Contributing

Thanks for helping. The most useful contribution is a wrong answer.

#Report a wrong answer

lxask "what you asked"
lxask --bad "what was wrong, or the right command"
lxask --share            # prints the report and a link to a prefilled issue

Open the link, remove anything private, and submit. Wrong answers become regression cases (evalsets/regressions.jsonl) that every later model must pass. You can also open a "Bad answer" issue by hand.

#Report a bug or suggest something

Open a "Something else" issue: what you ran, what happened, what you expected, and lxask --version. For security problems, see Security.

#Change the code

• Setup, building and tests: docs/development.md.
• The rules (checks, repairs, routing) are in rust/crates/lxask-core, and the Python training and evaluation scripts call the same code through lxask._core. Rust tests: cargo test --manifest-path rust/Cargo.toml -p lxask-core; Python tests: pytest (no model needed).
• A change to a rule should not change answers it is not about. Run the evaluation before and after (scripts/eval.py, see docs/development.md) and say in the pull request what moved.
• Keep commits small and their messages about why.

#Training data

The training data written with OpenAI models is not in this repository (see NOTICE and docs/development.md). scripts/generate_openai.py makes it again with your own API key; read OpenAI's terms on using model output for training before you do.

#Licence

By contributing you agree that your contribution is under the MIT License of this repository (LICENSE).