Two failing runs across refaire (run 1709) and rememed (run 1624)
both crash at the same step (`Set up Docker Buildx`) with a node error
on dist/index.js line 1, AND both reference the same act cache dir:
/root/.cache/act/6a647958c11e138a6cfcaf32d2b372bc8e0c97871d617bfb441d003d505b77cf
act keys remote-action cache entries by repo URL alone — pinning to
`@v3.10.0` doesn't help, every game that uses `docker/setup-buildx-
action` lands in the same dir. When you push N games at once on
home-runner-1, the act-runner does parallel `git clone` ops into that
shared dir; the loser's pull aborts ("worktree contains unstaged
changes") and leaves dist/ half-written, so the next job's `node
dist/index.js` throws on line 1 → step fails. That's the entire flake.
Fix: drop the two remote actions that were racing — setup-buildx-action
and build-push-action — and replace them with inline `docker buildx
create` + `docker buildx build --push` shell. Nothing is fetched from
GitHub at runtime, no cache dir is shared, the failure mode disappears.
Same image, same tags, same registry mirror, same cache-from/cache-to
shape, same secret-files mount (`--secret id=...,src=...`). Each job
gets a uniquely-named builder (`builder-<game>-<api-core|web>`) and a
teardown step so the runner host's docker state doesn't accumulate
abandoned builders.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
enricobuehler
0c5f9a347a