mentat-assistant

Command line tool to programmatically call Mentat, the AI coding assistant and let it run a custom series of prompts and commands by using the Mentat Python SDK

Usage (without docker):

`python3 assistant.py [-v] [--promptsdir /path/to/prompts] command [ARGS]`

The assistant will look up prompts/command.xml and run it. It'll use any other args for substituting place holders in the defined prompt and ask for any missing arguments. Additionally you can fill variables with a converter from other arguments. Currently there is one converter to resolve fully qualified PHP class names to file names so you can use the result either in your prompt or to automatically add files into the context. You can define your own custom prompts you want Mentat to run by adding them to the prompts/ folder. Look at prompts/generate-unit-tests.xml for an example of the XML structure.

Setup without docker

1. Clone this repository
2. Add a file .env with your OpenAI API key (OPENAI_API_KEY=sk-... - see .env.example)
3. Install the necessary python packages via pip install -r requirements.txt

Setup using docker

Alternatively you can run the assistant with the docker-based wrapper provided in bin/assistant:

1. Clone this repository
2. Add a file .env with your OpenAI API key (OPENAI_API_KEY=sk-... - see .env.example)
3. Make ./bin/assistant executable (on Linux: chmod +x ./bin/assistant)

Usage (same command line options as described previously):

`./bin/assistant [-v] [--promptsdir /path/to/prompts] command [ARGS]`

To simplify usage so that you don't have to call it with the full path /path/to/mentat/bin/assistant every time:

• Create a symlink from bin/assistant into your $PATH so that you can easily run the wrapper script from anywhere on your computer, e.g. ln -s /path/to/mentat-assistant/bin/assistant ~/.local/bin/assistant

Now you can run assistant from anywhere and it always operates on your current directory.

Configuration

Both setup variants will use your Mentat configuration. E.g. you can create a ~/.mentat/.mentat_config.json with

{
    "model": "gpt-4-1106-preview"
}

and Mentat will always use this model. Alternatively add a per-project .mentat-config.json into the directory you want to work on.

Contributing

This code is licensed unter the MIT License - you're warmly invited to contribute!

Code quality

To make sure to have some good level of code quality please run

• black . to auto-format the code with the black auto-formatter
• python3 -m unittest discover -s test/ to run all tests
• flake8 . to check for any linting issues
• mypy . for static type checking

However currently there are not that many type annotations and mypy --strict . would report a number of issues. Currently I don't intend to fix that as that would make the code harder to read and overly complicated for a non-mission-critical script. Instead I have plans to implement XML schema validation to ensure the XML command definition files are well-formatted with all required attributes.

With GitHub Actions these four commands are run on any push or pull request in the repository, see the Github workflow

Explanations on the Docker wrapper

Dockerfile: The container is based on Debian and is also available on Docker Hub: holybiber/mentat-assistant.

docker-compose.yml: Docker compose setup, makes mentat configuration available to the container

bin/assistant: Wrapper for docker compose so that we can run the assistant easily

See also

https://github.com/holybiber/mentat