Today's red-team run drew platform/devlog/devlog.mjs as its target. That file builds the site's development analyses page from the experiments log, the process backlog and the run records. redteam.py campaign seeded 8 faults, one at a time, ran the suite after each and restored the file. The suite caught 4.
| Line | Fault | Why it got through |
|---|---|---|
| 112 | || to && in the doc sort | no test had two docs on the same day, or an undated one |
| 167 | per-day track count starts at 1 | no test read by_day[...].tracks |
| 189 | b.count - a.count to + | the most-reported file was also the first one tallied |
| 189 | || to && in the same sort | no test had a tie on count |
The ranking gap is worth a sentence. The old test's top file happened to be the first one tallied, so a comparator that always returned "keep the order" still passed. The new test tallies two once-reported files first, then a twice-reported one, and expects [m.md 2, b.md 1, z.md 1]. All four faults now fail a test (redteam.py rerun: 8 of 8).
Then the question behind the gaps
Who runs this suite? No workflow does. portfolio-pages.yml runs devlog/build.mjs to build the page but never node --test on it. The repository already had a tool for this exact hole, platform/ci/unwired_checks.py, which runs every platform/ci check and Python suite that no pull-request workflow names. It didn't look at node suites anywhere else in platform/.
Matching names wasn't good enough. Workflows say node --test "*.test.mjs" with working-directory: platform/pages written after the run: line, or node --test platform/privacy/*.test.mjs, or cd platform/edge && node --test e.test.mjs. The new code finds each node --test line, takes the step it sits in, reads that step's working directory and any cd dir && in front of it, and expands the arguments as globs from there. A suite counts as run only if one of those globs reaches its file.
Of the 157 *.test.mjs files under platform/, 123 are run by some workflow. The other 34 are run by none: 1 in devlog, 27 in distribution, 6 in proposal.
One of them was red
platform/proposal/term-ids.test.mjs checks that every bid calculator with a deposit-and-terms panel carries those terms into its share link and its saved draft. It failed on locksmith-bid-calculator (save).
The page was fine. Earlier today PR #15774 changed the page's save from saveBid(localStorage, …) to saveBid(pageStorage(), …), so that a browser with site data blocked no longer throws. The guard's regex had the literal localStorage written into it. It went red the moment that PR merged, and nothing ran it. The regex now accepts any storage handle.
This is the kind of red that does damage slowly. A guard nobody runs still reads as protection to the next person who opens the file.
What now runs on its own
unwired_checks.py runs on every venture pull request through run_venture_tests.py --since. It now runs those 34 suites too, six at a time. That adds about 30 seconds, most of it android.test.mjs (34 s alone). distribution/ios.test.mjs is a named skip because one of its tests plans the whole catalogue and takes over 120 s. When a workflow starts running one of these suites, it drops out of the set without anyone editing a list. On its first run the new check caught term-ids.test.mjs.
The last gap is still open: a PR that touches only platform/ doesn't trigger ventures-tests.yml. That workflow file is control plane, so the fix needs the owner's approval. It is already in the red-team open list (2026-11-13).