web_search_preview tool call is not limited by max_turns option

[jihun-im-open]

Please read this first

Yes

Describe the bug

max_turns of run_streamed() is ignored for web_search_preview tool

Debug information

• Agents SDK version: v0.0.15
• Python version: Python 3.11

Repro steps

https://gist.github.com/jihun-im-open/b1351fb44320f83d72d50ab9f98a50da

Expected behavior

MaxTurn Error should be raised

[rm-openai]

All the web_search_preview calls happen on the server in one "turn", so this is expected. Is your goal to have at most 1 call to search?

[jihun-im-open]

I have encountered more than 6 sets of the following web_search_call events and those are regarded as a single turn:

• response.web_search_call.in_progress
• response.web_search_call.searching
• response.web_search_call.completed

Here is full response stream:
https://gist.github.com/jihun-im-open/2468aca698f6dd34b226156008fa23bd

My intention is that it would be great if max_turns could limit the count of web_search_call.xxx events. This is because web_search_call could also fall into an almost infinite loop.

[rm-openai]

Gotcha. We can't do this today, because max_turns is a local (SDK) param but the web searches happen in the API. An alternate approach is to set a max_tokens instead, which will force the model to use fewer tokens and hence fewer web searches?

[jihun-im-open]

Thanks @rm-openai ,

I see that we cannot control the API.

I solved this problem by counting ResponseWebSearchCallInProgressEvent events and canceling RunResultStreaming if a threshold is exceeded.

[rm-openai]

Makes sense. I'll also look into adding more configurabilty to the API. thanks for the feedback.