Development and integration¶

Checkout setup¶

Install uv and Git, then run:

git clone https://github.com/stroblme/wren.git
cd wren
uv sync --locked
uv run wren "turn on the kitchen light"

The checkout pins Python 3.14 in .python-version; the package supports Python 3.12 and newer. uv can download the pinned interpreter if needed. The default dev group includes the training extra, Fluksio, pytest, Ruff, ty and study dependencies. No GPU is required for installation training.

For an inference-only checkout, use uv sync --locked --no-dev and run commands with uv run --no-dev wren .... For CPU training, the default setup is sufficient. Python fit uses JAX's available backend; set JAX_PLATFORMS=cpu before starting Python or the engine to force CPU use.

Tracked training¶

From the checkout, start the local engine in its own terminal:

uv run fluksio serve

Keep it running, then use a second terminal in the same directory:

uv run wren --train home.json

Follow the How-to to create home.json. The command trains through wren.flow, records parameters, per-step loss, evaluation scores and the pack, and downloads home.bin. Local engine state is stored in the ignored .fluksio/ directory.

To retrain a copy of the demo installation without replacing packaged data:

cp wren/home.json home.json
uv run wren --train home.json
uv run wren --targets home.json --model home.bin "turn on the kitchen light"

home.json and home.bin are your local outputs; keep them out of a source change unless you intend to publish a model.

GPU training¶

For a compatible NVIDIA GPU and driver, select the cuda12 or cuda13 dependency group. For example:

uv sync --locked --group cuda12
uv run --group cuda12 fluksio serve --gpus 1

Then run the tracked training command in a second terminal. Include the selected group in later uv sync commands, which otherwise remove those dependencies.

Checks and documentation¶

Run the same checks as CI:

uv run ruff check .
uv run ruff format --check .
uv run ty check
uv run pytest -q
uv build

The runtime tests compile runtime/wren.c, so a C compiler must be on PATH. The tests include small training checks but do not retrain the shipped demo model.

Preview and build the documentation from the repository root:

uvx zensical==0.0.72 serve
uvx zensical==0.0.72 build --clean --strict

The site is configured in zensical.toml, built into site/ and published by .github/workflows/docs.yml. Edit the Markdown in docs/; edit README.md for the repository overview. Ignored working notes are excluded from the site. To update the architecture drawing, edit docs/architecture.d2 and render it with d2 docs/architecture.d2 docs/architecture.svg.

Wren in Fluksio¶

Fluksio's wren automation node runs Wren on typed or spoken commands. Its wired devices define the targets; your installation's model interprets commands and Fluksio applies calls or sends unhandled work to its assistant. Register this checkout's training flow on the engine used by the node:

uv run fluksio sync wren.flow

The node's Train button submits its target definitions, examples and tests to that flow. See the Fluksio node reference for node settings and engine setup.

Wren 2.0's library API uses list[Target] and returns list[Call], with confidence on each call. Code using the old House or Device API must migrate; the demo home is now JSON data rather than a built-in device model.