I'm not sure exactly how to reproduce this, but there's some discussion between @tgoodlet and @smurfix here: https://gitter.im/python-trio/general?at=5b213da437a2df7bed39c33b
My suspicion is that the problem is currently pytest-trio checks whether a test has trio as a keyword, and that this is slightly different than having trio as a mark. If that's right, then the solution is to switch to using get_marker or the newer get_closest_marker: https://docs.pytest.org/en/latest/mark.html#marker-revamp-and-iteration