self-bench 0.3.6 → 0.3.8

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 (1126) hide show
  1. package/.dockerignore +11 -1
  2. package/Dockerfile +9 -9
  3. package/Dockerfile.sandbox +5 -4
  4. package/README.md +27 -12
  5. package/bun.lock +462 -95
  6. package/compose.yaml +64 -20
  7. package/dist/agent-feed.d.ts +8 -0
  8. package/dist/agent-feed.d.ts.map +1 -0
  9. package/dist/agent-feed.js +56 -0
  10. package/dist/agent-feed.js.map +1 -0
  11. package/dist/api/http.d.ts +18 -0
  12. package/dist/api/http.d.ts.map +1 -0
  13. package/dist/api/http.js +111 -0
  14. package/dist/api/http.js.map +1 -0
  15. package/dist/api/run-request.d.ts +37 -0
  16. package/dist/api/run-request.d.ts.map +1 -0
  17. package/dist/api/run-request.js +59 -0
  18. package/dist/api/run-request.js.map +1 -0
  19. package/dist/api/status.d.ts +4 -0
  20. package/dist/api/status.d.ts.map +1 -0
  21. package/dist/api/status.js +36 -0
  22. package/dist/api/status.js.map +1 -0
  23. package/dist/api/viewer-routes.d.ts +9 -0
  24. package/dist/api/viewer-routes.d.ts.map +1 -0
  25. package/dist/api/viewer-routes.js +98 -0
  26. package/dist/api/viewer-routes.js.map +1 -0
  27. package/dist/api-main.js +5 -1
  28. package/dist/api-main.js.map +1 -1
  29. package/dist/api.d.ts +7 -25
  30. package/dist/api.d.ts.map +1 -1
  31. package/dist/api.js +80 -148
  32. package/dist/api.js.map +1 -1
  33. package/dist/archive.d.ts +14 -0
  34. package/dist/archive.d.ts.map +1 -0
  35. package/dist/archive.js +114 -0
  36. package/dist/archive.js.map +1 -0
  37. package/dist/artifacts/common.d.ts +15 -0
  38. package/dist/artifacts/common.d.ts.map +1 -0
  39. package/dist/artifacts/common.js +58 -0
  40. package/dist/artifacts/common.js.map +1 -0
  41. package/dist/artifacts/gcs.d.ts +18 -0
  42. package/dist/artifacts/gcs.d.ts.map +1 -0
  43. package/dist/artifacts/gcs.js +178 -0
  44. package/dist/artifacts/gcs.js.map +1 -0
  45. package/dist/artifacts/local.d.ts +17 -0
  46. package/dist/artifacts/local.d.ts.map +1 -0
  47. package/dist/artifacts/local.js +205 -0
  48. package/dist/artifacts/local.js.map +1 -0
  49. package/dist/artifacts/types.d.ts +18 -0
  50. package/dist/artifacts/types.d.ts.map +1 -0
  51. package/dist/artifacts/types.js +2 -0
  52. package/dist/artifacts/types.js.map +1 -0
  53. package/dist/artifacts.d.ts +5 -28
  54. package/dist/artifacts.d.ts.map +1 -1
  55. package/dist/artifacts.js +5 -302
  56. package/dist/artifacts.js.map +1 -1
  57. package/dist/audit.d.ts.map +1 -1
  58. package/dist/audit.js +8 -2
  59. package/dist/audit.js.map +1 -1
  60. package/dist/auth/api-key-routes.d.ts +17 -0
  61. package/dist/auth/api-key-routes.d.ts.map +1 -0
  62. package/dist/auth/api-key-routes.js +57 -0
  63. package/dist/auth/api-key-routes.js.map +1 -0
  64. package/dist/auth/api-keys.d.ts +47 -0
  65. package/dist/auth/api-keys.d.ts.map +1 -0
  66. package/dist/auth/api-keys.js +102 -0
  67. package/dist/auth/api-keys.js.map +1 -0
  68. package/dist/auth/config.d.ts +17 -0
  69. package/dist/auth/config.d.ts.map +1 -0
  70. package/dist/auth/config.js +38 -0
  71. package/dist/auth/config.js.map +1 -0
  72. package/dist/auth/cookies.d.ts +12 -0
  73. package/dist/auth/cookies.d.ts.map +1 -0
  74. package/dist/auth/cookies.js +45 -0
  75. package/dist/auth/cookies.js.map +1 -0
  76. package/dist/auth/crypto.d.ts +19 -0
  77. package/dist/auth/crypto.d.ts.map +1 -0
  78. package/dist/auth/crypto.js +48 -0
  79. package/dist/auth/crypto.js.map +1 -0
  80. package/dist/auth/github.d.ts +37 -0
  81. package/dist/auth/github.d.ts.map +1 -0
  82. package/dist/auth/github.js +127 -0
  83. package/dist/auth/github.js.map +1 -0
  84. package/dist/auth/routes.d.ts +41 -0
  85. package/dist/auth/routes.d.ts.map +1 -0
  86. package/dist/auth/routes.js +153 -0
  87. package/dist/auth/routes.js.map +1 -0
  88. package/dist/auth/session-expired.d.ts +4 -0
  89. package/dist/auth/session-expired.d.ts.map +1 -0
  90. package/dist/auth/session-expired.js +17 -0
  91. package/dist/auth/session-expired.js.map +1 -0
  92. package/dist/auth/session.d.ts +23 -0
  93. package/dist/auth/session.d.ts.map +1 -0
  94. package/dist/auth/session.js +56 -0
  95. package/dist/auth/session.js.map +1 -0
  96. package/dist/auth/users.d.ts +49 -0
  97. package/dist/auth/users.d.ts.map +1 -0
  98. package/dist/auth/users.js +125 -0
  99. package/dist/auth/users.js.map +1 -0
  100. package/dist/build-metadata.js +1 -1
  101. package/dist/cli/api-client.d.ts +8 -0
  102. package/dist/cli/api-client.d.ts.map +1 -0
  103. package/dist/cli/api-client.js +71 -0
  104. package/dist/cli/api-client.js.map +1 -0
  105. package/dist/cli/associate.d.ts +2 -0
  106. package/dist/cli/associate.d.ts.map +1 -0
  107. package/dist/cli/associate.js +51 -0
  108. package/dist/cli/associate.js.map +1 -0
  109. package/dist/cli/help.d.ts +2 -0
  110. package/dist/cli/help.d.ts.map +1 -0
  111. package/dist/cli/help.js +52 -0
  112. package/dist/cli/help.js.map +1 -0
  113. package/dist/cli/main.d.ts +2 -0
  114. package/dist/cli/main.d.ts.map +1 -0
  115. package/dist/cli/main.js +55 -0
  116. package/dist/cli/main.js.map +1 -0
  117. package/dist/cli/replay.d.ts +3 -0
  118. package/dist/cli/replay.d.ts.map +1 -0
  119. package/dist/cli/replay.js +59 -0
  120. package/dist/cli/replay.js.map +1 -0
  121. package/dist/cli/repository.d.ts +6 -0
  122. package/dist/cli/repository.d.ts.map +1 -0
  123. package/dist/cli/repository.js +36 -0
  124. package/dist/cli/repository.js.map +1 -0
  125. package/dist/cli/run.d.ts +2 -0
  126. package/dist/cli/run.d.ts.map +1 -0
  127. package/dist/cli/run.js +93 -0
  128. package/dist/cli/run.js.map +1 -0
  129. package/dist/cli/stack-environment.d.ts +26 -0
  130. package/dist/cli/stack-environment.d.ts.map +1 -0
  131. package/dist/cli/stack-environment.js +86 -0
  132. package/dist/cli/stack-environment.js.map +1 -0
  133. package/dist/cli/stack.d.ts +6 -0
  134. package/dist/cli/stack.d.ts.map +1 -0
  135. package/dist/cli/stack.js +131 -0
  136. package/dist/cli/stack.js.map +1 -0
  137. package/dist/cli/values.d.ts +8 -0
  138. package/dist/cli/values.d.ts.map +1 -0
  139. package/dist/cli/values.js +30 -0
  140. package/dist/cli/values.js.map +1 -0
  141. package/dist/cli/view.d.ts +2 -0
  142. package/dist/cli/view.d.ts.map +1 -0
  143. package/dist/cli/view.js +40 -0
  144. package/dist/cli/view.js.map +1 -0
  145. package/dist/cli.js +2 -429
  146. package/dist/cli.js.map +1 -1
  147. package/dist/config.d.ts +16 -2
  148. package/dist/config.d.ts.map +1 -1
  149. package/dist/config.js +60 -22
  150. package/dist/config.js.map +1 -1
  151. package/dist/contracts/common.d.ts +22 -0
  152. package/dist/contracts/common.d.ts.map +1 -0
  153. package/dist/contracts/common.js +29 -0
  154. package/dist/contracts/common.js.map +1 -0
  155. package/dist/contracts/results.d.ts +66 -0
  156. package/dist/contracts/results.d.ts.map +1 -0
  157. package/dist/contracts/results.js +2 -0
  158. package/dist/contracts/results.js.map +1 -0
  159. package/dist/contracts/run.d.ts +145 -0
  160. package/dist/contracts/run.d.ts.map +1 -0
  161. package/dist/contracts/run.js +74 -0
  162. package/dist/contracts/run.js.map +1 -0
  163. package/dist/contracts/task.d.ts +232 -0
  164. package/dist/contracts/task.d.ts.map +1 -0
  165. package/dist/contracts/task.js +198 -0
  166. package/dist/contracts/task.js.map +1 -0
  167. package/dist/contracts/verify.d.ts +331 -0
  168. package/dist/contracts/verify.d.ts.map +1 -0
  169. package/dist/contracts/verify.js +84 -0
  170. package/dist/contracts/verify.js.map +1 -0
  171. package/dist/contracts.d.ts +5 -243
  172. package/dist/contracts.d.ts.map +1 -1
  173. package/dist/contracts.js +5 -128
  174. package/dist/contracts.js.map +1 -1
  175. package/dist/db/client.d.ts +12 -0
  176. package/dist/db/client.d.ts.map +1 -0
  177. package/dist/db/client.js +16 -0
  178. package/dist/db/client.js.map +1 -0
  179. package/dist/db/schema.d.ts +1256 -0
  180. package/dist/db/schema.d.ts.map +1 -0
  181. package/dist/db/schema.js +128 -0
  182. package/dist/db/schema.js.map +1 -0
  183. package/dist/environment.d.ts +9 -0
  184. package/dist/environment.d.ts.map +1 -0
  185. package/dist/environment.js +101 -0
  186. package/dist/environment.js.map +1 -0
  187. package/dist/evaluation/account.d.ts +30 -0
  188. package/dist/evaluation/account.d.ts.map +1 -0
  189. package/dist/evaluation/account.js +34 -0
  190. package/dist/evaluation/account.js.map +1 -0
  191. package/dist/evaluation/activities.d.ts +9 -0
  192. package/dist/evaluation/activities.d.ts.map +1 -0
  193. package/dist/evaluation/activities.js +40 -0
  194. package/dist/evaluation/activities.js.map +1 -0
  195. package/dist/evaluation/catalog-pricing.d.ts +3 -0
  196. package/dist/evaluation/catalog-pricing.d.ts.map +1 -0
  197. package/dist/evaluation/catalog-pricing.js +37 -0
  198. package/dist/evaluation/catalog-pricing.js.map +1 -0
  199. package/dist/evaluation/catalog.d.ts +17 -0
  200. package/dist/evaluation/catalog.d.ts.map +1 -0
  201. package/dist/evaluation/catalog.js +126 -0
  202. package/dist/evaluation/catalog.js.map +1 -0
  203. package/dist/evaluation/codex-login-process.d.ts +12 -0
  204. package/dist/evaluation/codex-login-process.d.ts.map +1 -0
  205. package/dist/evaluation/codex-login-process.js +121 -0
  206. package/dist/evaluation/codex-login-process.js.map +1 -0
  207. package/dist/evaluation/codex-login-routes.d.ts +5 -0
  208. package/dist/evaluation/codex-login-routes.d.ts.map +1 -0
  209. package/dist/evaluation/codex-login-routes.js +55 -0
  210. package/dist/evaluation/codex-login-routes.js.map +1 -0
  211. package/dist/evaluation/codex-login.d.ts +42 -0
  212. package/dist/evaluation/codex-login.d.ts.map +1 -0
  213. package/dist/evaluation/codex-login.js +143 -0
  214. package/dist/evaluation/codex-login.js.map +1 -0
  215. package/dist/evaluation/comparisons.d.ts +65 -0
  216. package/dist/evaluation/comparisons.d.ts.map +1 -0
  217. package/dist/evaluation/comparisons.js +161 -0
  218. package/dist/evaluation/comparisons.js.map +1 -0
  219. package/dist/evaluation/config.d.ts +43 -0
  220. package/dist/evaluation/config.d.ts.map +1 -0
  221. package/dist/evaluation/config.js +116 -0
  222. package/dist/evaluation/config.js.map +1 -0
  223. package/dist/evaluation/cost.d.ts +3 -0
  224. package/dist/evaluation/cost.d.ts.map +1 -0
  225. package/dist/evaluation/cost.js +95 -0
  226. package/dist/evaluation/cost.js.map +1 -0
  227. package/dist/evaluation/credential-execution.d.ts +10 -0
  228. package/dist/evaluation/credential-execution.d.ts.map +1 -0
  229. package/dist/evaluation/credential-execution.js +65 -0
  230. package/dist/evaluation/credential-execution.js.map +1 -0
  231. package/dist/evaluation/credentials.d.ts +40 -0
  232. package/dist/evaluation/credentials.d.ts.map +1 -0
  233. package/dist/evaluation/credentials.js +146 -0
  234. package/dist/evaluation/credentials.js.map +1 -0
  235. package/dist/evaluation/encrypted-records.d.ts +15 -0
  236. package/dist/evaluation/encrypted-records.d.ts.map +1 -0
  237. package/dist/evaluation/encrypted-records.js +57 -0
  238. package/dist/evaluation/encrypted-records.js.map +1 -0
  239. package/dist/evaluation/gateway-execution.d.ts +18 -0
  240. package/dist/evaluation/gateway-execution.d.ts.map +1 -0
  241. package/dist/evaluation/gateway-execution.js +38 -0
  242. package/dist/evaluation/gateway-execution.js.map +1 -0
  243. package/dist/evaluation/harbor-cost.d.ts +3 -0
  244. package/dist/evaluation/harbor-cost.d.ts.map +1 -0
  245. package/dist/evaluation/harbor-cost.js +39 -0
  246. package/dist/evaluation/harbor-cost.js.map +1 -0
  247. package/dist/evaluation/harbor_e2b.py +20 -0
  248. package/dist/evaluation/harbor_gateway.py +22 -0
  249. package/dist/evaluation/harnesses.d.ts +8 -0
  250. package/dist/evaluation/harnesses.d.ts.map +1 -0
  251. package/dist/evaluation/harnesses.js +10 -0
  252. package/dist/evaluation/harnesses.js.map +1 -0
  253. package/dist/evaluation/model-options.d.ts +9 -0
  254. package/dist/evaluation/model-options.d.ts.map +1 -0
  255. package/dist/evaluation/model-options.js +81 -0
  256. package/dist/evaluation/model-options.js.map +1 -0
  257. package/dist/evaluation/org-credential-routes.d.ts +5 -0
  258. package/dist/evaluation/org-credential-routes.d.ts.map +1 -0
  259. package/dist/evaluation/org-credential-routes.js +56 -0
  260. package/dist/evaluation/org-credential-routes.js.map +1 -0
  261. package/dist/evaluation/org-records.d.ts +4 -0
  262. package/dist/evaluation/org-records.d.ts.map +1 -0
  263. package/dist/evaluation/org-records.js +14 -0
  264. package/dist/evaluation/org-records.js.map +1 -0
  265. package/dist/evaluation/output.d.ts +9 -0
  266. package/dist/evaluation/output.d.ts.map +1 -0
  267. package/dist/evaluation/output.js +171 -0
  268. package/dist/evaluation/output.js.map +1 -0
  269. package/dist/evaluation/platform-routes.d.ts +5 -0
  270. package/dist/evaluation/platform-routes.d.ts.map +1 -0
  271. package/dist/evaluation/platform-routes.js +123 -0
  272. package/dist/evaluation/platform-routes.js.map +1 -0
  273. package/dist/evaluation/profile-routes.d.ts +16 -0
  274. package/dist/evaluation/profile-routes.d.ts.map +1 -0
  275. package/dist/evaluation/profile-routes.js +47 -0
  276. package/dist/evaluation/profile-routes.js.map +1 -0
  277. package/dist/evaluation/profiles.d.ts +48 -0
  278. package/dist/evaluation/profiles.d.ts.map +1 -0
  279. package/dist/evaluation/profiles.js +134 -0
  280. package/dist/evaluation/profiles.js.map +1 -0
  281. package/dist/evaluation/providers.d.ts +11 -0
  282. package/dist/evaluation/providers.d.ts.map +1 -0
  283. package/dist/evaluation/providers.js +30 -0
  284. package/dist/evaluation/providers.js.map +1 -0
  285. package/dist/evaluation/routes.d.ts +23 -0
  286. package/dist/evaluation/routes.d.ts.map +1 -0
  287. package/dist/evaluation/routes.js +212 -0
  288. package/dist/evaluation/routes.js.map +1 -0
  289. package/dist/evaluation/runner.d.ts +16 -0
  290. package/dist/evaluation/runner.d.ts.map +1 -0
  291. package/dist/evaluation/runner.js +241 -0
  292. package/dist/evaluation/runner.js.map +1 -0
  293. package/dist/evaluation/store.d.ts +8 -0
  294. package/dist/evaluation/store.d.ts.map +1 -0
  295. package/dist/evaluation/store.js +60 -0
  296. package/dist/evaluation/store.js.map +1 -0
  297. package/dist/evaluation/types.d.ts +98 -0
  298. package/dist/evaluation/types.d.ts.map +1 -0
  299. package/dist/evaluation/types.js +2 -0
  300. package/dist/evaluation/types.js.map +1 -0
  301. package/dist/evaluation/worker-main.d.ts +2 -0
  302. package/dist/evaluation/worker-main.d.ts.map +1 -0
  303. package/dist/evaluation/worker-main.js +26 -0
  304. package/dist/evaluation/worker-main.js.map +1 -0
  305. package/dist/evaluation/worker-records.d.ts +5 -0
  306. package/dist/evaluation/worker-records.d.ts.map +1 -0
  307. package/dist/evaluation/worker-records.js +11 -0
  308. package/dist/evaluation/worker-records.js.map +1 -0
  309. package/dist/evaluation/workflow.d.ts +3 -0
  310. package/dist/evaluation/workflow.d.ts.map +1 -0
  311. package/dist/evaluation/workflow.js +20 -0
  312. package/dist/evaluation/workflow.js.map +1 -0
  313. package/dist/execution-environment.d.ts +3 -0
  314. package/dist/execution-environment.d.ts.map +1 -0
  315. package/dist/execution-environment.js +9 -0
  316. package/dist/execution-environment.js.map +1 -0
  317. package/dist/extension-authoring.bundle.js +347 -0
  318. package/dist/extension-verifier.bundle.js +73 -0
  319. package/dist/extensions/authoring.d.ts.map +1 -1
  320. package/dist/extensions/authoring.js +70 -59
  321. package/dist/extensions/authoring.js.map +1 -1
  322. package/dist/extensions/shared/deliverable.d.ts +33 -0
  323. package/dist/extensions/shared/deliverable.d.ts.map +1 -0
  324. package/dist/extensions/shared/deliverable.js +123 -0
  325. package/dist/extensions/shared/deliverable.js.map +1 -0
  326. package/dist/extensions/shared/mailbox.d.ts +33 -0
  327. package/dist/extensions/shared/mailbox.d.ts.map +1 -0
  328. package/dist/extensions/shared/mailbox.js +102 -0
  329. package/dist/extensions/shared/mailbox.js.map +1 -0
  330. package/dist/extensions/shared/schemas.d.ts +90 -0
  331. package/dist/extensions/shared/schemas.d.ts.map +1 -0
  332. package/dist/extensions/shared/schemas.js +71 -0
  333. package/dist/extensions/shared/schemas.js.map +1 -0
  334. package/dist/extensions/shared/static-check.d.ts +38 -0
  335. package/dist/extensions/shared/static-check.d.ts.map +1 -0
  336. package/dist/extensions/shared/static-check.js +52 -0
  337. package/dist/extensions/shared/static-check.js.map +1 -0
  338. package/dist/extensions/verifier.d.ts +4 -0
  339. package/dist/extensions/verifier.d.ts.map +1 -0
  340. package/dist/extensions/verifier.js +70 -0
  341. package/dist/extensions/verifier.js.map +1 -0
  342. package/dist/github.d.ts +1 -1
  343. package/dist/github.d.ts.map +1 -1
  344. package/dist/github.js +3 -3
  345. package/dist/github.js.map +1 -1
  346. package/dist/harbor-environment.d.ts +13 -1
  347. package/dist/harbor-environment.d.ts.map +1 -1
  348. package/dist/harbor-environment.js +37 -4
  349. package/dist/harbor-environment.js.map +1 -1
  350. package/dist/harbor-results.d.ts +2 -0
  351. package/dist/harbor-results.d.ts.map +1 -1
  352. package/dist/harbor-results.js +2 -0
  353. package/dist/harbor-results.js.map +1 -1
  354. package/dist/harbor-task/compiler.d.ts +10 -0
  355. package/dist/harbor-task/compiler.d.ts.map +1 -0
  356. package/dist/harbor-task/compiler.js +174 -0
  357. package/dist/harbor-task/compiler.js.map +1 -0
  358. package/dist/harbor-task/constants.d.ts +3 -0
  359. package/dist/harbor-task/constants.d.ts.map +1 -0
  360. package/dist/harbor-task/constants.js +3 -0
  361. package/dist/harbor-task/constants.js.map +1 -0
  362. package/dist/harbor-task/dependencies.d.ts +3 -0
  363. package/dist/harbor-task/dependencies.d.ts.map +1 -0
  364. package/dist/harbor-task/dependencies.js +20 -0
  365. package/dist/harbor-task/dependencies.js.map +1 -0
  366. package/dist/harbor-task/paths.d.ts +6 -0
  367. package/dist/harbor-task/paths.d.ts.map +1 -0
  368. package/dist/harbor-task/paths.js +37 -0
  369. package/dist/harbor-task/paths.js.map +1 -0
  370. package/dist/harbor-task/render.d.ts +11 -0
  371. package/dist/harbor-task/render.d.ts.map +1 -0
  372. package/dist/harbor-task/render.js +194 -0
  373. package/dist/harbor-task/render.js.map +1 -0
  374. package/dist/harbor-task/runtime/command.sh +26 -0
  375. package/dist/harbor-task/runtime/junit.py +69 -0
  376. package/dist/harbor-task/runtime-assets.d.ts +2 -0
  377. package/dist/harbor-task/runtime-assets.d.ts.map +1 -0
  378. package/dist/harbor-task/runtime-assets.js +8 -0
  379. package/dist/harbor-task/runtime-assets.js.map +1 -0
  380. package/dist/harbor-task/test-patch.d.ts +4 -0
  381. package/dist/harbor-task/test-patch.d.ts.map +1 -0
  382. package/dist/harbor-task/test-patch.js +15 -0
  383. package/dist/harbor-task/test-patch.js.map +1 -0
  384. package/dist/harbor-task/verifier.d.ts +4 -0
  385. package/dist/harbor-task/verifier.d.ts.map +1 -0
  386. package/dist/harbor-task/verifier.js +105 -0
  387. package/dist/harbor-task/verifier.js.map +1 -0
  388. package/dist/harbor-task.d.ts +3 -11
  389. package/dist/harbor-task.d.ts.map +1 -1
  390. package/dist/harbor-task.js +2 -463
  391. package/dist/harbor-task.js.map +1 -1
  392. package/dist/log-excerpt.d.ts +15 -0
  393. package/dist/log-excerpt.d.ts.map +1 -0
  394. package/dist/log-excerpt.js +104 -0
  395. package/dist/log-excerpt.js.map +1 -0
  396. package/dist/patch-check.d.ts +26 -0
  397. package/dist/patch-check.d.ts.map +1 -0
  398. package/dist/patch-check.js +137 -0
  399. package/dist/patch-check.js.map +1 -0
  400. package/dist/pi-event-feed.d.ts +16 -0
  401. package/dist/pi-event-feed.d.ts.map +1 -0
  402. package/dist/pi-event-feed.js +86 -0
  403. package/dist/pi-event-feed.js.map +1 -0
  404. package/dist/pi-session.d.ts +32 -0
  405. package/dist/pi-session.d.ts.map +1 -0
  406. package/dist/pi-session.js +175 -0
  407. package/dist/pi-session.js.map +1 -0
  408. package/dist/process.d.ts +5 -0
  409. package/dist/process.d.ts.map +1 -1
  410. package/dist/process.js +31 -9
  411. package/dist/process.js.map +1 -1
  412. package/dist/provenance/github.d.ts +4 -0
  413. package/dist/provenance/github.d.ts.map +1 -0
  414. package/dist/provenance/github.js +65 -0
  415. package/dist/provenance/github.js.map +1 -0
  416. package/dist/provenance/local.d.ts +4 -0
  417. package/dist/provenance/local.d.ts.map +1 -0
  418. package/dist/provenance/local.js +82 -0
  419. package/dist/provenance/local.js.map +1 -0
  420. package/dist/provenance/redact.d.ts +2 -0
  421. package/dist/provenance/redact.d.ts.map +1 -0
  422. package/dist/provenance/redact.js +28 -0
  423. package/dist/provenance/redact.js.map +1 -0
  424. package/dist/provenance/selection.d.ts +4 -0
  425. package/dist/provenance/selection.d.ts.map +1 -0
  426. package/dist/provenance/selection.js +29 -0
  427. package/dist/provenance/selection.js.map +1 -0
  428. package/dist/provenance/session.d.ts +3 -0
  429. package/dist/provenance/session.d.ts.map +1 -0
  430. package/dist/provenance/session.js +214 -0
  431. package/dist/provenance/session.js.map +1 -0
  432. package/dist/provenance/shared.d.ts +7 -0
  433. package/dist/provenance/shared.d.ts.map +1 -0
  434. package/dist/provenance/shared.js +10 -0
  435. package/dist/provenance/shared.js.map +1 -0
  436. package/dist/provenance/types.d.ts +34 -0
  437. package/dist/provenance/types.d.ts.map +1 -0
  438. package/dist/provenance/types.js +22 -0
  439. package/dist/provenance/types.js.map +1 -0
  440. package/dist/provenance-associations/apply.d.ts +3 -0
  441. package/dist/provenance-associations/apply.d.ts.map +1 -0
  442. package/dist/provenance-associations/apply.js +66 -0
  443. package/dist/provenance-associations/apply.js.map +1 -0
  444. package/dist/provenance-associations/shared.d.ts +45 -0
  445. package/dist/provenance-associations/shared.d.ts.map +1 -0
  446. package/dist/provenance-associations/shared.js +42 -0
  447. package/dist/provenance-associations/shared.js.map +1 -0
  448. package/dist/provenance-associations.d.ts +3 -30
  449. package/dist/provenance-associations.d.ts.map +1 -1
  450. package/dist/provenance-associations.js +4 -103
  451. package/dist/provenance-associations.js.map +1 -1
  452. package/dist/provenance.d.ts +7 -41
  453. package/dist/provenance.d.ts.map +1 -1
  454. package/dist/provenance.js +6 -432
  455. package/dist/provenance.js.map +1 -1
  456. package/dist/providers.d.ts +25 -2
  457. package/dist/providers.d.ts.map +1 -1
  458. package/dist/providers.js +23 -2
  459. package/dist/providers.js.map +1 -1
  460. package/dist/repair.d.ts +1 -6
  461. package/dist/repair.d.ts.map +1 -1
  462. package/dist/repair.js +17 -27
  463. package/dist/repair.js.map +1 -1
  464. package/dist/review/assets/index-BHHztQ9x.js +1679 -0
  465. package/dist/review/assets/index-KQ1eFQ0t.css +1 -0
  466. package/dist/review/index.html +11 -8
  467. package/dist/runtime/command.sh +26 -0
  468. package/dist/runtime/junit.py +69 -0
  469. package/dist/sandbox/contracts.d.ts +36 -1
  470. package/dist/sandbox/contracts.d.ts.map +1 -1
  471. package/dist/sandbox/contracts.js +3 -0
  472. package/dist/sandbox/contracts.js.map +1 -1
  473. package/dist/sandbox/index.d.ts +1 -1
  474. package/dist/sandbox/index.d.ts.map +1 -1
  475. package/dist/sandbox/index.js +13 -2
  476. package/dist/sandbox/index.js.map +1 -1
  477. package/dist/sandbox/live.d.ts +18 -0
  478. package/dist/sandbox/live.d.ts.map +1 -0
  479. package/dist/sandbox/live.js +55 -0
  480. package/dist/sandbox/live.js.map +1 -0
  481. package/dist/sandbox/output-retry.d.ts +16 -0
  482. package/dist/sandbox/output-retry.d.ts.map +1 -0
  483. package/dist/sandbox/output-retry.js +36 -0
  484. package/dist/sandbox/output-retry.js.map +1 -0
  485. package/dist/sandbox/programs/author.js +16 -6
  486. package/dist/sandbox/programs/author.js.map +1 -1
  487. package/dist/sandbox/programs/check.d.ts +3 -0
  488. package/dist/sandbox/programs/check.d.ts.map +1 -0
  489. package/dist/sandbox/programs/check.js +76 -0
  490. package/dist/sandbox/programs/check.js.map +1 -0
  491. package/dist/sandbox/programs/prepare-task.d.ts +12 -0
  492. package/dist/sandbox/programs/prepare-task.d.ts.map +1 -0
  493. package/dist/sandbox/programs/{prepare-repair.js → prepare-task.js} +11 -9
  494. package/dist/sandbox/programs/prepare-task.js.map +1 -0
  495. package/dist/sandbox/programs/verifier.d.ts +3 -0
  496. package/dist/sandbox/programs/verifier.d.ts.map +1 -0
  497. package/dist/sandbox/programs/verifier.js +9 -0
  498. package/dist/sandbox/programs/verifier.js.map +1 -0
  499. package/dist/sandbox/providers/docker/executor.d.ts +4 -1
  500. package/dist/sandbox/providers/docker/executor.d.ts.map +1 -1
  501. package/dist/sandbox/providers/docker/executor.js +75 -25
  502. package/dist/sandbox/providers/docker/executor.js.map +1 -1
  503. package/dist/sandbox/providers/e2b/bytes.d.ts +3 -0
  504. package/dist/sandbox/providers/e2b/bytes.d.ts.map +1 -0
  505. package/dist/sandbox/providers/e2b/bytes.js +8 -0
  506. package/dist/sandbox/providers/e2b/bytes.js.map +1 -0
  507. package/dist/sandbox/providers/e2b/cleanup.d.ts +8 -0
  508. package/dist/sandbox/providers/e2b/cleanup.d.ts.map +1 -0
  509. package/dist/sandbox/providers/e2b/cleanup.js +124 -0
  510. package/dist/sandbox/providers/e2b/cleanup.js.map +1 -0
  511. package/dist/sandbox/providers/e2b/command.d.ts +23 -0
  512. package/dist/sandbox/providers/e2b/command.d.ts.map +1 -0
  513. package/dist/sandbox/providers/e2b/command.js +160 -0
  514. package/dist/sandbox/providers/e2b/command.js.map +1 -0
  515. package/dist/sandbox/providers/e2b/config.d.ts +13 -0
  516. package/dist/sandbox/providers/e2b/config.d.ts.map +1 -0
  517. package/dist/sandbox/providers/e2b/config.js +2 -0
  518. package/dist/sandbox/providers/e2b/config.js.map +1 -0
  519. package/dist/sandbox/providers/e2b/executor.d.ts +15 -0
  520. package/dist/sandbox/providers/e2b/executor.d.ts.map +1 -0
  521. package/dist/sandbox/providers/e2b/executor.js +207 -0
  522. package/dist/sandbox/providers/e2b/executor.js.map +1 -0
  523. package/dist/sandbox/providers/e2b/lifecycle.d.ts +8 -0
  524. package/dist/sandbox/providers/e2b/lifecycle.d.ts.map +1 -0
  525. package/dist/sandbox/providers/e2b/lifecycle.js +37 -0
  526. package/dist/sandbox/providers/e2b/lifecycle.js.map +1 -0
  527. package/dist/sandbox/providers/e2b/outcome.d.ts +11 -0
  528. package/dist/sandbox/providers/e2b/outcome.d.ts.map +1 -0
  529. package/dist/sandbox/providers/e2b/outcome.js +72 -0
  530. package/dist/sandbox/providers/e2b/outcome.js.map +1 -0
  531. package/dist/sandbox/providers/e2b/stage-files.d.ts +10 -0
  532. package/dist/sandbox/providers/e2b/stage-files.d.ts.map +1 -0
  533. package/dist/sandbox/providers/e2b/stage-files.js +45 -0
  534. package/dist/sandbox/providers/e2b/stage-files.js.map +1 -0
  535. package/dist/sandbox/providers/e2b/startup.d.ts +10 -0
  536. package/dist/sandbox/providers/e2b/startup.d.ts.map +1 -0
  537. package/dist/sandbox/providers/e2b/startup.js +26 -0
  538. package/dist/sandbox/providers/e2b/startup.js.map +1 -0
  539. package/dist/sandbox/providers/e2b/types.d.ts +9 -0
  540. package/dist/sandbox/providers/e2b/types.d.ts.map +1 -0
  541. package/dist/sandbox/providers/e2b/types.js +2 -0
  542. package/dist/sandbox/providers/e2b/types.js.map +1 -0
  543. package/dist/sandbox/providers/e2b/validation.d.ts +14 -0
  544. package/dist/sandbox/providers/e2b/validation.d.ts.map +1 -0
  545. package/dist/sandbox/providers/e2b/validation.js +60 -0
  546. package/dist/sandbox/providers/e2b/validation.js.map +1 -0
  547. package/dist/{modal-auth.d.ts → sandbox/providers/modal/auth.d.ts} +1 -1
  548. package/dist/sandbox/providers/modal/auth.d.ts.map +1 -0
  549. package/dist/{modal-auth.js → sandbox/providers/modal/auth.js} +1 -1
  550. package/dist/sandbox/providers/modal/auth.js.map +1 -0
  551. package/dist/sandbox/providers/modal/executor.d.ts +4 -1
  552. package/dist/sandbox/providers/modal/executor.d.ts.map +1 -1
  553. package/dist/sandbox/providers/modal/executor.js +33 -9
  554. package/dist/sandbox/providers/modal/executor.js.map +1 -1
  555. package/dist/sandbox/providers/modal/live.d.ts +7 -0
  556. package/dist/sandbox/providers/modal/live.d.ts.map +1 -0
  557. package/dist/sandbox/providers/modal/live.js +50 -0
  558. package/dist/sandbox/providers/modal/live.js.map +1 -0
  559. package/dist/sandbox/providers/vercel/command.d.ts +6 -0
  560. package/dist/sandbox/providers/vercel/command.d.ts.map +1 -1
  561. package/dist/sandbox/providers/vercel/command.js +44 -11
  562. package/dist/sandbox/providers/vercel/command.js.map +1 -1
  563. package/dist/sandbox/providers/vercel/executor.d.ts +6 -7
  564. package/dist/sandbox/providers/vercel/executor.d.ts.map +1 -1
  565. package/dist/sandbox/providers/vercel/executor.js +21 -187
  566. package/dist/sandbox/providers/vercel/executor.js.map +1 -1
  567. package/dist/sandbox/providers/vercel/lifecycle.d.ts +13 -0
  568. package/dist/sandbox/providers/vercel/lifecycle.d.ts.map +1 -0
  569. package/dist/sandbox/providers/vercel/lifecycle.js +140 -0
  570. package/dist/sandbox/providers/vercel/lifecycle.js.map +1 -0
  571. package/dist/sandbox/providers/vercel/validation.d.ts +11 -0
  572. package/dist/sandbox/providers/vercel/validation.d.ts.map +1 -0
  573. package/dist/sandbox/providers/vercel/validation.js +30 -0
  574. package/dist/sandbox/providers/vercel/validation.js.map +1 -0
  575. package/dist/sandbox/remote-files.d.ts +11 -0
  576. package/dist/sandbox/remote-files.d.ts.map +1 -0
  577. package/dist/sandbox/remote-files.js +41 -0
  578. package/dist/sandbox/remote-files.js.map +1 -0
  579. package/dist/sandbox/request-validation.d.ts +5 -0
  580. package/dist/sandbox/request-validation.d.ts.map +1 -0
  581. package/dist/sandbox/request-validation.js +43 -0
  582. package/dist/sandbox/request-validation.js.map +1 -0
  583. package/dist/sandbox/supervisor.d.ts +42 -0
  584. package/dist/sandbox/supervisor.d.ts.map +1 -0
  585. package/dist/sandbox/supervisor.js +126 -0
  586. package/dist/sandbox/supervisor.js.map +1 -0
  587. package/dist/sandbox/timeout.d.ts +6 -1
  588. package/dist/sandbox/timeout.d.ts.map +1 -1
  589. package/dist/sandbox/timeout.js +12 -0
  590. package/dist/sandbox/timeout.js.map +1 -1
  591. package/dist/sandbox-author.bundle.js +47 -14879
  592. package/dist/{sandbox-validation-repair.bundle.js → sandbox-check.bundle.js} +7261 -1893
  593. package/dist/sandbox-verifier.bundle.js +307 -0
  594. package/dist/setup/e2b/build.d.ts +24 -0
  595. package/dist/setup/e2b/build.d.ts.map +1 -0
  596. package/dist/setup/e2b/build.js +46 -0
  597. package/dist/setup/e2b/build.js.map +1 -0
  598. package/dist/setup/e2b/cli.d.ts +2 -0
  599. package/dist/setup/e2b/cli.d.ts.map +1 -0
  600. package/dist/setup/e2b/cli.js +51 -0
  601. package/dist/setup/e2b/cli.js.map +1 -0
  602. package/dist/setup/e2b/index.d.ts +5 -0
  603. package/dist/setup/e2b/index.d.ts.map +1 -0
  604. package/dist/setup/e2b/index.js +4 -0
  605. package/dist/setup/e2b/index.js.map +1 -0
  606. package/dist/setup/e2b/template.d.ts +4 -0
  607. package/dist/setup/e2b/template.d.ts.map +1 -0
  608. package/dist/setup/e2b/template.js +53 -0
  609. package/dist/setup/e2b/template.js.map +1 -0
  610. package/dist/setup/vercel/cli-helpers.d.ts +22 -0
  611. package/dist/setup/vercel/cli-helpers.d.ts.map +1 -0
  612. package/dist/setup/vercel/cli-helpers.js +87 -0
  613. package/dist/setup/vercel/cli-helpers.js.map +1 -0
  614. package/dist/setup/vercel/cli-runner.d.ts +14 -0
  615. package/dist/setup/vercel/cli-runner.d.ts.map +1 -0
  616. package/dist/setup/vercel/cli-runner.js +24 -0
  617. package/dist/setup/vercel/cli-runner.js.map +1 -0
  618. package/dist/setup/vercel/cli-types.d.ts +24 -0
  619. package/dist/setup/vercel/cli-types.d.ts.map +1 -0
  620. package/dist/setup/vercel/cli-types.js +18 -0
  621. package/dist/setup/vercel/cli-types.js.map +1 -0
  622. package/dist/setup/vercel/cli.d.ts +5 -31
  623. package/dist/setup/vercel/cli.d.ts.map +1 -1
  624. package/dist/setup/vercel/cli.js +15 -125
  625. package/dist/setup/vercel/cli.js.map +1 -1
  626. package/dist/setup/vercel/profile-files.d.ts +6 -0
  627. package/dist/setup/vercel/profile-files.d.ts.map +1 -0
  628. package/dist/setup/vercel/profile-files.js +100 -0
  629. package/dist/setup/vercel/profile-files.js.map +1 -0
  630. package/dist/setup/vercel/profile-schema.d.ts +48 -0
  631. package/dist/setup/vercel/profile-schema.d.ts.map +1 -0
  632. package/dist/setup/vercel/profile-schema.js +39 -0
  633. package/dist/setup/vercel/profile-schema.js.map +1 -0
  634. package/dist/setup/vercel/profile.d.ts +3 -19
  635. package/dist/setup/vercel/profile.d.ts.map +1 -1
  636. package/dist/setup/vercel/profile.js +4 -130
  637. package/dist/setup/vercel/profile.js.map +1 -1
  638. package/dist/setup/vercel/selection.d.ts +18 -0
  639. package/dist/setup/vercel/selection.d.ts.map +1 -0
  640. package/dist/setup/vercel/selection.js +113 -0
  641. package/dist/setup/vercel/selection.js.map +1 -0
  642. package/dist/setup/vercel/setup-types.d.ts +34 -0
  643. package/dist/setup/vercel/setup-types.d.ts.map +1 -0
  644. package/dist/setup/vercel/setup-types.js +2 -0
  645. package/dist/setup/vercel/setup-types.js.map +1 -0
  646. package/dist/setup/vercel/setup.d.ts +3 -28
  647. package/dist/setup/vercel/setup.d.ts.map +1 -1
  648. package/dist/setup/vercel/setup.js +3 -111
  649. package/dist/setup/vercel/setup.js.map +1 -1
  650. package/dist/site/batch-activity.d.ts +5 -0
  651. package/dist/site/batch-activity.d.ts.map +1 -0
  652. package/dist/site/batch-activity.js +31 -0
  653. package/dist/site/batch-activity.js.map +1 -0
  654. package/dist/site/batch-progress.d.ts +16 -0
  655. package/dist/site/batch-progress.d.ts.map +1 -0
  656. package/dist/site/batch-progress.js +52 -0
  657. package/dist/site/batch-progress.js.map +1 -0
  658. package/dist/site/batch-routes.d.ts +30 -0
  659. package/dist/site/batch-routes.d.ts.map +1 -0
  660. package/dist/site/batch-routes.js +117 -0
  661. package/dist/site/batch-routes.js.map +1 -0
  662. package/dist/site/batch-start.d.ts +60 -0
  663. package/dist/site/batch-start.d.ts.map +1 -0
  664. package/dist/site/batch-start.js +39 -0
  665. package/dist/site/batch-start.js.map +1 -0
  666. package/dist/site/connected-repos.d.ts +25 -0
  667. package/dist/site/connected-repos.d.ts.map +1 -0
  668. package/dist/site/connected-repos.js +109 -0
  669. package/dist/site/connected-repos.js.map +1 -0
  670. package/dist/site/evaluation-start.d.ts +4 -0
  671. package/dist/site/evaluation-start.d.ts.map +1 -0
  672. package/dist/site/evaluation-start.js +19 -0
  673. package/dist/site/evaluation-start.js.map +1 -0
  674. package/dist/site/generation-config.d.ts +4 -0
  675. package/dist/site/generation-config.d.ts.map +1 -0
  676. package/dist/site/generation-config.js +16 -0
  677. package/dist/site/generation-config.js.map +1 -0
  678. package/dist/site/generation-credentials.d.ts +6 -0
  679. package/dist/site/generation-credentials.d.ts.map +1 -0
  680. package/dist/site/generation-credentials.js +101 -0
  681. package/dist/site/generation-credentials.js.map +1 -0
  682. package/dist/site/generation-run.d.ts +6 -0
  683. package/dist/site/generation-run.d.ts.map +1 -0
  684. package/dist/site/generation-run.js +22 -0
  685. package/dist/site/generation-run.js.map +1 -0
  686. package/dist/site/generation-settings.d.ts +80 -0
  687. package/dist/site/generation-settings.d.ts.map +1 -0
  688. package/dist/site/generation-settings.js +84 -0
  689. package/dist/site/generation-settings.js.map +1 -0
  690. package/dist/site/generation-subscription.d.ts +3 -0
  691. package/dist/site/generation-subscription.d.ts.map +1 -0
  692. package/dist/site/generation-subscription.js +38 -0
  693. package/dist/site/generation-subscription.js.map +1 -0
  694. package/dist/site/github-repos.d.ts +35 -0
  695. package/dist/site/github-repos.d.ts.map +1 -0
  696. package/dist/site/github-repos.js +121 -0
  697. package/dist/site/github-repos.js.map +1 -0
  698. package/dist/site/pr-candidate.d.ts +26 -0
  699. package/dist/site/pr-candidate.d.ts.map +1 -0
  700. package/dist/site/pr-candidate.js +101 -0
  701. package/dist/site/pr-candidate.js.map +1 -0
  702. package/dist/site/pr-list.d.ts +13 -0
  703. package/dist/site/pr-list.d.ts.map +1 -0
  704. package/dist/site/pr-list.js +24 -0
  705. package/dist/site/pr-list.js.map +1 -0
  706. package/dist/site/pr-routes.d.ts +26 -0
  707. package/dist/site/pr-routes.d.ts.map +1 -0
  708. package/dist/site/pr-routes.js +139 -0
  709. package/dist/site/pr-routes.js.map +1 -0
  710. package/dist/site/repo-store.d.ts +40 -0
  711. package/dist/site/repo-store.d.ts.map +1 -0
  712. package/dist/site/repo-store.js +74 -0
  713. package/dist/site/repo-store.js.map +1 -0
  714. package/dist/site/run-store.d.ts +17 -0
  715. package/dist/site/run-store.d.ts.map +1 -0
  716. package/dist/site/run-store.js +42 -0
  717. package/dist/site/run-store.js.map +1 -0
  718. package/dist/site/runtime.d.ts +27 -0
  719. package/dist/site/runtime.d.ts.map +1 -0
  720. package/dist/site/runtime.js +97 -0
  721. package/dist/site/runtime.js.map +1 -0
  722. package/dist/site/task-deletion.d.ts +7 -0
  723. package/dist/site/task-deletion.d.ts.map +1 -0
  724. package/dist/site/task-deletion.js +29 -0
  725. package/dist/site/task-deletion.js.map +1 -0
  726. package/dist/site/task-start-reservation.d.ts +4 -0
  727. package/dist/site/task-start-reservation.d.ts.map +1 -0
  728. package/dist/site/task-start-reservation.js +18 -0
  729. package/dist/site/task-start-reservation.js.map +1 -0
  730. package/dist/site/task-start.d.ts +33 -0
  731. package/dist/site/task-start.d.ts.map +1 -0
  732. package/dist/site/task-start.js +50 -0
  733. package/dist/site/task-start.js.map +1 -0
  734. package/dist/site/task-status.d.ts +36 -0
  735. package/dist/site/task-status.d.ts.map +1 -0
  736. package/dist/site/task-status.js +113 -0
  737. package/dist/site/task-status.js.map +1 -0
  738. package/dist/site/task-store.d.ts +78 -0
  739. package/dist/site/task-store.d.ts.map +1 -0
  740. package/dist/site/task-store.js +186 -0
  741. package/dist/site/task-store.js.map +1 -0
  742. package/dist/site/task-sync.d.ts +30 -0
  743. package/dist/site/task-sync.d.ts.map +1 -0
  744. package/dist/site/task-sync.js +121 -0
  745. package/dist/site/task-sync.js.map +1 -0
  746. package/dist/site/tasks.d.ts +47 -0
  747. package/dist/site/tasks.d.ts.map +1 -0
  748. package/dist/site/tasks.js +164 -0
  749. package/dist/site/tasks.js.map +1 -0
  750. package/dist/site/temporal-status.d.ts +8 -0
  751. package/dist/site/temporal-status.d.ts.map +1 -0
  752. package/dist/site/temporal-status.js +41 -0
  753. package/dist/site/temporal-status.js.map +1 -0
  754. package/dist/site/tenant.d.ts +4 -0
  755. package/dist/site/tenant.d.ts.map +1 -0
  756. package/dist/site/tenant.js +6 -0
  757. package/dist/site/tenant.js.map +1 -0
  758. package/dist/static-check.d.ts +38 -0
  759. package/dist/static-check.d.ts.map +1 -0
  760. package/dist/static-check.js +131 -0
  761. package/dist/static-check.js.map +1 -0
  762. package/dist/submission-hash.d.ts +18 -0
  763. package/dist/submission-hash.d.ts.map +1 -0
  764. package/dist/submission-hash.js +16 -0
  765. package/dist/submission-hash.js.map +1 -0
  766. package/dist/subscription-auth.d.ts +0 -7
  767. package/dist/subscription-auth.d.ts.map +1 -1
  768. package/dist/subscription-auth.js +3 -23
  769. package/dist/subscription-auth.js.map +1 -1
  770. package/dist/temporal/activities/agent-feed.d.ts +5 -0
  771. package/dist/temporal/activities/agent-feed.d.ts.map +1 -0
  772. package/dist/temporal/activities/agent-feed.js +41 -0
  773. package/dist/temporal/activities/agent-feed.js.map +1 -0
  774. package/dist/temporal/activities/agent-scripts.d.ts +16 -0
  775. package/dist/temporal/activities/agent-scripts.d.ts.map +1 -0
  776. package/dist/temporal/activities/agent-scripts.js +162 -0
  777. package/dist/temporal/activities/agent-scripts.js.map +1 -0
  778. package/dist/temporal/activities/authoring-round.d.ts +13 -0
  779. package/dist/temporal/activities/authoring-round.d.ts.map +1 -0
  780. package/dist/temporal/activities/authoring-round.js +160 -0
  781. package/dist/temporal/activities/authoring-round.js.map +1 -0
  782. package/dist/temporal/activities/constants.d.ts +5 -0
  783. package/dist/temporal/activities/constants.d.ts.map +1 -0
  784. package/dist/temporal/activities/constants.js +5 -0
  785. package/dist/temporal/activities/constants.js.map +1 -0
  786. package/dist/temporal/activities/discovery.d.ts +17 -0
  787. package/dist/temporal/activities/discovery.d.ts.map +1 -0
  788. package/dist/temporal/activities/discovery.js +153 -0
  789. package/dist/temporal/activities/discovery.js.map +1 -0
  790. package/dist/temporal/activities/drafts.d.ts +15 -0
  791. package/dist/temporal/activities/drafts.d.ts.map +1 -0
  792. package/dist/temporal/activities/drafts.js +51 -0
  793. package/dist/temporal/activities/drafts.js.map +1 -0
  794. package/dist/temporal/activities/excluded-source-prs.d.ts +9 -0
  795. package/dist/temporal/activities/excluded-source-prs.d.ts.map +1 -0
  796. package/dist/temporal/activities/excluded-source-prs.js +132 -0
  797. package/dist/temporal/activities/excluded-source-prs.js.map +1 -0
  798. package/dist/temporal/activities/export.d.ts +38 -0
  799. package/dist/temporal/activities/export.d.ts.map +1 -0
  800. package/dist/temporal/activities/export.js +79 -0
  801. package/dist/temporal/activities/export.js.map +1 -0
  802. package/dist/temporal/activities/factory.d.ts +5 -0
  803. package/dist/temporal/activities/factory.d.ts.map +1 -0
  804. package/dist/temporal/activities/factory.js +26 -0
  805. package/dist/temporal/activities/factory.js.map +1 -0
  806. package/dist/temporal/activities/gate-logs.d.ts +20 -0
  807. package/dist/temporal/activities/gate-logs.d.ts.map +1 -0
  808. package/dist/temporal/activities/gate-logs.js +62 -0
  809. package/dist/temporal/activities/gate-logs.js.map +1 -0
  810. package/dist/temporal/activities/generation-runtime.d.ts +6 -0
  811. package/dist/temporal/activities/generation-runtime.d.ts.map +1 -0
  812. package/dist/temporal/activities/generation-runtime.js +56 -0
  813. package/dist/temporal/activities/generation-runtime.js.map +1 -0
  814. package/dist/temporal/activities/harbor.d.ts +11 -0
  815. package/dist/temporal/activities/harbor.d.ts.map +1 -0
  816. package/dist/temporal/activities/harbor.js +112 -0
  817. package/dist/temporal/activities/harbor.js.map +1 -0
  818. package/dist/temporal/activities/modal-build-log.d.ts +12 -0
  819. package/dist/temporal/activities/modal-build-log.d.ts.map +1 -0
  820. package/dist/temporal/activities/modal-build-log.js +57 -0
  821. package/dist/temporal/activities/modal-build-log.js.map +1 -0
  822. package/dist/temporal/activities/prompts-authoring.d.ts +6 -0
  823. package/dist/temporal/activities/prompts-authoring.d.ts.map +1 -0
  824. package/dist/temporal/activities/prompts-authoring.js +87 -0
  825. package/dist/temporal/activities/prompts-authoring.js.map +1 -0
  826. package/dist/temporal/activities/prompts-verifier.d.ts +17 -0
  827. package/dist/temporal/activities/prompts-verifier.d.ts.map +1 -0
  828. package/dist/temporal/activities/prompts-verifier.js +66 -0
  829. package/dist/temporal/activities/prompts-verifier.js.map +1 -0
  830. package/dist/temporal/activities/provenance.d.ts +4 -0
  831. package/dist/temporal/activities/provenance.d.ts.map +1 -0
  832. package/dist/temporal/activities/provenance.js +18 -0
  833. package/dist/temporal/activities/provenance.js.map +1 -0
  834. package/dist/temporal/activities/replay.d.ts +20 -0
  835. package/dist/temporal/activities/replay.d.ts.map +1 -0
  836. package/dist/temporal/activities/replay.js +135 -0
  837. package/dist/temporal/activities/replay.js.map +1 -0
  838. package/dist/temporal/activities/round-outcome.d.ts +63 -0
  839. package/dist/temporal/activities/round-outcome.d.ts.map +1 -0
  840. package/dist/temporal/activities/round-outcome.js +127 -0
  841. package/dist/temporal/activities/round-outcome.js.map +1 -0
  842. package/dist/temporal/activities/runtime.d.ts +38 -0
  843. package/dist/temporal/activities/runtime.d.ts.map +1 -0
  844. package/dist/temporal/activities/runtime.js +146 -0
  845. package/dist/temporal/activities/runtime.js.map +1 -0
  846. package/dist/temporal/activities/session-verify.d.ts +45 -0
  847. package/dist/temporal/activities/session-verify.d.ts.map +1 -0
  848. package/dist/temporal/activities/session-verify.js +91 -0
  849. package/dist/temporal/activities/session-verify.js.map +1 -0
  850. package/dist/temporal/activities/submissions.d.ts +9 -0
  851. package/dist/temporal/activities/submissions.d.ts.map +1 -0
  852. package/dist/temporal/activities/submissions.js +28 -0
  853. package/dist/temporal/activities/submissions.js.map +1 -0
  854. package/dist/temporal/activities/task-compiler.d.ts +21 -0
  855. package/dist/temporal/activities/task-compiler.d.ts.map +1 -0
  856. package/dist/temporal/activities/task-compiler.js +63 -0
  857. package/dist/temporal/activities/task-compiler.js.map +1 -0
  858. package/dist/temporal/activities/types.d.ts +57 -0
  859. package/dist/temporal/activities/types.d.ts.map +1 -0
  860. package/dist/temporal/activities/types.js +2 -0
  861. package/dist/temporal/activities/types.js.map +1 -0
  862. package/dist/temporal/activities/verifier-material.d.ts +18 -0
  863. package/dist/temporal/activities/verifier-material.d.ts.map +1 -0
  864. package/dist/temporal/activities/verifier-material.js +49 -0
  865. package/dist/temporal/activities/verifier-material.js.map +1 -0
  866. package/dist/temporal/activities/verifier-outcome.d.ts +19 -0
  867. package/dist/temporal/activities/verifier-outcome.d.ts.map +1 -0
  868. package/dist/temporal/activities/verifier-outcome.js +70 -0
  869. package/dist/temporal/activities/verifier-outcome.js.map +1 -0
  870. package/dist/temporal/activities/verifier-round.d.ts +8 -0
  871. package/dist/temporal/activities/verifier-round.d.ts.map +1 -0
  872. package/dist/temporal/activities/verifier-round.js +110 -0
  873. package/dist/temporal/activities/verifier-round.js.map +1 -0
  874. package/dist/temporal/activities/verify-harbor.d.ts +15 -0
  875. package/dist/temporal/activities/verify-harbor.d.ts.map +1 -0
  876. package/dist/temporal/activities/verify-harbor.js +193 -0
  877. package/dist/temporal/activities/verify-harbor.js.map +1 -0
  878. package/dist/temporal/activities/verify.d.ts +11 -0
  879. package/dist/temporal/activities/verify.d.ts.map +1 -0
  880. package/dist/temporal/activities/verify.js +151 -0
  881. package/dist/temporal/activities/verify.js.map +1 -0
  882. package/dist/temporal/activities.d.ts +3 -51
  883. package/dist/temporal/activities.d.ts.map +1 -1
  884. package/dist/temporal/activities.js +2 -886
  885. package/dist/temporal/activities.js.map +1 -1
  886. package/dist/temporal/connection.d.ts.map +1 -1
  887. package/dist/temporal/connection.js +3 -2
  888. package/dist/temporal/connection.js.map +1 -1
  889. package/dist/temporal/worker-main.js +19 -3
  890. package/dist/temporal/worker-main.js.map +1 -1
  891. package/dist/temporal/workflow/activity-proxies.d.ts +3 -0
  892. package/dist/temporal/workflow/activity-proxies.d.ts.map +1 -0
  893. package/dist/temporal/workflow/activity-proxies.js +45 -0
  894. package/dist/temporal/workflow/activity-proxies.js.map +1 -0
  895. package/dist/temporal/workflow/authoring-stage.d.ts +5 -0
  896. package/dist/temporal/workflow/authoring-stage.d.ts.map +1 -0
  897. package/dist/temporal/workflow/authoring-stage.js +64 -0
  898. package/dist/temporal/workflow/authoring-stage.js.map +1 -0
  899. package/dist/temporal/workflow/candidate-tracker.d.ts +28 -0
  900. package/dist/temporal/workflow/candidate-tracker.d.ts.map +1 -0
  901. package/dist/temporal/workflow/candidate-tracker.js +85 -0
  902. package/dist/temporal/workflow/candidate-tracker.js.map +1 -0
  903. package/dist/temporal/workflow/candidate.d.ts +6 -0
  904. package/dist/temporal/workflow/candidate.d.ts.map +1 -0
  905. package/dist/temporal/workflow/candidate.js +54 -0
  906. package/dist/temporal/workflow/candidate.js.map +1 -0
  907. package/dist/temporal/workflow/discovery.d.ts +19 -0
  908. package/dist/temporal/workflow/discovery.d.ts.map +1 -0
  909. package/dist/temporal/workflow/discovery.js +137 -0
  910. package/dist/temporal/workflow/discovery.js.map +1 -0
  911. package/dist/temporal/workflow/failures.d.ts +5 -0
  912. package/dist/temporal/workflow/failures.d.ts.map +1 -0
  913. package/dist/temporal/workflow/failures.js +28 -0
  914. package/dist/temporal/workflow/failures.js.map +1 -0
  915. package/dist/temporal/workflow/run.d.ts +5 -0
  916. package/dist/temporal/workflow/run.d.ts.map +1 -0
  917. package/dist/temporal/workflow/run.js +154 -0
  918. package/dist/temporal/workflow/run.js.map +1 -0
  919. package/dist/temporal/workflow/stage.d.ts +26 -0
  920. package/dist/temporal/workflow/stage.d.ts.map +1 -0
  921. package/dist/temporal/workflow/stage.js +16 -0
  922. package/dist/temporal/workflow/stage.js.map +1 -0
  923. package/dist/temporal/workflow/verification-stage.d.ts +11 -0
  924. package/dist/temporal/workflow/verification-stage.d.ts.map +1 -0
  925. package/dist/temporal/workflow/verification-stage.js +20 -0
  926. package/dist/temporal/workflow/verification-stage.js.map +1 -0
  927. package/dist/temporal/workflow.d.ts +12 -4
  928. package/dist/temporal/workflow.d.ts.map +1 -1
  929. package/dist/temporal/workflow.js +38 -445
  930. package/dist/temporal/workflow.js.map +1 -1
  931. package/dist/verifier-fix.d.ts +20 -0
  932. package/dist/verifier-fix.d.ts.map +1 -0
  933. package/dist/verifier-fix.js +62 -0
  934. package/dist/verifier-fix.js.map +1 -0
  935. package/dist/verify-report.d.ts +9 -0
  936. package/dist/verify-report.d.ts.map +1 -0
  937. package/dist/verify-report.js +135 -0
  938. package/dist/verify-report.js.map +1 -0
  939. package/dist/viewer/archived-keys.d.ts +9 -0
  940. package/dist/viewer/archived-keys.d.ts.map +1 -0
  941. package/dist/viewer/archived-keys.js +36 -0
  942. package/dist/viewer/archived-keys.js.map +1 -0
  943. package/dist/viewer/archived.d.ts +20 -0
  944. package/dist/viewer/archived.d.ts.map +1 -0
  945. package/dist/viewer/archived.js +238 -0
  946. package/dist/viewer/archived.js.map +1 -0
  947. package/dist/viewer/artifacts.d.ts +8 -0
  948. package/dist/viewer/artifacts.d.ts.map +1 -0
  949. package/dist/viewer/artifacts.js +96 -0
  950. package/dist/viewer/artifacts.js.map +1 -0
  951. package/dist/viewer/bundle.d.ts +9 -0
  952. package/dist/viewer/bundle.d.ts.map +1 -0
  953. package/dist/viewer/bundle.js +79 -0
  954. package/dist/viewer/bundle.js.map +1 -0
  955. package/dist/viewer/candidates.d.ts +9 -0
  956. package/dist/viewer/candidates.d.ts.map +1 -0
  957. package/dist/viewer/candidates.js +137 -0
  958. package/dist/viewer/candidates.js.map +1 -0
  959. package/dist/viewer/local-server.d.ts +11 -0
  960. package/dist/viewer/local-server.d.ts.map +1 -0
  961. package/dist/viewer/local-server.js +64 -0
  962. package/dist/viewer/local-server.js.map +1 -0
  963. package/dist/viewer/task-files.d.ts +10 -0
  964. package/dist/viewer/task-files.d.ts.map +1 -0
  965. package/dist/viewer/task-files.js +120 -0
  966. package/dist/viewer/task-files.js.map +1 -0
  967. package/dist/viewer/types.d.ts +78 -0
  968. package/dist/viewer/types.d.ts.map +1 -0
  969. package/dist/viewer/types.js +36 -0
  970. package/dist/viewer/types.js.map +1 -0
  971. package/docs/api.md +168 -0
  972. package/docs/evaluations.md +42 -15
  973. package/docs/operations.md +254 -26
  974. package/docs/solver-evaluations.md +63 -0
  975. package/docs/task-construction.md +21 -11
  976. package/drizzle/0000_glossy_johnny_blaze.sql +97 -0
  977. package/drizzle/0001_evaluation_records.sql +5 -0
  978. package/drizzle/0002_task_deletion.sql +1 -0
  979. package/drizzle/0003_drop_repo_runs.sql +3 -0
  980. package/drizzle/0004_preserve_batch_runs.sql +8 -0
  981. package/drizzle/0005_api_keys.sql +15 -0
  982. package/drizzle/meta/0000_snapshot.json +776 -0
  983. package/drizzle/meta/0001_snapshot.json +807 -0
  984. package/drizzle/meta/0002_snapshot.json +813 -0
  985. package/drizzle/meta/0003_snapshot.json +726 -0
  986. package/drizzle/meta/0004_snapshot.json +813 -0
  987. package/drizzle/meta/0005_snapshot.json +919 -0
  988. package/drizzle/meta/_journal.json +48 -0
  989. package/package.json +39 -25
  990. package/src/extensions/authoring.ts +83 -72
  991. package/src/extensions/shared/deliverable.ts +153 -0
  992. package/src/extensions/shared/mailbox.ts +128 -0
  993. package/src/extensions/shared/schemas.ts +101 -0
  994. package/src/extensions/shared/static-check.ts +103 -0
  995. package/src/extensions/verifier.ts +88 -0
  996. package/src/skills/selfbench/SKILL.md +71 -13
  997. package/biome.json +0 -24
  998. package/dist/agent-smoke-main.d.ts +0 -3
  999. package/dist/agent-smoke-main.d.ts.map +0 -1
  1000. package/dist/agent-smoke-main.js +0 -55
  1001. package/dist/agent-smoke-main.js.map +0 -1
  1002. package/dist/agent-smoke.d.ts +0 -18
  1003. package/dist/agent-smoke.d.ts.map +0 -1
  1004. package/dist/agent-smoke.js +0 -103
  1005. package/dist/agent-smoke.js.map +0 -1
  1006. package/dist/codex-review.d.ts +0 -41
  1007. package/dist/codex-review.d.ts.map +0 -1
  1008. package/dist/codex-review.js +0 -193
  1009. package/dist/codex-review.js.map +0 -1
  1010. package/dist/eval-main.d.ts +0 -3
  1011. package/dist/eval-main.d.ts.map +0 -1
  1012. package/dist/eval-main.js +0 -81
  1013. package/dist/eval-main.js.map +0 -1
  1014. package/dist/evaluate.d.ts +0 -25
  1015. package/dist/evaluate.d.ts.map +0 -1
  1016. package/dist/evaluate.js +0 -214
  1017. package/dist/evaluate.js.map +0 -1
  1018. package/dist/modal-auth.d.ts.map +0 -1
  1019. package/dist/modal-auth.js.map +0 -1
  1020. package/dist/reaudit-main.d.ts +0 -3
  1021. package/dist/reaudit-main.d.ts.map +0 -1
  1022. package/dist/reaudit-main.js +0 -158
  1023. package/dist/reaudit-main.js.map +0 -1
  1024. package/dist/repair-main.d.ts +0 -3
  1025. package/dist/repair-main.d.ts.map +0 -1
  1026. package/dist/repair-main.js +0 -171
  1027. package/dist/repair-main.js.map +0 -1
  1028. package/dist/review/assets/index-CZ3G759O.js +0 -1568
  1029. package/dist/review/assets/index-DgSLzX8l.css +0 -1
  1030. package/dist/sandbox/programs/prepare-repair.d.ts +0 -7
  1031. package/dist/sandbox/programs/prepare-repair.d.ts.map +0 -1
  1032. package/dist/sandbox/programs/prepare-repair.js.map +0 -1
  1033. package/dist/sandbox/programs/repair.d.ts +0 -3
  1034. package/dist/sandbox/programs/repair.d.ts.map +0 -1
  1035. package/dist/sandbox/programs/repair.js +0 -100
  1036. package/dist/sandbox/programs/repair.js.map +0 -1
  1037. package/dist/sandbox/programs/review.d.ts +0 -2
  1038. package/dist/sandbox/programs/review.d.ts.map +0 -1
  1039. package/dist/sandbox/programs/review.js +0 -18
  1040. package/dist/sandbox/programs/review.js.map +0 -1
  1041. package/dist/sandbox/programs/validation-repair.d.ts +0 -3
  1042. package/dist/sandbox/programs/validation-repair.d.ts.map +0 -1
  1043. package/dist/sandbox/programs/validation-repair.js +0 -115
  1044. package/dist/sandbox/programs/validation-repair.js.map +0 -1
  1045. package/dist/sandbox-repair.bundle.js +0 -323
  1046. package/dist/sandbox-review.bundle.js +0 -14462
  1047. package/dist/validate-main.d.ts +0 -3
  1048. package/dist/validate-main.d.ts.map +0 -1
  1049. package/dist/validate-main.js +0 -143
  1050. package/dist/validate-main.js.map +0 -1
  1051. package/dist/validation-repair.d.ts +0 -10
  1052. package/dist/validation-repair.d.ts.map +0 -1
  1053. package/dist/validation-repair.js +0 -72
  1054. package/dist/validation-repair.js.map +0 -1
  1055. package/review/index.html +0 -14
  1056. package/review/src/App.tsx +0 -338
  1057. package/review/src/export-loader.test.ts +0 -57
  1058. package/review/src/export-loader.ts +0 -151
  1059. package/review/src/main.tsx +0 -15
  1060. package/review/src/styles.css +0 -162
  1061. package/review/src/types.ts +0 -47
  1062. package/review/tsconfig.json +0 -16
  1063. package/review/vite.config.ts +0 -20
  1064. package/scripts/build-sandbox-programs.ts +0 -22
  1065. package/scripts/verify-package.ts +0 -102
  1066. package/scripts/write-build-metadata.ts +0 -27
  1067. package/src/agent-smoke-main.ts +0 -64
  1068. package/src/agent-smoke.ts +0 -138
  1069. package/src/api-main.ts +0 -12
  1070. package/src/api.ts +0 -291
  1071. package/src/artifacts.ts +0 -357
  1072. package/src/audit.ts +0 -106
  1073. package/src/build-metadata.ts +0 -3
  1074. package/src/cli.ts +0 -504
  1075. package/src/codex-review.ts +0 -229
  1076. package/src/config.ts +0 -239
  1077. package/src/contracts.ts +0 -233
  1078. package/src/coupling.ts +0 -259
  1079. package/src/eval-main.ts +0 -93
  1080. package/src/evaluate.ts +0 -300
  1081. package/src/github.ts +0 -26
  1082. package/src/harbor-environment.ts +0 -17
  1083. package/src/harbor-results.ts +0 -144
  1084. package/src/harbor-task.ts +0 -533
  1085. package/src/hash.ts +0 -5
  1086. package/src/modal-auth.ts +0 -11
  1087. package/src/parallel.ts +0 -24
  1088. package/src/process.ts +0 -167
  1089. package/src/project-paths.ts +0 -15
  1090. package/src/provenance-associations.ts +0 -311
  1091. package/src/provenance.ts +0 -561
  1092. package/src/providers.ts +0 -22
  1093. package/src/reaudit-main.ts +0 -198
  1094. package/src/repair-main.ts +0 -209
  1095. package/src/repair.ts +0 -55
  1096. package/src/run-wait.ts +0 -40
  1097. package/src/sandbox/contracts.ts +0 -52
  1098. package/src/sandbox/index.ts +0 -23
  1099. package/src/sandbox/programs/author.ts +0 -19
  1100. package/src/sandbox/programs/prepare-repair.ts +0 -44
  1101. package/src/sandbox/programs/repair.ts +0 -132
  1102. package/src/sandbox/programs/review.ts +0 -19
  1103. package/src/sandbox/programs/validation-repair.ts +0 -153
  1104. package/src/sandbox/providers/docker/executor.ts +0 -115
  1105. package/src/sandbox/providers/modal/executor.ts +0 -217
  1106. package/src/sandbox/providers/vercel/command.ts +0 -127
  1107. package/src/sandbox/providers/vercel/executor.ts +0 -448
  1108. package/src/sandbox/providers/vercel/fetch.ts +0 -45
  1109. package/src/sandbox/timeout.ts +0 -56
  1110. package/src/setup/vercel/cli.ts +0 -403
  1111. package/src/setup/vercel/index.ts +0 -3
  1112. package/src/setup/vercel/probe.ts +0 -232
  1113. package/src/setup/vercel/profile.ts +0 -322
  1114. package/src/setup/vercel/runtime-image.ts +0 -204
  1115. package/src/setup/vercel/setup.ts +0 -419
  1116. package/src/subscription-auth.ts +0 -80
  1117. package/src/temporal/activities.ts +0 -1348
  1118. package/src/temporal/connection.ts +0 -23
  1119. package/src/temporal/worker-main.ts +0 -28
  1120. package/src/temporal/workflow.ts +0 -568
  1121. package/src/terminal-prompts.ts +0 -197
  1122. package/src/terminal-reporter.ts +0 -126
  1123. package/src/validate-main.ts +0 -177
  1124. package/src/validation-repair.ts +0 -94
  1125. package/tsconfig.build.json +0 -13
  1126. package/tsconfig.json +0 -21
@@ -1 +0,0 @@
1
- :root{--bg: #f4f5f7;--surface: #ffffff;--border: #d8dce3;--text: #1a1d23;--muted: #6b7280;--accent: #2563eb;--danger: #dc2626;--success: #059669;--code-bg: #f1f3f5;--font: -apple-system, BlinkMacSystemFont, "Segoe UI", system-ui, sans-serif;--mono: "SF Mono", "Cascadia Code", "Fira Code", monospace}@media(prefers-color-scheme:dark){:root{--bg: #0d1117;--surface: #161b22;--border: #30363d;--text: #e6edf3;--muted: #8b949e;--accent: #58a6ff;--danger: #f85149;--success: #3fb950;--code-bg: #1c2128}}*{margin:0;padding:0;box-sizing:border-box}body{font-family:var(--font);background:var(--bg);color:var(--text);font-size:14px;line-height:1.5}.app-shell{min-height:100vh;display:flex;flex-direction:column}.topbar{display:flex;justify-content:space-between;align-items:center;padding:16px 24px;border-bottom:1px solid var(--border);background:var(--surface)}.eyebrow{font-size:11px;font-weight:600;letter-spacing:.08em;text-transform:uppercase;color:var(--muted)}.topbar h1{font-size:18px;font-weight:600;margin-top:2px}.topbar-actions{display:flex;gap:8px;align-items:center}.token-input{padding:6px 10px;border:1px solid var(--border);border-radius:6px;background:var(--surface);color:var(--text);font-size:12px;width:180px}.error-banner{padding:12px 24px;background:var(--danger);color:#fff;font-size:13px}.workspace{flex:1;display:grid;grid-template-columns:280px 1fr;gap:16px;padding:16px 24px}.surface{background:var(--surface);border:1px solid var(--border);border-radius:8px}.run-list{padding:16px;overflow-y:auto;max-height:calc(100vh - 100px)}.section-heading{display:flex;justify-content:space-between;align-items:baseline;margin-bottom:12px;padding-bottom:8px;border-bottom:1px solid var(--border)}.section-heading h2{font-size:13px;font-weight:600}.section-heading span{font-size:11px;color:var(--muted)}.run-row{width:100%;display:flex;flex-direction:column;gap:2px;padding:10px 12px;margin-bottom:6px;border:1px solid var(--border);border-radius:6px;background:var(--surface);cursor:pointer;text-align:left;font:inherit;color:inherit}.run-row:hover{border-color:var(--accent)}.run-row.selected{border-color:var(--accent);box-shadow:0 0 0 1px var(--accent)}.run-row strong{font-size:13px}.run-row span{font-size:11px;color:var(--muted)}.run-row small{font-size:10px;color:var(--muted)}.content{display:flex;flex-direction:column;gap:16px}.empty{padding:64px 32px;text-align:center;display:flex;flex-direction:column;align-items:center;gap:8px;min-height:400px;justify-content:center}.empty h2{font-size:16px;font-weight:600}.empty p{color:var(--muted)}.summary{display:flex;justify-content:space-between;align-items:center;padding:16px 20px}.summary h2{font-size:16px;font-weight:600}.summary .muted{font-size:12px;color:var(--muted);margin-top:2px}.review-grid{display:grid;grid-template-columns:240px 1fr;gap:16px}.task-list{padding:16px;max-height:calc(100vh - 280px);overflow-y:auto}.task-row{width:100%;display:flex;justify-content:space-between;align-items:center;padding:8px 10px;margin-bottom:4px;border:1px solid var(--border);border-radius:5px;background:var(--surface);cursor:pointer;text-align:left;font:inherit;color:inherit;font-size:12px}.task-row:hover,.task-row.selected{border-color:var(--accent)}.task-row.removed{opacity:.5}.task-row.removed strong{text-decoration:line-through}.task-row span{font-size:10px;color:var(--muted)}.task-detail{padding:20px 24px;max-height:calc(100vh - 280px);overflow-y:auto}.detail-header{display:flex;justify-content:space-between;align-items:flex-start;margin-bottom:20px;padding-bottom:16px;border-bottom:1px solid var(--border)}.detail-header h2{font-size:18px;font-weight:600}.detail-block{margin-bottom:20px}.detail-block h3{font-size:12px;font-weight:600;letter-spacing:.04em;text-transform:uppercase;color:var(--muted);margin-bottom:6px}.detail-block pre{padding:12px 16px;background:var(--code-bg);border-radius:6px;font-family:var(--mono);font-size:12px;white-space:pre-wrap;word-break:break-all;max-height:400px;overflow-y:auto;line-height:1.6}.patch-files{display:grid;gap:10px;padding:8px;max-height:600px;overflow:auto;background:var(--code-bg);border-radius:6px}.patch-files>div{overflow:hidden;border:1px solid var(--border);border-radius:4px}.button{padding:7px 14px;border:1px solid var(--border);border-radius:6px;background:var(--surface);color:var(--text);font-size:12px;font-weight:500;cursor:pointer;font-family:var(--font)}.button:hover{border-color:var(--accent);color:var(--accent)}.button.primary{background:var(--accent);color:#fff;border-color:var(--accent)}.button.primary:hover{opacity:.9;color:#fff}.button.danger{color:var(--danger);border-color:var(--danger)}.button.danger:hover{background:var(--danger);color:#fff}.button.undo{color:var(--success);border-color:var(--success)}.button.undo:hover{background:var(--success);color:#fff}.muted{color:var(--muted)}
@@ -1,7 +0,0 @@
1
- export interface PreparedRepairTask {
2
- readonly extractedDirectory: string;
3
- readonly taskDirectory: string;
4
- readonly repositoryDirectory: string;
5
- }
6
- export declare function prepareRepairTask(archivePath: string, workDirectory?: string): Promise<PreparedRepairTask>;
7
- //# sourceMappingURL=prepare-repair.d.ts.map
@@ -1 +0,0 @@
1
- {"version":3,"file":"prepare-repair.d.ts","sourceRoot":"","sources":["../../../src/sandbox/programs/prepare-repair.ts"],"names":[],"mappings":"AAIA,MAAM,WAAW,kBAAkB;IACjC,QAAQ,CAAC,kBAAkB,EAAE,MAAM,CAAC;IACpC,QAAQ,CAAC,aAAa,EAAE,MAAM,CAAC;IAC/B,QAAQ,CAAC,mBAAmB,EAAE,MAAM,CAAC;CACtC;AAED,wBAAsB,iBAAiB,CACrC,WAAW,EAAE,MAAM,EACnB,aAAa,SAAU,GACtB,OAAO,CAAC,kBAAkB,CAAC,CA8B7B"}
@@ -1 +0,0 @@
1
- {"version":3,"file":"prepare-repair.js","sourceRoot":"","sources":["../../../src/sandbox/programs/prepare-repair.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,kBAAkB,CAAC;AACjD,OAAO,EAAE,IAAI,EAAE,MAAM,WAAW,CAAC;AACjC,OAAO,EAAE,UAAU,EAAE,MAAM,kBAAkB,CAAC;AAQ9C,MAAM,CAAC,KAAK,UAAU,iBAAiB,CACrC,WAAmB,EACnB,aAAa,GAAG,OAAO;IAEvB,MAAM,kBAAkB,GAAG,IAAI,CAAC,aAAa,EAAE,MAAM,CAAC,CAAC;IACvD,MAAM,mBAAmB,GAAG,IAAI,CAAC,aAAa,EAAE,MAAM,CAAC,CAAC;IACxD,MAAM,OAAO,CAAC,GAAG,CAAC,CAAC,KAAK,CAAC,kBAAkB,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC,EAAE,KAAK,CAAC,mBAAmB,CAAC,CAAC,CAAC,CAAC;IAChG,MAAM,UAAU,CAAC,KAAK,EAAE,CAAC,MAAM,EAAE,WAAW,EAAE,IAAI,EAAE,kBAAkB,CAAC,CAAC,CAAC;IACzE,MAAM,aAAa,GAAG,MAAM,MAAM,CAAC,IAAI,CAAC,kBAAkB,EAAE,gBAAgB,CAAC,CAAC,CAAC,IAAI,CACjF,GAAG,EAAE,CAAC,kBAAkB,EACxB,GAAG,EAAE,CAAC,IAAI,CAAC,kBAAkB,EAAE,aAAa,CAAC,CAC9C,CAAC;IAEF,MAAM,UAAU,CAAC,KAAK,EAAE;QACtB,MAAM;QACN,IAAI,CAAC,aAAa,EAAE,mBAAmB,CAAC;QACxC,IAAI;QACJ,mBAAmB;KACpB,CAAC,CAAC;IACH,MAAM,UAAU,CAAC,KAAK,EAAE,CAAC,IAAI,EAAE,mBAAmB,EAAE,MAAM,EAAE,IAAI,CAAC,CAAC,CAAC;IACnE,MAAM,UAAU,CAAC,KAAK,EAAE,CAAC,IAAI,EAAE,mBAAmB,EAAE,QAAQ,EAAE,WAAW,EAAE,WAAW,CAAC,CAAC,CAAC;IACzF,MAAM,UAAU,CAAC,KAAK,EAAE,CAAC,IAAI,EAAE,mBAAmB,EAAE,QAAQ,EAAE,YAAY,EAAE,iBAAiB,CAAC,CAAC,CAAC;IAChG,MAAM,UAAU,CAAC,KAAK,EAAE,CAAC,IAAI,EAAE,mBAAmB,EAAE,KAAK,EAAE,IAAI,CAAC,CAAC,CAAC;IAClE,MAAM,UAAU,CAAC,KAAK,EAAE,CAAC,IAAI,EAAE,mBAAmB,EAAE,QAAQ,EAAE,KAAK,EAAE,MAAM,CAAC,CAAC,CAAC;IAC9E,MAAM,UAAU,CAAC,KAAK,EAAE;QACtB,IAAI;QACJ,mBAAmB;QACnB,OAAO;QACP,IAAI,CAAC,aAAa,EAAE,kBAAkB,CAAC;KACxC,CAAC,CAAC;IACH,MAAM,UAAU,CAAC,KAAK,EAAE,CAAC,IAAI,EAAE,mBAAmB,EAAE,KAAK,EAAE,IAAI,EAAE,OAAO,CAAC,CAAC,CAAC;IAE3E,OAAO,EAAE,kBAAkB,EAAE,aAAa,EAAE,mBAAmB,EAAE,CAAC;AACpE,CAAC"}
@@ -1,3 +0,0 @@
1
- #!/usr/bin/env node
2
- export {};
3
- //# sourceMappingURL=repair.d.ts.map
@@ -1 +0,0 @@
1
- {"version":3,"file":"repair.d.ts","sourceRoot":"","sources":["../../../src/sandbox/programs/repair.ts"],"names":[],"mappings":""}
@@ -1,100 +0,0 @@
1
- #!/usr/bin/env node
2
- import { chmod, mkdir, readFile, writeFile } from "node:fs/promises";
3
- import { homedir, tmpdir } from "node:os";
4
- import { join } from "node:path";
5
- import { sha256 } from "../../hash.js";
6
- import { runCommand } from "../../process.js";
7
- import { assertRepairPaths, patchPaths, repairPrompt } from "../../repair.js";
8
- import { prepareRepairTask } from "./prepare-repair.js";
9
- const [archivePath, reviewPath, outputArchive, outputReport] = process.argv.slice(2);
10
- if (!archivePath || !reviewPath || !outputArchive || !outputReport) {
11
- throw new Error("usage: sandbox-repair TASK.tar.gz REVIEW.json OUTPUT.tar.gz OUTPUT-REPORT.json");
12
- }
13
- const { extractedDirectory, taskDirectory, repositoryDirectory } = await prepareRepairTask(archivePath);
14
- const [instruction, originalPatch, review] = await Promise.all([
15
- readFile(join(taskDirectory, "instruction.md"), "utf8"),
16
- readFile(join(taskDirectory, "tests/test.patch"), "utf8"),
17
- readFile(reviewPath, "utf8"),
18
- ]);
19
- const manifestPath = join(taskDirectory, ".selfbench-manifest.json");
20
- const manifest = JSON.parse(await readFile(manifestPath, "utf8"));
21
- const taskId = typeof manifest.taskId === "string" ? manifest.taskId : "unknown-task";
22
- const allowedPaths = patchPaths(originalPatch);
23
- const codexAuth = process.env.SELFBENCH_CODEX_AUTH_JSON;
24
- const apiKey = process.env.OPENAI_API_KEY;
25
- if (!apiKey && !codexAuth) {
26
- fail("OPENAI_API_KEY or SELFBENCH_CODEX_AUTH_JSON is required");
27
- }
28
- const codexHome = join(homedir(), ".codex");
29
- const authPath = join(codexHome, "auth.json");
30
- await mkdir(codexHome, { recursive: true });
31
- await writeFile(authPath, codexAuth ?? `${JSON.stringify({ OPENAI_API_KEY: apiKey })}\n`);
32
- await chmod(authPath, 0o600);
33
- const promptPath = join(tmpdir(), `selfbench-repair-${taskId}.md`);
34
- await writeFile(promptPath, repairPrompt({
35
- taskId,
36
- authenticRequest: instruction,
37
- couplingReport: review,
38
- allowedPaths,
39
- }));
40
- const codex = await runCommand("codex", [
41
- "exec",
42
- "--model",
43
- process.env.SELFBENCH_REPAIR_MODEL ?? "gpt-5.6-sol",
44
- "--dangerously-bypass-approvals-and-sandbox",
45
- "--ephemeral",
46
- "--ignore-user-config",
47
- "--json",
48
- "-C",
49
- repositoryDirectory,
50
- "-",
51
- ], {
52
- allowFailure: true,
53
- timeoutMs: 90 * 60 * 1000,
54
- env: process.env,
55
- input: await readFile(promptPath, "utf8"),
56
- onOutput: (stream, chunk) => {
57
- (stream === "stdout" ? process.stdout : process.stderr).write(chunk);
58
- },
59
- });
60
- if (codex.exitCode !== 0) {
61
- throw new Error(`Codex repair exited ${codex.exitCode}: ${codex.stderr.slice(-2_000)}`);
62
- }
63
- const [tracked, untracked] = await Promise.all([
64
- runCommand("git", ["-C", repositoryDirectory, "diff", "--name-only", "HEAD"]),
65
- runCommand("git", ["-C", repositoryDirectory, "ls-files", "--others", "--exclude-standard"]),
66
- ]);
67
- const changedPaths = [...tracked.stdout.split("\n"), ...untracked.stdout.split("\n")]
68
- .filter(Boolean)
69
- .sort();
70
- assertRepairPaths(originalPatch, changedPaths);
71
- const repaired = await runCommand("git", ["-C", repositoryDirectory, "diff", "--binary", "HEAD"]);
72
- if (!repaired.stdout.startsWith("diff --git ")) {
73
- throw new Error(`repair produced no held-out test patch; status=${JSON.stringify(changedPaths)}; Codex tail=${codex.stdout.slice(-4_000)}`);
74
- }
75
- if (repaired.stdout === originalPatch) {
76
- throw new Error("repair left the held-out test patch unchanged");
77
- }
78
- await writeFile(join(taskDirectory, "tests/test.patch"), repaired.stdout);
79
- await writeFile(manifestPath, `${JSON.stringify({
80
- ...manifest,
81
- testPatchSha256: sha256(repaired.stdout),
82
- repair: {
83
- model: process.env.SELFBENCH_REPAIR_MODEL ?? "gpt-5.6-sol",
84
- originalTestPatchSha256: sha256(originalPatch),
85
- },
86
- }, null, 2)}\n`);
87
- await runCommand("tar", ["-czf", outputArchive, "-C", extractedDirectory, "."]);
88
- await writeFile(outputReport, `${JSON.stringify({
89
- schemaVersion: 1,
90
- taskId,
91
- model: process.env.SELFBENCH_REPAIR_MODEL ?? "gpt-5.6-sol",
92
- changedPaths,
93
- originalTestPatchSha256: sha256(originalPatch),
94
- repairedTestPatchSha256: sha256(repaired.stdout),
95
- codexOutputTail: codex.stdout.slice(-4_000),
96
- }, null, 2)}\n`);
97
- function fail(message) {
98
- throw new Error(message);
99
- }
100
- //# sourceMappingURL=repair.js.map
@@ -1 +0,0 @@
1
- {"version":3,"file":"repair.js","sourceRoot":"","sources":["../../../src/sandbox/programs/repair.ts"],"names":[],"mappings":";AAEA,OAAO,EAAE,KAAK,EAAE,KAAK,EAAE,QAAQ,EAAE,SAAS,EAAE,MAAM,kBAAkB,CAAC;AACrE,OAAO,EAAE,OAAO,EAAE,MAAM,EAAE,MAAM,SAAS,CAAC;AAC1C,OAAO,EAAE,IAAI,EAAE,MAAM,WAAW,CAAC;AACjC,OAAO,EAAE,MAAM,EAAE,MAAM,eAAe,CAAC;AACvC,OAAO,EAAE,UAAU,EAAE,MAAM,kBAAkB,CAAC;AAC9C,OAAO,EAAE,iBAAiB,EAAE,UAAU,EAAE,YAAY,EAAE,MAAM,iBAAiB,CAAC;AAC9E,OAAO,EAAE,iBAAiB,EAAE,MAAM,qBAAqB,CAAC;AAExD,MAAM,CAAC,WAAW,EAAE,UAAU,EAAE,aAAa,EAAE,YAAY,CAAC,GAAG,OAAO,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;AACrF,IAAI,CAAC,WAAW,IAAI,CAAC,UAAU,IAAI,CAAC,aAAa,IAAI,CAAC,YAAY,EAAE,CAAC;IACnE,MAAM,IAAI,KAAK,CAAC,gFAAgF,CAAC,CAAC;AACpG,CAAC;AAED,MAAM,EAAE,kBAAkB,EAAE,aAAa,EAAE,mBAAmB,EAAE,GAC9D,MAAM,iBAAiB,CAAC,WAAW,CAAC,CAAC;AAEvC,MAAM,CAAC,WAAW,EAAE,aAAa,EAAE,MAAM,CAAC,GAAG,MAAM,OAAO,CAAC,GAAG,CAAC;IAC7D,QAAQ,CAAC,IAAI,CAAC,aAAa,EAAE,gBAAgB,CAAC,EAAE,MAAM,CAAC;IACvD,QAAQ,CAAC,IAAI,CAAC,aAAa,EAAE,kBAAkB,CAAC,EAAE,MAAM,CAAC;IACzD,QAAQ,CAAC,UAAU,EAAE,MAAM,CAAC;CAC7B,CAAC,CAAC;AACH,MAAM,YAAY,GAAG,IAAI,CAAC,aAAa,EAAE,0BAA0B,CAAC,CAAC;AACrE,MAAM,QAAQ,GAAG,IAAI,CAAC,KAAK,CAAC,MAAM,QAAQ,CAAC,YAAY,EAAE,MAAM,CAAC,CAA4B,CAAC;AAC7F,MAAM,MAAM,GAAG,OAAO,QAAQ,CAAC,MAAM,KAAK,QAAQ,CAAC,CAAC,CAAC,QAAQ,CAAC,MAAM,CAAC,CAAC,CAAC,cAAc,CAAC;AACtF,MAAM,YAAY,GAAG,UAAU,CAAC,aAAa,CAAC,CAAC;AAE/C,MAAM,SAAS,GAAG,OAAO,CAAC,GAAG,CAAC,yBAAyB,CAAC;AACxD,MAAM,MAAM,GAAG,OAAO,CAAC,GAAG,CAAC,cAAc,CAAC;AAC1C,IAAI,CAAC,MAAM,IAAI,CAAC,SAAS,EAAE,CAAC;IAC1B,IAAI,CAAC,yDAAyD,CAAC,CAAC;AAClE,CAAC;AACD,MAAM,SAAS,GAAG,IAAI,CAAC,OAAO,EAAE,EAAE,QAAQ,CAAC,CAAC;AAC5C,MAAM,QAAQ,GAAG,IAAI,CAAC,SAAS,EAAE,WAAW,CAAC,CAAC;AAC9C,MAAM,KAAK,CAAC,SAAS,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC,CAAC;AAC5C,MAAM,SAAS,CAAC,QAAQ,EAAE,SAAS,IAAI,GAAG,IAAI,CAAC,SAAS,CAAC,EAAE,cAAc,EAAE,MAAM,EAAE,CAAC,IAAI,CAAC,CAAC;AAC1F,MAAM,KAAK,CAAC,QAAQ,EAAE,KAAK,CAAC,CAAC;AAC7B,MAAM,UAAU,GAAG,IAAI,CAAC,MAAM,EAAE,EAAE,oBAAoB,MAAM,KAAK,CAAC,CAAC;AACnE,MAAM,SAAS,CACb,UAAU,EACV,YAAY,CAAC;IACX,MAAM;IACN,gBAAgB,EAAE,WAAW;IAC7B,cAAc,EAAE,MAAM;IACtB,YAAY;CACb,CAAC,CACH,CAAC;AAEF,MAAM,KAAK,GAAG,MAAM,UAAU,CAC5B,OAAO,EACP;IACE,MAAM;IACN,SAAS;IACT,OAAO,CAAC,GAAG,CAAC,sBAAsB,IAAI,aAAa;IACnD,4CAA4C;IAC5C,aAAa;IACb,sBAAsB;IACtB,QAAQ;IACR,IAAI;IACJ,mBAAmB;IACnB,GAAG;CACJ,EACD;IACE,YAAY,EAAE,IAAI;IAClB,SAAS,EAAE,EAAE,GAAG,EAAE,GAAG,IAAI;IACzB,GAAG,EAAE,OAAO,CAAC,GAAG;IAChB,KAAK,EAAE,MAAM,QAAQ,CAAC,UAAU,EAAE,MAAM,CAAC;IACzC,QAAQ,EAAE,CAAC,MAAM,EAAE,KAAK,EAAE,EAAE;QAC1B,CAAC,MAAM,KAAK,QAAQ,CAAC,CAAC,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC;IACvE,CAAC;CACF,CACF,CAAC;AACF,IAAI,KAAK,CAAC,QAAQ,KAAK,CAAC,EAAE,CAAC;IACzB,MAAM,IAAI,KAAK,CAAC,uBAAuB,KAAK,CAAC,QAAQ,KAAK,KAAK,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC;AAC1F,CAAC;AAED,MAAM,CAAC,OAAO,EAAE,SAAS,CAAC,GAAG,MAAM,OAAO,CAAC,GAAG,CAAC;IAC7C,UAAU,CAAC,KAAK,EAAE,CAAC,IAAI,EAAE,mBAAmB,EAAE,MAAM,EAAE,aAAa,EAAE,MAAM,CAAC,CAAC;IAC7E,UAAU,CAAC,KAAK,EAAE,CAAC,IAAI,EAAE,mBAAmB,EAAE,UAAU,EAAE,UAAU,EAAE,oBAAoB,CAAC,CAAC;CAC7F,CAAC,CAAC;AACH,MAAM,YAAY,GAAG,CAAC,GAAG,OAAO,CAAC,MAAM,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,GAAG,SAAS,CAAC,MAAM,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;KAClF,MAAM,CAAC,OAAO,CAAC;KACf,IAAI,EAAE,CAAC;AACV,iBAAiB,CAAC,aAAa,EAAE,YAAY,CAAC,CAAC;AAC/C,MAAM,QAAQ,GAAG,MAAM,UAAU,CAAC,KAAK,EAAE,CAAC,IAAI,EAAE,mBAAmB,EAAE,MAAM,EAAE,UAAU,EAAE,MAAM,CAAC,CAAC,CAAC;AAClG,IAAI,CAAC,QAAQ,CAAC,MAAM,CAAC,UAAU,CAAC,aAAa,CAAC,EAAE,CAAC;IAC/C,MAAM,IAAI,KAAK,CACb,kDAAkD,IAAI,CAAC,SAAS,CAAC,YAAY,CAAC,gBAAgB,KAAK,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,KAAK,CAAC,EAAE,CAC3H,CAAC;AACJ,CAAC;AACD,IAAI,QAAQ,CAAC,MAAM,KAAK,aAAa,EAAE,CAAC;IACtC,MAAM,IAAI,KAAK,CAAC,+CAA+C,CAAC,CAAC;AACnE,CAAC;AAED,MAAM,SAAS,CAAC,IAAI,CAAC,aAAa,EAAE,kBAAkB,CAAC,EAAE,QAAQ,CAAC,MAAM,CAAC,CAAC;AAC1E,MAAM,SAAS,CACb,YAAY,EACZ,GAAG,IAAI,CAAC,SAAS,CACf;IACE,GAAG,QAAQ;IACX,eAAe,EAAE,MAAM,CAAC,QAAQ,CAAC,MAAM,CAAC;IACxC,MAAM,EAAE;QACN,KAAK,EAAE,OAAO,CAAC,GAAG,CAAC,sBAAsB,IAAI,aAAa;QAC1D,uBAAuB,EAAE,MAAM,CAAC,aAAa,CAAC;KAC/C;CACF,EACD,IAAI,EACJ,CAAC,CACF,IAAI,CACN,CAAC;AACF,MAAM,UAAU,CAAC,KAAK,EAAE,CAAC,MAAM,EAAE,aAAa,EAAE,IAAI,EAAE,kBAAkB,EAAE,GAAG,CAAC,CAAC,CAAC;AAChF,MAAM,SAAS,CACb,YAAY,EACZ,GAAG,IAAI,CAAC,SAAS,CACf;IACE,aAAa,EAAE,CAAC;IAChB,MAAM;IACN,KAAK,EAAE,OAAO,CAAC,GAAG,CAAC,sBAAsB,IAAI,aAAa;IAC1D,YAAY;IACZ,uBAAuB,EAAE,MAAM,CAAC,aAAa,CAAC;IAC9C,uBAAuB,EAAE,MAAM,CAAC,QAAQ,CAAC,MAAM,CAAC;IAChD,eAAe,EAAE,KAAK,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,KAAK,CAAC;CAC5C,EACD,IAAI,EACJ,CAAC,CACF,IAAI,CACN,CAAC;AAEF,SAAS,IAAI,CAAC,OAAe;IAC3B,MAAM,IAAI,KAAK,CAAC,OAAO,CAAC,CAAC;AAC3B,CAAC"}
@@ -1,2 +0,0 @@
1
- export {};
2
- //# sourceMappingURL=review.d.ts.map
@@ -1 +0,0 @@
1
- {"version":3,"file":"review.d.ts","sourceRoot":"","sources":["../../../src/sandbox/programs/review.ts"],"names":[],"mappings":""}
@@ -1,18 +0,0 @@
1
- import { readFile, writeFile } from "node:fs/promises";
2
- import { reviewCouplingWithCodex } from "../../codex-review.js";
3
- const apiKey = process.env.OPENAI_API_KEY;
4
- const authJson = process.env.SELFBENCH_PI_AUTH_JSON;
5
- const outputPath = process.env.SELFBENCH_REVIEW_OUTPUT;
6
- if (!apiKey && !authJson) {
7
- throw new Error("OPENAI_API_KEY or SELFBENCH_PI_AUTH_JSON is required");
8
- }
9
- if (!outputPath) {
10
- throw new Error("SELFBENCH_REVIEW_OUTPUT is required");
11
- }
12
- const review = await reviewCouplingWithCodex({
13
- ...(apiKey ? { apiKey } : {}),
14
- ...(authJson ? { authJson } : {}),
15
- prompt: await readFile("/work/review-input.md", "utf8"),
16
- });
17
- await writeFile(outputPath, `${JSON.stringify(review, null, 2)}\n`, { flag: "wx" });
18
- //# sourceMappingURL=review.js.map
@@ -1 +0,0 @@
1
- {"version":3,"file":"review.js","sourceRoot":"","sources":["../../../src/sandbox/programs/review.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,QAAQ,EAAE,SAAS,EAAE,MAAM,kBAAkB,CAAC;AACvD,OAAO,EAAE,uBAAuB,EAAE,MAAM,uBAAuB,CAAC;AAEhE,MAAM,MAAM,GAAG,OAAO,CAAC,GAAG,CAAC,cAAc,CAAC;AAC1C,MAAM,QAAQ,GAAG,OAAO,CAAC,GAAG,CAAC,sBAAsB,CAAC;AACpD,MAAM,UAAU,GAAG,OAAO,CAAC,GAAG,CAAC,uBAAuB,CAAC;AACvD,IAAI,CAAC,MAAM,IAAI,CAAC,QAAQ,EAAE,CAAC;IACzB,MAAM,IAAI,KAAK,CAAC,sDAAsD,CAAC,CAAC;AAC1E,CAAC;AACD,IAAI,CAAC,UAAU,EAAE,CAAC;IAChB,MAAM,IAAI,KAAK,CAAC,qCAAqC,CAAC,CAAC;AACzD,CAAC;AAED,MAAM,MAAM,GAAG,MAAM,uBAAuB,CAAC;IAC3C,GAAG,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;IAC7B,GAAG,CAAC,QAAQ,CAAC,CAAC,CAAC,EAAE,QAAQ,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;IACjC,MAAM,EAAE,MAAM,QAAQ,CAAC,uBAAuB,EAAE,MAAM,CAAC;CACxD,CAAC,CAAC;AACH,MAAM,SAAS,CAAC,UAAU,EAAE,GAAG,IAAI,CAAC,SAAS,CAAC,MAAM,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,EAAE,IAAI,EAAE,IAAI,EAAE,CAAC,CAAC"}
@@ -1,3 +0,0 @@
1
- #!/usr/bin/env node
2
- export {};
3
- //# sourceMappingURL=validation-repair.d.ts.map
@@ -1 +0,0 @@
1
- {"version":3,"file":"validation-repair.d.ts","sourceRoot":"","sources":["../../../src/sandbox/programs/validation-repair.ts"],"names":[],"mappings":""}
@@ -1,115 +0,0 @@
1
- #!/usr/bin/env node
2
- import { chmod, mkdir, readFile, writeFile } from "node:fs/promises";
3
- import { homedir, tmpdir } from "node:os";
4
- import { join } from "node:path";
5
- import { taskDefinitionSchema } from "../../contracts.js";
6
- import { runCommand } from "../../process.js";
7
- import { assertValidationRepair, validationRepairPaths, validationRepairPrompt, } from "../../validation-repair.js";
8
- import { prepareRepairTask } from "./prepare-repair.js";
9
- const [archivePath, definitionPath, diagnosticsPath, outputDefinition, outputPatch, outputReport] = process.argv.slice(2);
10
- if (!archivePath ||
11
- !definitionPath ||
12
- !diagnosticsPath ||
13
- !outputDefinition ||
14
- !outputPatch ||
15
- !outputReport) {
16
- throw new Error("usage: sandbox-validation-repair TASK.tar.gz DEFINITION.json DIAGNOSTICS.txt OUTPUT-DEFINITION.json OUTPUT-PATCH OUTPUT-REPORT.json");
17
- }
18
- const { taskDirectory, repositoryDirectory } = await prepareRepairTask(archivePath);
19
- const [instruction, originalPatch, diagnostics, originalDefinitionBytes] = await Promise.all([
20
- readFile(join(taskDirectory, "instruction.md"), "utf8"),
21
- readFile(join(taskDirectory, "tests/test.patch"), "utf8"),
22
- readFile(diagnosticsPath, "utf8"),
23
- readFile(definitionPath, "utf8"),
24
- ]);
25
- const originalDefinition = taskDefinitionSchema.parse(JSON.parse(originalDefinitionBytes));
26
- const allowedPaths = validationRepairPaths(originalPatch);
27
- await Promise.all([
28
- writeFile("/work/definition.json", `${JSON.stringify(originalDefinition, null, 2)}\n`),
29
- writeFile("/work/gold.patch", await readFile(join(taskDirectory, "solution/gold.patch"))),
30
- writeFile("/work/test.sh", await readFile(join(taskDirectory, "tests/test.sh"))),
31
- ]);
32
- const piHome = join(homedir(), ".pi/agent");
33
- const piAuth = process.env.SELFBENCH_PI_AUTH_JSON;
34
- if (!process.env.OPENAI_API_KEY && !piAuth) {
35
- fail("OPENAI_API_KEY or SELFBENCH_PI_AUTH_JSON is required");
36
- }
37
- await mkdir(piHome, { recursive: true });
38
- await Promise.all([
39
- ...(piAuth
40
- ? [
41
- writeFile(join(piHome, "auth.json"), piAuth).then(() => chmod(join(piHome, "auth.json"), 0o600)),
42
- ]
43
- : []),
44
- writeFile(join(piHome, "settings.json"), `${JSON.stringify({ transport: "auto" })}\n`),
45
- ]);
46
- const promptPath = join(tmpdir(), `selfbench-validation-repair-${originalDefinition.taskId}.md`);
47
- await writeFile(promptPath, validationRepairPrompt({
48
- definition: originalDefinition,
49
- authenticRequest: instruction,
50
- diagnostics,
51
- allowedPaths,
52
- }));
53
- const pi = await runCommand("pi", [
54
- "--print",
55
- "--mode",
56
- "json",
57
- "--no-session",
58
- "--no-approve",
59
- "--no-skills",
60
- "--no-prompt-templates",
61
- "--no-context-files",
62
- "--no-extensions",
63
- "--provider",
64
- process.env.OPENAI_API_KEY ? "openai" : "openai-codex",
65
- "--model",
66
- process.env.SELFBENCH_REPAIR_MODEL ?? "gpt-5.6-sol",
67
- "--thinking",
68
- "high",
69
- "--tools",
70
- "read,bash,grep,find,ls",
71
- `@${promptPath}`,
72
- ], {
73
- allowFailure: true,
74
- timeoutMs: 90 * 60 * 1000,
75
- cwd: repositoryDirectory,
76
- env: process.env,
77
- onOutput: (stream, chunk) => {
78
- (stream === "stdout" ? process.stdout : process.stderr).write(chunk);
79
- },
80
- });
81
- if (pi.exitCode !== 0) {
82
- throw new Error(`Pi validation repair exited ${pi.exitCode}: ${pi.stderr.slice(-2_000)}`);
83
- }
84
- const repairedDefinition = taskDefinitionSchema.parse(JSON.parse(await readFile("/work/definition.json", "utf8")));
85
- const [tracked, untracked] = await Promise.all([
86
- runCommand("git", ["-C", repositoryDirectory, "diff", "--name-only", "HEAD"]),
87
- runCommand("git", ["-C", repositoryDirectory, "ls-files", "--others", "--exclude-standard"]),
88
- ]);
89
- const changedPaths = [...tracked.stdout.split("\n"), ...untracked.stdout.split("\n")]
90
- .filter(Boolean)
91
- .sort();
92
- assertValidationRepair(originalDefinition, repairedDefinition, originalPatch, changedPaths);
93
- const repaired = await runCommand("git", ["-C", repositoryDirectory, "diff", "--binary", "HEAD"]);
94
- const repairedPatch = repaired.stdout.startsWith("diff --git ") ? repaired.stdout : originalPatch;
95
- if (repairedPatch === originalPatch &&
96
- JSON.stringify(repairedDefinition) === JSON.stringify(originalDefinition)) {
97
- throw new Error("validation repair left the task unchanged");
98
- }
99
- await Promise.all([
100
- writeFile(outputDefinition, `${JSON.stringify(repairedDefinition, null, 2)}\n`),
101
- writeFile(outputPatch, repairedPatch),
102
- writeFile(outputReport, `${JSON.stringify({
103
- schemaVersion: 1,
104
- taskId: originalDefinition.taskId,
105
- model: process.env.SELFBENCH_REPAIR_MODEL ?? "gpt-5.6-sol",
106
- changedPaths,
107
- definitionChanged: JSON.stringify(repairedDefinition) !== JSON.stringify(originalDefinition),
108
- testsChanged: repairedPatch !== originalPatch,
109
- piOutputTail: pi.stdout.slice(-4_000),
110
- }, null, 2)}\n`),
111
- ]);
112
- function fail(message) {
113
- throw new Error(message);
114
- }
115
- //# sourceMappingURL=validation-repair.js.map
@@ -1 +0,0 @@
1
- {"version":3,"file":"validation-repair.js","sourceRoot":"","sources":["../../../src/sandbox/programs/validation-repair.ts"],"names":[],"mappings":";AAEA,OAAO,EAAE,KAAK,EAAE,KAAK,EAAE,QAAQ,EAAE,SAAS,EAAE,MAAM,kBAAkB,CAAC;AACrE,OAAO,EAAE,OAAO,EAAE,MAAM,EAAE,MAAM,SAAS,CAAC;AAC1C,OAAO,EAAE,IAAI,EAAE,MAAM,WAAW,CAAC;AACjC,OAAO,EAAE,oBAAoB,EAAE,MAAM,oBAAoB,CAAC;AAC1D,OAAO,EAAE,UAAU,EAAE,MAAM,kBAAkB,CAAC;AAC9C,OAAO,EACL,sBAAsB,EACtB,qBAAqB,EACrB,sBAAsB,GACvB,MAAM,4BAA4B,CAAC;AACpC,OAAO,EAAE,iBAAiB,EAAE,MAAM,qBAAqB,CAAC;AAExD,MAAM,CAAC,WAAW,EAAE,cAAc,EAAE,eAAe,EAAE,gBAAgB,EAAE,WAAW,EAAE,YAAY,CAAC,GAC/F,OAAO,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;AACxB,IACE,CAAC,WAAW;IACZ,CAAC,cAAc;IACf,CAAC,eAAe;IAChB,CAAC,gBAAgB;IACjB,CAAC,WAAW;IACZ,CAAC,YAAY,EACb,CAAC;IACD,MAAM,IAAI,KAAK,CACb,qIAAqI,CACtI,CAAC;AACJ,CAAC;AAED,MAAM,EAAE,aAAa,EAAE,mBAAmB,EAAE,GAAG,MAAM,iBAAiB,CAAC,WAAW,CAAC,CAAC;AACpF,MAAM,CAAC,WAAW,EAAE,aAAa,EAAE,WAAW,EAAE,uBAAuB,CAAC,GAAG,MAAM,OAAO,CAAC,GAAG,CAAC;IAC3F,QAAQ,CAAC,IAAI,CAAC,aAAa,EAAE,gBAAgB,CAAC,EAAE,MAAM,CAAC;IACvD,QAAQ,CAAC,IAAI,CAAC,aAAa,EAAE,kBAAkB,CAAC,EAAE,MAAM,CAAC;IACzD,QAAQ,CAAC,eAAe,EAAE,MAAM,CAAC;IACjC,QAAQ,CAAC,cAAc,EAAE,MAAM,CAAC;CACjC,CAAC,CAAC;AACH,MAAM,kBAAkB,GAAG,oBAAoB,CAAC,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC,uBAAuB,CAAC,CAAC,CAAC;AAC3F,MAAM,YAAY,GAAG,qBAAqB,CAAC,aAAa,CAAC,CAAC;AAE1D,MAAM,OAAO,CAAC,GAAG,CAAC;IAChB,SAAS,CAAC,uBAAuB,EAAE,GAAG,IAAI,CAAC,SAAS,CAAC,kBAAkB,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,CAAC;IACtF,SAAS,CAAC,kBAAkB,EAAE,MAAM,QAAQ,CAAC,IAAI,CAAC,aAAa,EAAE,qBAAqB,CAAC,CAAC,CAAC;IACzF,SAAS,CAAC,eAAe,EAAE,MAAM,QAAQ,CAAC,IAAI,CAAC,aAAa,EAAE,eAAe,CAAC,CAAC,CAAC;CACjF,CAAC,CAAC;AAEH,MAAM,MAAM,GAAG,IAAI,CAAC,OAAO,EAAE,EAAE,WAAW,CAAC,CAAC;AAC5C,MAAM,MAAM,GAAG,OAAO,CAAC,GAAG,CAAC,sBAAsB,CAAC;AAClD,IAAI,CAAC,OAAO,CAAC,GAAG,CAAC,cAAc,IAAI,CAAC,MAAM,EAAE,CAAC;IAC3C,IAAI,CAAC,sDAAsD,CAAC,CAAC;AAC/D,CAAC;AACD,MAAM,KAAK,CAAC,MAAM,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC,CAAC;AACzC,MAAM,OAAO,CAAC,GAAG,CAAC;IAChB,GAAG,CAAC,MAAM;QACR,CAAC,CAAC;YACE,SAAS,CAAC,IAAI,CAAC,MAAM,EAAE,WAAW,CAAC,EAAE,MAAM,CAAC,CAAC,IAAI,CAAC,GAAG,EAAE,CACrD,KAAK,CAAC,IAAI,CAAC,MAAM,EAAE,WAAW,CAAC,EAAE,KAAK,CAAC,CACxC;SACF;QACH,CAAC,CAAC,EAAE,CAAC;IACP,SAAS,CAAC,IAAI,CAAC,MAAM,EAAE,eAAe,CAAC,EAAE,GAAG,IAAI,CAAC,SAAS,CAAC,EAAE,SAAS,EAAE,MAAM,EAAE,CAAC,IAAI,CAAC;CACvF,CAAC,CAAC;AACH,MAAM,UAAU,GAAG,IAAI,CAAC,MAAM,EAAE,EAAE,+BAA+B,kBAAkB,CAAC,MAAM,KAAK,CAAC,CAAC;AACjG,MAAM,SAAS,CACb,UAAU,EACV,sBAAsB,CAAC;IACrB,UAAU,EAAE,kBAAkB;IAC9B,gBAAgB,EAAE,WAAW;IAC7B,WAAW;IACX,YAAY;CACb,CAAC,CACH,CAAC;AAEF,MAAM,EAAE,GAAG,MAAM,UAAU,CACzB,IAAI,EACJ;IACE,SAAS;IACT,QAAQ;IACR,MAAM;IACN,cAAc;IACd,cAAc;IACd,aAAa;IACb,uBAAuB;IACvB,oBAAoB;IACpB,iBAAiB;IACjB,YAAY;IACZ,OAAO,CAAC,GAAG,CAAC,cAAc,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,cAAc;IACtD,SAAS;IACT,OAAO,CAAC,GAAG,CAAC,sBAAsB,IAAI,aAAa;IACnD,YAAY;IACZ,MAAM;IACN,SAAS;IACT,wBAAwB;IACxB,IAAI,UAAU,EAAE;CACjB,EACD;IACE,YAAY,EAAE,IAAI;IAClB,SAAS,EAAE,EAAE,GAAG,EAAE,GAAG,IAAI;IACzB,GAAG,EAAE,mBAAmB;IACxB,GAAG,EAAE,OAAO,CAAC,GAAG;IAChB,QAAQ,EAAE,CAAC,MAAM,EAAE,KAAK,EAAE,EAAE;QAC1B,CAAC,MAAM,KAAK,QAAQ,CAAC,CAAC,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC;IACvE,CAAC;CACF,CACF,CAAC;AACF,IAAI,EAAE,CAAC,QAAQ,KAAK,CAAC,EAAE,CAAC;IACtB,MAAM,IAAI,KAAK,CAAC,+BAA+B,EAAE,CAAC,QAAQ,KAAK,EAAE,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC;AAC5F,CAAC;AAED,MAAM,kBAAkB,GAAG,oBAAoB,CAAC,KAAK,CACnD,IAAI,CAAC,KAAK,CAAC,MAAM,QAAQ,CAAC,uBAAuB,EAAE,MAAM,CAAC,CAAC,CAC5D,CAAC;AACF,MAAM,CAAC,OAAO,EAAE,SAAS,CAAC,GAAG,MAAM,OAAO,CAAC,GAAG,CAAC;IAC7C,UAAU,CAAC,KAAK,EAAE,CAAC,IAAI,EAAE,mBAAmB,EAAE,MAAM,EAAE,aAAa,EAAE,MAAM,CAAC,CAAC;IAC7E,UAAU,CAAC,KAAK,EAAE,CAAC,IAAI,EAAE,mBAAmB,EAAE,UAAU,EAAE,UAAU,EAAE,oBAAoB,CAAC,CAAC;CAC7F,CAAC,CAAC;AACH,MAAM,YAAY,GAAG,CAAC,GAAG,OAAO,CAAC,MAAM,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,GAAG,SAAS,CAAC,MAAM,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;KAClF,MAAM,CAAC,OAAO,CAAC;KACf,IAAI,EAAE,CAAC;AACV,sBAAsB,CAAC,kBAAkB,EAAE,kBAAkB,EAAE,aAAa,EAAE,YAAY,CAAC,CAAC;AAC5F,MAAM,QAAQ,GAAG,MAAM,UAAU,CAAC,KAAK,EAAE,CAAC,IAAI,EAAE,mBAAmB,EAAE,MAAM,EAAE,UAAU,EAAE,MAAM,CAAC,CAAC,CAAC;AAClG,MAAM,aAAa,GAAG,QAAQ,CAAC,MAAM,CAAC,UAAU,CAAC,aAAa,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,MAAM,CAAC,CAAC,CAAC,aAAa,CAAC;AAClG,IACE,aAAa,KAAK,aAAa;IAC/B,IAAI,CAAC,SAAS,CAAC,kBAAkB,CAAC,KAAK,IAAI,CAAC,SAAS,CAAC,kBAAkB,CAAC,EACzE,CAAC;IACD,MAAM,IAAI,KAAK,CAAC,2CAA2C,CAAC,CAAC;AAC/D,CAAC;AAED,MAAM,OAAO,CAAC,GAAG,CAAC;IAChB,SAAS,CAAC,gBAAgB,EAAE,GAAG,IAAI,CAAC,SAAS,CAAC,kBAAkB,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,CAAC;IAC/E,SAAS,CAAC,WAAW,EAAE,aAAa,CAAC;IACrC,SAAS,CACP,YAAY,EACZ,GAAG,IAAI,CAAC,SAAS,CACf;QACE,aAAa,EAAE,CAAC;QAChB,MAAM,EAAE,kBAAkB,CAAC,MAAM;QACjC,KAAK,EAAE,OAAO,CAAC,GAAG,CAAC,sBAAsB,IAAI,aAAa;QAC1D,YAAY;QACZ,iBAAiB,EACf,IAAI,CAAC,SAAS,CAAC,kBAAkB,CAAC,KAAK,IAAI,CAAC,SAAS,CAAC,kBAAkB,CAAC;QAC3E,YAAY,EAAE,aAAa,KAAK,aAAa;QAC7C,YAAY,EAAE,EAAE,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,KAAK,CAAC;KACtC,EACD,IAAI,EACJ,CAAC,CACF,IAAI,CACN;CACF,CAAC,CAAC;AAEH,SAAS,IAAI,CAAC,OAAe;IAC3B,MAAM,IAAI,KAAK,CAAC,OAAO,CAAC,CAAC;AAC3B,CAAC"}
@@ -1,323 +0,0 @@
1
- #!/usr/bin/env node
2
-
3
- // src/sandbox/programs/repair.ts
4
- import { chmod, mkdir as mkdir2, readFile, writeFile } from "node:fs/promises";
5
- import { homedir, tmpdir } from "node:os";
6
- import { join as join2 } from "node:path";
7
-
8
- // src/hash.ts
9
- import { createHash } from "node:crypto";
10
- function sha256(value) {
11
- return createHash("sha256").update(value).digest("hex");
12
- }
13
-
14
- // src/process.ts
15
- import { spawn } from "node:child_process";
16
-
17
- class InactivityTimeoutError extends Error {
18
- scope;
19
- timeoutMs;
20
- constructor(scope, timeoutMs) {
21
- super(`${scope} produced no output for ${timeoutMs}ms`);
22
- this.scope = scope;
23
- this.timeoutMs = timeoutMs;
24
- this.name = "InactivityTimeoutError";
25
- }
26
- }
27
- async function runCommand(command, args, options = {}) {
28
- return await new Promise((resolve, reject) => {
29
- const child = spawn(command, args, {
30
- cwd: options.cwd,
31
- env: options.env ?? process.env,
32
- stdio: [options.input === undefined ? "ignore" : "pipe", "pipe", "pipe"]
33
- });
34
- const stdout = new RollingOutput;
35
- const stderr = new RollingOutput;
36
- if (!child.stdout || !child.stderr) {
37
- reject(new Error(`failed to capture ${command} output`));
38
- return;
39
- }
40
- let inactivityTimer;
41
- let forceKillTimer;
42
- let inactivityError;
43
- const clearInactivityTimer = () => {
44
- if (inactivityTimer) {
45
- clearTimeout(inactivityTimer);
46
- inactivityTimer = undefined;
47
- }
48
- };
49
- const terminate = () => {
50
- child.kill("SIGTERM");
51
- forceKillTimer ??= setTimeout(() => child.kill("SIGKILL"), 5000);
52
- forceKillTimer.unref();
53
- };
54
- const armInactivityTimer = () => {
55
- clearInactivityTimer();
56
- if (!options.inactivityTimeoutMs) {
57
- return;
58
- }
59
- inactivityTimer = setTimeout(() => {
60
- inactivityError = new InactivityTimeoutError(`${command} ${args.join(" ")}`, options.inactivityTimeoutMs ?? 0);
61
- terminate();
62
- }, options.inactivityTimeoutMs);
63
- inactivityTimer.unref();
64
- };
65
- const capture = (stream, output, chunk) => {
66
- output.push(chunk);
67
- options.onOutput?.(stream, chunk);
68
- armInactivityTimer();
69
- };
70
- child.stdout.on("data", (chunk) => capture("stdout", stdout, chunk));
71
- child.stderr.on("data", (chunk) => capture("stderr", stderr, chunk));
72
- child.on("error", reject);
73
- const abort = () => terminate();
74
- if (options.signal?.aborted) {
75
- abort();
76
- } else {
77
- options.signal?.addEventListener("abort", abort, { once: true });
78
- }
79
- const timeout = options.timeoutMs ? setTimeout(() => terminate(), options.timeoutMs) : undefined;
80
- timeout?.unref();
81
- armInactivityTimer();
82
- child.on("close", (exitCode, signal) => {
83
- options.signal?.removeEventListener("abort", abort);
84
- clearInactivityTimer();
85
- if (forceKillTimer) {
86
- clearTimeout(forceKillTimer);
87
- }
88
- if (timeout) {
89
- clearTimeout(timeout);
90
- }
91
- const result = {
92
- stdout: stdout.text(),
93
- stderr: stderr.text(),
94
- exitCode: exitCode ?? (signal ? 128 : 1)
95
- };
96
- if (inactivityError) {
97
- reject(inactivityError);
98
- return;
99
- }
100
- if (result.exitCode !== 0 && !options.allowFailure) {
101
- const detail = result.stderr.trim() || result.stdout.trim();
102
- reject(new Error(`${command} exited ${result.exitCode}${detail ? `: ${detail}` : ""}`));
103
- return;
104
- }
105
- resolve(result);
106
- });
107
- if (child.stdin && options.input !== undefined) {
108
- child.stdin.end(options.input);
109
- }
110
- });
111
- }
112
-
113
- class RollingOutput {
114
- #limit = 8 * 1024 * 1024;
115
- #chunks = [];
116
- #size = 0;
117
- #truncated = false;
118
- push(chunk) {
119
- if (chunk.byteLength >= this.#limit) {
120
- this.#chunks.length = 0;
121
- this.#chunks.push(chunk.subarray(chunk.byteLength - this.#limit));
122
- this.#size = this.#limit;
123
- this.#truncated = true;
124
- return;
125
- }
126
- this.#chunks.push(chunk);
127
- this.#size += chunk.byteLength;
128
- while (this.#size > this.#limit) {
129
- const first = this.#chunks.shift();
130
- if (!first) {
131
- break;
132
- }
133
- const overflow = this.#size - this.#limit;
134
- if (first.byteLength > overflow) {
135
- this.#chunks.unshift(first.subarray(overflow));
136
- this.#size -= overflow;
137
- } else {
138
- this.#size -= first.byteLength;
139
- }
140
- this.#truncated = true;
141
- }
142
- }
143
- text() {
144
- const value = Buffer.concat(this.#chunks, this.#size).toString("utf8");
145
- return this.#truncated ? `[selfbench: earlier output truncated]
146
- ${value}` : value;
147
- }
148
- }
149
-
150
- // src/repair.ts
151
- function patchPaths(patch) {
152
- const paths = new Set;
153
- for (const line of patch.split(`
154
- `)) {
155
- const match = /^diff --git a\/(.+) b\/(.+)$/.exec(line);
156
- if (match?.[2]) {
157
- paths.add(match[2]);
158
- }
159
- }
160
- return [...paths].sort();
161
- }
162
- function assertRepairPaths(originalTestPatch, changedPaths) {
163
- const allowed = new Set(patchPaths(originalTestPatch));
164
- if (allowed.size === 0) {
165
- throw new Error("original held-out test patch changes no files");
166
- }
167
- const outside = changedPaths.filter((path) => !allowed.has(path));
168
- if (outside.length > 0) {
169
- throw new Error(`repair changed files outside the held-out tests: ${outside.join(", ")}`);
170
- }
171
- }
172
- function repairPrompt(input) {
173
- return `Repair the held-out tests for SelfBench task ${input.taskId}.
174
-
175
- The repository is the exact base snapshot with the current held-out test patch already applied. You may edit only these existing test files:
176
-
177
- ${input.allowedPaths.map((path) => `- ${path}`).join(`
178
- `)}
179
-
180
- The authentic engineer request is:
181
-
182
- <authentic_request>
183
- ${input.authenticRequest.trim()}
184
- </authentic_request>
185
-
186
- The independent coupling report is:
187
-
188
- <coupling_report>
189
- ${input.couplingReport.trim()}
190
- </coupling_report>
191
-
192
- Rewrite the tests so they verify every material requested behavior through stable public boundaries without requiring the gold patch's exact internal names, helper structure, error prose, mock input shape, response presentation, or incidental implementation choices. Preserve meaningful negative, authorization, compatibility, and regression coverage. Do not edit application code, the request, or the gold patch. Do not delete assertions merely to silence the report. A coherent alternative implementation must be able to pass, while the unchanged base must still fail the fail-to-pass tests and the gold implementation must still pass.
193
-
194
- The final working tree must retain a non-empty test patch relative to HEAD and every originally added fail-to-pass test must still have an equivalent behavioral test. Do not reset, revert, delete, or commit the test patch. If an exact public response field is the only stable seam, prefer asserting the underlying externally visible behavior through existing endpoints, persistence, or follow-up actions. If no rigorous uncoupled test is possible within the allowed files, leave the current tests in place and explain the blocker in your final message instead of removing coverage.
195
-
196
- Inspect /work/task/solution/gold.patch only to understand the intended behavior and available seams, never to copy its private structure into assertions. Inspect /work/task/tests/test.sh for the verifier command. Run focused tests when feasible. Finish only after the working tree contains the repaired test files and no other changes.`;
197
- }
198
-
199
- // src/sandbox/programs/prepare-repair.ts
200
- import { access, mkdir } from "node:fs/promises";
201
- import { join } from "node:path";
202
- async function prepareRepairTask(archivePath, workDirectory = "/work") {
203
- const extractedDirectory = join(workDirectory, "task");
204
- const repositoryDirectory = join(workDirectory, "repo");
205
- await Promise.all([mkdir(extractedDirectory, { recursive: true }), mkdir(repositoryDirectory)]);
206
- await runCommand("tar", ["-xzf", archivePath, "-C", extractedDirectory]);
207
- const taskDirectory = await access(join(extractedDirectory, "instruction.md")).then(() => extractedDirectory, () => join(extractedDirectory, "harbor-task"));
208
- await runCommand("tar", [
209
- "-xzf",
210
- join(taskDirectory, "tests/repo.tar.gz"),
211
- "-C",
212
- repositoryDirectory
213
- ]);
214
- await runCommand("git", ["-C", repositoryDirectory, "init", "-q"]);
215
- await runCommand("git", ["-C", repositoryDirectory, "config", "user.name", "SelfBench"]);
216
- await runCommand("git", ["-C", repositoryDirectory, "config", "user.email", "selfbench@local"]);
217
- await runCommand("git", ["-C", repositoryDirectory, "add", "-A"]);
218
- await runCommand("git", ["-C", repositoryDirectory, "commit", "-qm", "base"]);
219
- await runCommand("git", [
220
- "-C",
221
- repositoryDirectory,
222
- "apply",
223
- join(taskDirectory, "tests/test.patch")
224
- ]);
225
- await runCommand("git", ["-C", repositoryDirectory, "add", "-N", "--all"]);
226
- return { extractedDirectory, taskDirectory, repositoryDirectory };
227
- }
228
-
229
- // src/sandbox/programs/repair.ts
230
- var [archivePath, reviewPath, outputArchive, outputReport] = process.argv.slice(2);
231
- if (!archivePath || !reviewPath || !outputArchive || !outputReport) {
232
- throw new Error("usage: sandbox-repair TASK.tar.gz REVIEW.json OUTPUT.tar.gz OUTPUT-REPORT.json");
233
- }
234
- var { extractedDirectory, taskDirectory, repositoryDirectory } = await prepareRepairTask(archivePath);
235
- var [instruction, originalPatch, review] = await Promise.all([
236
- readFile(join2(taskDirectory, "instruction.md"), "utf8"),
237
- readFile(join2(taskDirectory, "tests/test.patch"), "utf8"),
238
- readFile(reviewPath, "utf8")
239
- ]);
240
- var manifestPath = join2(taskDirectory, ".selfbench-manifest.json");
241
- var manifest = JSON.parse(await readFile(manifestPath, "utf8"));
242
- var taskId = typeof manifest.taskId === "string" ? manifest.taskId : "unknown-task";
243
- var allowedPaths = patchPaths(originalPatch);
244
- var codexAuth = process.env.SELFBENCH_CODEX_AUTH_JSON;
245
- var apiKey = process.env.OPENAI_API_KEY;
246
- if (!apiKey && !codexAuth) {
247
- fail("OPENAI_API_KEY or SELFBENCH_CODEX_AUTH_JSON is required");
248
- }
249
- var codexHome = join2(homedir(), ".codex");
250
- var authPath = join2(codexHome, "auth.json");
251
- await mkdir2(codexHome, { recursive: true });
252
- await writeFile(authPath, codexAuth ?? `${JSON.stringify({ OPENAI_API_KEY: apiKey })}
253
- `);
254
- await chmod(authPath, 384);
255
- var promptPath = join2(tmpdir(), `selfbench-repair-${taskId}.md`);
256
- await writeFile(promptPath, repairPrompt({
257
- taskId,
258
- authenticRequest: instruction,
259
- couplingReport: review,
260
- allowedPaths
261
- }));
262
- var codex = await runCommand("codex", [
263
- "exec",
264
- "--model",
265
- process.env.SELFBENCH_REPAIR_MODEL ?? "gpt-5.6-sol",
266
- "--dangerously-bypass-approvals-and-sandbox",
267
- "--ephemeral",
268
- "--ignore-user-config",
269
- "--json",
270
- "-C",
271
- repositoryDirectory,
272
- "-"
273
- ], {
274
- allowFailure: true,
275
- timeoutMs: 90 * 60 * 1000,
276
- env: process.env,
277
- input: await readFile(promptPath, "utf8"),
278
- onOutput: (stream, chunk) => {
279
- (stream === "stdout" ? process.stdout : process.stderr).write(chunk);
280
- }
281
- });
282
- if (codex.exitCode !== 0) {
283
- throw new Error(`Codex repair exited ${codex.exitCode}: ${codex.stderr.slice(-2000)}`);
284
- }
285
- var [tracked, untracked] = await Promise.all([
286
- runCommand("git", ["-C", repositoryDirectory, "diff", "--name-only", "HEAD"]),
287
- runCommand("git", ["-C", repositoryDirectory, "ls-files", "--others", "--exclude-standard"])
288
- ]);
289
- var changedPaths = [...tracked.stdout.split(`
290
- `), ...untracked.stdout.split(`
291
- `)].filter(Boolean).sort();
292
- assertRepairPaths(originalPatch, changedPaths);
293
- var repaired = await runCommand("git", ["-C", repositoryDirectory, "diff", "--binary", "HEAD"]);
294
- if (!repaired.stdout.startsWith("diff --git ")) {
295
- throw new Error(`repair produced no held-out test patch; status=${JSON.stringify(changedPaths)}; Codex tail=${codex.stdout.slice(-4000)}`);
296
- }
297
- if (repaired.stdout === originalPatch) {
298
- throw new Error("repair left the held-out test patch unchanged");
299
- }
300
- await writeFile(join2(taskDirectory, "tests/test.patch"), repaired.stdout);
301
- await writeFile(manifestPath, `${JSON.stringify({
302
- ...manifest,
303
- testPatchSha256: sha256(repaired.stdout),
304
- repair: {
305
- model: process.env.SELFBENCH_REPAIR_MODEL ?? "gpt-5.6-sol",
306
- originalTestPatchSha256: sha256(originalPatch)
307
- }
308
- }, null, 2)}
309
- `);
310
- await runCommand("tar", ["-czf", outputArchive, "-C", extractedDirectory, "."]);
311
- await writeFile(outputReport, `${JSON.stringify({
312
- schemaVersion: 1,
313
- taskId,
314
- model: process.env.SELFBENCH_REPAIR_MODEL ?? "gpt-5.6-sol",
315
- changedPaths,
316
- originalTestPatchSha256: sha256(originalPatch),
317
- repairedTestPatchSha256: sha256(repaired.stdout),
318
- codexOutputTail: codex.stdout.slice(-4000)
319
- }, null, 2)}
320
- `);
321
- function fail(message) {
322
- throw new Error(message);
323
- }