Repro: /v1/responses returns invalid output_text for Google models (missing text field on tool calls)

README.md

How to run

npm init -y && npm install ai @ai-sdk/openai zod tsx
POE_API_KEY=your-key npx tsx poe-repro.ts

Bug

/v1/responses returns invalid output_text for Google models when making tool calls:

{"type": "output_text", "annotations": []}

Missing required text field. Should be:

{"type": "output_text", "text": "", "annotations": []}

Also, top-level text is {} instead of {"format": {"type": "text"}}.

Affects: gemini-3.1-pro, gemini-3-flash, gemini-2.5-pro, gemini-2.5-flash, gemini-2.5-flash-lite, gemini-2.0-flash. Not affected: OpenAI models (gpt-5.2, o3, etc.) work correctly.

poe-repro.ts

import { generateText, tool } from "ai";
import { createOpenAI } from "@ai-sdk/openai";
import { z } from "zod";

const poe = createOpenAI({
  apiKey: process.env.POE_API_KEY,
  baseURL: "https://api.poe.com/v1",
});

// Works fine
const ok = await generateText({
  model: poe.responses("gemini-3.1-pro"),
  prompt: "Say hello in 3 words",
});
console.log("text:", ok.text);

// Fails: AI_TypeValidationError — expected string, received undefined
// at output[0].content[0].text
const fail = await generateText({
  model: poe.responses("gemini-3.1-pro"),
  prompt: "What is the weather in SF? Use the getWeather tool.",
  tools: {
    getWeather: tool({
      description: "Get weather",
      parameters: z.object({ location: z.string() }),
    }),
  },
});
console.log("tools:", fail.toolCalls.length);

stdin

## How to run

```bash
npm init -y && npm install ai @ai-sdk/openai zod
POE_API_KEY=your-key npx tsx poe-repro.ts
```

## Bug

`/v1/responses` returns invalid `output_text` for Google models when making tool calls:

```json
{"type": "output_text", "annotations": []}
```

Missing required `text` field. Should be:

```json
{"type": "output_text", "text": "", "annotations": []}
```

Also, top-level `text` is `{}` instead of `{"format": {"type": "text"}}`.

Affects: gemini-3.1-pro, gemini-3-flash, gemini-2.5-pro, gemini-2.5-flash, gemini-2.5-flash-lite, gemini-2.0-flash.
Not affected: OpenAI models (gpt-5.2, o3, etc.) work correctly.