feat(007): robust review dispatch — per-type lock + post-completion hook

[zbigniewsobiecki]

Summary

Addresses the silent review drop observed on MNG-122/PR-572: the review agent was never dispatched because the work-item lock's total-concurrency cap blocked it while 2 other agents were enqueued for the same work item. The blocked check-suite-success webhooks were silently discarded with no retry.

Spec 007 delivers two intersecting fixes across two plans.

Plan 1 — Lock infrastructure (per-agent-type)

Removes MAX_WORK_ITEM_CONCURRENCY total cap from isWorkItemLocked. Only the per-type cap (MAX_SAME_TYPE_PER_WORK_ITEM = 1) remains. Different agent types can now run concurrently on the same work item.

• src/router/work-item-lock.ts — deleted MAX_WORK_ITEM_CONCURRENCY, removed total-count checks (in-memory + DB), simplified getInMemoryCounts → getInMemorySameTypeCount (single key lookup, no iteration), removed the dbTotal query (saves one DB round-trip per lock check), deleted dead keyPrefix helper.
• src/router/webhook-processor.ts — enriched lock-skip log with source and blockedAgentType.
• Net: -33 lines (simplification).

Plan 2 — Post-completion review dispatch

When an implementation agent succeeds with a PR, the execution pipeline now checks CI status and fires the review agent before the container exits.

• src/triggers/shared/agent-execution.ts — new tryDispatchPostCompletionReview function. Uses claimReviewDispatch with the same dedup key format as check-suite-success, so the two paths cannot double-enqueue. Uses the recursive runAgentExecutionPipeline pattern (same as splitting → backlog-manager chain). Best-effort: errors are logged as warn but never break the implementation pipeline.

Tests

• Plan 1: +2 new cross-type concurrency tests, ~6 updated existing tests reflecting the total-cap removal. 16 work-item-lock tests total.
• Plan 2: +7 new tests covering: fires review on success + green CI, skips for non-implementation, skips on failure, skips when no prUrl, skips when CI not green, skips when already dispatched, swallows errors gracefully.
• Total: 7859 unit tests pass. Lint + typecheck + build clean.

Documentation

CLAUDE.md updated under "Agent triggers" with:

• Per-agent-type lock semantics note
• Post-completion review dispatch documentation

Out of scope

• ❌ Enabling pr-opened for review on any project (per-project config decision)
• ❌ Lock persistence to Redis/DB (in-memory with 30-min TTL is sufficient)
• ❌ Generalizing the hook to other chains (splitting → planning, etc.)
• ❌ Distributed lock across multiple router instances (single-process deployment)

Test plan

• npm test (7859 pass)
• npm run typecheck, npm run lint clean
• After deploy: trigger a Linear-backed implementation run on llmist, watch the review agent fire from the post-completion hook within seconds of implementation completion — no reliance on check-suite-success webhook timing.

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 98.95833% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/triggers/shared/agent-execution.ts	98.71%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!