(v5) Double answers / Multiple tool-calls issue.

[pavlo-tkhir]

Hi,

I experience a weird issue where sometimes LMM calls the same tool multiple times. It also sometimes replies, calls the tool, and then replies again. So the user gets double replies. The replies are a bit different, so I can't write a workaround to exclude them from rendering. Is there anything I'm doing wrong or can do differently to prevent this behavior?

These multiple tool calls are within one AI answer message.

{
    "id": "aitxt-8Ui2pZdJ5DQTbSSC4IjxxP3S",
    "role": "assistant",
    "parts": [
      {
        "type": "step-start"
      },
      {
        "text": "This image is a concise guide to \"Good Strategy,\" stating that it's more than just a plan. According to the graphic, a good strategy involves three key steps:\n\n1.  **Diagnosing the problem:** This implies understanding the core issue that needs to be addressed.\n2.  **Defining guiding principles:** These are the foundational ideas or rules that will direct the approach.\n3.  **Taking coherent action:** This refers to implementing actions that are consistent with the diagnosis and guiding principles.\n\nThe image is labeled \"Strategic Thinking 101,\" suggesting it's a basic lesson on the topic.\n",
        "type": "text",
        "state": "done"
      },
      {
        "type": "tool-generateChatTitle",
        "input": {
          "title": "Good Strategy Explained"
        },
        "state": "output-available",
        "output": {
          "title": "Good Strategy Explained",
          "success": true
        },
        "toolCallId": "dhjpnmJ5JrhSNmkR"
      },
      {
        "type": "step-start"
      },
      {
        "text": "This image is an educational graphic titled \"Good Strategy ≠ Just a Plan\". It outlines three essential components for a good strategy:\n\n1.  **Diagnose the problem** (with a target emoji 🎯)\n2.  **Define guiding principles** (with a compass emoji 🧭)\n3.  **Take coherent action** (with a rocket emoji 🚀)\n\nThe bottom of the image also includes the text \"Strategic Thinking 101\". It serves as a brief guide to the fundamental elements of effective strategic thinking.",
        "type": "text",
        "state": "done"
      }
    ]
  }

And here is my backend code:

@Post("message")
async chat(@Body() model: ChatMessageModel, @Res() res: Response) {
  // ... extract old messages from the DB

  const stream = createUIMessageStream({
    execute({ writer }) {
      const result = streamText({
        model: google("gemini-2.5-flash"),
        system: "You have to do the following:
                  - Call the generateChatTitle tool to update the title. Current chat title is '${chatSession.name}'. This is the most accurate title from the database. If you see any tool calls to update the title with a different values, ignore them.
                  - Reply to the user.
        ",
        messages: convertToModelMessages(uiChatMessages),
        stopWhen: stepCountIs(5),
        tools: {
          generateChatTitle: tool({
            description: "Generate a concise, descriptive title for the chat session based on the conversation.",
            inputSchema: z.object({
              title: z.string().describe("A concise title (3-6 words) that describes the main topic or purpose of the conversation"),
            }),
            execute: async ({ title }) => {
              // .. update title in the DB ..
              return { success: true, title };
            },
          }),
        },
      });
      writer.merge(
        result.toUIMessageStream({
          sendReasoning: true,
          messageMetadata: messageMetadataProvider(llmModel, startTime),
        })
      );
    },
    onFinish: (data) => {
      // store response message in the DB
    },
  });
  pipeUIMessageStreamToResponse({ response: res, stream });
}

I have nest.js backend

• "@ai-sdk/google": "^2.0.0-beta.6"
• "ai": "^5.0.0-beta.4"

and React frontend

• "@ai-sdk/react": "^2.0.0-beta.4",

Thanks!

[lgrammel]

Can you create an issue? thank you!

[khanakia]

I believe it's related to this #7169

[pavlo-tkhir]

I have created a bug report here #7261

The relation to #7169 is unlikely (although not excluded completely) because it does not happen every time. Maybe once in 20-30 executions. Also, this is not a UI related issue. The response message contains an array of duplicated responses already when it hits the onFinish callback, and it is all saved into the database.

[lgrammel]

Please add originalMessages to either createUIMessageStream or toUIMessageStream (depending on what you are using on the outer level) so that the correct message id is set