n * 优雅退出,但这不是唯一的硬边界——即便 adapter 完全忽略它,运行器也会用\n * `Effect.timeoutTo` 兜底强制收尾(停 Sandbox 容器)。\n */\n readonly signal: AbortSignal;\n /** 本次 attempt 用的模型名,透传自 experiment 的 `model` 字段。sandbox 型 agent 通常在 before 里用它写配置,direct 型通常在 send 里用它选模型。 */\n readonly model?: string;\n /** 模型推理努力程度;归属同 model——实验决定,省略时不覆盖 agent 原生默认。 */\n readonly reasoningEffort?: string;\n /**\n * experiment 的 `flags` 字段原样透传,内容和结构完全由 experiment 作者自定义\n * (如 `{ webResearch: true }`、`{ systemPrompt: \"...\" }`)。adapter 按自己的约定\n * 读取其中的字段;框架本身不解释、不校验它的内容。命名特意避开 CLI 解析出的\n * `flag`(跑法层面的 --timeout/--budget 等),两者是不相关的概念。\n */\n readonly flags: ExperimentFlags;\n /**\n * 路径推导出的实验 id(与结果归属 `runWho` / `AgentRun.experimentId` 同源);不经\n * experiment 跑(如脱离 CLI、直接构造 `AgentRun` 的场景)时为 undefined。典型用途:\n * SandboxLayer command 按实验隔离跨 attempt 的基础设施内容,或 adapter 按实验切换鉴权 / 路由。\n * 与 `flags`(实验条件的\n * 具体取值)是两个维度——这里只是「跑的是哪个实验」的稳定标识,不携带条件内容。\n */\n readonly experimentId?: string;\n /**\n * 当前 Attempt 对应的 eval ID。NiceEval Runner 始终从 discovery 后的评估用例身份填入;\n * 第三方直接构造 AgentContext 时可省略。Adapter 可用它定位与题目同目录的只读宿主资产,\n * 但不能据此绕过 Sandbox 的隐藏判据隔离。\n */\n readonly evalId?: string;\n /** 当前 Attempt 的评估组;未分组评估用例省略。 */\n readonly evalGroup?: {\n readonly id: string;\n readonly definitionHash: string;\n };\n /** Runner 填入的当前 Attempt 引用;第三方直接构造上下文时可省略。 */\n readonly attempt?: AttemptRef;\n readonly session: AgentSession;\n /**\n * 仅当配置了 OTel 接入时有(agent 的 `tracing` 块 / config 的 telemetry 存在):\n * 本次运行的 OTLP traces 接收信息(endpoint + env-based 导出 env)。\n * 怎么把它交给 CLI 由 agent 的 `tracing` 块声明:env-based 的把 ctx.telemetry.env\n * spread 进 send;file-based 的在 tracing.configure 里写配置。远程 HTTP 接入的 send\n * 只需要把 headers spread 进请求头(每轮一个新 traceparent);端点是启动期配置\n * (defineConfig({ telemetry: { port } }))固定的,不从这里传。\n */\n readonly telemetry?: Telemetry;\n /**\n * 作用域反馈:报告此刻正在做什么(turn / tool / 安装进度)。短命状态——Human profile\n * 更新 active 行,`agent`/`ci` 不逐条打印,也不进最终结果;不要每个 token/delta 都调用。\n * runner 按当前回调所处的生命周期阶段(agent.before / agent.run / agent.after)归因,\n * 调用方不能冒充其它阶段(见 docs/feature/experiments/library.md)。\n */\n progress(update: ProgressUpdate): void;\n /**\n * 作用域反馈:报告运行结束后仍应保留的问题(协议降级、数据不完整、cleanup 问题)。\n * 永久事件,落进 attempt 的 diagnostics 并进各 profile 的永久输出;`dedupeKey` 去重。\n * 即使 level 为 \"error\" 也不改变 Turn.status / verdict——无法继续时抛异常。\n */\n diagnostic(input: DiagnosticInput): void;\n /**\n * 显式 timeout breadcrumb。它也更新 Human active 行,但最近若干条会在超时失败时\n * 并入结果的 error 信息;不得用它承载 user message、tool input 或其它只应短命显示的文本。\n */\n log(msg: string): void;\n}\n/** 仅 Sandbox Agent 可见的上下文;Direct Agent 不接收也不伪造 Sandbox。 */\nexport interface SandboxAgentContext extends AgentContext {\n readonly sandbox: Sandbox;\n}\n/** Per-send context; before and after deliberately never receive a turn identity. */\nexport interface AgentSendContext extends SandboxAgentContext {\n readonly turnId: import(\"../record/family/source-receipt/model.ts\").TurnId;\n}\nexport interface SandboxAgentSendContext extends AgentSendContext {\n}\n/**\n * Agent `before` 的上下文:沙箱上下文再加一条宿主侧安装清单回执通道。只有 `before` 收得到它——\n * 安装事实在这一步产生,`send` / `after` 不再补写。\n */\nexport interface AgentBeforeContext extends SandboxAgentContext {\n /**\n * 交回这次实际安装了什么(Skill / 原生 Plugin / MCP / 原生配置文件)。清单是宿主侧内存对象,\n * 运行器直接存成 attempt artifact `agent-setup.json`;沙箱里不落任何框架文件\n * (见 docs/observability.md「宿主侧行为断言:t.o11y」)。什么都没装就别调——空清单不生成\n * artifact。同一次 `before` 内多次调用后写覆盖先写。\n */\n reportInstallation(manifest: AgentInstallationManifest): void;\n}\n/**\n * agent 自己的沙箱运行时生命周期(每个 Attempt 一次,与「每轮 send」分开):\n * CLI 安装属于 runner 的 agent.ensure;`before` 只写 model/base/auth 等本 Attempt 运行配置,`send` 只管把一轮 prompt\n * 跑起来(第一次 fresh / 后续 resume)+ 解析 transcript,`after` 清理。\n * 运行器在 layer prepare → agent.ensure → baseline 后、第一次 send 前调一次 `before`,不返回值;\n * `after` 当且仅当本 attempt 走到过 `before` 时点才执行(`before` 抛错不豁免——半初始化\n * 的现场同样要扫尾),在 finally 里跑。要把 `before` 里创建的句柄传给 `after`,以\n * `sandbox` 实例为键存取(同一个 Agent 实例服务并发 attempt,不要用实例字段或模块变量)。\n */\nexport interface SandboxAgentBeforeContext extends AgentBeforeContext {\n}\nexport type AgentBefore = (sandbox: Sandbox, ctx: AgentBeforeContext) => MaybePromise<void>;\nexport type AgentAfter = (sandbox: Sandbox, ctx: SandboxAgentContext) => MaybePromise<void>;\nexport type DirectAgentBefore = (ctx: AgentContext) => MaybePromise<void>;\nexport type DirectAgentAfter = (ctx: AgentContext) => MaybePromise<void>;\n/**\n * 本 agent 的原生 OTLP span → canonical GenAI semconv 的薄 mapper。\n * 由 adapter 声明(和 tracing.env 一样属于「连到谁」的特殊性),core 只调接口——\n * 省略时 core 走通用 heuristic 兜底,不按 agent 名字分支。\n */\nexport type SpanMapper = (spans: TraceSpan[]) => TraceSpan[];\ninterface AgentBase {\n readonly name: string;\n /** Agent-specialized Adapter Eval factories bound to the session context contract. */\n readonly defineEval: typeof import(\"../define.ts\").defineEval;\n readonly defineScoreEval: typeof import(\"../define.ts\").defineScoreEval;\n /** 该 Adapter 的常态证据覆盖声明;六通道必填。 */\n readonly evidenceCoverage: EvidenceCoverage;\n /** 原生 span → canonical 的薄 mapper;省略走通用 heuristic。只影响瀑布图。 */\n spanMapper?: SpanMapper;\n /**\n * 可选 send 执行失败分类器:归类 adapter reject 的 `SendFailure`;可信的 failed Turn\n * 是领域结果,不进入这条链。\n * 返回 `undefined` 表示不认识、回落保守兜底。抛出点的内建失败声明优先。\n * 分类器只声明决策轴与诊断词,不影响重试策略(次数、\n * 退避对所有 agent 一致);抛错按 `undefined` 回落并被吞掉,不掩盖原始失败。形状与分类链、\n * 执行体时序见 docs/feature/error-classification/architecture.md。\n */\n classifySendFailure?: SendFailureClassifier;\n}\n// ───────────────────────── Agent Ensure / Installer ─────────────────────────\n/**\n * Agent 安装身份:纯数据,进 configHash 与 `run.json`。\n * 与 sandbox case 身份正交——改 Agent 版本不重建任务 BuildKey。\n * 见 docs/feature/adapters/architecture/agent-ensure.md。\n */\nexport interface AgentIdentity {\n /** Agent 名(与 adapter 显示名对齐,如 `codex`)。 */\n readonly agent: string;\n /** 精确版本;禁止 `latest` 等无钉死取值——无稳定身份不参与可携带结果。 */\n readonly version: string;\n /** 配方修订(内置用数字字符串,自定义可用任意稳定串)。 */\n readonly revision: string;\n}\n/**\n * 三种安装模式,全部显式;失败后不允许在模式之间静默降级。\n * - `staged`:制品在题面网络外准备,经文件 API 送入(内置默认)。\n * - `sandbox-network`:显式声明用沙箱内网络安装。\n * - `verify-only`:只接受预装命中;探测失败立即 errored。\n */\nexport type AgentInstallMode = \"staged\" | \"sandbox-network\" | \"verify-only\";\n/** staged payload 的完成态目标平台;Linux 的 libc 是身份的一部分,不能留成缺值。 */\nexport type AgentArtifactPlatform = {\n readonly _tag: \"Linux\";\n readonly os: \"linux\";\n readonly arch: string;\n readonly libc: \"gnu\" | \"musl\";\n} | {\n readonly _tag: \"Darwin\";\n readonly os: \"darwin\";\n readonly arch: string;\n} | {\n readonly _tag: \"Windows\";\n readonly os: \"windows\";\n readonly arch: string;\n};\n/**\n * Run 级准备好的锁定制品。digest + 实际 platform 只属于运行 provenance,不进入\n * configHash 或 Eval fingerprint;静态 installer identity/revision 与 ProviderPlan target 才是配置身份。\n * 安装时经主 Sandbox 文件 API 上传,不依赖题面网络。\n */\nexport interface AgentStagedArtifact {\n readonly platform: AgentArtifactPlatform;\n /** digest-backed 不可变内容句柄;不向 Attempt 泄露宿主路径。 */\n readonly content: RegisteredSandboxContent;\n /** 沙箱内落点,位于 workdir 外的 Agent 自有目录。 */\n readonly targetPath: string;\n /** 安装协议必须显式,不从缺失字段推断历史默认。 */\n readonly install: AgentArtifactInstallShape;\n}\nexport type AgentArtifactInstallShape = {\n readonly kind: \"npm-tarball\";\n} | {\n readonly kind: \"self-contained\";\n readonly binPath: string;\n};\n/** `agent.ensure` 的取值:区分 probe 命中与本次安装。 */\nexport type AgentEnsureOutcome = \"hit\" | \"installed\";\n/**\n * Adapter 声明的纯 Ensure 义务:身份与只读 probe。probe 的非零退出是未命中,不是\n * Agent 任务失败;Runner 决定是否交给 identity 精确匹配的 installer。\n */\nexport interface AgentEnsure {\n readonly identity: AgentIdentity;\n readonly probe: StableSandboxCommand;\n /** Human-only transient labels; omitted adapters do not add an ensure detail. */\n readonly progress?: {\n readonly checking?: string;\n readonly ready?: string;\n };\n}\nexport interface AgentInstallContext {\n readonly identity: AgentIdentity;\n readonly targetPlatform: AgentArtifactPlatform;\n readonly signal: AbortSignal;\n progress(update: ProgressUpdate): void;\n}\nexport interface AgentArtifactContext {\n readonly targetPlatform: AgentArtifactPlatform;\n readonly signal: AbortSignal;\n}\nexport interface StagedAgentInstallContext extends AgentInstallContext {\n readonly artifact: AgentStagedArtifact;\n}\ninterface AgentInstallerBase {\n /** 必须与某条 ensure identity 完全相同;Runner 不做近似匹配。 */\n readonly identity: AgentIdentity;\n /** 支持的目标平台键(`linux-x64-gnu` 等);省略表示 installer 自己支持所有平台。 */\n readonly platforms?: readonly string[];\n /** Human-only transient label; omitted installers do not add an install detail. */\n readonly progress?: {\n readonly installing?: string;\n };\n /** @internal 只由内置 cache-safe staged installer 写入。 */\n readonly [AGENT_DOCKERFILE_CACHE_SAFE]?: true;\n}\nexport type AgentInstaller = (AgentInstallerBase & {\n readonly installMode: \"staged\";\n prepareArtifact(context: AgentArtifactContext): MaybePromise<AgentStagedArtifact>;\n install(sandbox: SandboxCommandTarget, context: StagedAgentInstallContext): MaybePromise<void>;\n}) | (AgentInstallerBase & {\n readonly installMode: \"sandbox-network\";\n readonly prepareArtifact?: never;\n install(sandbox: SandboxCommandTarget, context: AgentInstallContext): MaybePromise<void>;\n}) | (AgentInstallerBase & {\n readonly installMode: \"verify-only\";\n readonly prepareArtifact?: never;\n readonly install?: never;\n});\n/** 在 NiceEval 管理的 Sandbox 内驱动 CLI 的 Agent。 */\nexport interface SandboxAgent extends AgentBase {\n readonly kind: \"sandbox\";\n /** Adapter-owned command-only contribution to the unified Sandbox action DAG. */\n readonly sandbox?: SandboxLayer<\"command-only\">;\n /** Runner 在 author layers 后、before 前按声明顺序执行。 */\n readonly ensure: readonly AgentEnsure[];\n /** 官方或第三方随 adapter 提供的安装层;仅 identity 精确匹配时可接手。 */\n readonly installers: readonly AgentInstaller[];\n before?: AgentBefore;\n tracing?: AgentTracing;\n send(input: TurnInput, ctx: SandboxAgentSendContext): Promise<Turn>;\n after?: AgentAfter;\n}\n/** 由 runner 直接调用函数、SDK 或服务端点的 Agent;目标进程可以在本地或远端。 */\nexport interface DirectAgent extends AgentBase {\n readonly kind: \"direct\";\n before?: DirectAgentBefore;\n tracing?: Omit<AgentTracing, \"configure\">;\n send(input: TurnInput, ctx: AgentSendContext): Promise<Turn>;\n after?: DirectAgentAfter;\n}\n/** 注册表里的 Agent;判别值描述驱动拓扑,不描述目标所在位置。 */\nexport type Agent = SandboxAgent | DirectAgent;\n/** `defineSandboxAgent()` 的入参形状(见 src/define.ts)——`kind: \"sandbox\"` 由 define 固定填入,不由用户声明。 */\nexport interface SandboxAgentDef {\n /** agent 的显示名/标识,原样进入 `Agent.name`——不是注册表查找 key,只用于展示、结果归属与去重指纹。 */\n name: string;\n /** 该 Adapter 的常态证据覆盖声明;完整采集可用 `completeEvidenceCoverage`。 */\n evidenceCoverage: EvidenceCoverage;\n /** Adapter 自有的 Sandbox action;Agent 不能选择或贡献 template。 */\n sandbox?: SandboxLayer<\"command-only\">;\n /** 单条或数组都按声明顺序规范化为 Agent layer。 */\n ensure: AgentEnsure | readonly AgentEnsure[];\n /**\n * 配对安装层;省略表示这个 adapter 只提供 probe 协议,未命中时 Runner 在\n * `agent.ensure` 点名缺失的精确 identity。工厂会把省略规范化成空数组。\n */\n installers?: readonly AgentInstaller[];\n /**\n * 每条 Attempt 一次(不是每轮 send 一次):写 config.toml / 鉴权配置(model/base/auth 等\n * 本 Attempt 内不变的运行配置)。CLI 的 probe、安装与复检属于 `agent.ensure`。运行器在 Sandbox\n * 备好(layer prepare/ensure/baseline 之后)、第一次 send 前调用一次,不返回值。\n */\n before?: AgentBefore;\n /** OTLP 导出配置:Sandbox 里怎么让 CLI 把 trace 发到 endpoint(env / 配置文件),从 before 拆出。 */\n tracing?: AgentTracing;\n /** 原生 span → canonical 的薄 mapper;省略走通用 heuristic。只影响瀑布图。 */\n spanMapper?: SpanMapper;\n /** 每轮一次:跑 prompt(fresh / resume)+ 解析成 events。作者 callback 保持 Promise ABI。 */\n send(input: TurnInput, ctx: SandboxAgentSendContext): Promise<Turn>;\n /** 可选 send 执行失败分类器:见 `Agent.classifySendFailure`。 */\n classifySendFailure?: SendFailureClassifier;\n /** Sandbox 销毁前的清理,当且仅当本 attempt 走到过 `before` 时点才执行(`before` 抛错不豁免),\n * 在 finally 里跑一次。 */\n after?: AgentAfter;\n}\n/** `defineAgent()` 的入参形状(见 src/define.ts)——`kind: \"direct\"` 由 define 固定填入,不由用户声明。 */\nexport interface DirectAgentDef {\n /** agent 的显示名/标识,原样进入 `Agent.name`——不是注册表查找 key,只用于展示、结果归属与去重指纹。 */\n name: string;\n /** 该 Adapter 的常态证据覆盖声明;完整采集可用 `completeEvidenceCoverage`。 */\n evidenceCoverage: EvidenceCoverage;\n /**\n * 每个 attempt 一次。Direct Agent 不接收 Sandbox;常用于建立连接、鉴权等一次性准备。\n */\n before?: DirectAgentBefore;\n /** OTLP 导出配置:被测对象怎么把 trace 发到 endpoint(env-based 注入)。 */\n tracing?: Omit<AgentTracing, \"configure\">;\n /** 原生 span → canonical 的薄 mapper;省略走通用 heuristic。只影响瀑布图。 */\n spanMapper?: SpanMapper;\n /** 每轮一次:把一轮 prompt 直接发给函数、SDK 或服务端点,解析响应成 events。作者 callback 保持 Promise ABI。 */\n send(input: TurnInput, ctx: AgentSendContext): Promise<Turn>;\n /** 可选 send 执行失败分类器:见 `Agent.classifySendFailure`。 */\n classifySendFailure?: SendFailureClassifier;\n /** 运行结束前的清理,当且仅当本 attempt 走到过 `before` 时点才执行(`before` 抛错不豁免),\n * 在 finally 里跑一次。 */\n after?: DirectAgentAfter;\n}\n"]}