@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 +1,170 @@
1
+ import { z } from "zod";
2
+ import { malformedToolEventCount, toolInvocations, toolResultValue, } from "../../core/index.js";
3
+ import { deepEquals, resolvePointer, weightThresholdFields } from "./helpers.js";
4
+ import { inspectSafeRegexes, testSafeRegexGroups } from "./safe-regex.js";
5
+ const ValueMatchSchema = z
6
+ .object({
7
+ jsonPointer: z.string().default(""),
8
+ equals: z.unknown().optional(),
9
+ contains: z.string().min(1).optional(),
10
+ matches: z.string().max(2_000).optional(),
11
+ flags: z.string().max(16).optional(),
12
+ })
13
+ .strict()
14
+ .refine((config) => [config.equals, config.contains, config.matches].filter((value) => value !== undefined)
15
+ .length === 1, { message: "set exactly one of equals, contains, or matches" })
16
+ .superRefine(async (config, ctx) => {
17
+ if (config.matches === undefined) {
18
+ if (config.flags !== undefined) {
19
+ ctx.addIssue({
20
+ code: "custom",
21
+ path: ["flags"],
22
+ message: "flags requires matches",
23
+ });
24
+ }
25
+ return;
26
+ }
27
+ const [inspection] = await inspectSafeRegexes([config.matches], config.flags ?? "");
28
+ if (inspection && "error" in inspection) {
29
+ ctx.addIssue({
30
+ code: "custom",
31
+ path: ["matches"],
32
+ message: inspection.error,
33
+ });
34
+ }
35
+ });
36
+ const MAX_TOOL_MATCH_CHARS = 1024 * 1024;
37
+ const MAX_TOOL_MATCH_TOTAL_CHARS = 8 * 1024 * 1024;
38
+ export const ToolInvocationConfigSchema = z
39
+ .object({
40
+ type: z.literal("tool-invocation"),
41
+ name: z.string(),
42
+ /** Filter paired results before choosing first, last, or any. */
43
+ result: z.enum(["success", "error", "any"]).default("success"),
44
+ occurrence: z.enum(["first", "last", "any"]).default("any"),
45
+ input: ValueMatchSchema.optional(),
46
+ output: ValueMatchSchema.optional(),
47
+ ...weightThresholdFields,
48
+ })
49
+ .strict();
50
+ const valueText = (value) => {
51
+ let text;
52
+ if (typeof value === "string") {
53
+ text = value;
54
+ }
55
+ else {
56
+ try {
57
+ text = JSON.stringify(value) ?? String(value);
58
+ }
59
+ catch {
60
+ return null;
61
+ }
62
+ }
63
+ return text.length <= MAX_TOOL_MATCH_CHARS ? text : null;
64
+ };
65
+ const prepareValueMatch = (root, match) => {
66
+ const value = resolvePointer(root, match.jsonPointer);
67
+ if (match.equals !== undefined)
68
+ return { matched: deepEquals(value, match.equals) };
69
+ if (match.contains !== undefined) {
70
+ if (value === undefined)
71
+ return { matched: false };
72
+ return { matched: valueText(value)?.includes(match.contains) ?? false };
73
+ }
74
+ if (match.matches !== undefined) {
75
+ if (value === undefined)
76
+ return { matched: false };
77
+ const text = valueText(value);
78
+ return text === null ? { matched: false } : { matched: true, regexText: text };
79
+ }
80
+ return { matched: false };
81
+ };
82
+ const selectOccurrences = (values, occurrence) => {
83
+ if (occurrence === "first")
84
+ return values.slice(0, 1);
85
+ if (occurrence === "last")
86
+ return values.slice(-1);
87
+ return values;
88
+ };
89
+ export const toolInvocationAssertion = {
90
+ type: "tool-invocation",
91
+ configSchema: ToolInvocationConfigSchema,
92
+ evaluate: async (config, run, ctx) => {
93
+ const malformed = malformedToolEventCount(run.events, ["tool-call", "tool-result"]);
94
+ if (malformed > 0) {
95
+ const errorMessage = `${malformed} malformed standardized tool event${malformed === 1 ? "" : "s"} captured`;
96
+ return {
97
+ passed: false,
98
+ score: 0,
99
+ comment: errorMessage,
100
+ errorMessage,
101
+ };
102
+ }
103
+ const invocations = toolInvocations(run.events, config.name);
104
+ const selected = selectOccurrences(invocations, config.occurrence);
105
+ const paired = selected.filter((invocation) => invocation.result !== undefined);
106
+ const relevant = paired.filter(({ result }) => {
107
+ if (config.result === "any")
108
+ return true;
109
+ return result.isError === (config.result === "error");
110
+ });
111
+ const candidates = [];
112
+ let regexTextChars = 0;
113
+ for (const invocation of relevant) {
114
+ const input = config.input
115
+ ? prepareValueMatch(invocation.call.input, config.input)
116
+ : { matched: true };
117
+ if (!input.matched)
118
+ continue;
119
+ const output = config.output
120
+ ? prepareValueMatch(toolResultValue(invocation.result), config.output)
121
+ : { matched: true };
122
+ if (!output.matched)
123
+ continue;
124
+ regexTextChars += (input.regexText?.length ?? 0) + (output.regexText?.length ?? 0);
125
+ if (regexTextChars > MAX_TOOL_MATCH_TOTAL_CHARS) {
126
+ throw new RangeError(`tool invocation regex evidence exceeds the ${MAX_TOOL_MATCH_TOTAL_CHARS}-character assertion limit`);
127
+ }
128
+ candidates.push({
129
+ ...(input.regexText !== undefined ? { inputRegexText: input.regexText } : {}),
130
+ ...(output.regexText !== undefined ? { outputRegexText: output.regexText } : {}),
131
+ });
132
+ }
133
+ const groups = [];
134
+ if (config.input?.matches !== undefined) {
135
+ groups.push({
136
+ pattern: config.input.matches,
137
+ flags: config.input.flags ?? "",
138
+ texts: candidates.map(({ inputRegexText }) => inputRegexText ?? ""),
139
+ });
140
+ }
141
+ if (config.output?.matches !== undefined) {
142
+ groups.push({
143
+ pattern: config.output.matches,
144
+ flags: config.output.flags ?? "",
145
+ texts: candidates.map(({ outputRegexText }) => outputRegexText ?? ""),
146
+ });
147
+ }
148
+ const regexResults = groups.length > 0 ? await testSafeRegexGroups(groups, { abortSignal: ctx?.abortSignal }) : [];
149
+ let groupIndex = 0;
150
+ const inputResults = config.input?.matches !== undefined ? regexResults[groupIndex++] : undefined;
151
+ const outputResults = config.output?.matches !== undefined ? regexResults[groupIndex] : undefined;
152
+ const matches = candidates.filter((_candidate, index) => (inputResults?.[index] ?? true) && (outputResults?.[index] ?? true));
153
+ const passed = matches.length > 0;
154
+ return {
155
+ passed,
156
+ score: passed ? 1 : 0,
157
+ comment: passed
158
+ ? `'${config.name}' has a matching ${config.result} invocation`
159
+ : `'${config.name}' has no matching ${config.result} invocation`,
160
+ detail: {
161
+ calls: invocations.length,
162
+ selected: selected.length,
163
+ paired: paired.length,
164
+ resultMatches: relevant.length,
165
+ matches: matches.length,
166
+ },
167
+ };
168
+ },
169
+ };
170
+ //# sourceMappingURL=tool-invocation.js.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"tool-invocation.js","sourceRoot":"","sources":["../../../src/assertions/core/tool-invocation.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AACxB,OAAO,EAEL,uBAAuB,EACvB,eAAe,EACf,eAAe,GAChB,MAAM,qBAAqB,CAAC;AAC7B,OAAO,EAAE,UAAU,EAAE,cAAc,EAAE,qBAAqB,EAAE,MAAM,cAAc,CAAC;AACjF,OAAO,EAAE,kBAAkB,EAAuB,mBAAmB,EAAE,MAAM,iBAAiB,CAAC;AAE/F,MAAM,gBAAgB,GAAG,CAAC;KACvB,MAAM,CAAC;IACN,WAAW,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,OAAO,CAAC,EAAE,CAAC;IACnC,MAAM,EAAE,CAAC,CAAC,OAAO,EAAE,CAAC,QAAQ,EAAE;IAC9B,QAAQ,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,QAAQ,EAAE;IACtC,OAAO,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,GAAG,CAAC,KAAK,CAAC,CAAC,QAAQ,EAAE;IACzC,KAAK,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC,QAAQ,EAAE;CACrC,CAAC;KACD,MAAM,EAAE;KACR,MAAM,CACL,CAAC,MAAM,EAAE,EAAE,CACT,CAAC,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,QAAQ,EAAE,MAAM,CAAC,OAAO,CAAC,CAAC,MAAM,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,KAAK,SAAS,CAAC;KACpF,MAAM,KAAK,CAAC,EACjB,EAAE,OAAO,EAAE,iDAAiD,EAAE,CAC/D;KACA,WAAW,CAAC,KAAK,EAAE,MAAM,EAAE,GAAG,EAAE,EAAE;IACjC,IAAI,MAAM,CAAC,OAAO,KAAK,SAAS,EAAE,CAAC;QACjC,IAAI,MAAM,CAAC,KAAK,KAAK,SAAS,EAAE,CAAC;YAC/B,GAAG,CAAC,QAAQ,CAAC;gBACX,IAAI,EAAE,QAAQ;gBACd,IAAI,EAAE,CAAC,OAAO,CAAC;gBACf,OAAO,EAAE,wBAAwB;aAClC,CAAC,CAAC;QACL,CAAC;QACD,OAAO;IACT,CAAC;IACD,MAAM,CAAC,UAAU,CAAC,GAAG,MAAM,kBAAkB,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,EAAE,MAAM,CAAC,KAAK,IAAI,EAAE,CAAC,CAAC;IACpF,IAAI,UAAU,IAAI,OAAO,IAAI,UAAU,EAAE,CAAC;QACxC,GAAG,CAAC,QAAQ,CAAC;YACX,IAAI,EAAE,QAAQ;YACd,IAAI,EAAE,CAAC,SAAS,CAAC;YACjB,OAAO,EAAE,UAAU,CAAC,KAAK;SAC1B,CAAC,CAAC;IACL,CAAC;AACH,CAAC,CAAC,CAAC;AAEL,MAAM,oBAAoB,GAAG,IAAI,GAAG,IAAI,CAAC;AACzC,MAAM,0BAA0B,GAAG,CAAC,GAAG,IAAI,GAAG,IAAI,CAAC;AAEnD,MAAM,CAAC,MAAM,0BAA0B,GAAG,CAAC;KACxC,MAAM,CAAC;IACN,IAAI,EAAE,CAAC,CAAC,OAAO,CAAC,iBAAiB,CAAC;IAClC,IAAI,EAAE,CAAC,CAAC,MAAM,EAAE;IAChB,iEAAiE;IACjE,MAAM,EAAE,CAAC,CAAC,IAAI,CAAC,CAAC,SAAS,EAAE,OAAO,EAAE,KAAK,CAAC,CAAC,CAAC,OAAO,CAAC,SAAS,CAAC;IAC9D,UAAU,EAAE,CAAC,CAAC,IAAI,CAAC,CAAC,OAAO,EAAE,MAAM,EAAE,KAAK,CAAC,CAAC,CAAC,OAAO,CAAC,KAAK,CAAC;IAC3D,KAAK,EAAE,gBAAgB,CAAC,QAAQ,EAAE;IAClC,MAAM,EAAE,gBAAgB,CAAC,QAAQ,EAAE;IACnC,GAAG,qBAAqB;CACzB,CAAC;KACD,MAAM,EAAE,CAAC;AAGZ,MAAM,SAAS,GAAG,CAAC,KAAc,EAAiB,EAAE;IAClD,IAAI,IAAY,CAAC;IACjB,IAAI,OAAO,KAAK,KAAK,QAAQ,EAAE,CAAC;QAC9B,IAAI,GAAG,KAAK,CAAC;IACf,CAAC;SAAM,CAAC;QACN,IAAI,CAAC;YACH,IAAI,GAAG,IAAI,CAAC,SAAS,CAAC,KAAK,CAAC,IAAI,MAAM,CAAC,KAAK,CAAC,CAAC;QAChD,CAAC;QAAC,MAAM,CAAC;YACP,OAAO,IAAI,CAAC;QACd,CAAC;IACH,CAAC;IACD,OAAO,IAAI,CAAC,MAAM,IAAI,oBAAoB,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,IAAI,CAAC;AAC3D,CAAC,CAAC;AAOF,MAAM,iBAAiB,GAAG,CAAC,IAAa,EAAE,KAAiB,EAAsB,EAAE;IACjF,MAAM,KAAK,GAAG,cAAc,CAAC,IAAI,EAAE,KAAK,CAAC,WAAW,CAAC,CAAC;IACtD,IAAI,KAAK,CAAC,MAAM,KAAK,SAAS;QAAE,OAAO,EAAE,OAAO,EAAE,UAAU,CAAC,KAAK,EAAE,KAAK,CAAC,MAAM,CAAC,EAAE,CAAC;IACpF,IAAI,KAAK,CAAC,QAAQ,KAAK,SAAS,EAAE,CAAC;QACjC,IAAI,KAAK,KAAK,SAAS;YAAE,OAAO,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC;QACnD,OAAO,EAAE,OAAO,EAAE,SAAS,CAAC,KAAK,CAAC,EAAE,QAAQ,CAAC,KAAK,CAAC,QAAQ,CAAC,IAAI,KAAK,EAAE,CAAC;IAC1E,CAAC;IACD,IAAI,KAAK,CAAC,OAAO,KAAK,SAAS,EAAE,CAAC;QAChC,IAAI,KAAK,KAAK,SAAS;YAAE,OAAO,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC;QACnD,MAAM,IAAI,GAAG,SAAS,CAAC,KAAK,CAAC,CAAC;QAC9B,OAAO,IAAI,KAAK,IAAI,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,IAAI,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC;IACjF,CAAC;IACD,OAAO,EAAE,OAAO,EAAE,KAAK,EAAE,CAAC;AAC5B,CAAC,CAAC;AAEF,MAAM,iBAAiB,GAAG,CAAI,MAAW,EAAE,UAA8C,EAAO,EAAE;IAChG,IAAI,UAAU,KAAK,OAAO;QAAE,OAAO,MAAM,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;IACtD,IAAI,UAAU,KAAK,MAAM;QAAE,OAAO,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC;IACnD,OAAO,MAAM,CAAC;AAChB,CAAC,CAAC;AAEF,MAAM,CAAC,MAAM,uBAAuB,GAAwC;IAC1E,IAAI,EAAE,iBAAiB;IACvB,YAAY,EAAE,0BAA0B;IACxC,QAAQ,EAAE,KAAK,EAAE,MAAM,EAAE,GAAG,EAAE,GAAG,EAAE,EAAE;QACnC,MAAM,SAAS,GAAG,uBAAuB,CAAC,GAAG,CAAC,MAAM,EAAE,CAAC,WAAW,EAAE,aAAa,CAAC,CAAC,CAAC;QACpF,IAAI,SAAS,GAAG,CAAC,EAAE,CAAC;YAClB,MAAM,YAAY,GAAG,GAAG,SAAS,qCAAqC,SAAS,KAAK,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,GAAG,WAAW,CAAC;YAC5G,OAAO;gBACL,MAAM,EAAE,KAAK;gBACb,KAAK,EAAE,CAAC;gBACR,OAAO,EAAE,YAAY;gBACrB,YAAY;aACb,CAAC;QACJ,CAAC;QACD,MAAM,WAAW,GAAG,eAAe,CAAC,GAAG,CAAC,MAAM,EAAE,MAAM,CAAC,IAAI,CAAC,CAAC;QAC7D,MAAM,QAAQ,GAAG,iBAAiB,CAAC,WAAW,EAAE,MAAM,CAAC,UAAU,CAAC,CAAC;QACnE,MAAM,MAAM,GAAG,QAAQ,CAAC,MAAM,CAC5B,CACE,UAAU,EAGV,EAAE,CAAC,UAAU,CAAC,MAAM,KAAK,SAAS,CACrC,CAAC;QACF,MAAM,QAAQ,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC,EAAE,MAAM,EAAE,EAAE,EAAE;YAC5C,IAAI,MAAM,CAAC,MAAM,KAAK,KAAK;gBAAE,OAAO,IAAI,CAAC;YACzC,OAAO,MAAM,CAAC,OAAO,KAAK,CAAC,MAAM,CAAC,MAAM,KAAK,OAAO,CAAC,CAAC;QACxD,CAAC,CAAC,CAAC;QACH,MAAM,UAAU,GAGX,EAAE,CAAC;QACR,IAAI,cAAc,GAAG,CAAC,CAAC;QACvB,KAAK,MAAM,UAAU,IAAI,QAAQ,EAAE,CAAC;YAClC,MAAM,KAAK,GAAG,MAAM,CAAC,KAAK;gBACxB,CAAC,CAAC,iBAAiB,CAAC,UAAU,CAAC,IAAI,CAAC,KAAK,EAAE,MAAM,CAAC,KAAK,CAAC;gBACxD,CAAC,CAAC,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC;YACtB,IAAI,CAAC,KAAK,CAAC,OAAO;gBAAE,SAAS;YAC7B,MAAM,MAAM,GAAG,MAAM,CAAC,MAAM;gBAC1B,CAAC,CAAC,iBAAiB,CAAC,eAAe,CAAC,UAAU,CAAC,MAAM,CAAC,EAAE,MAAM,CAAC,MAAM,CAAC;gBACtE,CAAC,CAAC,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC;YACtB,IAAI,CAAC,MAAM,CAAC,OAAO;gBAAE,SAAS;YAC9B,cAAc,IAAI,CAAC,KAAK,CAAC,SAAS,EAAE,MAAM,IAAI,CAAC,CAAC,GAAG,CAAC,MAAM,CAAC,SAAS,EAAE,MAAM,IAAI,CAAC,CAAC,CAAC;YACnF,IAAI,cAAc,GAAG,0BAA0B,EAAE,CAAC;gBAChD,MAAM,IAAI,UAAU,CAClB,8CAA8C,0BAA0B,4BAA4B,CACrG,CAAC;YACJ,CAAC;YACD,UAAU,CAAC,IAAI,CAAC;gBACd,GAAG,CAAC,KAAK,CAAC,SAAS,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,cAAc,EAAE,KAAK,CAAC,SAAS,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;gBAC7E,GAAG,CAAC,MAAM,CAAC,SAAS,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,eAAe,EAAE,MAAM,CAAC,SAAS,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;aACjF,CAAC,CAAC;QACL,CAAC;QAED,MAAM,MAAM,GAAqB,EAAE,CAAC;QACpC,IAAI,MAAM,CAAC,KAAK,EAAE,OAAO,KAAK,SAAS,EAAE,CAAC;YACxC,MAAM,CAAC,IAAI,CAAC;gBACV,OAAO,EAAE,MAAM,CAAC,KAAK,CAAC,OAAO;gBAC7B,KAAK,EAAE,MAAM,CAAC,KAAK,CAAC,KAAK,IAAI,EAAE;gBAC/B,KAAK,EAAE,UAAU,CAAC,GAAG,CAAC,CAAC,EAAE,cAAc,EAAE,EAAE,EAAE,CAAC,cAAc,IAAI,EAAE,CAAC;aACpE,CAAC,CAAC;QACL,CAAC;QACD,IAAI,MAAM,CAAC,MAAM,EAAE,OAAO,KAAK,SAAS,EAAE,CAAC;YACzC,MAAM,CAAC,IAAI,CAAC;gBACV,OAAO,EAAE,MAAM,CAAC,MAAM,CAAC,OAAO;gBAC9B,KAAK,EAAE,MAAM,CAAC,MAAM,CAAC,KAAK,IAAI,EAAE;gBAChC,KAAK,EAAE,UAAU,CAAC,GAAG,CAAC,CAAC,EAAE,eAAe,EAAE,EAAE,EAAE,CAAC,eAAe,IAAI,EAAE,CAAC;aACtE,CAAC,CAAC;QACL,CAAC;QACD,MAAM,YAAY,GAChB,MAAM,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,MAAM,mBAAmB,CAAC,MAAM,EAAE,EAAE,WAAW,EAAE,GAAG,EAAE,WAAW,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;QAChG,IAAI,UAAU,GAAG,CAAC,CAAC;QACnB,MAAM,YAAY,GAChB,MAAM,CAAC,KAAK,EAAE,OAAO,KAAK,SAAS,CAAC,CAAC,CAAC,YAAY,CAAC,UAAU,EAAE,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;QAC/E,MAAM,aAAa,GACjB,MAAM,CAAC,MAAM,EAAE,OAAO,KAAK,SAAS,CAAC,CAAC,CAAC,YAAY,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;QAC9E,MAAM,OAAO,GAAG,UAAU,CAAC,MAAM,CAC/B,CAAC,UAAU,EAAE,KAAK,EAAE,EAAE,CAAC,CAAC,YAAY,EAAE,CAAC,KAAK,CAAC,IAAI,IAAI,CAAC,IAAI,CAAC,aAAa,EAAE,CAAC,KAAK,CAAC,IAAI,IAAI,CAAC,CAC3F,CAAC;QACF,MAAM,MAAM,GAAG,OAAO,CAAC,MAAM,GAAG,CAAC,CAAC;QAClC,OAAO;YACL,MAAM;YACN,KAAK,EAAE,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;YACrB,OAAO,EAAE,MAAM;gBACb,CAAC,CAAC,IAAI,MAAM,CAAC,IAAI,oBAAoB,MAAM,CAAC,MAAM,aAAa;gBAC/D,CAAC,CAAC,IAAI,MAAM,CAAC,IAAI,qBAAqB,MAAM,CAAC,MAAM,aAAa;YAClE,MAAM,EAAE;gBACN,KAAK,EAAE,WAAW,CAAC,MAAM;gBACzB,QAAQ,EAAE,QAAQ,CAAC,MAAM;gBACzB,MAAM,EAAE,MAAM,CAAC,MAAM;gBACrB,aAAa,EAAE,QAAQ,CAAC,MAAM;gBAC9B,OAAO,EAAE,OAAO,CAAC,MAAM;aACxB;SACF,CAAC;IACJ,CAAC;CACF,CAAC"}
@@ -0,0 +1,59 @@
1
+ import { z } from "zod";
2
+ import { type AssertionImpl } from "../../core/index.js";
3
+ export declare const ToolResultIsErrorConfigSchema: z.ZodObject<{
4
+ id: z.ZodOptional<z.ZodString>;
5
+ weight: z.ZodOptional<z.ZodNumber>;
6
+ threshold: z.ZodOptional<z.ZodNumber>;
7
+ type: z.ZodLiteral<"tool-result-is-error">;
8
+ name: z.ZodOptional<z.ZodString>;
9
+ expected: z.ZodDefault<z.ZodBoolean>;
10
+ occurrence: z.ZodDefault<z.ZodEnum<{
11
+ any: "any";
12
+ first: "first";
13
+ last: "last";
14
+ }>>;
15
+ }, z.core.$strict>;
16
+ export type ToolResultIsErrorConfig = z.infer<typeof ToolResultIsErrorConfigSchema>;
17
+ export declare const toolResultIsErrorAssertion: AssertionImpl<ToolResultIsErrorConfig>;
18
+ export declare const ToolResultContainsConfigSchema: z.ZodObject<{
19
+ id: z.ZodOptional<z.ZodString>;
20
+ weight: z.ZodOptional<z.ZodNumber>;
21
+ threshold: z.ZodOptional<z.ZodNumber>;
22
+ type: z.ZodLiteral<"tool-result-contains">;
23
+ name: z.ZodOptional<z.ZodString>;
24
+ text: z.ZodString;
25
+ occurrence: z.ZodDefault<z.ZodEnum<{
26
+ any: "any";
27
+ first: "first";
28
+ last: "last";
29
+ }>>;
30
+ }, z.core.$strict>;
31
+ export type ToolResultContainsConfig = z.infer<typeof ToolResultContainsConfigSchema>;
32
+ export declare const toolResultContainsAssertion: AssertionImpl<ToolResultContainsConfig>;
33
+ export declare const ToolResultJsonPointerEqualsConfigSchema: z.ZodObject<{
34
+ id: z.ZodOptional<z.ZodString>;
35
+ weight: z.ZodOptional<z.ZodNumber>;
36
+ threshold: z.ZodOptional<z.ZodNumber>;
37
+ type: z.ZodLiteral<"tool-result-json-pointer-equals">;
38
+ name: z.ZodString;
39
+ jsonPointer: z.ZodDefault<z.ZodString>;
40
+ value: z.ZodOptional<z.ZodUnknown>;
41
+ contains: z.ZodOptional<z.ZodString>;
42
+ occurrence: z.ZodDefault<z.ZodEnum<{
43
+ any: "any";
44
+ first: "first";
45
+ last: "last";
46
+ }>>;
47
+ }, z.core.$strict>;
48
+ export type ToolResultJsonPointerEqualsConfig = z.infer<typeof ToolResultJsonPointerEqualsConfigSchema>;
49
+ export declare const toolResultJsonPointerEqualsAssertion: AssertionImpl<ToolResultJsonPointerEqualsConfig>;
50
+ export declare const NoErrorResultsConfigSchema: z.ZodObject<{
51
+ id: z.ZodOptional<z.ZodString>;
52
+ weight: z.ZodOptional<z.ZodNumber>;
53
+ threshold: z.ZodOptional<z.ZodNumber>;
54
+ type: z.ZodLiteral<"no-error-results">;
55
+ allowEmpty: z.ZodDefault<z.ZodBoolean>;
56
+ }, z.core.$strict>;
57
+ export type NoErrorResultsConfig = z.infer<typeof NoErrorResultsConfigSchema>;
58
+ export declare const noErrorResultsAssertion: AssertionImpl<NoErrorResultsConfig>;
59
+ //# sourceMappingURL=tool-result-checks.d.ts.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"tool-result-checks.d.ts","sourceRoot":"","sources":["../../../src/assertions/core/tool-result-checks.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AACxB,OAAO,EACL,KAAK,aAAa,EAKnB,MAAM,qBAAqB,CAAC;AAwB7B,eAAO,MAAM,6BAA6B;;;;;;;;;;;;kBAW/B,CAAC;AACZ,MAAM,MAAM,uBAAuB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,6BAA6B,CAAC,CAAC;AAEpF,eAAO,MAAM,0BAA0B,EAAE,aAAa,CAAC,uBAAuB,CA2B7E,CAAC;AAEF,eAAO,MAAM,8BAA8B;;;;;;;;;;;;kBAShC,CAAC;AACZ,MAAM,MAAM,wBAAwB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,8BAA8B,CAAC,CAAC;AAMtF,eAAO,MAAM,2BAA2B,EAAE,aAAa,CAAC,wBAAwB,CAyB/E,CAAC;AAEF,eAAO,MAAM,uCAAuC;;;;;;;;;;;;;;kBAmBhD,CAAC;AACL,MAAM,MAAM,iCAAiC,GAAG,CAAC,CAAC,KAAK,CACrD,OAAO,uCAAuC,CAC/C,CAAC;AAEF,eAAO,MAAM,oCAAoC,EAAE,aAAa,CAAC,iCAAiC,CAgC/F,CAAC;AAEJ,eAAO,MAAM,0BAA0B;;;;;;kBAO5B,CAAC;AACZ,MAAM,MAAM,oBAAoB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,0BAA0B,CAAC,CAAC;AAE9E,eAAO,MAAM,uBAAuB,EAAE,aAAa,CAAC,oBAAoB,CAyBvE,CAAC"}
@@ -0,0 +1,193 @@
1
+ import { z } from "zod";
2
+ import { malformedToolEventCount, toolResults, toolResultValue, } from "../../core/index.js";
3
+ import { deepEquals, resolvePointer, weightThresholdFields } from "./helpers.js";
4
+ const OccurrenceSchema = z.enum(["first", "last", "any"]);
5
+ const select = (results, occurrence) => {
6
+ if (occurrence === "first")
7
+ return results.slice(0, 1);
8
+ if (occurrence === "last")
9
+ return results.slice(-1);
10
+ return results;
11
+ };
12
+ const malformedResultEvidence = (run) => {
13
+ const malformed = malformedToolEventCount(run.events, ["tool-result"]);
14
+ if (malformed === 0)
15
+ return null;
16
+ const errorMessage = `${malformed} malformed tool result event${malformed === 1 ? "" : "s"} captured`;
17
+ return {
18
+ passed: false,
19
+ score: 0,
20
+ comment: errorMessage,
21
+ errorMessage,
22
+ };
23
+ };
24
+ export const ToolResultIsErrorConfigSchema = z
25
+ .object({
26
+ type: z.literal("tool-result-is-error"),
27
+ /** Tool whose result to inspect; omit to consider every result. */
28
+ name: z.string().optional(),
29
+ /** The `isError` value to require. Default true (assert the tool DID error). */
30
+ expected: z.boolean().default(true),
31
+ /** Which matching result(s) to require it of. Default "last". */
32
+ occurrence: OccurrenceSchema.default("last"),
33
+ ...weightThresholdFields,
34
+ })
35
+ .strict();
36
+ export const toolResultIsErrorAssertion = {
37
+ type: "tool-result-is-error",
38
+ configSchema: ToolResultIsErrorConfigSchema,
39
+ evaluate: async (config, run) => {
40
+ const malformed = malformedResultEvidence(run);
41
+ if (malformed)
42
+ return malformed;
43
+ const results = toolResults(run.events, config.name);
44
+ const label = config.name ? `'${config.name}'` : "any tool";
45
+ if (results.length === 0) {
46
+ return {
47
+ passed: false,
48
+ score: 0,
49
+ comment: `no result captured for ${label}`,
50
+ };
51
+ }
52
+ const relevant = config.occurrence === "any" ? results : select(results, config.occurrence);
53
+ const passed = config.occurrence === "any"
54
+ ? relevant.some((result) => result.isError === config.expected)
55
+ : relevant.every((result) => result.isError === config.expected);
56
+ return {
57
+ passed,
58
+ score: passed ? 1 : 0,
59
+ comment: `${label} isError expected ${config.expected} (${config.occurrence})`,
60
+ detail: { isError: relevant.map((result) => result.isError) },
61
+ };
62
+ },
63
+ };
64
+ export const ToolResultContainsConfigSchema = z
65
+ .object({
66
+ type: z.literal("tool-result-contains"),
67
+ name: z.string().optional(),
68
+ /** Substring the result's text (or stringified content) must contain. */
69
+ text: z.string().min(1),
70
+ occurrence: OccurrenceSchema.default("any"),
71
+ ...weightThresholdFields,
72
+ })
73
+ .strict();
74
+ const resultText = (result) => result.text ||
75
+ (typeof result.content === "string" ? result.content : JSON.stringify(result.content ?? ""));
76
+ export const toolResultContainsAssertion = {
77
+ type: "tool-result-contains",
78
+ configSchema: ToolResultContainsConfigSchema,
79
+ evaluate: async (config, run) => {
80
+ const malformed = malformedResultEvidence(run);
81
+ if (malformed)
82
+ return malformed;
83
+ const results = toolResults(run.events, config.name);
84
+ const label = config.name ? `'${config.name}'` : "any tool";
85
+ if (results.length === 0) {
86
+ return {
87
+ passed: false,
88
+ score: 0,
89
+ comment: `no result captured for ${label}`,
90
+ };
91
+ }
92
+ const relevant = config.occurrence === "any" ? results : select(results, config.occurrence);
93
+ const hit = relevant.filter((result) => resultText(result).includes(config.text));
94
+ const passed = hit.length > 0;
95
+ return {
96
+ passed,
97
+ score: passed ? 1 : 0,
98
+ comment: `${label} result ${passed ? "contains" : "does not contain"} ${JSON.stringify(config.text)}`,
99
+ detail: { matches: hit.length },
100
+ };
101
+ },
102
+ };
103
+ export const ToolResultJsonPointerEqualsConfigSchema = z
104
+ .object({
105
+ type: z.literal("tool-result-json-pointer-equals"),
106
+ name: z.string(),
107
+ /**
108
+ * RFC 6901 pointer into the result's `structuredContent`, falling back to
109
+ * the result text parsed as JSON. Defaults to "" (the whole value).
110
+ */
111
+ jsonPointer: z.string().default(""),
112
+ /** Structural (key-order-insensitive) equality against the value. */
113
+ value: z.unknown().optional(),
114
+ /** Substring containment against the stringified value. */
115
+ contains: z.string().min(1).optional(),
116
+ occurrence: OccurrenceSchema.default("last"),
117
+ ...weightThresholdFields,
118
+ })
119
+ .strict()
120
+ .refine((config) => (config.value === undefined) !== (config.contains === undefined), {
121
+ message: "set exactly one of value or contains",
122
+ });
123
+ export const toolResultJsonPointerEqualsAssertion = {
124
+ type: "tool-result-json-pointer-equals",
125
+ configSchema: ToolResultJsonPointerEqualsConfigSchema,
126
+ evaluate: async (config, run) => {
127
+ const malformed = malformedResultEvidence(run);
128
+ if (malformed)
129
+ return malformed;
130
+ const results = toolResults(run.events, config.name);
131
+ if (results.length === 0) {
132
+ return {
133
+ passed: false,
134
+ score: 0,
135
+ comment: `no result captured for '${config.name}'`,
136
+ };
137
+ }
138
+ const relevant = config.occurrence === "any" ? results : select(results, config.occurrence);
139
+ const matched = relevant.some((result) => {
140
+ const value = resolvePointer(toolResultValue(result), config.jsonPointer);
141
+ if (config.value !== undefined)
142
+ return deepEquals(value, config.value);
143
+ if (config.contains !== undefined) {
144
+ if (value === undefined)
145
+ return false;
146
+ const stringified = typeof value === "string" ? value : JSON.stringify(value);
147
+ return stringified.includes(config.contains);
148
+ }
149
+ return value !== undefined;
150
+ });
151
+ return {
152
+ passed: matched,
153
+ score: matched ? 1 : 0,
154
+ comment: `'${config.name}' result at '${config.jsonPointer || "/"}' ${matched ? "matched" : "did not match"}`,
155
+ };
156
+ },
157
+ };
158
+ export const NoErrorResultsConfigSchema = z
159
+ .object({
160
+ type: z.literal("no-error-results"),
161
+ /** Permit a run with no captured calls/results. Defaults false to prevent vacuous passes. */
162
+ allowEmpty: z.boolean().default(false),
163
+ ...weightThresholdFields,
164
+ })
165
+ .strict();
166
+ export const noErrorResultsAssertion = {
167
+ type: "no-error-results",
168
+ configSchema: NoErrorResultsConfigSchema,
169
+ evaluate: async (config, run) => {
170
+ const malformed = malformedResultEvidence(run);
171
+ if (malformed)
172
+ return malformed;
173
+ const results = toolResults(run.events);
174
+ if (results.length === 0 && !config.allowEmpty) {
175
+ return {
176
+ passed: false,
177
+ score: 0,
178
+ comment: "no tool results captured",
179
+ };
180
+ }
181
+ const offenders = results.filter((result) => result.isError);
182
+ const passed = offenders.length === 0;
183
+ return {
184
+ passed,
185
+ score: passed ? 1 : 0,
186
+ comment: passed
187
+ ? "no tool returned an error result"
188
+ : `error results from: ${offenders.map((result) => result.name).join(", ")}`,
189
+ detail: { offenders: offenders.map((result) => result.name) },
190
+ };
191
+ },
192
+ };
193
+ //# sourceMappingURL=tool-result-checks.js.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"tool-result-checks.js","sourceRoot":"","sources":["../../../src/assertions/core/tool-result-checks.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AACxB,OAAO,EAEL,uBAAuB,EAEvB,WAAW,EACX,eAAe,GAChB,MAAM,qBAAqB,CAAC;AAC7B,OAAO,EAAE,UAAU,EAAE,cAAc,EAAE,qBAAqB,EAAE,MAAM,cAAc,CAAC;AAEjF,MAAM,gBAAgB,GAAG,CAAC,CAAC,IAAI,CAAC,CAAC,OAAO,EAAE,MAAM,EAAE,KAAK,CAAC,CAAC,CAAC;AAG1D,MAAM,MAAM,GAAG,CAAC,OAA4B,EAAE,UAAsB,EAAuB,EAAE;IAC3F,IAAI,UAAU,KAAK,OAAO;QAAE,OAAO,OAAO,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;IACvD,IAAI,UAAU,KAAK,MAAM;QAAE,OAAO,OAAO,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC;IACpD,OAAO,OAAO,CAAC;AACjB,CAAC,CAAC;AAEF,MAAM,uBAAuB,GAAG,CAAC,GAA6C,EAAE,EAAE;IAChF,MAAM,SAAS,GAAG,uBAAuB,CAAC,GAAG,CAAC,MAAM,EAAE,CAAC,aAAa,CAAC,CAAC,CAAC;IACvE,IAAI,SAAS,KAAK,CAAC;QAAE,OAAO,IAAI,CAAC;IACjC,MAAM,YAAY,GAAG,GAAG,SAAS,+BAA+B,SAAS,KAAK,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,GAAG,WAAW,CAAC;IACtG,OAAO;QACL,MAAM,EAAE,KAAc;QACtB,KAAK,EAAE,CAAU;QACjB,OAAO,EAAE,YAAY;QACrB,YAAY;KACb,CAAC;AACJ,CAAC,CAAC;AAEF,MAAM,CAAC,MAAM,6BAA6B,GAAG,CAAC;KAC3C,MAAM,CAAC;IACN,IAAI,EAAE,CAAC,CAAC,OAAO,CAAC,sBAAsB,CAAC;IACvC,mEAAmE;IACnE,IAAI,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,QAAQ,EAAE;IAC3B,gFAAgF;IAChF,QAAQ,EAAE,CAAC,CAAC,OAAO,EAAE,CAAC,OAAO,CAAC,IAAI,CAAC;IACnC,iEAAiE;IACjE,UAAU,EAAE,gBAAgB,CAAC,OAAO,CAAC,MAAM,CAAC;IAC5C,GAAG,qBAAqB;CACzB,CAAC;KACD,MAAM,EAAE,CAAC;AAGZ,MAAM,CAAC,MAAM,0BAA0B,GAA2C;IAChF,IAAI,EAAE,sBAAsB;IAC5B,YAAY,EAAE,6BAA6B;IAC3C,QAAQ,EAAE,KAAK,EAAE,MAAM,EAAE,GAAG,EAAE,EAAE;QAC9B,MAAM,SAAS,GAAG,uBAAuB,CAAC,GAAG,CAAC,CAAC;QAC/C,IAAI,SAAS;YAAE,OAAO,SAAS,CAAC;QAChC,MAAM,OAAO,GAAG,WAAW,CAAC,GAAG,CAAC,MAAM,EAAE,MAAM,CAAC,IAAI,CAAC,CAAC;QACrD,MAAM,KAAK,GAAG,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,IAAI,MAAM,CAAC,IAAI,GAAG,CAAC,CAAC,CAAC,UAAU,CAAC;QAC5D,IAAI,OAAO,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;YACzB,OAAO;gBACL,MAAM,EAAE,KAAK;gBACb,KAAK,EAAE,CAAC;gBACR,OAAO,EAAE,0BAA0B,KAAK,EAAE;aAC3C,CAAC;QACJ,CAAC;QACD,MAAM,QAAQ,GAAG,MAAM,CAAC,UAAU,KAAK,KAAK,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,OAAO,EAAE,MAAM,CAAC,UAAU,CAAC,CAAC;QAC5F,MAAM,MAAM,GACV,MAAM,CAAC,UAAU,KAAK,KAAK;YACzB,CAAC,CAAC,QAAQ,CAAC,IAAI,CAAC,CAAC,MAAM,EAAE,EAAE,CAAC,MAAM,CAAC,OAAO,KAAK,MAAM,CAAC,QAAQ,CAAC;YAC/D,CAAC,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC,MAAM,EAAE,EAAE,CAAC,MAAM,CAAC,OAAO,KAAK,MAAM,CAAC,QAAQ,CAAC,CAAC;QACrE,OAAO;YACL,MAAM;YACN,KAAK,EAAE,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;YACrB,OAAO,EAAE,GAAG,KAAK,qBAAqB,MAAM,CAAC,QAAQ,KAAK,MAAM,CAAC,UAAU,GAAG;YAC9E,MAAM,EAAE,EAAE,OAAO,EAAE,QAAQ,CAAC,GAAG,CAAC,CAAC,MAAM,EAAE,EAAE,CAAC,MAAM,CAAC,OAAO,CAAC,EAAE;SAC9D,CAAC;IACJ,CAAC;CACF,CAAC;AAEF,MAAM,CAAC,MAAM,8BAA8B,GAAG,CAAC;KAC5C,MAAM,CAAC;IACN,IAAI,EAAE,CAAC,CAAC,OAAO,CAAC,sBAAsB,CAAC;IACvC,IAAI,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,QAAQ,EAAE;IAC3B,yEAAyE;IACzE,IAAI,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,GAAG,CAAC,CAAC,CAAC;IACvB,UAAU,EAAE,gBAAgB,CAAC,OAAO,CAAC,KAAK,CAAC;IAC3C,GAAG,qBAAqB;CACzB,CAAC;KACD,MAAM,EAAE,CAAC;AAGZ,MAAM,UAAU,GAAG,CAAC,MAAyB,EAAU,EAAE,CACvD,MAAM,CAAC,IAAI;IACX,CAAC,OAAO,MAAM,CAAC,OAAO,KAAK,QAAQ,CAAC,CAAC,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC,CAAC,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC,OAAO,IAAI,EAAE,CAAC,CAAC,CAAC;AAE/F,MAAM,CAAC,MAAM,2BAA2B,GAA4C;IAClF,IAAI,EAAE,sBAAsB;IAC5B,YAAY,EAAE,8BAA8B;IAC5C,QAAQ,EAAE,KAAK,EAAE,MAAM,EAAE,GAAG,EAAE,EAAE;QAC9B,MAAM,SAAS,GAAG,uBAAuB,CAAC,GAAG,CAAC,CAAC;QAC/C,IAAI,SAAS;YAAE,OAAO,SAAS,CAAC;QAChC,MAAM,OAAO,GAAG,WAAW,CAAC,GAAG,CAAC,MAAM,EAAE,MAAM,CAAC,IAAI,CAAC,CAAC;QACrD,MAAM,KAAK,GAAG,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,IAAI,MAAM,CAAC,IAAI,GAAG,CAAC,CAAC,CAAC,UAAU,CAAC;QAC5D,IAAI,OAAO,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;YACzB,OAAO;gBACL,MAAM,EAAE,KAAK;gBACb,KAAK,EAAE,CAAC;gBACR,OAAO,EAAE,0BAA0B,KAAK,EAAE;aAC3C,CAAC;QACJ,CAAC;QACD,MAAM,QAAQ,GAAG,MAAM,CAAC,UAAU,KAAK,KAAK,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,OAAO,EAAE,MAAM,CAAC,UAAU,CAAC,CAAC;QAC5F,MAAM,GAAG,GAAG,QAAQ,CAAC,MAAM,CAAC,CAAC,MAAM,EAAE,EAAE,CAAC,UAAU,CAAC,MAAM,CAAC,CAAC,QAAQ,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC;QAClF,MAAM,MAAM,GAAG,GAAG,CAAC,MAAM,GAAG,CAAC,CAAC;QAC9B,OAAO;YACL,MAAM;YACN,KAAK,EAAE,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;YACrB,OAAO,EAAE,GAAG,KAAK,WAAW,MAAM,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC,CAAC,kBAAkB,IAAI,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC,IAAI,CAAC,EAAE;YACrG,MAAM,EAAE,EAAE,OAAO,EAAE,GAAG,CAAC,MAAM,EAAE;SAChC,CAAC;IACJ,CAAC;CACF,CAAC;AAEF,MAAM,CAAC,MAAM,uCAAuC,GAAG,CAAC;KACrD,MAAM,CAAC;IACN,IAAI,EAAE,CAAC,CAAC,OAAO,CAAC,iCAAiC,CAAC;IAClD,IAAI,EAAE,CAAC,CAAC,MAAM,EAAE;IAChB;;;OAGG;IACH,WAAW,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,OAAO,CAAC,EAAE,CAAC;IACnC,qEAAqE;IACrE,KAAK,EAAE,CAAC,CAAC,OAAO,EAAE,CAAC,QAAQ,EAAE;IAC7B,2DAA2D;IAC3D,QAAQ,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,QAAQ,EAAE;IACtC,UAAU,EAAE,gBAAgB,CAAC,OAAO,CAAC,MAAM,CAAC;IAC5C,GAAG,qBAAqB;CACzB,CAAC;KACD,MAAM,EAAE;KACR,MAAM,CAAC,CAAC,MAAM,EAAE,EAAE,CAAC,CAAC,MAAM,CAAC,KAAK,KAAK,SAAS,CAAC,KAAK,CAAC,MAAM,CAAC,QAAQ,KAAK,SAAS,CAAC,EAAE;IACpF,OAAO,EAAE,sCAAsC;CAChD,CAAC,CAAC;AAKL,MAAM,CAAC,MAAM,oCAAoC,GAC/C;IACE,IAAI,EAAE,iCAAiC;IACvC,YAAY,EAAE,uCAAuC;IACrD,QAAQ,EAAE,KAAK,EAAE,MAAM,EAAE,GAAG,EAAE,EAAE;QAC9B,MAAM,SAAS,GAAG,uBAAuB,CAAC,GAAG,CAAC,CAAC;QAC/C,IAAI,SAAS;YAAE,OAAO,SAAS,CAAC;QAChC,MAAM,OAAO,GAAG,WAAW,CAAC,GAAG,CAAC,MAAM,EAAE,MAAM,CAAC,IAAI,CAAC,CAAC;QACrD,IAAI,OAAO,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;YACzB,OAAO;gBACL,MAAM,EAAE,KAAK;gBACb,KAAK,EAAE,CAAC;gBACR,OAAO,EAAE,2BAA2B,MAAM,CAAC,IAAI,GAAG;aACnD,CAAC;QACJ,CAAC;QACD,MAAM,QAAQ,GAAG,MAAM,CAAC,UAAU,KAAK,KAAK,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,OAAO,EAAE,MAAM,CAAC,UAAU,CAAC,CAAC;QAC5F,MAAM,OAAO,GAAG,QAAQ,CAAC,IAAI,CAAC,CAAC,MAAM,EAAE,EAAE;YACvC,MAAM,KAAK,GAAG,cAAc,CAAC,eAAe,CAAC,MAAM,CAAC,EAAE,MAAM,CAAC,WAAW,CAAC,CAAC;YAC1E,IAAI,MAAM,CAAC,KAAK,KAAK,SAAS;gBAAE,OAAO,UAAU,CAAC,KAAK,EAAE,MAAM,CAAC,KAAK,CAAC,CAAC;YACvE,IAAI,MAAM,CAAC,QAAQ,KAAK,SAAS,EAAE,CAAC;gBAClC,IAAI,KAAK,KAAK,SAAS;oBAAE,OAAO,KAAK,CAAC;gBACtC,MAAM,WAAW,GAAG,OAAO,KAAK,KAAK,QAAQ,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,IAAI,CAAC,SAAS,CAAC,KAAK,CAAC,CAAC;gBAC9E,OAAO,WAAW,CAAC,QAAQ,CAAC,MAAM,CAAC,QAAQ,CAAC,CAAC;YAC/C,CAAC;YACD,OAAO,KAAK,KAAK,SAAS,CAAC;QAC7B,CAAC,CAAC,CAAC;QACH,OAAO;YACL,MAAM,EAAE,OAAO;YACf,KAAK,EAAE,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;YACtB,OAAO,EAAE,IAAI,MAAM,CAAC,IAAI,gBAAgB,MAAM,CAAC,WAAW,IAAI,GAAG,KAAK,OAAO,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,eAAe,EAAE;SAC9G,CAAC;IACJ,CAAC;CACF,CAAC;AAEJ,MAAM,CAAC,MAAM,0BAA0B,GAAG,CAAC;KACxC,MAAM,CAAC;IACN,IAAI,EAAE,CAAC,CAAC,OAAO,CAAC,kBAAkB,CAAC;IACnC,6FAA6F;IAC7F,UAAU,EAAE,CAAC,CAAC,OAAO,EAAE,CAAC,OAAO,CAAC,KAAK,CAAC;IACtC,GAAG,qBAAqB;CACzB,CAAC;KACD,MAAM,EAAE,CAAC;AAGZ,MAAM,CAAC,MAAM,uBAAuB,GAAwC;IAC1E,IAAI,EAAE,kBAAkB;IACxB,YAAY,EAAE,0BAA0B;IACxC,QAAQ,EAAE,KAAK,EAAE,MAAM,EAAE,GAAG,EAAE,EAAE;QAC9B,MAAM,SAAS,GAAG,uBAAuB,CAAC,GAAG,CAAC,CAAC;QAC/C,IAAI,SAAS;YAAE,OAAO,SAAS,CAAC;QAChC,MAAM,OAAO,GAAG,WAAW,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC;QACxC,IAAI,OAAO,CAAC,MAAM,KAAK,CAAC,IAAI,CAAC,MAAM,CAAC,UAAU,EAAE,CAAC;YAC/C,OAAO;gBACL,MAAM,EAAE,KAAK;gBACb,KAAK,EAAE,CAAC;gBACR,OAAO,EAAE,0BAA0B;aACpC,CAAC;QACJ,CAAC;QACD,MAAM,SAAS,GAAG,OAAO,CAAC,MAAM,CAAC,CAAC,MAAM,EAAE,EAAE,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC;QAC7D,MAAM,MAAM,GAAG,SAAS,CAAC,MAAM,KAAK,CAAC,CAAC;QACtC,OAAO;YACL,MAAM;YACN,KAAK,EAAE,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;YACrB,OAAO,EAAE,MAAM;gBACb,CAAC,CAAC,kCAAkC;gBACpC,CAAC,CAAC,uBAAuB,SAAS,CAAC,GAAG,CAAC,CAAC,MAAM,EAAE,EAAE,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE;YAC9E,MAAM,EAAE,EAAE,SAAS,EAAE,SAAS,CAAC,GAAG,CAAC,CAAC,MAAM,EAAE,EAAE,CAAC,MAAM,CAAC,IAAI,CAAC,EAAE;SAC9D,CAAC;IACJ,CAAC;CACF,CAAC"}
@@ -0,0 +1,36 @@
1
+ import { z } from "zod";
2
+ import { type AssertionImpl } from "../../core/index.js";
3
+ export declare const ServerInfoEqualsConfigSchema: z.ZodObject<{
4
+ type: z.ZodLiteral<"server-info-equals">;
5
+ name: z.ZodOptional<z.ZodString>;
6
+ version: z.ZodOptional<z.ZodString>;
7
+ }, z.core.$strict>;
8
+ export type ServerInfoEqualsConfig = z.infer<typeof ServerInfoEqualsConfigSchema>;
9
+ export declare const serverInfoEqualsAssertion: AssertionImpl<ServerInfoEqualsConfig>;
10
+ export declare const ToolListedConfigSchema: z.ZodObject<{
11
+ type: z.ZodLiteral<"tool-listed">;
12
+ name: z.ZodString;
13
+ expected: z.ZodDefault<z.ZodBoolean>;
14
+ }, z.core.$strict>;
15
+ export type ToolListedConfig = z.infer<typeof ToolListedConfigSchema>;
16
+ export declare const toolListedAssertion: AssertionImpl<ToolListedConfig>;
17
+ export declare const ToolsListedCountConfigSchema: z.ZodObject<{
18
+ type: z.ZodLiteral<"tools-listed-count">;
19
+ min: z.ZodOptional<z.ZodNumber>;
20
+ max: z.ZodOptional<z.ZodNumber>;
21
+ }, z.core.$strict>;
22
+ export type ToolsListedCountConfig = z.infer<typeof ToolsListedCountConfigSchema>;
23
+ export declare const toolsListedCountAssertion: AssertionImpl<ToolsListedCountConfig>;
24
+ export declare const CapabilityAbsentConfigSchema: z.ZodObject<{
25
+ type: z.ZodLiteral<"capability-absent">;
26
+ capability: z.ZodEnum<{
27
+ completions: "completions";
28
+ logging: "logging";
29
+ prompts: "prompts";
30
+ resources: "resources";
31
+ tools: "tools";
32
+ }>;
33
+ }, z.core.$strict>;
34
+ export type CapabilityAbsentConfig = z.infer<typeof CapabilityAbsentConfigSchema>;
35
+ export declare const capabilityAbsentAssertion: AssertionImpl<CapabilityAbsentConfig>;
36
+ //# sourceMappingURL=discovery-checks.d.ts.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"discovery-checks.d.ts","sourceRoot":"","sources":["../../../src/assertions/mcp/discovery-checks.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AACxB,OAAO,EACL,KAAK,aAAa,EAKnB,MAAM,qBAAqB,CAAC;AAG7B,eAAO,MAAM,4BAA4B;;;;kBAWrC,CAAC;AACL,MAAM,MAAM,sBAAsB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,4BAA4B,CAAC,CAAC;AAElF,eAAO,MAAM,yBAAyB,EAAE,aAAa,CAAC,sBAAsB,CAuB3E,CAAC;AAEF,eAAO,MAAM,sBAAsB;;;;kBAQxB,CAAC;AACZ,MAAM,MAAM,gBAAgB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,sBAAsB,CAAC,CAAC;AAEtE,eAAO,MAAM,mBAAmB,EAAE,aAAa,CAAC,gBAAgB,CAgB/D,CAAC;AAEF,eAAO,MAAM,4BAA4B;;;;kBAYrC,CAAC;AACL,MAAM,MAAM,sBAAsB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,4BAA4B,CAAC,CAAC;AAElF,eAAO,MAAM,yBAAyB,EAAE,aAAa,CAAC,sBAAsB,CAgB3E,CAAC;AAEF,eAAO,MAAM,4BAA4B;;;;;;;;;kBAM9B,CAAC;AACZ,MAAM,MAAM,sBAAsB,GAAG,CAAC,CAAC,KAAK,CAAC,OAAO,4BAA4B,CAAC,CAAC;AAElF,eAAO,MAAM,yBAAyB,EAAE,aAAa,CAAC,sBAAsB,CAkB3E,CAAC"}
@@ -0,0 +1,122 @@
1
+ import { z } from "zod";
2
+ import { hasMcpInit, hasMcpToolsList, listedTools, mcpInit, } from "../../core/index.js";
3
+ import { inRange } from "../core/helpers.js";
4
+ export const ServerInfoEqualsConfigSchema = z
5
+ .object({
6
+ type: z.literal("server-info-equals"),
7
+ /** Expected `serverInfo.name` from `initialize`. */
8
+ name: z.string().optional(),
9
+ /** Expected `serverInfo.version` from `initialize`. */
10
+ version: z.string().optional(),
11
+ })
12
+ .strict()
13
+ .refine((c) => c.name !== undefined || c.version !== undefined, {
14
+ message: "set at least one of name or version",
15
+ });
16
+ export const serverInfoEqualsAssertion = {
17
+ type: "server-info-equals",
18
+ configSchema: ServerInfoEqualsConfigSchema,
19
+ evaluate: async (config, run) => {
20
+ const info = mcpInit(run.events)?.serverInfo;
21
+ if (!info) {
22
+ return { passed: false, score: 0, comment: "no initialize metadata captured" };
23
+ }
24
+ const mismatches = [];
25
+ if (config.name !== undefined && info.name !== config.name) {
26
+ mismatches.push(`name '${info.name}' != '${config.name}'`);
27
+ }
28
+ if (config.version !== undefined && info.version !== config.version) {
29
+ mismatches.push(`version '${info.version}' != '${config.version}'`);
30
+ }
31
+ const passed = mismatches.length === 0;
32
+ return {
33
+ passed,
34
+ score: passed ? 1 : 0,
35
+ comment: passed ? `serverInfo ${info.name}@${info.version} matched` : mismatches.join("; "),
36
+ detail: { serverInfo: info },
37
+ };
38
+ },
39
+ };
40
+ export const ToolListedConfigSchema = z
41
+ .object({
42
+ type: z.literal("tool-listed"),
43
+ /** Tool name whose presence is checked. */
44
+ name: z.string(),
45
+ /** Whether `tools/list` should include the tool. Default true. */
46
+ expected: z.boolean().default(true),
47
+ })
48
+ .strict();
49
+ export const toolListedAssertion = {
50
+ type: "tool-listed",
51
+ configSchema: ToolListedConfigSchema,
52
+ evaluate: async (config, run) => {
53
+ if (!hasMcpToolsList(run.events)) {
54
+ return { passed: false, score: 0, comment: "no tools/list result captured" };
55
+ }
56
+ const tools = listedTools(run.events);
57
+ const listed = tools.some((t) => t.name === config.name);
58
+ const passed = listed === config.expected;
59
+ return {
60
+ passed,
61
+ score: passed ? 1 : 0,
62
+ comment: `tool '${config.name}' is ${listed ? "listed" : `not among ${tools.length} listed tools`}`,
63
+ };
64
+ },
65
+ };
66
+ export const ToolsListedCountConfigSchema = z
67
+ .object({
68
+ type: z.literal("tools-listed-count"),
69
+ min: z.number().int().nonnegative().optional(),
70
+ max: z.number().int().nonnegative().optional(),
71
+ })
72
+ .strict()
73
+ .refine((c) => c.min !== undefined || c.max !== undefined, {
74
+ message: "set at least one of min or max",
75
+ })
76
+ .refine((c) => c.min === undefined || c.max === undefined || c.min <= c.max, {
77
+ message: "min must be <= max",
78
+ });
79
+ export const toolsListedCountAssertion = {
80
+ type: "tools-listed-count",
81
+ configSchema: ToolsListedCountConfigSchema,
82
+ evaluate: async (config, run) => {
83
+ if (!hasMcpToolsList(run.events)) {
84
+ return { passed: false, score: 0, comment: "no tools/list result captured" };
85
+ }
86
+ const count = listedTools(run.events).length;
87
+ const { passed, comment } = inRange(count, config.min, config.max);
88
+ return {
89
+ passed,
90
+ score: passed ? 1 : 0,
91
+ comment: `${count} tool${count === 1 ? "" : "s"} listed (${comment})`,
92
+ detail: { count },
93
+ };
94
+ },
95
+ };
96
+ export const CapabilityAbsentConfigSchema = z
97
+ .object({
98
+ type: z.literal("capability-absent"),
99
+ /** Server capability that must NOT be advertised. */
100
+ capability: z.enum(["resources", "prompts", "logging", "completions", "tools"]),
101
+ })
102
+ .strict();
103
+ export const capabilityAbsentAssertion = {
104
+ type: "capability-absent",
105
+ configSchema: CapabilityAbsentConfigSchema,
106
+ evaluate: async (config, run) => {
107
+ if (!hasMcpInit(run.events)) {
108
+ return { passed: false, score: 0, comment: "no initialize metadata captured" };
109
+ }
110
+ const caps = mcpInit(run.events)?.capabilities ?? {};
111
+ const value = caps[config.capability];
112
+ const present = value !== undefined && value !== false && value !== null;
113
+ return {
114
+ passed: !present,
115
+ score: present ? 0 : 1,
116
+ comment: present
117
+ ? `server advertises '${config.capability}'`
118
+ : `server does not advertise '${config.capability}'`,
119
+ };
120
+ },
121
+ };
122
+ //# sourceMappingURL=discovery-checks.js.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"discovery-checks.js","sourceRoot":"","sources":["../../../src/assertions/mcp/discovery-checks.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AACxB,OAAO,EAEL,UAAU,EACV,eAAe,EACf,WAAW,EACX,OAAO,GACR,MAAM,qBAAqB,CAAC;AAC7B,OAAO,EAAE,OAAO,EAAE,MAAM,oBAAoB,CAAC;AAE7C,MAAM,CAAC,MAAM,4BAA4B,GAAG,CAAC;KAC1C,MAAM,CAAC;IACN,IAAI,EAAE,CAAC,CAAC,OAAO,CAAC,oBAAoB,CAAC;IACrC,oDAAoD;IACpD,IAAI,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,QAAQ,EAAE;IAC3B,uDAAuD;IACvD,OAAO,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,QAAQ,EAAE;CAC/B,CAAC;KACD,MAAM,EAAE;KACR,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,SAAS,IAAI,CAAC,CAAC,OAAO,KAAK,SAAS,EAAE;IAC9D,OAAO,EAAE,qCAAqC;CAC/C,CAAC,CAAC;AAGL,MAAM,CAAC,MAAM,yBAAyB,GAA0C;IAC9E,IAAI,EAAE,oBAAoB;IAC1B,YAAY,EAAE,4BAA4B;IAC1C,QAAQ,EAAE,KAAK,EAAE,MAAM,EAAE,GAAG,EAAE,EAAE;QAC9B,MAAM,IAAI,GAAG,OAAO,CAAC,GAAG,CAAC,MAAM,CAAC,EAAE,UAAU,CAAC;QAC7C,IAAI,CAAC,IAAI,EAAE,CAAC;YACV,OAAO,EAAE,MAAM,EAAE,KAAK,EAAE,KAAK,EAAE,CAAC,EAAE,OAAO,EAAE,iCAAiC,EAAE,CAAC;QACjF,CAAC;QACD,MAAM,UAAU,GAAa,EAAE,CAAC;QAChC,IAAI,MAAM,CAAC,IAAI,KAAK,SAAS,IAAI,IAAI,CAAC,IAAI,KAAK,MAAM,CAAC,IAAI,EAAE,CAAC;YAC3D,UAAU,CAAC,IAAI,CAAC,SAAS,IAAI,CAAC,IAAI,SAAS,MAAM,CAAC,IAAI,GAAG,CAAC,CAAC;QAC7D,CAAC;QACD,IAAI,MAAM,CAAC,OAAO,KAAK,SAAS,IAAI,IAAI,CAAC,OAAO,KAAK,MAAM,CAAC,OAAO,EAAE,CAAC;YACpE,UAAU,CAAC,IAAI,CAAC,YAAY,IAAI,CAAC,OAAO,SAAS,MAAM,CAAC,OAAO,GAAG,CAAC,CAAC;QACtE,CAAC;QACD,MAAM,MAAM,GAAG,UAAU,CAAC,MAAM,KAAK,CAAC,CAAC;QACvC,OAAO;YACL,MAAM;YACN,KAAK,EAAE,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;YACrB,OAAO,EAAE,MAAM,CAAC,CAAC,CAAC,cAAc,IAAI,CAAC,IAAI,IAAI,IAAI,CAAC,OAAO,UAAU,CAAC,CAAC,CAAC,UAAU,CAAC,IAAI,CAAC,IAAI,CAAC;YAC3F,MAAM,EAAE,EAAE,UAAU,EAAE,IAAI,EAAE;SAC7B,CAAC;IACJ,CAAC;CACF,CAAC;AAEF,MAAM,CAAC,MAAM,sBAAsB,GAAG,CAAC;KACpC,MAAM,CAAC;IACN,IAAI,EAAE,CAAC,CAAC,OAAO,CAAC,aAAa,CAAC;IAC9B,2CAA2C;IAC3C,IAAI,EAAE,CAAC,CAAC,MAAM,EAAE;IAChB,kEAAkE;IAClE,QAAQ,EAAE,CAAC,CAAC,OAAO,EAAE,CAAC,OAAO,CAAC,IAAI,CAAC;CACpC,CAAC;KACD,MAAM,EAAE,CAAC;AAGZ,MAAM,CAAC,MAAM,mBAAmB,GAAoC;IAClE,IAAI,EAAE,aAAa;IACnB,YAAY,EAAE,sBAAsB;IACpC,QAAQ,EAAE,KAAK,EAAE,MAAM,EAAE,GAAG,EAAE,EAAE;QAC9B,IAAI,CAAC,eAAe,CAAC,GAAG,CAAC,MAAM,CAAC,EAAE,CAAC;YACjC,OAAO,EAAE,MAAM,EAAE,KAAK,EAAE,KAAK,EAAE,CAAC,EAAE,OAAO,EAAE,+BAA+B,EAAE,CAAC;QAC/E,CAAC;QACD,MAAM,KAAK,GAAG,WAAW,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC;QACtC,MAAM,MAAM,GAAG,KAAK,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,MAAM,CAAC,IAAI,CAAC,CAAC;QACzD,MAAM,MAAM,GAAG,MAAM,KAAK,MAAM,CAAC,QAAQ,CAAC;QAC1C,OAAO;YACL,MAAM;YACN,KAAK,EAAE,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;YACrB,OAAO,EAAE,SAAS,MAAM,CAAC,IAAI,QAAQ,MAAM,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,aAAa,KAAK,CAAC,MAAM,eAAe,EAAE;SACpG,CAAC;IACJ,CAAC;CACF,CAAC;AAEF,MAAM,CAAC,MAAM,4BAA4B,GAAG,CAAC;KAC1C,MAAM,CAAC;IACN,IAAI,EAAE,CAAC,CAAC,OAAO,CAAC,oBAAoB,CAAC;IACrC,GAAG,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,GAAG,EAAE,CAAC,WAAW,EAAE,CAAC,QAAQ,EAAE;IAC9C,GAAG,EAAE,CAAC,CAAC,MAAM,EAAE,CAAC,GAAG,EAAE,CAAC,WAAW,EAAE,CAAC,QAAQ,EAAE;CAC/C,CAAC;KACD,MAAM,EAAE;KACR,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,GAAG,KAAK,SAAS,IAAI,CAAC,CAAC,GAAG,KAAK,SAAS,EAAE;IACzD,OAAO,EAAE,gCAAgC;CAC1C,CAAC;KACD,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,GAAG,KAAK,SAAS,IAAI,CAAC,CAAC,GAAG,KAAK,SAAS,IAAI,CAAC,CAAC,GAAG,IAAI,CAAC,CAAC,GAAG,EAAE;IAC3E,OAAO,EAAE,oBAAoB;CAC9B,CAAC,CAAC;AAGL,MAAM,CAAC,MAAM,yBAAyB,GAA0C;IAC9E,IAAI,EAAE,oBAAoB;IAC1B,YAAY,EAAE,4BAA4B;IAC1C,QAAQ,EAAE,KAAK,EAAE,MAAM,EAAE,GAAG,EAAE,EAAE;QAC9B,IAAI,CAAC,eAAe,CAAC,GAAG,CAAC,MAAM,CAAC,EAAE,CAAC;YACjC,OAAO,EAAE,MAAM,EAAE,KAAK,EAAE,KAAK,EAAE,CAAC,EAAE,OAAO,EAAE,+BAA+B,EAAE,CAAC;QAC/E,CAAC;QACD,MAAM,KAAK,GAAG,WAAW,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC,MAAM,CAAC;QAC7C,MAAM,EAAE,MAAM,EAAE,OAAO,EAAE,GAAG,OAAO,CAAC,KAAK,EAAE,MAAM,CAAC,GAAG,EAAE,MAAM,CAAC,GAAG,CAAC,CAAC;QACnE,OAAO;YACL,MAAM;YACN,KAAK,EAAE,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;YACrB,OAAO,EAAE,GAAG,KAAK,QAAQ,KAAK,KAAK,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,GAAG,YAAY,OAAO,GAAG;YACrE,MAAM,EAAE,EAAE,KAAK,EAAE;SAClB,CAAC;IACJ,CAAC;CACF,CAAC;AAEF,MAAM,CAAC,MAAM,4BAA4B,GAAG,CAAC;KAC1C,MAAM,CAAC;IACN,IAAI,EAAE,CAAC,CAAC,OAAO,CAAC,mBAAmB,CAAC;IACpC,qDAAqD;IACrD,UAAU,EAAE,CAAC,CAAC,IAAI,CAAC,CAAC,WAAW,EAAE,SAAS,EAAE,SAAS,EAAE,aAAa,EAAE,OAAO,CAAC,CAAC;CAChF,CAAC;KACD,MAAM,EAAE,CAAC;AAGZ,MAAM,CAAC,MAAM,yBAAyB,GAA0C;IAC9E,IAAI,EAAE,mBAAmB;IACzB,YAAY,EAAE,4BAA4B;IAC1C,QAAQ,EAAE,KAAK,EAAE,MAAM,EAAE,GAAG,EAAE,EAAE;QAC9B,IAAI,CAAC,UAAU,CAAC,GAAG,CAAC,MAAM,CAAC,EAAE,CAAC;YAC5B,OAAO,EAAE,MAAM,EAAE,KAAK,EAAE,KAAK,EAAE,CAAC,EAAE,OAAO,EAAE,iCAAiC,EAAE,CAAC;QACjF,CAAC;QACD,MAAM,IAAI,GAAG,OAAO,CAAC,GAAG,CAAC,MAAM,CAAC,EAAE,YAAY,IAAI,EAAE,CAAC;QACrD,MAAM,KAAK,GAAI,IAAgC,CAAC,MAAM,CAAC,UAAU,CAAC,CAAC;QACnE,MAAM,OAAO,GAAG,KAAK,KAAK,SAAS,IAAI,KAAK,KAAK,KAAK,IAAI,KAAK,KAAK,IAAI,CAAC;QACzE,OAAO;YACL,MAAM,EAAE,CAAC,OAAO;YAChB,KAAK,EAAE,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;YACtB,OAAO,EAAE,OAAO;gBACd,CAAC,CAAC,sBAAsB,MAAM,CAAC,UAAU,GAAG;gBAC5C,CAAC,CAAC,8BAA8B,MAAM,CAAC,UAAU,GAAG;SACvD,CAAC;IACJ,CAAC;CACF,CAAC"}
@@ -0,0 +1,10 @@
1
+ import type { AssertionImpl, Registry } from "../../core/index.js";
2
+ export * from "./discovery-checks.js";
3
+ /**
4
+ * The result assertions remain in this compatibility collection for
5
+ * embedders that historically registered only the MCP assertion bundle.
6
+ * Their implementations are generic and also ship in `allCoreAssertions`.
7
+ */
8
+ export declare const allMcpAssertions: AssertionImpl[];
9
+ export declare const registerMcpAssertions: (registry: Registry) => void;
10
+ //# sourceMappingURL=index.d.ts.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/assertions/mcp/index.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,aAAa,EAAE,QAAQ,EAAE,MAAM,qBAAqB,CAAC;AAcnE,cAAc,uBAAuB,CAAC;AAEtC;;;;GAIG;AACH,eAAO,MAAM,gBAAgB,EAAE,aAAa,EAS3C,CAAC;AAEF,eAAO,MAAM,qBAAqB,aAAc,QAAQ,KAAG,IAI1D,CAAC"}