AutoCodeRover
Autonomous software engineering agent for program repair.
About
AutoCodeRover is a project structure-aware autonomous software engineering agent from the National University of Singapore's Automated Program Repair group, aimed at autonomous program improvement. It iteratively retrieves relevant program context through AST-level code search APIs rather than plain string matching, then generates patches, integrating software debugging capabilities including spectrum-based fault localization when test cases are available.
On SWE-bench it reported resolving 37.3% of SWE-bench lite tasks and 46.2% of SWE-bench verified tasks at pass@1, with each task costing under $0.70 (researcher-reported figures). A successor, AutoCodeRover-v2, infers program specifications such as function summaries, reproducer tests, and feedback on generated patches to help developers judge patch correctness. The work was published at ISSTA 2024.
What it can do
- AST-level project-structure-aware code retrieval
- Autonomous patch generation for GitHub issues
- Spectrum-based fault localization with test suites
- AutoCodeRover-v2 spec inference (summaries, reproducers)
- SWE-bench evaluated repair performance
Setup
Clone https://github.com/nus-apr/auto-code-rover, install dependencies, and configure an LLM API key. See the arXiv paper (2404.05427) for the method details.
Prerequisites
Python environment, an LLM API key, and a target repo with (optionally) a test suite.
Use cases
Works with
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Verified via vendor site/docs/GitHub and reputable coverage, wave 107; Trending score 76/100 from researcher assessment.
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.