Costing an Uncapped Tool Result

A research agent's average run is 20 steps. On roughly 1 run in 5, a search_documents call matches an unusually large set and returns about 6,000 lines of text (assume ~15 tokens per line). The oversized result typically lands around step 6, so it is re-encoded in the prompt for the remaining ~14 steps. Input tokens cost $3 per million. The team runs about 500 of these tasks a day.

1. Compute the extra input-token cost caused by one oversized result, and the daily cost across all affected runs.
2. Compute what changes if the harness caps tool results at 2,000 tokens with a pointer to the full result on disk.
3. An engineer objects that capping "hides information from the model and will hurt quality." Give the strongest version of that objection and then the design that answers it.