.Net: Bug: Unexpected Accumulation of Previous Assistant Messages in Final Response Using InvokeStreamingAsync

[nels-ai-source]

Describe the bug
When using the InvokeStreamingAsync method of ChatCompletionAgent, and receiving multi-turn conversation responses in the OnIntermediateMessage, it was found that the final assistant message included previous assistant messages. It appears that the last message accumulates all previous assistant messages. I am not sure if this is a bug, but it has caused confusion during usage.

To Reproduce
Steps to reproduce the behavior:

Define a ChatAgent with provided tools.

ChatCompletionAgent agent = new()
{
    Instructions = processRequest.AgentRequest.Instructions,
    Kernel = processRequest.Kernel,
    HistoryReducer = new ChatHistoryTruncationReducer(6, 4),
    Arguments = new KernelArguments(new OpenAIPromptExecutionSettings
    {
        Temperature = 0,
        FunctionChoiceBehavior = FunctionChoiceBehavior.Auto(functions: processRequest.AgentRequest.Functions, autoInvoke: processRequest.AgentRequest.Functions?.Count > 0, options: new() { RetainArgumentTypes = true })
    })
};

Receive messages in OnIntermediateMessage and notice overlapping data.
Expected behavior
Each message should be independent and non-repetitive.

Platform

Language: C#
Source: 1.56.0
IDE: Visual Studio
OS: Windows

[nels-ai-source]

@crickman
In the ChatCompletionAgent.InternalInvokeStreamingAsync method, the StringBuilder builder accumulates all message content throughout the entire streaming response process. When the method concludes, builder.ToString() is used to create the final assistant message, causing the last message to contain the content of all previous assistant messages.

Problem Analysis:

1. At line 425, each streaming message is appended to the same StringBuilder: builder.Append(message.ToString())
2. At line 442, the entire accumulated content is sent as a new message: new(role ?? AuthorRole.Assistant, builder.ToString())

history.json

[crickman]

@westey-m - Can you review what's going on here and see if it aligns to your intent. I'm not 100% on all of the considerations here.

[westey-m]

@nels-ai-source, the intent of OnIntermediateMessage is to be able to subscribe to complete messages. This is separate to the streaming responses that you would receive as part of the main InvokeStreamingAsync response. If you consume both you will receive the response text as individual updates via the InvokeStreamingAsync response and as a complete message via OnIntermediateMessage and this is by design.

Is this causing you issues? What problem are you trying to solve?

[nels-ai-source]

Hi @westey-m ,

Thank you very much for your detailed and patient explanation — it was very helpful in understanding the design intent behind OnIntermediateMessage and how it interacts with the streaming response.

I’ve had some time to investigate further, and I’d like to share an observation regarding the final message in the chat history. After the agent.InvokeStreamingAsync call completes, I retrieve the latest message with the role assistant from the thread.ChatHistory. However, I noticed that this final message includes all previous assistant messages, effectively concatenating them into one.

I’ve attached a screenshot for clarity. In the image, you can see that:

The final message in the history contains not just the most recent assistant response,
But also includes earlier assistant messages from the same conversation.
This behavior leads to duplicated or overlapping content, which can be confusing during usage, especially when building conversation history or persisting messages.

I wanted to confirm whether this is by design or if it might be an unintended side effect of how the messages are collected and stored in the chat history.

Thank you again for your time and support. I truly appreciate it.

Best regards,