rocm Verification Before Completion
Ensures fixes and changes are actually validated before declaring success. Prevents "it compiles so it works" thinking.
Core Principle
Never mark work done based on absence of errors. Verify the positive — confirm the expected behavior exists.
Verification Checklist
Before marking any task complete, verify:
1. Build Verification
- Code compiles without new warnings
- All existing tests still pass
- Package builds successfully (if applicable)
2. Behavioral Verification
- The fix actually addresses the reported problem (not just a related symptom)
- Expected output is produced (not just "no error")
- Edge cases considered: empty input, null handles, max values, concurrent access
3. Regression Check
- No existing functionality broken by the change
- API contract preserved (same inputs → compatible outputs)
- No new warnings from static analysis or formatters
4. Integration Verification
- Changes work across all affected layers (C → Python → CLI if applicable)
- Both system-installed and pip-installed paths work (for Python changes)
- Both RPM and DEB packaging still work (for build changes)
Anti-Patterns
| Anti-Pattern | Reality |
|---|---|
| "It compiles" | Compilation proves syntax, not correctness |
| "Tests pass" | Tests may not cover the changed path |
| "No errors in output" | Silent failures are the worst failures |
| "Works on my machine" | Check both install contexts |
| "The diff looks right" | Run the code, don't just read it |
When to Use
- After fixing a review finding
- After completing a build/install cycle
- After refactoring code
- Before marking a subagent task as done