Calibrate the model against real token counts

A document's token count is the prompt with the document minus the same prompt without it. Claude Code adds its own context to every call, and caches most of it, so the one decision that matters is reading all three input fields (input_tokens, cache_creation_input_tokens, cache_read_input_tokens) and subtracting a baseline, rather than trusting input_tokens, which on a warm session is a handful of tokens. Everything runs on your Claude subscription through claude -p; there is no API key involved.