@larsoner You put the original checks in, though I hesitate to ping you for a review as I want you to review my sphinx-gallery PR ;-)

I don't think this can sensibly be tested w/o merging to main. Perhaps@ksunden has a few minutes to see if it looks reasonable over the next little while?

It seems to work, CodeQL and linting are still run, which I think is fine (they tend to finish before the doc build anyway and are checking things like examples which may be cases where you would use this).

The one that we may want to add config to leave out is appveyor. that is pretty consistently the bottleneck as we get fewer parallel tasks from them, so they get backed up over multiple PRs.

Thedefault handles the standard set we had wanted, but we may need to write our own (possibly including the standard set as I think it overrides, not appends) now that we want another.

(We can remove/squash out my empty test commit if it's desired, but that was a successful attempt at testing, for what it's worth)

Cool - I would use this a lot, just to save heating up the server farm where CI lives.

I wonder if we could do even better/easier for us all and if the only files that are updated in a PR are indocs/* ortutorials/* only send to circle (i.e. anything that isn't tested code can ignore the tests). That would avoid needing to remember a commit message.

Something involving a path filter?https://docs.github.com/en/actions/using-workflows/workflow-syntax-for-github-actions#onpushpull_requestpull_request_targetpathspaths-ignore

That would be very cool - the [skip *] commands are used by very few people so far as I can tell

And if at some point you want to speed up your doc only build, you could render only the examples that have been modified by the PR, e.g.

https://github.com/mne-tools/mne-python/blob/11bc09911a2dbb6b73a86b31d8140ccd8a93b1c3/tools/circleci_download.sh#L14

But if your API pages take longer than your examples it might not be worth the effort

I'll go ahead and merge this as is. It works, and helps.

I may follow up with something making appveyor respect this key, as that is often the worst offender in terms of waiting for ci to complete.