# AGENTS.md Rules for AI coding assistants working in this repository. - **dora >= 1.0.** Verified against `dora-cli 1.0.1`, which CI pins in `.gitea/workflows/validate.yml`. - **`uv` for everything:** `uv sync --group dev`, `uv run pytest`, `uv run ruff check .`, `uv run python ...`. Never `pip` into the ambient environment. - **`ruff` is the linter and the formatter.** `uv run ruff check .` must be clean. - **Markdown is linted too**, by `markdownlint-cli2` in CI. Its line length is set in `.markdownlint.json`. - **numpy-style docstrings.** - **Dependencies must ship prebuilt wheels for linux-x86_64, linux-aarch64 and macos-arm64.** `uv sync` is the whole install: no system packages, no compiler. - **Nothing outside `dora_joy/gamepad.py` may touch pygame.** Everything else sees a plain `Sample`, which is what lets it be tested with no pad plugged in. - **Nothing is shaped on the way out.** No deadzone, no expo, no clamping, no rescaling of a trigger. Raw SDL or nothing: what a reading *means* belongs to whatever acts on it, and two places shaping one number is a bug nobody can see. - **Do not teach this node what a control is for.** No `GRASP_BUTTON` variable, no semantic output ports. This is a source; it names no consumer. - **Publish on every tick, changed or not.** The stream doubles as a heartbeat, so silence means the node died and nothing else. Do not add publish-on-change. - **`present` is never inferred from zeroed axes.** An absent pad is not a centred pad, and a consumer driving something that moves needs the difference. - **The Arrow schema is declared, never inferred.** pyarrow reads an empty `axes` list as `list`, so an inferred schema would change type the moment a pad is unplugged. - **Do not commit or push.** Leave changes in the working tree and wait to be asked. - **Do not mention AI tools in commit messages or documentation.**