-k expression ignores directory paths

[VladimirPodolian]

• a detailed description of the bug or problem you are having

The -k expression does not work with directory paths. For example, pytest -k '/common' does not collect any tests from the /common folder.

---

• output of pip list from the virtual environment you are using

Package    Version
---------- -------
attrs      25.4.0
iniconfig  2.3.0
packaging  26.0
pip        25.3
pluggy     1.6.0
py         1.11.0
Pygments   2.19.2
pytest     9.0.2
setuptools 80.9.0
tomli      2.4.0
wheel      0.45.1

---

• pytest and operating system versions
  MacOS 26.2 (25C56)
  pytest 9.0.2

---

• minimal example if possible

Files tree:
└── tests
├── common
│   └── test_common.py
├── desktop
│   └── test_desktop.py
└── mobile
└── mobile_test.py

# test_common.py:
def test_case_common():
    pass

# test_desktop.py:
def test_case_1():
    pass

# test_mobile.py:
def test_case_2():
    pass

Outputs:

/check-pytest [5]> pytest -k '/common'
============================= test session starts =============================
platform darwin -- Python 3.12.4, pytest-9.0.2, pluggy-1.6.0
rootdir: /check-pytest
collected 3 items / 3 deselected / 0 selected         
============================= 3 deselected in 0.01s =============================

[RonnyPfannschmidt]

slashes where never supported in mark or keyword expressions

[VladimirPodolian]

@RonnyPfannschmidt Got it, but in version 7.1.1 they were supported. By mistake probably

[RonnyPfannschmidt]

Reopening as this needs some research

[bluetech]

This changed in 385796b (issue #7777). The specific reason can be seen by looking at the collection tree before and after. Before:

<Module foo/test_it.py>
  <Function test_it>

After:

<Dir pytest>
  <Dir foo>
    <Module test_it.py>
      <Function test_it>

What KeywordMatcher does (the underlying "matcher" for -k) is here:

pytest/src/_pytest/mark/__init__.py

Lines 169 to 198 in 16f0764

	@classmethod
	def from_item(cls, item: Item) -> KeywordMatcher:
	mapped_names = set()
	# Add the names of the current item and any parent items,
	# except the Session and root Directory's which are not
	# interesting for matching.
	import pytest
	for node in item.listchain():
	if isinstance(node, pytest.Session):
	continue
	if isinstance(node, pytest.Directory) and isinstance(
	node.parent, pytest.Session
	):
	continue
	mapped_names.add(node.name)
	# Add the names added as extra keywords to current or parent items.
	mapped_names.update(item.listextrakeywords())
	# Add the names attached to the current function through direct assignment.
	function_obj = getattr(item, "function", None)
	if function_obj:
	mapped_names.update(function_obj.__dict__)
	# Add the markers to the keywords as we no longer handle them correctly.
	mapped_names.update(mark.name for mark in item.iter_markers())
	return cls(mapped_names)

The relevant piece is that it goes over all of the node's hierarchy and adds their name. As can be see above, before the names were foo/test_it.py and test_it, and now they are foo, test_it.py and test_it. That's why foo still matches but foo/ doesn't.

I would say that what @RonnyPfannschmidt said is correct - by the way the code is written, there was no intention to support / specifically, otherwise it would just have added the nodeid or paths or such. That said, if this issue had been reported after 8.0.0 was released, I would have considered it a regression and fixed it. But since ~2 years have passed since that release without more users complaining (AFAIK), I am now more inclined to treat this as a feature request that is considered on merit rather than a regression bug (sorry, I know this is unsatisfying).