Paper 4 · Question 33

A: The system prompt at 8k is a small fraction of the budget. Reducing it by 6k would help marginally but misidentifies the problem — conversation history and tool results are consuming the vast majority of the budget.

B: Clearing history would free tokens but destroy conversation continuity. The customer would need to repeat their issue. Selective summarisation is far preferable to wholesale deletion.

C: max_tokens caps the response length but cannot exceed the remaining context budget. If 193k of 200k is consumed by input, increasing max_tokens beyond 7k would have no effect.

D: Token budget is shared: system prompt + conversation history + tool results + available for response = total budget. With 193k consumed, only 7k remains for the response. The fix targets the largest movable allocations: summarise older history or trim verbose tool output.