@apralabs/apra-fleet 0.3.5 → 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.
- package/README.md +245 -257
- package/dist/agents/backlog-groomer.md +287 -0
- package/dist/agents/deployer.md +52 -11
- package/dist/agents/doer.md +32 -7
- package/dist/agents/integ-test-runner.md +151 -61
- package/dist/agents/plan-reviewer.md +74 -17
- package/dist/agents/planner.md +94 -14
- package/dist/agents/regression-test-runner.md +216 -0
- package/dist/agents/reviewer.md +21 -6
- package/dist/agents/schemas/backlog-groomer-input.json +35 -0
- package/dist/agents/schemas/backlog-groomer-output.json +164 -0
- package/dist/agents/schemas/integ-test-runner-output.json +34 -0
- package/dist/agents/schemas/regression-test-runner-input.json +15 -0
- package/dist/agents/schemas/regression-test-runner-output.json +50 -0
- package/dist/agents/schemas/reviewer-output.json +4 -0
- package/dist/cli/agent-transform.d.ts +1 -0
- package/dist/cli/agent-transform.d.ts.map +1 -1
- package/dist/cli/agent-transform.js +77 -0
- package/dist/cli/agent-transform.js.map +1 -1
- package/dist/cli/auth.d.ts +3 -0
- package/dist/cli/auth.d.ts.map +1 -1
- package/dist/cli/auth.js +181 -5
- package/dist/cli/auth.js.map +1 -1
- package/dist/cli/config.d.ts +5 -1
- package/dist/cli/config.d.ts.map +1 -1
- package/dist/cli/config.js +7 -1
- package/dist/cli/config.js.map +1 -1
- package/dist/cli/dolt-install.d.ts +91 -0
- package/dist/cli/dolt-install.d.ts.map +1 -0
- package/dist/cli/dolt-install.js +334 -0
- package/dist/cli/dolt-install.js.map +1 -0
- package/dist/cli/install-guard.d.ts +29 -0
- package/dist/cli/install-guard.d.ts.map +1 -0
- package/dist/cli/install-guard.js +153 -0
- package/dist/cli/install-guard.js.map +1 -0
- package/dist/cli/install.d.ts +44 -3
- package/dist/cli/install.d.ts.map +1 -1
- package/dist/cli/install.js +604 -108
- package/dist/cli/install.js.map +1 -1
- package/dist/cli/join.d.ts +12 -0
- package/dist/cli/join.d.ts.map +1 -0
- package/dist/cli/join.js +87 -0
- package/dist/cli/join.js.map +1 -0
- package/dist/cli/register-member.d.ts +2 -0
- package/dist/cli/register-member.d.ts.map +1 -0
- package/dist/cli/register-member.js +193 -0
- package/dist/cli/register-member.js.map +1 -0
- package/dist/cli/restart.d.ts +2 -0
- package/dist/cli/restart.d.ts.map +1 -0
- package/dist/cli/restart.js +7 -0
- package/dist/cli/restart.js.map +1 -0
- package/dist/cli/secret.js +8 -2
- package/dist/cli/secret.js.map +1 -1
- package/dist/cli/spoke.d.ts +35 -0
- package/dist/cli/spoke.d.ts.map +1 -0
- package/dist/cli/spoke.js +212 -0
- package/dist/cli/spoke.js.map +1 -0
- package/dist/cli/start.d.ts +2 -0
- package/dist/cli/start.d.ts.map +1 -0
- package/dist/cli/start.js +92 -0
- package/dist/cli/start.js.map +1 -0
- package/dist/cli/status.d.ts +2 -0
- package/dist/cli/status.d.ts.map +1 -0
- package/dist/cli/status.js +83 -0
- package/dist/cli/status.js.map +1 -0
- package/dist/cli/stop.d.ts +2 -0
- package/dist/cli/stop.d.ts.map +1 -0
- package/dist/cli/stop.js +51 -0
- package/dist/cli/stop.js.map +1 -0
- package/dist/cli/uninstall.d.ts.map +1 -1
- package/dist/cli/uninstall.js +109 -9
- package/dist/cli/uninstall.js.map +1 -1
- package/dist/cli/update.d.ts.map +1 -1
- package/dist/cli/update.js +10 -1
- package/dist/cli/update.js.map +1 -1
- package/dist/cli/workflow-assets.d.ts +37 -0
- package/dist/cli/workflow-assets.d.ts.map +1 -0
- package/dist/cli/workflow-assets.js +144 -0
- package/dist/cli/workflow-assets.js.map +1 -0
- package/dist/cli/workflow.d.ts +75 -0
- package/dist/cli/workflow.d.ts.map +1 -0
- package/dist/cli/workflow.js +397 -0
- package/dist/cli/workflow.js.map +1 -0
- package/dist/hub-service/activity.d.ts +30 -0
- package/dist/hub-service/activity.d.ts.map +1 -0
- package/dist/hub-service/activity.js +20 -0
- package/dist/hub-service/activity.js.map +1 -0
- package/dist/hub-service/audit-log.d.ts +19 -0
- package/dist/hub-service/audit-log.d.ts.map +1 -0
- package/dist/hub-service/audit-log.js +13 -0
- package/dist/hub-service/audit-log.js.map +1 -0
- package/dist/hub-service/db/pool.d.ts +19 -0
- package/dist/hub-service/db/pool.d.ts.map +1 -0
- package/dist/hub-service/db/pool.js +53 -0
- package/dist/hub-service/db/pool.js.map +1 -0
- package/dist/hub-service/enrollment.d.ts +29 -0
- package/dist/hub-service/enrollment.d.ts.map +1 -0
- package/dist/hub-service/enrollment.js +48 -0
- package/dist/hub-service/enrollment.js.map +1 -0
- package/dist/hub-service/envelope-routes.d.ts +46 -0
- package/dist/hub-service/envelope-routes.d.ts.map +1 -0
- package/dist/hub-service/envelope-routes.js +94 -0
- package/dist/hub-service/envelope-routes.js.map +1 -0
- package/dist/hub-service/handlers/installers.d.ts +14 -0
- package/dist/hub-service/handlers/installers.d.ts.map +1 -0
- package/dist/hub-service/handlers/installers.js +8 -0
- package/dist/hub-service/handlers/installers.js.map +1 -0
- package/dist/hub-service/http-server.d.ts +31 -0
- package/dist/hub-service/http-server.d.ts.map +1 -0
- package/dist/hub-service/http-server.js +711 -0
- package/dist/hub-service/http-server.js.map +1 -0
- package/dist/hub-service/hub-jwt.d.ts +30 -0
- package/dist/hub-service/hub-jwt.d.ts.map +1 -0
- package/dist/hub-service/hub-jwt.js +105 -0
- package/dist/hub-service/hub-jwt.js.map +1 -0
- package/dist/hub-service/jwt-revocation.d.ts +20 -0
- package/dist/hub-service/jwt-revocation.d.ts.map +1 -0
- package/dist/hub-service/jwt-revocation.js +18 -0
- package/dist/hub-service/jwt-revocation.js.map +1 -0
- package/dist/hub-service/machines.d.ts +16 -0
- package/dist/hub-service/machines.d.ts.map +1 -0
- package/dist/hub-service/machines.js +14 -0
- package/dist/hub-service/machines.js.map +1 -0
- package/dist/hub-service/main.d.ts +3 -0
- package/dist/hub-service/main.d.ts.map +1 -0
- package/dist/hub-service/main.js +64 -0
- package/dist/hub-service/main.js.map +1 -0
- package/dist/hub-service/member-tokens.d.ts +11 -0
- package/dist/hub-service/member-tokens.d.ts.map +1 -0
- package/dist/hub-service/member-tokens.js +49 -0
- package/dist/hub-service/member-tokens.js.map +1 -0
- package/dist/hub-service/member-view.d.ts +20 -0
- package/dist/hub-service/member-view.d.ts.map +1 -0
- package/dist/hub-service/member-view.js +76 -0
- package/dist/hub-service/member-view.js.map +1 -0
- package/dist/hub-service/members.d.ts +44 -0
- package/dist/hub-service/members.d.ts.map +1 -0
- package/dist/hub-service/members.js +25 -0
- package/dist/hub-service/members.js.map +1 -0
- package/dist/hub-service/presence.d.ts +40 -0
- package/dist/hub-service/presence.d.ts.map +1 -0
- package/dist/hub-service/presence.js +44 -0
- package/dist/hub-service/presence.js.map +1 -0
- package/dist/hub-service/project-view.d.ts +12 -0
- package/dist/hub-service/project-view.d.ts.map +1 -0
- package/dist/hub-service/project-view.js +34 -0
- package/dist/hub-service/project-view.js.map +1 -0
- package/dist/hub-service/projects.d.ts +48 -0
- package/dist/hub-service/projects.d.ts.map +1 -0
- package/dist/hub-service/projects.js +80 -0
- package/dist/hub-service/projects.js.map +1 -0
- package/dist/hub-service/relay-queue.d.ts +115 -0
- package/dist/hub-service/relay-queue.d.ts.map +1 -0
- package/dist/hub-service/relay-queue.js +148 -0
- package/dist/hub-service/relay-queue.js.map +1 -0
- package/dist/hub-service/session-jwt.d.ts +10 -0
- package/dist/hub-service/session-jwt.d.ts.map +1 -0
- package/dist/hub-service/session-jwt.js +69 -0
- package/dist/hub-service/session-jwt.js.map +1 -0
- package/dist/hub-service/usage.d.ts +40 -0
- package/dist/hub-service/usage.d.ts.map +1 -0
- package/dist/hub-service/usage.js +25 -0
- package/dist/hub-service/usage.js.map +1 -0
- package/dist/hub-service/users.d.ts +56 -0
- package/dist/hub-service/users.d.ts.map +1 -0
- package/dist/hub-service/users.js +72 -0
- package/dist/hub-service/users.js.map +1 -0
- package/dist/hub-service/workspaces.d.ts +20 -0
- package/dist/hub-service/workspaces.d.ts.map +1 -0
- package/dist/hub-service/workspaces.js +10 -0
- package/dist/hub-service/workspaces.js.map +1 -0
- package/dist/index.js +179 -132
- package/dist/index.js.map +1 -1
- package/dist/os/linux.d.ts +13 -0
- package/dist/os/linux.d.ts.map +1 -1
- package/dist/os/linux.js +62 -3
- package/dist/os/linux.js.map +1 -1
- package/dist/os/os-commands.d.ts +10 -0
- package/dist/os/os-commands.d.ts.map +1 -1
- package/dist/os/windows.d.ts +1 -0
- package/dist/os/windows.d.ts.map +1 -1
- package/dist/os/windows.js +21 -1
- package/dist/os/windows.js.map +1 -1
- package/dist/paths.d.ts +23 -0
- package/dist/paths.d.ts.map +1 -1
- package/dist/paths.js +29 -0
- package/dist/paths.js.map +1 -1
- package/dist/providers/agy.d.ts +9 -2
- package/dist/providers/agy.d.ts.map +1 -1
- package/dist/providers/agy.js +127 -14
- package/dist/providers/agy.js.map +1 -1
- package/dist/providers/claude.d.ts +6 -2
- package/dist/providers/claude.d.ts.map +1 -1
- package/dist/providers/claude.js +248 -9
- package/dist/providers/claude.js.map +1 -1
- package/dist/providers/codex.d.ts +3 -2
- package/dist/providers/codex.d.ts.map +1 -1
- package/dist/providers/codex.js +7 -0
- package/dist/providers/codex.js.map +1 -1
- package/dist/providers/copilot.d.ts +3 -2
- package/dist/providers/copilot.d.ts.map +1 -1
- package/dist/providers/copilot.js +8 -1
- package/dist/providers/copilot.js.map +1 -1
- package/dist/providers/gemini.d.ts +3 -2
- package/dist/providers/gemini.d.ts.map +1 -1
- package/dist/providers/gemini.js +11 -1
- package/dist/providers/gemini.js.map +1 -1
- package/dist/providers/index.d.ts +1 -0
- package/dist/providers/index.d.ts.map +1 -1
- package/dist/providers/index.js +3 -0
- package/dist/providers/index.js.map +1 -1
- package/dist/providers/none.d.ts +51 -0
- package/dist/providers/none.d.ts.map +1 -0
- package/dist/providers/none.js +103 -0
- package/dist/providers/none.js.map +1 -0
- package/dist/providers/opencode.d.ts +4 -2
- package/dist/providers/opencode.d.ts.map +1 -1
- package/dist/providers/opencode.js +49 -0
- package/dist/providers/opencode.js.map +1 -1
- package/dist/providers/provider.d.ts +79 -1
- package/dist/providers/provider.d.ts.map +1 -1
- package/dist/providers/provider.js +10 -0
- package/dist/providers/provider.js.map +1 -1
- package/dist/services/agent-provisioner.d.ts +3 -3
- package/dist/services/agent-provisioner.d.ts.map +1 -1
- package/dist/services/agent-provisioner.js +12 -5
- package/dist/services/agent-provisioner.js.map +1 -1
- package/dist/services/auth-socket.d.ts.map +1 -1
- package/dist/services/auth-socket.js +2 -0
- package/dist/services/auth-socket.js.map +1 -1
- package/dist/services/budget-awareness.d.ts +130 -0
- package/dist/services/budget-awareness.d.ts.map +1 -0
- package/dist/services/budget-awareness.js +129 -0
- package/dist/services/budget-awareness.js.map +1 -0
- package/dist/services/cloud-sync.d.ts +35 -0
- package/dist/services/cloud-sync.d.ts.map +1 -0
- package/dist/services/cloud-sync.js +90 -0
- package/dist/services/cloud-sync.js.map +1 -0
- package/dist/services/context-admission.d.ts +82 -0
- package/dist/services/context-admission.d.ts.map +1 -0
- package/dist/services/context-admission.js +136 -0
- package/dist/services/context-admission.js.map +1 -0
- package/dist/services/event-bus.d.ts +27 -0
- package/dist/services/event-bus.d.ts.map +1 -0
- package/dist/services/event-bus.js +20 -0
- package/dist/services/event-bus.js.map +1 -0
- package/dist/services/file-chunker.d.ts +38 -0
- package/dist/services/file-chunker.d.ts.map +1 -0
- package/dist/services/file-chunker.js +63 -0
- package/dist/services/file-chunker.js.map +1 -0
- package/dist/services/file-transfer-relay.d.ts +55 -0
- package/dist/services/file-transfer-relay.d.ts.map +1 -0
- package/dist/services/file-transfer-relay.js +115 -0
- package/dist/services/file-transfer-relay.js.map +1 -0
- package/dist/services/github-app.d.ts.map +1 -1
- package/dist/services/github-app.js +1 -0
- package/dist/services/github-app.js.map +1 -1
- package/dist/services/http-transport.d.ts +25 -0
- package/dist/services/http-transport.d.ts.map +1 -0
- package/dist/services/http-transport.js +371 -0
- package/dist/services/http-transport.js.map +1 -0
- package/dist/services/hub-client.d.ts +69 -0
- package/dist/services/hub-client.d.ts.map +1 -0
- package/dist/services/hub-client.js +217 -0
- package/dist/services/hub-client.js.map +1 -0
- package/dist/services/jwt.d.ts +17 -0
- package/dist/services/jwt.d.ts.map +1 -0
- package/dist/services/jwt.js +70 -0
- package/dist/services/jwt.js.map +1 -0
- package/dist/services/known-sessions.d.ts +8 -0
- package/dist/services/known-sessions.d.ts.map +1 -0
- package/dist/services/known-sessions.js +40 -0
- package/dist/services/known-sessions.js.map +1 -0
- package/dist/services/model-pricing.d.ts +36 -0
- package/dist/services/model-pricing.d.ts.map +1 -0
- package/dist/services/model-pricing.js +108 -0
- package/dist/services/model-pricing.js.map +1 -0
- package/dist/services/orphan-recovery.d.ts +56 -0
- package/dist/services/orphan-recovery.d.ts.map +1 -0
- package/dist/services/orphan-recovery.js +111 -0
- package/dist/services/orphan-recovery.js.map +1 -0
- package/dist/services/pending-responses.d.ts +23 -0
- package/dist/services/pending-responses.d.ts.map +1 -0
- package/dist/services/pending-responses.js +52 -0
- package/dist/services/pending-responses.js.map +1 -0
- package/dist/services/registry.d.ts +4 -1
- package/dist/services/registry.d.ts.map +1 -1
- package/dist/services/registry.js +3 -0
- package/dist/services/registry.js.map +1 -1
- package/dist/services/relay-context.d.ts +23 -0
- package/dist/services/relay-context.d.ts.map +1 -0
- package/dist/services/relay-context.js +8 -0
- package/dist/services/relay-context.js.map +1 -0
- package/dist/services/relay-executor.d.ts +68 -0
- package/dist/services/relay-executor.d.ts.map +1 -0
- package/dist/services/relay-executor.js +70 -0
- package/dist/services/relay-executor.js.map +1 -0
- package/dist/services/relay-request.d.ts +65 -0
- package/dist/services/relay-request.d.ts.map +1 -0
- package/dist/services/relay-request.js +129 -0
- package/dist/services/relay-request.js.map +1 -0
- package/dist/services/relay-strategy.d.ts +19 -0
- package/dist/services/relay-strategy.d.ts.map +1 -0
- package/dist/services/relay-strategy.js +111 -0
- package/dist/services/relay-strategy.js.map +1 -0
- package/dist/services/service-manager/index.d.ts +5 -0
- package/dist/services/service-manager/index.d.ts.map +1 -0
- package/dist/services/service-manager/index.js +66 -0
- package/dist/services/service-manager/index.js.map +1 -0
- package/dist/services/service-manager/linux.d.ts +10 -0
- package/dist/services/service-manager/linux.d.ts.map +1 -0
- package/dist/services/service-manager/linux.js +97 -0
- package/dist/services/service-manager/linux.js.map +1 -0
- package/dist/services/service-manager/macos.d.ts +10 -0
- package/dist/services/service-manager/macos.d.ts.map +1 -0
- package/dist/services/service-manager/macos.js +94 -0
- package/dist/services/service-manager/macos.js.map +1 -0
- package/dist/services/service-manager/types.d.ts +18 -0
- package/dist/services/service-manager/types.d.ts.map +1 -0
- package/dist/services/service-manager/types.js +5 -0
- package/dist/services/service-manager/types.js.map +1 -0
- package/dist/services/service-manager/windows.d.ts +10 -0
- package/dist/services/service-manager/windows.d.ts.map +1 -0
- package/dist/services/service-manager/windows.js +73 -0
- package/dist/services/service-manager/windows.js.map +1 -0
- package/dist/services/session-registry.d.ts +70 -0
- package/dist/services/session-registry.d.ts.map +1 -0
- package/dist/services/session-registry.js +91 -0
- package/dist/services/session-registry.js.map +1 -0
- package/dist/services/singleton.d.ts +15 -0
- package/dist/services/singleton.d.ts.map +1 -0
- package/dist/services/singleton.js +98 -0
- package/dist/services/singleton.js.map +1 -0
- package/dist/services/sprint-coordination.d.ts +168 -0
- package/dist/services/sprint-coordination.d.ts.map +1 -0
- package/dist/services/sprint-coordination.js +564 -0
- package/dist/services/sprint-coordination.js.map +1 -0
- package/dist/services/ssh.d.ts.map +1 -1
- package/dist/services/ssh.js +70 -3
- package/dist/services/ssh.js.map +1 -1
- package/dist/services/stall/log-path-resolver.d.ts.map +1 -1
- package/dist/services/stall/log-path-resolver.js +8 -1
- package/dist/services/stall/log-path-resolver.js.map +1 -1
- package/dist/services/stall/stall-detector.d.ts.map +1 -1
- package/dist/services/stall/stall-detector.js +39 -7
- package/dist/services/stall/stall-detector.js.map +1 -1
- package/dist/services/stall/stall-poller.d.ts +16 -0
- package/dist/services/stall/stall-poller.d.ts.map +1 -1
- package/dist/services/stall/stall-poller.js +136 -18
- package/dist/services/stall/stall-poller.js.map +1 -1
- package/dist/services/strategy.d.ts.map +1 -1
- package/dist/services/strategy.js +55 -5
- package/dist/services/strategy.js.map +1 -1
- package/dist/services/substitution-engine.d.ts +20 -0
- package/dist/services/substitution-engine.d.ts.map +1 -0
- package/dist/services/substitution-engine.js +92 -0
- package/dist/services/substitution-engine.js.map +1 -0
- package/dist/services/task-cleanup.d.ts.map +1 -1
- package/dist/services/task-cleanup.js +1 -9
- package/dist/services/task-cleanup.js.map +1 -1
- package/dist/services/token-issuer.d.ts +47 -0
- package/dist/services/token-issuer.d.ts.map +1 -0
- package/dist/services/token-issuer.js +63 -0
- package/dist/services/token-issuer.js.map +1 -0
- package/dist/services/tool-registry.d.ts +3 -0
- package/dist/services/tool-registry.d.ts.map +1 -0
- package/dist/services/tool-registry.js +139 -0
- package/dist/services/tool-registry.js.map +1 -0
- package/dist/services/user-config.d.ts +14 -0
- package/dist/services/user-config.d.ts.map +1 -1
- package/dist/services/user-config.js +55 -0
- package/dist/services/user-config.js.map +1 -1
- package/dist/services/vcs/types.d.ts +44 -0
- package/dist/services/vcs/types.d.ts.map +1 -1
- package/dist/skills/fleet-sprint-cli/SKILL.md +103 -0
- package/dist/skills/pm/cost.md +3 -0
- package/dist/skills/pm/doer-reviewer-loop.md +3 -0
- package/dist/tools/check-status.d.ts.map +1 -1
- package/dist/tools/check-status.js +8 -2
- package/dist/tools/check-status.js.map +1 -1
- package/dist/tools/child-id-allocator.d.ts +45 -0
- package/dist/tools/child-id-allocator.d.ts.map +1 -0
- package/dist/tools/child-id-allocator.js +62 -0
- package/dist/tools/child-id-allocator.js.map +1 -0
- package/dist/tools/cloud-control.d.ts +2 -2
- package/dist/tools/compose-permissions.d.ts +12 -4
- package/dist/tools/compose-permissions.d.ts.map +1 -1
- package/dist/tools/compose-permissions.js +55 -4
- package/dist/tools/compose-permissions.js.map +1 -1
- package/dist/tools/dolt-push-mutex.d.ts +44 -0
- package/dist/tools/dolt-push-mutex.d.ts.map +1 -0
- package/dist/tools/dolt-push-mutex.js +83 -0
- package/dist/tools/dolt-push-mutex.js.map +1 -0
- package/dist/tools/execute-command.d.ts +11 -1
- package/dist/tools/execute-command.d.ts.map +1 -1
- package/dist/tools/execute-command.js +22 -5
- package/dist/tools/execute-command.js.map +1 -1
- package/dist/tools/execute-prompt.d.ts +72 -12
- package/dist/tools/execute-prompt.d.ts.map +1 -1
- package/dist/tools/execute-prompt.js +962 -71
- package/dist/tools/execute-prompt.js.map +1 -1
- package/dist/tools/get-member-model-pricing.d.ts +21 -0
- package/dist/tools/get-member-model-pricing.d.ts.map +1 -0
- package/dist/tools/get-member-model-pricing.js +29 -0
- package/dist/tools/get-member-model-pricing.js.map +1 -0
- package/dist/tools/list-members.d.ts.map +1 -1
- package/dist/tools/list-members.js +39 -7
- package/dist/tools/list-members.js.map +1 -1
- package/dist/tools/member-detail.d.ts +1 -1
- package/dist/tools/member-detail.d.ts.map +1 -1
- package/dist/tools/member-detail.js +8 -2
- package/dist/tools/member-detail.js.map +1 -1
- package/dist/tools/member-reservation.d.ts +29 -0
- package/dist/tools/member-reservation.d.ts.map +1 -0
- package/dist/tools/member-reservation.js +71 -0
- package/dist/tools/member-reservation.js.map +1 -0
- package/dist/tools/provision-vcs-auth.d.ts +7 -7
- package/dist/tools/provision-vcs-auth.js +1 -1
- package/dist/tools/provision-vcs-auth.js.map +1 -1
- package/dist/tools/register-member.d.ts +17 -5
- package/dist/tools/register-member.d.ts.map +1 -1
- package/dist/tools/register-member.js +144 -5
- package/dist/tools/register-member.js.map +1 -1
- package/dist/tools/remove-member.d.ts +1 -1
- package/dist/tools/report-status.d.ts +33 -0
- package/dist/tools/report-status.d.ts.map +1 -0
- package/dist/tools/report-status.js +44 -0
- package/dist/tools/report-status.js.map +1 -0
- package/dist/tools/respond-to-message.d.ts +21 -0
- package/dist/tools/respond-to-message.d.ts.map +1 -0
- package/dist/tools/respond-to-message.js +23 -0
- package/dist/tools/respond-to-message.js.map +1 -0
- package/dist/tools/send-files.d.ts +3 -0
- package/dist/tools/send-files.d.ts.map +1 -1
- package/dist/tools/send-files.js +88 -8
- package/dist/tools/send-files.js.map +1 -1
- package/dist/tools/send-message.d.ts +27 -0
- package/dist/tools/send-message.d.ts.map +1 -0
- package/dist/tools/send-message.js +46 -0
- package/dist/tools/send-message.js.map +1 -0
- package/dist/tools/shutdown-server.d.ts +2 -0
- package/dist/tools/shutdown-server.d.ts.map +1 -1
- package/dist/tools/shutdown-server.js +19 -0
- package/dist/tools/shutdown-server.js.map +1 -1
- package/dist/tools/update-member.d.ts +8 -8
- package/dist/tools/update-member.d.ts.map +1 -1
- package/dist/tools/update-member.js +11 -1
- package/dist/tools/update-member.js.map +1 -1
- package/dist/types.d.ts +15 -3
- package/dist/types.d.ts.map +1 -1
- package/dist/utils/agent-helpers.d.ts +2 -1
- package/dist/utils/agent-helpers.d.ts.map +1 -1
- package/dist/utils/agent-helpers.js +7 -2
- package/dist/utils/agent-helpers.js.map +1 -1
- package/dist/utils/pid-helpers.d.ts +15 -0
- package/dist/utils/pid-helpers.d.ts.map +1 -1
- package/dist/utils/pid-helpers.js +23 -0
- package/dist/utils/pid-helpers.js.map +1 -1
- package/dist/utils/process-utils.d.ts +3 -0
- package/dist/utils/process-utils.d.ts.map +1 -0
- package/dist/utils/process-utils.js +42 -0
- package/dist/utils/process-utils.js.map +1 -0
- package/dist/utils/prompt-errors.d.ts +2 -1
- package/dist/utils/prompt-errors.d.ts.map +1 -1
- package/dist/utils/prompt-errors.js +13 -3
- package/dist/utils/prompt-errors.js.map +1 -1
- package/dist/utils/workspace-trust.d.ts +24 -0
- package/dist/utils/workspace-trust.d.ts.map +1 -0
- package/dist/utils/workspace-trust.js +35 -0
- package/dist/utils/workspace-trust.js.map +1 -0
- package/dist/workflows/auto-sprint.js +450 -87
- package/package.json +35 -7
- package/packages/apra-fleet-client/package.json +26 -0
- package/packages/apra-fleet-client/src/client/api.mjs +396 -0
- package/packages/apra-fleet-client/src/client/client.mjs +146 -0
- package/packages/apra-fleet-client/src/client/errors.mjs +71 -0
- package/packages/apra-fleet-client/src/client/factory.mjs +65 -0
- package/packages/apra-fleet-client/src/client/server-resolution.mjs +265 -0
- package/packages/apra-fleet-client/src/client/transport.mjs +327 -0
- package/packages/apra-fleet-se/apra-pm/agents/backlog-groomer.md +287 -0
- package/packages/apra-fleet-se/apra-pm/agents/ci-watcher.md +89 -0
- package/packages/apra-fleet-se/apra-pm/agents/deployer.md +139 -0
- package/packages/apra-fleet-se/apra-pm/agents/doer.md +225 -0
- package/packages/apra-fleet-se/apra-pm/agents/harvester.md +169 -0
- package/packages/apra-fleet-se/apra-pm/agents/integ-test-runner.md +316 -0
- package/packages/apra-fleet-se/apra-pm/agents/plan-reviewer.md +188 -0
- package/packages/apra-fleet-se/apra-pm/agents/planner.md +239 -0
- package/packages/apra-fleet-se/apra-pm/agents/regression-test-runner.md +216 -0
- package/packages/apra-fleet-se/apra-pm/agents/reviewer.md +174 -0
- package/packages/apra-fleet-se/apra-pm/agents/schemas/backlog-groomer-input.json +35 -0
- package/packages/apra-fleet-se/apra-pm/agents/schemas/backlog-groomer-output.json +164 -0
- package/packages/apra-fleet-se/apra-pm/agents/schemas/ci-watcher-input.json +26 -0
- package/packages/apra-fleet-se/apra-pm/agents/schemas/ci-watcher-output.json +18 -0
- package/packages/apra-fleet-se/apra-pm/agents/schemas/deployer-input.json +19 -0
- package/packages/apra-fleet-se/apra-pm/agents/schemas/deployer-output.json +17 -0
- package/packages/apra-fleet-se/apra-pm/agents/schemas/doer-input.json +25 -0
- package/packages/apra-fleet-se/apra-pm/agents/schemas/doer-output.json +22 -0
- package/packages/apra-fleet-se/apra-pm/agents/schemas/harvester-input.json +31 -0
- package/packages/apra-fleet-se/apra-pm/agents/schemas/harvester-output.json +18 -0
- package/packages/apra-fleet-se/apra-pm/agents/schemas/integ-test-runner-input.json +20 -0
- package/packages/apra-fleet-se/apra-pm/agents/schemas/integ-test-runner-output.json +61 -0
- package/packages/apra-fleet-se/apra-pm/agents/schemas/plan-reviewer-input.json +15 -0
- package/packages/apra-fleet-se/apra-pm/agents/schemas/plan-reviewer-output.json +30 -0
- package/packages/apra-fleet-se/apra-pm/agents/schemas/regression-test-runner-input.json +15 -0
- package/packages/apra-fleet-se/apra-pm/agents/schemas/regression-test-runner-output.json +50 -0
- package/packages/apra-fleet-se/apra-pm/agents/schemas/reviewer-input.json +25 -0
- package/packages/apra-fleet-se/apra-pm/agents/schemas/reviewer-output.json +38 -0
- package/packages/apra-fleet-se/apra-pm/skills/pm/SKILL.md +402 -0
- package/packages/apra-fleet-se/apra-pm/skills/pm/beads.md +142 -0
- package/packages/apra-fleet-se/apra-pm/skills/pm/cost.md +157 -0
- package/packages/apra-fleet-se/apra-pm/skills/pm/doer-reviewer-loop.md +300 -0
- package/packages/apra-fleet-se/apra-pm/skills/pm/fleet-addendum.md +121 -0
- package/packages/apra-fleet-se/apra-pm/skills/pm/simple-sprint.md +65 -0
- package/packages/apra-fleet-se/apra-pm/skills/pm/sprint.md +327 -0
- package/packages/apra-fleet-se/apra-pm/skills/pm/worktrees.md +92 -0
- package/packages/apra-fleet-se/apra-pm/version.json +3 -0
- package/packages/apra-fleet-se/bin/cli.mjs +857 -0
- package/packages/apra-fleet-se/bin/serve.mjs +383 -0
- package/packages/apra-fleet-se/fleet-sprint/conflict-ladder.mjs +178 -0
- package/packages/apra-fleet-se/fleet-sprint/contracts.mjs +914 -0
- package/packages/apra-fleet-se/fleet-sprint/dispatch-safety-guard.mjs +177 -0
- package/packages/apra-fleet-se/fleet-sprint/docs/README.md +191 -0
- package/packages/apra-fleet-se/fleet-sprint/docs/coach-role-design.md +151 -0
- package/packages/apra-fleet-se/fleet-sprint/docs/dolt-tier2-runbook.md +127 -0
- package/packages/apra-fleet-se/fleet-sprint/docs/fleet-sprint-diagram.md +94 -0
- package/packages/apra-fleet-se/fleet-sprint/docs/stabilization-log.md +1125 -0
- package/packages/apra-fleet-se/fleet-sprint/dolt-literal-guard.mjs +72 -0
- package/packages/apra-fleet-se/fleet-sprint/dolt-recovery-path-b.mjs +339 -0
- package/packages/apra-fleet-se/fleet-sprint/dolt-recovery-tier2.mjs +360 -0
- package/packages/apra-fleet-se/fleet-sprint/dolt-recovery.mjs +363 -0
- package/packages/apra-fleet-se/fleet-sprint/dolt-sync.mjs +1244 -0
- package/packages/apra-fleet-se/fleet-sprint/errors.mjs +572 -0
- package/packages/apra-fleet-se/fleet-sprint/full-db-fetch-guard.mjs +118 -0
- package/packages/apra-fleet-se/fleet-sprint/runner.js +9447 -0
- package/packages/apra-fleet-se/fleet-sprint/skills/fleet-sprint-cli/SKILL.md +103 -0
- package/packages/apra-fleet-se/fleet-sprint/sprint-lock.mjs +140 -0
- package/packages/apra-fleet-se/fleet-sprint/vcs-module.mjs +435 -0
- package/packages/apra-fleet-se/fleet-sprint/vcs-providers/azure-devops.mjs +68 -0
- package/packages/apra-fleet-se/fleet-sprint/vcs-providers/bitbucket.mjs +53 -0
- package/packages/apra-fleet-se/fleet-sprint/vcs-providers/dolt.mjs +146 -0
- package/packages/apra-fleet-se/fleet-sprint/vcs-providers/generic-git.mjs +176 -0
- package/packages/apra-fleet-se/fleet-sprint/vcs-providers/github.mjs +222 -0
- package/packages/apra-fleet-se/fleet-sprint/vcs-providers/index.mjs +218 -0
- package/packages/apra-fleet-se/fleet-sprint/viewer-extensions.mjs +882 -0
- package/packages/apra-fleet-se/package.json +28 -0
- package/packages/apra-fleet-se/workflow.json +1 -0
- package/packages/apra-fleet-workflow/package.json +30 -0
- package/packages/apra-fleet-workflow/src/viewer/command-output-cap.mjs +112 -0
- package/packages/apra-fleet-workflow/src/viewer/debounced-writer.mjs +153 -0
- package/packages/apra-fleet-workflow/src/viewer/html-utils.mjs +31 -0
- package/packages/apra-fleet-workflow/src/viewer/index.mjs +1325 -0
- package/packages/apra-fleet-workflow/src/viewer/lean-state.mjs +237 -0
- package/packages/apra-fleet-workflow/src/viewer/run-state-paths.mjs +130 -0
- package/packages/apra-fleet-workflow/src/viewer/sprint-state-paths.mjs +29 -0
- package/packages/apra-fleet-workflow/src/workflow/engine.mjs +200 -0
- package/packages/apra-fleet-workflow/src/workflow/errors.mjs +146 -0
- package/packages/apra-fleet-workflow/src/workflow/index.mjs +1723 -0
- package/packages/apra-fleet-workflow/src/workflow/journal.mjs +291 -0
- package/packages/apra-fleet-workflow/src/workflow/pricing.mjs +99 -0
- package/packages/apra-fleet-workflow/src/workflow/vetting.mjs +111 -0
- package/scripts/agy-transcript-reader.js +1 -0
- package/scripts/fleet-statusline.sh +17 -13
- package/skills/fleet/SKILL.md +37 -2
- package/skills/fleet/beads.md +14 -7
- package/skills/fleet/permissions.md +50 -0
- package/skills/fleet/profiles/base-dev.json +1 -0
- package/skills/fleet/profiles/base-reviewer.json +1 -0
- package/skills/fleet/profiles/node.json +2 -2
- package/version.json +1 -1
- package/scripts/agy-settings-merge.js +0 -27
- /package/{dist → packages/apra-fleet-se/apra-pm}/agents/_shared/GRAPH-SEMANTICS.md +0 -0
|
@@ -0,0 +1,1125 @@
|
|
|
1
|
+
# Auto-Sprint Stabilization Log
|
|
2
|
+
|
|
3
|
+
Running log of issues found while driving the real `apra-fleet-eft` sprint
|
|
4
|
+
(member: fleet-rev, branch `auto-sprint/eft-service`, base `feat/fleet-reorg`)
|
|
5
|
+
per `docs/README.md`, and the fixes made for each. Newest entries at the
|
|
6
|
+
bottom. Each entry: symptom -> root cause -> fix -> evidence.
|
|
7
|
+
|
|
8
|
+
## Loop iteration 1 (2026-07-19)
|
|
9
|
+
|
|
10
|
+
### Issue 1: Transport closed kills the whole sprint at start+~304s, every run
|
|
11
|
+
|
|
12
|
+
- **Symptom**: every CLI sprint run died ~5 minutes in with
|
|
13
|
+
`FleetTransportError: Transport closed`, regardless of which dispatch
|
|
14
|
+
happened to be in flight at that moment (`sprint-logs/auto-sprint-relaunch2.log`,
|
|
15
|
+
`relaunch3`, `relaunch4`).
|
|
16
|
+
- **Root cause** (deterministic, confirmed by timestamps): the
|
|
17
|
+
`StreamableHttpTransport` persistent SSE GET stream is normally silent
|
|
18
|
+
(JSON-RPC responses arrive over each POST's own SSE response). Node's
|
|
19
|
+
built-in fetch (undici) enforces a default ~300s idle `bodyTimeout` on
|
|
20
|
+
response bodies, so the single-shot GET stream died at start+~304s
|
|
21
|
+
(fleet server log: stream opened 23:37:40, `Transport closed` 23:42:44).
|
|
22
|
+
`transport.mjs`'s `finally` block then emitted `close`, and
|
|
23
|
+
`McpClient`'s close handler rejected EVERY in-flight request.
|
|
24
|
+
- **Fix**: `packages/apra-fleet-client/src/client/transport.mjs` -- the
|
|
25
|
+
persistent GET stream is now a self-reconnecting background loop. An
|
|
26
|
+
idle-death is treated as an expected, recoverable event (quiet reopen with
|
|
27
|
+
capped backoff); `close` only surfaces on deliberate `stop()` or 5
|
|
28
|
+
consecutive reconnect failures. Regression test:
|
|
29
|
+
`packages/apra-fleet-client/test/transport-reconnect.test.mjs` (real local
|
|
30
|
+
SSE server, kills the first stream, asserts a second GET arrives and no
|
|
31
|
+
`close`/`error` fires).
|
|
32
|
+
|
|
33
|
+
### Issue 2: busy-lock race -- orphaned remote session holds the member lock for minutes
|
|
34
|
+
|
|
35
|
+
- **Symptom**: after any client-side abandonment of a dispatch (e.g. Issue
|
|
36
|
+
1's transport close), every subsequent dispatch to that member failed with
|
|
37
|
+
`execute_prompt is already running for "fleet-rev"` until the sprint
|
|
38
|
+
aborted. A blind timed backoff (up to ~110s) lost the race every time.
|
|
39
|
+
- **Root cause**: the client-side failure does not cancel the REMOTE
|
|
40
|
+
session -- the member's CLI keeps running (fleet server log: the orphaned
|
|
41
|
+
plan-reviewer repair session ran 235s after the client gave up at 4.6s,
|
|
42
|
+
completing with a perfectly good 12,976-token response that was
|
|
43
|
+
discarded) and holds the server's one-dispatch-per-member lock the whole
|
|
44
|
+
time. Busy is transient-but-SLOW: minutes, not seconds.
|
|
45
|
+
- **Fix**: `packages/apra-fleet-workflow/src/workflow/index.mjs` -- agent()
|
|
46
|
+
now busy-waits by default: on a `busy` dispatch rejection it polls
|
|
47
|
+
(cheap, side-effect-free re-dispatch) every `busyPollMs` (default 15s) up
|
|
48
|
+
to `busyWaitMs` (default 10 min) before surfacing the busy
|
|
49
|
+
AgentDispatchError. Tests:
|
|
50
|
+
`packages/apra-fleet-workflow/test/apra-fleet-workflow-busy-and-empty.test.mjs`.
|
|
51
|
+
(The runner-level Planner timed-backoff loop from earlier the same day is
|
|
52
|
+
retained as an outer safety net but should rarely trigger now.)
|
|
53
|
+
|
|
54
|
+
### Issue 3: empty execute_prompt response misreported as "LLM returned invalid JSON" (apra-fleet-eft.14)
|
|
55
|
+
|
|
56
|
+
- **Symptom**:
|
|
57
|
+
`[Agent API Error] Schema-invalid output ... Unexpected token '\ud83d', "[clipboard-emoji] Respons"... is not valid JSON`
|
|
58
|
+
-- looked like a Unicode/JSON bug. It is not: the emoji is just the first
|
|
59
|
+
char the JSON parser saw.
|
|
60
|
+
- **Root cause**: the fleet server returned success (exit 0, no isError)
|
|
61
|
+
whose text was EXACTLY the display wrapper `"[clipboard-emoji] Response from
|
|
62
|
+
fleet-rev:\n\n"` -- 29 chars, empty parsed result, no Tokens/session
|
|
63
|
+
footer (the server-side log line for that dispatch also lacks its usual
|
|
64
|
+
`in=`/`out=` token counts: `exit=0 elapsed=65241ms`). The member CLI
|
|
65
|
+
produced no parseable output; the wrapper-only text then went to schema
|
|
66
|
+
extraction, which correctly found no JSON and misclassified the failure.
|
|
67
|
+
This is bead apra-fleet-eft.14 (first seen on the Planner, now confirmed
|
|
68
|
+
on the plan-reviewer). The underlying server-side parse gap (provider CLI
|
|
69
|
+
exits 0 with nothing parsed) is still open -- tracked in eft.14.
|
|
70
|
+
- **Fix (client-side resilience)**:
|
|
71
|
+
`packages/apra-fleet-workflow/src/workflow/index.mjs` -- agent() now
|
|
72
|
+
detects a wrapper-only response (header + optional Tokens/session footer,
|
|
73
|
+
nothing else) and throws a typed
|
|
74
|
+
`AgentDispatchError` with `reason: 'empty_response'` instead of feeding
|
|
75
|
+
it to schema extraction (or returning it as a garbage success for
|
|
76
|
+
no-schema calls). Every runner dispatch site already catches
|
|
77
|
+
AgentDispatchError and degrades the round gracefully. Tests: same new
|
|
78
|
+
test file as Issue 2.
|
|
79
|
+
|
|
80
|
+
### Issue 4: transport-closed rejections were untyped
|
|
81
|
+
|
|
82
|
+
- **Symptom**: downstream had to sniff the string 'Transport closed'.
|
|
83
|
+
- **Fix**: new `TransportClosedError` (`code: 'TRANSPORT_CLOSED'`) in
|
|
84
|
+
`packages/apra-fleet-client/src/client/errors.mjs`, used by McpClient's
|
|
85
|
+
close handler. Convention documented in that file: new
|
|
86
|
+
execute_prompt/execute_command failure kinds get a typed class there
|
|
87
|
+
(transport/request-level) or an AgentDispatchError `reason` code
|
|
88
|
+
(dispatch-level via structuredContent) -- never a bare Error message to
|
|
89
|
+
sniff.
|
|
90
|
+
|
|
91
|
+
### Earlier same-day fixes (pre-loop, first crash analysis)
|
|
92
|
+
|
|
93
|
+
- **FleetTransportError uncaught at schema dispatch sites**: every
|
|
94
|
+
combined-catch site (Reviewer, Plan Reviewer, Streak Assignment,
|
|
95
|
+
Deployer, Integ Test Runner, Final Review x2, Harvester) only handled
|
|
96
|
+
AgentOutputError/AgentDispatchError; a raw transport failure propagated
|
|
97
|
+
and killed the sprint. All sites now also catch FleetTransportError and
|
|
98
|
+
degrade to the same failed-round handling. (runner.js, commit 3daa8fe)
|
|
99
|
+
- **finalizeAbort() rev-list against a nonexistent local base ref**: the
|
|
100
|
+
abort path ran `git rev-list --count <base>..<branch>` assuming the raw
|
|
101
|
+
`--base` value resolves locally on the member; exit 128 ("unknown
|
|
102
|
+
revision") made finalizeAbort itself throw and the [ABORTED] PR was never
|
|
103
|
+
raised. Now fetches the base branch and diffs against
|
|
104
|
+
`origin/<baseBranch>`. (runner.js, commit 3daa8fe)
|
|
105
|
+
- **Planner dispatch single blind retry**: replaced with a bounded
|
|
106
|
+
retry-with-backoff loop (commits 96eaf25, fbb7619); superseded in
|
|
107
|
+
practice by Issue 2's busy-wait but retained as an outer net.
|
|
108
|
+
- **Stale test assertions (4 failures)**: roleMap tests now exclude
|
|
109
|
+
per-member `bd dolt pull/push` sync brackets from the
|
|
110
|
+
orchestrator-routing assertion (verified correct via live dispatch-log
|
|
111
|
+
probe); R12 no-auto-merge check excludes the safe `git merge --ff-only`
|
|
112
|
+
self-sync; run1/run2/multidoer bd recordings re-recorded via
|
|
113
|
+
`npm run test:record`. (commit 827dabe)
|
|
114
|
+
|
|
115
|
+
### Issue 5: G-pull fetch hard-fails on a brand-new (not-yet-pushed) sprint branch
|
|
116
|
+
|
|
117
|
+
- **Symptom**: `mock-sprint-ensure-branch-fetch-failure.test.mjs` failing
|
|
118
|
+
with `[Sync] G-pull fetch failed ... couldn't find remote ref
|
|
119
|
+
auto-sprint/mock-branchnotexist` (110s of transient retries first).
|
|
120
|
+
- **Root cause**: syncMemberBefore() treated EVERY fetch failure as an
|
|
121
|
+
error, but a sprint branch just created locally from base (first G-push
|
|
122
|
+
hasn't happened) legitimately does not exist on the remote yet -- there
|
|
123
|
+
is nothing to pull, which is a no-op, not a failure. Any real first-cycle
|
|
124
|
+
sprint on a fresh branch would have died on its first sync bracket.
|
|
125
|
+
- **Fix**: runner.js syncMemberBefore() -- the exact `couldn't find remote
|
|
126
|
+
ref` git message (and only that message, mirroring Ensure Sprint Branch's
|
|
127
|
+
own fetch-fallback rationale) now skips the pull half of the bracket with
|
|
128
|
+
a log line instead of throwing.
|
|
129
|
+
|
|
130
|
+
### Issue 6: injected git-push failure now surfaces as GitSyncError, not CommandError
|
|
131
|
+
|
|
132
|
+
- **Symptom**: `mock-sprint-finalization-git-push-failure.test.mjs`
|
|
133
|
+
asserting `CommandError` failed with `GitSyncError`.
|
|
134
|
+
- **Root cause**: not a bug -- an architecture shift. Before the sync
|
|
135
|
+
brackets, the first `git push` a sprint issued was the finalization
|
|
136
|
+
publish (plain command() -> CommandError). The brackets now G-push after
|
|
137
|
+
every code-writing dispatch, so an injected `/^git push/` failure is hit
|
|
138
|
+
first by a bracket push, surfacing as its typed GitSyncError.
|
|
139
|
+
- **Fix**: test updated to accept either typed error (both are
|
|
140
|
+
never-swallowed surfaces of the same failure), with the
|
|
141
|
+
underlying-git-text assertion retained so a silent swallow still fails.
|
|
142
|
+
|
|
143
|
+
### Issue 7: recording/snapshot drift from the sync-bracket work
|
|
144
|
+
|
|
145
|
+
- `apra-fleet-mock-sprint-abortprfail.jsonl` was a crashed mid-scenario
|
|
146
|
+
capture (exitCode null entry) -> re-recorded via
|
|
147
|
+
`node scripts/run-tests.mjs record test/mock-sprint-abort-pr.test.mjs`.
|
|
148
|
+
- Both golden transcripts regenerated (UPDATE_GOLDEN=1) after verifying
|
|
149
|
+
the diff is exactly the expected shape: sync-bracket commands (git
|
|
150
|
+
fetch / merge --ff-only / bd dolt pull before each dispatch, bd dolt
|
|
151
|
+
push after) interleaving an otherwise-identical dispatch sequence.
|
|
152
|
+
|
|
153
|
+
**Iteration 1 result**: full apra-fleet-se suite 500/500 green (was 10
|
|
154
|
+
failing). apra-fleet-client 21/21 + new reconnect test. apra-fleet-workflow
|
|
155
|
+
core suites green + 8 new busy/empty tests; 2 pre-existing eft-WIP
|
|
156
|
+
conditions confirmed NOT caused by this work (debounced-writer 3 failures,
|
|
157
|
+
sprint-state file-level exit hang -- both reproduce with all stabilization
|
|
158
|
+
changes stashed; they belong to the in-flight eft.2.1 feature work).
|
|
159
|
+
|
|
160
|
+
## Loop iteration 2 (2026-07-19)
|
|
161
|
+
|
|
162
|
+
### Issue 8: POST-stream idle timeout kills every >300s-silent dispatch ("terminated")
|
|
163
|
+
|
|
164
|
+
- **Symptom** (run 5, `sprint-logs/auto-sprint-relaunch5.log`): long
|
|
165
|
+
dispatches failed with `[Agent API Error] terminated`, then the retry hit
|
|
166
|
+
`busy`, busy-waited out the orphan, re-dispatched -- and the redo ALSO
|
|
167
|
+
died with `terminated`. The sprint survived (Issues 1-3 fixes all held)
|
|
168
|
+
but ground in a terminated -> retry -> busy-wait -> duplicate-dispatch
|
|
169
|
+
loop, doing each long planner run twice and discarding the results.
|
|
170
|
+
- **Root cause**: the Issue 1 fix covered the persistent GET stream, but
|
|
171
|
+
each POST's own SSE response stream has the same undici ~300s idle
|
|
172
|
+
bodyTimeout. `claude -p` prints nothing until the turn completes, so any
|
|
173
|
+
dispatch quieter than 300s+ loses its response stream mid-flight
|
|
174
|
+
("terminated" is undici's body-timeout error message). Server log
|
|
175
|
+
evidence: planner R2 pid 15590 ran 675s (00:44:40 -> 00:55:55, exit=0
|
|
176
|
+
out=16285) -- client stream killed at ~00:49:40, full result discarded,
|
|
177
|
+
duplicate planner pid 18865 dispatched 00:55:55 by the busy-waiting
|
|
178
|
+
retry, which then also exceeded 300s and got terminated in turn.
|
|
179
|
+
- **Fix**: `packages/apra-fleet-client` now depends on `undici` explicitly
|
|
180
|
+
and `transport.mjs` routes ALL StreamableHttpTransport fetches (init
|
|
181
|
+
POST, persistent GET, send POST) through undici's fetch with a shared
|
|
182
|
+
`Agent({ headersTimeout: 0, bodyTimeout: 0 })` dispatcher. No unbounded
|
|
183
|
+
hangs result: McpClient's per-request timeout still bounds every
|
|
184
|
+
JSON-RPC call, and the GET stream keeps its reconnect loop. Guard test:
|
|
185
|
+
`packages/apra-fleet-client/test/transport-idle-timeout-guard.test.mjs`
|
|
186
|
+
(asserts the wiring at source level -- a behavioral test would need a
|
|
187
|
+
>300s silent stream; the existing reconnect test exercises the undici
|
|
188
|
+
fetch path against a real server).
|
|
189
|
+
- **Decision note**: server-side SSE keepalives on POST responses would fix
|
|
190
|
+
this for ALL clients and remain a good hardening follow-up, but need a
|
|
191
|
+
fleet-server rebuild+redeploy; the client-side dispatcher fix is
|
|
192
|
+
deterministic and deploy-free, so it ships first.
|
|
193
|
+
|
|
194
|
+
### Server-side eft.14 classification (committed, deploy deferred)
|
|
195
|
+
|
|
196
|
+
- `src/tools/execute-prompt.ts` now classifies an exit-0-with-empty-result
|
|
197
|
+
dispatch as a typed failure (`structuredContent: { isError: true,
|
|
198
|
+
reason: 'empty_response' }` + stderr tail in the text) instead of
|
|
199
|
+
returning a bare display wrapper as success. Tests added in
|
|
200
|
+
`tests/execute-prompt.test.ts`; full root vitest suite green (2318
|
|
201
|
+
passed). Deployment needs a fleet-server rebuild+reinstall+restart, which
|
|
202
|
+
kills live member sessions -- deferred to a natural stop; the workflow
|
|
203
|
+
layer's Issue 3 detection handles it in the meantime.
|
|
204
|
+
|
|
205
|
+
## Loop iteration 3 (2026-07-19)
|
|
206
|
+
|
|
207
|
+
Run 6 (the first on the Issue 8 transport fix) got dramatically further
|
|
208
|
+
than any prior run: Plan C1 approved in round 1 (the previously
|
|
209
|
+
empty-response-prone plan-reviewer completed normally), the full Develop
|
|
210
|
+
C1 R1 phase ran multi-minute doer streaks back-to-back with ZERO transport
|
|
211
|
+
failures (beads closed: eft.6.4, 9.1, 4.6, more; one streak correctly
|
|
212
|
+
degraded on a VERIFY/still-open mismatch), and the sprint reached Review
|
|
213
|
+
C1 R1 -- where it found the next two systemic issues:
|
|
214
|
+
|
|
215
|
+
### Issue 9: reviewer budgets undersized for a full-cycle review + infra failures tripped the contract-violation guard
|
|
216
|
+
|
|
217
|
+
- **Symptom** (run 6): the reviewer ran out of turns (num_turns=51 after
|
|
218
|
+
~12 min of legitimate review work, typed max_turns_exhausted); the retry
|
|
219
|
+
then hit the 930s client-side request timeout. BOTH synthesized
|
|
220
|
+
CHANGES_NEEDED fallback verdicts had empty reopenIds/newTasks, so
|
|
221
|
+
isReviewerContractViolation() -- designed to catch a GENUINE
|
|
222
|
+
self-contradictory LLM verdict -- counted two "contract violations" and
|
|
223
|
+
aborted the sprint with ReviewerContractViolationError.
|
|
224
|
+
- **Root cause**: two distinct gaps. (a) The reviewer dispatch had no
|
|
225
|
+
explicit max_turns and no resume path, so a big-cycle review
|
|
226
|
+
deterministically dies at the fleet default and a fresh retry re-dies
|
|
227
|
+
the same way. (b) dispatchReview() synthesized infrastructure-failure
|
|
228
|
+
verdicts in the same shape as LLM verdicts, so the contract-violation
|
|
229
|
+
guard could not tell them apart.
|
|
230
|
+
- **Fix** (runner.js dispatchReview): explicit
|
|
231
|
+
`max_turns: BASE_REVIEWER_MAX_TURNS (60)`; on max_turns_exhausted the
|
|
232
|
+
SAME session is resumed once with a doubled budget (mirrors
|
|
233
|
+
dispatchDoerResume -- the session already holds the review context).
|
|
234
|
+
Synthesized failure verdicts now carry `dispatchFailed: true` and are
|
|
235
|
+
returned as a DEGRADED round (counting toward the bounded stall budget
|
|
236
|
+
like every other role) after one infrastructure retry -- they can no
|
|
237
|
+
longer throw ReviewerContractViolationError, which is reserved for
|
|
238
|
+
genuine schema-valid self-contradicting LLM verdicts. Structural test
|
|
239
|
+
baselines bumped (11 agent sites / 10 withGitSync brackets), reviewer
|
|
240
|
+
scenario recordings re-recorded; se suite 602/602.
|
|
241
|
+
|
|
242
|
+
### Issue 10: agent() silently dropped max_total_s from the dispatch payload
|
|
243
|
+
|
|
244
|
+
- **Symptom** (run 6): the reviewer retry timed out client-side at exactly
|
|
245
|
+
930000ms despite the runner passing `max_total_s: 3600` (which
|
|
246
|
+
deriveTimeoutMs prefers -- it should have produced a 3630s request
|
|
247
|
+
timeout).
|
|
248
|
+
- **Root cause**: the payload object built inside
|
|
249
|
+
`packages/apra-fleet-workflow/src/workflow/index.mjs` agent() forwarded
|
|
250
|
+
timeout_s/max_turns/etc. but NOT max_total_s, so (a) the server never
|
|
251
|
+
received the hard wall-clock ceiling for ANY workflow dispatch, and (b)
|
|
252
|
+
the derived client timeout always fell back to timeout_s+30s = 930s,
|
|
253
|
+
killing any legitimately-slow dispatch at ~15.5 min while the remote
|
|
254
|
+
session kept running (orphan + busy lock).
|
|
255
|
+
- **Fix**: `max_total_s: opts.max_total_s` added to the payload, with
|
|
256
|
+
JSDoc documenting both server- and client-side effects.
|
|
257
|
+
|
|
258
|
+
### Issue 11: dirty member working tree kills the next sprint at Setup
|
|
259
|
+
|
|
260
|
+
- **Symptom** (run 7): `Ensure Sprint Branch` died immediately with
|
|
261
|
+
`error: Your local changes to the following files would be overwritten
|
|
262
|
+
by checkout` -- fleet-rev had an uncommitted runner.js modification left
|
|
263
|
+
by an interrupted dispatch from run 5/6.
|
|
264
|
+
- **Root cause**: any infrastructure-killed dispatch (transport drop,
|
|
265
|
+
timeout, stop_prompt) predictably leaves the member's tree dirty with
|
|
266
|
+
whatever the agent had in flight; `git checkout -B` refuses to proceed.
|
|
267
|
+
A stabilized sprint loop guarantees such orphans will keep happening, so
|
|
268
|
+
Setup must expect them.
|
|
269
|
+
- **Fix** (runner.js Ensure Sprint Branch): the checkout is now failSoft;
|
|
270
|
+
on the specific "would be overwritten" failure the orphaned WIP is
|
|
271
|
+
preserved in a named stash (`git stash push -u -m "auto-sprint[branch]
|
|
272
|
+
auto-stash of orphaned WIP..."`) and the checkout retried once. Any
|
|
273
|
+
other checkout failure still aborts loudly. Happy path (clean tree)
|
|
274
|
+
issues no extra commands. The run-7 orphan itself (a one-line eft.8.x
|
|
275
|
+
error-shape WIP) was inspected and hand-stashed on fleet-rev with a
|
|
276
|
+
descriptive message; its bead is still open so a future streak redoes
|
|
277
|
+
it properly. Command-count baseline 26 -> 28.
|
|
278
|
+
|
|
279
|
+
### Feature bug found by the loop, routed to the sprint backlog
|
|
280
|
+
|
|
281
|
+
- The run-6 doer's new supervisor-lifecycle test (eft.4.6, written and
|
|
282
|
+
verified on macOS fleet-rev) fails on Windows: live-child re-adoption
|
|
283
|
+
recovers the port from the process command line via a ps reader that
|
|
284
|
+
returns nothing usable on Windows. Filed as a P1 bug bead under
|
|
285
|
+
apra-fleet-eft.4 (2026-07-19) so the sprint's own doers fix it --
|
|
286
|
+
stabilization-loop scope stays on the sprint machinery, not the eft
|
|
287
|
+
feature. se suite otherwise green (667 pass / 2 fail, both this bug).
|
|
288
|
+
|
|
289
|
+
## Loop iteration 4 (2026-07-19)
|
|
290
|
+
|
|
291
|
+
Run 8 was the best run yet: TWO full cycles of plan -> develop -> review,
|
|
292
|
+
with the plan approved first-round both times, every reviewer round up to
|
|
293
|
+
C2 returning a genuine actionable verdict (real catches: a red test suite
|
|
294
|
+
from a doer regression, uncommitted freshest work, a stale guard-test
|
|
295
|
+
baseline), the doer resume path recovering every turn-limit exhaustion,
|
|
296
|
+
and Cycle 1 evaluation rolling into Cycle 2 cleanly. It died in Review C2
|
|
297
|
+
R1 on two NEW issue types:
|
|
298
|
+
|
|
299
|
+
### Issue 12: server inactivity timeout (900s) makes max_total_s unreachable for silent CLIs
|
|
300
|
+
|
|
301
|
+
- **Symptom** (run 8): the C2 reviewer was killed with "Command timed out
|
|
302
|
+
after 900000ms of inactivity" despite max_total_s: 3600.
|
|
303
|
+
- **Root cause**: `claude -p` prints nothing until the turn completes, so
|
|
304
|
+
server-side INACTIVITY == total runtime for every agent dispatch; the
|
|
305
|
+
900s inactivity timer always fires before the 3600s wall-clock ceiling
|
|
306
|
+
can matter. Earlier reviews survived only by finishing under 15 min.
|
|
307
|
+
- **Fix**: every agent dispatch site in runner.js now sets
|
|
308
|
+
`timeout_s: 3600` equal to its `max_total_s: 3600` -- the hard ceiling
|
|
309
|
+
is the real limit; inactivity adds nothing for a silent-until-done CLI.
|
|
310
|
+
(Also: BASE_DOER_MAX_TURNS 50 -> 100, since in run 8 EVERY streak --
|
|
311
|
+
even single-bead ones -- exhausted 50 turns and paid a resume
|
|
312
|
+
round-trip; resumes now escalate 200 -> 400. Assertion ladder updated in
|
|
313
|
+
mock-sprint-doer-max-turns.test.mjs.)
|
|
314
|
+
|
|
315
|
+
### Issue 13: sporadic client 'fetch failed' + sync-bracket failures are sprint-fatal at the review site
|
|
316
|
+
|
|
317
|
+
- **Symptom** (run 8): sporadic `[Command API Error] fetch failed`
|
|
318
|
+
throughout the run (dashboard badge computation, a doer bracket, and
|
|
319
|
+
finally the review-retry G-pull). The last one surfaced as GitSyncError
|
|
320
|
+
from dispatchReview's sync bracket -- which the reviewer catch did not
|
|
321
|
+
handle -- and killed the sprint. The fleet server was healthy the whole
|
|
322
|
+
time.
|
|
323
|
+
- **Root cause**: three stacked gaps. (a) The pooled undici dispatcher
|
|
324
|
+
(Issue 8 fix) can reuse a socket the server just decided to close
|
|
325
|
+
(Node http default keepAliveTimeout 5000ms) -- the request then fails at
|
|
326
|
+
the socket level as 'fetch failed'. (b) classifyGitFailure() rated
|
|
327
|
+
"Transport failure while executing command: fetch failed" as 'unknown'
|
|
328
|
+
(never retried) instead of transient. (c) dispatchReview degraded
|
|
329
|
+
dispatch/transport errors but not sync-bracket GitSyncError/
|
|
330
|
+
DoltSyncError.
|
|
331
|
+
- **Fix**: (a) transport.mjs -- dispatcher keepAliveTimeout: 4000 (below
|
|
332
|
+
the server's 5000ms) plus a bounded connection-level retry in send()
|
|
333
|
+
(2 retries, 500ms/2s; only for rejections where no response was
|
|
334
|
+
received, so a re-send cannot double-execute). (b) the transport-failure
|
|
335
|
+
strings added to GIT_TRANSIENT_PATTERNS so runGitStep retries them.
|
|
336
|
+
(c) dispatchReview's catch now also degrades GitSyncError/DoltSyncError
|
|
337
|
+
(dispatchFailed round); GitDivergedError/DoltDivergedError still
|
|
338
|
+
propagate -- real divergence is an integrity problem, not a blip.
|
|
339
|
+
|
|
340
|
+
Rebased over 10 doer commits from run 8 (including the doer's own
|
|
341
|
+
eft.9.7 claim-loop restructure of dispatchDoer -- conflict resolved
|
|
342
|
+
keeping the doer's structure with the new timeouts applied). se suite
|
|
343
|
+
759 pass / 2 fail (both the beaded Windows lifecycle bug); client 22/22.
|
|
344
|
+
|
|
345
|
+
## Loop iteration 5 (2026-07-19)
|
|
346
|
+
|
|
347
|
+
Run 9 aborted at planning (SprintPlanRejectedError, 3 rounds): R1-R2
|
|
348
|
+
flagged all-children-closed features as undecomposed (the planner resolved
|
|
349
|
+
this itself by R3 -- no runner change needed), and R3's sole blocker was
|
|
350
|
+
an operator-injected bug bead (eft.4.7) missing the `model` metadata the
|
|
351
|
+
plan contract requires. Operator lesson recorded (bead injections must
|
|
352
|
+
carry model metadata and be created member-side + dolt-pushed; the
|
|
353
|
+
orchestrator's local clone is stale/conflicted outside sprint brackets).
|
|
354
|
+
Run 10 relaunched after fixing the metadata; its plan R1 caught the run-9
|
|
355
|
+
planner's duplicate re-creations of closed eft.13 work and R2 approved a
|
|
356
|
+
deduplicated DAG -- the plan gate working as designed.
|
|
357
|
+
|
|
358
|
+
### Final Review hardening (turn budget, resume, feedback-to-beads)
|
|
359
|
+
|
|
360
|
+
- **Gap 1**: dispatchFinalReview had no explicit max_turns and no resume
|
|
361
|
+
path -- an epic-wide review would deterministically die at the fleet
|
|
362
|
+
default (50 turns) twice (the retry restarts from scratch) and flip the
|
|
363
|
+
sprint to a FAIL whose notes carry only the error text.
|
|
364
|
+
- **Gap 2**: the finalVerdict schema was {verdict, notes} only; a FAIL's
|
|
365
|
+
actionable findings reached the PR body and analysis doc but NEVER
|
|
366
|
+
beads -- invisible to the next sprint's planner, which reads only
|
|
367
|
+
beads.
|
|
368
|
+
- **Fix** (runner.js + contracts.mjs): FINAL_REVIEW_MAX_TURNS = 60 with
|
|
369
|
+
one same-session resume at 120 turns (mirrors doer/reviewer);
|
|
370
|
+
finalVerdict gains optional newTasks (same item shape as
|
|
371
|
+
reviewerVerdict), the prompt instructs findings-as-newTasks on FAIL,
|
|
372
|
+
and the orchestrator applies them via the same validateNewTask
|
|
373
|
+
allowlist + child-id allocator + D-push path as per-round reviewer
|
|
374
|
+
newTasks. Structural baselines 12 agent sites / 11 brackets; goldens
|
|
375
|
+
regenerated (diff = the Final Review prompt/schema line only); se
|
|
376
|
+
suite 759 pass / 2 fail (the beaded Windows lifecycle bug).
|
|
377
|
+
|
|
378
|
+
### Issue 15: streak-assignment candidates with shortened bead ids silently lose the whole grouping
|
|
379
|
+
|
|
380
|
+
- **Symptom** (run 8, Develop C2 R1; operator-reported): the streak
|
|
381
|
+
assignment LLM returned suffix-stripped ids
|
|
382
|
+
(`{"streaks":[["8.4","8.6"],["9.3","9.7"],...]}` for
|
|
383
|
+
apra-fleet-eft.8.4 etc.). selectStreaks() correctly rejected the
|
|
384
|
+
candidate (`unknown/non-ready bead id '8.4'`) and fell back to
|
|
385
|
+
one-bead-per-streak -- safe, but the model's sequencing intent
|
|
386
|
+
(8.4+8.6 same-doer, 9.3+9.7 same-doer) was silently discarded. On a
|
|
387
|
+
multi-doer fleet that fallback would PARALLELIZE beads the model said
|
|
388
|
+
must run sequentially.
|
|
389
|
+
- **Design stance (per operator)**: bead ids from different scopes need
|
|
390
|
+
not share any prefix, so suffix-matching/prefix-inference repair is
|
|
391
|
+
forbidden -- a shortened id is genuinely ambiguous and must stay
|
|
392
|
+
rejected. The fix is to stop the model shortening ids and to recover
|
|
393
|
+
via RE-ASK, never via guessing.
|
|
394
|
+
- **Fix** (runner.js): (a) buildStreakAssignmentPrompt now demands
|
|
395
|
+
verbatim ids (full prefix, character for character, explicit warning
|
|
396
|
+
that shortened ids are rejected); (b) one bounded semantic-repair
|
|
397
|
+
re-ask -- when a schema-VALID candidate fails selectStreaks(), the
|
|
398
|
+
agent is re-asked once with the exact rejection reason before the
|
|
399
|
+
one-bead-per-streak fallback engages. Structural baselines: 13 agent
|
|
400
|
+
sites; the repair re-ask is the second documented no-bracket exemption
|
|
401
|
+
(same pure-compute grouping task as the first).
|
|
402
|
+
|
|
403
|
+
### Issue 14: pending-closure features misread as "undecomposed" burn entire plan phases
|
|
404
|
+
|
|
405
|
+
- **Symptom**: run 9 (cycle 1) and run 10 (cycle 2) each exhausted all 3
|
|
406
|
+
planning rounds -- SprintPlanRejectedError -- because plan-reviewer
|
|
407
|
+
dispatches flagged features whose children are ALL CLOSED (eft.1/3/4/5/
|
|
408
|
+
7/8 by run 10) as "undecomposed, zero child tasks, no model metadata"
|
|
409
|
+
and demanded re-decomposition. The planner cannot sensibly satisfy that
|
|
410
|
+
without recreating closed work -- run 9's planner did exactly that, and
|
|
411
|
+
the NEXT review round correctly rejected the duplicates (thrash both
|
|
412
|
+
ways).
|
|
413
|
+
- **Root cause**: verdict inconsistency across reviewer dispatches on the
|
|
414
|
+
same DAG state. The reviewer's child queries default to OPEN issues, so
|
|
415
|
+
an all-children-closed feature looks childless; some dispatches
|
|
416
|
+
correctly read it as "eligible for closure, non-blocking" (run 8 C2),
|
|
417
|
+
others fail hard criteria on it. Feature closure belongs to the
|
|
418
|
+
integ-test phase, so mid-sprint this state is NORMAL and long-lived.
|
|
419
|
+
- **Fix** (runner.js): pin the interpretation in both prompts.
|
|
420
|
+
buildPlanReviewerPrompt now instructs: check closed children (bd list
|
|
421
|
+
--all) before flagging a feature undecomposed; all-children-closed =
|
|
422
|
+
pending feature-closure = non-blocking note only, never a criteria
|
|
423
|
+
failure, never a re-decomposition demand. buildPlannerPrompt's
|
|
424
|
+
re-planning preamble gets the symmetric guard: leave such features
|
|
425
|
+
alone even if review feedback appears to demand decomposition.
|
|
426
|
+
|
|
427
|
+
### Issue 16: closed beads invisible in the viewer sprint tree
|
|
428
|
+
|
|
429
|
+
- **Symptom** (operator-reported): the sprint viewer showed only open
|
|
430
|
+
items under the sprint tree, making progress impossible to judge -- by
|
|
431
|
+
run 11 most eft work was CLOSED, so the tree looked nearly empty.
|
|
432
|
+
- **Root cause**: updateDashboard fetched beads via bdListScoped('--json').
|
|
433
|
+
Any non-empty argument string routes to a second plain `bd list`
|
|
434
|
+
invocation, which defaults to open/in_progress only. The no-args path
|
|
435
|
+
bdListScoped('') reuses the shared `bd list --all --limit 0 --json`
|
|
436
|
+
fetch, which includes closed beads.
|
|
437
|
+
- **Fix** (runner.js): updateDashboard now calls bdListScoped('') so the
|
|
438
|
+
dashboard payload carries the full DAG including closed items. Golden
|
|
439
|
+
transcripts regenerated (the dashboard section of every golden changed).
|
|
440
|
+
|
|
441
|
+
### Issue 17: reviewer pins CHANGES_NEEDED by reopening deferred P3 features
|
|
442
|
+
|
|
443
|
+
- **Symptom**: run 11 cycles 2-3 -- reviews returned CHANGES_NEEDED with
|
|
444
|
+
reopenIds ['apra-fleet-eft.10','apra-fleet-eft.11' (,'.12')] round
|
|
445
|
+
after round. Those are P3 features, deferred by design under the sprint
|
|
446
|
+
goal constraint (P1/P2). The completion gate needs zero open
|
|
447
|
+
goal-priority beads AND a final APPROVED verdict, so a reviewer that
|
|
448
|
+
keeps reopening below-goal beads pins the verdict and starves the gate.
|
|
449
|
+
- **Root cause**: buildReviewerPrompt never told the reviewer what the
|
|
450
|
+
sprint goal was. The reviewer read the epic's full child list, found P3
|
|
451
|
+
features with no commits, and (correctly, from its blind viewpoint)
|
|
452
|
+
demanded the missing work.
|
|
453
|
+
- **Fix** (runner.js), two layers:
|
|
454
|
+
(a) buildReviewerPrompt now takes the goal and carries a SPRINT SCOPE
|
|
455
|
+
instruction: features/beads below the goal priority are deferred by
|
|
456
|
+
design, must not fail the verdict, must not appear in reopenIds.
|
|
457
|
+
(b) Deterministic guard at the reopen loop: when reopenIds is
|
|
458
|
+
non-empty, fetch the scoped DAG once (bdListScoped('')) and SKIP any
|
|
459
|
+
id whose bead priority is numerically below the goal (priority >
|
|
460
|
+
goalMax), logging the skip. The lookup is conditional on reopenIds
|
|
461
|
+
being non-empty so the common empty path issues no extra bd call.
|
|
462
|
+
- **Test fallout lesson**: the first (unconditional) lookup shape caused
|
|
463
|
+
35 bd-replay drift failures; the conditional shape confined drift to
|
|
464
|
+
the 14 scenario files that genuinely exercise reopenIds. Those were
|
|
465
|
+
re-recorded (SUMMARY pass=37 fail=0) and the golden bd recordings +
|
|
466
|
+
transcripts re-recorded/regenerated; full suite green (789/789).
|
|
467
|
+
|
|
468
|
+
### Issue 18: Deploy fails every cycle on a permission self-check; Integ Test never runs, so features can never close
|
|
469
|
+
|
|
470
|
+
- **Symptom**: run 11 C1-C3 -- "Deploy FAILED this cycle: Stopped at
|
|
471
|
+
Step 0 (permission check)... Skipping Integration Test phase." Every
|
|
472
|
+
cycle evaluation then reports the same ~11 open P2+ beads. Those 11
|
|
473
|
+
are the pending-closure FEATURE nodes themselves (all children
|
|
474
|
+
closed): the integ-test-runner is the only role that closes verified
|
|
475
|
+
features, so with Deploy failing its pre-flight, the completion gate
|
|
476
|
+
(zero open goal beads + APPROVED) is mathematically unreachable and
|
|
477
|
+
every run rides to maxCycles.
|
|
478
|
+
- **Root cause**: deploy.md's Permissions section requires Bash(gh run *),
|
|
479
|
+
Bash(gh release *), Bash(mkdir *), Bash(*apra-fleet-installer-*
|
|
480
|
+
install *), Bash(*apra-fleet-installer-* --version) in the member's
|
|
481
|
+
permissions.allow. fleet-rev's .claude/settings.json had no
|
|
482
|
+
permissions key at all, and settings.local.json (composed at
|
|
483
|
+
onboarding, BEFORE deploy.md existed -- the playbook is sprint-authored
|
|
484
|
+
content) covered only Bash(gh:*). Nothing re-composes member
|
|
485
|
+
permissions when the sprint's own doers author a playbook that carries
|
|
486
|
+
new permission requirements.
|
|
487
|
+
- **Fix** (member-side, live, no sprint restart): added the five
|
|
488
|
+
deploy.md prefixes (plus a colon-syntax mkdir variant) to BOTH
|
|
489
|
+
.claude/settings.json and .claude/settings.local.json in fleet-rev's
|
|
490
|
+
sprint workspace. Both files are gitignored, so the member's working
|
|
491
|
+
tree stays clean. Applied while run 11 was in Review C4 R1, before
|
|
492
|
+
Deploy C4 fired.
|
|
493
|
+
- **Durability caveat (watched)**: if fleet-rev's permissions are ever
|
|
494
|
+
re-composed (compose_permissions / re-onboarding), these entries are
|
|
495
|
+
lost. The structural fix belongs in the permission profile for the
|
|
496
|
+
deployer/integ-test roles, or a pre-Deploy permission-ensure step in
|
|
497
|
+
the runner; deferred until Integ Test is observed actually running.
|
|
498
|
+
|
|
499
|
+
### Run 11 outcome: StalledSprintError, reopen-thrash on the deferred P3 trio
|
|
500
|
+
|
|
501
|
+
- Run 11 ended at Deploy C4 with StalledSprintError: closed-count history
|
|
502
|
+
[57, 62, 62, 62], reopen-thrash detected on apra-fleet-eft.10/.11/.12
|
|
503
|
+
(each reopened more than 3 times). This is exactly the Issue 17
|
|
504
|
+
oscillation -- the stall detector (correctly) refused to burn cycle 5
|
|
505
|
+
on it. It also means Deploy C4 never ran, so the Issue 18 permission
|
|
506
|
+
fix gets its first live test in run 12. finalizeAbort worked as
|
|
507
|
+
designed (idempotent [ABORTED] PR, terminal history record). Run 11's
|
|
508
|
+
content record: 4 cycles, zero infrastructure failures, closed beads
|
|
509
|
+
57->62, every P1/P2 feature at all-children-closed pending closure.
|
|
510
|
+
|
|
511
|
+
### Issue 19: deploy.md is non-idempotent and hardcodes the Windows artifact
|
|
512
|
+
|
|
513
|
+
- **Symptom**: run 12 Deploy C1 -- with Issue 18's permissions in place the
|
|
514
|
+
deployer executed deploy.md for the first time ever, and failed mid-run:
|
|
515
|
+
`gh run download ... error extracting "apra-fleet-installer-win-x64.exe":
|
|
516
|
+
file exists` (/tmp/fleet-deploy already held the artifact from the same
|
|
517
|
+
deploy session's earlier attempt).
|
|
518
|
+
- **Root cause**, two content defects in the sprint-authored playbook:
|
|
519
|
+
(a) NON-IDEMPOTENT: `mkdir -p /tmp/fleet-deploy` then download -- any
|
|
520
|
+
re-run (or step retry) collides with the previous download, and gh run
|
|
521
|
+
download has no overwrite flag. (b) WRONG PLATFORM: the Deploy and smoke
|
|
522
|
+
test sections hardcode apra-fleet-installer-win-x64.exe and
|
|
523
|
+
apra-fleet.exe -- authored from the orchestrator's Windows viewpoint,
|
|
524
|
+
but the runbook executes ON THE MEMBER (fleet-rev is macOS ARM). Even a
|
|
525
|
+
clean download would have died at the install step. The playbook's own
|
|
526
|
+
Platform binaries table already listed the right artifact names.
|
|
527
|
+
- **Fix** (deploy.md, committed to the sprint branch): Deploy section now
|
|
528
|
+
selects the artifact by `uname -s` (Darwin/Linux/fallback-Windows),
|
|
529
|
+
does `rm -rf /tmp/fleet-deploy` before mkdir, and `chmod +x` before
|
|
530
|
+
install; smoke test tries both binary names; manual/rollback sections
|
|
531
|
+
parameterized on the platform table. Permissions section gains
|
|
532
|
+
`Bash(rm -rf /tmp/fleet-deploy*)`, `Bash(chmod +x /tmp/fleet-deploy*)`,
|
|
533
|
+
and widens the version check to `Bash(*apra-fleet* --version)` (the old
|
|
534
|
+
prefix only covered the installer, not the installed binary the smoke
|
|
535
|
+
test runs). Same prefixes added live to fleet-rev's settings files, and
|
|
536
|
+
the stale /tmp/fleet-deploy was removed on the member.
|
|
537
|
+
|
|
538
|
+
### Issue 20: no deployable artifact exists for darwin-x64 members at all
|
|
539
|
+
|
|
540
|
+
- **Symptom**: run 12 Deploy C2 -- the Issue 19 platform selection worked,
|
|
541
|
+
but install failed with exit 127 'bad CPU type in executable':
|
|
542
|
+
fleet-rev is an INTEL Mac (Darwin x86_64) and main's CI matrix builds
|
|
543
|
+
only win-x64, linux-x64, darwin-arm64. There is no working deploy path
|
|
544
|
+
for this member: the arm64 binary cannot execute, and the fallback
|
|
545
|
+
apra-fleet-tarball flow is broken on main (the tarball omits
|
|
546
|
+
install.cjs and scripts/ which install.sh requires, and the legacy
|
|
547
|
+
tarball installer never creates ~/.apra-fleet/bin/, so the smoke test
|
|
548
|
+
could never pass) -- a main-repo packaging defect worth its own bead
|
|
549
|
+
on that backlog.
|
|
550
|
+
- **Disposition**: this is eft CONTENT (the deploy/integ runbooks are the
|
|
551
|
+
epic's own deliverables), not sprint-process infrastructure, so it was
|
|
552
|
+
filed as a sprint-scoped P2 bug bead apra-fleet-eft.15 per the bead
|
|
553
|
+
injection contract (created member-side with model metadata, dolt
|
|
554
|
+
pushed). The bead's preferred option is a build-from-source fallback
|
|
555
|
+
in deploy.md (npm run build:binary when no artifact matches
|
|
556
|
+
uname -s/-m), which is self-contained on the branch. Being P2 it
|
|
557
|
+
correctly holds the completion gate open until Deploy actually works.
|
|
558
|
+
|
|
559
|
+
### Issue 21: childless bug beads are planned as directly-dispatchable, but doers refuse non-task beads
|
|
560
|
+
|
|
561
|
+
- **Symptom**: run 12 C3 -- the planner assigned the injected bug bead
|
|
562
|
+
apra-fleet-eft.15 directly to a doer streak and the plan-reviewer
|
|
563
|
+
APPROVED that plan; the doer then skipped it per its contract ("doers
|
|
564
|
+
may only claim issue_type=task"), reporting BLOCKED, and the runner
|
|
565
|
+
marked the streak FAILED. Left alone this repeats every cycle the bug
|
|
566
|
+
stays open and the stall detector eventually kills the run -- the same
|
|
567
|
+
thrash shape as Issue 17, one layer down.
|
|
568
|
+
- **Root cause**: a planning-side blind spot symmetric to Issue 14.
|
|
569
|
+
Features get decomposed into task children, but an injected bug that
|
|
570
|
+
arrives mid-sprint as a childless leaf is treated as dispatchable
|
|
571
|
+
as-is by both planner and plan-reviewer; only the doer contract knows
|
|
572
|
+
it is not.
|
|
573
|
+
- **Fix** (runner.js, mirror-image prompt pins): buildPlannerPrompt now
|
|
574
|
+
states that doers can only claim task-type beads and any open childless
|
|
575
|
+
bug in scope must be decomposed into task-type children (with model
|
|
576
|
+
metadata and a [test] task where testable) during planning;
|
|
577
|
+
buildPlanReviewerPrompt gains the matching DISPATCHABILITY criterion --
|
|
578
|
+
a plan leaving an open childless bug-type leaf is CHANGES_NEEDED, with
|
|
579
|
+
an explicit carve-out for pending-closure features. Goldens
|
|
580
|
+
regenerated; full se suite green (863/0).
|
|
581
|
+
- **Rollout**: the pin reaches the sprint at the next relaunch. Run 12's
|
|
582
|
+
C4 planner may still self-correct from the doer's BLOCKED note, which
|
|
583
|
+
recommended exactly this decomposition.
|
|
584
|
+
|
|
585
|
+
### Run 12 outcome: StalledSprintError (flat closed count), Issues 19-21 discovered
|
|
586
|
+
|
|
587
|
+
- Run 12 ended at the C3 evaluation with StalledSprintError: closed-count
|
|
588
|
+
history [62, 62, 62], no reopen-thrash. C1 was a clean sweep (plan +
|
|
589
|
+
3 reviews APPROVED first try -- Issues 14 and 17 both holding), and
|
|
590
|
+
Deploy executed deploy.md for the first time in the campaign (Issue 18
|
|
591
|
+
fix working), which surfaced Issues 19 (non-idempotent, Windows-only
|
|
592
|
+
playbook) and 20 (no darwin-x64 artifact exists; filed as bead
|
|
593
|
+
apra-fleet-eft.15). C2/C3 then closed nothing: the only non-feature
|
|
594
|
+
open bead was eft.15, which doers refused every round as issue_type=
|
|
595
|
+
bug (Issue 21), and the stall detector correctly aborted before Plan
|
|
596
|
+
C4 could decompose it. Run 13 launches with the Issue 21 planner/
|
|
597
|
+
plan-reviewer pins, so its FIRST planning pass must decompose eft.15
|
|
598
|
+
into workable task children -- unblocking Deploy, then Integ Test,
|
|
599
|
+
then feature closure.
|
|
600
|
+
|
|
601
|
+
### Issue 22: integ-test-playbook.md permissions never provisioned (Issue 18's class, second member)
|
|
602
|
+
|
|
603
|
+
- **Symptom**: run 13 Integ Test C1 -- the FIRST integ-test dispatch of
|
|
604
|
+
the campaign (Deploy C1 finally succeeded via eft.15's source-build
|
|
605
|
+
fallback) refused to run at Step 0a: five of the six permission
|
|
606
|
+
prefixes integ-test-playbook.md requires were missing from fleet-rev's
|
|
607
|
+
allowlist (only Bash(mkdir *) was present, from the Issue 18 fix).
|
|
608
|
+
featuresClosed: 0 -- the cycle's feature-closure opportunity was lost.
|
|
609
|
+
- **Root cause**: same class as Issue 18. Sprint-authored playbooks
|
|
610
|
+
(deploy.md, integ-test-playbook.md) declare their own Permissions
|
|
611
|
+
sections, but nothing in onboarding or the sprint provisions those
|
|
612
|
+
prefixes onto the member; each playbook's Step 0 self-check then
|
|
613
|
+
correctly refuses.
|
|
614
|
+
- **Fix** (member-side, live): added the five missing prefixes to both
|
|
615
|
+
settings files on fleet-rev before Integ Test C2.
|
|
616
|
+
- **Structural fix candidate (watched)**: this is now the third
|
|
617
|
+
member-side permission patch (deploy Step 0, deploy rm/chmod/version,
|
|
618
|
+
integ Step 0a). The runner should get a pre-phase permission-ensure
|
|
619
|
+
step: parse the playbook's Permissions section and merge any missing
|
|
620
|
+
prefixes into the member's settings before dispatching deployer /
|
|
621
|
+
integ-test-runner. Defer until the current permission set proves
|
|
622
|
+
sufficient end-to-end, then implement once.
|
|
623
|
+
|
|
624
|
+
### Issue 23: member workspace untrusted -- ALL settings.json permissions silently ignored
|
|
625
|
+
|
|
626
|
+
- **Symptom**: run 13 Integ Test C2 -- the first real playbook dispatch
|
|
627
|
+
exhausted its turn budget, and the error carried the actual cause:
|
|
628
|
+
"Ignoring 18 permissions.allow entries from .claude/settings.json:
|
|
629
|
+
this workspace has not been trusted." Every playbook command was being
|
|
630
|
+
denied, so the agent burned its turns fighting refusals.
|
|
631
|
+
- **Root cause**: the Claude CLI only honors a workspace's
|
|
632
|
+
.claude/settings.json after the interactive trust dialog has been
|
|
633
|
+
accepted for that directory -- recorded in ~/.claude.json under
|
|
634
|
+
projects[<path>].hasTrustDialogAccepted. Headless members never see
|
|
635
|
+
the dialog, so all the Issue 18/19/22 permission provisioning into
|
|
636
|
+
settings.json was silently inert. (settings.local.json entries also
|
|
637
|
+
gate on the same trust flag.)
|
|
638
|
+
- **Fix** (member-side, live): set projects["/Users/akhil/git/
|
|
639
|
+
apra-fleet"].hasTrustDialogAccepted = true in ~/.claude.json on
|
|
640
|
+
fleet-rev. Onboarding lesson for the fleet: provisioning permissions
|
|
641
|
+
onto a member is only effective if the workspace is also trusted --
|
|
642
|
+
worth folding into compose_permissions / member onboarding.
|
|
643
|
+
|
|
644
|
+
### Issue 24: integ-test-runner had no turn-exhaustion resume ladder
|
|
645
|
+
|
|
646
|
+
- **Symptom**: same dispatch as Issue 23 -- max_turns_exhausted was
|
|
647
|
+
caught and degraded to passed:false (correct), but the cycle's entire
|
|
648
|
+
feature-closure opportunity was lost with no recovery attempt.
|
|
649
|
+
- **Fix** (runner.js): INTEG_TEST_MAX_TURNS = 100 (doer-sized -- the
|
|
650
|
+
runner owns the sandbox lifecycle plus full functional suites) with
|
|
651
|
+
the same same-session resume-and-continue ladder the reviewer/final-
|
|
652
|
+
review dispatches use (resume at 200 turns, sandbox kept, then the
|
|
653
|
+
existing degrade path). Structural baselines updated: 14 agent sites,
|
|
654
|
+
12 withGitSync brackets, 6 pushBeads brackets (integ once + resume).
|
|
655
|
+
Full se suite green (863/0). Applies from run 14 (run 13 already
|
|
656
|
+
loaded its runner; its C3+ integ attempts rely on the Issue 23 trust
|
|
657
|
+
fix removing the turn waste).
|
|
658
|
+
|
|
659
|
+
### Issue 25: universal turn-exhaustion resume (operator directive)
|
|
660
|
+
|
|
661
|
+
- **Directive**: every max_turns-based dispatch failure, in ANY phase,
|
|
662
|
+
must be answered with a same-session resume at doubled turns and a
|
|
663
|
+
"continue where you left off" prompt -- never treated as a terminal
|
|
664
|
+
failure on the first exhaustion.
|
|
665
|
+
- **Coverage before**: doer (100->200->400), reviewer (60->120), final
|
|
666
|
+
review (60->120), integ-test-runner (100->200, Issue 24). Uncovered:
|
|
667
|
+
planner, plan-reviewer, deployer, harvester -- each would degrade or
|
|
668
|
+
fail its round on a single exhaustion.
|
|
669
|
+
- **Fix** (runner.js): all four remaining member-dispatch sites gain
|
|
670
|
+
explicit turn budgets and the same resume ladder -- planner 100->200
|
|
671
|
+
(doer-sized, builds the whole DAG), plan-reviewer 60->120, deployer
|
|
672
|
+
60->120 (the source-build fallback deploy runs npm ci + two builds),
|
|
673
|
+
harvester 60->120. Each resume repeats the role's withGitSync/dolt
|
|
674
|
+
bracket semantics (planner and harvester resumes carry pushBeads;
|
|
675
|
+
harvester resume carries pushCode).
|
|
676
|
+
- **Deliberate exception**: the two streak-assignment agent() sites stay
|
|
677
|
+
ladder-free -- they are pure-compute schema asks (no member run) whose
|
|
678
|
+
failure degrades safely to one-bead-per-streak, and they already have
|
|
679
|
+
their own bounded semantic-repair re-ask (Issue 15).
|
|
680
|
+
- **Structural baselines**: 18 agent sites, 16 withGitSync brackets,
|
|
681
|
+
8 pushBeads brackets, 4 pushCode brackets.
|
|
682
|
+
- **Rollout**: run 14 launched minutes before this change (it has the
|
|
683
|
+
Issue 24 integ ladder, which is the one that matters for its critical
|
|
684
|
+
path); full coverage from run 15.
|
|
685
|
+
|
|
686
|
+
### Run 14 outcome: full 5-cycle lifecycle, Final Verdict FAIL on the Issue 26 harness gap
|
|
687
|
+
|
|
688
|
+
- Run 14 was the campaign's most productive run and the first to complete
|
|
689
|
+
the whole sprint lifecycle: 5 cycles + Final Review + Harvest +
|
|
690
|
+
Publish PR (idempotent -- updated the existing PR). The integ phase ran
|
|
691
|
+
for the first time end to end and CLOSED 8 features in cycle 1, then
|
|
692
|
+
drove a real find-fix loop through genuine PRODUCT bugs across cycles:
|
|
693
|
+
eft.16 (dolt no-remote), eft.18 (canary self-provision), eft.19 (P0:
|
|
694
|
+
dev-mode install missing undici), eft.20 (P1: state.json corruption --
|
|
695
|
+
doubled-quote breaking JSON.parse in the eft.2 persistence path,
|
|
696
|
+
stalling auto-sprint after Plan), eft.21 (teardown ~-reresolution).
|
|
697
|
+
Every one fixed with tests; unit suite ended green (2337 passing).
|
|
698
|
+
- Final Verdict: FAIL -- correctly. The headline deliverable (auto-sprint
|
|
699
|
+
running a sprint end-to-end via the integ smoke test's Part 2 Test
|
|
700
|
+
scenario) never completed in any cycle. Root cause = Issue 26 below.
|
|
701
|
+
Final Review filed a newTask gating epic-close on a green Part 2 run.
|
|
702
|
+
- Two run-14-only gaps (both fixed for run 15): the Harvester hit turn
|
|
703
|
+
exhaustion with NO resume (run 14 launched minutes before the Issue 25
|
|
704
|
+
universal ladder), so docs were not harvested; and the completion gate
|
|
705
|
+
was never reachable because of Issue 26.
|
|
706
|
+
|
|
707
|
+
### Issue 26: integ-test-runner cannot complete its own playbook's Part 2 (no MCP tools)
|
|
708
|
+
|
|
709
|
+
- **Symptom**: integ Part 2 "Test scenario" step 1 registered a member
|
|
710
|
+
via the `register_member` MCP tool, but the integ-test-runner agent is
|
|
711
|
+
granted only [Read, Bash, Grep, Glob] -- no MCP. Cycles C2/C3 only got
|
|
712
|
+
past it by hand-rolling a Bash HTTP/JSON-RPC call to the MCP server (the
|
|
713
|
+
workaround CLAUDE.md forbids); C4/C5 correctly refused and reported the
|
|
714
|
+
gap, so Part 2 never completed and the sprint could not sign off.
|
|
715
|
+
- **Root cause**: the agent's tool grant contradicted the playbook it
|
|
716
|
+
owns end to end. `register_member` had no CLI equivalent, so Part 2
|
|
717
|
+
step 1 was not shell-drivable at all.
|
|
718
|
+
- **Fix** (chosen by operator: add the CLI, not grant MCP -- keeps the
|
|
719
|
+
fix in apra-fleet, no apra-pm submodule/upstream PR, and makes Part 2
|
|
720
|
+
fully shell-only per the playbook's own Setup/Reset/Teardown principle):
|
|
721
|
+
new `register-member` CLI subcommand (src/cli/register-member.ts, wired
|
|
722
|
+
into src/index.ts) wrapping the SAME shared registerMember() function
|
|
723
|
+
the MCP tool uses (src/tools/register-member.ts:110) through the same
|
|
724
|
+
Zod schema -- no logic fork. integ-test-playbook.md Part 2 step 1
|
|
725
|
+
reworded to `node dist/index.js register-member --type local ...` via
|
|
726
|
+
Bash; the shell-vs-MCP notes updated (every section is now
|
|
727
|
+
shell-drivable). 9 new CLI tests; full suite green. Commit 60c518a.
|
|
728
|
+
- **Note**: local+claude registration still bootstraps an interactive
|
|
729
|
+
session needing the running server (Setup starts it) and spawns a real
|
|
730
|
+
claude process -- identical to the MCP path, faithfully wrapped, not
|
|
731
|
+
diverged.
|
|
732
|
+
|
|
733
|
+
### Observed while dispatching the 0ei hotfix (separate track): stale busy lock
|
|
734
|
+
|
|
735
|
+
- fleet-dev's execute_prompt lock returned {"isError":true,"reason":
|
|
736
|
+
"busy"} for ~20 minutes while the member session was idle with no live
|
|
737
|
+
process (token counts frozen); stop_prompt reported "no active session
|
|
738
|
+
was running" and cleared it. A dispatch-layer lock leak on the fleet
|
|
739
|
+
server (v0.3.5_8a1ef3) -- likely left by an earlier abandoned/killed
|
|
740
|
+
dispatch. Watch for recurrence; if it repeats, file a server bead (lock
|
|
741
|
+
release on child-process reap, or a TTL/liveness check on
|
|
742
|
+
inFlightAgents).
|
|
743
|
+
|
|
744
|
+
### Still open / watched
|
|
745
|
+
|
|
746
|
+
- **apra-fleet-eft.14 (server)**: why does the provider CLI sometimes exit
|
|
747
|
+
0 with no parseable result/usage/sessionId? Client-side typed handling
|
|
748
|
+
is in place (Issue 3); the server-side parse gap needs its own
|
|
749
|
+
investigation.
|
|
750
|
+
- **apra-fleet-eft.15 (perf)**: repeated near-identical project-wide
|
|
751
|
+
`bd list` dumps within a phase (clearly visible in the fleet server log,
|
|
752
|
+
~1.5-2s each); P2, not sprint-fatal.
|
|
753
|
+
- **POST-stream idle timeout risk**: OBSERVED in run 5 and fixed as Issue
|
|
754
|
+
8 (iteration 2). Server-side SSE keepalives remain an optional hardening
|
|
755
|
+
follow-up.
|
|
756
|
+
- **apra-fleet-qv1**: parallel doer streaks to a single member are unsafe
|
|
757
|
+
(pre-existing bead, visible in the server log's bead dumps); single-member
|
|
758
|
+
sprints serialize doers so not currently hit.
|
|
759
|
+
|
|
760
|
+
### Process change (2026-07-20): launch via the delivered CLI surface
|
|
761
|
+
|
|
762
|
+
From run 16 onward the stabilization loop launches sprints through the
|
|
763
|
+
delivered product surface instead of the raw engine path:
|
|
764
|
+
|
|
765
|
+
npm run build
|
|
766
|
+
node dist/index.js install (refresh ~/.apra-fleet workflow runtime)
|
|
767
|
+
node dist/index.js workflow auto-sprint --issue apra-fleet-eft \
|
|
768
|
+
--members fleet-rev --branch auto-sprint/eft-service \
|
|
769
|
+
--base feat/fleet-reorg --viewer-port 18300
|
|
770
|
+
|
|
771
|
+
Rationale: every relaunch now also exercises the dev-mode install asset
|
|
772
|
+
pipeline, the workflow launcher (version-skew check, ADR server
|
|
773
|
+
resolution, verbatim arg passthrough), matching how end users run
|
|
774
|
+
sprints. An install/launcher regression now blocks the loop by design --
|
|
775
|
+
that is surfaced signal, not friction to route around. The integ-test
|
|
776
|
+
playbook's smoke test (Part 2 step 3) already launches via
|
|
777
|
+
'apra-fleet workflow auto-sprint', so both lanes now run the same
|
|
778
|
+
delivered surface. Version bumped to 0.4.0 on this branch (the reorg/eft
|
|
779
|
+
line ships as 0.4.0; 0.3.5 was released from main).
|
|
780
|
+
|
|
781
|
+
### Issue 27 (run 15 C1): bare "continue" resume prompts lose per-dispatch scope
|
|
782
|
+
|
|
783
|
+
Observed live twice in run 15 C1. (a) The integ runner exhausted 100 turns,
|
|
784
|
+
was resumed with a generic continue-prompt, and returned featuresClosed: 0
|
|
785
|
+
with "this dispatch's .fleet-task.md carried no explicit feature-id list"
|
|
786
|
+
-- correct behavior per its missing-scope contract, because a resumed
|
|
787
|
+
dispatch DELIVERS A NEW PROMPT ARTIFACT to the member, replacing the
|
|
788
|
+
original one that carried the feature-id scope. (b) The resumed doer for
|
|
789
|
+
one streak drifted onto a previously-finished bead ("no new work to
|
|
790
|
+
continue") for the same reason: its continue-prompt never restated the
|
|
791
|
+
assigned bead ids.
|
|
792
|
+
|
|
793
|
+
Fix: the three resume sites whose contracts depend on per-dispatch scope
|
|
794
|
+
(integ-test-runner feature ids; doer bead ids + branch; reviewer bead ids
|
|
795
|
+
+ branch/base) now restate that scope verbatim inside the resume prompt.
|
|
796
|
+
The five other resume sites (planner, plan-reviewer, deployer, final
|
|
797
|
+
review, harvester) self-derive their scope or carry it in-session and are
|
|
798
|
+
left as-is. Engine suite: 887 pass / 0 fail.
|
|
799
|
+
|
|
800
|
+
### Issue 28 (run 15 C2/C3): undispatchable beads wasted a full doer dispatch per round
|
|
801
|
+
|
|
802
|
+
Observed live: 4 of 6 streaks in one Develop round were bug-type beads a
|
|
803
|
+
doer is contract-bound to refuse -- each still cost a complete LLM dispatch
|
|
804
|
+
that returned "skip" deterministically. Three engine holes, three fixes:
|
|
805
|
+
(a) readyLeafBeads built its decomposed-parent exclusion set from OPEN
|
|
806
|
+
children only, so the moment a bug's task children all closed it re-entered
|
|
807
|
+
doer seeding as a fake leaf -- now built from the any-status dump
|
|
808
|
+
(bdListScoped('') -- no new bd command); (b) streak seeding now mirrors the
|
|
809
|
+
doer contract engine-side, dropping non-task beads before dispatch, with
|
|
810
|
+
target issues exempt (the eft.24 childless-leaf-target sprint case must
|
|
811
|
+
stay dispatchable); (c) pending-closure BUGS (all task children closed) had
|
|
812
|
+
no closure owner at all -- doers refuse, reviewers may not close, integ
|
|
813
|
+
prompts only named features -- so they lingered open at goal forever; the
|
|
814
|
+
integ dispatch prompt now names them for verify-and-close.
|
|
815
|
+
|
|
816
|
+
### Issue 29 (run 15 C3 R3): '-tier'-suffixed model metadata 404-failed a whole streak
|
|
817
|
+
|
|
818
|
+
An out-of-band injected bead carried metadata {"model": "standard-tier"};
|
|
819
|
+
the engine passed it verbatim to execute_prompt, the server tier map missed
|
|
820
|
+
it, and it reached the provider CLI as a literal model name: claude --model
|
|
821
|
+
standard-tier -> api_error_status 404, streak lost. Fix: normalizeTierToken()
|
|
822
|
+
containment-normalizes any value holding exactly ONE tier word (any
|
|
823
|
+
case/shape) at the single engine read site; zero or multiple matches pass
|
|
824
|
+
through untouched so explicit model ids are never mangled. The planner
|
|
825
|
+
prompt's defensive '-tier' legalese was trimmed at the same time -- enforce
|
|
826
|
+
in code, not prompts. Data fixed member-side (eft.26/27 -> "standard").
|
|
827
|
+
|
|
828
|
+
### Issue 30 (run 15 C4, preemptive): 60-min hard ceiling would false-fail a long integ pass
|
|
829
|
+
|
|
830
|
+
Found by inspection while Integ Test C4 was in flight -- the first integ
|
|
831
|
+
pass expected to run LONG (real suites + a full sandbox smoke sprint),
|
|
832
|
+
which is exactly when the ceiling matters. The integ dispatch set
|
|
833
|
+
timeout_s: 3600 AND max_total_s: 3600. The first is an inactivity timer
|
|
834
|
+
(resets on member stdout; the playbook's background-sprint poll loop keeps
|
|
835
|
+
it fed). The second is a HARD kill at elapsed time regardless of activity
|
|
836
|
+
or progress. A timer kill surfaces as a plain AgentDispatchError, which
|
|
837
|
+
the catch converts to passed:false -- a FALSE integ failure -- and the
|
|
838
|
+
resume ladder only matches reason max_turns_exhausted, so the killed run
|
|
839
|
+
gets no resume even if minutes from done. Compounding: two such cycles
|
|
840
|
+
with no other closes trip stall-abort and kill the whole run.
|
|
841
|
+
|
|
842
|
+
Note the sprint-level stall detector is NOT a risk here: it is cycle-
|
|
843
|
+
granular (closed-bead high-water mark at cycle evaluation), structurally
|
|
844
|
+
unable to fire mid-phase.
|
|
845
|
+
|
|
846
|
+
Fix: integ runner max_total_s raised to 7200 while keeping timeout_s at
|
|
847
|
+
3600 -- a genuinely hung runner still dies after 60 min of silence; an
|
|
848
|
+
active long pass gets 2h of headroom. Other roles keep 3600/3600 (deploy,
|
|
849
|
+
plan, review are not plausibly hour-plus when healthy). Takes effect run
|
|
850
|
+
16; run 15's in-flight C4 still carries the 60-min ceiling, so a C4 integ
|
|
851
|
+
result of passed:false with a dispatch-failed summary near the 60-min mark
|
|
852
|
+
should be read as THIS signature, not a real test failure.
|
|
853
|
+
|
|
854
|
+
### Issue 31 (run 15 verdict FAIL, root cause of eft.30): D-push gate ran AFTER the push attempt
|
|
855
|
+
|
|
856
|
+
The single most serious finding of run 15: in doltPushAfter(), `bd dolt
|
|
857
|
+
push` was attempted FIRST and isMemberSyncRemoteConfigured() was consulted
|
|
858
|
+
only on the FAILURE path (eft.30.2), to downgrade an unclassifiable failure
|
|
859
|
+
to a benign skip. But bd auto-provisions a Dolt-level remote from git's own
|
|
860
|
+
origin on the push attempt itself, so on a clone with valid credentials the
|
|
861
|
+
push SUCCEEDS against the real shared remote -- neutralization verified
|
|
862
|
+
clean seconds earlier is simply re-armed. Observed live in integ C4 AND C5:
|
|
863
|
+
a sandbox reached the real fleet-e2e-toy remote and was stopped only by the
|
|
864
|
+
machine's missing GitHub credentials, not by design. The C5 R3 reviewer
|
|
865
|
+
reopened eft.30 with exactly this diagnosis; the final FAIL verdict cited
|
|
866
|
+
it as the run's core defect.
|
|
867
|
+
|
|
868
|
+
Fix: the push is now gated BEFORE being issued -- when the member's
|
|
869
|
+
bd-level sync.remote is positively confirmed absent/neutralized, no `bd
|
|
870
|
+
dolt push` command is issued at all (same benign no-remote skip result).
|
|
871
|
+
isMemberSyncRemoteConfigured fails CLOSED (any inconclusive read reports
|
|
872
|
+
configured), so a real, actively-synced clone always still pushes; the
|
|
873
|
+
eft.30.2 failure-path downgrade stays as defense-in-depth. New test pins
|
|
874
|
+
the load-bearing assertion: with sync.remote absent, ZERO push commands are
|
|
875
|
+
ISSUED (not merely "the failure was benign"); a stateful-stub test keeps
|
|
876
|
+
the failure-path branch covered. Cost: one `bd config get sync.remote`
|
|
877
|
+
read per D-push bracket. Goldens refreshed (8 pre-gate reads now appear in
|
|
878
|
+
the happy-path transcript); arg-contract member-routing tests exclude the
|
|
879
|
+
new read alongside `bd dolt *` (same per-member bracket class). Engine
|
|
880
|
+
suite: 960 pass / 0 fail; root vitest green minus the 2 known Windows-local
|
|
881
|
+
teardown flakes.
|
|
882
|
+
|
|
883
|
+
### Issue 32 (run 15 integ C4/C5): a live-but-silent hung dispatch costs a full hour, and no watchdog CAN fire
|
|
884
|
+
|
|
885
|
+
The eft.28 recurrence signature: the sandbox planner dispatch hung with
|
|
886
|
+
its PID alive, dashboard state frozen 2m15s+, zero output. "Why did no
|
|
887
|
+
watchdog call it stalled?" -- because with `claude -p` there is no signal
|
|
888
|
+
to watch: the CLI prints nothing until the turn completes (Issue 12), so
|
|
889
|
+
server-side inactivity == total runtime and an output-based watchdog
|
|
890
|
+
cannot distinguish a hang from silent work. The engine DOES self-heal (the
|
|
891
|
+
dispatch dies at timeout_s and the retry/resume ladders take over) but at
|
|
892
|
+
the cost of the full dispatch budget: 60 minutes lost per hang, fatal
|
|
893
|
+
inside a 10-minute canary smoke test.
|
|
894
|
+
|
|
895
|
+
Engine fix: the budget is now an arg. `dispatch_timeout_s` (default 3600,
|
|
896
|
+
floor 60; CLI `--dispatch-timeout-s`) is applied as BOTH timeout_s and
|
|
897
|
+
max_total_s at every dispatch site (the Issue 12 equality is preserved at
|
|
898
|
+
any value), with the integ-test ceiling at 2x (preserving Issue 30). The
|
|
899
|
+
integ playbook's sandbox canary launch now passes 900 -- a hang there
|
|
900
|
+
costs 15 minutes and the sprint's own ladders recover. TRUE liveness
|
|
901
|
+
detection (is the member's claude process making progress?) requires a
|
|
902
|
+
member-side signal the orchestrator does not have; that remains the open
|
|
903
|
+
eft.28 product bead (execute-prompt.ts), now with two cycles of fresh
|
|
904
|
+
repro evidence.
|
|
905
|
+
|
|
906
|
+
Also found while editing: the playbook's smoke step named a
|
|
907
|
+
`skip_dolt_push: true` arg that has never existed in the engine
|
|
908
|
+
(validateArgs would reject it; the CLI's strict parser has no such flag)
|
|
909
|
+
-- removed, with sandbox isolation now genuinely enforced by the Issue 31
|
|
910
|
+
pre-gate instead. Tests: arg-contract suite pins default/validation and
|
|
911
|
+
end-to-end plumbing (every mock dispatch carries the overridden budget).
|
|
912
|
+
Engine suite: 963 pass / 0 fail.
|
|
913
|
+
|
|
914
|
+
### Issue 33 (run 15 C5): a regressed bug with closed children is unreachable by every role
|
|
915
|
+
|
|
916
|
+
Observed end-to-end in run 15 C5: eft.28 and eft.30 carried fresh
|
|
917
|
+
"recurred despite the landed fixes" evidence from integ C4, but their fix
|
|
918
|
+
children were all closed -- so the plan reviewer classified them as
|
|
919
|
+
pending-closure ("do NOT re-decompose"), the planner left them alone,
|
|
920
|
+
doers refused them as non-task beads (three full develop rounds of nine
|
|
921
|
+
refusals each), and the integ runner may only verify-and-close. Nobody in
|
|
922
|
+
the role graph had authority to create NEW fix tasks for a regression;
|
|
923
|
+
the run ended FAIL on exactly those two beads.
|
|
924
|
+
|
|
925
|
+
Fix: a REGRESSION EXCEPTION added to both sides of the Issue 14/21 rule
|
|
926
|
+
pair. buildPlannerPrompt (delta cycles): an open bug whose children are
|
|
927
|
+
all closed but whose notes record the defect reproducing AFTER closure is
|
|
928
|
+
a regression -- read the latest evidence and create NEW fix + [test] task
|
|
929
|
+
children targeting the residual mechanism (never duplicating the closed
|
|
930
|
+
fix). buildPlanReviewerPrompt: such a bug with no new open task children
|
|
931
|
+
makes the plan NOT approvable (CHANGES_NEEDED); bugs without post-closure
|
|
932
|
+
recurrence stay under pending-closure as before. Goldens refreshed.
|
|
933
|
+
Engine suite: 963 pass / 0 fail.
|
|
934
|
+
|
|
935
|
+
### Issue 34 (run 16 C1 R2): Issue 28's seeding filter missed the in-loop site
|
|
936
|
+
|
|
937
|
+
Observed live: bug bead eft.37 (filed mid-run, after the plan phase) was
|
|
938
|
+
seeded into its own doer streak in Develop C1 R2. Root cause: the Issue 28
|
|
939
|
+
dispatchability filter was applied only to the PRE-LOOP readyBeads list
|
|
940
|
+
(the round-zero check); the IN-LOOP currentReady list -- the one that
|
|
941
|
+
actually feeds the streak-assignment prompt each round -- had no filter,
|
|
942
|
+
so any non-task bead born after planning (reviewer newTasks are created
|
|
943
|
+
as tasks, but out-of-band bug filings are not) reaches a doer and burns a
|
|
944
|
+
dispatch on a contract-bound refusal. Fix: same filter (with the same
|
|
945
|
+
target-issue exemption) applied at the in-loop site. Honest note: no test
|
|
946
|
+
ever pinned the filter directly -- that is exactly why the second site
|
|
947
|
+
slipped; the run-16 live sprint is the current verification, and a
|
|
948
|
+
focused seeding-filter unit remains worth adding when the runner's
|
|
949
|
+
seeding is next factored testably. Engine suite: 963 pass; the 2
|
|
950
|
+
dolt-sync-discipline real-subprocess failures are the known Windows EBUSY
|
|
951
|
+
temp-dir-cleanup flake (same class PR #332 fixed elsewhere), aggravated
|
|
952
|
+
by the live sprint's dolt server -- unrelated to this change.
|
|
953
|
+
|
|
954
|
+
## Issue 35: install boundary used a stale binary (9 commits behind tip)
|
|
955
|
+
|
|
956
|
+
Observed: run 17's installed runtime (built at 4286ef0) lacked eft.38's
|
|
957
|
+
viewer more-button plus engine fixes it was meant to validate (Issue 34
|
|
958
|
+
in-loop filter, eft.28.3 client dispatch watchdog, eft.34/35 D-pull
|
|
959
|
+
gates, eft.27.4 stdout cap). Root cause: the SEA binary was built hours
|
|
960
|
+
before the install boundary and never rebuilt from the branch tip; the
|
|
961
|
+
delivered-binary convention runs the INSTALLED runtime, so workspace
|
|
962
|
+
commits after the build silently do not participate.
|
|
963
|
+
|
|
964
|
+
Rule adopted: at every run boundary, git fetch + fast-forward the
|
|
965
|
+
workspace to origin tip, run the suites, `npm run build:binary`, THEN
|
|
966
|
+
`install --force` -- and verify `apra-fleet version` reports the tip
|
|
967
|
+
short-SHA before relaunching. Never install a binary whose embedded
|
|
968
|
+
version does not match the branch tip.
|
|
969
|
+
|
|
970
|
+
## Issue 36: eft.41.1/41.2 fix pair killed every real sprint launch
|
|
971
|
+
|
|
972
|
+
Observed: run 18 died at t=0 with the new 41.2 backstop's own error
|
|
973
|
+
("entry did not execute ... Nothing happened"). cli.mjs self-executes
|
|
974
|
+
via its isMainModule() guard but never declared the launcher contract
|
|
975
|
+
`export const selfExecuting = true`; its async main() starts, the import
|
|
976
|
+
resolves, the backstop sees no flag and no exported callable, errors,
|
|
977
|
+
and returns 1 -- killing the sprint it just started. All platforms, all
|
|
978
|
+
real launches. It escaped the doer's validation because `--help` calls
|
|
979
|
+
process.exit(0) during import, before the backstop runs -- the only path
|
|
980
|
+
CI smoke exercises (hello-world is safe the same way).
|
|
981
|
+
|
|
982
|
+
Fix: declare `export const selfExecuting = true` in cli.mjs (d0c000a).
|
|
983
|
+
Verified: run 19 launches and runs on the delivered binary. Lesson
|
|
984
|
+
recorded on eft.41: the [test] child must exercise a REAL-ARGS launch
|
|
985
|
+
through the launcher import path, not only --help; a backstop that fires
|
|
986
|
+
on "no observable execution" must account for guarded self-executing
|
|
987
|
+
modules that neither export nor exit.
|
|
988
|
+
|
|
989
|
+
## Issue 37: sprint branches are a CI blind spot
|
|
990
|
+
|
|
991
|
+
Observed: ci.yml triggers only on PRs to main and pushes to main, so the
|
|
992
|
+
sprint branch accumulated ~20 commits with ZERO CI signal; the first-ever
|
|
993
|
+
run (manual workflow_dispatch, 29803574606) immediately caught a broken
|
|
994
|
+
intermediate state no member could see. Rule: at every install boundary,
|
|
995
|
+
manually dispatch CI on the sprint branch (gh workflow run --ref <branch>)
|
|
996
|
+
and treat its result as a gate input. Candidate engine fix: add a push
|
|
997
|
+
trigger for auto-sprint/** or have the runner dispatch CI after each
|
|
998
|
+
review-approved round.
|
|
999
|
+
|
|
1000
|
+
## Issue 38: never descope one half of a sequenced migration
|
|
1001
|
+
|
|
1002
|
+
Observed: eft.37.1 (core rename) landed with an explicit note that the se
|
|
1003
|
+
package cannot import until eft.37.2 updates consumers; the orchestrator
|
|
1004
|
+
then descoped the whole eft.37 family to P3 -- stranding the branch in a
|
|
1005
|
+
knowingly-broken intermediate state that CI (Issue 37) exposed: all 10
|
|
1006
|
+
supervisor test files failed at import. Rule: before cutting scope, check
|
|
1007
|
+
each candidate for landed-but-incomplete sequenced work (a sibling commit
|
|
1008
|
+
already on the branch that depends on it); restore the minimal completing
|
|
1009
|
+
task (37.2) rather than reverting landed work.
|
|
1010
|
+
|
|
1011
|
+
## Issue 39: stale in_progress beads are gate-blocking zombies
|
|
1012
|
+
|
|
1013
|
+
Observed: eft.13.4 sat in_progress for 2 days after its doer streak was
|
|
1014
|
+
interrupted -- its work had actually LANDED (salvage commit 8b6c987) but
|
|
1015
|
+
the bead was never closed, and in_progress beads do not re-enter the ready
|
|
1016
|
+
set. A full audit found 5 more stale claims (eft.19/7/15/12/17.1), several
|
|
1017
|
+
with all work complete. Rules: audit Started/Updated timestamps, not
|
|
1018
|
+
status glyphs; cross-reference claimed beads against landed commits before
|
|
1019
|
+
believing any status; engine candidate: a streak that ends abnormally must
|
|
1020
|
+
reset its claimed beads to open (or the runner sweeps claims older than a
|
|
1021
|
+
cycle at plan time).
|
|
1022
|
+
|
|
1023
|
+
## Issue 40: closed-bead notes are a feedback black hole
|
|
1024
|
+
|
|
1025
|
+
Observed: recurring user feedback (viewer scroll, LLM-output more button)
|
|
1026
|
+
was appended as notes to CLOSED beads (eft.27 family, eft.38) where no
|
|
1027
|
+
planner or verifier ever looks; the defects survived multiple 'fixed'
|
|
1028
|
+
cycles. eft.38's closing unit test fabricated an activity shape (inline
|
|
1029
|
+
output + Truncated flag) the real pipeline never produces, so the feature
|
|
1030
|
+
shipped rendering zero buttons. Rules: post-closure regressions REOPEN the
|
|
1031
|
+
old bead with evidence (never file a duplicate -- lineage matters); search
|
|
1032
|
+
all statuses AND all prefixes before filing anything (two open scroll bugs
|
|
1033
|
+
sat outside the epic, invisible to every sprint); UX beads must carry
|
|
1034
|
+
machine-verifiable acceptance (DOM measurement or HTTP check) so integ can
|
|
1035
|
+
verify closures without a human eyeball.
|
|
1036
|
+
|
|
1037
|
+
## Issue 41: the bead-injection contract applies to the orchestrator too
|
|
1038
|
+
|
|
1039
|
+
Observed: orchestrator-injected eft.18.5 was rejected by the plan reviewer
|
|
1040
|
+
for exactly the defects the injection contract predicts: no model metadata
|
|
1041
|
+
key, oversized scope (playbook rewrite + multi-file test retirement in one
|
|
1042
|
+
task), and an open task under a closed parent. The planner had to spend a
|
|
1043
|
+
round restructuring it (18.5/.6/.7 + 18.8). Rule: orchestrator-side beads
|
|
1044
|
+
follow the same contract as member-side ones -- model metadata set, single
|
|
1045
|
+
-concern task size, correct parentage -- or they cost a plan round.
|
|
1046
|
+
|
|
1047
|
+
## Issue 42: structural isolation beats runtime neutralization
|
|
1048
|
+
|
|
1049
|
+
Observed: the smoke test's entire recurring bug family (eft.18/25/30/31/
|
|
1050
|
+
39/47 -- neutralize dance, origin breakage, canary misses, Dolt remote
|
|
1051
|
+
re-adoption) stems from one design divergence: it hydrates beads from the
|
|
1052
|
+
live shared Dolt remote (bd bootstrap) and then tries to clean up, while
|
|
1053
|
+
the e2e suite prevents the class structurally (no-db JSONL seeding, remote
|
|
1054
|
+
never configured). Adopted design (eft.18.5-.8, user-amended): seed like
|
|
1055
|
+
e2e (bd init --from-jsonl, hardcoded canary gh-toy-4ef) but keep the
|
|
1056
|
+
workflow's D-push/D-pull under test against a sandbox-local file://
|
|
1057
|
+
THROWAWAY remote; git origin points at a local mirror so bd's
|
|
1058
|
+
push-time remote auto-provisioning can only ever derive sandbox paths; the
|
|
1059
|
+
four-check guard script is retargeted (all remotes must resolve inside the
|
|
1060
|
+
sandbox), never deleted. Safety invariant: the real remote URL must not
|
|
1061
|
+
appear anywhere in sandbox git or beads config at any point.
|
|
1062
|
+
## Issue 43: Smoke-test rehearsal on a second platform finds holes the author platform cannot (2026-07-21)
|
|
1063
|
+
|
|
1064
|
+
**What happened:** Before run 19's Integ C3 executed the rewritten smoke-test
|
|
1065
|
+
playbook on fleet-rev (macOS), the orchestrator rehearsed the same playbook
|
|
1066
|
+
on the Windows host in an isolated sandbox (own HOME, port, throwaway
|
|
1067
|
+
remotes), with real model dispatches. The rehearsal surfaced six holes, four
|
|
1068
|
+
of which would have failed the real run:
|
|
1069
|
+
|
|
1070
|
+
1. (doc) Playbook assumes a built repo: `dist/` present and `packages/apra-fleet-se/apra-pm`
|
|
1071
|
+
submodule initialized. Fresh checkout fails at `install` step 6/12.
|
|
1072
|
+
Fix: prerequisites block (recursive clone/`git submodule update --init
|
|
1073
|
+
--recursive`, `npm install && npm run build`).
|
|
1074
|
+
2. (product, Windows) `start` trusts `schtasks /query` for "service
|
|
1075
|
+
installed" -- schtasks is per-user global, not HOME-scoped, so a
|
|
1076
|
+
sandboxed start finds the operator's REAL ApraFleet task and starts
|
|
1077
|
+
that; the sandbox server never spawns. The real server survived only
|
|
1078
|
+
via its own singleton guard. macOS is immune (plist existence check
|
|
1079
|
+
under the overridden HOME).
|
|
1080
|
+
3. (playbook, all platforms) Fresh sandbox HOME has no git identity; bd's
|
|
1081
|
+
seed commit fails exit 128 and the toy doer's first commit would too.
|
|
1082
|
+
Fix: seed `git config --global user.name/email` in Setup.
|
|
1083
|
+
4. (product, all platforms) `auth --oauth` wrote a credentials file with
|
|
1084
|
+
ONLY `claudeAiOauth.accessToken`; the Claude CLI requires at least
|
|
1085
|
+
`expiresAt` and treats the file as "Not logged in". Every dispatch
|
|
1086
|
+
failed auth. Fix (validated in rehearsal): `auth --oauth` accepts a
|
|
1087
|
+
full claudeAiOauth JSON object as the secret; the playbook seeds the
|
|
1088
|
+
whole object from REAL_HOME, preserving real expiry + refresh token.
|
|
1089
|
+
The eft.48.2 checker passed while dispatch failed -- it verifies the
|
|
1090
|
+
file lands, not that Claude accepts it; harden with expiresAt check.
|
|
1091
|
+
5. (product+engine, all platforms) Auth failures are retried and can hang:
|
|
1092
|
+
the non-interactive dispatch path fails fast (8s, classified 'auth'),
|
|
1093
|
+
but (a) the interactive path waits the full --dispatch-timeout-s with
|
|
1094
|
+
no output classification; (b) register_member launches the member's
|
|
1095
|
+
live session BEFORE credentials are provisioned (playbook order:
|
|
1096
|
+
register step 1, provision step 3), creating a permanently
|
|
1097
|
+
unauthenticated zombie session that a later file fix cannot heal; and
|
|
1098
|
+
(c) the engine's dispatch retry wrapper retries an error that begins
|
|
1099
|
+
"Authentication failed" five times -- isRetryable() already says
|
|
1100
|
+
'auth' is terminal, but the engine never consults it. A transport
|
|
1101
|
+
timeout string was also fed into the schema-repair loop as if it were
|
|
1102
|
+
agent output.
|
|
1103
|
+
6. (doc) The "every step is shell-drivable" claim has no shell-drivable
|
|
1104
|
+
recovery: there is no remove-member/relaunch-session CLI path, so an
|
|
1105
|
+
unauthenticated member cannot be relaunched without restarting the
|
|
1106
|
+
sandbox server.
|
|
1107
|
+
|
|
1108
|
+
**Lesson:** A playbook exercised only on the platform and machine where it
|
|
1109
|
+
was written inherits that machine's ambient state (running server, service
|
|
1110
|
+
registrations, git identity, credential shape). One rehearsal on a second
|
|
1111
|
+
platform with a genuinely cold HOME is cheap relative to an integ cycle and
|
|
1112
|
+
converts "integ failed, cycle N+1" into "root-caused beads before integ
|
|
1113
|
+
completed". Retry policy must be category-driven: an auth failure can never
|
|
1114
|
+
be solved by retry.
|
|
1115
|
+
|
|
1116
|
+
**Outcome:** Holes 3/4 (playbook+product) fixed in 98ff2a1/ae8b870, hole
|
|
1117
|
+
5's engine piece (category-driven retry: never retry auth/trust) in
|
|
1118
|
+
5e5ab49, all pushed to the sprint branch before Cycle 4's develop rounds.
|
|
1119
|
+
Integ C3 on fleet-rev independently root-caused hole 4 with the identical
|
|
1120
|
+
diagnosis minutes later (its evidence lives on apra-fleet-eft.48; the
|
|
1121
|
+
silent retry hang is apra-fleet-eft.50), corroborating the rehearsal
|
|
1122
|
+
without either run seeing the other's analysis. Hole 5's remaining
|
|
1123
|
+
server-side pieces (interactive-path output classification,
|
|
1124
|
+
register-launch auth health check) and hole 2 (Windows schtasks leak)
|
|
1125
|
+
stay open as beads for the sprint.
|