@prismatic-io/lux 0.0.1

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (636) hide show
  1. package/LICENSE +21 -0
  2. package/lib/answerers/claude-code/index.d.ts +17 -0
  3. package/lib/answerers/claude-code/index.d.ts.map +1 -0
  4. package/lib/answerers/claude-code/index.js +280 -0
  5. package/lib/answerers/claude-code/index.js.map +1 -0
  6. package/lib/answerers/persona/index.d.ts +40 -0
  7. package/lib/answerers/persona/index.d.ts.map +1 -0
  8. package/lib/answerers/persona/index.js +265 -0
  9. package/lib/answerers/persona/index.js.map +1 -0
  10. package/lib/answerers/scripted/index.d.ts +8 -0
  11. package/lib/answerers/scripted/index.d.ts.map +1 -0
  12. package/lib/answerers/scripted/index.js +163 -0
  13. package/lib/answerers/scripted/index.js.map +1 -0
  14. package/lib/answerers/shared.d.ts +13 -0
  15. package/lib/answerers/shared.d.ts.map +1 -0
  16. package/lib/answerers/shared.js +17 -0
  17. package/lib/answerers/shared.js.map +1 -0
  18. package/lib/answerers/terminal/index.d.ts +9 -0
  19. package/lib/answerers/terminal/index.d.ts.map +1 -0
  20. package/lib/answerers/terminal/index.js +97 -0
  21. package/lib/answerers/terminal/index.js.map +1 -0
  22. package/lib/assertions/core/command-exits-zero.d.ts +21 -0
  23. package/lib/assertions/core/command-exits-zero.d.ts.map +1 -0
  24. package/lib/assertions/core/command-exits-zero.js +87 -0
  25. package/lib/assertions/core/command-exits-zero.js.map +1 -0
  26. package/lib/assertions/core/event-checks.d.ts +57 -0
  27. package/lib/assertions/core/event-checks.d.ts.map +1 -0
  28. package/lib/assertions/core/event-checks.js +123 -0
  29. package/lib/assertions/core/event-checks.js.map +1 -0
  30. package/lib/assertions/core/file-checks.d.ts +103 -0
  31. package/lib/assertions/core/file-checks.d.ts.map +1 -0
  32. package/lib/assertions/core/file-checks.js +288 -0
  33. package/lib/assertions/core/file-checks.js.map +1 -0
  34. package/lib/assertions/core/helpers.d.ts +66 -0
  35. package/lib/assertions/core/helpers.d.ts.map +1 -0
  36. package/lib/assertions/core/helpers.js +170 -0
  37. package/lib/assertions/core/helpers.js.map +1 -0
  38. package/lib/assertions/core/index.d.ts +15 -0
  39. package/lib/assertions/core/index.d.ts.map +1 -0
  40. package/lib/assertions/core/index.js +56 -0
  41. package/lib/assertions/core/index.js.map +1 -0
  42. package/lib/assertions/core/json-pointer-equals.d.ts +19 -0
  43. package/lib/assertions/core/json-pointer-equals.d.ts.map +1 -0
  44. package/lib/assertions/core/json-pointer-equals.js +71 -0
  45. package/lib/assertions/core/json-pointer-equals.js.map +1 -0
  46. package/lib/assertions/core/output-checks.d.ts +46 -0
  47. package/lib/assertions/core/output-checks.d.ts.map +1 -0
  48. package/lib/assertions/core/output-checks.js +121 -0
  49. package/lib/assertions/core/output-checks.js.map +1 -0
  50. package/lib/assertions/core/predicate.d.ts +14 -0
  51. package/lib/assertions/core/predicate.d.ts.map +1 -0
  52. package/lib/assertions/core/predicate.js +55 -0
  53. package/lib/assertions/core/predicate.js.map +1 -0
  54. package/lib/assertions/core/resource-checks.d.ts +71 -0
  55. package/lib/assertions/core/resource-checks.d.ts.map +1 -0
  56. package/lib/assertions/core/resource-checks.js +159 -0
  57. package/lib/assertions/core/resource-checks.js.map +1 -0
  58. package/lib/assertions/core/safe-regex.d.ts +23 -0
  59. package/lib/assertions/core/safe-regex.d.ts.map +1 -0
  60. package/lib/assertions/core/safe-regex.js +102 -0
  61. package/lib/assertions/core/safe-regex.js.map +1 -0
  62. package/lib/assertions/core/tool-called-with.d.ts +21 -0
  63. package/lib/assertions/core/tool-called-with.d.ts.map +1 -0
  64. package/lib/assertions/core/tool-called-with.js +69 -0
  65. package/lib/assertions/core/tool-called-with.js.map +1 -0
  66. package/lib/assertions/core/tool-calls.d.ts +22 -0
  67. package/lib/assertions/core/tool-calls.d.ts.map +1 -0
  68. package/lib/assertions/core/tool-calls.js +17 -0
  69. package/lib/assertions/core/tool-calls.js.map +1 -0
  70. package/lib/assertions/core/tool-invocation.d.ts +36 -0
  71. package/lib/assertions/core/tool-invocation.d.ts.map +1 -0
  72. package/lib/assertions/core/tool-invocation.js +170 -0
  73. package/lib/assertions/core/tool-invocation.js.map +1 -0
  74. package/lib/assertions/core/tool-result-checks.d.ts +59 -0
  75. package/lib/assertions/core/tool-result-checks.d.ts.map +1 -0
  76. package/lib/assertions/core/tool-result-checks.js +193 -0
  77. package/lib/assertions/core/tool-result-checks.js.map +1 -0
  78. package/lib/assertions/mcp/discovery-checks.d.ts +36 -0
  79. package/lib/assertions/mcp/discovery-checks.d.ts.map +1 -0
  80. package/lib/assertions/mcp/discovery-checks.js +122 -0
  81. package/lib/assertions/mcp/discovery-checks.js.map +1 -0
  82. package/lib/assertions/mcp/index.d.ts +10 -0
  83. package/lib/assertions/mcp/index.d.ts.map +1 -0
  84. package/lib/assertions/mcp/index.js +25 -0
  85. package/lib/assertions/mcp/index.js.map +1 -0
  86. package/lib/assertions/rubric/index.d.ts +104 -0
  87. package/lib/assertions/rubric/index.d.ts.map +1 -0
  88. package/lib/assertions/rubric/index.js +191 -0
  89. package/lib/assertions/rubric/index.js.map +1 -0
  90. package/lib/assertions/rubric/internal.d.ts +14 -0
  91. package/lib/assertions/rubric/internal.d.ts.map +1 -0
  92. package/lib/assertions/rubric/internal.js +335 -0
  93. package/lib/assertions/rubric/internal.js.map +1 -0
  94. package/lib/authoring.d.ts +3 -0
  95. package/lib/authoring.d.ts.map +1 -0
  96. package/lib/authoring.js +3 -0
  97. package/lib/authoring.js.map +1 -0
  98. package/lib/cli/bin.d.ts +3 -0
  99. package/lib/cli/bin.d.ts.map +1 -0
  100. package/lib/cli/bin.js +15 -0
  101. package/lib/cli/bin.js.map +1 -0
  102. package/lib/cli/command-runtime.d.ts +5 -0
  103. package/lib/cli/command-runtime.d.ts.map +1 -0
  104. package/lib/cli/command-runtime.js +30 -0
  105. package/lib/cli/command-runtime.js.map +1 -0
  106. package/lib/cli/init.d.ts +21 -0
  107. package/lib/cli/init.d.ts.map +1 -0
  108. package/lib/cli/init.js +340 -0
  109. package/lib/cli/init.js.map +1 -0
  110. package/lib/cli/program.d.ts +22 -0
  111. package/lib/cli/program.d.ts.map +1 -0
  112. package/lib/cli/program.js +757 -0
  113. package/lib/cli/program.js.map +1 -0
  114. package/lib/cli/render/format.d.ts +14 -0
  115. package/lib/cli/render/format.d.ts.map +1 -0
  116. package/lib/cli/render/format.js +71 -0
  117. package/lib/cli/render/format.js.map +1 -0
  118. package/lib/cli/render/reporter.d.ts +73 -0
  119. package/lib/cli/render/reporter.d.ts.map +1 -0
  120. package/lib/cli/render/reporter.js +519 -0
  121. package/lib/cli/render/reporter.js.map +1 -0
  122. package/lib/cli/render/tty.d.ts +77 -0
  123. package/lib/cli/render/tty.d.ts.map +1 -0
  124. package/lib/cli/render/tty.js +283 -0
  125. package/lib/cli/render/tty.js.map +1 -0
  126. package/lib/cli/run-options.d.ts +27 -0
  127. package/lib/cli/run-options.d.ts.map +1 -0
  128. package/lib/cli/run-options.js +58 -0
  129. package/lib/cli/run-options.js.map +1 -0
  130. package/lib/cli/skills.d.ts +38 -0
  131. package/lib/cli/skills.d.ts.map +1 -0
  132. package/lib/cli/skills.js +84 -0
  133. package/lib/cli/skills.js.map +1 -0
  134. package/lib/core/annotation.d.ts +56 -0
  135. package/lib/core/annotation.d.ts.map +1 -0
  136. package/lib/core/annotation.js +92 -0
  137. package/lib/core/annotation.js.map +1 -0
  138. package/lib/core/answerer.d.ts +68 -0
  139. package/lib/core/answerer.d.ts.map +1 -0
  140. package/lib/core/answerer.js +17 -0
  141. package/lib/core/answerer.js.map +1 -0
  142. package/lib/core/artifact-evidence.d.ts +50 -0
  143. package/lib/core/artifact-evidence.d.ts.map +1 -0
  144. package/lib/core/artifact-evidence.js +185 -0
  145. package/lib/core/artifact-evidence.js.map +1 -0
  146. package/lib/core/artifact-snapshot.d.ts +37 -0
  147. package/lib/core/artifact-snapshot.d.ts.map +1 -0
  148. package/lib/core/artifact-snapshot.js +339 -0
  149. package/lib/core/artifact-snapshot.js.map +1 -0
  150. package/lib/core/assertion.d.ts +129 -0
  151. package/lib/core/assertion.d.ts.map +1 -0
  152. package/lib/core/assertion.js +75 -0
  153. package/lib/core/assertion.js.map +1 -0
  154. package/lib/core/case.d.ts +79 -0
  155. package/lib/core/case.d.ts.map +1 -0
  156. package/lib/core/case.js +90 -0
  157. package/lib/core/case.js.map +1 -0
  158. package/lib/core/conversation-turn.d.ts +23 -0
  159. package/lib/core/conversation-turn.d.ts.map +1 -0
  160. package/lib/core/conversation-turn.js +31 -0
  161. package/lib/core/conversation-turn.js.map +1 -0
  162. package/lib/core/driver.d.ts +64 -0
  163. package/lib/core/driver.d.ts.map +1 -0
  164. package/lib/core/driver.js +18 -0
  165. package/lib/core/driver.js.map +1 -0
  166. package/lib/core/emitter.d.ts +20 -0
  167. package/lib/core/emitter.d.ts.map +1 -0
  168. package/lib/core/emitter.js +34 -0
  169. package/lib/core/emitter.js.map +1 -0
  170. package/lib/core/events.d.ts +202 -0
  171. package/lib/core/events.d.ts.map +1 -0
  172. package/lib/core/events.js +107 -0
  173. package/lib/core/events.js.map +1 -0
  174. package/lib/core/exec.d.ts +82 -0
  175. package/lib/core/exec.d.ts.map +1 -0
  176. package/lib/core/exec.js +168 -0
  177. package/lib/core/exec.js.map +1 -0
  178. package/lib/core/experiment.d.ts +285 -0
  179. package/lib/core/experiment.d.ts.map +1 -0
  180. package/lib/core/experiment.js +389 -0
  181. package/lib/core/experiment.js.map +1 -0
  182. package/lib/core/glob.d.ts +3 -0
  183. package/lib/core/glob.d.ts.map +1 -0
  184. package/lib/core/glob.js +32 -0
  185. package/lib/core/glob.js.map +1 -0
  186. package/lib/core/hash.d.ts +8 -0
  187. package/lib/core/hash.d.ts.map +1 -0
  188. package/lib/core/hash.js +214 -0
  189. package/lib/core/hash.js.map +1 -0
  190. package/lib/core/index.d.ts +29 -0
  191. package/lib/core/index.d.ts.map +1 -0
  192. package/lib/core/index.js +29 -0
  193. package/lib/core/index.js.map +1 -0
  194. package/lib/core/lifecycle-fixtures.d.ts +20 -0
  195. package/lib/core/lifecycle-fixtures.d.ts.map +1 -0
  196. package/lib/core/lifecycle-fixtures.js +109 -0
  197. package/lib/core/lifecycle-fixtures.js.map +1 -0
  198. package/lib/core/lifecycle-hooks.d.ts +52 -0
  199. package/lib/core/lifecycle-hooks.d.ts.map +1 -0
  200. package/lib/core/lifecycle-hooks.js +42 -0
  201. package/lib/core/lifecycle-hooks.js.map +1 -0
  202. package/lib/core/mcp-events.d.ts +44 -0
  203. package/lib/core/mcp-events.d.ts.map +1 -0
  204. package/lib/core/mcp-events.js +23 -0
  205. package/lib/core/mcp-events.js.map +1 -0
  206. package/lib/core/model-cli.d.ts +57 -0
  207. package/lib/core/model-cli.d.ts.map +1 -0
  208. package/lib/core/model-cli.js +237 -0
  209. package/lib/core/model-cli.js.map +1 -0
  210. package/lib/core/observer.d.ts +31 -0
  211. package/lib/core/observer.d.ts.map +1 -0
  212. package/lib/core/observer.js +25 -0
  213. package/lib/core/observer.js.map +1 -0
  214. package/lib/core/platform-process.d.ts +33 -0
  215. package/lib/core/platform-process.d.ts.map +1 -0
  216. package/lib/core/platform-process.js +214 -0
  217. package/lib/core/platform-process.js.map +1 -0
  218. package/lib/core/registry.d.ts +22 -0
  219. package/lib/core/registry.d.ts.map +1 -0
  220. package/lib/core/registry.js +45 -0
  221. package/lib/core/registry.js.map +1 -0
  222. package/lib/core/rubric.d.ts +18 -0
  223. package/lib/core/rubric.d.ts.map +1 -0
  224. package/lib/core/rubric.js +49 -0
  225. package/lib/core/rubric.js.map +1 -0
  226. package/lib/core/run.d.ts +419 -0
  227. package/lib/core/run.d.ts.map +1 -0
  228. package/lib/core/run.js +126 -0
  229. package/lib/core/run.js.map +1 -0
  230. package/lib/core/statistics.d.ts +39 -0
  231. package/lib/core/statistics.d.ts.map +1 -0
  232. package/lib/core/statistics.js +163 -0
  233. package/lib/core/statistics.js.map +1 -0
  234. package/lib/core/subject.d.ts +4 -0
  235. package/lib/core/subject.d.ts.map +1 -0
  236. package/lib/core/subject.js +46 -0
  237. package/lib/core/subject.js.map +1 -0
  238. package/lib/core/tool-events.d.ts +69 -0
  239. package/lib/core/tool-events.d.ts.map +1 -0
  240. package/lib/core/tool-events.js +242 -0
  241. package/lib/core/tool-events.js.map +1 -0
  242. package/lib/core/usage.d.ts +78 -0
  243. package/lib/core/usage.d.ts.map +1 -0
  244. package/lib/core/usage.js +104 -0
  245. package/lib/core/usage.js.map +1 -0
  246. package/lib/core/wire.d.ts +83 -0
  247. package/lib/core/wire.d.ts.map +1 -0
  248. package/lib/core/wire.js +101 -0
  249. package/lib/core/wire.js.map +1 -0
  250. package/lib/core/zod.d.ts +10 -0
  251. package/lib/core/zod.d.ts.map +1 -0
  252. package/lib/core/zod.js +18 -0
  253. package/lib/core/zod.js.map +1 -0
  254. package/lib/drivers/claude-code/index.d.ts +55 -0
  255. package/lib/drivers/claude-code/index.d.ts.map +1 -0
  256. package/lib/drivers/claude-code/index.js +779 -0
  257. package/lib/drivers/claude-code/index.js.map +1 -0
  258. package/lib/drivers/claude-code/parse-events.d.ts +46 -0
  259. package/lib/drivers/claude-code/parse-events.d.ts.map +1 -0
  260. package/lib/drivers/claude-code/parse-events.js +387 -0
  261. package/lib/drivers/claude-code/parse-events.js.map +1 -0
  262. package/lib/drivers/codex/app-events.d.ts +16 -0
  263. package/lib/drivers/codex/app-events.d.ts.map +1 -0
  264. package/lib/drivers/codex/app-events.js +184 -0
  265. package/lib/drivers/codex/app-events.js.map +1 -0
  266. package/lib/drivers/codex/config.d.ts +55 -0
  267. package/lib/drivers/codex/config.d.ts.map +1 -0
  268. package/lib/drivers/codex/config.js +125 -0
  269. package/lib/drivers/codex/config.js.map +1 -0
  270. package/lib/drivers/codex/index.d.ts +8 -0
  271. package/lib/drivers/codex/index.d.ts.map +1 -0
  272. package/lib/drivers/codex/index.js +904 -0
  273. package/lib/drivers/codex/index.js.map +1 -0
  274. package/lib/drivers/mcp/index.d.ts +52 -0
  275. package/lib/drivers/mcp/index.d.ts.map +1 -0
  276. package/lib/drivers/mcp/index.js +282 -0
  277. package/lib/drivers/mcp/index.js.map +1 -0
  278. package/lib/drivers/mcp/transport.d.ts +50 -0
  279. package/lib/drivers/mcp/transport.d.ts.map +1 -0
  280. package/lib/drivers/mcp/transport.js +190 -0
  281. package/lib/drivers/mcp/transport.js.map +1 -0
  282. package/lib/drivers/shared/artifacts.d.ts +4 -0
  283. package/lib/drivers/shared/artifacts.d.ts.map +1 -0
  284. package/lib/drivers/shared/artifacts.js +81 -0
  285. package/lib/drivers/shared/artifacts.js.map +1 -0
  286. package/lib/drivers/shared/secure-copy.d.ts +2 -0
  287. package/lib/drivers/shared/secure-copy.d.ts.map +1 -0
  288. package/lib/drivers/shared/secure-copy.js +35 -0
  289. package/lib/drivers/shared/secure-copy.js.map +1 -0
  290. package/lib/drivers/subprocess/index.d.ts +24 -0
  291. package/lib/drivers/subprocess/index.d.ts.map +1 -0
  292. package/lib/drivers/subprocess/index.js +451 -0
  293. package/lib/drivers/subprocess/index.js.map +1 -0
  294. package/lib/index.d.ts +40 -0
  295. package/lib/index.d.ts.map +1 -0
  296. package/lib/index.js +48 -0
  297. package/lib/index.js.map +1 -0
  298. package/lib/optimization/reflective.d.ts +27 -0
  299. package/lib/optimization/reflective.d.ts.map +1 -0
  300. package/lib/optimization/reflective.js +257 -0
  301. package/lib/optimization/reflective.js.map +1 -0
  302. package/lib/orchestrator/annotation-loader.d.ts +65 -0
  303. package/lib/orchestrator/annotation-loader.d.ts.map +1 -0
  304. package/lib/orchestrator/annotation-loader.js +73 -0
  305. package/lib/orchestrator/annotation-loader.js.map +1 -0
  306. package/lib/orchestrator/annotation-store.d.ts +3 -0
  307. package/lib/orchestrator/annotation-store.d.ts.map +1 -0
  308. package/lib/orchestrator/annotation-store.js +100 -0
  309. package/lib/orchestrator/annotation-store.js.map +1 -0
  310. package/lib/orchestrator/authored-dependencies.d.ts +3 -0
  311. package/lib/orchestrator/authored-dependencies.d.ts.map +1 -0
  312. package/lib/orchestrator/authored-dependencies.js +146 -0
  313. package/lib/orchestrator/authored-dependencies.js.map +1 -0
  314. package/lib/orchestrator/campaign-lifecycle.d.ts +105 -0
  315. package/lib/orchestrator/campaign-lifecycle.d.ts.map +1 -0
  316. package/lib/orchestrator/campaign-lifecycle.js +141 -0
  317. package/lib/orchestrator/campaign-lifecycle.js.map +1 -0
  318. package/lib/orchestrator/candidate-integrity.d.ts +4 -0
  319. package/lib/orchestrator/candidate-integrity.d.ts.map +1 -0
  320. package/lib/orchestrator/candidate-integrity.js +14 -0
  321. package/lib/orchestrator/candidate-integrity.js.map +1 -0
  322. package/lib/orchestrator/candidates.d.ts +13 -0
  323. package/lib/orchestrator/candidates.d.ts.map +1 -0
  324. package/lib/orchestrator/candidates.js +577 -0
  325. package/lib/orchestrator/candidates.js.map +1 -0
  326. package/lib/orchestrator/compare.d.ts +80 -0
  327. package/lib/orchestrator/compare.d.ts.map +1 -0
  328. package/lib/orchestrator/compare.js +270 -0
  329. package/lib/orchestrator/compare.js.map +1 -0
  330. package/lib/orchestrator/comparison-identity.d.ts +10 -0
  331. package/lib/orchestrator/comparison-identity.d.ts.map +1 -0
  332. package/lib/orchestrator/comparison-identity.js +39 -0
  333. package/lib/orchestrator/comparison-identity.js.map +1 -0
  334. package/lib/orchestrator/config.d.ts +121 -0
  335. package/lib/orchestrator/config.d.ts.map +1 -0
  336. package/lib/orchestrator/config.js +176 -0
  337. package/lib/orchestrator/config.js.map +1 -0
  338. package/lib/orchestrator/content-bound-json.d.ts +5 -0
  339. package/lib/orchestrator/content-bound-json.d.ts.map +1 -0
  340. package/lib/orchestrator/content-bound-json.js +59 -0
  341. package/lib/orchestrator/content-bound-json.js.map +1 -0
  342. package/lib/orchestrator/discover.d.ts +45 -0
  343. package/lib/orchestrator/discover.d.ts.map +1 -0
  344. package/lib/orchestrator/discover.js +98 -0
  345. package/lib/orchestrator/discover.js.map +1 -0
  346. package/lib/orchestrator/doctor.d.ts +141 -0
  347. package/lib/orchestrator/doctor.d.ts.map +1 -0
  348. package/lib/orchestrator/doctor.js +525 -0
  349. package/lib/orchestrator/doctor.js.map +1 -0
  350. package/lib/orchestrator/driver-session.d.ts +177 -0
  351. package/lib/orchestrator/driver-session.d.ts.map +1 -0
  352. package/lib/orchestrator/driver-session.js +396 -0
  353. package/lib/orchestrator/driver-session.js.map +1 -0
  354. package/lib/orchestrator/evaluator-identity.d.ts +7 -0
  355. package/lib/orchestrator/evaluator-identity.d.ts.map +1 -0
  356. package/lib/orchestrator/evaluator-identity.js +42 -0
  357. package/lib/orchestrator/evaluator-identity.js.map +1 -0
  358. package/lib/orchestrator/experiment-budget.d.ts +11 -0
  359. package/lib/orchestrator/experiment-budget.d.ts.map +1 -0
  360. package/lib/orchestrator/experiment-budget.js +51 -0
  361. package/lib/orchestrator/experiment-budget.js.map +1 -0
  362. package/lib/orchestrator/experiment-context.d.ts +43 -0
  363. package/lib/orchestrator/experiment-context.d.ts.map +1 -0
  364. package/lib/orchestrator/experiment-context.js +9 -0
  365. package/lib/orchestrator/experiment-context.js.map +1 -0
  366. package/lib/orchestrator/experiment-contracts.d.ts +290 -0
  367. package/lib/orchestrator/experiment-contracts.d.ts.map +1 -0
  368. package/lib/orchestrator/experiment-contracts.js +38 -0
  369. package/lib/orchestrator/experiment-contracts.js.map +1 -0
  370. package/lib/orchestrator/experiment-corpus.d.ts +21 -0
  371. package/lib/orchestrator/experiment-corpus.d.ts.map +1 -0
  372. package/lib/orchestrator/experiment-corpus.js +276 -0
  373. package/lib/orchestrator/experiment-corpus.js.map +1 -0
  374. package/lib/orchestrator/experiment-evaluation.d.ts +44 -0
  375. package/lib/orchestrator/experiment-evaluation.d.ts.map +1 -0
  376. package/lib/orchestrator/experiment-evaluation.js +542 -0
  377. package/lib/orchestrator/experiment-evaluation.js.map +1 -0
  378. package/lib/orchestrator/experiment-finish.d.ts +6 -0
  379. package/lib/orchestrator/experiment-finish.d.ts.map +1 -0
  380. package/lib/orchestrator/experiment-finish.js +19 -0
  381. package/lib/orchestrator/experiment-finish.js.map +1 -0
  382. package/lib/orchestrator/experiment-identity.d.ts +6 -0
  383. package/lib/orchestrator/experiment-identity.d.ts.map +1 -0
  384. package/lib/orchestrator/experiment-identity.js +257 -0
  385. package/lib/orchestrator/experiment-identity.js.map +1 -0
  386. package/lib/orchestrator/experiment-loader.d.ts +175 -0
  387. package/lib/orchestrator/experiment-loader.d.ts.map +1 -0
  388. package/lib/orchestrator/experiment-loader.js +77 -0
  389. package/lib/orchestrator/experiment-loader.js.map +1 -0
  390. package/lib/orchestrator/experiment-report.d.ts +450 -0
  391. package/lib/orchestrator/experiment-report.d.ts.map +1 -0
  392. package/lib/orchestrator/experiment-report.js +570 -0
  393. package/lib/orchestrator/experiment-report.js.map +1 -0
  394. package/lib/orchestrator/experiment-runs.d.ts +379 -0
  395. package/lib/orchestrator/experiment-runs.d.ts.map +1 -0
  396. package/lib/orchestrator/experiment-runs.js +796 -0
  397. package/lib/orchestrator/experiment-runs.js.map +1 -0
  398. package/lib/orchestrator/experiment-runtime-identity.d.ts +41 -0
  399. package/lib/orchestrator/experiment-runtime-identity.d.ts.map +1 -0
  400. package/lib/orchestrator/experiment-runtime-identity.js +411 -0
  401. package/lib/orchestrator/experiment-runtime-identity.js.map +1 -0
  402. package/lib/orchestrator/experiment-selection.d.ts +25 -0
  403. package/lib/orchestrator/experiment-selection.d.ts.map +1 -0
  404. package/lib/orchestrator/experiment-selection.js +286 -0
  405. package/lib/orchestrator/experiment-selection.js.map +1 -0
  406. package/lib/orchestrator/experiment.d.ts +11 -0
  407. package/lib/orchestrator/experiment.d.ts.map +1 -0
  408. package/lib/orchestrator/experiment.js +639 -0
  409. package/lib/orchestrator/experiment.js.map +1 -0
  410. package/lib/orchestrator/fixture-identity.d.ts +7 -0
  411. package/lib/orchestrator/fixture-identity.d.ts.map +1 -0
  412. package/lib/orchestrator/fixture-identity.js +55 -0
  413. package/lib/orchestrator/fixture-identity.js.map +1 -0
  414. package/lib/orchestrator/fs-read.d.ts +19 -0
  415. package/lib/orchestrator/fs-read.d.ts.map +1 -0
  416. package/lib/orchestrator/fs-read.js +113 -0
  417. package/lib/orchestrator/fs-read.js.map +1 -0
  418. package/lib/orchestrator/grader-experiment.d.ts +190 -0
  419. package/lib/orchestrator/grader-experiment.d.ts.map +1 -0
  420. package/lib/orchestrator/grader-experiment.js +579 -0
  421. package/lib/orchestrator/grader-experiment.js.map +1 -0
  422. package/lib/orchestrator/grading.d.ts +31 -0
  423. package/lib/orchestrator/grading.d.ts.map +1 -0
  424. package/lib/orchestrator/grading.js +186 -0
  425. package/lib/orchestrator/grading.js.map +1 -0
  426. package/lib/orchestrator/index.d.ts +22 -0
  427. package/lib/orchestrator/index.d.ts.map +1 -0
  428. package/lib/orchestrator/index.js +30 -0
  429. package/lib/orchestrator/index.js.map +1 -0
  430. package/lib/orchestrator/lifecycle-hook-execution.d.ts +4 -0
  431. package/lib/orchestrator/lifecycle-hook-execution.d.ts.map +1 -0
  432. package/lib/orchestrator/lifecycle-hook-execution.js +67 -0
  433. package/lib/orchestrator/lifecycle-hook-execution.js.map +1 -0
  434. package/lib/orchestrator/list-runs.d.ts +33 -0
  435. package/lib/orchestrator/list-runs.d.ts.map +1 -0
  436. package/lib/orchestrator/list-runs.js +308 -0
  437. package/lib/orchestrator/list-runs.js.map +1 -0
  438. package/lib/orchestrator/loader.d.ts +3 -0
  439. package/lib/orchestrator/loader.d.ts.map +1 -0
  440. package/lib/orchestrator/loader.js +17 -0
  441. package/lib/orchestrator/loader.js.map +1 -0
  442. package/lib/orchestrator/optimization-lifecycle.d.ts +12 -0
  443. package/lib/orchestrator/optimization-lifecycle.d.ts.map +1 -0
  444. package/lib/orchestrator/optimization-lifecycle.js +92 -0
  445. package/lib/orchestrator/optimization-lifecycle.js.map +1 -0
  446. package/lib/orchestrator/orchestrator.d.ts +144 -0
  447. package/lib/orchestrator/orchestrator.d.ts.map +1 -0
  448. package/lib/orchestrator/orchestrator.js +632 -0
  449. package/lib/orchestrator/orchestrator.js.map +1 -0
  450. package/lib/orchestrator/owned-lock.d.ts +17 -0
  451. package/lib/orchestrator/owned-lock.d.ts.map +1 -0
  452. package/lib/orchestrator/owned-lock.js +149 -0
  453. package/lib/orchestrator/owned-lock.js.map +1 -0
  454. package/lib/orchestrator/package-version.d.ts +2 -0
  455. package/lib/orchestrator/package-version.d.ts.map +1 -0
  456. package/lib/orchestrator/package-version.js +4 -0
  457. package/lib/orchestrator/package-version.js.map +1 -0
  458. package/lib/orchestrator/provenance.d.ts +5 -0
  459. package/lib/orchestrator/provenance.d.ts.map +1 -0
  460. package/lib/orchestrator/provenance.js +137 -0
  461. package/lib/orchestrator/provenance.js.map +1 -0
  462. package/lib/orchestrator/regrade.d.ts +80 -0
  463. package/lib/orchestrator/regrade.d.ts.map +1 -0
  464. package/lib/orchestrator/regrade.js +433 -0
  465. package/lib/orchestrator/regrade.js.map +1 -0
  466. package/lib/orchestrator/rubric-loader.d.ts +7 -0
  467. package/lib/orchestrator/rubric-loader.d.ts.map +1 -0
  468. package/lib/orchestrator/rubric-loader.js +65 -0
  469. package/lib/orchestrator/rubric-loader.js.map +1 -0
  470. package/lib/orchestrator/run-dir.d.ts +40 -0
  471. package/lib/orchestrator/run-dir.d.ts.map +1 -0
  472. package/lib/orchestrator/run-dir.js +161 -0
  473. package/lib/orchestrator/run-dir.js.map +1 -0
  474. package/lib/orchestrator/run-execution.d.ts +30 -0
  475. package/lib/orchestrator/run-execution.d.ts.map +1 -0
  476. package/lib/orchestrator/run-execution.js +531 -0
  477. package/lib/orchestrator/run-execution.js.map +1 -0
  478. package/lib/orchestrator/run-lifecycle.d.ts +115 -0
  479. package/lib/orchestrator/run-lifecycle.d.ts.map +1 -0
  480. package/lib/orchestrator/run-lifecycle.js +136 -0
  481. package/lib/orchestrator/run-lifecycle.js.map +1 -0
  482. package/lib/orchestrator/suite-compare.d.ts +76 -0
  483. package/lib/orchestrator/suite-compare.d.ts.map +1 -0
  484. package/lib/orchestrator/suite-compare.js +159 -0
  485. package/lib/orchestrator/suite-compare.js.map +1 -0
  486. package/lib/orchestrator/suite-lifecycle.d.ts +157 -0
  487. package/lib/orchestrator/suite-lifecycle.d.ts.map +1 -0
  488. package/lib/orchestrator/suite-lifecycle.js +253 -0
  489. package/lib/orchestrator/suite-lifecycle.js.map +1 -0
  490. package/lib/orchestrator/suite-runs.d.ts +91 -0
  491. package/lib/orchestrator/suite-runs.d.ts.map +1 -0
  492. package/lib/orchestrator/suite-runs.js +157 -0
  493. package/lib/orchestrator/suite-runs.js.map +1 -0
  494. package/lib/orchestrator/suite-summary.d.ts +109 -0
  495. package/lib/orchestrator/suite-summary.d.ts.map +1 -0
  496. package/lib/orchestrator/suite-summary.js +213 -0
  497. package/lib/orchestrator/suite-summary.js.map +1 -0
  498. package/lib/orchestrator/trace-limits.d.ts +20 -0
  499. package/lib/orchestrator/trace-limits.d.ts.map +1 -0
  500. package/lib/orchestrator/trace-limits.js +28 -0
  501. package/lib/orchestrator/trace-limits.js.map +1 -0
  502. package/lib/orchestrator/validate-plugin-config.d.ts +10 -0
  503. package/lib/orchestrator/validate-plugin-config.d.ts.map +1 -0
  504. package/lib/orchestrator/validate-plugin-config.js +26 -0
  505. package/lib/orchestrator/validate-plugin-config.js.map +1 -0
  506. package/package.json +84 -0
  507. package/skills/lux/SKILL.md +85 -0
  508. package/skills/lux/references/cli.md +102 -0
  509. package/skills/lux/references/eval-authoring.md +105 -0
  510. package/skills/lux-answerer/SKILL.md +169 -0
  511. package/src/answerers/claude-code/index.ts +342 -0
  512. package/src/answerers/persona/index.ts +327 -0
  513. package/src/answerers/scripted/index.ts +186 -0
  514. package/src/answerers/shared.ts +23 -0
  515. package/src/answerers/terminal/index.ts +117 -0
  516. package/src/assertions/core/command-exits-zero.ts +94 -0
  517. package/src/assertions/core/event-checks.ts +138 -0
  518. package/src/assertions/core/file-checks.ts +336 -0
  519. package/src/assertions/core/helpers.ts +203 -0
  520. package/src/assertions/core/index.ts +80 -0
  521. package/src/assertions/core/json-pointer-equals.ts +84 -0
  522. package/src/assertions/core/output-checks.ts +134 -0
  523. package/src/assertions/core/predicate.ts +67 -0
  524. package/src/assertions/core/resource-checks.ts +180 -0
  525. package/src/assertions/core/safe-regex.ts +196 -0
  526. package/src/assertions/core/tool-called-with.ts +68 -0
  527. package/src/assertions/core/tool-calls.ts +28 -0
  528. package/src/assertions/core/tool-invocation.ts +198 -0
  529. package/src/assertions/core/tool-result-checks.ts +211 -0
  530. package/src/assertions/mcp/discovery-checks.ts +139 -0
  531. package/src/assertions/mcp/index.ts +37 -0
  532. package/src/assertions/rubric/index.ts +258 -0
  533. package/src/assertions/rubric/internal.ts +367 -0
  534. package/src/authoring.ts +2 -0
  535. package/src/cli/bin.ts +14 -0
  536. package/src/cli/command-runtime.ts +34 -0
  537. package/src/cli/init.ts +410 -0
  538. package/src/cli/program.ts +992 -0
  539. package/src/cli/render/format.ts +69 -0
  540. package/src/cli/render/reporter.ts +643 -0
  541. package/src/cli/render/tty.ts +314 -0
  542. package/src/cli/run-options.ts +82 -0
  543. package/src/cli/skills.ts +129 -0
  544. package/src/core/annotation.ts +104 -0
  545. package/src/core/answerer.ts +75 -0
  546. package/src/core/artifact-evidence.ts +249 -0
  547. package/src/core/artifact-snapshot.ts +429 -0
  548. package/src/core/assertion.ts +156 -0
  549. package/src/core/case.ts +103 -0
  550. package/src/core/conversation-turn.ts +47 -0
  551. package/src/core/driver.ts +75 -0
  552. package/src/core/emitter.ts +38 -0
  553. package/src/core/events.ts +135 -0
  554. package/src/core/exec.ts +263 -0
  555. package/src/core/experiment.ts +467 -0
  556. package/src/core/glob.ts +31 -0
  557. package/src/core/hash.ts +237 -0
  558. package/src/core/index.ts +28 -0
  559. package/src/core/lifecycle-fixtures.ts +145 -0
  560. package/src/core/lifecycle-hooks.ts +101 -0
  561. package/src/core/mcp-events.ts +57 -0
  562. package/src/core/model-cli.ts +271 -0
  563. package/src/core/observer.ts +50 -0
  564. package/src/core/platform-process.ts +284 -0
  565. package/src/core/registry.ts +57 -0
  566. package/src/core/rubric.ts +52 -0
  567. package/src/core/run.ts +144 -0
  568. package/src/core/statistics.ts +194 -0
  569. package/src/core/subject.ts +50 -0
  570. package/src/core/tool-events.ts +292 -0
  571. package/src/core/usage.ts +123 -0
  572. package/src/core/wire.ts +135 -0
  573. package/src/core/zod.ts +20 -0
  574. package/src/drivers/claude-code/index.ts +866 -0
  575. package/src/drivers/claude-code/parse-events.ts +464 -0
  576. package/src/drivers/codex/app-events.ts +218 -0
  577. package/src/drivers/codex/config.ts +128 -0
  578. package/src/drivers/codex/index.ts +995 -0
  579. package/src/drivers/mcp/index.ts +351 -0
  580. package/src/drivers/mcp/transport.ts +222 -0
  581. package/src/drivers/shared/artifacts.ts +86 -0
  582. package/src/drivers/shared/secure-copy.ts +40 -0
  583. package/src/drivers/subprocess/index.ts +490 -0
  584. package/src/index.ts +594 -0
  585. package/src/optimization/reflective.ts +319 -0
  586. package/src/orchestrator/annotation-loader.ts +97 -0
  587. package/src/orchestrator/annotation-store.ts +112 -0
  588. package/src/orchestrator/authored-dependencies.ts +193 -0
  589. package/src/orchestrator/campaign-lifecycle.ts +181 -0
  590. package/src/orchestrator/candidate-integrity.ts +18 -0
  591. package/src/orchestrator/candidates.ts +683 -0
  592. package/src/orchestrator/compare.ts +358 -0
  593. package/src/orchestrator/comparison-identity.ts +45 -0
  594. package/src/orchestrator/config.ts +205 -0
  595. package/src/orchestrator/content-bound-json.ts +71 -0
  596. package/src/orchestrator/discover.ts +132 -0
  597. package/src/orchestrator/doctor.ts +781 -0
  598. package/src/orchestrator/driver-session.ts +507 -0
  599. package/src/orchestrator/evaluator-identity.ts +45 -0
  600. package/src/orchestrator/experiment-budget.ts +61 -0
  601. package/src/orchestrator/experiment-context.ts +41 -0
  602. package/src/orchestrator/experiment-contracts.ts +43 -0
  603. package/src/orchestrator/experiment-corpus.ts +416 -0
  604. package/src/orchestrator/experiment-evaluation.ts +749 -0
  605. package/src/orchestrator/experiment-finish.ts +33 -0
  606. package/src/orchestrator/experiment-identity.ts +356 -0
  607. package/src/orchestrator/experiment-loader.ts +105 -0
  608. package/src/orchestrator/experiment-report.ts +788 -0
  609. package/src/orchestrator/experiment-runs.ts +960 -0
  610. package/src/orchestrator/experiment-runtime-identity.ts +465 -0
  611. package/src/orchestrator/experiment-selection.ts +467 -0
  612. package/src/orchestrator/experiment.ts +991 -0
  613. package/src/orchestrator/fixture-identity.ts +67 -0
  614. package/src/orchestrator/fs-read.ts +126 -0
  615. package/src/orchestrator/grader-experiment.ts +856 -0
  616. package/src/orchestrator/grading.ts +260 -0
  617. package/src/orchestrator/index.ts +186 -0
  618. package/src/orchestrator/lifecycle-hook-execution.ts +90 -0
  619. package/src/orchestrator/list-runs.ts +384 -0
  620. package/src/orchestrator/loader.ts +19 -0
  621. package/src/orchestrator/optimization-lifecycle.ts +126 -0
  622. package/src/orchestrator/orchestrator.ts +919 -0
  623. package/src/orchestrator/owned-lock.ts +147 -0
  624. package/src/orchestrator/package-version.ts +5 -0
  625. package/src/orchestrator/provenance.ts +140 -0
  626. package/src/orchestrator/regrade.ts +580 -0
  627. package/src/orchestrator/rubric-loader.ts +74 -0
  628. package/src/orchestrator/run-dir.ts +209 -0
  629. package/src/orchestrator/run-execution.ts +694 -0
  630. package/src/orchestrator/run-lifecycle.ts +174 -0
  631. package/src/orchestrator/suite-compare.ts +195 -0
  632. package/src/orchestrator/suite-lifecycle.ts +335 -0
  633. package/src/orchestrator/suite-runs.ts +202 -0
  634. package/src/orchestrator/suite-summary.ts +276 -0
  635. package/src/orchestrator/trace-limits.ts +65 -0
  636. package/src/orchestrator/validate-plugin-config.ts +34 -0
@@ -0,0 +1,202 @@
1
+ import { existsSync } from "node:fs";
2
+ import { mkdir, readdir, rename, writeFile } from "node:fs/promises";
3
+ import { basename, join, relative, resolve } from "node:path";
4
+ import { z } from "zod";
5
+ import { safeReadJson } from "./fs-read.js";
6
+
7
+ const SUITES_DIR = ".suites";
8
+ const SUITE_FILE = "suite.json";
9
+
10
+ const SuiteItemSchema = z.object({
11
+ caseId: z.string(),
12
+ canonicalCaseId: z.string(),
13
+ status: z.enum(["pending", "running", "completed"]),
14
+ runDir: z.string().nullable(),
15
+ exitReason: z.union([z.literal("done"), z.string()]).nullable(),
16
+ casePassed: z.boolean().nullable(),
17
+ errorMessage: z.exactOptional(z.string()),
18
+ });
19
+
20
+ export const SuiteRunSchema = z.object({
21
+ schemaVersion: z.literal(1),
22
+ suiteId: z.string(),
23
+ startedAt: z.string(),
24
+ endedAt: z.string().nullable(),
25
+ status: z.enum(["running", "passed", "failed", "aborted"]),
26
+ aborted: z.boolean(),
27
+ concurrency: z.number().int().positive(),
28
+ profileId: z.exactOptional(z.string()),
29
+ experimentId: z.exactOptional(z.string()),
30
+ experimentUnitId: z.exactOptional(z.string()),
31
+ experimentAttemptId: z.exactOptional(z.string()),
32
+ variantId: z.exactOptional(z.string()),
33
+ candidateId: z.exactOptional(z.string()),
34
+ subjectHash: z.exactOptional(z.string()),
35
+ errorMessage: z.exactOptional(z.string()),
36
+ fixtures: z
37
+ .array(
38
+ z.object({
39
+ name: z.string(),
40
+ pluginVersion: z.string(),
41
+ }),
42
+ )
43
+ .optional(),
44
+ items: z.array(SuiteItemSchema),
45
+ });
46
+ export type SuiteRun = z.infer<typeof SuiteRunSchema>;
47
+ export type SuiteRunItem = z.infer<typeof SuiteItemSchema>;
48
+
49
+ const timestampId = (date: Date): string => date.toISOString().replace(/[:.]/g, "-");
50
+
51
+ const writeJsonAtomic = async (path: string, value: unknown): Promise<void> => {
52
+ const temporary = `${path}.${process.pid}.tmp`;
53
+ await writeFile(temporary, `${JSON.stringify(value, null, 2)}\n`);
54
+ await rename(temporary, path);
55
+ };
56
+
57
+ export type OpenedSuiteRun = {
58
+ path: string;
59
+ manifest: SuiteRun;
60
+ write: () => Promise<void>;
61
+ relativeRunDir: (runDir: string) => string;
62
+ };
63
+
64
+ export const openSuiteRun = async (options: {
65
+ runsRoot: string;
66
+ concurrency: number;
67
+ items: { id: string; caseId?: string | undefined }[];
68
+ startedAt?: Date | undefined;
69
+ profileId?: string | undefined;
70
+ experimentId?: string | undefined;
71
+ experimentUnitId?: string | undefined;
72
+ experimentAttemptId?: string | undefined;
73
+ variantId?: string | undefined;
74
+ candidateId?: string | undefined;
75
+ subjectHash?: string | undefined;
76
+ fixtures?: Array<{ name: string; pluginVersion: string }> | undefined;
77
+ }): Promise<OpenedSuiteRun> => {
78
+ const startedAt = options.startedAt ?? new Date();
79
+ const suitesRoot = join(resolve(options.runsRoot), SUITES_DIR);
80
+ await mkdir(suitesRoot, { recursive: true });
81
+ const baseId = timestampId(startedAt);
82
+ let suiteId = baseId;
83
+ let path = join(suitesRoot, suiteId);
84
+ for (let n = 2; ; n++) {
85
+ try {
86
+ await mkdir(path);
87
+ break;
88
+ } catch (err) {
89
+ if ((err as NodeJS.ErrnoException).code !== "EEXIST") throw err;
90
+ suiteId = `${baseId}-${n}`;
91
+ path = join(suitesRoot, suiteId);
92
+ }
93
+ }
94
+ const manifest: SuiteRun = {
95
+ schemaVersion: 1,
96
+ suiteId,
97
+ startedAt: startedAt.toISOString(),
98
+ endedAt: null,
99
+ status: "running",
100
+ aborted: false,
101
+ concurrency: options.concurrency,
102
+ ...(options.profileId ? { profileId: options.profileId } : {}),
103
+ ...(options.experimentId ? { experimentId: options.experimentId } : {}),
104
+ ...(options.experimentUnitId ? { experimentUnitId: options.experimentUnitId } : {}),
105
+ ...(options.experimentAttemptId ? { experimentAttemptId: options.experimentAttemptId } : {}),
106
+ ...(options.variantId ? { variantId: options.variantId } : {}),
107
+ ...(options.candidateId ? { candidateId: options.candidateId } : {}),
108
+ ...(options.subjectHash ? { subjectHash: options.subjectHash } : {}),
109
+ ...(options.fixtures && options.fixtures.length > 0 ? { fixtures: options.fixtures } : {}),
110
+ items: options.items.map((item) => ({
111
+ caseId: item.id,
112
+ canonicalCaseId: item.caseId ?? item.id,
113
+ status: "pending",
114
+ runDir: null,
115
+ exitReason: null,
116
+ casePassed: null,
117
+ })),
118
+ };
119
+ const write = (): Promise<void> => writeJsonAtomic(join(path, SUITE_FILE), manifest);
120
+ await write();
121
+ return { path, manifest, write, relativeRunDir: (runDir) => relative(path, runDir) };
122
+ };
123
+
124
+ export const loadSuiteRun = async (path: string): Promise<SuiteRun | null> =>
125
+ safeReadJson(join(path, SUITE_FILE), SuiteRunSchema);
126
+
127
+ export const listSuiteRuns = async (
128
+ runsRoot: string,
129
+ ): Promise<Array<SuiteRun & { suiteDir: string }>> => {
130
+ const root = join(runsRoot, SUITES_DIR);
131
+ if (!existsSync(root)) return [];
132
+ const dirs = (await readdir(root, { withFileTypes: true }))
133
+ .filter((entry) => entry.isDirectory())
134
+ .map((entry) => join(root, entry.name));
135
+ const manifests = await Promise.all(
136
+ dirs.map(async (suiteDir) => ({ suiteDir, suite: await loadSuiteRun(suiteDir) })),
137
+ );
138
+ return manifests
139
+ .flatMap(({ suiteDir, suite }) => (suite ? [{ ...suite, suiteDir }] : []))
140
+ .sort((a, b) => b.suiteId.localeCompare(a.suiteId));
141
+ };
142
+
143
+ const resolveSuiteRunDir = (suiteDir: string, runDir: string): string => resolve(suiteDir, runDir);
144
+
145
+ export const formatSuiteRunsTable = (suites: Array<SuiteRun & { suiteDir: string }>): string => {
146
+ if (suites.length === 0) return "(no suite runs found)";
147
+ const rows = [
148
+ ["status", "suite", "cases", "passed", "failed", "ungraded", "pending"],
149
+ ...suites.map((suite) => [
150
+ suite.status.toUpperCase(),
151
+ basename(suite.suiteDir),
152
+ String(suite.items.length),
153
+ String(suite.items.filter((item) => item.casePassed === true).length),
154
+ String(
155
+ suite.items.filter(
156
+ (item) =>
157
+ item.status === "completed" &&
158
+ (item.exitReason !== "done" || item.casePassed === false),
159
+ ).length,
160
+ ),
161
+ String(
162
+ suite.items.filter(
163
+ (item) =>
164
+ item.status === "completed" && item.exitReason === "done" && item.casePassed === null,
165
+ ).length,
166
+ ),
167
+ String(suite.items.filter((item) => item.status !== "completed").length),
168
+ ]),
169
+ ];
170
+ const widths = (rows[0] ?? []).map((_, column) =>
171
+ Math.max(...rows.map((row) => row[column]?.length ?? 0)),
172
+ );
173
+ return rows
174
+ .map((row) => row.map((cell, column) => cell.padEnd(widths[column] ?? 0)).join(" "))
175
+ .join("\n");
176
+ };
177
+
178
+ export const formatSuiteRunDetail = (suiteDir: string, suite: SuiteRun): string => {
179
+ const lines = [
180
+ `suiteDir: ${suiteDir}`,
181
+ `suite: ${suite.suiteId}`,
182
+ `status: ${suite.status}`,
183
+ `started: ${suite.startedAt}`,
184
+ `ended: ${suite.endedAt ?? "—"}`,
185
+ `cases: ${suite.items.length}`,
186
+ ...(suite.errorMessage ? [`error: ${suite.errorMessage}`] : []),
187
+ "",
188
+ ];
189
+ for (const item of suite.items) {
190
+ const status = (() => {
191
+ if (item.status !== "completed") return item.status.toUpperCase();
192
+ if (item.exitReason !== "done" || item.casePassed === false) return "FAIL";
193
+ if (item.casePassed === true) return "PASS";
194
+ return "UNGRADED";
195
+ })();
196
+ lines.push(
197
+ `${status.padEnd(8)} ${item.caseId}${item.runDir ? ` ${resolveSuiteRunDir(suiteDir, item.runDir)}` : ""}`,
198
+ );
199
+ if (item.errorMessage) lines.push(` ${item.errorMessage}`);
200
+ }
201
+ return lines.join("\n");
202
+ };
@@ -0,0 +1,276 @@
1
+ import { rename, writeFile } from "node:fs/promises";
2
+ import { join } from "node:path";
3
+ import { z } from "zod";
4
+ import { harnessUsage, isUsageEmpty, type RunMetadata } from "../core/index.js";
5
+ import { runComparisonIdentity } from "./comparison-identity.js";
6
+ import { safeReadJson } from "./fs-read.js";
7
+ import type { SuiteCaseResult } from "./orchestrator.js";
8
+ import type { SuiteRun } from "./suite-runs.js";
9
+
10
+ const SUMMARY_FILE = "suite-summary.json";
11
+
12
+ export const MetricSummarySchema = z.object({
13
+ n: z.number().int().nonnegative(),
14
+ min: z.number(),
15
+ max: z.number(),
16
+ mean: z.number(),
17
+ median: z.number(),
18
+ p90: z.number(),
19
+ });
20
+ export type MetricSummary = z.infer<typeof MetricSummarySchema>;
21
+
22
+ const OutcomeSummarySchema = z.object({
23
+ attempts: z.number().int().nonnegative(),
24
+ completed: z.number().int().nonnegative(),
25
+ passed: z.number().int().nonnegative(),
26
+ failed: z.number().int().nonnegative(),
27
+ ungraded: z.number().int().nonnegative(),
28
+ pending: z.number().int().nonnegative(),
29
+ passRate: z.number().min(0).max(1),
30
+ });
31
+ export type OutcomeSummary = z.infer<typeof OutcomeSummarySchema>;
32
+
33
+ const CaseSummarySchema = OutcomeSummarySchema.extend({
34
+ canonicalCaseId: z.string(),
35
+ metrics: z.record(z.string(), MetricSummarySchema),
36
+ });
37
+ export type CaseSummary = z.infer<typeof CaseSummarySchema>;
38
+
39
+ export const SuiteSampleSchema = z.object({
40
+ index: z.number().int().nonnegative(),
41
+ caseId: z.string(),
42
+ canonicalCaseId: z.string(),
43
+ exitReason: z.string().nullable(),
44
+ passed: z.boolean().nullable(),
45
+ identityHash: z.string().nullable(),
46
+ score: z.exactOptional(z.number()),
47
+ metrics: z.record(z.string(), z.number()),
48
+ });
49
+ export type SuiteSample = z.infer<typeof SuiteSampleSchema>;
50
+
51
+ export const SuiteSummarySchema = z.object({
52
+ schemaVersion: z.literal(1),
53
+ suiteId: z.string(),
54
+ generatedAt: z.string(),
55
+ totals: OutcomeSummarySchema,
56
+ metrics: z.record(z.string(), MetricSummarySchema),
57
+ cases: z.array(CaseSummarySchema),
58
+ samples: z.array(SuiteSampleSchema),
59
+ });
60
+ export type SuiteSummary = z.infer<typeof SuiteSummarySchema>;
61
+
62
+ type MetricSamples = Map<string, number[]>;
63
+
64
+ const add = (samples: MetricSamples, name: string, value: number | undefined): void => {
65
+ if (value === undefined || !Number.isFinite(value)) return;
66
+ const values = samples.get(name) ?? [];
67
+ values.push(value);
68
+ samples.set(name, values);
69
+ };
70
+
71
+ const cacheHitRate = (usage: {
72
+ input: number;
73
+ cacheRead: number;
74
+ cacheCreation: number;
75
+ }): number | undefined => {
76
+ const promptTokens = usage.input + usage.cacheRead + usage.cacheCreation;
77
+ return promptTokens > 0 ? usage.cacheRead / promptTokens : undefined;
78
+ };
79
+
80
+ const collectMetadata = (samples: MetricSamples, metadata: RunMetadata): void => {
81
+ const started = Date.parse(metadata.startedAt);
82
+ const ended = Date.parse(metadata.endedAt);
83
+ if (!Number.isNaN(started) && !Number.isNaN(ended)) {
84
+ add(samples, "durationMs", Math.max(0, ended - started));
85
+ }
86
+
87
+ if (!isUsageEmpty(metadata.usage.agent)) {
88
+ add(samples, "agent.input", metadata.usage.agent.input);
89
+ add(samples, "agent.output", metadata.usage.agent.output);
90
+ add(samples, "agent.cacheRead", metadata.usage.agent.cacheRead);
91
+ add(samples, "agent.cacheCreation", metadata.usage.agent.cacheCreation);
92
+ add(samples, "agent.cacheHitRate", cacheHitRate(metadata.usage.agent));
93
+ add(
94
+ samples,
95
+ "agent.total",
96
+ metadata.usage.agent.input +
97
+ metadata.usage.agent.output +
98
+ metadata.usage.agent.cacheRead +
99
+ metadata.usage.agent.cacheCreation,
100
+ );
101
+ }
102
+ const harness = harnessUsage(metadata.usage);
103
+ if (!isUsageEmpty(harness)) {
104
+ add(samples, "harness.input", harness.input);
105
+ add(samples, "harness.output", harness.output);
106
+ add(samples, "harness.cacheRead", harness.cacheRead);
107
+ add(samples, "harness.cacheCreation", harness.cacheCreation);
108
+ add(samples, "harness.cacheHitRate", cacheHitRate(harness));
109
+ add(
110
+ samples,
111
+ "harness.total",
112
+ harness.input + harness.output + harness.cacheRead + harness.cacheCreation,
113
+ );
114
+ }
115
+ add(samples, "interrupts.ask", metadata.interruptCounts.ask);
116
+ add(samples, "interrupts.approve", metadata.interruptCounts.approve);
117
+ add(samples, "interrupts.custom", metadata.interruptCounts.custom);
118
+ add(samples, "toolCalls", metadata.toolOutcomes?.total ?? metadata.toolCallCount);
119
+ add(
120
+ samples,
121
+ "toolCalls.reported",
122
+ metadata.toolOutcomes?.reportedCalls ?? metadata.toolCallCount,
123
+ );
124
+ add(samples, "toolCalls.succeeded", metadata.toolOutcomes?.succeeded);
125
+ add(samples, "toolCalls.failed", metadata.toolOutcomes?.failed);
126
+ add(samples, "toolCalls.unresolved", metadata.toolOutcomes?.unresolved);
127
+ add(samples, "toolCalls.unmatchedResults", metadata.toolOutcomes?.unmatchedResults);
128
+ add(samples, "toolCalls.malformedEvents", metadata.toolOutcomes?.malformedEvents);
129
+ const settledToolCalls =
130
+ (metadata.toolOutcomes?.succeeded ?? 0) + (metadata.toolOutcomes?.failed ?? 0);
131
+ if (settledToolCalls > 0 && metadata.toolOutcomes) {
132
+ add(samples, "toolCalls.failureRate", metadata.toolOutcomes.failed / settledToolCalls);
133
+ }
134
+ add(samples, "artifacts.count", metadata.artifacts?.count);
135
+ add(samples, "artifacts.bytes", metadata.artifacts?.bytes);
136
+ add(samples, "cost.usd", metadata.cost?.usd);
137
+ add(samples, "phase.agentMs", metadata.phaseDurationMs?.agent);
138
+ add(samples, "phase.collectMs", metadata.phaseDurationMs?.collect);
139
+ add(samples, "phase.gradingMs", metadata.phaseDurationMs?.grading);
140
+ };
141
+
142
+ const percentile = (sorted: readonly number[], fraction: number): number => {
143
+ const index = Math.max(0, Math.ceil(sorted.length * fraction) - 1);
144
+ return sorted[index] ?? 0;
145
+ };
146
+
147
+ export const summarizeNumbers = (values: readonly number[]): MetricSummary => {
148
+ const sorted = [...values].sort((a, b) => a - b);
149
+ const middle = Math.floor(sorted.length / 2);
150
+ const median =
151
+ sorted.length % 2 === 0
152
+ ? ((sorted[middle - 1] ?? 0) + (sorted[middle] ?? 0)) / 2
153
+ : (sorted[middle] ?? 0);
154
+ return {
155
+ n: sorted.length,
156
+ min: sorted[0] ?? 0,
157
+ max: sorted.at(-1) ?? 0,
158
+ mean: sorted.reduce((sum, value) => sum + value, 0) / Math.max(sorted.length, 1),
159
+ median,
160
+ p90: percentile(sorted, 0.9),
161
+ };
162
+ };
163
+
164
+ const summarizeMetrics = (samples: MetricSamples): Record<string, MetricSummary> =>
165
+ Object.fromEntries(
166
+ [...samples.entries()]
167
+ .sort(([a], [b]) => a.localeCompare(b))
168
+ .map(([name, values]) => [name, summarizeNumbers(values)]),
169
+ );
170
+
171
+ const summarizeOutcomes = (
172
+ entries: Array<{
173
+ item: SuiteRun["items"][number];
174
+ result: SuiteCaseResult | undefined;
175
+ }>,
176
+ ): OutcomeSummary => {
177
+ const completed = entries.filter((entry) => entry.item.status === "completed").length;
178
+ const passed = entries.filter((entry) => entry.result?.casePassed === true).length;
179
+ const failed = entries.filter(
180
+ (entry) =>
181
+ entry.item.status === "completed" &&
182
+ (entry.result?.exitReason !== "done" || entry.result.casePassed === false),
183
+ ).length;
184
+ const ungraded = entries.filter(
185
+ (entry) =>
186
+ entry.item.status === "completed" &&
187
+ entry.result?.exitReason === "done" &&
188
+ entry.result.casePassed === null,
189
+ ).length;
190
+ return {
191
+ attempts: entries.length,
192
+ completed,
193
+ passed,
194
+ failed,
195
+ ungraded,
196
+ pending: entries.length - completed,
197
+ passRate: entries.length === 0 ? 0 : passed / entries.length,
198
+ };
199
+ };
200
+
201
+ export const buildSuiteSummary = (
202
+ suite: SuiteRun,
203
+ results: readonly (SuiteCaseResult | undefined)[],
204
+ generatedAt: Date = new Date(),
205
+ ): SuiteSummary => {
206
+ const entries = suite.items.map((item, index) => ({ item, result: results[index] }));
207
+ const allMetrics: MetricSamples = new Map();
208
+ for (const entry of entries) {
209
+ const result = entry.result?.result;
210
+ if (!result) continue;
211
+ add(allMetrics, "score", result.grading.caseScore ?? undefined);
212
+ collectMetadata(allMetrics, result.metadata);
213
+ }
214
+
215
+ const byCase = new Map<string, typeof entries>();
216
+ for (const entry of entries) {
217
+ const group = byCase.get(entry.item.canonicalCaseId) ?? [];
218
+ group.push(entry);
219
+ byCase.set(entry.item.canonicalCaseId, group);
220
+ }
221
+
222
+ const cases = [...byCase.entries()]
223
+ .sort(([a], [b]) => a.localeCompare(b))
224
+ .map(([canonicalCaseId, caseEntries]) => {
225
+ const metrics: MetricSamples = new Map();
226
+ for (const entry of caseEntries) {
227
+ const result = entry.result?.result;
228
+ if (!result) continue;
229
+ add(metrics, "score", result.grading.caseScore ?? undefined);
230
+ collectMetadata(metrics, result.metadata);
231
+ }
232
+ return {
233
+ canonicalCaseId,
234
+ ...summarizeOutcomes(caseEntries),
235
+ metrics: summarizeMetrics(metrics),
236
+ };
237
+ });
238
+
239
+ return {
240
+ schemaVersion: 1,
241
+ suiteId: suite.suiteId,
242
+ generatedAt: generatedAt.toISOString(),
243
+ totals: summarizeOutcomes(entries),
244
+ metrics: summarizeMetrics(allMetrics),
245
+ cases,
246
+ samples: entries.map((entry, index) => {
247
+ const samples: MetricSamples = new Map();
248
+ const result = entry.result?.result;
249
+ if (result) collectMetadata(samples, result.metadata);
250
+ return {
251
+ index,
252
+ caseId: entry.item.caseId,
253
+ canonicalCaseId: entry.item.canonicalCaseId,
254
+ exitReason: entry.result?.exitReason ?? null,
255
+ passed: entry.result?.casePassed ?? null,
256
+ identityHash: result ? runComparisonIdentity(result.metadata) : null,
257
+ ...(result?.grading.caseScore !== null && result?.grading.caseScore !== undefined
258
+ ? { score: result.grading.caseScore }
259
+ : {}),
260
+ metrics: Object.fromEntries(
261
+ [...samples.entries()].map(([name, values]) => [name, values[0] ?? 0]),
262
+ ),
263
+ };
264
+ }),
265
+ };
266
+ };
267
+
268
+ export const writeSuiteSummary = async (suiteDir: string, summary: SuiteSummary): Promise<void> => {
269
+ const path = join(suiteDir, SUMMARY_FILE);
270
+ const temporary = `${path}.${process.pid}.tmp`;
271
+ await writeFile(temporary, `${JSON.stringify(summary, null, 2)}\n`);
272
+ await rename(temporary, path);
273
+ };
274
+
275
+ export const loadSuiteSummary = async (suiteDir: string): Promise<SuiteSummary | null> =>
276
+ safeReadJson(join(suiteDir, SUMMARY_FILE), SuiteSummarySchema);
@@ -0,0 +1,65 @@
1
+ import type { RunEvent } from "../core/index.js";
2
+
3
+ export const MAX_RUN_TRACE_BYTES = 10 * 1024 * 1024;
4
+ export const MAX_RUN_TRACE_EVENTS = 50_000;
5
+
6
+ export type RunTraceLimits = {
7
+ maxBytes: number;
8
+ maxEvents: number;
9
+ };
10
+
11
+ export type RunTraceLimitFailure = {
12
+ code: "run-trace-limit";
13
+ limit: "bytes" | "events";
14
+ maxBytes: number;
15
+ maxEvents: number;
16
+ reason: string;
17
+ };
18
+
19
+ export const isRunTraceLimit = (value: unknown): value is Pick<RunTraceLimitFailure, "code"> =>
20
+ typeof value === "object" &&
21
+ value !== null &&
22
+ "code" in value &&
23
+ value.code === "run-trace-limit";
24
+
25
+ export const DEFAULT_RUN_TRACE_LIMITS: RunTraceLimits = {
26
+ maxBytes: MAX_RUN_TRACE_BYTES,
27
+ maxEvents: MAX_RUN_TRACE_EVENTS,
28
+ };
29
+
30
+ export const runTraceLimitFailure = (
31
+ limit: RunTraceLimitFailure["limit"],
32
+ limits: RunTraceLimits,
33
+ ): RunTraceLimitFailure => ({
34
+ code: "run-trace-limit",
35
+ limit,
36
+ maxBytes: limits.maxBytes,
37
+ maxEvents: limits.maxEvents,
38
+ reason:
39
+ limit === "events"
40
+ ? `run trace exceeded the ${limits.maxEvents}-event limit`
41
+ : `run trace exceeded the ${limits.maxBytes}-byte limit`,
42
+ });
43
+
44
+ export const runTraceLimitDiagnostic = (
45
+ failure: RunTraceLimitFailure,
46
+ ts: number = Date.now(),
47
+ ): RunEvent => ({
48
+ type: "error",
49
+ exitReason: "error",
50
+ reason: failure.reason,
51
+ ts,
52
+ });
53
+
54
+ const serializedEventBytes = (event: RunEvent): number =>
55
+ Buffer.byteLength(`${JSON.stringify(event)}\n`);
56
+
57
+ export const terminalDiagnosticReserveBytes = (limits: RunTraceLimits): number =>
58
+ Math.max(
59
+ serializedEventBytes(
60
+ runTraceLimitDiagnostic(runTraceLimitFailure("bytes", limits), Number.MAX_SAFE_INTEGER),
61
+ ),
62
+ serializedEventBytes(
63
+ runTraceLimitDiagnostic(runTraceLimitFailure("events", limits), Number.MAX_SAFE_INTEGER),
64
+ ),
65
+ );
@@ -0,0 +1,34 @@
1
+ import type { z } from "zod";
2
+ import { formatZodIssues } from "../core/index.js";
3
+
4
+ /**
5
+ * Centralized plugin/assertion config validation. Both the orchestrator
6
+ * (driver/answerer config) and the grader (assertion config) need to
7
+ * validate untrusted config blobs against a plugin-supplied Zod schema
8
+ * with consistent error formatting; this is the one place that does it.
9
+ */
10
+ export const validatePluginConfig = (
11
+ schema: z.ZodTypeAny | undefined,
12
+ config: unknown,
13
+ contextLabel: string,
14
+ ): unknown => {
15
+ if (!schema) return config;
16
+ const parsed = schema.safeParse(config ?? {});
17
+ if (!parsed.success) {
18
+ throw new Error(`invalid config for ${contextLabel}: ${formatZodIssues(parsed.error)}`);
19
+ }
20
+ return parsed.data;
21
+ };
22
+
23
+ export const validatePluginConfigAsync = async (
24
+ schema: z.ZodTypeAny | undefined,
25
+ config: unknown,
26
+ contextLabel: string,
27
+ ): Promise<unknown> => {
28
+ if (!schema) return config;
29
+ const parsed = await schema.safeParseAsync(config ?? {});
30
+ if (!parsed.success) {
31
+ throw new Error(`invalid config for ${contextLabel}: ${formatZodIssues(parsed.error)}`);
32
+ }
33
+ return parsed.data;
34
+ };