Six test files nothing ran, and the naming rule that hid them

Six node test suites for our browser checks matched no CI job because they were named test_x.mjs, not x.test.mjs. Now every pull request runs them.

This repository has a script whose whole job is to find tests that no CI job runs. It is called unwired_checks.py, and it already caught 30 Python checks, 71 Python test files and 34 node suites that had been sitting outside every workflow. A pull request runs whatever it finds, so a check added without CI wiring is still covered from the day it lands.

It looked for node suites by one pattern: *.test.mjs. Six suites in platform/ci are named the other way round, test_check_enter_reload.mjs and five siblings. No workflow named them and the pattern skipped them, so nothing ran them. They test the helpers behind our browser sweeps: the one that checks whether pressing Enter in a calculator wipes what you typed, the one that checks whether "1,200" is read as 1, and four more.

What we found when we ran them

All six pass: 22 tests, under 3 seconds in all. So no bug was hiding. The gap is that one could have been, and the suites guard checkers that runs keep extending.

One suite, for the "1,200" check, starts a real Chromium. The CI machine has no Playwright, so run there as written it would fail on a missing module rather than on a real fault. It now checks for Playwright first and reports a named skip when it's missing ("playwright is not installed here"), which we checked by hiding the module.

The change

unwired_checks.py now matches both *.test.mjs and test_*.mjs, prints the right cd ... && node --test ... line to reproduce either kind, and has a unit test that fails if a test_*.mjs suite goes missing from its list again.

What is still uncovered

The 14 browser sweeps themselves (check_enter_reload.mjs and the rest) are also run by no CI job. Each takes minutes and needs a browser, so they belong in the nightly browser-check job. That job is a workflow file, and workflow edits here wait for the owner's approval, so that gap is written down for a later run rather than fixed in this one.

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