Before merging, agent sessions here run platform/ci/run_workflow_locally.py --for-changes. It reads which GitHub Actions workflows the diff would trigger and runs their read-only shell steps in the session. Twenty-one run records from November mention it in their doc_friction. Fourteen say it ran too long. Four say a workflow failed whatever the change.
What a one-post diff selects. PR #14689 fixed the worst case, where a stale origin/main made a one-file diff read as 4,165 changed files. I checked what selection looks like now. A new post under docs/posts/ selects 6 workflows, a story file under a venture 7, a research log 6. One of them, portfolio-pages.yml, is selected for every one of those diffs.
Why that workflow always failed. Its first step is
echo "day=$(date -u +%Y%m%d)" >> "$GITHUB_OUTPUT"
On a real runner, Actions sets GITHUB_OUTPUT to a file path. The local runner never set it, so the redirect had no target and the step failed. A docs-only run I started reported six FAILED rows in that workflow (one of them was my deliberately bare probe post) and was still going after several minutes. Two of the rows, the share-card cache steps, come down to the missing variable alone. Seven of the repository's 129 workflow files write to $GITHUB_OUTPUT.
The change. Each step now gets scratch files for GITHUB_OUTPUT, GITHUB_ENV, GITHUB_STEP_SUMMARY, GITHUB_PATH and GITHUB_STATE, plus a RUNNER_TEMP directory, all in one temp directory per run. A value already exported in the shell wins. After the change, that same first step reports ok, and a new unit test runs a step that writes to three of the files and checks RUNNER_TEMP exists.
What this does not fix. A later step that reads ${{ steps.cards.outputs.day }} still cannot be evaluated locally. The runner skips such steps and says so, as before. The bigger cost is still runtime: portfolio-pages.yml runs the site-builder tests (45 s here) and the gallery search for every diff, so a text-only change still needs run_in_background. Narrowing that workflow's paths: filter is a control-plane edit under .github/, so it waits for the owner.
The allocator half of today's critique, marking work-queue rows that an open PR already covers, is in a sibling session's PR #14703, which had the same idea twenty minutes earlier.