Fix spurious 'no result set' warnings and dropped execution tags under pytest-forked

[dahtey-bk]

Context

Customer escalation: a suite using pytest --forked (required for their tests — removing it causes deadlocks) logged WARNING - Test <nodeid> has no result set at finalization for every test after switching to this collector.

Root cause

pytest-forked runs each test's protocol in a forked child process with log=False, then replays the serialized reports in the parent:

1. The parent creates the in-flight test record at pytest_runtest_logstart, then forks.
2. In the child, pytest_runtest_makereport fires (setup/call/teardown), but pytest_runtest_logreport — the only hook where we set the pass/fail result — never fires there because of log=False. So at teardown the child finalizes its inherited copy of the in-flight record with result=None and emits the warning.
3. The child exits and its entire copy of the plugin state is discarded. The parent replays the reports, sets the result, and finalizes correctly.

So the warning was pure noise from a doomed process copy — uploaded results, durations, and split history were never affected. Reproduced and confirmed with PID-stamped debug logs (parent pid creates + correctly finalizes; child pid emits the warning).

The same parent/child state split hid a second, real bug: execution_tag markers were silently dropped under --forked, because tags were only applied in pytest_runtest_makereport, which only runs in the discarded child.

Changes

src/buildkite_test_collector/pytest_plugin/buildkite_plugin.py

• Record the owning process pid (_owner_pid) at plugin construction. finalize_test is now a no-op in any other process (i.e. a fork-per-test child), eliminating the warning at its source. Safe with xdist: each worker runs its own pytest_configure, so plugin pid == worker pid.
• Capture execution_tag markers into _tags_by_nodeid during pytest_collection_modifyitems (which runs in the owning process, before any fork) and apply them in finalize_test. Tag application in pytest_runtest_makereport is kept so runtime-added markers still work in non-forked runs; re-applying the same values is idempotent.

Tests

• Unit tests: finalize skipped in a simulated forked child, finalize unchanged in the owning process, collection-time tag capture + application (including ignoring malformed markers).
• New integration test (tests/buildkite_test_collector/test_integration_forked.py) running real pytest --forked end to end, asserting no warnings, correct results, and correct tags. Skips cleanly where pytest-forked/os.fork are unavailable.

Dependencies

• pytest-forked added to the dev extra (non-Windows only) for the integration test; uv.lock regenerated.

Behaviour matrix (verified locally on pytest 8.4.1 and 9.1.1)

Mode	Warnings before → after	Tags before → after
plain	none → none	✅ → ✅
--forked	one per test → none	❌ dropped → ✅
-n 2 (xdist)	none → none	✅ (worker uploads) → ✅
-n 2 --forked	one per test → none	❌ dropped → ✅ (worker uploads)

Results (pass/fail/skip), failure reasons, and durations verified identical across all modes. Full suite passes (uv run pytest), pylint 10.00/10.

🤖 Generated with Claude Code

[socket-security]

Review the following changes in direct dependencies. Learn more about Socket for GitHub.

Diff	Package	Supply Chain Security	Vulnerability	Quality	Maintenance	License
	pytest-forked@​1.6.0

View full report

[omehegan]

This looks fine to me, but I don't have the ability to merge to this repo.

[omehegan]

@gchan we don't have permission to merge to this repo. If you're happy, can you merge and possibly do a release as well?