My previous fix for this (#4151) was to use TRAVIS_BUILD_DIR which is already defined and is the same as MPL_REPO_DIR. Now we have a mix of MPL_REPO_DIR and TRAVIS_BUILD_DIR. I still think using TRAVIS_BUILD_DIR is clearer.

@matthew-brett I think you misunderstand. We need MPL_REPO_DIR because the pep8 tests need to know where the examples, which are not installed, are. You never changed the PEP8 test to read from TRAVIS_BUILD_DIR thus the merge of your pull request resulted in the pep8 job on the examples being marked as known fail. This sets the variable only for the purpose of running the pep8 test correctly. Which is also what the comment on the same line tells you. The original issue was that MPL_REPO_DIR was used for other stuff where it was not intended.

And IMHO special casing the PEP8 test to try reading TRAVIS_BUILD_DIR would be significantly less clear than defining this extra variable in the travis file

Just to be clear - you agree that the (already defined by travis) TRAVIS_BUILD_DIR is the matplotlib source directory and is therefore equal to MPL_REPO_DIR?

Ah sorry - I see - sorry - the matplotlib code is using MPL_REPO_DIR - that makes sense.

And so I'd have a vague preference for only using MPL_REPO_DIR in the pep8 section to be clear that it is specifically needed for that section, and using TRAVIS_BUILD_DIR for the travis stuff (cd $TRAVIS_BUILD_DIR).

This change just sets the variable it still uses TRAVIS_BUILD_DIR to do the navigation. There is no specific pep8 section that is just controlled by another env variable but we could wrap setting MPL_REPO_DIR in a if TEST_ARGS=--pep8 section to make it more clear.

Sorry - you're completely right - no quarrel at all about this change - I was confused by an earlier pull request.