PyHealer: Agentic Self-Repair for Python Program Execution

PyHealer is an LLM-backed, agentic execution error repair pipeline for Python projects. It automatically analyzes execution failures, infers dependencies and Python versions, builds isolated Docker environments, and applies iterative fixes using specialized LLM agents.

The system is intended for research, experimentation, and automated program repair. Key features include:

LLM-driven API extraction

Automatic dependency and Python version inference

Docker-based isolated execution

Agentic feedback loop with specialized agents (Dependency, Code, Python, System)

Meta-Agent for failure classification

Iterative repair attempts

Structured JSON reports and metrics

The pipeline relies on pre-built knowledge artifacts for dependency resolution, Python feature mapping, and system error correction. Users can run PyHealer on target Python projects or benchmark datasets (e.g., Gistable Hg2.9k) to automatically detect and repair execution errors.

Prerequisites: Python 3.8+, Docker, and the Ollama local LLM runtime with the codellama:70b model.

License: (Specify your license here, e.g., MIT, Apache 2.0)

Repository / Files Structure:

PyHealer.py – main script

Example_Project/ – sample target project

Knowledge Data/ – knowledge artifacts for repair agents

Analysis Report of PyHealer Repairs

This file "24 code-level repairs generated by PyHealer - 16 repairs were purely corrective.zip" contains the detailed analysis reports for 24 code repairs performed by PyHealer on real-world Python Gists. Each report documents the original purpose of the program, the issue identified, the repair applied by PyHealer, and the resulting semantic impact.

The analyzed repairs are categorized according to the nature of the modifications:

16 repairs (66.7%) were purely corrective, addressing issues such as undefined variables, missing imports, incorrect attribute accesses, Python-version compatibility problems, and other implementation defects. These repairs fully preserved the intended behavior of the original programs.

3 repairs (12.5%) involved minor simplifications, such as removing redundant constructs, modernizing syntax, or simplifying implementation details while maintaining the same functionality and observable behavior.

5 repairs (20.8%) repairs required more substantial modifications and are analyzed individually with respect to their functional impact and behavioral preservation.

For each repair, the report provides:

The Gist identifier.

A concise description of the original purpose of the code.

A summary of the change performed by PyHealer.

An assessment of the semantic impact of the repair.

These reports support the evaluation of PyHealer's effectiveness in repairing real-world Python programs while preserving their intended functionality.