RepoRepair

RepoRepair is a novel documentation-driven approach for repository-level automated program repair, which leverages hierarchically generated code documentation to achieve precise fault localization and cost-effective patch generation across diverse programming languages. Evaluated on both SWE-bench Lite and Multimodal benchmarks, it achieves state-of-the-art repair rate (45.67% on Lite, 37.13% on Multimodal) while maintaining superior cost efficiency.

Key Features

• 📚 Documentation-Aware:
  Uses LLM-generated code documentation for cross-file context understanding.
• 🌐 Language-Agnostic Design:
  Supports JavaScript/TypeScript and Python repositories through AST-based parsing and generalized documentation generation.
• 💰 Cost Efficiency:
  Achieves average repair costs of $0.44 on SWE-bench Lite and $0.56 on SWE-bench Multimodal.

Performance Highlights

Metric	RepoRepair	Agentless Lite	Improvement
Lite Results
%Resolved	45.67%	32.33%	+13.34%
Avg. Cost/Repair	$0.44	$0.21	+0.23
Multimodal Results
%Resolved	37.13%	25.34%	+11.79%
Avg. Cost/Repair	$0.56	$0.38	+0.18

*Agentless Lite uses different model configurations across benchmarks

Installation

git clone https://github.com/ZhongQiangDev/RepoRepair.git
cd RepoRepair
pip install -r requirements.txt  # Requires Python 3.9+

Usage

1. Resource Download

# Download issues and repositories
python issue_diff_download.py

python issue_repo_download.py
python unzip.py

• Use Selenium to fetch the repository's compressed file from GitHub.

2. Repository Parsing

# Parse code and analyze dependencies
python CodeParser.py  # Uses Tree-sitter for PY/JS/TS parsing
python DependencyGraph.py
python generate_doc_meta.py  # Output: repo_doc_meta/

3. Code Documentation Generation

# Generate documentation at different levels
python generate_document_func.py  # Output: repo_document_func/

python generate_document_file.py  # Output: repo_document_file/

• Cloud resources are released in Google Drive: https://drive.google.com/file/d/1jAWcQy3HM-Fu37r1CAoBO5MerSGFluRp/view?usp=sharing, https://drive.google.com/file/d/1rAiVn4o5FK-OHqPoGoc-GoDWXxJqQp5I/view?usp=sharing.

4. File Retrieval

# Analyze and retrieve relevant files
python ps_cause_analyze.py  # Output: problem_statement_analysis/

python file_retrival.py  # Uses LangChain, output: repo_file_rag/

5. Localization

# Hierarchical localization
python file_localization.py  # Output: buggy_files/

python func_localization.py  # Output: buggy_elements/

6. Repair

# Generate patches
python bug_repair.py  # Output: bug_repair/

Directory Structure

├── repo_doc_meta/              # Parsed repository metadata
├── repo_document_func/         # Function-level documentation
├── repo_document_file/         # File-level documentation
├── problem_statement_analysis/ # Issue analysis results
├── repo_file_rag/              # Retrieved files
├── buggy_files/                # Localized problematic files  
├── buggy_elements/             # Localized functions/classes
└── bug_repair/                 # Generated patches