[SaulMoro]

Summary

CI is slow from work, not from queueing: runners start in 3–13 s. Two jobs set the wall time. On main and internal PRs it was the credentialed E2E job, which ran every E2E file on one runner under the fixture-repo lock. On every PR it is the macOS unit leg, whose suite spends 43% of its test time in 44.6k git processes.

 e2e  (fixture token, one runner, e2e-fixture-repo lock)
-  every E2E file                        913 s median on main
+  src/__tests__/e2e/e2e.test.ts only    the one file that uses the fixture repo
 fork-e2e-shard × 3  (no credentials)
-  fork PRs only
+  every PR and push                     slowest shard 374–449 s
 lint-and-test × 6
-  --coverage on every leg
+  --coverage on Ubuntu Node 20 only     the one leg that uploads it

Work the unit suite did for nothing:

 locateExclude(dir)       a config dir not created yet (.codex, .kiro/settings)
-  start git there, fail, answer null, uncached: again on every call
+  answer null without git                 same answer, 35% fewer git processes in mcp-reconcile and uninstall
 updateFileLocked         tests: _setExcludeLockRetryMs(1)   mcp-reconcile
 runLocalDriver           tests: _setDriverPollMs(20)        4 workflow driver files

Production keeps the lock retry and the poll interval.

Evidence

• Before: E2E (GitHub provider, full surface): 913 s median over 40 main pushes.
  After: slowest credential-free shard: 449 s, 398 s and 419 s in this PR's runs (-51%).
• Before: slow macOS unit legs: 838 s and 849 s job time (Vitest 807–814 s). Run.
  After: 713 s and 729 s (Vitest 675–698 s), -15%. Run.
• Per file on macOS, before → after: mcp-reconcile.test.ts 136–145 s → 95–102 s; workflow-workspace.test.ts, touched only by the coverage change, 160–172 s → 123–133 s.
• Coverage alone, paired on one macOS runner with the order swapped: about 67 s per leg.
• Locally: mcp-reconcile.test.ts + uninstall.test.ts start 10918 → 7115 git processes; 51–52 s → 44–45 s and 17–18 s → 16 s, three runs each. Unit suite 12180 passed; tsc and lint clean.
• Real CLI (npm run build, then npm run test:e2e): git-exclude-worktrees-915, git-exclude-delivered-915, claude-mcp-local-915: 73 passed.

Not measured: the credentialed e2e job, which fork PRs skip, so its first run is the merge to main. macOS runners vary: the same commit takes 300 s on some and 700–850 s on others; the numbers above compare the slow ones.

Ruled out: raising Vitest workers on the 3-vCPU macOS runner. 6 workers halved the leg, but 8 timed out submodule tests in workflow-workspace.test.ts, so the headroom is too thin to rely on.

Merge Danger

Door: two-way

Blast Radius: CI, and teamai pull / teamai uninstall when they look up a git exclude file for an agent directory that does not exist (same answer, without starting git).

If the narrowed e2e job fails on main, a revert restores the single-runner job.

[github-actions]

No confident findings in the specified PR diff.

The description documents sufficient testing: it records two credential-free E2E shard runs and links the CI run, alongside unit-test results. The explicitly deferred credentialed E2E run is not blocking; the review rules require one representative real-CLI run, not the full provider × agent matrix.

Review was read-only; I did not run, build, or install PR code.

[github-actions]

No confident findings in the specified PR diff. The earlier review raised no findings, so none require a resolution update.

The description documents sufficient testing: two credential-free E2E shard runs, a linked CI run, and unit-test results. The explicitly deferred credentialed E2E run is not blocking; one representative real-CLI run suffices under the review rules.

Review was read-only. I did not run, build, or install PR code.

[jeff-r2026]

LGTM. CI/test-only change; approving per maintainer decision.