What this template provides¶

Python Package Skeleton¶

• The src directory contains the code that will be packaged and deployed on the user system. That code is in a directory with the same name as the package.
• Using a separate src directory helps avoid accidentally importing the package when you're working in the top level directory of the repository.
• A simple python module (dummy.py), and a separate module providing a command line interface to that module (cli.py) are included as examples.
• Any files in the src/package_data/ directory will also be packaged and deployed.
• What files are included in or excluded from the package on the user's system is controlled by the [tool.hatch.build.targets.wheel] options in pyproject.toml. We build with hatchling, which packages every file under the specified packages directory (both .py and non-Python files) -- there's no separate MANIFEST.in to maintain.
• The CLI is deployed using project.scripts defined in pyproject.toml.
• We use hatch-vcs (configured under [tool.hatch.version]) to obtain the package's version directly from git tags, rather than storing it in the repository and manually updating it.
• README.md is read in and used for the package's long_description. This is what is displayed on the PyPI page for the package.
• By default we create several sets of "extras" -- additional optional package dependencies that can be installed in special circumstances: dev, docs, lint, tests, and types. The packages listed there are used in development, building the docs, linting, running the tests, and type checking (respectively) but aren't required for a normal user who is just installing the package from pip or conda. These are defined under the project.optional-dependencies section of pyproject.toml.

Environment & Task Management with Pixi¶

• We use pixi to manage the development environment and the tasks used to test, lint, format, and document the project.
• Run pixi install once to create the environment described in pyproject.toml (under [tool.pixi.*]), then use pixi run <task> to run any of the tasks defined under [tool.pixi.tasks].
• The most important tasks are:
  • pixi run test -- run all the unit and integration tests under tests/ with pytest and report test coverage.
  • pixi run lint -- run ruff and ty to catch errors and style issues.
  • pixi run format -- automatically reformat the code and other files using ruff, taplo, mdformat, and prettier.
  • pixi run docs -- build the documentation with zensical.
• There's a single default pixi environment that contains everything needed for local development, since splitting a small template repository into many environments adds more complexity than it saves.
• This package is installed into that environment in editable mode via [tool.pixi.pypi-dependencies], which explicitly lists every extra from project.optional-dependencies (dev, docs, lint, tests, types) that should be pulled in, rather than relying on pixi's implicit behavior of matching same-named pixi features to extras. The one named pixi feature that still exists, lint, exists only to add a few non-Python formatting tools (taplo, prettier, mdformat, etc.) that aren't published to PyPI and so can't be expressed as an extra.

Devcontainer¶

• .devcontainer/devcontainer.json defines a basic, editor-agnostic development container: the same ghcr.io/prefix-dev/pixi image used in docker/Dockerfile, with git added (the base image doesn't include it, and pixi install needs it to derive the package version from git tags) and pixi install / pixi run prek install run automatically once the container starts. It works with VS Code, JetBrains Gateway, GitHub Codespaces, or the standalone devcontainer CLI -- useful for giving a coding agent (or a human) an isolated, reproducible, disposable sandbox to work in instead of your host machine.

Pytest Testing Framework¶

• A skeleton pytest testing setup is included in the tests/ directory.
• Tests are split into unit and integration categories.
• Session-wide test fixtures, additional command line options, and other pytest configuration can be added to tests/conftest.py.
• Exactly what pytest commands are run during continuous integration is controlled by the pixi tasks defined in pyproject.toml.
• Pytest can also be run manually without going through the pixi task, but still in the pixi environment by prefixing the command with pixi run. For example pixi run pytest --no-cov tests/unit Running pytest on its own is a good way to debug a specific new or failing test quickly, but we should always use pixi run test for actual testing.

Git Pre-commit Hooks¶

• A variety of sanity checks are defined as git pre-commit hooks -- they run any time you try to make a commit, to catch common issues before they are saved. Many of these hooks are taken from the excellent pre-commit project.
• The hooks are configured in .pre-commit-config.yaml, and run using prek, a much faster, dependency-free tool that reads that same standard config format.
• For them to run automatically when you try to make a commit, you must install the hooks in your cloned repository first by running pixi run prek install. This only has to be done once.
• These checks are run as part of our CI, and the CI will fail if the hooks fail.
• We also use the pre-commit.ci service to run the same checks on any code that is pushed to GitHub, and to apply standard code formatting to the PR in case it hasn't been run locally prior to being committed.
• Run pixi run prek-update to bump the hook rev pins in .pre-commit-config.yaml to their latest versions. The update-lockfiles GitHub Action runs this (along with pixi update for pixi.lock) weekly and opens a PR with the changes.

Code Formatting & Linting¶

To avoid the tedium of meticulously formatting all the code ourselves, and to ensure a standard style of formatting and syntactical idioms across the codebase, we use the ruff code linter and formatter, which runs both as a pre-commit hook and via pixi run format / pixi run lint. These can be integrated directly into your text editor or IDE with the appropriate plugins. The ruff linter / formatter has a huge array of configuration options and different kinds of checks it can run, which are defined under the tool.ruff section of pyproject.toml.

We also have a custom hook that clears Jupyter notebook outputs prior to committing.

Type Checking¶

We use ty, an extremely fast Rust-based type checker from Astral (the ruff/uv team). It's configured under the tool.ty.src/tool.ty.environment sections of pyproject.toml and run via pixi run lint. It also runs as a pre-commit hook, but as a local hook rather than the official astral-sh/ty-pre-commit one -- that hook shells out to uv check, which needs network access to resolve dependencies, and pre-commit.ci disables network access while running hooks. Running it through the pixi environment (which already has every dependency installed) sidesteps that, but since pre-commit.ci still can't run it, it's separately enforced by a dedicated step in the pytest GitHub Actions workflow.

Code & Documentation Linters¶

To catch errors before commits are made, and to ensure uniform formatting across the codebase, we also use linters outside of ruff. They don't change the code or documentation files, but they will raise an error or warning when something doesn't look right so you can fix it.

• pre-commit has a collection of built-in checks that use pygrep to search Python files for common problems, as well as language agnostic problems like accidentally checking large binary files into the repository or having unresolved merge conflicts.
• hadolint checks Dockerfiles for errors and violations of best practices. It runs as a pre-commit hook.
• actionlint checks the GitHub Actions workflow files for errors. It runs as a pre-commit hook.
• shellcheck checks shell scripts for common bugs and portability issues. It runs as a pre-commit hook.
• markdownlint and mdformat check and reformat the Markdown documentation. The mdformat-mkdocs plugin keeps mdformat from mangling Zensical/MkDocs-flavored syntax, like the snippet-include lines mentioned above.

Test Coverage¶

• We use the pytest coverage plugin to measure and record what percentage of our codebase is being tested, and to identify which modules, functions, and individual lines of code are not being exercised by the tests.
• When you run pixi run test, a summary of the test coverage will be printed at the end of the tests (assuming they succeed). The full details of the test coverage are written to coverage.xml.
• There are some configuration options for this process set in the tool.coverage.report section of pyproject.toml.
• When the tests are run via the pytest workflow in GitHub Actions, the test coverage data from the coverage.xml output is uploaded to a service called CodeCov that saves historical data about our test coverage, and provides a nice visual representation of the data -- identifying which subpackages, modules, and individual lines are being tested. For example, here are the results for the cheshire repo.
• The connection to CodeCov is configured in the .codecov.yml YAML file. Uploads authenticate with the catalyst-cooperative org's shared "Global Upload Token," stored as an organization-level CODECOV_TOKEN secret in GitHub, so individual repos don't need their own CodeCov token minted and stored separately.
• CodeCov also adds a couple of test coverage checks to any pull request, to alert us if a PR reduces overall test coverage (which we would like to avoid).

Documentation Builds¶

• We build our documentation using Zensical, a modern Markdown-based static site generator from the Material for MkDocs team.
• Standalone docs files are stored under the docs/ directory as Markdown, and the Zensical configuration lives in zensical.toml at the top of the repository.
• The top level documentation page (docs/index.md) simply embeds this README.md verbatim using Zensical's pymdownx.snippets syntax (--8<-- "README.md"); docs/license.md embeds LICENSE.txt the same way. docs/code_of_conduct.md and docs/release_notes.md are standalone Markdown files.
• docs/reference.md holds the API reference, rendered from docstrings by mkdocstrings (configured under [project.plugins.mkdocstrings...] in zensical.toml, currently a preliminary Zensical integration). Add a module.path line there for any new module that should show up in the API reference -- it isn't generated automatically.
• Build the docs with pixi run docs, which wipes the previously generated site/ directory and rebuilds everything from scratch, or preview them locally with pixi run docs-serve.

Documentation Publishing¶

• We publish our documentation to GitHub Pages.
• When you push to main the docs GitHub Actions workflow builds the site with Zensical and deploys it automatically.
• To enable this for a new repository, go to the repo's Settings -> Pages, and under "Build and deployment" set the source to "GitHub Actions."

Dependabot¶

We use GitHub's Dependabot to automatically update the versions of the GitHub Actions that we employ, configured in .github/dependabot.yml. Our Python dependencies are refreshed separately, by the weekly update-lockfiles GitHub Action described below, instead of by Dependabot.

GitHub Actions¶

Under .github/workflows are YAML files that configure the GitHub Actions associated with the repository. We use GitHub Actions to:

• Run continuous integration with pixi run test and upload test coverage to CodeCov.
• Build the documentation with Zensical and deploy it to GitHub Pages.
• Build a Docker container using the docker-build-push action for every commit and PR, once pytest has passed for it, to catch Dockerfile breakage early. It's only pushed to Docker Hub for main and version tags, so branches and PRs don't clutter the registry with images nobody will pull.
• Release a new version of the package on PyPI when a version tag is pushed to main.
• Approve and enable auto-merge on bot PRs from pre-commit.ci and Dependabot, using gh pr merge --auto, which respects our merge queue and required status checks.
• Refresh pixi.lock and the rev pins in .pre-commit-config.yaml weekly, opening a PR with the changes so CI can confirm the updated dependencies still work.