Invalid JSON payload: “text” at system_instruction.parts[0] when using Gemini model via @openai/agents-extensions

[Codewithsaffy]

Describe the bug

When creating an Agent with the instructions field and using the Gemini model (gemini-2.0-flash) via @openai/agents-extensions (which wraps @ai-sdk/google), the system prompt is serialized incorrectly. Instead of sending each part’s text as a string, the SDK wraps it in an array. The Gemini endpoint then rejects the request with:

APICallError [AI_APICallError]: Invalid JSON payload received. Unknown name "text" at 'system_instruction.parts[0]': Proto field is not repeating, cannot start list.

---

Debug information

• Agents SDK version: v0.0.2

• @openai/agents-extensions version: v0.0.2

• @ai-sdk/google version: v1.2.19

• Runtime environment:

  • Node.js v22.12.0
  • OS: Windows 10 Pro (build 2004)

---

Repro steps

1. Create a new folder and run:

   npm init -y

2. Install these dependencies:

   npm install @openai/agents @openai/agents-extensions @ai-sdk/google dotenv

3. Create a file named main.ts (or main.js) with the following content:

   import { Agent, run } from '@openai/agents';
   import { google } from '@ai-sdk/google';
   import { aisdk } from '@openai/agents-extensions';
   import { setTracingDisabled } from '@openai/agents';
   import dotenv from 'dotenv';
   
   dotenv.config();
   setTracingDisabled(true);
   
   // Initialize Gemini model via Vercel AI SDK
   const model = aisdk(google('gemini-2.0-flash'));
   
   // Pass an `instructions` string to the Agent constructor
   const agent = new Agent({
     name: 'My Agent',
     instructions: 'You are a helpful assistant.',
     model,
   });
   
   (async () => {
     const result = await run(agent, 'What is the capital of France?');
     console.log(result.finalOutput);
   })();

4. Create a .env file and set your Gemini API key (e.g. GOOGLE_API_KEY=<your_key>).

5. Run the script:

   node main.ts

6. Observe the error:

   APICallError [AI_APICallError]: Invalid JSON payload received. Unknown name "text" at 'system_instruction.parts[0]': Proto field is not repeating, cannot start list.
       at …/provider-utils/src/response-handler.ts:59:16
   …
   requestBodyValues:
     contents: [{ role: 'user', parts: [{ text: 'What is the capital of France?' }] }]
     systemInstruction: { parts: [{ text: [ [Object] ] }] }
   …
   statusCode: 400
   

---

Expected behavior

The SDK should serialize the instructions string into a single-element parts array where parts[0].text is a plain string, for example:

"systemInstruction": {
  "parts": [
    { "text": "You are a helpful assistant." }
  ]
}

With this correct format, Gemini’s API would accept the request and return a valid response.

[prashantbhudwal]

const client = new OpenAI({
apiKey: process.env.GOOGLE_GENERATIVE_AI_API_KEY,
baseURL: "https://generativelanguage.googleapis.com/v1beta/openai/",
});

setDefaultOpenAIClient(client);
setOpenAIAPI("chat_completions");

[prashantbhudwal]

The issue is agent sdk defualts to responses api, gemini works with openai only with chat completions api. The text is new param in responses api.

[seratch]

Hi @prashantbhudwal, thanks for this really helpful bug report! I've confirmed the issue and am working on its fix.

[Codewithsaffy]

thanks @prashantbhudwal