AutoCodeRover — AI Bug Fixing with AST-Aware Code Search

Introduction

AutoCodeRover is an autonomous software engineering agent that resolves GitHub issues by combining large language models with program structure-aware code search. Unlike agents that search codebases via plain text matching, AutoCodeRover navigates the abstract syntax tree to find relevant methods, classes, and code paths. This structure-aware approach produces more accurate patches at a fraction of the cost of manual debugging.

What AutoCodeRover Does

• Automatically generates patches for GitHub issues and bug reports
• Navigates codebases through abstract syntax tree analysis, not just text search
• Uses statistical fault localization with test suites to pinpoint bug locations
• Supports multiple LLM backends including GPT-4, Claude, and open-source models
• Operates in three modes: GitHub issues, local repos, and SWE-bench tasks

Architecture Overview

AutoCodeRover works in two phases. In the context retrieval phase, the agent uses AST-aware search APIs to explore the project structure, examining classes, methods, and call graphs to gather relevant code context. In the patch generation phase, it synthesizes a fix based on the collected context. When test suites are available, the system applies statistical fault localization to narrow down the likely bug location before generating patches. The entire pipeline runs in a Docker container for reproducibility.

Self-Hosting & Configuration

• Clone the repository and build the Docker image from the provided Dockerfile
• Set API keys for your chosen LLM provider via environment variables
• Configure the target repository, commit hash, and issue link as CLI arguments
• Output patches are written to a configurable output directory
• Supports GPT-4o, Claude, Llama, Groq, and AWS Bedrock as backends

Key Features

• AST-aware code search that understands program structure, not just text
• Statistical fault localization using existing test suites for precise bug targeting
• Multi-model support across commercial and open-source LLM providers
• Reproducible execution via Docker with deterministic environment setup
• SWE-bench integration for benchmarking against standardized bug-fixing tasks

Comparison with Similar Tools

• SWE-agent — uses a custom shell interface for code navigation; AutoCodeRover uses AST search
• Aider — interactive pair programming tool; AutoCodeRover is fully autonomous
• OpenHands — general-purpose agent; AutoCodeRover specializes in bug fixing with fault localization
• GitHub Copilot — code completion assistant; AutoCodeRover handles end-to-end issue resolution
• Agentless — simpler patch generation without AST search; lower accuracy on complex bugs

FAQ

Q: What programming languages does AutoCodeRover support? A: It primarily targets Python repositories due to its AST tooling, but the LLM-based approach can handle other languages with reduced structural analysis.

Q: How much does it cost per issue? A: Under $0.70 per task on average when using GPT-4o, based on SWE-bench evaluations.

Q: What is SWE-bench? A: SWE-bench is a standardized benchmark of real GitHub issues from popular Python projects, used to evaluate automated software engineering agents.

Q: Can I use AutoCodeRover with private repositories? A: Yes. Point it at a local clone of your private repository using the local mode.

Sources

• https://github.com/nus-apr/auto-code-rover