Agentchat refactor

[ekzhu]

1. Make events internal to group chat and log the agent messages directly with simplified implementation. Basic usage case will just use run_stream to see the stream of messages rather than relying on logs to get agent chats. Decoupling presentation logics from execution.

2. Move message thread to team from group chat manager, so the team can maintain the message thread state for resume.

# Create team
terminate = TokenUsageTermination(...)
team = Swarm(..., termination_condition=terminate)

# First run
result1 = await team.run("task ...")

# Reset termination condition and run again with the same team and agent state.
await terminate.reset()
result2 = await team.run("task ...")

3. Expose termination stop reason in task result

result = await team.run("some task")
print(result.stop_reason)

Text 'TERMINATE' mentioned

4. Expose inner messages to termination conditions, so condition on inner messages, such as token usage, can be applied correctly.

Next step:

1. document resume of team by calling run again.
2. inject agent & team state through built-in memory module Memory Interface for AgentChat agents #4039

[victordibia]

Good first implementation.
+1 on returning stop_message.
Useful for applications to keep track of reasons the task ended. (similar to finish_reason in the oai api)