`generateText` with `Output.object()` + `tools` - model returns tool call arguments as text instead of structured `tool_calls`

[egorprnn]

Environment

• @openrouter/ai-sdk-provider: 2.2.3
• ai (Vercel AI SDK): 6.0.79
• Models tested: google/gemini-3-pro-preview, openai/gpt-5.2

Description

When using generateText with both output: Output.object({ schema }) and tools, the model returns tool call arguments as plain text in message.content with finish_reason: "tool_calls", but without an actual tool_calls array in the response.

This happens because Output.object() adds response_format: { type: "json_schema", ... } to every step of the tool loop (including steps where the model should be making tool calls). The model receives conflicting instructions — "respond strictly in this JSON schema" AND "here are tools you can call" — and resolves the conflict by dumping tool arguments into the content field as text while still signaling finish_reason: "tool_calls".

Observed behavior

response.text = '{"email":"abraeng@gmail.com"}'  // tool call args leaked as text
response.finishReason = 'tool-calls'
response.rawFinishReason = 'tool_calls'

No tools are actually executed. The do...while loop in generateText exits immediately because clientToolCalls.length === 0 (no type: "tool-call" parts in the provider response content). The stopWhen callback is never even evaluated.

Expected behavior

The model should either:

1. Return a proper tool_calls array so the AI SDK can parse and execute the tools, OR
2. The provider should not send response_format: { type: "json_schema" } on steps where tools are present (or at least document this incompatibility)

Root cause (traced through provider source)

1. Output.object() sets responseFormat: { type: 'json', schema } on every doGenerate call
2. The provider maps this to response_format: { type: "json_schema", json_schema: { schema, strict: true } } (line ~3077 in dist)
3. Both response_format and tools are sent in the same API request body
4. The upstream model returns tool call args in message.content (text) + finish_reason: "tool_calls", but no message.tool_calls
5. Provider creates a { type: "text" } content part from message.content and maps finish reason to "tool-calls", but no { type: "tool-call" } parts exist
6. AI SDK finds zero tool-call parts → loop exits → tools never execute

Reproduction

import { generateText, Output } from 'ai';
import { createOpenRouter } from '@openrouter/ai-sdk-provider';
import { z } from 'zod';

const openrouter = createOpenRouter({ apiKey: '...' });

const response = await generateText({
  model: openrouter.chat('google/gemini-3-pro-preview'),
  system: 'You are a helpful assistant. Use tools when needed.',
  messages: [{ role: 'user', content: 'Look up the email abraeng@gmail.com' }],
  tools: {
    lookupEmail: {
      description: 'Look up information about an email address',
      parameters: z.object({ email: z.string() }),
      execute: async ({ email }) => ({ found: true, name: 'Test User' }),
    },
  },
  output: Output.object({
    schema: z.object({
      answer: z.string(),
    }),
  }),
  stopWhen: ({ steps }) => {
    const lastStep = steps.at(-1);
    return Boolean(lastStep?.text && lastStep?.finishReason !== 'tool-calls') || steps.length > 10;
  },
});

// Expected: tools execute, response.text contains final JSON answer
// Actual: response.text = '{"email":"abraeng@gmail.com"}', response.finishReason = 'tool-calls'
console.log(response.text, response.finishReason, response.rawFinishReason);