Integrating a Vercel AI SDK Agent with Dunetrace

Quick Start

npm install dunetrace ai

import { Dunetrace, wrapGenerateText } from "dunetrace";
import { generateText } from "ai";
import { openai } from "@ai-sdk/openai";

const dt = new Dunetrace();   // local dev, no API key needed
const instrumentedGenerateText = wrapGenerateText(generateText);   // patch once, at startup

await dt.run("my-agent", { model: "gpt-4o" }, async (run) => {
  const result = await instrumentedGenerateText({ model: openai("gpt-4o"), prompt: "Hi" });
  run.finalAnswer();
});

await dt.shutdown();

What this does

wrapGenerateText/wrapStreamText hook into the AI SDK's step lifecycle (onStepEnd) — every LLM call and tool call inside a dt.run() context is captured automatically, including multi-step tool loops. Your own onStepStart/onStepEnd/onEnd callbacks are preserved and still run.

Streaming

wrapStreamText works the same way — events fire as the stream is consumed:

const instrumentedStreamText = wrapStreamText(streamText);
const result = instrumentedStreamText({ model: openai("gpt-4o"), prompt });
for await (const chunk of result.textStream) {
  process.stdout.write(chunk);
}

Verification

Run your agent, then open the dashboard at http://localhost:3000 — the run should appear within ~15 seconds. To trigger a detector signal without a real LLM call:

SCENARIO=failures python packages/sdk-py/examples/decorator_agent.py

Advanced (optional)

Manual option injection

If you'd rather not wrap the imports, merge Dunetrace's callbacks into your own options:

import { instrumentGenerateTextOptions } from "dunetrace";
const result = await generateText(instrumentGenerateTextOptions({ model: openai("gpt-4o"), prompt }));

Full run wrapper

traceGenerateText/traceStreamText open and close the run for you — useful for a single call with no other run-scoped logic:

import { traceGenerateText } from "dunetrace";
const result = await traceGenerateText(dt, "my-agent", { userInput: prompt }, generateText, {
  model: openai("gpt-4o"), prompt,
});

traceStreamText drains the stream internally before closing the run — use it when you only need the final result.text/result.usage. For incremental streaming to a client, use wrapStreamText inside an explicit dt.run() instead.

Next.js App Router

export async function POST(req: Request) {
  const { prompt } = await req.json();
  const result = await dt.run("chat-api", { model: "gpt-4o", userInput: prompt }, async () =>
    instrumentedStreamText({ model: openai("gpt-4o"), prompt })
  );
  return result.toTextStreamResponse();
}

Troubleshooting

• No events in the dashboard — confirm the wrapped call happens inside dt.run() (or use traceGenerateText); call await dt.shutdown() before exit
• No events from a streamed run — onStepEnd only fires once the stream is consumed; make sure you read result.textStream
• Type errors — install a matching ai peer dependency (npm install ai@^7)