Agent Skills: Python Environment

Python Environment

Always use uv. Never use bare python/pip, never search for a Python interpreter, never create a virtual environment by hand. uv run resolves the environment from the nearest pyproject.toml up the directory tree.

If pyproject.toml is not present in the project root, the environment has not been bootstrapped — run /bayesian-workflow:setup.

After setup, run every script from the project root:

uv run python experiments/experiment_X/fit/posterior_fit.py --model ... --data ... --out ...

The environment provides arviz, cmdstanpy, numpy, pandas, pyarrow, matplotlib, scipy, and seaborn. scikit-learn, statsmodels, and deep-learning frameworks are intentionally excluded — the workflow is Stan-based (see orchestration > Technical stack). Add a dependency to pyproject.toml only if a model genuinely requires it.

Stan execution

There is no shared library to import. The three canonical recipes — posterior inference, prior predictive simulation, fake-data simulation with a recovery check — are runnable reference scripts in the fit-pipeline skill (references/posterior_fit.py, prior_predictive.py, fake_data.py). Copy the one you need into the stage directory, adapt the marked section (for the posterior fit, build_stan_data()), run it with uv run, and keep it. The artifact contract those scripts honor is documented in the same skill; do not change file names or JSON layouts per experiment.

Script structure

Write small, focused scripts — not monolithic files. Separate concerns:

experiments/experiment_X/
  fit/
    posterior_fit.py  # adapted copy of the fit-pipeline reference (entry point)
    plots.py          # diagnostic visualisation
  model.stan

Each script should:

• Do one thing well
• Be runnable via uv run python script.py from the project root
• Start from the fit-pipeline reference whenever it samples, converts, or writes fit artifacts, instead of re-deriving those steps
• Use paths relative to the project root and write outputs into the canonical folder structure

Long runs

For fits likely to exceed a few minutes, launch detached and poll the log rather than blocking one foreground command:

nohup uv run python experiments/experiment_X/fit/posterior_fit.py ... > experiments/experiment_X/fit/run.log 2>&1 &

Progress output stays off inside the reference scripts (show_progress=False, show_console=False); never turn it on — it floods the transcript.