What is a token budget and why is it mandatory?

Every retrieval call to the daemon takes a hard token_budget, and the daemon trims the result to fit. There is no unbudgeted retrieval path.

Why mandatory? Because the failure mode it prevents is the most common one in agent systems: a retrieval tool that returns "everything relevant" and quietly floods the context window, degrading the model's reasoning and inflating cost. With budgets, the ceiling is explicit per call: the caller decides what a lookup is worth and the daemon honours it exactly.

The economics hold up under measurement: on the ScoreCrux benchmark, budgeted retrieval returns far fewer tokens than naive top-K, with answer quality essentially unchanged.

Practical guidance: keep budgets tight by default and raise them deliberately for calls that genuinely need depth. A budget you raised on purpose is a decision; a missing budget is an outage waiting for scale.