@apralabs/apra-fleet 0.4.2 → 0.4.3
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 +38 -5
- package/dist/agents/backlog-groomer.md +25 -0
- package/dist/agents/ci-watcher.md +11 -0
- package/dist/agents/deployer.md +37 -2
- package/dist/agents/doer.md +36 -12
- package/dist/agents/harvester.md +11 -0
- package/dist/agents/integ-test-runner.md +11 -0
- package/dist/agents/kb-reconciler.md +9 -0
- package/dist/agents/plan-reviewer.md +110 -22
- package/dist/agents/planner.md +125 -37
- package/dist/agents/regression-test-runner.md +11 -0
- package/dist/agents/reviewer.md +13 -2
- package/dist/agents/schemas/plan-reviewer-output.json +37 -0
- package/dist/cli/agent-transform.d.ts +83 -2
- package/dist/cli/agent-transform.d.ts.map +1 -1
- package/dist/cli/agent-transform.js +266 -21
- package/dist/cli/agent-transform.js.map +1 -1
- package/dist/cli/auth.d.ts.map +1 -1
- package/dist/cli/auth.js +28 -9
- package/dist/cli/auth.js.map +1 -1
- package/dist/cli/config.d.ts +7 -5
- package/dist/cli/config.d.ts.map +1 -1
- package/dist/cli/config.js +44 -34
- package/dist/cli/config.js.map +1 -1
- package/dist/cli/install.d.ts +33 -0
- package/dist/cli/install.d.ts.map +1 -1
- package/dist/cli/install.js +290 -23
- package/dist/cli/install.js.map +1 -1
- package/dist/cli/kb-directives.d.ts.map +1 -1
- package/dist/cli/kb-directives.js +2 -1
- package/dist/cli/kb-directives.js.map +1 -1
- package/dist/cli/register-member.d.ts.map +1 -1
- package/dist/cli/register-member.js +6 -2
- package/dist/cli/register-member.js.map +1 -1
- package/dist/commands/kb-server.d.ts +3 -0
- package/dist/commands/kb-server.d.ts.map +1 -1
- package/dist/commands/kb-server.js +43 -3
- package/dist/commands/kb-server.js.map +1 -1
- package/dist/index.js +3 -3
- package/dist/os/linux.d.ts +6 -0
- package/dist/os/linux.d.ts.map +1 -1
- package/dist/os/linux.js +35 -2
- package/dist/os/linux.js.map +1 -1
- package/dist/os/os-commands.d.ts +61 -2
- package/dist/os/os-commands.d.ts.map +1 -1
- package/dist/os/windows-gitbash.d.ts +5 -0
- package/dist/os/windows-gitbash.d.ts.map +1 -1
- package/dist/os/windows-gitbash.js +15 -0
- package/dist/os/windows-gitbash.js.map +1 -1
- package/dist/os/windows.d.ts +6 -0
- package/dist/os/windows.d.ts.map +1 -1
- package/dist/os/windows.js +46 -3
- package/dist/os/windows.js.map +1 -1
- package/dist/providers/agy.d.ts +115 -6
- package/dist/providers/agy.d.ts.map +1 -1
- package/dist/providers/agy.js +569 -53
- package/dist/providers/agy.js.map +1 -1
- package/dist/providers/claude.d.ts +81 -3
- package/dist/providers/claude.d.ts.map +1 -1
- package/dist/providers/claude.js +344 -10
- package/dist/providers/claude.js.map +1 -1
- package/dist/providers/codex.d.ts +4 -2
- package/dist/providers/codex.d.ts.map +1 -1
- package/dist/providers/codex.js +19 -2
- package/dist/providers/codex.js.map +1 -1
- package/dist/providers/copilot.d.ts +4 -2
- package/dist/providers/copilot.d.ts.map +1 -1
- package/dist/providers/copilot.js +20 -2
- package/dist/providers/copilot.js.map +1 -1
- package/dist/providers/none.d.ts +4 -2
- package/dist/providers/none.d.ts.map +1 -1
- package/dist/providers/none.js +17 -2
- package/dist/providers/none.js.map +1 -1
- package/dist/providers/opencode.d.ts +3 -1
- package/dist/providers/opencode.d.ts.map +1 -1
- package/dist/providers/opencode.js +22 -2
- package/dist/providers/opencode.js.map +1 -1
- package/dist/providers/provider.d.ts +201 -6
- package/dist/providers/provider.d.ts.map +1 -1
- package/dist/providers/provider.js +56 -0
- package/dist/providers/provider.js.map +1 -1
- package/dist/services/agy-project.d.ts +102 -0
- package/dist/services/agy-project.d.ts.map +1 -0
- package/dist/services/agy-project.js +356 -0
- package/dist/services/agy-project.js.map +1 -0
- package/dist/services/auth-socket.js +4 -4
- package/dist/services/auth-web.js +1 -1
- package/dist/services/cloud/lifecycle.d.ts.map +1 -1
- package/dist/services/cloud/lifecycle.js +33 -4
- package/dist/services/cloud/lifecycle.js.map +1 -1
- package/dist/services/credential-cleanup.d.ts.map +1 -1
- package/dist/services/credential-cleanup.js +35 -17
- package/dist/services/credential-cleanup.js.map +1 -1
- package/dist/services/exit-drain.d.ts +44 -0
- package/dist/services/exit-drain.d.ts.map +1 -0
- package/dist/services/exit-drain.js +52 -0
- package/dist/services/exit-drain.js.map +1 -0
- package/dist/services/github-app.d.ts.map +1 -1
- package/dist/services/github-app.js +32 -4
- package/dist/services/github-app.js.map +1 -1
- package/dist/services/http-transport.d.ts +1 -0
- package/dist/services/http-transport.d.ts.map +1 -1
- package/dist/services/http-transport.js +48 -1
- package/dist/services/http-transport.js.map +1 -1
- package/dist/services/knowledge/kb-config.d.ts +42 -0
- package/dist/services/knowledge/kb-config.d.ts.map +1 -0
- package/dist/services/knowledge/kb-config.js +54 -0
- package/dist/services/knowledge/kb-config.js.map +1 -0
- package/dist/services/knowledge/kb-providers.d.ts +7 -1
- package/dist/services/knowledge/kb-providers.d.ts.map +1 -1
- package/dist/services/knowledge/kb-providers.js +191 -17
- package/dist/services/knowledge/kb-providers.js.map +1 -1
- package/dist/services/knowledge/require-sqlite-project.d.ts +32 -0
- package/dist/services/knowledge/require-sqlite-project.d.ts.map +1 -0
- package/dist/services/knowledge/require-sqlite-project.js +38 -0
- package/dist/services/knowledge/require-sqlite-project.js.map +1 -0
- package/dist/services/model-pricing.d.ts.map +1 -1
- package/dist/services/model-pricing.js +7 -3
- package/dist/services/model-pricing.js.map +1 -1
- package/dist/services/secret-token.d.ts +38 -0
- package/dist/services/secret-token.d.ts.map +1 -0
- package/dist/services/secret-token.js +51 -0
- package/dist/services/secret-token.js.map +1 -0
- package/dist/services/service-manager/types.d.ts +3 -1
- package/dist/services/service-manager/types.d.ts.map +1 -1
- package/dist/services/service-manager/windows.d.ts +14 -2
- package/dist/services/service-manager/windows.d.ts.map +1 -1
- package/dist/services/service-manager/windows.js +65 -7
- package/dist/services/service-manager/windows.js.map +1 -1
- package/dist/services/ssh.d.ts.map +1 -1
- package/dist/services/ssh.js +58 -2
- package/dist/services/ssh.js.map +1 -1
- package/dist/services/stall/read-log-tail.d.ts.map +1 -1
- package/dist/services/stall/read-log-tail.js +20 -10
- package/dist/services/stall/read-log-tail.js.map +1 -1
- package/dist/services/stall/stall-detector.d.ts +53 -4
- package/dist/services/stall/stall-detector.d.ts.map +1 -1
- package/dist/services/stall/stall-detector.js +212 -34
- package/dist/services/stall/stall-detector.js.map +1 -1
- package/dist/services/strategy.d.ts.map +1 -1
- package/dist/services/strategy.js +79 -2
- package/dist/services/strategy.js.map +1 -1
- package/dist/services/substitution-engine.js +2 -2
- package/dist/services/substitution-engine.js.map +1 -1
- package/dist/services/tool-registry.d.ts.map +1 -1
- package/dist/services/tool-registry.js +5 -3
- package/dist/services/tool-registry.js.map +1 -1
- package/dist/services/vcs/azure-devops.d.ts +66 -1
- package/dist/services/vcs/azure-devops.d.ts.map +1 -1
- package/dist/services/vcs/azure-devops.js +192 -21
- package/dist/services/vcs/azure-devops.js.map +1 -1
- package/dist/services/vcs/types.d.ts +1 -1
- package/dist/skills/fleet-sprint-cli/SKILL.md +18 -0
- package/dist/skills/pm/SKILL.md +1 -1
- package/dist/skills/pm/cost.md +0 -3
- package/dist/skills/pm/fleet-addendum.md +2 -2
- package/dist/tools/check-status.d.ts +7 -4
- package/dist/tools/check-status.d.ts.map +1 -1
- package/dist/tools/check-status.js +2 -0
- package/dist/tools/check-status.js.map +1 -1
- package/dist/tools/compose-permissions.d.ts +6 -0
- package/dist/tools/compose-permissions.d.ts.map +1 -1
- package/dist/tools/compose-permissions.js +182 -34
- package/dist/tools/compose-permissions.js.map +1 -1
- package/dist/tools/credential-store-set.js +1 -1
- package/dist/tools/execute-command.d.ts.map +1 -1
- package/dist/tools/execute-command.js +31 -20
- package/dist/tools/execute-command.js.map +1 -1
- package/dist/tools/execute-prompt.d.ts +15 -2
- package/dist/tools/execute-prompt.d.ts.map +1 -1
- package/dist/tools/execute-prompt.js +1196 -892
- package/dist/tools/execute-prompt.js.map +1 -1
- package/dist/tools/kb-capture.d.ts +2 -2
- package/dist/tools/kb-capture.d.ts.map +1 -1
- package/dist/tools/kb-capture.js +29 -5
- package/dist/tools/kb-capture.js.map +1 -1
- package/dist/tools/kb-export.d.ts +2 -2
- package/dist/tools/kb-export.d.ts.map +1 -1
- package/dist/tools/kb-export.js +3 -28
- package/dist/tools/kb-export.js.map +1 -1
- package/dist/tools/kb-feedback.d.ts.map +1 -1
- package/dist/tools/kb-feedback.js +3 -1
- package/dist/tools/kb-feedback.js.map +1 -1
- package/dist/tools/kb-freshness-sweep.d.ts.map +1 -1
- package/dist/tools/kb-freshness-sweep.js +3 -1
- package/dist/tools/kb-freshness-sweep.js.map +1 -1
- package/dist/tools/kb-import.d.ts.map +1 -1
- package/dist/tools/kb-import.js +2 -1
- package/dist/tools/kb-import.js.map +1 -1
- package/dist/tools/kb-list.d.ts.map +1 -1
- package/dist/tools/kb-list.js +3 -1
- package/dist/tools/kb-list.js.map +1 -1
- package/dist/tools/kb-query.d.ts +2 -2
- package/dist/tools/kb-reconcile-prefilter.d.ts.map +1 -1
- package/dist/tools/kb-reconcile-prefilter.js +3 -1
- package/dist/tools/kb-reconcile-prefilter.js.map +1 -1
- package/dist/tools/kb-resolve-contradiction.d.ts.map +1 -1
- package/dist/tools/kb-resolve-contradiction.js +3 -1
- package/dist/tools/kb-resolve-contradiction.js.map +1 -1
- package/dist/tools/kb-setup.d.ts.map +1 -1
- package/dist/tools/kb-setup.js +125 -20
- package/dist/tools/kb-setup.js.map +1 -1
- package/dist/tools/kb-stats.d.ts.map +1 -1
- package/dist/tools/kb-stats.js +64 -43
- package/dist/tools/kb-stats.js.map +1 -1
- package/dist/tools/list-members.d.ts +2 -2
- package/dist/tools/list-members.d.ts.map +1 -1
- package/dist/tools/list-members.js +1 -0
- package/dist/tools/list-members.js.map +1 -1
- package/dist/tools/member-detail.d.ts +2 -2
- package/dist/tools/member-detail.d.ts.map +1 -1
- package/dist/tools/member-detail.js +17 -0
- package/dist/tools/member-detail.js.map +1 -1
- package/dist/tools/member-reservation.d.ts +58 -1
- package/dist/tools/member-reservation.d.ts.map +1 -1
- package/dist/tools/member-reservation.js +61 -20
- package/dist/tools/member-reservation.js.map +1 -1
- package/dist/tools/provision-auth.d.ts +80 -1
- package/dist/tools/provision-auth.d.ts.map +1 -1
- package/dist/tools/provision-auth.js +142 -54
- package/dist/tools/provision-auth.js.map +1 -1
- package/dist/tools/provision-vcs-auth.d.ts +91 -1
- package/dist/tools/provision-vcs-auth.d.ts.map +1 -1
- package/dist/tools/provision-vcs-auth.js +209 -49
- package/dist/tools/provision-vcs-auth.js.map +1 -1
- package/dist/tools/register-member.d.ts +12 -9
- package/dist/tools/register-member.d.ts.map +1 -1
- package/dist/tools/register-member.js +99 -12
- package/dist/tools/register-member.js.map +1 -1
- package/dist/tools/remove-member.d.ts.map +1 -1
- package/dist/tools/remove-member.js +27 -2
- package/dist/tools/remove-member.js.map +1 -1
- package/dist/tools/send-email.d.ts +2 -2
- package/dist/tools/setup-git-app.d.ts.map +1 -1
- package/dist/tools/setup-git-app.js +11 -7
- package/dist/tools/setup-git-app.js.map +1 -1
- 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 +23 -1
- package/dist/tools/shutdown-server.js.map +1 -1
- package/dist/tools/update-member.d.ts +12 -9
- package/dist/tools/update-member.d.ts.map +1 -1
- package/dist/tools/update-member.js +43 -11
- package/dist/tools/update-member.js.map +1 -1
- package/dist/tools/vcs-credential-exec.d.ts +127 -0
- package/dist/tools/vcs-credential-exec.d.ts.map +1 -0
- package/dist/tools/vcs-credential-exec.js +197 -0
- package/dist/tools/vcs-credential-exec.js.map +1 -0
- package/dist/types.d.ts +13 -0
- package/dist/types.d.ts.map +1 -1
- package/dist/utils/log-helpers.d.ts.map +1 -1
- package/dist/utils/log-helpers.js +2 -2
- package/dist/utils/log-helpers.js.map +1 -1
- package/dist/utils/prompt-errors.d.ts.map +1 -1
- package/dist/utils/prompt-errors.js +7 -1
- package/dist/utils/prompt-errors.js.map +1 -1
- package/dist/utils/shell-escape.d.ts +16 -0
- package/dist/utils/shell-escape.d.ts.map +1 -1
- package/dist/utils/shell-escape.js +22 -2
- package/dist/utils/shell-escape.js.map +1 -1
- package/dist/utils/vcs-provider-detect.d.ts +65 -0
- package/dist/utils/vcs-provider-detect.d.ts.map +1 -0
- package/dist/utils/vcs-provider-detect.js +109 -0
- package/dist/utils/vcs-provider-detect.js.map +1 -0
- package/dist/utils/workspace-trust.d.ts +20 -17
- package/dist/utils/workspace-trust.d.ts.map +1 -1
- package/dist/utils/workspace-trust.js +87 -1
- package/dist/utils/workspace-trust.js.map +1 -1
- package/dist/workflows/auto-sprint.js +176 -15
- package/package.json +6 -4
- package/packages/apra-fleet-client/src/client/api.mjs +319 -17
- package/packages/apra-fleet-client/src/client/errors.mjs +6 -2
- package/packages/apra-fleet-se/apra-pm/agents/_shared/GRAPH-SEMANTICS.md +10 -2
- package/packages/apra-fleet-se/apra-pm/agents/backlog-groomer.md +25 -0
- package/packages/apra-fleet-se/apra-pm/agents/ci-watcher.md +11 -0
- package/packages/apra-fleet-se/apra-pm/agents/deployer.md +37 -2
- package/packages/apra-fleet-se/apra-pm/agents/doer.md +36 -12
- package/packages/apra-fleet-se/apra-pm/agents/harvester.md +11 -0
- package/packages/apra-fleet-se/apra-pm/agents/integ-test-runner.md +11 -0
- package/packages/apra-fleet-se/apra-pm/agents/kb-reconciler.md +9 -0
- package/packages/apra-fleet-se/apra-pm/agents/plan-reviewer.md +110 -22
- package/packages/apra-fleet-se/apra-pm/agents/planner.md +125 -37
- package/packages/apra-fleet-se/apra-pm/agents/regression-test-runner.md +11 -0
- package/packages/apra-fleet-se/apra-pm/agents/reviewer.md +13 -2
- package/packages/apra-fleet-se/apra-pm/agents/schemas/plan-reviewer-output.json +37 -0
- package/packages/apra-fleet-se/apra-pm/skills/pm/SKILL.md +1 -1
- package/packages/apra-fleet-se/apra-pm/skills/pm/cost.md +0 -3
- package/packages/apra-fleet-se/apra-pm/skills/pm/fleet-addendum.md +2 -2
- package/packages/apra-fleet-se/bin/cli.mjs +149 -1
- package/packages/apra-fleet-se/bin/serve.mjs +78 -4
- package/packages/apra-fleet-se/fleet-sprint/abort.mjs +481 -0
- package/packages/apra-fleet-se/fleet-sprint/beads-children.mjs +491 -0
- package/packages/apra-fleet-se/fleet-sprint/beads-identity-check.mjs +282 -0
- package/packages/apra-fleet-se/fleet-sprint/beads-identity.mjs +208 -0
- package/packages/apra-fleet-se/fleet-sprint/beads-scope.mjs +638 -0
- package/packages/apra-fleet-se/fleet-sprint/beads-transitions.mjs +301 -0
- package/packages/apra-fleet-se/fleet-sprint/branch-ensure.mjs +96 -0
- package/packages/apra-fleet-se/fleet-sprint/conflict-ladder.mjs +8 -0
- package/packages/apra-fleet-se/fleet-sprint/contracts.mjs +61 -14
- package/packages/apra-fleet-se/fleet-sprint/coordination.mjs +516 -0
- package/packages/apra-fleet-se/fleet-sprint/dispatch-failure.mjs +184 -0
- package/packages/apra-fleet-se/fleet-sprint/dispatch-role.mjs +1121 -0
- package/packages/apra-fleet-se/fleet-sprint/dispatch-safety-guard.mjs +216 -4
- package/packages/apra-fleet-se/fleet-sprint/docs/README.md +1 -1
- package/packages/apra-fleet-se/fleet-sprint/docs/dolt-sync-redesign.md +276 -0
- package/packages/apra-fleet-se/fleet-sprint/docs/escalate-to-llm-design.md +1 -1
- package/packages/apra-fleet-se/fleet-sprint/docs/run-sprint-cycle-prelude-decision.md +215 -0
- package/packages/apra-fleet-se/fleet-sprint/dolt-literal-guard.mjs +56 -1
- package/packages/apra-fleet-se/fleet-sprint/dolt-sync.mjs +996 -32
- package/packages/apra-fleet-se/fleet-sprint/errors.mjs +267 -0
- package/packages/apra-fleet-se/fleet-sprint/explicit-id-create-guard.mjs +501 -0
- package/packages/apra-fleet-se/fleet-sprint/fatal-diagnostics.mjs +134 -0
- package/packages/apra-fleet-se/fleet-sprint/full-db-fetch-guard.mjs +102 -0
- package/packages/apra-fleet-se/fleet-sprint/git-sync.mjs +556 -0
- package/packages/apra-fleet-se/fleet-sprint/git-topology.mjs +495 -0
- package/packages/apra-fleet-se/fleet-sprint/guarded-modules.mjs +564 -0
- package/packages/apra-fleet-se/fleet-sprint/inline-ladder-guard.mjs +216 -0
- package/packages/apra-fleet-se/fleet-sprint/kb.mjs +728 -0
- package/packages/apra-fleet-se/fleet-sprint/mcp-result.mjs +75 -0
- package/packages/apra-fleet-se/fleet-sprint/member-provisioning.mjs +252 -0
- package/packages/apra-fleet-se/fleet-sprint/member-sync.mjs +581 -0
- package/packages/apra-fleet-se/fleet-sprint/member-target.mjs +93 -0
- package/packages/apra-fleet-se/fleet-sprint/newtask-text.mjs +266 -0
- package/packages/apra-fleet-se/fleet-sprint/parent-notes-staleness.mjs +201 -0
- package/packages/apra-fleet-se/fleet-sprint/phases/deploy.mjs +136 -0
- package/packages/apra-fleet-se/fleet-sprint/phases/develop.mjs +822 -0
- package/packages/apra-fleet-se/fleet-sprint/phases/ensure-sprint-branch.mjs +294 -0
- package/packages/apra-fleet-se/fleet-sprint/phases/final-review.mjs +403 -0
- package/packages/apra-fleet-se/fleet-sprint/phases/harvest.mjs +180 -0
- package/packages/apra-fleet-se/fleet-sprint/phases/integ-test.mjs +358 -0
- package/packages/apra-fleet-se/fleet-sprint/phases/plan.mjs +390 -0
- package/packages/apra-fleet-se/fleet-sprint/phases/publish-pr.mjs +323 -0
- package/packages/apra-fleet-se/fleet-sprint/phases/re-review.mjs +232 -0
- package/packages/apra-fleet-se/fleet-sprint/phases/regression-test.mjs +174 -0
- package/packages/apra-fleet-se/fleet-sprint/phases/replan.mjs +197 -0
- package/packages/apra-fleet-se/fleet-sprint/phases/review.mjs +305 -0
- package/packages/apra-fleet-se/fleet-sprint/prompts.mjs +559 -0
- package/packages/apra-fleet-se/fleet-sprint/role-policies.mjs +1416 -0
- package/packages/apra-fleet-se/fleet-sprint/round-session.mjs +100 -0
- package/packages/apra-fleet-se/fleet-sprint/runner.js +1467 -9798
- package/packages/apra-fleet-se/fleet-sprint/se-posix.mjs +17 -12
- package/packages/apra-fleet-se/fleet-sprint/shell-command-guard.mjs +304 -0
- package/packages/apra-fleet-se/fleet-sprint/skills/fleet-sprint-cli/SKILL.md +18 -0
- package/packages/apra-fleet-se/fleet-sprint/skills/fleet-supervisor/SKILL.md +21 -2
- package/packages/apra-fleet-se/fleet-sprint/sprint-args.mjs +454 -0
- package/packages/apra-fleet-se/fleet-sprint/sprint-report.mjs +239 -0
- package/packages/apra-fleet-se/fleet-sprint/sprint-state.mjs +163 -0
- package/packages/apra-fleet-se/fleet-sprint/unbracketed-push-guard.mjs +234 -0
- package/packages/apra-fleet-se/fleet-sprint/usage-limit-controller.mjs +211 -0
- package/packages/apra-fleet-se/fleet-sprint/vcs-auth.mjs +1279 -0
- package/packages/apra-fleet-se/fleet-sprint/vcs-module.mjs +153 -19
- package/packages/apra-fleet-se/fleet-sprint/vcs-providers/azure-devops.mjs +71 -21
- package/packages/apra-fleet-se/fleet-sprint/vcs-providers/bitbucket.mjs +28 -0
- package/packages/apra-fleet-se/fleet-sprint/vcs-providers/github.mjs +131 -0
- package/packages/apra-fleet-se/fleet-sprint/vcs-providers/index.mjs +140 -16
- package/packages/apra-fleet-se/fleet-sprint/vcs-providers/shell-helpers.mjs +1 -0
- package/packages/apra-fleet-se/fleet-sprint/viewer-extensions.mjs +77 -2
- package/packages/apra-fleet-se/fleet-sprint/worklists.mjs +559 -0
- package/packages/apra-fleet-se/package.json +1 -1
- package/packages/apra-fleet-workflow/src/viewer/index.mjs +17 -6
- package/packages/apra-fleet-workflow/src/workflow/index.mjs +49 -7
- package/skills/fleet/SKILL.md +8 -8
- package/skills/fleet/auth-azdevops.md +14 -3
- package/skills/fleet/auth-bitbucket.md +2 -2
- package/skills/fleet/auth-github.md +5 -4
- package/skills/fleet/onboarding.md +3 -3
- package/skills/fleet/permissions.md +5 -1
- package/skills/fleet/profiles/node.json +1 -1
- package/skills/fleet/troubleshooting.md +3 -3
- package/version.json +1 -1
|
@@ -0,0 +1,1121 @@
|
|
|
1
|
+
import { policyFor, PRE_DISPATCH_STEPS_IN_BRACKET } from './role-policies.mjs';
|
|
2
|
+
import {
|
|
3
|
+
isNonRetryableDispatchError,
|
|
4
|
+
isAuthDispatchError,
|
|
5
|
+
isPostDispatchSyncFailure,
|
|
6
|
+
isInfraDispatchFailure,
|
|
7
|
+
isUsageLimitDispatchError,
|
|
8
|
+
usageLimitOf,
|
|
9
|
+
} from './errors.mjs';
|
|
10
|
+
import { AgentOutputError, AgentDispatchError, FleetTransportError, WorkflowError } from '@apralabs/apra-fleet-workflow';
|
|
11
|
+
|
|
12
|
+
// =============================================================================
|
|
13
|
+
// apra-fleet-3swo.5.3 -- dispatchRole(ctx, roleName, opts): the ONE dispatch
|
|
14
|
+
// engine that executes a role's ladder from fleet-sprint/role-policies.mjs's
|
|
15
|
+
// data table, replacing the hand-written agent() ladders runner.js used to
|
|
16
|
+
// spell out once per role.
|
|
17
|
+
//
|
|
18
|
+
// WHAT MOVED HERE: for every role role-policies.mjs marks `migrated: true`,
|
|
19
|
+
// this module owns the whole ladder -- the agent() dispatch itself, the
|
|
20
|
+
// withGitSync(...) bracket around it, the withDispatchWatchdog(...) race, the
|
|
21
|
+
// max_turns-exhaustion resume, the bounded retry ladder (backoff, LLM-auth
|
|
22
|
+
// self-heal, non-retryable abort, post-dispatch-sync-failure abort), the
|
|
23
|
+
// bounded semantic-repair re-ask, and the degrade that decides what the
|
|
24
|
+
// caller gets back once the attempts are spent. runner.js keeps only what is
|
|
25
|
+
// genuinely NOT policy: the prompt text, the presentation labels, the runtime
|
|
26
|
+
// bindings a policy names, and what the caller does with the outcome.
|
|
27
|
+
//
|
|
28
|
+
// apra-fleet-3swo.5.4 finished the job on the DEGRADE side: the failure TEXT
|
|
29
|
+
// a degraded value carries used to arrive as a per-call `synthesizedNotes`
|
|
30
|
+
// map of note builders, i.e. six runner.js call sites each spelling out the
|
|
31
|
+
// same two sentences. It is now `degrade.noteTemplates` data in
|
|
32
|
+
// role-policies.mjs, rendered by renderDegradeNote() below -- ONE
|
|
33
|
+
// implementation of verdict/report fallback text for every role.
|
|
34
|
+
//
|
|
35
|
+
// WHY IT IS NOT A PASS-THROUGH: a dispatchRole that merely wrapped the
|
|
36
|
+
// original inline ladders would leave every behavioural axis duplicated per
|
|
37
|
+
// role, which is the condition role-policies.mjs was built to remove. There
|
|
38
|
+
// is exactly ONE agent() call site in this file, ONE withGitSync(...) bracket
|
|
39
|
+
// call and ONE withDispatchWatchdog(...) race; every per-role difference
|
|
40
|
+
// between them is read out of the policy row rather than written out again.
|
|
41
|
+
// fleet-sprint/inline-ladder-guard.mjs enforces the other half of that: a
|
|
42
|
+
// migrated role that still has an inline agent() ladder anywhere in the
|
|
43
|
+
// guarded-module set is a violation.
|
|
44
|
+
//
|
|
45
|
+
// EVERYTHING RUNNER-SIDE IS INJECTED, NEVER IMPORTED (same discipline as
|
|
46
|
+
// git-sync.mjs): agent(), withGitSync(), withDispatchWatchdog(), the member
|
|
47
|
+
// resolver, the session guard, the auth self-heal hook, the model-tier map,
|
|
48
|
+
// the budget map and the sprint's log() all arrive through `ctx`. This module
|
|
49
|
+
// deliberately never imports runner.js back -- runner.js imports it, so
|
|
50
|
+
// importing runner here would be a module cycle.
|
|
51
|
+
//
|
|
52
|
+
// TURN BASES live here rather than in runner.js because the dispatch that
|
|
53
|
+
// consumes them lives here. They keep their original runner.js NAMES
|
|
54
|
+
// (PLANNER_MAX_TURNS, ...) because role-policies.mjs records each policy's
|
|
55
|
+
// turn budget symbolically, by the name of the constant that supplies it --
|
|
56
|
+
// test/role-policies-table.test.mjs re-derives `maxTurns.value` by looking
|
|
57
|
+
// that name up in real source, so renaming one here would break that proof
|
|
58
|
+
// rather than silently drift.
|
|
59
|
+
// =============================================================================
|
|
60
|
+
|
|
61
|
+
// -----------------------------------------------------------------------------
|
|
62
|
+
// Turn bases. One per migrated ladder, named exactly as role-policies.mjs's
|
|
63
|
+
// `maxTurns.base` records them.
|
|
64
|
+
// -----------------------------------------------------------------------------
|
|
65
|
+
|
|
66
|
+
/** Planner turn base: it builds the whole epic DAG, so it gets a doer-sized budget. */
|
|
67
|
+
const PLANNER_MAX_TURNS = 500;
|
|
68
|
+
/** Plan-reviewer turn base: reviewer-sized, like every other review dispatch. */
|
|
69
|
+
const PLAN_REVIEWER_MAX_TURNS = 500;
|
|
70
|
+
/** Scoped replan planner turn base. */
|
|
71
|
+
const SCOPED_REPLAN_PLANNER_MAX_TURNS = 500;
|
|
72
|
+
/** Scoped replan plan-reviewer turn base. */
|
|
73
|
+
const SCOPED_REPLAN_REVIEWER_MAX_TURNS = 500;
|
|
74
|
+
/** Harvester turn base: it writes docs/changelog across the whole epic. */
|
|
75
|
+
const HARVESTER_MAX_TURNS = 500;
|
|
76
|
+
/** Deployer turn base: it runs real deploy commands per a runbook. */
|
|
77
|
+
const DEPLOYER_MAX_TURNS = 500;
|
|
78
|
+
/**
|
|
79
|
+
* Regression turn base: the real functional suite alone spends roughly one
|
|
80
|
+
* turn per liveness poll for the better part of an hour, and this single
|
|
81
|
+
* dispatch carries both it and the sandbox smoke sprint.
|
|
82
|
+
*/
|
|
83
|
+
const REGRESSION_TEST_MAX_TURNS = 500;
|
|
84
|
+
/**
|
|
85
|
+
* Integ turn base: a compliant runner spends ~1 turn per liveness poll, but a
|
|
86
|
+
* real per-feature pass also spends fast, sub-poll turns, so the WALL-CLOCK
|
|
87
|
+
* ceiling (max_total_s) is what should bind first on any progressing run.
|
|
88
|
+
*/
|
|
89
|
+
const INTEG_TEST_MAX_TURNS = 500;
|
|
90
|
+
/**
|
|
91
|
+
* Final review turn base: it reads the full diff/evidence across an entire
|
|
92
|
+
* epic's worth of closed tasks, and a timeout here flips a whole sprint's
|
|
93
|
+
* outcome to FAIL.
|
|
94
|
+
*/
|
|
95
|
+
const FINAL_REVIEW_MAX_TURNS = 500;
|
|
96
|
+
/**
|
|
97
|
+
* Per-round reviewer turn base: a full-cycle review can genuinely exhaust the
|
|
98
|
+
* fleet's default turn budget, and a fresh retry deterministically hits the
|
|
99
|
+
* same wall -- hence the explicit budget plus the same-session resume.
|
|
100
|
+
*/
|
|
101
|
+
const BASE_REVIEWER_MAX_TURNS = 500;
|
|
102
|
+
/**
|
|
103
|
+
* Doer streak turn base. A streak runs a full impl+test+commit cycle, which is
|
|
104
|
+
* categorically heavier than a one-shot prompt; the resume ladder escalates
|
|
105
|
+
* from here by doubling, bounded by the row's resumeAttempts.
|
|
106
|
+
*/
|
|
107
|
+
const BASE_DOER_MAX_TURNS = 500;
|
|
108
|
+
|
|
109
|
+
/**
|
|
110
|
+
* Every turn-base constant a policy's `maxTurns.base` may name, keyed by that
|
|
111
|
+
* name. A policy naming a base absent from this map is a table/engine
|
|
112
|
+
* mismatch and throws at dispatch time rather than silently dispatching with
|
|
113
|
+
* no turn budget at all.
|
|
114
|
+
*/
|
|
115
|
+
export const TURN_BASES = Object.freeze({
|
|
116
|
+
PLANNER_MAX_TURNS,
|
|
117
|
+
PLAN_REVIEWER_MAX_TURNS,
|
|
118
|
+
SCOPED_REPLAN_PLANNER_MAX_TURNS,
|
|
119
|
+
SCOPED_REPLAN_REVIEWER_MAX_TURNS,
|
|
120
|
+
HARVESTER_MAX_TURNS,
|
|
121
|
+
DEPLOYER_MAX_TURNS,
|
|
122
|
+
REGRESSION_TEST_MAX_TURNS,
|
|
123
|
+
INTEG_TEST_MAX_TURNS,
|
|
124
|
+
FINAL_REVIEW_MAX_TURNS,
|
|
125
|
+
BASE_REVIEWER_MAX_TURNS,
|
|
126
|
+
BASE_DOER_MAX_TURNS,
|
|
127
|
+
});
|
|
128
|
+
|
|
129
|
+
/**
|
|
130
|
+
* The ladder's real timed backoff sleeps exist purely for production
|
|
131
|
+
* busy-lock resilience; a hermetic mock run has no busy-lock to wait out, so
|
|
132
|
+
* the harness sets this flag to exercise the full ladder LOGIC with zero
|
|
133
|
+
* wall-clock. The delay values and the "waiting Ns" log line are unchanged
|
|
134
|
+
* either way, so observable behaviour matches. Same env var the planner's
|
|
135
|
+
* inline ladder read before this migration.
|
|
136
|
+
*/
|
|
137
|
+
function instantRetryBackoff() {
|
|
138
|
+
return process.env.APRA_FLEET_MOCK_INSTANT_RETRY_BACKOFF === '1';
|
|
139
|
+
}
|
|
140
|
+
|
|
141
|
+
// -----------------------------------------------------------------------------
|
|
142
|
+
// Policy -> runtime resolution. Each function turns ONE policy field into the
|
|
143
|
+
// value the dispatch actually needs, and throws (rather than defaulting) when
|
|
144
|
+
// the table names something the engine cannot resolve.
|
|
145
|
+
// -----------------------------------------------------------------------------
|
|
146
|
+
|
|
147
|
+
/**
|
|
148
|
+
* The member a dispatch routes to, resolved from `policy.member`:
|
|
149
|
+
* 'role' -- ctx.getMemberForRole(member.role)
|
|
150
|
+
* 'pool-head' -- bindings[member.binding], a runner-local pool head
|
|
151
|
+
* 'runtime' -- bindings[member.binding], a runner-local binding
|
|
152
|
+
*
|
|
153
|
+
* @param {object} ctx
|
|
154
|
+
* @param {{kind:string, role?:string, binding?:string}} member
|
|
155
|
+
* @param {Record<string, any>} bindings
|
|
156
|
+
* @returns {string}
|
|
157
|
+
*/
|
|
158
|
+
export function resolveMember(ctx, member, bindings = {}) {
|
|
159
|
+
if (member && member.kind === 'role') return ctx.getMemberForRole(member.role);
|
|
160
|
+
if (member && (member.kind === 'pool-head' || member.kind === 'runtime')) {
|
|
161
|
+
const value = bindings[member.binding];
|
|
162
|
+
if (value === undefined) {
|
|
163
|
+
throw new Error(
|
|
164
|
+
`dispatch-role: member binding '${member.binding}' was not supplied in opts.bindings -- a ` +
|
|
165
|
+
`'${member.kind}'-kind member is a runner-local value the engine cannot resolve on its own.`
|
|
166
|
+
);
|
|
167
|
+
}
|
|
168
|
+
return value;
|
|
169
|
+
}
|
|
170
|
+
throw new Error(`dispatch-role: unresolvable member kind ${JSON.stringify(member && member.kind)}.`);
|
|
171
|
+
}
|
|
172
|
+
|
|
173
|
+
/**
|
|
174
|
+
* The model tier a dispatch is priced at, resolved from `policy.model`:
|
|
175
|
+
* 'fixed' -- ctx.fixedRoleTier[model.key]
|
|
176
|
+
* 'per-bead' -- bindings[model.binding] (the doer's declared tier)
|
|
177
|
+
* 'inherited' -- undefined: a resume of an already-priced dispatch passes
|
|
178
|
+
* no tier of its own.
|
|
179
|
+
*/
|
|
180
|
+
export function resolveModelTier(ctx, model, bindings = {}) {
|
|
181
|
+
if (!model) return undefined;
|
|
182
|
+
if (model.kind === 'fixed') return ctx.fixedRoleTier[model.key];
|
|
183
|
+
if (model.kind === 'per-bead') return bindings[model.binding];
|
|
184
|
+
if (model.kind === 'inherited') return undefined;
|
|
185
|
+
throw new Error(`dispatch-role: unresolvable model kind ${JSON.stringify(model.kind)}.`);
|
|
186
|
+
}
|
|
187
|
+
|
|
188
|
+
/**
|
|
189
|
+
* A budget (`timeouts.timeoutS` / `timeouts.maxTotalS`) resolved from its
|
|
190
|
+
* SYMBOLIC name against ctx.budgets. null means "pass nothing and take the
|
|
191
|
+
* transport default", which is a policy value in its own right (streak
|
|
192
|
+
* assignment), not a missing one.
|
|
193
|
+
*/
|
|
194
|
+
export function resolveBudget(ctx, name) {
|
|
195
|
+
if (name === null || name === undefined) return undefined;
|
|
196
|
+
const value = ctx.budgets ? ctx.budgets[name] : undefined;
|
|
197
|
+
if (value === undefined) {
|
|
198
|
+
throw new Error(`dispatch-role: policy names budget '${name}', which ctx.budgets does not supply.`);
|
|
199
|
+
}
|
|
200
|
+
return value;
|
|
201
|
+
}
|
|
202
|
+
|
|
203
|
+
/**
|
|
204
|
+
* The turn budget a dispatch passes, resolved from `policy.maxTurns`:
|
|
205
|
+
* null -- no max_turns at all (the transport default)
|
|
206
|
+
* 'constant' -- TURN_BASES[base] * multiplier
|
|
207
|
+
* 'runtime' -- bindings[binding], computed per resume attempt
|
|
208
|
+
*/
|
|
209
|
+
export function resolveMaxTurns(maxTurns, bindings = {}) {
|
|
210
|
+
if (!maxTurns) return undefined;
|
|
211
|
+
if (maxTurns.kind === 'runtime') return bindings[maxTurns.binding];
|
|
212
|
+
return turnBaseValue(maxTurns.base) * maxTurns.multiplier;
|
|
213
|
+
}
|
|
214
|
+
|
|
215
|
+
function turnBaseValue(name) {
|
|
216
|
+
const base = TURN_BASES[name];
|
|
217
|
+
if (typeof base !== 'number') {
|
|
218
|
+
throw new Error(
|
|
219
|
+
`dispatch-role: policy names turn base '${name}', which is not one of TURN_BASES ` +
|
|
220
|
+
`(${Object.keys(TURN_BASES).join(', ')}).`
|
|
221
|
+
);
|
|
222
|
+
}
|
|
223
|
+
return base;
|
|
224
|
+
}
|
|
225
|
+
|
|
226
|
+
/**
|
|
227
|
+
* The resume argument, resolved from `policy.resumeArg`:
|
|
228
|
+
* 'same-session' -- true (an in-dispatch continuation of the session just run)
|
|
229
|
+
* 'fresh-session' -- false (a self-contained one-off instruction)
|
|
230
|
+
* 'round-session' -- opts.resumeArg, the runner's per-round session lookup
|
|
231
|
+
* 'worklist' -- opts.resumeArg, the doer's per-worklist resume argument
|
|
232
|
+
* null/absent -- false, NOT undefined (see below)
|
|
233
|
+
*
|
|
234
|
+
* NEVER RETURNS `undefined`. The dispatch below strips undefined options, so
|
|
235
|
+
* an undefined here means the `resume` field is omitted from the payload
|
|
236
|
+
* entirely -- and an omitted `resume` is not neutral: execute_prompt's own
|
|
237
|
+
* default is `resume: true`, a best-effort reattachment to whatever session
|
|
238
|
+
* that member last ran. A dispatch that never declared an intent to continue
|
|
239
|
+
* anything would then silently inherit an unrelated, possibly enormous prior
|
|
240
|
+
* context. Falling back to `false` states the safe intent explicitly instead
|
|
241
|
+
* of letting the transport chain decide it. Every row of the policy table
|
|
242
|
+
* also names its own resumeArg, so this fallback is a backstop for a future
|
|
243
|
+
* row that forgets, not the table's working default.
|
|
244
|
+
*/
|
|
245
|
+
export function resolveResumeArg(resumeArg, opts) {
|
|
246
|
+
if (!resumeArg) return false;
|
|
247
|
+
if (resumeArg.kind === 'same-session') return true;
|
|
248
|
+
if (resumeArg.kind === 'fresh-session') return false;
|
|
249
|
+
// A runner-supplied argument: already `false` or an explicit session id at
|
|
250
|
+
// every call site today. `?? false` keeps the never-undefined guarantee if
|
|
251
|
+
// a caller ever omits it.
|
|
252
|
+
return opts.resumeArg ?? false;
|
|
253
|
+
}
|
|
254
|
+
|
|
255
|
+
/**
|
|
256
|
+
* A watchdog label, rebuilt from `policy.watchdog.label`'s segment list: a
|
|
257
|
+
* plain string segment is literal text, an `{ expr }` segment is a runner
|
|
258
|
+
* expression. The only expressions the table uses are turn-base arithmetic
|
|
259
|
+
* (`PLANNER_MAX_TURNS * 2`), evaluated against TURN_BASES, so the watchdog's
|
|
260
|
+
* kill-path identity stays owned by the policy table rather than passed in
|
|
261
|
+
* as an opaque caller string.
|
|
262
|
+
*/
|
|
263
|
+
export function resolveWatchdogLabel(segments, bindings = {}) {
|
|
264
|
+
return segments
|
|
265
|
+
.map((segment) => {
|
|
266
|
+
if (typeof segment === 'string') return segment;
|
|
267
|
+
return String(evalLabelExpr(segment.expr, bindings));
|
|
268
|
+
})
|
|
269
|
+
.join('');
|
|
270
|
+
}
|
|
271
|
+
|
|
272
|
+
function evalLabelExpr(expr, bindings) {
|
|
273
|
+
if (Object.prototype.hasOwnProperty.call(bindings, expr)) return bindings[expr];
|
|
274
|
+
const m = /^([A-Za-z_$][\w$]*)(?:\s*\*\s*(\d+))?$/.exec(String(expr).trim());
|
|
275
|
+
if (m && Object.prototype.hasOwnProperty.call(TURN_BASES, m[1])) {
|
|
276
|
+
return m[2] ? TURN_BASES[m[1]] * Number(m[2]) : TURN_BASES[m[1]];
|
|
277
|
+
}
|
|
278
|
+
throw new Error(
|
|
279
|
+
`dispatch-role: watchdog label expression ${JSON.stringify(expr)} is neither a supplied binding nor ` +
|
|
280
|
+
'turn-base arithmetic over TURN_BASES.'
|
|
281
|
+
);
|
|
282
|
+
}
|
|
283
|
+
|
|
284
|
+
/**
|
|
285
|
+
* The schema object a dispatch validates its output against. Policies record
|
|
286
|
+
* a schema by NAME ('planReviewerVerdict'); ctx.schemas maps that name to the
|
|
287
|
+
* real contracts.mjs schema, so this module never imports the contracts and a
|
|
288
|
+
* policy naming an unknown schema fails loudly instead of dispatching
|
|
289
|
+
* unvalidated.
|
|
290
|
+
*/
|
|
291
|
+
function resolveSchema(ctx, schemaName) {
|
|
292
|
+
if (!schemaName) return undefined;
|
|
293
|
+
const schema = ctx.schemas ? ctx.schemas[schemaName] : undefined;
|
|
294
|
+
if (!schema) {
|
|
295
|
+
throw new Error(`dispatch-role: policy names schema '${schemaName}', which ctx.schemas does not supply.`);
|
|
296
|
+
}
|
|
297
|
+
return schema;
|
|
298
|
+
}
|
|
299
|
+
|
|
300
|
+
// -----------------------------------------------------------------------------
|
|
301
|
+
// Degrade: what the caller gets back once a ladder's attempts are spent, per
|
|
302
|
+
// role-policies.mjs's DEGRADE_KINDS.
|
|
303
|
+
// -----------------------------------------------------------------------------
|
|
304
|
+
|
|
305
|
+
/** Every constructor NAME on an error's prototype chain, nearest first. */
|
|
306
|
+
function errorClassNames(err) {
|
|
307
|
+
const names = [];
|
|
308
|
+
for (let proto = err; proto; proto = Object.getPrototypeOf(proto)) {
|
|
309
|
+
const name = proto.constructor && proto.constructor.name;
|
|
310
|
+
if (name) names.push(name);
|
|
311
|
+
if (name === 'Error') break;
|
|
312
|
+
}
|
|
313
|
+
return names;
|
|
314
|
+
}
|
|
315
|
+
|
|
316
|
+
/**
|
|
317
|
+
* Classifies a ladder failure into one of role-policies.mjs's ERROR_CLASSES.
|
|
318
|
+
* The first two are universal:
|
|
319
|
+
* 'schema' -- AgentOutputError: agent()'s own bounded schema-repair loop
|
|
320
|
+
* was exhausted, so no schema-valid output ever arrived;
|
|
321
|
+
* 'dispatch' -- AgentDispatchError/FleetTransportError: the dispatch channel
|
|
322
|
+
* itself failed (a dropped transport is exactly as transient
|
|
323
|
+
* and non-schema as a dispatch error -- neither may abort the
|
|
324
|
+
* whole sprint);
|
|
325
|
+
* null -- anything else, i.e. an unrecognised error class.
|
|
326
|
+
*
|
|
327
|
+
* The remaining three exist only for the roles whose POLICY asks for them,
|
|
328
|
+
* which is what keeps them variances-as-data rather than engine branches:
|
|
329
|
+
* 'infra' -- `degrade.classifiesInfraFailures` (the integ runner): a
|
|
330
|
+
* dispatch that produced no result envelope at all is not a
|
|
331
|
+
* test verdict and must never be recorded as one.
|
|
332
|
+
* 'dispatch' -- also any class named in `degrade.extraDispatchErrors` (the
|
|
333
|
+
* per-round reviewer's own read-side sync-bracket failures).
|
|
334
|
+
* 'sync' -- `degrade.classifiesSyncFailures` (the regression phase):
|
|
335
|
+
* the git/beads sync AROUND the dispatch failed, which for
|
|
336
|
+
* the one phase whose whole job is mutating beads is a
|
|
337
|
+
* routine outcome worth reporting distinctly.
|
|
338
|
+
* 'unknown' -- `degrade.classifiesUnrecognisedErrors` (the regression
|
|
339
|
+
* catch-all): even an unrecognised class gets a degrade path
|
|
340
|
+
* rather than aborting an informational phase.
|
|
341
|
+
*
|
|
342
|
+
* Called with no policy it behaves exactly as the two-class version did.
|
|
343
|
+
*
|
|
344
|
+
* @param {Error} err
|
|
345
|
+
* @param {object} [policy] the ROLE_POLICIES row whose degrade is in force
|
|
346
|
+
*/
|
|
347
|
+
export function classifyLadderError(err, policy) {
|
|
348
|
+
const degrade = policy ? policy.degrade : null;
|
|
349
|
+
if (err instanceof AgentOutputError) return 'schema';
|
|
350
|
+
// BEFORE the generic dispatch class: an infra failure IS an
|
|
351
|
+
// AgentDispatchError, and the whole point of the class is telling the two
|
|
352
|
+
// apart for the one role that must not conflate them.
|
|
353
|
+
if (degrade && degrade.classifiesInfraFailures
|
|
354
|
+
&& err instanceof AgentDispatchError && isInfraDispatchFailure(err)) {
|
|
355
|
+
return 'infra';
|
|
356
|
+
}
|
|
357
|
+
if (err instanceof AgentDispatchError || err instanceof FleetTransportError) return 'dispatch';
|
|
358
|
+
if (degrade && degrade.extraDispatchErrors.length > 0) {
|
|
359
|
+
const names = errorClassNames(err);
|
|
360
|
+
if (degrade.extraDispatchErrors.some((name) => names.includes(name))) return 'dispatch';
|
|
361
|
+
}
|
|
362
|
+
if (degrade && degrade.classifiesSyncFailures
|
|
363
|
+
&& (isPostDispatchSyncFailure(err) || err instanceof WorkflowError)) {
|
|
364
|
+
return 'sync';
|
|
365
|
+
}
|
|
366
|
+
if (degrade && degrade.classifiesUnrecognisedErrors) return 'unknown';
|
|
367
|
+
return null;
|
|
368
|
+
}
|
|
369
|
+
|
|
370
|
+
/**
|
|
371
|
+
* Thrown by the engine itself when a SCHEMA-VALID result is rejected by the
|
|
372
|
+
* caller's own semantic validator under `retry.retryOnInvalidResult`. It is
|
|
373
|
+
* never surfaced to the caller: the ladder either retries the attempt or
|
|
374
|
+
* converts it through `opts.onResultRejected`.
|
|
375
|
+
*/
|
|
376
|
+
export class LadderResultRejectedError extends Error {
|
|
377
|
+
constructor(reason) {
|
|
378
|
+
super(`dispatch-role: the ladder's result was rejected by its own validator (${reason}).`);
|
|
379
|
+
this.name = 'LadderResultRejectedError';
|
|
380
|
+
this.reason = reason;
|
|
381
|
+
}
|
|
382
|
+
}
|
|
383
|
+
|
|
384
|
+
/**
|
|
385
|
+
* The value a degrade fabricates for one failed attempt, from
|
|
386
|
+
* `policy.degrade`. Only 'synthesized-verdict' fabricates anything on the
|
|
387
|
+
* planning side: 'fallback-value', 'defer-to-next-cycle' and 'non-approval'
|
|
388
|
+
* all hand the caller `null` and let it apply its own deterministic fallback
|
|
389
|
+
* (selectStreaks' one-bead-per-streak grouping; deferring the flagged beads
|
|
390
|
+
* to the next cycle; treating the scoped round as not approved).
|
|
391
|
+
*
|
|
392
|
+
* `degrade.marker` is stamped on every synthesized value so a degraded result
|
|
393
|
+
* is never mistaken for a genuine one, and `degrade.neverSynthesizes` is what
|
|
394
|
+
* role-policies-table.test.mjs holds this shape to: no degrade path anywhere
|
|
395
|
+
* may fabricate an approval.
|
|
396
|
+
*/
|
|
397
|
+
function synthesizeDegradedValue(policy, opts, err, errorClass) {
|
|
398
|
+
const degrade = policy.degrade;
|
|
399
|
+
// `degrade.classes` -- not the degrade KIND -- decides whether this
|
|
400
|
+
// ladder fabricates anything for THIS error class. A ladder that
|
|
401
|
+
// recognises a class but does not list it (streak assignment's
|
|
402
|
+
// 'fallback-value', the scoped replan's 'defer-to-next-cycle') hands the
|
|
403
|
+
// caller null on purpose.
|
|
404
|
+
if (!degrade.synthesized || !degrade.classes.includes(errorClass)) return null;
|
|
405
|
+
const value = { ...degrade.synthesized, [degrade.notesField]: renderDegradeNote(policy, err, errorClass) };
|
|
406
|
+
if (degrade.marker) value[degrade.marker] = true;
|
|
407
|
+
return value;
|
|
408
|
+
}
|
|
409
|
+
|
|
410
|
+
/**
|
|
411
|
+
* apra-fleet-3swo.5.4: the ONE place a degraded value's failure TEXT is
|
|
412
|
+
* built, for every role.
|
|
413
|
+
*
|
|
414
|
+
* Each runner.js dispatch site used to hand the engine its own map of
|
|
415
|
+
* per-class note builders, so six call sites carried six near-identical
|
|
416
|
+
* copies of the same two sentences and any change to the wording had six
|
|
417
|
+
* places to drift. The wording is now `degrade.noteTemplates` DATA in
|
|
418
|
+
* role-policies.mjs and this function is the only thing that reads it.
|
|
419
|
+
*
|
|
420
|
+
* A class the ladder fabricates for but has no template for THROWS rather
|
|
421
|
+
* than falling back to a generic sentence: a silent generic note is exactly
|
|
422
|
+
* how a role's distinct degrade text (the regression phase's "may not have
|
|
423
|
+
* reached the shared remote") would quietly disappear while every test that
|
|
424
|
+
* only checks `typeof notes === 'string'` kept passing.
|
|
425
|
+
*
|
|
426
|
+
* `{message}` falls back to String(err) so a non-Error throw (which only the
|
|
427
|
+
* regression catch-all's 'unknown' class can even reach) still reads as
|
|
428
|
+
* something rather than as an empty note.
|
|
429
|
+
*/
|
|
430
|
+
function renderDegradeNote(policy, err, errorClass) {
|
|
431
|
+
const { noteTemplates } = policy.degrade;
|
|
432
|
+
const template = noteTemplates && noteTemplates[errorClass];
|
|
433
|
+
if (typeof template !== 'string') {
|
|
434
|
+
throw new Error(
|
|
435
|
+
`dispatch-role: policy '${policy.role}' fabricates a value for error class '${errorClass}' but records ` +
|
|
436
|
+
'no degrade.noteTemplates entry for it -- add the template to role-policies.mjs rather than letting the ' +
|
|
437
|
+
'degrade write an empty note.'
|
|
438
|
+
);
|
|
439
|
+
}
|
|
440
|
+
return template.replace(/\{(message|name)\}/g, (_, field) => {
|
|
441
|
+
if (field === 'name') return String((err && err.name) || 'Error');
|
|
442
|
+
const message = err && err.message;
|
|
443
|
+
return message ? String(message) : String(err);
|
|
444
|
+
});
|
|
445
|
+
}
|
|
446
|
+
|
|
447
|
+
/**
|
|
448
|
+
* The DEGRADE_STEPS a policy runs on each attempt that degrades (as opposed
|
|
449
|
+
* to `postResult`, which runs only after a SUCCESSFUL attempt). The one step
|
|
450
|
+
* today is the per-round reviewer's 'clear-round-session': a failed round's
|
|
451
|
+
* session must not be resumed by the next round, and a successful one must.
|
|
452
|
+
*/
|
|
453
|
+
async function runDegradeSteps(ctx, policy, err, errorClass) {
|
|
454
|
+
for (const step of policy.degrade.steps) {
|
|
455
|
+
const hook = ctx.steps && ctx.steps[step];
|
|
456
|
+
if (typeof hook !== 'function') {
|
|
457
|
+
throw new Error(`dispatch-role: policy names degrade step '${step}', which ctx.steps does not supply.`);
|
|
458
|
+
}
|
|
459
|
+
await hook({ policy, error: err, errorClass });
|
|
460
|
+
}
|
|
461
|
+
}
|
|
462
|
+
|
|
463
|
+
// -----------------------------------------------------------------------------
|
|
464
|
+
// The engine.
|
|
465
|
+
// -----------------------------------------------------------------------------
|
|
466
|
+
|
|
467
|
+
/**
|
|
468
|
+
* Runs `roleName`'s dispatch ladder exactly as role-policies.mjs describes it.
|
|
469
|
+
*
|
|
470
|
+
* @param {object} ctx runner-supplied primitives (see this file's header):
|
|
471
|
+
* agent, withGitSync, withDispatchWatchdog, log, getMemberForRole,
|
|
472
|
+
* memberSessionGuard, onLlmAuthFailure, fixedRoleTier, budgets, schemas,
|
|
473
|
+
* isNoMutationDispatchFailure, invalidateAllBeadsCache, steps.
|
|
474
|
+
* @param {string} roleName a key of ROLE_POLICIES.
|
|
475
|
+
* @param {object} opts per-call values the policy deliberately does not carry:
|
|
476
|
+
* prompt -- the built prompt (required)
|
|
477
|
+
* resumePrompt -- the prompt a 'max-turns-resume' secondary sends
|
|
478
|
+
* repairPrompt -- (reason) => prompt, for a 'semantic-repair-re-ask'
|
|
479
|
+
* validate -- (value) => { ok, reason, result }: the semantic
|
|
480
|
+
* validation `postResult: ['select-streaks-validate']`
|
|
481
|
+
* names. Its `result` comes back as `validation`.
|
|
482
|
+
* resumeArg -- runtime value for a 'round-session'/'worklist' resumeArg
|
|
483
|
+
* onSessionId -- session-id recorder
|
|
484
|
+
* bindings -- runner-local values a policy names by binding
|
|
485
|
+
* roleLabel -- human role name used in this engine's log lines
|
|
486
|
+
* label/resumeLabel/repairLabel -- presentation labels (NOT policy: see
|
|
487
|
+
* role-policies.mjs's "deliberately out of scope" note)
|
|
488
|
+
* attemptOptions -- ({ attempt, skipPreDispatchSync }) => extra withGitSync options
|
|
489
|
+
* afterAttempt -- async (value) => void, run INSIDE the attempt's try so
|
|
490
|
+
* a failure in it is classified by the same ladder
|
|
491
|
+
* onResultRejected -- (reason) => Error, the error a `retryOnInvalidResult`
|
|
492
|
+
* ladder throws once its budget is spent on results its
|
|
493
|
+
* own validator keeps rejecting
|
|
494
|
+
* prepare -- async ({dispatch, resumeAttempt}) => {prompt, label,
|
|
495
|
+
* resumeArg, bindings}: per-dispatch values that are only
|
|
496
|
+
* knowable INSIDE the git-sync bracket, after this
|
|
497
|
+
* dispatch's in-bracket pre-dispatch steps have run (the
|
|
498
|
+
* doer's streak narrows when the claim finds beads another
|
|
499
|
+
* sprint already holds, and its prompt/label/tier follow)
|
|
500
|
+
* bindings -- runner-local values a policy names by binding; a
|
|
501
|
+
* FUNCTION ({resumeAttempt}) => bindings when a value
|
|
502
|
+
* differs per resume (the doer's escalating turn budget)
|
|
503
|
+
* @returns {Promise<{ok:boolean, value:any, error:Error|null, degraded:boolean,
|
|
504
|
+
* inconclusive:{reason:string,message:string}|null,
|
|
505
|
+
* attempts:number, resumesIssued:number,
|
|
506
|
+
* stepResults:Record<string,any>, validation:any}>}
|
|
507
|
+
*/
|
|
508
|
+
export async function dispatchRole(ctx, roleName, opts = {}) {
|
|
509
|
+
// Destructured so this module's ONE dispatch reads `agent(` -- the literal
|
|
510
|
+
// token dispatch-safety-guard.mjs and inline-ladder-guard.mjs scan for. A
|
|
511
|
+
// `ctx.agent(` call would fall outside both guards' call-site regexes and
|
|
512
|
+
// leave this file's only dispatch silently unguarded.
|
|
513
|
+
const { agent } = ctx;
|
|
514
|
+
// ctx.policies: the table this run reads its row from, defaulting to the
|
|
515
|
+
// real frozen ROLE_POLICIES. It exists so a test can splice ONE field of
|
|
516
|
+
// ONE row and observe the engine behave differently -- which is the only
|
|
517
|
+
// way to prove a per-role variance (the integ INCONCLUSIVE record, the
|
|
518
|
+
// regression catch-all, the final-review heal short-circuit, the
|
|
519
|
+
// deployer's pre-dispatch prompt invariant) is really driven by POLICY
|
|
520
|
+
// DATA rather than by a branch in this file that happens to agree with
|
|
521
|
+
// the table. Production never passes it.
|
|
522
|
+
const policy = ctx.policies ? policyFor(roleName, ctx.policies) : policyFor(roleName);
|
|
523
|
+
/**
|
|
524
|
+
* The runner-local values a policy names by binding. A FUNCTION when a
|
|
525
|
+
* value differs per resume attempt -- the doer's escalating turn budget is
|
|
526
|
+
* computed from the attempt number, so the table records it as a
|
|
527
|
+
* 'runtime' turn budget and the runner supplies the arithmetic.
|
|
528
|
+
*/
|
|
529
|
+
const bindingsFor = (resumeAttempt = 0) => (typeof opts.bindings === 'function'
|
|
530
|
+
? opts.bindings({ resumeAttempt })
|
|
531
|
+
: (opts.bindings || {}));
|
|
532
|
+
const bindings = bindingsFor(0);
|
|
533
|
+
const member = resolveMember(ctx, policy.member, bindings);
|
|
534
|
+
const roleLabel = opts.roleLabel || policy.role;
|
|
535
|
+
const retry = policy.retry;
|
|
536
|
+
const backoffMs = retry.backoffMs;
|
|
537
|
+
const attempts = retry.attempts;
|
|
538
|
+
|
|
539
|
+
// --- ONE dispatch, however the policy shapes it ------------------------
|
|
540
|
+
// Every per-dispatch difference (member, persona, tier, budgets, turn
|
|
541
|
+
// budget, schema, resume argument) is resolved from the policy row rather
|
|
542
|
+
// than written out again per role. member_name is spelled at the call site
|
|
543
|
+
// itself, never folded into `options`, because dispatch-safety-guard.mjs
|
|
544
|
+
// requires every agent() call site to name its member explicitly.
|
|
545
|
+
//
|
|
546
|
+
// EVERYTHING IS RESOLVED INSIDE `invoke`, i.e. INSIDE the git-sync bracket
|
|
547
|
+
// when there is one. That matters for exactly one reason, and it is a real
|
|
548
|
+
// one: an in-bracket pre-dispatch step can CHANGE what the dispatch should
|
|
549
|
+
// say. The doer claims its beads after the bracket's D-pull, which is what
|
|
550
|
+
// reveals that another sprint already holds some of them; the streak
|
|
551
|
+
// narrows, and its prompt, its label and the tier it is priced at all
|
|
552
|
+
// follow. Resolving options before the bracket opened would dispatch the
|
|
553
|
+
// pre-claim scope.
|
|
554
|
+
const runDispatch = (dispatch, defaultPrompt, defaultLabel, attemptOpts, extra = {}) => {
|
|
555
|
+
const invoke = async () => {
|
|
556
|
+
// In-bracket pre-dispatch steps, then the caller's own preparation.
|
|
557
|
+
const inBracket = await runPreDispatchSteps(ctx, dispatch, member, opts, extra, 'in-bracket');
|
|
558
|
+
if (inBracket) return inBracket.value;
|
|
559
|
+
const prepared = opts.prepare ? await opts.prepare({ dispatch, ...extra }) : {};
|
|
560
|
+
const localBindings = { ...bindingsFor(extra.resumeAttempt || 0), ...(prepared.bindings || {}) };
|
|
561
|
+
const localOpts = prepared.resumeArg !== undefined ? { ...opts, resumeArg: prepared.resumeArg } : opts;
|
|
562
|
+
const options = {
|
|
563
|
+
agentType: dispatch.agentType ?? undefined,
|
|
564
|
+
model: resolveModelTier(ctx, dispatch.model, localBindings),
|
|
565
|
+
timeout_s: resolveBudget(ctx, dispatch.timeouts.timeoutS),
|
|
566
|
+
max_total_s: resolveBudget(ctx, dispatch.timeouts.maxTotalS),
|
|
567
|
+
max_turns: resolveMaxTurns(dispatch.maxTurns, localBindings),
|
|
568
|
+
schema: resolveSchema(ctx, dispatch.schema),
|
|
569
|
+
resume: resolveResumeArg(dispatch.resumeArg, localOpts),
|
|
570
|
+
onSessionId: opts.onSessionId,
|
|
571
|
+
label: prepared.label ?? defaultLabel,
|
|
572
|
+
};
|
|
573
|
+
for (const key of Object.keys(options)) {
|
|
574
|
+
if (options[key] === undefined) delete options[key];
|
|
575
|
+
}
|
|
576
|
+
const inFlight = agent(prepared.prompt ?? defaultPrompt, {
|
|
577
|
+
...options,
|
|
578
|
+
member_name: member,
|
|
579
|
+
});
|
|
580
|
+
if (!dispatch.watchdog.armed) return inFlight;
|
|
581
|
+
// Raced against a client-side watchdog so a frozen-but-alive
|
|
582
|
+
// member session can never leave this await silently hanging past
|
|
583
|
+
// its budget -- needed in ADDITION to, not instead of, the
|
|
584
|
+
// server-side timeout_s/max_total_s above.
|
|
585
|
+
return ctx.withDispatchWatchdog(inFlight, {
|
|
586
|
+
timeoutS: resolveBudget(ctx, dispatch.watchdog.timeoutS),
|
|
587
|
+
member,
|
|
588
|
+
label: resolveWatchdogLabel(dispatch.watchdog.label, localBindings),
|
|
589
|
+
log: ctx.log,
|
|
590
|
+
});
|
|
591
|
+
};
|
|
592
|
+
// `dispatch.bracket`, not `policy.bracket`: a secondary dispatch
|
|
593
|
+
// inherits its main dispatch's bracket by spread today, but reading it
|
|
594
|
+
// off the dispatch actually being run is what keeps that an inherited
|
|
595
|
+
// DEFAULT rather than a hard-wired assumption the table cannot
|
|
596
|
+
// override.
|
|
597
|
+
if (!dispatch.bracket.wrapped) return invoke();
|
|
598
|
+
return ctx.withGitSync(member, dispatch.bracket.pushCode === true, invoke, {
|
|
599
|
+
pushBeads: dispatch.bracket.pushBeads === true,
|
|
600
|
+
...(attemptOpts || {}),
|
|
601
|
+
});
|
|
602
|
+
};
|
|
603
|
+
|
|
604
|
+
// --- one ATTEMPT: the main dispatch, plus the resume it may need --------
|
|
605
|
+
// TWO conditions reach the SAME secondary dispatch, and the policy says
|
|
606
|
+
// which of them this ladder honours:
|
|
607
|
+
// retry.maxTurnsResume -- the turn budget was spent. Universal.
|
|
608
|
+
// retry.infraResumeAttempts -- the member CLI died mid-turn and lost
|
|
609
|
+
// its result envelope, so the run may have
|
|
610
|
+
// made real progress and merely failed to
|
|
611
|
+
// report it. The integ runner alone, and
|
|
612
|
+
// bounded at one recovery.
|
|
613
|
+
const runAttempt = async (attemptOpts) => {
|
|
614
|
+
const preShortCircuit = await runPreDispatchSteps(ctx, policy, member, opts, {});
|
|
615
|
+
if (preShortCircuit) return preShortCircuit.value;
|
|
616
|
+
try {
|
|
617
|
+
return await runDispatch(policy, opts.prompt, opts.label, attemptOpts);
|
|
618
|
+
} catch (err) {
|
|
619
|
+
const secondary = policy.secondary;
|
|
620
|
+
const isTurnExhaustion = err instanceof AgentDispatchError
|
|
621
|
+
&& err.details && err.details.reason === 'max_turns_exhausted';
|
|
622
|
+
const isInfra = err instanceof AgentDispatchError && isInfraDispatchFailure(err);
|
|
623
|
+
const resumesOnExhaustion = retry.maxTurnsResume && isTurnExhaustion;
|
|
624
|
+
const resumesOnInfra = retry.infraResumeAttempts > 0 && isInfra && !isTurnExhaustion;
|
|
625
|
+
if (!secondary || secondary.kind !== 'max-turns-resume' || (!resumesOnExhaustion && !resumesOnInfra)) {
|
|
626
|
+
throw err;
|
|
627
|
+
}
|
|
628
|
+
// A BOUNDED ladder of resumes, not a single one: retry.resumeAttempts
|
|
629
|
+
// says how many, and retry.turnEscalation='double' is performed by
|
|
630
|
+
// the runtime turn budget the policy names (the runner's bindings
|
|
631
|
+
// thunk is handed the attempt number). A resume that fails for any
|
|
632
|
+
// reason OTHER than spending its turns again is not something more
|
|
633
|
+
// turns can fix, so the escalation stops there.
|
|
634
|
+
let resumeAttempt = 0;
|
|
635
|
+
let lastResumeErr = err;
|
|
636
|
+
while (resumeAttempt < Math.max(retry.resumeAttempts, resumesOnInfra ? 1 : 0)) {
|
|
637
|
+
resumeAttempt += 1;
|
|
638
|
+
// A resume follows an agent that DID run (max_turns_exhausted is
|
|
639
|
+
// a resumable partial-work case, not a no-mutation failure), so
|
|
640
|
+
// it always runs the full pre-dispatch sync -- no attemptOpts.
|
|
641
|
+
//
|
|
642
|
+
// Run BEFORE announcing the resume: a pre-dispatch step may find
|
|
643
|
+
// the work already DONE (the doer's 'verify-streak-closed': a
|
|
644
|
+
// streak whose beads all closed before the turn limit hit is a
|
|
645
|
+
// success that merely missed its VERIFY checkpoint), and
|
|
646
|
+
// announcing a resume that never happens reads as a wasted
|
|
647
|
+
// dispatch in the log.
|
|
648
|
+
const shortCircuit = await runPreDispatchSteps(ctx, secondary, member, opts, { resumeAttempt });
|
|
649
|
+
if (shortCircuit) return shortCircuit.value;
|
|
650
|
+
if (resumeAttempt === 1 && resumesOnExhaustion) {
|
|
651
|
+
ctx.log(
|
|
652
|
+
`${roleLabel} exhausted its turn limit ` +
|
|
653
|
+
`(max_turns=${resolveMaxTurns(policy.maxTurns, bindingsFor(0))}) -- resuming the same ` +
|
|
654
|
+
`session with max_turns=${resolveMaxTurns(secondary.maxTurns, bindingsFor(1))}.`
|
|
655
|
+
);
|
|
656
|
+
} else if (resumeAttempt === 1) {
|
|
657
|
+
ctx.log(
|
|
658
|
+
`${roleLabel}: infrastructure dispatch failure ` +
|
|
659
|
+
`(${(err.details && err.details.reason) || 'unknown'}) -- the member produced no result ` +
|
|
660
|
+
'envelope. This is NOT a failure verdict; resuming the same session once to recover ' +
|
|
661
|
+
'before recording anything.'
|
|
662
|
+
);
|
|
663
|
+
} else {
|
|
664
|
+
ctx.log(
|
|
665
|
+
`${roleLabel} exhausted its turn limit again -- resuming the same session with ` +
|
|
666
|
+
`max_turns=${resolveMaxTurns(secondary.maxTurns, bindingsFor(resumeAttempt))} ` +
|
|
667
|
+
`(attempt ${resumeAttempt}/${retry.resumeAttempts}).`
|
|
668
|
+
);
|
|
669
|
+
}
|
|
670
|
+
resumesIssued += 1;
|
|
671
|
+
try {
|
|
672
|
+
return await runDispatch(secondary, opts.resumePrompt, opts.resumeLabel, undefined, { resumeAttempt });
|
|
673
|
+
} catch (resumeErr) {
|
|
674
|
+
lastResumeErr = resumeErr;
|
|
675
|
+
const exhaustedAgain = resumeErr instanceof AgentDispatchError
|
|
676
|
+
&& resumeErr.details && resumeErr.details.reason === 'max_turns_exhausted';
|
|
677
|
+
if (retry.maxTurnsResume && exhaustedAgain) continue;
|
|
678
|
+
throw resumeErr;
|
|
679
|
+
}
|
|
680
|
+
}
|
|
681
|
+
// The escalating ladder is spent. A ladder that declares
|
|
682
|
+
// abortAfterSpentResumeLadder has thereby PROVED the task is too
|
|
683
|
+
// large for one streak, so another full attempt at the same size
|
|
684
|
+
// would only reproduce it; the caller is told so instead.
|
|
685
|
+
resumeLadderSpent = retry.abortAfterSpentResumeLadder;
|
|
686
|
+
throw lastResumeErr;
|
|
687
|
+
}
|
|
688
|
+
};
|
|
689
|
+
|
|
690
|
+
// A usage-limit re-dispatch after the controller reported the member is
|
|
691
|
+
// usable again (apra-fleet-hzeb.4.2). It mirrors runAttempt's resume path:
|
|
692
|
+
// it runs the resume dispatch's own pre-dispatch steps, then the resume
|
|
693
|
+
// dispatch itself, so a role with a resume ladder continues the SAME member
|
|
694
|
+
// session it hit the limit on (warm context kept) rather than starting a
|
|
695
|
+
// fresh one. Only reached when the failing attempt reported a sessionId and
|
|
696
|
+
// the role actually has a max-turns resume secondary; otherwise the caller
|
|
697
|
+
// re-runs runAttempt fresh.
|
|
698
|
+
const redispatchViaUsageLimitResume = async (attemptOpts) => {
|
|
699
|
+
const secondary = policy.secondary;
|
|
700
|
+
const shortCircuit = await runPreDispatchSteps(ctx, secondary, member, opts, { resumeAttempt: 1 });
|
|
701
|
+
if (shortCircuit) return shortCircuit.value;
|
|
702
|
+
return runDispatch(secondary, opts.resumePrompt ?? opts.prompt, opts.resumeLabel ?? opts.label, attemptOpts, { resumeAttempt: 1 });
|
|
703
|
+
};
|
|
704
|
+
|
|
705
|
+
let lastErr = null;
|
|
706
|
+
let value;
|
|
707
|
+
let ok = false;
|
|
708
|
+
let degradedValue = null;
|
|
709
|
+
let skipPreDispatchSyncNext = false;
|
|
710
|
+
// Set once an auth self-heal succeeded under retry.authSelfHealShortCircuits:
|
|
711
|
+
// the attempt it grants is the ladder's last, whatever it returns.
|
|
712
|
+
let healRetryIsFinal = false;
|
|
713
|
+
// How many attempts the ladder really made, and how many resume dispatches
|
|
714
|
+
// it issued. Surfaced on the outcome because a caller can legitimately
|
|
715
|
+
// need to know (the doer records whether a streak was retried).
|
|
716
|
+
let attemptsMade = 0;
|
|
717
|
+
let resumesIssued = 0;
|
|
718
|
+
let stepResults = {};
|
|
719
|
+
// Set when an attempt's escalating resume ladder ran out under a policy
|
|
720
|
+
// that treats that as terminal. Reset per attempt.
|
|
721
|
+
let resumeLadderSpent = false;
|
|
722
|
+
// apra-fleet-hzeb.4.2: pending usage-limit re-dispatch mode for the NEXT
|
|
723
|
+
// loop pass. null = normal attempt; 'fresh' = re-run runAttempt after a
|
|
724
|
+
// paused-then-resumed member; 'resume' = continue the SAME session via the
|
|
725
|
+
// resume ladder. A usage-limit re-dispatch does NOT consume a ladder
|
|
726
|
+
// attempt (see the hook below), so it also skips the retry backoff.
|
|
727
|
+
let usageLimitRedispatch = null;
|
|
728
|
+
|
|
729
|
+
for (let attempt = 1; attempt <= attempts; attempt++) {
|
|
730
|
+
if (!usageLimitRedispatch && backoffMs && backoffMs[attempt - 1] > 0) {
|
|
731
|
+
ctx.log(
|
|
732
|
+
`${roleLabel} dispatch: waiting ${backoffMs[attempt - 1] / 1000}s before retry attempt ` +
|
|
733
|
+
`${attempt}/${backoffMs.length}...`
|
|
734
|
+
);
|
|
735
|
+
if (!instantRetryBackoff()) {
|
|
736
|
+
await new Promise((resolve) => setTimeout(resolve, backoffMs[attempt - 1]));
|
|
737
|
+
}
|
|
738
|
+
}
|
|
739
|
+
attemptsMade = attempt;
|
|
740
|
+
resumeLadderSpent = false;
|
|
741
|
+
try {
|
|
742
|
+
// A generic retry resumes onto the streak branch's REMOTE TIP so it
|
|
743
|
+
// builds on already-published work instead of re-implementing the
|
|
744
|
+
// task as a divergent, content-identical duplicate commit that can
|
|
745
|
+
// never fast-forward. Only after a failure that may have committed:
|
|
746
|
+
// an auth/trust failure provably ran nothing, so its retry keeps
|
|
747
|
+
// plain ff-only pre-dispatch sync.
|
|
748
|
+
const retryFollowsPossibleMutation = attempt > 1 && lastErr
|
|
749
|
+
&& !isNonRetryableDispatchError(lastErr);
|
|
750
|
+
const attemptOpts = {
|
|
751
|
+
...(retry.skipPreDispatchSyncOnNoMutation ? { skipPreDispatchSync: skipPreDispatchSyncNext } : {}),
|
|
752
|
+
...(retry.resumeOntoRemoteTipOnRetry && retryFollowsPossibleMutation
|
|
753
|
+
? { resumeOntoRemoteTip: true }
|
|
754
|
+
: {}),
|
|
755
|
+
...(opts.attemptOptions ? opts.attemptOptions({ attempt }) : {}),
|
|
756
|
+
};
|
|
757
|
+
// A usage-limit resume re-dispatch continues the same session via
|
|
758
|
+
// the resume ladder; every other pass (including a 'fresh'
|
|
759
|
+
// usage-limit re-dispatch) is a normal full attempt. Read and clear
|
|
760
|
+
// the mode BEFORE the await so a failing re-dispatch never leaves a
|
|
761
|
+
// stale mode armed for a later attempt -- the hook re-sets it if the
|
|
762
|
+
// re-dispatch hits the limit again.
|
|
763
|
+
const redispatchMode = usageLimitRedispatch;
|
|
764
|
+
usageLimitRedispatch = null;
|
|
765
|
+
value = redispatchMode === 'resume'
|
|
766
|
+
? await redispatchViaUsageLimitResume(attemptOpts)
|
|
767
|
+
: await runAttempt(attemptOpts);
|
|
768
|
+
if (opts.afterAttempt) await opts.afterAttempt(value);
|
|
769
|
+
// A postResult step may REJECT the result (the reviewer's
|
|
770
|
+
// contract guard: a CHANGES_NEEDED verdict with nothing for the
|
|
771
|
+
// orchestrator to act on is self-contradictory). Steps run in
|
|
772
|
+
// order and the rejection throws immediately, so a later step
|
|
773
|
+
// never acts on a result the ladder has already judged unusable.
|
|
774
|
+
stepResults = await runPostResultSteps(ctx, policy, value, opts, member);
|
|
775
|
+
ok = true;
|
|
776
|
+
lastErr = null;
|
|
777
|
+
degradedValue = null;
|
|
778
|
+
break;
|
|
779
|
+
} catch (err) {
|
|
780
|
+
lastErr = err;
|
|
781
|
+
|
|
782
|
+
// A rejected-but-schema-valid result is NOT a degrade: the
|
|
783
|
+
// dispatch worked, its answer was unusable. Retry the whole
|
|
784
|
+
// attempt inside the same budget -- a nudge cannot fix a verdict
|
|
785
|
+
// that contradicts itself, only a fresh review can -- then hand
|
|
786
|
+
// the caller its own error rather than a fabricated value.
|
|
787
|
+
if (err instanceof LadderResultRejectedError) {
|
|
788
|
+
if (retry.retryOnInvalidResult && attempt < attempts) {
|
|
789
|
+
ctx.log(
|
|
790
|
+
`${roleLabel}: result rejected (${err.reason}) -- re-running the dispatch ` +
|
|
791
|
+
`(attempt ${attempt + 1} of ${attempts}) before treating this as a distinct failure.`
|
|
792
|
+
);
|
|
793
|
+
continue;
|
|
794
|
+
}
|
|
795
|
+
throw opts.onResultRejected ? opts.onResultRejected(err.reason) : err;
|
|
796
|
+
}
|
|
797
|
+
|
|
798
|
+
// The turn ALREADY RAN and its output is committed on the member's
|
|
799
|
+
// own clone -- only the post-dispatch sync failed, and withGitSync
|
|
800
|
+
// already retried that step on its own. Re-dispatching would spawn
|
|
801
|
+
// a second session for the same phase on top of completed work.
|
|
802
|
+
//
|
|
803
|
+
// This ENDS the ladder but still runs the degrade below, because
|
|
804
|
+
// "the dispatch ran and its writes are local" is itself something
|
|
805
|
+
// a caller may need reported: the regression phase, whose whole job
|
|
806
|
+
// is filing carry-over beads, has to say in its summary that they
|
|
807
|
+
// may not have reached the shared remote. A ladder that fabricates
|
|
808
|
+
// nothing for this class (the planner, the doer) is unaffected --
|
|
809
|
+
// its degrade.classes simply does not list it.
|
|
810
|
+
const completedButSyncFailed =
|
|
811
|
+
(retry.skipRedispatchOnPostDispatchSyncFailure && isPostDispatchSyncFailure(err))
|
|
812
|
+
|| resumeLadderSpent;
|
|
813
|
+
if (retry.skipRedispatchOnPostDispatchSyncFailure && isPostDispatchSyncFailure(err)) {
|
|
814
|
+
ctx.log(
|
|
815
|
+
`${roleLabel} dispatch COMPLETED but its post-dispatch sync failed: ${err.message} Aborting ` +
|
|
816
|
+
'retries WITHOUT re-dispatching -- the turn already ran and its writes are local; fix the sync ' +
|
|
817
|
+
'and re-run.'
|
|
818
|
+
);
|
|
819
|
+
}
|
|
820
|
+
|
|
821
|
+
// Only a provably no-mutation dispatch failure leaves the workspace
|
|
822
|
+
// unchanged, so only then may the next attempt skip its
|
|
823
|
+
// pre-dispatch sync. Any other error re-arms it.
|
|
824
|
+
if (!completedButSyncFailed && retry.skipPreDispatchSyncOnNoMutation) {
|
|
825
|
+
skipPreDispatchSyncNext = ctx.isNoMutationDispatchFailure(err);
|
|
826
|
+
}
|
|
827
|
+
|
|
828
|
+
// Usage-limit pause/resume (apra-fleet-hzeb.4.2). A provider
|
|
829
|
+
// rate/usage limit is neither a failure of this role nor something a
|
|
830
|
+
// retry ladder can out-wait on its own budget -- it needs a real
|
|
831
|
+
// wall-clock wait until the quota window resets. When the policy
|
|
832
|
+
// arms usageLimitPause and the runner wired an onUsageLimit
|
|
833
|
+
// controller, hand the signal to it: it pauses the whole run through
|
|
834
|
+
// the engine's cooperative pause primitive (releasing reservations,
|
|
835
|
+
// showing the resume time on the dashboard), waits, re-probes the
|
|
836
|
+
// member on a bounded backoff, and reports whether the member is
|
|
837
|
+
// usable again. A resumed member RE-DISPATCHES this role WITHOUT
|
|
838
|
+
// consuming one of the ladder's real attempts; an exhausted wait
|
|
839
|
+
// gives up with the controller's typed UsageLimitWaitExhaustedError
|
|
840
|
+
// (registered in abort.mjs's isTypedAbortError, so it routes through
|
|
841
|
+
// finalizeAbort()/an [ABORTED] PR like every other terminal abort).
|
|
842
|
+
//
|
|
843
|
+
// This runs OUTSIDE any withGitSync bracket: the failing dispatch's
|
|
844
|
+
// bracket has fully unwound by the time its error reaches this catch
|
|
845
|
+
// (openSyncBracketCount is back to 0), which is what lets the
|
|
846
|
+
// engine's clean-state pause guard actually engage the pause -- a
|
|
847
|
+
// pause requested mid-bracket would be deferred forever.
|
|
848
|
+
if (!completedButSyncFailed && retry.usageLimitPause
|
|
849
|
+
&& isUsageLimitDispatchError(err) && typeof ctx.onUsageLimit === 'function') {
|
|
850
|
+
const failedSessionId = (err.details && err.details.sessionId) || null;
|
|
851
|
+
const outcome = await ctx.onUsageLimit({
|
|
852
|
+
member,
|
|
853
|
+
roleLabel,
|
|
854
|
+
signal: usageLimitOf(err),
|
|
855
|
+
sessionId: failedSessionId,
|
|
856
|
+
tier: resolveModelTier(ctx, policy.model, bindings),
|
|
857
|
+
});
|
|
858
|
+
if (outcome && outcome.resumed) {
|
|
859
|
+
// The member is usable again: re-dispatch this role without
|
|
860
|
+
// charging a ladder attempt. Prefer resuming the reported
|
|
861
|
+
// session when the role actually HAS a resume ladder (warm
|
|
862
|
+
// context kept); otherwise re-dispatch fresh.
|
|
863
|
+
const canResumeSession = Boolean(failedSessionId) && retry.resumeAttempts > 0
|
|
864
|
+
&& policy.secondary && policy.secondary.kind === 'max-turns-resume';
|
|
865
|
+
usageLimitRedispatch = canResumeSession ? 'resume' : 'fresh';
|
|
866
|
+
attempt -= 1;
|
|
867
|
+
continue;
|
|
868
|
+
}
|
|
869
|
+
// The controller waited/re-probed to its budget and the member
|
|
870
|
+
// is still limited: its typed give-up error is this ladder's
|
|
871
|
+
// last word and propagates straight out (no degrade).
|
|
872
|
+
throw (outcome && outcome.error) || err;
|
|
873
|
+
}
|
|
874
|
+
|
|
875
|
+
// Auth/workspace-trust failures are deterministic -- no retry can
|
|
876
|
+
// succeed -- so a ladder that declares abortOnNonRetryable ends
|
|
877
|
+
// rather than burning its remaining attempts on the same wall.
|
|
878
|
+
// Deliberately WITHOUT a degrade: a ladder that aborts here has a
|
|
879
|
+
// credentials/trust problem to fix, not a result to fabricate.
|
|
880
|
+
if (!completedButSyncFailed && retry.abortOnNonRetryable && isNonRetryableDispatchError(err)) {
|
|
881
|
+
// An LLM-auth failure (unlike workspace trust, which self-heal
|
|
882
|
+
// cannot fix) gets one bounded self-heal attempt first.
|
|
883
|
+
if (retry.authSelfHeal && isAuthDispatchError(err) && typeof ctx.onLlmAuthFailure === 'function') {
|
|
884
|
+
const healed = await ctx.onLlmAuthFailure({
|
|
885
|
+
member, label: `${roleLabel} dispatch`, error: err.message,
|
|
886
|
+
});
|
|
887
|
+
if (healed) {
|
|
888
|
+
ctx.log(`${roleLabel} dispatch: LLM auth self-heal succeeded -- retrying.`);
|
|
889
|
+
// retry.authSelfHealShortCircuits: the healed retry is
|
|
890
|
+
// this ladder's LAST word. A role whose dispatch is
|
|
891
|
+
// the single most expensive in the sprint (final
|
|
892
|
+
// review) must not fall through to the generic retry
|
|
893
|
+
// as well -- that fires a SECOND full review and
|
|
894
|
+
// silently discards the healed verdict, which for a
|
|
895
|
+
// PASS/FAIL gate can invert the sprint's outcome.
|
|
896
|
+
healRetryIsFinal = retry.authSelfHealShortCircuits;
|
|
897
|
+
continue;
|
|
898
|
+
}
|
|
899
|
+
}
|
|
900
|
+
ctx.log(
|
|
901
|
+
`${roleLabel} dispatch threw a non-retryable error (auth/trust): ${err.message}. Aborting ` +
|
|
902
|
+
"retries -- fix the member's credentials/trust and re-run."
|
|
903
|
+
);
|
|
904
|
+
// A ladder whose own failure legitimately fails the whole
|
|
905
|
+
// sprint propagates instead of degrading: with the dispatch
|
|
906
|
+
// channel walled off there is no judgement to fabricate.
|
|
907
|
+
if (retry.rethrowsUnhealedNonRetryable) throw err;
|
|
908
|
+
break;
|
|
909
|
+
}
|
|
910
|
+
|
|
911
|
+
// A ladder that does NOT abort on a non-retryable error still
|
|
912
|
+
// self-heals an unhealed LLM-auth failure, because the SAME member
|
|
913
|
+
// is dispatched again later in the cycle and would otherwise walk
|
|
914
|
+
// into the identical wall.
|
|
915
|
+
if (!completedButSyncFailed && !retry.abortOnNonRetryable && retry.authSelfHeal
|
|
916
|
+
&& isAuthDispatchError(err) && typeof ctx.onLlmAuthFailure === 'function') {
|
|
917
|
+
await ctx.onLlmAuthFailure({ member, label: `${roleLabel} dispatch`, error: err.message });
|
|
918
|
+
}
|
|
919
|
+
|
|
920
|
+
// RUN-level control signals are not failures of this role at all,
|
|
921
|
+
// so they keep propagating even through a catch-all that swallows
|
|
922
|
+
// everything else: honouring a cancellation outranks finishing an
|
|
923
|
+
// informational phase, and swallowing a blown spend ceiling would
|
|
924
|
+
// let the run keep spending past a limit the operator set.
|
|
925
|
+
if (policy.degrade.rethrowsRunControlSignals.length > 0) {
|
|
926
|
+
const names = errorClassNames(err);
|
|
927
|
+
if (policy.degrade.rethrowsRunControlSignals.some((name) => names.includes(name))) {
|
|
928
|
+
ctx.log(`${roleLabel}: run-level ${names[0] || 'control signal'} received -- not retrying this dispatch: ${err.message}`);
|
|
929
|
+
throw err;
|
|
930
|
+
}
|
|
931
|
+
}
|
|
932
|
+
|
|
933
|
+
const errorClass = classifyLadderError(err, policy);
|
|
934
|
+
// An unrecognised error class still propagates rather than being
|
|
935
|
+
// degraded silently -- except under a degrade that declares
|
|
936
|
+
// otherwise, and except under 'fatal', whose whole contract is to
|
|
937
|
+
// accumulate and rethrow after the ladder is spent.
|
|
938
|
+
if (errorClass === null && policy.degrade.kind !== 'fatal'
|
|
939
|
+
&& policy.degrade.rethrowsUnrecognisedErrors) {
|
|
940
|
+
throw err;
|
|
941
|
+
}
|
|
942
|
+
if (errorClass === 'schema') {
|
|
943
|
+
ctx.log(`${roleLabel}: schema-repair exhausted, degrading (${policy.degrade.kind}): ${err.message}`);
|
|
944
|
+
} else if (errorClass === 'dispatch') {
|
|
945
|
+
ctx.log(`${roleLabel}: agent dispatch failed, degrading (${policy.degrade.kind}): ${err.message}`);
|
|
946
|
+
} else if (errorClass === 'infra') {
|
|
947
|
+
ctx.log(
|
|
948
|
+
`${roleLabel}: infrastructure dispatch failure persisted -- degrading ` +
|
|
949
|
+
`(${policy.degrade.kind}), NOT recording a failure verdict: ${err.message}`
|
|
950
|
+
);
|
|
951
|
+
} else if (errorClass === 'sync') {
|
|
952
|
+
ctx.log(
|
|
953
|
+
`${roleLabel}: the git/beads sync around the dispatch FAILED (${err.name}: ${err.message}) -- ` +
|
|
954
|
+
`degrading (${policy.degrade.kind}). Anything this phase filed may still be local-only.`
|
|
955
|
+
);
|
|
956
|
+
} else if (errorClass === 'unknown') {
|
|
957
|
+
ctx.log(
|
|
958
|
+
`${roleLabel}: unexpected error, degrading (${policy.degrade.kind}) -- this phase never aborts ` +
|
|
959
|
+
`the sprint: ${err && err.stack ? err.stack : err}`
|
|
960
|
+
);
|
|
961
|
+
}
|
|
962
|
+
await runDegradeSteps(ctx, policy, err, errorClass);
|
|
963
|
+
degradedValue = synthesizeDegradedValue(policy, opts, err, errorClass);
|
|
964
|
+
const isLastAttempt = attempt === attempts || healRetryIsFinal || completedButSyncFailed;
|
|
965
|
+
ctx.log(
|
|
966
|
+
`${roleLabel} dispatch threw: ${err.message}.` +
|
|
967
|
+
(isLastAttempt ? ' Retries exhausted.' : ` Retrying (attempt ${attempt + 1} of ${attempts}).`)
|
|
968
|
+
);
|
|
969
|
+
if (healRetryIsFinal || completedButSyncFailed) break;
|
|
970
|
+
}
|
|
971
|
+
}
|
|
972
|
+
|
|
973
|
+
if (!ok) {
|
|
974
|
+
// The one FATAL degrade in the table: there is no sprint without a
|
|
975
|
+
// plan, so an exhausted planner ladder rethrows rather than
|
|
976
|
+
// synthesizing one.
|
|
977
|
+
if (policy.degrade.kind === 'fatal' && lastErr) throw lastErr;
|
|
978
|
+
return {
|
|
979
|
+
ok: false,
|
|
980
|
+
value: degradedValue,
|
|
981
|
+
error: lastErr,
|
|
982
|
+
degraded: true,
|
|
983
|
+
// An 'inconclusive' degrade whose terminal failure really was an
|
|
984
|
+
// envelope-less infra fault reports WHY there is no verdict, so
|
|
985
|
+
// the caller records "no evidence" instead of "the tests failed".
|
|
986
|
+
// Driven by degrade.kind: any other kind reports null here, which
|
|
987
|
+
// is what makes the INCONCLUSIVE variance policy data rather than
|
|
988
|
+
// a branch the integ caller happens to take.
|
|
989
|
+
inconclusive: inconclusiveOf(policy, lastErr),
|
|
990
|
+
attempts: attemptsMade,
|
|
991
|
+
resumesIssued,
|
|
992
|
+
stepResults: {},
|
|
993
|
+
validation: opts.validate ? opts.validate(degradedValue).result : null,
|
|
994
|
+
};
|
|
995
|
+
}
|
|
996
|
+
|
|
997
|
+
// --- semantic-repair re-ask -------------------------------------------
|
|
998
|
+
// ONE bounded attempt layered on top of agent()'s own schema-repair loop:
|
|
999
|
+
// a candidate can be schema-valid yet semantically invalid, and dropping
|
|
1000
|
+
// it straight to the caller's deterministic fallback silently discards
|
|
1001
|
+
// real intent. Re-ask once with the exact validation failure; only then
|
|
1002
|
+
// let the caller fall back. Guarded, never looped.
|
|
1003
|
+
let validation = null;
|
|
1004
|
+
if (opts.validate) {
|
|
1005
|
+
let verdictOfCandidate = opts.validate(value);
|
|
1006
|
+
if (!verdictOfCandidate.ok && retry.semanticRepairReAsks > 0 && value) {
|
|
1007
|
+
ctx.log(
|
|
1008
|
+
`${roleLabel}: candidate rejected (${verdictOfCandidate.reason}) -- re-asking once with the ` +
|
|
1009
|
+
'validation failure before falling back.'
|
|
1010
|
+
);
|
|
1011
|
+
try {
|
|
1012
|
+
value = await runDispatch(policy.secondary, opts.repairPrompt(verdictOfCandidate.reason), opts.repairLabel);
|
|
1013
|
+
verdictOfCandidate = opts.validate(value);
|
|
1014
|
+
} catch (repairErr) {
|
|
1015
|
+
if (classifyLadderError(repairErr, policy) === null) throw repairErr;
|
|
1016
|
+
ctx.log(`${roleLabel} (semantic repair): dispatch failed (${repairErr.message}) -- falling back.`);
|
|
1017
|
+
}
|
|
1018
|
+
}
|
|
1019
|
+
validation = verdictOfCandidate.result;
|
|
1020
|
+
}
|
|
1021
|
+
|
|
1022
|
+
return {
|
|
1023
|
+
ok: true,
|
|
1024
|
+
value,
|
|
1025
|
+
error: null,
|
|
1026
|
+
degraded: false,
|
|
1027
|
+
inconclusive: null,
|
|
1028
|
+
attempts: attemptsMade,
|
|
1029
|
+
resumesIssued,
|
|
1030
|
+
stepResults,
|
|
1031
|
+
validation,
|
|
1032
|
+
};
|
|
1033
|
+
}
|
|
1034
|
+
|
|
1035
|
+
/**
|
|
1036
|
+
* The "there is no verdict, and here is why" record an 'inconclusive' degrade
|
|
1037
|
+
* returns, or null for every other degrade kind (and for an inconclusive
|
|
1038
|
+
* ladder whose terminal failure was NOT an infra fault -- that one really did
|
|
1039
|
+
* fail, and says so).
|
|
1040
|
+
*/
|
|
1041
|
+
function inconclusiveOf(policy, lastErr) {
|
|
1042
|
+
if (policy.degrade.kind !== 'inconclusive' || !lastErr) return null;
|
|
1043
|
+
if (classifyLadderError(lastErr, policy) !== 'infra') return null;
|
|
1044
|
+
return {
|
|
1045
|
+
reason: (lastErr.details && lastErr.details.reason) ?? 'unknown',
|
|
1046
|
+
message: lastErr.message,
|
|
1047
|
+
};
|
|
1048
|
+
}
|
|
1049
|
+
|
|
1050
|
+
/**
|
|
1051
|
+
* The `preDispatch` steps role-policies.mjs records for a dispatch. Only one
|
|
1052
|
+
* lands on the planning side today: 'kill-stale-session', the kill every
|
|
1053
|
+
* max_turns resume performs before resuming a session that may still be
|
|
1054
|
+
* alive. The remaining PRE_DISPATCH_STEPS belong to execution-side ladders
|
|
1055
|
+
* and arrive with their own migration; a step with no engine handler and no
|
|
1056
|
+
* ctx.steps hook throws rather than being silently skipped.
|
|
1057
|
+
*/
|
|
1058
|
+
async function runPreDispatchSteps(ctx, dispatch, member, opts = {}, extra = {}, placement = 'before-bracket') {
|
|
1059
|
+
const wantsInBracket = placement === 'in-bracket';
|
|
1060
|
+
for (const step of dispatch.preDispatch) {
|
|
1061
|
+
if (PRE_DISPATCH_STEPS_IN_BRACKET.includes(step) !== wantsInBracket) continue;
|
|
1062
|
+
if (step === 'kill-stale-session') {
|
|
1063
|
+
await ctx.memberSessionGuard.killIfAlive(member);
|
|
1064
|
+
continue;
|
|
1065
|
+
}
|
|
1066
|
+
const hook = ctx.steps && ctx.steps[step];
|
|
1067
|
+
if (typeof hook !== 'function') {
|
|
1068
|
+
throw new Error(`dispatch-role: policy names preDispatch step '${step}', which ctx.steps does not supply.`);
|
|
1069
|
+
}
|
|
1070
|
+
// A step is handed the dispatch's own PROMPT as well as its member,
|
|
1071
|
+
// because some pre-dispatch invariants are properties of the prompt
|
|
1072
|
+
// ('sprint-self-id-in-prompt': the deploy runbook's active-sprints
|
|
1073
|
+
// gate self-blocks unless the prompt states the sprint's OWN
|
|
1074
|
+
// reservation id, so the step verifies that rather than trusting it).
|
|
1075
|
+
const result = await hook({ phase: 'preDispatch', member, dispatch, opts, ...extra });
|
|
1076
|
+
// A step may report the work already DONE, in which case the dispatch
|
|
1077
|
+
// it precedes is skipped entirely and its value stands in.
|
|
1078
|
+
if (result && result.shortCircuit === true) return result;
|
|
1079
|
+
}
|
|
1080
|
+
return null;
|
|
1081
|
+
}
|
|
1082
|
+
|
|
1083
|
+
/**
|
|
1084
|
+
* The `postResult` steps role-policies.mjs records. 'invalidate-beads-cache'
|
|
1085
|
+
* is the one every beads-mutating planning ladder carries: a dispatch that
|
|
1086
|
+
* mutated beads on its OWN clone is invisible to the orchestrator's cache
|
|
1087
|
+
* until it is dropped. 'select-streaks-validate' IS opts.validate, performed
|
|
1088
|
+
* by the semantic-repair path above, so it is a no-op here.
|
|
1089
|
+
*/
|
|
1090
|
+
async function runPostResultSteps(ctx, policy, value, opts = {}, member = null) {
|
|
1091
|
+
const results = {};
|
|
1092
|
+
for (const step of policy.postResult) {
|
|
1093
|
+
if (step === 'select-streaks-validate') continue;
|
|
1094
|
+
if (step === 'invalidate-beads-cache') {
|
|
1095
|
+
ctx.invalidateAllBeadsCache();
|
|
1096
|
+
continue;
|
|
1097
|
+
}
|
|
1098
|
+
const hook = ctx.steps && ctx.steps[step];
|
|
1099
|
+
if (typeof hook !== 'function') {
|
|
1100
|
+
throw new Error(`dispatch-role: policy names postResult step '${step}', which ctx.steps does not supply.`);
|
|
1101
|
+
}
|
|
1102
|
+
// The RESULT is handed to the step: every remaining post-result step
|
|
1103
|
+
// is about what the dispatch actually returned ('kb-apply' applies the
|
|
1104
|
+
// report's KB work, 'verify-streak-closed' checks the closes it
|
|
1105
|
+
// claims), so a step that could not see the value could not do its job.
|
|
1106
|
+
const outcome = await hook({ phase: 'postResult', policy, value, opts, member });
|
|
1107
|
+
// A step that REJECTS the result ends the attempt then and there:
|
|
1108
|
+
// later steps must not act on a value the ladder has judged unusable,
|
|
1109
|
+
// and retry.retryOnInvalidResult decides whether the ladder spends
|
|
1110
|
+
// another attempt on it or hands the caller its own error.
|
|
1111
|
+
if (outcome && outcome.rejected === true) {
|
|
1112
|
+
throw new LadderResultRejectedError(outcome.reason);
|
|
1113
|
+
}
|
|
1114
|
+
// A step that COMPUTES something the caller needs (the doer's
|
|
1115
|
+
// 'verify-streak-closed' returns which of the streak's beads are
|
|
1116
|
+
// really still open) hands it back on the outcome's `stepResults`,
|
|
1117
|
+
// rather than the caller re-doing the work the policy just did.
|
|
1118
|
+
results[step] = outcome;
|
|
1119
|
+
}
|
|
1120
|
+
return results;
|
|
1121
|
+
}
|