claude-code - 💡(How to fix) Fix Sudden reasoning quality degradation — model behavior change April 9-10, 2026? [3 comments, 3 participants]

Summary

Experienced a sudden, significant degradation in Claude Code reasoning quality and instruction-following between April 9th ~4PM MDT and April 10th, 2026. No client-side changes explain it.

Symptoms

• Simple bug fixes that previously took 1-2 attempts now take 7-10 rounds of failures
• Model stops invoking skills/slash commands (/plan, /bug, /learn) that it previously used consistently
• Model stops using TaskCreate for progress tracking despite system reminders
• Model defaults to general-purpose agents instead of routing to specialists (dispatch rules in context are ignored)
• Model implements directly in main context instead of delegating to subagents
• Overall "reasoning quality" feels noticeably lower — more surface-level responses, less deep analysis

Environment

• Claude Code version: 2.1.81 (installed March 23, 2026)
• Model: claude-opus-4-6 (1M context)
• Platform: Linux (Debian 13, AMD Ryzen 9 5900X)
• No client update occurred between the working and degraded states
• Governance rules, agent definitions, and memory files were stable — the behavioral change was sudden, not gradual

What was ruled out

• Client update: binary unchanged since March 23
• Settings changes: some hooks were added but the degradation pattern (reasoning quality, not just skill invocation) predates those changes
• Context size: governance context is ~57K-76K tokens (~6-13% of 1M budget), well within limits
• Hook performance: all hooks under 200ms, no blocking

Question

Was there a server-side model update or configuration change to claude-opus-4-6 between April 9th and April 10th, 2026 that could explain this behavioral shift? The degradation is consistent across multiple independent sessions and projects.

Impact

This is a production workflow — the governance layer (18 specialized agents, 21 rule files, 15 skills, 52 scripts) was built over weeks and relies on consistent instruction-following. A sudden regression in the model's ability to follow complex system prompts breaks the entire workflow.