Self-Evolution (OpenClaw)

Teaches the OpenClaw agent to run self-modification cycles on its own workspace: one cycle = one coherent transformation ending in a commit (and optional restart). The agent uses the workspace repo as the target of change; constitution/identity files (if present) are protected.

When this skill applies: User says "evolve", "run evolution", "self-improvement cycle", or the system enqueues an evolution task. The skill is loaded when eligible (e.g. git on PATH per metadata.openclaw.requires).

One Cycle = One Transformation

Each evolution cycle must produce at most one coherent change (or an explicit refusal with reason). No "analyze only" cycles — evolution is defined as a commit.

Cycle Steps (in order)

1. Assessment — Read the workspace codebase. Identify where change has maximum leverage (complexity, clarity, performance, or alignment with constitution).
2. Selection — Choose one transformation. Prefer high leverage over small tweaks.
3. Implementation — Implement fully. No "80% done" — complete and clean.
4. Smoke test — Run tests or a minimal verification before commit. If the project has tests (e.g. npm test, pytest), they must pass before push.
5. Multi-model review (optional) — For significant changes (new modules, architecture, security): send diff or key files to 2–3 different LLMs with a clear review prompt; treat as advice, not mandatory.
6. Constitution check — If the workspace has a constitution (e.g. BIBLE.md, SOUL.md): verify the change does not violate it. Identity/core files are untouchable unless the constitution allows amendments.
7. Commit (+ optional restart) — Bump version/changelog if applicable, commit with a clear message, push. Request restart only after a successful push so the new code loads.

Invariants

• Evolution = commit. A cycle that only produces a report or analysis is not evolution; the next cycle should either complete the change or explicitly close the direction.
• One transformation per cycle. Prevents scope creep and keeps history interpretable.
• Constitution overrides convenience. Do not delete or weaken identity/core documents to "simplify" — that is amnesia, not evolution.
• Tests gate push. If the project has automated tests, failing tests block push; fix or revert.

Circuit Breaker (if running continuously)

When evolution runs in a loop (e.g. queue empty → enqueue next evolution):

• After N consecutive failed cycles (e.g. no commit, or cost/rounds below threshold), pause evolution and notify the user. Require explicit resume.
• When remaining budget is below a reserve (e.g. for normal conversation), stop evolution and notify.

Tools the agent uses

The OpenClaw agent relies on:

• Read workspace: list and read files (including constitution and identity).
• Write workspace: write file(s), then commit (and optionally push) via git or exec.
• Run tests: exec test command (e.g. npm test, pytest); interpret pass/fail.
• Optional: multi-model review (e.g. LLM task or external API), restart request.

Ensure the session has access to git and the workspace; sandbox settings may need to allow exec and file writes.

Output per cycle

After each cycle, report briefly to the user:

• What was changed (file(s), intent).
• Whether multi-model review was used and outcome.
• New version or commit ref if applicable.
• If the cycle was refused: reason.

Detailed reference

For full step-by-step checklist and constitution constraints, read the bundled reference: {baseDir}/references/cycle.md.