This repository's red-team track seeds faults into code and counts how many the tests catch. When a run finds a real defect it cannot fix in the same pull request, it writes the defect into its run record, and prompts/redteam.py open lists every such entry still unpatched. The red-team prompt tells each new run to fix that list first.
On 2026-11-15 the list held 12 entries, and three red-team runs that day wrote the same complaint into their records: each had checked all 12 by hand and found nothing to do. Every entry was a change to CI workflows or prompt tooling. Those paths need the owner's approval (they are listed in .github/CODEOWNERS), and each fix was already sitting in an open pull request waiting for it.
What changed
platform/ci/owner_held_prs.py, which lists the open pull requests that wait on the owner, now takes --redteam. For each open entry it names the held pull requests that carry the fix. A PR counts as a match if the entry names it (#15945) or if the PR's added lines contain a file or identifier the entry mentions (check_csv_formula_guard.py, current_approvers). Entries with no match are flagged as the ones a red-team run can actually fix. The tool reads the open list through redteam.py's own function, so it never disagrees with it. It also leaves that owner-gated file untouched, which let this change merge without waiting in the same queue.
Run against the live repository it printed:
12 open red-team entries (prompts/redteam.py open); 12 have a fix waiting on
the owner in a held PR, 0 have none and are a red-team run's to fix.
For all 12 entries, the pull request written for that entry was among the matches. I checked each one against the PR titles by hand. Four of the workflow-wiring entries also matched #15871, which wires four of the same checks into one workflow at once, and #14808 and #14877 both add the same insecure-reference check. Those are real overlaps for the owner to settle, not false matches.
The queue behind it
The list was full because the approval queue is long. The same tool counted 61 open same-repository pull requests that touch an owner-gated path. 36 of them are over a week old, 12 are over two weeks old, and the oldest is 38 days old. 22 of the 61 change nothing else, so they are pure control-plane edits waiting on one review. That queue, not the red team's diligence, is why the open list stops shrinking. The new flag at least makes sure a run notices in seconds instead of spending ten minutes finding out.