The hook ran `pnpm test`, the whole coverage suite, and the agent flow in AGENTS.md ran that same suite again before committing: one commit verified identical content twice, about 45s of it.
`pnpm run check` (scripts/check.sh) runs lint, typecheck and coverage concurrently, prints each tool's complete output (also kept in .tmp/check-logs) and fails when any of them does. A green run fingerprints the worktree into .tmp/check.stamp, so the hook skips the suite for content that just passed; any tracked or untracked edit, or a stamp older than an hour (gitignored inputs like lib/ and node_modules can drift unseen), re-runs it, and `--full` always does. The hook gains lint and typecheck at no extra wall time: 3.2s and 1.4s hide inside the 23s test run.
.gitignore now lists the local tool state (.codegraph, .idea, .zcode): it is not a gate input, and its churn would otherwise invalidate every stamp.
Verified: green gate on 146 files / 2305 tests, hook skip in 0.1s, full re-run after any edit, and a stubbed failing tool exits 1 without stamping.
bowenliang123
e0b42ee6a7