| name | evolution-loop |
| description | Run a bounded repository improvement loop using explicit fitness criteria, small reversible mutations, validation evidence, and human approval gates. |
Evolution Loop
Use this skill when the user wants FullSelfDriving to improve a repository, not just audit it.
Loop
- Define explicit fitness criteria.
- Inspect existing repo instructions and validation commands.
- Generate small candidate changes.
- Validate candidates with the narrowest meaningful checks.
- Score by evidence.
- Keep the best patch and discard weak candidates.
- Hand off with commands run, artifacts, risks, and next candidate improvements.
Fitness Criteria
Common criteria:
- tests pass
- lint/typecheck pass
- coverage improves without meaningless tests
- app health check passes
- screenshot or trace confirms UI behavior
- complexity decreases
- debug evidence becomes easier to collect
- isolation improves
Guardrails
- Do not delete, skip, or weaken tests to improve scores.
- Do not hide errors.
- Do not add mocks outside legitimate test seams.
- Do not run unbounded loops.
- Do not make destructive changes without explicit approval.
- Prefer one well-validated improvement over many speculative edits.