@xoxoai/checkmate 0.4.28 → 0.5.1

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 (397) hide show
  1. package/README.md +259 -224
  2. package/bin/checkmate.js +1 -1
  3. package/dist/ai/client.d.ts +45 -30
  4. package/dist/ai/client.d.ts.map +1 -1
  5. package/dist/ai/client.js +198 -129
  6. package/dist/ai/client.js.map +1 -1
  7. package/dist/ai/message-handler.d.ts +5 -7
  8. package/dist/ai/message-handler.d.ts.map +1 -1
  9. package/dist/ai/message-handler.js +38 -21
  10. package/dist/ai/message-handler.js.map +1 -1
  11. package/dist/ai/message-history.d.ts +4 -6
  12. package/dist/ai/message-history.d.ts.map +1 -1
  13. package/dist/ai/message-history.js +18 -21
  14. package/dist/ai/message-history.js.map +1 -1
  15. package/dist/ai/rate-limit-policy.d.ts +5 -3
  16. package/dist/ai/rate-limit-policy.d.ts.map +1 -1
  17. package/dist/ai/rate-limit-policy.js +24 -7
  18. package/dist/ai/rate-limit-policy.js.map +1 -1
  19. package/dist/ai/tool-response-handler.d.ts +11 -16
  20. package/dist/ai/tool-response-handler.d.ts.map +1 -1
  21. package/dist/ai/tool-response-handler.js +59 -28
  22. package/dist/ai/tool-response-handler.js.map +1 -1
  23. package/dist/ai/turn-processor.d.ts +34 -0
  24. package/dist/ai/turn-processor.d.ts.map +1 -0
  25. package/dist/ai/turn-processor.js +107 -0
  26. package/dist/ai/turn-processor.js.map +1 -0
  27. package/dist/api/allocate-run-identity.d.ts +8 -0
  28. package/dist/api/allocate-run-identity.d.ts.map +1 -0
  29. package/dist/api/allocate-run-identity.js +10 -0
  30. package/dist/api/allocate-run-identity.js.map +1 -0
  31. package/dist/api/execute-prepared-run.d.ts +22 -0
  32. package/dist/api/execute-prepared-run.d.ts.map +1 -0
  33. package/dist/api/execute-prepared-run.js +160 -0
  34. package/dist/api/execute-prepared-run.js.map +1 -0
  35. package/dist/api/index.d.ts +25 -0
  36. package/dist/api/index.d.ts.map +1 -0
  37. package/dist/api/index.js +69 -0
  38. package/dist/api/index.js.map +1 -0
  39. package/dist/api/prepare-run.d.ts +44 -0
  40. package/dist/api/prepare-run.d.ts.map +1 -0
  41. package/dist/api/prepare-run.js +261 -0
  42. package/dist/api/prepare-run.js.map +1 -0
  43. package/dist/cli/diagnostic-relay.d.ts +6 -0
  44. package/dist/cli/diagnostic-relay.d.ts.map +1 -0
  45. package/dist/cli/diagnostic-relay.js +6 -0
  46. package/dist/cli/diagnostic-relay.js.map +1 -0
  47. package/dist/cli/main.d.ts +21 -0
  48. package/dist/cli/main.d.ts.map +1 -0
  49. package/dist/cli/main.js +166 -0
  50. package/dist/cli/main.js.map +1 -0
  51. package/dist/cli/protocol.d.ts +57 -0
  52. package/dist/cli/protocol.d.ts.map +1 -0
  53. package/dist/cli/protocol.js +116 -0
  54. package/dist/cli/protocol.js.map +1 -0
  55. package/dist/cli/reconcile-result.d.ts +13 -0
  56. package/dist/cli/reconcile-result.d.ts.map +1 -0
  57. package/dist/cli/reconcile-result.js +300 -0
  58. package/dist/cli/reconcile-result.js.map +1 -0
  59. package/dist/cli/run-parent.d.ts +37 -0
  60. package/dist/cli/run-parent.d.ts.map +1 -0
  61. package/dist/cli/run-parent.js +491 -0
  62. package/dist/cli/run-parent.js.map +1 -0
  63. package/dist/cli/signals.d.ts +36 -0
  64. package/dist/cli/signals.d.ts.map +1 -0
  65. package/dist/cli/signals.js +85 -0
  66. package/dist/cli/signals.js.map +1 -0
  67. package/dist/cli/static-commands.d.ts +11 -0
  68. package/dist/cli/static-commands.d.ts.map +1 -0
  69. package/dist/cli/static-commands.js +83 -0
  70. package/dist/cli/static-commands.js.map +1 -0
  71. package/dist/cli/worker-stream-guard.d.ts +20 -0
  72. package/dist/cli/worker-stream-guard.d.ts.map +1 -0
  73. package/dist/cli/worker-stream-guard.js +41 -0
  74. package/dist/cli/worker-stream-guard.js.map +1 -0
  75. package/dist/cli/worker.d.ts +3 -0
  76. package/dist/cli/worker.d.ts.map +1 -0
  77. package/dist/cli/worker.js +111 -0
  78. package/dist/cli/worker.js.map +1 -0
  79. package/dist/config/ingestion.d.ts +58 -0
  80. package/dist/config/ingestion.d.ts.map +1 -0
  81. package/dist/config/ingestion.js +283 -0
  82. package/dist/config/ingestion.js.map +1 -0
  83. package/dist/config/manifest.d.ts +20 -0
  84. package/dist/config/manifest.d.ts.map +1 -0
  85. package/dist/config/manifest.js +91 -0
  86. package/dist/config/manifest.js.map +1 -0
  87. package/dist/config/model-egress.d.ts +10 -0
  88. package/dist/config/model-egress.d.ts.map +1 -0
  89. package/dist/config/model-egress.js +62 -0
  90. package/dist/config/model-egress.js.map +1 -0
  91. package/dist/config/package-resolution.d.ts +4 -0
  92. package/dist/config/package-resolution.d.ts.map +1 -0
  93. package/dist/config/package-resolution.js +24 -0
  94. package/dist/config/package-resolution.js.map +1 -0
  95. package/dist/config/policy.d.ts +44 -0
  96. package/dist/config/policy.d.ts.map +1 -0
  97. package/dist/config/policy.js +141 -0
  98. package/dist/config/policy.js.map +1 -0
  99. package/dist/config/record.d.ts +4 -0
  100. package/dist/config/record.d.ts.map +1 -0
  101. package/dist/config/record.js +10 -0
  102. package/dist/config/record.js.map +1 -0
  103. package/dist/config/secrets.d.ts +11 -0
  104. package/dist/config/secrets.d.ts.map +1 -0
  105. package/dist/config/secrets.js +42 -0
  106. package/dist/config/secrets.js.map +1 -0
  107. package/dist/contracts/diagnostics.d.ts +4 -0
  108. package/dist/contracts/diagnostics.d.ts.map +1 -0
  109. package/dist/contracts/diagnostics.js +31 -0
  110. package/dist/contracts/diagnostics.js.map +1 -0
  111. package/dist/contracts/serialize.d.ts +2 -0
  112. package/dist/contracts/serialize.d.ts.map +1 -0
  113. package/dist/contracts/serialize.js +90 -0
  114. package/dist/contracts/serialize.js.map +1 -0
  115. package/dist/contracts/types.d.ts +316 -0
  116. package/dist/contracts/types.d.ts.map +1 -0
  117. package/dist/contracts/types.js +74 -0
  118. package/dist/contracts/types.js.map +1 -0
  119. package/dist/contracts/validator.d.ts +8 -0
  120. package/dist/contracts/validator.d.ts.map +1 -0
  121. package/dist/contracts/validator.js +71 -0
  122. package/dist/contracts/validator.js.map +1 -0
  123. package/dist/driver.d.ts +97 -0
  124. package/dist/driver.d.ts.map +1 -0
  125. package/dist/driver.js +24 -0
  126. package/dist/driver.js.map +1 -0
  127. package/dist/drivers/descriptor.d.ts +15 -0
  128. package/dist/drivers/descriptor.d.ts.map +1 -0
  129. package/dist/drivers/descriptor.js +122 -0
  130. package/dist/drivers/descriptor.js.map +1 -0
  131. package/dist/drivers/loader.d.ts +18 -0
  132. package/dist/drivers/loader.d.ts.map +1 -0
  133. package/dist/drivers/loader.js +97 -0
  134. package/dist/drivers/loader.js.map +1 -0
  135. package/dist/drivers/web/checkmate-driver.json +44 -0
  136. package/dist/drivers/web/index.d.ts +6 -0
  137. package/dist/drivers/web/index.d.ts.map +1 -0
  138. package/dist/drivers/web/index.js +27 -0
  139. package/dist/drivers/web/index.js.map +1 -0
  140. package/dist/drivers/web/session.d.ts +22 -0
  141. package/dist/drivers/web/session.d.ts.map +1 -0
  142. package/dist/drivers/web/session.js +130 -0
  143. package/dist/drivers/web/session.js.map +1 -0
  144. package/dist/drivers/web/tools/network-request-recorder.d.ts +46 -0
  145. package/dist/drivers/web/tools/network-request-recorder.d.ts.map +1 -0
  146. package/dist/drivers/web/tools/network-request-recorder.js +164 -0
  147. package/dist/drivers/web/tools/network-request-recorder.js.map +1 -0
  148. package/dist/{tools/browser → drivers/web/tools}/screenshot-service.d.ts +1 -1
  149. package/dist/drivers/web/tools/screenshot-service.d.ts.map +1 -0
  150. package/dist/drivers/web/tools/screenshot-service.js.map +1 -0
  151. package/dist/drivers/web/tools/snapshot-filter/index.d.ts.map +1 -0
  152. package/dist/drivers/web/tools/snapshot-filter/index.js.map +1 -0
  153. package/dist/drivers/web/tools/snapshot-filter/semantic-scorer.d.ts.map +1 -0
  154. package/dist/drivers/web/tools/snapshot-filter/semantic-scorer.js.map +1 -0
  155. package/dist/drivers/web/tools/snapshot-filter/snapshot-filter.d.ts +5 -0
  156. package/dist/drivers/web/tools/snapshot-filter/snapshot-filter.d.ts.map +1 -0
  157. package/dist/drivers/web/tools/snapshot-filter/snapshot-filter.js +56 -0
  158. package/dist/drivers/web/tools/snapshot-filter/snapshot-filter.js.map +1 -0
  159. package/dist/drivers/web/tools/snapshot-filter/tree-reconstructor.d.ts.map +1 -0
  160. package/dist/drivers/web/tools/snapshot-filter/tree-reconstructor.js.map +1 -0
  161. package/dist/drivers/web/tools/snapshot-service.d.ts +31 -0
  162. package/dist/drivers/web/tools/snapshot-service.d.ts.map +1 -0
  163. package/dist/{tools/browser → drivers/web/tools}/snapshot-service.js +36 -7
  164. package/dist/drivers/web/tools/snapshot-service.js.map +1 -0
  165. package/dist/drivers/web/tools/tool.d.ts +76 -0
  166. package/dist/drivers/web/tools/tool.d.ts.map +1 -0
  167. package/dist/drivers/web/tools/tool.js +562 -0
  168. package/dist/drivers/web/tools/tool.js.map +1 -0
  169. package/dist/drivers/web/tools/transient-state-tracker.d.ts +26 -0
  170. package/dist/drivers/web/tools/transient-state-tracker.d.ts.map +1 -0
  171. package/dist/{tools/browser → drivers/web/tools}/transient-state-tracker.js +54 -19
  172. package/dist/drivers/web/tools/transient-state-tracker.js.map +1 -0
  173. package/dist/drivers/web/tools/types.d.ts +11 -0
  174. package/dist/drivers/web/tools/types.d.ts.map +1 -0
  175. package/dist/drivers/web/tools/types.js +2 -0
  176. package/dist/drivers/web/tools/types.js.map +1 -0
  177. package/dist/evidence/atomic-file.d.ts +23 -0
  178. package/dist/evidence/atomic-file.d.ts.map +1 -0
  179. package/dist/evidence/atomic-file.js +104 -0
  180. package/dist/evidence/atomic-file.js.map +1 -0
  181. package/dist/evidence/layout.d.ts +30 -0
  182. package/dist/evidence/layout.d.ts.map +1 -0
  183. package/dist/evidence/layout.js +152 -0
  184. package/dist/evidence/layout.js.map +1 -0
  185. package/dist/evidence/retention.d.ts +5 -0
  186. package/dist/evidence/retention.d.ts.map +1 -0
  187. package/dist/evidence/retention.js +10 -0
  188. package/dist/evidence/retention.js.map +1 -0
  189. package/dist/evidence/store.d.ts +101 -0
  190. package/dist/evidence/store.d.ts.map +1 -0
  191. package/dist/evidence/store.js +375 -0
  192. package/dist/evidence/store.js.map +1 -0
  193. package/dist/evidence/terminal-finalizer.d.ts +19 -0
  194. package/dist/evidence/terminal-finalizer.d.ts.map +1 -0
  195. package/dist/evidence/terminal-finalizer.js +56 -0
  196. package/dist/evidence/terminal-finalizer.js.map +1 -0
  197. package/dist/index.d.ts +8 -1
  198. package/dist/index.d.ts.map +1 -1
  199. package/dist/index.js +11 -1
  200. package/dist/index.js.map +1 -1
  201. package/dist/logging/invocation-logger.d.ts +4 -0
  202. package/dist/logging/invocation-logger.d.ts.map +1 -0
  203. package/dist/logging/invocation-logger.js +25 -0
  204. package/dist/logging/invocation-logger.js.map +1 -0
  205. package/dist/logging/types.d.ts +8 -0
  206. package/dist/logging/types.d.ts.map +1 -0
  207. package/dist/logging/types.js +7 -0
  208. package/dist/logging/types.js.map +1 -0
  209. package/dist/redaction/diagnostic-sanitizer.d.ts +7 -0
  210. package/dist/redaction/diagnostic-sanitizer.d.ts.map +1 -0
  211. package/dist/redaction/diagnostic-sanitizer.js +42 -0
  212. package/dist/redaction/diagnostic-sanitizer.js.map +1 -0
  213. package/dist/redaction/redactor.d.ts +40 -0
  214. package/dist/redaction/redactor.d.ts.map +1 -0
  215. package/dist/redaction/redactor.js +176 -0
  216. package/dist/redaction/redactor.js.map +1 -0
  217. package/dist/redaction/scrub.d.ts +29 -0
  218. package/dist/redaction/scrub.d.ts.map +1 -0
  219. package/dist/redaction/scrub.js +54 -0
  220. package/dist/redaction/scrub.js.map +1 -0
  221. package/dist/runtime/config.d.ts +15 -0
  222. package/dist/runtime/config.d.ts.map +1 -0
  223. package/dist/runtime/config.js +2 -0
  224. package/dist/runtime/config.js.map +1 -0
  225. package/dist/runtime/driver-boundary.d.ts +16 -0
  226. package/dist/runtime/driver-boundary.d.ts.map +1 -0
  227. package/dist/runtime/driver-boundary.js +71 -0
  228. package/dist/runtime/driver-boundary.js.map +1 -0
  229. package/dist/runtime/driver-session.d.ts +9 -0
  230. package/dist/runtime/driver-session.d.ts.map +1 -0
  231. package/dist/runtime/driver-session.js +64 -0
  232. package/dist/runtime/driver-session.js.map +1 -0
  233. package/dist/runtime/internal-step-evidence.d.ts +30 -0
  234. package/dist/runtime/internal-step-evidence.d.ts.map +1 -0
  235. package/dist/runtime/internal-step-evidence.js +82 -0
  236. package/dist/runtime/internal-step-evidence.js.map +1 -0
  237. package/dist/runtime/result-adapter.d.ts +9 -0
  238. package/dist/runtime/result-adapter.d.ts.map +1 -0
  239. package/dist/runtime/result-adapter.js +77 -0
  240. package/dist/runtime/result-adapter.js.map +1 -0
  241. package/dist/runtime/runner.d.ts +35 -88
  242. package/dist/runtime/runner.d.ts.map +1 -1
  243. package/dist/runtime/runner.js +60 -74
  244. package/dist/runtime/runner.js.map +1 -1
  245. package/dist/runtime/scenario-control.d.ts +40 -0
  246. package/dist/runtime/scenario-control.d.ts.map +1 -0
  247. package/dist/runtime/scenario-control.js +94 -0
  248. package/dist/runtime/scenario-control.js.map +1 -0
  249. package/dist/runtime/scenario-runner.d.ts +39 -0
  250. package/dist/runtime/scenario-runner.d.ts.map +1 -0
  251. package/dist/runtime/scenario-runner.js +273 -0
  252. package/dist/runtime/scenario-runner.js.map +1 -0
  253. package/dist/runtime/scenario-state.d.ts +55 -0
  254. package/dist/runtime/scenario-state.d.ts.map +1 -0
  255. package/dist/runtime/scenario-state.js +128 -0
  256. package/dist/runtime/scenario-state.js.map +1 -0
  257. package/dist/runtime/step-execution.d.ts +24 -14
  258. package/dist/runtime/step-execution.d.ts.map +1 -1
  259. package/dist/runtime/step-execution.js +130 -36
  260. package/dist/runtime/step-execution.js.map +1 -1
  261. package/dist/runtime/text.d.ts +10 -0
  262. package/dist/runtime/text.d.ts.map +1 -0
  263. package/dist/runtime/text.js +25 -0
  264. package/dist/runtime/text.js.map +1 -0
  265. package/dist/runtime/types.d.ts +50 -76
  266. package/dist/runtime/types.d.ts.map +1 -1
  267. package/dist/runtime/usage-tracker.d.ts +37 -0
  268. package/dist/runtime/usage-tracker.d.ts.map +1 -0
  269. package/dist/runtime/usage-tracker.js +110 -0
  270. package/dist/runtime/usage-tracker.js.map +1 -0
  271. package/dist/tools/define-agent-tool.d.ts +3 -28
  272. package/dist/tools/define-agent-tool.d.ts.map +1 -1
  273. package/dist/tools/define-agent-tool.js +3 -28
  274. package/dist/tools/define-agent-tool.js.map +1 -1
  275. package/dist/tools/dispatcher.d.ts +11 -3
  276. package/dist/tools/dispatcher.d.ts.map +1 -1
  277. package/dist/tools/dispatcher.js +50 -8
  278. package/dist/tools/dispatcher.js.map +1 -1
  279. package/dist/tools/registry.d.ts +8 -10
  280. package/dist/tools/registry.d.ts.map +1 -1
  281. package/dist/tools/registry.js +19 -11
  282. package/dist/tools/registry.js.map +1 -1
  283. package/dist/tools/step/result-tool.d.ts.map +1 -1
  284. package/dist/tools/step/result-tool.js +8 -6
  285. package/dist/tools/step/result-tool.js.map +1 -1
  286. package/dist/tools/tool-contract.d.ts +1 -1
  287. package/dist/tools/tool-contract.d.ts.map +1 -1
  288. package/dist/tools/tool-contract.js.map +1 -1
  289. package/dist/tools/types.d.ts +77 -6
  290. package/dist/tools/types.d.ts.map +1 -1
  291. package/dist/tools/types.js.map +1 -1
  292. package/docs/CLI.md +59 -0
  293. package/docs/CONFIGURATION.md +91 -0
  294. package/docs/DRIVERS.md +75 -0
  295. package/docs/EVIDENCE.md +62 -0
  296. package/package.json +51 -59
  297. package/schemas/checkmate-config.v1.json +176 -0
  298. package/schemas/describe-result.v1.json +414 -0
  299. package/schemas/driver-descriptor.v1.json +66 -0
  300. package/schemas/run-request.v1.json +67 -0
  301. package/schemas/run-result.v1.json +333 -0
  302. package/schemas/validation-result.v1.json +86 -0
  303. package/dist/ai/response-processor.d.ts +0 -20
  304. package/dist/ai/response-processor.d.ts.map +0 -1
  305. package/dist/ai/response-processor.js +0 -62
  306. package/dist/ai/response-processor.js.map +0 -1
  307. package/dist/ai/token-pricing.d.ts +0 -13
  308. package/dist/ai/token-pricing.d.ts.map +0 -1
  309. package/dist/ai/token-pricing.js +0 -348
  310. package/dist/ai/token-pricing.js.map +0 -1
  311. package/dist/ai/token-tracker.d.ts +0 -33
  312. package/dist/ai/token-tracker.d.ts.map +0 -1
  313. package/dist/ai/token-tracker.js +0 -138
  314. package/dist/ai/token-tracker.js.map +0 -1
  315. package/dist/cli/create-examples.d.ts +0 -21
  316. package/dist/cli/create-examples.d.ts.map +0 -1
  317. package/dist/cli/create-examples.js +0 -116
  318. package/dist/cli/create-examples.js.map +0 -1
  319. package/dist/cli.d.ts +0 -15
  320. package/dist/cli.d.ts.map +0 -1
  321. package/dist/cli.js +0 -69
  322. package/dist/cli.js.map +0 -1
  323. package/dist/config/runtime-config.d.ts +0 -21
  324. package/dist/config/runtime-config.d.ts.map +0 -1
  325. package/dist/config/runtime-config.js +0 -98
  326. package/dist/config/runtime-config.js.map +0 -1
  327. package/dist/core.d.ts +0 -8
  328. package/dist/core.d.ts.map +0 -1
  329. package/dist/core.js +0 -4
  330. package/dist/core.js.map +0 -1
  331. package/dist/integrations/salesforce/authenticator.d.ts +0 -27
  332. package/dist/integrations/salesforce/authenticator.d.ts.map +0 -1
  333. package/dist/integrations/salesforce/authenticator.js +0 -27
  334. package/dist/integrations/salesforce/authenticator.js.map +0 -1
  335. package/dist/integrations/salesforce/cli-handler.d.ts +0 -14
  336. package/dist/integrations/salesforce/cli-handler.d.ts.map +0 -1
  337. package/dist/integrations/salesforce/cli-handler.js +0 -50
  338. package/dist/integrations/salesforce/cli-handler.js.map +0 -1
  339. package/dist/logging/index.d.ts +0 -2
  340. package/dist/logging/index.d.ts.map +0 -1
  341. package/dist/logging/index.js +0 -4
  342. package/dist/logging/index.js.map +0 -1
  343. package/dist/logging/logger.d.ts +0 -5
  344. package/dist/logging/logger.d.ts.map +0 -1
  345. package/dist/logging/logger.js +0 -15
  346. package/dist/logging/logger.js.map +0 -1
  347. package/dist/playwright.d.ts +0 -102
  348. package/dist/playwright.d.ts.map +0 -1
  349. package/dist/playwright.js +0 -116
  350. package/dist/playwright.js.map +0 -1
  351. package/dist/runtime/extension.d.ts +0 -274
  352. package/dist/runtime/extension.d.ts.map +0 -1
  353. package/dist/runtime/extension.js +0 -171
  354. package/dist/runtime/extension.js.map +0 -1
  355. package/dist/salesforce.d.ts +0 -71
  356. package/dist/salesforce.d.ts.map +0 -1
  357. package/dist/salesforce.js +0 -73
  358. package/dist/salesforce.js.map +0 -1
  359. package/dist/tools/browser/screenshot-service.d.ts.map +0 -1
  360. package/dist/tools/browser/screenshot-service.js.map +0 -1
  361. package/dist/tools/browser/snapshot-filter/index.d.ts.map +0 -1
  362. package/dist/tools/browser/snapshot-filter/index.js.map +0 -1
  363. package/dist/tools/browser/snapshot-filter/semantic-scorer.d.ts.map +0 -1
  364. package/dist/tools/browser/snapshot-filter/semantic-scorer.js.map +0 -1
  365. package/dist/tools/browser/snapshot-filter/snapshot-filter.d.ts +0 -4
  366. package/dist/tools/browser/snapshot-filter/snapshot-filter.d.ts.map +0 -1
  367. package/dist/tools/browser/snapshot-filter/snapshot-filter.js +0 -57
  368. package/dist/tools/browser/snapshot-filter/snapshot-filter.js.map +0 -1
  369. package/dist/tools/browser/snapshot-filter/tree-reconstructor.d.ts.map +0 -1
  370. package/dist/tools/browser/snapshot-filter/tree-reconstructor.js.map +0 -1
  371. package/dist/tools/browser/snapshot-service.d.ts +0 -19
  372. package/dist/tools/browser/snapshot-service.d.ts.map +0 -1
  373. package/dist/tools/browser/snapshot-service.js.map +0 -1
  374. package/dist/tools/browser/tool.d.ts +0 -34
  375. package/dist/tools/browser/tool.d.ts.map +0 -1
  376. package/dist/tools/browser/tool.js +0 -226
  377. package/dist/tools/browser/tool.js.map +0 -1
  378. package/dist/tools/browser/transient-state-tracker.d.ts +0 -16
  379. package/dist/tools/browser/transient-state-tracker.d.ts.map +0 -1
  380. package/dist/tools/browser/transient-state-tracker.js.map +0 -1
  381. package/dist/tools/salesforce/login-tool.d.ts +0 -7
  382. package/dist/tools/salesforce/login-tool.d.ts.map +0 -1
  383. package/dist/tools/salesforce/login-tool.js +0 -25
  384. package/dist/tools/salesforce/login-tool.js.map +0 -1
  385. package/docs/EXTENSIONS.md +0 -232
  386. package/docs/GUIDE.md +0 -465
  387. package/docs/ROADMAP.md +0 -47
  388. package/playwright.config.ts +0 -39
  389. package/test/examples/salesforce/trial-dev-org.spec.ts +0 -82
  390. package/test/examples/web/website-testing.spec.ts +0 -281
  391. /package/dist/{tools/browser → drivers/web/tools}/screenshot-service.js +0 -0
  392. /package/dist/{tools/browser → drivers/web/tools}/snapshot-filter/index.d.ts +0 -0
  393. /package/dist/{tools/browser → drivers/web/tools}/snapshot-filter/index.js +0 -0
  394. /package/dist/{tools/browser → drivers/web/tools}/snapshot-filter/semantic-scorer.d.ts +0 -0
  395. /package/dist/{tools/browser → drivers/web/tools}/snapshot-filter/semantic-scorer.js +0 -0
  396. /package/dist/{tools/browser → drivers/web/tools}/snapshot-filter/tree-reconstructor.d.ts +0 -0
  397. /package/dist/{tools/browser → drivers/web/tools}/snapshot-filter/tree-reconstructor.js +0 -0
@@ -1,17 +1,27 @@
1
- import { Step } from './types.js';
2
- import { AiClient } from '../ai/client.js';
3
- import { ExtensionHost } from './extension.js';
1
+ import type { AiClient } from '../ai/client.js';
2
+ import type { RuntimeConfig } from './config.js';
3
+ import type { DriverSession, StepIntent } from '../driver.js';
4
+ import type { RuntimeLogger } from '../logging/types.js';
5
+ import { DiagnosticSanitizer } from '../redaction/diagnostic-sanitizer.js';
6
+ import type { ToolRegistry } from '../tools/registry.js';
7
+ import type { StepControl } from './scenario-control.js';
8
+ import type { InternalStepReport } from './types.js';
9
+ import { ScenarioUsageTracker } from './usage-tracker.js';
10
+ export type StepExecutionDependencies = {
11
+ config: RuntimeConfig;
12
+ aiClient: AiClient;
13
+ toolRegistry: ToolRegistry;
14
+ driverSession: DriverSession;
15
+ usageTracker: ScenarioUsageTracker;
16
+ driverId: string;
17
+ redact: boolean;
18
+ diagnosticSanitizer: DiagnosticSanitizer;
19
+ logger: RuntimeLogger;
20
+ };
21
+ /** Authoritative driver-backed model/tool loop. */
4
22
  export declare class StepExecution {
5
- private readonly aiClient;
6
- private readonly extensionHost;
7
- private stepResult;
8
- private resolveStepResult;
9
- private readonly stepResultPromise;
10
- constructor(aiClient: AiClient, extensionHost: ExtensionHost);
11
- run(step: Step): Promise<void>;
12
- private logStepStart;
13
- private logStepFinish;
14
- private assertStepResult;
15
- private assertionMessage;
23
+ private readonly dependencies;
24
+ constructor(dependencies: StepExecutionDependencies);
25
+ run(step: StepIntent, control: StepControl): Promise<InternalStepReport>;
16
26
  }
17
27
  //# sourceMappingURL=step-execution.d.ts.map
@@ -1 +1 @@
1
- {"version":3,"file":"step-execution.d.ts","sourceRoot":"","sources":["../../src/runtime/step-execution.ts"],"names":[],"mappings":"AAEA,OAAO,EAAE,IAAI,EAAoD,MAAM,YAAY,CAAA;AAGnF,OAAO,EAAE,QAAQ,EAAE,MAAM,iBAAiB,CAAA;AAC1C,OAAO,EAAE,aAAa,EAAE,MAAM,gBAAgB,CAAA;AAE9C,qBAAa,aAAa;IAMxB,OAAO,CAAC,QAAQ,CAAC,QAAQ;IACzB,OAAO,CAAC,QAAQ,CAAC,aAAa;IAN/B,OAAO,CAAC,UAAU,CAA4C;IAC9D,OAAO,CAAC,iBAAiB,CAAoB;IAC7C,OAAO,CAAC,QAAQ,CAAC,iBAAiB,CAAmB;gBAGnC,QAAQ,EAAE,QAAQ,EAClB,aAAa,EAAE,aAAa;IAOxC,GAAG,CAAC,IAAI,EAAE,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC;IAoBpC,OAAO,CAAC,YAAY;IAIpB,OAAO,CAAC,aAAa;IAIrB,OAAO,CAAC,gBAAgB;IAIxB,OAAO,CAAC,gBAAgB;CAGxB"}
1
+ {"version":3,"file":"step-execution.d.ts","sourceRoot":"","sources":["../../src/runtime/step-execution.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,QAAQ,EAAE,MAAM,iBAAiB,CAAA;AAI/C,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,aAAa,CAAA;AAChD,OAAO,KAAK,EAAE,aAAa,EAAE,UAAU,EAAE,MAAM,cAAc,CAAA;AAC7D,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,qBAAqB,CAAA;AACxD,OAAO,EAAE,mBAAmB,EAAE,MAAM,sCAAsC,CAAA;AAG1E,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,sBAAsB,CAAA;AAIxD,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,uBAAuB,CAAA;AACxD,OAAO,KAAK,EAAE,kBAAkB,EAA6B,MAAM,YAAY,CAAA;AAC/E,OAAO,EAGN,oBAAoB,EAEpB,MAAM,oBAAoB,CAAA;AAE3B,MAAM,MAAM,yBAAyB,GAAG;IACvC,MAAM,EAAE,aAAa,CAAA;IACrB,QAAQ,EAAE,QAAQ,CAAA;IAClB,YAAY,EAAE,YAAY,CAAA;IAC1B,aAAa,EAAE,aAAa,CAAA;IAC5B,YAAY,EAAE,oBAAoB,CAAA;IAClC,QAAQ,EAAE,MAAM,CAAA;IAChB,MAAM,EAAE,OAAO,CAAA;IACf,mBAAmB,EAAE,mBAAmB,CAAA;IACxC,MAAM,EAAE,aAAa,CAAA;CACrB,CAAA;AAED,mDAAmD;AACnD,qBAAa,aAAa;IACb,OAAO,CAAC,QAAQ,CAAC,YAAY;gBAAZ,YAAY,EAAE,yBAAyB;IAE9D,GAAG,CAAC,IAAI,EAAE,UAAU,EAAE,OAAO,EAAE,WAAW,GAAG,OAAO,CAAC,kBAAkB,CAAC;CAoI9E"}
@@ -1,49 +1,143 @@
1
- import { expect } from '@playwright/test';
2
- import { logger } from '../logging/index.js';
3
- import { STEP_START_USER_PROMPT, STEP_SYSTEM_PROMPT } from '../ai/prompts.js';
4
1
  import { MessageHistory } from '../ai/message-history.js';
2
+ import { STEP_START_USER_PROMPT, STEP_SYSTEM_PROMPT } from '../ai/prompts.js';
3
+ import { TurnProcessor } from '../ai/turn-processor.js';
4
+ import { ToolDispatchError } from '../tools/dispatcher.js';
5
+ import { LoopDetector } from '../tools/loop-detector.js';
6
+ import { DriverBoundaryError } from './driver-boundary.js';
7
+ import { buildInitialDriverContext, buildPostToolDriverContext, isEphemeralDriverMessage } from './driver-session.js';
8
+ import { InternalStepEvidence } from './internal-step-evidence.js';
9
+ import { ProviderUsageInvalidError, ProviderUsageUnavailableError, TokenBudgetExceededError, } from './usage-tracker.js';
10
+ /** Authoritative driver-backed model/tool loop. */
5
11
  export class StepExecution {
6
- aiClient;
7
- extensionHost;
8
- stepResult = { passed: false, actual: '' };
9
- resolveStepResult;
10
- stepResultPromise;
11
- constructor(aiClient, extensionHost) {
12
- this.aiClient = aiClient;
13
- this.extensionHost = extensionHost;
14
- this.stepResultPromise = new Promise((resolve) => {
15
- this.resolveStepResult = resolve;
16
- });
12
+ dependencies;
13
+ constructor(dependencies) {
14
+ this.dependencies = dependencies;
17
15
  }
18
- async run(step) {
19
- this.logStepStart(step);
16
+ async run(step, control) {
17
+ const { config, aiClient, toolRegistry, driverSession, usageTracker, driverId, redact, diagnosticSanitizer, logger, } = this.dependencies;
18
+ const evidence = new InternalStepEvidence(step, driverId, redact, diagnosticSanitizer, control.now);
19
+ const checkpoint = usageTracker.beginStep();
20
+ const messages = [];
21
+ const ephemeralMessages = new Set();
22
+ const turnProcessor = new TurnProcessor({
23
+ config,
24
+ toolRegistry,
25
+ loopDetector: new LoopDetector(config.loopMaxRepetitions),
26
+ evidence,
27
+ logger,
28
+ });
29
+ messages.push(...new MessageHistory().buildInitialMessages({
30
+ systemPrompt: STEP_SYSTEM_PROMPT([...driverSession.instructions]),
31
+ userPrompt: STEP_START_USER_PROMPT(step),
32
+ }));
33
+ logger.info(`step started: ${step.id}`);
34
+ let turns = 0;
20
35
  try {
21
- await this.aiClient.initialize(step, this.resolveStepResult);
22
- const initialMessages = new MessageHistory().buildInitialMessages({
23
- systemPrompt: STEP_SYSTEM_PROMPT(this.extensionHost.getInstructions()),
24
- userPrompt: STEP_START_USER_PROMPT(step),
25
- contextMessages: await this.extensionHost.buildInitialMessages(step),
26
- });
27
- await this.aiClient.sendMessage(initialMessages);
28
- this.stepResult = await this.stepResultPromise;
29
- this.assertStepResult(step);
36
+ try {
37
+ appendDriverContext(messages, ephemeralMessages, await buildInitialDriverContext(driverSession, step, control));
38
+ }
39
+ catch (error) {
40
+ return finish(evidence, usageTracker, checkpoint, failureFrom(error, control, turns, diagnosticSanitizer), logger);
41
+ }
42
+ for (;;) {
43
+ if (turns >= config.turnCap) {
44
+ return finish(evidence, usageTracker, checkpoint, { outcome: 'failed', reason: 'turn-cap-exceeded', turns }, logger);
45
+ }
46
+ const expired = control.poll();
47
+ if (expired.expired) {
48
+ return finish(evidence, usageTracker, checkpoint, { outcome: 'failed', reason: expired.reason, turns }, logger);
49
+ }
50
+ turns++;
51
+ try {
52
+ const { response, assistantMessages } = await aiClient.send(messages, {
53
+ step,
54
+ signal: control.signal,
55
+ });
56
+ usageTracker.record(response.usage);
57
+ assertControlLive(control, 'provider-response');
58
+ messages.push(...assistantMessages);
59
+ const outcome = await turnProcessor.process({ response, step, turn: turns, control });
60
+ assertControlLive(control, 'turn-processing');
61
+ if (outcome.kind === 'assertion') {
62
+ return finish(evidence, usageTracker, checkpoint, {
63
+ outcome: outcome.passed ? 'passed' : 'failed',
64
+ reason: outcome.passed ? 'met-expectation' : 'failed-expectation',
65
+ actual: outcome.actual,
66
+ turns,
67
+ }, logger);
68
+ }
69
+ if (outcome.kind === 'stuck') {
70
+ return finish(evidence, usageTracker, checkpoint, {
71
+ outcome: 'failed',
72
+ reason: 'loop-detected',
73
+ actual: 'the model repeated the same tool calls without reaching a result',
74
+ turns,
75
+ }, logger);
76
+ }
77
+ dropEphemeralMessages(messages, ephemeralMessages);
78
+ messages.push(...outcome.messages);
79
+ appendDriverContext(messages, ephemeralMessages, await buildPostToolDriverContext(driverSession, step, turns, outcome.toolResults, control));
80
+ }
81
+ catch (error) {
82
+ return finish(evidence, usageTracker, checkpoint, failureFrom(error, control, turns, diagnosticSanitizer), logger);
83
+ }
84
+ }
30
85
  }
31
- catch (error) {
32
- throw new Error(`\nFailed to execute action:\n${step.action}`, { cause: error });
86
+ finally {
87
+ control.dispose();
33
88
  }
34
- this.logStepFinish();
35
89
  }
36
- logStepStart(step) {
37
- logger.info(`step started:\n${JSON.stringify(step, null, 2).replaceAll(' ', '').trim()}`);
90
+ }
91
+ function finish(evidence, usageTracker, checkpoint, termination, logger) {
92
+ if (termination.reason === 'step-timeout' || termination.reason === 'scenario-timeout') {
93
+ evidence.recordDiagnostic({
94
+ code: 'expired-boundary',
95
+ path: '',
96
+ message: termination.actual ?? `Step execution expired with ${termination.reason}`,
97
+ });
38
98
  }
39
- logStepFinish() {
40
- logger.info('step finished');
99
+ const report = evidence.buildReport(termination, usageTracker.stepUsage(checkpoint));
100
+ logger.info(`step finished: ${report.step.id} ${report.outcome} (${report.category} / ${report.reason})`);
101
+ return report;
102
+ }
103
+ function appendDriverContext(messages, ephemeralMessages, context) {
104
+ for (const message of context) {
105
+ messages.push(message);
106
+ if (isEphemeralDriverMessage(message))
107
+ ephemeralMessages.add(message);
41
108
  }
42
- assertStepResult(step) {
43
- expect(this.stepResult.passed, this.assertionMessage(step)).toBeTruthy();
109
+ }
110
+ function dropEphemeralMessages(messages, ephemeralMessages) {
111
+ if (ephemeralMessages.size === 0)
112
+ return;
113
+ const retained = messages.filter((message) => !ephemeralMessages.has(message));
114
+ messages.length = 0;
115
+ messages.push(...retained);
116
+ ephemeralMessages.clear();
117
+ }
118
+ function assertControlLive(control, operation) {
119
+ const expired = control.poll();
120
+ if (expired.expired)
121
+ throw new DriverBoundaryError(operation, expired.reason);
122
+ }
123
+ function failureFrom(error, control, turns, sanitizer) {
124
+ const actual = sanitizer.error(error);
125
+ const expired = control.poll();
126
+ if (expired.expired)
127
+ return { outcome: 'failed', reason: expired.reason, actual, turns };
128
+ if (error instanceof DriverBoundaryError)
129
+ return { outcome: 'failed', reason: error.reason, actual, turns };
130
+ if (error instanceof TokenBudgetExceededError) {
131
+ return { outcome: 'failed', reason: 'token-budget-exceeded', actual, turns };
44
132
  }
45
- assertionMessage(step) {
46
- return this.stepResult.passed ? step.expect : `\nExpected: ${step.expect}\nActual: ${this.stepResult.actual}`;
133
+ if (error instanceof ProviderUsageUnavailableError || error instanceof ProviderUsageInvalidError) {
134
+ return { outcome: 'failed', reason: 'provider-error', actual, turns };
47
135
  }
136
+ if (error instanceof ToolDispatchError)
137
+ return { outcome: 'failed', reason: 'tool-error', actual, turns };
138
+ return { outcome: 'failed', reason: driverInfraReason(error), actual, turns };
139
+ }
140
+ function driverInfraReason(error) {
141
+ return error instanceof ToolDispatchError ? 'tool-error' : 'provider-error';
48
142
  }
49
143
  //# sourceMappingURL=step-execution.js.map
@@ -1 +1 @@
1
- {"version":3,"file":"step-execution.js","sourceRoot":"","sources":["../../src/runtime/step-execution.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,MAAM,EAAE,MAAM,kBAAkB,CAAA;AACzC,OAAO,EAAE,MAAM,EAAE,MAAM,qBAAqB,CAAA;AAE5C,OAAO,EAAE,sBAAsB,EAAE,kBAAkB,EAAE,MAAM,kBAAkB,CAAA;AAC7E,OAAO,EAAE,cAAc,EAAE,MAAM,0BAA0B,CAAA;AAIzD,MAAM,OAAO,aAAa;IAMP;IACA;IANV,UAAU,GAAe,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,EAAE,EAAE,EAAE,CAAA;IACtD,iBAAiB,CAAoB;IAC5B,iBAAiB,CAAmB;IAErD,YACkB,QAAkB,EAClB,aAA4B;QAD5B,aAAQ,GAAR,QAAQ,CAAU;QAClB,kBAAa,GAAb,aAAa,CAAe;QAE7C,IAAI,CAAC,iBAAiB,GAAG,IAAI,OAAO,CAAa,CAAC,OAAO,EAAE,EAAE;YAC5D,IAAI,CAAC,iBAAiB,GAAG,OAAO,CAAA;QACjC,CAAC,CAAC,CAAA;IACH,CAAC;IAED,KAAK,CAAC,GAAG,CAAC,IAAU;QACnB,IAAI,CAAC,YAAY,CAAC,IAAI,CAAC,CAAA;QAEvB,IAAI,CAAC;YACJ,MAAM,IAAI,CAAC,QAAQ,CAAC,UAAU,CAAC,IAAI,EAAE,IAAI,CAAC,iBAAiB,CAAC,CAAA;YAC5D,MAAM,eAAe,GAAG,IAAI,cAAc,EAAE,CAAC,oBAAoB,CAAC;gBACjE,YAAY,EAAE,kBAAkB,CAAC,IAAI,CAAC,aAAa,CAAC,eAAe,EAAE,CAAC;gBACtE,UAAU,EAAE,sBAAsB,CAAC,IAAI,CAAC;gBACxC,eAAe,EAAE,MAAM,IAAI,CAAC,aAAa,CAAC,oBAAoB,CAAC,IAAI,CAAC;aACpE,CAAC,CAAA;YACF,MAAM,IAAI,CAAC,QAAQ,CAAC,WAAW,CAAC,eAAe,CAAC,CAAA;YAChD,IAAI,CAAC,UAAU,GAAG,MAAM,IAAI,CAAC,iBAAiB,CAAA;YAC9C,IAAI,CAAC,gBAAgB,CAAC,IAAI,CAAC,CAAA;QAC5B,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YAChB,MAAM,IAAI,KAAK,CAAC,gCAAgC,IAAI,CAAC,MAAM,EAAE,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,CAAC,CAAA;QACjF,CAAC;QAED,IAAI,CAAC,aAAa,EAAE,CAAA;IACrB,CAAC;IAEO,YAAY,CAAC,IAAU;QAC9B,MAAM,CAAC,IAAI,CAAC,kBAAkB,IAAI,CAAC,SAAS,CAAC,IAAI,EAAE,IAAI,EAAE,CAAC,CAAC,CAAC,UAAU,CAAC,IAAI,EAAE,EAAE,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,CAAA;IAC3F,CAAC;IAEO,aAAa;QACpB,MAAM,CAAC,IAAI,CAAC,eAAe,CAAC,CAAA;IAC7B,CAAC;IAEO,gBAAgB,CAAC,IAAU;QAClC,MAAM,CAAC,IAAI,CAAC,UAAU,CAAC,MAAM,EAAE,IAAI,CAAC,gBAAgB,CAAC,IAAI,CAAC,CAAC,CAAC,UAAU,EAAE,CAAA;IACzE,CAAC;IAEO,gBAAgB,CAAC,IAAU;QAClC,OAAO,IAAI,CAAC,UAAU,CAAC,MAAM,CAAC,CAAC,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,eAAe,IAAI,CAAC,MAAM,aAAa,IAAI,CAAC,UAAU,CAAC,MAAM,EAAE,CAAA;IAC9G,CAAC;CACD"}
1
+ {"version":3,"file":"step-execution.js","sourceRoot":"","sources":["../../src/runtime/step-execution.ts"],"names":[],"mappings":"AAEA,OAAO,EAAE,cAAc,EAAE,MAAM,0BAA0B,CAAA;AACzD,OAAO,EAAE,sBAAsB,EAAE,kBAAkB,EAAE,MAAM,kBAAkB,CAAA;AAC7E,OAAO,EAAE,aAAa,EAAE,MAAM,yBAAyB,CAAA;AAKvD,OAAO,EAAE,iBAAiB,EAAE,MAAM,wBAAwB,CAAA;AAC1D,OAAO,EAAE,YAAY,EAAE,MAAM,2BAA2B,CAAA;AAExD,OAAO,EAAE,mBAAmB,EAAE,MAAM,sBAAsB,CAAA;AAC1D,OAAO,EAAE,yBAAyB,EAAE,0BAA0B,EAAE,wBAAwB,EAAE,MAAM,qBAAqB,CAAA;AACrH,OAAO,EAAE,oBAAoB,EAAgC,MAAM,6BAA6B,CAAA;AAGhG,OAAO,EACN,yBAAyB,EACzB,6BAA6B,EAE7B,wBAAwB,GACxB,MAAM,oBAAoB,CAAA;AAc3B,mDAAmD;AACnD,MAAM,OAAO,aAAa;IACI;IAA7B,YAA6B,YAAuC;QAAvC,iBAAY,GAAZ,YAAY,CAA2B;IAAG,CAAC;IAExE,KAAK,CAAC,GAAG,CAAC,IAAgB,EAAE,OAAoB;QAC/C,MAAM,EACL,MAAM,EACN,QAAQ,EACR,YAAY,EACZ,aAAa,EACb,YAAY,EACZ,QAAQ,EACR,MAAM,EACN,mBAAmB,EACnB,MAAM,GACN,GAAG,IAAI,CAAC,YAAY,CAAA;QACrB,MAAM,QAAQ,GAAG,IAAI,oBAAoB,CAAC,IAAI,EAAE,QAAQ,EAAE,MAAM,EAAE,mBAAmB,EAAE,OAAO,CAAC,GAAG,CAAC,CAAA;QACnG,MAAM,UAAU,GAAG,YAAY,CAAC,SAAS,EAAE,CAAA;QAC3C,MAAM,QAAQ,GAAiC,EAAE,CAAA;QACjD,MAAM,iBAAiB,GAAG,IAAI,GAAG,EAA8B,CAAA;QAC/D,MAAM,aAAa,GAAG,IAAI,aAAa,CAAC;YACvC,MAAM;YACN,YAAY;YACZ,YAAY,EAAE,IAAI,YAAY,CAAC,MAAM,CAAC,kBAAkB,CAAC;YACzD,QAAQ;YACR,MAAM;SACN,CAAC,CAAA;QACF,QAAQ,CAAC,IAAI,CACZ,GAAG,IAAI,cAAc,EAAE,CAAC,oBAAoB,CAAC;YAC5C,YAAY,EAAE,kBAAkB,CAAC,CAAC,GAAG,aAAa,CAAC,YAAY,CAAC,CAAC;YACjE,UAAU,EAAE,sBAAsB,CAAC,IAAI,CAAC;SACxC,CAAC,CACF,CAAA;QACD,MAAM,CAAC,IAAI,CAAC,iBAAiB,IAAI,CAAC,EAAE,EAAE,CAAC,CAAA;QAEvC,IAAI,KAAK,GAAG,CAAC,CAAA;QACb,IAAI,CAAC;YACJ,IAAI,CAAC;gBACJ,mBAAmB,CAClB,QAAQ,EACR,iBAAiB,EACjB,MAAM,yBAAyB,CAAC,aAAa,EAAE,IAAI,EAAE,OAAO,CAAC,CAC7D,CAAA;YACF,CAAC;YAAC,OAAO,KAAK,EAAE,CAAC;gBAChB,OAAO,MAAM,CACZ,QAAQ,EACR,YAAY,EACZ,UAAU,EACV,WAAW,CAAC,KAAK,EAAE,OAAO,EAAE,KAAK,EAAE,mBAAmB,CAAC,EACvD,MAAM,CACN,CAAA;YACF,CAAC;YAED,SAAS,CAAC;gBACT,IAAI,KAAK,IAAI,MAAM,CAAC,OAAO,EAAE,CAAC;oBAC7B,OAAO,MAAM,CACZ,QAAQ,EACR,YAAY,EACZ,UAAU,EACV,EAAE,OAAO,EAAE,QAAQ,EAAE,MAAM,EAAE,mBAAmB,EAAE,KAAK,EAAE,EACzD,MAAM,CACN,CAAA;gBACF,CAAC;gBACD,MAAM,OAAO,GAAG,OAAO,CAAC,IAAI,EAAE,CAAA;gBAC9B,IAAI,OAAO,CAAC,OAAO,EAAE,CAAC;oBACrB,OAAO,MAAM,CACZ,QAAQ,EACR,YAAY,EACZ,UAAU,EACV,EAAE,OAAO,EAAE,QAAQ,EAAE,MAAM,EAAE,OAAO,CAAC,MAAM,EAAE,KAAK,EAAE,EACpD,MAAM,CACN,CAAA;gBACF,CAAC;gBAED,KAAK,EAAE,CAAA;gBACP,IAAI,CAAC;oBACJ,MAAM,EAAE,QAAQ,EAAE,iBAAiB,EAAE,GAAG,MAAM,QAAQ,CAAC,IAAI,CAAC,QAAQ,EAAE;wBACrE,IAAI;wBACJ,MAAM,EAAE,OAAO,CAAC,MAAM;qBACtB,CAAC,CAAA;oBACF,YAAY,CAAC,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAA;oBACnC,iBAAiB,CAAC,OAAO,EAAE,mBAAmB,CAAC,CAAA;oBAC/C,QAAQ,CAAC,IAAI,CAAC,GAAG,iBAAiB,CAAC,CAAA;oBACnC,MAAM,OAAO,GAAG,MAAM,aAAa,CAAC,OAAO,CAAC,EAAE,QAAQ,EAAE,IAAI,EAAE,IAAI,EAAE,KAAK,EAAE,OAAO,EAAE,CAAC,CAAA;oBACrF,iBAAiB,CAAC,OAAO,EAAE,iBAAiB,CAAC,CAAA;oBAE7C,IAAI,OAAO,CAAC,IAAI,KAAK,WAAW,EAAE,CAAC;wBAClC,OAAO,MAAM,CACZ,QAAQ,EACR,YAAY,EACZ,UAAU,EACV;4BACC,OAAO,EAAE,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,QAAQ;4BAC7C,MAAM,EAAE,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,iBAAiB,CAAC,CAAC,CAAC,oBAAoB;4BACjE,MAAM,EAAE,OAAO,CAAC,MAAM;4BACtB,KAAK;yBACL,EACD,MAAM,CACN,CAAA;oBACF,CAAC;oBACD,IAAI,OAAO,CAAC,IAAI,KAAK,OAAO,EAAE,CAAC;wBAC9B,OAAO,MAAM,CACZ,QAAQ,EACR,YAAY,EACZ,UAAU,EACV;4BACC,OAAO,EAAE,QAAQ;4BACjB,MAAM,EAAE,eAAe;4BACvB,MAAM,EAAE,kEAAkE;4BAC1E,KAAK;yBACL,EACD,MAAM,CACN,CAAA;oBACF,CAAC;oBAED,qBAAqB,CAAC,QAAQ,EAAE,iBAAiB,CAAC,CAAA;oBAClD,QAAQ,CAAC,IAAI,CAAC,GAAG,OAAO,CAAC,QAAQ,CAAC,CAAA;oBAClC,mBAAmB,CAClB,QAAQ,EACR,iBAAiB,EACjB,MAAM,0BAA0B,CAAC,aAAa,EAAE,IAAI,EAAE,KAAK,EAAE,OAAO,CAAC,WAAW,EAAE,OAAO,CAAC,CAC1F,CAAA;gBACF,CAAC;gBAAC,OAAO,KAAK,EAAE,CAAC;oBAChB,OAAO,MAAM,CACZ,QAAQ,EACR,YAAY,EACZ,UAAU,EACV,WAAW,CAAC,KAAK,EAAE,OAAO,EAAE,KAAK,EAAE,mBAAmB,CAAC,EACvD,MAAM,CACN,CAAA;gBACF,CAAC;YACF,CAAC;QACF,CAAC;gBAAS,CAAC;YACV,OAAO,CAAC,OAAO,EAAE,CAAA;QAClB,CAAC;IACF,CAAC;CACD;AAED,SAAS,MAAM,CACd,QAA8B,EAC9B,YAAkC,EAClC,UAAyD,EACzD,WAAoC,EACpC,MAAqB;IAErB,IAAI,WAAW,CAAC,MAAM,KAAK,cAAc,IAAI,WAAW,CAAC,MAAM,KAAK,kBAAkB,EAAE,CAAC;QACxF,QAAQ,CAAC,gBAAgB,CAAC;YACzB,IAAI,EAAE,kBAAkB;YACxB,IAAI,EAAE,EAAE;YACR,OAAO,EAAE,WAAW,CAAC,MAAM,IAAI,+BAA+B,WAAW,CAAC,MAAM,EAAE;SAClF,CAAC,CAAA;IACH,CAAC;IACD,MAAM,MAAM,GAAG,QAAQ,CAAC,WAAW,CAAC,WAAW,EAAE,YAAY,CAAC,SAAS,CAAC,UAAU,CAAC,CAAC,CAAA;IACpF,MAAM,CAAC,IAAI,CAAC,kBAAkB,MAAM,CAAC,IAAI,CAAC,EAAE,IAAI,MAAM,CAAC,OAAO,KAAK,MAAM,CAAC,QAAQ,MAAM,MAAM,CAAC,MAAM,GAAG,CAAC,CAAA;IACzG,OAAO,MAAM,CAAA;AACd,CAAC;AAED,SAAS,mBAAmB,CAC3B,QAAsC,EACtC,iBAAkD,EAClD,OAAqC;IAErC,KAAK,MAAM,OAAO,IAAI,OAAO,EAAE,CAAC;QAC/B,QAAQ,CAAC,IAAI,CAAC,OAAO,CAAC,CAAA;QACtB,IAAI,wBAAwB,CAAC,OAAO,CAAC;YAAE,iBAAiB,CAAC,GAAG,CAAC,OAAO,CAAC,CAAA;IACtE,CAAC;AACF,CAAC;AAED,SAAS,qBAAqB,CAC7B,QAAsC,EACtC,iBAAkD;IAElD,IAAI,iBAAiB,CAAC,IAAI,KAAK,CAAC;QAAE,OAAM;IACxC,MAAM,QAAQ,GAAG,QAAQ,CAAC,MAAM,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,CAAC,iBAAiB,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,CAAA;IAC9E,QAAQ,CAAC,MAAM,GAAG,CAAC,CAAA;IACnB,QAAQ,CAAC,IAAI,CAAC,GAAG,QAAQ,CAAC,CAAA;IAC1B,iBAAiB,CAAC,KAAK,EAAE,CAAA;AAC1B,CAAC;AAED,SAAS,iBAAiB,CAAC,OAAoB,EAAE,SAAiB;IACjE,MAAM,OAAO,GAAG,OAAO,CAAC,IAAI,EAAE,CAAA;IAC9B,IAAI,OAAO,CAAC,OAAO;QAAE,MAAM,IAAI,mBAAmB,CAAC,SAAS,EAAE,OAAO,CAAC,MAAM,CAAC,CAAA;AAC9E,CAAC;AAED,SAAS,WAAW,CACnB,KAAc,EACd,OAAoB,EACpB,KAAa,EACb,SAA8B;IAE9B,MAAM,MAAM,GAAG,SAAS,CAAC,KAAK,CAAC,KAAK,CAAC,CAAA;IACrC,MAAM,OAAO,GAAG,OAAO,CAAC,IAAI,EAAE,CAAA;IAC9B,IAAI,OAAO,CAAC,OAAO;QAAE,OAAO,EAAE,OAAO,EAAE,QAAQ,EAAE,MAAM,EAAE,OAAO,CAAC,MAAM,EAAE,MAAM,EAAE,KAAK,EAAE,CAAA;IACxF,IAAI,KAAK,YAAY,mBAAmB;QAAE,OAAO,EAAE,OAAO,EAAE,QAAQ,EAAE,MAAM,EAAE,KAAK,CAAC,MAAM,EAAE,MAAM,EAAE,KAAK,EAAE,CAAA;IAC3G,IAAI,KAAK,YAAY,wBAAwB,EAAE,CAAC;QAC/C,OAAO,EAAE,OAAO,EAAE,QAAQ,EAAE,MAAM,EAAE,uBAAuB,EAAE,MAAM,EAAE,KAAK,EAAE,CAAA;IAC7E,CAAC;IACD,IAAI,KAAK,YAAY,6BAA6B,IAAI,KAAK,YAAY,yBAAyB,EAAE,CAAC;QAClG,OAAO,EAAE,OAAO,EAAE,QAAQ,EAAE,MAAM,EAAE,gBAAgB,EAAE,MAAM,EAAE,KAAK,EAAE,CAAA;IACtE,CAAC;IACD,IAAI,KAAK,YAAY,iBAAiB;QAAE,OAAO,EAAE,OAAO,EAAE,QAAQ,EAAE,MAAM,EAAE,YAAY,EAAE,MAAM,EAAE,KAAK,EAAE,CAAA;IACzG,OAAO,EAAE,OAAO,EAAE,QAAQ,EAAE,MAAM,EAAE,iBAAiB,CAAC,KAAK,CAAC,EAAE,MAAM,EAAE,KAAK,EAAE,CAAA;AAC9E,CAAC;AAED,SAAS,iBAAiB,CAAC,KAAc;IACxC,OAAO,KAAK,YAAY,iBAAiB,CAAC,CAAC,CAAC,YAAY,CAAC,CAAC,CAAC,gBAAgB,CAAA;AAC5E,CAAC"}
@@ -0,0 +1,10 @@
1
+ /**
2
+ * Removes the common leading indentation from a multi-line string.
3
+ *
4
+ * Steps are written as template literals in a spec file, so `action` and `expect` arrive
5
+ * carrying whatever indentation the surrounding code happened to have. That indentation is
6
+ * an artifact of the source layout, not of the step, and it should not reach the report or
7
+ * the failure message.
8
+ */
9
+ export declare function dedent(value: string): string;
10
+ //# sourceMappingURL=text.d.ts.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"text.d.ts","sourceRoot":"","sources":["../../src/runtime/text.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AACH,wBAAgB,MAAM,CAAC,KAAK,EAAE,MAAM,GAAG,MAAM,CAoB5C"}
@@ -0,0 +1,25 @@
1
+ /**
2
+ * Removes the common leading indentation from a multi-line string.
3
+ *
4
+ * Steps are written as template literals in a spec file, so `action` and `expect` arrive
5
+ * carrying whatever indentation the surrounding code happened to have. That indentation is
6
+ * an artifact of the source layout, not of the step, and it should not reach the report or
7
+ * the failure message.
8
+ */
9
+ export function dedent(value) {
10
+ const lines = value.split('\n').map((line) => line.replace(/\s+$/, ''));
11
+ while (lines.length > 0 && lines[0].trim() === '') {
12
+ lines.shift();
13
+ }
14
+ while (lines.length > 0 && lines[lines.length - 1].trim() === '') {
15
+ lines.pop();
16
+ }
17
+ const indent = lines
18
+ .filter((line) => line.trim() !== '')
19
+ .reduce((smallest, line) => Math.min(smallest, line.length - line.trimStart().length), Number.MAX_SAFE_INTEGER);
20
+ if (indent === Number.MAX_SAFE_INTEGER) {
21
+ return '';
22
+ }
23
+ return lines.map((line) => line.slice(indent)).join('\n');
24
+ }
25
+ //# sourceMappingURL=text.js.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"text.js","sourceRoot":"","sources":["../../src/runtime/text.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AACH,MAAM,UAAU,MAAM,CAAC,KAAa;IACnC,MAAM,KAAK,GAAG,KAAK,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,OAAO,CAAC,MAAM,EAAE,EAAE,CAAC,CAAC,CAAA;IAEvE,OAAO,KAAK,CAAC,MAAM,GAAG,CAAC,IAAI,KAAK,CAAC,CAAC,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE,EAAE,CAAC;QACnD,KAAK,CAAC,KAAK,EAAE,CAAA;IACd,CAAC;IAED,OAAO,KAAK,CAAC,MAAM,GAAG,CAAC,IAAI,KAAK,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE,EAAE,CAAC;QAClE,KAAK,CAAC,GAAG,EAAE,CAAA;IACZ,CAAC;IAED,MAAM,MAAM,GAAG,KAAK;SAClB,MAAM,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,IAAI,EAAE,KAAK,EAAE,CAAC;SACpC,MAAM,CAAC,CAAC,QAAQ,EAAE,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,GAAG,CAAC,QAAQ,EAAE,IAAI,CAAC,MAAM,GAAG,IAAI,CAAC,SAAS,EAAE,CAAC,MAAM,CAAC,EAAE,MAAM,CAAC,gBAAgB,CAAC,CAAA;IAEhH,IAAI,MAAM,KAAK,MAAM,CAAC,gBAAgB,EAAE,CAAC;QACxC,OAAO,EAAE,CAAA;IACV,CAAC;IAED,OAAO,KAAK,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAA;AAC1D,CAAC"}
@@ -1,79 +1,53 @@
1
- /**
2
- * A single natural-language test step executed by Checkmate.
3
- *
4
- * `action` should describe what the agent needs to do.
5
- * `expect` should describe the expected result after the action finishes.
6
- * `search` can bias snapshot filtering toward specific keywords.
7
- * `topPercent` controls how much of the scored page snapshot is kept.
8
- *
9
- * @example
10
- * ```ts
11
- * const step: Step = {
12
- * action: "Open the pricing page and click the Pro plan",
13
- * expect: "The checkout page for the Pro plan is displayed",
14
- * topPercent: 15,
15
- * }
16
- * ```
17
- */
18
- export interface Step {
19
- /**
20
- * What the agent should do in the browser.
21
- */
22
- action: string;
23
- /**
24
- * What should be true after the action is complete.
25
- */
26
- expect: string;
27
- /**
28
- * Optional keyword hints for snapshot filtering.
29
- *
30
- * When provided, these terms are prioritized over semantic `action + expect` matching.
31
- */
32
- search?: string[];
33
- /**
34
- * Optional percentage of the highest-scoring snapshot elements to keep.
35
- *
36
- * This value is expressed as a real percent from `1` to `100`.
37
- * For example, `10` keeps the top 10% of scored elements.
38
- *
39
- * @example
40
- * ```ts
41
- * topPercent: 20
42
- * ```
43
- */
44
- topPercent?: number;
45
- }
46
- /**
47
- * Final result returned by a step execution.
48
- *
49
- * @example
50
- * ```ts
51
- * const result: StepResult = {
52
- * passed: true,
53
- * actual: 'Checkout page is visible',
54
- * }
55
- * ```
56
- */
57
- export interface StepResult {
58
- /**
59
- * Whether the step passed.
60
- */
1
+ import type { ChatCompletionMessageParam } from 'openai/resources/chat/completions';
2
+ import type { Diagnostic, Usage } from '../contracts/types.js';
3
+ import type { StepIntent } from '../driver.js';
4
+ import type { ToolExecution } from '../tools/types.js';
5
+ export type Step = Pick<StepIntent, 'action' | 'expect'>;
6
+ export type StepAssertion = {
61
7
  passed: boolean;
62
- /**
63
- * The observed result collected during execution.
64
- */
65
8
  actual: string;
66
- }
67
- export type StepResultPromise = Promise<StepResult>;
68
- /**
69
- * Callback used internally to resolve a running step result.
70
- *
71
- * @example
72
- * ```ts
73
- * const resolve: ResolveStepResult = (result) => {
74
- * console.log(result.actual)
75
- * }
76
- * ```
77
- */
78
- export type ResolveStepResult = (result: StepResult) => void;
9
+ };
10
+ export type InternalTerminationReason = 'met-expectation' | 'failed-expectation' | 'loop-detected' | 'turn-cap-exceeded' | 'step-timeout' | 'scenario-timeout' | 'tool-error' | 'provider-error' | 'token-budget-exceeded' | 'interrupted' | 'internal-error';
11
+ export type InternalStepToolCall = {
12
+ turn: number;
13
+ driverId: string;
14
+ name: string;
15
+ arguments: unknown;
16
+ status: 'ok' | 'error';
17
+ };
18
+ export type InternalStepReport = {
19
+ step: StepIntent;
20
+ outcome: 'passed' | 'failed';
21
+ category: StepCategory;
22
+ reason: InternalTerminationReason;
23
+ actual?: string;
24
+ turns: number;
25
+ durationMs: number;
26
+ usage: Usage;
27
+ toolCalls: InternalStepToolCall[];
28
+ transcript: TranscriptEntry[];
29
+ diagnostics: Diagnostic[];
30
+ };
31
+ export type StepCategory = 'app' | 'model' | 'infra';
32
+ export type TranscriptEntry = {
33
+ turn: number;
34
+ role: 'assistant' | 'tool';
35
+ content: string;
36
+ };
37
+ export type ContextMessage = {
38
+ message: ChatCompletionMessageParam;
39
+ ephemeral?: boolean;
40
+ };
41
+ export type TurnOutcome = {
42
+ kind: 'continue';
43
+ toolResults: ToolExecution[];
44
+ messages: ChatCompletionMessageParam[];
45
+ } | {
46
+ kind: 'assertion';
47
+ passed: boolean;
48
+ actual: string;
49
+ } | {
50
+ kind: 'stuck';
51
+ reason: 'loop-detected';
52
+ };
79
53
  //# sourceMappingURL=types.d.ts.map
@@ -1 +1 @@
1
- {"version":3,"file":"types.d.ts","sourceRoot":"","sources":["../../src/runtime/types.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AACH,MAAM,WAAW,IAAI;IACpB;;OAEG;IACH,MAAM,EAAE,MAAM,CAAA;IAEd;;OAEG;IACH,MAAM,EAAE,MAAM,CAAA;IAEd;;;;OAIG;IACH,MAAM,CAAC,EAAE,MAAM,EAAE,CAAA;IAEjB;;;;;;;;;;OAUG;IACH,UAAU,CAAC,EAAE,MAAM,CAAA;CACnB;AAED;;;;;;;;;;GAUG;AACH,MAAM,WAAW,UAAU;IAC1B;;OAEG;IACH,MAAM,EAAE,OAAO,CAAA;IAEf;;OAEG;IACH,MAAM,EAAE,MAAM,CAAA;CACd;AAED,MAAM,MAAM,iBAAiB,GAAG,OAAO,CAAC,UAAU,CAAC,CAAA;AAEnD;;;;;;;;;GASG;AACH,MAAM,MAAM,iBAAiB,GAAG,CAAC,MAAM,EAAE,UAAU,KAAK,IAAI,CAAA"}
1
+ {"version":3,"file":"types.d.ts","sourceRoot":"","sources":["../../src/runtime/types.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,0BAA0B,EAAE,MAAM,mCAAmC,CAAA;AACnF,OAAO,KAAK,EAAE,UAAU,EAAE,KAAK,EAAE,MAAM,uBAAuB,CAAA;AAC9D,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,cAAc,CAAA;AAC9C,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,mBAAmB,CAAA;AAEtD,MAAM,MAAM,IAAI,GAAG,IAAI,CAAC,UAAU,EAAE,QAAQ,GAAG,QAAQ,CAAC,CAAA;AAExD,MAAM,MAAM,aAAa,GAAG;IAAE,MAAM,EAAE,OAAO,CAAC;IAAC,MAAM,EAAE,MAAM,CAAA;CAAE,CAAA;AAE/D,MAAM,MAAM,yBAAyB,GAClC,iBAAiB,GACjB,oBAAoB,GACpB,eAAe,GACf,mBAAmB,GACnB,cAAc,GACd,kBAAkB,GAClB,YAAY,GACZ,gBAAgB,GAChB,uBAAuB,GACvB,aAAa,GACb,gBAAgB,CAAA;AAEnB,MAAM,MAAM,oBAAoB,GAAG;IAClC,IAAI,EAAE,MAAM,CAAA;IACZ,QAAQ,EAAE,MAAM,CAAA;IAChB,IAAI,EAAE,MAAM,CAAA;IACZ,SAAS,EAAE,OAAO,CAAA;IAClB,MAAM,EAAE,IAAI,GAAG,OAAO,CAAA;CACtB,CAAA;AAED,MAAM,MAAM,kBAAkB,GAAG;IAChC,IAAI,EAAE,UAAU,CAAA;IAChB,OAAO,EAAE,QAAQ,GAAG,QAAQ,CAAA;IAC5B,QAAQ,EAAE,YAAY,CAAA;IACtB,MAAM,EAAE,yBAAyB,CAAA;IACjC,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,KAAK,EAAE,MAAM,CAAA;IACb,UAAU,EAAE,MAAM,CAAA;IAClB,KAAK,EAAE,KAAK,CAAA;IACZ,SAAS,EAAE,oBAAoB,EAAE,CAAA;IACjC,UAAU,EAAE,eAAe,EAAE,CAAA;IAC7B,WAAW,EAAE,UAAU,EAAE,CAAA;CACzB,CAAA;AAED,MAAM,MAAM,YAAY,GAAG,KAAK,GAAG,OAAO,GAAG,OAAO,CAAA;AAEpD,MAAM,MAAM,eAAe,GAAG;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,IAAI,EAAE,WAAW,GAAG,MAAM,CAAC;IAAC,OAAO,EAAE,MAAM,CAAA;CAAE,CAAA;AAE3F,MAAM,MAAM,cAAc,GAAG;IAAE,OAAO,EAAE,0BAA0B,CAAC;IAAC,SAAS,CAAC,EAAE,OAAO,CAAA;CAAE,CAAA;AAEzF,MAAM,MAAM,WAAW,GACpB;IAAE,IAAI,EAAE,UAAU,CAAC;IAAC,WAAW,EAAE,aAAa,EAAE,CAAC;IAAC,QAAQ,EAAE,0BAA0B,EAAE,CAAA;CAAE,GAC1F;IAAE,IAAI,EAAE,WAAW,CAAC;IAAC,MAAM,EAAE,OAAO,CAAC;IAAC,MAAM,EAAE,MAAM,CAAA;CAAE,GACtD;IAAE,IAAI,EAAE,OAAO,CAAC;IAAC,MAAM,EAAE,eAAe,CAAA;CAAE,CAAA"}
@@ -0,0 +1,37 @@
1
+ import type { ChatCompletion } from 'openai/resources/chat/completions';
2
+ import type { Usage } from '../contracts/types.js';
3
+ export type UsageCheckpoint = Readonly<{
4
+ promptTokens: number;
5
+ cachedPromptTokens: number;
6
+ completionTokens: number;
7
+ reportedResponses: number;
8
+ missingResponses: number;
9
+ }>;
10
+ export declare class TokenBudgetExceededError extends Error {
11
+ readonly budgetTokens: number;
12
+ readonly totalTokens: number;
13
+ constructor(budgetTokens: number, totalTokens: number);
14
+ }
15
+ export declare class ProviderUsageUnavailableError extends Error {
16
+ constructor();
17
+ }
18
+ export declare class ProviderUsageInvalidError extends Error {
19
+ constructor(message: string);
20
+ }
21
+ export declare class ScenarioUsageTracker {
22
+ private readonly budgetTokens?;
23
+ private promptTokens;
24
+ private cachedPromptTokens;
25
+ private completionTokens;
26
+ private reportedResponses;
27
+ private missingResponses;
28
+ constructor(budgetTokens?: number);
29
+ beginStep(): UsageCheckpoint;
30
+ record(usage: ChatCompletion['usage']): void;
31
+ stepUsage(checkpoint: UsageCheckpoint): Usage;
32
+ usage(): Usage & {
33
+ state: 'complete' | 'partial' | 'unavailable';
34
+ };
35
+ private checkpoint;
36
+ }
37
+ //# sourceMappingURL=usage-tracker.d.ts.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"usage-tracker.d.ts","sourceRoot":"","sources":["../../src/runtime/usage-tracker.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,mCAAmC,CAAA;AACvE,OAAO,KAAK,EAAE,KAAK,EAAE,MAAM,uBAAuB,CAAA;AAElD,MAAM,MAAM,eAAe,GAAG,QAAQ,CAAC;IACtC,YAAY,EAAE,MAAM,CAAA;IACpB,kBAAkB,EAAE,MAAM,CAAA;IAC1B,gBAAgB,EAAE,MAAM,CAAA;IACxB,iBAAiB,EAAE,MAAM,CAAA;IACzB,gBAAgB,EAAE,MAAM,CAAA;CACxB,CAAC,CAAA;AAEF,qBAAa,wBAAyB,SAAQ,KAAK;IAEjD,QAAQ,CAAC,YAAY,EAAE,MAAM;IAC7B,QAAQ,CAAC,WAAW,EAAE,MAAM;gBADnB,YAAY,EAAE,MAAM,EACpB,WAAW,EAAE,MAAM;CAK7B;AAED,qBAAa,6BAA8B,SAAQ,KAAK;;CAKvD;AAED,qBAAa,yBAA0B,SAAQ,KAAK;gBACvC,OAAO,EAAE,MAAM;CAI3B;AAED,qBAAa,oBAAoB;IAOpB,OAAO,CAAC,QAAQ,CAAC,YAAY,CAAC;IAN1C,OAAO,CAAC,YAAY,CAAI;IACxB,OAAO,CAAC,kBAAkB,CAAI;IAC9B,OAAO,CAAC,gBAAgB,CAAI;IAC5B,OAAO,CAAC,iBAAiB,CAAI;IAC7B,OAAO,CAAC,gBAAgB,CAAI;gBAEC,YAAY,CAAC,EAAE,MAAM;IAMlD,SAAS,IAAI,eAAe;IAI5B,MAAM,CAAC,KAAK,EAAE,cAAc,CAAC,OAAO,CAAC,GAAG,IAAI;IAgC5C,SAAS,CAAC,UAAU,EAAE,eAAe,GAAG,KAAK;IAQ7C,KAAK,IAAI,KAAK,GAAG;QAAE,KAAK,EAAE,UAAU,GAAG,SAAS,GAAG,aAAa,CAAA;KAAE;IAYlE,OAAO,CAAC,UAAU;CASlB"}
@@ -0,0 +1,110 @@
1
+ export class TokenBudgetExceededError extends Error {
2
+ budgetTokens;
3
+ totalTokens;
4
+ constructor(budgetTokens, totalTokens) {
5
+ super(`Scenario token budget of ${budgetTokens} was exceeded by response total ${totalTokens}`);
6
+ this.budgetTokens = budgetTokens;
7
+ this.totalTokens = totalTokens;
8
+ this.name = 'TokenBudgetExceededError';
9
+ }
10
+ }
11
+ export class ProviderUsageUnavailableError extends Error {
12
+ constructor() {
13
+ super('Provider response omitted usage while a scenario token budget is active');
14
+ this.name = 'ProviderUsageUnavailableError';
15
+ }
16
+ }
17
+ export class ProviderUsageInvalidError extends Error {
18
+ constructor(message) {
19
+ super(`Provider response contained invalid usage: ${message}`);
20
+ this.name = 'ProviderUsageInvalidError';
21
+ }
22
+ }
23
+ export class ScenarioUsageTracker {
24
+ budgetTokens;
25
+ promptTokens = 0;
26
+ cachedPromptTokens = 0;
27
+ completionTokens = 0;
28
+ reportedResponses = 0;
29
+ missingResponses = 0;
30
+ constructor(budgetTokens) {
31
+ this.budgetTokens = budgetTokens;
32
+ if (budgetTokens !== undefined && (!Number.isSafeInteger(budgetTokens) || budgetTokens < 1)) {
33
+ throw new Error('Scenario token budget must be a positive integer');
34
+ }
35
+ }
36
+ beginStep() {
37
+ return this.checkpoint();
38
+ }
39
+ record(usage) {
40
+ if (!usage) {
41
+ this.missingResponses++;
42
+ if (this.budgetTokens !== undefined)
43
+ throw new ProviderUsageUnavailableError();
44
+ return;
45
+ }
46
+ const promptTokens = requiredTokenCount(usage.prompt_tokens, 'prompt_tokens');
47
+ const completionTokens = requiredTokenCount(usage.completion_tokens, 'completion_tokens');
48
+ const totalTokens = requiredTokenCount(usage.total_tokens, 'total_tokens');
49
+ if (totalTokens !== promptTokens + completionTokens) {
50
+ throw new ProviderUsageInvalidError('total_tokens must equal prompt_tokens plus completion_tokens');
51
+ }
52
+ const cachedPromptTokens = optionalTokenCount(usage.prompt_tokens_details?.cached_tokens, 'prompt_tokens_details.cached_tokens');
53
+ if (cachedPromptTokens > promptTokens) {
54
+ throw new ProviderUsageInvalidError('cached_tokens cannot exceed prompt_tokens');
55
+ }
56
+ this.promptTokens += promptTokens;
57
+ this.cachedPromptTokens += cachedPromptTokens;
58
+ this.completionTokens += completionTokens;
59
+ this.reportedResponses++;
60
+ const scenarioTotalTokens = this.promptTokens + this.completionTokens;
61
+ if (this.budgetTokens !== undefined && scenarioTotalTokens > this.budgetTokens) {
62
+ throw new TokenBudgetExceededError(this.budgetTokens, scenarioTotalTokens);
63
+ }
64
+ }
65
+ stepUsage(checkpoint) {
66
+ return usageFrom({
67
+ promptTokens: this.promptTokens - checkpoint.promptTokens,
68
+ cachedPromptTokens: this.cachedPromptTokens - checkpoint.cachedPromptTokens,
69
+ completionTokens: this.completionTokens - checkpoint.completionTokens,
70
+ });
71
+ }
72
+ usage() {
73
+ const state = this.reportedResponses === 0 ? 'unavailable' : this.missingResponses > 0 ? 'partial' : 'complete';
74
+ return {
75
+ ...usageFrom({
76
+ promptTokens: this.promptTokens,
77
+ cachedPromptTokens: this.cachedPromptTokens,
78
+ completionTokens: this.completionTokens,
79
+ }),
80
+ state,
81
+ };
82
+ }
83
+ checkpoint() {
84
+ return {
85
+ promptTokens: this.promptTokens,
86
+ cachedPromptTokens: this.cachedPromptTokens,
87
+ completionTokens: this.completionTokens,
88
+ reportedResponses: this.reportedResponses,
89
+ missingResponses: this.missingResponses,
90
+ };
91
+ }
92
+ }
93
+ function requiredTokenCount(value, field) {
94
+ if (!Number.isSafeInteger(value) || value < 0) {
95
+ throw new ProviderUsageInvalidError(`${field} must be a non-negative safe integer`);
96
+ }
97
+ return value;
98
+ }
99
+ function optionalTokenCount(value, field) {
100
+ return value === undefined ? 0 : requiredTokenCount(value, field);
101
+ }
102
+ function usageFrom(input) {
103
+ return {
104
+ promptTokens: input.promptTokens,
105
+ cachedPromptTokens: Math.min(input.cachedPromptTokens, input.promptTokens),
106
+ completionTokens: input.completionTokens,
107
+ totalTokens: input.promptTokens + input.completionTokens,
108
+ };
109
+ }
110
+ //# sourceMappingURL=usage-tracker.js.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"usage-tracker.js","sourceRoot":"","sources":["../../src/runtime/usage-tracker.ts"],"names":[],"mappings":"AAWA,MAAM,OAAO,wBAAyB,SAAQ,KAAK;IAExC;IACA;IAFV,YACU,YAAoB,EACpB,WAAmB;QAE5B,KAAK,CAAC,4BAA4B,YAAY,mCAAmC,WAAW,EAAE,CAAC,CAAA;QAHtF,iBAAY,GAAZ,YAAY,CAAQ;QACpB,gBAAW,GAAX,WAAW,CAAQ;QAG5B,IAAI,CAAC,IAAI,GAAG,0BAA0B,CAAA;IACvC,CAAC;CACD;AAED,MAAM,OAAO,6BAA8B,SAAQ,KAAK;IACvD;QACC,KAAK,CAAC,yEAAyE,CAAC,CAAA;QAChF,IAAI,CAAC,IAAI,GAAG,+BAA+B,CAAA;IAC5C,CAAC;CACD;AAED,MAAM,OAAO,yBAA0B,SAAQ,KAAK;IACnD,YAAY,OAAe;QAC1B,KAAK,CAAC,8CAA8C,OAAO,EAAE,CAAC,CAAA;QAC9D,IAAI,CAAC,IAAI,GAAG,2BAA2B,CAAA;IACxC,CAAC;CACD;AAED,MAAM,OAAO,oBAAoB;IAOH;IANrB,YAAY,GAAG,CAAC,CAAA;IAChB,kBAAkB,GAAG,CAAC,CAAA;IACtB,gBAAgB,GAAG,CAAC,CAAA;IACpB,iBAAiB,GAAG,CAAC,CAAA;IACrB,gBAAgB,GAAG,CAAC,CAAA;IAE5B,YAA6B,YAAqB;QAArB,iBAAY,GAAZ,YAAY,CAAS;QACjD,IAAI,YAAY,KAAK,SAAS,IAAI,CAAC,CAAC,MAAM,CAAC,aAAa,CAAC,YAAY,CAAC,IAAI,YAAY,GAAG,CAAC,CAAC,EAAE,CAAC;YAC7F,MAAM,IAAI,KAAK,CAAC,kDAAkD,CAAC,CAAA;QACpE,CAAC;IACF,CAAC;IAED,SAAS;QACR,OAAO,IAAI,CAAC,UAAU,EAAE,CAAA;IACzB,CAAC;IAED,MAAM,CAAC,KAA8B;QACpC,IAAI,CAAC,KAAK,EAAE,CAAC;YACZ,IAAI,CAAC,gBAAgB,EAAE,CAAA;YACvB,IAAI,IAAI,CAAC,YAAY,KAAK,SAAS;gBAAE,MAAM,IAAI,6BAA6B,EAAE,CAAA;YAC9E,OAAM;QACP,CAAC;QAED,MAAM,YAAY,GAAG,kBAAkB,CAAC,KAAK,CAAC,aAAa,EAAE,eAAe,CAAC,CAAA;QAC7E,MAAM,gBAAgB,GAAG,kBAAkB,CAAC,KAAK,CAAC,iBAAiB,EAAE,mBAAmB,CAAC,CAAA;QACzF,MAAM,WAAW,GAAG,kBAAkB,CAAC,KAAK,CAAC,YAAY,EAAE,cAAc,CAAC,CAAA;QAC1E,IAAI,WAAW,KAAK,YAAY,GAAG,gBAAgB,EAAE,CAAC;YACrD,MAAM,IAAI,yBAAyB,CAAC,8DAA8D,CAAC,CAAA;QACpG,CAAC;QAED,MAAM,kBAAkB,GAAG,kBAAkB,CAC5C,KAAK,CAAC,qBAAqB,EAAE,aAAa,EAC1C,qCAAqC,CACrC,CAAA;QACD,IAAI,kBAAkB,GAAG,YAAY,EAAE,CAAC;YACvC,MAAM,IAAI,yBAAyB,CAAC,2CAA2C,CAAC,CAAA;QACjF,CAAC;QACD,IAAI,CAAC,YAAY,IAAI,YAAY,CAAA;QACjC,IAAI,CAAC,kBAAkB,IAAI,kBAAkB,CAAA;QAC7C,IAAI,CAAC,gBAAgB,IAAI,gBAAgB,CAAA;QACzC,IAAI,CAAC,iBAAiB,EAAE,CAAA;QAExB,MAAM,mBAAmB,GAAG,IAAI,CAAC,YAAY,GAAG,IAAI,CAAC,gBAAgB,CAAA;QACrE,IAAI,IAAI,CAAC,YAAY,KAAK,SAAS,IAAI,mBAAmB,GAAG,IAAI,CAAC,YAAY,EAAE,CAAC;YAChF,MAAM,IAAI,wBAAwB,CAAC,IAAI,CAAC,YAAY,EAAE,mBAAmB,CAAC,CAAA;QAC3E,CAAC;IACF,CAAC;IAED,SAAS,CAAC,UAA2B;QACpC,OAAO,SAAS,CAAC;YAChB,YAAY,EAAE,IAAI,CAAC,YAAY,GAAG,UAAU,CAAC,YAAY;YACzD,kBAAkB,EAAE,IAAI,CAAC,kBAAkB,GAAG,UAAU,CAAC,kBAAkB;YAC3E,gBAAgB,EAAE,IAAI,CAAC,gBAAgB,GAAG,UAAU,CAAC,gBAAgB;SACrE,CAAC,CAAA;IACH,CAAC;IAED,KAAK;QACJ,MAAM,KAAK,GAAG,IAAI,CAAC,iBAAiB,KAAK,CAAC,CAAC,CAAC,CAAC,aAAa,CAAC,CAAC,CAAC,IAAI,CAAC,gBAAgB,GAAG,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,UAAU,CAAA;QAC/G,OAAO;YACN,GAAG,SAAS,CAAC;gBACZ,YAAY,EAAE,IAAI,CAAC,YAAY;gBAC/B,kBAAkB,EAAE,IAAI,CAAC,kBAAkB;gBAC3C,gBAAgB,EAAE,IAAI,CAAC,gBAAgB;aACvC,CAAC;YACF,KAAK;SACL,CAAA;IACF,CAAC;IAEO,UAAU;QACjB,OAAO;YACN,YAAY,EAAE,IAAI,CAAC,YAAY;YAC/B,kBAAkB,EAAE,IAAI,CAAC,kBAAkB;YAC3C,gBAAgB,EAAE,IAAI,CAAC,gBAAgB;YACvC,iBAAiB,EAAE,IAAI,CAAC,iBAAiB;YACzC,gBAAgB,EAAE,IAAI,CAAC,gBAAgB;SACvC,CAAA;IACF,CAAC;CACD;AAED,SAAS,kBAAkB,CAAC,KAAc,EAAE,KAAa;IACxD,IAAI,CAAC,MAAM,CAAC,aAAa,CAAC,KAAK,CAAC,IAAK,KAAgB,GAAG,CAAC,EAAE,CAAC;QAC3D,MAAM,IAAI,yBAAyB,CAAC,GAAG,KAAK,sCAAsC,CAAC,CAAA;IACpF,CAAC;IACD,OAAO,KAAe,CAAA;AACvB,CAAC;AAED,SAAS,kBAAkB,CAAC,KAAc,EAAE,KAAa;IACxD,OAAO,KAAK,KAAK,SAAS,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,kBAAkB,CAAC,KAAK,EAAE,KAAK,CAAC,CAAA;AAClE,CAAC;AAED,SAAS,SAAS,CAAC,KAAqF;IACvG,OAAO;QACN,YAAY,EAAE,KAAK,CAAC,YAAY;QAChC,kBAAkB,EAAE,IAAI,CAAC,GAAG,CAAC,KAAK,CAAC,kBAAkB,EAAE,KAAK,CAAC,YAAY,CAAC;QAC1E,gBAAgB,EAAE,KAAK,CAAC,gBAAgB;QACxC,WAAW,EAAE,KAAK,CAAC,YAAY,GAAG,KAAK,CAAC,gBAAgB;KACxD,CAAA;AACF,CAAC"}