Agents in this repository check a change with platform/ci/run_workflow_locally.py, which runs the GitHub Actions workflows a diff would trigger on the agent's own machine. Twelve sessions this week said the same thing about it: any change to a product, a blog post or a research page selected the site workflow, and its site-builder tests ran past the 300-second budget, so the run ended with no verdict.
The suite is not slow in CI, because CI keeps the share-card images (about 7,300 of them) in actions/cache between runs. The local runner skips every uses: step, cache included, and gave runner.temp a fresh temporary directory on each run. So every local run drew every card again, four times over, and left about 350 MB of images in /tmp.
The fix gives a cache restored under runner.temp a local stand-in: a directory under ~/.cache that survives the process. That is no less safe than the real cache, since a workflow that restores one must already cope with an older entry. Measured in a 4-core container:
| run | site-builder step |
|---|---|
| cold (first run in a container) | 330 s |
| with the kept cache | 116 s |
While timing it, a worse problem turned up. The runner skips jobs whose if: cannot be true on a pull request, but it gave up on any condition containing ||. The deploy job's condition has a || inside parentheses, then github.event_name != 'pull_request', so the runner ran the deploy job, and its IndexNow step got as far as indexnow.mjs --submit before failing on a missing file. The parser now splits && only outside parentheses and quotes. Across all workflows, the steps that stop running locally are all deploys or writes that CI never runs on a pull request: two Cloudflare Worker deploys, IndexNow submissions, a branch delete and an access-policy update.
The details and per-file timings are in docs/history/run-workflow-locally-card-cache-2026-11-14.md.