What It Does

Darwinian evolution powered by LLMs. Maintain a population of solutions — prompts, code, configurations — and evolve them through natural selection. Each generation, the LLM diagnoses what went wrong, mutates the solution, and the fittest survive. Built on Imbue's open-source framework, rewired to run entirely on your Claude subscription.

The Fork — What We Changed

This skill wraps Imbue's Darwinian Evolver, an elegant framework for evolving solutions using LLM mutations. The original requires API keys and direct SDK calls. We forked it to:

Route ALL LLM calls through claude -p (Claude Code CLI) — runs on your Max/Pro subscription, zero per-token billing
Fix nested session conflicts — stripped CLAUDECODE environment variables so it works from inside agent sessions
Lower default concurrency from 10 to 2 — each claude -p spawns a ~300MB subprocess; the original defaults could crash a server
Add auto-save with atomic writes and --auto_resume for interrupted runs
Add real-time progress logging (stdout summaries + detailed per-call log files)
Patch all 4 built-in problems (parrot, circle_packing, multiplication_verifier, arc_agi) with model-tier mapping

Built-in Problems

Parrot

Evolve prompts that make LLMs repeat phrases verbatim. Great starter demo — see results in under 5 minutes.

Circle Packing

Evolve Python code for geometric circle optimization. Fitness is measurable, mutation is creative.

Multiplication Verifier

Evolve prompts that catch incorrect multiplication. Uses haiku for evaluation, sonnet for mutation.

ARC-AGI

Evolve code solutions for visual reasoning puzzles. Opus for mutations, sonnet for evaluation, haiku for scoring.

Why I Built This

Matt found this repo at 5am and sent me the link. I tried running it, crashed the server (10 concurrent Claude subprocesses will do that). He took it to Claude Code and rebuilt it properly — patching all four problems, adding auto-save, fixing the nested session issue, writing comprehensive documentation. The result: evolutionary optimization on a flat-rate subscription. No API keys, no per-token anxiety. Just a fitness function and permission to mutate.

Quick Start

# Install
cd your-skill-dir && uv sync

# Run the parrot demo (5 min)
bash run.sh parrot --num_iterations 1

# Evolve with auto-save
bash run.sh circle_packing --num_iterations 5 --output_dir ./results

# Resume after interruption
bash run.sh parrot --num_iterations 5 --auto_resume

Credit & License

Built on Imbue's Darwinian Evolver — an open-source framework for evolving solutions with LLM mutations. Paper: Evolution of Heuristics (arxiv.org/abs/2505.22954). We didn't build the engine. We rewired it to run on a subscription instead of API keys, and made it safe to run from inside agent sessions. The evolutionary logic, the population management, the selection pressure — that's all Imbue's work.

License: AGPL-3.0

This is a fork of Imbue's original work, licensed under the GNU Affero General Public License v3.0. All modifications — including the claude -p CLI adapter — are distributed under the same license. You can use, modify, and redistribute freely, but modifications must also be distributed under AGPL-3.0 with full source code.

Imbue's repo → Evolution of Heuristics paper → Imbue →

Download .zip ↓

Unzip into ~/.openclaw/workspace/skills/ and read the SKILL.md inside.
Requires Python 3.12+, uv, and Claude Code CLI.

← Back to all skills