LLM

A tool's own breakdown of its prompt can show no change for a trim that removed 5,095 tokens. The input-token count in the API's usage field is the number to budget with, and a fixed minimal prompt is how to read it cleanly.