Repository navigation
ci: the release job installs the declared test extra, so tagging works again - #55
Merged
Merged
Conversation
Pushing v2.4.0 failed. Every release test job died 0.85s in with `ModuleNotFoundError: No module named 'hypothesis'` while collecting `tests/biology/test_pass_isolation.py`, and `-x` turned that into one line of output. Nothing was published: the publish job needs the test job, so sdist, wheels and the PyPI upload were all skipped and the version number is still free. The cause is dependency drift with a long fuse. `build-wheels.yml` installed `pytest pandas` by name and then the built wheel with the `[mcp]` extra alone, while `test.yml` installs it with `[test,mcp]`. That hand-written list was accurate until a test module imported something outside it. `hypothesis` entered `tests/support/pipelines/registry.py` on 2026-10-31, so every release has been broken since -- invisibly, because no release was attempted between then and now, and because the suite CI actually runs was green the whole time. So the release job now installs the extras rather than restating them: `[test]` is where pytest, pandas, hypothesis, pytest-xdist and airr are declared, and `pyproject.toml` stays the only place that knows the list. A pin in the CI contract enforces it, since the thing that makes this class of bug expensive is that a tag is the first place it shows. It asserts the release job installs the wheel with `[test]` and names no test dependency by hand. Verified by pointing it at the pre-fix workflow: it fails and quotes the offending line.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Pushing
v2.4.0failed. Nothing was published — the publish jobneeds the test job, so sdist, wheels and the PyPI upload were all
skipped. PyPI still returns 404 for 2.4.0, the latest release is still
v2.3.1, and the version number is free.
What happened
Every release test job died 0.85s in:
pytest -xreports a collection error as a single failure, which is whytwelve platforms produced one line of output each.
Why it was invisible
build-wheels.ymlinstalledpytest pandasby name and then the builtwheel with the
[mcp]extra alone.test.ymlinstalls the same wheelwith
[test,mcp]. The hand-written list was accurate until a testmodule imported something outside it —
hypothesisenteredtests/support/pipelines/registry.pyon 2026-10-31.So every release since 2026-10-31 would have failed, and nobody
could know: no release was attempted in between, and the suite CI
actually runs was green throughout. The first evidence of the break was
a failed release.
This is not fallout from #54. #54 was 13/13 green on the exact tree that
was tagged.
The fix
The release job installs the extras rather than restating them.
[test]is where
pytest,pandas,hypothesis,pytest-xdistandairraredeclared, so
pyproject.tomlstays the only place that knows the list,and this particular drift cannot recur.
The pin
What makes this class of bug expensive is that a tag is the first place
it shows, so a guard belongs in the suite rather than in the release.
tests/support/meta/test_ci_contract.pygains a pin asserting therelease job installs the wheel with
[test]and names no testdependency by hand.
Verified by pointing it at the pre-fix workflow — it fails and quotes
the offending line:
Verification
pytest tests/support/meta/test_ci_contract.py— 2 passedruff check tests/ src/ tools/— cleanAfter this merges
The
v2.4.0tag points at the #54 merge commit, which does not carrythis fix, so it has to be moved to the new master before the release
will run. Nothing has consumed that tag — no PyPI artifact, no GitHub
release — so moving it costs nothing, but it is your call.