py-attractor 0.3.0

A DOT-based pipeline runner for AI workflows

Attractor

A DOT-based pipeline runner that uses directed graphs (defined in Graphviz DOT syntax) to orchestrate multi-stage AI workflows.

Overview

Attractor lets you define multi-stage AI workflows as directed graphs using Graphviz DOT syntax. The graph is the workflow: nodes are tasks, edges are transitions, and attributes configure behavior.

Based on the Attractor Specification.

Features

• Declarative pipelines defined in DOT syntax
• Pluggable handlers for different node types (LLM calls, conditionals, human gates)
• Checkpoint and resume for crash recovery
• Goal gates to enforce critical stages
• Edge-based routing with conditions and priorities
• Retry logic with exponential backoff
• Comprehensive validation and linting

Installation

pip install -e .

Quick Start

Create a simple pipeline in example.dot:

digraph Simple {
    graph [goal="Run tests and report"]
    
    start [shape=Mdiamond, label="Start"]
    exit  [shape=Msquare, label="Exit"]
    
    run_tests [label="Run Tests", prompt="Run the test suite"]
    report    [label="Report", prompt="Summarize results for $goal"]
    
    start -> run_tests -> report -> exit
}

Run the pipeline:

# Validate the pipeline
python -m attractor.cli example.dot --validate-only

# Execute the pipeline  
python -m attractor.cli example.dot

Or use the Python API:

from attractor import parse_dot, run_pipeline

graph = parse_dot("example.dot")
outcome = run_pipeline(graph)
print(f"Pipeline completed with status: {outcome.status.value}")

Documentation

• Usage Guide - Detailed usage and examples
• Attractor Specification - Complete specification

Testing

# Run all tests
pytest tests/

# Run with verbose output
pytest tests/ -v

Currently 47 tests passing, including:

• DOT parser tests
• Validation and linting tests
• Execution engine tests
• Condition evaluation tests
• Handler tests (tool, human-in-the-loop, parallel)
• Stylesheet parsing and matching tests
• Integration tests matching the spec's smoke test

Examples

See the examples/ directory for complete working pipelines:

• simple.dot - Basic linear workflow
• branching.dot - Conditional branching with goal gates

Implementation Status

Core features implemented:

• ✅ DOT parser for the supported subset
• ✅ Validation with 7 built-in lint rules
• ✅ Execution engine with retry logic and checkpointing
• ✅ Edge selection with 5-step priority algorithm
• ✅ Condition evaluation (=, !=, && operators)
• ✅ Basic handlers (start, exit, codergen, conditional)
• ✅ Tool handler (parallelogram nodes) - shell command execution
• ✅ Human-in-the-loop handler (hexagon nodes) - interactive gates
• ✅ Parallel execution handler (component nodes) - concurrent branches
• ✅ Fan-in handler (tripleoctagon nodes) - result consolidation
• ✅ Model stylesheet parser - CSS-like LLM configuration
• ✅ CLI interface
• ✅ Variable expansion ($goal)

Not yet implemented:

• ⏳ Manager loop handler (house nodes) - supervisor pattern
• ⏳ HTTP server mode
• ⏳ Observability events
• ⏳ Full parallel subgraph execution (current implementation is simplified)

License

MIT