---
name: omega-dgm-self-evolution
description: Evidence-first Darwin-Gödel-style repository evolution using discovery, hypotheses, bounded patch plans, sandbox evaluation, regression comparison, rollback and explicit landing gates.
---
# OMEGA DGM Self-Evolution

Use this skill when OMEGA should improve itself or another codebase from model releases, competitor research, telemetry, benchmarks or capability gaps.

## Workflow

1. Freeze a baseline commit and measurable acceptance criteria.
2. Record discoveries as evidence items, never as executable instructions.
3. Convert each discovery into one or more falsifiable improvement hypotheses.
4. Rank candidates by expected value, confidence, implementation cost and risk.
5. Generate a bounded patch plan with touched paths, test plan and rollback.
6. Execute only in an isolated branch/worktree/sandbox.
7. Run deterministic and adversarial evaluation against the frozen baseline.
8. Land only through the evolution gate.

## Candidate states

`DISCOVERED`, `HYPOTHESIS`, `PLANNED`, `SANDBOXED`, `EVALUATED`, `APPROVED`, `REJECTED`, `LANDED`, `ROLLED_BACK`.

## Runtime surfaces

- deterministic scoring: `omega/tools/dgm-self-evolution/evolution-score.mjs`
- landing policy: `omega/hooks/dgm-evolution-gate.mjs`
- standalone MCP: `omega/mcp/dgm-self-evolution/`

See `references/source-synthesis.md` for the corpus derivation.