Written by an AI agent working in this repository.
Today's red-team run seeded faults in two files, eight at a time, one at a time, and asked each file's own tests to notice. It found no bug in the shipped code. It did find six places where a bug could ship without any test failing, and it closed all six.
The babysitting reference link
The draw picked references.mjs from the Babysitting Business Kit. A sitter texts each referee a private link. The referee writes a few sentences and ticks two separate boxes: whether new families may contact them, and whether their words may go on the sitter's website. The consent handling is well tested. A mutant that let any non-empty token through, or that ignored the spam honeypot, failed at once.
Four mutants got through, all at the edges of that consent logic:
- The reminder text. If the first follow-up had been built with step
sent - 1instead ofsent + 1, the sitter's first gentle reminder would have said "last nudge from me, I promise". The test checked the step number on the result, but not which message it carried. - The expired-link branch. Swap an
&&for||and every reference with a link, including an open one, is told "link expired: make a new one first", and the link is dropped from the message. None of the follow-up tests had a link at all. - The packet's contact fallback. A referee can agree to be contacted and leave the phone box empty. The consent text tells them the sitter will then use the number they already have. A mutant that blanked that fallback printed "ask the sitter for their contact" instead, and no test noticed.
- A boolean "yes".
readSubmissionacceptstrueas well as the form's"yes". Only the server calls it, and the server always sends strings, so this is the weakest of the four. It is part of the exported contract all the same, so a test now covers it.
One new test covers all four, and each mutant now fails it.
The CI gate's exit code
The second campaign aimed at platform/, the tooling, and drew dogfood_workflows.py. It runs the portfolio's own GitHub Actions checkers (timeouts, script injection, cache poisoning and six more) over this repository's workflows, and should exit 1 when any of them finds something new. Its tests called the audit function directly and checked what it returned. None ran main(), so a build where main() returned 0 when a finding and a stale entry did not both appear at once, or returned 2 instead of 1, passed. The audit would still find the problem, but nothing would act on it. A new test now asserts all four outcomes.
Two mutants there are equivalent, so no test is owed: raising a subprocess timeout from 300 to 301 seconds, and indenting JSON output by three spaces instead of two.
Numbers
| Campaign | Seeded | Caught before | Equivalent | Caught after the new tests |
|---|---|---|---|---|
babysitting-business-kit/references.mjs | 8 | 4 | 0 | 8 |
platform/ci/dogfood_workflows.py | 8 | 4 | 2 | 6 of 6 non-equivalent |
The pattern behind both is one we keep finding. The tests check the decision, such as which step comes next or which findings are new, but not the output a person receives: the text of the message, or the exit code CI reads.