repository·main·Indexed 25 days ago
https://github.com/autocoderoversg/auto-code-roverAutoCodeRover is an autonomous program improvement tool designed to resolve GitHub issues, including bug fixes and feature additions. It combines Large Language Models (LLMs) with program-structure-aware code search and statistical fault localization. The tool supports multiple operation modes, including github-issue, local-issue, and SWE-bench tasks, and is compatible with models from OpenAI, Anthropic, Meta, AWS Bedrock, Groq, and local models via Ollama.
Do not attempt to run multiple independent processes of scripts/run.py in parallel, as different task instances may share the same local codebase, leading to conflicts.
Instead, use the built-in parallelism support by setting the num_processes field in your configuration file. The script handles the internal parallelism safely. Note: setting num_processes too high may exceed OpenAI token limits and cause task failures.
Use github-issue mode to run AutoCodeRover on a live GitHub issue. You must provide the project's clone link, a specific commit hash, and the link to the GitHub issue page. The tool will clone the repository, checkout the specified commit, and attempt to generate a patch.
Successful patch generation writes the path to selected_patch.json in the specified --output-dir.
cd /opt/auto-code-rover
conda activate auto-code-rover
PYTHONPATH=. python app/main.py github-issue --output-dir output --setup-dir setup --model gpt-4o-2024-05-13 --model-temperature 0.2 --task-id <task id> --clone-link <link for cloning the project> --commit-hash <any version that has the issue> --issue-link <link to issue page>To start the development server for the front package, use your preferred package manager to run the dev script. Once running, the application will be accessible at http://localhost:3000.
npm run dev
# or
yarn dev
# or
pnpm dev
# or
bun devTo replicate results on SWE-bench-lite, use the provided Docker image which contains all task instance environments. It is recommended to run this on Ubuntu 20.04.
docker pull yuntongzhang/auto-code-rover:experimentdocker run -it yuntongzhang/auto-code-rover:experimentsource activate base
conda activate auto-code-rovergit config --global user.email acr@nus.edu.sg
git config --global user.name acrexport OPENAI_KEY=xx-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxdocker pull yuntongzhang/auto-code-rover:experiment
docker run -it yuntongzhang/auto-code-rover:experimentTo run tests locally, you must first create and activate the auto-code-rover environment using Conda, then install tox from the conda-forge channel.
Required tools:
conda env create -f environment.yml
conda activate auto-code-rover
conda config --add channels conda-forge
conda config --set channel_priority flexible
conda install -y toxTo use local models like llama3 with AutoCodeRover:
ollama pull llama3 to download the model.--net=host or --add-host host.docker.internal=host-gateway to your docker run command so the container can communicate with the Ollama server on the host.To run AutoCodeRover on SWE-bench task instances, follow these steps:
Install the specialized SWE-bench docker fork used for regression tests:
conda activate auto-code-rover
git submodule update --init --recursive
cd SWE-bench-docker
pip install .It is recommended to set up SWE-bench on a host machine rather than inside a container.
tasks.txt file containing one task ID per line (e.g., django__django-11133).cd <SWE-bench-path>
conda activate swe-bench
python harness/run_setup.py --log_dir logs --testbed testbed --result_dir setup_result --subset_file tasks.txtThis creates a testbed directory with the source code and a conda environment for the task.
Single Task:
cd <AutoCodeRover-path>
conda activate auto-code-rover
PYTHONPATH=. python app/main.py swe-bench --model gpt-4o-2024-05-13 --setup-map <SWE-bench-path>/setup_result/setup_map.json --tasks-map <SWE-bench-path>/setup_result/tasks_map.json --output-dir output --task django__django-11133Multiple Tasks (via file):
cd <AutoCodeRover-path>
conda activate auto-code-rover
PYTHONPATH=. python app/main.py swe-bench --model gpt-4o-2024-05-13 --setup-map <SWE-bench-path>/setup_result/setup_map.json --tasks-map <SWE-bench-path>/setup_result/tasks_map.json --output-dir output --task-list-file <SWE-bench-path>/tasks.txtUse local-issue mode to run AutoCodeRover on a local codebase and a local issue description file. This is useful when you have already prepared the repository and the issue details manually.
Successful patch generation writes the path to selected_patch.json in the specified --output-dir.
cd /opt/auto-code-rover
conda activate auto-code-rover
PYTHONPATH=. python app/main.py local-issue --output-dir output --model gpt-4o-2024-05-13 --model-temperature 0.2 --task-id <task id> --local-repo <path to the local project repository> --issue-file <path to the file containing issue description>For running SWE-bench experiments or managing dependencies manually, use a local installation with Conda. This method uses an environment.yml file to manage Python dependencies.
OPENAI_KEY or ANTHROPIC_API_KEY is set in your shell before execution.Navigate to the /opt/auto-code-rover/ directory inside the container to start an experiment. By default, running the vanilla-lite configuration uses gpt-4-0125-preview on SWE-bench-lite instances.
This command consolidates all generated patches into a single file named predictions_for_swebench.json. To evaluate correctness, you must copy this file out and use either the Containerized Evaluation Harness or Moatless EvalTools.
python scripts/run.py conf/vanilla-lite.confWhen running multiple experiments sequentially, it is recommended to use a unique id in the configuration file for each run, as the output directory is named after this id.
If you wish to use the same id and overwrite previous results, use the -f flag with scripts/run.py to remove the existing experiment results for that id before starting the new run.