@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 +1,336 @@
1
+ import { glob } from "tinyglobby";
2
+ import { z } from "zod";
3
+ import { type AssertionImpl, readInspectedFileBytes } from "../../core/index.js";
4
+ import {
5
+ baseAssertionFields,
6
+ globEscapesAssertionRoot,
7
+ inRange,
8
+ inspectAssertionPath,
9
+ LocationSchema,
10
+ resolveLocationRoot,
11
+ } from "./helpers.js";
12
+ import {
13
+ countSafeRegex,
14
+ inspectSafeRegexes,
15
+ testSafeRegex,
16
+ testSafeRegexSet,
17
+ } from "./safe-regex.js";
18
+
19
+ const MAX_REGEX_FILE_BYTES = 5 * 1024 * 1024;
20
+ const MAX_TEXT_FILE_BYTES = 10 * 1024 * 1024;
21
+ const MAX_REGEX_PATTERNS = 50;
22
+ const MAX_REGEX_PATTERN_CHARS = 2_000;
23
+
24
+ export const FileExistsConfigSchema = z
25
+ .object({
26
+ type: z.literal("file-exists"),
27
+ path: z.string(),
28
+ in: LocationSchema,
29
+ ...baseAssertionFields,
30
+ })
31
+ .strict();
32
+ export type FileExistsConfig = z.infer<typeof FileExistsConfigSchema>;
33
+
34
+ export const fileExistsAssertion: AssertionImpl<FileExistsConfig> = {
35
+ type: "file-exists",
36
+ configSchema: FileExistsConfigSchema,
37
+ evaluate: async (config, run) => {
38
+ const { stats, issue } = await inspectAssertionPath(run, config.in, config.path);
39
+ const passed = stats?.isFile() ?? false;
40
+ return {
41
+ passed,
42
+ score: passed ? 1 : 0,
43
+ comment: passed
44
+ ? `${config.path} exists`
45
+ : (issue ?? (stats ? `${config.path} is not a file` : `${config.path} not found`)),
46
+ };
47
+ },
48
+ };
49
+
50
+ export const FileContainsConfigSchema = z
51
+ .object({
52
+ type: z.literal("file-contains"),
53
+ path: z.string(),
54
+ text: z.string(),
55
+ in: LocationSchema,
56
+ ...baseAssertionFields,
57
+ })
58
+ .strict();
59
+ export type FileContainsConfig = z.infer<typeof FileContainsConfigSchema>;
60
+
61
+ export const fileContainsAssertion: AssertionImpl<FileContainsConfig> = {
62
+ type: "file-contains",
63
+ configSchema: FileContainsConfigSchema,
64
+ evaluate: async (config, run) => {
65
+ const { target, stats, issue } = await inspectAssertionPath(run, config.in, config.path);
66
+ if (!stats?.isFile()) {
67
+ return {
68
+ passed: false,
69
+ score: 0,
70
+ comment: issue ?? (stats ? `${config.path} is not a file` : `${config.path} not found`),
71
+ };
72
+ }
73
+ if (stats.size > MAX_TEXT_FILE_BYTES) {
74
+ return {
75
+ passed: false,
76
+ score: 0,
77
+ comment: `${config.path} exceeds the ${MAX_TEXT_FILE_BYTES}-byte text assertion limit`,
78
+ };
79
+ }
80
+ const content = (
81
+ await readInspectedFileBytes(target, stats, MAX_TEXT_FILE_BYTES, config.path)
82
+ ).toString("utf8");
83
+ const passed = content.includes(config.text);
84
+ return {
85
+ passed,
86
+ score: passed ? 1 : 0,
87
+ comment: passed
88
+ ? `${config.path} contains '${config.text}'`
89
+ : `${config.path} does not contain '${config.text}'`,
90
+ };
91
+ },
92
+ };
93
+
94
+ export const FileMatchesConfigSchema = z
95
+ .object({
96
+ type: z.literal("file-matches"),
97
+ path: z.string(),
98
+ pattern: z.string().max(MAX_REGEX_PATTERN_CHARS),
99
+ flags: z.string().max(16).optional(),
100
+ in: LocationSchema,
101
+ ...baseAssertionFields,
102
+ })
103
+ .strict()
104
+ .superRefine(async (config, ctx) => {
105
+ const [inspection] = await inspectSafeRegexes([config.pattern], config.flags ?? "");
106
+ if (inspection && "error" in inspection) {
107
+ ctx.addIssue({ code: "custom", path: ["pattern"], message: inspection.error });
108
+ }
109
+ });
110
+ export type FileMatchesConfig = z.infer<typeof FileMatchesConfigSchema>;
111
+
112
+ export const fileMatchesAssertion: AssertionImpl<FileMatchesConfig> = {
113
+ type: "file-matches",
114
+ configSchema: FileMatchesConfigSchema,
115
+ evaluate: async (config, run, ctx) => {
116
+ const inspected = await readAssertionFile(config, run);
117
+ if ("failure" in inspected) return inspected.failure;
118
+ const passed = await testSafeRegex(config.pattern, config.flags ?? "", inspected.content, {
119
+ abortSignal: ctx?.abortSignal,
120
+ });
121
+ return {
122
+ passed,
123
+ score: passed ? 1 : 0,
124
+ comment: passed
125
+ ? `${config.path} matches /${config.pattern}/${config.flags ?? ""}`
126
+ : `${config.path} does not match /${config.pattern}/${config.flags ?? ""}`,
127
+ };
128
+ },
129
+ };
130
+
131
+ const readAssertionFile = async (
132
+ config: { in: z.infer<typeof LocationSchema>; path: string },
133
+ run: Parameters<AssertionImpl["evaluate"]>[1],
134
+ ): Promise<{ content: string } | { failure: { passed: false; score: 0; comment: string } }> => {
135
+ const { target, stats, issue } = await inspectAssertionPath(run, config.in, config.path);
136
+ if (!stats?.isFile()) {
137
+ return {
138
+ failure: {
139
+ passed: false,
140
+ score: 0,
141
+ comment: issue ?? (stats ? `${config.path} is not a file` : `${config.path} not found`),
142
+ },
143
+ };
144
+ }
145
+ if (stats.size > MAX_REGEX_FILE_BYTES) {
146
+ return {
147
+ failure: {
148
+ passed: false,
149
+ score: 0,
150
+ comment: `${config.path} exceeds the ${MAX_REGEX_FILE_BYTES}-byte regex assertion limit`,
151
+ },
152
+ };
153
+ }
154
+ return {
155
+ content: (
156
+ await readInspectedFileBytes(target, stats, MAX_REGEX_FILE_BYTES, config.path)
157
+ ).toString("utf8"),
158
+ };
159
+ };
160
+
161
+ const regexSetPassed = (
162
+ mode: FileRegexSetConfig["mode"],
163
+ matched: number,
164
+ total: number,
165
+ ): boolean => {
166
+ if (mode === "all") return matched === total;
167
+ if (mode === "any") return matched > 0;
168
+ return matched === 0;
169
+ };
170
+
171
+ export const FileRegexSetConfigSchema = z
172
+ .object({
173
+ type: z.literal("file-regex-set"),
174
+ path: z.string(),
175
+ in: LocationSchema,
176
+ patterns: z.array(z.string().max(MAX_REGEX_PATTERN_CHARS)).min(1).max(MAX_REGEX_PATTERNS),
177
+ mode: z.enum(["all", "any", "none"]),
178
+ flags: z.string().max(16).optional(),
179
+ ...baseAssertionFields,
180
+ })
181
+ .strict()
182
+ .superRefine(async (config, ctx) => {
183
+ const inspections = await inspectSafeRegexes(config.patterns, config.flags ?? "");
184
+ for (const [index, inspection] of inspections.entries()) {
185
+ if ("error" in inspection) {
186
+ ctx.addIssue({
187
+ code: "custom",
188
+ path: ["patterns", index],
189
+ message: inspection.error,
190
+ });
191
+ }
192
+ }
193
+ });
194
+ export type FileRegexSetConfig = z.infer<typeof FileRegexSetConfigSchema>;
195
+
196
+ export const fileRegexSetAssertion: AssertionImpl<FileRegexSetConfig> = {
197
+ type: "file-regex-set",
198
+ configSchema: FileRegexSetConfigSchema,
199
+ evaluate: async (config, run, ctx) => {
200
+ const inspected = await readAssertionFile(config, run);
201
+ if ("failure" in inspected) return inspected.failure;
202
+ const matches = await testSafeRegexSet(config.patterns, config.flags ?? "", inspected.content, {
203
+ abortSignal: ctx?.abortSignal,
204
+ });
205
+ const matched = matches.filter(Boolean).length;
206
+ const passed = regexSetPassed(config.mode, matched, matches.length);
207
+ return {
208
+ passed,
209
+ score: passed ? 1 : 0,
210
+ comment: `${matched}/${matches.length} regex patterns matched ${config.path}; expected ${config.mode}`,
211
+ detail: {
212
+ matched,
213
+ total: matches.length,
214
+ unmatchedPatterns: config.patterns.filter((_pattern, index) => !matches[index]),
215
+ },
216
+ };
217
+ },
218
+ };
219
+
220
+ export const FileRegexCountConfigSchema = z
221
+ .object({
222
+ type: z.literal("file-regex-count"),
223
+ path: z.string(),
224
+ in: LocationSchema,
225
+ pattern: z.string().max(MAX_REGEX_PATTERN_CHARS),
226
+ flags: z.string().max(16).optional(),
227
+ min: z.number().int().nonnegative().optional(),
228
+ max: z.number().int().nonnegative().optional(),
229
+ ...baseAssertionFields,
230
+ })
231
+ .strict()
232
+ .refine((config) => config.min !== undefined || config.max !== undefined, {
233
+ message: "set at least one of min or max",
234
+ })
235
+ .refine((config) => config.max !== undefined || (config.min ?? 0) > 0, {
236
+ message: "min alone must be greater than zero",
237
+ })
238
+ .refine(
239
+ (config) => config.min === undefined || config.max === undefined || config.min <= config.max,
240
+ { message: "min must be <= max" },
241
+ )
242
+ .superRefine(async (config, ctx) => {
243
+ const [inspection] = await inspectSafeRegexes([config.pattern], config.flags ?? "");
244
+ if (!inspection) return;
245
+ if ("error" in inspection) {
246
+ ctx.addIssue({ code: "custom", path: ["pattern"], message: inspection.error });
247
+ return;
248
+ }
249
+ if (inspection.matchesEmpty) {
250
+ ctx.addIssue({
251
+ code: "custom",
252
+ path: ["pattern"],
253
+ message: "count pattern must not match an empty string",
254
+ });
255
+ }
256
+ });
257
+ export type FileRegexCountConfig = z.infer<typeof FileRegexCountConfigSchema>;
258
+
259
+ export const fileRegexCountAssertion: AssertionImpl<FileRegexCountConfig> = {
260
+ type: "file-regex-count",
261
+ configSchema: FileRegexCountConfigSchema,
262
+ evaluate: async (config, run, ctx) => {
263
+ const inspected = await readAssertionFile(config, run);
264
+ if ("failure" in inspected) return inspected.failure;
265
+ const stopAfter = config.max !== undefined ? config.max + 1 : (config.min ?? 1);
266
+ const { count, exact } = await countSafeRegex(
267
+ config.pattern,
268
+ config.flags ?? "",
269
+ inspected.content,
270
+ stopAfter,
271
+ { abortSignal: ctx?.abortSignal },
272
+ );
273
+ const { passed, comment } = inRange(count, config.min, config.max);
274
+ return {
275
+ passed,
276
+ score: passed ? 1 : 0,
277
+ comment: `${config.path} has ${exact ? "" : "at least "}${count} matches for /${config.pattern}/${config.flags ?? ""} (${comment})`,
278
+ detail: { count, exact },
279
+ };
280
+ },
281
+ };
282
+
283
+ export const GlobCountConfigSchema = z
284
+ .object({
285
+ type: z.literal("glob-count"),
286
+ glob: z.string(),
287
+ min: z.number().int().nonnegative().optional(),
288
+ max: z.number().int().nonnegative().optional(),
289
+ in: LocationSchema,
290
+ ...baseAssertionFields,
291
+ })
292
+ .strict()
293
+ // With neither bound the assertion is vacuous — it would always pass.
294
+ .refine((c) => c.min !== undefined || c.max !== undefined, {
295
+ message: "set at least one of min or max",
296
+ })
297
+ .refine((c) => c.min === undefined || c.max === undefined || c.min <= c.max, {
298
+ message: "min must be <= max",
299
+ });
300
+ export type GlobCountConfig = z.infer<typeof GlobCountConfigSchema>;
301
+
302
+ export const globCountAssertion: AssertionImpl<GlobCountConfig> = {
303
+ type: "glob-count",
304
+ configSchema: GlobCountConfigSchema,
305
+ evaluate: async (config, run) => {
306
+ const root = resolveLocationRoot(run, config.in);
307
+ if (globEscapesAssertionRoot(config.glob)) {
308
+ return {
309
+ passed: false,
310
+ score: 0,
311
+ comment: `${config.glob} escapes the selected assertion root`,
312
+ };
313
+ }
314
+ const matches = await glob(config.glob, {
315
+ cwd: root,
316
+ onlyFiles: true,
317
+ followSymbolicLinks: false,
318
+ });
319
+ const inspected = await Promise.all(
320
+ matches.map(async (path) => ({
321
+ path,
322
+ inspection: await inspectAssertionPath(run, config.in, path),
323
+ })),
324
+ );
325
+ const files = inspected
326
+ .filter(({ inspection }) => inspection.stats?.isFile())
327
+ .map(({ path }) => path);
328
+ const { passed, comment } = inRange(files.length, config.min, config.max);
329
+ return {
330
+ passed,
331
+ score: passed ? 1 : 0,
332
+ comment: `${files.length} matches for ${config.glob} (${comment})`,
333
+ detail: { count: files.length, files: files.slice(0, 10) },
334
+ };
335
+ },
336
+ };
@@ -0,0 +1,203 @@
1
+ import type { Stats } from "node:fs";
2
+ import { lstat } from "node:fs/promises";
3
+ import { isAbsolute, join, relative, resolve, sep } from "node:path";
4
+ import { z } from "zod";
5
+ import { inspectRunArtifact, runArtifactsRoot } from "../../core/artifact-evidence.js";
6
+ import type { Run, RunEvent } from "../../core/index.js";
7
+
8
+ const FsLocationSchema = z.enum(["artifacts", "cwd"]);
9
+ export type FsLocation = z.infer<typeof FsLocationSchema>;
10
+
11
+ export const resolveLocationRoot = (run: Run, where: FsLocation = "artifacts"): string =>
12
+ where === "cwd" ? run.runDir : runArtifactsRoot(run);
13
+
14
+ export type AssertionPathInspection = {
15
+ target: string;
16
+ stats?: Stats;
17
+ issue?: string;
18
+ };
19
+
20
+ const pathEscapesRoot = (root: string, target: string): boolean => {
21
+ const rel = relative(root, target);
22
+ return rel === ".." || rel.startsWith(`..${sep}`) || isAbsolute(rel);
23
+ };
24
+
25
+ export const inspectAssertionPath = async (
26
+ run: Run,
27
+ where: FsLocation,
28
+ authoredPath: string,
29
+ ): Promise<AssertionPathInspection> => {
30
+ if (where === "artifacts") {
31
+ const { target, stats, issue } = await inspectRunArtifact(run, authoredPath);
32
+ return { target, ...(stats ? { stats } : {}), ...(issue ? { issue } : {}) };
33
+ }
34
+ const root = resolve(resolveLocationRoot(run, where));
35
+ const target = resolve(root, authoredPath);
36
+ if (pathEscapesRoot(root, target)) {
37
+ return { target, issue: `${authoredPath} escapes the selected assertion root` };
38
+ }
39
+
40
+ const rel = relative(root, target);
41
+ const components = [
42
+ root,
43
+ ...rel
44
+ .split(sep)
45
+ .filter(Boolean)
46
+ .map((_, index, parts) => join(root, ...parts.slice(0, index + 1))),
47
+ ];
48
+ for (const [index, component] of components.entries()) {
49
+ let stats: Stats;
50
+ try {
51
+ stats = await lstat(component);
52
+ } catch (err) {
53
+ const code = (err as NodeJS.ErrnoException).code;
54
+ if (code === "ENOENT" || code === "ENOTDIR") {
55
+ return { target, issue: `${authoredPath} not found` };
56
+ }
57
+ throw err;
58
+ }
59
+ if (stats.isSymbolicLink()) {
60
+ return { target, issue: `${authoredPath} resolves through a symlink` };
61
+ }
62
+ if (index < components.length - 1 && !stats.isDirectory()) {
63
+ return { target, issue: `${authoredPath} not found` };
64
+ }
65
+ if (index === components.length - 1) return { target, stats };
66
+ }
67
+ return { target, issue: `${authoredPath} not found` };
68
+ };
69
+
70
+ export const globEscapesAssertionRoot = (pattern: string): boolean =>
71
+ isAbsolute(pattern) || pattern.split(/[\\/]/).includes("..");
72
+
73
+ const COMMAND_TOOLCHAIN_ENV = [
74
+ "PATH",
75
+ "PATHEXT",
76
+ "HOME",
77
+ "USERPROFILE",
78
+ "SYSTEMROOT",
79
+ "COMSPEC",
80
+ "SHELL",
81
+ "TMPDIR",
82
+ "TMP",
83
+ "TEMP",
84
+ "LANG",
85
+ "LANGUAGE",
86
+ "LC_ALL",
87
+ "LC_CTYPE",
88
+ "TERM",
89
+ "COLORTERM",
90
+ "NO_COLOR",
91
+ "FORCE_COLOR",
92
+ ] as const;
93
+
94
+ export const commandEnvironment = (
95
+ parent: NodeJS.ProcessEnv,
96
+ inherit: true | readonly string[] = [],
97
+ ): NodeJS.ProcessEnv => {
98
+ if (inherit === true) return { ...parent };
99
+ const names = new Set<string>([...COMMAND_TOOLCHAIN_ENV, ...inherit]);
100
+ return Object.fromEntries(
101
+ [...names].flatMap((name) => {
102
+ const value = parent[name];
103
+ return value === undefined ? [] : [[name, value] as const];
104
+ }),
105
+ );
106
+ };
107
+
108
+ /**
109
+ * Resolve an RFC 6901 JSON pointer against a parsed value; undefined when it
110
+ * does not resolve. Deviates from the RFC for arrays: any own property
111
+ * resolves, so `/length` works even though the RFC only allows index tokens.
112
+ */
113
+ export const resolvePointer = (root: unknown, pointer: string): unknown => {
114
+ if (pointer === "") return root;
115
+ if (!pointer.startsWith("/")) return undefined;
116
+ const parts = pointer
117
+ .slice(1)
118
+ .split("/")
119
+ .map((p) => p.replace(/~1/g, "/").replace(/~0/g, "~"));
120
+ let node: unknown = root;
121
+ for (const part of parts) {
122
+ if (node === null || typeof node !== "object") return undefined;
123
+ // Own properties only — a pointer like `/constructor` must not walk the
124
+ // prototype chain into inherited values.
125
+ if (!Object.hasOwn(node, part)) return undefined;
126
+ node = (node as Record<string, unknown>)[part];
127
+ }
128
+ return node;
129
+ };
130
+
131
+ export const inRange = (
132
+ value: number,
133
+ min?: number,
134
+ max?: number,
135
+ ): { passed: boolean; comment: string } => {
136
+ if (min !== undefined && value < min) {
137
+ return { passed: false, comment: `${value} < min ${min}` };
138
+ }
139
+ if (max !== undefined && value > max) {
140
+ return { passed: false, comment: `${value} > max ${max}` };
141
+ }
142
+ const bounds = [
143
+ min !== undefined ? `>= ${min}` : undefined,
144
+ max !== undefined ? `<= ${max}` : undefined,
145
+ ]
146
+ .filter((b) => b !== undefined)
147
+ .join(", ");
148
+ return {
149
+ passed: true,
150
+ comment: bounds ? `${value} ${bounds}` : `${value}`,
151
+ };
152
+ };
153
+
154
+ /** Structural equality for JSON-shaped values — object key order is irrelevant. */
155
+ export const deepEquals = (a: unknown, b: unknown): boolean => {
156
+ if (Object.is(a, b)) return true;
157
+ if (typeof a !== "object" || typeof b !== "object" || a === null || b === null) return false;
158
+ if (Array.isArray(a) || Array.isArray(b)) {
159
+ if (!Array.isArray(a) || !Array.isArray(b) || a.length !== b.length) return false;
160
+ return a.every((v, i) => deepEquals(v, b[i]));
161
+ }
162
+ const aEntries = Object.entries(a);
163
+ if (aEntries.length !== Object.keys(b).length) return false;
164
+ return aEntries.every(
165
+ ([k, v]) => Object.hasOwn(b, k) && deepEquals(v, (b as Record<string, unknown>)[k]),
166
+ );
167
+ };
168
+
169
+ /**
170
+ * Zod fragment for the framework-level fields every assertion shares
171
+ * (weight, threshold, name). Spread into per-assertion config schemas:
172
+ *
173
+ * z.object({ type: z.literal("foo"), ...baseAssertionFields, ... })
174
+ *
175
+ * Mirrors {@link Assertion} from the core module while letting each built-in
176
+ * assertion reject misspelled type-specific fields at its own schema boundary.
177
+ */
178
+ export const baseAssertionFields = {
179
+ id: z.string().min(1).optional(),
180
+ weight: z.number().positive().optional(),
181
+ threshold: z.number().min(0).max(1).optional(),
182
+ name: z.string().optional(),
183
+ } as const;
184
+
185
+ /**
186
+ * Where on the run dir an assertion looks for files. Defaults to
187
+ * `artifacts` (the standard place drivers drop captured files).
188
+ */
189
+ export const LocationSchema = FsLocationSchema.default("artifacts");
190
+
191
+ // `name` is intentionally excluded: for tool assertions the `name` field is
192
+ // the tool's name, which then doubles as the display label.
193
+ export const weightThresholdFields = {
194
+ id: baseAssertionFields.id,
195
+ weight: baseAssertionFields.weight,
196
+ threshold: baseAssertionFields.threshold,
197
+ } as const;
198
+
199
+ export type ProgressEvent = Extract<RunEvent, { type: "progress" }>;
200
+ export type ToolCallEvent = ProgressEvent & { kind: "tool-call" };
201
+
202
+ export const isToolCallEvent = (e: RunEvent): e is ToolCallEvent =>
203
+ e.type === "progress" && e.kind === "tool-call";
@@ -0,0 +1,80 @@
1
+ import type { AssertionImpl, Registry } from "../../core/index.js";
2
+ import { commandExitsZeroAssertion } from "./command-exits-zero.js";
3
+ import {
4
+ interruptCountAssertion,
5
+ toolCallCountAssertion,
6
+ toolCalledAssertion,
7
+ } from "./event-checks.js";
8
+ import {
9
+ fileContainsAssertion,
10
+ fileExistsAssertion,
11
+ fileMatchesAssertion,
12
+ fileRegexCountAssertion,
13
+ fileRegexSetAssertion,
14
+ globCountAssertion,
15
+ } from "./file-checks.js";
16
+ import { jsonPointerEqualsAssertion } from "./json-pointer-equals.js";
17
+ import { containsAssertion, notContainsAssertion, regexAssertion } from "./output-checks.js";
18
+ import { predicateAssertion } from "./predicate.js";
19
+ import {
20
+ artifactBytesUnderAssertion,
21
+ costUnderAssertion,
22
+ latencyUnderAssertion,
23
+ runSucceededAssertion,
24
+ tokensUnderAssertion,
25
+ } from "./resource-checks.js";
26
+ import { toolCalledWithAssertion } from "./tool-called-with.js";
27
+ import { toolInvocationAssertion } from "./tool-invocation.js";
28
+ import {
29
+ noErrorResultsAssertion,
30
+ toolResultContainsAssertion,
31
+ toolResultIsErrorAssertion,
32
+ toolResultJsonPointerEqualsAssertion,
33
+ } from "./tool-result-checks.js";
34
+
35
+ export * from "./command-exits-zero.js";
36
+ export * from "./event-checks.js";
37
+ export * from "./file-checks.js";
38
+ export * from "./json-pointer-equals.js";
39
+ export * from "./output-checks.js";
40
+ export * from "./predicate.js";
41
+ export * from "./resource-checks.js";
42
+ export * from "./tool-called-with.js";
43
+ export * from "./tool-calls.js";
44
+ export * from "./tool-invocation.js";
45
+ export * from "./tool-result-checks.js";
46
+
47
+ export const allCoreAssertions: AssertionImpl[] = [
48
+ fileExistsAssertion,
49
+ fileContainsAssertion,
50
+ fileMatchesAssertion,
51
+ fileRegexSetAssertion,
52
+ fileRegexCountAssertion,
53
+ globCountAssertion,
54
+ toolCalledAssertion,
55
+ toolCalledWithAssertion,
56
+ toolInvocationAssertion,
57
+ toolCallCountAssertion,
58
+ toolResultIsErrorAssertion,
59
+ toolResultContainsAssertion,
60
+ toolResultJsonPointerEqualsAssertion,
61
+ noErrorResultsAssertion,
62
+ interruptCountAssertion,
63
+ commandExitsZeroAssertion,
64
+ containsAssertion,
65
+ notContainsAssertion,
66
+ regexAssertion,
67
+ jsonPointerEqualsAssertion,
68
+ latencyUnderAssertion,
69
+ artifactBytesUnderAssertion,
70
+ tokensUnderAssertion,
71
+ costUnderAssertion,
72
+ runSucceededAssertion,
73
+ predicateAssertion,
74
+ ];
75
+
76
+ export const registerCoreAssertions = (registry: Registry): void => {
77
+ for (const a of allCoreAssertions) {
78
+ if (!registry.getAssertion(a.type)) registry.registerAssertion(a);
79
+ }
80
+ };
@@ -0,0 +1,84 @@
1
+ import { z } from "zod";
2
+ import { type AssertionImpl, readInspectedFileBytes } from "../../core/index.js";
3
+ import {
4
+ baseAssertionFields,
5
+ deepEquals,
6
+ inspectAssertionPath,
7
+ LocationSchema,
8
+ resolvePointer,
9
+ } from "./helpers.js";
10
+
11
+ const MAX_JSON_FILE_BYTES = 10 * 1024 * 1024;
12
+
13
+ export const JsonPointerEqualsConfigSchema = z
14
+ .object({
15
+ type: z.literal("json-pointer-equals"),
16
+ /** File path to read (parsed as JSON). */
17
+ path: z.string(),
18
+ /** Where to look for the file. */
19
+ in: LocationSchema,
20
+ /** RFC 6901 JSON pointer. */
21
+ pointer: z.string(),
22
+ /** Expected value at the pointer. */
23
+ value: z.unknown(),
24
+ ...baseAssertionFields,
25
+ })
26
+ .strict()
27
+ // An omitted `value` would silently pass against any missing pointer
28
+ // (undefined === undefined), so its presence is required outright.
29
+ .refine((c) => "value" in c, { message: "value is required" });
30
+ export type JsonPointerEqualsConfig = z.infer<typeof JsonPointerEqualsConfigSchema>;
31
+
32
+ export const jsonPointerEqualsAssertion: AssertionImpl<JsonPointerEqualsConfig> = {
33
+ type: "json-pointer-equals",
34
+ configSchema: JsonPointerEqualsConfigSchema,
35
+ evaluate: async (config, run) => {
36
+ const { target, stats, issue } = await inspectAssertionPath(run, config.in, config.path);
37
+ if (!stats?.isFile()) {
38
+ return {
39
+ passed: false,
40
+ score: 0,
41
+ comment: issue ?? (stats ? `${config.path} is not a file` : `${config.path} not found`),
42
+ };
43
+ }
44
+ if (stats.size > MAX_JSON_FILE_BYTES) {
45
+ return {
46
+ passed: false,
47
+ score: 0,
48
+ comment: `${config.path} exceeds the ${MAX_JSON_FILE_BYTES}-byte JSON assertion limit`,
49
+ };
50
+ }
51
+ let parsed: unknown;
52
+ try {
53
+ const contents = await readInspectedFileBytes(
54
+ target,
55
+ stats,
56
+ MAX_JSON_FILE_BYTES,
57
+ config.path,
58
+ );
59
+ parsed = JSON.parse(contents.toString("utf8"));
60
+ } catch (err) {
61
+ return {
62
+ passed: false,
63
+ score: 0,
64
+ comment: `${config.path} is not valid JSON: ${err instanceof Error ? err.message : String(err)}`,
65
+ };
66
+ }
67
+ const resolved = resolvePointer(parsed, config.pointer);
68
+ if (resolved === undefined) {
69
+ return {
70
+ passed: false,
71
+ score: 0,
72
+ comment: `${config.pointer} does not resolve in ${config.path}`,
73
+ };
74
+ }
75
+ const passed = deepEquals(resolved, config.value);
76
+ return {
77
+ passed,
78
+ score: passed ? 1 : 0,
79
+ comment: passed
80
+ ? `${config.pointer} matches expected value`
81
+ : `${config.pointer}: expected ${JSON.stringify(config.value)}, got ${JSON.stringify(resolved)}`,
82
+ };
83
+ },
84
+ };