feat(loops): enforce durable dispatch leases and retry backoff

[ncolesummers]

Summary

• persist durable dispatch ownership and enforce one active run per repository issue across loop types
• admit development runs through manifest-resolved concurrency groups, defer over-cap work, and drain queued runs after terminal release
• add bounded fixed/exponential stage retries plus trace-linked retry runs for stalled and timed-out executions
• centralize dispatch/retry telemetry and document the durable policy in proposed ADR 0017

Why

The manifest already declared concurrency and retry policy, but persistent development-run creation did not enforce either contract. Concurrent deliveries could therefore exceed the configured cap, overlap work for one issue, or lose retry intent across process restarts.

This change moves admission, lease ownership, queue ordering, finalization, and retry eligibility into database-backed transactions so restart behavior follows durable state rather than in-memory or step-status inference.

Operator and developer impact

• over-cap runs remain inspectable as queued work instead of being dropped
• terminal runs release their lease once, then best-effort promote the oldest eligible queued work
• stage failures retain forensic history while bounded retries become eligible at manifest-computed times
• stalled and timed-out runs remain immutable and link to a new planning run when retry budget remains
• research runs participate in cross-loop issue exclusivity, while full research lifecycle enforcement remains deliberately deferred

Review focus

• lock ordering, partial uniqueness, and migration preflight behavior
• total-attempt accounting across stage retries and linked terminal retries
• terminal lease release, replay repair, and queue-drain transaction boundaries
• trace propagation and low-cardinality telemetry attributes
• the deliberate absence of a hosted supervisor cadence

Validation

• bun run validate
  • 75 Vitest files, 470 tests
  • Storybook production build
  • 27 Playwright tests, including responsive and accessibility coverage
• bun run build
• focused dispatch/retry/reconciliation/migration/observability suite: 11 files, 100 tests
• git diff --check

No rendered UI changed, so Issue #96-specific browser screenshots were not applicable.

Known verification boundaries

• PGlite covers the transaction and invariant behavior but cannot prove true multi-session PostgreSQL FOR UPDATE scheduling; ADR 0017 records that native-Postgres concurrency coverage is still needed.
• The supplied Claude review artifact did not expose the resolved model identifier, so strict Opus 5 verification remains a review-process gap.
• No hosted polling cadence is added because the manifest does not define one.

Closes #96

Co-authored with OpenAI Codex (GPT-5).

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
loopworks	Ready	Preview, Comment	Jul 27, 2026 12:56am

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.