Improve robustness when handling "thinking" content by using a dedicated helper to extract the thinking text. This ensures wrapped or nested thinking objects are handled correctly instead of relying on a direct string extraction, reducing parsing errors for complex payloads.