Python: Bug: Chat history summarizer is not working

[pintuiitbhi]

Describe the bug
Chat history summarizer is not working here. It still passes the complete chat history to llm api call.
This is one of the agent. Each agent has its own kernel instance.
Its is used in GroupChatOrchestation as one of agent memeber.

summarization_reducer = ChatHistorySummarizationReducer(
        service=kernel.get_service(),
        target_count=3,
        threshold_count=2,
        auto_reduce=True,
        include_function_content_in_summary=True,
    )
    system_message = """
    Summarize the chat history to reduce its size while retaining key information.
    The summary should be concise, capturing the essence of the conversation without losing important details.
    """

    summarization_reducer.add_system_message(system_message)
    return ChatCompletionAgent(
        kernel=kernel,
        name='agent1',
        prompt_template_config=prompt_config,
        arguments=KernelArguments(
            chat_history=summarization_reducer,
            settings=PromptExecutionSettings(
                temperature=temperature
            ),
        ),
    )

To Reproduce
Steps to reproduce the behavior:
Just create a multi turn agent collaboration using 2 or 3 agents and observer in debug log message of semantic kernel when it is making llm chat completion api call using openai base client.

Expected behavior
It should summarize then chat history should be passed to llm request body

Platform

• Language: Python
• Source: main branch of repository
• AI model: gpt-4o Azure
• IDE: VS Code
• OS: Windows

[moonbox3]

Hi @pintuiitbhi, when using the chat history reducer, there are two paths:

1. the caller needs to explicitly handle reducing the history.
2. the history will be reduced automatically if configured with the bool (auto_reduce), and if messages are added via add_message_async().

Right now, there is no handling in the new orchestration patterns, to allow one to pass in a chat history reducer. We will need to look at how we can allow on to incorporate this.

Tagging @TaoChenOSU for visibility.

[pintuiitbhi]

Got it.

In real-world multi-agent applications, passing the full chat history in every LLM call leads to token overflow issues and inefficiencies. Managing context size—via summarization, truncation, or memory—is essential and should be part of the initial design.

Currently, GroupChatOrchestration sends the entire history with each request, which can impact performance and scalability. Addressing this will be key to broader adoption.

Another issue of similar nature

[TaoChenOSU]

Hi @pintuiitbhi!

You raised a valuable feature request. Currently, the chat history reducer only works on the ChatCompletionAgent, whereas the orchestrations rely on a higher-level abstraction that is the AgentThread. We will see how to make it possible to specify a thread for an agent in an orchestration.