Reasoning token extractor for o1, o3, DeepSeek-R1, and Claude Extended Thinking. Separates internal scratchpad reasoning from client-facing output and detects cognitive self-corrections.