Adversarial tester

A green suite written by the same agent that wrote the code proves little: the same blind spots wrote both. This skill hands the code to a second agent with a clean context, whose job is to make green fail. Many real problems cannot be made into a failing test (a race that needs exact timing, a case the criteria never decided, a design that breaks the next change). The agent reports those too, labeled REASONED, so a human can check them.

How to run it