Python: Bug: when terminate a function(tool) call in an Agent, the Groupchat cannot catch functioncallresult in chathistory

[HarryWhy]

Describe the bug
by using

@kernel.filter(FilterTypes.AUTO_FUNCTION_INVOCATION)
async def auto_function_invocation_filter(context: AutoFunctionInvocationContext,next) -> None:
  await next(context)
  context.terminate = True

to terminate a function call just after it gets result.

it terminated after tool gets result, so the 3rd step is not triggered and the Agent finish its job, but when the next agent B is invoked(be selected by group manager as next participant). checking on the chat history(context), the agent A response content is '[]' - an empty list. seems the agent doesn't use tool result as its response in this case.

To Reproduce
running a multi-agent GroupChatOrchestration() in my project.

when an agent A in the groupchat is invoked, the agent will:
1st step: call GPT to suggest a tool;
2nd step: tool response with it result to agent; then as @kernel.filter(FilterTypes.AUTO_FUNCTION_INVOCATION).. terminated the subsequence process, Agent A is job is done.
3rd step: the next agent B is invoked by group manager, check on agent B's messages that send to GPT, Agent A 's response content is '[]'- an empty list.

Expected behavior
now it is running like this:
when an agent A in the groupchat is invoked, the agent will:
1st step: call GPT to suggest a tool;
2nd step: tool response with it result to agent;
3rd step: Agent uses tool's result and sends a subsequence request to GTP to generate the agent (final)response.

what we want is:
we want the 3rd step be ignored and tool result can be used as the Agent's response and can be seen , used by other agent in the rest of agents' invoke in this group chat.

the reason is :
because the 2nd step we are retrieving lots of structure data, we just want the agent to use tool's result as the agent (final)response.(in fact we added instruction to agent prompt to let him do nothing but just use tool's result). but now the 3rd step costs lots of tokens and seconds but just bring tool's result in and return tool's result back(as we request him to do like this).

Platform

• Language: Python
• Source: semantic-kernel==1.35.0
• AI model: OpenAI:GPT-4.1]
• IDE:VS Code
• OS: Windows

Additional context
Add any other context about the problem here.

[moonbox3]

@HarryWhy do you have some minimal sample code we can use to repro it? I want to make sure we're looking at the same thing, and we don't make assumptions. Thanks.

[HarryWhy]

@HarryWhy do you have some minimal sample code we can use to repro it? I want to make sure we're looking at the same thing, and we don't make assumptions. Thanks.

@moonbox3 sure! here is a sample code:
sample.txt

the sample code will terminate the math agent's subsequence tool call/llm call after times function is done,
and then i check final Assistant gpt completions call from here:

i can see that now it is final Assistant be invoked and math assistant content is an empty []:

and if i turn off the terminate:

then i can see the math assistant content:

[TaoChenOSU]

Hi @HarryWhy,

Thank you for you question!

When you terminate the auto function invocation loop, the agent does not produce a final response. In a normal flow, the agent will use the function call result to generate a response.

In your scenario, it sounds like you want Agent A to execute the tool but Agent B to produce a response. Is there any reason why you don't want Agent A to give a response before Agent B takes over?

[HarryWhy]

Hi @HarryWhy,

Thank you for you question!

When you terminate the auto function invocation loop, the agent does not produce a final response. In a normal flow, the agent will use the function call result to generate a response.

In your scenario, it sounds like you want Agent A to execute the tool but Agent B to produce a response. Is there any reason why you don't want Agent A to give a response before Agent B takes over?

hi @TaoChenOSU , actually that's not the case. we just want the agent can be stopped once tool calling is done and use tool's result as the agent response. now it is not working like that.
i attached a sample code you can have a try, thank you!

[jerschi]

I'm running into the same problem, but am not using a GroupChatAgent.
Due to latency, I want my ChatCompletionAgent to just return the result of the tool call, vs. going over the result and then returning that exact result.
I'm using semantic-kernel==1.35.0 right now.
Took this from the samples and adapted it to my use case:

@kernel.filter(FilterTypes.AUTO_FUNCTION_INVOCATION)
async def auto_function_invocation_filter(context: AutoFunctionInvocationContext, next):
    """A filter that will be called for each function call in the response."""

    print("\nAuto function invocation filter")
    print(f"Function: {context.function.name}")
    print(f"Request sequence: {context.request_sequence_index}")
    print(f"Function sequence: {context.function_sequence_index}")

    function_calls = context.chat_history.messages[-1].items
    print(f"Number of function calls: {len(function_calls)}")

    await next(context)

    if context.function.plugin_name in ("APlugin", "BPlugin"):
        print("Returning plugin result directly")
        print(f"Plugin name: {context.function.plugin_name}")

        context.terminate = True

The termination works, but it returns an empty string. For debugging I printed the context.function_result right before termination and it has the result I am expecting.