Streaming partial messages for UI integration.

[jackgerrits]

We need to ensure the streaming use case is adequately supported in the API. While agent-agent communication likely does not need streaming, we need to ensure agent-UI supports streaming.

User can:

• Create a UI which can show partial model output
• Interrupt an in progress streaming response

[ekzhu]

We should use a closure agent to collect the partial messages and pump them through to the front end via a websocket.

The partial message type is application defined.

We can show case this using a cookbook or a sample directory.

[ekzhu]

Related to #3970 and #3969, a simple solution is to just have the agent and team return a streaming iterator.

[SuMiaoALi]

Related to #3970 and #3969, a simple solution is to just have the agent and team return a streaming iterator.与 #3970 和 #3969 相关，一个简单的解决方案是让代理和团队返回一个流式迭代器。

I totally agree!
Agent token streaming output is very important.
In some cases, streaming output may not be needed, such as the interaction between tool_caller and tool_executor;
But when interacting with user agents, it is very necessary. LLM output is streaming and it will be time-consuming. We need to let users see the first token output as soon as possible.
I think the fastest way to achieve this is that any Agent that interacts with UserProxyAgent should return an iterator or generator instead of full chat_messages
So that we can implement streaming output to the UI ourselves.
Looking forward to it soon. Thanks