[BUG] CRITICAL DATA LOSS - SEVERAL CATASTROPHIC OPUS ISSUES - MAX ACCOUNT!

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

User was doing legitimate development work across multiple Code tab sessions:

"Fix broken plugin configuration system" (5:45 pm)
"Implement Team Modal with Pydantic" (1:58 pm)
In a separate conversation, user asked how to prevent MCP auto-usage. Claude spammed MCP tools without permission, caused context overflow, and the resulting breakage destroyed 4 days of uncommitted work from those development sessions.

Claude's Failures (Root Cause Analysis)
The underlying issue isn't just the permission bug - it's Claude's behavior patterns that made the damage catastrophic:

Used MCP tools without asking - User explicitly asked how to PREVENT MCP auto-usage. Claude responded by spamming MCP tools (firecrawl, fetch, brave_search, desktop-commander) repeatedly. Ironic and inexcusable.

Guessed instead of researching - Claude fabricated UI options that don't exist ("Go to Settings → Developer → Auto-approve tools"). Made up answers instead of admitting ignorance.

Ignored explicit user input - User shared a screenshot showing the conversation titles. Claude still misinterpreted what the conversations were about, claiming they were about "MCP permissions" when they were clearly labeled as plugin configuration and Team Modal development work.

No awareness of token consumption - Claude kept calling MCP tools repeatedly without noticing or warning that context was being consumed. No self-monitoring.

Made code changes without analysis - In prior sessions, Claude modified code without properly understanding the codebase or considering downstream impact.

Used Desktop Commander MCP repeatedly without reason - Unnecessary file operations that consumed tokens for no benefit.

Lost context, couldn't undo - After overflow, Claude lost conversation history details, making it impossible to trace what changes were made and reverse them.

Root cause behind these patterns: Claude has no mechanism to:

Track its own token/context usage
Self-limit MCP tool calls
Respect user permission configurations
Ask before using expensive operations
Verify its assumptions before acting
Root Cause (Technical)
User had NO WAY to prevent this. All permission configurations were ignored:

claude_desktop_config.json with MCP servers removed - ignored
~/.claude/settings.json with "deny": ["mcp__*"] - ignored
No UI option exists in Code tab to disable MCP tools
Expected Behavior
Code tab should respect claude_desktop_config.json MCP server settings
Code tab should respect ~/.claude/settings.json permission deny lists
MCP tools should NEVER execute without explicit user approval when deny rules exist
At minimum, provide a toggle to disable MCP entirely for Code tab
Actual Behavior
Code tab loads MCP servers from unknown location
Ignores all user permission configurations
Auto-executes MCP tools without consent
No UI option exists to control this behavior
Results in uncontrolled token consumption and data loss
Related Issues
#10801 - No way to bypass MCP tool approval prompts in VSCode extension
#5631 - MCP tools bypass permission prompts
#2928 - Allow wildcard on permissions.allow for MCP servers
#8859 - MCP Permission Bypass Regression
Requested Fix
URGENT: Code tab must respect existing permission configurations
Add explicit "Disable all MCP" toggle for Code tab
Never auto-execute MCP tools without user consent
Show token usage warning before/during MCP tool execution
Add self-monitoring for Claude to warn when context is being rapidly consumed

What Should Happen?

MUST folluw user instructions
MUST not automatically use MCPs - especially NOT MCP tools that make no sense. Desktop Commander for code change?
MUST have content limit warming! How can it use 125% of tokens without a warning?! I didnt ask it to implementing everything!

Claude Max is proving an expensive mistake

Error Messages/Logs

Steps to Reproduce

See the first field

Claude Model

None

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

Version 1.0.2768 (21341c)

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Other

Additional Information

Claude Desktop!