Completion Verify

Run the quality gates a task must clear before you tell the user it is done.

Usage

/completion-verify              # verify current uncommitted work
/completion-verify --staged      # verify staged changes only

Behavior

Phase 1: Diff scope

git diff --stat
git diff --name-only

Confirm the changed files match what the task actually required. Flag anything unrelated.

Phase 2: Build

npm run build 2>&1 | tail -20 || pnpm build 2>&1 | tail -20

Stop and fix before continuing if the build fails.

Phase 3: Type check

npx --no-install tsc --noEmit 2>&1 | head -30
# or: pyright . 2>&1 | head -30

Phase 4: Lint

npm run lint 2>&1 | head -30
# or: ruff check . 2>&1 | head -30

Phase 5: Tests

npm test 2>&1 | tail -50
# or: pytest -q 2>&1 | tail -50

Report pass/fail counts. Never disable or skip a failing test to make this pass.

Phase 6: Quick security check

git diff | grep -E "sk-|api[_-]?key|password\s*=|BEGIN.*PRIVATE KEY" | head -10

Phase 7: Verdict

COMPLETION CHECK
Scope:    N files, all within task
Build:    PASS/FAIL
Types:    PASS/FAIL (N errors)
Lint:     PASS/FAIL (N warnings)
Tests:    PASS/FAIL (X/Y passed)
Security: PASS/FAIL (N findings)

Verdict: READY / NOT READY

Only claim the task is done when every gate that applies to the project passes.

Examples

Clean finish:

/completion-verify
→ 3 files changed, build PASS, types PASS, tests 42/42, no findings
  Verdict: READY

Caught issue:

/completion-verify
→ tests 40/42 (2 failed in UserService.test.ts)
  Verdict: NOT READY — fix failing tests before claiming done

Token Optimization

Expected range: 500–1,500 tokens (initial), 100–300 tokens (cache hit)

Caching: Caches detected build/test/lint commands in .claude/cache/completion-verify/config.json for 7 days, invalidated when package.json (or equivalent manifest) changes.

Early exit: If git diff --stat is empty, reports "nothing to verify" and exits immediately.

Patterns used: Git diff scope default, caching, grep-before-read for the security check, early exit.

Edge Cases

• No test suite: Report tests as N/A, note the gap, still verify build/types/lint.
• No build step (e.g. interpreted scripts): Skip Phase 2, note it as N/A.
• Monorepo: Scope build/test/lint to the packages touched by the diff, not the whole repo.
• Nothing staged or changed: Exit early with "nothing to verify".

Safety

Read-only: runs build/lint/test commands and greps the diff, but makes no file changes itself. If any command mutates files (e.g. an autofix lint script), do not run it here — surface the failure and let the user choose to fix it.

See Also

• /loop-verify — an independent checker that gates loop-produced changes with an APPROVE/REJECT/ESCALATE_HUMAN verdict; use that when a separate agent must validate another agent's automated change. /completion-verify is for verifying your own session's work before claiming completion.
• /test — full context-aware test runner with failure analysis and auto-fix; /completion-verify calls tests as one gate among several, not a replacement for it.