ArXiv Paper Β Β Β Β Β Β Website Β Β Β Β Β Β Discord server
Note
This is a public version of the AutoCodeRover project. Check the latest results on our website.
- [July 18, 2024] Added support for fix localization output mode and requirements.txt
- [June 20, 2024] AutoCodeRover 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.
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 now resolves 30.67% of issues (pass@1) in SWE-bench lite! AutoCodeRover achieved this efficacy while being economical - each task costs less than $0.7 and is completed within 7 mins!
[April 08, 2024] First release of AutoCodeRover resolves 19% of issues in SWE-bench lite (pass@1), improving over the current state-of-the-art efficacy of AI software engineers.
AutoCodeRover works in two stages:
- π Context retrieval: The LLM is provided with code search APIs to navigate the codebase and collect relevant context.
- π Patch generation: The LLM tries to write a patch, based on retrieved context.
AutoCodeRover has two unique features:
- Code search APIs are Program Structure Aware. Instead of searching over files by plain string matching, AutoCodeRover searches for relevant code context (methods/classes) in the abstract syntax tree.
- When a test suite is available, AutoCodeRover can take advantage of test cases to achieve an even higher repair rate, by performing statistical fault localization.
AutoCodeRover: Autonomous Program Improvement [arXiv 2404.05427]
For referring to our work, please cite and mention:
@misc{zhang2024autocoderover,
title={AutoCodeRover: Autonomous Program Improvement},
author={Yuntong Zhang and Haifeng Ruan and Zhiyu Fan and Abhik Roychoudhury},
year={2024},
eprint={2404.05427},
archivePrefix={arXiv},
primaryClass={cs.SE}
}
As an example, AutoCodeRover successfully fixed issue #32347 of Django. See the demo video for the full process:
acr-final.mp4
AutoCodeRover can resolve even more issues, if test cases are available. See an example in the video:
acr_enhancement-final.mp4
We recommend running AutoCodeRover in a Docker container.
Set the OPENAI_KEY
env var to your OpenAI key:
export OPENAI_KEY=sk-YOUR-OPENAI-API-KEY-HERE
For Anthropic model, Set the ANTHROPIC_API_KEY
env var can be found here
export ANTHROPIC_API_KEY=sk-ant-api...
The same with GROQ_API_KEY
Build and start the docker image:
docker build -f Dockerfile -t acr .
docker run -it -e OPENAI_KEY="${OPENAI_KEY:-OPENAI_API_KEY}" -p 3000:3000 -p 5000:5000 acr
Alternatively, you can use Dockerfile.scratch
which supports arm64 (Apple silicon) and ppc in addition to amd64.
Dockerfile.scratch
will build both SWE-bench (from https://github.com/yuntongzhang/SWE-bench.git) and ACR.
docker build -f Dockerfile.scratch -t acr .
There are build args for customizing the build in Dockerfile.scratch
like this:
docker build --build-arg [email protected] --build-arg GIT_NAME=your_id \
--build-arg SWE_BENCH_REPO=https://github.com/your_id/SWE-bench.git \
-f Dockerfile.scratch -t acr .
After setting up, we can run ACR in three modes:
- GitHub issue mode: Run ACR on a live GitHub issue by providing a link to the issue page.
- Local issue mode: Run ACR on a local repository and a file containing the issue description.
- SWE-bench mode: Run ACR on SWE-bench task instances.
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 printed in the end.
Web UI is also provided for visualization of the issue fixing process. In the docker shell, run the following command:
cd /opt/auto-code-rover/demo_vis/
bash run.sh
then open the url localhost:3000
in the web explorer.
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 printed in the end.
This mode is for running ACR on existing issue tasks contained in SWE-bench.
In the docker container, we need to first set up the tasks to run in SWE-bench (e.g., django__django-11133
). The list of all tasks can be found in conf/swe_lite_tasks.txt
.
The tasks need to be put in a file, one per line:
cd /opt/SWE-bench
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:
cd /opt/SWE-bench
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 their ids in tasks.txt
and follow the same steps.
Before running the task (django__django-11133
here), make sure it has been set up as mentioned above.
cd /opt/auto-code-rover
conda activate auto-code-rover
PYTHONPATH=. python app/main.py swe-bench --model gpt-4o-2024-05-13 --setup-map ../SWE-bench/setup_result/setup_map.json --tasks-map ../SWE-bench/setup_result/tasks_map.json --output-dir output --task django__django-11133
The output of the run can then be found in output/
. For example, the patch generated for django__django-11133
can be found at a location like this: output/applicable_patch/django__django-11133_yyyy-MM-dd_HH-mm-ss/extracted_patch_1.diff
(the date-time field in the directory name will be different depending on when the experiment was run).
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 /opt/auto-code-rover
conda activate auto-code-rover
PYTHONPATH=. python app/main.py swe-bench --model gpt-4o-2024-05-13 --setup-map ../SWE-bench/setup_result/setup_map.json --tasks-map ../SWE-bench/setup_result/tasks_map.json --output-dir output --task-list-file /opt/SWE-bench/tasks.txt
NOTE: make sure that the tasks in tasks.txt
have all been set up in SWE-bench. See the steps above.
Alternatively, a config file can be used to specify all parameters and tasks to run. See conf/vanilla-lite.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/vanilla-lite.conf
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 | 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 | |
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:
- Before using the llama3 models, please install ollama and download the corresponding models with ollama (e.g.
ollama pull llama3
). - 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.
- 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 thedocker run
command when starting the ACR container, so that ACR can communicate with the ollama server on the host machine.
Please refer to EXPERIMENT.md for information on experiment replication.
For any queries, you are welcome to open an issue.
Alternatively, contact us at: {yuntong,hruan,zhiyufan}@comp.nus.edu.sg.
This work was partially supported by a Singapore Ministry of Education (MoE) Tier 3 grant "Automated Program Repair", MOE-MOET32021-0001.