An end-to-end LLM-powered repository-level code repair pipeline open-sourced by JD, featuring intelligent patch generation and collaborative test verification on SWE-bench benchmark, achieving 74.6% resolve rate on SWE-bench Verified with 30-50% lower resource consumption.

JoyCode Agent#

Overview#

JoyCode Agent (JoyCode SWE-bench Agent Pipeline) is an automated code repair system open-sourced by JD OpenSource. It automates the software defect repair workflow—from understanding Issue descriptions, locating code problems, generating repair patches, to creating verification test cases and executing closed-loop verification—forming a complete "Problem → Fix → Verify" automation chain.

Core Capabilities#

High Performance & Cost Efficiency#

Achieves 74.6% resolve rate on SWE-bench Verified dataset (373/500)
30-50% lower resource consumption compared to top competitors

Patch-Test Collaborative Generation#

Intelligent Test Generation: Automatically creates Fail2Pass and Pass2Pass tests with pre-validation
Collaborative Verification: Patch and test co-generation for comprehensive validation
Closed-loop Iteration: "Generate → Validate → Refine" cycle replaces single-generation mode

Intelligent Failure Attribution#

Root cause analysis: Precisely distinguish Patch issues vs. test issues
Targeted retry: Experience-driven retry strategy based on failure analysis
CSR-driven learning: Historical success pattern retrieval optimization

Multi-Agent Architecture#

Specialized agents: Testing Agent, Patch Agent, CSR Agent, Decision Agent
ReAct-based "Observe-Think-Act" cycle
LLM-driven Patch voting selection mechanism

Engineering Features#

Containerized execution: Isolated environment based on SWE-bench Docker images
Repository-level understanding: Multi-file coordination and cross-module reasoning
Complete logging: Full trajectory recording with optional compression
Multi-LLM support: Flexible model configuration for different pipeline stages

Architecture#

Pipeline Stages#

Container Setup: Pull and start SWE-bench Docker image
Test Generation (optional): Create and validate tests on original code
Agent Execution: Use LLM Agent via cli.py to generate Patch
Verification: Run tests and evaluate Patch quality
Post-processing (optional): Trajectory compression, similarity matching, intelligent retry

Repository Structure#

joycode/
├── run_patch_pipeline.py           # Main entry
├── cli.py                         # Core Agent CLI
├── test_case_generator/           # Test generation logic
├── test/                          # Test execution & verification
├── utils/docker_utils.py          # Container management
├── llm_server/                    # LLM integration layer
└── vote.py                        # Patch voting system

Installation & Quick Start#

Requirements#

Python 3.11+
Docker (access to docker.1ms.run required)
LLM API keys (OpenAI, Anthropic, etc.)

Installation#

git clone https://github.com/jd-opensource/joycode-agent.git
cd joycode
conda create -n joycode python=3.11
conda activate joycode
pip install -r requirements.txt

Quick Start#

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

# Single instance
python run_patch_pipeline.py --problem-id django__django-11099 --num-processes 1

# Batch processing
python run_patch_pipeline.py --num-examples 10 --num-processes 4

Performance#

Results on SWE-bench Verified dataset:

Repository	Resolve Rate
pytest-dev/pytest	17/19 (89.47%)
scikit-learn/scikit-learn	28/32 (87.5%)
pydata/xarray	19/22 (86.36%)
django/django	178/231 (77.06%)
sympy/sympy	57/75 (76.0%)
sphinx-doc/sphinx	29/44 (65.91%)
matplotlib/matplotlib	25/34 (73.53%)
astropy/astropy	13/22 (59.09%)
Total	373/500 (74.6%)

Use Cases#

Automatic open-source project Issue repair
Verification test case generation
Repository-level code patch generation and verification
Automated code repair pipeline construction
Enterprise CI/CD integration for bug fixing
AI programming research and SWE-bench baseline comparison

JoyCode Agent