I really dislike the fix, commit, and wait loop that is involved with CI. I decided to fix that by shortening the loop. I did that by mocking the GH Actions control pane: the runner is the official GH runner, but the API is a mock.
What you get is caching in ~0 ms. Pause on failure. Let your agent fix it and retry, without pushing!
It's easy for humans, but even easier to AI to validate that your actually going to pass CI - the result is that an agent won't tell you it's done if CI doesn't pass.