fix(router): coalesce uses unique jobIds + name-as-coalesceKey to stop silently dropping events

[zbigniewsobiecki]

🚨 Live regression fix

PR #1226's coalesce flow silently drops PM status-changed events when a prior coalesced job for the same (projectId, workItemId) is in any state other than 'delayed' / 'waiting'. Verified live on prod 2026-04-29 16:13:25 UTC: user moved MNG-422 from planning to splitting; webhook decision logged Coalesced dispatch scheduled: splitting agent for work item MNG-422; no worker ever spawned, splitting agent silently dropped.

Root cause

scheduleCoalescedJob reused a deterministic jobId = coalesce:${coalesceKey}. BullMQ's add(name, data, { jobId }) is a silent no-op when a job with that id already exists in the completed/failed/active set, and the queue keeps completed jobs for 24h via removeOnComplete: { age: 86400 }. The if/else-if chain handled 'delayed' | 'waiting' | 'active' but FELL THROUGH for 'completed' / 'failed' / 'paused' / 'waiting-children' — silently dropping the new event. And even the 'active' early return (activeExists: true) caused the caller to drop the event with a "Coalesced dispatch skipped: active job already running" decision reason — also wrong, because the user's status change is real new intent.

Fix

• Each schedule produces a UNIQUE jobId (coalesce_<safeKey>_<ts>_<rand>). BullMQ never silently drops it.
• coalesceKey is stored as the BullMQ "job name" — the supersede pass reads only delayed/waiting jobs and filters by name to find prior pending events to remove.
• Active/completed/failed prior jobs are NOT consulted: an active job is busy with the previous unit of work and the new event becomes its own delayed dispatch behind it; completed/failed jobs are done and the new event is real new intent.
• The new jobId format is colon-free because BullMQ rejects custom ids with colons unless they have exactly 3 colon-separated parts (legacy repeatable-job compatibility). The colon-free form also avoids the Docker-container-name issue patched earlier today.
• Caller (webhook-processor.ts) drops the activeExists branch entirely. Supersede + lock-cleanup loop preserved for delayed/waiting case.

Test plan

• Unit: tests/unit/router/queue.test.ts — 8 tests covering no-prior-job, delayed/waiting supersede, COMPLETED/FAILED/ACTIVE non-blocking (the headline MNG-422 regression pins), name-filter-isolation, unique-id-per-call, colon-free-format.
• Unit: tests/unit/router/webhook-processor.test.ts — existing activeExists test rewritten to pin the new contract: active jobs do NOT block, locks ARE marked, decisionReason is "scheduled" not "skipped".
• Integration: tests/integration/coalesce-bullmq.test.ts — real-Redis tests updated for the unique-id + name-filter contract; supersede on delayed/waiting verified end-to-end. Completed/failed scenarios documented as covered by unit suite (moving real BullMQ jobs to completed/failed requires worker-lock-token plumbing not worth duplicating).
• Full unit suite: 475 files / 8715 tests green.
• Integration: coalesce-bullmq.test.ts 4/4 green.
• Typecheck + lint clean.
• Post-deploy: move a Linear issue through planning → splitting on prod ucho; verify both agents fire (planning runs; while it's still running, move to splitting; splitting fires once the planning worker exits OR fires its own delayed dispatch independently after the 10s window).

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!