[Bug]: chat_messages + agent.run() issues

[alecsolder]

Describe the bug

When creating an agent and specifying chat_messages, those messages are only sent to the LLM if the sender matches:

ag2/autogen/agentchat/conversable_agent.py

Line 1298 in 681d551

reply = self.generate_reply(messages=self.chat_messages[sender], sender=sender)

This is hard when using .run() because it creates the other agent with

ag2/autogen/agentchat/conversable_agent.py

Line 3333 in 681d551

with self._create_executor(

which would make it impossible to get those messages included.

Not sure exactly how to solve this, but the end state is that there should just be a way to add in messages to the history that will be sent to the LLM.

Steps to reproduce

No response

Model Used

No response

Expected Behavior

No response

Screenshots and logs

No response

Additional Information

No response

[rjambrecic]

Hi @alecsolder ,

if I understood correctly you want to be able to do:

assistant = AssistantAgent(
    name="Assistant",
    llm_config=llm_config,
)

chat_result = assistant.run(
    message="what is two plus two?",
    max_turns=1,
    user_input=False,
)

user (to Assistant):

what is two plus two?

--------------------------------------------------------------------------------
Assistant (to user):

Two plus two equals four.

--------------------------------------------------------------------------------

And then send another message but with preserved chat history:

chat_result = assistant.run(
    message="What was my last question?",
    max_turns=1,
    user_input=False,
)

user (to Assistant):

What was my last question?

--------------------------------------------------------------------------------
Assistant (to user):

Your last question was, "what is two plus two?"
--------------------------------------------------------------------------------

Is that correct?