ci(cmux-next): implicit batch queue for feat-cmux-next

[azooz2003-bit]

Summary

PRs into feat-cmux-next land one at a time today, and each re-runs the full CI on every push. The coordinator also stacks open UI PRs by hand, regenerates the agent-pane bundle and builds the stack on the fleet. This PR makes that loop an implicit merge queue. Agents open PRs and the queue lands them.

.github/workflows/cmux-next-batch.yml runs scripts/ci/next_batch.py. docs/ci/cmux-next-batch.md has the full rules.

• Trigger: PR events and feat-cmux-next pushes debounce into one batch: 2 minutes of quiet, at most 10 minutes after the first event. A running batch is never cancelled; a newer dispatch waits.

• Eligible: not a draft, a branch in this repo, no hold/exploration/needs a call label, head newer than 5 days, and no red check outside the heavy tier (pending is fine). Only PRs by the batch authors (default teamleaderleo) are auto-landed; anyone else opts in with batch-queue.

• Stack: up to 12 PRs in PR-number order on the feat-cmux-next head. Generated files are never resolved by hand. A conflict keeps the stack's copy, and the generator rebuilds it once:

  • web bundles on Linux;
  • SDK bindings from the merged cmux-tui/spec, which merges key by key;
  • Swift exports (action contracts, settings schema, MDM) on a mini.

  Any other conflict drops that PR and comments on it, once per head.

• Validate once: dispatch cmux-next.yml on next-batch/<run>-<n>, which runs every tier (enabled by ci(cmux-next): run every tier when dispatched on a next-batch branch #17528). In parallel, a mini submits a fleet --production build of the stack. Jobs already red on feat-cmux-next don't count, and failed jobs get one rerun before any PR is blamed.

• Land or bisect:

  • Green: each PR lands with main's gh-merge-green --squash after a comment linking the batch. --override names the batch only when the PR's own heavy check is red.
  • Red: halving over stack prefixes finds the culprit. It gets a comment with the failing jobs, and the batch reruns without it.

• Report: the job summary, plus a sticky comment on the HQ issue named by CMUX_NEXT_BATCH_STICKY.

Controller logic runs on Blacksmith Linux, or for now as next_batch.py serve on a workstation under the operator's gh login (no App yet). In serve mode, its pushes and merges start CI like anyone's, and the fleet build uses the host's cmux-ci. Generators run the stack's code, so they never run on the controller. regen-linux (web bundles, SDK, webviews formatting) and regen (Swift exports, on a mini) check out the stack with a read-only token and return a patch. The controller applies and pushes that patch.

• Formatting never blocks a landing: a PR changing webviews/ whose only red is web / react-apps-check or web / Web status stays eligible. Before landing it, regen-linux runs bun run check:fix (formatter and lint autofix) on its head, and the queue pushes the fix to its branch.
• Mini refusals: a job a runner refused before writing a result is dispatched once more.

Limits

• The job token merges and pushes, and GitHub starts no workflows for those events. After landing, the batch dispatches cmux-next.yml, cmux-tui-artifacts.yml and the next batch on feat-cmux-next. The routing App has no contents permission, so it can't do this either.
• If two batch PRs both changed a generated file, GitHub can't squash the second after the first lands. The queue comments on that PR, which needs a merge of feat-cmux-next plus regeneration. Pushing that fix automatically would need a token whose pushes start CI.
• The dev-build page from feat: publish fleet dev builds as update tracks #17473 isn't on feat-cmux-next yet, so the build link goes only to the sticky report.

Testing

• python3 -m unittest tests.test_next_batch: 50 tests. They cover the serve debounce, remote regeneration, the local fleet build, formatting before landing, eligibility, debounce timing, JSON key merge, real-git stacking (clean, generated conflict, source conflict, spec merge, clean-but-both-changed generated file), bisect over 1 to 8 PRs (at most 3 probes), sticky history, landing override and the workflow's concurrency and trust shape. They failed on 7d49a1a (no controller) and pass on 168b6a9.
• actionlint: clean. python3 scripts/verify-local.py --affected origin/feat-cmux-next: 8 of 9 pass; localization parity fails with 27,403 errors in catalogs this PR doesn't touch.
• Local stack of the 11 open lane PRs (next_batch.py stack --no-regen), 24 s:
  • fix(cmux-next): keep hidden WebViews out of scene captures #17370 is dropped: it conflicts with feat-cmux-next.
  • cmux-next sidebar: head the workspace list with Projects #17526 is dropped: it conflicts with cmux-next sidebar: opt-in tab counts and a per-workspace tabs chevron #17520 in sidebar source.
  • Agent-pane and settings-schema conflicts were marked for regeneration.
  • regenerate-web-bundles.sh rebuilt them in 40 s.
• End-to-end on real open PRs, next-batch/20261006 (11 requested PRs, feat-cmux-next 7cf558be8de1):
  • Stack, 20 s:
    • Merged fix(cmux-next): Cmd-T keeps the pane's content until the agent page paints; new terminal tabs titled with their folder #17527, cmux-next: one model picker, a clickable context ring and an aligned composer footer #17537, fix(cmux-next): drop a workspace on a pane to merge it; quiet refused zones; steady drop zones #17542, cmux-next: Home's first run shows only actions; onboarding Done opens New Tab #17553, ci(cmux-next): stop rescue churn, one Mac package build per PR #17554 and cmux-next: live reload Ghostty config #17559, with 24 agent-pane conflicts and 3 Swift export conflicts kept for regeneration and 2 xcstrings key unions.
    • Dropped cmux-next sidebar: opt-in tab counts and a per-workspace tabs chevron #17520, cmux-next: size the composer plus menu to content #17534, cmux-next: run ! commands in place instead of jumping to a terminal #17541, fix(cmux-next): App Store pages join Back/Forward; centered detail page without capsules #17551 and cmux-next sidebar: drag a workspace onto another to group them #17552. Each conflicts with feat-cmux-next itself in source, and each got a comment.
  • Regeneration:
    • Web bundles on Linux (bun 1.4.2): about 4 min, including dispatch, giving f16f40ea1d1b.
    • Swift exports on a mini, third attempt, giving 954248bfe072. The first attempt ran the stack's copy of the generator script, which the stack doesn't have yet (fixed: the script now comes from the workflow ref). The second was refused at runner setup (fixed: one more dispatch).
  • Fleet production build: c8f563681211aa65c6e94453, tag next-batch-f16f40ea. It went from submitted to green in 2 min 57 s. The submission first failed because cmux-ci refuses --backend-mode local with --production (fixed).
  • Landing: not done yet; see the PR thread.

Changelog

none

🤖 Generated with Claude Code

---

Migrated from #17535 after correcting the PR author identity. The original head commit 405bf26 is preserved.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

---

Summary by cubic

Adds an implicit merge queue for feat-cmux-next (docs/ci/cmux-next-batch.md has the full rules). PR events debounce into one batch that stacks eligible PRs, rebuilds conflicted generated files, validates the stack once across every cmux-next tier plus a fleet production build, then lands each PR with gh-merge-green or bisects a red stack to its culprit. PRs no longer land one at a time with full CI on every push.

Batch rules

• Eligible: not a draft, branch in this repo, no hold/exploration/needs a call label, head newer than 5 days, no red check outside the heavy tier; batch authors (default teamleaderleo) auto-land, others opt in with batch-queue.
• PRs touching .github/workflows/ are never queued; they land by hand.
• A generated-file conflict keeps the stack's copy and the generator rebuilds it once; other conflicts drop the PR and comment on it.
• A webviews-changing PR whose only red check is web formatting stays eligible; the queue runs bun run check:fix on its head and pushes the fix before landing.
• The token's merges and pushes start no workflows, so the batch dispatches cmux-next.yml, cmux-tui-artifacts.yml, and the next batch after landing.
• If two batch PRs changed the same generated file, the second can't be squashed after the first lands; the queue comments and that PR needs a manual merge and regen.

^{Written for commit 405bf26. Summary will update on new commits.}

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

• Configuration used: Repository: manaflow-ai/cmux/.coderabbit.yaml
• Review profile: ASSERTIVE
• Plan: Advanced
• Run ID: 45b4583a-7ec1-4242-b68c-787ce34e2759

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}