Why results are capped in size

Every tool's results are capped in size before being handed back to the model. A search that would technically return thousands of matches is truncated, with an explicit note that truncation happened, rather than flooding the rest of the turn with more than could usefully be reasoned about.

This exists to protect the quality of everything that happens afterward in the same turn: an overly broad, uncapped result would crowd out the model's ability to hold the actually relevant parts of the conversation in mind for the rest of its reasoning.

In practice this rarely changes the outcome of a search; if the first attempt is too broad to be useful, the model narrows it on the next step the same way a person would after noticing the same thing.

The truncation note itself is informative on its own: seeing that a search was cut off at its cap is often the first real signal that a search term was too generic, a strong prompt to narrow it rather than something to work around.

The cap applies per tool call, not across an entire turn; a turn that legitimately needs several separate searches is not penalized for the number of searches it makes, only for any single one of them returning an excessively broad result.