claude-evolve

Automated algorithm evolution using AI. Start with a base algorithm, let Claude evolve better variants autonomously.

Install & Quick Start

# Install
npm install -g claude-evolve

# Set up project
claude-evolve setup

# Generate initial ideas
claude-evolve ideate

# Start evolution (runs forever until Ctrl+C)
claude-evolve run

How It Works

1. Write your problem in evolution/BRIEF.md
2. Create base algorithm in evolution/algorithm.py
3. Define evaluation in evolution/evaluator.py
4. Generate ideas - Claude creates algorithm variations
5. Evolve automatically - System tests variations, keeps best ones, generates new ideas

Evolution runs indefinitely until you stop it. Perfect for overnight optimization.

Commands

Core Commands

claude-evolve           # Interactive menu
claude-evolve setup     # Initialize workspace
claude-evolve ideate    # Generate new algorithm ideas
claude-evolve run       # Start evolution loop (runs forever)
claude-evolve analyze   # View results and progress with charts
claude-evolve status    # Quick progress overview
claude-evolve autostatus # Live auto-updating status display

Management Commands

claude-evolve edit      # Manage candidate statuses
claude-evolve config    # View/edit configuration
claude-evolve cleanup   # Clean up old lock files

Maintenance Commands

claude-evolve clean-invalid      # Remove invalid candidates
claude-evolve cleanup-duplicates # Remove duplicate entries
claude-evolve csv-fix           # Fix CSV formatting issues

Working with Multiple Projects

# Use different working directory
claude-evolve --working-dir=my-project run
claude-evolve --working-dir=experiments/trading ideate

Project Structure

your-project/
├── evolution/
│   ├── BRIEF.md           # Problem description
│   ├── algorithm.py       # Base algorithm
│   ├── evaluator.py       # Performance measurement
│   ├── config.yaml        # Settings
│   ├── evolution.csv      # Progress tracking
│   └── evolution_*.py     # Generated variants

Evaluator Requirements

Your evaluator.py must output a performance score to stdout. The system supports multiple output formats:

Format 1: Simple numeric value

# Just print a single number
print(1.234)

Format 2: JSON with performance/score field

# JSON with 'performance' field (recommended)
print('{"performance": 1.234, "accuracy": 0.95, "latency": 45.2}')

# OR JSON with 'score' field
print('{"score": 1.234, "precision": 0.88, "recall": 0.92}')

Format 3: SCORE: prefix (legacy)

# For backward compatibility
print("SCORE: 1.234")

Important notes:

• Higher scores = better performance
• When using JSON, all fields are saved to the CSV for analysis
• The performance or score field is required for evolution decisions
• Return value of 0 doesn't mean failure - it's just a low score
• Actual failures should exit with non-zero status code

Configuration

Edit evolution/config.yaml:

# Files
algorithm_file: "algorithm.py"
evaluator_file: "evaluator.py" 
evolution_csv: "evolution.csv"

# Evolution strategy
ideation_strategies:
  total_ideas: 15
  novel_exploration: 3    # Creative new approaches
  hill_climbing: 5        # Parameter tuning
  structural_mutation: 3  # Architecture changes
  crossover_hybrid: 4     # Combine best features

# Auto-generate new ideas when queue empty
auto_ideate: true

# Parallel execution
parallel:
  enabled: false
  max_workers: 4

Requirements

• Node.js 14+
• Python 3.x
• Claude CLI
• Bash shell (Git Bash on Windows)

Tips

• Start simple - Basic algorithm, let evolution add complexity
• Monitor progress - Check evolution.csv for performance trends
• Guide evolution - Add manual ideas when stuck in local optima
• Let it run - Evolution works best over long periods

Common Issues

Too many failures? Check your evaluator handles edge cases and outputs valid scores.

Stuck in local optimum? Increase novel_exploration in config.yaml or add manual ideas.

Evaluator crashes? Make sure dependencies are installed and error handling is robust.

License

MIT