预发布自我审查
Before committing significant changes, run through this checklist. Catch issues before CI catches them — or worse, before users do.
When to Use
Run this skill when any of these are true:
- Changed >100 lines of code
- Modified >5 files
- Touched core infrastructure (executor, storage, main dispatch loops)
- Introduced new abstractions or public APIs
- Uncertain about edge cases or design choices
Quick gate:
git diff --stat HEAD | tail -5 # How much changed?
git diff --name-only HEAD | wc -l # How many files?
Skip If
- Trivial change (≤50 lines, ≤3 files)
- Pure documentation or lesson updates
- Test-only changes (tests are their own review)
- Simple one-liner bug fix with obvious, limited scope
- Mechanical refactoring with no logic changes
The Checklist
1. Edge Cases
- [ ] Null/empty inputs handled?
- [ ] Boundary conditions tested (off-by-one, empty lists, zero)?
- [ ] Concurrent access or race conditions considered?
- [ ] Unicode/encoding issues if processing text?
2. Error Paths
- [ ] All exceptions caught at the right level (not too broad)?
- [ ] Error messages informative (what failed, why, what to do)?
- [ ] Failure modes fail gracefully, not silently?
- [ ] No bare `except:` or swallowed errors?
3. Test Coverage
- [ ] Critical paths have tests?
- [ ] Edge cases identified above have tests?
- [ ] Tests actually run and pass?
- [ ] No test stubs left at `pass` or `...`?
4. Documentation
- [ ] Docstrings/comments match current behavior?
- [ ] README updated if public interface changed?
- [ ] Commit message captures what changed and why?
- [ ] Relevant architecture docs still accurate?
5. Code Quality
- [ ] Naming is clear and consistent with surrounding code?
- [ ] No obvious code smells (God functions, magic numbers, hardcoded paths)?
- [ ] Follows existing patterns in the file/package?
- [ ] Static analysis / typecheck passes?
One-Liner Smoke Test
# Run before committing (adapt to your stack):
make format && make typecheck && pytest -x -q && echo "✅ Clean to land"
If any step fails, fix it before committing. Don't ship red.
Anti-Patterns
| Anti-Pattern | Fix |
|---|---|
| "CI will catch it" | CI is a last resort, not a substitute for your own judgment. Catch it first. |
| "I'll add tests later" | Later doesn't come. Write the test now or accept the gap permanently. |
| "It worked when I tested manually" | Manual tests aren't reproducible. Write an automated test so future sessions can verify it too. |
| "The error message is a TODO" | TODO error messages ship as TODO error messages. Write the real message now. |
| "I can see from the code it handles this case" | If it's not tested, a future refactor will break it invisibly. |
Why This Matters
Autonomous agents are trained to produce outputs good enough to avoid correction. Without a structured review gate, autonomous code commits bypass the human-in-the-loop check that normally catches regressions and edge cases. This skill acts as that gate — five targeted lenses run in sequence so nothing obvious slips through.
The payoff compounds: each review pass that catches a real issue is a CI round-trip saved, a user bug avoided, and a revert PR not needed.
Related
- Self-Push skill — quality/laziness check that complements this structural review
- Multi-Lens Review skill — deeper parallel review for high-stakes changes
