Repository navigation
feat(observability): export standalone queue, ownership and execution signals #1252
Description
Why
Dispatch needs to reveal eligible work, waiting assignments, hand-outs, stale ownership, PR-fix settlement and sync/grooming health independently of Courier. Implement the signal and state catalogue in #1237 / PR #1239.
Work
- Implement a bounded, read-only snapshot/exporter for queue depth/age, eligible and withheld work, assignment vs observed active execution, Lease/AgentWork, PR-fix hand-out/ack/reclaim, operational PARKED vs human BLOCKED, and sync/scheduler/groomer freshness.
- Separate task execution signals from groomer/heartbeat/legacy AgentRuns via the accepted allowlist; don't infer a task pickup from a row written only at report completion.
- Use the actual persisted IssueSyncRun statuses (
running,completed,failed) and an absent-series/no-success baseline. Honor fresh-sync/eligibility guards when claiming a fleet has stopped. - Expose the inputs required for aged specific unadmitted assignments and PR-fix overdue unacked hand-outs. A continuously nonzero aggregate count in a busy queue is not evidence that one item is stranded.
- Windowed database-derived failure metrics are gauges, not monotonic counters. Express window-specific counts clearly; never apply counter-rate/
deltarules that lose window events. - Bounded enum labels only. No repo, issue, PR, worker identity, arbitrary error text or secrets in metric dimensions. No coupling to Courier phases or home-ops/Grafana installation.
Acceptance
- Healthy idle, busy steady churn, genuinely stranded assignment, no first successful sync, failing discovery, inactive worker, and held dependency produce distinguishable observed/unknown states.
- Tests prove a fleet with overlapping but promptly claimed assignments is not reported as 'aged stranded'; an actual aged assignment is. Windowed audit-failure arrivals at equal counts are still observable.
- Requesting/exporting metrics is read-only; no claim release, requeue, status mutation, credential changes or mandatory backend.
Dependencies
Design complete as of 2026-10-18: #1237 is closed and #1239 merged. The implementation is READY at P3. No technical prerequisite remains; Dispatch's own priority/aging scheduler decides ordering relative to other ready tasks. Coordinate with #1202/#1203 lifecycle-event foundation; do not duplicate it. Dashboards and executable alerts are separate.
Activity
its-saffron commentedon Oct 19, 2026
Grooming note: this is moving off
status/readypending verification, not because the ask is unclear — it is well bounded with concrete acceptance criteria. The design-complete claim in the Dependencies section (#1237 closed, #1239 merged) and the state of #1202/#1203 could not be confirmed from evidence captured this pass, and no standalone exporter implementation was verifiable onmain. Once the design catalogue and the current metrics surface (e.g.docs/observability.md,prisma/schema.prisma, the API route layer) are captured, it can be re-groomed to ready.feat(observability): export standalone queue, ownership and execution…