@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 +1,198 @@
1
+ import { z } from "zod";
2
+ import {
3
+ type AssertionImpl,
4
+ malformedToolEventCount,
5
+ toolInvocations,
6
+ toolResultValue,
7
+ } from "../../core/index.js";
8
+ import { deepEquals, resolvePointer, weightThresholdFields } from "./helpers.js";
9
+ import { inspectSafeRegexes, type SafeRegexGroup, testSafeRegexGroups } from "./safe-regex.js";
10
+
11
+ const ValueMatchSchema = z
12
+ .object({
13
+ jsonPointer: z.string().default(""),
14
+ equals: z.unknown().optional(),
15
+ contains: z.string().min(1).optional(),
16
+ matches: z.string().max(2_000).optional(),
17
+ flags: z.string().max(16).optional(),
18
+ })
19
+ .strict()
20
+ .refine(
21
+ (config) =>
22
+ [config.equals, config.contains, config.matches].filter((value) => value !== undefined)
23
+ .length === 1,
24
+ { message: "set exactly one of equals, contains, or matches" },
25
+ )
26
+ .superRefine(async (config, ctx) => {
27
+ if (config.matches === undefined) {
28
+ if (config.flags !== undefined) {
29
+ ctx.addIssue({
30
+ code: "custom",
31
+ path: ["flags"],
32
+ message: "flags requires matches",
33
+ });
34
+ }
35
+ return;
36
+ }
37
+ const [inspection] = await inspectSafeRegexes([config.matches], config.flags ?? "");
38
+ if (inspection && "error" in inspection) {
39
+ ctx.addIssue({
40
+ code: "custom",
41
+ path: ["matches"],
42
+ message: inspection.error,
43
+ });
44
+ }
45
+ });
46
+ type ValueMatch = z.infer<typeof ValueMatchSchema>;
47
+ const MAX_TOOL_MATCH_CHARS = 1024 * 1024;
48
+ const MAX_TOOL_MATCH_TOTAL_CHARS = 8 * 1024 * 1024;
49
+
50
+ export const ToolInvocationConfigSchema = z
51
+ .object({
52
+ type: z.literal("tool-invocation"),
53
+ name: z.string(),
54
+ /** Filter paired results before choosing first, last, or any. */
55
+ result: z.enum(["success", "error", "any"]).default("success"),
56
+ occurrence: z.enum(["first", "last", "any"]).default("any"),
57
+ input: ValueMatchSchema.optional(),
58
+ output: ValueMatchSchema.optional(),
59
+ ...weightThresholdFields,
60
+ })
61
+ .strict();
62
+ export type ToolInvocationConfig = z.infer<typeof ToolInvocationConfigSchema>;
63
+
64
+ const valueText = (value: unknown): string | null => {
65
+ let text: string;
66
+ if (typeof value === "string") {
67
+ text = value;
68
+ } else {
69
+ try {
70
+ text = JSON.stringify(value) ?? String(value);
71
+ } catch {
72
+ return null;
73
+ }
74
+ }
75
+ return text.length <= MAX_TOOL_MATCH_CHARS ? text : null;
76
+ };
77
+
78
+ type PreparedValueMatch = {
79
+ matched: boolean;
80
+ regexText?: string;
81
+ };
82
+
83
+ const prepareValueMatch = (root: unknown, match: ValueMatch): PreparedValueMatch => {
84
+ const value = resolvePointer(root, match.jsonPointer);
85
+ if (match.equals !== undefined) return { matched: deepEquals(value, match.equals) };
86
+ if (match.contains !== undefined) {
87
+ if (value === undefined) return { matched: false };
88
+ return { matched: valueText(value)?.includes(match.contains) ?? false };
89
+ }
90
+ if (match.matches !== undefined) {
91
+ if (value === undefined) return { matched: false };
92
+ const text = valueText(value);
93
+ return text === null ? { matched: false } : { matched: true, regexText: text };
94
+ }
95
+ return { matched: false };
96
+ };
97
+
98
+ const selectOccurrences = <T>(values: T[], occurrence: ToolInvocationConfig["occurrence"]): T[] => {
99
+ if (occurrence === "first") return values.slice(0, 1);
100
+ if (occurrence === "last") return values.slice(-1);
101
+ return values;
102
+ };
103
+
104
+ export const toolInvocationAssertion: AssertionImpl<ToolInvocationConfig> = {
105
+ type: "tool-invocation",
106
+ configSchema: ToolInvocationConfigSchema,
107
+ evaluate: async (config, run, ctx) => {
108
+ const malformed = malformedToolEventCount(run.events, ["tool-call", "tool-result"]);
109
+ if (malformed > 0) {
110
+ const errorMessage = `${malformed} malformed standardized tool event${malformed === 1 ? "" : "s"} captured`;
111
+ return {
112
+ passed: false,
113
+ score: 0,
114
+ comment: errorMessage,
115
+ errorMessage,
116
+ };
117
+ }
118
+ const invocations = toolInvocations(run.events, config.name);
119
+ const selected = selectOccurrences(invocations, config.occurrence);
120
+ const paired = selected.filter(
121
+ (
122
+ invocation,
123
+ ): invocation is typeof invocation & {
124
+ result: NonNullable<typeof invocation.result>;
125
+ } => invocation.result !== undefined,
126
+ );
127
+ const relevant = paired.filter(({ result }) => {
128
+ if (config.result === "any") return true;
129
+ return result.isError === (config.result === "error");
130
+ });
131
+ const candidates: Array<{
132
+ inputRegexText?: string;
133
+ outputRegexText?: string;
134
+ }> = [];
135
+ let regexTextChars = 0;
136
+ for (const invocation of relevant) {
137
+ const input = config.input
138
+ ? prepareValueMatch(invocation.call.input, config.input)
139
+ : { matched: true };
140
+ if (!input.matched) continue;
141
+ const output = config.output
142
+ ? prepareValueMatch(toolResultValue(invocation.result), config.output)
143
+ : { matched: true };
144
+ if (!output.matched) continue;
145
+ regexTextChars += (input.regexText?.length ?? 0) + (output.regexText?.length ?? 0);
146
+ if (regexTextChars > MAX_TOOL_MATCH_TOTAL_CHARS) {
147
+ throw new RangeError(
148
+ `tool invocation regex evidence exceeds the ${MAX_TOOL_MATCH_TOTAL_CHARS}-character assertion limit`,
149
+ );
150
+ }
151
+ candidates.push({
152
+ ...(input.regexText !== undefined ? { inputRegexText: input.regexText } : {}),
153
+ ...(output.regexText !== undefined ? { outputRegexText: output.regexText } : {}),
154
+ });
155
+ }
156
+
157
+ const groups: SafeRegexGroup[] = [];
158
+ if (config.input?.matches !== undefined) {
159
+ groups.push({
160
+ pattern: config.input.matches,
161
+ flags: config.input.flags ?? "",
162
+ texts: candidates.map(({ inputRegexText }) => inputRegexText ?? ""),
163
+ });
164
+ }
165
+ if (config.output?.matches !== undefined) {
166
+ groups.push({
167
+ pattern: config.output.matches,
168
+ flags: config.output.flags ?? "",
169
+ texts: candidates.map(({ outputRegexText }) => outputRegexText ?? ""),
170
+ });
171
+ }
172
+ const regexResults =
173
+ groups.length > 0 ? await testSafeRegexGroups(groups, { abortSignal: ctx?.abortSignal }) : [];
174
+ let groupIndex = 0;
175
+ const inputResults =
176
+ config.input?.matches !== undefined ? regexResults[groupIndex++] : undefined;
177
+ const outputResults =
178
+ config.output?.matches !== undefined ? regexResults[groupIndex] : undefined;
179
+ const matches = candidates.filter(
180
+ (_candidate, index) => (inputResults?.[index] ?? true) && (outputResults?.[index] ?? true),
181
+ );
182
+ const passed = matches.length > 0;
183
+ return {
184
+ passed,
185
+ score: passed ? 1 : 0,
186
+ comment: passed
187
+ ? `'${config.name}' has a matching ${config.result} invocation`
188
+ : `'${config.name}' has no matching ${config.result} invocation`,
189
+ detail: {
190
+ calls: invocations.length,
191
+ selected: selected.length,
192
+ paired: paired.length,
193
+ resultMatches: relevant.length,
194
+ matches: matches.length,
195
+ },
196
+ };
197
+ },
198
+ };
@@ -0,0 +1,211 @@
1
+ import { z } from "zod";
2
+ import {
3
+ type AssertionImpl,
4
+ malformedToolEventCount,
5
+ type ToolResultPayload,
6
+ toolResults,
7
+ toolResultValue,
8
+ } from "../../core/index.js";
9
+ import { deepEquals, resolvePointer, weightThresholdFields } from "./helpers.js";
10
+
11
+ const OccurrenceSchema = z.enum(["first", "last", "any"]);
12
+ type Occurrence = z.infer<typeof OccurrenceSchema>;
13
+
14
+ const select = (results: ToolResultPayload[], occurrence: Occurrence): ToolResultPayload[] => {
15
+ if (occurrence === "first") return results.slice(0, 1);
16
+ if (occurrence === "last") return results.slice(-1);
17
+ return results;
18
+ };
19
+
20
+ const malformedResultEvidence = (run: Parameters<AssertionImpl["evaluate"]>[1]) => {
21
+ const malformed = malformedToolEventCount(run.events, ["tool-result"]);
22
+ if (malformed === 0) return null;
23
+ const errorMessage = `${malformed} malformed tool result event${malformed === 1 ? "" : "s"} captured`;
24
+ return {
25
+ passed: false as const,
26
+ score: 0 as const,
27
+ comment: errorMessage,
28
+ errorMessage,
29
+ };
30
+ };
31
+
32
+ export const ToolResultIsErrorConfigSchema = z
33
+ .object({
34
+ type: z.literal("tool-result-is-error"),
35
+ /** Tool whose result to inspect; omit to consider every result. */
36
+ name: z.string().optional(),
37
+ /** The `isError` value to require. Default true (assert the tool DID error). */
38
+ expected: z.boolean().default(true),
39
+ /** Which matching result(s) to require it of. Default "last". */
40
+ occurrence: OccurrenceSchema.default("last"),
41
+ ...weightThresholdFields,
42
+ })
43
+ .strict();
44
+ export type ToolResultIsErrorConfig = z.infer<typeof ToolResultIsErrorConfigSchema>;
45
+
46
+ export const toolResultIsErrorAssertion: AssertionImpl<ToolResultIsErrorConfig> = {
47
+ type: "tool-result-is-error",
48
+ configSchema: ToolResultIsErrorConfigSchema,
49
+ evaluate: async (config, run) => {
50
+ const malformed = malformedResultEvidence(run);
51
+ if (malformed) return malformed;
52
+ const results = toolResults(run.events, config.name);
53
+ const label = config.name ? `'${config.name}'` : "any tool";
54
+ if (results.length === 0) {
55
+ return {
56
+ passed: false,
57
+ score: 0,
58
+ comment: `no result captured for ${label}`,
59
+ };
60
+ }
61
+ const relevant = config.occurrence === "any" ? results : select(results, config.occurrence);
62
+ const passed =
63
+ config.occurrence === "any"
64
+ ? relevant.some((result) => result.isError === config.expected)
65
+ : relevant.every((result) => result.isError === config.expected);
66
+ return {
67
+ passed,
68
+ score: passed ? 1 : 0,
69
+ comment: `${label} isError expected ${config.expected} (${config.occurrence})`,
70
+ detail: { isError: relevant.map((result) => result.isError) },
71
+ };
72
+ },
73
+ };
74
+
75
+ export const ToolResultContainsConfigSchema = z
76
+ .object({
77
+ type: z.literal("tool-result-contains"),
78
+ name: z.string().optional(),
79
+ /** Substring the result's text (or stringified content) must contain. */
80
+ text: z.string().min(1),
81
+ occurrence: OccurrenceSchema.default("any"),
82
+ ...weightThresholdFields,
83
+ })
84
+ .strict();
85
+ export type ToolResultContainsConfig = z.infer<typeof ToolResultContainsConfigSchema>;
86
+
87
+ const resultText = (result: ToolResultPayload): string =>
88
+ result.text ||
89
+ (typeof result.content === "string" ? result.content : JSON.stringify(result.content ?? ""));
90
+
91
+ export const toolResultContainsAssertion: AssertionImpl<ToolResultContainsConfig> = {
92
+ type: "tool-result-contains",
93
+ configSchema: ToolResultContainsConfigSchema,
94
+ evaluate: async (config, run) => {
95
+ const malformed = malformedResultEvidence(run);
96
+ if (malformed) return malformed;
97
+ const results = toolResults(run.events, config.name);
98
+ const label = config.name ? `'${config.name}'` : "any tool";
99
+ if (results.length === 0) {
100
+ return {
101
+ passed: false,
102
+ score: 0,
103
+ comment: `no result captured for ${label}`,
104
+ };
105
+ }
106
+ const relevant = config.occurrence === "any" ? results : select(results, config.occurrence);
107
+ const hit = relevant.filter((result) => resultText(result).includes(config.text));
108
+ const passed = hit.length > 0;
109
+ return {
110
+ passed,
111
+ score: passed ? 1 : 0,
112
+ comment: `${label} result ${passed ? "contains" : "does not contain"} ${JSON.stringify(config.text)}`,
113
+ detail: { matches: hit.length },
114
+ };
115
+ },
116
+ };
117
+
118
+ export const ToolResultJsonPointerEqualsConfigSchema = z
119
+ .object({
120
+ type: z.literal("tool-result-json-pointer-equals"),
121
+ name: z.string(),
122
+ /**
123
+ * RFC 6901 pointer into the result's `structuredContent`, falling back to
124
+ * the result text parsed as JSON. Defaults to "" (the whole value).
125
+ */
126
+ jsonPointer: z.string().default(""),
127
+ /** Structural (key-order-insensitive) equality against the value. */
128
+ value: z.unknown().optional(),
129
+ /** Substring containment against the stringified value. */
130
+ contains: z.string().min(1).optional(),
131
+ occurrence: OccurrenceSchema.default("last"),
132
+ ...weightThresholdFields,
133
+ })
134
+ .strict()
135
+ .refine((config) => (config.value === undefined) !== (config.contains === undefined), {
136
+ message: "set exactly one of value or contains",
137
+ });
138
+ export type ToolResultJsonPointerEqualsConfig = z.infer<
139
+ typeof ToolResultJsonPointerEqualsConfigSchema
140
+ >;
141
+
142
+ export const toolResultJsonPointerEqualsAssertion: AssertionImpl<ToolResultJsonPointerEqualsConfig> =
143
+ {
144
+ type: "tool-result-json-pointer-equals",
145
+ configSchema: ToolResultJsonPointerEqualsConfigSchema,
146
+ evaluate: async (config, run) => {
147
+ const malformed = malformedResultEvidence(run);
148
+ if (malformed) return malformed;
149
+ const results = toolResults(run.events, config.name);
150
+ if (results.length === 0) {
151
+ return {
152
+ passed: false,
153
+ score: 0,
154
+ comment: `no result captured for '${config.name}'`,
155
+ };
156
+ }
157
+ const relevant = config.occurrence === "any" ? results : select(results, config.occurrence);
158
+ const matched = relevant.some((result) => {
159
+ const value = resolvePointer(toolResultValue(result), config.jsonPointer);
160
+ if (config.value !== undefined) return deepEquals(value, config.value);
161
+ if (config.contains !== undefined) {
162
+ if (value === undefined) return false;
163
+ const stringified = typeof value === "string" ? value : JSON.stringify(value);
164
+ return stringified.includes(config.contains);
165
+ }
166
+ return value !== undefined;
167
+ });
168
+ return {
169
+ passed: matched,
170
+ score: matched ? 1 : 0,
171
+ comment: `'${config.name}' result at '${config.jsonPointer || "/"}' ${matched ? "matched" : "did not match"}`,
172
+ };
173
+ },
174
+ };
175
+
176
+ export const NoErrorResultsConfigSchema = z
177
+ .object({
178
+ type: z.literal("no-error-results"),
179
+ /** Permit a run with no captured calls/results. Defaults false to prevent vacuous passes. */
180
+ allowEmpty: z.boolean().default(false),
181
+ ...weightThresholdFields,
182
+ })
183
+ .strict();
184
+ export type NoErrorResultsConfig = z.infer<typeof NoErrorResultsConfigSchema>;
185
+
186
+ export const noErrorResultsAssertion: AssertionImpl<NoErrorResultsConfig> = {
187
+ type: "no-error-results",
188
+ configSchema: NoErrorResultsConfigSchema,
189
+ evaluate: async (config, run) => {
190
+ const malformed = malformedResultEvidence(run);
191
+ if (malformed) return malformed;
192
+ const results = toolResults(run.events);
193
+ if (results.length === 0 && !config.allowEmpty) {
194
+ return {
195
+ passed: false,
196
+ score: 0,
197
+ comment: "no tool results captured",
198
+ };
199
+ }
200
+ const offenders = results.filter((result) => result.isError);
201
+ const passed = offenders.length === 0;
202
+ return {
203
+ passed,
204
+ score: passed ? 1 : 0,
205
+ comment: passed
206
+ ? "no tool returned an error result"
207
+ : `error results from: ${offenders.map((result) => result.name).join(", ")}`,
208
+ detail: { offenders: offenders.map((result) => result.name) },
209
+ };
210
+ },
211
+ };
@@ -0,0 +1,139 @@
1
+ import { z } from "zod";
2
+ import {
3
+ type AssertionImpl,
4
+ hasMcpInit,
5
+ hasMcpToolsList,
6
+ listedTools,
7
+ mcpInit,
8
+ } from "../../core/index.js";
9
+ import { inRange } from "../core/helpers.js";
10
+
11
+ export const ServerInfoEqualsConfigSchema = z
12
+ .object({
13
+ type: z.literal("server-info-equals"),
14
+ /** Expected `serverInfo.name` from `initialize`. */
15
+ name: z.string().optional(),
16
+ /** Expected `serverInfo.version` from `initialize`. */
17
+ version: z.string().optional(),
18
+ })
19
+ .strict()
20
+ .refine((c) => c.name !== undefined || c.version !== undefined, {
21
+ message: "set at least one of name or version",
22
+ });
23
+ export type ServerInfoEqualsConfig = z.infer<typeof ServerInfoEqualsConfigSchema>;
24
+
25
+ export const serverInfoEqualsAssertion: AssertionImpl<ServerInfoEqualsConfig> = {
26
+ type: "server-info-equals",
27
+ configSchema: ServerInfoEqualsConfigSchema,
28
+ evaluate: async (config, run) => {
29
+ const info = mcpInit(run.events)?.serverInfo;
30
+ if (!info) {
31
+ return { passed: false, score: 0, comment: "no initialize metadata captured" };
32
+ }
33
+ const mismatches: string[] = [];
34
+ if (config.name !== undefined && info.name !== config.name) {
35
+ mismatches.push(`name '${info.name}' != '${config.name}'`);
36
+ }
37
+ if (config.version !== undefined && info.version !== config.version) {
38
+ mismatches.push(`version '${info.version}' != '${config.version}'`);
39
+ }
40
+ const passed = mismatches.length === 0;
41
+ return {
42
+ passed,
43
+ score: passed ? 1 : 0,
44
+ comment: passed ? `serverInfo ${info.name}@${info.version} matched` : mismatches.join("; "),
45
+ detail: { serverInfo: info },
46
+ };
47
+ },
48
+ };
49
+
50
+ export const ToolListedConfigSchema = z
51
+ .object({
52
+ type: z.literal("tool-listed"),
53
+ /** Tool name whose presence is checked. */
54
+ name: z.string(),
55
+ /** Whether `tools/list` should include the tool. Default true. */
56
+ expected: z.boolean().default(true),
57
+ })
58
+ .strict();
59
+ export type ToolListedConfig = z.infer<typeof ToolListedConfigSchema>;
60
+
61
+ export const toolListedAssertion: AssertionImpl<ToolListedConfig> = {
62
+ type: "tool-listed",
63
+ configSchema: ToolListedConfigSchema,
64
+ evaluate: async (config, run) => {
65
+ if (!hasMcpToolsList(run.events)) {
66
+ return { passed: false, score: 0, comment: "no tools/list result captured" };
67
+ }
68
+ const tools = listedTools(run.events);
69
+ const listed = tools.some((t) => t.name === config.name);
70
+ const passed = listed === config.expected;
71
+ return {
72
+ passed,
73
+ score: passed ? 1 : 0,
74
+ comment: `tool '${config.name}' is ${listed ? "listed" : `not among ${tools.length} listed tools`}`,
75
+ };
76
+ },
77
+ };
78
+
79
+ export const ToolsListedCountConfigSchema = z
80
+ .object({
81
+ type: z.literal("tools-listed-count"),
82
+ min: z.number().int().nonnegative().optional(),
83
+ max: z.number().int().nonnegative().optional(),
84
+ })
85
+ .strict()
86
+ .refine((c) => c.min !== undefined || c.max !== undefined, {
87
+ message: "set at least one of min or max",
88
+ })
89
+ .refine((c) => c.min === undefined || c.max === undefined || c.min <= c.max, {
90
+ message: "min must be <= max",
91
+ });
92
+ export type ToolsListedCountConfig = z.infer<typeof ToolsListedCountConfigSchema>;
93
+
94
+ export const toolsListedCountAssertion: AssertionImpl<ToolsListedCountConfig> = {
95
+ type: "tools-listed-count",
96
+ configSchema: ToolsListedCountConfigSchema,
97
+ evaluate: async (config, run) => {
98
+ if (!hasMcpToolsList(run.events)) {
99
+ return { passed: false, score: 0, comment: "no tools/list result captured" };
100
+ }
101
+ const count = listedTools(run.events).length;
102
+ const { passed, comment } = inRange(count, config.min, config.max);
103
+ return {
104
+ passed,
105
+ score: passed ? 1 : 0,
106
+ comment: `${count} tool${count === 1 ? "" : "s"} listed (${comment})`,
107
+ detail: { count },
108
+ };
109
+ },
110
+ };
111
+
112
+ export const CapabilityAbsentConfigSchema = z
113
+ .object({
114
+ type: z.literal("capability-absent"),
115
+ /** Server capability that must NOT be advertised. */
116
+ capability: z.enum(["resources", "prompts", "logging", "completions", "tools"]),
117
+ })
118
+ .strict();
119
+ export type CapabilityAbsentConfig = z.infer<typeof CapabilityAbsentConfigSchema>;
120
+
121
+ export const capabilityAbsentAssertion: AssertionImpl<CapabilityAbsentConfig> = {
122
+ type: "capability-absent",
123
+ configSchema: CapabilityAbsentConfigSchema,
124
+ evaluate: async (config, run) => {
125
+ if (!hasMcpInit(run.events)) {
126
+ return { passed: false, score: 0, comment: "no initialize metadata captured" };
127
+ }
128
+ const caps = mcpInit(run.events)?.capabilities ?? {};
129
+ const value = (caps as Record<string, unknown>)[config.capability];
130
+ const present = value !== undefined && value !== false && value !== null;
131
+ return {
132
+ passed: !present,
133
+ score: present ? 0 : 1,
134
+ comment: present
135
+ ? `server advertises '${config.capability}'`
136
+ : `server does not advertise '${config.capability}'`,
137
+ };
138
+ },
139
+ };
@@ -0,0 +1,37 @@
1
+ import type { AssertionImpl, Registry } from "../../core/index.js";
2
+ import {
3
+ noErrorResultsAssertion,
4
+ toolResultContainsAssertion,
5
+ toolResultIsErrorAssertion,
6
+ toolResultJsonPointerEqualsAssertion,
7
+ } from "../core/tool-result-checks.js";
8
+ import {
9
+ capabilityAbsentAssertion,
10
+ serverInfoEqualsAssertion,
11
+ toolListedAssertion,
12
+ toolsListedCountAssertion,
13
+ } from "./discovery-checks.js";
14
+
15
+ export * from "./discovery-checks.js";
16
+
17
+ /**
18
+ * The result assertions remain in this compatibility collection for
19
+ * embedders that historically registered only the MCP assertion bundle.
20
+ * Their implementations are generic and also ship in `allCoreAssertions`.
21
+ */
22
+ export const allMcpAssertions: AssertionImpl[] = [
23
+ toolResultIsErrorAssertion,
24
+ toolResultContainsAssertion,
25
+ toolResultJsonPointerEqualsAssertion,
26
+ noErrorResultsAssertion,
27
+ serverInfoEqualsAssertion,
28
+ toolListedAssertion,
29
+ toolsListedCountAssertion,
30
+ capabilityAbsentAssertion,
31
+ ];
32
+
33
+ export const registerMcpAssertions = (registry: Registry): void => {
34
+ for (const a of allMcpAssertions) {
35
+ if (!registry.getAssertion(a.type)) registry.registerAssertion(a);
36
+ }
37
+ };