What broke when we ran agents on small local models

Five real failures from running Bees agents on 3B and 4B models with llama.cpp: a tool-call format that changed mid-run, tools the model was told to use but never shown, a tool cap that killed long runs, a thinking model that outran a timeout, and pr