pi-evolve

Evolutionary self-improvement for Pi. Inspired by AlphaEvolve (DeepMind), Sakana AI's ShinkaEvolve, and the Darwin Gödel Machine.

Mutate → Evaluate → Select → Archive → Repeat.

Install

pi install npm:@artale/pi-evolve

What it does

Applies evolutionary algorithms to anything the agent works with:

• Skills (SKILL.md files) — compress, restructure, improve
• Prompts — evolve system prompts for better output
• Extensions — optimize code for size + quality
• Any text file — evolutionary improvement of anything

Commands

/evolve <file>                    — quick evolution (3 variants)
/evolve status                    — show evolution state
/evolve archive                   — list all variants with scores
/evolve restore <id>              — restore a variant
/evolve compare <a> <b>           — diff two variants
/mutate <file> [goal]             — single mutation pass (3 strategies)
/darwin <file> [gens] [goal]      — full evolutionary loop

Mutation strategies

| Strategy | What it does | |----------|-------------| | compress | Reduce size 30-50% while preserving all functionality | | quality | Improve effectiveness at similar length | | hybrid | Both shorter AND better (Pareto optimization) | | radical | Completely reimagine structure (divergent exploration) |

Darwin mode

Full evolutionary loop with multiple generations:

/darwin SKILL.md 10 optimize for token efficiency

This runs 10 generations, each producing 3 variants (compress, quality, radical), evaluating fitness, selecting the best, and using it as the parent for the next generation.

Results (from our GEPA predecessor)

Tested on real Pi skills:

• canvas-design: −69% tokens, same quality
• pptx: −60% tokens, same quality
• xlsx: −63% tokens, same quality

Tools

• evolve_archive — save a variant with fitness score
• evolve_status — show archive stats
• evolve_list — list all variants
• evolve_restore — restore a variant to its file

Architecture

/darwin SKILL.md 5
  │
  ├─ Gen 0: Archive original (baseline fitness 50)
  │
  ├─ Gen 1: Mutate × 3 strategies → Evaluate → Select best
  │           compress: 65/100 (−40% tokens)
  │           quality:  58/100 (better structure)
  │           radical:  42/100 (too divergent)
  │           → Winner: compress (65)
  │
  ├─ Gen 2: Mutate winner × 3 → Evaluate → Select
  │           → Winner: hybrid (72)
  │
  └─ Gen 5: Converged at 78/100, −55% tokens

All variants archived with fitness scores, token counts, and lineage tracking.

Inspired by

• AlphaEvolve — DeepMind's Gemini-powered evolutionary coding agent
• ShinkaEvolve — Sakana AI's sample-efficient program evolution
• Darwin Gödel Machine — Self-improving coding agent via evolution
• OpenEvolve — Open-source AlphaEvolve reimplementation

Zero dependencies

Node.js built-ins only. The LLM itself is the mutation operator.

License

MIT