@stigmer/runner 3.15.1 → 3.15.2-dev.20260913223433

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 (518) hide show
  1. package/README.md +2 -4
  2. package/dist/.build-fingerprint +1 -1
  3. package/dist/__test-utils__/execution-record-fixture.d.ts +14 -0
  4. package/dist/__test-utils__/execution-record-fixture.js +4 -0
  5. package/dist/__test-utils__/execution-record-fixture.js.map +1 -1
  6. package/dist/__test-utils__/file-review-projection.d.ts +40 -0
  7. package/dist/__test-utils__/file-review-projection.js +136 -0
  8. package/dist/__test-utils__/file-review-projection.js.map +1 -0
  9. package/dist/__test-utils__/harness-contract/types.d.ts +23 -1
  10. package/dist/__test-utils__/harness-contract/types.js +15 -0
  11. package/dist/__test-utils__/harness-contract/types.js.map +1 -1
  12. package/dist/__test-utils__/hermetic-activity.d.ts +44 -12
  13. package/dist/__test-utils__/hermetic-activity.js +52 -14
  14. package/dist/__test-utils__/hermetic-activity.js.map +1 -1
  15. package/dist/__test-utils__/model-registry-fixture.d.ts +32 -7
  16. package/dist/__test-utils__/model-registry-fixture.js +31 -7
  17. package/dist/__test-utils__/model-registry-fixture.js.map +1 -1
  18. package/dist/__test-utils__/turn-input-fixture.d.ts +1 -1
  19. package/dist/__test-utils__/turn-input-fixture.js +1 -1
  20. package/dist/__test-utils__/turn-input-fixture.js.map +1 -1
  21. package/dist/activities/call-agent.js +2 -2
  22. package/dist/activities/call-agent.js.map +1 -1
  23. package/dist/activities/execute-cursor/__test-utils__/contract-subject.js +45 -3
  24. package/dist/activities/execute-cursor/__test-utils__/contract-subject.js.map +1 -1
  25. package/dist/activities/execute-cursor/__test-utils__/hermetic-cursor.d.ts +0 -9
  26. package/dist/activities/execute-cursor/__test-utils__/hermetic-cursor.js +2 -41
  27. package/dist/activities/execute-cursor/__test-utils__/hermetic-cursor.js.map +1 -1
  28. package/dist/activities/execute-cursor/adapter.d.ts +7 -5
  29. package/dist/activities/execute-cursor/adapter.js +7 -5
  30. package/dist/activities/execute-cursor/adapter.js.map +1 -1
  31. package/dist/activities/execute-cursor/approval-policy.d.ts +5 -2
  32. package/dist/activities/execute-cursor/approval-policy.js +5 -2
  33. package/dist/activities/execute-cursor/approval-policy.js.map +1 -1
  34. package/dist/activities/execute-cursor/approval-state.d.ts +16 -8
  35. package/dist/activities/execute-cursor/approval-state.js +6 -3
  36. package/dist/activities/execute-cursor/approval-state.js.map +1 -1
  37. package/dist/activities/execute-cursor/cas-observations.d.ts +14 -2
  38. package/dist/activities/execute-cursor/cas-observations.js +19 -2
  39. package/dist/activities/execute-cursor/cas-observations.js.map +1 -1
  40. package/dist/activities/execute-cursor/cursor-capabilities.d.ts +8 -4
  41. package/dist/activities/execute-cursor/cursor-capabilities.js +21 -4
  42. package/dist/activities/execute-cursor/cursor-capabilities.js.map +1 -1
  43. package/dist/activities/execute-cursor/message-translator.d.ts +12 -20
  44. package/dist/activities/execute-cursor/message-translator.js +19 -32
  45. package/dist/activities/execute-cursor/message-translator.js.map +1 -1
  46. package/dist/activities/execute-cursor/prompt-builder.d.ts +48 -58
  47. package/dist/activities/execute-cursor/prompt-builder.js +76 -95
  48. package/dist/activities/execute-cursor/prompt-builder.js.map +1 -1
  49. package/dist/activities/execute-cursor/turn-boundary.d.ts +27 -55
  50. package/dist/activities/execute-cursor/turn-boundary.js +30 -90
  51. package/dist/activities/execute-cursor/turn-boundary.js.map +1 -1
  52. package/dist/activities/execute-cursor/turn-settle.d.ts +11 -7
  53. package/dist/activities/execute-cursor/turn-settle.js +37 -49
  54. package/dist/activities/execute-cursor/turn-settle.js.map +1 -1
  55. package/dist/activities/execute-cursor/turn-setup.d.ts +16 -24
  56. package/dist/activities/execute-cursor/turn-setup.js +29 -59
  57. package/dist/activities/execute-cursor/turn-setup.js.map +1 -1
  58. package/dist/activities/execute-cursor/turn-stream.d.ts +3 -8
  59. package/dist/activities/execute-cursor/turn-stream.js +9 -22
  60. package/dist/activities/execute-cursor/turn-stream.js.map +1 -1
  61. package/dist/activities/execute-cursor/turn.js +5 -1
  62. package/dist/activities/execute-cursor/turn.js.map +1 -1
  63. package/dist/activities/execute-deep-agent/__test-utils__/contract-subject.d.ts +93 -0
  64. package/dist/activities/execute-deep-agent/__test-utils__/contract-subject.js +348 -0
  65. package/dist/activities/execute-deep-agent/__test-utils__/contract-subject.js.map +1 -0
  66. package/dist/activities/execute-deep-agent/__test-utils__/hermetic-deep-agent.d.ts +170 -0
  67. package/dist/activities/execute-deep-agent/__test-utils__/hermetic-deep-agent.js +225 -0
  68. package/dist/activities/execute-deep-agent/__test-utils__/hermetic-deep-agent.js.map +1 -0
  69. package/dist/activities/execute-deep-agent/__test-utils__/hitl-script.d.ts +24 -0
  70. package/dist/activities/execute-deep-agent/__test-utils__/hitl-script.js +41 -0
  71. package/dist/activities/execute-deep-agent/__test-utils__/hitl-script.js.map +1 -0
  72. package/dist/activities/execute-deep-agent/__test-utils__/scripted-model-module.d.ts +54 -0
  73. package/dist/activities/execute-deep-agent/__test-utils__/scripted-model-module.js +72 -0
  74. package/dist/activities/execute-deep-agent/__test-utils__/scripted-model-module.js.map +1 -0
  75. package/dist/activities/execute-deep-agent/__test-utils__/scripted-model.d.ts +187 -23
  76. package/dist/activities/execute-deep-agent/__test-utils__/scripted-model.js +238 -27
  77. package/dist/activities/execute-deep-agent/__test-utils__/scripted-model.js.map +1 -1
  78. package/dist/activities/execute-deep-agent/adapter.d.ts +41 -0
  79. package/dist/activities/execute-deep-agent/adapter.js +78 -0
  80. package/dist/activities/execute-deep-agent/adapter.js.map +1 -0
  81. package/dist/activities/execute-deep-agent/cas-capture-observer.d.ts +8 -0
  82. package/dist/activities/execute-deep-agent/cas-capture-observer.js +9 -0
  83. package/dist/activities/execute-deep-agent/cas-capture-observer.js.map +1 -1
  84. package/dist/activities/execute-deep-agent/deep-agent-capabilities.d.ts +31 -0
  85. package/dist/activities/execute-deep-agent/deep-agent-capabilities.js +40 -0
  86. package/dist/activities/execute-deep-agent/deep-agent-capabilities.js.map +1 -0
  87. package/dist/activities/execute-deep-agent/execution-state.d.ts +4 -11
  88. package/dist/activities/execute-deep-agent/execution-state.js +4 -12
  89. package/dist/activities/execute-deep-agent/execution-state.js.map +1 -1
  90. package/dist/activities/execute-deep-agent/hitl.d.ts +49 -47
  91. package/dist/activities/execute-deep-agent/hitl.js +77 -103
  92. package/dist/activities/execute-deep-agent/hitl.js.map +1 -1
  93. package/dist/activities/execute-deep-agent/prompt-builder.d.ts +51 -23
  94. package/dist/activities/execute-deep-agent/prompt-builder.js +115 -76
  95. package/dist/activities/execute-deep-agent/prompt-builder.js.map +1 -1
  96. package/dist/activities/execute-deep-agent/status-builder-shared.d.ts +18 -32
  97. package/dist/activities/execute-deep-agent/status-builder-shared.js +10 -65
  98. package/dist/activities/execute-deep-agent/status-builder-shared.js.map +1 -1
  99. package/dist/activities/execute-deep-agent/streaming-side-effects.d.ts +10 -8
  100. package/dist/activities/execute-deep-agent/streaming-side-effects.js +12 -14
  101. package/dist/activities/execute-deep-agent/streaming-side-effects.js.map +1 -1
  102. package/dist/activities/execute-deep-agent/subagent-tracker.d.ts +36 -13
  103. package/dist/activities/execute-deep-agent/subagent-tracker.js +76 -39
  104. package/dist/activities/execute-deep-agent/subagent-tracker.js.map +1 -1
  105. package/dist/activities/execute-deep-agent/subagent-transformer.d.ts +25 -28
  106. package/dist/activities/execute-deep-agent/subagent-transformer.js +29 -82
  107. package/dist/activities/execute-deep-agent/subagent-transformer.js.map +1 -1
  108. package/dist/activities/execute-deep-agent/subagent-wiring.d.ts +3 -2
  109. package/dist/activities/execute-deep-agent/subagent-wiring.js +2 -2
  110. package/dist/activities/execute-deep-agent/subagent-wiring.js.map +1 -1
  111. package/dist/activities/execute-deep-agent/turn-settle.d.ts +42 -0
  112. package/dist/activities/execute-deep-agent/turn-settle.js +146 -0
  113. package/dist/activities/execute-deep-agent/turn-settle.js.map +1 -0
  114. package/dist/activities/execute-deep-agent/turn-setup.d.ts +212 -0
  115. package/dist/activities/execute-deep-agent/turn-setup.js +430 -0
  116. package/dist/activities/execute-deep-agent/turn-setup.js.map +1 -0
  117. package/dist/activities/execute-deep-agent/turn-stream.d.ts +97 -0
  118. package/dist/activities/execute-deep-agent/turn-stream.js +242 -0
  119. package/dist/activities/execute-deep-agent/turn-stream.js.map +1 -0
  120. package/dist/activities/execute-deep-agent/turn.d.ts +42 -0
  121. package/dist/activities/execute-deep-agent/turn.js +116 -0
  122. package/dist/activities/execute-deep-agent/turn.js.map +1 -0
  123. package/dist/activities/execute-deep-agent/v3-event-recorder.d.ts +6 -5
  124. package/dist/activities/execute-deep-agent/v3-event-recorder.js +6 -5
  125. package/dist/activities/execute-deep-agent/v3-event-recorder.js.map +1 -1
  126. package/dist/activities/execute-deep-agent/v3-events.d.ts +1 -1
  127. package/dist/activities/execute-deep-agent/v3-events.js +1 -1
  128. package/dist/activities/execute-deep-agent/v3-status-builder.d.ts +60 -17
  129. package/dist/activities/execute-deep-agent/v3-status-builder.js +72 -57
  130. package/dist/activities/execute-deep-agent/v3-status-builder.js.map +1 -1
  131. package/dist/harness/approval-decisions.d.ts +61 -0
  132. package/dist/harness/approval-decisions.js +102 -0
  133. package/dist/harness/approval-decisions.js.map +1 -0
  134. package/dist/harness/capabilities.d.ts +16 -5
  135. package/dist/harness/capabilities.js +9 -0
  136. package/dist/harness/capabilities.js.map +1 -1
  137. package/dist/harness/capture.d.ts +144 -0
  138. package/dist/harness/capture.js +219 -0
  139. package/dist/harness/capture.js.map +1 -0
  140. package/dist/harness/persist-chokepoint.d.ts +19 -4
  141. package/dist/harness/persist-chokepoint.js +17 -4
  142. package/dist/harness/persist-chokepoint.js.map +1 -1
  143. package/dist/harness/run-turn.js +115 -29
  144. package/dist/harness/run-turn.js.map +1 -1
  145. package/dist/harness/terminal-table.d.ts +40 -11
  146. package/dist/harness/terminal-table.js +58 -19
  147. package/dist/harness/terminal-table.js.map +1 -1
  148. package/dist/harness/turn-context.d.ts +141 -76
  149. package/dist/harness/turn-context.js +194 -99
  150. package/dist/harness/turn-context.js.map +1 -1
  151. package/dist/harness/types.d.ts +65 -12
  152. package/dist/harness/types.js +8 -2
  153. package/dist/harness/types.js.map +1 -1
  154. package/dist/harness-adapters.d.ts +12 -8
  155. package/dist/harness-adapters.js +17 -9
  156. package/dist/harness-adapters.js.map +1 -1
  157. package/dist/middleware/approval-gate.d.ts +1 -1
  158. package/dist/middleware/approval-gate.js +1 -1
  159. package/dist/middleware/approval-gate.js.map +1 -1
  160. package/dist/middleware/cost-advisory.d.ts +39 -0
  161. package/dist/middleware/{cost-cap.js → cost-advisory.js} +37 -55
  162. package/dist/middleware/cost-advisory.js.map +1 -0
  163. package/dist/middleware/index.d.ts +23 -12
  164. package/dist/middleware/index.js +26 -18
  165. package/dist/middleware/index.js.map +1 -1
  166. package/dist/middleware/types.d.ts +3 -2
  167. package/dist/runner-manager.js +1 -5
  168. package/dist/runner-manager.js.map +1 -1
  169. package/dist/runner.js +1 -5
  170. package/dist/runner.js.map +1 -1
  171. package/dist/shared/attachment-resolver.d.ts +41 -18
  172. package/dist/shared/attachment-resolver.js +178 -46
  173. package/dist/shared/attachment-resolver.js.map +1 -1
  174. package/dist/shared/attachment-zip.d.ts +85 -0
  175. package/dist/shared/attachment-zip.js +185 -0
  176. package/dist/shared/attachment-zip.js.map +1 -0
  177. package/dist/shared/connect-backfill.js +5 -4
  178. package/dist/shared/connect-backfill.js.map +1 -1
  179. package/dist/shared/cost-guard.d.ts +3 -3
  180. package/dist/shared/cost-guard.js +3 -3
  181. package/dist/shared/execution-status-writer.d.ts +5 -5
  182. package/dist/shared/execution-status-writer.js +5 -5
  183. package/dist/shared/extract-structured-output.d.ts +8 -7
  184. package/dist/shared/extract-structured-output.js +8 -7
  185. package/dist/shared/extract-structured-output.js.map +1 -1
  186. package/dist/shared/filereview/cas-progress.d.ts +1 -18
  187. package/dist/shared/filereview/cas-progress.js +4 -0
  188. package/dist/shared/filereview/cas-progress.js.map +1 -1
  189. package/dist/shared/filereview/cas-touched.d.ts +56 -0
  190. package/dist/shared/filereview/cas-touched.js +63 -0
  191. package/dist/shared/filereview/cas-touched.js.map +1 -0
  192. package/dist/shared/filereview/index.d.ts +2 -1
  193. package/dist/shared/filereview/index.js +2 -1
  194. package/dist/shared/filereview/index.js.map +1 -1
  195. package/dist/shared/mcp-resolver.d.ts +4 -5
  196. package/dist/shared/mcp-resolver.js +4 -5
  197. package/dist/shared/mcp-resolver.js.map +1 -1
  198. package/dist/shared/persist-decision.d.ts +3 -4
  199. package/dist/shared/persist-decision.js +3 -4
  200. package/dist/shared/persist-decision.js.map +1 -1
  201. package/dist/shared/plan-mode-permissions.d.ts +2 -2
  202. package/dist/shared/plan-mode-permissions.js +2 -2
  203. package/dist/shared/plan-mode-prompt.d.ts +1 -1
  204. package/dist/shared/plan-mode-prompt.js +1 -1
  205. package/dist/shared/prompt-sections.d.ts +163 -0
  206. package/dist/shared/prompt-sections.js +164 -0
  207. package/dist/shared/prompt-sections.js.map +1 -0
  208. package/dist/shared/skill-resolver.d.ts +4 -2
  209. package/dist/shared/skill-resolver.js +4 -2
  210. package/dist/shared/skill-resolver.js.map +1 -1
  211. package/dist/shared/tool-rounds.d.ts +32 -0
  212. package/dist/shared/tool-rounds.js +43 -0
  213. package/dist/shared/tool-rounds.js.map +1 -1
  214. package/dist/shared/tool-row.d.ts +34 -0
  215. package/dist/shared/tool-row.js +52 -0
  216. package/dist/shared/tool-row.js.map +1 -1
  217. package/dist/shared/workspace/session-provision.d.ts +3 -3
  218. package/dist/shared/workspace/session-provision.js +3 -3
  219. package/dist/shared/workspace/stigmer-link.d.ts +3 -3
  220. package/dist/shared/workspace/stigmer-link.js +3 -3
  221. package/dist/shared/workspace/writeback-coordinator.d.ts +15 -22
  222. package/dist/shared/workspace/writeback-coordinator.js +16 -53
  223. package/dist/shared/workspace/writeback-coordinator.js.map +1 -1
  224. package/dist/tools/index.d.ts +1 -1
  225. package/dist/tools/index.js +1 -1
  226. package/package.json +4 -4
  227. package/src/__test-utils__/__tests__/harness-contract-self-check.test.ts +1 -0
  228. package/src/__test-utils__/execution-record-fixture.ts +18 -2
  229. package/src/__test-utils__/file-review-projection.ts +156 -0
  230. package/src/__test-utils__/git-workspace-fixture.ts +58 -0
  231. package/src/__test-utils__/harness-boot-order-child.ts +7 -2
  232. package/src/__test-utils__/harness-contract/contract.ts +33 -20
  233. package/src/__test-utils__/harness-contract/recording-sink.ts +11 -1
  234. package/src/__test-utils__/harness-contract/runtime-contract.ts +328 -10
  235. package/src/__test-utils__/harness-contract/scripted-adapter.ts +38 -6
  236. package/src/__test-utils__/harness-contract/types.ts +27 -3
  237. package/src/__test-utils__/hermetic-activity.ts +52 -14
  238. package/src/__test-utils__/model-registry-fixture.ts +32 -7
  239. package/src/__test-utils__/module-specifiers.ts +24 -0
  240. package/src/__test-utils__/turn-input-fixture.ts +2 -2
  241. package/src/__tests__/harness-boot-order.test.ts +9 -2
  242. package/src/activities/call-agent.ts +2 -2
  243. package/src/activities/execute-cursor/__test-utils__/contract-subject.ts +47 -3
  244. package/src/activities/execute-cursor/__test-utils__/hermetic-cursor.ts +2 -48
  245. package/src/activities/execute-cursor/__tests__/approval-decisions-agree.test.ts +1 -1
  246. package/src/activities/execute-cursor/__tests__/approval-gate.test.ts +6 -1
  247. package/src/activities/execute-cursor/__tests__/build-prompt.test.ts +52 -27
  248. package/src/activities/execute-cursor/__tests__/cas-observations.test.ts +25 -0
  249. package/src/activities/execute-cursor/__tests__/goldens/prompt.enhanced.everything.prompt.md +166 -0
  250. package/src/activities/execute-cursor/__tests__/goldens/prompt.enhanced.plan-mode.prompt.md +186 -0
  251. package/src/activities/execute-cursor/__tests__/goldens/prompt.hitl.recovery.prompt.md +191 -0
  252. package/src/activities/execute-cursor/__tests__/goldens/prompt.hitl.reinvocation.prompt.md +13 -0
  253. package/src/activities/execute-cursor/__tests__/goldens/prompt.resumed.prefixed.prompt.md +32 -0
  254. package/src/activities/execute-cursor/__tests__/goldens/prompt.resumed.raw.prompt.md +1 -0
  255. package/src/activities/execute-cursor/__tests__/hermetic/file-review-capture.test.ts +1 -2
  256. package/src/activities/execute-cursor/__tests__/hermetic/goldens/sdk-error-at-create.status.json +4 -0
  257. package/src/activities/execute-cursor/__tests__/hermetic/harness-contract.test.ts +24 -4
  258. package/src/activities/execute-cursor/__tests__/hermetic/stream-self-stop-arms.test.ts +6 -5
  259. package/src/activities/execute-cursor/__tests__/hermetic/thrown-error-arms.test.ts +9 -0
  260. package/src/activities/execute-cursor/__tests__/hitl-resume-history.test.ts +11 -8
  261. package/src/activities/execute-cursor/__tests__/prompt-goldens.test.ts +229 -0
  262. package/src/activities/execute-cursor/__tests__/same-identity-reproposal.test.ts +3 -3
  263. package/src/activities/execute-cursor/__tests__/sequential-gate-resume.test.ts +2 -2
  264. package/src/activities/execute-cursor/__tests__/turn-boundary.test.ts +78 -234
  265. package/src/activities/execute-cursor/__tests__/turn-stream.test.ts +1 -26
  266. package/src/activities/execute-cursor/adapter.ts +7 -5
  267. package/src/activities/execute-cursor/approval-policy.ts +5 -2
  268. package/src/activities/execute-cursor/approval-state.ts +16 -8
  269. package/src/activities/execute-cursor/cas-observations.ts +20 -2
  270. package/src/activities/execute-cursor/cursor-capabilities.ts +24 -5
  271. package/src/activities/execute-cursor/message-translator.ts +29 -41
  272. package/src/activities/execute-cursor/prompt-builder.ts +99 -157
  273. package/src/activities/execute-cursor/turn-boundary.ts +34 -140
  274. package/src/activities/execute-cursor/turn-settle.ts +37 -51
  275. package/src/activities/execute-cursor/turn-setup.ts +30 -70
  276. package/src/activities/execute-cursor/turn-stream.ts +8 -31
  277. package/src/activities/execute-cursor/turn.ts +5 -1
  278. package/src/activities/execute-deep-agent/__test-utils__/contract-subject.ts +391 -0
  279. package/src/activities/execute-deep-agent/__test-utils__/hermetic-deep-agent.ts +313 -0
  280. package/src/activities/execute-deep-agent/__test-utils__/hitl-script.ts +48 -0
  281. package/src/activities/execute-deep-agent/__test-utils__/scripted-model-module.ts +85 -0
  282. package/src/activities/execute-deep-agent/__test-utils__/scripted-model.ts +383 -36
  283. package/src/activities/execute-deep-agent/__tests__/adapter-graph-is-sdk-free.test.ts +76 -0
  284. package/src/activities/execute-deep-agent/__tests__/adapter-is-temporal-free.test.ts +47 -0
  285. package/src/activities/execute-deep-agent/__tests__/adapter.test.ts +80 -0
  286. package/src/activities/execute-deep-agent/__tests__/cas-capture-backend.test.ts +1 -1
  287. package/src/activities/execute-deep-agent/__tests__/execution-state-extended.test.ts +0 -2
  288. package/src/activities/execute-deep-agent/__tests__/execution-state.test.ts +0 -3
  289. package/src/activities/execute-deep-agent/__tests__/goldens/system-prompt.build-from-plan.prompt.md +172 -0
  290. package/src/activities/execute-deep-agent/__tests__/goldens/system-prompt.everything.prompt.md +167 -0
  291. package/src/activities/execute-deep-agent/__tests__/goldens/system-prompt.minimal.prompt.md +36 -0
  292. package/src/activities/execute-deep-agent/__tests__/goldens/system-prompt.plan-mode.prompt.md +178 -0
  293. package/src/activities/execute-deep-agent/__tests__/goldens/system-prompt.skills-below-threshold.prompt.md +158 -0
  294. package/src/activities/execute-deep-agent/__tests__/hermetic/file-review-capture.test.ts +210 -0
  295. package/src/activities/execute-deep-agent/__tests__/hermetic/goldens/approve-all-lease.turn1.status.json +74 -0
  296. package/src/activities/execute-deep-agent/__tests__/hermetic/goldens/approve-all-lease.turn2.status.json +104 -0
  297. package/src/activities/execute-deep-agent/__tests__/hermetic/goldens/approve.turn2.status.json +78 -0
  298. package/src/activities/execute-deep-agent/__tests__/hermetic/goldens/file-review-capture.status.json +138 -0
  299. package/src/activities/execute-deep-agent/__tests__/hermetic/goldens/gate.turn1.status.json +68 -0
  300. package/src/activities/execute-deep-agent/__tests__/hermetic/goldens/inline-artifact.turn1.status.json +154 -0
  301. package/src/activities/execute-deep-agent/__tests__/hermetic/goldens/inline-artifact.turn2.status.json +116 -0
  302. package/src/activities/execute-deep-agent/__tests__/hermetic/goldens/pause.loop.after-tool.status.json +69 -0
  303. package/src/activities/execute-deep-agent/__tests__/hermetic/goldens/pause.loop.mid-tool.status.json +67 -0
  304. package/src/activities/execute-deep-agent/__tests__/hermetic/goldens/plain-turn.status.json +47 -0
  305. package/src/activities/execute-deep-agent/__tests__/hermetic/goldens/platform-stop.status.json +68 -0
  306. package/src/activities/execute-deep-agent/__tests__/hermetic/goldens/recursion-limit.persisted.status.json +280 -0
  307. package/src/activities/execute-deep-agent/__tests__/hermetic/goldens/reject.turn2.status.json +77 -0
  308. package/src/activities/execute-deep-agent/__tests__/hermetic/goldens/resolution-error.status.json +18 -0
  309. package/src/activities/execute-deep-agent/__tests__/hermetic/goldens/sequential-gates.turn2.status.json +99 -0
  310. package/src/activities/execute-deep-agent/__tests__/hermetic/goldens/sequential-gates.turn3.status.json +109 -0
  311. package/src/activities/execute-deep-agent/__tests__/hermetic/goldens/skip.turn2.status.json +76 -0
  312. package/src/activities/execute-deep-agent/__tests__/hermetic/goldens/structured-output.text-fallback.status.json +51 -0
  313. package/src/activities/execute-deep-agent/__tests__/hermetic/goldens/structured-output.tier2.status.json +51 -0
  314. package/src/activities/execute-deep-agent/__tests__/hermetic/goldens/structured-output.tool-strategy.status.json +56 -0
  315. package/src/activities/execute-deep-agent/__tests__/hermetic/goldens/sub-agent-delegation.status.json +90 -0
  316. package/src/activities/execute-deep-agent/__tests__/hermetic/goldens/tool-call.status.json +70 -0
  317. package/src/activities/execute-deep-agent/__tests__/hermetic/goldens/worker-shutdown.status.json +71 -0
  318. package/src/activities/execute-deep-agent/__tests__/hermetic/harness-contract.test.ts +144 -0
  319. package/src/activities/execute-deep-agent/__tests__/hermetic/hitl-approve-all-lease.test.ts +184 -0
  320. package/src/activities/execute-deep-agent/__tests__/hermetic/hitl-round-trips.test.ts +203 -0
  321. package/src/activities/execute-deep-agent/__tests__/hermetic/hitl-sequential-gates.test.ts +170 -0
  322. package/src/activities/execute-deep-agent/__tests__/hermetic/inline-artifact-across-gate.test.ts +175 -0
  323. package/src/activities/execute-deep-agent/__tests__/hermetic/pause-vs-shutdown.test.ts +206 -0
  324. package/src/activities/execute-deep-agent/__tests__/hermetic/plain-turn.test.ts +142 -0
  325. package/src/activities/execute-deep-agent/__tests__/hermetic/platform-stop.test.ts +126 -0
  326. package/src/activities/execute-deep-agent/__tests__/hermetic/recursion-limit.test.ts +139 -0
  327. package/src/activities/execute-deep-agent/__tests__/hermetic/resolution-error.test.ts +122 -0
  328. package/src/activities/execute-deep-agent/__tests__/hermetic/structured-output.test.ts +236 -0
  329. package/src/activities/execute-deep-agent/__tests__/hermetic/sub-agent-delegation.test.ts +161 -0
  330. package/src/activities/execute-deep-agent/__tests__/hermetic/tool-call.test.ts +157 -0
  331. package/src/activities/execute-deep-agent/__tests__/hermetic/workspace-lock-timeout.test.ts +152 -0
  332. package/src/activities/execute-deep-agent/__tests__/hitl.test.ts +50 -122
  333. package/src/activities/execute-deep-agent/__tests__/inline-publisher.test.ts +8 -6
  334. package/src/activities/execute-deep-agent/__tests__/plan-mode-path-normalization.test.ts +2 -2
  335. package/src/activities/execute-deep-agent/__tests__/prompt-builder.test.ts +35 -35
  336. package/src/activities/execute-deep-agent/__tests__/prompt-goldens.test.ts +247 -0
  337. package/src/activities/execute-deep-agent/__tests__/scripted-model.test.ts +329 -0
  338. package/src/activities/execute-deep-agent/__tests__/status-builder-shared.test.ts +1 -20
  339. package/src/activities/execute-deep-agent/__tests__/streaming-side-effects.test.ts +78 -0
  340. package/src/activities/execute-deep-agent/__tests__/subagent-gitignored-capture.test.ts +1 -1
  341. package/src/activities/execute-deep-agent/__tests__/subagent-plan-mode-permissions.test.ts +2 -2
  342. package/src/activities/execute-deep-agent/__tests__/subagent-tracker.test.ts +57 -31
  343. package/src/activities/execute-deep-agent/__tests__/subagent-transformer.test.ts +19 -79
  344. package/src/activities/execute-deep-agent/__tests__/subagent-wiring.test.ts +15 -15
  345. package/src/activities/execute-deep-agent/__tests__/summarization-verification.test.ts +1 -1
  346. package/src/activities/execute-deep-agent/__tests__/turn-stream.test.ts +391 -0
  347. package/src/activities/execute-deep-agent/__tests__/v3-status-builder.test.ts +343 -38
  348. package/src/activities/execute-deep-agent/__tests__/vision-input.test.ts +5 -5
  349. package/src/activities/execute-deep-agent/adapter.ts +94 -0
  350. package/src/activities/execute-deep-agent/cas-capture-observer.ts +11 -0
  351. package/src/activities/execute-deep-agent/deep-agent-capabilities.ts +43 -0
  352. package/src/activities/execute-deep-agent/execution-state.ts +4 -13
  353. package/src/activities/execute-deep-agent/hitl.ts +100 -123
  354. package/src/activities/execute-deep-agent/prompt-builder.ts +141 -115
  355. package/src/activities/execute-deep-agent/status-builder-shared.ts +21 -84
  356. package/src/activities/execute-deep-agent/streaming-side-effects.ts +13 -19
  357. package/src/activities/execute-deep-agent/subagent-tracker.ts +77 -42
  358. package/src/activities/execute-deep-agent/subagent-transformer.ts +46 -105
  359. package/src/activities/execute-deep-agent/subagent-wiring.ts +5 -4
  360. package/src/activities/execute-deep-agent/turn-settle.ts +164 -0
  361. package/src/activities/execute-deep-agent/turn-setup.ts +541 -0
  362. package/src/activities/execute-deep-agent/turn-stream.ts +313 -0
  363. package/src/activities/execute-deep-agent/turn.ts +131 -0
  364. package/src/activities/execute-deep-agent/v3-event-recorder.ts +6 -5
  365. package/src/activities/execute-deep-agent/v3-events.ts +1 -1
  366. package/src/activities/execute-deep-agent/v3-status-builder.ts +84 -63
  367. package/src/harness/__tests__/approval-decisions.test.ts +116 -0
  368. package/src/harness/__tests__/capture.test.ts +268 -0
  369. package/src/harness/__tests__/mount-skills.test.ts +169 -0
  370. package/src/harness/__tests__/persist-chokepoint.test.ts +38 -2
  371. package/src/harness/__tests__/run-turn-deterministic.test.ts +94 -0
  372. package/src/harness/__tests__/run-turn.test.ts +39 -23
  373. package/src/harness/__tests__/turn-attachments-vision.test.ts +80 -0
  374. package/src/harness/__tests__/turn-context.test.ts +73 -26
  375. package/src/harness/approval-decisions.ts +104 -0
  376. package/src/harness/capabilities.ts +16 -5
  377. package/src/harness/capture.ts +301 -0
  378. package/src/harness/persist-chokepoint.ts +24 -4
  379. package/src/harness/run-turn.ts +120 -29
  380. package/src/harness/terminal-table.ts +61 -20
  381. package/src/harness/turn-context.ts +208 -111
  382. package/src/harness/types.ts +66 -12
  383. package/src/harness-adapters.ts +17 -9
  384. package/src/middleware/__tests__/approval-gate.test.ts +1 -1
  385. package/src/middleware/__tests__/cost-advisory.test.ts +113 -0
  386. package/src/middleware/approval-gate.ts +2 -2
  387. package/src/middleware/{cost-cap.ts → cost-advisory.ts} +40 -70
  388. package/src/middleware/index.ts +29 -29
  389. package/src/middleware/types.ts +3 -2
  390. package/src/runner-manager.ts +0 -8
  391. package/src/runner.ts +0 -8
  392. package/src/shared/__tests__/attachment-naming.test.ts +6 -19
  393. package/src/shared/__tests__/attachment-resolver.test.ts +158 -8
  394. package/src/shared/__tests__/attachment-zip.test.ts +236 -0
  395. package/src/shared/__tests__/bedrock-adapter.test.ts +1 -1
  396. package/src/shared/__tests__/blueprint-resolver.test.ts +43 -0
  397. package/src/shared/__tests__/cost-guard.test.ts +3 -3
  398. package/src/shared/__tests__/prompt-sections.test.ts +167 -0
  399. package/src/shared/__tests__/secret-leak-scan.test.ts +2 -2
  400. package/src/{activities/execute-deep-agent → shared}/__tests__/stamp-flowed-rows.test.ts +33 -10
  401. package/src/shared/__tests__/vertex-adapter.test.ts +1 -1
  402. package/src/shared/attachment-resolver.ts +218 -55
  403. package/src/shared/attachment-zip.ts +281 -0
  404. package/src/shared/connect-backfill.ts +9 -11
  405. package/src/shared/cost-guard.ts +3 -3
  406. package/src/shared/execution-status-writer.ts +5 -5
  407. package/src/shared/extract-structured-output.ts +8 -7
  408. package/src/shared/filereview/__tests__/capture.test.ts +86 -4
  409. package/src/shared/filereview/__tests__/cas-progress.test.ts +2 -1
  410. package/src/shared/filereview/cas-progress.ts +5 -19
  411. package/src/shared/filereview/cas-touched.ts +89 -0
  412. package/src/shared/filereview/index.ts +4 -2
  413. package/src/shared/mcp-resolver.ts +4 -5
  414. package/src/shared/persist-decision.ts +3 -4
  415. package/src/shared/plan-mode-permissions.ts +2 -2
  416. package/src/shared/plan-mode-prompt.ts +1 -1
  417. package/src/shared/prompt-sections.ts +256 -0
  418. package/src/shared/skill-resolver.ts +4 -2
  419. package/src/shared/tool-rounds.ts +48 -0
  420. package/src/shared/tool-row.ts +61 -0
  421. package/src/shared/workspace/__tests__/stigmer-link.test.ts +7 -22
  422. package/src/shared/workspace/__tests__/writeback-coordinator.test.ts +20 -61
  423. package/src/shared/workspace/session-provision.ts +3 -3
  424. package/src/shared/workspace/stigmer-link.ts +3 -3
  425. package/src/shared/workspace/writeback-coordinator.ts +16 -64
  426. package/src/tools/index.ts +1 -1
  427. package/dist/activities/execute-cursor/capture-flow.d.ts +0 -142
  428. package/dist/activities/execute-cursor/capture-flow.js +0 -285
  429. package/dist/activities/execute-cursor/capture-flow.js.map +0 -1
  430. package/dist/activities/execute-cursor/command-provenance.d.ts +0 -48
  431. package/dist/activities/execute-cursor/command-provenance.js +0 -38
  432. package/dist/activities/execute-cursor/command-provenance.js.map +0 -1
  433. package/dist/activities/execute-deep-agent/attachment-injector.d.ts +0 -134
  434. package/dist/activities/execute-deep-agent/attachment-injector.js +0 -386
  435. package/dist/activities/execute-deep-agent/attachment-injector.js.map +0 -1
  436. package/dist/activities/execute-deep-agent/command-provenance.d.ts +0 -61
  437. package/dist/activities/execute-deep-agent/command-provenance.js +0 -72
  438. package/dist/activities/execute-deep-agent/command-provenance.js.map +0 -1
  439. package/dist/activities/execute-deep-agent/environment.d.ts +0 -24
  440. package/dist/activities/execute-deep-agent/environment.js +0 -58
  441. package/dist/activities/execute-deep-agent/environment.js.map +0 -1
  442. package/dist/activities/execute-deep-agent/event-recorder.d.ts +0 -21
  443. package/dist/activities/execute-deep-agent/event-recorder.js +0 -67
  444. package/dist/activities/execute-deep-agent/event-recorder.js.map +0 -1
  445. package/dist/activities/execute-deep-agent/index.d.ts +0 -15
  446. package/dist/activities/execute-deep-agent/index.js +0 -918
  447. package/dist/activities/execute-deep-agent/index.js.map +0 -1
  448. package/dist/activities/execute-deep-agent/mcp-gate.d.ts +0 -28
  449. package/dist/activities/execute-deep-agent/mcp-gate.js +0 -22
  450. package/dist/activities/execute-deep-agent/mcp-gate.js.map +0 -1
  451. package/dist/activities/execute-deep-agent/post-stream.d.ts +0 -23
  452. package/dist/activities/execute-deep-agent/post-stream.js +0 -71
  453. package/dist/activities/execute-deep-agent/post-stream.js.map +0 -1
  454. package/dist/activities/execute-deep-agent/setup.d.ts +0 -136
  455. package/dist/activities/execute-deep-agent/setup.js +0 -767
  456. package/dist/activities/execute-deep-agent/setup.js.map +0 -1
  457. package/dist/activities/execute-deep-agent/stamp-flowed-rows.d.ts +0 -36
  458. package/dist/activities/execute-deep-agent/stamp-flowed-rows.js +0 -56
  459. package/dist/activities/execute-deep-agent/stamp-flowed-rows.js.map +0 -1
  460. package/dist/activities/execute-deep-agent/status-builder.d.ts +0 -93
  461. package/dist/activities/execute-deep-agent/status-builder.js +0 -365
  462. package/dist/activities/execute-deep-agent/status-builder.js.map +0 -1
  463. package/dist/activities/execute-deep-agent/streaming-terminal.d.ts +0 -21
  464. package/dist/activities/execute-deep-agent/streaming-terminal.js +0 -80
  465. package/dist/activities/execute-deep-agent/streaming-terminal.js.map +0 -1
  466. package/dist/activities/execute-deep-agent/streaming-v3.d.ts +0 -13
  467. package/dist/activities/execute-deep-agent/streaming-v3.js +0 -174
  468. package/dist/activities/execute-deep-agent/streaming-v3.js.map +0 -1
  469. package/dist/activities/execute-deep-agent/streaming.d.ts +0 -81
  470. package/dist/activities/execute-deep-agent/streaming.js +0 -160
  471. package/dist/activities/execute-deep-agent/streaming.js.map +0 -1
  472. package/dist/middleware/cost-cap.d.ts +0 -22
  473. package/dist/middleware/cost-cap.js.map +0 -1
  474. package/dist/middleware/graceful-stop.d.ts +0 -17
  475. package/dist/middleware/graceful-stop.js +0 -63
  476. package/dist/middleware/graceful-stop.js.map +0 -1
  477. package/dist/shared/skill-writer.d.ts +0 -75
  478. package/dist/shared/skill-writer.js +0 -207
  479. package/dist/shared/skill-writer.js.map +0 -1
  480. package/src/activities/execute-cursor/__tests__/capture-flow.test.ts +0 -1032
  481. package/src/activities/execute-cursor/__tests__/command-provenance.test.ts +0 -240
  482. package/src/activities/execute-cursor/__tests__/progress-substrate.test.ts +0 -169
  483. package/src/activities/execute-cursor/capture-flow.ts +0 -379
  484. package/src/activities/execute-cursor/command-provenance.ts +0 -73
  485. package/src/activities/execute-deep-agent/__tests__/attachment-injector.test.ts +0 -1161
  486. package/src/activities/execute-deep-agent/__tests__/command-provenance.test.ts +0 -252
  487. package/src/activities/execute-deep-agent/__tests__/environment.test.ts +0 -108
  488. package/src/activities/execute-deep-agent/__tests__/event-recorder.test.ts +0 -150
  489. package/src/activities/execute-deep-agent/__tests__/hitl-integration.test.ts +0 -226
  490. package/src/activities/execute-deep-agent/__tests__/hitl-reject.test.ts +0 -318
  491. package/src/activities/execute-deep-agent/__tests__/hitl-resume-approve-all.test.ts +0 -351
  492. package/src/activities/execute-deep-agent/__tests__/hitl-resume-history.test.ts +0 -293
  493. package/src/activities/execute-deep-agent/__tests__/index.test.ts +0 -99
  494. package/src/activities/execute-deep-agent/__tests__/mcp-gate.test.ts +0 -44
  495. package/src/activities/execute-deep-agent/__tests__/post-stream.test.ts +0 -112
  496. package/src/activities/execute-deep-agent/__tests__/sequential-gate-resume.test.ts +0 -358
  497. package/src/activities/execute-deep-agent/__tests__/status-builder.test.ts +0 -1933
  498. package/src/activities/execute-deep-agent/__tests__/streaming-terminal.test.ts +0 -40
  499. package/src/activities/execute-deep-agent/__tests__/streaming-v3.test.ts +0 -597
  500. package/src/activities/execute-deep-agent/__tests__/streaming.test.ts +0 -508
  501. package/src/activities/execute-deep-agent/attachment-injector.ts +0 -627
  502. package/src/activities/execute-deep-agent/command-provenance.ts +0 -102
  503. package/src/activities/execute-deep-agent/environment.ts +0 -76
  504. package/src/activities/execute-deep-agent/event-recorder.ts +0 -95
  505. package/src/activities/execute-deep-agent/index.ts +0 -1072
  506. package/src/activities/execute-deep-agent/mcp-gate.ts +0 -37
  507. package/src/activities/execute-deep-agent/post-stream.ts +0 -109
  508. package/src/activities/execute-deep-agent/setup.ts +0 -1081
  509. package/src/activities/execute-deep-agent/stamp-flowed-rows.ts +0 -64
  510. package/src/activities/execute-deep-agent/status-builder.ts +0 -481
  511. package/src/activities/execute-deep-agent/streaming-terminal.ts +0 -106
  512. package/src/activities/execute-deep-agent/streaming-v3.ts +0 -277
  513. package/src/activities/execute-deep-agent/streaming.ts +0 -313
  514. package/src/middleware/__tests__/cost-cap.test.ts +0 -192
  515. package/src/middleware/__tests__/graceful-stop.test.ts +0 -105
  516. package/src/middleware/graceful-stop.ts +0 -86
  517. package/src/shared/__tests__/skill-writer.test.ts +0 -372
  518. package/src/shared/skill-writer.ts +0 -266
@@ -0,0 +1,225 @@
1
+ /**
2
+ * The native (deep-agent) slice of the hermetic activity driver: build the
3
+ * fixture record, bind the scripted model and the `Config`, run the REAL
4
+ * `ExecuteDeepAgent` activity under `MockActivityEnvironment` with the REAL
5
+ * deepagents graph, and hand back what it persisted.
6
+ *
7
+ * Layering (the `approval-contract/` + `gateway-substrate.ts` split): everything
8
+ * harness-agnostic lives in `src/__test-utils__/hermetic-activity.ts` (the
9
+ * activity environment, the execution record behind the client, the temp
10
+ * `~/.stigmer`, the ticking clock, the shutdown signal) and its fixtures
11
+ * (`execution-record-fixture.ts`, `model-registry-fixture.ts`,
12
+ * `git-workspace-fixture.ts`). This module adds only what is native's: the
13
+ * scripted model as the one LLM (`scripted-model.ts`, reached through the
14
+ * `shared/model-client.ts` double), the native `Config` slice with its
15
+ * checkpointer, the per-scenario reset of module caches and of the session's
16
+ * durable state, and the production wire shape of the activity input.
17
+ *
18
+ * What the two `vi.mock` calls a scenario file must carry look like — they are
19
+ * hoisted by vitest and so cannot live here:
20
+ *
21
+ * ```ts
22
+ * vi.mock("../../../../shared/model-client.js", async () =>
23
+ * (await import("../../__test-utils__/scripted-model-module.js")).scriptedModelClientModule());
24
+ * vi.mock("../../../../client/stigmer-client.js", async () =>
25
+ * (await import("../../../../__test-utils__/hermetic-activity.js")).hermeticStigmerClientModule());
26
+ * ```
27
+ *
28
+ * Two facts about this harness the driver encodes, both read from production:
29
+ *
30
+ * - `thread_id` is `thread-{sessionId}` on EVERY invocation, the first
31
+ * included (`invoke-agent-execution.ts` runs `EnsureThread` once and passes
32
+ * the same id to `firstInvoke` and every `reinvoke`; `ensure-thread.ts`
33
+ * derives it from the session). Cursor's driver sends `""` on a first turn
34
+ * because Cursor's engine mints its own id; copying that here would key the
35
+ * sqlite checkpoint on `""` in turn 1 and look it up under `thread-…` in
36
+ * turn 2. A scenario overrides the default only to stage the empty-id edge.
37
+ * - The `memory` checkpointer is recreated empty per invocation
38
+ * (`checkpointer/factory.ts`), so a HITL reinvocation REPLAYS the graph; the
39
+ * `sqlite` checkpointer keys one file per session under `~/.stigmer`, so a
40
+ * reinvocation RESUMES from the interrupt — the durable posture with no
41
+ * network. A scenario picks one; the HITL arms record both.
42
+ *
43
+ * The ONE knowledge of production this driver holds is `activityFactory`: how
44
+ * to obtain the `ExecuteDeepAgent` function from a `Config`. Since S3 M2a it
45
+ * is the harness registry's: the native adapter booted and bound to its
46
+ * byte-pinned activity name through `createHarnessActivities`, exactly the
47
+ * way the composition roots bind it since S3 M2b through `harness-adapters.ts`
48
+ * (and the way the Cursor driver binds its own, `hermetic-cursor.ts`). The
49
+ * orchestrator this adapter replaced (`index.ts`, `setup.ts`) was deleted at
50
+ * M2b; the goldens moved once, with the M2a flip, each hunk quoted with its
51
+ * ruling in its test's header (Q-M2a-1), and the switch of the roots moved
52
+ * none (Q-M2b-4).
53
+ *
54
+ * `boot` registers the deepagents harness profiles as production does
55
+ * (S3 M2a F-M2a-14: the orchestrator-driven runs never did, so deepagents
56
+ * auto-injected its ungated `general-purpose` sub-agent into every M0
57
+ * scenario; the scripted model never delegated to it).
58
+ *
59
+ * Determinism beyond S0's: the clock ticks once per model turn (the scripted
60
+ * model's `onTurn`), so status timestamps and the persist cadence land on the
61
+ * same instants every run; a scenario's interruption (pause, worker shutdown)
62
+ * is scheduled on a model turn through the same hook, never on a timer.
63
+ */
64
+ import { rmSync } from "node:fs";
65
+ import { dirname, join } from "node:path";
66
+ import { getCheckpointDbPath } from "../../../shared/workspace/platform-dir.js";
67
+ import { _resetRegistryCache } from "../../../shared/model-registry.js";
68
+ import { _resetPricingCache } from "../../../shared/model-pricing-data.js";
69
+ import { bindHermeticClient, runActivityHermetically, } from "../../../__test-utils__/hermetic-activity.js";
70
+ import { executionRecordFixture } from "../../../__test-utils__/execution-record-fixture.js";
71
+ import { FIXTURE_NATIVE_MODEL } from "../../../__test-utils__/model-registry-fixture.js";
72
+ import { ScriptedModel } from "./scripted-model.js";
73
+ import { bindScriptedModel } from "./scripted-model-module.js";
74
+ // ---------------------------------------------------------------------------
75
+ // Fixture ids — fixed so goldens are readable and byte-stable
76
+ // ---------------------------------------------------------------------------
77
+ const SESSION_ID = "ses_hermetic_0001";
78
+ export const FIXTURE = {
79
+ org: "hermetic-org",
80
+ executionId: "aex_hermetic_0001",
81
+ sessionId: SESSION_ID,
82
+ agentInstanceId: "ain_hermetic_0001",
83
+ agentId: "agt_hermetic_0001",
84
+ agentName: "hermetic-agent",
85
+ /** The pinned model; in the registry document as the native entry. */
86
+ model: FIXTURE_NATIVE_MODEL,
87
+ /** The LangGraph thread id the control plane sends on every invocation (`ensure-thread.ts`). */
88
+ threadId: `thread-${SESSION_ID}`,
89
+ };
90
+ /** The four resources of one execution, wired by id into a chain, under the native fixture ids. */
91
+ export function deepAgentExecutionRecord(options) {
92
+ return executionRecordFixture({
93
+ ...options,
94
+ modelName: options.modelName ?? FIXTURE.model,
95
+ ids: {
96
+ org: FIXTURE.org,
97
+ executionId: FIXTURE.executionId,
98
+ sessionId: FIXTURE.sessionId,
99
+ agentInstanceId: FIXTURE.agentInstanceId,
100
+ agentId: FIXTURE.agentId,
101
+ agentName: FIXTURE.agentName,
102
+ },
103
+ });
104
+ }
105
+ /** The OSS/local runner posture: no proxy, local mode, the scenario's checkpointer. */
106
+ export function hermeticDeepAgentConfig(env, checkpointerType) {
107
+ return {
108
+ taskQueue: "hermetic-test-queue",
109
+ temporalAddress: "localhost:7233",
110
+ temporalNamespace: "default",
111
+ stigmerBackendEndpoint: "http://localhost:7234",
112
+ mcpBridgeEndpoint: null,
113
+ stigmerToken: null,
114
+ cursorApiKey: "",
115
+ workspaceRootDir: env.workspaceRootDir,
116
+ mode: "local",
117
+ proxyEndpoint: null,
118
+ maxConcurrentActivities: 1,
119
+ idleTimeoutSeconds: null,
120
+ cloudModeEnabled: false,
121
+ checkpointerType,
122
+ checkpointerProxyEndpoint: null,
123
+ artifactProxyEndpoint: null,
124
+ primaryModel: FIXTURE.model,
125
+ cursorStreamStallTimeoutMs: 180_000,
126
+ agentResolveTimeoutMs: 120_000,
127
+ workspaceLockTimeoutMs: 900_000,
128
+ };
129
+ }
130
+ /**
131
+ * The primary workspace directory a session with NO workspace entries runs in
132
+ * (`session-root.ts`: `<workspaceRootDir>/sessions/<sessionId>`). A scenario
133
+ * seeds the files its script reads here.
134
+ */
135
+ export function sessionWorkspaceDir(env) {
136
+ return join(env.workspaceRootDir, "sessions", FIXTURE.sessionId);
137
+ }
138
+ /**
139
+ * The adapter through the registry: `createDeepAgentAdapter()` booted on the
140
+ * scenario's config, bound to `ExecuteDeepAgent` by `createHarnessActivities`
141
+ * over the turn runtime. Imported lazily so the scenario file's `vi.mock`
142
+ * declarations are in force before the engine slice loads the model client
143
+ * and before the registry loads the `StigmerClient` module.
144
+ */
145
+ export const runtimeActivityFactory = async (config) => {
146
+ const { createDeepAgentAdapter } = await import("../adapter.js");
147
+ const { createHarnessActivities } = await import("../../../harness/registry.js");
148
+ const adapter = createDeepAgentAdapter();
149
+ await adapter.boot(config);
150
+ const activities = await createHarnessActivities([{ harness: "deep-agent", adapter }], config);
151
+ return activities.ExecuteDeepAgent;
152
+ };
153
+ /**
154
+ * Drop every module-level cache the harness keeps across invocations and
155
+ * every durable trace the session left behind (its sqlite checkpoint, its
156
+ * session workspace), so a scenario starts where a new session starts.
157
+ * Called at the START of a scenario, never between its invocations: a HITL
158
+ * scenario relies on the sqlite checkpoint surviving from turn 1 into turn 2
159
+ * exactly as it does in a live runner.
160
+ */
161
+ export function resetDeepAgentScenarioState(env) {
162
+ _resetRegistryCache();
163
+ _resetPricingCache();
164
+ rmSync(dirname(getCheckpointDbPath(FIXTURE.sessionId)), { recursive: true, force: true });
165
+ rmSync(sessionWorkspaceDir(env), { recursive: true, force: true });
166
+ }
167
+ /** One scenario: the record, the one LLM, the runner it runs on, and the invocation in flight. */
168
+ export class DeepAgentScenario {
169
+ env;
170
+ clock;
171
+ record;
172
+ config;
173
+ /** The controls of the invocation in flight; set by {@link runDeepAgentTurn} before the activity starts. */
174
+ controls;
175
+ model;
176
+ constructor(env, clock, record,
177
+ /** The runner this scenario runs on: the OSS posture plus the scenario's knobs. */
178
+ config, script, onTurn) {
179
+ this.env = env;
180
+ this.clock = clock;
181
+ this.record = record;
182
+ this.config = config;
183
+ this.model = new ScriptedModel(script, [], {
184
+ onTurn: async (info) => {
185
+ this.clock.tick();
186
+ if (onTurn && this.controls)
187
+ await onTurn(info, this.controls);
188
+ },
189
+ });
190
+ }
191
+ }
192
+ /** Bind the record and the model for a scenario; resets the harness's module and session state first. */
193
+ export function beginDeepAgentScenario(options) {
194
+ resetDeepAgentScenarioState(options.env);
195
+ const scenario = new DeepAgentScenario(options.env, options.clock, options.record, { ...hermeticDeepAgentConfig(options.env, options.checkpointer ?? "memory"), ...options.config }, options.script, options.onTurn);
196
+ bindScriptedModel(scenario.model);
197
+ bindHermeticClient(options.record.client(options.clientOverrides));
198
+ return scenario;
199
+ }
200
+ /**
201
+ * Run ONE `ExecuteDeepAgent` invocation for the scenario. Built per invocation
202
+ * (the factory constructs its client, which is the one bound at
203
+ * `beginDeepAgentScenario`).
204
+ */
205
+ export async function runDeepAgentTurn(scenario, options = {}) {
206
+ const activity = await (options.activityFactory ?? runtimeActivityFactory)(scenario.config);
207
+ const threadId = options.threadId ?? FIXTURE.threadId;
208
+ const turnSeq = options.turnSeq ?? 0;
209
+ const input = {
210
+ execution_id: scenario.record.executionId,
211
+ thread_id: threadId,
212
+ turn_seq: turnSeq,
213
+ };
214
+ const args = options.positional
215
+ ? [scenario.record.executionId, threadId]
216
+ : [input];
217
+ return runActivityHermetically(activity, args, {
218
+ taskQueue: scenario.config.taskQueue,
219
+ onControls: (controls) => {
220
+ scenario.controls = controls;
221
+ options.onControls?.(controls);
222
+ },
223
+ });
224
+ }
225
+ //# sourceMappingURL=hermetic-deep-agent.js.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"hermetic-deep-agent.js","sourceRoot":"","sources":["../../../../src/activities/execute-deep-agent/__test-utils__/hermetic-deep-agent.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA8DG;AAEH,OAAO,EAAE,MAAM,EAAE,MAAM,SAAS,CAAC;AACjC,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,MAAM,WAAW,CAAC;AAI1C,OAAO,EAAE,mBAAmB,EAAE,MAAM,2CAA2C,CAAC;AAChF,OAAO,EAAE,mBAAmB,EAAE,MAAM,mCAAmC,CAAC;AACxE,OAAO,EAAE,kBAAkB,EAAE,MAAM,uCAAuC,CAAC;AAC3E,OAAO,EAGL,kBAAkB,EAClB,uBAAuB,GAIxB,MAAM,8CAA8C,CAAC;AACtD,OAAO,EAAE,sBAAsB,EAA+B,MAAM,qDAAqD,CAAC;AAC1H,OAAO,EAAE,oBAAoB,EAAE,MAAM,mDAAmD,CAAC;AACzF,OAAO,EAAE,aAAa,EAA8C,MAAM,qBAAqB,CAAC;AAChG,OAAO,EAAE,iBAAiB,EAAE,MAAM,4BAA4B,CAAC;AAE/D,8EAA8E;AAC9E,8DAA8D;AAC9D,8EAA8E;AAE9E,MAAM,UAAU,GAAG,mBAAmB,CAAC;AAEvC,MAAM,CAAC,MAAM,OAAO,GAAG;IACrB,GAAG,EAAE,cAAc;IACnB,WAAW,EAAE,mBAAmB;IAChC,SAAS,EAAE,UAAU;IACrB,eAAe,EAAE,mBAAmB;IACpC,OAAO,EAAE,mBAAmB;IAC5B,SAAS,EAAE,gBAAgB;IAC3B,sEAAsE;IACtE,KAAK,EAAE,oBAAoB;IAC3B,gGAAgG;IAChG,QAAQ,EAAE,UAAU,UAAU,EAAE;CACxB,CAAC;AASX,mGAAmG;AACnG,MAAM,UAAU,wBAAwB,CAAC,OAA+B;IACtE,OAAO,sBAAsB,CAAC;QAC5B,GAAG,OAAO;QACV,SAAS,EAAE,OAAO,CAAC,SAAS,IAAI,OAAO,CAAC,KAAK;QAC7C,GAAG,EAAE;YACH,GAAG,EAAE,OAAO,CAAC,GAAG;YAChB,WAAW,EAAE,OAAO,CAAC,WAAW;YAChC,SAAS,EAAE,OAAO,CAAC,SAAS;YAC5B,eAAe,EAAE,OAAO,CAAC,eAAe;YACxC,OAAO,EAAE,OAAO,CAAC,OAAO;YACxB,SAAS,EAAE,OAAO,CAAC,SAAS;SAC7B;KACF,CAAC,CAAC;AACL,CAAC;AAQD,uFAAuF;AACvF,MAAM,UAAU,uBAAuB,CAAC,GAAwB,EAAE,gBAAsC;IACtG,OAAO;QACL,SAAS,EAAE,qBAAqB;QAChC,eAAe,EAAE,gBAAgB;QACjC,iBAAiB,EAAE,SAAS;QAC5B,sBAAsB,EAAE,uBAAuB;QAC/C,iBAAiB,EAAE,IAAI;QACvB,YAAY,EAAE,IAAI;QAClB,YAAY,EAAE,EAAE;QAChB,gBAAgB,EAAE,GAAG,CAAC,gBAAgB;QACtC,IAAI,EAAE,OAAO;QACb,aAAa,EAAE,IAAI;QACnB,uBAAuB,EAAE,CAAC;QAC1B,kBAAkB,EAAE,IAAI;QACxB,gBAAgB,EAAE,KAAK;QACvB,gBAAgB;QAChB,yBAAyB,EAAE,IAAI;QAC/B,qBAAqB,EAAE,IAAI;QAC3B,YAAY,EAAE,OAAO,CAAC,KAAK;QAC3B,0BAA0B,EAAE,OAAO;QACnC,qBAAqB,EAAE,OAAO;QAC9B,sBAAsB,EAAE,OAAO;KAChC,CAAC;AACJ,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,mBAAmB,CAAC,GAAwB;IAC1D,OAAO,IAAI,CAAC,GAAG,CAAC,gBAAgB,EAAE,UAAU,EAAE,OAAO,CAAC,SAAS,CAAC,CAAC;AACnE,CAAC;AAYD;;;;;;GAMG;AACH,MAAM,CAAC,MAAM,sBAAsB,GAAoB,KAAK,EAAE,MAAM,EAAE,EAAE;IACtE,MAAM,EAAE,sBAAsB,EAAE,GAAG,MAAM,MAAM,CAAC,eAAe,CAAC,CAAC;IACjE,MAAM,EAAE,uBAAuB,EAAE,GAAG,MAAM,MAAM,CAAC,8BAA8B,CAAC,CAAC;IACjF,MAAM,OAAO,GAAG,sBAAsB,EAAE,CAAC;IACzC,MAAM,OAAO,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;IAC3B,MAAM,UAAU,GAAG,MAAM,uBAAuB,CAAC,CAAC,EAAE,OAAO,EAAE,YAAY,EAAE,OAAO,EAAE,CAAC,EAAE,MAAM,CAAC,CAAC;IAC/F,OAAO,UAAU,CAAC,gBAAiB,CAAC;AACtC,CAAC,CAAC;AAEF;;;;;;;GAOG;AACH,MAAM,UAAU,2BAA2B,CAAC,GAAwB;IAClE,mBAAmB,EAAE,CAAC;IACtB,kBAAkB,EAAE,CAAC;IACrB,MAAM,CAAC,OAAO,CAAC,mBAAmB,CAAC,OAAO,CAAC,SAAS,CAAC,CAAC,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,KAAK,EAAE,IAAI,EAAE,CAAC,CAAC;IAC1F,MAAM,CAAC,mBAAmB,CAAC,GAAG,CAAC,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,KAAK,EAAE,IAAI,EAAE,CAAC,CAAC;AACrE,CAAC;AAED,kGAAkG;AAClG,MAAM,OAAO,iBAAiB;IAMjB;IACA;IACA;IAEA;IATX,4GAA4G;IAC5G,QAAQ,CAAiC;IAChC,KAAK,CAAgB;IAE9B,YACW,GAAwB,EACxB,KAAoB,EACpB,MAAuB;IAChC,mFAAmF;IAC1E,MAAc,EACvB,MAAsB,EACtB,MAA0C;QANjC,QAAG,GAAH,GAAG,CAAqB;QACxB,UAAK,GAAL,KAAK,CAAe;QACpB,WAAM,GAAN,MAAM,CAAiB;QAEvB,WAAM,GAAN,MAAM,CAAQ;QAIvB,IAAI,CAAC,KAAK,GAAG,IAAI,aAAa,CAAC,MAAM,EAAE,EAAE,EAAE;YACzC,MAAM,EAAE,KAAK,EAAE,IAAI,EAAE,EAAE;gBACrB,IAAI,CAAC,KAAK,CAAC,IAAI,EAAE,CAAC;gBAClB,IAAI,MAAM,IAAI,IAAI,CAAC,QAAQ;oBAAE,MAAM,MAAM,CAAC,IAAI,EAAE,IAAI,CAAC,QAAQ,CAAC,CAAC;YACjE,CAAC;SACF,CAAC,CAAC;IACL,CAAC;CACF;AA0BD,yGAAyG;AACzG,MAAM,UAAU,sBAAsB,CAAC,OAAiC;IACtE,2BAA2B,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC;IACzC,MAAM,QAAQ,GAAG,IAAI,iBAAiB,CACpC,OAAO,CAAC,GAAG,EACX,OAAO,CAAC,KAAK,EACb,OAAO,CAAC,MAAM,EACd,EAAE,GAAG,uBAAuB,CAAC,OAAO,CAAC,GAAG,EAAE,OAAO,CAAC,YAAY,IAAI,QAAQ,CAAC,EAAE,GAAG,OAAO,CAAC,MAAM,EAAE,EAChG,OAAO,CAAC,MAAM,EACd,OAAO,CAAC,MAAM,CACf,CAAC;IACF,iBAAiB,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC;IAClC,kBAAkB,CAAC,OAAO,CAAC,MAAM,CAAC,MAAM,CAAC,OAAO,CAAC,eAAe,CAAC,CAAC,CAAC;IACnE,OAAO,QAAQ,CAAC;AAClB,CAAC;AAaD;;;;GAIG;AACH,MAAM,CAAC,KAAK,UAAU,gBAAgB,CACpC,QAA2B,EAC3B,UAAgC,EAAE;IAElC,MAAM,QAAQ,GAAG,MAAM,CAAC,OAAO,CAAC,eAAe,IAAI,sBAAsB,CAAC,CAAC,QAAQ,CAAC,MAAM,CAAC,CAAC;IAC5F,MAAM,QAAQ,GAAG,OAAO,CAAC,QAAQ,IAAI,OAAO,CAAC,QAAQ,CAAC;IACtD,MAAM,OAAO,GAAG,OAAO,CAAC,OAAO,IAAI,CAAC,CAAC;IACrC,MAAM,KAAK,GAAyB;QAClC,YAAY,EAAE,QAAQ,CAAC,MAAM,CAAC,WAAW;QACzC,SAAS,EAAE,QAAQ;QACnB,QAAQ,EAAE,OAAO;KAClB,CAAC;IACF,MAAM,IAAI,GAA6C,OAAO,CAAC,UAAU;QACvE,CAAC,CAAC,CAAC,QAAQ,CAAC,MAAM,CAAC,WAAW,EAAE,QAAQ,CAAC;QACzC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC;IACZ,OAAO,uBAAuB,CAAC,QAAQ,EAAE,IAAI,EAAE;QAC7C,SAAS,EAAE,QAAQ,CAAC,MAAM,CAAC,SAAS;QACpC,UAAU,EAAE,CAAC,QAAQ,EAAE,EAAE;YACvB,QAAQ,CAAC,QAAQ,GAAG,QAAQ,CAAC;YAC7B,OAAO,CAAC,UAAU,EAAE,CAAC,QAAQ,CAAC,CAAC;QACjC,CAAC;KACF,CAAC,CAAC;AACL,CAAC"}
@@ -0,0 +1,24 @@
1
+ /**
2
+ * The one gated opening every HITL arm shares, so their first-invocation
3
+ * statuses are byte-identical and can match ONE golden file
4
+ * (`__tests__/hermetic/goldens/gate.turn1.status.json`). Matching the same
5
+ * file from the approve, skip, reject and sequential-gates arms is itself an
6
+ * assertion: the pending decision is invisible to the first invocation.
7
+ *
8
+ * Why `execute`: under capture mode (on whenever a workspace dir and artifact
9
+ * storage exist — every hermetic run) file writes FLOW into the CAS ledger
10
+ * and are reviewed post hoc; shell is the built-in that still interrupts
11
+ * (`middleware/approval-gate.ts`, `shared/tool-kind.ts` SHELL). `echo` is a
12
+ * real subprocess with byte-stable output, the same tool
13
+ * `__tests__/shell-execute-gate.test.ts` drives through the real gate.
14
+ */
15
+ import type { ScriptedToolCall, ScriptedTurn } from "./scripted-model.js";
16
+ export declare const EXECUTE_CALL_A: ScriptedToolCall;
17
+ export declare const EXECUTE_CALL_B: ScriptedToolCall;
18
+ /** The gate's copy for a SHELL built-in (`CATEGORY_APPROVAL_MESSAGE.shell` over the args). */
19
+ export declare function executeApprovalMessage(call: ScriptedToolCall): string;
20
+ /** Round 0 of every gated arm: one text block and the gated `execute` A. */
21
+ export declare const GATED_OPENING_TURN: ScriptedTurn;
22
+ /** The closing text after the decision has been carried out (or declined). */
23
+ export declare const CLOSING_TURN: ScriptedTurn;
24
+ export declare const GATE_TURN1_GOLDEN = "./goldens/gate.turn1.status.json";
@@ -0,0 +1,41 @@
1
+ /**
2
+ * The one gated opening every HITL arm shares, so their first-invocation
3
+ * statuses are byte-identical and can match ONE golden file
4
+ * (`__tests__/hermetic/goldens/gate.turn1.status.json`). Matching the same
5
+ * file from the approve, skip, reject and sequential-gates arms is itself an
6
+ * assertion: the pending decision is invisible to the first invocation.
7
+ *
8
+ * Why `execute`: under capture mode (on whenever a workspace dir and artifact
9
+ * storage exist — every hermetic run) file writes FLOW into the CAS ledger
10
+ * and are reviewed post hoc; shell is the built-in that still interrupts
11
+ * (`middleware/approval-gate.ts`, `shared/tool-kind.ts` SHELL). `echo` is a
12
+ * real subprocess with byte-stable output, the same tool
13
+ * `__tests__/shell-execute-gate.test.ts` drives through the real gate.
14
+ */
15
+ export const EXECUTE_CALL_A = {
16
+ id: "call-hermetic-exec-0001",
17
+ name: "execute",
18
+ args: { command: "echo hermetic-a" },
19
+ };
20
+ export const EXECUTE_CALL_B = {
21
+ id: "call-hermetic-exec-0002",
22
+ name: "execute",
23
+ args: { command: "echo hermetic-b" },
24
+ };
25
+ /** The gate's copy for a SHELL built-in (`CATEGORY_APPROVAL_MESSAGE.shell` over the args). */
26
+ export function executeApprovalMessage(call) {
27
+ return `Execute command: ${String(call.args.command)}`;
28
+ }
29
+ /** Round 0 of every gated arm: one text block and the gated `execute` A. */
30
+ export const GATED_OPENING_TURN = {
31
+ text: "I will run the command.",
32
+ toolCalls: [EXECUTE_CALL_A],
33
+ usage: { inputTokens: 1_400, outputTokens: 50 },
34
+ };
35
+ /** The closing text after the decision has been carried out (or declined). */
36
+ export const CLOSING_TURN = {
37
+ text: "Done with the command.",
38
+ usage: { inputTokens: 1_600, outputTokens: 20 },
39
+ };
40
+ export const GATE_TURN1_GOLDEN = "./goldens/gate.turn1.status.json";
41
+ //# sourceMappingURL=hitl-script.js.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"hitl-script.js","sourceRoot":"","sources":["../../../../src/activities/execute-deep-agent/__test-utils__/hitl-script.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AAIH,MAAM,CAAC,MAAM,cAAc,GAAqB;IAC9C,EAAE,EAAE,yBAAyB;IAC7B,IAAI,EAAE,SAAS;IACf,IAAI,EAAE,EAAE,OAAO,EAAE,iBAAiB,EAAE;CACrC,CAAC;AAEF,MAAM,CAAC,MAAM,cAAc,GAAqB;IAC9C,EAAE,EAAE,yBAAyB;IAC7B,IAAI,EAAE,SAAS;IACf,IAAI,EAAE,EAAE,OAAO,EAAE,iBAAiB,EAAE;CACrC,CAAC;AAEF,8FAA8F;AAC9F,MAAM,UAAU,sBAAsB,CAAC,IAAsB;IAC3D,OAAO,oBAAoB,MAAM,CAAC,IAAI,CAAC,IAAI,CAAC,OAAO,CAAC,EAAE,CAAC;AACzD,CAAC;AAED,4EAA4E;AAC5E,MAAM,CAAC,MAAM,kBAAkB,GAAiB;IAC9C,IAAI,EAAE,yBAAyB;IAC/B,SAAS,EAAE,CAAC,cAAc,CAAC;IAC3B,KAAK,EAAE,EAAE,WAAW,EAAE,KAAK,EAAE,YAAY,EAAE,EAAE,EAAE;CAChD,CAAC;AAEF,8EAA8E;AAC9E,MAAM,CAAC,MAAM,YAAY,GAAiB;IACxC,IAAI,EAAE,wBAAwB;IAC9B,KAAK,EAAE,EAAE,WAAW,EAAE,KAAK,EAAE,YAAY,EAAE,EAAE,EAAE;CAChD,CAAC;AAEF,MAAM,CAAC,MAAM,iBAAiB,GAAG,kCAAkC,CAAC"}
@@ -0,0 +1,54 @@
1
+ /**
2
+ * The module-boundary double for `shared/model-client.ts` — how a
3
+ * `ScriptedModel` reaches the REAL native engine build.
4
+ *
5
+ * `buildChatModel` is the runner's single construction point for LangChain
6
+ * chat models, and a native turn reaches it three ways: the primary model
7
+ * (`turn-setup.ts` `buildEngine`), every sub-agent's `modelFactory` (the same
8
+ * `buildModelFor`), and the runtime's tier-2 structured-output extractor
9
+ * (`shared/extract-structured-output.ts`). None of them takes a model as an
10
+ * argument, so the one seam a hermetic run has is this module — the native analog of the Cursor
11
+ * driver's `vi.mock("@cursor/sdk")`: the vendor engine doubled where the
12
+ * runner imports it, and nothing else (S3 M0 ruling Q-M0-1, amending
13
+ * Q-S3-11).
14
+ *
15
+ * `vi.mock` is hoisted and must live in the test file; the factory can
16
+ * `await import()` this module:
17
+ *
18
+ * ```ts
19
+ * vi.mock("../../../../shared/model-client.js", async () =>
20
+ * (await import("../../__test-utils__/scripted-model-module.js")).scriptedModelClientModule(),
21
+ * );
22
+ * ```
23
+ *
24
+ * Every build the activity requests is recorded (`builds`) so a scenario can
25
+ * assert WHAT production asked for — the registry id, the service tier, the
26
+ * proxy posture — while the model handed back is the scenario's script. The
27
+ * provider is inferred from the requested name exactly as the real module
28
+ * does, so a golden's provider-labelled copy reads as in production.
29
+ *
30
+ * A scenario binds ONE model (every build of every execution gets it) or a
31
+ * RESOLVER over the build's options: production scopes every build to its
32
+ * execution through `headerScope.executionId` (the proxy's attribution
33
+ * header), so a test that runs several executions at once — the harness
34
+ * contract kit's concurrency invariant — routes each build to that
35
+ * execution's own engine through the id production already sends, and
36
+ * through nothing added for the test (S3 M3).
37
+ */
38
+ import type { BuildChatModelOptions, BuiltChatModel } from "../../../shared/model-client.js";
39
+ import type { ScriptedModel } from "./scripted-model.js";
40
+ /** Which `ScriptedModel` answers one `buildChatModel` call. */
41
+ export type ScriptedModelResolver = (opts: BuildChatModelOptions) => ScriptedModel;
42
+ /**
43
+ * Bind the model every `buildChatModel` call hands back for the current
44
+ * scenario — or a resolver that picks one per call — and forget the previous
45
+ * scenario's build log. Called by the driver at `beginDeepAgentScenario`,
46
+ * before the activity factory runs.
47
+ */
48
+ export declare function bindScriptedModel(model: ScriptedModel | ScriptedModelResolver): void;
49
+ /** The `buildChatModel` requests the activity made in this scenario, in order. */
50
+ export declare function recordedModelBuilds(): readonly BuildChatModelOptions[];
51
+ /** The factory a test passes to `vi.mock(".../shared/model-client.js", ...)`. */
52
+ export declare function scriptedModelClientModule(): {
53
+ buildChatModel: (opts: BuildChatModelOptions) => Promise<BuiltChatModel>;
54
+ };
@@ -0,0 +1,72 @@
1
+ /**
2
+ * The module-boundary double for `shared/model-client.ts` — how a
3
+ * `ScriptedModel` reaches the REAL native engine build.
4
+ *
5
+ * `buildChatModel` is the runner's single construction point for LangChain
6
+ * chat models, and a native turn reaches it three ways: the primary model
7
+ * (`turn-setup.ts` `buildEngine`), every sub-agent's `modelFactory` (the same
8
+ * `buildModelFor`), and the runtime's tier-2 structured-output extractor
9
+ * (`shared/extract-structured-output.ts`). None of them takes a model as an
10
+ * argument, so the one seam a hermetic run has is this module — the native analog of the Cursor
11
+ * driver's `vi.mock("@cursor/sdk")`: the vendor engine doubled where the
12
+ * runner imports it, and nothing else (S3 M0 ruling Q-M0-1, amending
13
+ * Q-S3-11).
14
+ *
15
+ * `vi.mock` is hoisted and must live in the test file; the factory can
16
+ * `await import()` this module:
17
+ *
18
+ * ```ts
19
+ * vi.mock("../../../../shared/model-client.js", async () =>
20
+ * (await import("../../__test-utils__/scripted-model-module.js")).scriptedModelClientModule(),
21
+ * );
22
+ * ```
23
+ *
24
+ * Every build the activity requests is recorded (`builds`) so a scenario can
25
+ * assert WHAT production asked for — the registry id, the service tier, the
26
+ * proxy posture — while the model handed back is the scenario's script. The
27
+ * provider is inferred from the requested name exactly as the real module
28
+ * does, so a golden's provider-labelled copy reads as in production.
29
+ *
30
+ * A scenario binds ONE model (every build of every execution gets it) or a
31
+ * RESOLVER over the build's options: production scopes every build to its
32
+ * execution through `headerScope.executionId` (the proxy's attribution
33
+ * header), so a test that runs several executions at once — the harness
34
+ * contract kit's concurrency invariant — routes each build to that
35
+ * execution's own engine through the id production already sends, and
36
+ * through nothing added for the test (S3 M3).
37
+ */
38
+ import { stripProviderPrefix, tryInferProvider } from "../../../shared/llm-proxy.js";
39
+ let resolve;
40
+ const builds = [];
41
+ /**
42
+ * Bind the model every `buildChatModel` call hands back for the current
43
+ * scenario — or a resolver that picks one per call — and forget the previous
44
+ * scenario's build log. Called by the driver at `beginDeepAgentScenario`,
45
+ * before the activity factory runs.
46
+ */
47
+ export function bindScriptedModel(model) {
48
+ resolve = typeof model === "function" ? model : () => model;
49
+ builds.length = 0;
50
+ }
51
+ /** The `buildChatModel` requests the activity made in this scenario, in order. */
52
+ export function recordedModelBuilds() {
53
+ return builds;
54
+ }
55
+ /** The factory a test passes to `vi.mock(".../shared/model-client.js", ...)`. */
56
+ export function scriptedModelClientModule() {
57
+ return {
58
+ buildChatModel: async (opts) => {
59
+ if (!resolve) {
60
+ throw new Error("scripted-model-module: no ScriptedModel bound — call bindScriptedModel(model) " +
61
+ "before constructing the activities");
62
+ }
63
+ builds.push(opts);
64
+ return {
65
+ model: resolve(opts),
66
+ provider: tryInferProvider(opts.modelName) ?? "anthropic",
67
+ apiModelId: stripProviderPrefix(opts.modelName),
68
+ };
69
+ },
70
+ };
71
+ }
72
+ //# sourceMappingURL=scripted-model-module.js.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"scripted-model-module.js","sourceRoot":"","sources":["../../../../src/activities/execute-deep-agent/__test-utils__/scripted-model-module.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAoCG;AAGH,OAAO,EAAE,mBAAmB,EAAE,gBAAgB,EAAE,MAAM,8BAA8B,CAAC;AAMrF,IAAI,OAA0C,CAAC;AAC/C,MAAM,MAAM,GAA4B,EAAE,CAAC;AAE3C;;;;;GAKG;AACH,MAAM,UAAU,iBAAiB,CAAC,KAA4C;IAC5E,OAAO,GAAG,OAAO,KAAK,KAAK,UAAU,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,GAAG,EAAE,CAAC,KAAK,CAAC;IAC5D,MAAM,CAAC,MAAM,GAAG,CAAC,CAAC;AACpB,CAAC;AAED,kFAAkF;AAClF,MAAM,UAAU,mBAAmB;IACjC,OAAO,MAAM,CAAC;AAChB,CAAC;AAED,iFAAiF;AACjF,MAAM,UAAU,yBAAyB;IAGvC,OAAO;QACL,cAAc,EAAE,KAAK,EAAE,IAAI,EAAE,EAAE;YAC7B,IAAI,CAAC,OAAO,EAAE,CAAC;gBACb,MAAM,IAAI,KAAK,CACb,gFAAgF;oBAC9E,oCAAoC,CACvC,CAAC;YACJ,CAAC;YACD,MAAM,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;YAClB,OAAO;gBACL,KAAK,EAAE,OAAO,CAAC,IAAI,CAAC;gBACpB,QAAQ,EAAE,gBAAgB,CAAC,IAAI,CAAC,SAAS,CAAC,IAAI,WAAW;gBACzD,UAAU,EAAE,mBAAmB,CAAC,IAAI,CAAC,SAAS,CAAC;aAChD,CAAC;QACJ,CAAC;KACF,CAAC;AACJ,CAAC"}