- Replace ThinkingAdapterMaxTokens with a more flexible DefaultMaxTokens map - Add support for model-specific default max tokens configuration - Update relay and web interface to use the new configuration approach - Implement a fallback mechanism for default max tokens