JoyCode Agent addresses the challenge of automated software repair for real-world open-source issues. It provides an LLM-powered pipeline for generating patches, creating and verifying tests, and intelligently retrying fixes. This system offers a high-performance, cost-efficient solution for developers and researchers seeking to automate software maintenance tasks, achieving state-of-the-art results on the SWE-bench dataset.

How It Works

This project employs an end-to-end LLM-driven pipeline for robust code repair. Its core innovation lies in patch-test co-generation, where tests are automatically created alongside patches, enabling comprehensive validation through a closed-loop "Generate → Validate → Refine" cycle. Intelligent failure attribution and targeted retry strategies, powered by a multi-agent architecture (including specialized Testing, Patch, CSR, and Decision agents), allow for precise root cause analysis and optimized repair attempts, mimicking human developer workflows.

Quick Start & Requirements

Installation involves cloning the repository, creating a Conda environment with Python 3.11, and installing dependencies via pip install -r requirements.txt. Key prerequisites include Docker with access to docker.1ms.run and API keys for LLM services (e.g., OpenAI, Anthropic). Users must configure LLM details in llm_server/model_config.json and specify target instances in instance_id.txt. The primary execution command is python run_patch_pipeline.py.

Highlighted Details

• Achieved a 74.6% resolution rate on the SWE-bench Verified split, ranking second globally.
• Offers 30-50% lower resource consumption compared to top competitors, providing an exceptional cost-performance ratio.
• Features smart test generation (Fail2Pass, Pass2Pass) with pre-validation and collaborative verification of patches and tests.
• Employs intelligent failure attribution for precise root cause analysis and targeted retry strategies.
• Utilizes a multi-agent architecture with specialized roles and a React-based "Observe-Think-Act" workflow.
• Supports containerized execution within isolated Docker environments for repository-level understanding and multi-file coordination.

Maintenance & Community

The provided README does not contain specific details regarding community channels (e.g., Discord, Slack), active contributors, sponsorships, or a public roadmap.

Licensing & Compatibility

The project is licensed under the permissive MIT License, allowing for broad compatibility, including commercial use and integration into closed-source projects without copyleft restrictions.

Limitations & Caveats

Successful operation is contingent on correctly configured LLM API access and a functional Docker environment capable of pulling images from docker.1ms.run. Performance is benchmarked on the SWE-bench dataset, and real-world effectiveness may vary. The system's complexity, involving multiple agents and containerized environments, may require significant technical expertise for setup and troubleshooting.