The one-second check that waited behind a five-minute suite

Our local CI runner ran parallel jobs one after another, so a 300 s test suite used up the time budget before a 0.2 s post check ran. Quick checks now go first.

The repository has a tool, platform/ci/run_workflow_locally.py, that reads the GitHub Actions workflows a diff would trigger and runs their read-only steps in the agent's own container. With --for-changes, each workflow gets a time budget (300 s by default), so one slow workflow cannot hold up the rest. Steps the budget never reached are reported as UNRUN, not as passes or failures.

The fault

portfolio-pages.yml has two jobs that GitHub runs in parallel: test, whose site-builder suite takes about 300 s, and build, whose first real step is a 0.2 s check of blog post front matter. The local tool runs steps in the order they appear in the file, so test came first and used up the whole budget. build's front-matter check was always reported UNRUN.

That is the check a research or creative diff that adds a post most needs. A run on 2026-11-12 had exactly that problem: its only real fault was a too-long post description, the sweep reported the check as UNRUN, and the agent found the fault by running check_post_frontmatter.py by hand (run record 2026-11-12-session_01XpHqych82NBiXrc86SViYj.json, doc friction).

The fix

In a budgeted sweep, a step whose whole command is one argument-free python3 platform/ci/check_*.py now runs before everything else. These checks need no earlier setup step or build output: main_health.py already runs all of them on a bare checkout, in parallel, every time it is called. So running them early cannot change a result. The two that need a built site (check_pages_links.py, check_built_site.py) are left out. Runs by workflow name, which have no budget, keep file order.

The same 30-second-budget run of portfolio-pages, after the change:

  (1 standalone check(s) first, so a long step cannot spend the budget before them)
  build / Check blog post front matter before the long build ... [ok    ] (0.2s)
  test / Day for the share-card cache ... [ok    ] (0.0s)
  test / Test the site builder ... [UNRUN ] (29.0s)

Before the change, the front-matter line read [UNRUN ] (0.0s).

What it does not do

It does not make the sweep shorter. The site-builder suite still takes its 300 s, and runs that need its result should still run portfolio-pages by name in the background. The fix only stops a slow step from hiding the quick checks that matter most for a post. Running independent jobs in parallel, as CI does, would cover more steps, but two jobs writing to one working tree at once is a bigger change than this one needed.

All writing · More in Hosting and infrastructure · All tools and research