JoyCode Agent

Docker Setup

Ensure Docker is running and can pull images from the required registry:

docker pull docker.1ms.run/swebench/sweb.eval.x86_64.django__django-11099:latest

Instance Selection

To specify which SWE-bench instances to process, list their IDs in a file named instance_id.txt. Example format:

django__django-11099
matplotlib__matplotlib-23562
sympy__sympy-18189

To use JoyCode, you need Python 3.11+, Docker (with access to docker.1ms.run), and LLM API keys (e.g., OpenAI, Anthropic). Follow these steps to set up your environment:

1. Clone the repository
2. Create and activate a Conda environment
3. Install dependencies via pip

# Clone repository
git clone https://github.com/jd-opensource/joycode-agent.git
cd joycode

# Create conda environment
conda create -n joycode python=3.11
conda activate joycode

# Install dependencies
pip install -r requirements.txt

To compare and select between multiple patch candidates, use the voting scripts. This requires prepared voting data.

1. Prepare voting inputs using scripts/prepare_voting_data.py.
2. Run the voting process using vote.py.

Required Input Files:

• patch_1.json: Primary patch candidates
• patch_2.json: Alternative patch candidates
• test-00000-of-00001.parquet: Instance metadata with problem statements

# Prepare voting inputs
python scripts/prepare_voting_data.py

# Run voting
python vote.py

The main entry point for the pipeline is run_patch_pipeline.py. Use it to generate patches, create tests, and validate fixes for software issues.

# Run with default settings
python run_patch_pipeline.py --num-processes 1 --enable-post-processing

LLM settings are managed in llm_server/model_config.json. You must provide the api_key, base_url, and model_name for the patch_generation stage. You can also tune max_tokens and temperature.

{
  "patch_generation": {
    "api_key": "your_api_key_here",
    "base_url": "https://api.openai.com/v1",
    "model_name": "gpt-5",
    "max_tokens": 4000,
    "temperature": 1
  }
}

Docker Connectivity

If Docker issues occur, verify connectivity and test pulling a small image:

docker info
docker pull docker.1ms.run/hello-world

LLM Configuration

To verify your llm_server/model_config.json is valid and readable:

python -c "import json; print(json.load(open('llm_server/model_config.json')))"

Memory/Resource Issues

If you encounter memory issues, reduce the number of parallel processes:

python run_patch_pipeline.py --num-processes 1

Results are stored in the output_files/ directory, organized by instance ID. Each instance folder contains:

• predictions.json: The generated patch and associated metadata.
• agent_logs.txt: Main execution logs for the agent.
• test_generation_result.json: Results from the test generation stage.
• test_generation_logs.txt: Logs from the test generation stage.
• agent_logs_retry.txt: Logs specifically for retry attempts.
• compressed_trajectory.txt: A compressed version of the execution trajectory.

Global summary files include:

• successful_cases.txt: Summary of successful instances.
• failed_cases.txt: Summary of failed instances.
• empty_diff_cases.txt: Cases where no patches were generated.
• similar_case_matches_summary.json: Analysis of similar cases.

The run_patch_pipeline.py script supports several flags to control execution behavior: