Custom fetch and headers options not forwarded to AI SDK providers

[iozfiliz]

Environment Information

Stagehand:

• Language/SDK: TypeScript
• Stagehand version: 3.0.3 (also affects 3.0.1, 3.0.2)

AI Provider:

• Provider: OpenAI (also affects Anthropic, Google, and all AI SDK providers)
• Model: Any model using the "provider/model" format (e.g., openai/gpt-4o-mini)

Issue Description

When providing custom fetch and headers options in ClientOptions, they are not forwarded to the underlying AI SDK provider. This prevents important use cases like:

• Proxy authentication - Adding custom auth headers (e.g., X-LLM-Request-Token)
• Request logging - Intercepting requests for monitoring/debugging
• Custom retry logic - Implementing application-specific retry behavior
• Rate limiting - Custom rate limiting implementations

The options are accepted by Stagehand's TypeScript interface but silently ignored during execution, which is misleading to users.

Steps to Reproduce

1. Create a Stagehand instance with custom fetch and headers
2. Use an AI SDK provider model (e.g., openai/gpt-4o-mini)
3. Make an act() or extract() call
4. Observe that custom fetch function is never called

Minimal Reproduction Code

import { Stagehand } from '@browserbase/stagehand';

// Custom fetch function
const customFetch = async (url: string, options?: RequestInit) => {
  console.log('✅ Custom fetch called:', url);
  const headers = new Headers(options?.headers);
  headers.set('X-Custom-Header', 'my-value');
  return fetch(url, { ...options, headers });
};

// Initialize Stagehand with custom fetch
const stagehand = new Stagehand({
  model: {
    modelName: 'openai/gpt-4o-mini',
    apiKey: process.env.OPENAI_API_KEY,
    fetch: customFetch,  // ← This is ignored!
    headers: {
      'X-Custom-Header': 'test'  // ← This is also ignored!
    }
  }
});

await stagehand.init();
await stagehand.act('Click the login button');

// Expected: "✅ Custom fetch called: ..." logged
// Actual: Nothing logged (default fetch used instead)

Error Messages / Log trace

No error is thrown. The options are silently ignored, which makes debugging difficult.

Root Cause

File: packages/core/lib/v3/llm/LLMProvider.ts

The getAISDKLanguageModel() function (lines 96-128) only accepts 4 parameters and doesn't forward fetch and headers:

// Current (buggy) implementation
export function getAISDKLanguageModel(
  subProvider: string,
  subModelName: string,
  apiKey?: string,
  baseURL?: string,  // ← Only 4 parameters!
) {
  const providerConfig: { apiKey: string; baseURL?: string } = { apiKey };
  if (baseURL) {
    providerConfig.baseURL = baseURL;
  }
  // ❌ fetch and headers never added to providerConfig
  const provider = creator(providerConfig);
  return provider(subModelName);
}

Call site (lines 146-151) also doesn't pass these options:

const languageModel = getAISDKLanguageModel(
  subProvider,
  subModelName,
  clientOptions?.apiKey,
  clientOptions?.baseURL,
  // ❌ MISSING: clientOptions?.headers
  // ❌ MISSING: clientOptions?.fetch
);

Expected Behavior

Custom fetch and headers options should be forwarded to the AI SDK provider, enabling:

• Custom authentication via fetch interceptors
• Request logging and monitoring
• Proxy integration with custom headers
• Custom retry and error handling logic

Additional Context

• ClientOptions type already includes fetch and headers (defined as OpenAIClientOptions | AnthropicClientOptions)
• All underlying AI SDK providers (OpenAI, Anthropic, Google, etc.) support these options
• The baseURL option IS forwarded correctly (inconsistent behavior)
• This is a documented feature of the Vercel AI SDK that Stagehand should expose

Use Case: We need this for LLM proxy authentication where all requests must include custom authentication headers. Currently using a runtime patch as a workaround.

Related Issues

None found (searched for "fetch", "headers", "proxy")

---

Note: I have a PR ready with the fix and test case.