That PR is the one that introduced the behavior to skip on [ci doc], prior to that it wasn't accounted for.

I tested it in that PR by adding an empty commit and it didn't run gha (appveyor still ran, I believe)

As far as I can tell, it worked for the most recent commit of #26669 (a84ab3d).

Note that it does only take into account the most recent commit message, not all commits of the PR.

Also, only the tests workflow is skipped, not all gha workflows, so linting/pr cleanliness, mypy, etc still do run, but those are all pretty quick checks.

Quoting myself from #25383 (comment)

the GitHub Actions tests still ran!

Yeah, I've been digging around the docs for that, as I've noticed it before. From what I can tell, head_commit is a property of a push event, but not of a pull_request event. So possibly it just can't work for PR CI?

Hmmm.... I mean just looking at #26669, that does not show the tests as having run for the PR status checks, so it doesn't seem to be a universal thing, but perhaps it doesn't work on initial opening of the PR, but does for subsequent pushes?

Would have to test it out to see the boundaries.

#26669 only touches files under doc and galleries, so will also be affected by these lines:

2065f14 ran all the Github tests despite the commit message being [ci doc]

However, I did not make the initial commit have this in the commit message. Is that the difference?

Sorry, I meant #25175 broke this.

the logic is indeed different: if github.event_name == 'workflow_dispatch' is True then the workflow runs, even if '[ci doc]' is in the commit message. Is it possible workflow_dispatch is True in these runs? From the docs it seems it should not be, but??

"workflow_dispatch" is the keyword for "manually run action", i.e. running the action by clicking the "Run workflow" button in the GHActions web UI.

I think if you are asking explicitly to run a workflow, it should run, regardless of whether the last commit has a [ci skip] in it or not, so I think that logic is sound.

I do wonder if "ignore paths" causes empty commits to skip, which made it look like it was working when I was testing it, though.

The paths_ignore was added at #25261, so more recently than the #25145 testing.