Issue: Claude Code Unusable for Complex Engineering Tasks Since February Update

Data analysis results have shown that Claude Code's thinking token redaction update degraded the model's ability to perform complex engineering. Analysis of 17,871 thinking blocks and 234,760 tool call records confirmed that the redaction measure, which began in mid-February, is directly linked to the model's reasoning quality.

Key findings:

• Sharp decline in Thinking Depth: The median thinking length, which was about 2,200 characters before the redaction of thinking content was introduced, decreased by 73% to about 600 characters after the redaction measure.
• Change in tool usage patterns: The model shifted from a 'Research-first' approach, in which it sufficiently reads and modifies code, to an 'Edit-first' approach, attempting edits without investigation.
• Deterioration in quality and user experience:
  • Cases of 'Stop hook' violations, which detect premature termination and ownership avoidance, surged from 0 to over 10 per day on average.
  • Frustration indicators in user prompts increased by 68%.
  • The number of prompts per session decreased by 22%, reducing task continuity.

In conclusion, the data proves that the model's extended thinking is not merely an auxiliary feature, but a structurally essential element for multi-step research, convention adherence, and precise code modification.