@apralabs/apra-fleet 0.3.6 → 0.4.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (569) hide show
  1. package/README.md +245 -257
  2. package/dist/agents/backlog-groomer.md +287 -0
  3. package/dist/agents/deployer.md +52 -11
  4. package/dist/agents/doer.md +32 -7
  5. package/dist/agents/integ-test-runner.md +151 -61
  6. package/dist/agents/plan-reviewer.md +74 -17
  7. package/dist/agents/planner.md +94 -14
  8. package/dist/agents/regression-test-runner.md +216 -0
  9. package/dist/agents/reviewer.md +21 -6
  10. package/dist/agents/schemas/backlog-groomer-input.json +35 -0
  11. package/dist/agents/schemas/backlog-groomer-output.json +164 -0
  12. package/dist/agents/schemas/integ-test-runner-output.json +34 -0
  13. package/dist/agents/schemas/regression-test-runner-input.json +15 -0
  14. package/dist/agents/schemas/regression-test-runner-output.json +50 -0
  15. package/dist/agents/schemas/reviewer-output.json +4 -0
  16. package/dist/cli/agent-transform.d.ts +1 -0
  17. package/dist/cli/agent-transform.d.ts.map +1 -1
  18. package/dist/cli/agent-transform.js +77 -0
  19. package/dist/cli/agent-transform.js.map +1 -1
  20. package/dist/cli/auth.d.ts +3 -0
  21. package/dist/cli/auth.d.ts.map +1 -1
  22. package/dist/cli/auth.js +181 -5
  23. package/dist/cli/auth.js.map +1 -1
  24. package/dist/cli/config.d.ts +5 -1
  25. package/dist/cli/config.d.ts.map +1 -1
  26. package/dist/cli/config.js +7 -1
  27. package/dist/cli/config.js.map +1 -1
  28. package/dist/cli/dolt-install.d.ts +91 -0
  29. package/dist/cli/dolt-install.d.ts.map +1 -0
  30. package/dist/cli/dolt-install.js +334 -0
  31. package/dist/cli/dolt-install.js.map +1 -0
  32. package/dist/cli/install-guard.d.ts +29 -0
  33. package/dist/cli/install-guard.d.ts.map +1 -0
  34. package/dist/cli/install-guard.js +153 -0
  35. package/dist/cli/install-guard.js.map +1 -0
  36. package/dist/cli/install.d.ts +44 -3
  37. package/dist/cli/install.d.ts.map +1 -1
  38. package/dist/cli/install.js +604 -108
  39. package/dist/cli/install.js.map +1 -1
  40. package/dist/cli/join.d.ts +12 -0
  41. package/dist/cli/join.d.ts.map +1 -0
  42. package/dist/cli/join.js +87 -0
  43. package/dist/cli/join.js.map +1 -0
  44. package/dist/cli/register-member.d.ts +2 -0
  45. package/dist/cli/register-member.d.ts.map +1 -0
  46. package/dist/cli/register-member.js +193 -0
  47. package/dist/cli/register-member.js.map +1 -0
  48. package/dist/cli/restart.d.ts +2 -0
  49. package/dist/cli/restart.d.ts.map +1 -0
  50. package/dist/cli/restart.js +7 -0
  51. package/dist/cli/restart.js.map +1 -0
  52. package/dist/cli/spoke.d.ts +35 -0
  53. package/dist/cli/spoke.d.ts.map +1 -0
  54. package/dist/cli/spoke.js +212 -0
  55. package/dist/cli/spoke.js.map +1 -0
  56. package/dist/cli/start.d.ts +2 -0
  57. package/dist/cli/start.d.ts.map +1 -0
  58. package/dist/cli/start.js +92 -0
  59. package/dist/cli/start.js.map +1 -0
  60. package/dist/cli/status.d.ts +2 -0
  61. package/dist/cli/status.d.ts.map +1 -0
  62. package/dist/cli/status.js +83 -0
  63. package/dist/cli/status.js.map +1 -0
  64. package/dist/cli/stop.d.ts +2 -0
  65. package/dist/cli/stop.d.ts.map +1 -0
  66. package/dist/cli/stop.js +51 -0
  67. package/dist/cli/stop.js.map +1 -0
  68. package/dist/cli/uninstall.d.ts.map +1 -1
  69. package/dist/cli/uninstall.js +109 -9
  70. package/dist/cli/uninstall.js.map +1 -1
  71. package/dist/cli/update.d.ts.map +1 -1
  72. package/dist/cli/update.js +10 -1
  73. package/dist/cli/update.js.map +1 -1
  74. package/dist/cli/workflow-assets.d.ts +37 -0
  75. package/dist/cli/workflow-assets.d.ts.map +1 -0
  76. package/dist/cli/workflow-assets.js +144 -0
  77. package/dist/cli/workflow-assets.js.map +1 -0
  78. package/dist/cli/workflow.d.ts +75 -0
  79. package/dist/cli/workflow.d.ts.map +1 -0
  80. package/dist/cli/workflow.js +397 -0
  81. package/dist/cli/workflow.js.map +1 -0
  82. package/dist/hub-service/activity.d.ts +30 -0
  83. package/dist/hub-service/activity.d.ts.map +1 -0
  84. package/dist/hub-service/activity.js +20 -0
  85. package/dist/hub-service/activity.js.map +1 -0
  86. package/dist/hub-service/audit-log.d.ts +19 -0
  87. package/dist/hub-service/audit-log.d.ts.map +1 -0
  88. package/dist/hub-service/audit-log.js +13 -0
  89. package/dist/hub-service/audit-log.js.map +1 -0
  90. package/dist/hub-service/db/pool.d.ts +19 -0
  91. package/dist/hub-service/db/pool.d.ts.map +1 -0
  92. package/dist/hub-service/db/pool.js +53 -0
  93. package/dist/hub-service/db/pool.js.map +1 -0
  94. package/dist/hub-service/enrollment.d.ts +29 -0
  95. package/dist/hub-service/enrollment.d.ts.map +1 -0
  96. package/dist/hub-service/enrollment.js +48 -0
  97. package/dist/hub-service/enrollment.js.map +1 -0
  98. package/dist/hub-service/envelope-routes.d.ts +46 -0
  99. package/dist/hub-service/envelope-routes.d.ts.map +1 -0
  100. package/dist/hub-service/envelope-routes.js +94 -0
  101. package/dist/hub-service/envelope-routes.js.map +1 -0
  102. package/dist/hub-service/handlers/installers.d.ts +14 -0
  103. package/dist/hub-service/handlers/installers.d.ts.map +1 -0
  104. package/dist/hub-service/handlers/installers.js +8 -0
  105. package/dist/hub-service/handlers/installers.js.map +1 -0
  106. package/dist/hub-service/http-server.d.ts +31 -0
  107. package/dist/hub-service/http-server.d.ts.map +1 -0
  108. package/dist/hub-service/http-server.js +711 -0
  109. package/dist/hub-service/http-server.js.map +1 -0
  110. package/dist/hub-service/hub-jwt.d.ts +30 -0
  111. package/dist/hub-service/hub-jwt.d.ts.map +1 -0
  112. package/dist/hub-service/hub-jwt.js +105 -0
  113. package/dist/hub-service/hub-jwt.js.map +1 -0
  114. package/dist/hub-service/jwt-revocation.d.ts +20 -0
  115. package/dist/hub-service/jwt-revocation.d.ts.map +1 -0
  116. package/dist/hub-service/jwt-revocation.js +18 -0
  117. package/dist/hub-service/jwt-revocation.js.map +1 -0
  118. package/dist/hub-service/machines.d.ts +16 -0
  119. package/dist/hub-service/machines.d.ts.map +1 -0
  120. package/dist/hub-service/machines.js +14 -0
  121. package/dist/hub-service/machines.js.map +1 -0
  122. package/dist/hub-service/main.d.ts +3 -0
  123. package/dist/hub-service/main.d.ts.map +1 -0
  124. package/dist/hub-service/main.js +64 -0
  125. package/dist/hub-service/main.js.map +1 -0
  126. package/dist/hub-service/member-tokens.d.ts +11 -0
  127. package/dist/hub-service/member-tokens.d.ts.map +1 -0
  128. package/dist/hub-service/member-tokens.js +49 -0
  129. package/dist/hub-service/member-tokens.js.map +1 -0
  130. package/dist/hub-service/member-view.d.ts +20 -0
  131. package/dist/hub-service/member-view.d.ts.map +1 -0
  132. package/dist/hub-service/member-view.js +76 -0
  133. package/dist/hub-service/member-view.js.map +1 -0
  134. package/dist/hub-service/members.d.ts +44 -0
  135. package/dist/hub-service/members.d.ts.map +1 -0
  136. package/dist/hub-service/members.js +25 -0
  137. package/dist/hub-service/members.js.map +1 -0
  138. package/dist/hub-service/presence.d.ts +40 -0
  139. package/dist/hub-service/presence.d.ts.map +1 -0
  140. package/dist/hub-service/presence.js +44 -0
  141. package/dist/hub-service/presence.js.map +1 -0
  142. package/dist/hub-service/project-view.d.ts +12 -0
  143. package/dist/hub-service/project-view.d.ts.map +1 -0
  144. package/dist/hub-service/project-view.js +34 -0
  145. package/dist/hub-service/project-view.js.map +1 -0
  146. package/dist/hub-service/projects.d.ts +48 -0
  147. package/dist/hub-service/projects.d.ts.map +1 -0
  148. package/dist/hub-service/projects.js +80 -0
  149. package/dist/hub-service/projects.js.map +1 -0
  150. package/dist/hub-service/relay-queue.d.ts +115 -0
  151. package/dist/hub-service/relay-queue.d.ts.map +1 -0
  152. package/dist/hub-service/relay-queue.js +148 -0
  153. package/dist/hub-service/relay-queue.js.map +1 -0
  154. package/dist/hub-service/session-jwt.d.ts +10 -0
  155. package/dist/hub-service/session-jwt.d.ts.map +1 -0
  156. package/dist/hub-service/session-jwt.js +69 -0
  157. package/dist/hub-service/session-jwt.js.map +1 -0
  158. package/dist/hub-service/usage.d.ts +40 -0
  159. package/dist/hub-service/usage.d.ts.map +1 -0
  160. package/dist/hub-service/usage.js +25 -0
  161. package/dist/hub-service/usage.js.map +1 -0
  162. package/dist/hub-service/users.d.ts +56 -0
  163. package/dist/hub-service/users.d.ts.map +1 -0
  164. package/dist/hub-service/users.js +72 -0
  165. package/dist/hub-service/users.js.map +1 -0
  166. package/dist/hub-service/workspaces.d.ts +20 -0
  167. package/dist/hub-service/workspaces.d.ts.map +1 -0
  168. package/dist/hub-service/workspaces.js +10 -0
  169. package/dist/hub-service/workspaces.js.map +1 -0
  170. package/dist/index.js +179 -132
  171. package/dist/index.js.map +1 -1
  172. package/dist/os/linux.d.ts +13 -0
  173. package/dist/os/linux.d.ts.map +1 -1
  174. package/dist/os/linux.js +62 -3
  175. package/dist/os/linux.js.map +1 -1
  176. package/dist/os/os-commands.d.ts +10 -0
  177. package/dist/os/os-commands.d.ts.map +1 -1
  178. package/dist/os/windows.d.ts +1 -0
  179. package/dist/os/windows.d.ts.map +1 -1
  180. package/dist/os/windows.js +21 -1
  181. package/dist/os/windows.js.map +1 -1
  182. package/dist/paths.d.ts +23 -0
  183. package/dist/paths.d.ts.map +1 -1
  184. package/dist/paths.js +29 -0
  185. package/dist/paths.js.map +1 -1
  186. package/dist/providers/agy.d.ts +9 -2
  187. package/dist/providers/agy.d.ts.map +1 -1
  188. package/dist/providers/agy.js +127 -14
  189. package/dist/providers/agy.js.map +1 -1
  190. package/dist/providers/claude.d.ts +6 -2
  191. package/dist/providers/claude.d.ts.map +1 -1
  192. package/dist/providers/claude.js +248 -9
  193. package/dist/providers/claude.js.map +1 -1
  194. package/dist/providers/codex.d.ts +3 -2
  195. package/dist/providers/codex.d.ts.map +1 -1
  196. package/dist/providers/codex.js +7 -0
  197. package/dist/providers/codex.js.map +1 -1
  198. package/dist/providers/copilot.d.ts +3 -2
  199. package/dist/providers/copilot.d.ts.map +1 -1
  200. package/dist/providers/copilot.js +8 -1
  201. package/dist/providers/copilot.js.map +1 -1
  202. package/dist/providers/gemini.d.ts +3 -2
  203. package/dist/providers/gemini.d.ts.map +1 -1
  204. package/dist/providers/gemini.js +11 -1
  205. package/dist/providers/gemini.js.map +1 -1
  206. package/dist/providers/index.d.ts +1 -0
  207. package/dist/providers/index.d.ts.map +1 -1
  208. package/dist/providers/index.js +3 -0
  209. package/dist/providers/index.js.map +1 -1
  210. package/dist/providers/none.d.ts +51 -0
  211. package/dist/providers/none.d.ts.map +1 -0
  212. package/dist/providers/none.js +103 -0
  213. package/dist/providers/none.js.map +1 -0
  214. package/dist/providers/opencode.d.ts +4 -2
  215. package/dist/providers/opencode.d.ts.map +1 -1
  216. package/dist/providers/opencode.js +49 -0
  217. package/dist/providers/opencode.js.map +1 -1
  218. package/dist/providers/provider.d.ts +79 -1
  219. package/dist/providers/provider.d.ts.map +1 -1
  220. package/dist/providers/provider.js +10 -0
  221. package/dist/providers/provider.js.map +1 -1
  222. package/dist/services/agent-provisioner.d.ts +3 -3
  223. package/dist/services/agent-provisioner.d.ts.map +1 -1
  224. package/dist/services/agent-provisioner.js +12 -5
  225. package/dist/services/agent-provisioner.js.map +1 -1
  226. package/dist/services/auth-socket.d.ts.map +1 -1
  227. package/dist/services/auth-socket.js +2 -0
  228. package/dist/services/auth-socket.js.map +1 -1
  229. package/dist/services/budget-awareness.d.ts +130 -0
  230. package/dist/services/budget-awareness.d.ts.map +1 -0
  231. package/dist/services/budget-awareness.js +129 -0
  232. package/dist/services/budget-awareness.js.map +1 -0
  233. package/dist/services/cloud-sync.d.ts +35 -0
  234. package/dist/services/cloud-sync.d.ts.map +1 -0
  235. package/dist/services/cloud-sync.js +90 -0
  236. package/dist/services/cloud-sync.js.map +1 -0
  237. package/dist/services/context-admission.d.ts +82 -0
  238. package/dist/services/context-admission.d.ts.map +1 -0
  239. package/dist/services/context-admission.js +136 -0
  240. package/dist/services/context-admission.js.map +1 -0
  241. package/dist/services/event-bus.d.ts +27 -0
  242. package/dist/services/event-bus.d.ts.map +1 -0
  243. package/dist/services/event-bus.js +20 -0
  244. package/dist/services/event-bus.js.map +1 -0
  245. package/dist/services/file-chunker.d.ts +38 -0
  246. package/dist/services/file-chunker.d.ts.map +1 -0
  247. package/dist/services/file-chunker.js +63 -0
  248. package/dist/services/file-chunker.js.map +1 -0
  249. package/dist/services/file-transfer-relay.d.ts +55 -0
  250. package/dist/services/file-transfer-relay.d.ts.map +1 -0
  251. package/dist/services/file-transfer-relay.js +115 -0
  252. package/dist/services/file-transfer-relay.js.map +1 -0
  253. package/dist/services/github-app.d.ts.map +1 -1
  254. package/dist/services/github-app.js +1 -0
  255. package/dist/services/github-app.js.map +1 -1
  256. package/dist/services/http-transport.d.ts +25 -0
  257. package/dist/services/http-transport.d.ts.map +1 -0
  258. package/dist/services/http-transport.js +371 -0
  259. package/dist/services/http-transport.js.map +1 -0
  260. package/dist/services/hub-client.d.ts +69 -0
  261. package/dist/services/hub-client.d.ts.map +1 -0
  262. package/dist/services/hub-client.js +217 -0
  263. package/dist/services/hub-client.js.map +1 -0
  264. package/dist/services/jwt.d.ts +17 -0
  265. package/dist/services/jwt.d.ts.map +1 -0
  266. package/dist/services/jwt.js +70 -0
  267. package/dist/services/jwt.js.map +1 -0
  268. package/dist/services/known-sessions.d.ts +8 -0
  269. package/dist/services/known-sessions.d.ts.map +1 -0
  270. package/dist/services/known-sessions.js +40 -0
  271. package/dist/services/known-sessions.js.map +1 -0
  272. package/dist/services/model-pricing.d.ts +36 -0
  273. package/dist/services/model-pricing.d.ts.map +1 -0
  274. package/dist/services/model-pricing.js +108 -0
  275. package/dist/services/model-pricing.js.map +1 -0
  276. package/dist/services/orphan-recovery.d.ts +56 -0
  277. package/dist/services/orphan-recovery.d.ts.map +1 -0
  278. package/dist/services/orphan-recovery.js +111 -0
  279. package/dist/services/orphan-recovery.js.map +1 -0
  280. package/dist/services/pending-responses.d.ts +23 -0
  281. package/dist/services/pending-responses.d.ts.map +1 -0
  282. package/dist/services/pending-responses.js +52 -0
  283. package/dist/services/pending-responses.js.map +1 -0
  284. package/dist/services/registry.d.ts +4 -1
  285. package/dist/services/registry.d.ts.map +1 -1
  286. package/dist/services/registry.js +3 -0
  287. package/dist/services/registry.js.map +1 -1
  288. package/dist/services/relay-context.d.ts +23 -0
  289. package/dist/services/relay-context.d.ts.map +1 -0
  290. package/dist/services/relay-context.js +8 -0
  291. package/dist/services/relay-context.js.map +1 -0
  292. package/dist/services/relay-executor.d.ts +68 -0
  293. package/dist/services/relay-executor.d.ts.map +1 -0
  294. package/dist/services/relay-executor.js +70 -0
  295. package/dist/services/relay-executor.js.map +1 -0
  296. package/dist/services/relay-request.d.ts +65 -0
  297. package/dist/services/relay-request.d.ts.map +1 -0
  298. package/dist/services/relay-request.js +129 -0
  299. package/dist/services/relay-request.js.map +1 -0
  300. package/dist/services/relay-strategy.d.ts +19 -0
  301. package/dist/services/relay-strategy.d.ts.map +1 -0
  302. package/dist/services/relay-strategy.js +111 -0
  303. package/dist/services/relay-strategy.js.map +1 -0
  304. package/dist/services/service-manager/index.d.ts +5 -0
  305. package/dist/services/service-manager/index.d.ts.map +1 -0
  306. package/dist/services/service-manager/index.js +66 -0
  307. package/dist/services/service-manager/index.js.map +1 -0
  308. package/dist/services/service-manager/linux.d.ts +10 -0
  309. package/dist/services/service-manager/linux.d.ts.map +1 -0
  310. package/dist/services/service-manager/linux.js +97 -0
  311. package/dist/services/service-manager/linux.js.map +1 -0
  312. package/dist/services/service-manager/macos.d.ts +10 -0
  313. package/dist/services/service-manager/macos.d.ts.map +1 -0
  314. package/dist/services/service-manager/macos.js +94 -0
  315. package/dist/services/service-manager/macos.js.map +1 -0
  316. package/dist/services/service-manager/types.d.ts +18 -0
  317. package/dist/services/service-manager/types.d.ts.map +1 -0
  318. package/dist/services/service-manager/types.js +5 -0
  319. package/dist/services/service-manager/types.js.map +1 -0
  320. package/dist/services/service-manager/windows.d.ts +10 -0
  321. package/dist/services/service-manager/windows.d.ts.map +1 -0
  322. package/dist/services/service-manager/windows.js +73 -0
  323. package/dist/services/service-manager/windows.js.map +1 -0
  324. package/dist/services/session-registry.d.ts +70 -0
  325. package/dist/services/session-registry.d.ts.map +1 -0
  326. package/dist/services/session-registry.js +91 -0
  327. package/dist/services/session-registry.js.map +1 -0
  328. package/dist/services/singleton.d.ts +15 -0
  329. package/dist/services/singleton.d.ts.map +1 -0
  330. package/dist/services/singleton.js +98 -0
  331. package/dist/services/singleton.js.map +1 -0
  332. package/dist/services/sprint-coordination.d.ts +168 -0
  333. package/dist/services/sprint-coordination.d.ts.map +1 -0
  334. package/dist/services/sprint-coordination.js +564 -0
  335. package/dist/services/sprint-coordination.js.map +1 -0
  336. package/dist/services/ssh.d.ts.map +1 -1
  337. package/dist/services/ssh.js +70 -3
  338. package/dist/services/ssh.js.map +1 -1
  339. package/dist/services/stall/log-path-resolver.d.ts.map +1 -1
  340. package/dist/services/stall/log-path-resolver.js +8 -1
  341. package/dist/services/stall/log-path-resolver.js.map +1 -1
  342. package/dist/services/stall/stall-detector.d.ts.map +1 -1
  343. package/dist/services/stall/stall-detector.js +39 -7
  344. package/dist/services/stall/stall-detector.js.map +1 -1
  345. package/dist/services/stall/stall-poller.d.ts +16 -0
  346. package/dist/services/stall/stall-poller.d.ts.map +1 -1
  347. package/dist/services/stall/stall-poller.js +136 -18
  348. package/dist/services/stall/stall-poller.js.map +1 -1
  349. package/dist/services/strategy.d.ts.map +1 -1
  350. package/dist/services/strategy.js +55 -5
  351. package/dist/services/strategy.js.map +1 -1
  352. package/dist/services/substitution-engine.d.ts +20 -0
  353. package/dist/services/substitution-engine.d.ts.map +1 -0
  354. package/dist/services/substitution-engine.js +92 -0
  355. package/dist/services/substitution-engine.js.map +1 -0
  356. package/dist/services/task-cleanup.d.ts.map +1 -1
  357. package/dist/services/task-cleanup.js +1 -9
  358. package/dist/services/task-cleanup.js.map +1 -1
  359. package/dist/services/token-issuer.d.ts +47 -0
  360. package/dist/services/token-issuer.d.ts.map +1 -0
  361. package/dist/services/token-issuer.js +63 -0
  362. package/dist/services/token-issuer.js.map +1 -0
  363. package/dist/services/tool-registry.d.ts +3 -0
  364. package/dist/services/tool-registry.d.ts.map +1 -0
  365. package/dist/services/tool-registry.js +139 -0
  366. package/dist/services/tool-registry.js.map +1 -0
  367. package/dist/services/user-config.d.ts +14 -0
  368. package/dist/services/user-config.d.ts.map +1 -1
  369. package/dist/services/user-config.js +55 -0
  370. package/dist/services/user-config.js.map +1 -1
  371. package/dist/services/vcs/types.d.ts +44 -0
  372. package/dist/services/vcs/types.d.ts.map +1 -1
  373. package/dist/skills/fleet-sprint-cli/SKILL.md +103 -0
  374. package/dist/skills/pm/cost.md +3 -0
  375. package/dist/skills/pm/doer-reviewer-loop.md +3 -0
  376. package/dist/tools/check-status.d.ts.map +1 -1
  377. package/dist/tools/check-status.js +8 -2
  378. package/dist/tools/check-status.js.map +1 -1
  379. package/dist/tools/child-id-allocator.d.ts +45 -0
  380. package/dist/tools/child-id-allocator.d.ts.map +1 -0
  381. package/dist/tools/child-id-allocator.js +62 -0
  382. package/dist/tools/child-id-allocator.js.map +1 -0
  383. package/dist/tools/cloud-control.d.ts +2 -2
  384. package/dist/tools/compose-permissions.d.ts +12 -4
  385. package/dist/tools/compose-permissions.d.ts.map +1 -1
  386. package/dist/tools/compose-permissions.js +55 -4
  387. package/dist/tools/compose-permissions.js.map +1 -1
  388. package/dist/tools/dolt-push-mutex.d.ts +44 -0
  389. package/dist/tools/dolt-push-mutex.d.ts.map +1 -0
  390. package/dist/tools/dolt-push-mutex.js +83 -0
  391. package/dist/tools/dolt-push-mutex.js.map +1 -0
  392. package/dist/tools/execute-command.d.ts +11 -1
  393. package/dist/tools/execute-command.d.ts.map +1 -1
  394. package/dist/tools/execute-command.js +22 -5
  395. package/dist/tools/execute-command.js.map +1 -1
  396. package/dist/tools/execute-prompt.d.ts +72 -12
  397. package/dist/tools/execute-prompt.d.ts.map +1 -1
  398. package/dist/tools/execute-prompt.js +962 -71
  399. package/dist/tools/execute-prompt.js.map +1 -1
  400. package/dist/tools/get-member-model-pricing.d.ts +21 -0
  401. package/dist/tools/get-member-model-pricing.d.ts.map +1 -0
  402. package/dist/tools/get-member-model-pricing.js +29 -0
  403. package/dist/tools/get-member-model-pricing.js.map +1 -0
  404. package/dist/tools/list-members.d.ts.map +1 -1
  405. package/dist/tools/list-members.js +39 -7
  406. package/dist/tools/list-members.js.map +1 -1
  407. package/dist/tools/member-detail.d.ts +1 -1
  408. package/dist/tools/member-detail.d.ts.map +1 -1
  409. package/dist/tools/member-detail.js +8 -2
  410. package/dist/tools/member-detail.js.map +1 -1
  411. package/dist/tools/member-reservation.d.ts +29 -0
  412. package/dist/tools/member-reservation.d.ts.map +1 -0
  413. package/dist/tools/member-reservation.js +71 -0
  414. package/dist/tools/member-reservation.js.map +1 -0
  415. package/dist/tools/provision-vcs-auth.d.ts +7 -7
  416. package/dist/tools/provision-vcs-auth.js +1 -1
  417. package/dist/tools/provision-vcs-auth.js.map +1 -1
  418. package/dist/tools/register-member.d.ts +17 -5
  419. package/dist/tools/register-member.d.ts.map +1 -1
  420. package/dist/tools/register-member.js +144 -5
  421. package/dist/tools/register-member.js.map +1 -1
  422. package/dist/tools/remove-member.d.ts +1 -1
  423. package/dist/tools/report-status.d.ts +33 -0
  424. package/dist/tools/report-status.d.ts.map +1 -0
  425. package/dist/tools/report-status.js +44 -0
  426. package/dist/tools/report-status.js.map +1 -0
  427. package/dist/tools/respond-to-message.d.ts +21 -0
  428. package/dist/tools/respond-to-message.d.ts.map +1 -0
  429. package/dist/tools/respond-to-message.js +23 -0
  430. package/dist/tools/respond-to-message.js.map +1 -0
  431. package/dist/tools/send-files.d.ts +3 -0
  432. package/dist/tools/send-files.d.ts.map +1 -1
  433. package/dist/tools/send-files.js +88 -8
  434. package/dist/tools/send-files.js.map +1 -1
  435. package/dist/tools/send-message.d.ts +27 -0
  436. package/dist/tools/send-message.d.ts.map +1 -0
  437. package/dist/tools/send-message.js +46 -0
  438. package/dist/tools/send-message.js.map +1 -0
  439. package/dist/tools/shutdown-server.d.ts +2 -0
  440. package/dist/tools/shutdown-server.d.ts.map +1 -1
  441. package/dist/tools/shutdown-server.js +19 -0
  442. package/dist/tools/shutdown-server.js.map +1 -1
  443. package/dist/tools/update-member.d.ts +8 -8
  444. package/dist/tools/update-member.d.ts.map +1 -1
  445. package/dist/tools/update-member.js +11 -1
  446. package/dist/tools/update-member.js.map +1 -1
  447. package/dist/types.d.ts +15 -3
  448. package/dist/types.d.ts.map +1 -1
  449. package/dist/utils/agent-helpers.d.ts +2 -1
  450. package/dist/utils/agent-helpers.d.ts.map +1 -1
  451. package/dist/utils/agent-helpers.js +7 -2
  452. package/dist/utils/agent-helpers.js.map +1 -1
  453. package/dist/utils/pid-helpers.d.ts +15 -0
  454. package/dist/utils/pid-helpers.d.ts.map +1 -1
  455. package/dist/utils/pid-helpers.js +23 -0
  456. package/dist/utils/pid-helpers.js.map +1 -1
  457. package/dist/utils/process-utils.d.ts +3 -0
  458. package/dist/utils/process-utils.d.ts.map +1 -0
  459. package/dist/utils/process-utils.js +42 -0
  460. package/dist/utils/process-utils.js.map +1 -0
  461. package/dist/utils/prompt-errors.d.ts +2 -1
  462. package/dist/utils/prompt-errors.d.ts.map +1 -1
  463. package/dist/utils/prompt-errors.js +13 -3
  464. package/dist/utils/prompt-errors.js.map +1 -1
  465. package/dist/utils/workspace-trust.d.ts +24 -0
  466. package/dist/utils/workspace-trust.d.ts.map +1 -0
  467. package/dist/utils/workspace-trust.js +35 -0
  468. package/dist/utils/workspace-trust.js.map +1 -0
  469. package/dist/workflows/auto-sprint.js +450 -87
  470. package/package.json +35 -7
  471. package/packages/apra-fleet-client/package.json +26 -0
  472. package/packages/apra-fleet-client/src/client/api.mjs +396 -0
  473. package/packages/apra-fleet-client/src/client/client.mjs +146 -0
  474. package/packages/apra-fleet-client/src/client/errors.mjs +71 -0
  475. package/packages/apra-fleet-client/src/client/factory.mjs +65 -0
  476. package/packages/apra-fleet-client/src/client/server-resolution.mjs +265 -0
  477. package/packages/apra-fleet-client/src/client/transport.mjs +327 -0
  478. package/packages/apra-fleet-se/apra-pm/agents/backlog-groomer.md +287 -0
  479. package/packages/apra-fleet-se/apra-pm/agents/ci-watcher.md +89 -0
  480. package/packages/apra-fleet-se/apra-pm/agents/deployer.md +139 -0
  481. package/packages/apra-fleet-se/apra-pm/agents/doer.md +225 -0
  482. package/packages/apra-fleet-se/apra-pm/agents/harvester.md +169 -0
  483. package/packages/apra-fleet-se/apra-pm/agents/integ-test-runner.md +316 -0
  484. package/packages/apra-fleet-se/apra-pm/agents/plan-reviewer.md +188 -0
  485. package/packages/apra-fleet-se/apra-pm/agents/planner.md +239 -0
  486. package/packages/apra-fleet-se/apra-pm/agents/regression-test-runner.md +216 -0
  487. package/packages/apra-fleet-se/apra-pm/agents/reviewer.md +174 -0
  488. package/packages/apra-fleet-se/apra-pm/agents/schemas/backlog-groomer-input.json +35 -0
  489. package/packages/apra-fleet-se/apra-pm/agents/schemas/backlog-groomer-output.json +164 -0
  490. package/packages/apra-fleet-se/apra-pm/agents/schemas/ci-watcher-input.json +26 -0
  491. package/packages/apra-fleet-se/apra-pm/agents/schemas/ci-watcher-output.json +18 -0
  492. package/packages/apra-fleet-se/apra-pm/agents/schemas/deployer-input.json +19 -0
  493. package/packages/apra-fleet-se/apra-pm/agents/schemas/deployer-output.json +17 -0
  494. package/packages/apra-fleet-se/apra-pm/agents/schemas/doer-input.json +25 -0
  495. package/packages/apra-fleet-se/apra-pm/agents/schemas/doer-output.json +22 -0
  496. package/packages/apra-fleet-se/apra-pm/agents/schemas/harvester-input.json +31 -0
  497. package/packages/apra-fleet-se/apra-pm/agents/schemas/harvester-output.json +18 -0
  498. package/packages/apra-fleet-se/apra-pm/agents/schemas/integ-test-runner-input.json +20 -0
  499. package/packages/apra-fleet-se/apra-pm/agents/schemas/integ-test-runner-output.json +61 -0
  500. package/packages/apra-fleet-se/apra-pm/agents/schemas/plan-reviewer-input.json +15 -0
  501. package/packages/apra-fleet-se/apra-pm/agents/schemas/plan-reviewer-output.json +30 -0
  502. package/packages/apra-fleet-se/apra-pm/agents/schemas/regression-test-runner-input.json +15 -0
  503. package/packages/apra-fleet-se/apra-pm/agents/schemas/regression-test-runner-output.json +50 -0
  504. package/packages/apra-fleet-se/apra-pm/agents/schemas/reviewer-input.json +25 -0
  505. package/packages/apra-fleet-se/apra-pm/agents/schemas/reviewer-output.json +38 -0
  506. package/packages/apra-fleet-se/apra-pm/skills/pm/SKILL.md +402 -0
  507. package/packages/apra-fleet-se/apra-pm/skills/pm/beads.md +142 -0
  508. package/packages/apra-fleet-se/apra-pm/skills/pm/cost.md +157 -0
  509. package/packages/apra-fleet-se/apra-pm/skills/pm/doer-reviewer-loop.md +300 -0
  510. package/packages/apra-fleet-se/apra-pm/skills/pm/fleet-addendum.md +121 -0
  511. package/packages/apra-fleet-se/apra-pm/skills/pm/simple-sprint.md +65 -0
  512. package/packages/apra-fleet-se/apra-pm/skills/pm/sprint.md +327 -0
  513. package/packages/apra-fleet-se/apra-pm/skills/pm/worktrees.md +92 -0
  514. package/packages/apra-fleet-se/apra-pm/version.json +3 -0
  515. package/packages/apra-fleet-se/bin/cli.mjs +857 -0
  516. package/packages/apra-fleet-se/bin/serve.mjs +383 -0
  517. package/packages/apra-fleet-se/fleet-sprint/conflict-ladder.mjs +178 -0
  518. package/packages/apra-fleet-se/fleet-sprint/contracts.mjs +914 -0
  519. package/packages/apra-fleet-se/fleet-sprint/dispatch-safety-guard.mjs +177 -0
  520. package/packages/apra-fleet-se/fleet-sprint/docs/README.md +191 -0
  521. package/packages/apra-fleet-se/fleet-sprint/docs/coach-role-design.md +151 -0
  522. package/packages/apra-fleet-se/fleet-sprint/docs/dolt-tier2-runbook.md +127 -0
  523. package/packages/apra-fleet-se/fleet-sprint/docs/fleet-sprint-diagram.md +94 -0
  524. package/packages/apra-fleet-se/fleet-sprint/docs/stabilization-log.md +1125 -0
  525. package/packages/apra-fleet-se/fleet-sprint/dolt-literal-guard.mjs +72 -0
  526. package/packages/apra-fleet-se/fleet-sprint/dolt-recovery-path-b.mjs +339 -0
  527. package/packages/apra-fleet-se/fleet-sprint/dolt-recovery-tier2.mjs +360 -0
  528. package/packages/apra-fleet-se/fleet-sprint/dolt-recovery.mjs +363 -0
  529. package/packages/apra-fleet-se/fleet-sprint/dolt-sync.mjs +1244 -0
  530. package/packages/apra-fleet-se/fleet-sprint/errors.mjs +572 -0
  531. package/packages/apra-fleet-se/fleet-sprint/full-db-fetch-guard.mjs +118 -0
  532. package/packages/apra-fleet-se/fleet-sprint/runner.js +9447 -0
  533. package/packages/apra-fleet-se/fleet-sprint/skills/fleet-sprint-cli/SKILL.md +103 -0
  534. package/packages/apra-fleet-se/fleet-sprint/sprint-lock.mjs +140 -0
  535. package/packages/apra-fleet-se/fleet-sprint/vcs-module.mjs +435 -0
  536. package/packages/apra-fleet-se/fleet-sprint/vcs-providers/azure-devops.mjs +68 -0
  537. package/packages/apra-fleet-se/fleet-sprint/vcs-providers/bitbucket.mjs +53 -0
  538. package/packages/apra-fleet-se/fleet-sprint/vcs-providers/dolt.mjs +146 -0
  539. package/packages/apra-fleet-se/fleet-sprint/vcs-providers/generic-git.mjs +176 -0
  540. package/packages/apra-fleet-se/fleet-sprint/vcs-providers/github.mjs +222 -0
  541. package/packages/apra-fleet-se/fleet-sprint/vcs-providers/index.mjs +218 -0
  542. package/packages/apra-fleet-se/fleet-sprint/viewer-extensions.mjs +882 -0
  543. package/packages/apra-fleet-se/package.json +28 -0
  544. package/packages/apra-fleet-se/workflow.json +1 -0
  545. package/packages/apra-fleet-workflow/package.json +30 -0
  546. package/packages/apra-fleet-workflow/src/viewer/command-output-cap.mjs +112 -0
  547. package/packages/apra-fleet-workflow/src/viewer/debounced-writer.mjs +153 -0
  548. package/packages/apra-fleet-workflow/src/viewer/html-utils.mjs +31 -0
  549. package/packages/apra-fleet-workflow/src/viewer/index.mjs +1325 -0
  550. package/packages/apra-fleet-workflow/src/viewer/lean-state.mjs +237 -0
  551. package/packages/apra-fleet-workflow/src/viewer/run-state-paths.mjs +130 -0
  552. package/packages/apra-fleet-workflow/src/viewer/sprint-state-paths.mjs +29 -0
  553. package/packages/apra-fleet-workflow/src/workflow/engine.mjs +200 -0
  554. package/packages/apra-fleet-workflow/src/workflow/errors.mjs +146 -0
  555. package/packages/apra-fleet-workflow/src/workflow/index.mjs +1723 -0
  556. package/packages/apra-fleet-workflow/src/workflow/journal.mjs +291 -0
  557. package/packages/apra-fleet-workflow/src/workflow/pricing.mjs +99 -0
  558. package/packages/apra-fleet-workflow/src/workflow/vetting.mjs +111 -0
  559. package/scripts/agy-transcript-reader.js +1 -0
  560. package/scripts/fleet-statusline.sh +17 -13
  561. package/skills/fleet/SKILL.md +37 -2
  562. package/skills/fleet/beads.md +14 -7
  563. package/skills/fleet/permissions.md +50 -0
  564. package/skills/fleet/profiles/base-dev.json +1 -0
  565. package/skills/fleet/profiles/base-reviewer.json +1 -0
  566. package/skills/fleet/profiles/node.json +2 -2
  567. package/version.json +1 -1
  568. package/scripts/agy-settings-merge.js +0 -27
  569. /package/{dist → packages/apra-fleet-se/apra-pm}/agents/_shared/GRAPH-SEMANTICS.md +0 -0
@@ -6,41 +6,119 @@ import { z } from 'zod';
6
6
  import { getStrategy } from '../services/strategy.js';
7
7
  import { getOsCommands } from '../os/index.js';
8
8
  import { getProvider } from '../providers/index.js';
9
- import { getAgentOS, touchAgent } from '../utils/agent-helpers.js';
9
+ import { getAgentOS, touchAgent, getStoredPid } from '../utils/agent-helpers.js';
10
10
  import { updateAgent } from '../services/registry.js';
11
11
  import { memberIdentifier, resolveMember } from '../utils/resolve-member.js';
12
- import { isRetryable, authErrorAdvice } from '../utils/prompt-errors.js';
12
+ import { isRetryable, authErrorAdvice, workspaceNotTrustedAdvice } from '../utils/prompt-errors.js';
13
13
  import { buildAuthEnvPrefix } from '../utils/auth-env.js';
14
14
  import { writeStatusline } from '../services/statusline.js';
15
15
  import { getModelOverride } from '../services/user-config.js';
16
16
  import { ensureCloudReady } from '../services/cloud/lifecycle.js';
17
17
  import { getStallDetector, resolveSessionLogPath } from '../services/stall/index.js';
18
- import { provisionAgents, remoteAgentsDir } from '../services/agent-provisioner.js';
18
+ import { provisionAgents, remoteAgentsDir, loadCanonicalAgentSet } from '../services/agent-provisioner.js';
19
19
  import { escapeWindowsArg, escapeDoubleQuoted } from '../os/os-commands.js';
20
20
  import { resolveTilde } from './execute-command.js';
21
21
  import { clearStoredPid } from '../utils/agent-helpers.js';
22
- import { tryKillPid } from '../utils/pid-helpers.js';
23
- import { LogScope, maskSecrets, truncateForLog } from '../utils/log-helpers.js';
22
+ import { tryKillPid, isPidAlive } from '../utils/pid-helpers.js';
23
+ import { recoverOrphanedDispatch, isRemoteProcessAlive } from '../services/orphan-recovery.js';
24
+ import { seedWorkspaceTrust } from '../utils/workspace-trust.js';
25
+ import { durableOutputPath } from '../os/linux.js';
26
+ import { LogScope, maskSecrets, truncateForLog, logWarn } from '../utils/log-helpers.js';
24
27
  import { getLogPreviewChars } from '../services/user-config.js';
28
+ import { validateSubstitutionKeys, applySubstitutions } from '../services/substitution-engine.js';
29
+ import { sessionRegistry } from '../services/session-registry.js';
30
+ import { getTokenIssuer } from '../services/token-issuer.js';
31
+ import { resolveExpectedDemand, checkContextAdmission, recordSessionUsage } from '../services/context-admission.js';
32
+ import { recordKnownSession, isKnownSession } from '../services/known-sessions.js';
33
+ import { resolveBudgetScope, evaluateBudget, recordAndEvaluate } from '../services/budget-awareness.js';
34
+ import { sendMessage } from './send-message.js';
35
+ import { registerPending } from '../services/pending-responses.js';
36
+ import { isMaxTurnsResponse } from '../providers/provider.js';
25
37
  export const executePromptSchema = z.object({
26
38
  ...memberIdentifier,
27
39
  prompt: z.string().describe('The prompt to send to the LLM on the remote member'),
28
- resume: z.boolean().default(true).describe('Resume the previous session if one exists (default: true)'),
29
- timeout_s: z.number().default(300).describe('Inactivity timeout in seconds — the command is killed after this many seconds without any stdout/stderr output (default: 300s / 5 minutes)'),
30
- max_total_s: z.number().optional().describe('Hard ceiling in seconds — the command is killed after this total elapsed time regardless of activity. If omitted, there is no total time limit.'),
40
+ resume: z.union([z.boolean(), z.string()]).default(true).describe('Session-resume control (default: true). ' +
41
+ 'true = best-effort resume of the member\'s stored last session (a stale/unknown stored session transparently falls back to a fresh session). ' +
42
+ 'false = always start a fresh session. ' +
43
+ 'A session-id STRING = EXPLICIT resume of exactly that session, preferred over the member\'s stored session -- the caller asserts this prompt depends on that session\'s prior context, so an unknown/expired id is a TERMINAL error ' +
44
+ '(structured {isError, reason: "session_not_found"}, NO LLM call, and NO fresh-session fallback) rather than a silent wrong-context dispatch.'),
45
+ timeout_s: z.number().default(300).describe('Inactivity timeout in seconds -- the command is killed after this many seconds without any stdout/stderr output (default: 300s / 5 minutes)'),
46
+ max_total_s: z.number().optional().describe('Hard ceiling in seconds -- the command is killed after this total elapsed time regardless of activity. If omitted, there is no total time limit.'),
31
47
  max_turns: z.number().min(1).max(500).optional().describe('Max turns for claude -p (default: 50)'),
32
- dangerously_skip_permissions: z.boolean().default(false).describe('DEPRECATED: use update_member(unattended="dangerous") instead. This field is ignored and will be removed in a future version.'),
33
- model: z.string().optional().describe('Model tier ("cheap", "standard", "premium") or a specific model ID for power users. Prefer tier names — the server resolves them to the correct model per provider. If omitted, defaults to the standard tier. Applies to both new and resumed sessions.'),
34
- });
35
- function buildFailureMessage(agentName, result, provider) {
48
+ model: z.string().optional().describe('Model tier ("cheap", "standard", "premium") or a specific model ID for power users. Prefer tier names -- the server resolves them to the correct model per provider. If omitted, defaults to the standard tier. Applies to both new and resumed sessions.'),
49
+ substitutions: z.record(z.string(), z.string()).optional().describe('Optional map of token name to replacement value. ' +
50
+ 'When provided, every occurrence of {{name}} in the prompt is replaced before the prompt is staged on the member. ' +
51
+ 'Keys must match [A-Za-z_][A-Za-z0-9_]*. Missing tokens cause the call to fail with no CLI invoked. ' +
52
+ 'Extra keys are silently ignored. Values are never logged.'),
53
+ agent: z.string().optional().describe('Optional agent name to activate. ' +
54
+ 'For Claude: invokes claude --agent <name>. ' +
55
+ 'For Gemini: prepends @<name> to the prompt on every dispatch. ' +
56
+ 'For AGY: prepends @<name> to the prompt on every dispatch (same as Gemini). ' +
57
+ 'Substitution runs before the @<name> prepend. ' +
58
+ 'Agent file must exist at the provider-specific path on the member: ' +
59
+ 'Claude: <workFolder>/.claude/agents/<name>.md or ~/.claude/agents/<name>.md; ' +
60
+ 'Gemini: <workFolder>/.gemini/agents/<name>.md or ~/.gemini/agents/<name>.md; ' +
61
+ 'AGY: <workFolder>/.gemini/antigravity-cli/agents/<name>.md or ~/.gemini/antigravity-cli/agents/<name>.md -- ' +
62
+ 'the call is rejected with a clear error if neither is present.'),
63
+ sprint_id: z.string().optional().describe('Opaque identity of the sprint issuing this dispatch (apra-fleet-eft.29.1). ' +
64
+ 'When provided, the server-side member-reservation check (see reservedBy below) ' +
65
+ 'compares this value directly against the reservation instead of falling back to ' +
66
+ 'this server process\'s APRA_FLEET_SPRINT_ID env var -- the same per-call value ' +
67
+ 'the caller passed to member_reservation reserve/release. Callers that never set a ' +
68
+ 'reservation, or that reserved and dispatch in the same process, should omit this; ' +
69
+ 'omitting it preserves the pre-existing env-var-based behavior exactly.'),
70
+ expected_context_tokens: z.number().positive().optional().describe('Optional estimate (apra-fleet-eft.81.1) of how many tokens this dispatch will add ' +
71
+ 'to the target session\'s context. When set (or context_size is set), the server ' +
72
+ 'compares it against the session\'s remaining context-window headroom BEFORE ' +
73
+ 'invoking the LLM: too little headroom rejects the call with ' +
74
+ '{reason: "insufficient_context_headroom", detail: {demand, headroom, window}} and no ' +
75
+ 'spawn; a fit that lands inside the safety margin still proceeds but attaches a ' +
76
+ 'structured contextWarning. Wins over context_size when both are set. Omitting both ' +
77
+ 'fields disables the check entirely -- pre-existing behavior is unchanged.'),
78
+ context_size: z.enum(['S', 'M', 'L']).optional().describe('Optional size-bucket shorthand for expected_context_tokens (apra-fleet-eft.81.1): ' +
79
+ 'S/M/L map to configured token estimates (fleet defaults, overridable via ' +
80
+ 'config.json\'s contextAdmission.sizeBucketTokens). Ignored when expected_context_tokens ' +
81
+ 'is also set. The auto-sprint engine supplies this from the dispatched task\'s size ' +
82
+ 'metadata.'),
83
+ }).strict();
84
+ function buildFailureMessage(agentName, result, provider, parsed) {
36
85
  const output = result.stderr || result.stdout;
37
- const category = provider.classifyError(output);
86
+ // apra-fleet-p4f.2: prefer the already-parsed structured signal over the
87
+ // stderr/stdout regex scan -- a max_turns-exhausted transcript can still
88
+ // have auth-like noise in stderr (a stale warning, an unrelated retry
89
+ // message, etc.) that would otherwise misclassify it as an auth failure.
90
+ const category = isMaxTurnsResponse(parsed) ? 'max_turns' : provider.classifyError(output);
91
+ if (category === 'max_turns') {
92
+ return `[FAIL] Prompt on "${agentName}" was stopped after exhausting its turn limit (max_turns), not a genuine failure -- the model ran out of turns before finishing:
93
+ ${output}`;
94
+ }
38
95
  return category === 'auth'
39
96
  ? authErrorAdvice(agentName)
40
- : `❌ Prompt failed on "${agentName}":
97
+ : `[FAIL] Prompt failed on "${agentName}":
41
98
  ${output}`;
42
99
  }
43
100
  const SERVER_RETRY_DELAY_MS = 5000;
101
+ // A prompt written whole into a single remote exec command line can exceed
102
+ // the SSH exec channel's / Windows CreateProcess's command-line ceiling once
103
+ // the Windows path's UTF-16LE + base64 -EncodedCommand encoding (~2.67x
104
+ // inflation) is applied -- observed live (2026-07-30, fleet-win-dev1): a
105
+ // Review dispatch embedding full `bd show --json` output (description +
106
+ // acceptance criteria) for 5 beads produced a garbled "Unable to exec ..."
107
+ // response instead of real review output, because the encoded command line
108
+ // was too long for the remote shell to exec at all. Doer dispatches (small,
109
+ // just branch + bead ids) never hit this; only the larger Review-phase
110
+ // prompts do -- and only on Windows, where the encoding overhead is worst.
111
+ // Chunking keeps every single exec command line bounded regardless of total
112
+ // prompt size, on both OS paths (POSIX has a much higher real-world ceiling
113
+ // but shares the same underlying single-exec-command-line hazard).
114
+ const REMOTE_PROMPT_CHUNK_CHARS = 4000;
115
+ function chunkContent(content) {
116
+ const chunks = [];
117
+ for (let i = 0; i < content.length; i += REMOTE_PROMPT_CHUNK_CHARS) {
118
+ chunks.push(content.slice(i, i + REMOTE_PROMPT_CHUNK_CHARS));
119
+ }
120
+ return chunks.length > 0 ? chunks : [''];
121
+ }
44
122
  async function writePromptFile(agent, strategy, promptFilePath, content) {
45
123
  if (agent.agentType === 'local') {
46
124
  fs.writeFileSync(promptFilePath, content, 'utf-8');
@@ -49,24 +127,41 @@ async function writePromptFile(agent, strategy, promptFilePath, content) {
49
127
  const agentOs = getAgentOS(agent);
50
128
  const promptFileName = path.basename(promptFilePath);
51
129
  const remoteDir = path.dirname(promptFilePath);
130
+ const chunks = chunkContent(content);
52
131
  if (agentOs === 'windows') {
53
132
  const escapedFolder = escapeWindowsArg(remoteDir);
54
- const psScript = `New-Item -Path '${escapedFolder}' -ItemType Directory -Force | Out-Null; Set-Location "${escapedFolder}"; Set-Content -Path "${promptFileName}" -Value '${content.replace(/'/g, "''")}' -NoNewline -Encoding UTF8`;
55
- const encoded = Buffer.from(psScript, 'utf16le').toString('base64');
56
- await strategy.execCommand(`powershell -EncodedCommand ${encoded}`);
133
+ for (let i = 0; i < chunks.length; i++) {
134
+ const setup = i === 0 ? `New-Item -Path '${escapedFolder}' -ItemType Directory -Force | Out-Null; ` : '';
135
+ const cmdlet = i === 0 ? 'Set-Content' : 'Add-Content';
136
+ const psScript = `${setup}Set-Location "${escapedFolder}"; ${cmdlet} -Path "${promptFileName}" -Value '${chunks[i].replace(/'/g, "''")}' -NoNewline -Encoding UTF8`;
137
+ const encoded = Buffer.from(psScript, 'utf16le').toString('base64');
138
+ // eslint-disable-next-line no-await-in-loop -- each chunk must land before the next appends
139
+ await strategy.execCommand(`powershell -EncodedCommand ${encoded}`);
140
+ }
57
141
  }
58
142
  else {
59
- const b64 = Buffer.from(content).toString('base64');
60
143
  const escapedFolder = escapeDoubleQuoted(remoteDir);
61
- await strategy.execCommand(`mkdir -p "${escapedFolder}" && cd "${escapedFolder}" && echo '${b64}' | base64 -d > ${promptFileName}`);
144
+ for (let i = 0; i < chunks.length; i++) {
145
+ const b64 = Buffer.from(chunks[i]).toString('base64');
146
+ const redirect = i === 0 ? '>' : '>>';
147
+ const mkdirPrefix = i === 0 ? `mkdir -p "${escapedFolder}" && ` : '';
148
+ // eslint-disable-next-line no-await-in-loop -- each chunk must land before the next appends
149
+ await strategy.execCommand(`${mkdirPrefix}cd "${escapedFolder}" && echo '${b64}' | base64 -d ${redirect} ${promptFileName}`);
150
+ }
62
151
  }
63
152
  }
64
- async function deletePromptFile(agent, strategy, promptFilePath) {
153
+ async function deletePromptFile(agent, strategy, promptFilePath, extraPaths = []) {
65
154
  if (agent.agentType === 'local') {
66
155
  try {
67
156
  fs.unlinkSync(promptFilePath);
68
157
  }
69
158
  catch { /* ignore */ }
159
+ for (const p of extraPaths) {
160
+ try {
161
+ fs.unlinkSync(p);
162
+ }
163
+ catch { /* ignore */ }
164
+ }
70
165
  return;
71
166
  }
72
167
  const agentOs = getAgentOS(agent);
@@ -80,7 +175,10 @@ async function deletePromptFile(agent, strategy, promptFilePath) {
80
175
  }
81
176
  else {
82
177
  const escapedFolder = escapeDoubleQuoted(remoteDir);
83
- await strategy.execCommand(`cd "${escapedFolder}" && rm -f ${promptFileName}`).catch(() => { });
178
+ // apra-fleet-6z8.1: the durable stdout mirror is cleaned up in the SAME
179
+ // round trip as the prompt file -- no extra exec per dispatch.
180
+ const extras = extraPaths.map(p => ` "${escapeDoubleQuoted(p)}"`).join('');
181
+ await strategy.execCommand(`cd "${escapedFolder}" && rm -f ${promptFileName}${extras}`).catch(() => { });
84
182
  }
85
183
  }
86
184
  export function resolveModelForTier(agent, tier, provider) {
@@ -92,6 +190,20 @@ export function resolveModelForTier(agent, tier, provider) {
92
190
  return provider.modelForTier(tier);
93
191
  }
94
192
  const SECURE_TOKEN_RE = /\{\{secure\.[a-zA-Z0-9_-]{1,64}\}\}/;
193
+ /**
194
+ * The sprint id this server process dispatches on behalf of, or undefined when
195
+ * run outside a sprint (e.g. a manual cli.mjs invocation). Sourced from
196
+ * APRA_FLEET_SPRINT_ID, which the auto-sprint spawner stamps into the per-sprint
197
+ * server's environment (apra-fleet-eft.10.3). Read on every dispatch so a
198
+ * reservation set/cleared mid-run is observed without a restart.
199
+ */
200
+ export function currentSprintId() {
201
+ const raw = process.env.APRA_FLEET_SPRINT_ID;
202
+ if (typeof raw !== 'string')
203
+ return undefined;
204
+ const trimmed = raw.trim();
205
+ return trimmed.length > 0 ? trimmed : undefined;
206
+ }
95
207
  export const inFlightAgents = new Set();
96
208
  // Member ids whose remote agent files (planner.md, doer.md, _shared/, schemas/, ...)
97
209
  // have already been probed/refreshed this server process uptime -- the #336
@@ -128,17 +240,192 @@ async function ensureAgentFilesProvisioned(agent) {
128
240
  }
129
241
  }
130
242
  // All exit paths from executePrompt clear busy state via the finally block (inFlightAgents.delete + writeStatusline):
131
- // (a) normal success: result.code === 0 → finally sets idle and removes agent from inFlight
132
- // (b) non-zero exit from execCommand: result.code !== 0 → finally sets idle and removes agent from inFlight
133
- // (c) exception in try block (auth, network, crash) → catch records error type; finally sets offline or idle
134
- // (d) AbortSignal/MCP client cancellation → abortHandler kills PID, execCommand resolves, finally clears
135
- // (e) stale session retry → retried without session ID; finally clears on success or failure
136
- // (f) server overload retry → retried after delay; finally clears on success or failure
243
+ // (a) normal success: result.code === 0 -> finally sets idle and removes agent from inFlight
244
+ // (b) non-zero exit from execCommand: result.code !== 0 -> finally sets idle and removes agent from inFlight
245
+ // (c) exception in try block (auth, network, crash) -> catch records error type; finally sets offline or idle
246
+ // (d) AbortSignal/MCP client cancellation -> abortHandler kills PID, execCommand resolves, finally clears
247
+ // (e) stale session retry -> retried without session ID; finally clears on success or failure
248
+ // (f) server overload retry -> retried after delay; finally clears on success or failure
137
249
  // (g) early returns before inFlightAgents.add: busy state never entered
250
+ /**
251
+ * apra-fleet-idb: liveness probe for a busy-locked member's backing process.
252
+ * An inFlightAgents entry can outlive the process it was guarding -- a child
253
+ * reaped without its 'close'/'error' handler ever firing (e.g. a hung SSH
254
+ * channel that never signals exit, so clearStoredPid never runs), or an
255
+ * interactive member's underlying claude process dying/disconnecting after
256
+ * its session was registered -- permanently wedging the member: every future
257
+ * dispatch would keep returning busy even though fleet_status's own
258
+ * independent live-process check (src/tools/check-status.ts's
259
+ * fleetProcessCheck) would report idle. Returns the confirmed-dead pid (used
260
+ * only for the release warning) when the lock should self-heal, or undefined
261
+ * when the member is genuinely busy.
262
+ *
263
+ * Conservative on ambiguity, deliberately: NO captured pid at all (neither an
264
+ * interactive session pid nor a subprocess pid) is treated as still busy, not
265
+ * as evidence of staleness -- a dispatch that has not reached its pid-capture
266
+ * step yet must never be raced by a concurrent "self-heal" attempt. Only a
267
+ * DEFINITIVE dead-pid reading releases the lock.
268
+ */
269
+ async function findDeadLockPid(agent, workspaceId) {
270
+ // Interactive sessions are always local (register_member's interactive
271
+ // bootstrap is gated to isLocal members) -- the same local
272
+ // process.kill(pid, 0) probe the eft.28.1 dead-session guard uses further
273
+ // below applies directly here. Falls back to the durable lastKnownPid
274
+ // anchor exactly like that guard does, so a disconnected-then-reconnected
275
+ // session (pid lost on the live SessionState) is still checkable.
276
+ const session = sessionRegistry.get(workspaceId, agent.id);
277
+ const interactivePid = session?.pid ?? sessionRegistry.lastKnownPid(workspaceId, agent.id);
278
+ if (interactivePid !== undefined && !isPidAlive(interactivePid)) {
279
+ return interactivePid;
280
+ }
281
+ // Subprocess dispatch pid (local strategy or remote-over-SSH/relay). A
282
+ // remote pid lives in a DIFFERENT machine's pid namespace -- process.kill()
283
+ // -based isPidAlive is meaningless there (it would almost always read back
284
+ // ESRCH for a pid that simply does not exist on THIS machine, wrongly
285
+ // declaring a genuinely-alive remote session dead and racing a duplicate
286
+ // dispatch onto it). Probe it the same way orphan-recovery's lease-of-life
287
+ // gate does instead: a fresh, independent SSH round trip, never the
288
+ // (possibly wedged) channel that produced the stale lock.
289
+ const subprocessPid = getStoredPid(agent.id);
290
+ if (subprocessPid !== undefined) {
291
+ const alive = agent.agentType === 'local'
292
+ ? isPidAlive(subprocessPid)
293
+ : await isRemoteProcessAlive(getStrategy(agent), subprocessPid);
294
+ if (!alive)
295
+ return subprocessPid;
296
+ }
297
+ return undefined;
298
+ }
299
+ // apra-fleet-eft.28.1: how often the interactive wait re-checks that the
300
+ // target member's claude process is still alive. This is the dispatch-level
301
+ // liveness/no-progress bound -- the member's process can die AFTER the
302
+ // pre-dispatch liveness check above (e.g. mid-turn, right after send_message
303
+ // lands) with no further signal ever arriving, so the wait for
304
+ // respond_to_message must not be the sole backstop up to the full timeout_s
305
+ // (which can be 3600s). A short, fixed poll interval keeps the surfaced
306
+ // error well under the playbook's <10min budget regardless of how large
307
+ // timeout_s is.
308
+ const INTERACTIVE_LIVENESS_POLL_MS = 5000;
309
+ /** Raised by waitForInteractiveResponse when the member's claude process is
310
+ * confirmed dead while a response is still pending -- distinguished from a
311
+ * plain "nobody answered in time" timeout so the caller can surface a more
312
+ * actionable terminal error. */
313
+ class InteractiveSessionDiedError extends Error {
314
+ }
315
+ /**
316
+ * Waits for the member's respond_to_message reply, racing that wait against
317
+ * a periodic PID-liveness poll of the same session. Resolves/rejects as soon
318
+ * as either side settles -- a dead PID short-circuits the wait instead of
319
+ * letting it run out the full timeoutMs.
320
+ */
321
+ function waitForInteractiveResponse(agent, workspaceId, msgid, timeoutMs, scope) {
322
+ return new Promise((resolve, reject) => {
323
+ let settled = false;
324
+ const poller = setInterval(() => {
325
+ if (settled)
326
+ return;
327
+ const session = sessionRegistry.get(workspaceId, agent.id);
328
+ // apra-fleet-eft.50.1: fall back to the durable launch-pid anchor when the
329
+ // live session lost its pid on a reconnect, so this in-flight poll can
330
+ // still detect a dead persistent process on a retry that reused the
331
+ // reconnected session -- not only when the session still carries its pid.
332
+ const pid = session?.pid ?? sessionRegistry.lastKnownPid(workspaceId, agent.id);
333
+ if (pid !== undefined && !isPidAlive(pid)) {
334
+ settled = true;
335
+ clearInterval(poller);
336
+ scope.info(`[interactive] member process pid=${pid} died while awaiting a response -- aborting wait`);
337
+ sessionRegistry.unregister(workspaceId, agent.id);
338
+ reject(new InteractiveSessionDiedError(`member claude process (pid ${pid}) died while waiting for a response`));
339
+ }
340
+ }, INTERACTIVE_LIVENESS_POLL_MS);
341
+ registerPending(msgid, timeoutMs).then((res) => {
342
+ if (settled)
343
+ return;
344
+ settled = true;
345
+ clearInterval(poller);
346
+ resolve(res);
347
+ }, (err) => {
348
+ if (settled)
349
+ return;
350
+ settled = true;
351
+ clearInterval(poller);
352
+ reject(err);
353
+ });
354
+ });
355
+ }
356
+ /**
357
+ * Interactive routing (apra-fleet-2xs.8): pushes the prompt to a connected
358
+ * member's live session via send_message and waits for that member to call
359
+ * respond_to_message with the matching reply_to. No subprocess, no SSH, no
360
+ * prompt file, no stall detector on a log file -- the session is a
361
+ * long-lived interactive process, not something this call spawns or owns.
362
+ */
363
+ async function executePromptInteractive(agent, renderedPrompt, input, workspaceId, heuristicWarningSuffix) {
364
+ const timeoutS = input.timeout_s ?? 300;
365
+ const scope = new LogScope('execute_prompt', `[interactive] timeout=${timeoutS}s ${truncateForLog(maskSecrets(input.prompt))}`, agent);
366
+ const sendResult = await sendMessage({ member_id: agent.id, content: renderedPrompt }, workspaceId);
367
+ const parsed = JSON.parse(sendResult);
368
+ if (parsed.error) {
369
+ scope.abort(`send failed: ${parsed.error}`);
370
+ return `[FAIL] Failed to deliver prompt to "${agent.friendlyName}" (interactive session): ${parsed.error}`;
371
+ }
372
+ try {
373
+ const response = await waitForInteractiveResponse(agent, workspaceId, parsed.msgid, timeoutS * 1000, scope);
374
+ scope.ok('interactive response received');
375
+ let output = `[RESULT] Response from ${agent.friendlyName}:\n\n${response}`;
376
+ if (heuristicWarningSuffix)
377
+ output += heuristicWarningSuffix;
378
+ return output;
379
+ }
380
+ catch (err) {
381
+ if (err instanceof InteractiveSessionDiedError) {
382
+ scope.abort(err.message);
383
+ return `[ERROR] "${agent.friendlyName}"'s interactive claude process died while this dispatch was waiting for a response (${err.message}). The prompt was delivered but nothing will ever answer it -- re-launch the member (re-run register_member) before retrying.`;
384
+ }
385
+ scope.abort(`interactive timeout: ${err.message}`);
386
+ // apra-fleet-eft.74.2: self-heal on interactive-route timeout. A session
387
+ // with no verifiable live pid that just timed out is a phantom channel
388
+ // (the eft.74 wedge): re-routing the NEXT execute_prompt to it would
389
+ // silently re-burn the full timeout_s, forever (observed 5x 900s). Evict
390
+ // it here so the next dispatch finds no interactive session and falls back
391
+ // to the subprocess path. A session that DOES have a verifiably live pid is
392
+ // left registered -- the member is alive, merely slow, so a later dispatch
393
+ // may legitimately reach it interactively again.
394
+ const timedOutSession = sessionRegistry.get(workspaceId, agent.id);
395
+ const timedOutPid = timedOutSession?.pid ?? sessionRegistry.lastKnownPid(workspaceId, agent.id);
396
+ const hasVerifiableLivePid = timedOutPid !== undefined && isPidAlive(timedOutPid);
397
+ if (!hasVerifiableLivePid) {
398
+ sessionRegistry.unregister(workspaceId, agent.id);
399
+ scope.info(`[interactive] timed-out session for "${agent.friendlyName}" has no verifiable live pid (pid=${timedOutPid ?? 'none'}) -- evicting so the next dispatch falls back to subprocess`);
400
+ }
401
+ return `[FAIL] Timed out waiting for "${agent.friendlyName}" to respond (interactive session, ${timeoutS}s). The prompt was delivered; the member may still respond late, but this call has given up waiting.`;
402
+ }
403
+ }
138
404
  export async function executePrompt(input, extra) {
139
405
  if (SECURE_TOKEN_RE.test(input.prompt)) {
140
406
  return 'error: execute_prompt prompt contains {{secure.NAME}} token. Secrets must never be passed to LLM prompts. Use execute_command with {{secure.NAME}} instead.';
141
407
  }
408
+ // Validate substitution keys before any I/O or member resolution.
409
+ if (input.substitutions !== undefined) {
410
+ const keyCheck = validateSubstitutionKeys('execute_prompt', input.substitutions);
411
+ if (!keyCheck.ok)
412
+ return keyCheck.error;
413
+ }
414
+ // Apply substitutions to the prompt string (or emit heuristic warning when omitted).
415
+ let renderedPrompt = input.prompt;
416
+ let heuristicWarningSuffix = '';
417
+ if (input.substitutions !== undefined) {
418
+ const result = applySubstitutions('execute_prompt', [{ label: 'prompt', content: input.prompt }], input.substitutions);
419
+ if (!result.ok)
420
+ return result.error;
421
+ renderedPrompt = result.outputs[0];
422
+ }
423
+ else {
424
+ const warnResult = applySubstitutions('execute_prompt', [{ label: 'prompt', content: input.prompt }], undefined);
425
+ if (warnResult.ok && warnResult.warning) {
426
+ heuristicWarningSuffix = `\n\n[WARN] ${warnResult.warning}`;
427
+ }
428
+ }
142
429
  const promptFileName = `.fleet-task.md`;
143
430
  const agentOrError = resolveMember(input.member_id, input.member_name);
144
431
  if (typeof agentOrError === 'string')
@@ -148,15 +435,181 @@ export async function executePrompt(input, extra) {
148
435
  agent = await ensureCloudReady(agentOrError); // auto-start if stopped
149
436
  }
150
437
  catch (err) {
151
- return `❌ Failed to execute prompt on "${agentOrError.friendlyName}": ${err.message}`;
438
+ return `[FAIL] Failed to execute prompt on "${agentOrError.friendlyName}": ${err.message}`;
439
+ }
440
+ // Server-side member reservation enforcement (apra-fleet-eft.10.3): a member
441
+ // reserved by a DIFFERENT sprint may not be dispatched to. Mirrors the
442
+ // inFlightAgents busy-rejection error path -- the error names the owning
443
+ // sprint. A dispatch from the owning sprint (matching APRA_FLEET_SPRINT_ID)
444
+ // or against an unreserved member proceeds unchanged, so behavior with no
445
+ // reservations is identical to before. Checked before any busy state is
446
+ // entered, closing the manual-CLI bypass the ledger alone could not.
447
+ //
448
+ // apra-fleet-eft.29.1: currentSprintId() alone (APRA_FLEET_SPRINT_ID on
449
+ // THIS server process) is only correct when the launcher spawns a private
450
+ // per-sprint server and stamps its env -- it is not for the eft.7.1
451
+ // CLI/shared-fleet-server path, where cli.mjs attaches to an existing,
452
+ // long-lived HTTP singleton it never spawns and so can never stamp. There,
453
+ // APRA_FLEET_SPRINT_ID is unset (or stale from a different run), so a
454
+ // member reserved by ITS OWN owning sprint was rejected as if from another
455
+ // sprint. Prefer the per-call `sprint_id` -- the exact same opaque token
456
+ // the caller already passed to member_reservation reserve/release (see
457
+ // createMemberReservationClient / sprintMutexId in
458
+ // packages/apra-fleet-se/fleet-sprint/runner.js) -- and fall back to
459
+ // currentSprintId() only when the caller omits it, so existing
460
+ // env-var-based callers/tests are unaffected.
461
+ const owningSprint = agent.reservedBy ?? null;
462
+ const dispatchSprintId = input.sprint_id ?? currentSprintId();
463
+ if (owningSprint && owningSprint !== dispatchSprintId) {
464
+ return {
465
+ text: `[-] Member "${agent.friendlyName}" is reserved by sprint "${owningSprint}" and cannot accept a dispatch from another sprint. Wait for that sprint to release it, or force-release the reservation to recover.`,
466
+ structuredContent: { isError: true, reason: 'reserved' },
467
+ };
152
468
  }
153
469
  if (inFlightAgents.has(agent.id)) {
154
- return `❌ execute_prompt is already running for "${agent.friendlyName}". Wait for the current call to finish before sending another.`;
470
+ // apra-fleet-idb: before honoring the busy rejection, verify the locked
471
+ // session actually still has a live backing process -- see
472
+ // findDeadLockPid's docstring for the full rationale. This is what keeps
473
+ // fleet_status (which decides busy/idle from its own independent live
474
+ // process check) and this dispatch gate from ever disagreeing: a stale
475
+ // lock self-heals here instead of permanently wedging the member.
476
+ const staleLockPid = await findDeadLockPid(agent, getTokenIssuer().workspaceId());
477
+ if (staleLockPid !== undefined) {
478
+ logWarn('busy_lock', `orphaned busy-lock for "${agent.friendlyName}" -- locked pid=${staleLockPid} is confirmed dead; releasing the stale lock and proceeding with this dispatch instead of rejecting it as busy`, agent);
479
+ inFlightAgents.delete(agent.id);
480
+ getStallDetector().remove(agent.id);
481
+ writeStatusline(new Map([[agent.id, 'idle']]));
482
+ }
483
+ else {
484
+ return {
485
+ text: `[FAIL] execute_prompt is already running for "${agent.friendlyName}". Wait for the current call to finish before sending another.`,
486
+ structuredContent: { isError: true, reason: 'busy' },
487
+ };
488
+ }
489
+ }
490
+ // No-LLM members (apra-fleet-us9.14) are plain command executors -- neither
491
+ // execute_prompt mode applies (there is no LLM to prompt in either a
492
+ // subprocess or an interactive session). Rejected here, before any busy
493
+ // state is entered, rather than relying on NoneProvider's methods to throw
494
+ // deeper in either dispatch path.
495
+ if (agent.llmProvider === 'none') {
496
+ return `[FAIL] "${agent.friendlyName}" has no LLM provider (llm_provider: "none") -- it is a plain command executor. Use execute_command instead.`;
497
+ }
498
+ // Interactive routing (apra-fleet-2xs.8/us9.8, docs/cloud-fleet-architecture.md
499
+ // section 6): if this member has a live MCP session connected right now,
500
+ // route via send_message + wait-for-response instead of spawning a
501
+ // subprocess. Decided tier-2-locally against THIS machine's session
502
+ // registry only (never caller/hub-side state, per apra-fleet-2xs.8's own
503
+ // scope note) -- so behavior is unaffected by whether execute_prompt is
504
+ // invoked directly (Phase 1) or relayed through a future hub. Falls
505
+ // through to the unchanged subprocess/SSH path below for every member
506
+ // without a live session (the common case today, and always for members
507
+ // that never opt into an interactive session).
508
+ //
509
+ // Gated by capability, not provider name (apra-fleet-cqa, eft.74 follow-up):
510
+ // mode (b) -- server-push mid-session prompt injection -- was POC-proven on
511
+ // Claude via the provider-branded `notifications/claude/channel` capability
512
+ // (apra-fleet-us9.9's survey, docs/interactive-injection-provider-survey.md),
513
+ // but the routing decision itself must be provider-agnostic: whatever the
514
+ // provider, a session is only an interactive-routing candidate if it
515
+ // actually declared that capability at MCP initialize time (recorded as
516
+ // SessionState.channelCapable below). Gemini/Codex are confirmed [FAIL]
517
+ // today (no equivalent push mechanism) and so never end up channelCapable in
518
+ // practice, but that is a fact about what each provider adapter currently
519
+ // advertises, not a name-based pre-filter here -- any provider that
520
+ // implements the same MCP channel capability is picked up automatically. A
521
+ // member CAN still have a live sessionRegistry entry (registerMcpEndpoint
522
+ // gives it basic MCP tool access, apra-fleet-fnz.1-3) without that meaning
523
+ // it can receive or act on this push -- routing to it anyway would silently
524
+ // spend the full timeout_s waiting for a response that can never arrive.
525
+ const workspaceId = getTokenIssuer().workspaceId();
526
+ const rawSession = sessionRegistry.get(workspaceId, agent.id);
527
+ // apra-fleet-eft.74.1: interactive routing requires the EXPLICIT channel
528
+ // opt-in handshake, not mere JWT registration. A plain subprocess
529
+ // connect-back (a Doer that opened an MCP tool-access session with a member
530
+ // JWT but never declared the `claude/channel` capability) registers a live
531
+ // `server` here, yet can never receive the `notifications/claude/channel`
532
+ // push -- routing to it would enqueue a message nothing reads and burn the
533
+ // full timeout_s on every later dispatch (the eft.74 wedge). Only a
534
+ // channel-capable session is an interactive-routing candidate; anything else
535
+ // (including that Doer's live tool session, which must be left untouched)
536
+ // falls through to the unchanged subprocess path below.
537
+ let interactiveSession = rawSession?.channelCapable ? rawSession : undefined;
538
+ // apra-fleet-eft.50.1: resolve the pid to test FRESH on every dispatch
539
+ // (never cached from a prior attempt) and fall back to the durable
540
+ // launch-pid anchor when this reused session lost its own pid on a
541
+ // reconnect. This is what re-arms the dead-session guard on a retry attempt
542
+ // 2+ exactly as on attempt 1: the specific eft.50 ordering (attempt 1 fails
543
+ // clean, attempt 2 targets a now-dead reconnected session) used to slip
544
+ // through here because the reconnected SessionState had pid=undefined, so the
545
+ // check below was skipped and the caller hung on the dead channel.
546
+ const interactivePid = interactiveSession?.pid
547
+ ?? sessionRegistry.lastKnownPid(workspaceId, agent.id);
548
+ if (interactiveSession?.server && interactivePid !== undefined && !isPidAlive(interactivePid)) {
549
+ // apra-fleet-eft.28.1/eft.28.5: never reuse a persistent interactive
550
+ // session whose underlying member claude process has already died. Before
551
+ // eft.28.1, a dead launch-time process (e.g. it crashed before ever
552
+ // producing a plan) left a `server` entry in sessionRegistry that looked
553
+ // reusable -- send_message would happily enqueue to it, but nothing would
554
+ // ever call respond_to_message, so the caller silently burned the full
555
+ // timeout_s (observed up to 3600s in apra-fleet-eft.28) with zero
556
+ // fleet-server log output and no watchdog coverage.
557
+ //
558
+ // eft.28.5 changes what happens once the death is detected: instead of
559
+ // surfacing a terminal dispatch_failed error that forces a manual
560
+ // register_member, EVICT the dead session and FALL THROUGH to a fresh
561
+ // non-interactive (subprocess) dispatch below -- i.e. re-dispatch fresh
562
+ // instead of blocking on waitForInteractiveResponse. The bug in
563
+ // apra-fleet-eft.28 was precisely that a dead session was reused "rather
564
+ // than detecting its death and spawning a fresh dispatch"; this does the
565
+ // spawning. If the fresh subprocess dispatch itself cannot start it
566
+ // returns its own terminal error, so nothing ever hangs.
567
+ //
568
+ // The liveness check now fires for connect-back interactive sessions too:
569
+ // http-transport carries the launch-time pid forward across re-registration
570
+ // (eft.28.5), so `pid` is no longer undefined for a member that registered
571
+ // via register_member and then connected back -- the exact real-fleet
572
+ // repro that evaded eft.28.1.
573
+ //
574
+ // apra-fleet-eft.50.1: eft.28.5's carry-forward still lost the pid when a
575
+ // reconnect happened AFTER the prior SessionState was already unregistered
576
+ // (priorPid lookup found nothing), so a retry attempt 2+ reused a
577
+ // pid=undefined session and hung. `interactivePid` above now back-stops
578
+ // that with sessionRegistry.lastKnownPid, the durable per-member launch-pid
579
+ // anchor, so this guard re-arms on EVERY dispatch attempt that reuses an
580
+ // interactive session, not just the first. It stays undefined only for
581
+ // sessions that never had a captured PID at all (e.g. tests, or a provider
582
+ // that never went through register_member's local spawn path); those are
583
+ // left to the pre-existing interactive behavior, unchanged.
584
+ const deadScope = new LogScope('execute_prompt', `[interactive] session liveness check pid=${interactivePid}`, agent);
585
+ sessionRegistry.unregister(workspaceId, agent.id);
586
+ deadScope.info(`member claude process (pid ${interactivePid}) for "${agent.friendlyName}" is dead -- evicting the stale interactive session and re-dispatching fresh (non-interactive)`);
587
+ interactiveSession = undefined;
588
+ }
589
+ if (interactiveSession?.server) {
590
+ inFlightAgents.add(agent.id);
591
+ writeStatusline(new Map([[agent.id, 'busy']]));
592
+ try {
593
+ return await executePromptInteractive(agent, renderedPrompt, input, workspaceId, heuristicWarningSuffix);
594
+ }
595
+ finally {
596
+ inFlightAgents.delete(agent.id);
597
+ writeStatusline(new Map([[agent.id, 'idle']]));
598
+ }
155
599
  }
156
600
  inFlightAgents.add(agent.id);
157
601
  await ensureAgentFilesProvisioned(agent);
158
602
  const stallDetector = getStallDetector();
159
603
  let clearedByStall = false;
604
+ // apra-fleet-3c9.1: a CONFIRMED stall must not only kill the remote pid but
605
+ // also cancel the in-flight strategy.execCommand() promise. Before this, the
606
+ // client kept waiting out its full deriveTimeoutMs deadline after the
607
+ // server-side work had already died (the 60.5-min hung dispatch in
608
+ // apra-fleet-3c9). onStall aborts this controller; its signal is merged into
609
+ // the signal handed to every execCommand below (see dispatchSignal), so a
610
+ // confirmed stall settles the pending dispatch immediately and surfaces a
611
+ // typed 'stalled' error instead of hanging.
612
+ const stallAbortController = new AbortController();
160
613
  stallDetector.add(agent.id, {
161
614
  sessionId: null,
162
615
  logFilePath: null,
@@ -174,6 +627,21 @@ export async function executePrompt(input, extra) {
174
627
  // a new execute_prompt that may have already claimed the member.
175
628
  inFlightAgents.delete(agent.id);
176
629
  clearedByStall = true;
630
+ // apra-fleet-6z8.2: a CONFIRMED stall means the remote turn made no
631
+ // progress of any kind for the whole threshold. Clearing bookkeeping
632
+ // alone left that wedged process running indefinitely on the member --
633
+ // burning its LLM session, holding its work folder, and colliding with
634
+ // whatever dispatch takes the member next. Kill the tracked pid too.
635
+ // Best-effort and never awaited: onStall is a fire-and-forget callback
636
+ // from the poll loop, and tryKillPid already swallows its own errors.
637
+ void tryKillPid(agent, strategy, cmds).catch(() => { });
638
+ // apra-fleet-3c9.1: killing the remote pid alone left the pending
639
+ // execCommand promise still awaiting its full deadline. Abort it now so
640
+ // the dispatch settles promptly and returns a typed 'stalled' error.
641
+ try {
642
+ stallAbortController.abort();
643
+ }
644
+ catch { /* best-effort */ }
177
645
  },
178
646
  });
179
647
  const tmpDir = agent.agentType === 'local' ? os.tmpdir() : '/tmp';
@@ -209,14 +677,48 @@ export async function executePrompt(input, extra) {
209
677
  else {
210
678
  resolvedModel = tiers[resolvedModel] ?? resolvedModel;
211
679
  }
212
- const deprecationWarning = input.dangerously_skip_permissions
213
- ? '⚠️ DEPRECATION: dangerously_skip_permissions is deprecated and ignored. Use update_member(unattended="dangerous") instead.\n\n'
214
- : '';
215
680
  const scope = new LogScope('execute_prompt', `[${resolvedModel}] resume=${input.resume} timeout=${input.timeout_s ?? 300}s ${truncateForLog(maskSecrets(input.prompt), getLogPreviewChars())}`, agent);
216
- const resuming = !!(input.resume && agent.sessionId && provider.supportsResume());
681
+ // Resume semantics (apra-fleet-eft.78.1). `resume` is boolean | string:
682
+ // - true -> best-effort resume of the member's stored last session; a
683
+ // stale/unknown stored session transparently retries fresh.
684
+ // - false -> always a fresh session.
685
+ // - string -> EXPLICIT session-id resume: resume exactly this id, preferring
686
+ // it over agent.sessionId. The caller asserts the prompt depends
687
+ // on that session's context, so an unknown/expired id is a
688
+ // TERMINAL session_not_found (handled just below) and NO
689
+ // fresh-session fallback is ever applied (see the retry paths).
690
+ const explicitResumeId = typeof input.resume === 'string' && input.resume.length > 0 ? input.resume : undefined;
691
+ const resumeRequested = input.resume === true || explicitResumeId !== undefined;
692
+ const resumeTargetId = explicitResumeId ?? agent.sessionId;
693
+ // An explicit-id resume must never silently degrade to a fresh session: that
694
+ // is exactly the wrong-context dispatch this feature forbids. resume=true and
695
+ // resume=false keep their pre-existing transparent recovery.
696
+ const allowFreshSessionFallback = explicitResumeId === undefined;
697
+ const resuming = !!(resumeRequested && resumeTargetId && provider.supportsResume());
217
698
  const mintedId = (provider.name === 'claude' || provider.name === 'gemini' || provider.name === 'agy')
218
- ? (resuming ? agent.sessionId : uuid())
219
- : (resuming ? agent.sessionId : undefined);
699
+ ? (resuming ? resumeTargetId : uuid())
700
+ : (resuming ? resumeTargetId : undefined);
701
+ // Terminal session-not-found gate for explicit-id resumes (apra-fleet-eft.78.1).
702
+ // Checked BEFORE any spawn (writePromptFile / the LLM invocation): if the
703
+ // caller named a session the server has never issued for this member -- and it
704
+ // is not the member's currently-stored session -- there is no context to
705
+ // resume. Reject with a structured session_not_found and make NO LLM call,
706
+ // so an orchestrator can rebuild a self-contained prompt and re-dispatch fresh
707
+ // deliberately, rather than getting a silent blank-session response.
708
+ if (explicitResumeId !== undefined) {
709
+ const resumable = provider.supportsResume()
710
+ && (isKnownSession(agent.id, explicitResumeId) || explicitResumeId === agent.sessionId);
711
+ if (!resumable) {
712
+ scope.abort(`explicit resume rejected -- session "${explicitResumeId}" is unknown/expired (no LLM call)`);
713
+ inFlightAgents.delete(agent.id);
714
+ stallDetector.remove(agent.id);
715
+ writeStatusline(new Map([[agent.id, 'idle']]));
716
+ return {
717
+ text: `[FAIL] execute_prompt on "${agent.friendlyName}" rejected -- session "${explicitResumeId}" cannot be resumed (unknown or expired). No LLM call was made. Rebuild the context and re-dispatch with a full, self-contained prompt (resume=false), or resume=true for best-effort recovery.`,
718
+ structuredContent: { isError: true, reason: 'session_not_found', sessionId: explicitResumeId },
719
+ };
720
+ }
721
+ }
220
722
  const promptOpts = {
221
723
  folder: resolvedWorkFolder,
222
724
  promptFile: promptFileName,
@@ -227,15 +729,152 @@ export async function executePrompt(input, extra) {
227
729
  tier: resolvedTier,
228
730
  maxTurns: input.max_turns,
229
731
  inv: scope.getInv(),
732
+ agentName: input.agent,
230
733
  };
231
734
  const claudeCmd = authPrefix + cmds.buildAgentPromptCommand(provider, promptOpts);
735
+ // apra-fleet-6z8.1: the per-invocation durable stdout mirror the unix prompt
736
+ // wrapper tees to (see durableOutputPath / buildAgentPromptCommand). Windows
737
+ // members have no such companion tee, so recovery is skipped for them.
738
+ const durablePath = getAgentOS(agent) === 'windows' ? undefined : durableOutputPath(scope.getInv());
739
+ const dispatchStartedAt = Date.now();
232
740
  const timeoutMs = (input.timeout_s ?? 300) * 1000;
233
741
  const maxTotalMs = input.max_total_s !== undefined ? input.max_total_s * 1000 : undefined;
742
+ // apra-fleet-y8q.1: every retry below (dispatch-exception, stale-session,
743
+ // server-overloaded) re-dispatches with a FRESH session but used to reuse the
744
+ // SAME full timeoutMs/maxTotalMs as the original attempt -- so a single
745
+ // dispatch could burn up to ~2x max_total_s server-side (original attempt +
746
+ // one full-budget retry), well past what the client's deriveTimeoutMs()
747
+ // (packages/apra-fleet-client/src/client/api.mjs) budgets for the whole
748
+ // tools/call (max_total_s*1000 + a fixed grace margin). That let the
749
+ // client's hard timeout fire before the server's own retry-and-report path
750
+ // ever got a chance, hiding a clean typed server error behind a raw client
751
+ // transport timeout. Share ONE deadline budget across the original attempt
752
+ // and any single retry: cap a retry's maxTotalMs (and its inactivity
753
+ // timeoutMs, so it can't independently outlast the shared ceiling) to
754
+ // whatever remains of max_total_s since dispatchStartedAt, and skip the
755
+ // retry entirely once that budget is exhausted -- so total wall-clock time
756
+ // for this call never exceeds max_total_s, which is exactly what the client
757
+ // is prepared to wait for. When max_total_s is absent there is no hard
758
+ // ceiling to share, so retries keep their full timeout_s (unchanged,
759
+ // pre-existing behavior).
760
+ function retryBudget() {
761
+ if (maxTotalMs === undefined)
762
+ return { timeoutMs, maxTotalMs: undefined, exhausted: false };
763
+ const remaining = Math.max(0, maxTotalMs - (Date.now() - dispatchStartedAt));
764
+ return { timeoutMs: Math.min(timeoutMs, remaining), maxTotalMs: remaining, exhausted: remaining <= 0 };
765
+ }
766
+ // Agent file validation -- verify named agent exists before any CLI invocation
767
+ if (input.agent) {
768
+ const provName = provider.name;
769
+ // AGY uses ~/.gemini/antigravity-cli/ as its config root, not ~/.agy/
770
+ const agentRelDir = provName === 'agy' ? '.gemini/antigravity-cli/agents' : `.${provName}/agents`;
771
+ let agentFound = false;
772
+ if (agent.agentType === 'local') {
773
+ const projPath = path.join(resolvedWorkFolder, agentRelDir, `${input.agent}.md`);
774
+ const userPath = path.join(os.homedir(), agentRelDir, `${input.agent}.md`);
775
+ agentFound = fs.existsSync(projPath) || fs.existsSync(userPath);
776
+ if (!agentFound) {
777
+ inFlightAgents.delete(agent.id);
778
+ stallDetector.remove(agent.id);
779
+ writeStatusline(new Map([[agent.id, 'idle']]));
780
+ return `execute_prompt: agent "${input.agent}" not found.\n\nExpected at:\n ${projPath.replace(/\\/g, '/')}\n ${userPath.replace(/\\/g, '/')}`;
781
+ }
782
+ }
783
+ else {
784
+ // Canonical PM role agents (planner/doer/reviewer/plan-reviewer/...) are
785
+ // already guaranteed present in ~/${agentRelDir} by
786
+ // ensureAgentFilesProvisioned() above (ln ~576, which ran provisionAgents()
787
+ // for this exact member earlier in this same call) -- trust that instead
788
+ // of re-probing the remote here. This also SIDESTEPS the bug this check
789
+ // used to have: the old code hand-rolled a POSIX-only
790
+ // `test -f ... || test -f ...` command run via strategy.execCommand(),
791
+ // which throws a PowerShell parser error (not a POSIX shell) on every
792
+ // Windows remote -- a nonzero exit that this check misread as "agent not
793
+ // found" even when the file genuinely existed (apra-fleet P0 bug, fleet-
794
+ // sprint via a Windows remote member always failed plan-review with
795
+ // "agent 'plan-reviewer' not found").
796
+ const canonicalRelPath = `${input.agent}.md`;
797
+ agentFound = remoteAgentsDir(provName) !== null
798
+ && loadCanonicalAgentSet(provName).some((f) => f.relPath === canonicalRelPath);
799
+ if (!agentFound) {
800
+ // Not part of the canonical PM set (e.g. a project-local custom
801
+ // agent) -- fall back to a REAL existence probe, built platform-aware
802
+ // via this repo's own getOsCommands() abstraction (src/os/*.ts,
803
+ // already used the same way by list-members.ts/member-detail.ts for
804
+ // credential-file checks) instead of a single hardcoded shell dialect.
805
+ const cmds = getOsCommands(getAgentOS(agent));
806
+ const projPath = `${resolvedWorkFolder}/${agentRelDir}/${input.agent}.md`;
807
+ const userPath = `~/${agentRelDir}/${input.agent}.md`;
808
+ const [projResult, userResult] = await Promise.all([
809
+ strategy.execCommand(cmds.credentialFileCheck(projPath), 10000),
810
+ strategy.execCommand(cmds.credentialFileCheck(userPath), 10000),
811
+ ]);
812
+ agentFound = projResult.stdout.includes('found') || userResult.stdout.includes('found');
813
+ }
814
+ if (!agentFound) {
815
+ inFlightAgents.delete(agent.id);
816
+ stallDetector.remove(agent.id);
817
+ writeStatusline(new Map([[agent.id, 'idle']]));
818
+ return `execute_prompt: agent "${input.agent}" not found on "${agent.friendlyName}".\n\nExpected at:\n ${resolvedWorkFolder}/${agentRelDir}/${input.agent}.md\n ~/${agentRelDir}/${input.agent}.md`;
819
+ }
820
+ }
821
+ }
822
+ // Context-headroom admission control (apra-fleet-eft.81.1): a declared
823
+ // demand is checked against this session's remaining headroom BEFORE any
824
+ // CLI is spawned. No declared demand (both fields omitted) means this
825
+ // block is skipped entirely -- pre-existing behavior is unchanged.
826
+ const contextDemand = resolveExpectedDemand(input.expected_context_tokens, input.context_size);
827
+ let contextWarning;
828
+ if (contextDemand !== undefined) {
829
+ const admission = checkContextAdmission({
830
+ provider: provider.name,
831
+ resolvedModel,
832
+ sessionId: mintedId,
833
+ demand: contextDemand,
834
+ });
835
+ if (!admission.allowed) {
836
+ scope.abort(`insufficient context headroom: demand=${admission.detail.demand} headroom=${admission.detail.headroom} window=${admission.detail.window}`);
837
+ inFlightAgents.delete(agent.id);
838
+ stallDetector.remove(agent.id);
839
+ writeStatusline(new Map([[agent.id, 'idle']]));
840
+ return {
841
+ text: `[FAIL] execute_prompt on "${agent.friendlyName}" rejected -- insufficient context headroom (demand=${admission.detail.demand}, headroom=${admission.detail.headroom}, window=${admission.detail.window}). Start a fresh session, shrink the task, or split it.`,
842
+ structuredContent: { isError: true, reason: 'insufficient_context_headroom', detail: admission.detail },
843
+ };
844
+ }
845
+ if (admission.warning) {
846
+ contextWarning = { message: admission.warning, detail: admission.detail };
847
+ }
848
+ }
849
+ // Usage/budget admission (apra-fleet-eft.80.2): when a budget is configured
850
+ // for this member (or its workspace), a hard-threshold crossing rejects the
851
+ // NEW dispatch BEFORE any LLM is spawned -- never killing an in-flight call.
852
+ // No configured budget (the common case) means resolveBudgetScope returns
853
+ // undefined and this block is skipped entirely -- behavior is unchanged.
854
+ const budgetScope = resolveBudgetScope([agent.id, workspaceId]);
855
+ if (budgetScope) {
856
+ const preBudget = await evaluateBudget({ scope: budgetScope, agent, provider });
857
+ if (preBudget?.exhausted) {
858
+ scope.abort(`budget exhausted: scope=${budgetScope} spent=${preBudget.block.spent} budget=${preBudget.block.budget} source=${preBudget.block.source}`);
859
+ inFlightAgents.delete(agent.id);
860
+ stallDetector.remove(agent.id);
861
+ writeStatusline(new Map([[agent.id, 'idle']]));
862
+ return {
863
+ text: `[FAIL] execute_prompt on "${agent.friendlyName}" rejected -- budget exhausted (scope=${budgetScope}, spent=${preBudget.block.spent}, budget=${preBudget.block.budget} ${preBudget.block.unit}, source=${preBudget.block.source}). No LLM call was made; raise or reset the budget to resume.`,
864
+ structuredContent: { isError: true, reason: 'budget_exhausted', budgetUsage: preBudget.block },
865
+ };
866
+ }
867
+ }
234
868
  // Kill any leftover session from a previous (possibly zombie) execute_prompt call
235
869
  await tryKillPid(agent, strategy, cmds);
236
- // Write the prompt to the unique prompt file before execution
237
- await writePromptFile(agent, strategy, promptFilePath, input.prompt);
870
+ // Write the rendered prompt (with substitutions applied) to the prompt file before execution
871
+ await writePromptFile(agent, strategy, promptFilePath, renderedPrompt);
872
+ // apra-fleet-6z8.1: remembered for the lease-of-life gate below -- the pid
873
+ // outlives the SSH channel that reported it, and is the only way to tell a
874
+ // fabricated "exit 0 / empty" close apart from a real one.
875
+ let capturedPid;
238
876
  const onPidCaptured = (pid) => {
877
+ capturedPid = pid;
239
878
  scope.info(`pid=${pid}`);
240
879
  if (mintedId) {
241
880
  try {
@@ -254,52 +893,264 @@ export async function executePrompt(input, extra) {
254
893
  tryKillPid(agent, strategy, cmds).catch(() => { });
255
894
  };
256
895
  extra?.signal?.addEventListener('abort', abortHandler);
896
+ // apra-fleet-3c9.1: the signal handed to execCommand fires on EITHER the MCP
897
+ // client's cancellation OR a confirmed stall (stallAbortController). Merging
898
+ // them means a stall aborts the pending dispatch exactly as a client cancel
899
+ // would, while a live (non-stalled) dispatch -- whose controller is never
900
+ // aborted -- is left completely untouched.
901
+ const dispatchSignal = extra?.signal
902
+ ? AbortSignal.any([extra.signal, stallAbortController.signal])
903
+ : stallAbortController.signal;
257
904
  // Mark agent as busy in statusline
258
905
  writeStatusline(new Map([[agent.id, 'busy']]));
259
906
  let _epExitCode = 'error';
260
907
  let _epError;
261
908
  let _epUsage;
262
909
  let _epOffline = false;
910
+ // apra-fleet-6a7.1: gates the exit-0/empty-stdout workspace_not_trusted
911
+ // self-heal-and-retry below to exactly one attempt per call, mirroring
912
+ // runGitStep's authHealAttempted shape (fleet-sprint/runner.js:616) -- a
913
+ // repeat trust failure after the heal is terminal, never looped.
914
+ let trustHealAttempted = false;
263
915
  try {
264
- let result = await strategy.execCommand(claudeCmd, timeoutMs, maxTotalMs, onPidCaptured, extra?.signal);
265
- let parsed = provider.parseResponse(result);
266
- if (parsed.usage)
267
- _epUsage = parsed.usage;
268
- // Stale session retry — fresh session ID, no resume
269
- if (result.code !== 0 && input.resume && agent.sessionId) {
270
- scope.info(`[${resolvedModel}] retrying — stale session`);
271
- await tryKillPid(agent, strategy, cmds);
272
- const freshOpts = { ...promptOpts, sessionId: (provider.name === 'claude' || provider.name === 'gemini' || provider.name === 'agy') ? uuid() : undefined, resuming: false };
273
- const retryCmd = authPrefix + cmds.buildAgentPromptCommand(provider, freshOpts);
274
- result = await strategy.execCommand(retryCmd, timeoutMs, maxTotalMs, onPidCaptured, extra?.signal);
275
- parsed = provider.parseResponse(result);
276
- if (parsed.usage)
277
- _epUsage = parsed.usage;
278
- }
279
- // Server/overloaded error retry — single attempt after delay
280
- if (result.code !== 0 && isRetryable(provider.classifyError(result.stderr || result.stdout))) {
281
- scope.info(`[${resolvedModel}] retrying — server overloaded`);
916
+ let result;
917
+ try {
918
+ result = await strategy.execCommand(claudeCmd, timeoutMs, maxTotalMs, onPidCaptured, dispatchSignal);
919
+ }
920
+ catch (dispatchErr) {
921
+ // apra-fleet-02s.1: a genuine command-execution exception (e.g. an
922
+ // inactivity timeout, or any other error strategy.execCommand throws)
923
+ // used to be unconditionally unretried here -- it bypasses both retry
924
+ // mechanisms below, since those only fire on a non-throwing nonzero
925
+ // exit, never on a thrown exception. Retry once with a fresh session
926
+ // before giving up, mirroring the stale-session/server-overloaded
927
+ // retries' bounded, single-attempt shape. Skip the retry if the client
928
+ // itself cancelled the request -- there is nothing to recover from a
929
+ // deliberate cancellation.
930
+ if (extra?.signal?.aborted)
931
+ throw dispatchErr;
932
+ // apra-fleet-3c9.1: a stall-triggered abort is terminal. onStall already
933
+ // killed the remote process and its session is gone, so retrying in a
934
+ // fresh session would just re-dispatch onto a member we just tore down.
935
+ // Let the exception surface to the outer catch, which classifies it as a
936
+ // typed 'stalled' error instead of retrying or hanging.
937
+ if (stallAbortController.signal.aborted)
938
+ throw dispatchErr;
939
+ // apra-fleet-eft.78.1: an explicit-id resume must NOT retry in a fresh
940
+ // session (that would run a context-dependent delta prompt with no
941
+ // context). Let the exception surface as dispatch_failed instead.
942
+ if (!allowFreshSessionFallback)
943
+ throw dispatchErr;
944
+ // apra-fleet-y8q.1: no budget left to share with a retry (the original
945
+ // attempt already consumed the whole max_total_s) -- retrying here would
946
+ // just re-burn a fresh full budget past what the client is waiting for.
947
+ // Let the original exception surface instead of retrying blind.
948
+ const budget = retryBudget();
949
+ if (budget.exhausted)
950
+ throw dispatchErr;
951
+ scope.info(`[${resolvedModel}] retrying -- dispatch exception: ${dispatchErr.message}`);
282
952
  await tryKillPid(agent, strategy, cmds);
283
- await new Promise(r => setTimeout(r, SERVER_RETRY_DELAY_MS));
284
953
  const freshOpts = { ...promptOpts, sessionId: (provider.name === 'claude' || provider.name === 'gemini' || provider.name === 'agy') ? uuid() : undefined, resuming: false };
285
954
  const retryCmd = authPrefix + cmds.buildAgentPromptCommand(provider, freshOpts);
286
- result = await strategy.execCommand(retryCmd, timeoutMs, maxTotalMs, onPidCaptured, extra?.signal);
287
- parsed = provider.parseResponse(result);
288
- if (parsed.usage)
289
- _epUsage = parsed.usage;
955
+ result = await strategy.execCommand(retryCmd, budget.timeoutMs, budget.maxTotalMs, onPidCaptured, dispatchSignal);
956
+ }
957
+ let parsed = provider.parseResponse(result);
958
+ if (parsed.usage)
959
+ _epUsage = parsed.usage;
960
+ // apra-fleet-eft.40.3: workspace-not-trusted degrades composed permissions
961
+ // (project-scoped allow entries silently dropped) without killing the CLI process --
962
+ // from there, unattended -p cannot auto-approve or prompt, so tools get denied and
963
+ // the turn eventually fails. Blindly falling through to the stale-session /
964
+ // server-overloaded retries below just repeats the same degraded dispatch against a
965
+ // workspace that is still untrusted, and can walk into eft.28's dead-session hang.
966
+ // Classify and fail fast here, before either retry path, naming
967
+ // ensureWorkspaceTrusted (apra-fleet-eft.40.1/40.2) as the remediation.
968
+ if (result.code !== 0 && provider.classifyError(result.stderr || result.stdout) === 'workspace_not_trusted') {
969
+ return {
970
+ text: `[FAIL] ${workspaceNotTrustedAdvice(agent.friendlyName)}\n${result.stderr || result.stdout}`,
971
+ structuredContent: { isError: true, reason: 'workspace_not_trusted' },
972
+ };
973
+ }
974
+ // Stale session retry -- fresh session ID, no resume. apra-fleet-eft.78.1:
975
+ // ONLY the best-effort resume=true mode gets this transparent retry-fresh
976
+ // recovery. An explicit session-id resume (string) deliberately does NOT --
977
+ // its caller asserted context dependence, so a not-found id is terminal.
978
+ if (result.code !== 0 && input.resume === true && agent.sessionId) {
979
+ // apra-fleet-y8q.1: share the remaining max_total_s budget with this
980
+ // retry too -- skip it outright once exhausted (see retryBudget above).
981
+ const staleBudget = retryBudget();
982
+ if (!staleBudget.exhausted) {
983
+ scope.info(`[${resolvedModel}] retrying -- stale session`);
984
+ await tryKillPid(agent, strategy, cmds);
985
+ const freshOpts = { ...promptOpts, sessionId: (provider.name === 'claude' || provider.name === 'gemini' || provider.name === 'agy') ? uuid() : undefined, resuming: false };
986
+ const retryCmd = authPrefix + cmds.buildAgentPromptCommand(provider, freshOpts);
987
+ result = await strategy.execCommand(retryCmd, staleBudget.timeoutMs, staleBudget.maxTotalMs, onPidCaptured, dispatchSignal);
988
+ parsed = provider.parseResponse(result);
989
+ if (parsed.usage)
990
+ _epUsage = parsed.usage;
991
+ }
992
+ }
993
+ // Server/overloaded error retry -- single attempt after delay. Skipped for
994
+ // an explicit-id resume (apra-fleet-eft.78.1): the retry starts a fresh
995
+ // session, which would discard the exact context the caller asked to resume.
996
+ if (result.code !== 0 && allowFreshSessionFallback && isRetryable(provider.classifyError(result.stderr || result.stdout))) {
997
+ // apra-fleet-y8q.1: share the remaining max_total_s budget with this
998
+ // retry too -- skip it outright once exhausted (see retryBudget above).
999
+ const overloadBudget = retryBudget();
1000
+ if (!overloadBudget.exhausted) {
1001
+ scope.info(`[${resolvedModel}] retrying -- server overloaded`);
1002
+ await tryKillPid(agent, strategy, cmds);
1003
+ await new Promise(r => setTimeout(r, SERVER_RETRY_DELAY_MS));
1004
+ const freshOpts = { ...promptOpts, sessionId: (provider.name === 'claude' || provider.name === 'gemini' || provider.name === 'agy') ? uuid() : undefined, resuming: false };
1005
+ const retryCmd = authPrefix + cmds.buildAgentPromptCommand(provider, freshOpts);
1006
+ result = await strategy.execCommand(retryCmd, overloadBudget.timeoutMs, overloadBudget.maxTotalMs, onPidCaptured, dispatchSignal);
1007
+ parsed = provider.parseResponse(result);
1008
+ if (parsed.usage)
1009
+ _epUsage = parsed.usage;
1010
+ }
290
1011
  }
291
1012
  _epExitCode = result.code;
292
1013
  if (result.code !== 0) {
293
- return buildFailureMessage(agent.friendlyName, result, provider);
1014
+ // apra-fleet-391: surface an auth failure as a STRUCTURED reason (not
1015
+ // just prose in `text`) so callers -- notably fleet-sprint's
1016
+ // isAuthDispatchError -- can key off it directly instead of regexing
1017
+ // the message string. Overloading 'nonzero_exit' for this case was
1018
+ // what made auth self-heal impossible to wire reliably upstream.
1019
+ const failureCategory = isMaxTurnsResponse(parsed)
1020
+ ? 'max_turns'
1021
+ : provider.classifyError(result.stderr || result.stdout);
1022
+ return {
1023
+ text: buildFailureMessage(agent.friendlyName, result, provider, parsed),
1024
+ structuredContent: {
1025
+ isError: true,
1026
+ reason: failureCategory === 'max_turns'
1027
+ ? 'max_turns_exhausted'
1028
+ : failureCategory === 'auth'
1029
+ ? 'auth'
1030
+ : failureCategory === 'server'
1031
+ ? 'server'
1032
+ : failureCategory === 'overloaded'
1033
+ ? 'overloaded'
1034
+ : failureCategory === 'workspace_not_trusted'
1035
+ ? 'workspace_not_trusted'
1036
+ : 'nonzero_exit',
1037
+ // apra-fleet-63x.1: a nonzero exit -- most commonly max_turns_exhausted,
1038
+ // where the CLI ran real turns and burned real tokens before hitting its
1039
+ // ceiling -- still has a REAL parsed usage figure sitting in _epUsage
1040
+ // (captured just above from `parsed.usage` regardless of exit code).
1041
+ // This branch used to return no `usage` field at all on any failure,
1042
+ // so FleetWorkflow.agent() (packages/apra-fleet-workflow/src/workflow/
1043
+ // index.mjs, apra-fleet-202.3) saw hasRealUsage=false and never priced
1044
+ // it, silently under-counting a sprint's tracked spend (observed: a
1045
+ // 10-hour run with dozens of max_turns exhaustions reporting
1046
+ // stats.totalCost of $0). Attach it here whenever it's available so
1047
+ // the caller can record the real partial cost instead of nothing.
1048
+ ...(_epUsage ? { usage: { input_tokens: _epUsage.input_tokens, output_tokens: _epUsage.output_tokens, total_tokens: _epUsage.input_tokens + _epUsage.output_tokens } } : {}),
1049
+ },
1050
+ };
1051
+ }
1052
+ // Exit 0 but the provider parser extracted NOTHING (no result text)
1053
+ // -- observed live (apra-fleet-eft.14, 2026-07-19 stabilization loop):
1054
+ // the claude CLI can die silently mid-turn (member-side session
1055
+ // transcript stops at a tool_result with no final assistant message)
1056
+ // and still exit 0 with EMPTY stdout, so parseResponse falls through
1057
+ // to its plain-text fallback with result: ''. Returning that as a
1058
+ // success used to hand callers a display wrapper with nothing inside,
1059
+ // which schema-extraction layers then misreported as "LLM returned
1060
+ // invalid JSON". Classify it at the source as a typed dispatch error
1061
+ // instead, with stderr's tail attached for diagnosis.
1062
+ //
1063
+ // apra-fleet-6z8.1 (lease-of-life gate): before declaring that, cross-check
1064
+ // whether the captured pid is STILL ALIVE via a fresh short exec. ssh.ts
1065
+ // substitutes code 0 when the channel closes without ever receiving an
1066
+ // 'exit' event, so "exit 0 + empty" is ALSO exactly what a torn-down
1067
+ // channel over a perfectly healthy, still-running turn looks like (live
1068
+ // evidence: pid 89858 running 2+ minutes past the close). Declaring
1069
+ // empty_response there orphans the CLI and invites a duplicate concurrent
1070
+ // dispatch of the same scope. A confirmed-dead pid keeps today's behavior
1071
+ // verbatim.
1072
+ if (!parsed.result || parsed.result.trim() === '') {
1073
+ const recovery = await recoverOrphanedDispatch({
1074
+ strategy,
1075
+ cmds,
1076
+ pid: capturedPid,
1077
+ durablePath,
1078
+ unsupported: getAgentOS(agent) === 'windows',
1079
+ maxWaitMs: maxTotalMs !== undefined ? Math.max(maxTotalMs - (Date.now() - dispatchStartedAt), 0) : undefined,
1080
+ scope,
1081
+ });
1082
+ if (recovery.status === 'timeout') {
1083
+ scope.info(`orphan recovery timed out after ${Math.round((recovery.waitedMs ?? 0) / 1000)}s -- pid killed`);
1084
+ return {
1085
+ text: `[FAIL] execute_prompt on "${agent.friendlyName}" lost its SSH channel while the member CLI (pid ${capturedPid}) was still running, and that process was still alive after the recovery window (${Math.round((recovery.waitedMs ?? 0) / 1000)}s). The process has been killed; no result was recovered. This is NOT an empty response -- do not treat it as a failed turn without checking the member's session transcript first.`,
1086
+ structuredContent: { isError: true, reason: 'orphan_recovery_timeout' },
1087
+ };
1088
+ }
1089
+ if (recovery.status === 'recovered' && recovery.stdout) {
1090
+ // Feed the durable output through the normal provider parse path,
1091
+ // exactly as if it had arrived on the original channel.
1092
+ const recoveredParsed = provider.parseResponse({ stdout: recovery.stdout, stderr: result.stderr ?? '', code: 0 });
1093
+ if (recoveredParsed.result && recoveredParsed.result.trim() !== '') {
1094
+ scope.info(`recovered the real result from the durable output file after a false-alarm empty_response (waited ${Math.round((recovery.waitedMs ?? 0) / 1000)}s)`);
1095
+ parsed = recoveredParsed;
1096
+ if (parsed.usage)
1097
+ _epUsage = parsed.usage;
1098
+ }
1099
+ }
1100
+ // apra-fleet-6a7.1: the code!==0 workspace_not_trusted classification
1101
+ // above (~line 1048) never sees this branch -- exit 0 + empty stdout is
1102
+ // a SEPARATE path, but live evidence (apra-fleet-2g2, fleet-win-dev1,
1103
+ // 2026-08-02) showed it can be caused by the exact same trust-gate
1104
+ // failure: exited 0, empty stdout, and the stderr tail contained the
1105
+ // exact matched phrase, so the code!==0 classifier never ran. Classify
1106
+ // it here too and -- unlike that path, which only advises a manual fix
1107
+ // -- self-heal via the same seedWorkspaceTrust/ensureWorkspaceTrusted
1108
+ // path compose_permissions already uses, then retry the dispatch
1109
+ // exactly once (gated by trustHealAttempted, mirroring runGitStep's
1110
+ // onAuthFailure shape: a repeat failure after the heal is terminal, not
1111
+ // looped).
1112
+ if ((!parsed.result || parsed.result.trim() === '')
1113
+ && provider.classifyError(result.stderr || result.stdout) === 'workspace_not_trusted'
1114
+ && !trustHealAttempted) {
1115
+ trustHealAttempted = true;
1116
+ scope.info(`[${resolvedModel}] exit-0/empty-stdout classified as workspace_not_trusted -- self-healing (seedWorkspaceTrust) once, then retrying the dispatch`);
1117
+ await seedWorkspaceTrust(agent, strategy, 'execute_prompt');
1118
+ const healBudget = retryBudget();
1119
+ if (!healBudget.exhausted) {
1120
+ await tryKillPid(agent, strategy, cmds);
1121
+ const freshOpts = { ...promptOpts, sessionId: (provider.name === 'claude' || provider.name === 'gemini' || provider.name === 'agy') ? uuid() : undefined, resuming: false };
1122
+ const retryCmd = authPrefix + cmds.buildAgentPromptCommand(provider, freshOpts);
1123
+ result = await strategy.execCommand(retryCmd, healBudget.timeoutMs, healBudget.maxTotalMs, onPidCaptured, dispatchSignal);
1124
+ parsed = provider.parseResponse(result);
1125
+ if (parsed.usage)
1126
+ _epUsage = parsed.usage;
1127
+ }
1128
+ }
1129
+ if (!parsed.result || parsed.result.trim() === '') {
1130
+ const stderrTail = (result.stderr || '').trim().slice(-500);
1131
+ const trustClassified = provider.classifyError(result.stderr || result.stdout) === 'workspace_not_trusted';
1132
+ return {
1133
+ text: trustClassified
1134
+ ? `[FAIL] ${workspaceNotTrustedAdvice(agent.friendlyName)}\n${stderrTail}`
1135
+ : `[FAIL] execute_prompt on "${agent.friendlyName}" exited 0 but produced no parseable output (empty result -- the member CLI likely died mid-turn without printing its result envelope).${stderrTail ? `\n[stderr tail]\n${stderrTail}` : ''}`,
1136
+ structuredContent: { isError: true, reason: trustClassified ? 'workspace_not_trusted' : 'empty_response' },
1137
+ };
1138
+ }
294
1139
  }
295
1140
  // Session-id assertion: returned id must match the one we minted/resumed
296
- if (mintedId && parsed.sessionId && parsed.sessionId !== mintedId) {
1141
+ // Exception: AGY owns its session generation, so we always accept its returned ID
1142
+ if (provider.name !== 'agy' && mintedId && parsed.sessionId && parsed.sessionId !== mintedId) {
297
1143
  scope.info(`session-id mismatch: expected=${mintedId} got=${parsed.sessionId} -- not persisting`);
298
1144
  touchAgent(agent.id, undefined);
299
1145
  }
300
1146
  else {
301
- touchAgent(agent.id, mintedId ?? parsed.sessionId);
1147
+ touchAgent(agent.id, parsed.sessionId ?? mintedId);
302
1148
  }
1149
+ // apra-fleet-eft.78.1: mark the session this dispatch actually landed on as
1150
+ // known/resumable for this member, so a later explicit-id resume of it
1151
+ // passes the terminal session_not_found gate above instead of being
1152
+ // rejected as unknown.
1153
+ recordKnownSession(agent.id, parsed.sessionId ?? mintedId);
303
1154
  if (mintedId) {
304
1155
  try {
305
1156
  stallDetector.update(agent.id, {
@@ -319,25 +1170,65 @@ export async function executePrompt(input, extra) {
319
1170
  output: prev.output + parsed.usage.output_tokens,
320
1171
  },
321
1172
  });
1173
+ // Per-session cumulative tracking for context-headroom admission
1174
+ // (apra-fleet-eft.81.1) -- separate from the per-member lifetime total
1175
+ // just above. Keyed by whichever session id this dispatch actually
1176
+ // landed on (the returned id, falling back to the minted/resumed one),
1177
+ // so the NEXT admission check on this same session sees this dispatch's
1178
+ // usage already accounted for.
1179
+ recordSessionUsage(parsed.sessionId ?? mintedId, parsed.usage);
1180
+ }
1181
+ // Usage/budget accounting (apra-fleet-eft.80.2): account this dispatch's
1182
+ // token usage against the scope's budget (self-metered only when the
1183
+ // provider has no native getUsage()), then attach the usage block to the
1184
+ // result when the warning band is crossed. Nothing is attached below the
1185
+ // band, and nothing happens at all when no budget is configured.
1186
+ let budgetUsage;
1187
+ if (budgetScope && parsed.usage) {
1188
+ const postBudget = await recordAndEvaluate({ scope: budgetScope, agent, provider, tier: resolvedTier, usage: parsed.usage });
1189
+ if (postBudget?.warned)
1190
+ budgetUsage = postBudget.block;
322
1191
  }
323
- let output = `${deprecationWarning}📋 Response from ${agent.friendlyName}:
1192
+ let output = `[RESULT] Response from ${agent.friendlyName}:
324
1193
 
325
1194
  ${parsed.result}`;
326
- if (parsed.usage)
327
- output += `
328
- Tokens: input=${parsed.usage.input_tokens} output=${parsed.usage.output_tokens}`;
329
1195
  if (parsed.sessionId)
330
1196
  output += `
331
1197
 
332
1198
  ---
333
1199
  session: ${parsed.sessionId}`;
334
- return output;
1200
+ if (heuristicWarningSuffix)
1201
+ output += heuristicWarningSuffix;
1202
+ return {
1203
+ text: output,
1204
+ structuredContent: {
1205
+ response: parsed.result,
1206
+ ...(_epUsage ? { usage: { input_tokens: _epUsage.input_tokens, output_tokens: _epUsage.output_tokens, total_tokens: _epUsage.input_tokens + _epUsage.output_tokens } } : {}),
1207
+ ...(parsed.sessionId ? { sessionId: parsed.sessionId } : {}),
1208
+ ...(contextWarning ? { contextWarning } : {}),
1209
+ ...(budgetUsage ? { budgetUsage } : {}),
1210
+ },
1211
+ };
335
1212
  }
336
1213
  catch (err) {
1214
+ // apra-fleet-3c9.1: a confirmed stall aborted the in-flight execCommand (and
1215
+ // NOT the MCP client). Surface it as a typed 'stalled' error so the dispatch
1216
+ // settles here -- well under the client hard timeout -- instead of being
1217
+ // mislabeled dispatch_failed or waiting out the full deadline.
1218
+ if (stallAbortController.signal.aborted && !extra?.signal?.aborted) {
1219
+ _epError = 'dispatch aborted by confirmed stall';
1220
+ return {
1221
+ text: `[FAIL] execute_prompt on "${agent.friendlyName}" was aborted after a confirmed stall -- the remote turn made no progress for the stall threshold, its process was killed, and the in-flight dispatch was cancelled immediately rather than waiting out the client timeout.`,
1222
+ structuredContent: { isError: true, reason: 'stalled' },
1223
+ };
1224
+ }
337
1225
  // Only mark offline for genuine SSH/network connection failures, not for cancellations
338
1226
  _epOffline = !!(err.message && /ssh|network|econnrefused|ehostunreach|connection timed out/i.test(err.message));
339
1227
  _epError = err.message;
340
- return `❌ Failed to execute prompt on "${agent.friendlyName}": ${err.message}`;
1228
+ return {
1229
+ text: `[FAIL] Failed to execute prompt on "${agent.friendlyName}": ${err.message}`,
1230
+ structuredContent: { isError: true, reason: 'dispatch_failed' },
1231
+ };
341
1232
  }
342
1233
  finally {
343
1234
  extra?.signal?.removeEventListener('abort', abortHandler);
@@ -348,14 +1239,14 @@ session: ${parsed.sessionId}`;
348
1239
  scope.fail(`exit=${_epExitCode}${_epTok}`);
349
1240
  else
350
1241
  scope.ok(`exit=0${_epTok}`);
351
- // Skip if stall detector already cleared state — a new execute_prompt may have
1242
+ // Skip if stall detector already cleared state -- a new execute_prompt may have
352
1243
  // claimed inFlightAgents and set busy again; clobbering it here would be wrong.
353
1244
  if (!clearedByStall) {
354
1245
  writeStatusline(new Map([[agent.id, _epOffline ? 'offline' : 'idle']]));
355
1246
  inFlightAgents.delete(agent.id);
356
1247
  }
357
1248
  stallDetector.remove(agent.id);
358
- await deletePromptFile(agent, strategy, promptFilePath);
1249
+ await deletePromptFile(agent, strategy, promptFilePath, durablePath ? [durablePath] : []);
359
1250
  }
360
1251
  }
361
1252
  //# sourceMappingURL=execute-prompt.js.map