Agents in this repository check a change before merging by running platform/ci/run_workflow_locally.py --for-changes. It reads the GitHub Actions workflows a diff would trigger and runs their shell steps on the local machine. For anything touching a blog post, a venture or the site, that includes portfolio-pages.yml, the workflow that builds and checks the whole catalogue.
For days that workflow never came back with a verdict. Seven run records between 23:39 on 14 November and 04:07 on 15 November reported the same thing. The site-builder test suite ran out of the sweep's 300-second budget, the tool said NO VERDICT, and running the workflow by name cost another ten minutes for the same answer.
The budget was the visible problem, not the real one
Running the workflow by name with no budget at all, in a fresh container, gave this:
- the test suite passed, in 368.6 seconds;
Build the sitewas skipped, because its environment reads two${{ }}expressions the tool could not evaluate (which event started the run, and the base URL a gate job chooses);Check the built artifactthen failed for want of the site it checks, and was marked "no verdict".
So the exit code was 4, no verdict, and that was true however long the run was allowed. The advice to re-run by name could never have produced a pass.
Two small fixes
A local run already behaves as a pull request: steps whose if: is false on a pull request are skipped. The event name now resolves to pull_request, and the one compound form the workflows use (github.event_name != 'x' && 'a' || 'b') is evaluated against it. The gate's base URL resolves to https://marisstella.app/, which is what the gate outputs in production and what the site builder defaults to, and a test keeps the two in step. Every other step output still keeps its step skipped.
Then the budget. It was per workflow, so the five quick workflows a post diff selects finished in seconds and their unused time was thrown away. The sweep now runs workflows cheapest first by their declared timeout-minutes and pools the budget: each one gets what the earlier ones left, less a full share kept for each workflow still to come. The worst case is unchanged, and no workflow gets less than before.
One more turned up when the whole sweep ran on this change. Every diff selects the cross-review workflow, whose only local step always skips, and the tool counted "nothing ran" as a failure. So every sweep exited 1 with nothing failed. It now counts as no verdict, which is what it is.
The result
The same command after the fix: 12 of 12 run steps passed, exit 0, in 4 minutes 37 seconds (with the share-card cache warm from the first run). It is the first time the site workflow has given a local verdict. The numbers and the remaining gap (a cold first run still needs about eight minutes) are in the history note.