fix #14004 - connect conftests to nodeids/nodes instead of matching string prefixes - ensure we connect in the collect phase in case of directory confusion

[RonnyPfannschmidt]

fixes #14004

the key issue was that nodeid/baseid for fixtures of such conftest.py would be "" instead of something fitting

ths pr is a chain of changes

• 694e2f6 is a quickfix i volunteer to backport
• the rest migrates to assigning fixtures to nodes nodes as @bluetech introduced directory nodes that make string matching unnecessary and deprecates it

[Copilot]

Pull request overview

This PR fixes conftest fixture scoping when testpaths points outside rootdir using relative paths (issue #14004). The fix migrates from fragile string prefix matching to robust node-based matching for fixture scoping. Conftest fixtures are now parsed during Directory collection, using the Directory node's nodeid for proper scoping instead of calculating string-based nodeids during plugin registration.

• Implements deferred conftest parsing tied to Directory collection
• Adds node-based fixture matching alongside string-based fallback for compatibility
• Deprecates baseid/nodeid string parameters in favor of node parameter

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
testing/test_conftest.py	Adds comprehensive regression test for fixture scoping with testpaths outside rootdir
src/_pytest/fixtures.py	Implements deferred conftest parsing, node-based matching, and deprecates string-based APIs
src/_pytest/unittest.py	Updates fixture registration calls to use new node parameter
src/_pytest/python.py	Updates fixture registration calls to use new node parameter
changelog/14004.deprecation.rst	Documents deprecation of baseid/nodeid string parameters
changelog/14004.bugfix.rst	Documents the fixture scoping fix

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[bluetech]

Interesting!

The "tests outside of the rootdir" scenario is ill-advised but I guess we need to support it.

---

Regarding the first commit, I'm looking at the test test_conftest_fixture_scoping_with_testpaths_outside_rootdir, and first thing I wanted to understand is the collection tree and nodeid's generated in this case (when pytest is run from the rootdir pytester.path/sdk). The result is (I added the nodeid reprs after the names):

<Dir tests> '::tests'
  <Dir sdk> ''
    <Dir inner> 'inner'
      <Module test_inner.py> 'inner/test_inner.py'
        <Function test_inner> 'inner/test_inner.py::test_inner'
    <Module test_outer.py> 'test_outer.py'
      <Function test_outer> 'test_outer.py::test_outer'

That seems odd and broken to me, there's obviously a confusion going on here. Generally the intention is for nodeid's to be relative to the rootdir, so what happens when the collection path is not under the rootdir? I've already written about this issue in #11245.

The commit "fix #14004 - connect conftests to nodeids/nodes instead of matching string prefixes - ensure we connect in the collect phase in case of directory confusion" works around the issue using some "fixup" code in fixtures.py. I think it's a bit confusing, particularly the function name "_get_nodeid_for_path_outside_rootpath", it basically invents a "fake" nodeid because the existing nodeid is broken, but it's not an actual ID of a node...

I think that if we fix the underlying issue #11245, that would be a better way to go? In that issue we had a couple of suggestions.

Also regarding the implementation itself, I fear that the loop over the config.args might create an O(n^2) situation which could be problematic if lot's of args are given (I know some users pass lot's of args). I haven't investigated in detail, just something to look out for.

---

Regarding the "fix: assign conftest fixtures to Directory nodes during collection" idea and subsequent changes, I think I like the direction, I definitely think we should to try to integrate conftests with the collection tree better, and get away from the current path matching. Thanks for taking a look at this.

I need to look into it more, but can you describe the interaction with initial conftests? Are they associated with a Directory or are they "global"?

The _pending_conftests adds a bit of tricky state, I wonder if we can't pass the node directly instead of relying on the collection report hook? I'm thinking here:

pytest/src/_pytest/runner.py

Lines 383 to 399 in 0e9db5f

	def pytest_make_collect_report(collector: Collector) -> CollectReport:
	def collect() -> list[Item | Collector]:
	# Before collecting, if this is a Directory, load the conftests.
	# If a conftest import fails to load, it is considered a collection
	# error of the Directory collector. This is why it's done inside of the
	# CallInfo wrapper.
	#
	# Note: initial conftests are loaded early, not here.
	if isinstance(collector, Directory):
	collector.config.pluginmanager._loadconftestmodules(
	collector.path,
	collector.config.getoption("importmode"),
	rootpath=collector.config.rootpath,
	consider_namespace_packages=collector.config.getini(
	"consider_namespace_packages"
	),
	)

[RonnyPfannschmidt]

the main confusion thats happening is that initial conftest files are loaded before we have a node, so we cant parse factories until we see them again in collection

the case of tets outside of a rootdir is not as ill advised as one might think the moment one ships integration testsuites as installed python packages and suddenly many tests collect as part of site-packages instead of the cwd

[bluetech]

The "add: nodes.norm_sep helper for nodeid path separator normalization" change looks sensible, you can submit and merge it to main with my review if you'd like. As an internal refactoring it doesn't need a changelog entry though.

Can you submit "improve: compute meaningful nodeids for paths outside rootdir" in its own PR? I have some thoughts but it's a substantial change so will be better considered separately before this PR.

[bluetech]

the main confusion thats happening is that initial conftest files are loaded before we have a node, so we cant parse factories until we see them again in collection

But do all initial conftests will have a node? Conftests are looked up the filesystem tree (up to confcutdir) while nodes are not. So I figure it's possible to have a conftest plugin without a corresponding Directory, though I haven't verified this so might be wrong about that.

[RonnyPfannschmidt]

As far as I understand any conftest not having a attached node isn't a valid fixture source

[bluetech]

You may be right but that would surprise me! I'd imagine these fixtures would be "global" (empty/None baseid, don't remember which one of them). Are you able to verify this?