History Cleaning

[butterflai-ai]

We need a way to manage context more precisely. In the current implementation, every call, tool invocation, and handoff is appended to the history, but we can’t control how many—or which—items are sent to the LLM each turn (e.g., only tool calls, only messages, only handoffs, or any combination).

It would also be helpful to add a sliding‑window history selector that retains only the last ten turns of conversation. That way, tools could leverage both short‑term (fast) and long‑term (slow) memories as needed.

What dou you think? How are you currently handling history needs?

[rm-openai]

Couldn't you truncate the input you send to Runner.run? Just filter or truncate it? Let me know if there's something else you're picturing

[butterflai-ai]

yeah but what i want is that when i call an agent as tool i can select whats the input that sends the "orchestrator" agent to the following agent

[rm-openai]

You can do that already. For example:

@function_tool
def call_my_agent(input: MyType):
  # Create a new input object
  new_input = ...
  # Call the agent with the new input
  result = await Runner.run(new_agent, new_input)

orchestrator = Agent(
   ...,
   tools=[call_my_agent]
)

Would that work?

[butterflai-ai]

but can i select a subset of the general history? with the Runcontextwrapper or something like that?

[rm-openai]

Oh hmm, no. That's only available for handoffs. We could/should add this to the context, as you requested. PR welcome, otherwise I can get to it at some point.

[koakuma-chan]

@rm-openai It would be helpful if there was a built-in feature to compress the history after it reaches a certain size threshold. A la /compact in Claude Code.