oh-my-knowledge 0.30.0 → 0.32.0

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 (1084) hide show
  1. package/README.md +33 -672
  2. package/README.zh.md +45 -684
  3. package/dist/analysis/coverage-analyzer.d.ts.map +1 -0
  4. package/dist/analysis/coverage-analyzer.js.map +1 -0
  5. package/dist/analysis/failure-clusterer.d.ts.map +1 -0
  6. package/dist/analysis/failure-clusterer.js.map +1 -0
  7. package/dist/analysis/gap-analyzer.d.ts +121 -0
  8. package/dist/analysis/gap-analyzer.d.ts.map +1 -0
  9. package/dist/analysis/gap-analyzer.js +471 -0
  10. package/dist/analysis/gap-analyzer.js.map +1 -0
  11. package/dist/analysis/hedging-classifier.d.ts.map +1 -0
  12. package/dist/analysis/hedging-classifier.js.map +1 -0
  13. package/dist/analysis/report-diagnostics.d.ts +34 -0
  14. package/dist/analysis/report-diagnostics.d.ts.map +1 -0
  15. package/dist/analysis/report-diagnostics.js +753 -0
  16. package/dist/analysis/report-diagnostics.js.map +1 -0
  17. package/dist/analysis/sample-diagnostics.d.ts.map +1 -0
  18. package/dist/analysis/sample-diagnostics.js.map +1 -0
  19. package/dist/analysis/saturation.d.ts.map +1 -0
  20. package/dist/analysis/saturation.js.map +1 -0
  21. package/dist/authoring/evolver.d.ts +102 -0
  22. package/dist/authoring/evolver.d.ts.map +1 -0
  23. package/dist/authoring/evolver.js +681 -0
  24. package/dist/authoring/evolver.js.map +1 -0
  25. package/dist/authoring/generator.d.ts +38 -0
  26. package/dist/authoring/generator.d.ts.map +1 -0
  27. package/dist/authoring/generator.js +720 -0
  28. package/dist/authoring/generator.js.map +1 -0
  29. package/dist/authoring/sample-fixer.d.ts +51 -0
  30. package/dist/authoring/sample-fixer.d.ts.map +1 -0
  31. package/dist/authoring/sample-fixer.js +213 -0
  32. package/dist/authoring/sample-fixer.js.map +1 -0
  33. package/dist/cli/commands/doctor.d.ts +26 -0
  34. package/dist/cli/commands/doctor.d.ts.map +1 -0
  35. package/dist/cli/commands/doctor.js +302 -0
  36. package/dist/cli/commands/doctor.js.map +1 -0
  37. package/dist/cli/commands/eval/gold/compare.d.ts +17 -0
  38. package/dist/cli/commands/eval/gold/compare.d.ts.map +1 -0
  39. package/dist/cli/commands/eval/gold/compare.js +91 -0
  40. package/dist/cli/commands/eval/gold/compare.js.map +1 -0
  41. package/dist/cli/commands/eval/gold/index.d.ts +9 -0
  42. package/dist/cli/commands/eval/gold/index.d.ts.map +1 -0
  43. package/dist/cli/commands/eval/gold/index.js +34 -0
  44. package/dist/cli/commands/eval/gold/index.js.map +1 -0
  45. package/dist/cli/commands/eval/gold/init.d.ts +11 -0
  46. package/dist/cli/commands/eval/gold/init.d.ts.map +1 -0
  47. package/dist/cli/commands/eval/gold/init.js +51 -0
  48. package/dist/cli/commands/eval/gold/init.js.map +1 -0
  49. package/dist/cli/commands/eval/gold/validate.d.ts +12 -0
  50. package/dist/cli/commands/eval/gold/validate.d.ts.map +1 -0
  51. package/dist/cli/commands/eval/gold/validate.js +46 -0
  52. package/dist/cli/commands/eval/gold/validate.js.map +1 -0
  53. package/dist/cli/commands/eval/index.d.ts +54 -0
  54. package/dist/cli/commands/eval/index.d.ts.map +1 -0
  55. package/dist/cli/commands/eval/index.js +465 -0
  56. package/dist/cli/commands/eval/index.js.map +1 -0
  57. package/dist/cli/commands/evolve.d.ts +37 -0
  58. package/dist/cli/commands/evolve.d.ts.map +1 -0
  59. package/dist/cli/commands/evolve.js +283 -0
  60. package/dist/cli/commands/evolve.js.map +1 -0
  61. package/dist/cli/commands/init.d.ts +16 -0
  62. package/dist/cli/commands/init.d.ts.map +1 -0
  63. package/dist/cli/commands/init.js +152 -0
  64. package/dist/cli/commands/init.js.map +1 -0
  65. package/dist/cli/commands/observe/inbox.d.ts +23 -0
  66. package/dist/cli/commands/observe/inbox.d.ts.map +1 -0
  67. package/dist/cli/commands/observe/inbox.js +260 -0
  68. package/dist/cli/commands/observe/inbox.js.map +1 -0
  69. package/dist/cli/commands/observe/index.d.ts +22 -0
  70. package/dist/cli/commands/observe/index.d.ts.map +1 -0
  71. package/dist/cli/commands/observe/index.js +118 -0
  72. package/dist/cli/commands/observe/index.js.map +1 -0
  73. package/dist/cli/commands/observe/ingest.d.ts +13 -0
  74. package/dist/cli/commands/observe/ingest.d.ts.map +1 -0
  75. package/dist/cli/commands/observe/ingest.js +71 -0
  76. package/dist/cli/commands/observe/ingest.js.map +1 -0
  77. package/dist/cli/commands/observe/show.d.ts +13 -0
  78. package/dist/cli/commands/observe/show.d.ts.map +1 -0
  79. package/dist/cli/commands/observe/show.js +49 -0
  80. package/dist/cli/commands/observe/show.js.map +1 -0
  81. package/dist/cli/commands/sample.d.ts +38 -0
  82. package/dist/cli/commands/sample.d.ts.map +1 -0
  83. package/dist/cli/commands/sample.js +487 -0
  84. package/dist/cli/commands/sample.js.map +1 -0
  85. package/dist/cli/commands/studio.d.ts +23 -0
  86. package/dist/cli/commands/studio.d.ts.map +1 -0
  87. package/dist/cli/commands/studio.js +158 -0
  88. package/dist/cli/commands/studio.js.map +1 -0
  89. package/dist/cli/index.d.ts.map +1 -0
  90. package/dist/cli/index.js +29 -0
  91. package/dist/cli/index.js.map +1 -0
  92. package/dist/cli/lib/cli-exit.d.ts +16 -0
  93. package/dist/cli/lib/cli-exit.d.ts.map +1 -0
  94. package/dist/cli/lib/cli-exit.js +20 -0
  95. package/dist/cli/lib/cli-exit.js.map +1 -0
  96. package/dist/cli/lib/cmd-flags.d.ts +224 -0
  97. package/dist/cli/lib/cmd-flags.d.ts.map +1 -0
  98. package/dist/cli/lib/cmd-flags.js +46 -0
  99. package/dist/cli/lib/cmd-flags.js.map +1 -0
  100. package/dist/cli/lib/i18n-dict/common.d.ts +4 -0
  101. package/dist/cli/lib/i18n-dict/common.d.ts.map +1 -0
  102. package/dist/cli/lib/i18n-dict/common.js +67 -0
  103. package/dist/cli/lib/i18n-dict/common.js.map +1 -0
  104. package/dist/cli/lib/i18n-dict/evolve.d.ts +4 -0
  105. package/dist/cli/lib/i18n-dict/evolve.d.ts.map +1 -0
  106. package/dist/cli/lib/i18n-dict/evolve.js +43 -0
  107. package/dist/cli/lib/i18n-dict/evolve.js.map +1 -0
  108. package/dist/cli/lib/i18n-dict/gen.d.ts +4 -0
  109. package/dist/cli/lib/i18n-dict/gen.d.ts.map +1 -0
  110. package/dist/cli/lib/i18n-dict/gen.js +63 -0
  111. package/dist/cli/lib/i18n-dict/gen.js.map +1 -0
  112. package/dist/cli/lib/i18n-dict/help.d.ts +4 -0
  113. package/dist/cli/lib/i18n-dict/help.d.ts.map +1 -0
  114. package/dist/cli/lib/i18n-dict/help.js +281 -0
  115. package/dist/cli/lib/i18n-dict/help.js.map +1 -0
  116. package/dist/cli/lib/i18n-dict/init.d.ts +4 -0
  117. package/dist/cli/lib/i18n-dict/init.d.ts.map +1 -0
  118. package/dist/cli/lib/i18n-dict/init.js +27 -0
  119. package/dist/cli/lib/i18n-dict/init.js.map +1 -0
  120. package/dist/cli/lib/i18n-dict/run.d.ts +4 -0
  121. package/dist/cli/lib/i18n-dict/run.d.ts.map +1 -0
  122. package/dist/cli/lib/i18n-dict/run.js +143 -0
  123. package/dist/cli/lib/i18n-dict/run.js.map +1 -0
  124. package/dist/cli/lib/i18n-dict/types.d.ts +5 -0
  125. package/dist/cli/lib/i18n-dict/types.d.ts.map +1 -0
  126. package/dist/cli/lib/i18n-dict/types.js +2 -0
  127. package/dist/cli/lib/i18n-dict/types.js.map +1 -0
  128. package/dist/cli/lib/i18n-dict.d.ts +63 -0
  129. package/dist/cli/lib/i18n-dict.d.ts.map +1 -0
  130. package/dist/cli/lib/i18n-dict.js +67 -0
  131. package/dist/cli/lib/i18n-dict.js.map +1 -0
  132. package/dist/cli/lib/i18n.d.ts +29 -0
  133. package/dist/cli/lib/i18n.d.ts.map +1 -0
  134. package/dist/cli/lib/i18n.js +58 -0
  135. package/dist/cli/lib/i18n.js.map +1 -0
  136. package/dist/cli/lib/parse-run-config/judge-models.d.ts +24 -0
  137. package/dist/cli/lib/parse-run-config/judge-models.d.ts.map +1 -0
  138. package/dist/cli/lib/parse-run-config/judge-models.js +55 -0
  139. package/dist/cli/lib/parse-run-config/judge-models.js.map +1 -0
  140. package/dist/cli/lib/parse-run-config/samples-discovery.d.ts +19 -0
  141. package/dist/cli/lib/parse-run-config/samples-discovery.d.ts.map +1 -0
  142. package/dist/cli/lib/parse-run-config/samples-discovery.js +53 -0
  143. package/dist/cli/lib/parse-run-config/samples-discovery.js.map +1 -0
  144. package/dist/cli/lib/parse-run-config/variant-resolution.d.ts +18 -0
  145. package/dist/cli/lib/parse-run-config/variant-resolution.d.ts.map +1 -0
  146. package/dist/cli/lib/parse-run-config/variant-resolution.js +59 -0
  147. package/dist/cli/lib/parse-run-config/variant-resolution.js.map +1 -0
  148. package/dist/cli/lib/parse-run-config.d.ts +93 -0
  149. package/dist/cli/lib/parse-run-config.d.ts.map +1 -0
  150. package/dist/cli/lib/parse-run-config.js +157 -0
  151. package/dist/cli/lib/parse-run-config.js.map +1 -0
  152. package/dist/cli/lib/progress.d.ts.map +1 -0
  153. package/dist/cli/lib/progress.js.map +1 -0
  154. package/dist/cli/lib/resolve-skill-input.d.ts +8 -0
  155. package/dist/cli/lib/resolve-skill-input.d.ts.map +1 -0
  156. package/dist/cli/lib/resolve-skill-input.js +38 -0
  157. package/dist/cli/lib/resolve-skill-input.js.map +1 -0
  158. package/dist/cli/lib/run-tally.d.ts +18 -0
  159. package/dist/cli/lib/run-tally.d.ts.map +1 -0
  160. package/dist/cli/lib/run-tally.js.map +1 -0
  161. package/dist/cli/lib/shared.d.ts +12 -0
  162. package/dist/cli/lib/shared.d.ts.map +1 -0
  163. package/dist/cli/lib/shared.js +26 -0
  164. package/dist/cli/lib/shared.js.map +1 -0
  165. package/dist/cli/lib/update-check.d.ts +9 -0
  166. package/dist/cli/lib/update-check.d.ts.map +1 -0
  167. package/dist/cli/lib/update-check.js +113 -0
  168. package/dist/cli/lib/update-check.js.map +1 -0
  169. package/dist/cli/oclif/base-command.d.ts +8 -0
  170. package/dist/cli/oclif/base-command.d.ts.map +1 -0
  171. package/dist/cli/oclif/base-command.js +27 -0
  172. package/dist/cli/oclif/base-command.js.map +1 -0
  173. package/dist/cli/oclif/help.d.ts +16 -0
  174. package/dist/cli/oclif/help.d.ts.map +1 -0
  175. package/dist/cli/oclif/help.js +33 -0
  176. package/dist/cli/oclif/help.js.map +1 -0
  177. package/dist/cli/oclif/i18n.d.ts +16 -0
  178. package/dist/cli/oclif/i18n.d.ts.map +1 -0
  179. package/dist/cli/oclif/i18n.js +62 -0
  180. package/dist/cli/oclif/i18n.js.map +1 -0
  181. package/dist/cli/oclif/parsers.d.ts +10 -0
  182. package/dist/cli/oclif/parsers.d.ts.map +1 -0
  183. package/dist/cli/oclif/parsers.js +102 -0
  184. package/dist/cli/oclif/parsers.js.map +1 -0
  185. package/dist/cli/oclif/projection.d.ts +23 -0
  186. package/dist/cli/oclif/projection.d.ts.map +1 -0
  187. package/dist/cli/oclif/projection.js +52 -0
  188. package/dist/cli/oclif/projection.js.map +1 -0
  189. package/dist/cli/oclif/run.d.ts +2 -0
  190. package/dist/cli/oclif/run.d.ts.map +1 -0
  191. package/dist/cli/oclif/run.js +11 -0
  192. package/dist/cli/oclif/run.js.map +1 -0
  193. package/dist/diagnosis/observe-mapper.d.ts +68 -0
  194. package/dist/diagnosis/observe-mapper.d.ts.map +1 -0
  195. package/dist/diagnosis/observe-mapper.js +276 -0
  196. package/dist/diagnosis/observe-mapper.js.map +1 -0
  197. package/dist/diagnosis/observe-producer.d.ts +9 -0
  198. package/dist/diagnosis/observe-producer.d.ts.map +1 -0
  199. package/dist/diagnosis/observe-producer.js +199 -0
  200. package/dist/diagnosis/observe-producer.js.map +1 -0
  201. package/dist/diagnosis/studio-projection.d.ts +7 -0
  202. package/dist/diagnosis/studio-projection.d.ts.map +1 -0
  203. package/dist/diagnosis/studio-projection.js +84 -0
  204. package/dist/diagnosis/studio-projection.js.map +1 -0
  205. package/dist/diagnosis/types.d.ts +4 -0
  206. package/dist/diagnosis/types.d.ts.map +1 -0
  207. package/dist/diagnosis/types.js +20 -0
  208. package/dist/diagnosis/types.js.map +1 -0
  209. package/dist/doctor/fixer.d.ts +12 -0
  210. package/dist/doctor/fixer.d.ts.map +1 -0
  211. package/dist/doctor/fixer.js +326 -0
  212. package/dist/doctor/fixer.js.map +1 -0
  213. package/dist/doctor/health/builtin-dimensions.d.ts.map +1 -0
  214. package/dist/doctor/health/builtin-dimensions.js.map +1 -0
  215. package/dist/doctor/health/composer.d.ts.map +1 -0
  216. package/dist/doctor/health/composer.js +296 -0
  217. package/dist/doctor/health/composer.js.map +1 -0
  218. package/dist/doctor/health/dimension-registry.d.ts.map +1 -0
  219. package/dist/doctor/health/dimension-registry.js.map +1 -0
  220. package/dist/doctor/health/dimension-spec.d.ts +47 -0
  221. package/dist/doctor/health/dimension-spec.d.ts.map +1 -0
  222. package/dist/doctor/health/dimension-spec.js.map +1 -0
  223. package/dist/doctor/health/parser.d.ts +26 -0
  224. package/dist/doctor/health/parser.d.ts.map +1 -0
  225. package/dist/doctor/health/parser.js +228 -0
  226. package/dist/doctor/health/parser.js.map +1 -0
  227. package/dist/doctor/health/prompt-builder.d.ts +3 -0
  228. package/dist/doctor/health/prompt-builder.d.ts.map +1 -0
  229. package/dist/doctor/health/prompt-builder.js +2 -0
  230. package/dist/doctor/health/prompt-builder.js.map +1 -0
  231. package/dist/doctor/health/register.d.ts.map +1 -0
  232. package/dist/doctor/health/register.js.map +1 -0
  233. package/dist/doctor/index.d.ts.map +1 -0
  234. package/dist/doctor/index.js +246 -0
  235. package/dist/doctor/index.js.map +1 -0
  236. package/dist/doctor/messages.d.ts +21 -0
  237. package/dist/doctor/messages.d.ts.map +1 -0
  238. package/dist/doctor/messages.js +217 -0
  239. package/dist/doctor/messages.js.map +1 -0
  240. package/dist/doctor/preflight.d.ts.map +1 -0
  241. package/dist/doctor/preflight.js.map +1 -0
  242. package/dist/doctor/renderer.d.ts +17 -0
  243. package/dist/doctor/renderer.d.ts.map +1 -0
  244. package/dist/doctor/renderer.js +122 -0
  245. package/dist/doctor/renderer.js.map +1 -0
  246. package/dist/doctor/rules.d.ts.map +1 -0
  247. package/dist/doctor/rules.js +289 -0
  248. package/dist/doctor/rules.js.map +1 -0
  249. package/dist/eval-core/bootstrap.d.ts.map +1 -0
  250. package/dist/eval-core/bootstrap.js.map +1 -0
  251. package/dist/eval-core/cache.d.ts.map +1 -0
  252. package/dist/eval-core/cache.js.map +1 -0
  253. package/dist/eval-core/comparability.d.ts.map +1 -0
  254. package/dist/eval-core/comparability.js.map +1 -0
  255. package/dist/eval-core/dependency-checker.d.ts +37 -0
  256. package/dist/eval-core/dependency-checker.d.ts.map +1 -0
  257. package/dist/eval-core/dependency-checker.js.map +1 -0
  258. package/dist/eval-core/evaluation-execution.d.ts.map +1 -0
  259. package/dist/eval-core/evaluation-execution.js.map +1 -0
  260. package/dist/eval-core/evaluation-job.d.ts.map +1 -0
  261. package/dist/eval-core/evaluation-job.js.map +1 -0
  262. package/dist/eval-core/evaluation-reporting.d.ts.map +1 -0
  263. package/dist/eval-core/evaluation-reporting.js.map +1 -0
  264. package/dist/eval-core/execution-strategy.d.ts.map +1 -0
  265. package/dist/eval-core/execution-strategy.js +165 -0
  266. package/dist/eval-core/execution-strategy.js.map +1 -0
  267. package/dist/eval-core/fact-checker.d.ts.map +1 -0
  268. package/dist/eval-core/fact-checker.js.map +1 -0
  269. package/dist/eval-core/layer-gates.d.ts.map +1 -0
  270. package/dist/eval-core/layer-gates.js.map +1 -0
  271. package/dist/eval-core/mock-hook.cjs +215 -0
  272. package/dist/eval-core/mocks-runtime.d.ts +100 -0
  273. package/dist/eval-core/mocks-runtime.d.ts.map +1 -0
  274. package/dist/eval-core/mocks-runtime.js +559 -0
  275. package/dist/eval-core/mocks-runtime.js.map +1 -0
  276. package/dist/eval-core/schema.d.ts.map +1 -0
  277. package/dist/eval-core/schema.js.map +1 -0
  278. package/dist/eval-core/statistics.d.ts.map +1 -0
  279. package/dist/eval-core/statistics.js.map +1 -0
  280. package/dist/eval-core/task-planner.d.ts.map +1 -0
  281. package/dist/eval-core/task-planner.js.map +1 -0
  282. package/dist/eval-core/verdict.d.ts.map +1 -0
  283. package/dist/eval-core/verdict.js.map +1 -0
  284. package/dist/eval-workflows/batch-evaluation-workflow.d.ts.map +1 -0
  285. package/dist/eval-workflows/batch-evaluation-workflow.js.map +1 -0
  286. package/dist/eval-workflows/evaluation-pipeline/preflight-warnings.d.ts +22 -0
  287. package/dist/eval-workflows/evaluation-pipeline/preflight-warnings.d.ts.map +1 -0
  288. package/dist/eval-workflows/evaluation-pipeline/preflight-warnings.js +67 -0
  289. package/dist/eval-workflows/evaluation-pipeline/preflight-warnings.js.map +1 -0
  290. package/dist/eval-workflows/evaluation-pipeline/report-finalize.d.ts +32 -0
  291. package/dist/eval-workflows/evaluation-pipeline/report-finalize.d.ts.map +1 -0
  292. package/dist/eval-workflows/evaluation-pipeline/report-finalize.js +57 -0
  293. package/dist/eval-workflows/evaluation-pipeline/report-finalize.js.map +1 -0
  294. package/dist/eval-workflows/evaluation-pipeline/run-state.d.ts +60 -0
  295. package/dist/eval-workflows/evaluation-pipeline/run-state.d.ts.map +1 -0
  296. package/dist/eval-workflows/evaluation-pipeline/run-state.js +88 -0
  297. package/dist/eval-workflows/evaluation-pipeline/run-state.js.map +1 -0
  298. package/dist/eval-workflows/evaluation-pipeline/test-set-hash.d.ts +21 -0
  299. package/dist/eval-workflows/evaluation-pipeline/test-set-hash.d.ts.map +1 -0
  300. package/dist/eval-workflows/evaluation-pipeline/test-set-hash.js +46 -0
  301. package/dist/eval-workflows/evaluation-pipeline/test-set-hash.js.map +1 -0
  302. package/dist/eval-workflows/evaluation-pipeline.d.ts +97 -0
  303. package/dist/eval-workflows/evaluation-pipeline.d.ts.map +1 -0
  304. package/dist/eval-workflows/evaluation-pipeline.js +179 -0
  305. package/dist/eval-workflows/evaluation-pipeline.js.map +1 -0
  306. package/dist/eval-workflows/evaluation-preparation.d.ts.map +1 -0
  307. package/dist/eval-workflows/evaluation-preparation.js.map +1 -0
  308. package/dist/eval-workflows/messages.d.ts +4 -0
  309. package/dist/eval-workflows/messages.d.ts.map +1 -0
  310. package/dist/eval-workflows/messages.js +30 -0
  311. package/dist/eval-workflows/messages.js.map +1 -0
  312. package/dist/eval-workflows/run-evaluation.d.ts.map +1 -0
  313. package/dist/eval-workflows/run-evaluation.js +528 -0
  314. package/dist/eval-workflows/run-evaluation.js.map +1 -0
  315. package/dist/executors/anthropic-api.d.ts.map +1 -0
  316. package/dist/executors/anthropic-api.js.map +1 -0
  317. package/dist/executors/claude-cli.d.ts.map +1 -0
  318. package/dist/executors/claude-cli.js +181 -0
  319. package/dist/executors/claude-cli.js.map +1 -0
  320. package/dist/executors/claude-sdk-trace.d.ts.map +1 -0
  321. package/dist/executors/claude-sdk-trace.js.map +1 -0
  322. package/dist/executors/claude-sdk.d.ts.map +1 -0
  323. package/dist/executors/claude-sdk.js.map +1 -0
  324. package/dist/executors/codex-cli-trace.d.ts.map +1 -0
  325. package/dist/executors/codex-cli-trace.js.map +1 -0
  326. package/dist/executors/codex-cli.d.ts.map +1 -0
  327. package/dist/executors/codex-cli.js.map +1 -0
  328. package/dist/executors/codex-sdk.d.ts.map +1 -0
  329. package/dist/executors/codex-sdk.js.map +1 -0
  330. package/dist/executors/gemini.d.ts.map +1 -0
  331. package/dist/executors/gemini.js.map +1 -0
  332. package/dist/executors/index.d.ts.map +1 -0
  333. package/dist/executors/index.js.map +1 -0
  334. package/dist/executors/openai-api.d.ts.map +1 -0
  335. package/dist/executors/openai-api.js.map +1 -0
  336. package/dist/executors/runtime-fingerprint.d.ts.map +1 -0
  337. package/dist/executors/runtime-fingerprint.js.map +1 -0
  338. package/dist/executors/script.d.ts.map +1 -0
  339. package/dist/executors/script.js.map +1 -0
  340. package/dist/executors/shared.d.ts +194 -0
  341. package/dist/executors/shared.d.ts.map +1 -0
  342. package/dist/executors/shared.js +253 -0
  343. package/dist/executors/shared.js.map +1 -0
  344. package/dist/grading/assertions.d.ts.map +1 -0
  345. package/dist/grading/assertions.js.map +1 -0
  346. package/dist/grading/debias-validate.d.ts.map +1 -0
  347. package/dist/grading/debias-validate.js.map +1 -0
  348. package/dist/grading/diagnostic.d.ts.map +1 -0
  349. package/dist/grading/diagnostic.js.map +1 -0
  350. package/dist/grading/gold-cli.d.ts.map +1 -0
  351. package/dist/grading/gold-cli.js.map +1 -0
  352. package/dist/grading/gold-dataset.d.ts.map +1 -0
  353. package/dist/grading/gold-dataset.js.map +1 -0
  354. package/dist/grading/human-gold.d.ts.map +1 -0
  355. package/dist/grading/human-gold.js.map +1 -0
  356. package/dist/grading/index.d.ts.map +1 -0
  357. package/dist/grading/index.js.map +1 -0
  358. package/dist/grading/judge.d.ts.map +1 -0
  359. package/dist/grading/judge.js.map +1 -0
  360. package/dist/grading/layered-scores.d.ts.map +1 -0
  361. package/dist/grading/layered-scores.js.map +1 -0
  362. package/dist/inputs/eval-config.d.ts.map +1 -0
  363. package/dist/inputs/eval-config.js.map +1 -0
  364. package/dist/inputs/load-samples.d.ts.map +1 -0
  365. package/dist/inputs/load-samples.js.map +1 -0
  366. package/dist/inputs/mcp-resolver.d.ts.map +1 -0
  367. package/dist/inputs/mcp-resolver.js.map +1 -0
  368. package/dist/inputs/skill-loader.d.ts.map +1 -0
  369. package/dist/inputs/skill-loader.js +261 -0
  370. package/dist/inputs/skill-loader.js.map +1 -0
  371. package/dist/inputs/url-fetcher.d.ts.map +1 -0
  372. package/dist/inputs/url-fetcher.js.map +1 -0
  373. package/dist/observability/experience-frontmatter.d.ts +37 -0
  374. package/dist/observability/experience-frontmatter.d.ts.map +1 -0
  375. package/dist/observability/experience-frontmatter.js +100 -0
  376. package/dist/observability/experience-frontmatter.js.map +1 -0
  377. package/dist/observability/experience.d.ts +23 -0
  378. package/dist/observability/experience.d.ts.map +1 -0
  379. package/dist/observability/experience.js +3438 -0
  380. package/dist/observability/experience.js.map +1 -0
  381. package/dist/observability/feedback-matchers.d.ts +37 -0
  382. package/dist/observability/feedback-matchers.d.ts.map +1 -0
  383. package/dist/observability/feedback-matchers.js +232 -0
  384. package/dist/observability/feedback-matchers.js.map +1 -0
  385. package/dist/observability/feedback-projection.d.ts +21 -0
  386. package/dist/observability/feedback-projection.d.ts.map +1 -0
  387. package/dist/observability/feedback-projection.js +20 -0
  388. package/dist/observability/feedback-projection.js.map +1 -0
  389. package/dist/observability/inbox-view-model.d.ts +43 -0
  390. package/dist/observability/inbox-view-model.d.ts.map +1 -0
  391. package/dist/observability/inbox-view-model.js +180 -0
  392. package/dist/observability/inbox-view-model.js.map +1 -0
  393. package/dist/observability/inbox.d.ts +28 -0
  394. package/dist/observability/inbox.d.ts.map +1 -0
  395. package/dist/observability/inbox.js +765 -0
  396. package/dist/observability/inbox.js.map +1 -0
  397. package/dist/observability/problem-patterns.d.ts +12 -0
  398. package/dist/observability/problem-patterns.d.ts.map +1 -0
  399. package/dist/observability/problem-patterns.js +244 -0
  400. package/dist/observability/problem-patterns.js.map +1 -0
  401. package/dist/observability/prompts/llm-enhanced-review.prompt.md +224 -0
  402. package/dist/observability/resolved-review.d.ts +37 -0
  403. package/dist/observability/resolved-review.d.ts.map +1 -0
  404. package/dist/observability/resolved-review.js +552 -0
  405. package/dist/observability/resolved-review.js.map +1 -0
  406. package/dist/observability/review-state.d.ts +20 -0
  407. package/dist/observability/review-state.d.ts.map +1 -0
  408. package/dist/observability/review-state.js +169 -0
  409. package/dist/observability/review-state.js.map +1 -0
  410. package/dist/observability/skill-chain-advisories.d.ts +16 -0
  411. package/dist/observability/skill-chain-advisories.d.ts.map +1 -0
  412. package/dist/observability/skill-chain-advisories.js +69 -0
  413. package/dist/observability/skill-chain-advisories.js.map +1 -0
  414. package/dist/observability/skill-chain.d.ts +7 -0
  415. package/dist/observability/skill-chain.d.ts.map +1 -0
  416. package/dist/observability/skill-chain.js +474 -0
  417. package/dist/observability/skill-chain.js.map +1 -0
  418. package/dist/observability/skill-health-analyzer.d.ts.map +1 -0
  419. package/dist/observability/skill-health-analyzer.js.map +1 -0
  420. package/dist/observability/soft-standards/constants.d.ts +5 -0
  421. package/dist/observability/soft-standards/constants.d.ts.map +1 -0
  422. package/dist/observability/soft-standards/constants.js +10 -0
  423. package/dist/observability/soft-standards/constants.js.map +1 -0
  424. package/dist/observability/soft-standards/index.d.ts +16 -0
  425. package/dist/observability/soft-standards/index.d.ts.map +1 -0
  426. package/dist/observability/soft-standards/index.js +15 -0
  427. package/dist/observability/soft-standards/index.js.map +1 -0
  428. package/dist/observability/soft-standards/llm-extractor.d.ts +6 -0
  429. package/dist/observability/soft-standards/llm-extractor.d.ts.map +1 -0
  430. package/dist/observability/soft-standards/llm-extractor.js +692 -0
  431. package/dist/observability/soft-standards/llm-extractor.js.map +1 -0
  432. package/dist/observability/soft-standards/runtime-evaluator.d.ts +12 -0
  433. package/dist/observability/soft-standards/runtime-evaluator.d.ts.map +1 -0
  434. package/dist/observability/soft-standards/runtime-evaluator.js +385 -0
  435. package/dist/observability/soft-standards/runtime-evaluator.js.map +1 -0
  436. package/dist/observability/soft-standards/skill-standards-store.d.ts +11 -0
  437. package/dist/observability/soft-standards/skill-standards-store.d.ts.map +1 -0
  438. package/dist/observability/soft-standards/skill-standards-store.js +158 -0
  439. package/dist/observability/soft-standards/skill-standards-store.js.map +1 -0
  440. package/dist/observability/soft-standards/types.d.ts +212 -0
  441. package/dist/observability/soft-standards/types.d.ts.map +1 -0
  442. package/dist/observability/soft-standards/types.js +2 -0
  443. package/dist/observability/soft-standards/types.js.map +1 -0
  444. package/dist/observability/text-signals.d.ts +14 -0
  445. package/dist/observability/text-signals.d.ts.map +1 -0
  446. package/dist/observability/text-signals.js +116 -0
  447. package/dist/observability/text-signals.js.map +1 -0
  448. package/dist/observability/trace-adapter.d.ts.map +1 -0
  449. package/dist/observability/trace-adapter.js.map +1 -0
  450. package/dist/observability/trace-attribution.d.ts +56 -0
  451. package/dist/observability/trace-attribution.d.ts.map +1 -0
  452. package/dist/observability/trace-attribution.js +200 -0
  453. package/dist/observability/trace-attribution.js.map +1 -0
  454. package/dist/observability/trace-segmenter.d.ts +52 -0
  455. package/dist/observability/trace-segmenter.d.ts.map +1 -0
  456. package/dist/observability/trace-segmenter.js +322 -0
  457. package/dist/observability/trace-segmenter.js.map +1 -0
  458. package/dist/observability/trace-source.d.ts +92 -0
  459. package/dist/observability/trace-source.d.ts.map +1 -0
  460. package/dist/observability/trace-source.js +502 -0
  461. package/dist/observability/trace-source.js.map +1 -0
  462. package/dist/renderer/html-renderer.d.ts.map +1 -0
  463. package/dist/renderer/html-renderer.js.map +1 -0
  464. package/dist/renderer/layout.d.ts.map +1 -0
  465. package/dist/renderer/layout.js.map +1 -0
  466. package/dist/renderer/observation-inbox/helpers.d.ts +14 -0
  467. package/dist/renderer/observation-inbox/helpers.d.ts.map +1 -0
  468. package/dist/renderer/observation-inbox/helpers.js +75 -0
  469. package/dist/renderer/observation-inbox/helpers.js.map +1 -0
  470. package/dist/renderer/observation-inbox/styles.d.ts +2 -0
  471. package/dist/renderer/observation-inbox/styles.d.ts.map +1 -0
  472. package/dist/renderer/observation-inbox/styles.js +5184 -0
  473. package/dist/renderer/observation-inbox/styles.js.map +1 -0
  474. package/dist/renderer/observation-inbox-renderer.d.ts +5 -0
  475. package/dist/renderer/observation-inbox-renderer.d.ts.map +1 -0
  476. package/dist/renderer/observation-inbox-renderer.js +7182 -0
  477. package/dist/renderer/observation-inbox-renderer.js.map +1 -0
  478. package/dist/renderer/skill-detail-renderer.d.ts +13 -0
  479. package/dist/renderer/skill-detail-renderer.d.ts.map +1 -0
  480. package/dist/renderer/skill-detail-renderer.js +1244 -0
  481. package/dist/renderer/skill-detail-renderer.js.map +1 -0
  482. package/dist/renderer/skill-health-renderer.d.ts.map +1 -0
  483. package/dist/renderer/skill-health-renderer.js.map +1 -0
  484. package/dist/renderer/skill-list-renderer.d.ts +3 -0
  485. package/dist/renderer/skill-list-renderer.d.ts.map +1 -0
  486. package/dist/renderer/skill-list-renderer.js +334 -0
  487. package/dist/renderer/skill-list-renderer.js.map +1 -0
  488. package/dist/renderer/summary.d.ts +68 -0
  489. package/dist/renderer/summary.d.ts.map +1 -0
  490. package/dist/renderer/summary.js +1896 -0
  491. package/dist/renderer/summary.js.map +1 -0
  492. package/dist/renderer/table.d.ts.map +1 -0
  493. package/dist/renderer/table.js.map +1 -0
  494. package/dist/renderer/test-view.d.ts.map +1 -0
  495. package/dist/renderer/test-view.js +905 -0
  496. package/dist/renderer/test-view.js.map +1 -0
  497. package/dist/renderer/trends.d.ts.map +1 -0
  498. package/dist/renderer/trends.js.map +1 -0
  499. package/dist/server/job-store.d.ts.map +1 -0
  500. package/dist/server/job-store.js.map +1 -0
  501. package/dist/server/report-server.d.ts.map +1 -0
  502. package/dist/server/report-server.js +892 -0
  503. package/dist/server/report-server.js.map +1 -0
  504. package/dist/server/report-store.d.ts.map +1 -0
  505. package/dist/server/report-store.js.map +1 -0
  506. package/dist/server/skill-index.d.ts +14 -0
  507. package/dist/server/skill-index.d.ts.map +1 -0
  508. package/dist/server/skill-index.js +381 -0
  509. package/dist/server/skill-index.js.map +1 -0
  510. package/dist/server/skill-insights.d.ts +32 -0
  511. package/dist/server/skill-insights.d.ts.map +1 -0
  512. package/dist/server/skill-insights.js +788 -0
  513. package/dist/server/skill-insights.js.map +1 -0
  514. package/dist/shared/hard-rules.d.ts +42 -0
  515. package/dist/shared/hard-rules.d.ts.map +1 -0
  516. package/dist/shared/hard-rules.js +242 -0
  517. package/dist/shared/hard-rules.js.map +1 -0
  518. package/dist/shared/llm-prompts/index.d.ts +14 -0
  519. package/dist/shared/llm-prompts/index.d.ts.map +1 -0
  520. package/dist/shared/llm-prompts/index.js +22 -0
  521. package/dist/shared/llm-prompts/index.js.map +1 -0
  522. package/dist/shared/llm-prompts/skill-health.d.ts +22 -0
  523. package/dist/shared/llm-prompts/skill-health.d.ts.map +1 -0
  524. package/dist/shared/llm-prompts/skill-health.js +170 -0
  525. package/dist/shared/llm-prompts/skill-health.js.map +1 -0
  526. package/dist/shared/time.d.ts.map +1 -0
  527. package/dist/shared/time.js.map +1 -0
  528. package/dist/shared/tool-search.d.ts.map +1 -0
  529. package/dist/shared/tool-search.js.map +1 -0
  530. package/dist/types/dependencies.d.ts +23 -0
  531. package/dist/types/dependencies.d.ts.map +1 -0
  532. package/dist/types/dependencies.js +2 -0
  533. package/dist/types/dependencies.js.map +1 -0
  534. package/dist/types/diagnosis.d.ts +88 -0
  535. package/dist/types/diagnosis.d.ts.map +1 -0
  536. package/dist/types/diagnosis.js +2 -0
  537. package/dist/types/diagnosis.js.map +1 -0
  538. package/dist/types/doctor.d.ts +157 -0
  539. package/dist/types/doctor.d.ts.map +1 -0
  540. package/dist/types/doctor.js.map +1 -0
  541. package/dist/types/eval.d.ts +377 -0
  542. package/dist/types/eval.d.ts.map +1 -0
  543. package/dist/types/eval.js.map +1 -0
  544. package/dist/types/executor.d.ts.map +1 -0
  545. package/dist/types/executor.js.map +1 -0
  546. package/dist/types/index.d.ts +12 -0
  547. package/dist/types/index.d.ts.map +1 -0
  548. package/dist/types/index.js +12 -0
  549. package/dist/types/index.js.map +1 -0
  550. package/dist/types/judge.d.ts.map +1 -0
  551. package/dist/types/judge.js.map +1 -0
  552. package/dist/types/observability.d.ts +865 -0
  553. package/dist/types/observability.d.ts.map +1 -0
  554. package/dist/types/observability.js +14 -0
  555. package/dist/types/observability.js.map +1 -0
  556. package/dist/types/report.d.ts +560 -0
  557. package/dist/types/report.d.ts.map +1 -0
  558. package/dist/types/report.js.map +1 -0
  559. package/dist/types/shared.d.ts.map +1 -0
  560. package/dist/types/shared.js.map +1 -0
  561. package/dist/types/skill-index.d.ts +142 -0
  562. package/dist/types/skill-index.d.ts.map +1 -0
  563. package/dist/types/skill-index.js +2 -0
  564. package/dist/types/skill-index.js.map +1 -0
  565. package/dist/types/storage.d.ts.map +1 -0
  566. package/dist/types/storage.js.map +1 -0
  567. package/dist/util/safe-slice.d.ts +23 -0
  568. package/dist/util/safe-slice.d.ts.map +1 -0
  569. package/dist/util/safe-slice.js +33 -0
  570. package/dist/util/safe-slice.js.map +1 -0
  571. package/package.json +40 -8
  572. package/dist/src/analysis/coverage-analyzer.d.ts.map +0 -1
  573. package/dist/src/analysis/coverage-analyzer.js.map +0 -1
  574. package/dist/src/analysis/failure-clusterer.d.ts.map +0 -1
  575. package/dist/src/analysis/failure-clusterer.js.map +0 -1
  576. package/dist/src/analysis/gap-analyzer.d.ts +0 -121
  577. package/dist/src/analysis/gap-analyzer.d.ts.map +0 -1
  578. package/dist/src/analysis/gap-analyzer.js +0 -471
  579. package/dist/src/analysis/gap-analyzer.js.map +0 -1
  580. package/dist/src/analysis/hedging-classifier.d.ts.map +0 -1
  581. package/dist/src/analysis/hedging-classifier.js.map +0 -1
  582. package/dist/src/analysis/report-diagnostics.d.ts +0 -34
  583. package/dist/src/analysis/report-diagnostics.d.ts.map +0 -1
  584. package/dist/src/analysis/report-diagnostics.js +0 -753
  585. package/dist/src/analysis/report-diagnostics.js.map +0 -1
  586. package/dist/src/analysis/sample-diagnostics.d.ts.map +0 -1
  587. package/dist/src/analysis/sample-diagnostics.js.map +0 -1
  588. package/dist/src/analysis/saturation.d.ts.map +0 -1
  589. package/dist/src/analysis/saturation.js.map +0 -1
  590. package/dist/src/authoring/evolver.d.ts +0 -78
  591. package/dist/src/authoring/evolver.d.ts.map +0 -1
  592. package/dist/src/authoring/evolver.js +0 -320
  593. package/dist/src/authoring/evolver.js.map +0 -1
  594. package/dist/src/authoring/generator.d.ts +0 -35
  595. package/dist/src/authoring/generator.d.ts.map +0 -1
  596. package/dist/src/authoring/generator.js +0 -704
  597. package/dist/src/authoring/generator.js.map +0 -1
  598. package/dist/src/authoring/sample-fixer.d.ts +0 -45
  599. package/dist/src/authoring/sample-fixer.d.ts.map +0 -1
  600. package/dist/src/authoring/sample-fixer.js +0 -186
  601. package/dist/src/authoring/sample-fixer.js.map +0 -1
  602. package/dist/src/cli/cli-exit.d.ts +0 -15
  603. package/dist/src/cli/cli-exit.d.ts.map +0 -1
  604. package/dist/src/cli/cli-exit.js +0 -19
  605. package/dist/src/cli/cli-exit.js.map +0 -1
  606. package/dist/src/cli/commands/_shared.d.ts +0 -12
  607. package/dist/src/cli/commands/_shared.d.ts.map +0 -1
  608. package/dist/src/cli/commands/_shared.js +0 -26
  609. package/dist/src/cli/commands/_shared.js.map +0 -1
  610. package/dist/src/cli/commands/doctor.d.ts +0 -2
  611. package/dist/src/cli/commands/doctor.d.ts.map +0 -1
  612. package/dist/src/cli/commands/doctor.js +0 -174
  613. package/dist/src/cli/commands/doctor.js.map +0 -1
  614. package/dist/src/cli/commands/eval-gold.d.ts +0 -2
  615. package/dist/src/cli/commands/eval-gold.d.ts.map +0 -1
  616. package/dist/src/cli/commands/eval-gold.js +0 -137
  617. package/dist/src/cli/commands/eval-gold.js.map +0 -1
  618. package/dist/src/cli/commands/eval-runner.d.ts +0 -2
  619. package/dist/src/cli/commands/eval-runner.d.ts.map +0 -1
  620. package/dist/src/cli/commands/eval-runner.js +0 -299
  621. package/dist/src/cli/commands/eval-runner.js.map +0 -1
  622. package/dist/src/cli/commands/eval.d.ts +0 -2
  623. package/dist/src/cli/commands/eval.d.ts.map +0 -1
  624. package/dist/src/cli/commands/eval.js +0 -11
  625. package/dist/src/cli/commands/eval.js.map +0 -1
  626. package/dist/src/cli/commands/evolve.d.ts +0 -2
  627. package/dist/src/cli/commands/evolve.d.ts.map +0 -1
  628. package/dist/src/cli/commands/evolve.js +0 -132
  629. package/dist/src/cli/commands/evolve.js.map +0 -1
  630. package/dist/src/cli/commands/init.d.ts +0 -2
  631. package/dist/src/cli/commands/init.d.ts.map +0 -1
  632. package/dist/src/cli/commands/init.js +0 -110
  633. package/dist/src/cli/commands/init.js.map +0 -1
  634. package/dist/src/cli/commands/observe.d.ts +0 -2
  635. package/dist/src/cli/commands/observe.d.ts.map +0 -1
  636. package/dist/src/cli/commands/observe.js +0 -247
  637. package/dist/src/cli/commands/observe.js.map +0 -1
  638. package/dist/src/cli/commands/registry.d.ts +0 -11
  639. package/dist/src/cli/commands/registry.d.ts.map +0 -1
  640. package/dist/src/cli/commands/registry.js +0 -32
  641. package/dist/src/cli/commands/registry.js.map +0 -1
  642. package/dist/src/cli/commands/sample.d.ts +0 -15
  643. package/dist/src/cli/commands/sample.d.ts.map +0 -1
  644. package/dist/src/cli/commands/sample.js +0 -415
  645. package/dist/src/cli/commands/sample.js.map +0 -1
  646. package/dist/src/cli/commands/studio.d.ts +0 -2
  647. package/dist/src/cli/commands/studio.d.ts.map +0 -1
  648. package/dist/src/cli/commands/studio.js +0 -84
  649. package/dist/src/cli/commands/studio.js.map +0 -1
  650. package/dist/src/cli/i18n-dict.d.ts +0 -54
  651. package/dist/src/cli/i18n-dict.d.ts.map +0 -1
  652. package/dist/src/cli/i18n-dict.js +0 -1125
  653. package/dist/src/cli/i18n-dict.js.map +0 -1
  654. package/dist/src/cli/i18n.d.ts +0 -24
  655. package/dist/src/cli/i18n.d.ts.map +0 -1
  656. package/dist/src/cli/i18n.js +0 -53
  657. package/dist/src/cli/i18n.js.map +0 -1
  658. package/dist/src/cli/index.d.ts.map +0 -1
  659. package/dist/src/cli/index.js +0 -61
  660. package/dist/src/cli/index.js.map +0 -1
  661. package/dist/src/cli/parse-run-config.d.ts +0 -90
  662. package/dist/src/cli/parse-run-config.d.ts.map +0 -1
  663. package/dist/src/cli/parse-run-config.js +0 -302
  664. package/dist/src/cli/parse-run-config.js.map +0 -1
  665. package/dist/src/cli/parse-strict.d.ts +0 -7
  666. package/dist/src/cli/parse-strict.d.ts.map +0 -1
  667. package/dist/src/cli/parse-strict.js +0 -26
  668. package/dist/src/cli/parse-strict.js.map +0 -1
  669. package/dist/src/cli/progress.d.ts.map +0 -1
  670. package/dist/src/cli/progress.js.map +0 -1
  671. package/dist/src/cli/run-tally.d.ts +0 -18
  672. package/dist/src/cli/run-tally.d.ts.map +0 -1
  673. package/dist/src/cli/run-tally.js.map +0 -1
  674. package/dist/src/cli/update-check.d.ts +0 -3
  675. package/dist/src/cli/update-check.d.ts.map +0 -1
  676. package/dist/src/cli/update-check.js +0 -37
  677. package/dist/src/cli/update-check.js.map +0 -1
  678. package/dist/src/doctor/health/builtin-dimensions.d.ts.map +0 -1
  679. package/dist/src/doctor/health/builtin-dimensions.js.map +0 -1
  680. package/dist/src/doctor/health/composer.d.ts.map +0 -1
  681. package/dist/src/doctor/health/composer.js +0 -293
  682. package/dist/src/doctor/health/composer.js.map +0 -1
  683. package/dist/src/doctor/health/dimension-registry.d.ts.map +0 -1
  684. package/dist/src/doctor/health/dimension-registry.js.map +0 -1
  685. package/dist/src/doctor/health/dimension-spec.d.ts +0 -47
  686. package/dist/src/doctor/health/dimension-spec.d.ts.map +0 -1
  687. package/dist/src/doctor/health/dimension-spec.js.map +0 -1
  688. package/dist/src/doctor/health/parser.d.ts +0 -27
  689. package/dist/src/doctor/health/parser.d.ts.map +0 -1
  690. package/dist/src/doctor/health/parser.js +0 -191
  691. package/dist/src/doctor/health/parser.js.map +0 -1
  692. package/dist/src/doctor/health/prompt-builder.d.ts +0 -22
  693. package/dist/src/doctor/health/prompt-builder.d.ts.map +0 -1
  694. package/dist/src/doctor/health/prompt-builder.js +0 -170
  695. package/dist/src/doctor/health/prompt-builder.js.map +0 -1
  696. package/dist/src/doctor/health/register.d.ts.map +0 -1
  697. package/dist/src/doctor/health/register.js.map +0 -1
  698. package/dist/src/doctor/html-renderer.d.ts +0 -20
  699. package/dist/src/doctor/html-renderer.d.ts.map +0 -1
  700. package/dist/src/doctor/html-renderer.js +0 -376
  701. package/dist/src/doctor/html-renderer.js.map +0 -1
  702. package/dist/src/doctor/index.d.ts.map +0 -1
  703. package/dist/src/doctor/index.js +0 -245
  704. package/dist/src/doctor/index.js.map +0 -1
  705. package/dist/src/doctor/preflight.d.ts.map +0 -1
  706. package/dist/src/doctor/preflight.js.map +0 -1
  707. package/dist/src/doctor/renderer.d.ts +0 -17
  708. package/dist/src/doctor/renderer.d.ts.map +0 -1
  709. package/dist/src/doctor/renderer.js +0 -123
  710. package/dist/src/doctor/renderer.js.map +0 -1
  711. package/dist/src/doctor/rules.d.ts.map +0 -1
  712. package/dist/src/doctor/rules.js +0 -289
  713. package/dist/src/doctor/rules.js.map +0 -1
  714. package/dist/src/eval-core/bootstrap.d.ts.map +0 -1
  715. package/dist/src/eval-core/bootstrap.js.map +0 -1
  716. package/dist/src/eval-core/cache.d.ts.map +0 -1
  717. package/dist/src/eval-core/cache.js.map +0 -1
  718. package/dist/src/eval-core/comparability.d.ts.map +0 -1
  719. package/dist/src/eval-core/comparability.js.map +0 -1
  720. package/dist/src/eval-core/dependency-checker.d.ts +0 -58
  721. package/dist/src/eval-core/dependency-checker.d.ts.map +0 -1
  722. package/dist/src/eval-core/dependency-checker.js.map +0 -1
  723. package/dist/src/eval-core/evaluation-execution.d.ts.map +0 -1
  724. package/dist/src/eval-core/evaluation-execution.js.map +0 -1
  725. package/dist/src/eval-core/evaluation-job.d.ts.map +0 -1
  726. package/dist/src/eval-core/evaluation-job.js.map +0 -1
  727. package/dist/src/eval-core/evaluation-reporting.d.ts.map +0 -1
  728. package/dist/src/eval-core/evaluation-reporting.js.map +0 -1
  729. package/dist/src/eval-core/execution-strategy.d.ts.map +0 -1
  730. package/dist/src/eval-core/execution-strategy.js +0 -165
  731. package/dist/src/eval-core/execution-strategy.js.map +0 -1
  732. package/dist/src/eval-core/fact-checker.d.ts.map +0 -1
  733. package/dist/src/eval-core/fact-checker.js.map +0 -1
  734. package/dist/src/eval-core/layer-gates.d.ts.map +0 -1
  735. package/dist/src/eval-core/layer-gates.js.map +0 -1
  736. package/dist/src/eval-core/mock-hook.cjs +0 -203
  737. package/dist/src/eval-core/mocks-runtime.d.ts +0 -96
  738. package/dist/src/eval-core/mocks-runtime.d.ts.map +0 -1
  739. package/dist/src/eval-core/mocks-runtime.js +0 -323
  740. package/dist/src/eval-core/mocks-runtime.js.map +0 -1
  741. package/dist/src/eval-core/schema.d.ts.map +0 -1
  742. package/dist/src/eval-core/schema.js.map +0 -1
  743. package/dist/src/eval-core/statistics.d.ts.map +0 -1
  744. package/dist/src/eval-core/statistics.js.map +0 -1
  745. package/dist/src/eval-core/task-planner.d.ts.map +0 -1
  746. package/dist/src/eval-core/task-planner.js.map +0 -1
  747. package/dist/src/eval-core/verdict.d.ts.map +0 -1
  748. package/dist/src/eval-core/verdict.js.map +0 -1
  749. package/dist/src/eval-workflows/batch-evaluation-workflow.d.ts.map +0 -1
  750. package/dist/src/eval-workflows/batch-evaluation-workflow.js.map +0 -1
  751. package/dist/src/eval-workflows/evaluation-pipeline.d.ts +0 -109
  752. package/dist/src/eval-workflows/evaluation-pipeline.d.ts.map +0 -1
  753. package/dist/src/eval-workflows/evaluation-pipeline.js +0 -373
  754. package/dist/src/eval-workflows/evaluation-pipeline.js.map +0 -1
  755. package/dist/src/eval-workflows/evaluation-preparation.d.ts.map +0 -1
  756. package/dist/src/eval-workflows/evaluation-preparation.js.map +0 -1
  757. package/dist/src/eval-workflows/run-evaluation.d.ts.map +0 -1
  758. package/dist/src/eval-workflows/run-evaluation.js +0 -528
  759. package/dist/src/eval-workflows/run-evaluation.js.map +0 -1
  760. package/dist/src/executors/anthropic-api.d.ts.map +0 -1
  761. package/dist/src/executors/anthropic-api.js.map +0 -1
  762. package/dist/src/executors/claude-cli.d.ts.map +0 -1
  763. package/dist/src/executors/claude-cli.js +0 -179
  764. package/dist/src/executors/claude-cli.js.map +0 -1
  765. package/dist/src/executors/claude-sdk-trace.d.ts.map +0 -1
  766. package/dist/src/executors/claude-sdk-trace.js.map +0 -1
  767. package/dist/src/executors/claude-sdk.d.ts.map +0 -1
  768. package/dist/src/executors/claude-sdk.js.map +0 -1
  769. package/dist/src/executors/codex-cli-trace.d.ts.map +0 -1
  770. package/dist/src/executors/codex-cli-trace.js.map +0 -1
  771. package/dist/src/executors/codex-cli.d.ts.map +0 -1
  772. package/dist/src/executors/codex-cli.js.map +0 -1
  773. package/dist/src/executors/codex-sdk.d.ts.map +0 -1
  774. package/dist/src/executors/codex-sdk.js.map +0 -1
  775. package/dist/src/executors/gemini.d.ts.map +0 -1
  776. package/dist/src/executors/gemini.js.map +0 -1
  777. package/dist/src/executors/index.d.ts.map +0 -1
  778. package/dist/src/executors/index.js.map +0 -1
  779. package/dist/src/executors/openai-api.d.ts.map +0 -1
  780. package/dist/src/executors/openai-api.js.map +0 -1
  781. package/dist/src/executors/runtime-fingerprint.d.ts.map +0 -1
  782. package/dist/src/executors/runtime-fingerprint.js.map +0 -1
  783. package/dist/src/executors/script.d.ts.map +0 -1
  784. package/dist/src/executors/script.js.map +0 -1
  785. package/dist/src/executors/shared.d.ts +0 -194
  786. package/dist/src/executors/shared.d.ts.map +0 -1
  787. package/dist/src/executors/shared.js +0 -256
  788. package/dist/src/executors/shared.js.map +0 -1
  789. package/dist/src/grading/assertions.d.ts.map +0 -1
  790. package/dist/src/grading/assertions.js.map +0 -1
  791. package/dist/src/grading/debias-validate.d.ts.map +0 -1
  792. package/dist/src/grading/debias-validate.js.map +0 -1
  793. package/dist/src/grading/diagnostic.d.ts.map +0 -1
  794. package/dist/src/grading/diagnostic.js.map +0 -1
  795. package/dist/src/grading/gold-cli.d.ts.map +0 -1
  796. package/dist/src/grading/gold-cli.js.map +0 -1
  797. package/dist/src/grading/gold-dataset.d.ts.map +0 -1
  798. package/dist/src/grading/gold-dataset.js.map +0 -1
  799. package/dist/src/grading/human-gold.d.ts.map +0 -1
  800. package/dist/src/grading/human-gold.js.map +0 -1
  801. package/dist/src/grading/index.d.ts.map +0 -1
  802. package/dist/src/grading/index.js.map +0 -1
  803. package/dist/src/grading/judge.d.ts.map +0 -1
  804. package/dist/src/grading/judge.js.map +0 -1
  805. package/dist/src/grading/layered-scores.d.ts.map +0 -1
  806. package/dist/src/grading/layered-scores.js.map +0 -1
  807. package/dist/src/inputs/eval-config.d.ts.map +0 -1
  808. package/dist/src/inputs/eval-config.js.map +0 -1
  809. package/dist/src/inputs/load-samples.d.ts.map +0 -1
  810. package/dist/src/inputs/load-samples.js.map +0 -1
  811. package/dist/src/inputs/mcp-resolver.d.ts.map +0 -1
  812. package/dist/src/inputs/mcp-resolver.js.map +0 -1
  813. package/dist/src/inputs/skill-loader.d.ts.map +0 -1
  814. package/dist/src/inputs/skill-loader.js +0 -253
  815. package/dist/src/inputs/skill-loader.js.map +0 -1
  816. package/dist/src/inputs/url-fetcher.d.ts.map +0 -1
  817. package/dist/src/inputs/url-fetcher.js.map +0 -1
  818. package/dist/src/observability/experience.d.ts +0 -260
  819. package/dist/src/observability/experience.d.ts.map +0 -1
  820. package/dist/src/observability/experience.js +0 -1233
  821. package/dist/src/observability/experience.js.map +0 -1
  822. package/dist/src/observability/inbox-view-model.d.ts +0 -27
  823. package/dist/src/observability/inbox-view-model.d.ts.map +0 -1
  824. package/dist/src/observability/inbox-view-model.js +0 -135
  825. package/dist/src/observability/inbox-view-model.js.map +0 -1
  826. package/dist/src/observability/inbox.d.ts +0 -125
  827. package/dist/src/observability/inbox.d.ts.map +0 -1
  828. package/dist/src/observability/inbox.js +0 -741
  829. package/dist/src/observability/inbox.js.map +0 -1
  830. package/dist/src/observability/problem-patterns.d.ts +0 -51
  831. package/dist/src/observability/problem-patterns.d.ts.map +0 -1
  832. package/dist/src/observability/problem-patterns.js +0 -205
  833. package/dist/src/observability/problem-patterns.js.map +0 -1
  834. package/dist/src/observability/review-state.d.ts +0 -54
  835. package/dist/src/observability/review-state.d.ts.map +0 -1
  836. package/dist/src/observability/review-state.js +0 -131
  837. package/dist/src/observability/review-state.js.map +0 -1
  838. package/dist/src/observability/skill-chain-advisories.d.ts +0 -25
  839. package/dist/src/observability/skill-chain-advisories.d.ts.map +0 -1
  840. package/dist/src/observability/skill-chain-advisories.js +0 -69
  841. package/dist/src/observability/skill-chain-advisories.js.map +0 -1
  842. package/dist/src/observability/skill-chain.d.ts +0 -61
  843. package/dist/src/observability/skill-chain.d.ts.map +0 -1
  844. package/dist/src/observability/skill-chain.js +0 -233
  845. package/dist/src/observability/skill-chain.js.map +0 -1
  846. package/dist/src/observability/skill-health-analyzer.d.ts.map +0 -1
  847. package/dist/src/observability/skill-health-analyzer.js.map +0 -1
  848. package/dist/src/observability/text-signals.d.ts +0 -7
  849. package/dist/src/observability/text-signals.d.ts.map +0 -1
  850. package/dist/src/observability/text-signals.js +0 -22
  851. package/dist/src/observability/text-signals.js.map +0 -1
  852. package/dist/src/observability/trace-adapter.d.ts.map +0 -1
  853. package/dist/src/observability/trace-adapter.js.map +0 -1
  854. package/dist/src/observability/trace-attribution.d.ts +0 -56
  855. package/dist/src/observability/trace-attribution.d.ts.map +0 -1
  856. package/dist/src/observability/trace-attribution.js +0 -200
  857. package/dist/src/observability/trace-attribution.js.map +0 -1
  858. package/dist/src/observability/trace-segmenter.d.ts +0 -52
  859. package/dist/src/observability/trace-segmenter.d.ts.map +0 -1
  860. package/dist/src/observability/trace-segmenter.js +0 -320
  861. package/dist/src/observability/trace-segmenter.js.map +0 -1
  862. package/dist/src/observability/trace-source.d.ts +0 -99
  863. package/dist/src/observability/trace-source.d.ts.map +0 -1
  864. package/dist/src/observability/trace-source.js +0 -492
  865. package/dist/src/observability/trace-source.js.map +0 -1
  866. package/dist/src/renderer/html-renderer.d.ts.map +0 -1
  867. package/dist/src/renderer/html-renderer.js.map +0 -1
  868. package/dist/src/renderer/layout.d.ts.map +0 -1
  869. package/dist/src/renderer/layout.js.map +0 -1
  870. package/dist/src/renderer/observation-inbox-renderer.d.ts +0 -4
  871. package/dist/src/renderer/observation-inbox-renderer.d.ts.map +0 -1
  872. package/dist/src/renderer/observation-inbox-renderer.js +0 -5376
  873. package/dist/src/renderer/observation-inbox-renderer.js.map +0 -1
  874. package/dist/src/renderer/skill-detail-renderer.d.ts +0 -15
  875. package/dist/src/renderer/skill-detail-renderer.d.ts.map +0 -1
  876. package/dist/src/renderer/skill-detail-renderer.js +0 -1234
  877. package/dist/src/renderer/skill-detail-renderer.js.map +0 -1
  878. package/dist/src/renderer/skill-health-renderer.d.ts.map +0 -1
  879. package/dist/src/renderer/skill-health-renderer.js.map +0 -1
  880. package/dist/src/renderer/skill-list-renderer.d.ts +0 -4
  881. package/dist/src/renderer/skill-list-renderer.d.ts.map +0 -1
  882. package/dist/src/renderer/skill-list-renderer.js +0 -327
  883. package/dist/src/renderer/skill-list-renderer.js.map +0 -1
  884. package/dist/src/renderer/summary.d.ts +0 -68
  885. package/dist/src/renderer/summary.d.ts.map +0 -1
  886. package/dist/src/renderer/summary.js +0 -1896
  887. package/dist/src/renderer/summary.js.map +0 -1
  888. package/dist/src/renderer/table.d.ts.map +0 -1
  889. package/dist/src/renderer/table.js.map +0 -1
  890. package/dist/src/renderer/test-view.d.ts.map +0 -1
  891. package/dist/src/renderer/test-view.js +0 -905
  892. package/dist/src/renderer/test-view.js.map +0 -1
  893. package/dist/src/renderer/trends.d.ts.map +0 -1
  894. package/dist/src/renderer/trends.js.map +0 -1
  895. package/dist/src/server/job-store.d.ts.map +0 -1
  896. package/dist/src/server/job-store.js.map +0 -1
  897. package/dist/src/server/report-server.d.ts.map +0 -1
  898. package/dist/src/server/report-server.js +0 -801
  899. package/dist/src/server/report-server.js.map +0 -1
  900. package/dist/src/server/report-store.d.ts.map +0 -1
  901. package/dist/src/server/report-store.js.map +0 -1
  902. package/dist/src/server/skill-index.d.ts +0 -77
  903. package/dist/src/server/skill-index.d.ts.map +0 -1
  904. package/dist/src/server/skill-index.js +0 -331
  905. package/dist/src/server/skill-index.js.map +0 -1
  906. package/dist/src/server/skill-insights.d.ts +0 -92
  907. package/dist/src/server/skill-insights.d.ts.map +0 -1
  908. package/dist/src/server/skill-insights.js +0 -676
  909. package/dist/src/server/skill-insights.js.map +0 -1
  910. package/dist/src/shared/hard-rules.d.ts +0 -40
  911. package/dist/src/shared/hard-rules.d.ts.map +0 -1
  912. package/dist/src/shared/hard-rules.js +0 -200
  913. package/dist/src/shared/hard-rules.js.map +0 -1
  914. package/dist/src/shared/time.d.ts.map +0 -1
  915. package/dist/src/shared/time.js.map +0 -1
  916. package/dist/src/shared/tool-search.d.ts.map +0 -1
  917. package/dist/src/shared/tool-search.js.map +0 -1
  918. package/dist/src/types/doctor.d.ts +0 -155
  919. package/dist/src/types/doctor.d.ts.map +0 -1
  920. package/dist/src/types/doctor.js.map +0 -1
  921. package/dist/src/types/eval.d.ts +0 -372
  922. package/dist/src/types/eval.d.ts.map +0 -1
  923. package/dist/src/types/eval.js.map +0 -1
  924. package/dist/src/types/executor.d.ts.map +0 -1
  925. package/dist/src/types/executor.js.map +0 -1
  926. package/dist/src/types/index.d.ts +0 -8
  927. package/dist/src/types/index.d.ts.map +0 -1
  928. package/dist/src/types/index.js +0 -8
  929. package/dist/src/types/index.js.map +0 -1
  930. package/dist/src/types/judge.d.ts.map +0 -1
  931. package/dist/src/types/judge.js.map +0 -1
  932. package/dist/src/types/report.d.ts +0 -551
  933. package/dist/src/types/report.d.ts.map +0 -1
  934. package/dist/src/types/report.js.map +0 -1
  935. package/dist/src/types/shared.d.ts.map +0 -1
  936. package/dist/src/types/shared.js.map +0 -1
  937. package/dist/src/types/storage.d.ts.map +0 -1
  938. package/dist/src/types/storage.js.map +0 -1
  939. package/dist/src/util/safe-slice.d.ts +0 -23
  940. package/dist/src/util/safe-slice.d.ts.map +0 -1
  941. package/dist/src/util/safe-slice.js +0 -33
  942. package/dist/src/util/safe-slice.js.map +0 -1
  943. /package/dist/{src/analysis → analysis}/coverage-analyzer.d.ts +0 -0
  944. /package/dist/{src/analysis → analysis}/coverage-analyzer.js +0 -0
  945. /package/dist/{src/analysis → analysis}/failure-clusterer.d.ts +0 -0
  946. /package/dist/{src/analysis → analysis}/failure-clusterer.js +0 -0
  947. /package/dist/{src/analysis → analysis}/hedging-classifier.d.ts +0 -0
  948. /package/dist/{src/analysis → analysis}/hedging-classifier.js +0 -0
  949. /package/dist/{src/analysis → analysis}/sample-diagnostics.d.ts +0 -0
  950. /package/dist/{src/analysis → analysis}/sample-diagnostics.js +0 -0
  951. /package/dist/{src/analysis → analysis}/saturation.d.ts +0 -0
  952. /package/dist/{src/analysis → analysis}/saturation.js +0 -0
  953. /package/dist/{src/cli → cli}/index.d.ts +0 -0
  954. /package/dist/{src/cli → cli/lib}/progress.d.ts +0 -0
  955. /package/dist/{src/cli → cli/lib}/progress.js +0 -0
  956. /package/dist/{src/cli → cli/lib}/run-tally.js +0 -0
  957. /package/dist/{src/doctor → doctor}/health/builtin-dimensions.d.ts +0 -0
  958. /package/dist/{src/doctor → doctor}/health/builtin-dimensions.js +0 -0
  959. /package/dist/{src/doctor → doctor}/health/composer.d.ts +0 -0
  960. /package/dist/{src/doctor → doctor}/health/dimension-registry.d.ts +0 -0
  961. /package/dist/{src/doctor → doctor}/health/dimension-registry.js +0 -0
  962. /package/dist/{src/doctor → doctor}/health/dimension-spec.js +0 -0
  963. /package/dist/{src/doctor → doctor}/health/register.d.ts +0 -0
  964. /package/dist/{src/doctor → doctor}/health/register.js +0 -0
  965. /package/dist/{src/doctor → doctor}/index.d.ts +0 -0
  966. /package/dist/{src/doctor → doctor}/preflight.d.ts +0 -0
  967. /package/dist/{src/doctor → doctor}/preflight.js +0 -0
  968. /package/dist/{src/doctor → doctor}/rules.d.ts +0 -0
  969. /package/dist/{src/eval-core → eval-core}/bootstrap.d.ts +0 -0
  970. /package/dist/{src/eval-core → eval-core}/bootstrap.js +0 -0
  971. /package/dist/{src/eval-core → eval-core}/cache.d.ts +0 -0
  972. /package/dist/{src/eval-core → eval-core}/cache.js +0 -0
  973. /package/dist/{src/eval-core → eval-core}/comparability.d.ts +0 -0
  974. /package/dist/{src/eval-core → eval-core}/comparability.js +0 -0
  975. /package/dist/{src/eval-core → eval-core}/dependency-checker.js +0 -0
  976. /package/dist/{src/eval-core → eval-core}/evaluation-execution.d.ts +0 -0
  977. /package/dist/{src/eval-core → eval-core}/evaluation-execution.js +0 -0
  978. /package/dist/{src/eval-core → eval-core}/evaluation-job.d.ts +0 -0
  979. /package/dist/{src/eval-core → eval-core}/evaluation-job.js +0 -0
  980. /package/dist/{src/eval-core → eval-core}/evaluation-reporting.d.ts +0 -0
  981. /package/dist/{src/eval-core → eval-core}/evaluation-reporting.js +0 -0
  982. /package/dist/{src/eval-core → eval-core}/execution-strategy.d.ts +0 -0
  983. /package/dist/{src/eval-core → eval-core}/fact-checker.d.ts +0 -0
  984. /package/dist/{src/eval-core → eval-core}/fact-checker.js +0 -0
  985. /package/dist/{src/eval-core → eval-core}/layer-gates.d.ts +0 -0
  986. /package/dist/{src/eval-core → eval-core}/layer-gates.js +0 -0
  987. /package/dist/{src/eval-core → eval-core}/schema.d.ts +0 -0
  988. /package/dist/{src/eval-core → eval-core}/schema.js +0 -0
  989. /package/dist/{src/eval-core → eval-core}/statistics.d.ts +0 -0
  990. /package/dist/{src/eval-core → eval-core}/statistics.js +0 -0
  991. /package/dist/{src/eval-core → eval-core}/task-planner.d.ts +0 -0
  992. /package/dist/{src/eval-core → eval-core}/task-planner.js +0 -0
  993. /package/dist/{src/eval-core → eval-core}/verdict.d.ts +0 -0
  994. /package/dist/{src/eval-core → eval-core}/verdict.js +0 -0
  995. /package/dist/{src/eval-workflows → eval-workflows}/batch-evaluation-workflow.d.ts +0 -0
  996. /package/dist/{src/eval-workflows → eval-workflows}/batch-evaluation-workflow.js +0 -0
  997. /package/dist/{src/eval-workflows → eval-workflows}/evaluation-preparation.d.ts +0 -0
  998. /package/dist/{src/eval-workflows → eval-workflows}/evaluation-preparation.js +0 -0
  999. /package/dist/{src/eval-workflows → eval-workflows}/run-evaluation.d.ts +0 -0
  1000. /package/dist/{src/executors → executors}/anthropic-api.d.ts +0 -0
  1001. /package/dist/{src/executors → executors}/anthropic-api.js +0 -0
  1002. /package/dist/{src/executors → executors}/claude-cli.d.ts +0 -0
  1003. /package/dist/{src/executors → executors}/claude-sdk-trace.d.ts +0 -0
  1004. /package/dist/{src/executors → executors}/claude-sdk-trace.js +0 -0
  1005. /package/dist/{src/executors → executors}/claude-sdk.d.ts +0 -0
  1006. /package/dist/{src/executors → executors}/claude-sdk.js +0 -0
  1007. /package/dist/{src/executors → executors}/codex-cli-trace.d.ts +0 -0
  1008. /package/dist/{src/executors → executors}/codex-cli-trace.js +0 -0
  1009. /package/dist/{src/executors → executors}/codex-cli.d.ts +0 -0
  1010. /package/dist/{src/executors → executors}/codex-cli.js +0 -0
  1011. /package/dist/{src/executors → executors}/codex-sdk.d.ts +0 -0
  1012. /package/dist/{src/executors → executors}/codex-sdk.js +0 -0
  1013. /package/dist/{src/executors → executors}/gemini.d.ts +0 -0
  1014. /package/dist/{src/executors → executors}/gemini.js +0 -0
  1015. /package/dist/{src/executors → executors}/index.d.ts +0 -0
  1016. /package/dist/{src/executors → executors}/index.js +0 -0
  1017. /package/dist/{src/executors → executors}/openai-api.d.ts +0 -0
  1018. /package/dist/{src/executors → executors}/openai-api.js +0 -0
  1019. /package/dist/{src/executors → executors}/runtime-fingerprint.d.ts +0 -0
  1020. /package/dist/{src/executors → executors}/runtime-fingerprint.js +0 -0
  1021. /package/dist/{src/executors → executors}/script.d.ts +0 -0
  1022. /package/dist/{src/executors → executors}/script.js +0 -0
  1023. /package/dist/{src/grading → grading}/assertions.d.ts +0 -0
  1024. /package/dist/{src/grading → grading}/assertions.js +0 -0
  1025. /package/dist/{src/grading → grading}/debias-validate.d.ts +0 -0
  1026. /package/dist/{src/grading → grading}/debias-validate.js +0 -0
  1027. /package/dist/{src/grading → grading}/diagnostic.d.ts +0 -0
  1028. /package/dist/{src/grading → grading}/diagnostic.js +0 -0
  1029. /package/dist/{src/grading → grading}/gold-cli.d.ts +0 -0
  1030. /package/dist/{src/grading → grading}/gold-cli.js +0 -0
  1031. /package/dist/{src/grading → grading}/gold-dataset.d.ts +0 -0
  1032. /package/dist/{src/grading → grading}/gold-dataset.js +0 -0
  1033. /package/dist/{src/grading → grading}/human-gold.d.ts +0 -0
  1034. /package/dist/{src/grading → grading}/human-gold.js +0 -0
  1035. /package/dist/{src/grading → grading}/index.d.ts +0 -0
  1036. /package/dist/{src/grading → grading}/index.js +0 -0
  1037. /package/dist/{src/grading → grading}/judge.d.ts +0 -0
  1038. /package/dist/{src/grading → grading}/judge.js +0 -0
  1039. /package/dist/{src/grading → grading}/layered-scores.d.ts +0 -0
  1040. /package/dist/{src/grading → grading}/layered-scores.js +0 -0
  1041. /package/dist/{src/inputs → inputs}/eval-config.d.ts +0 -0
  1042. /package/dist/{src/inputs → inputs}/eval-config.js +0 -0
  1043. /package/dist/{src/inputs → inputs}/load-samples.d.ts +0 -0
  1044. /package/dist/{src/inputs → inputs}/load-samples.js +0 -0
  1045. /package/dist/{src/inputs → inputs}/mcp-resolver.d.ts +0 -0
  1046. /package/dist/{src/inputs → inputs}/mcp-resolver.js +0 -0
  1047. /package/dist/{src/inputs → inputs}/skill-loader.d.ts +0 -0
  1048. /package/dist/{src/inputs → inputs}/url-fetcher.d.ts +0 -0
  1049. /package/dist/{src/inputs → inputs}/url-fetcher.js +0 -0
  1050. /package/dist/{src/observability → observability}/skill-health-analyzer.d.ts +0 -0
  1051. /package/dist/{src/observability → observability}/skill-health-analyzer.js +0 -0
  1052. /package/dist/{src/observability → observability}/trace-adapter.d.ts +0 -0
  1053. /package/dist/{src/observability → observability}/trace-adapter.js +0 -0
  1054. /package/dist/{src/renderer → renderer}/html-renderer.d.ts +0 -0
  1055. /package/dist/{src/renderer → renderer}/html-renderer.js +0 -0
  1056. /package/dist/{src/renderer → renderer}/layout.d.ts +0 -0
  1057. /package/dist/{src/renderer → renderer}/layout.js +0 -0
  1058. /package/dist/{src/renderer → renderer}/skill-health-renderer.d.ts +0 -0
  1059. /package/dist/{src/renderer → renderer}/skill-health-renderer.js +0 -0
  1060. /package/dist/{src/renderer → renderer}/table.d.ts +0 -0
  1061. /package/dist/{src/renderer → renderer}/table.js +0 -0
  1062. /package/dist/{src/renderer → renderer}/test-view.d.ts +0 -0
  1063. /package/dist/{src/renderer → renderer}/trends.d.ts +0 -0
  1064. /package/dist/{src/renderer → renderer}/trends.js +0 -0
  1065. /package/dist/{src/server → server}/job-store.d.ts +0 -0
  1066. /package/dist/{src/server → server}/job-store.js +0 -0
  1067. /package/dist/{src/server → server}/report-server.d.ts +0 -0
  1068. /package/dist/{src/server → server}/report-store.d.ts +0 -0
  1069. /package/dist/{src/server → server}/report-store.js +0 -0
  1070. /package/dist/{src/shared → shared}/time.d.ts +0 -0
  1071. /package/dist/{src/shared → shared}/time.js +0 -0
  1072. /package/dist/{src/shared → shared}/tool-search.d.ts +0 -0
  1073. /package/dist/{src/shared → shared}/tool-search.js +0 -0
  1074. /package/dist/{src/types → types}/doctor.js +0 -0
  1075. /package/dist/{src/types → types}/eval.js +0 -0
  1076. /package/dist/{src/types → types}/executor.d.ts +0 -0
  1077. /package/dist/{src/types → types}/executor.js +0 -0
  1078. /package/dist/{src/types → types}/judge.d.ts +0 -0
  1079. /package/dist/{src/types → types}/judge.js +0 -0
  1080. /package/dist/{src/types → types}/report.js +0 -0
  1081. /package/dist/{src/types → types}/shared.d.ts +0 -0
  1082. /package/dist/{src/types → types}/shared.js +0 -0
  1083. /package/dist/{src/types → types}/storage.d.ts +0 -0
  1084. /package/dist/{src/types → types}/storage.js +0 -0
@@ -1,905 +0,0 @@
1
- /**
2
- * 单测视角渲染器 — 把 EvaluationReport 当作 vitest 风格的测试结果看:
3
- *
4
- * 每个 sample = 一个 test case
5
- * ┌─ 用例设计(测试契约): prompt / rubric / 期望 assertions / 工具调用 mock
6
- * ├─ 执行轨迹: thinking + tool calls + 断言绑定到 turn
7
- * └─ 断言汇总: 每条 assertion 在哪一步触发,pass/fail
8
- *
9
- * 数据全部来自 EvaluationReport,不引入新数据源:
10
- * - 设计 ← report.sampleSnapshots[sample_id]
11
- * - 执行 ← VariantResult.turns / .toolCalls / .mockStats / .fullOutput
12
- * - 断言 ← VariantResult.assertions.details (已经有 passed/failed)
13
- *
14
- * 旧 report 缺 sampleSnapshots 时降级显示("无设计快照,只有执行结果")。
15
- */
16
- import { e } from './layout.js';
17
- // ──────────────────────────────────────────────────────────────────
18
- // Assertion → Turn 绑定:扫执行轨迹找断言匹配的 turn 编号
19
- // ──────────────────────────────────────────────────────────────────
20
- /** Tool call 的稳定签名,用于在 turns 里反查 turn 序号。 */
21
- export function toolCallSignature(tc) {
22
- return `${tc.tool}::${JSON.stringify(tc.input)}`;
23
- }
24
- /** 走 turns 数组,给每个 tool call 标一个 turn 编号(1-based)。导出供单测验证。 */
25
- export function indexToolCallsByTurn(turns) {
26
- const map = new Map();
27
- if (!turns)
28
- return map;
29
- turns.forEach((turn, i) => {
30
- if (turn.toolCalls) {
31
- for (const tc of turn.toolCalls) {
32
- const sig = toolCallSignature(tc);
33
- if (!map.has(sig))
34
- map.set(sig, i + 1);
35
- }
36
- }
37
- });
38
- return map;
39
- }
40
- /** 字符串解析:`tool_input_contains` / `tool_output_contains` 用 `Tool:substring` 格式。 */
41
- export function parseToolColonValue(value) {
42
- const idx = value.indexOf(':');
43
- if (idx < 0)
44
- return { tool: '', needle: value };
45
- return { tool: value.slice(0, idx), needle: value.slice(idx + 1) };
46
- }
47
- /** 把单条 assertion 绑定到某一步。匹配不到返回 null(contains/regex 等作用于 fullOutput
48
- * 的天然没有 turn 绑定)。导出供单测。
49
- *
50
- * 局限:mock_hit 的 turn 绑定靠"该 tool 的第 N 次调用 = 第 N 个 mock"启发式,
51
- * 和真实 mock matcher(按 match 规则)可能不一致——状态机 mock 场景下 turn 标签可能近似。
52
- * pass/fail 仍以 AssertionDetail.passed 为准。 */
53
- export function findAssertionTurn(assertion, detail, toolCalls, turnByCallSig) {
54
- const type = assertion.type;
55
- if (type === 'tool_input_contains' || type === 'tool_output_contains' || type === 'tool_input_not_contains') {
56
- const value = String(assertion.value ?? detail.value ?? '');
57
- const { tool, needle } = parseToolColonValue(value);
58
- for (const tc of toolCalls) {
59
- if (tool && tc.tool !== tool)
60
- continue;
61
- const target = type === 'tool_output_contains'
62
- ? JSON.stringify(tc.output ?? '')
63
- : JSON.stringify(tc.input ?? '');
64
- if (target.includes(needle)) {
65
- return { turn: turnByCallSig.get(toolCallSignature(tc)) ?? null, matchedCall: tc };
66
- }
67
- }
68
- return { turn: null };
69
- }
70
- if (type === 'tools_called' || type === 'tools_not_called') {
71
- const targets = (assertion.values ?? [String(assertion.value ?? '')]).filter(Boolean);
72
- for (const tc of toolCalls) {
73
- if (targets.includes(tc.tool)) {
74
- return { turn: turnByCallSig.get(toolCallSignature(tc)) ?? null, matchedCall: tc };
75
- }
76
- }
77
- return { turn: null };
78
- }
79
- if (type === 'mock_hit') {
80
- // mock_hit 的 value 形如 "Bash:2",对应 mockStats.perMock 的 key。
81
- // 不靠 mockStats 反推 turn(没有那个映射),靠 assertion.value 解析的 tool + 第 N 个该 tool 的调用。
82
- const value = String(assertion.value ?? detail.value ?? '');
83
- const { tool, needle } = parseToolColonValue(value);
84
- const targetIndex = parseInt(needle, 10);
85
- if (!Number.isFinite(targetIndex))
86
- return { turn: null };
87
- let counter = 0;
88
- for (const tc of toolCalls) {
89
- if (tc.tool !== tool)
90
- continue;
91
- counter += 1;
92
- if (counter === targetIndex) {
93
- return { turn: turnByCallSig.get(toolCallSignature(tc)) ?? null, matchedCall: tc };
94
- }
95
- }
96
- return { turn: null };
97
- }
98
- // contains / not_contains / regex 作用于 fullOutput,无 turn 绑定
99
- return { turn: null };
100
- }
101
- // ──────────────────────────────────────────────────────────────────
102
- // 渲染:per-sample 卡片
103
- // ──────────────────────────────────────────────────────────────────
104
- /** 用例的整体 verdict:全部 assertion 通过 → pass;任一失败 → fail;无 assertion → 视为 pass(谨慎)。 */
105
- export function sampleVerdict(details) {
106
- if (!details || details.length === 0)
107
- return 'pass';
108
- return details.every((d) => d.passed) ? 'pass' : 'fail';
109
- }
110
- /**
111
- * 判断 sample 是否是"诱错样本(tripwire)"——视觉上要跟普通 fail 区分开。
112
- *
113
- * 识别来源(任一即算):
114
- * 1. snapshot.tripwire === true(generator/作者显式标记)
115
- * 2. diagnostic.rootCause 包含 'tripwire_intentional'(LLM 诊断时自己识破的)
116
- *
117
- * 用途:UI 把诱错样本的失败渲染成紫色 "🪤 诱错触发" 而非红色 "✗ FAIL",
118
- * 避免使用者误以为是要修的真 bug。
119
- */
120
- export function isTripwireSample(snapshot, variant) {
121
- if (snapshot?.tripwire === true)
122
- return true;
123
- const rc = variant.diagnostic?.rootCause || [];
124
- return rc.includes('tripwire_intentional');
125
- }
126
- /** 作用于 LLM 全文输出的 assertion 类型(不绑定到具体 turn,而是检查 fullOutput / final answer)。 */
127
- const OUTPUT_LEVEL_ASSERTION_TYPES = new Set([
128
- 'contains', 'not_contains', 'regex',
129
- 'min_length', 'max_length', 'length',
130
- 'json_schema', 'starts_with', 'ends_with',
131
- 'rouge_n_min', 'bleu_min', 'levenshtein_max',
132
- 'embedding_similarity_min', 'answer_relevancy', 'faithfulness',
133
- ]);
134
- function renderAssertionLine(d, turn, lang) {
135
- const icon = d.passed ? '✅' : '❌';
136
- const valueRepr = typeof d.value === 'object' ? JSON.stringify(d.value) : String(d.value);
137
- let turnTag;
138
- if (turn !== null) {
139
- turnTag = `<span class="tv-turn-tag">@ Turn ${turn}</span>`;
140
- }
141
- else if (OUTPUT_LEVEL_ASSERTION_TYPES.has(d.type)) {
142
- turnTag = `<span class="tv-turn-tag">${lang === 'zh' ? '@ 全文输出' : '@ full output'}</span>`;
143
- }
144
- else if (d.passed && d.type === 'tools_not_called') {
145
- turnTag = `<span class="tv-turn-tag tv-turn-tag--ok">${lang === 'zh' ? '未触发(通过)' : 'never triggered (pass)'}</span>`;
146
- }
147
- else {
148
- turnTag = `<span class="tv-turn-tag tv-turn-tag--miss">${lang === 'zh' ? '未匹配到任何 turn' : 'no matching turn'}</span>`;
149
- }
150
- return `<div class="tv-assert ${d.passed ? 'tv-assert--pass' : 'tv-assert--fail'}">
151
- ${icon} <code>${e(d.type)}</code> <code class="tv-assert-value">${e(valueRepr)}</code>${turnTag}
152
- ${d.message ? `<div class="tv-assert-msg">${e(d.message)}</div>` : ''}
153
- </div>`;
154
- }
155
- function renderDesignSection(snapshot, lang) {
156
- // 注意:本函数返回的内容会被外层 `<details><summary>用例设计</summary>` 包起来
157
- //(见 renderSubSections 的 .tv-section),所以这里不再重复"📋 用例设计"标题。
158
- if (!snapshot) {
159
- return `<div class="tv-empty">${lang === 'zh' ? '该 report 早于单测视角支持时间,缺设计快照' : 'Report predates test-view support; no design snapshot'}</div>`;
160
- }
161
- const meta = [];
162
- if (snapshot.capability && snapshot.capability.length > 0) {
163
- const capLabel = lang === 'zh' ? '能力(capability)' : 'capability';
164
- meta.push(`<span class="tv-meta-pill">${capLabel}: ${snapshot.capability.map(e).join(' · ')}</span>`);
165
- }
166
- if (snapshot.construct) {
167
- const lbl = lang === 'zh' ? '验证目标(construct)' : 'construct';
168
- meta.push(`<span class="tv-meta-pill">${lbl}: ${e(constructValueLabel(snapshot.construct, lang))}</span>`);
169
- }
170
- if (snapshot.difficulty) {
171
- const lbl = lang === 'zh' ? '难度(difficulty)' : 'difficulty';
172
- meta.push(`<span class="tv-meta-pill">${lbl}: ${e(difficultyValueLabel(snapshot.difficulty, lang))}</span>`);
173
- }
174
- if (snapshot.provenance) {
175
- const lbl = lang === 'zh' ? '来源(provenance)' : 'provenance';
176
- meta.push(`<span class="tv-meta-pill">${lbl}: ${e(provenanceValueLabel(snapshot.provenance, lang))}</span>`);
177
- }
178
- const expectedAssertions = (snapshot.assertions ?? []).map((a) => renderExpectedAssertion(a)).join('');
179
- const mocksBlock = snapshot.mocks && snapshot.mocks.length > 0
180
- ? `<details class="tv-collapsible"><summary>${lang === 'zh' ? `工具调用 Mock(${snapshot.mocks.length} 条)` : `Tool-call mocks (${snapshot.mocks.length})`}</summary>
181
- <div class="tv-mocks">${snapshot.mocks.map((m, i) => renderMock(m, i + 1)).join('')}</div>
182
- </details>`
183
- : '';
184
- return `<div class="tv-section--design-body">
185
- <div class="tv-prompt"><strong>${lang === 'zh' ? '提示词(Prompt):' : 'Prompt:'}</strong> <span>${e(snapshot.prompt)}</span></div>
186
- ${snapshot.rubric ? `<div class="tv-rubric"><strong>${lang === 'zh' ? '评分标准(Rubric):' : 'Rubric:'}</strong> <span>${e(snapshot.rubric)}</span></div>` : ''}
187
- ${snapshot.context ? `<details class="tv-collapsible"><summary>${lang === 'zh' ? '上下文(Context)' : 'Context'}</summary><pre class="tv-pre">${e(snapshot.context)}</pre></details>` : ''}
188
- ${meta.length > 0 ? `<div class="tv-meta-pills">${meta.join(' ')}</div>` : ''}
189
- ${expectedAssertions ? `<div class="tv-expected">
190
- <div class="tv-expected-h">${lang === 'zh' ? '期望(assertions):' : 'Expected (assertions):'}</div>
191
- ${expectedAssertions}
192
- </div>` : ''}
193
- ${mocksBlock}
194
- </div>`;
195
- }
196
- /** construct 取值常见三种,zh 显示纯中文;非枚举的自由值原样回显(作者自定义,无法翻译)。 */
197
- function constructValueLabel(v, lang) {
198
- if (lang !== 'zh')
199
- return v;
200
- const map = {
201
- 'necessity': '必要性',
202
- 'capability': '能力',
203
- 'quality': '质量',
204
- };
205
- return map[v] ?? v;
206
- }
207
- function difficultyValueLabel(v, lang) {
208
- if (lang !== 'zh')
209
- return v;
210
- const map = {
211
- 'easy': '容易',
212
- 'medium': '中等',
213
- 'hard': '困难',
214
- };
215
- return map[v] ?? v;
216
- }
217
- function provenanceValueLabel(v, lang) {
218
- if (lang !== 'zh')
219
- return v;
220
- const map = {
221
- 'human': '人工编写',
222
- 'llm-generated': 'LLM 生成',
223
- 'production-trace': '线上轨迹',
224
- };
225
- return map[v] ?? v;
226
- }
227
- function renderExpectedAssertion(a) {
228
- const typeLabel = e(a.type);
229
- const valueRepr = a.values
230
- ? JSON.stringify(a.values)
231
- : a.value !== undefined
232
- ? String(a.value)
233
- : a.pattern
234
- ? `regex /${a.pattern}/${a.flags ?? ''}`
235
- : '';
236
- const weight = typeof a.weight === 'number' ? ` (w=${a.weight})` : '';
237
- return `<div class="tv-expect-line">• <code>${typeLabel}</code> <code class="tv-assert-value">${e(valueRepr)}</code>${weight}</div>`;
238
- }
239
- function renderMock(m, n) {
240
- const match = m.match ? JSON.stringify(m.match) : '(any)';
241
- let returnRepr;
242
- if (m.return !== undefined) {
243
- returnRepr = typeof m.return === 'string' ? m.return : JSON.stringify(m.return);
244
- }
245
- else if (m.return_seq) {
246
- returnRepr = `seq(${m.return_seq.length})`;
247
- }
248
- else if (m.return_file) {
249
- returnRepr = `file: ${m.return_file}`;
250
- }
251
- else {
252
- returnRepr = '(no return)';
253
- }
254
- return `<div class="tv-mock-line">
255
- <span class="tv-mock-idx">#${n}</span>
256
- <code class="tv-mock-tool">${e(m.tool)}</code>
257
- <code class="tv-mock-match">${e(match)}</code>
258
- <span class="tv-mock-arrow">→</span>
259
- <code class="tv-mock-return">${e(returnRepr.length > 200 ? returnRepr.slice(0, 200) + '…' : returnRepr)}</code>
260
- </div>`;
261
- }
262
- function renderExecutionTrace(variant, detailsByTurn, matchedMockNumbersByTurn, lang) {
263
- const turns = variant.turns ?? [];
264
- if (turns.length === 0) {
265
- return `<div class="tv-section tv-section--trace">
266
- <div class="tv-section-h">${lang === 'zh' ? '🎬 执行轨迹' : '🎬 Execution trace'}</div>
267
- <div class="tv-empty">${lang === 'zh' ? '该 variant 无 turn 轨迹(executor 未上报)' : 'No turn trace from executor'}</div>
268
- </div>`;
269
- }
270
- const stepsHtml = turns.map((turn, i) => {
271
- const turnNum = i + 1;
272
- const isAssistant = turn.role === 'assistant';
273
- const thinkingBlock = isAssistant && turn.content
274
- ? `<details class="tv-thinking"><summary>${lang === 'zh' ? '展开 thinking' : 'show thinking'} (${turn.content.length} ${lang === 'zh' ? '字' : 'chars'})</summary><div class="tv-thinking-body">${e(turn.content)}</div></details>`
275
- : '';
276
- const calls = (turn.toolCalls ?? []).map((tc) => renderToolCall(tc, turnNum, matchedMockNumbersByTurn.get(turnNum) ?? [])).join('');
277
- const assertsForTurn = (detailsByTurn.get(turnNum) ?? []);
278
- const turnAssertsHtml = assertsForTurn.map((d) => `<div class="tv-trace-assert ${d.passed ? 'tv-assert--pass' : 'tv-assert--fail'}">${d.passed ? '✅' : '❌'} <code>${e(d.type)}</code> <code class="tv-assert-value">${e(typeof d.value === 'object' ? JSON.stringify(d.value) : String(d.value))}</code></div>`).join('');
279
- const durationTag = turn.durationMs ? `<span class="tv-turn-dur">${(turn.durationMs / 1000).toFixed(1)}s</span>` : '';
280
- return `<div class="tv-turn">
281
- <div class="tv-turn-h">Turn ${turnNum} · <code>${e(turn.role)}</code> ${durationTag}</div>
282
- ${thinkingBlock}
283
- ${calls}
284
- ${turnAssertsHtml}
285
- </div>`;
286
- }).join('');
287
- const finalOutput = variant.fullOutput
288
- ? `<div class="tv-final"><div class="tv-final-h">${lang === 'zh' ? '最终回答' : 'Final answer'}</div><pre class="tv-pre">${e(variant.fullOutput.length > 2000 ? variant.fullOutput.slice(0, 2000) + '…' : variant.fullOutput)}</pre></div>`
289
- : '';
290
- return `<div class="tv-section tv-section--trace">
291
- <div class="tv-section-h">${lang === 'zh' ? '🎬 执行轨迹' : '🎬 Execution trace'}</div>
292
- ${stepsHtml}
293
- ${finalOutput}
294
- </div>`;
295
- }
296
- function renderToolCall(tc, _turnNum, mockHitsHere) {
297
- const inputRepr = typeof tc.input === 'string' ? tc.input : JSON.stringify(tc.input ?? '');
298
- const outputRepr = typeof tc.output === 'string' ? tc.output : JSON.stringify(tc.output ?? '');
299
- const mockBadge = mockHitsHere.length > 0
300
- ? `<span class="tv-mock-badge" title="matched mock #${mockHitsHere.join(',#')}">✓ mock#${mockHitsHere.join(',#')}</span>`
301
- : '';
302
- return `<div class="tv-call ${tc.success ? '' : 'tv-call--err'}">
303
- <div class="tv-call-h">🔧 <code>${e(tc.tool)}</code> ${mockBadge}</div>
304
- <details><summary>input</summary><pre class="tv-pre">${e(inputRepr.length > 1000 ? inputRepr.slice(0, 1000) + '…' : inputRepr)}</pre></details>
305
- <details><summary>output</summary><pre class="tv-pre">${e(outputRepr.length > 1000 ? outputRepr.slice(0, 1000) + '…' : outputRepr)}</pre></details>
306
- </div>`;
307
- }
308
- function renderAssertionsSummary(details, expected, toolCalls, turns, lang) {
309
- if (!details || details.length === 0) {
310
- return {
311
- html: `<div class="tv-section tv-section--asserts">
312
- <div class="tv-section-h">${lang === 'zh' ? '📊 断言结果' : '📊 Assertion results'}</div>
313
- <div class="tv-empty">${lang === 'zh' ? '该 sample 没有声明 assertion' : 'No assertions declared on this sample'}</div>
314
- </div>`,
315
- detailsByTurn: new Map(),
316
- matchedMocks: [],
317
- };
318
- }
319
- const turnByCallSig = indexToolCallsByTurn(turns);
320
- const detailsByTurn = new Map();
321
- const matchedMocks = [];
322
- const lines = [];
323
- details.forEach((d, i) => {
324
- const expectedAssertion = expected?.[i];
325
- const binding = expectedAssertion
326
- ? findAssertionTurn(expectedAssertion, d, toolCalls, turnByCallSig)
327
- : { turn: null };
328
- if (binding.turn !== null) {
329
- const arr = detailsByTurn.get(binding.turn) ?? [];
330
- arr.push(d);
331
- detailsByTurn.set(binding.turn, arr);
332
- }
333
- if (expectedAssertion?.type === 'mock_hit' && binding.turn !== null) {
334
- const value = String(expectedAssertion.value ?? d.value ?? '');
335
- const { needle } = parseToolColonValue(value);
336
- const idx = parseInt(needle, 10);
337
- if (Number.isFinite(idx))
338
- matchedMocks.push(idx);
339
- }
340
- lines.push(renderAssertionLine(d, binding.turn, lang));
341
- });
342
- const passCount = details.filter((d) => d.passed).length;
343
- return {
344
- html: `<div class="tv-section tv-section--asserts">
345
- <div class="tv-section-h">${lang === 'zh' ? '📊 断言结果' : '📊 Assertion results'} <span class="tv-section-stat">(${passCount}/${details.length})</span></div>
346
- ${lines.join('')}
347
- </div>`,
348
- detailsByTurn,
349
- matchedMocks,
350
- };
351
- }
352
- function buildMatchedMocksByTurn(matchedMockIndices, expected, toolCalls, turns) {
353
- // matchedMockIndices: list of mock indices(1-based)that have an associated turn-bound assertion.
354
- // 重新走一遍 toolCalls,把 mock idx 对应的 turn 编号收集起来。
355
- const out = new Map();
356
- if (matchedMockIndices.length === 0 || !expected)
357
- return out;
358
- const turnByCallSig = indexToolCallsByTurn(turns);
359
- const counters = new Map();
360
- // 给每个 toolCall 配上"它在该 tool 内的第几次调用"
361
- const callIndexByCall = new Map();
362
- for (const tc of toolCalls) {
363
- const next = (counters.get(tc.tool) ?? 0) + 1;
364
- counters.set(tc.tool, next);
365
- callIndexByCall.set(toolCallSignature(tc), next);
366
- }
367
- // 对每条 mock_hit 期望,找到匹配的 toolCall 的 turn
368
- for (const a of expected) {
369
- if (a.type !== 'mock_hit')
370
- continue;
371
- const value = String(a.value ?? '');
372
- const { tool, needle } = parseToolColonValue(value);
373
- const want = parseInt(needle, 10);
374
- if (!Number.isFinite(want))
375
- continue;
376
- if (!matchedMockIndices.includes(want))
377
- continue;
378
- for (const tc of toolCalls) {
379
- if (tc.tool !== tool)
380
- continue;
381
- const sig = toolCallSignature(tc);
382
- if (callIndexByCall.get(sig) === want) {
383
- const turn = turnByCallSig.get(sig);
384
- if (turn) {
385
- const arr = out.get(turn) ?? [];
386
- arr.push(want);
387
- out.set(turn, arr);
388
- }
389
- break;
390
- }
391
- }
392
- }
393
- return out;
394
- }
395
- /**
396
- * Diagnostic 区 — 与 Judge 评价彻底独立的功能性测试视角。
397
- *
398
- * 数据来源:variant.diagnostic(eval pipeline 在 failed sample 上跑独立 LLM 调用产出)。
399
- * --no-diagnostic 跑会缺这个字段,显示降级提示。
400
- *
401
- * 与 Judge 评价的区别:
402
- * - Judge 答"打几分"(主观质量),归 📊 评分 tab。
403
- * - Diagnostic 答"哪错了 + skill 怎么改"(可操作建议),归 ✅ 单测 tab。
404
- * 此 renderer 只负责后者,不读 llmReason / llmReasoning。
405
- *
406
- * 设计:fail 用例顶部默认展开;pass 用例不显示这个 section(诊断只对 fail 有意义)。
407
- */
408
- function renderDiagnosticSection(variant, verdict, lang) {
409
- // pass 用例不需要诊断
410
- if (verdict === 'pass')
411
- return '';
412
- const diag = variant.diagnostic;
413
- if (!diag) {
414
- return `<div class="tv-diag">
415
- <p class="tv-diag-text"><em>${lang === 'zh' ? '此用例无诊断(可能跑了 --no-diagnostic 或 report 早于诊断功能)。重跑 omk eval 即可看到针对性建议。' : 'No diagnostic on this failed case. Re-run omk eval to get targeted suggestions.'}</em></p>
416
- </div>`;
417
- }
418
- if (!diag.ok) {
419
- return `<div class="tv-diag">
420
- <p class="tv-diag-text"><em>${lang === 'zh' ? '诊断生成失败:' : 'Diagnostic generation failed:'} ${e(diag.error || '')}</em></p>
421
- ${diag.summary ? `<p class="tv-diag-text">${e(diag.summary)}</p>` : ''}
422
- </div>`;
423
- }
424
- // 单一建议块:三类 suggestion 中只挑非空的展示。优先级 sample > skill > none(用例bug 比 skill 改进比"无需改"更紧迫)。
425
- let suggestion = null;
426
- if (diag.suggestion.sample) {
427
- suggestion = { label: lang === 'zh' ? '◇ 建议改 sample' : '◇ Edit sample', body: diag.suggestion.sample };
428
- }
429
- else if (diag.suggestion.skill) {
430
- suggestion = { label: lang === 'zh' ? '◇ 建议改 skill' : '◇ Edit skill', body: diag.suggestion.skill };
431
- }
432
- else if (diag.suggestion.none) {
433
- suggestion = { label: lang === 'zh' ? '◇ 无需改动' : '◇ No fix needed', body: diag.suggestion.none };
434
- }
435
- const suggestionHtml = suggestion
436
- ? `<div class="tv-suggestion">
437
- <span class="tv-suggestion-label">${suggestion.label}</span>
438
- <p class="tv-suggestion-body">${e(suggestion.body)}</p>
439
- </div>`
440
- : '';
441
- // 失败模式标签 — 中文枚举,直接显著展示在 summary 下
442
- const failureModes = diag.failureModes ?? [];
443
- const failureModesHtml = failureModes.length > 0
444
- ? `<div class="tv-diag-modes">
445
- <span class="tv-diag-modes-label">${lang === 'zh' ? '失败模式:' : 'Failure modes:'}</span>
446
- ${failureModes.map((m) => `<span class="tv-diag-mode-pill tv-diag-mode--${e(failureModeSlug(m))}">${e(m)}</span>`).join('')}
447
- </div>`
448
- : '';
449
- // 工作流校验清单 — 默认展开,放在 suggestion 之上(最直接的"哪步对哪步错"视图)
450
- const workflowChecks = diag.workflowChecks ?? [];
451
- const workflowChecksHtml = workflowChecks.length > 0
452
- ? `<div class="tv-diag-workflow">
453
- <p class="tv-diag-workflow-title">${lang === 'zh' ? '工作流校验' : 'Workflow checklist'}</p>
454
- <ul class="tv-diag-workflow-list">
455
- ${workflowChecks.map((c) => `
456
- <li class="tv-diag-workflow-item tv-diag-workflow-item--${c.passed ? 'pass' : 'fail'}">
457
- <span class="tv-diag-workflow-icon">${c.passed ? '✓' : '✗'}</span>
458
- <div class="tv-diag-workflow-body">
459
- <span class="tv-diag-workflow-step">${e(c.step)}</span>
460
- ${c.evidence ? `<span class="tv-diag-workflow-evidence">${e(c.evidence)}</span>` : ''}
461
- </div>
462
- </li>
463
- `).join('')}
464
- </ul>
465
- </div>`
466
- : '';
467
- // 详情区(rootCause + expected/actual + 其余 suggestion)默认折叠
468
- const rootCausePills = diag.rootCause.length > 0
469
- ? `<p><strong>${lang === 'zh' ? '归因:' : 'Root cause:'}</strong>${diag.rootCause.map((c) => `<span class="tv-diag-rc-pill tv-diag-rc--${e(c)}">${e(rootCauseLabel(c, lang))}</span>`).join('')}</p>`
470
- : '';
471
- const detailsContent = [];
472
- if (rootCausePills)
473
- detailsContent.push(rootCausePills);
474
- if (diag.expected)
475
- detailsContent.push(`<p><strong>${lang === 'zh' ? '期望:' : 'Expected:'}</strong>${e(diag.expected)}</p>`);
476
- if (diag.actual)
477
- detailsContent.push(`<p><strong>${lang === 'zh' ? '实际:' : 'Actual:'}</strong>${e(diag.actual)}</p>`);
478
- // 其它 suggestion 槽位也露出来(避免完全隐藏)
479
- if (diag.suggestion.sample && suggestion?.label !== (lang === 'zh' ? '◇ 建议改 sample' : '◇ Edit sample')) {
480
- detailsContent.push(`<p><strong>${lang === 'zh' ? '另:改 sample' : 'Also: edit sample'}</strong>${e(diag.suggestion.sample)}</p>`);
481
- }
482
- if (diag.suggestion.skill && suggestion?.label !== (lang === 'zh' ? '◇ 建议改 skill' : '◇ Edit skill')) {
483
- detailsContent.push(`<p><strong>${lang === 'zh' ? '另:改 skill' : 'Also: edit skill'}</strong>${e(diag.suggestion.skill)}</p>`);
484
- }
485
- const detailsHtml = detailsContent.length > 0
486
- ? `<details class="tv-diag-details"><summary>${lang === 'zh' ? '展开归因细节' : 'show root cause / expected vs actual'}</summary>
487
- <div class="tv-diag-details-body">${detailsContent.join('')}</div>
488
- </details>`
489
- : '';
490
- // 部分内容(salvage)时挂个小提示,让用户知道诊断 LLM 输出被截断
491
- const partialBadge = diag.error?.startsWith('partial:')
492
- ? `<p class="tv-diag-partial"><em>${lang === 'zh' ? '⚠ 诊断 LLM 输出被截断,以下为部分内容' : '⚠ Diagnostic LLM output truncated; partial content shown'}</em></p>`
493
- : '';
494
- return `<div class="tv-diag">
495
- ${partialBadge}
496
- ${diag.summary ? `<p class="tv-diag-text">${e(diag.summary)}</p>` : ''}
497
- ${failureModesHtml}
498
- ${workflowChecksHtml}
499
- ${suggestionHtml}
500
- ${detailsHtml}
501
- </div>`;
502
- }
503
- /** 失败模式中文标签 → CSS-safe slug,用于挂 tv-diag-mode--<slug> 的颜色样式。 */
504
- function failureModeSlug(m) {
505
- const map = {
506
- '工作流跳步': 'workflow-skip',
507
- '硬编码值': 'hardcoded',
508
- '幻觉输出': 'hallucination',
509
- '工具误用': 'tool-misuse',
510
- '环境拦截': 'env-blocked',
511
- '误读约束': 'misread',
512
- '其他': 'other',
513
- };
514
- return map[m] ?? 'other';
515
- }
516
- /** rootCause 英文枚举 → 中文/英文短标签(给已有 rc-pill 用,UI 不再直接吐 enum 名)。 */
517
- function rootCauseLabel(c, lang) {
518
- const zh = {
519
- 'skill_doc_unclear': 'skill 文档模糊',
520
- 'skill_doc_missing': 'skill 缺失场景',
521
- 'llm_misread': 'LLM 误读',
522
- 'sample_design': 'sample 设计 bug',
523
- 'tripwire_intentional': '诱错样本(预期)',
524
- };
525
- const en = {
526
- 'skill_doc_unclear': 'skill doc unclear',
527
- 'skill_doc_missing': 'skill doc missing',
528
- 'llm_misread': 'LLM misread',
529
- 'sample_design': 'sample design bug',
530
- 'tripwire_intentional': 'tripwire (expected fail)',
531
- };
532
- return (lang === 'zh' ? zh : en)[c] ?? c;
533
- }
534
- /** 把 design / trace / assertions 三段子内容包成 <details>(默认折叠),供 pass row 展开 / fail card 折叠后查看用。 */
535
- function renderSubSections(d, lang) {
536
- const { snapshot, variant } = d;
537
- const toolCalls = variant.toolCalls ?? [];
538
- const turns = variant.turns;
539
- const summary = renderAssertionsSummary(variant.assertions?.details, snapshot?.assertions, toolCalls, turns, lang);
540
- const matchedByTurn = buildMatchedMocksByTurn(summary.matchedMocks, snapshot?.assertions, toolCalls, turns);
541
- const designHtml = renderDesignSection(snapshot, lang);
542
- const traceHtml = renderExecutionTrace(variant, summary.detailsByTurn, matchedByTurn, lang);
543
- const totalAsserts = variant.assertions?.total ?? 0;
544
- const passedAsserts = variant.assertions?.passed ?? 0;
545
- return `
546
- <details class="tv-section"><summary>${lang === 'zh' ? '用例设计' : 'Test design'}</summary><div class="tv-section-body">${designHtml}</div></details>
547
- <details class="tv-section"><summary>${lang === 'zh' ? '执行轨迹' : 'Execution trace'}<span class="tv-section-stat">(${turns?.length ?? 0} ${lang === 'zh' ? '步' : 'steps'} · ${toolCalls.length} ${lang === 'zh' ? '工具调用' : 'tool calls'})</span></summary><div class="tv-section-body">${traceHtml}</div></details>
548
- <details class="tv-section"><summary>${lang === 'zh' ? '断言结果' : 'Assertion results'}<span class="tv-section-stat">(${passedAsserts}/${totalAsserts})</span></summary><div class="tv-section-body">${summary.html}</div></details>
549
- `;
550
- }
551
- function buildCardData(sample_id, snapshot, variant) {
552
- const rawVerdict = sampleVerdict(variant.assertions?.details);
553
- const isTripwire = rawVerdict === 'fail' && isTripwireSample(snapshot, variant);
554
- return { sample_id, snapshot, variant, cardKind: isTripwire ? 'tripwire' : rawVerdict };
555
- }
556
- /** Header 公共部分:状态点 + sample_id + 标题 + 耗时。
557
- * 评分(compositeScore)和成本($cost)挪到详情区,列表只保留状态识别需要的最小信息。 */
558
- function renderCardHeader(d, lang) {
559
- const { sample_id, snapshot, variant, cardKind } = d;
560
- const dur = variant.timing?.execMs ?? variant.durationMs;
561
- const durTag = dur ? `${(dur / 1000).toFixed(1)}s` : '—';
562
- const titleSnippet = snapshot?.prompt
563
- ? snapshot.prompt.slice(0, 80).replace(/\n/g, ' ')
564
- : (lang === 'zh' ? '(无 prompt 快照)' : '(no prompt snapshot)');
565
- return `
566
- <span class="tv-dot tv-dot--${cardKind}"></span>
567
- <code class="tv-id">${e(sample_id)}</code>
568
- <span class="tv-title">${e(titleSnippet)}${snapshot?.prompt && snapshot.prompt.length > 80 ? '…' : ''}</span>
569
- <span class="tv-meta">${durTag}</span>
570
- `;
571
- }
572
- /** Pass 用例:单行折叠,点击展开看完整内容(design/trace/assertions)。 */
573
- function renderPassRow(d, lang) {
574
- return `<details class="tv-row tv-row--pass">
575
- <summary class="tv-row-h">${renderCardHeader(d, lang)}</summary>
576
- <div class="tv-row-body">${renderSubSections(d, lang)}</div>
577
- </details>`;
578
- }
579
- /** Fail / tripwire 卡片:默认展开,顶部诊断区,sub-section 折叠。 */
580
- function renderFailCard(d, lang) {
581
- const tripwireBanner = d.cardKind === 'tripwire'
582
- ? `<div class="tv-tripwire-banner">
583
- <div class="tv-tripwire-banner-title">${lang === 'zh' ? '◆ 这是诱错样本' : '◆ This is a tripwire sample'}</div>
584
- <div class="tv-tripwire-banner-body">${lang === 'zh'
585
- ? '用户 prompt 故意诱导 LLM 走错。LLM 失败 = 已记录的 LLM 边界数据点,不是要修的 bug。跨版本对比这条用例的 pass 率,可观察 LLM 抗诱导能力变化。'
586
- : 'The user prompt intentionally misdirects the LLM. A failure here is a recorded LLM-susceptibility data point — not a bug to fix.'}</div>
587
- </div>`
588
- : '';
589
- const diagnosticHtml = renderDiagnosticSection(d.variant, 'fail', lang);
590
- // 默认折叠(用户反馈:fail/tripwire 也别自动展开,跟 pass 行为一致)。
591
- // 卡片头部加一行诊断 summary 摘要,折叠态也能看到失败原因不用展开。
592
- const diagSummaryPreview = d.variant.diagnostic?.ok && d.variant.diagnostic.summary
593
- ? `<div class="tv-card-preview">${e(d.variant.diagnostic.summary.slice(0, 120))}${d.variant.diagnostic.summary.length > 120 ? '…' : ''}</div>`
594
- : '';
595
- return `<details class="tv-card tv-card--${d.cardKind}">
596
- <summary class="tv-card-h">${renderCardHeader(d, lang)}</summary>
597
- ${diagSummaryPreview}
598
- <div class="tv-card-body">
599
- ${tripwireBanner}
600
- ${diagnosticHtml}
601
- ${renderSubSections(d, lang)}
602
- </div>
603
- </details>`;
604
- }
605
- // ──────────────────────────────────────────────────────────────────
606
- // 主入口
607
- // ──────────────────────────────────────────────────────────────────
608
- export function renderTestView(report, lang) {
609
- const variants = report.meta.variants || [];
610
- const results = report.results || [];
611
- const snapshots = report.sampleSnapshots ?? {};
612
- if (variants.length === 0 || results.length === 0) {
613
- return `<div class="tv-root"><div class="tv-empty-page">${lang === 'zh' ? '无 variant 或 result,无法渲染单测视角' : 'No variants or results to render'}</div></div>`;
614
- }
615
- const hasSnapshots = Object.keys(snapshots).length > 0;
616
- const fallbackBanner = hasSnapshots ? '' : `<div class="tv-warn">${lang === 'zh' ? '⚠ 该 report 缺 sampleSnapshots(早期版本生成),只能展示执行结果,无设计信息' : '⚠ Report has no sampleSnapshots (older format); only execution results shown'}</div>`;
617
- const variantSections = variants.map((variant, idx) => {
618
- // 按 verdict 分三组:pass(单行) / fail(完整卡片) / tripwire(诱错卡片)。
619
- const passData = [];
620
- const failData = [];
621
- const tripwireData = [];
622
- for (const r of results) {
623
- const v = r.variants[variant];
624
- if (!v)
625
- continue;
626
- const d = buildCardData(r.sample_id, snapshots[r.sample_id], v);
627
- if (d.cardKind === 'pass')
628
- passData.push(d);
629
- else if (d.cardKind === 'tripwire')
630
- tripwireData.push(d);
631
- else
632
- failData.push(d);
633
- }
634
- const passSection = passData.length > 0
635
- ? `<div class="tv-group-sep">${lang === 'zh' ? '通过用例' : 'Passed'}</div>
636
- <div class="tv-rows">${passData.map((d) => renderPassRow(d, lang)).join('')}</div>`
637
- : '';
638
- const failSection = failData.length > 0
639
- ? `<div class="tv-group-sep">${lang === 'zh' ? '失败用例' : 'Failed'}</div>
640
- ${failData.map((d) => renderFailCard(d, lang)).join('')}`
641
- : '';
642
- const tripwireSection = tripwireData.length > 0
643
- ? `<div class="tv-group-sep">${lang === 'zh' ? '诱错样本' : 'Tripwire samples'}</div>
644
- ${tripwireData.map((d) => renderFailCard(d, lang)).join('')}`
645
- : '';
646
- const variantStats = computeVariantStats(report, variant);
647
- const tripwireStat = variantStats.tripwire > 0
648
- ? `<span class="tv-stat tv-stat--tripwire"><span class="tv-stat-num">${variantStats.tripwire}</span>${lang === 'zh' ? '诱错触发' : 'tripwire'}</span>`
649
- : '';
650
- return `<div class="tv-variant-panel" data-variant="${e(variant)}" ${idx === 0 ? '' : 'hidden'}>
651
- <div class="tv-variant-stats">
652
- <span class="tv-stat tv-stat--pass"><span class="tv-stat-num">${variantStats.passed}</span>${lang === 'zh' ? '通过' : 'passed'}</span>
653
- <span class="tv-stat tv-stat--fail"><span class="tv-stat-num">${variantStats.failed}</span>${lang === 'zh' ? '失败' : 'failed'}</span>
654
- ${tripwireStat}
655
- <span class="tv-stat"><span class="tv-stat-num">${variantStats.avgScore.toFixed(1)}</span>${lang === 'zh' ? '平均分' : 'avg / 5'}</span>
656
- </div>
657
- ${failSection}
658
- ${tripwireSection}
659
- ${passSection}
660
- </div>`;
661
- }).join('');
662
- // 只 1 个 variant 时不显示切换器(SOLO 模式没对照,显示 tab 反而困惑)。
663
- const variantTabsHtml = variants.length > 1
664
- ? `<div class="tv-variant-tabs">${variants.map((variant, idx) => `<button type="button" class="tv-vtab ${idx === 0 ? 'tv-vtab--active' : ''}" data-target="${e(variant)}" onclick="omkSwitchVariant(this)">${e(variant)}</button>`).join('')}</div>`
665
- : '';
666
- // 整个 test view 包在 tv-root 里,scoped 浅色岛(覆盖外层深色 layout)。
667
- return `<div class="tv-root">${fallbackBanner}
668
- ${variantTabsHtml}
669
- ${variantSections}</div>`;
670
- }
671
- function computeVariantStats(report, variant) {
672
- const snapshots = report.sampleSnapshots ?? {};
673
- let passed = 0;
674
- let failed = 0;
675
- let tripwire = 0;
676
- let scoreSum = 0;
677
- let scoreCount = 0;
678
- for (const r of report.results) {
679
- const v = r.variants[variant];
680
- if (!v)
681
- continue;
682
- const verdict = sampleVerdict(v.assertions?.details);
683
- if (verdict === 'pass') {
684
- passed += 1;
685
- }
686
- else if (isTripwireSample(snapshots[r.sample_id], v)) {
687
- tripwire += 1; // 诱错样本失败单独统计,不算"失败"也不算"通过"
688
- }
689
- else {
690
- failed += 1;
691
- }
692
- if (typeof v.compositeScore === 'number') {
693
- scoreSum += v.compositeScore;
694
- scoreCount += 1;
695
- }
696
- }
697
- return {
698
- passed,
699
- failed,
700
- tripwire,
701
- avgScore: scoreCount > 0 ? scoreSum / scoreCount : 0,
702
- };
703
- }
704
- /** 单测视角的 CSS。
705
- *
706
- * 设计原则:
707
- * - **scoped 浅色岛**:外层 layout 是深色主题(评分 tab 用),test view 在 .tv-root
708
- * 里覆盖一套浅色变量,实现"切到单测 tab = 切到杂志阅读模式"。
709
- * - **潘通启发的哑色调**:sage/dusty rose/mauve(饱和度 ~50%),不像 GitHub error log。
710
- * - **大字号 + 高行高**:正文 15px / 行高 1.75,代码 13px。中文优先 PingFang/YaHei。
711
- * - **subtle shadow 替代 border**:卡片间 16-24px 留白,box-shadow 浮起感。
712
- * - **default-collapsed**:通过的用例只显一行;失败/诱错展开诊断,其余 section 折叠。
713
- */
714
- export const TEST_VIEW_CSS = `
715
- /* tv-root: 浅色岛,scoped 覆盖外层深色 layout */
716
- .tv-root {
717
- --tv-bg:#fbfaf7; --tv-bg-card:#ffffff; --tv-bg-soft:rgba(58,58,58,.025);
718
- --tv-text:#3a3a3a; --tv-text-2:#7a7a7a; --tv-text-3:#a8a8a8;
719
- --tv-divider:#e8e3da;
720
- /* 加大 pass / fail / tripwire 的色相距离 + 暗度差,避免远看雷同。
721
- pass: 鼠尾草绿(暗一档,靠近林地);fail: 砖红土陶(暗一档,远离 pink);tripwire: 紫罗兰(独立色相)。 */
722
- --tv-pass:#5e8252; --tv-fail:#9c4a3f; --tv-tripwire:#7a6b89;
723
- --tv-pass-soft:rgba(94,130,82,.10); --tv-fail-soft:rgba(156,74,63,.10); --tv-tripwire-soft:rgba(122,107,137,.10);
724
- --tv-accent:#5a7a93; --tv-warn:#b08030;
725
- --tv-shadow:0 1px 3px rgba(58,58,58,.06);
726
- --tv-shadow-hover:0 2px 8px rgba(58,58,58,.10);
727
- background:var(--tv-bg); color:var(--tv-text);
728
- padding:32px 0; margin:-16px 0 0; border-radius:6px;
729
- font-family:-apple-system,BlinkMacSystemFont,"PingFang SC","Microsoft YaHei","Segoe UI",sans-serif;
730
- font-size:15px; line-height:1.75;
731
- }
732
- /* Variant tabs — 显式 reset 浏览器默认 button 样式(border-radius / outline / appearance)
733
- 防止 layout 那边的全局 button 规则给底部加了奇怪圆角和深色背景。 */
734
- .tv-root .tv-variant-tabs { display:flex;gap:0;margin:0 0 24px;border-bottom:1px solid var(--tv-divider) }
735
- .tv-root .tv-vtab { padding:12px 20px;background:transparent;border:none;border-radius:0;cursor:pointer;color:var(--tv-text-2);border-bottom:2px solid transparent;font-size:14px;font-family:inherit;font-weight:500;outline:none;appearance:none;-webkit-appearance:none;margin-bottom:-1px }
736
- .tv-root .tv-vtab:hover { color:var(--tv-text) }
737
- .tv-root .tv-vtab--active { color:var(--tv-text);border-bottom-color:var(--tv-accent);font-weight:600 }
738
- .tv-root .tv-vtab:focus-visible { outline:2px solid var(--tv-accent);outline-offset:-4px;border-radius:4px }
739
- /* tv-variant-panel: 不打横向 padding,内容直接对齐 .tv-root 边界,跟外层 omk-view-panel 的 h1/subtitle 同基线。 */
740
- /* Stats — 更扁平的数字+标签布局 */
741
- .tv-root .tv-variant-stats { display:flex;gap:32px;margin:0 0 24px;font-size:14px;color:var(--tv-text-2);align-items:baseline }
742
- .tv-root .tv-stat { display:flex;align-items:baseline;gap:6px }
743
- .tv-root .tv-stat-num { color:var(--tv-text);font-weight:600;font-size:18px;font-variant-numeric:tabular-nums }
744
- .tv-root .tv-stat--pass .tv-stat-num { color:var(--tv-pass) }
745
- .tv-root .tv-stat--fail .tv-stat-num { color:var(--tv-fail) }
746
- .tv-root .tv-stat--tripwire .tv-stat-num { color:var(--tv-tripwire) }
747
- /* Group separator */
748
- .tv-root .tv-group-sep { display:flex;align-items:center;gap:14px;margin:32px 0 16px;color:var(--tv-text-3);font-size:12px;text-transform:uppercase;letter-spacing:0.08em;font-weight:500 }
749
- .tv-root .tv-group-sep::after { content:'';flex:1;height:1px;background:var(--tv-divider) }
750
- /* Pass rows — one-line list,左边一道淡 sage 色条做"通过"暗示 */
751
- .tv-root .tv-rows { display:flex;flex-direction:column;gap:6px;margin:0 }
752
- .tv-root .tv-row { background:var(--tv-bg-card);border-radius:6px;box-shadow:var(--tv-shadow);transition:box-shadow .15s;border-left:3px solid var(--tv-pass) }
753
- .tv-root .tv-row:hover { box-shadow:var(--tv-shadow-hover) }
754
- .tv-root .tv-row[open] { box-shadow:var(--tv-shadow-hover) }
755
- .tv-root .tv-row-h { display:flex;align-items:center;gap:14px;padding:14px 20px;cursor:pointer;list-style:none;font-size:15px }
756
- .tv-root .tv-row-h::-webkit-details-marker { display:none }
757
- .tv-root .tv-row-h::after { content:'›';margin-left:8px;color:var(--tv-text-3);font-size:18px;transition:transform .15s }
758
- .tv-root .tv-row[open] .tv-row-h::after { transform:rotate(90deg) }
759
- /* Cards (fail / tripwire) — 默认折叠,跟 pass row 一致;头部下面加一行 preview 让折叠态也有信息密度 */
760
- .tv-root .tv-card { background:var(--tv-bg-card);border-radius:8px;box-shadow:var(--tv-shadow);margin:0 0 12px;transition:box-shadow .15s }
761
- .tv-root .tv-card:hover { box-shadow:var(--tv-shadow-hover) }
762
- .tv-root .tv-card--fail { border-left:3px solid var(--tv-fail) }
763
- .tv-root .tv-card--tripwire { border-left:3px solid var(--tv-tripwire) }
764
- .tv-root .tv-card-h { display:flex;align-items:center;gap:14px;padding:16px 22px;cursor:pointer;list-style:none;font-size:15px }
765
- .tv-root .tv-card-h::-webkit-details-marker { display:none }
766
- .tv-root .tv-card-h::after { content:'›';margin-left:8px;color:var(--tv-text-3);font-size:18px;transition:transform .15s }
767
- .tv-root .tv-card[open] .tv-card-h::after { transform:rotate(90deg) }
768
- .tv-root .tv-card[open] .tv-card-h { border-bottom:1px solid var(--tv-divider) }
769
- .tv-root .tv-card-preview { padding:0 22px 14px;color:var(--tv-text-2);font-size:14px;line-height:1.7;border-bottom:1px solid var(--tv-divider) }
770
- .tv-root .tv-card[open] .tv-card-preview { display:none }
771
- /* Status dots — 用形状 + 色 + 字符的三重区分,避免色盲 / 远看雷同。
772
- pass: 实心圆 sage;fail: 红 ✗ 字符(更显眼);tripwire: 紫色 ◆ 菱形。 */
773
- .tv-root .tv-dot { display:inline-flex;align-items:center;justify-content:center;width:18px;height:18px;flex-shrink:0;font-size:14px;font-weight:700;line-height:1 }
774
- .tv-root .tv-dot--pass::before { content:'';width:10px;height:10px;border-radius:50%;background:var(--tv-pass);box-shadow:0 0 0 3px var(--tv-pass-soft) }
775
- .tv-root .tv-dot--fail::before { content:'✗';color:var(--tv-fail);font-size:16px;font-weight:700 }
776
- .tv-root .tv-dot--tripwire::before { content:'◆';color:var(--tv-tripwire);font-size:13px }
777
- .tv-root .tv-id { font-family:"SF Mono","JetBrains Mono",Menlo,Consolas,monospace;font-size:13px;color:var(--tv-text-3) }
778
- .tv-root .tv-title { flex:1;font-size:15px;color:var(--tv-text);overflow:hidden;text-overflow:ellipsis;white-space:nowrap }
779
- .tv-root .tv-meta { font-size:13px;color:var(--tv-text-3);font-variant-numeric:tabular-nums;font-family:"SF Mono","JetBrains Mono",Menlo,monospace }
780
- /* Body padding */
781
- .tv-root .tv-card-body, .tv-root .tv-row-body { padding:22px }
782
- /* Tripwire banner — 卡片顶部的紫色解释带 */
783
- .tv-root .tv-tripwire-banner { margin:-22px -22px 20px;padding:18px 22px;background:rgba(139,124,147,.08);border-radius:8px 8px 0 0;border-bottom:1px solid var(--tv-divider) }
784
- .tv-root .tv-tripwire-banner-title { font-size:14px;font-weight:600;color:var(--tv-tripwire);margin-bottom:6px;letter-spacing:0.02em }
785
- .tv-root .tv-tripwire-banner-body { font-size:14px;color:var(--tv-text-2);line-height:1.7 }
786
- /* Diagnosis — 一段白话 + 单一蓝色建议块 + 折叠细节 */
787
- .tv-root .tv-diag { margin:4px 0 16px }
788
- .tv-root .tv-diag-text { font-size:15px;line-height:1.75;color:var(--tv-text);margin:0 0 16px }
789
- .tv-root .tv-suggestion { background:rgba(90,122,147,.06);border-left:3px solid var(--tv-accent);padding:14px 18px;margin:12px 0;border-radius:0 4px 4px 0 }
790
- .tv-root .tv-suggestion-label { display:inline-block;color:var(--tv-accent);font-weight:600;font-size:13px;margin-right:8px;letter-spacing:0.02em }
791
- .tv-root .tv-suggestion-body { margin:6px 0 0;font-size:14.5px;line-height:1.7;color:var(--tv-text) }
792
- .tv-root .tv-diag-details { margin-top:14px }
793
- .tv-root .tv-diag-details > summary { cursor:pointer;font-size:13px;color:var(--tv-text-2);list-style:none }
794
- .tv-root .tv-diag-details > summary::-webkit-details-marker { display:none }
795
- .tv-root .tv-diag-details > summary::before { content:'▸ ';color:var(--tv-text-3);display:inline-block;width:16px }
796
- .tv-root .tv-diag-details[open] > summary::before { content:'▾ ' }
797
- .tv-root .tv-diag-details-body { padding:10px 0 0 16px;font-size:14px;line-height:1.7;color:var(--tv-text-2) }
798
- .tv-root .tv-diag-details-body p { margin:6px 0 }
799
- .tv-root .tv-diag-details-body strong { color:var(--tv-text);margin-right:6px }
800
- /* RootCause inline pill */
801
- .tv-root .tv-diag-rc-pill { font-size:11px;padding:2px 8px;border-radius:10px;background:var(--tv-bg-soft);color:var(--tv-text-2);margin-right:6px;display:inline-block }
802
- .tv-root .tv-diag-rc--tripwire_intentional { background:rgba(139,124,147,.12);color:var(--tv-tripwire) }
803
- .tv-root .tv-diag-rc--sample_design { background:rgba(90,122,147,.10);color:var(--tv-accent) }
804
- .tv-root .tv-diag-rc--skill_doc_unclear, .tv-diag-rc--skill_doc_missing, .tv-root .tv-diag-rc--llm_misread { background:rgba(176,128,48,.10);color:var(--tv-warn) }
805
- /* Failure-mode chips — 中文标签,显著展示在 summary 下 */
806
- .tv-root .tv-diag-modes { display:flex;flex-wrap:wrap;align-items:center;gap:6px;margin:8px 0 14px }
807
- .tv-root .tv-diag-modes-label { font-size:12px;color:var(--tv-text-2);margin-right:4px }
808
- .tv-root .tv-diag-mode-pill { font-size:12px;padding:3px 10px;border-radius:11px;background:var(--tv-bg-soft);color:var(--tv-text);font-weight:500 }
809
- .tv-root .tv-diag-mode--workflow-skip { background:rgba(176,128,48,.12);color:var(--tv-warn) }
810
- .tv-root .tv-diag-mode--hardcoded { background:rgba(176,128,48,.12);color:var(--tv-warn) }
811
- .tv-root .tv-diag-mode--hallucination { background:rgba(180,75,75,.14);color:#b44b4b }
812
- .tv-root .tv-diag-mode--tool-misuse { background:rgba(176,128,48,.12);color:var(--tv-warn) }
813
- .tv-root .tv-diag-mode--env-blocked { background:rgba(90,122,147,.12);color:var(--tv-accent) }
814
- .tv-root .tv-diag-mode--misread { background:rgba(176,128,48,.12);color:var(--tv-warn) }
815
- .tv-root .tv-diag-mode--other { background:var(--tv-bg-soft);color:var(--tv-text-2) }
816
- /* Workflow checklist — 默认展开,放在 suggestion 之上,直观展示 LLM 走对哪几步 */
817
- .tv-root .tv-diag-workflow { margin:14px 0;padding:12px 14px;background:var(--tv-bg-soft);border-radius:6px }
818
- .tv-root .tv-diag-workflow-title { font-size:13px;font-weight:600;color:var(--tv-text);margin:0 0 10px;letter-spacing:0.02em }
819
- .tv-root .tv-diag-workflow-list { list-style:none;padding:0;margin:0;display:flex;flex-direction:column;gap:8px }
820
- .tv-root .tv-diag-workflow-item { display:flex;align-items:flex-start;gap:10px;line-height:1.6 }
821
- .tv-root .tv-diag-workflow-icon { flex-shrink:0;font-weight:700;font-size:14px;width:16px;text-align:center;margin-top:2px }
822
- .tv-root .tv-diag-workflow-item--pass .tv-diag-workflow-icon { color:#3a8a3a }
823
- .tv-root .tv-diag-workflow-item--fail .tv-diag-workflow-icon { color:#b44b4b }
824
- .tv-root .tv-diag-workflow-body { display:flex;flex-direction:column;gap:2px;flex:1 }
825
- .tv-root .tv-diag-workflow-step { font-size:14px;color:var(--tv-text) }
826
- .tv-root .tv-diag-workflow-evidence { font-size:12.5px;color:var(--tv-text-2);line-height:1.55 }
827
- .tv-root .tv-diag-workflow-item--fail .tv-diag-workflow-step { color:#b44b4b }
828
- /* Partial salvage 提示 */
829
- .tv-root .tv-diag-partial { font-size:12px;color:var(--tv-text-3);margin:0 0 8px;font-style:italic }
830
- /* Sub-sections — design / trace / assertions, 全部默认折叠 */
831
- .tv-root .tv-section { margin-top:6px;border-top:1px solid var(--tv-divider);padding-top:14px }
832
- .tv-root .tv-section > summary { cursor:pointer;font-size:14px;font-weight:500;color:var(--tv-text);padding:6px 0;list-style:none }
833
- .tv-root .tv-section > summary::-webkit-details-marker { display:none }
834
- .tv-root .tv-section > summary::before { content:'▸ ';color:var(--tv-text-3);display:inline-block;width:16px;font-weight:400 }
835
- .tv-root .tv-section[open] > summary::before { content:'▾ ' }
836
- .tv-root .tv-section-stat { font-weight:400;color:var(--tv-text-3);margin-left:6px;font-size:13px }
837
- .tv-root .tv-section-body { padding:14px 0 0 16px;font-size:14.5px;line-height:1.7 }
838
- /* Design fields */
839
- .tv-root .tv-prompt, .tv-root .tv-rubric { font-size:14.5px;line-height:1.75;margin:8px 0;color:var(--tv-text-2) }
840
- .tv-root .tv-prompt strong, .tv-root .tv-rubric strong { color:var(--tv-text);margin-right:6px;font-weight:600 }
841
- .tv-root .tv-meta-pills { margin:10px 0;display:flex;flex-wrap:wrap;gap:8px }
842
- .tv-root .tv-meta-pill { font-size:12px;padding:3px 10px;background:var(--tv-bg-soft);border-radius:10px;color:var(--tv-text-2);font-family:"SF Mono",Menlo,monospace }
843
- .tv-root .tv-expected { margin-top:12px }
844
- .tv-root .tv-expected-h { font-size:13px;color:var(--tv-text-2);margin-bottom:6px }
845
- .tv-root .tv-expect-line { font-size:13px;font-family:"SF Mono",Menlo,monospace;color:var(--tv-text-2);padding:3px 0;line-height:1.6 }
846
- .tv-root .tv-collapsible { margin-top:10px;font-size:13px }
847
- .tv-root .tv-collapsible > summary { cursor:pointer;color:var(--tv-text-2) }
848
- /* Mocks */
849
- .tv-root .tv-mocks { margin-top:8px }
850
- .tv-root .tv-mock-line { font-size:12px;font-family:"SF Mono",Menlo,monospace;padding:3px 0;display:flex;flex-wrap:wrap;gap:6px;align-items:center;line-height:1.6 }
851
- .tv-root .tv-mock-idx { color:var(--tv-text-3);font-weight:600 }
852
- .tv-root .tv-mock-tool { background:var(--tv-bg-soft);padding:2px 6px;border-radius:3px }
853
- .tv-root .tv-mock-match, .tv-root .tv-mock-return { color:var(--tv-text-2);overflow:hidden;text-overflow:ellipsis;max-width:340px;white-space:nowrap }
854
- .tv-root .tv-mock-arrow { color:var(--tv-text-3) }
855
- /* Trace turns */
856
- .tv-root .tv-turn { margin:10px 0;padding:12px 16px;background:var(--tv-bg-soft);border-radius:4px;border-left:2px solid var(--tv-divider) }
857
- .tv-root .tv-turn-h { font-size:14px;font-weight:500;color:var(--tv-text);margin-bottom:6px }
858
- .tv-root .tv-turn-dur { color:var(--tv-text-3);font-weight:400;font-size:12px;margin-left:6px }
859
- .tv-root .tv-thinking { margin:6px 0;font-size:13px }
860
- .tv-root .tv-thinking summary { cursor:pointer;color:var(--tv-text-3);font-style:italic }
861
- .tv-root .tv-thinking-body { padding:10px 14px;background:rgba(58,58,58,.04);border-radius:4px;color:var(--tv-text-2);white-space:pre-wrap;font-size:13px;line-height:1.7;margin-top:6px }
862
- /* Tool calls */
863
- .tv-root .tv-call { margin:6px 0;padding:10px 14px;background:rgba(58,58,58,.03);border-radius:4px;font-size:13px }
864
- .tv-root .tv-call--err { border-left:2px solid var(--tv-fail) }
865
- .tv-root .tv-call-h { font-weight:500;color:var(--tv-text);margin-bottom:4px;font-size:14px }
866
- .tv-root .tv-call details { margin:4px 0;font-size:12px }
867
- .tv-root .tv-call summary { cursor:pointer;color:var(--tv-text-3);padding:2px 0 }
868
- .tv-root .tv-pre { font-family:"SF Mono","JetBrains Mono",Menlo,Consolas,monospace;font-size:13px;background:rgba(58,58,58,.04);padding:10px 14px;border-radius:4px;overflow:auto;max-height:320px;white-space:pre-wrap;margin:4px 0;color:var(--tv-text-2);line-height:1.55 }
869
- .tv-root .tv-mock-badge { font-size:12px;color:var(--tv-pass);background:rgba(125,155,118,.12);padding:2px 8px;border-radius:10px;margin-left:8px }
870
- .tv-root .tv-trace-assert { font-size:13px;font-family:"SF Mono",Menlo,monospace;padding:3px 0;margin-top:6px;line-height:1.5 }
871
- .tv-root .tv-trace-assert.tv-assert--fail { color:var(--tv-fail) }
872
- .tv-root .tv-trace-assert.tv-assert--pass { color:var(--tv-text-2) }
873
- /* Final answer */
874
- .tv-root .tv-final { margin-top:14px }
875
- .tv-root .tv-final-h { font-size:14px;font-weight:500;color:var(--tv-text);margin-bottom:6px }
876
- /* Assertions */
877
- .tv-root .tv-assert { font-size:14px;font-family:"SF Mono",Menlo,monospace;padding:5px 0;line-height:1.6 }
878
- .tv-root .tv-assert--pass { color:var(--tv-text-2) }
879
- .tv-root .tv-assert--fail { color:var(--tv-fail) }
880
- .tv-root .tv-assert-value { background:rgba(58,58,58,.04);padding:2px 6px;border-radius:3px;margin:0 2px }
881
- .tv-root .tv-turn-tag { font-size:12px;color:var(--tv-text-3);margin-left:8px;font-style:italic;font-family:-apple-system,sans-serif }
882
- .tv-root .tv-turn-tag--ok { color:var(--tv-pass) }
883
- .tv-root .tv-turn-tag--miss { color:var(--tv-fail) }
884
- .tv-root .tv-assert-msg { font-size:13px;color:var(--tv-text-3);margin-left:20px;margin-top:3px }
885
- /* Empty/warn states */
886
- .tv-root .tv-empty, .tv-root .tv-empty-page { color:var(--tv-text-3);font-style:italic;padding:14px 0;font-size:14px }
887
- .tv-root .tv-warn { color:var(--tv-warn);background:rgba(176,128,48,.08);padding:14px 18px;border-radius:4px;margin:8px 24px;font-size:14px;line-height:1.7 }
888
- `;
889
- /** 单测视角配套 JS:variant 切换 + view tab 切换。挂在 layout 末尾。 */
890
- export const TEST_VIEW_JS = `
891
- function omkSwitchVariant(btn) {
892
- var target = btn.getAttribute('data-target');
893
- var tabs = btn.parentElement.querySelectorAll('.tv-vtab');
894
- tabs.forEach(function(t){ t.classList.toggle('tv-vtab--active', t === btn); });
895
- var panels = btn.parentElement.parentElement.querySelectorAll('.tv-variant-panel');
896
- panels.forEach(function(p){ p.hidden = (p.getAttribute('data-variant') !== target); });
897
- }
898
- function omkSwitchView(name) {
899
- var tabs = document.querySelectorAll('.omk-view-tab');
900
- tabs.forEach(function(t){ t.classList.toggle('omk-view-tab--active', t.getAttribute('data-view') === name); });
901
- var panels = document.querySelectorAll('.omk-view-panel');
902
- panels.forEach(function(p){ p.hidden = (p.getAttribute('data-view') !== name); });
903
- }
904
- `;
905
- //# sourceMappingURL=test-view.js.map