Simplify AgentChat: run_stream().to_console()

[ekzhu]

Have the return value of run_stream() directly calls itself with to_console() to avoid importing Console.

[aguynamedben]

I wish run_stream() had some sort of callback or hook for handling messages. I want to be able to (a) print pretty things to the console, (b) send messages via WebSockets, but right now I have to chose whether to wrap my stream in Console() or not. I'm ending up copying the implementation of Console() to my own version so I can both print to Console in a pretty way and also send to WebSockets.

It seems Console and/or run_stream() could have some sort of callback API that makes this easier?

Pseudo code, but maybe something like this?

stream = team.prepare_stream(task="make the logo big")
stream.add_message_handler(my_handler_func)
stream.run()

[aguynamedben]

Another way to address this could be to break the handy type checking and console display logic currently in _console.py into a system of helper functions. Then a user could

• Create an AutoGen stream
• Use the add_message_handler idea above with something like RxPY to create a "Subject" (Rx term) with type Subject[AgentEvent | ChatMessage | TaskResult]
• Subscriber A: Subscribe to subject, relay messages to console print helpers (i.e. to see what's happening in dev)
• Subscriber B: Subscribe to subject, relay messages to WebSockets (i.e. to interface with other systems)

The weirdest part about handling a stream is the type handling, and a lot of valuable type handling AND console printing is happenign inside _console.py. Maybe decomposing those roles would be valuable.

[ekzhu]

@aguynamedben thanks for the input. The current Console and stream handling is certainly much to be desired.

I created a new issue here: #6182 as a way to customize message handling and allow for reuse of handlers that are already in Console. Let us know if you would like to build it.