Event metadata

Your server speaks AG-UI. TanStack useChat still misses finishReason, the model id, or leftover usage. Put those extras in metadata.tanstack on the event. Spec fields stay at the top.

Copy this RUN_FINISHED event first:

{
  "type": "RUN_FINISHED",
  "threadId": "thread-1",
  "runId": "run-1",
  "usage": [
    {
      "inputTokens": 12,
      "outputTokens": 34,
      "totalTokens": 46
    }
  ],
  "metadata": {
    "tanstack": {
      "finishReason": "stop",
      "model": "gpt-5.5"
    }
  }
}

{
  "type": "RUN_FINISHED",
  "threadId": "thread-1",
  "runId": "run-1",
  "usage": [
    {
      "inputTokens": 12,
      "outputTokens": 34,
      "totalTokens": 46
    }
  ],
  "metadata": {
    "tanstack": {
      "finishReason": "stop",
      "model": "gpt-5.5"
    }
  }
}

The client copies metadata.tanstack onto the chunk. After chunk.type === "RUN_FINISHED", read chunk.metadata?.tanstack?.finishReason. In-process usage is TanStack TokenUsage (promptTokens). The wire uses the spec array (inputTokens).

Do now

Send these on every successful run:

finishReason is one of "stop", "length", "content_filter", "tool_calls", or null.

If you omit finishReason after a tool result, the client treats it as not "stop" and can send another turn.

Read it on the client

import { useChat, fetchServerSentEvents } from "@tanstack/ai-react";

const { messages } = useChat({
  connection: fetchServerSentEvents("/api/chat"),
  onChunk: (chunk) => {
    if (chunk.type === "RUN_FINISHED") {
      console.log(chunk.usage);
      console.log(chunk.metadata?.tanstack?.finishReason);
      console.log(chunk.metadata?.tanstack?.model);
    }
  },
});

import { useChat, fetchServerSentEvents } from "@tanstack/ai-react";

const { messages } = useChat({
  connection: fetchServerSentEvents("/api/chat"),
  onChunk: (chunk) => {
    if (chunk.type === "RUN_FINISHED") {
      console.log(chunk.usage);
      console.log(chunk.metadata?.tanstack?.finishReason);
      console.log(chunk.metadata?.tanstack?.model);
    }
  },
});

Do not import a helper. Check chunk.type, then read chunk.metadata?.tanstack.

RUN_ERROR

AG-UI RUN_ERROR has message and optional code at the top. Put correlation ids in metadata.tanstack:

{
  "type": "RUN_ERROR",
  "message": "Provider timeout",
  "code": "TIMEOUT",
  "metadata": {
    "tanstack": {
      "threadId": "thread-1",
      "runId": "run-1",
      "model": "gpt-5.5"
    }
  }
}

{
  "type": "RUN_ERROR",
  "message": "Provider timeout",
  "code": "TIMEOUT",
  "metadata": {
    "tanstack": {
      "threadId": "thread-1",
      "runId": "run-1",
      "model": "gpt-5.5"
    }
  }
}

Rate limit? Put the wait in metadata.tanstack.retryAfterMs, in milliseconds. Clients read it from there. See Rate limits.

Later

Add these when you use the matching feature:

• Leftover usage. Spec usage[] also accepts cachedInputTokens and reasoningTokens. Put cost and other leftover fields in metadata.tanstack.usage. The client rebuilds TanStack TokenUsage (promptTokens) from the array plus that leftover.
• Interrupt errors. On RUN_ERROR, set metadata.tanstack.interruptErrors so ChatClient can match a failed interrupt submit.
• Tool output error. On TOOL_CALL_RESULT, set metadata.tanstack.state to "output-error" when the tool result is an error payload.
• Message stamps. On wire messages, metadata.tanstack.createdAt is an ISO-8601 string.
• Signatures. Stream thinking signatures on REASONING_ENCRYPTED_VALUE. On the next-turn body, set spec encryptedValue on role: "reasoning" messages and on toolCalls. Other tool-call provider fields stay in metadata.tanstack.toolCallMetadata. See Thinking & Reasoning.

A TanStack chat() server already writes this bag. Use this page when you emit AG-UI events yourself.

See Stream Events for the event table, and AG-UI Client Compliance for the request body.