@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 +1,174 @@
1
+ import { assign, createActor, fromPromise, setup, toPromise } from "xstate";
2
+ import { isRunTraceLimit } from "./trace-limits.js";
3
+
4
+ export type RunFailurePhase = "setup" | "running" | "collection" | "grading" | "close" | "trace";
5
+
6
+ /**
7
+ * Side effects owned by the run lifecycle. The machine decides when each
8
+ * effect may run; implementations hold the filesystem and plugin details.
9
+ */
10
+ export type RunLifecycleEffects<Result> = {
11
+ prepare: () => Promise<void>;
12
+ start: () => Promise<void>;
13
+ waitForAgent: () => Promise<void>;
14
+ collect: () => Promise<void>;
15
+ grade: () => Promise<Result>;
16
+ recordFailure: (phase: RunFailurePhase, error: unknown) => Promise<void>;
17
+ close: (primaryError: unknown | null) => Promise<void>;
18
+ };
19
+
20
+ type ErasedEffects = RunLifecycleEffects<unknown>;
21
+ type RunLifecycleContext = {
22
+ effects: ErasedEffects;
23
+ phase: RunFailurePhase;
24
+ result: unknown;
25
+ error: unknown;
26
+ };
27
+ type RunLifecycleInput = { effects: ErasedEffects };
28
+ type RunLifecycleOutcome =
29
+ | { status: "succeeded"; result: unknown }
30
+ | { status: "failed"; error: unknown };
31
+
32
+ const invokeEffect = (effect: (effects: ErasedEffects) => Promise<unknown>) =>
33
+ fromPromise<unknown, ErasedEffects>(({ input }) => effect(input));
34
+
35
+ const runLifecycleMachine = setup({
36
+ types: {
37
+ context: {} as RunLifecycleContext,
38
+ input: {} as RunLifecycleInput,
39
+ output: {} as RunLifecycleOutcome,
40
+ },
41
+ actors: {
42
+ prepare: invokeEffect((effects) => effects.prepare()),
43
+ start: invokeEffect((effects) => effects.start()),
44
+ waitForAgent: invokeEffect((effects) => effects.waitForAgent()),
45
+ collect: invokeEffect((effects) => effects.collect()),
46
+ grade: invokeEffect((effects) => effects.grade()),
47
+ recordFailure: fromPromise<
48
+ void,
49
+ { effects: ErasedEffects; phase: RunFailurePhase; error: unknown }
50
+ >(({ input }) => input.effects.recordFailure(input.phase, input.error)),
51
+ closeSuccess: invokeEffect((effects) => effects.close(null)),
52
+ closeFailure: fromPromise<void, { effects: ErasedEffects; error: unknown }>(({ input }) =>
53
+ input.effects.close(input.error),
54
+ ),
55
+ },
56
+ actions: {
57
+ captureError: assign(({ context, event }) => {
58
+ const error = "error" in event ? event.error : new Error("run lifecycle failed");
59
+ return { error, phase: isRunTraceLimit(error) ? "trace" : context.phase };
60
+ }),
61
+ captureResult: assign({
62
+ result: ({ event }) => ("output" in event ? event.output : undefined),
63
+ }),
64
+ enterRunning: assign({ phase: "running" }),
65
+ enterCollection: assign({ phase: "collection" }),
66
+ enterGrading: assign({ phase: "grading" }),
67
+ enterClose: assign({ phase: "close" }),
68
+ },
69
+ }).createMachine({
70
+ id: "lux-run",
71
+ context: ({ input }) => ({
72
+ effects: input.effects,
73
+ phase: "setup",
74
+ result: undefined,
75
+ error: undefined,
76
+ }),
77
+ initial: "preparing",
78
+ states: {
79
+ preparing: {
80
+ invoke: {
81
+ src: "prepare",
82
+ input: ({ context }) => context.effects,
83
+ onDone: { target: "starting" },
84
+ onError: { target: "recordingFailure", actions: "captureError" },
85
+ },
86
+ },
87
+ starting: {
88
+ invoke: {
89
+ src: "start",
90
+ input: ({ context }) => context.effects,
91
+ onDone: { target: "running" },
92
+ onError: { target: "recordingFailure", actions: "captureError" },
93
+ },
94
+ },
95
+ recordingFailure: {
96
+ invoke: {
97
+ src: "recordFailure",
98
+ input: ({ context }) => ({
99
+ effects: context.effects,
100
+ phase: context.phase,
101
+ error: context.error,
102
+ }),
103
+ onDone: { target: "closingFailure" },
104
+ // Keep the original phase error as the terminal reason.
105
+ onError: { target: "closingFailure" },
106
+ },
107
+ },
108
+ running: {
109
+ entry: "enterRunning",
110
+ invoke: {
111
+ src: "waitForAgent",
112
+ input: ({ context }) => context.effects,
113
+ onDone: { target: "collecting" },
114
+ onError: { target: "recordingFailure", actions: "captureError" },
115
+ },
116
+ },
117
+ collecting: {
118
+ entry: "enterCollection",
119
+ invoke: {
120
+ src: "collect",
121
+ input: ({ context }) => context.effects,
122
+ onDone: { target: "grading" },
123
+ onError: { target: "recordingFailure", actions: "captureError" },
124
+ },
125
+ },
126
+ grading: {
127
+ entry: "enterGrading",
128
+ invoke: {
129
+ src: "grade",
130
+ input: ({ context }) => context.effects,
131
+ onDone: { target: "closingSuccess", actions: "captureResult" },
132
+ onError: { target: "recordingFailure", actions: "captureError" },
133
+ },
134
+ },
135
+ closingSuccess: {
136
+ entry: "enterClose",
137
+ invoke: {
138
+ src: "closeSuccess",
139
+ input: ({ context }) => context.effects,
140
+ onDone: { target: "succeeded" },
141
+ onError: { target: "recordingFailure", actions: "captureError" },
142
+ },
143
+ },
144
+ closingFailure: {
145
+ invoke: {
146
+ src: "closeFailure",
147
+ input: ({ context }) => ({ effects: context.effects, error: context.error }),
148
+ onDone: { target: "failed" },
149
+ onError: { target: "failed" },
150
+ },
151
+ },
152
+ succeeded: { type: "final" },
153
+ failed: { type: "final" },
154
+ },
155
+ output: ({ context }) => {
156
+ if (context.error !== undefined) {
157
+ return { status: "failed", error: context.error };
158
+ }
159
+ return { status: "succeeded", result: context.result };
160
+ },
161
+ });
162
+
163
+ export const executeRunLifecycle = async <Result>(
164
+ effects: RunLifecycleEffects<Result>,
165
+ ): Promise<Result> => {
166
+ const actor = createActor(runLifecycleMachine, { input: { effects } });
167
+ actor.start();
168
+ const outcome = await toPromise(actor);
169
+ if (outcome.status === "failed") throw outcome.error;
170
+ return outcome.result as Result;
171
+ };
172
+
173
+ // Internal export for transition-level tests.
174
+ export { runLifecycleMachine };
@@ -0,0 +1,195 @@
1
+ import { z } from "zod";
2
+ import {
3
+ loadSuiteSummary,
4
+ type MetricSummary,
5
+ MetricSummarySchema,
6
+ type SuiteSample,
7
+ type SuiteSummary,
8
+ summarizeNumbers,
9
+ } from "./suite-summary.js";
10
+
11
+ const ComparedOutcomeSchema = z
12
+ .object({
13
+ attempts: z.number().int().nonnegative(),
14
+ passRate: z.number().min(0).max(1),
15
+ })
16
+ .strict();
17
+
18
+ export const SuiteCaseComparisonSchema = z
19
+ .object({
20
+ canonicalCaseId: z.string(),
21
+ before: ComparedOutcomeSchema.nullable(),
22
+ after: ComparedOutcomeSchema.nullable(),
23
+ passRateDelta: z.number().nullable(),
24
+ regressed: z.boolean(),
25
+ pairedMetricDeltas: z.record(z.string(), MetricSummarySchema),
26
+ })
27
+ .strict();
28
+ export type SuiteCaseComparison = z.infer<typeof SuiteCaseComparisonSchema>;
29
+
30
+ export const SuiteComparisonSchema = z
31
+ .object({
32
+ pathA: z.string(),
33
+ pathB: z.string(),
34
+ passRateTolerance: z.number().nonnegative(),
35
+ before: ComparedOutcomeSchema,
36
+ after: ComparedOutcomeSchema,
37
+ passRateDelta: z.number(),
38
+ cases: z.array(SuiteCaseComparisonSchema),
39
+ identityMismatches: z.array(
40
+ z
41
+ .object({
42
+ canonicalCaseId: z.string(),
43
+ repetition: z.number().int().positive(),
44
+ before: z.string().nullable(),
45
+ after: z.string().nullable(),
46
+ })
47
+ .strict(),
48
+ ),
49
+ regressions: z.number().int().nonnegative(),
50
+ })
51
+ .strict();
52
+ export type SuiteComparison = z.infer<typeof SuiteComparisonSchema>;
53
+
54
+ const samplesByCase = (summary: SuiteSummary): Map<string, SuiteSample[]> => {
55
+ const out = new Map<string, SuiteSample[]>();
56
+ for (const sample of summary.samples) {
57
+ const entries = out.get(sample.canonicalCaseId) ?? [];
58
+ entries.push(sample);
59
+ out.set(sample.canonicalCaseId, entries);
60
+ }
61
+ return out;
62
+ };
63
+
64
+ const pairedMetricDeltas = (
65
+ before: readonly SuiteSample[],
66
+ after: readonly SuiteSample[],
67
+ ): Record<string, MetricSummary> => {
68
+ const deltas = new Map<string, number[]>();
69
+ for (let index = 0; index < Math.min(before.length, after.length); index++) {
70
+ const a = before[index];
71
+ const b = after[index];
72
+ if (!a || !b) continue;
73
+ const names = new Set([...Object.keys(a.metrics), ...Object.keys(b.metrics)]);
74
+ for (const name of names) {
75
+ const av = a.metrics[name];
76
+ const bv = b.metrics[name];
77
+ if (av === undefined || bv === undefined) continue;
78
+ const values = deltas.get(name) ?? [];
79
+ values.push(bv - av);
80
+ deltas.set(name, values);
81
+ }
82
+ }
83
+ return Object.fromEntries(
84
+ [...deltas.entries()]
85
+ .sort(([a], [b]) => a.localeCompare(b))
86
+ .map(([name, values]) => [name, summarizeNumbers(values)]),
87
+ );
88
+ };
89
+
90
+ export const compareSuiteSummaries = (
91
+ pathA: string,
92
+ a: SuiteSummary,
93
+ pathB: string,
94
+ b: SuiteSummary,
95
+ options: { passRateTolerance?: number } = {},
96
+ ): SuiteComparison => {
97
+ const passRateTolerance = options.passRateTolerance ?? 0;
98
+ const aCases = new Map(a.cases.map((entry) => [entry.canonicalCaseId, entry]));
99
+ const bCases = new Map(b.cases.map((entry) => [entry.canonicalCaseId, entry]));
100
+ const aSamples = samplesByCase(a);
101
+ const bSamples = samplesByCase(b);
102
+ const ids = new Set([...aCases.keys(), ...bCases.keys()]);
103
+ const identityMismatches = [...ids].sort().flatMap((canonicalCaseId) => {
104
+ const before = aSamples.get(canonicalCaseId) ?? [];
105
+ const after = bSamples.get(canonicalCaseId) ?? [];
106
+ return Array.from({ length: Math.max(before.length, after.length) }, (_, index) => {
107
+ const beforeHash = before[index]?.identityHash ?? null;
108
+ const afterHash = after[index]?.identityHash ?? null;
109
+ return beforeHash === afterHash && beforeHash !== null
110
+ ? []
111
+ : [
112
+ {
113
+ canonicalCaseId,
114
+ repetition: index + 1,
115
+ before: beforeHash,
116
+ after: afterHash,
117
+ },
118
+ ];
119
+ }).flat();
120
+ });
121
+ const cases = [...ids].sort().map((canonicalCaseId) => {
122
+ const before = aCases.get(canonicalCaseId);
123
+ const after = bCases.get(canonicalCaseId);
124
+ let passRateDelta: number | null = null;
125
+ if (before && after) {
126
+ passRateDelta = after.passRate - before.passRate;
127
+ } else if (before) {
128
+ passRateDelta = -before.passRate;
129
+ }
130
+ return {
131
+ canonicalCaseId,
132
+ before: before ? { attempts: before.attempts, passRate: before.passRate } : null,
133
+ after: after ? { attempts: after.attempts, passRate: after.passRate } : null,
134
+ passRateDelta,
135
+ regressed:
136
+ Boolean(before) && (!after || after.passRate + passRateTolerance < (before?.passRate ?? 0)),
137
+ pairedMetricDeltas: pairedMetricDeltas(
138
+ aSamples.get(canonicalCaseId) ?? [],
139
+ bSamples.get(canonicalCaseId) ?? [],
140
+ ),
141
+ };
142
+ });
143
+ return SuiteComparisonSchema.parse({
144
+ pathA,
145
+ pathB,
146
+ passRateTolerance,
147
+ before: { attempts: a.totals.attempts, passRate: a.totals.passRate },
148
+ after: { attempts: b.totals.attempts, passRate: b.totals.passRate },
149
+ passRateDelta: b.totals.passRate - a.totals.passRate,
150
+ cases,
151
+ identityMismatches,
152
+ regressions: cases.filter((entry) => entry.regressed).length,
153
+ });
154
+ };
155
+
156
+ export const compareSuiteRuns = async (
157
+ pathA: string,
158
+ pathB: string,
159
+ options: { passRateTolerance?: number } = {},
160
+ ): Promise<SuiteComparison> => {
161
+ const [a, b] = await Promise.all([loadSuiteSummary(pathA), loadSuiteSummary(pathB)]);
162
+ if (!a) throw new Error(`suite summary not found at ${pathA}`);
163
+ if (!b) throw new Error(`suite summary not found at ${pathB}`);
164
+ return compareSuiteSummaries(pathA, a, pathB, b, options);
165
+ };
166
+
167
+ export const hasSuiteRegression = (comparison: SuiteComparison): boolean =>
168
+ comparison.regressions > 0 || comparison.identityMismatches.length > 0;
169
+
170
+ const percent = (value: number): string => `${(value * 100).toFixed(1)}%`;
171
+
172
+ export const formatSuiteComparisonText = (comparison: SuiteComparison): string => {
173
+ const lines = [
174
+ `A: ${comparison.pathA}`,
175
+ `B: ${comparison.pathB}`,
176
+ "",
177
+ "Suite:",
178
+ ` pass rate: ${percent(comparison.before.passRate)} → ${percent(comparison.after.passRate)} (${comparison.passRateDelta >= 0 ? "+" : ""}${percent(comparison.passRateDelta)})`,
179
+ ` attempts: ${comparison.before.attempts} → ${comparison.after.attempts}`,
180
+ ` regressions: ${comparison.regressions}`,
181
+ ];
182
+ if (comparison.identityMismatches.length > 0) {
183
+ lines.push(` identity mismatches: ${comparison.identityMismatches.length}`);
184
+ for (const mismatch of comparison.identityMismatches) {
185
+ lines.push(` ! ${mismatch.canonicalCaseId} repetition ${mismatch.repetition}`);
186
+ }
187
+ }
188
+ for (const entry of comparison.cases) {
189
+ if (!entry.regressed) continue;
190
+ lines.push(
191
+ ` ↓ ${entry.canonicalCaseId}: ${entry.before ? percent(entry.before.passRate) : "—"} → ${entry.after ? percent(entry.after.passRate) : "missing"}`,
192
+ );
193
+ }
194
+ return lines.join("\n");
195
+ };
@@ -0,0 +1,335 @@
1
+ import { assign, createActor, fromCallback, fromPromise, sendTo, setup, toPromise } from "xstate";
2
+
3
+ type SuiteLifecycleStatus = "succeeded" | "failed" | "aborted";
4
+ type SuiteItemLifecycleStatus = "pending" | "running" | "completed";
5
+
6
+ type SuiteLifecycleItem<Result> = {
7
+ status: SuiteItemLifecycleStatus;
8
+ result: Result | null;
9
+ };
10
+
11
+ export type SuiteLifecycleFinalization<Result> = {
12
+ status: SuiteLifecycleStatus;
13
+ items: readonly SuiteLifecycleItem<Result>[];
14
+ results: readonly Result[];
15
+ error: unknown | null;
16
+ };
17
+
18
+ export type SuiteLifecycleEffects<Result> = {
19
+ itemStarted: (index: number) => Promise<void>;
20
+ runItem: (index: number) => Promise<Result>;
21
+ itemCompleted: (index: number, result: Result) => Promise<void>;
22
+ resultFailed: (result: Result) => boolean;
23
+ finalize: (finalization: SuiteLifecycleFinalization<Result>) => Promise<void>;
24
+ recordFinalizationFailure: (finalization: SuiteLifecycleFinalization<Result>) => Promise<void>;
25
+ };
26
+
27
+ type ErasedEffects = SuiteLifecycleEffects<unknown>;
28
+ type SuiteLifecycleInput = {
29
+ itemCount: number;
30
+ concurrency: number;
31
+ abortSignal?: AbortSignal;
32
+ effects: ErasedEffects;
33
+ };
34
+ type SuiteLifecycleContext = {
35
+ itemCount: number;
36
+ concurrency: number;
37
+ abortSignal?: AbortSignal;
38
+ effects: ErasedEffects;
39
+ items: SuiteLifecycleItem<unknown>[];
40
+ nextIndex: number;
41
+ runningCount: number;
42
+ aborted: boolean;
43
+ status: SuiteLifecycleStatus | null;
44
+ error: unknown | null;
45
+ };
46
+ type WorkerPoolInput = {
47
+ concurrency: number;
48
+ abortSignal?: AbortSignal;
49
+ effects: ErasedEffects;
50
+ };
51
+ type WorkerPoolCommand =
52
+ | { type: "startItem"; workerId: number; index: number }
53
+ | { type: "completeItem"; workerId: number; index: number; result: unknown };
54
+ type SuiteLifecycleEvent =
55
+ | { type: "workerReady"; workerId: number }
56
+ | { type: "itemFinished"; workerId: number; index: number; result: unknown }
57
+ | { type: "workerFailed"; index: number; error: unknown }
58
+ | { type: "abort" };
59
+
60
+ const workerPool = fromCallback<WorkerPoolCommand, WorkerPoolInput>(
61
+ ({ input, receive, sendBack }) => {
62
+ const executeCommand = async (event: WorkerPoolCommand): Promise<void> => {
63
+ try {
64
+ if (event.type === "startItem") {
65
+ await input.effects.itemStarted(event.index);
66
+ const result = await input.effects.runItem(event.index);
67
+ sendBack({
68
+ type: "itemFinished",
69
+ workerId: event.workerId,
70
+ index: event.index,
71
+ result,
72
+ });
73
+ return;
74
+ }
75
+ await input.effects.itemCompleted(event.index, event.result);
76
+ sendBack({ type: "workerReady", workerId: event.workerId });
77
+ } catch (error) {
78
+ sendBack({ type: "workerFailed", index: event.index, error });
79
+ }
80
+ };
81
+ receive((event) => void executeCommand(event));
82
+
83
+ for (let workerId = 0; workerId < input.concurrency; workerId++) {
84
+ sendBack({ type: "workerReady", workerId });
85
+ }
86
+
87
+ const abort = (): void => sendBack({ type: "abort" });
88
+ if (input.abortSignal?.aborted) abort();
89
+ else input.abortSignal?.addEventListener("abort", abort, { once: true });
90
+ return () => input.abortSignal?.removeEventListener("abort", abort);
91
+ },
92
+ );
93
+
94
+ const finalizationOf = (context: SuiteLifecycleContext): SuiteLifecycleFinalization<unknown> => ({
95
+ status: context.status ?? "failed",
96
+ items: context.items,
97
+ results: context.items.flatMap((item) =>
98
+ item.status === "completed" && item.result !== null ? [item.result] : [],
99
+ ),
100
+ error: context.error,
101
+ });
102
+
103
+ const selectedFinalStatus = (context: SuiteLifecycleContext): SuiteLifecycleStatus => {
104
+ if (context.error !== null) return "failed";
105
+ if (context.aborted || context.abortSignal?.aborted) return "aborted";
106
+ const resultFailed = context.items.some(
107
+ (item) =>
108
+ item.status === "completed" &&
109
+ item.result !== null &&
110
+ context.effects.resultFailed(item.result),
111
+ );
112
+ return resultFailed ? "failed" : "succeeded";
113
+ };
114
+
115
+ const suiteLifecycleMachine = setup({
116
+ types: {
117
+ context: {} as SuiteLifecycleContext,
118
+ input: {} as SuiteLifecycleInput,
119
+ events: {} as SuiteLifecycleEvent,
120
+ output: {} as SuiteLifecycleFinalization<unknown>,
121
+ },
122
+ actors: {
123
+ workerPool,
124
+ finalize: fromPromise<
125
+ void,
126
+ { effects: ErasedEffects; finalization: SuiteLifecycleFinalization<unknown> }
127
+ >(({ input }) => input.effects.finalize(input.finalization)),
128
+ recordFinalizationFailure: fromPromise<
129
+ void,
130
+ { effects: ErasedEffects; finalization: SuiteLifecycleFinalization<unknown> }
131
+ >(({ input }) => input.effects.recordFinalizationFailure(input.finalization)),
132
+ },
133
+ guards: {
134
+ canStartItem: ({ context }) =>
135
+ context.error === null &&
136
+ !context.aborted &&
137
+ !context.abortSignal?.aborted &&
138
+ context.nextIndex < context.itemCount,
139
+ readyToFinalize: ({ context }) =>
140
+ context.runningCount === 0 &&
141
+ (context.error !== null ||
142
+ context.aborted ||
143
+ context.abortSignal?.aborted ||
144
+ context.nextIndex >= context.itemCount),
145
+ failureLeavesNoRunningItems: ({ context, event }) =>
146
+ event.type === "workerFailed" &&
147
+ context.runningCount === (context.items[event.index]?.status === "running" ? 1 : 0),
148
+ abortLeavesNoRunningItems: ({ context }) => context.runningCount === 0,
149
+ finalizedAsAborted: ({ context }) => context.status === "aborted",
150
+ finalizedAsFailed: ({ context }) => context.status === "failed",
151
+ },
152
+ actions: {
153
+ dispatchNextItem: sendTo("workers", ({ context, event }) => {
154
+ if (event.type !== "workerReady") {
155
+ throw new Error("suite worker dispatch requires an available worker");
156
+ }
157
+ return {
158
+ type: "startItem",
159
+ workerId: event.workerId,
160
+ index: context.nextIndex,
161
+ };
162
+ }),
163
+ markNextItemRunning: assign(({ context }) => ({
164
+ items: context.items.map((item, index) =>
165
+ index === context.nextIndex ? { ...item, status: "running" as const } : item,
166
+ ),
167
+ nextIndex: context.nextIndex + 1,
168
+ runningCount: context.runningCount + 1,
169
+ })),
170
+ markItemCompleted: assign(({ context, event }) => {
171
+ if (event.type !== "itemFinished") return {};
172
+ return {
173
+ items: context.items.map((item, index) =>
174
+ index === event.index ? { status: "completed" as const, result: event.result } : item,
175
+ ),
176
+ runningCount: context.runningCount - 1,
177
+ };
178
+ }),
179
+ persistCompletedItem: sendTo("workers", ({ event }) => {
180
+ if (event.type !== "itemFinished") {
181
+ throw new Error("suite item completion requires a finished item");
182
+ }
183
+ return {
184
+ type: "completeItem",
185
+ workerId: event.workerId,
186
+ index: event.index,
187
+ result: event.result,
188
+ };
189
+ }),
190
+ rememberAbort: assign({ aborted: true }),
191
+ selectFinalStatus: assign(({ context }) => ({
192
+ aborted: context.aborted || Boolean(context.abortSignal?.aborted),
193
+ status: selectedFinalStatus(context),
194
+ })),
195
+ rememberFailure: assign(({ context, event }) => {
196
+ if (event.type !== "workerFailed") return {};
197
+ const wasRunning = context.items[event.index]?.status === "running";
198
+ return {
199
+ error: context.error ?? event.error,
200
+ runningCount: wasRunning ? context.runningCount - 1 : context.runningCount,
201
+ };
202
+ }),
203
+ rememberFinalizationFailure: assign(({ event }) => ({
204
+ error: "error" in event ? event.error : new Error("suite finalization failed"),
205
+ status: "failed" as const,
206
+ })),
207
+ },
208
+ }).createMachine({
209
+ id: "lux-suite",
210
+ context: ({ input }) => ({
211
+ itemCount: input.itemCount,
212
+ concurrency: input.concurrency,
213
+ ...(input.abortSignal ? { abortSignal: input.abortSignal } : {}),
214
+ effects: input.effects,
215
+ items: Array.from({ length: input.itemCount }, () => ({
216
+ status: "pending" as const,
217
+ result: null,
218
+ })),
219
+ nextIndex: 0,
220
+ runningCount: 0,
221
+ aborted: input.abortSignal?.aborted ?? false,
222
+ status: null,
223
+ error: null,
224
+ }),
225
+ initial: "running",
226
+ states: {
227
+ running: {
228
+ invoke: {
229
+ id: "workers",
230
+ src: "workerPool",
231
+ input: ({ context }) => ({
232
+ concurrency: context.concurrency,
233
+ ...(context.abortSignal ? { abortSignal: context.abortSignal } : {}),
234
+ effects: context.effects,
235
+ }),
236
+ },
237
+ on: {
238
+ workerReady: [
239
+ {
240
+ guard: "readyToFinalize",
241
+ target: "finalizing",
242
+ actions: "selectFinalStatus",
243
+ },
244
+ {
245
+ guard: "canStartItem",
246
+ actions: ["dispatchNextItem", "markNextItemRunning"],
247
+ },
248
+ ],
249
+ itemFinished: {
250
+ actions: ["markItemCompleted", "persistCompletedItem"],
251
+ },
252
+ workerFailed: [
253
+ {
254
+ guard: "failureLeavesNoRunningItems",
255
+ target: "finalizing",
256
+ actions: ["rememberFailure", "selectFinalStatus"],
257
+ },
258
+ { actions: "rememberFailure" },
259
+ ],
260
+ abort: [
261
+ {
262
+ guard: "abortLeavesNoRunningItems",
263
+ target: "finalizing",
264
+ actions: ["rememberAbort", "selectFinalStatus"],
265
+ },
266
+ { actions: "rememberAbort" },
267
+ ],
268
+ },
269
+ },
270
+ finalizing: {
271
+ invoke: {
272
+ src: "finalize",
273
+ input: ({ context }) => ({
274
+ effects: context.effects,
275
+ finalization: finalizationOf(context),
276
+ }),
277
+ onDone: [
278
+ { guard: "finalizedAsAborted", target: "aborted" },
279
+ { guard: "finalizedAsFailed", target: "failed" },
280
+ { target: "succeeded" },
281
+ ],
282
+ onError: {
283
+ target: "recordingFinalizationFailure",
284
+ actions: "rememberFinalizationFailure",
285
+ },
286
+ },
287
+ },
288
+ recordingFinalizationFailure: {
289
+ invoke: {
290
+ src: "recordFinalizationFailure",
291
+ input: ({ context }) => ({
292
+ effects: context.effects,
293
+ finalization: finalizationOf(context),
294
+ }),
295
+ onDone: { target: "failed" },
296
+ onError: { target: "failed" },
297
+ },
298
+ },
299
+ succeeded: { type: "final" },
300
+ failed: { type: "final" },
301
+ aborted: { type: "final" },
302
+ },
303
+ output: ({ context }) => finalizationOf(context),
304
+ });
305
+
306
+ export const validateSuiteConcurrency = (concurrency: number | undefined): number => {
307
+ const value = concurrency ?? 1;
308
+ if (!Number.isFinite(value) || !Number.isInteger(value) || value <= 0) {
309
+ throw new Error("suite concurrency must be a positive finite integer");
310
+ }
311
+ return value;
312
+ };
313
+
314
+ export const executeSuiteLifecycle = async <Result>(options: {
315
+ itemCount: number;
316
+ concurrency: number;
317
+ abortSignal?: AbortSignal;
318
+ effects: SuiteLifecycleEffects<Result>;
319
+ }): Promise<SuiteLifecycleFinalization<Result>> => {
320
+ const actor = createActor(suiteLifecycleMachine, {
321
+ input: {
322
+ itemCount: options.itemCount,
323
+ concurrency: options.concurrency,
324
+ ...(options.abortSignal ? { abortSignal: options.abortSignal } : {}),
325
+ effects: options.effects as ErasedEffects,
326
+ },
327
+ });
328
+ actor.start();
329
+ const outcome = await toPromise(actor);
330
+ if (outcome.error !== null) throw outcome.error;
331
+ return outcome as SuiteLifecycleFinalization<Result>;
332
+ };
333
+
334
+ // Internal export for transition-level tests.
335
+ export { suiteLifecycleMachine };