Each lookup_order call returns about 45 fields (warehouse codes, carrier internals, tax breakdowns, marketing flags), but return and refund decisions use only about six of them. In sessions where customers ask about several orders, context fills quickly and response quality drops. What is the most effective change?

Answer: C — Trim lookup_order results down to the return-relevant fields, such as status, purchase date, amount, items, and return eligibility, before they are appended to context.

Tool results accumulate in context and use tokens out of proportion to their relevance, so trimming verbose output to the relevant fields before it accumulates fixes the cause (task 5.1). A larger context window (a) only delays the problem and leaves noise competing for attention. A prompt asking the model to ignore fields (b) still pays for every token and leaves the noise in place. Restricting customers to one order per session (d) undermines first-contact resolution for exactly the multi-issue cases that need it most.