check

name	check
description	Run local quality checks (typecheck, lint, build, tests). Quick check during development.

Local Quality Check

Run all local quality checks. Use during development to verify the current state quickly.

cdk-real-drift (cdkrd) is developed solo, but on a GitHub remote and through PRs: work lands via wt-* worktree branches, and .claude/hooks/verify-pr-gate.sh gates gh pr create / gh pr merge. This skill is the LOCAL half of that flow — it runs no real-AWS deploy/destroy — and mirrors the CI workflow (.github/workflows/ci.yml), which runs typecheck / lint+format / build / unit tests on every push.

Steps

Run these sequentially and report results:

1. vp run typecheck — tsc --project tsconfig.json --noEmit.
2. vp check --fix — lint + Prettier formatting, with auto-fix. Use this, not vp run lint:fix: CI runs vp check (which includes formatting), and lint:fix does NOT touch formatting — so a lint:fix-only run can pass locally while CI fails with formatting issues on the same branch.
3. vp pack (tsdown ESM bundle to dist/) — and run it BEFORE step 4, because a fresh worktree has no dist/ and 13 tests fail without it (they spawn the built CLI). Either invocation rebuilds: build is cache: false in vite.config.ts and has been since the toolchain landed, so vp run build cannot serve a stale dist/. What HAS caused a false-negative live-test is a dist/ nobody rebuilt: a stale binary that lacked the change under test.
4. vp test run (Vitest unit tests; tests/integration/** is excluded by vite.config.ts). The run-task cache is a real foot-gun — PR go-to-k/cdk-real-drift#438's duplicate object key passed a CACHED vp run typecheck and reached main, which is why typecheck is now cache: false — but test reported a cache MISS on every run measured on 2026-08-19 (go-to-k/cdk-real-drift#1768). Prefer the direct form; do not treat a vp run test result as suspect on cache grounds alone.

When piping any of the above to tail / head / grep, check the actual output content for Error / Command failed markers — $? after a pipeline reflects the LAST stage (usually 0), NOT the build tool's exit. When in doubt, capture without piping: vp <cmd> > /tmp/out 2>&1; rc=$?; tail -3 /tmp/out; echo "[rc=$rc]".

Output

Report as a table:

Check	Result
typecheck (vp run typecheck)	pass/fail
lint + format (vp check --fix)	pass/fail
build (vp pack)	pass/fail
tests (N files, M tests)	pass/fail

If all pass, confirm "All checks passed." If any fail, show the error output and STOP — do not write the commit-gate marker.

Before treating a failure in a file you did NOT touch as a real (or peer-introduced) main regression, rule out a stale worktree cache. A long-lived worktree's tsc/oxc cache can REPLAY phantom errors from an earlier dependency/lockfile state (the inverse of the cache MASKING real ones). If CI on main is green and the failure is in code outside your diff, REPRODUCE it in a throwaway fresh worktree (git worktree add … main → pnpm install → vp check) before reporting "main is red" or opening a fix lane — a clean fresh worktree means the error was a local cache artifact, not a regression.

Commit-gate marker (on success only)

After all four checks pass, record the check marker so the markgate check gate is satisfied. The marker captures the current working-tree state of the gate's scope; any subsequent edit in that scope invalidates it and requires re-running /check. Since go-to-k/cdk-real-drift#1837 that scope is src/**, tests/**, the build inputs (package.json, pnpm-lock.yaml, pnpm-workspace.yaml, tsconfig*.json, vite.config.ts, .mise.toml), this gate's own definition (.markgate.yml), and the checker-INPUT files the unit suite reads from outside src/tests — .claude/skills/**, .claude/settings.json, .claude/hooks/**, scripts/**, .releaserc.json, .github/workflows/**, plus the hand-written markdown OUTSIDE those trees (README.md, DESIGN.md, CLAUDE.md, CONTRIBUTING.md, docs/**, demo/README.md), which tests/markdown-fmt-corruption-1771.test.ts reads.

A markdown edit therefore stales check — with exactly one exception. Three buckets: README.md / DESIGN.md / docs/** are also the docs gate's subject, so they need /check AND /check-docs; CLAUDE.md, CONTRIBUTING.md and demo/README.md are check-only inputs, because the docs gate includes just src/**, docs/**, README.md and DESIGN.md; and CHANGELOG.md is in NEITHER gate — it is machine-generated by semantic-release and is the one file the markdown scanner itself excludes, so an edit there stales nothing. (The scanner's population is every tracked *.md bar that one, so tests/integration/README.md is in it too — already covered by tests/**.) Read .markgate.yml for the authoritative list, not this sentence: each entry there names the test that reads it, and this copy has already gone stale once (go-to-k/cdk-real-drift#1837 widened the gate and had to repair the enumeration here in the same PR).

Run from the root of the tree you are WORKING in — the worktree, not the main checkout, whenever the lane lives in one. The marker store is .git/markgate, shared by every worktree, but the hashes come from the cwd's files, so setting it from the main checkout records main's content instead of yours: measured 2026-08-19, with the worktree dirty and the marker set from the main checkout, markgate verify check returns rc=1 from the worktree and rc=0 from main. It fails CLOSED, so the cost is a wasted gate cycle plus a "run /check first" message right after you ran it. Also set it in its OWN command, separate from the git commit — check-gate is a PreToolUse hook and judges the call before anything in it runs, so a markgate set … && git commit one-liner is blocked in full.

cdkrd pins markgate via mise, so use mise exec to avoid PATH issues when shims aren't active:

mise exec -- markgate set check

Skip this step if any check failed — a stale or missing marker correctly forces re-running /check after fixing the failure.