Development and testing

Development and testing

Setup

Rust 1.89+, Python 3.10+, and prek for the repository hooks:

prek install --hook-type pre-commit --hook-type pre-push

Formatting, Ruff (including docstring rules), YAML/TOML and workflow lint run on commit; Clippy runs on push. prek run --all-files checks everything, as CI does.

Tests

cargo test --workspace --locked                      # unit, integration, end-to-end, doctests
python -m pip install . && python -m unittest discover -s tests/python -v

Python tests need the built package installed. python -m unittest discover -s scripts/tests covers the repository scripts, including the install scripts against a local fake release (install.ps1 too, when pwsh is installed). Three layers:

1. Offline end-to-end (crates/cli/tests/e2e.rs): the real executable with a cross-platform Pixi stand-in on every CI host.
2. Native integration (ignored by default): real Pixi, uv, cargo, conda-lock, Grype, pypi.org, crates.io and the GitHub API, plus a download of the pinned uv. PIXI=… UV=… CARGO=… CONDA_LOCK=… CONDA_SOLVER=… GRYPE=… GITHUB_TOKEN=… cargo test --workspace -- --ignored, and PIXI=… for the native Python tests.
3. CLI/Python parity (tests/python/test_parity.py).

Documentation

Public Rust items and the public Python API must be documented (missing_docs, Ruff D); cargo doc runs with warnings as errors. Follow the glossary in CONTEXT.md. The CLI reference and Python API pages are generated: run python scripts/gen-reference.py (with the package installed) after changing help text or docstrings. CI's docs-reference job fails when they are stale or a wiki link is broken.

CI and merging

• CI builds and tests on Linux, Windows and macOS (Intel and Apple Silicon), checks Rust 1.89, the sdist, the conda recipe, the end-to-end tests and prek.
• Distribution: the crates job packages and verifies both crates as crates.io would (cargo publish --workspace --dry-run); install-scripts runs the install script tests on macOS and Windows.
• Coverage: Codecov requires 80% of changed lines covered, and the in-repo gate requires 75% total line coverage.
• Integration runs the native tests on pull requests (informational), on main, nightly (failures open a nightly-integration issue) and before releases.
• Latest tools runs weekly. A tool's candidate is its highest stable release published at least 7 days ago and newer than the tested one, so new releases settle first and nothing moves back. Integration runs once per candidate with only that tool changed. When it passes on every OS, scripts/tool-drift.py opens a pull request on latest-tools/<tool> that bumps the adapter's tested_versions, its pinned downloads in provision.rs (with the release's sha256 digests) and its pin in .github/tool-versions.json together; review and land it like any other. A candidate that breaks Integration, a pinned download that is gone or whose sha256 changed, or a bump that cannot be made opens a latest-tools issue instead (one per tool). cargo, which CI runs as stable Rust, is never bumped automatically. crates/core/tests/pins.rs checks the pins agree with the tested versions.
• Changes reach main through pull requests with signed commits; the required checks must pass.