AutoCodeRover: Autonomous Program Improvement

ArXiv Paper        Website        Discord server

Note

This is a public version of the AutoCodeRover project. Check the latest results on our website.

📣 Updates

• [November 21, 2024] AutoCodeRover(v20240620) achieves 46.20% efficacy on SWE-bench Verified and 24.89% on full SWE-bench.
• [August 14, 2024] On the SWE-bench Verified dataset released by OpenAI, AutoCodeRover(v20240620) achieves 38.40% efficacy, and AutoCodeRover(v20240408) achieves 28.8% efficacy. More details in the blog post from OpenAI and SWE-bench leaderboard.
• [July 18, 2024] AutoCodeRover now supports a new mode that outputs the list of potential fix locations.
• [June 20, 2024] AutoCodeRover(v20240620) now achieves 30.67% efficacy (pass@1) on SWE-bench-lite!
• [June 08, 2024] Added support for Gemini, Groq (thank you KasaiHarcore for the contribution!) and Anthropic models through AWS Bedrock (thank you JGalego for the contribution!).
• [April 29, 2024] Added support for Claude and Llama models. Find the list of supported models here! Support for more models coming soon.
• [April 19, 2024] AutoCodeRover now supports running on GitHub issues and local issues! Feel free to try it out and we welcome your feedback!

Discord - server for general discussion, questions, and feedback.

👋 Overview

AutoCodeRover is a fully automated approach for resolving GitHub issues (bug fixing and feature addition) where LLMs are combined with analysis and debugging capabilities to prioritize patch locations ultimately leading to a patch.

[Update on June 20, 2024] AutoCodeRover(v20240620) now resolves 30.67% of issues (pass@1) in SWE-bench lite! AutoCodeRover achieved this efficacy while being economical - each task costs less than $0.7 and is completed within 7 mins!

[April 08, 2024] First release of AutoCodeRover(v20240408) resolves 19% of issues in SWE-bench lite (pass@1), improving over the current state-of-the-art efficacy of AI software engineers.

AutoCodeRover works in two stages:

• 🔎 Context retrieval: The LLM is provided with code search APIs to navigate the codebase and collect relevant context.
• 💊 Patch generation: The LLM tries to write a patch, based on retrieved context.

✨ Highlights

AutoCodeRover has two unique features:

• Code search APIs are Program Structure Aware. Instead of searching over files by plain string matching, AutoCodeRover searches for relevant code context (methods/classes) in the abstract syntax tree.
• When a test suite is available, AutoCodeRover can take advantage of test cases to achieve an even higher repair rate, by performing statistical fault localization.

🗎 arXiv Paper

AutoCodeRover: Autonomous Program Improvement [arXiv 2404.05427]

For referring to our work, please cite and mention:

@inproceedings{zhang2024autocoderover,
    author = {Zhang, Yuntong and Ruan, Haifeng and Fan, Zhiyu and Roychoudhury, Abhik},
    title = {AutoCodeRover: Autonomous Program Improvement},
    year = {2024},
    isbn = {9798400706127},
    publisher = {Association for Computing Machinery},
    address = {New York, NY, USA},
    url = {https://doi.org/10.1145/3650212.3680384},
    doi = {10.1145/3650212.3680384},
    booktitle = {Proceedings of the 33rd ACM SIGSOFT International Symposium on Software Testing and Analysis},
    pages = {1592–1604},
    numpages = {13},
    keywords = {automatic program repair, autonomous software engineering, autonomous software improvement, large language model},
    location = {Vienna, Austria},
    series = {ISSTA 2024}
}

✔️ Example: Django Issue #32347

As an example, AutoCodeRover successfully fixed issue #32347 of Django. See the demo video for the full process:

acr-final.mp4

Enhancement: leveraging test cases

AutoCodeRover can resolve even more issues, if test cases are available. See an example in the video:

acr_enhancement-final.mp4

🚀 Setup & Running

Setup API key and environment

We recommend running AutoCodeRover in a Docker container.

Set the OPENAI_KEY env var to your OpenAI key:

export OPENAI_KEY=sk-YOUR-OPENAI-API-KEY-HERE

For Anthropic model, Set the ANTHROPIC_API_KEY env var can be found here

export ANTHROPIC_API_KEY=sk-ant-api...

The same with GROQ_API_KEY

Build and start the docker image for the AutoCodeRover tool:

docker build -f Dockerfile.minimal -t acr .
docker run -it -e OPENAI_KEY="${OPENAI_KEY:-OPENAI_API_KEY}" acr

Setup: local mode

Alternatively, you can have a local copy of AutoCodeRover and manage python dependencies with environment.yml. This is the recommended setup for running SWE-bench experiments with AutoCodeRover. With a working conda installation, do conda env create -f environment.yml. Similarly, set OPENAI_KEY or ANTHROPIC_API_KEY in your shell before running AutoCodeRover.

Running AutoCodeRover

You can run AutoCodeRover in three modes:

1. GitHub issue mode: Run ACR on a live GitHub issue by providing a link to the issue page.
2. Local issue mode: Run ACR on a local repository and a file containing the issue description.
3. SWE-bench mode: Run ACR on SWE-bench task instances. (local setup of ACR recommend.)

[GitHub issue mode] Set up and run on new GitHub issues

If you want to use AutoCodeRover for new GitHub issues in a project, prepare the following:

• Link to clone the project (used for git clone ...).
• Commit hash of the project version for AutoCodeRover to work on (used for git checkout ...).
• Link to the GitHub issue page.

Then, in the docker container (or your local copy of AutoCodeRover), run the following commands to set up the target project and generate patch:

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>

Here is an example command for running ACR on an issue from the langchain GitHub issue tracker:

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 langchain-20453 --clone-link https://github.com/langchain-ai/langchain.git --commit-hash cb6e5e5 --issue-link https://github.com/langchain-ai/langchain/issues/20453

The <task id> can be any string used to identify this issue.

If patch generation is successful, the path to the generated patch will be written to a file named selected_patch.json in the output directory.

[Local issue mode] Set up and run on local repositories and local issues

Instead of cloning a remote project and run ACR on an online issue, you can also prepare the local repository and issue beforehand, if that suits the use case.

For running ACR on a local issue and local codebase, prepare a local codebase and write an issue description into a file, and run the following commands:

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>

If patch generation is successful, the path to the generated patch will be written to a file named selected_patch.json in the output directory.

[SWE-bench mode] Set up and run on SWE-bench tasks

This mode is for running ACR on existing issue tasks contained in SWE-bench.

Set up

Install SWE-bench Docker

We use a fork of SWE-bench docker to run regression tests (not FAIL_TO_PASS tests, but all the tests in the buggy programs). To install this, run

conda activate auto-code-rover
git submodule update --init --recursive
cd SWE-bench-docker
pip install .

Setting up Testbed

For SWE-bench mode, we recommend setting up ACR on a host machine, instead of running it in docker mode.

Firstly, set up the SWE-bench task instances locally.

1. Clone this SWE-bench fork and follow the installation instruction to install dependencies.

2. Put the tasks to be run into a file, one per line:

cd <SWE-bench-path>
echo django__django-11133 > tasks.txt

Or if running on arm64 (e.g. Apple silicon), try this one which doesn't depend on Python 3.6 (which isn't supported in this env):

echo django__django-16041 > tasks.txt

Then, set up these tasks by running: 3. Set up these tasks in the file by running:

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.txt

Once the setup for this task is completed, the following two lines will be printed:

setup_map is saved to setup_result/setup_map.json
tasks_map is saved to setup_result/tasks_map.json

The testbed directory will now contain the cloned source code of the target project. A conda environment will also be created for this task instance.

If you want to set up multiple tasks together, put multiple ids in tasks.txt and follow the same steps.

Run a single task in SWE-bench

Before running the task (django__django-11133 here), make sure it has been set up as mentioned above.

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-11133

The output for a run (e.g. for django__django-11133) can be found at a location like this: output/applicable_patch/django__django-11133_yyyy-MM-dd_HH-mm-ss/ (the date-time field in the directory name will be different depending on when the experiment was run).

Path to the final generated patch is written in a file named selected_patch.json in the output directory.

Run multiple tasks in SWE-bench

First, put the id's of all tasks to run in a file, one per line. Suppose this file is tasks.txt, the tasks can be run with

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.txt

NOTE: make sure that the tasks in tasks.txt have all been set up in SWE-bench. See the steps above.

Using a config file

Alternatively, a config file can be used to specify all parameters and tasks to run. See conf/example.conf for an example. Also see EXPERIMENT.md for the details of the items in a conf file. A config file can be used by:

python scripts/run.py conf/example.conf

Using a different model

AutoCodeRover works with different foundation models. You can set the foundation model to be used with the --model command line argument.

The current list of supported models:

	Model	AutoCodeRover cmd line argument
OpenAI	gpt-4o-2024-08-06	--model gpt-4o-2024-08-06
	gpt-4o-2024-05-13	--model gpt-4o-2024-05-13
	gpt-4-turbo-2024-04-09	--model gpt-4-turbo-2024-04-09
	gpt-4-0125-preview	--model gpt-4-0125-preview
	gpt-4-1106-preview	--model gpt-4-1106-preview
	gpt-3.5-turbo-0125	--model gpt-3.5-turbo-0125
	gpt-3.5-turbo-1106	--model gpt-3.5-turbo-1106
	gpt-3.5-turbo-16k-0613	--model gpt-3.5-turbo-16k-0613
	gpt-3.5-turbo-0613	--model gpt-3.5-turbo-0613
	gpt-4-0613	--model gpt-4-0613
Anthropic	Claude 3.5 Sonnet	--model claude-3-5-sonnet-20240620
	Claude 3 Opus	--model claude-3-opus-20240229
	Claude 3 Sonnet	--model claude-3-sonnet-20240229
	Claude 3 Haiku	--model claude-3-haiku-20240307
Meta	Llama 3 70B	--model llama3:70b
	Llama 3 8B	--model llama3
AWS Bedrock	Claude 3 Opus	--model bedrock/anthropic.claude-3-opus-20240229-v1:0
	Claude 3 Sonnet	--model bedrock/anthropic.claude-3-sonnet-20240229-v1:0
	Claude 3 Haiku	--model bedrock/anthropic.claude-3-haiku-20240307-v1:0
	Claude 3.5 Sonnet	--model bedrock/anthropic.claude-3-5-sonnet-20241022-v2:0
	Nova Pro	--model bedrock/us.amazon.nova-pro-v1:0
	Nova Lite	--model bedrock/us.amazon.nova-lite-v1:0
	Nova Micro	--model bedrock/us.amazon.nova-micro-v1:0
LiteLLM	Any LiteLLM model	--model litellm-generic-<MODEL_NAME_HERE>
Groq	Llama 3 8B	--model groq/llama3-8b-8192
	Llama 3 70B	--model groq/llama3-70b-8192
	Llama 2 70B	--model groq/llama2-70b-4096
	Mixtral 8x7B	--model groq/mixtral-8x7b-32768
	Gemma 7B	--model groq/gemma-7b-it

Note

Using the Groq models on a free plan can cause the context limit to be exceeded, even on simple issues.

Note

Some notes on running ACR with local models such as llama3:

1. Before using the llama3 models, please install ollama and download the corresponding models with ollama (e.g. ollama pull llama3).
2. You can run ollama server on the host machine, and ACR in its container. ACR will attempt to communicate to the ollama server on host.
3. If your setup is ollama in host + ACR in its container, we recommend installing Docker Desktop on the host, in addition to the Docker Engine.
   • Docker Desktop contains Docker Engine, and also has a virtual machine which makes it easier to access the host ports from within a container. With Docker Desktop, this setup will work without additional effort.
   • When the docker installation is only Docker Engine, you may need to add either --net=host or --add-host host.docker.internal=host-gateway to the docker run command when starting the ACR container, so that ACR can communicate with the ollama server on the host machine. If you encounter any issue in the tool or experiment, you can contact us via email at [email protected], or through our discord server.

Experiment Replication

Please refer to EXPERIMENT.md for information on experiment replication.

✉️ Contacts

For any queries, you are welcome to open an issue.

Alternatively, contact us at: {yuntong,hruan,zhiyufan}@comp.nus.edu.sg.

Acknowledgements

This work was partially supported by a Singapore Ministry of Education (MoE) Tier 3 grant "Automated Program Repair", MOE-MOET32021-0001.