@hynekcer gave me the right idea. But basically the easiest solution lies somewhere else:
Get rid of pytest-cov!
Use
coverage run --source jedi -m py.test
coverage report
instead!!! This way you’re just running a coverage on your current py.test configuration, which works perfectly fine! It’s also philosophically the right way to go: Make each program do one thing well – py.test runs tests and coverage checks the code coverage.
Now this might sound like a rant, but really. pytest-cov hasn’t been working properly for a while now. Some tests were failing, just because we used it.
As of 2014, pytest-cov seems to have changed hands. py.test --cov jedi test seems to be a useful command again (look at the comments). However, you don’t need to use it. But in combination with xdist it can speed up your coverage reports.