git-bob

<!---->

git-bob uses AI to solve GitHub issues. It runs inside the GitHub CI, no need to install anything on your computer. Read more in the publication.

Under the hood it uses Anthropic's Claude or OpenAI's chatGPT or Google's Gemini to understand the text and pygithub to interact with the issues and pull requests. As its discussions are conserved, you can document how things were done using AI and others can learn how to prompt for the things you did. For example, the pair-plot discussion above is available online.

Disclaimer

git-bob is a research project aiming at streamlining GitHub interaction in software development projects. Under the hood it uses artificial intelligence / large language models to generate text and code fulfilling the user's requests. Users are responsible to verify the generated code according to good scientific practice.

When using git-bob you configure it to use an API key to access the AI models. You have to pay for the usage and must be careful in using the software. Do not use this technology if you are not aware of the costs and consequences.

[!CAUTION] When using the Anthropic, OpenAI, Google Gemini, Mistral or any other endpoint via git-bob, you are bound to the terms of service of the respective companies or organizations. The GitHub issues, pull requests and messages you enter are transferred to their servers and may be processed and stored there. Make sure to not submit any sensitive, confidential or personal data. Also using these services may cost money.

Installation as GitHub action

There is a detailed tutorial on how to install git-bob as GitHub action to your repository. In very short, to use git-bob in your GitHub repository, you need to * Copy the git-bob GitHub workflow in folder .github/workflows/ to your repository. * Make sure to replace pip install -e . with a specific git-bob version such as pip install git-bob==0.16.0. * If your project does not contain a requirements.txt file, remove the line pip install -r requirements.txt. * Configure the LLM you want to use in the workflow files by specifying the GIT_BOB_LLM_NAME GitHub repository secret. These were tested: * anthropic:claude-3-5-sonnet-20241022 * openai:gpt-4o-2024-08-06 * github_models:gpt-4o * github_models:meta-llama-3.1-405b-instruct * google:gemini-1.5-pro-002 * mistral:mistral-large-2411 (uses pixtral-12b-2409 for vision tasks) * deepseek:deepseek-chat * e-infra_cz:llama3.3:latest * configure a GitHub secret with the corresponding key from the LLM provider depending on the above configured LLM: * OPENAI_API_KEY: OpenAI (gpt) * ANTHROPIC_API_KEY: Anthropic (claude) * GH_MODELS_API_KEY: GitHub Models Marketplace * GOOGLE_API_KEY: Google AI * MISTRAL_API_KEY: Mistral * DEEPSEEK_API_KEY: DeepSeek * KISSKI_API_KEY: KISSKI * BLABLADOR_API_KEY: BLABLADOR * E_INFRA_CZ_API_KEY chat.ai.e-infra.cz docs * configure GitHub actions to run the workflow on issues and pull requests. Also give write-access to the Workflow using the GITHUB_TOKEN.

When using it in your repository, you can also set a custom system message, for example for: * General Data Science / Python Programming * Bio-Image Analysis * Giving advice on a specific repository / library * Manuscript writing

Furthermore, to guide discussions, you may want to setup issue templates, e.g. * General Python Programming Questions * Bio-Image Analysis * Statistics and Plotting

Installation as gitlab pipeline

Since version 0.10.1 git-bob has experimental support for gitlab. You find detailed instructions how to install it here.

Usage: Trigger words

To trigger git-bob, you need to comment on an issue or pull request with the comment trigger word (or aliases think about, review, respond):

git-bob comment

If you want to ask git-bob for a review of a pull-request, you can use the review trigger word. Also make sure mention explictly what you want to be reviewed.

git-bob review this PR. Check code quality and comments.

After some back-and-forth discussion, you can also use the solve trigger word (or aliases implement, apply) make git-bob solve an issue and send a pull-request. This trigger can also be used to modify code in pull requests.

git-bob solve

You can ask git-bob to implement a solution for testing, without sending a pull-request, using the try trigger: git-bob try

If you have multiple API-Key for different LLMs configured, you can specify the LLM in the command using the ask <LLM-Name> to trigger command:

git-bob ask claude-3-5-sonnet-20241022 to solve this issue.

If the issue is complex and should be split into sub-issues, you can use the following command:

git-bob split

If you have two GitHub secrets TWINE_USERNAME and TWINE_PASSWORD configured, you can also use the following command to publish a new version of your library to PyPI:

git-bob deploy

All trigger words can be combined with please and/or ,, which will make no difference to calling git-bob without these words:

git-bob, please ask gemini-1.5-pro-002 to solve this issue.

Recommended Workflow

Here's the recommended workflow for using git-bob:

1. Create an issue describing the problem or task.
2. Comment on the issue with git-bob comment, or git-bob think about this (an alias for comment) to trigger git-bob making a plan.
3. Respond to git-bob with any clarifications or additional information it requests.
4. Comment on the issue with git-bob solve or git-bob implement this (an alias for solve) to trigger git-bob.
5. Wait for git-bob to create a pull request (PR) addressing the issue.
6. Review the PR and comment on the PR or on the original issue if changes are needed.
7. Wait for git-bob to create new PR or modifying the existing PR with the requested changes.
8. Repeat steps 3-5 as necessary until the issue is resolved satisfactorily.

Supported file formats

Git-bob can interact with a variety of file formats. * Jupyter Notebooks (.ipynb), can also be executed * Text and code files (.txt, .tex, .py, .csv, ...) * Markdown files (.md) * Markup files (.yml, .yaml, .xml) * Word Document files (.docx) * PowerPoint files (.pptx, write-only) * Audio files (.mp3, write-only, OPENAIAPIKEY required) * SVG files (.svg) * images files (.png, .jpg, ..., for reading images, a vision-capable LLM must be chosen, for creating images DALL-E is used, OPENAIAPIKEY required)

Use-case examples

A huge variety of use-cases for git-bob are thinkable. Here are some examples. Many serve purely demonstrative purposes. Some were parts of real scientific data analysis projects.

• Question answering
  • About specific Python libraries
  • The capital of France
  • Asking for git-bob's capabilities
• Translation
  • Translating text in Jupyter notebooks in other Languages
  • Translating PowerPoint files
• Image Analysis
  • Nuclei segmentation
  • Cell segmentation
  • Resizing images
  • Natural image classification
• Programming
  • Caching
  • Fixing bugs in notebooks
  • Summarize code of a file
  • Summarize code in a repository
  • Programming Google Search
  • Implementing a generic Factory patters
  • Generating a QR-Code
  • Caching LLM prompt responses
  • Interactive bash inside a program
• Prompting
  • Prompting for SVG files
• Continuous Integration and Deployment
  • Configuring Github workflows
  • Configuring a Jupyter Book
• Data & Code Management
  • Write a Data Management Plan (DMP)
  • Research Data Management & Folder Structures
  • Documenting source code
  • Determining licenses of dependencies
  • Assisting scientific manuscript writing
  • Deleting files
  • Converting tables to key-value pairs
  • Exporting Google Scholar profile as bibtex
  • Deciding for file formats: JSON versus YAML
  • Generating Galaxy workflows
  • Count citations of given DOIs
  • Convert PDF documents to PNG images
  • Convert PDF documents to animated GIFs
  • Split PDFs
  • Convert SVG files to PNG images
  • Querying the arxiv
  • Retrieving meta-data of arxiv articles
  • Counting PowerPoint slides in Zenodo records
  • Search for duplicate entries in a yml file
  • Running multiple notebooks and trace failures
• Graphical User Interfaces
  • Interactive drawing on an ipcanvas
• Statistics
  • Filtering data and counting records
  • Selecting ranges of columns
  • Summarizing data and plotting
  • Writing text about well-known statistical methods
  • Drawing a decision tree for statistial tests
  • Converting Dice and Jaccard indices
  • Generating personal profile data of random scientists
• Plotting
  • Violing plots with simulated data
  • UMAPs with simulated data
  • Plotting a circle of triangles (in gitlab)
  • Pairplots with simulated data
  • Word clouds
  • Plot simulated income data
  • Ploting the Dunning Kruger effect
  • Reproducing a plot
  • Animated zooming into an image
• Science Communication
  • Making slides for Deep Learning training
  • Making slides for Research Data Management training
  • Visualizing how Fourier-Transform works
  • Making Jupyter Notebooks for training
  • Making slides about text
  • Visualizing processes using CPUs and GPUs
• Fun
  • Playing Tic-Tac-Toe
  • Story Telling for kids
  • PowerPoint Karaoke
  • Solving Advent of Code 2024 puzzles
  • Drawing a Christmas tree as SVG
  • Asking for the Meaning of Life
  • Creating audio files (also in German and French)
• git-bob refusing to help
  • Changing the working directory of the parent shell
  • Solving the Halting Problem [2nd attempt]
• Things that didn't work well
  • How to use aider from python
  • How to use the atproto API
  • Executing code in a sandbox
  • Summarize code in a repository
  • Drawing multiple trees in a SVG file
  • Drawing relationships between agents in a multi-agent system
  • Complex code refactoring
  • Debugging OpenCL error
  • Signing PDFs with visual signature AND cryptographic signing
  • Extending cell types in Jupyter notebooks

Installation for development

git clone https://github.com/haesleinhuepf/git-bob.git cd git-bob

Usage as command-line tool (for development)

You can also install git-bob locally and run it from the terminal. In this case, create a GitHub token and store it in an environment variable named GITHUB_API_KEY. Also create an environment variable GIT_BOB_LLM_NAME with the name of the LLM you want to use, e.g. "gpt-4o-2024-05-13" or "claude-3-5-sonnet-20241022" or "github_models:gpt-4o". Then you can install git-bob using pip:

bash pip install git-bob

You can then use git-bob from the terminal on repositories you have read/write access to. It is recommended to call it from the root folder of the repository you want to interact with.

bash git clone https://github.com/<organization>/<repository> cd <repository> git-bob <action> <organization>/<repository> <issue-number>

Available actions: * review-pull-request * comment-on-issue * solve-issue * split-issue

Limitations

git-bob is a research project and has limitations. It serves as basis for discussion and further development. Once LLMs become better, git-bob will become better as well.

At the moment, these limitations can be observed: * git-bob was tested for Python projects mostly. It seems to be able to process Java and C++ as well. * It can only execute code in Jupyter Notebooks. * It sometimes hallucinates, especially in code reviews. E.g. it claimed to have tested code, which was certainly not true. * It cannot solve issues where changing long files is required, as the output of the LLMs is limited by a maximum number of tokens (e.g. 16k for gpt-4o-2024-08-06). When using OpenAI's models it combines output of multiple requests to a maximum file length about 64k tokens. It may then miss some spaces or a line break where responses were stitched. When using GitHub models, the maximum file length is 4k tokens. When using Anthropic's Claude, the maximum file length is 8k tokens. * When changing multiple files, it may introduce conflicts between the files, as it does not know about the changed contents of the other files. * It has only limited logic to control who is allowed to trigger it. If you are a repository member, you can trigger it. If others send a pull request, a repository member must allow the action to run manually. * git-bob is incompatible with locally running open-source/-weight LLMs. This might make sense when being executed locally only. In the GitHub-CI this might be impossible. * Recently tested claude-3-5-sonnet-20241022, gpt-4o-2024-08-06, github_models:gpt-4o, github_models:meta-llama-3.1-405b-instruct and gemini-1.5-pro-002 produced useful results. * git-bob is not allowed to modify workflow files, because it also uses GitHub workflows. * As git-bob is installed as part of git-hub workflows, its download statistics might be misleading. There are not as many people downloading it as the numer of downloads suggest.

Extensibility

git-bob can be extended in multiple ways. All you need to do is to set up small python library which implements specific functions and exposes them using Pythons plugin system. You find an example implementation of the extensions described below in this respository.

Adding new trigger words

If you want to add new trigger words and corresponding python functions, you can do so by implementing a new trigger handler function with a predefined signature in a small python library. The function can have the arguments repository, issue, prompt_function and base_branch and if you do not need all of them, just leave them out and add **kwargs at the end of the argument list. E.g. if you want to add a new trigger word love, you can implement a new function like this.

python def love_github_issue(repository, issue, **kwargs): from git_bob._utilities import Config Config.git_utilities.add_comment(repository, issue, "I love this issue! <3")

Additionally, you need to configure your plugin's entry point in its setup.cfg file:

git_bob.triggers = love = my_library.my_python_file:love_github_issue

Adding new LLMs / prompt handlers

If you use institutional LLM-servers which are accessible from the internet (or from your gitlab-server), you can use them using git-bob by implementing a new prompt handler function with a predefined signature. E.g. if your LLM-server is openai-compatible, you can reuse the prompt_openai function, adjust parameters such as max_response_tokens, and the url of your LLM-server like this:

```python def promptmycustomllm(message: str, model=None, image=None): import os from gitbob.endpoints import promptopenai

model = model.replace("mycustomllm:", "") return promptopenai(message, model=model, image=image, baseurl="https://myserver/v1", apikey=os.environ.get("MYCUSTOMAPIKEY"), maxresponse_tokens=8192) ```

Additionally, you need to configure your plugin's entry point in its setup.cfg file:

git_bob.prompt_handlers = my_custom_llm = my_library.my_python_file:prompt_my_custom_llm

git-bob will then detect your plugin and can use it if the GIT_BOB_LLM_NAME secret is set to any model containing my_custom_llm. You could for example configure a llama model running on your LLM-server like this: my_custom_llm:llama3.3-70b.

Filtering extensions

If you wish to extend git-bob with custom triggers or prompt handlers, but avoid default triggers and prompt handlers, you can configure a filter in the git-bob.yml workflow file. Just overwrite this default regular expression accepting all extensions:

GIT_BOB_EXTENSIONS_FILTER_REGEXP: ".*"

If you want to only accept extensions starting with my_library, you can configure the filter like this:

GIT_BOB_EXTENSIONS_FILTER_REGEXP: "^my_library.*"

If you want to accept all extensions but not git-bob`s defaults, you can configure the filter like this:

GIT_BOB_EXTENSIONS_FILTER_REGEXP: "^(?!git_bob).*"

Similar projects

There are similar projects out there * Claide github action * Claude Engineer * BioChatter * aider * OpenDevin * Devika * GPT-Codemaster * GitHub Copilot Workspace * agentless * git-aid * SWE-agent * gh-gitgen

Contributing

Feedback and contributions are welcome! Just open an issue and let's discuss before you send a pull request. A human will respond and comment on your ideas!

Citation

If you use git-bob, please cite it: @misc{haase_2024_13928832, author = {Haase, Robert}, title = {{Towards Transparency and Knowledge Exchange in AI- assisted Data Analysis Code Generation}}, month = mar, year = 2025, publisher = {Nature Computational Science}, doi = {10.1038/s43588-025-00781-1}, url = {https://doi.org/10.1038/s43588-025-00781-1} }

Acknowledgements

We acknowledge the financial support by the Federal Ministry of Education and Research of Germany and by Sächsische Staatsministerium für Wissenschaft, Kultur und Tourismus in the programme Center of Excellence for AI-research „Center for Scalable Data Analytics and Artificial Intelligence Dresden/Leipzig", project identification number: ScaDS.AI