gate

Gate — Self-Healing Verification

Run checks, read errors, fix them, repeat. Prove code WORKS by executing it.

Anti-Shortcut Rules

1. NEVER declare PASS from reading source. Execute and observe output.
2. NEVER declare PASS without command output. "It should work" is not evidence.
3. If a check can't execute, report BLOCKED — never fake PASS.

Step 0: Detect the Stack and Get Commands

Read .claude/project.local.json (local, gitignored override) if present, else .claude/project.json, for build commands. If not, detect the stack and infer the right commands:

Stack marker	Lint	Typecheck	Build	Test
package.json	npm run lint	npm run typecheck / tsc --noEmit	npm run build	npm test
go.mod	golangci-lint run	(types checked at build)	go build ./...	go test ./...
pyproject.toml / requirements.txt	ruff check . / flake8	mypy .	(no compile step)	pytest
Gemfile	rubocop	(Sorbet if present)	bundle exec rake assets:precompile	bundle exec rspec
pom.xml	mvn checkstyle:check	(compiled)	mvn package	mvn test
Cargo.toml	cargo clippy	(compiled)	cargo build	cargo test

Also check Makefile — many projects centralize all commands there (make lint, make test, make build).

Or init from the plugin CLI if available:

${CLAUDE_PLUGIN_ROOT}/scripts/cli init

Optional: Audit pass

Before running the loop, gate can invoke /audit to clean the branch first:

/audit          # de-slop + DRY + dead code on changed files only

Run audit when: the user says "clean before shipping", the diff is large and messy, or this is a pre-release gate. Skip audit for hotfixes and small patches — the loop is enough.

Audit scope when invoked from gate: files changed on this branch only (git diff main...HEAD). Full-codebase audit is a separate invocation.

The Loop

Run up to 4 iterations:

1. Execute: lint → typecheck → build → test
2. All pass? → GATE PASSED — stop.
3. Any fail? → Read full error, fix it, run again.

Fix Rules

Error Type	Fix	Don't
Type errors	Fix the type, add the import	Use @ts-ignore
Build errors	Fix imports, exports, modules	Skip the check
Lint errors	Fix the actual issue	Blanket disable
Test failures	Fix the code or the test	Delete the test

NEVER change business logic during gate. Only fix types, imports, lint.

Circuit Breaker

After 4 iterations without full pass:

1. Report which checks still fail with last error output
2. Do NOT fake a PASS
3. Inform user or team lead

CLI Reference

Flag	What
-l	Lint only
-c	Typecheck only
-b	Build only
-t	Tests only
-a	App startup (dev server + health check)
-u	UI tests (Playwright/Cypress)
-g	Full gate: lint + typecheck + build + test
--all	Everything including app + UI