Possible to truncate tool call outputs before sending back to model?

[JoeBerg8]

I'm using the chat completions API and the version 0.6.1 of the agents sdk.

I want to know if it's possible to truncate a ToolCallOutputItem before the agents sdk sends it back to the model. The motivation here is to manage context token budgets. Occasionally, some of my data retrieval tools will return large amount of data and this can get me close to the context limit of a model over the course of a long conversation It would be great if I could truncate tool call outputs before the tool outputs get sent to the model. I'm aware of summarization techniques but that's not always what I want to do - sometimes a simple truncation would be better for my use case.

I'm happy to take a stab at a PR for this somewhere in the ToolCallOutputItem class in agents/items.py if it's not currently possible.

Thanks in advance!

[BrodaNoel]

Why don't you just truncate the data you return on your tool?

[JoeBerg8]

@BrodaNoel well because doing that isn't necessarily the best approach. I don't want to truncate if i don't need to, and if I need to truncate, I want to have control over what gets truncated. I don't want to apply the same truncation to each tool call. Some control over how/when tool outputs get truncated would be a better solution

[seratch]

The simplest approach would be to have additional logic before returning a large output data on the tool side, but if you need something else, using "on_tool_end" hooks may be helpful: https://openai.github.io/openai-agents-python/ref/lifecycle/#agents.lifecycle.RunHooksBase.on_tool_end

[JoeBerg8]

this looks interesting, thanks @seratch