@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 +1,507 @@
1
+ import { assign, createActor, fromPromise, setup, toPromise } from "xstate";
2
+ import {
3
+ type AgentDriver,
4
+ type Answer,
5
+ type Answerer,
6
+ type AnswererContext,
7
+ type AnswerHistoryEntry,
8
+ AnswerSchema,
9
+ type Done,
10
+ type ErrorEvent,
11
+ type Interrupt,
12
+ type Progress,
13
+ type RunEvent,
14
+ } from "../core/index.js";
15
+ import type { OpenedRunDir } from "./run-dir.js";
16
+ import type { RunTraceLimitFailure } from "./trace-limits.js";
17
+
18
+ export type DriverSessionOutcome =
19
+ | { kind: "done"; done: Done }
20
+ | {
21
+ kind: "error";
22
+ error: ErrorEvent;
23
+ operationalFailure?: {
24
+ phase: "answerer" | "driver-transport" | "trace" | "abort";
25
+ reason: string;
26
+ };
27
+ };
28
+
29
+ export type DriverSession = {
30
+ outcome: Promise<DriverSessionOutcome>;
31
+ fail: (reason: string) => void;
32
+ };
33
+
34
+ type DeferredInterrupt = {
35
+ interrupt: Interrupt;
36
+ complete: () => void;
37
+ };
38
+
39
+ type CompletedAnswer = {
40
+ answer: Answer;
41
+ durationMs: number;
42
+ };
43
+
44
+ export type DriverSessionEffects = {
45
+ answer: (interrupt: Interrupt) => Promise<CompletedAnswer>;
46
+ respond: (interrupt: Interrupt, completed: CompletedAnswer) => Promise<void>;
47
+ progress: (progress: Progress) => void;
48
+ terminal: (outcome: DriverSessionOutcome, alreadyRecorded: boolean) => void;
49
+ };
50
+
51
+ type DriverSessionContext = {
52
+ effects: DriverSessionEffects;
53
+ current: DeferredInterrupt | null;
54
+ queued: DeferredInterrupt[];
55
+ terminal: DriverSessionOutcome | null;
56
+ completedAnswer: CompletedAnswer | null;
57
+ };
58
+
59
+ type DriverSessionEvent =
60
+ | { type: "interrupt"; pending: DeferredInterrupt }
61
+ | { type: "progress"; progress: Progress }
62
+ | { type: "done"; done: Done; recorded?: boolean }
63
+ | { type: "error"; error: ErrorEvent; recorded?: boolean }
64
+ | { type: "abort"; recorded?: boolean }
65
+ | { type: "fail"; reason: string; recorded?: boolean }
66
+ | { type: "traceLimit"; failure: RunTraceLimitFailure };
67
+ type RecordableTerminalEvent = Extract<
68
+ DriverSessionEvent,
69
+ { type: "done" | "error" | "abort" | "fail" }
70
+ >;
71
+
72
+ class InteractionFailure extends Error {
73
+ readonly outcome: DriverSessionOutcome;
74
+ readonly terminalRecorded: boolean;
75
+
76
+ constructor(outcome: DriverSessionOutcome, terminalRecorded = false) {
77
+ super(outcome.kind === "error" ? outcome.error.reason : "driver interaction failed");
78
+ this.outcome = outcome;
79
+ this.terminalRecorded = terminalRecorded;
80
+ }
81
+ }
82
+
83
+ const interactionFailure = (error: unknown): DriverSessionOutcome =>
84
+ error instanceof InteractionFailure
85
+ ? error.outcome
86
+ : {
87
+ kind: "error",
88
+ error: { exitReason: "error", reason: "driver interaction failed" },
89
+ };
90
+
91
+ const doneOutcome = (event: DriverSessionEvent): DriverSessionOutcome =>
92
+ event.type === "done"
93
+ ? { kind: "done", done: event.done }
94
+ : {
95
+ kind: "error",
96
+ error: { exitReason: "error", reason: "invalid driver completion event" },
97
+ };
98
+
99
+ const errorOutcome = (event: DriverSessionEvent): DriverSessionOutcome => {
100
+ if (event.type === "error") return { kind: "error", error: event.error };
101
+ const reason = event.type === "fail" ? event.reason : "run aborted";
102
+ return {
103
+ kind: "error",
104
+ error: { exitReason: "error", reason },
105
+ ...(event.type === "abort" ? { operationalFailure: { phase: "abort" as const, reason } } : {}),
106
+ };
107
+ };
108
+
109
+ const traceLimitOutcome = (event: DriverSessionEvent): DriverSessionOutcome => {
110
+ const reason =
111
+ event.type === "traceLimit" ? event.failure.reason : "run trace exceeded its limit";
112
+ return {
113
+ kind: "error",
114
+ error: { exitReason: "error", reason },
115
+ operationalFailure: { phase: "trace", reason },
116
+ };
117
+ };
118
+
119
+ const driverSessionMachine = setup({
120
+ types: {
121
+ context: {} as DriverSessionContext,
122
+ input: {} as { effects: DriverSessionEffects },
123
+ events: {} as DriverSessionEvent,
124
+ output: {} as DriverSessionOutcome,
125
+ },
126
+ actors: {
127
+ answer: fromPromise<CompletedAnswer, { effects: DriverSessionEffects; interrupt: Interrupt }>(
128
+ ({ input }) => input.effects.answer(input.interrupt),
129
+ ),
130
+ respond: fromPromise<
131
+ void,
132
+ {
133
+ effects: DriverSessionEffects;
134
+ interrupt: Interrupt;
135
+ completed: CompletedAnswer;
136
+ }
137
+ >(({ input }) => input.effects.respond(input.interrupt, input.completed)),
138
+ },
139
+ guards: {
140
+ hasNoTerminal: ({ context }) => context.terminal === null,
141
+ hasDoneTerminal: ({ context }) => context.terminal?.kind === "done",
142
+ hasErrorTerminal: ({ context }) => context.terminal?.kind === "error",
143
+ hasTraceTerminal: ({ context }) =>
144
+ context.terminal?.kind === "error" && context.terminal.operationalFailure?.phase === "trace",
145
+ hasQueuedInterrupt: ({ context }) => context.queued.length > 0,
146
+ },
147
+ actions: {
148
+ beginInterrupt: assign({
149
+ current: ({ event }) => (event.type === "interrupt" ? event.pending : null),
150
+ }),
151
+ queueInterrupt: assign({
152
+ queued: ({ context, event }) =>
153
+ event.type === "interrupt" ? [...context.queued, event.pending] : context.queued,
154
+ }),
155
+ takeNextInterrupt: assign(({ context }) => ({
156
+ current: context.queued[0] ?? null,
157
+ queued: context.queued.slice(1),
158
+ })),
159
+ completeCurrentInterrupt: assign(({ context }) => {
160
+ context.current?.complete();
161
+ return { current: null, completedAnswer: null };
162
+ }),
163
+ releaseInterrupts: ({ context }) => {
164
+ context.current?.complete();
165
+ for (const pending of context.queued) pending.complete();
166
+ },
167
+ recordProgress: ({ context, event }) => {
168
+ if (event.type === "progress") context.effects.progress(event.progress);
169
+ },
170
+ recordDone: ({ context, event }) => {
171
+ context.effects.terminal(
172
+ doneOutcome(event),
173
+ event.type === "done" && event.recorded === true,
174
+ );
175
+ },
176
+ rememberDone: assign({
177
+ terminal: ({ event }) => doneOutcome(event),
178
+ }),
179
+ recordError: ({ context, event }) => {
180
+ context.effects.terminal(errorOutcome(event), "recorded" in event && event.recorded === true);
181
+ },
182
+ rememberError: assign({
183
+ terminal: ({ event }) => errorOutcome(event),
184
+ }),
185
+ recordTraceLimit: ({ context, event }) => {
186
+ context.effects.terminal(traceLimitOutcome(event), true);
187
+ },
188
+ rememberTraceLimit: assign({
189
+ terminal: ({ event }) => traceLimitOutcome(event),
190
+ }),
191
+ recordInteractionFailure: ({ context, event }) => {
192
+ if ("error" in event) {
193
+ context.effects.terminal(
194
+ interactionFailure(event.error),
195
+ event.error instanceof InteractionFailure && event.error.terminalRecorded,
196
+ );
197
+ }
198
+ },
199
+ rememberInteractionFailure: assign({
200
+ terminal: ({ event }) =>
201
+ "error" in event ? interactionFailure(event.error) : errorOutcome(event),
202
+ }),
203
+ rememberCompletedAnswer: assign({
204
+ completedAnswer: ({ event }) =>
205
+ "output" in event ? (event.output as CompletedAnswer) : null,
206
+ }),
207
+ },
208
+ }).createMachine({
209
+ id: "lux-driver-session",
210
+ context: ({ input }) => ({
211
+ effects: input.effects,
212
+ current: null,
213
+ queued: [],
214
+ terminal: null,
215
+ completedAnswer: null,
216
+ }),
217
+ initial: "waiting",
218
+ states: {
219
+ waiting: {
220
+ on: {
221
+ interrupt: { target: "answering", actions: "beginInterrupt" },
222
+ progress: { actions: "recordProgress" },
223
+ done: { target: "succeeded", actions: ["recordDone", "rememberDone"] },
224
+ error: { target: "failed", actions: ["recordError", "rememberError"] },
225
+ abort: { target: "failed", actions: ["recordError", "rememberError"] },
226
+ fail: { target: "failed", actions: ["recordError", "rememberError"] },
227
+ traceLimit: {
228
+ target: "failed",
229
+ actions: ["recordTraceLimit", "rememberTraceLimit"],
230
+ },
231
+ },
232
+ },
233
+ answering: {
234
+ invoke: {
235
+ src: "answer",
236
+ input: ({ context }) => {
237
+ if (!context.current) throw new Error("driver session has no current interrupt");
238
+ return { effects: context.effects, interrupt: context.current.interrupt };
239
+ },
240
+ onDone: [
241
+ {
242
+ guard: ({ context }) => context.terminal !== null,
243
+ target: "selecting",
244
+ actions: "completeCurrentInterrupt",
245
+ },
246
+ { target: "responding", actions: "rememberCompletedAnswer" },
247
+ ],
248
+ onError: [
249
+ {
250
+ guard: "hasTraceTerminal",
251
+ target: "selecting",
252
+ actions: "completeCurrentInterrupt",
253
+ },
254
+ {
255
+ target: "failed",
256
+ actions: ["recordInteractionFailure", "rememberInteractionFailure"],
257
+ },
258
+ ],
259
+ },
260
+ on: {
261
+ interrupt: { actions: "queueInterrupt" },
262
+ progress: { actions: "recordProgress" },
263
+ done: {
264
+ guard: "hasNoTerminal",
265
+ actions: ["recordDone", "rememberDone"],
266
+ },
267
+ error: {
268
+ guard: "hasNoTerminal",
269
+ actions: ["recordError", "rememberError"],
270
+ },
271
+ abort: {
272
+ guard: "hasNoTerminal",
273
+ actions: ["recordError", "rememberError"],
274
+ },
275
+ fail: {
276
+ guard: "hasNoTerminal",
277
+ actions: ["recordError", "rememberError"],
278
+ },
279
+ traceLimit: {
280
+ actions: ["recordTraceLimit", "rememberTraceLimit"],
281
+ },
282
+ },
283
+ },
284
+ responding: {
285
+ invoke: {
286
+ src: "respond",
287
+ input: ({ context }) => {
288
+ if (!context.current || !context.completedAnswer) {
289
+ throw new Error("driver session has no completed answer to send");
290
+ }
291
+ return {
292
+ effects: context.effects,
293
+ interrupt: context.current.interrupt,
294
+ completed: context.completedAnswer,
295
+ };
296
+ },
297
+ onDone: { target: "selecting", actions: "completeCurrentInterrupt" },
298
+ onError: [
299
+ {
300
+ guard: "hasTraceTerminal",
301
+ target: "selecting",
302
+ actions: "completeCurrentInterrupt",
303
+ },
304
+ {
305
+ target: "failed",
306
+ actions: ["recordInteractionFailure", "rememberInteractionFailure"],
307
+ },
308
+ ],
309
+ },
310
+ on: {
311
+ interrupt: { actions: "queueInterrupt" },
312
+ progress: { actions: "recordProgress" },
313
+ done: {
314
+ guard: "hasNoTerminal",
315
+ actions: ["recordDone", "rememberDone"],
316
+ },
317
+ error: {
318
+ guard: "hasNoTerminal",
319
+ actions: ["recordError", "rememberError"],
320
+ },
321
+ abort: {
322
+ guard: "hasNoTerminal",
323
+ actions: ["recordError", "rememberError"],
324
+ },
325
+ fail: {
326
+ guard: "hasNoTerminal",
327
+ actions: ["recordError", "rememberError"],
328
+ },
329
+ traceLimit: {
330
+ actions: ["recordTraceLimit", "rememberTraceLimit"],
331
+ },
332
+ },
333
+ },
334
+ selecting: {
335
+ always: [
336
+ { guard: "hasDoneTerminal", target: "succeeded" },
337
+ { guard: "hasErrorTerminal", target: "failed" },
338
+ {
339
+ guard: "hasQueuedInterrupt",
340
+ target: "answering",
341
+ actions: "takeNextInterrupt",
342
+ },
343
+ { target: "waiting" },
344
+ ],
345
+ },
346
+ succeeded: { type: "final", entry: "releaseInterrupts" },
347
+ failed: { type: "final", entry: "releaseInterrupts" },
348
+ },
349
+ output: ({ context }) =>
350
+ context.terminal ?? {
351
+ kind: "error",
352
+ error: { exitReason: "error", reason: "driver session ended without a terminal event" },
353
+ },
354
+ });
355
+
356
+ const messageOf = (error: unknown): string =>
357
+ error instanceof Error ? error.message : String(error);
358
+
359
+ const interactionError = (
360
+ phase: "answerer" | "driver-transport",
361
+ reason: string,
362
+ ): DriverSessionOutcome => ({
363
+ kind: "error",
364
+ error: { exitReason: "error", reason },
365
+ operationalFailure: { phase, reason },
366
+ });
367
+
368
+ export const createDriverSession = (options: {
369
+ driver: AgentDriver;
370
+ answerer: Answerer;
371
+ context: AnswererContext;
372
+ runDir: OpenedRunDir;
373
+ interruptCounts: { ask: number; approve: number; custom: number };
374
+ history: AnswerHistoryEntry[];
375
+ abortSignal?: AbortSignal;
376
+ onToolCall: () => void;
377
+ }): DriverSession => {
378
+ const { driver, answerer, context, runDir, interruptCounts, history, abortSignal, onToolCall } =
379
+ options;
380
+ const recordedInteractionFailure = (
381
+ phase: "answerer" | "driver-transport",
382
+ reason: string,
383
+ ): InteractionFailure => {
384
+ const outcome = interactionError(phase, reason);
385
+ const failure = runDir.events.append({
386
+ type: "error",
387
+ exitReason: "error",
388
+ reason,
389
+ ts: Date.now(),
390
+ });
391
+ return failure
392
+ ? new InteractionFailure(traceLimitOutcome({ type: "traceLimit", failure }), true)
393
+ : new InteractionFailure(outcome, true);
394
+ };
395
+ const rejectedTraceEvent = (failure: RunTraceLimitFailure): InteractionFailure =>
396
+ new InteractionFailure(traceLimitOutcome({ type: "traceLimit", failure }), true);
397
+ const effects: DriverSessionEffects = {
398
+ answer: async (interrupt) => {
399
+ const failure = runDir.events.append({ type: "interrupt", ...interrupt, ts: Date.now() });
400
+ if (failure) throw rejectedTraceEvent(failure);
401
+ interruptCounts[interrupt.kind]++;
402
+ const startedAt = Date.now();
403
+ try {
404
+ const answer = AnswerSchema.parse(await answerer.answer(interrupt, context));
405
+ if (answer.kind !== interrupt.kind || answer.id !== interrupt.id) {
406
+ throw new Error(
407
+ `answer correlation mismatch: expected ${interrupt.kind} interrupt '${interrupt.id}', received ${answer.kind} answer '${answer.id}'`,
408
+ );
409
+ }
410
+ return { answer, durationMs: Date.now() - startedAt };
411
+ } catch (error) {
412
+ const reason = `answerer failed: ${messageOf(error)}`;
413
+ throw recordedInteractionFailure("answerer", reason);
414
+ }
415
+ },
416
+ respond: async (interrupt, completed) => {
417
+ const failure = runDir.events.append({
418
+ type: "responded",
419
+ ...completed.answer,
420
+ ts: Date.now(),
421
+ });
422
+ if (failure) throw rejectedTraceEvent(failure);
423
+ history.push({ interrupt, ...completed });
424
+ try {
425
+ await driver.respond(completed.answer);
426
+ } catch (error) {
427
+ const reason = `driver respond failed: ${messageOf(error)}`;
428
+ throw recordedInteractionFailure("driver-transport", reason);
429
+ }
430
+ },
431
+ progress: (progress) => {
432
+ if (runDir.events.append({ type: "progress", ...progress })) return;
433
+ if (progress.kind === "tool-call") onToolCall();
434
+ },
435
+ terminal: (terminalOutcome, alreadyRecorded) => {
436
+ if (alreadyRecorded) return;
437
+ if (terminalOutcome.kind === "done") {
438
+ runDir.events.append({ type: "done", ...terminalOutcome.done, ts: Date.now() });
439
+ } else {
440
+ runDir.events.append({ type: "error", ...terminalOutcome.error, ts: Date.now() });
441
+ }
442
+ },
443
+ };
444
+
445
+ const actor = createActor(driverSessionMachine, { input: { effects } });
446
+ const onInterrupt = (interrupt: Interrupt): Promise<void> => {
447
+ if (actor.getSnapshot().status === "done") return Promise.resolve();
448
+ return new Promise((complete) => {
449
+ actor.send({ type: "interrupt", pending: { interrupt, complete } });
450
+ });
451
+ };
452
+ const onProgress = (progress: Progress): void => {
453
+ if (actor.getSnapshot().status === "active") actor.send({ type: "progress", progress });
454
+ };
455
+ const sendRecordedTerminal = (event: RunEvent, terminalEvent: RecordableTerminalEvent): void => {
456
+ if (actor.getSnapshot().status !== "active") return;
457
+ if (runDir.events.append(event)) return;
458
+ actor.send({ ...terminalEvent, recorded: true });
459
+ };
460
+ const onDone = (done: Done): void => {
461
+ sendRecordedTerminal({ type: "done", ...done, ts: Date.now() }, { type: "done", done });
462
+ };
463
+ const onError = (error: ErrorEvent): void => {
464
+ sendRecordedTerminal({ type: "error", ...error, ts: Date.now() }, { type: "error", error });
465
+ };
466
+ const onAbort = (): void => {
467
+ sendRecordedTerminal(
468
+ { type: "error", exitReason: "error", reason: "run aborted", ts: Date.now() },
469
+ { type: "abort" },
470
+ );
471
+ };
472
+ const onTraceLimit = (failure: RunTraceLimitFailure): void => {
473
+ if (actor.getSnapshot().status === "active") actor.send({ type: "traceLimit", failure });
474
+ };
475
+ let stopListeningForTraceLimit = (): void => {};
476
+ const unbind = (): void => {
477
+ driver.off("interrupt", onInterrupt);
478
+ driver.off("progress", onProgress);
479
+ driver.off("done", onDone);
480
+ driver.off("error", onError);
481
+ abortSignal?.removeEventListener("abort", onAbort);
482
+ stopListeningForTraceLimit();
483
+ };
484
+
485
+ driver.on("interrupt", onInterrupt);
486
+ driver.on("progress", onProgress);
487
+ driver.on("done", onDone);
488
+ driver.on("error", onError);
489
+ if (!abortSignal?.aborted) abortSignal?.addEventListener("abort", onAbort, { once: true });
490
+
491
+ actor.start();
492
+ stopListeningForTraceLimit = runDir.events.onLimitExceeded(onTraceLimit);
493
+ if (abortSignal?.aborted) onAbort();
494
+ const outcome = toPromise(actor).finally(unbind);
495
+ return {
496
+ outcome,
497
+ fail: (reason) => {
498
+ sendRecordedTerminal(
499
+ { type: "error", exitReason: "error", reason, ts: Date.now() },
500
+ { type: "fail", reason },
501
+ );
502
+ },
503
+ };
504
+ };
505
+
506
+ // Internal export for transition-level tests.
507
+ export { driverSessionMachine };
@@ -0,0 +1,45 @@
1
+ import { readFile } from "node:fs/promises";
2
+ import { dirname, extname, join, relative, resolve, sep } from "node:path";
3
+ import { fileURLToPath } from "node:url";
4
+ import { contentHash, valueHash } from "../core/index.js";
5
+ import { walkAuthoredFiles } from "./fs-read.js";
6
+
7
+ const currentDistributionRoot = (): string => {
8
+ const modulePath = fileURLToPath(import.meta.url);
9
+ const parent = dirname(dirname(modulePath));
10
+ const leaf = parent.split(sep).at(-1);
11
+ if (leaf !== "src" && leaf !== "lib") {
12
+ throw new Error(`cannot locate Lux evaluator distribution from ${modulePath}`);
13
+ }
14
+ return parent;
15
+ };
16
+
17
+ const isRuntimeImplementation = (path: string): boolean => {
18
+ if (path.endsWith(".test.ts") || path.endsWith(".test.js")) return false;
19
+ if (path.endsWith("test-helpers.ts") || path.endsWith("test-helpers.js")) return false;
20
+ if (path.split(sep).includes("fixtures")) return false;
21
+ const extension = extname(path);
22
+ return extension === ".ts" || extension === ".js" || extension === ".json";
23
+ };
24
+
25
+ /**
26
+ * Hash the exact Lux source/build tree executing the evaluation. Package
27
+ * versions alone are insufficient in development, where evaluator semantics
28
+ * can change without a release bump.
29
+ */
30
+ export const evaluatorImplementationHash = async (
31
+ distributionRoot: string = currentDistributionRoot(),
32
+ ): Promise<string> => {
33
+ const root = resolve(distributionRoot);
34
+ const files = (await walkAuthoredFiles(root)).filter(isRuntimeImplementation);
35
+ if (files.length === 0) {
36
+ throw new Error(`Lux evaluator distribution has no implementation files: ${root}`);
37
+ }
38
+ const evidence = await Promise.all(
39
+ files.map(async (path) => ({
40
+ path: relative(root, join(root, path)).split(sep).join("/"),
41
+ hash: contentHash(await readFile(join(root, path))),
42
+ })),
43
+ );
44
+ return valueHash(evidence);
45
+ };
@@ -0,0 +1,61 @@
1
+ import { type ExperimentCampaign, totalTokens } from "../core/index.js";
2
+ import type { CampaignExecutionOptions } from "./experiment-context.js";
3
+ import { experimentActiveElapsedMs, type OpenedExperimentRun } from "./experiment-runs.js";
4
+
5
+ export const throwIfAborted = (signal: AbortSignal | undefined): void => {
6
+ if (!signal?.aborted) return;
7
+ throw signal.reason instanceof Error ? signal.reason : new Error("experiment aborted");
8
+ };
9
+
10
+ export const boundedExecutionSignal = (
11
+ execution: CampaignExecutionOptions,
12
+ experiment: OpenedExperimentRun,
13
+ ): AbortSignal => {
14
+ const elapsed = experimentActiveElapsedMs(experiment.manifest);
15
+ const remaining = execution.loaded.campaign.budget.maxWallTimeMs - elapsed;
16
+ if (remaining <= 0) {
17
+ throw new Error(
18
+ `experiment exceeded its wall-time budget ${execution.loaded.campaign.budget.maxWallTimeMs}ms`,
19
+ );
20
+ }
21
+ const deadline = AbortSignal.timeout(Math.max(1, remaining));
22
+ return execution.abortSignal ? AbortSignal.any([execution.abortSignal, deadline]) : deadline;
23
+ };
24
+
25
+ export const ensureBudget = (
26
+ campaign: ExperimentCampaign,
27
+ experiment: OpenedExperimentRun,
28
+ additional: { metricCalls?: number; judgeCalls?: number; proposalCalls?: number } = {},
29
+ ): void => {
30
+ const usage = experiment.manifest.budgetUsage;
31
+ const elapsed = experimentActiveElapsedMs(experiment.manifest);
32
+ const checks: [boolean, string][] = [
33
+ [
34
+ usage.metricCalls + (additional.metricCalls ?? 0) > campaign.budget.maxMetricCalls,
35
+ `metric-call budget ${campaign.budget.maxMetricCalls}`,
36
+ ],
37
+ [
38
+ usage.judgeCalls + (additional.judgeCalls ?? 0) > campaign.budget.maxJudgeCalls,
39
+ `judge-call budget ${campaign.budget.maxJudgeCalls}`,
40
+ ],
41
+ [
42
+ usage.proposalCalls + (additional.proposalCalls ?? 0) > campaign.budget.maxProposalCalls,
43
+ `proposal-call budget ${campaign.budget.maxProposalCalls}`,
44
+ ],
45
+ [
46
+ totalTokens(usage.usage) > campaign.budget.maxTokens,
47
+ `token budget ${campaign.budget.maxTokens}`,
48
+ ],
49
+ [usage.costUsd > campaign.budget.maxCostUsd, `cost budget $${campaign.budget.maxCostUsd}`],
50
+ [
51
+ (usage.unpricedModelCalls ?? 0) > campaign.budget.maxUnpricedModelCalls,
52
+ `unpriced-model-call budget ${campaign.budget.maxUnpricedModelCalls}`,
53
+ ],
54
+ [
55
+ elapsed > campaign.budget.maxWallTimeMs,
56
+ `wall-time budget ${campaign.budget.maxWallTimeMs}ms`,
57
+ ],
58
+ ];
59
+ const exceeded = checks.find(([failed]) => failed);
60
+ if (exceeded) throw new Error(`experiment exceeded its ${exceeded[1]}`);
61
+ };
@@ -0,0 +1,41 @@
1
+ import type { Registry } from "../core/index.js";
2
+ import type { ResolvedLuxConfig } from "./config.js";
3
+ import type { discoverCases } from "./discover.js";
4
+ import type { LoadedExperimentCampaign } from "./experiment-loader.js";
5
+
6
+ export type EvaluationCorpus =
7
+ | {
8
+ kind: "agent";
9
+ cases: Awaited<ReturnType<typeof discoverCases>>["cases"];
10
+ hash: string;
11
+ }
12
+ | {
13
+ kind: "grader";
14
+ hash: string;
15
+ };
16
+
17
+ export type CampaignExecutionOptions = {
18
+ config: ResolvedLuxConfig;
19
+ registry: Registry;
20
+ loaded: LoadedExperimentCampaign;
21
+ runsRoot?: string;
22
+ resumeFrom?: string;
23
+ abortSignal?: AbortSignal;
24
+ onEvaluationUnit?: (unit: {
25
+ unitId: string;
26
+ candidateId: string;
27
+ profileId: string;
28
+ split: "train" | "validation" | "test";
29
+ }) => void | Promise<void>;
30
+ };
31
+
32
+ export const profileOrThrow = (config: ResolvedLuxConfig, profileId: string) => {
33
+ const profile = config.profiles?.[profileId];
34
+ if (!profile) {
35
+ const available = Object.keys(config.profiles ?? {}).sort();
36
+ throw new Error(
37
+ `experiment profile '${profileId}' is not configured${available.length > 0 ? `; available: ${available.join(", ")}` : ""}`,
38
+ );
39
+ }
40
+ return profile;
41
+ };
@@ -0,0 +1,43 @@
1
+ import { z } from "zod";
2
+ import { CandidateSchema, ExperimentBudgetSchema } from "../core/index.js";
3
+ import { ExperimentRunSchema } from "./experiment-runs.js";
4
+
5
+ export const CampaignExecutionResultSchema = z
6
+ .object({
7
+ experimentDir: z.string(),
8
+ champion: CandidateSchema.nullable(),
9
+ manifest: ExperimentRunSchema,
10
+ })
11
+ .strict();
12
+ export type CampaignExecutionResult = z.infer<typeof CampaignExecutionResultSchema>;
13
+
14
+ export const ExperimentCampaignPlanSchema = z
15
+ .object({
16
+ campaignId: z.string(),
17
+ mode: z.enum(["experiment", "optimize"]),
18
+ evaluationKind: z.enum(["agent", "grader"]),
19
+ subjectRoot: z.string(),
20
+ mutableComponents: z.array(z.string()),
21
+ profiles: z.array(z.string()),
22
+ splits: z.array(
23
+ z
24
+ .object({
25
+ split: z.enum(["train", "validation", "test"]),
26
+ examples: z.number().int().nonnegative(),
27
+ })
28
+ .strict(),
29
+ ),
30
+ projected: z
31
+ .object({
32
+ candidates: z.number().int().nonnegative(),
33
+ metricCalls: z.number().int().nonnegative(),
34
+ judgeCalls: z.number().int().nonnegative(),
35
+ proposalCalls: z.number().int().nonnegative(),
36
+ })
37
+ .strict(),
38
+ budget: ExperimentBudgetSchema,
39
+ fitsProjectedCallBudgets: z.boolean(),
40
+ warnings: z.array(z.string()),
41
+ })
42
+ .strict();
43
+ export type ExperimentCampaignPlan = z.infer<typeof ExperimentCampaignPlanSchema>;