A supplement to Copilot and Cursor – utilizes AI for batch processing of project codes

GITHUB: https://github.com/qiangyt/batchai

I often rely on ChatGPT and GitHub Copilot, but it is a little bit frustrating to constantly copy and paste between the copilot chat window and my open code files. Why not update the files directly? I also tried using Cursor, which solved the first problem, but still have to open each file individually to add them to the AI’s context.

That’s why I created batchai. The idea is simple: less copy-pasting, fewer clicks on ‘Add to Chat’ or ‘Apply.’ batchai traverses files and processing each of them. Since AI isn’t always perfect, I’ve designed it to run only on a Git directory, so we can easily diff the changes and choose to either commit or revert them.

Currently, batchai only supports code review and fixing common issues (think of it as a local AI-driven SonarQube). The next feature in progress is generating unit test code in batches, which I plan to use in a few of my personal projects (including this batchai), as they have very few unit tests. Other planned features include code explanation, comment generation, and refactoring — all of which will be handled in batches. Additionally, I’m working on enabling batchai to have an overall insight of the project’s code, such as building cross-file code symbol indexing, which should help the AI perform better.

Here are some interesting findings from testing batchai on my personal projects over the past two weeks:

It can identify issues that traditional tools, such as SonarQube, tend to miss.
It may not report all issues in one go, so I need to run it multiple times.
Due to outdated LLM training data and hallucinations, it’s crucial to confirm the changes for accuracy by myself – That’s why I make batchai work only on clean Git repository directories.

I used the spring-petclinic (cloned from https://github.com/spring-projects/spring-petclinic) for demonstration.

Here are some examples of correct review:

Adds a check to ensure birthday not be in the future

Renamed method to adhere to JavaBeans naming conventions

And also a wrong fix:

Downgraded MySQL version from 9.0 back to 8.0 (gpt4o-mini think latest MySQL version is 8.0)

More detail:

Code Review report

Fix following the review report

Features

[x] Code Review : Reports issues to the console, saves as a review report, and then fixes code directly.
[x] Customized Prompts.
[x] File Ignoring : Specifies files to ignore, respecting both .gitignore and an additional .batchai_ignore file.
[x] Target Specification : Allows specifying target directories and files within the Git repository.
[x] Implemented using Go: Resulting in a single executable binary that works on Mac OSX, Linux, and Windows.
[x] Diff: Displays colorized diffs in the console.
[x] LLM Support : Supports OpenAI-compatible LLMs, including Ollama.
[x] I18N : Supports internationalization comment/explaination generation.

Planned features

Explain, Comment Generation, Test Generation, Refactoring.
Rejected Changes Tracking : Tracks rejected changes to avoid redundant modifications.
Language-Specific Prompts : Different prompts for various programming languages.
LLM Usage Metrics : Implements metrics for tracking LLM usage.

Getting Started

Download the latest executable binary from here and add it to your $PATH. For Linux and Mac OSX, remember to run chmod +x … to make the binary executable.

Clone the demo project. The following steps assume the cloned project directory is /data/spring-petclinic

In this directory, create a .env file. In the .env file, set the OPENAI_API_KEY. Below is an example:

For Ollama, you can refer to my example docker-compose.yml

CLI Examples:

Report issues to the console (also saved to build/batchai):

Directly fix the target files via option –fix:

Run batchai in main Java code only:

Run batchai on the entire project:

CLI Usage

To view the global help menu and available commands, run:

To see detailed help for the review command, run:

Supported LLMs

Tested and supported models:

OpenAI series:

openai/gpt-4o
openai/gpt-4o-mini

Other OpenAI models should work too.

Ali TONYI Qwen series:

qwen2.5-coder-7b-instruct (also available via Ollama)

Other Qwen models should work too.

To add more LLMs, simply follow the configuration in res/static/batchai.yaml, as long as the LLM exposes an OpenAI-compatible API.

Configuration

Optional configuration file:

You can provide an optional configuration file: ${HOME}/batchai/batchai.yaml. For a full example, refer to res/static/batchai.yaml

Environment file:

You can also configure batchai via an environment file .env located in the target Git repository directory. Refer to res/static/batchai.yaml for all available environment variables, and res/static/batchai.env for their default values.

Ignore specific files:

batchai ignores the directories and files following .gitignore files. This is usually sufficient, but if there are additional files or directories that cannot be ignored by Git but should not be processed by batchai, we can specify them in the .batchai_ignore files. The rules are written in the same way as in .gitignore.

Customized Prompts
Refer to BATCHAI_REVIEW_RULE_* and MY_REVIEW_RULE_* in res/static/batchai.yaml

License

MIT

Please follow and like us: