humanish 0.106.1 → 0.107.0

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 (453) hide show
  1. package/AGENTS.md +4 -4
  2. package/ARCHITECTURE.md +17 -16
  3. package/CHANGELOG.md +14 -0
  4. package/CONTEXT.md +42 -39
  5. package/CONTRIBUTING.md +8 -8
  6. package/README.md +8 -3
  7. package/dist/actors/codex/app-server-client.js +2 -1
  8. package/dist/actors/codex/app-server-trace.d.ts +0 -1
  9. package/dist/actors/codex/app-server-trace.js +1 -3
  10. package/dist/actors/codex/app-server.js +2 -1
  11. package/dist/actors/codex/protocol-compat.d.ts +27 -0
  12. package/dist/actors/codex/protocol-compat.js +357 -0
  13. package/dist/actors/codex/protocol-contract.d.ts +61 -0
  14. package/dist/actors/codex/protocol-contract.js +471 -0
  15. package/dist/actors/codex/qualified-versions.d.ts +30 -3
  16. package/dist/actors/codex/qualified-versions.js +50 -5
  17. package/dist/actors/codex/restricted-executable.d.ts +24 -0
  18. package/dist/actors/codex/restricted-executable.js +97 -22
  19. package/dist/actors/codex/restricted-launch.d.ts +7 -1
  20. package/dist/actors/codex/restricted-launch.js +42 -3
  21. package/dist/actors/codex/restricted-notifications.d.ts +101 -0
  22. package/dist/actors/codex/restricted-notifications.js +277 -0
  23. package/dist/actors/codex/restricted-participant.d.ts +30 -0
  24. package/dist/actors/codex/restricted-participant.js +60 -2
  25. package/dist/actors/codex/restricted-policy.d.ts +9 -1
  26. package/dist/actors/codex/restricted-session.d.ts +14 -1
  27. package/dist/actors/codex/restricted-session.js +50 -17
  28. package/dist/actors/codex/restricted-transport.d.ts +29 -0
  29. package/dist/actors/codex/restricted-transport.js +139 -27
  30. package/dist/actors/codex/restricted-turn.d.ts +19 -4
  31. package/dist/actors/codex/restricted-turn.js +43 -30
  32. package/dist/actors/computer-use/actor.d.ts +1 -1
  33. package/dist/actors/computer-use/actor.js +4 -4
  34. package/dist/actors/computer-use/loop/backstop.js +1 -1
  35. package/dist/actors/computer-use/loop/ending.d.ts +2 -2
  36. package/dist/actors/computer-use/loop/ending.js +2 -2
  37. package/dist/actors/computer-use/loop/spend.d.ts +1 -1
  38. package/dist/actors/computer-use/loop/spend.js +3 -3
  39. package/dist/actors/computer-use/loop/trace.js +2 -2
  40. package/dist/actors/computer-use/loop/types.d.ts +14 -5
  41. package/dist/actors/computer-use/loop.js +4 -4
  42. package/dist/actors/computer-use/openai-provider.d.ts +1 -1
  43. package/dist/actors/computer-use/openai-provider.js +2 -2
  44. package/dist/actors/contract.d.ts +9 -7
  45. package/dist/actors/local-agent/cli.js +1 -1
  46. package/dist/actors/scripted-browser/actor.d.ts +2 -2
  47. package/dist/actors/scripted-browser/journey.js +15 -4
  48. package/dist/actors/scripted-browser/steps.js +18 -0
  49. package/dist/actors/scripted-browser/types.d.ts +11 -1
  50. package/dist/actors/stop-conditions.js +1 -3
  51. package/dist/analysis/automatic-completion.d.ts +13 -8
  52. package/dist/analysis/automatic-completion.js +14 -8
  53. package/dist/analysis/automatic-config.d.ts +1 -1
  54. package/dist/analysis/automatic-config.js +2 -2
  55. package/dist/analysis/automatic.d.ts +10 -4
  56. package/dist/analysis/automatic.js +17 -13
  57. package/dist/analysis/diagnostics.d.ts +23 -0
  58. package/dist/analysis/diagnostics.js +112 -0
  59. package/dist/analysis/evidence-sources.js +2 -2
  60. package/dist/analysis/evidence.js +6 -6
  61. package/dist/analysis/execute.d.ts +7 -1
  62. package/dist/analysis/execute.js +39 -12
  63. package/dist/analysis/job.d.ts +8 -5
  64. package/dist/analysis/job.js +15 -14
  65. package/dist/analysis/load.d.ts +1 -1
  66. package/dist/analysis/load.js +5 -5
  67. package/dist/analysis/provider.d.ts +9 -1
  68. package/dist/analysis/restricted-codex.d.ts +2 -0
  69. package/dist/analysis/restricted-codex.js +10 -1
  70. package/dist/analysis/service.d.ts +2 -0
  71. package/dist/analysis/service.js +31 -12
  72. package/dist/analysis/sharing.d.ts +2 -2
  73. package/dist/analysis/sharing.js +5 -5
  74. package/dist/analysis/store-executions.d.ts +5 -5
  75. package/dist/analysis/store-executions.js +12 -12
  76. package/dist/analysis/store.d.ts +4 -4
  77. package/dist/analysis/store.js +13 -13
  78. package/dist/cli/commands/analysis-signals.d.ts +2 -1
  79. package/dist/cli/commands/analysis-signals.js +4 -1
  80. package/dist/cli/commands/analyze.js +4 -2
  81. package/dist/cli/commands/comms.js +8 -1
  82. package/dist/cli/commands/lab-route-computer-use.d.ts +0 -2
  83. package/dist/cli/commands/lab-route-computer-use.js +26 -14
  84. package/dist/cli/commands/lab-route-preview.d.ts +1 -3
  85. package/dist/cli/commands/lab-route-preview.js +3 -4
  86. package/dist/cli/commands/lab-route-run.d.ts +6 -4
  87. package/dist/cli/commands/lab-route-run.js +6 -2
  88. package/dist/cli/commands/lab-route-scripted.d.ts +0 -2
  89. package/dist/cli/commands/lab-route-scripted.js +0 -1
  90. package/dist/cli/commands/lab-route-shared-world.d.ts +0 -2
  91. package/dist/cli/commands/lab-route-shared-world.js +0 -1
  92. package/dist/cli/commands/lab-route-terminal.d.ts +0 -2
  93. package/dist/cli/commands/lab-route-terminal.js +0 -1
  94. package/dist/cli/commands/lab-run.js +76 -45
  95. package/dist/cli/commands/lab-scorer.d.ts +0 -2
  96. package/dist/cli/commands/lab.js +12 -6
  97. package/dist/cli/commands/observe.js +22 -0
  98. package/dist/cli/commands/run-signals.d.ts +30 -0
  99. package/dist/cli/commands/run-signals.js +146 -0
  100. package/dist/cli/commands/runs.js +10 -8
  101. package/dist/cli/commands/watch.js +12 -10
  102. package/dist/cli/doctor.d.ts +3 -0
  103. package/dist/cli/doctor.js +10 -2
  104. package/dist/cli/io.d.ts +5 -4
  105. package/dist/cli/io.js +12 -8
  106. package/dist/cli/program.js +3 -3
  107. package/dist/cli/renamed-options.d.ts +11 -0
  108. package/dist/cli/renamed-options.js +18 -0
  109. package/dist/cli/telemetry.d.ts +2 -2
  110. package/dist/cli/telemetry.js +1 -1
  111. package/dist/comms/capture-surface.d.ts +1 -1
  112. package/dist/comms/catch-host.d.ts +1 -1
  113. package/dist/comms/connections.js +4 -4
  114. package/dist/comms/external-evidence.d.ts +27 -0
  115. package/dist/comms/external-evidence.js +93 -0
  116. package/dist/comms/lease-store.js +7 -7
  117. package/dist/comms/types.js +1 -1
  118. package/dist/evidence/browser-hygiene.d.ts +1 -1
  119. package/dist/evidence/browser-hygiene.js +1 -1
  120. package/dist/evidence/redaction.js +1 -1
  121. package/dist/evidence/secret-scrub.d.ts +9 -0
  122. package/dist/evidence/secret-scrub.js +96 -0
  123. package/dist/feedback/draft.js +7 -9
  124. package/dist/feedback/export-bundle.js +4 -4
  125. package/dist/feedback/export.js +9 -9
  126. package/dist/feedback/feedback.d.ts +1 -1
  127. package/dist/index.d.ts +3 -6
  128. package/dist/index.js +3 -6
  129. package/dist/keys/key-resolution.d.ts +3 -1
  130. package/dist/keys/key-resolution.js +18 -16
  131. package/dist/lab/adapter-extension.d.ts +28 -6
  132. package/dist/lab/adapter-extension.js +65 -126
  133. package/dist/lab/adapter-scorer-loader.d.ts +25 -5
  134. package/dist/lab/adapter-scorer-loader.js +24 -0
  135. package/dist/lab/composition-rules.js +20 -20
  136. package/dist/lab/config.js +28 -26
  137. package/dist/lab/device-presets.d.ts +2 -2
  138. package/dist/lab/discover.js +1 -3
  139. package/dist/lab/doctor.d.ts +44 -9
  140. package/dist/lab/doctor.js +149 -41
  141. package/dist/lab/init-plan.js +1 -3
  142. package/dist/lab/keys.js +1 -1
  143. package/dist/lab/lab-deps.d.ts +85 -0
  144. package/dist/lab/lab-deps.js +4 -0
  145. package/dist/lab/parse/actors.js +8 -7
  146. package/dist/lab/parse/comms.js +3 -2
  147. package/dist/lab/parse/execution.js +3 -2
  148. package/dist/lab/parse/subject-state.js +2 -1
  149. package/dist/lab/parse/subject.js +2 -1
  150. package/dist/lab/parse/values.d.ts +0 -1
  151. package/dist/lab/parse/values.js +0 -3
  152. package/dist/lab/persona-resolve.d.ts +3 -3
  153. package/dist/lab/persona-resolve.js +4 -4
  154. package/dist/lab/persona.js +1 -3
  155. package/dist/lab/plan-base.d.ts +0 -2
  156. package/dist/lab/plan-base.js +6 -4
  157. package/dist/lab/plan-participants.d.ts +5 -5
  158. package/dist/lab/plan-participants.js +2 -2
  159. package/dist/lab/plan-types.d.ts +22 -4
  160. package/dist/lab/plan.d.ts +3 -11
  161. package/dist/lab/plan.js +18 -42
  162. package/dist/lab/preflight-result.js +0 -2
  163. package/dist/lab/preflight.d.ts +0 -3
  164. package/dist/lab/preflight.js +1 -1
  165. package/dist/lab/requirements.d.ts +20 -0
  166. package/dist/lab/requirements.js +54 -0
  167. package/dist/lab/route-inputs.d.ts +23 -12
  168. package/dist/lab/route-inputs.js +48 -51
  169. package/dist/lab/routing.d.ts +19 -24
  170. package/dist/lab/routing.js +14 -47
  171. package/dist/lab/run-lab-events.d.ts +6 -11
  172. package/dist/lab/run-lab-events.js +3 -11
  173. package/dist/lab/run-lab-homes.d.ts +64 -0
  174. package/dist/lab/run-lab-homes.js +3 -0
  175. package/dist/lab/run-lab-options.d.ts +12 -66
  176. package/dist/lab/run-lab-options.js +45 -275
  177. package/dist/lab/summary.d.ts +9 -4
  178. package/dist/lab/summary.js +15 -8
  179. package/dist/lab/types.d.ts +132 -132
  180. package/dist/lab/validation.d.ts +15 -15
  181. package/dist/lab/validation.js +28 -28
  182. package/dist/lab/warnings.js +20 -19
  183. package/dist/observer/artifact.d.ts +1 -1
  184. package/dist/observer/artifact.js +3 -3
  185. package/dist/observer/data.js +2 -2
  186. package/dist/observer/render.js +2 -2
  187. package/dist/observer/run-routes.d.ts +1 -1
  188. package/dist/observer/run-routes.js +5 -5
  189. package/dist/observer/serve.d.ts +17 -0
  190. package/dist/observer/serve.js +34 -3
  191. package/dist/routes/computer-use/bundle-parts.d.ts +1 -1
  192. package/dist/routes/computer-use/bundle-parts.js +10 -3
  193. package/dist/routes/computer-use/bundle.d.ts +4 -11
  194. package/dist/routes/computer-use/bundle.js +9 -15
  195. package/dist/routes/computer-use/diagnostics.d.ts +1 -1
  196. package/dist/routes/computer-use/diagnostics.js +1 -1
  197. package/dist/routes/computer-use/e2b-desktop/comms.d.ts +10 -10
  198. package/dist/routes/computer-use/e2b-desktop/comms.js +11 -14
  199. package/dist/routes/computer-use/e2b-desktop/desktop.d.ts +2 -2
  200. package/dist/routes/computer-use/e2b-desktop/desktop.js +4 -7
  201. package/dist/routes/computer-use/e2b-desktop/fidelity.d.ts +13 -13
  202. package/dist/routes/computer-use/e2b-desktop/fidelity.js +11 -11
  203. package/dist/routes/computer-use/e2b-desktop/prepare.d.ts +3 -3
  204. package/dist/routes/computer-use/e2b-desktop/prepare.js +33 -36
  205. package/dist/routes/computer-use/e2b-desktop/start.d.ts +1 -1
  206. package/dist/routes/computer-use/e2b-desktop/start.js +24 -23
  207. package/dist/routes/computer-use/e2b-desktop/state.d.ts +5 -4
  208. package/dist/routes/computer-use/e2b-desktop/state.js +5 -3
  209. package/dist/routes/computer-use/e2b-desktop/teardown.d.ts +1 -1
  210. package/dist/routes/computer-use/e2b-desktop/teardown.js +29 -17
  211. package/dist/routes/computer-use/external-comms.d.ts +4 -4
  212. package/dist/routes/computer-use/external-comms.js +19 -43
  213. package/dist/routes/computer-use/fanout-bundle.d.ts +3 -3
  214. package/dist/routes/computer-use/fanout-bundle.js +25 -29
  215. package/dist/routes/computer-use/{fanout-lanes.js → fanout-records.js} +25 -19
  216. package/dist/routes/computer-use/in-process-desktop.d.ts +3 -10
  217. package/dist/routes/computer-use/in-process-desktop.js +3 -9
  218. package/dist/routes/computer-use/live-flush.d.ts +7 -5
  219. package/dist/routes/computer-use/live-flush.js +19 -27
  220. package/dist/routes/computer-use/{run-lanes.js → live-phase.js} +12 -11
  221. package/dist/routes/computer-use/local-tree-pack.d.ts +3 -3
  222. package/dist/routes/computer-use/local-tree-pack.js +4 -5
  223. package/dist/routes/computer-use/local-vm.d.ts +11 -8
  224. package/dist/routes/computer-use/local-vm.js +39 -66
  225. package/dist/routes/computer-use/participant-desktop.d.ts +8 -23
  226. package/dist/routes/computer-use/participant-desktop.js +5 -25
  227. package/dist/routes/computer-use/{lanes.d.ts → participant-execution.d.ts} +10 -10
  228. package/dist/routes/computer-use/{lanes.js → participant-execution.js} +35 -22
  229. package/dist/routes/computer-use/participant-facts.d.ts +7 -1
  230. package/dist/routes/computer-use/participant-facts.js +12 -1
  231. package/dist/routes/computer-use/participant-model.d.ts +19 -11
  232. package/dist/routes/computer-use/participant-model.js +49 -25
  233. package/dist/routes/computer-use/participant-prompt.d.ts +3 -3
  234. package/dist/routes/computer-use/participant-prompt.js +4 -4
  235. package/dist/routes/computer-use/participant-runs.d.ts +6 -6
  236. package/dist/routes/computer-use/participant-runs.js +8 -8
  237. package/dist/routes/computer-use/plan.d.ts +21 -29
  238. package/dist/routes/computer-use/plan.js +52 -60
  239. package/dist/routes/computer-use/preflight.d.ts +3 -5
  240. package/dist/routes/computer-use/preflight.js +20 -16
  241. package/dist/routes/computer-use/rerun-selection.js +10 -6
  242. package/dist/routes/computer-use/result.d.ts +1 -1
  243. package/dist/routes/computer-use/result.js +40 -30
  244. package/dist/routes/computer-use/route.d.ts +3 -4
  245. package/dist/routes/computer-use/route.js +21 -19
  246. package/dist/routes/computer-use/self-report.d.ts +2 -2
  247. package/dist/routes/computer-use/self-report.js +4 -4
  248. package/dist/routes/computer-use/setup.d.ts +8 -14
  249. package/dist/routes/computer-use/setup.js +96 -72
  250. package/dist/routes/computer-use/single-bundle.d.ts +8 -11
  251. package/dist/routes/computer-use/single-bundle.js +20 -25
  252. package/dist/routes/computer-use/subject-projection.d.ts +5 -7
  253. package/dist/routes/computer-use/subject-projection.js +20 -19
  254. package/dist/routes/computer-use/types.d.ts +169 -210
  255. package/dist/routes/computer-use/types.js +13 -5
  256. package/dist/routes/preview.js +0 -1
  257. package/dist/routes/scripted/bundle.d.ts +3 -6
  258. package/dist/routes/scripted/bundle.js +11 -16
  259. package/dist/routes/scripted/plan.d.ts +5 -7
  260. package/dist/routes/scripted/plan.js +6 -4
  261. package/dist/routes/scripted/result.d.ts +4 -1
  262. package/dist/routes/scripted/result.js +10 -6
  263. package/dist/routes/scripted/route.js +10 -11
  264. package/dist/routes/scripted/scenario.d.ts +1 -1
  265. package/dist/routes/scripted/scenario.js +1 -1
  266. package/dist/routes/scripted/setup.d.ts +3 -2
  267. package/dist/routes/scripted/setup.js +22 -21
  268. package/dist/routes/scripted/subject.d.ts +8 -3
  269. package/dist/routes/scripted/subject.js +9 -5
  270. package/dist/routes/scripted/surface-records.js +1 -1
  271. package/dist/routes/scripted/surface-sessions.d.ts +2 -2
  272. package/dist/routes/scripted/surface-sessions.js +4 -4
  273. package/dist/routes/scripted/types.d.ts +19 -36
  274. package/dist/routes/scripted/types.js +1 -2
  275. package/dist/routes/shared-world/{seat-records.js → bundle-records.js} +8 -8
  276. package/dist/routes/shared-world/bundle.d.ts +2 -2
  277. package/dist/routes/shared-world/bundle.js +22 -26
  278. package/dist/routes/shared-world/checkpoints.d.ts +4 -4
  279. package/dist/routes/shared-world/checkpoints.js +4 -4
  280. package/dist/routes/shared-world/comms.d.ts +5 -4
  281. package/dist/routes/shared-world/comms.js +28 -41
  282. package/dist/routes/shared-world/external-public.d.ts +5 -5
  283. package/dist/routes/shared-world/external-public.js +19 -20
  284. package/dist/routes/shared-world/handoff.d.ts +4 -4
  285. package/dist/routes/shared-world/handoff.js +50 -50
  286. package/dist/routes/shared-world/lobby-code.d.ts +5 -5
  287. package/dist/routes/shared-world/lobby-code.js +13 -13
  288. package/dist/routes/shared-world/{seats.d.ts → participant-specs.d.ts} +8 -6
  289. package/dist/routes/shared-world/{seats.js → participant-specs.js} +43 -48
  290. package/dist/routes/shared-world/plan.d.ts +2 -6
  291. package/dist/routes/shared-world/plan.js +1 -2
  292. package/dist/routes/shared-world/provenance.d.ts +4 -4
  293. package/dist/routes/shared-world/provenance.js +4 -4
  294. package/dist/routes/shared-world/provisioned.d.ts +3 -2
  295. package/dist/routes/shared-world/provisioned.js +46 -40
  296. package/dist/routes/shared-world/result.js +31 -19
  297. package/dist/routes/shared-world/route.js +37 -41
  298. package/dist/routes/shared-world/setup.d.ts +5 -4
  299. package/dist/routes/shared-world/setup.js +12 -14
  300. package/dist/routes/shared-world/types.d.ts +53 -126
  301. package/dist/routes/shared-world/types.js +2 -2
  302. package/dist/routes/terminal/adapter.d.ts +8 -8
  303. package/dist/routes/terminal/adapter.js +28 -125
  304. package/dist/routes/terminal/bundle.d.ts +7 -12
  305. package/dist/routes/terminal/bundle.js +24 -34
  306. package/dist/routes/terminal/credentials.d.ts +2 -2
  307. package/dist/routes/terminal/credentials.js +16 -16
  308. package/dist/routes/terminal/dry-run.d.ts +1 -1
  309. package/dist/routes/terminal/dry-run.js +9 -11
  310. package/dist/routes/terminal/encoding.d.ts +1 -1
  311. package/dist/routes/terminal/encoding.js +2 -2
  312. package/dist/routes/terminal/ledger.d.ts +6 -6
  313. package/dist/routes/terminal/ledger.js +9 -9
  314. package/dist/routes/terminal/live-finish.d.ts +1 -2
  315. package/dist/routes/terminal/live-finish.js +21 -20
  316. package/dist/routes/terminal/live-sandbox.d.ts +3 -2
  317. package/dist/routes/terminal/live-sandbox.js +20 -20
  318. package/dist/routes/terminal/participant-text.d.ts +23 -0
  319. package/dist/routes/terminal/participant-text.js +165 -0
  320. package/dist/routes/terminal/plan.d.ts +2 -6
  321. package/dist/routes/terminal/plan.js +13 -9
  322. package/dist/routes/terminal/recorder.d.ts +6 -1
  323. package/dist/routes/terminal/recorder.js +12 -4
  324. package/dist/routes/terminal/result.js +4 -4
  325. package/dist/routes/terminal/route.js +26 -26
  326. package/dist/routes/terminal/runtime.d.ts +3 -2
  327. package/dist/routes/terminal/runtime.js +3 -3
  328. package/dist/routes/terminal/sandbox.d.ts +3 -3
  329. package/dist/routes/terminal/sandbox.js +9 -9
  330. package/dist/routes/terminal/session.js +16 -14
  331. package/dist/routes/terminal/token-usage.js +5 -5
  332. package/dist/routes/terminal/trace.d.ts +5 -8
  333. package/dist/routes/terminal/trace.js +9 -64
  334. package/dist/routes/terminal/types.d.ts +85 -93
  335. package/dist/routes/terminal/types.js +9 -4
  336. package/dist/run/active-runs.d.ts +13 -0
  337. package/dist/run/active-runs.js +12 -0
  338. package/dist/run/bundle-shape.d.ts +2 -1
  339. package/dist/run/bundle-shape.js +4 -4
  340. package/dist/run/bundle.d.ts +63 -56
  341. package/dist/run/bundle.js +9 -8
  342. package/dist/run/cost-summary.js +3 -2
  343. package/dist/run/costs.d.ts +6 -6
  344. package/dist/run/costs.js +10 -10
  345. package/dist/run/detail.d.ts +5 -5
  346. package/dist/run/detail.js +6 -6
  347. package/dist/run/dry-run.js +5 -12
  348. package/dist/run/{study-files.d.ts → evidence-files.d.ts} +4 -4
  349. package/dist/run/{study-files.js → evidence-files.js} +5 -5
  350. package/dist/run/feedback-shape.js +1 -1
  351. package/dist/run/judge.d.ts +16 -4
  352. package/dist/run/judge.js +24 -7
  353. package/dist/run/locate.d.ts +2 -0
  354. package/dist/run/locate.js +3 -1
  355. package/dist/run/outcomes.d.ts +3 -3
  356. package/dist/run/outcomes.js +3 -3
  357. package/dist/run/participant-records.d.ts +2 -2
  358. package/dist/run/participant-records.js +2 -2
  359. package/dist/run/paths.js +1 -1
  360. package/dist/run/pricing.d.ts +19 -9
  361. package/dist/run/pricing.js +30 -18
  362. package/dist/run/projection.d.ts +13 -13
  363. package/dist/run/projection.js +8 -8
  364. package/dist/run/reclaim.d.ts +6 -0
  365. package/dist/run/reclaim.js +13 -3
  366. package/dist/run/results.d.ts +1 -1
  367. package/dist/run/run-index.d.ts +1 -1
  368. package/dist/run/run-index.js +11 -11
  369. package/dist/run/run.d.ts +4 -2
  370. package/dist/run/run.js +9 -5
  371. package/dist/run/sandbox-receipts.d.ts +4 -4
  372. package/dist/run/sandbox-receipts.js +4 -4
  373. package/dist/run/shared-world-evidence.d.ts +47 -47
  374. package/dist/run/shared-world-shape.d.ts +3 -3
  375. package/dist/run/shared-world-shape.js +11 -11
  376. package/dist/run/stats.d.ts +5 -5
  377. package/dist/run/stats.js +8 -8
  378. package/dist/run/status.d.ts +16 -6
  379. package/dist/run/status.js +31 -8
  380. package/dist/run/streams.d.ts +26 -19
  381. package/dist/run/terminal-contract.d.ts +5 -5
  382. package/dist/run/terminal-contract.js +4 -4
  383. package/dist/run/transient-comms-secrets.d.ts +5 -0
  384. package/dist/run/transient-comms-secrets.js +12 -0
  385. package/dist/run-lab.d.ts +36 -34
  386. package/dist/run-lab.js +41 -37
  387. package/dist/subject/desktop-cli.d.ts +1 -1
  388. package/dist/subject/desktop-cli.js +1 -1
  389. package/dist/subject/local-tree-archive.d.ts +1 -1
  390. package/dist/subject/local-tree-archive.js +4 -4
  391. package/dist/subject/local-tree.d.ts +2 -2
  392. package/dist/subject/local-tree.js +1 -1
  393. package/dist/subject/node-bootstrap.d.ts +1 -1
  394. package/dist/subject/node-bootstrap.js +4 -4
  395. package/dist/subject/runtime.js +5 -5
  396. package/dist/subject/serve.js +1 -1
  397. package/dist/subject/steps.d.ts +4 -4
  398. package/dist/subject/steps.js +5 -5
  399. package/dist/substrates/e2b/cdp-probe.d.ts +2 -2
  400. package/dist/substrates/e2b/cdp-probe.js +1 -1
  401. package/dist/substrates/e2b/desktop-browser.d.ts +3 -3
  402. package/dist/substrates/e2b/desktop-browser.js +4 -4
  403. package/dist/substrates/e2b/desktop-cdp.d.ts +7 -7
  404. package/dist/substrates/e2b/desktop-cdp.js +4 -4
  405. package/dist/substrates/e2b/desktop-geometry.d.ts +4 -4
  406. package/dist/substrates/e2b/desktop-geometry.js +3 -3
  407. package/dist/substrates/e2b/desktop-media.js +2 -2
  408. package/dist/substrates/e2b/sandbox.js +1 -1
  409. package/dist/substrates/e2b/sdk.js +2 -2
  410. package/dist/substrates/local/runtime-host.d.ts +2 -0
  411. package/dist/substrates/local/runtime-host.js +6 -0
  412. package/dist/tui/actions.d.ts +7 -5
  413. package/dist/tui/actions.js +18 -9
  414. package/dist/tui-app.js +42 -42
  415. package/dist/verify/actor.d.ts +10 -10
  416. package/dist/verify/actor.js +15 -15
  417. package/dist/verify/artifacts.d.ts +1 -1
  418. package/dist/verify/artifacts.js +3 -3
  419. package/dist/verify/costs.d.ts +2 -2
  420. package/dist/verify/costs.js +2 -2
  421. package/dist/verify/liveness.d.ts +1 -1
  422. package/dist/verify/liveness.js +4 -1
  423. package/dist/verify/rerun.d.ts +1 -1
  424. package/dist/verify/rerun.js +1 -1
  425. package/dist/verify/shared-world.js +2 -2
  426. package/dist/verify/verify.js +6 -6
  427. package/docs/README.md +1 -1
  428. package/docs/architecture/actor-contract.md +60 -22
  429. package/docs/architecture/desktop-sessions.md +9 -9
  430. package/docs/architecture/external-public-shared-world.md +5 -0
  431. package/docs/architecture/observer.md +4 -4
  432. package/docs/architecture/restricted-codex-analysis.md +36 -2
  433. package/docs/architecture/state-driven-executor.md +4 -5
  434. package/docs/architecture/task-protocol-support.md +3 -5
  435. package/docs/architecture/{terminal-product-lane.md → terminal-product-route.md} +24 -16
  436. package/docs/contracts/run-bundle.md +10 -8
  437. package/docs/contracts/schemas.md +109 -53
  438. package/docs/contracts/study-analysis.md +24 -0
  439. package/docs/decisions/0004-participant-desktops-stay-in-the-route.md +2 -1
  440. package/docs/product/automatic-analysis.md +1 -0
  441. package/docs/ramp/README.md +2 -1
  442. package/package.json +4 -4
  443. package/skills/humanish/SKILL.md +3 -2
  444. package/skills/humanish/agents/openai.yaml +1 -1
  445. package/dist/deprecated.d.ts +0 -90
  446. package/dist/deprecated.js +0 -106
  447. package/dist/lab/bag-overrides.d.ts +0 -18
  448. package/dist/lab/bag-overrides.js +0 -121
  449. package/dist/routes/computer-use/legacy-lane-spec.d.ts +0 -62
  450. package/dist/routes/computer-use/legacy-lane-spec.js +0 -39
  451. /package/dist/routes/computer-use/{fanout-lanes.d.ts → fanout-records.d.ts} +0 -0
  452. /package/dist/routes/computer-use/{run-lanes.d.ts → live-phase.d.ts} +0 -0
  453. /package/dist/routes/shared-world/{seat-records.d.ts → bundle-records.d.ts} +0 -0
@@ -1,13 +1,14 @@
1
- import { adapterScoreFailureMessage, applyBrowserAdapterHooks, } from "../../lab/adapter-extension.js";
1
+ import { adapterScoreFailureMessage, applyBrowserScorer } from "../../lab/adapter-extension.js";
2
2
  import { redactText } from "../../evidence/redaction.js";
3
3
  import { round6 } from "../../run/pricing.js";
4
4
  import { foldScorerFailures, judgeExecution, OUTCOME_POLICIES, participantHarnessFailed, resultOk, } from "../../run/judge.js";
5
- import { participantFactsOf, participantOutcomeOk } from "./participant-facts.js";
5
+ import { participantFactsOf, participantOutcomeOk, unreleasedSandboxFailures, } from "./participant-facts.js";
6
6
  import { summarizeCuaDiagnostics } from "./diagnostics.js";
7
- import { toParticipantResult } from "./lanes.js";
7
+ import { toParticipantResult } from "./participant-execution.js";
8
8
  import { buildCuaRunBundle, judgeComputerUseRun } from "./bundle.js";
9
9
  import { aggregateCuaSubject, projectParticipantSubjects, subjectProvenanceArg, } from "./subject-projection.js";
10
- import { CUA_ACTOR_LAB_SCHEMA, CUA_FANOUT_STRATEGY, } from "./types.js";
10
+ import { CUA_ACTOR_LAB_SCHEMA, CUA_FANOUT_STRATEGY, participantSubjectEnv, } from "./types.js";
11
+ import { plannedAppUrl } from "./plan.js";
11
12
  /** Aggregate participant counts for the result projection. */
12
13
  function buildParticipantSummary(outcomes, participantCount, participantPlan, dryRun) {
13
14
  if (dryRun || !outcomes) {
@@ -51,13 +52,14 @@ function buildParticipantSummary(outcomes, participantCount, participantPlan, dr
51
52
  }
52
53
  /**
53
54
  * The computer-use lab result for a finished run. The run passes only when the Observer rendered,
54
- * every lane passed (dry-run lanes pass as contracts), and no adapter score or declared scorer
55
+ * every participant passed (dry-run participants pass as contracts), and no adapter score or declared scorer
55
56
  * verdict failed; otherwise the error names the first reason.
56
57
  */
57
58
  function cuaLabResult(args) {
58
59
  const { labId, cwd, runId, appUrl, dryRun, participantRuns, outcomes, subjects, aggregateSubject, participantPlan, rerunLineage, bundle, observer, receivingWarnings, aggregateWarnings, adapterWarnings, } = args;
59
60
  const participantCount = participantRuns.length;
60
- // Lane-level pass: dry-run lanes are contract-ok; live lanes need a passed, engaged session.
61
+ // Participant-level pass: dry-run participants are contract-ok; live ones need a passed,
62
+ // engaged session.
61
63
  const participantOk = (outcome) => participantOutcomeOk(outcome, dryRun);
62
64
  const adapterFailure = adapterScoreFailureMessage(bundle);
63
65
  const ok = resultOk({
@@ -81,10 +83,7 @@ function cuaLabResult(args) {
81
83
  if (ok)
82
84
  return undefined;
83
85
  if (adapterFailure !== undefined) {
84
- return {
85
- code: "HUMANISH_CUA_LAB_FAILED",
86
- message: adapterFailure,
87
- };
86
+ return { code: "HUMANISH_CUA_LAB_FAILED", message: adapterFailure };
88
87
  }
89
88
  if (participantCount === 1) {
90
89
  const outcome = firstOutcome;
@@ -92,9 +91,10 @@ function cuaLabResult(args) {
92
91
  code: outcome?.failureCode ?? "HUMANISH_CUA_LAB_FAILED",
93
92
  message: outcome?.sessionError ??
94
93
  outcome?.providerCleanupError ??
94
+ outcome?.providerPolicyError ??
95
95
  (outcome?.noEngagement
96
96
  ? "Actor took no actions and produced no message (likely a blank/still-loading screen); not a credible goal_satisfied."
97
- : // The lane result (toParticipantResult) named this refusal; the N=1 envelope fell through to
97
+ : // The participant result (toParticipantResult) named this refusal; the N=1 envelope fell through to
98
98
  // "did not produce a terminal session", which is false — it produced one and refused it.
99
99
  outcome?.selfReportedBlocker
100
100
  ? "Actor reported goal_satisfied while its final message described a blocker or asked for missing instructions; not a credible pass."
@@ -164,10 +164,11 @@ function cuaLabResult(args) {
164
164
  };
165
165
  }
166
166
  /**
167
- * The run's execution failures: each lane whose session failed in the harness, and an Observer
168
- * that failed.
167
+ * The run's execution failures: each participant whose session failed in the harness, each provider whose
168
+ * cleanup is unconfirmed or that reported a disallowed item after its last request, each sandbox
169
+ * whose release is unconfirmed, and an Observer that failed.
169
170
  */
170
- function computerUseExecutionFailures(outcomes, observer) {
171
+ function computerUseExecutionFailures(outcomes, observer, runId) {
171
172
  return [
172
173
  ...(outcomes ?? [])
173
174
  .filter((outcome) => participantHarnessFailed(participantFactsOf(outcome)))
@@ -183,6 +184,15 @@ function computerUseExecutionFailures(outcomes, observer) {
183
184
  message: `${outcome.spec.planned.id}: ${outcome.providerCleanupError}`,
184
185
  },
185
186
  ]),
187
+ ...(outcomes ?? []).flatMap((outcome) => outcome.providerPolicyError === undefined
188
+ ? []
189
+ : [
190
+ {
191
+ kind: "provider-policy",
192
+ message: `${outcome.spec.planned.id}: ${outcome.providerPolicyError}`,
193
+ },
194
+ ]),
195
+ ...unreleasedSandboxFailures(outcomes, runId),
186
196
  ...(observer.ok
187
197
  ? []
188
198
  : [
@@ -194,33 +204,34 @@ function computerUseExecutionFailures(outcomes, observer) {
194
204
  ];
195
205
  }
196
206
  /**
197
- * execution.caps.maxUsd is enforced inside each lane's loop independently, so an N-lane fan-out can
198
- * spend up to N × maxUsd before any lane aborts, while the run cost summary reports the larger
207
+ * execution.caps.maxUsd is enforced inside each participant's loop independently, so an
208
+ * N-participant fan-out can spend up to N × maxUsd before any participant aborts, while the run cost summary reports the larger
199
209
  * aggregate. The warning names that ceiling, unless the study declared a shared maxTotalUsd budget.
200
210
  */
201
- function participantCapWarning(config, participantCount) {
202
- const capUsd = config.execution?.caps?.maxUsd;
211
+ function participantCapWarning(caps, participantCount) {
212
+ const capUsd = caps.maxUsd;
203
213
  if (capUsd === undefined || participantCount <= 1)
204
214
  return undefined;
205
- if (config.execution?.caps?.maxTotalUsd !== undefined)
215
+ if (caps.maxTotalUsd !== undefined)
206
216
  return undefined;
207
217
  return `execution.caps.maxUsd ($${capUsd}) is a PER-PARTICIPANT cap; ${participantCount} participants may spend up to ${participantCount} × $${capUsd} (~$${round6(capUsd * participantCount)} total) before any participant aborts. Set execution.caps.maxTotalUsd for a shared study budget.`;
208
218
  }
209
219
  /** Builds and publishes the final bundle, runs the adapter hooks, renders the Observer and returns the result. */
210
220
  export async function finishCuaRun(setup, finish, ran) {
211
- const { plan, input, config, cwd, streams, descriptor, subjectRoute, run } = setup;
221
+ const { plan, input, cwd, streams, descriptor, run } = setup;
212
222
  const { participantRuns, participantPlan, scrubKnownValues, bundleBase } = setup;
213
- const { hooks, rerunLineage, publicRepo, subjectArgs } = finish;
223
+ const { rerunLineage, publicRepo, subjectArgs } = finish;
214
224
  const { dryRun } = plan;
215
- const { appUrl, subjectEnvNames } = subjectRoute;
225
+ const appUrl = plannedAppUrl(plan.runner.subject);
226
+ const subjectEnvNames = [...participantSubjectEnv(plan.runner.subject)];
216
227
  const { runId } = run;
217
228
  const participantCount = participantRuns.length;
218
229
  const { outcomes, failFastReason, receiving, receivingWarnings, externalCommsWarnings } = ran;
219
- // Per-lane subject projections (invariant 5).
230
+ // Per-participant subject projections (invariant 5).
220
231
  const subjects = projectParticipantSubjects({ ...subjectArgs, outcomes, dryRun });
221
232
  const aggregate = aggregateCuaSubject({ subjects, outcomes, participantCount, dryRun });
222
233
  const aggregateSubject = aggregate.subject;
223
- const capWarning = participantCapWarning(config, participantCount);
234
+ const capWarning = participantCapWarning(plan.caps, participantCount);
224
235
  const aggregateWarnings = [
225
236
  ...externalCommsWarnings,
226
237
  ...(capWarning === undefined ? [] : [capWarning]),
@@ -239,8 +250,8 @@ export async function finishCuaRun(setup, finish, ran) {
239
250
  ...(failFastReason === undefined ? {} : { failFastReason }),
240
251
  });
241
252
  const adapterWarnings = [];
242
- const scorerResult = await applyBrowserAdapterHooks({
243
- hooks,
253
+ const scorerResult = await applyBrowserScorer({
254
+ scorer: input.scorer,
244
255
  bundle,
245
256
  context: {
246
257
  bundle,
@@ -248,13 +259,12 @@ export async function finishCuaRun(setup, finish, ran) {
248
259
  labId: plan.labId,
249
260
  runId,
250
261
  actor: descriptor.id,
251
- backend: "cua",
262
+ route: "computer-use",
252
263
  dryRun,
253
- laneCount: participantCount,
264
+ participantCount,
254
265
  },
255
266
  sanitize: (text) => redactText(scrubKnownValues(text)),
256
267
  warnings: adapterWarnings,
257
- hookLabel: "cuaHooks",
258
268
  ...(input.scorerProvenance === undefined ? {} : { scorerProvenance: input.scorerProvenance }),
259
269
  });
260
270
  // The one final verdict fold: scoring can only make the judged verdict stricter.
@@ -264,7 +274,7 @@ export async function finishCuaRun(setup, finish, ran) {
264
274
  const finished = await run.finish(bundle);
265
275
  const observer = await finished.renderObserver();
266
276
  streams.attachFinal(observer);
267
- const execution = judgeExecution(computerUseExecutionFailures(outcomes, observer), OUTCOME_POLICIES["computer-use"]);
277
+ const execution = judgeExecution(computerUseExecutionFailures(outcomes, observer, runId), OUTCOME_POLICIES["computer-use"]);
268
278
  const result = cuaLabResult({
269
279
  labId: plan.labId,
270
280
  cwd,
@@ -11,7 +11,7 @@ import { type ComputerUseRunInput, type CuaActorLabResult, type RunCuaActorLabOp
11
11
  export declare function runCuaActorLab(options: RunCuaActorLabOptions): Promise<CuaActorLabResult>;
12
12
  /**
13
13
  * A refused computer-use lab's result, at the refusal's stage: a before-scope refusal has its own
14
- * envelope and no analysis record; the others come after the cwd checks, and the lane cap after
14
+ * envelope and no analysis record; the others come after the cwd checks, and the participant cap after
15
15
  * the personas are read.
16
16
  */
17
17
  export declare function computerUseLabRefusal(options: RunCuaActorLabOptions, refusal: ComputerUseRefusal): Promise<CuaActorLabResult>;
@@ -26,8 +26,7 @@ export declare function admitComputerUsePlan(plan: ComputerUsePlan, input: Compu
26
26
  * Run a computer-use plan: its local checks, then the run. The run scope in runAdmittedCuaRun
27
27
  * finalizes any status record the run opened, on every exit, so a test or library caller does not
28
28
  * leave the 5 s status cadence writing into a directory something else is deleting (an unrelated
29
- * ENOTEMPTY). Participants and the brain (model and local agent) come from the plan. `config` is
30
- * still read for values the plan does not carry yet (persona ids, comms settings, the
31
- * per-participant cap) and by the compatibility hooks, which take the whole config.
29
+ * ENOTEMPTY). Participants, the brain, the subject, the caps and the residual config come from the
30
+ * plan. `config` is what the caller's createProvider and inProcess executor receive.
32
31
  */
33
32
  export declare function runComputerUsePlan(plan: ComputerUsePlan, input: ComputerUseRunInput, config: LabConfig): Promise<CuaActorLabResult>;
@@ -26,10 +26,11 @@ import path from "node:path";
26
26
  import { completeAutomaticAnalysis } from "../../analysis/automatic-completion.js";
27
27
  import { runScope } from "../../run/run.js";
28
28
  import { resolveAutomaticAnalysis } from "../../analysis/automatic-config.js";
29
- import { computerUseInputWithScorer } from "../../lab/route-inputs.js";
30
- import { planComputerUseLab } from "./plan.js";
29
+ import { browserRouteScorer } from "../../lab/adapter-scorer-loader.js";
30
+ import { withLateScorer } from "../../lab/route-inputs.js";
31
+ import { callerDrivingOf, planComputerUseLab } from "./plan.js";
31
32
  import { finishCuaRun } from "./result.js";
32
- import { runLabParticipants } from "./run-lanes.js";
33
+ import { runLabParticipants } from "./live-phase.js";
33
34
  import { admitCuaRun, refuseCuaLab, startCuaRun } from "./setup.js";
34
35
  import { CUA_ACTOR_LAB_SCHEMA, } from "./types.js";
35
36
  /**
@@ -38,12 +39,12 @@ import { CUA_ACTOR_LAB_SCHEMA, } from "./types.js";
38
39
  * run and its analysis.
39
40
  */
40
41
  export async function runCuaActorLab(options) {
41
- const { config, dryRun, lab, ...input } = options;
42
+ const { config, dryRun, ...input } = options;
42
43
  // planComputerUseLab makes every configuration refusal, in the order this route always has.
43
44
  const planned = planComputerUseLab(config, {
44
45
  dryRun,
45
- ...(lab === undefined ? {} : { lab }),
46
- ...(input.hooks === undefined ? {} : { hooks: input.hooks }),
46
+ hasRunSession: input.deps?.runSession !== undefined,
47
+ driving: callerDrivingOf(input),
47
48
  ...(input.countOverride === undefined ? {} : { countOverride: input.countOverride }),
48
49
  ...(input.rerun === undefined ? {} : { rerun: input.rerun }),
49
50
  });
@@ -53,7 +54,7 @@ export async function runCuaActorLab(options) {
53
54
  }
54
55
  /**
55
56
  * A refused computer-use lab's result, at the refusal's stage: a before-scope refusal has its own
56
- * envelope and no analysis record; the others come after the cwd checks, and the lane cap after
57
+ * envelope and no analysis record; the others come after the cwd checks, and the participant cap after
57
58
  * the personas are read.
58
59
  */
59
60
  export async function computerUseLabRefusal(options, refusal) {
@@ -72,9 +73,9 @@ export async function computerUseLabRefusal(options, refusal) {
72
73
  warnings: [],
73
74
  error: { code: refusal.code, message: refusal.message },
74
75
  };
75
- // The other refusals come after the cwd checks, and the lane cap after the personas are read.
76
+ // The other refusals come after the cwd checks, and the participant cap after the personas are read.
76
77
  const analysis = resolveAutomaticAnalysis(config.review?.analysis);
77
- return completeAutomaticAnalysis(await refuseCuaLab(options, refusal), undefined, analysis.ok ? analysis.config : undefined, options.automaticAnalysis, { trigger: config.review?.analysis === undefined ? "default" : "explicit" });
78
+ return completeAutomaticAnalysis(await refuseCuaLab(options, refusal), undefined, analysis.ok ? analysis.config : undefined, options, { trigger: config.review?.analysis === undefined ? "default" : "explicit" });
78
79
  }
79
80
  /**
80
81
  * runLab's step for a computer-use plan. It runs the plan's local checks (admitCuaRun: the
@@ -92,24 +93,22 @@ export async function admitComputerUsePlan(plan, input, config) {
92
93
  return {
93
94
  ok: true,
94
95
  run: async (scorer) => {
95
- const running = computerUseInputWithScorer(input, scorer);
96
- // The scorer's hooks wrap the admitted bag, so the runSession, provider and desktop the
97
- // checks admitted stay the ones the run uses.
98
- const admitted = { ...admission.admitted, hooks: running.hooks ?? admission.admitted.hooks };
99
- return cuaOutcome(await runAdmittedCuaRun(plan, running, admitted));
96
+ // The scorer joins the input only; the runSession, provider and desktop the checks admitted
97
+ // stay the ones the run uses.
98
+ const running = withLateScorer(input, scorer, browserRouteScorer);
99
+ return cuaOutcome(await runAdmittedCuaRun(plan, running, admission.admitted));
100
100
  },
101
101
  };
102
102
  }
103
103
  function cuaOutcome(result) {
104
- return { route: "computer-use", backend: "cua", result };
104
+ return { route: "computer-use", result };
105
105
  }
106
106
  /**
107
107
  * Run a computer-use plan: its local checks, then the run. The run scope in runAdmittedCuaRun
108
108
  * finalizes any status record the run opened, on every exit, so a test or library caller does not
109
109
  * leave the 5 s status cadence writing into a directory something else is deleting (an unrelated
110
- * ENOTEMPTY). Participants and the brain (model and local agent) come from the plan. `config` is
111
- * still read for values the plan does not carry yet (persona ids, comms settings, the
112
- * per-participant cap) and by the compatibility hooks, which take the whole config.
110
+ * ENOTEMPTY). Participants, the brain, the subject, the caps and the residual config come from the
111
+ * plan. `config` is what the caller's createProvider and inProcess executor receive.
113
112
  */
114
113
  export async function runComputerUsePlan(plan, input, config) {
115
114
  const admission = await admitCuaRun(plan, input, config);
@@ -129,9 +128,12 @@ function runAdmittedCuaRun(plan, input, admitted) {
129
128
  });
130
129
  }
131
130
  function completeCuaAnalysis(plan, input, result, finished) {
132
- return completeAutomaticAnalysis(result, finished, plan.analysis?.config, input.automaticAnalysis, {
131
+ // A local VM study whose cleanup is unconfirmed records a skip instead of analyzing.
132
+ const refusal = input.localVm?.analysisRefusal;
133
+ return completeAutomaticAnalysis(result, finished, plan.analysis?.config, input, {
133
134
  ...(plan.analysis === undefined ? {} : { trigger: plan.analysis.trigger }),
134
135
  preferLargerOutput: plan.analysis?.preferLargerOutput === true,
136
+ ...(refusal === undefined ? {} : { refusal }),
135
137
  });
136
138
  }
137
139
  async function runPlanInScope(plan, input, admitted, scope) {
@@ -1,12 +1,12 @@
1
1
  import type { CuaLoopResult } from "../../actors/computer-use/loop.js";
2
2
  import { type SessionEnding } from "../../run/judge.js";
3
3
  /**
4
- * A goal_satisfied lane counts as a self-reported blocker ONLY when its final narrative contradicts
4
+ * A goal_satisfied participant counts as a self-reported blocker ONLY when its final narrative contradicts
5
5
  * the goal AND the run's own stop predicate did NOT fire. A matched stopWhen is independent,
6
6
  * structured completion evidence, so it overrides a text scan of the free-form narrative — which can
7
7
  * otherwise trip on the subject app's OWN quoted copy (e.g. a relayed "cannot be undone" banner).
8
8
  * Resolved-arc segments never block the verdict (#453). Returns the offending reason, or undefined
9
- * when the lane is a clean pass. Exported for testing.
9
+ * when the participant is a clean pass. Exported for testing.
10
10
  */
11
11
  export declare function resolveSelfReportedBlocker(session: CuaLoopResult | undefined): string | undefined;
12
12
  /** A session's ending, reduced to what the judge's engagement and blocker rules read. */
@@ -20,7 +20,7 @@ function hasBlockerLanguage(text) {
20
20
  // Report-shaped language: what a participant writes when it finished AND has something to say.
21
21
  // Every one of the day's eleven drawDB reports (2026-09-01) opened a section "What confused me"
22
22
  // or "Accessibility defects:"; none contained a blocker word, so none became a feedback candidate
23
- // and a lane that had just replicated a keyboard-accessibility defect three times drafted "Live
23
+ // and a participant that had just replicated a keyboard-accessibility defect three times drafted "Live
24
24
  // study completed without a participant-reported finding". Friction is the INCLUSIVE scan; a
25
25
  // false positive here adds a candidate a person then reads, which is the cheap direction.
26
26
  const REPORTED_DEFECT_LANGUAGE = /\b(defects?|bugs?|accessibilit(y|ies)|inaccessible|not (keyboard|screen.?reader)[- ]?accessible|confus(ed|ing)|hesitat(ed|ion)|unexpected(ly)?|unclear|hard to (find|tell|see|read|reach)|no (visible )?focus|overlap(ped|ping|s)?|truncat(ed|es|ion)|cut off|did nothing|nothing happened|no effect)\b/;
@@ -94,7 +94,7 @@ function completionReasonBlocksVerdict(reason) {
94
94
  // with a parser error. A simpler SQL import succeeded."). The lookahead demands the retry flavor
95
95
  // on purpose: unrelated praise ("Separately, the search box worked") must never launder an
96
96
  // unresolved failure. "Login failed so I gave up" has no recovery anywhere and stays a blocker.
97
- // (#453 — the run-1 false negative: a defect report after demonstrated success failed the lane,
97
+ // (#453 — the run-1 false negative: a defect report after demonstrated success failed the participant,
98
98
  // an incentive inversion against exactly the participant behavior a study wants most.)
99
99
  const RESOLUTION_TERMS = /\b(succeed(?:ed|s)?|success(?:ful|fully)?|worked|works around|then worked|now works?|resolved|fixed|recovered|got it working|went through)\b/;
100
100
  const RETRY_RESOLUTION = /\b(simpler|simplified|retry(?:ing)?|retried|second (?:attempt|try)|another (?:attempt|try|approach)|different (?:approach|way|route)|instead|then|eventually|after that)\b[^.!?\n]{0,80}\b(succeed(?:ed|s)?|success(?:ful|fully)?|worked|went through|completed|passed)\b/;
@@ -152,12 +152,12 @@ function traceHasStopWhenMatch(session) {
152
152
  (item.title.startsWith("stopWhen matched") || item.title === "dwell window complete"));
153
153
  }
154
154
  /**
155
- * A goal_satisfied lane counts as a self-reported blocker ONLY when its final narrative contradicts
155
+ * A goal_satisfied participant counts as a self-reported blocker ONLY when its final narrative contradicts
156
156
  * the goal AND the run's own stop predicate did NOT fire. A matched stopWhen is independent,
157
157
  * structured completion evidence, so it overrides a text scan of the free-form narrative — which can
158
158
  * otherwise trip on the subject app's OWN quoted copy (e.g. a relayed "cannot be undone" banner).
159
159
  * Resolved-arc segments never block the verdict (#453). Returns the offending reason, or undefined
160
- * when the lane is a clean pass. Exported for testing.
160
+ * when the participant is a clean pass. Exported for testing.
161
161
  */
162
162
  export function resolveSelfReportedBlocker(session) {
163
163
  return session !== undefined && selfReportedBlocker(blockerFacts(session))
@@ -3,15 +3,15 @@ import type { CuaActorDescriptor } from "../../actors/registry.js";
3
3
  import type { LabCommsEmail, LabCommsExternal, LabConfig } from "../../lab/types.js";
4
4
  import { type RunRerunLineage } from "../../run/bundle.js";
5
5
  import { type LocalTreeArchive } from "../../subject/local-tree-archive.js";
6
- import { type ComputerUseRefusal, type CuaSubjectRoute } from "./plan.js";
6
+ import { type ComputerUseRefusal } from "./plan.js";
7
7
  import type { ComputerUsePlan } from "../../lab/plan-types.js";
8
8
  import { trackRuntimeStreams, type LiveTraceFlush } from "./live-flush.js";
9
9
  import { type CuaRunBundleBase } from "./bundle.js";
10
10
  import { projectParticipantSubjects } from "./subject-projection.js";
11
- import { type CuaActorLabErrorCode, type CuaActorLabHooks, type CuaActorLabResult, type CuaParticipantDeps, type CuaParticipantPlan, type DesktopParticipantRun, type CuaSubjectProjection, type CuaSubjectProvenanceArg, type ComputerUseRunInput, type RunCuaActorLabOptions } from "./types.js";
11
+ import { type CuaActorLabErrorCode, type CuaActorLabResult, type CuaParticipantDeps, type CuaParticipantPlan, type DesktopParticipantRun, type CuaSubjectProjection, type CuaSubjectProvenanceArg, type ComputerUseRunInput, type RunCuaActorLabOptions } from "./types.js";
12
12
  /**
13
- * The envelope of a refusal the plan left for after the cwd checks. The lane cap and in-process
14
- * fan-out refusals come after the committed personas are read, so a persona-file error still wins.
13
+ * The envelope of a refusal the plan left for after the cwd checks. The participant cap and
14
+ * in-process fan-out refusals come after the committed personas are read, so a persona-file error still wins.
15
15
  */
16
16
  export declare function refuseCuaLab(options: RunCuaActorLabOptions, refusal: ComputerUseRefusal): Promise<CuaActorLabResult>;
17
17
  export type AdmittedCuaRun = Extract<Awaited<ReturnType<typeof admitCuaRun>>, {
@@ -24,13 +24,9 @@ type StartedRun = Extract<Awaited<ReturnType<RunScope["startRun"]>>, {
24
24
  export interface CuaRunSetup {
25
25
  plan: ComputerUsePlan;
26
26
  input: ComputerUseRunInput;
27
- /** Read for values the plan does not carry yet, and by the compatibility hooks, which take the
28
- * whole config. Participants come from the plan. */
29
- config: LabConfig;
30
27
  /** The physical project root. */
31
28
  cwd: string;
32
29
  descriptor: CuaActorDescriptor;
33
- subjectRoute: CuaSubjectRoute;
34
30
  run: StartedRun;
35
31
  streams: ReturnType<typeof trackRuntimeStreams>;
36
32
  participantRuns: DesktopParticipantRun[];
@@ -66,7 +62,6 @@ export interface CuaParticipantsSetup {
66
62
  }
67
63
  /** What only finishCuaRun reads besides the participants' outcomes. */
68
64
  export interface CuaFinishFacts {
69
- hooks: CuaActorLabHooks;
70
65
  rerunLineage: RunRerunLineage | undefined;
71
66
  publicRepo: string | undefined;
72
67
  subjectArgs: Omit<Parameters<typeof projectParticipantSubjects>[0], "outcomes" | "dryRun">;
@@ -95,16 +90,16 @@ export declare function admitCuaRun(plan: ComputerUsePlan, input: ComputerUseRun
95
90
  config: LabConfig;
96
91
  dryRun: boolean;
97
92
  cwd: string;
98
- hooks: CuaActorLabHooks;
93
+ seams: import("../../lab/lab-deps.js").LabDeps;
99
94
  streams: {
100
- hooks: CuaActorLabHooks;
95
+ onStream: NonNullable<import("../../lab/run-lab-homes.js").RunLabHomes["onStream"]>;
101
96
  showIn: (observer: import("../../index.js").ObserverResult & {
102
97
  ok: true;
103
98
  }) => void;
104
99
  attachFinal: (observer: import("../../index.js").ObserverResult) => void;
105
100
  };
106
- env: NodeJS.ProcessEnv;
107
- subjectRoute: CuaSubjectRoute;
101
+ env: Record<string, string | undefined>;
102
+ appUrl: string;
108
103
  fail: (code: CuaActorLabErrorCode, message: string, actorLabel?: string) => CuaActorLabResult;
109
104
  refuse: (code: CuaActorLabErrorCode, message: string, actorLabel?: string | undefined) => {
110
105
  ok: false;
@@ -123,7 +118,6 @@ export declare function admitCuaRun(plan: ComputerUsePlan, input: ComputerUseRun
123
118
  knownSecretValues: string[];
124
119
  scrubKnownValues: (text: string) => string;
125
120
  publicRepo: string | undefined;
126
- hasGithubToken: boolean;
127
121
  localTreeArchive: LocalTreeArchive | undefined;
128
122
  localTreeArchiveBuffer: ArrayBuffer | undefined;
129
123
  };