.NET: [Bug]: AG-UI host generates as many text messages as completions

[tlecomte]

Description

When using the AG-UI hosting package, and an OpenAI model like GPT 5.4 mini, the AG-UI event stream contains as many as TEXT_MESSAGE_START/_CONTENT/_END as completions, causing an AG-UI client to display completions one at a time, each on its own line.

Code Sample

I believe this can be reproduced with https://github.com/microsoft/agent-framework/tree/main/dotnet/samples/05-end-to-end/AGUIWebChat

Error Messages / Stack Traces

Package Versions

Microsoft.Agents.AI.Hosting.AGUI.AspNetCore 1.6.1-preview.260514.1

.NET Version

.NET 10

Additional Context

No response

[github-actions]

🤖 Automated triage reproduction notes (agent-authored — trust but verify)

Agent analysis

Repro: ChatResponseUpdateAGUIExtensions.AsAGUIEventStreamAsync() in dotnet/src/Microsoft.Agents.AI.AGUI/Shared/ChatResponseUpdateAGUIExtensions.cs (lines 466-506) emits a new TEXT_MESSAGE_START/END block every time ChatResponseUpdate.MessageId changes. When OpenAI assigns different MessageIds per completion in a tool-calling loop (FunctionInvokingChatClient), each completion's text becomes a separate AG-UI text message. Minimal repro: feed AsAGUIEventStreamAsync a stream of ChatResponseUpdate objects with TextContent where consecutive groups have different MessageIds (simulating multiple LM completions), and observe that multiple TextMessageStartEvent/TextMessageEndEvent pairs are emitted instead of one.

• Failing test: dotnet/tests/Microsoft.Agents.AI.Hosting.AGUI.AspNetCore.UnitTests/Issue6010_MultipleTextMessagesPerCompletionTests.cs
• Files examined: dotnet/src/Microsoft.Agents.AI.AGUI/Shared/ChatResponseUpdateAGUIExtensions.cs, dotnet/src/Microsoft.Agents.AI.Hosting.AGUI.AspNetCore/AGUIEndpointRouteBuilderExtensions.cs, dotnet/src/Microsoft.Agents.AI.AGUI/AGUIChatClient.cs, dotnet/src/Microsoft.Agents.AI.AGUI/AGUIHttpService.cs, dotnet/src/Microsoft.Agents.AI/ChatClient/ChatClientAgent.cs, dotnet/tests/Microsoft.Agents.AI.Hosting.AGUI.AspNetCore.UnitTests/ChatResponseUpdateAGUIExtensionsTests.cs, dotnet/tests/Microsoft.Agents.AI.Hosting.AGUI.AspNetCore.IntegrationTests/BasicStreamingTests.cs
• Tests run: Issue6010_MultipleTextMessagesPerCompletionTests.AsAGUIEventStreamAsync_MultipleCompletions_WithToolCallsBetween_ProducesTooManyTextMessages, Issue6010_MultipleTextMessagesPerCompletionTests.AsAGUIEventStreamAsync_TwoCompletionsWithDifferentMessageIds_ProducesTwoTextMessages
• Reported version: 1.6.1-preview.260514.1

[javiercn]

AG-UI text-message boundaries are defined by messageId, not by the overall user turn or the number of model invocations.

The .NET adapter preserves the identity supplied by MEAI:

ChatResponseUpdate.MessageId
    → TextMessageStartEvent.MessageId

During a function-invocation loop, the model can produce multiple distinct assistant messages:

assistant: “I’ll check that.”
assistant: tool call
tool: result
assistant: “Here is the answer.”

Providers commonly assign different message IDs to the assistant responses produced by separate model completions. When those IDs differ, the correct AG-UI representation is a separate text-message lifecycle for each message:

TEXT_MESSAGE_START / CONTENT / END
TEXT_MESSAGE_START / CONTENT / END

AG-UI does not define a separate logical-turn identifier requiring those provider messages to be combined into one frontend message.

Automatically coalescing them in the server adapter would remove provider message identity and could incorrectly merge content across:

• tool-call boundaries;
• different agent authors;
• distinct assistant messages;
• replay/history boundaries.

Applications or frontend components may choose to visually group related assistant messages, but that is presentation policy rather than protocol conversion behavior.

The current adapter therefore behaves as intended by emitting one AG-UI text message for each distinct MEAI message ID. No framework or AG-UI SDK change is planned, and this issue can be closed as by design.