Extended Thinking Detection for API Calls

What: Added automatic detection of extended thinking content in conversation history to ensure accurate token counting and proper API parameter configuration.

How it works: When Claude Code needs to count tokens for API calls, it now scans through the message history to detect if any assistant responses contain thinking or redacted_thinking blocks. If detected, it automatically includes the appropriate thinking parameters in the API call.

Details:

• Detects both thinking and redacted_thinking content types
• Automatically adjusts token counting for conversations with extended thinking
• Ensures API calls include proper thinking budget parameters (1024 tokens)
• Improves accuracy when estimating costs for conversations using extended thinking
• Evidence: uS2() at line 376194 (new detection function), enhanced TZ1() at line 376212 and dS2() at line 376234