An update on recent Claude Code quality reports
Summary
Recent reports of performance degradation in Claude Code were traced to bugs within the tool's execution harness rather than the underlying LLM models. A logic error in session management caused the tool to lose context during active interactions, leading to perceived decreases in model quality.
Key Points
- Three separate issues within the Claude Code harness were identified as the cause of recent quality reports.
- A deployment on March 26, 2026, introduced a bug affecting session state management.
- The bug was part of a feature intended to clear "thinking" data from sessions idle for over one hour to reduce latency.
- The error caused the clearing mechanism to trigger on every turn for the remainder of the session after the initial hour of inactivity.
- The resulting behavior caused the model to exhibit increased forgetfulness and repetitive outputs.
Technical Details
The degradation was caused by a failure in the Claude Code harness's session management logic. The update was designed to optimize latency by purging older "thinking" tokens from sessions that had been inactive for more than 60 minutes. Instead of performing a single purge upon session resumption, the bug caused the purge to execute continuously on every subsequent turn. This prevented the model from maintaining a stable context window, effectively stripping away necessary reasoning history during active use.
Impact / Why It Matters
This incident demonstrates that the reliability of agentic systems is heavily dependent on the stability of the execution harness, not just the underlying model. Developers must implement robust monitoring for state management and session persistence to prevent harness-level bugs from mimicking model degradation.