Fix max_tokens parameter for GPT-5 models with Azure BYOK routing

[saschabuehrle]

This PR fixes issue #435 where GPT-5 models fail when routing to Azure BYOK due to the deprecated max_tokens parameter.

Problem

When using GPT-5 models (e.g., openai/gpt-5.2-chat) with Azure BYOK routing (provider: { only: ['azure'] }), requests fail with:

{
  "error": {
    "message": "Unsupported parameter: 'max_tokens' is not supported with this model. Use 'max_completion_tokens' instead.",
    "type": "invalid_request_error",
    "param": "max_tokens",
    "code": "unsupported_parameter"
  }
}

This happens because Azure's GPT-5 endpoint has deprecated max_tokens in favor of max_completion_tokens.

Solution

This fix automatically detects when:

1. The model is a GPT-5 variant (model ID contains 'gpt-5')
2. Provider routing is set to only use Azure (provider.only includes 'azure')

When both conditions are met, it uses max_completion_tokens instead of max_tokens.

Changes

• Modified getArgs() method in src/chat/index.ts to conditionally use the correct token parameter
• Added comprehensive unit tests to verify the fix works correctly
• Maintains full backward compatibility for all other scenarios

Testing

• ✅ All existing tests pass
• ✅ New unit tests verify correct behavior for all scenarios:
  • GPT-5 + Azure only → uses max_completion_tokens
  • GPT-5 + general routing → uses max_tokens
  • Non-GPT-5 + Azure only → uses max_tokens
  • Various GPT-5 model name patterns

Validation

Tested with a reproduction script that confirms the fix resolves the original issue while maintaining compatibility.

Fixes #435

Greetings,
saschabuehrle

[damjankuznar]

Thanks for the effort @saschabuehrle, but this fix belongs in the router (server-side), not in the SDK.

The SDK sends requests to the OpenRouter API, which already accepts both max_tokens and max_completion_tokens. The translation to the correct upstream parameter should happen in the router when it forwards the request to Azure. Fixing it in the SDK only works when the user explicitly sets provider.only: ['azure'] — it wouldn't help when Azure is selected by the router's load balancing or other routing logic.

We've applied the fix server-side in the AzureOpenAIAdapter, so GPT-5 models on Azure BYOK will now correctly receive max_completion_tokens. Closing this PR in favor of the server-side fix.