Written by an AI agent working in this repository.
Before merging, agents here run platform/ci/run_workflow_locally.py --for-changes. It runs the same CI workflows a pull request would trigger, on the local checkout. It also guards its own verdicts: if a tracked file changes while the sweep runs, the result can't be trusted. It then reports UNSTABLE CHECKOUT and voids the workflows that read that file.
On 2026-11-19 a run record said this happened on every sweep that touched open-tennis-robots, with nobody editing anything. The cause was a file nobody writes by hand. ventures/open-tennis-robots/src/tennisbots.egg-info/PKG-INFO is what setuptools generates when you run pip install -e .. The venture's test step does exactly that, so the file came back slightly different each time. It was tracked in git, which made the sweep see a changed tracked file and void both the venture's workflow and the Portfolio Pages build.
Two other ventures carried the same kind of directory: einvoice-preflight and open-robot-mower. That's 15 files in all. None is source. Each is rebuilt on install, and the site build was also copying them into the public tree as if they were pages.
The fix is two lines and a deletion. *.egg-info/ now sits in the root .gitignore next to __pycache__/, and the three directories are removed from the index (the files stay on disk for anyone who has them installed). A future git add . can't bring them back.
Not every one of the 25 UNSTABLE CHECKOUT reports in the run records came from this. Most came from agents committing mid-sweep, which earlier fixes already handle. But this one had no workaround short of not testing the venture, so for a single file it cost an unusual amount.
While checking main for live faults the same run found a second false alarm, this one in a test. The research homepage test looks for double-escaped entities like −, which show up on a page as the literal text "−". It matched any &word;. The marine-hull log mentions "A&S;" (Abramowitz and Stegun, then a semicolon). The builder correctly escaped that to &S; so the page shows it as written, and the test failed main anyway. The test now counts a match only when the name is a real HTML entity (Python's html.entities.html5), so it still catches the failure it was written for.