vercel/ai

Description

Describe the Bug

When a ToolMessage arrives via LangGraph's messages mode without a preceding AIMessageChunk (e.g. from handleChainEnd with the nostream tag), the adapter emits tool-output-available BEFORE tool-input-start. This violates the UI message stream contract and can cause the frontend's useChat hook to crash or render incorrectly.

To Reproduce

import { toUIMessageStream } from "@ai-sdk/langchain";
import { createUIMessageStream } from "ai";
import { ToolMessage } from "@langchain/core/messages";

// Simulate a bare ToolMessage in messages mode (no preceding AIMessageChunk)
async function* source() {
    yield [
        "messages",
        [
            new ToolMessage({
                tool_call_id: "call_1",
                content: "3 products found",
                name: "searchProducts",
                id: "tool-1",
            }),
            { langgraph_node: "executor" },
        ],
    ];
}

const uiStream = toUIMessageStream(source());
// ... collect chunks

Observed chunk order:

1. tool-output-available (call_1)
2. tool-input-start (call_1)   ← should come first

Expected chunk order:

1. tool-input-start (call_1)
2. tool-output-available (call_1)