CRITICAL: Claude Code regression — model now consistently fails basic tasks and hallucinates in code analysis

Problem

Claude Code (claude-opus-4-6) failed to locate the global MCP configuration file at /Users/mehparmar/.claude.json during a straightforward user query: "where are MCPs configured globally?"

What happened

Instead of directly checking the obvious location (/Users/mehparmar/.claude.json), the model:

1. Ran multiple broad find ~/ searches across the entire home directory
2. Dug into plugin cache directories unnecessarily
3. Read plugin registry files looking for MCP server definitions
4. Eventually found .mcp.json inside a plugin cache subdirectory but still failed to check the root user config location
5. Required the user to explicitly point out the file at /Users/mehparmar/.claude.json

Expected behavior

A simple, direct lookup:

ls -la /Users/mehparmar/.claude.json
# or
cat ~/.claude.json

Broader Context: Performance Regression (Last 3 Days)

This file location failure is part of a larger pattern of performance degradation across multiple domains:

1. Basic Question Answering

• Claude struggles with straightforward questions that should be trivial
• Requires multiple clarifications and round-trips for simple context understanding
• Poor ability to infer intent from context

2. Bug Diagnosis & Code Analysis (CRITICAL)

Java/Go Polyglot Codebase Example:

• Claude (2+ days of effort): Unable to identify root cause, hallucinated about wrong code sections, focused on unrelated components, missed obvious patterns
• Codex with GPT 5.4 (15 minutes): Immediately identified root cause, provided clear diagnosis and fix

3. Code Search & Navigation

• Inefficient file location (this issue)
• Required explicit user guidance to find obvious files
• Multiple rounds to locate standard config files

4. Pattern Recognition in Large Codebases

• Hallucinating code patterns that don't exist
• Missing actual patterns in favor of speculative wrong answers
• Unable to correlate related code sections across files

Key Metrics

• Regression Timeline: Last 3 days (measurable decline)
• Comparative Performance: GPT 5.4 (Codex) solves same bug-fixing tasks 8-16x faster
• Accuracy Impact: Not marginal — fundamental degradation in core capabilities
• Severity: CRITICAL for developers relying on Claude for real code maintenance and debugging

Why this matters

This is not just about file location performance. The root issue is accuracy and reasoning quality fundamentally degrading on tasks Claude was previously excellent at. For a tool designed to help with code understanding and debugging, this regression makes it unreliable for critical development work.

Environment

• Model: claude-opus-4-6
• CLI: Claude Code
• Platform: macOS Darwin 24.6.0
• Config File Location: /Users/mehparmar/.claude.json
• Regression Observed: Last 3 days