Package-level pytestmark in __init__.py no longer propagates to tests

[SuYanqi]

a detailed description of the bug or problem you are having

A package-level pytestmark = pytest.mark.skip(...) defined in __init__.py is no longer propagated to tests inside that package. Tests that should be skipped run and fail instead.

This is a regression introduced by PR #6197.

pytest and operating system versions

pytest 9.1.1
Python 3.11.13
Linux

minimal example if possible

from pathlib import Path
import shutil
import pytest

root = Path("/tmp/pytest_pkg_skip_test")
shutil.rmtree(root, ignore_errors=True)
package = root / "skippedpkg"
package.mkdir(parents=True)
(package / "__init__.py").write_text(
    "import pytest\n"
    "pytestmark = pytest.mark.skip(reason='package is disabled')\n"
)
(package / "test_skipped.py").write_text(
    "def test_should_be_skipped():\n"
    "    assert False\n"
)

result = pytest.main(["-q", str(package)])
print("exit code:", int(result))
shutil.rmtree(root, ignore_errors=True)

Expected: 1 skipped, exit code 0
Actual: 1 failed (assert False runs because skip marker is not applied), exit code 1

[Zuhef]

I can reproduce this, but I do not think it is a regression - the behaviour looks unchanged for as far back as I tested, so the fix is probably a policy decision rather than a straight bug fix.

Reproduction

Your script reproduces on current main (56b196e92): 1 failed, exit 1, where you expected 1 skipped.

Version matrix

I ran the same package against released versions in clean virtualenvs, so the only variable is the pytest version:

pytest	result
7.4.4	1 failed
8.3.5	1 failed
9.1.1	1 failed
main (56b196e)	1 failed

The marker is ignored in all of them, so there is no version where this worked that a bisect could land between.

Why #6197 is not quite the culprit

#6197 is in the history as 19a15a94e, "Fix incorrect discovery of non-test __init__.py files" (Nov 2019), and the changelog records it as:

:issue:6197: Revert "The first test in a package (__init__.py) marked with @pytest.mark.skip is now correctly skipped."

That commit removed 24 lines from src/_pytest/python.py and deleted the corresponding test from testing/test_skipping.py, in order to fix #6194. So the current behaviour was chosen deliberately roughly six years ago as the cost of not collecting non-test __init__.py files - it is the intended state of the code, not something that drifted.

Why I stopped short of a patch

Restoring propagation means collecting __init__.py as a module again in some form, which is the thing #6197 was written to stop. A patch that simply re-applies the reverted behaviour would likely reintroduce #6194, so this needs a maintainer's call on which of two documented behaviours wins.

If the desired outcome is that package-level marks apply, I would be glad to implement whichever direction you prefer:

1. Collect marks from __init__.py without collecting it as a test module - read pytestmark from the package during Package collection and fold it into the child modules' marks, leaving discovery untouched.
2. Document the current behaviour instead, since conftest.py already provides a supported way to apply marks package-wide (pytestmark in conftest.py, or pytest_collection_modifyitems), and close this as working-as-intended.

Option 1 keeps the reporter's expectation and avoids #6194 because nothing new is collected, only marks are read - but I did not want to assume that is wanted given the explicit revert.

[aryansk]

Working on this — will propose a fix in Package to propagate pytestmark from init.py without re-collecting it as a test module (to avoid reintroducing #6194). Will open a draft PR shortly.