[openvoiceos-bot]

The defect

The Run Tests step ran python -m pytest. The -m form puts the working
directory first on sys.path, and the working directory is the repository root.
For a flat-layout package the root holds the package directory, so import mypkg
resolved to the source tree and never to site-packages. The job built a wheel,
installed it, and then tested the checkout instead. A packaging defect passed in
every caller: a module left out of the wheel, a dependency not declared, package
data not shipped.

The proof

A control package whose wheel contains no code at all:

$ python -c "import zipfile; print(zipfile.ZipFile('dist/ctlpkg-0.0.1-py3-none-any.whl').namelist())"
['ctlpkg-0.0.1.dist-info/METADATA', 'ctlpkg-0.0.1.dist-info/WHEEL',
 'ctlpkg-0.0.1.dist-info/top_level.txt', 'ctlpkg-0.0.1.dist-info/RECORD']
$ ls .venv/lib/python3.12/site-packages/ | grep ctl
ctlpkg-0.0.1.dist-info          # the dist-info only, no module

The old command and the new one, same tree, same venv:

$ python -m pytest tests -v      # the command this workflow ran
tests/test_ctl.py::test_answer PASSED      1 passed        <- the defect
$ pytest tests -q                # the command it runs now
FAILED tests/test_ctl.py::test_answer - ModuleNotFoundError    1 failed

The reverse control, a correctly packaged repository with a test that asserts
where the import came from:

$ pytest tests -v
tests/test_ok.py::test_answer PASSED
tests/test_ok.py::test_imported_from_site_packages PASSED
tests/test_ok.py::test_relative_cwd_fixture_still_readable PASSED    3 passed
$ python -m pytest tests -q      # the old command, same repository
FAILED tests/test_ok.py::test_imported_from_site_packages     1 failed, 2 passed

The fix

One token: pytest, not python -m pytest. The console script adds no working
directory entry.

The working directory is deliberately not changed. A suite that reads a
repo-relative fixture path keeps working, and prepend import mode still puts the
test directory on sys.path, so a helper module beside the tests still resolves.
Both are controls in the new test file.

--import-mode=importlib was measured and is not needed: dropping -m is
sufficient, and importlib mode would stop a tests-local helper import from
resolving.

Blast radius, counted

callers of build-tests.yml in the checked-out clones:      192
of those, with a test_path set:                            187
of those, flat layout (a package directory at the root):   171   <- were affected

A screen for callers whose tests import a root-level name that the packaging
config never names, which is where a new red is possible:

screened: 187 in, 187 out
candidates: 4
  ml/voiceclonnx           tests   imports: conversion
  ml/wt-spike              test    imports: benchmarks
  ovos/ovos-skill-moon-game test   imports: __init__
  utils/powerguess         test    imports: train

The screen is a screen: a name that appears anywhere in the packaging config
counts as declared, so it can miss a case and it can raise a false one. It is
not a substitute for the sweep.

A limit worth naming

48 of the 182 caller test directories call sys.path themselves. Where a
conftest re-adds the repository root, this fix is defeated and the suite still
tests the source tree. Those repositories need the sys.path line removed, one
at a time, and that is separate work.

Canary runs

Three callers were pointed at this branch. One produced the evidence; the other
two are red for reasons that predate this change, and both are named below as
fleet work, not as a result of this one.

OpenVoiceOS/nebulento#57, run 36218046118: green on all five versions.
The executed step, read from the log of job 108337853495:

# ... the comment this change adds ...
pytest test/ -v  | tee pytest.log
...
rootdir: /home/runner/work/nebulento/nebulento
======================= 178 passed, 5 warnings in 23.05s =======================

The new command form ran, the rootdir is still the repository root, and a real
178-test suite stayed green. nebulento is flat layout and carries a second root
package (benchmark), so it is the shape most at risk.

TigreGotico/linux2mqtt#15: no evidence, the repository has no working CI.
7 of its 9 workflows that call a gh-automations reusable end in
startup_failure, and the 2 that call none (docker.yml,
conventional-label.yml) run. The same is true on its other branches, so it
predates this change. The cause is not this workflow: the repository default is
read and those 7 callers carry no permissions: block, so the forge refuses
the run before any job starts. Diagnosed and fixed in linux2mqtt#16.

VoiceGamez/ovos-skill-moon-game#21: no evidence, red before the tests.
The Install Built Package step fails to build fann2==1.0.7 (ld: cannot find -ldoublefann) on all five versions, so no job reaches Run Tests. Its
build-tests has failed this way since 2026-06. Filed separately.

That leaves the __init__ prediction in the screen above unverified by a
run: moon-game was the one repository that could have shown it, and it never
gets far enough to import anything.

Local test result

$ pytest test/test_build_tests_runs_the_wheel.py -v      6 passed
$ pytest test -q          779 passed, 6 failed
$ git stash && pytest test/test_lint_actionlint.py test/test_scripts.py -q
                          146 passed, 6 failed        # the same 6, on dev

The 6 are pre-existing and local: no actionlint binary and no
ovos-plugin-manager in this environment.

Mutation check: restoring python -m makes
test_the_workflow_does_not_invoke_pytest_through_dash_m fail, so the guard is
not a silent pass.

🤖 Generated with Claude Code

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Advanced

Run ID: 5b36a47a-9b6b-483d-938d-1cffcb763155

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}