feat: constraints mode for testing the oldest supported versions in CI - #46

Constraints mode for testing the oldest supported versions in CI

The action can already raise the lower bounds in pyproject.toml. This adds a second mode that helps projects check those lower bounds: mode: constraints writes a constraints file with exact SPEC 0 minimum versions, so a CI job can install and test the oldest supported combination. It does not touch the project file or open a PR.

Motivation: shap wants to replace a hand-maintained numpy==2.0 pin in its CI with floors taken from the SPEC 0 schedule (shap/shap#5263). That issue suggests extending this action instead of writing a separate script, so other projects can use it too. Made during the GC.OS hackathon in Ulm.

- uses: scientific-python/spec0-action@<commit>
  id: spec0
  with:
    mode: constraints
    packages: numpy scipy pandas scikit-learn matplotlib python
- uses: actions/setup-python@v6
  with:
    python-version: ${{ steps.spec0.outputs.python_version }}
- run: pip install -c constraints-spec0.txt . && pytest

Changes

• spec0_action/constraints.py: computes the floors on a reference date by applying all schedule entries up to that date in order (each entry only lists the packages that change), and formats them as exact pins. With >= the installer would pick the newest release, so the minimum would never be tested. Python is left out of the file and reported as an output instead.
• run_spec0_constraints.py: CLI with --packages, --output and --reference-date. Writes constraints_file, versions (JSON) and python_version to $GITHUB_OUTPUT.
• action.yaml: new inputs mode, packages, constraints_file, reference_date and the three outputs. In constraints mode the checkout, git setup, update and PR steps are skipped. update stays the default, so existing users are not affected.
• pyproject.toml: write-constraints pixi task.
• Readme section and a new job in test_action.yaml that runs the action in constraints mode and checks the file and outputs.

The run fails with a message naming the packages if the schedule has no floor for them on the reference date, and on invalid schedule data, instead of falling back to the newest version.

Testing

• 19 new tests in tests/test_constraints.py (floors across entries, start date boundaries, unsorted schedule, missing floors, invalid schedules, CLI outputs). Full suite: 101 passed.
• pre-commit passes.
• Checked with the current schedule: numpy 2.3.0, scipy 1.15.0, pandas 2.3.0, scikit-learn 1.7.0, matplotlib 3.11.0, ipython 8.32.0, Python 3.13. I then installed shap with that constraints file on Python 3.13 and ran its test suite: 386 passed, 0 failed.
• Ran the action for real in a shap CI job on my fork: it generated the pins, set up Python 3.13 from the python_version output, installed shap at those versions and the tests passed (387 passed, 0 failed): https://github.com/omar-radwan7/shap/actions/runs/41460204617/job/125326992471

AI disclosure

Written with Claude Code. I chose the task and the approach with the shap maintainers at the hackathon. Claude wrote the code and tests and ran them. I reviewed the changes before opening this PR.

Generated with Claude Code

https://claude.ai/code/session_01Q7ywk9uaZ3gLUcBP7UysV3