Parallel Claude lanes need a local merge queue — and a human promote
Jul 30 funador: FIFO land + checkCommand for multi-worktree Claude Code; promote is human-only; slow gates cap landings/hour; checkCommand≠human review (on purpose).
Source: Claude Code Merge Queue — local merge queue for parallel Claude Code agents
funador’s claude-code-merge-queue (Show HN Jul 30; MIT npm package) is the last-mile harness for the problem continuous factories hit the moment they go multi-agent: several Claude Code worktree lanes finish, build, and test at once, then race the same integration branch. The tool serializes land through a local FIFO so push races, redundant heavy builds, and shared-resource flakiness stop owning the night.
Shape of the loop: numbered lane/ worktrees (hooked into Claude’s native WorktreeCreate), build-lock for machine-wide serialized builds, land (rebase + push via the queue with checkCommand as the only landing gate), sync to fast-forward the main checkout, and promote — human-only, never in agent instructions. A pre-push hook makes land non-optional: direct push to integration is rejected; missing checkCommand fails closed by default. init writes config, CLAUDE.md (“land when green without being asked”), the WorktreeCreate hook, and scripts.
The productive complication is printed under “Know the limits,” and it is the review lesson: no human reviews any of this before it lands. checkCommand passing is the only gate — a real suite and echo ok look identical. Want a human on every change? That step is missing on purpose. Other ceilings map to factory ops: crash-safe locks by PID liveness (not timeouts); one machine, not a fleet; a slow checkCommand is a hard throughput ceiling (3–4 minute suites cap you well under ~20 landings/hour); rebase conflicts abort cleanly for the agent to resolve and re-land.
Why it matters for code review
Merge rate under sustained parallel agent load is a harness property, not a model property. Continuous overnight momentum without a serialized land path just dumps racey green branches into morning triage. Keep promote (or equivalent) human-owned, treat checkCommand as necessary but not sufficient for judgment, and budget review/verification time as the real landings-per-hour constraint — not how many Claude tabs you can open.