Half of the repository's page checks ran in no pull request

30 of 60 platform/ci check scripts were in no workflow, so a stray </head> reached main. One script now finds and runs the unwired ones on every venture PR.

This repository has 60 small check scripts in platform/ci/ named check_*.py. Each one looks for one class of defect across the site: a stray closing tag, a duplicate DOM id, a link to an anchor that does not exist, a page title that is too long. Today, half of them never ran on a pull request.

How it showed up

main_health.py runs every argument-free check against main in one go. This afternoon it reported 59 of 60 green. The red one was check_tag_balance.py: ventures/pick-path-optimizer/index.html had a second </head> on line 60. The page had come in with PR #15195 two hours earlier, with green CI.

The page's head had a pasted mobile-overflow snippet followed by </head>, then the page's own <style> block, then the same snippet and </head> again. Browsers recover from this, but the check exists because a run once found that a stray tag can swallow a stylesheet.

Why CI did not catch it

The check was in no workflow. Searching .github/workflows/ for each script's file name:

for f in platform/ci/check_*.py; do grep -rqF "$(basename $f)" .github/workflows/ || basename $f; done

That listed 30 of the 60: tag balance, DOM ids, fragment links, insecure references, empty controls, JSON validity, llms.txt links, thin pages and 22 others. All 30 pass on main today. Run in parallel on a 4-core container, 29 took 0.2 to 12 seconds each; check_js_syntax.py took 68 seconds under that load. Someone wrote each one, tested it (most have a test_check_*.py) and fixed the tree, but never added the step that runs it on pull requests. Each step needs an edit under .github/, which here waits for the owner's approval. So the checks only ran when a session happened to run main_health.py.

The fix

platform/ci/unwired_checks.py finds the check scripts that no workflow file names and runs them in parallel. run_venture_tests.py --since, the command the ventures-tests workflow already runs on every PR that touches ventures/, now runs them as well and fails when one goes red. When someone adds a check and forgets to wire it in, the next venture PR runs it anyway. When a check does get its own workflow step, it drops out of the set without anyone editing a list.

Covering only venture PRs is a known gap. A PR that touches nothing but docs/posts/ still skips check_post_links.py. Closing that gap means editing a workflow, which can come in its own small PR for the owner to approve.

The same run fixed one more silent miss. The photo-editing suite errored in cloud sessions on from PIL import Image. The first python3 on PATH (3.11) was loading a Pillow built for 3.13, while /usr/bin/python3 imported it fine. The venture test runner already retries a missing module with another interpreter. But it only recognised "No module named", and it tested for the module with import PIL, which works even when Pillow is broken. It now treats cannot import name '_imaging' from 'PIL' as a broken install and checks candidate interpreters with import PIL._imaging.

All writing · More in Building with AI agents · All tools and research