Verification Before Completion

Evidence Requirements

To declare something DONE, you MUST provide:

• Command run — The exact command you executed
• Output snippet — Relevant portion of output showing results
• Pass/fail tally — "5/5 tests passed" or "2 failures in auth module"
• Timestamp — When you verified it

Example:

✅ Verification Complete

**Command:** pytest tests/ -v
**Output:** 
  tests/auth/test_login.py::test_valid_user PASSED
  tests/auth/test_login.py::test_invalid_user PASSED
  tests/auth/test_password_reset.py PASSED
  5 passed in 0.24s

**Timestamp:** 2026-04-13 14:32 UTC

What "Works" Means by Task Type

Failing Verification

If verification fails:

1. Identify specific failures — Which tests? Which output doesn't match?
2. Route back to implementer — Describe failures with exact output
3. Implementer fixes — Changes implementation, re-runs verification
4. Iterate — Keep verifying until it passes
5. Never claim "mostly works" — Either it passes or it doesn't

Example failure report:

❌ Verification Failed

**Command:** pytest tests/ -v
**Output:**
  tests/auth/test_login.py::test_valid_user FAILED
  AssertionError: Expected 200, got 401
  tests/auth/test_password_reset.py PASSED
  1 failed, 1 passed

**Issue:** Login test failing with 401. Check token generation in implementer fix.

Common Pitfalls

❌ Claiming "all tests pass" when output shows failures ❌ Not running verification (just trusting the code looks right) ❌ Running only partial tests (run full suite relevant to task) ❌ Skipping manual verification for user-facing features ❌ Suppressing or simplifying failing checks to manufacture green result ❌ Saying "this should work" instead of actually running commands

✅ Always run the command, capture output ✅ Report outcomes faithfully (failures are OK, misreporting is not) ✅ Complete all relevant tests before marking done ✅ If you can't verify (no test exists, can't run), say so explicitly

Report Template

After verifying every completed task:

## Verification Summary

**Task:** [Task name]
**Branch:** [git branch if applicable]

### Tests Run
- Command: `[exact command]`
- Result: [X/Y passed, Z failed]
- Output: [relevant snippet]

### Manual Checks (if applicable)
- [Feature smoke test]: [result]
- [Integration point]: [result]

### Evidence Collected
- ✅ All specified tests pass
- ✅ Output matches expected
- ✅ No regressions detected

### Conclusion
**Status:** [DONE / NEEDS_FIX with specific issues]

Integration

This skill is invoked after subagent-driven-development completes all tasks.

It gates the final "task complete" state: you cannot declare done without verification evidence.

If verification fails, it routes back to implementer subagent with specific failures for rework.

Success Criteria

• ✅ Evidence provided for every passing task
• ✅ All relevant tests run and pass
• ✅ Manual verification completed for user-facing features
• ✅ Failures reported immediately (not hidden)
• ✅ No "probably works" claims — only "verified" or "not yet verified"

Philosophy

Verification before completion is not about extra bureaucracy. It's about confidence.

The benefit of running pytest -v is not just that tests pass — it's that YOU KNOW they passed because you watched it happen. No assumptions, no guesses, just evidence.

That evidence is what lets you confidently report "done" to the user instead of "probably done, let's hope."