Anthropic Fixes Claude Code Regression, Publishes Post-Mortem
Anthropic has confirmed and resolved three harness changes that degraded Claude Code performance between March 26 and April 10, 2026: (1) default reasoning was downgraded from high to medium; (2) a bug evicted thinking blocks every session turn rather than only those older than one hour; and (3) a system prompt change intended to reduce verbosity also reduced code quality. All three issues are fixed in Claude Code v2.1.116+. Usage limits have been reset for all subscribers. Anthropic published a full post-mortem at anthropic.com/engineering/april-23-postmortem.
Why It Matters
The post-mortem is notable because it quantifies how small harness changes — not model weight changes — caused measurable quality regression at scale. The community response, including an open-harness argument from @Vtrivedy10 on LangChain, signals that harness transparency is becoming a competitive dimension alongside model capability.