- 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
