[BUG] Claude Code gives completely unrelated hallucinated responses to simple prompts (e.g., "1+1" returns "4+5=9")

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?

Claude Code v2.1.76 (Opus 4.6) consistently returns completely unrelated, hallucinated responses — even to trivial prompts like "can you do 1+1?"

Examples from fresh sessions:

• "can you do 1+1?" → "4 + 5 = 9"
• "can you do 1+1?" → Analysis of anonymous-origins.py and hierarchical.py Python files
• Previous session: "1+1" → "5 + 5 = 10", then switched to Chinese, then explained Python's def keyword
• Asking to read a Slack URL (with Slack MCP connected) → hallucinated it was a Launchpad Ubuntu package page, then a GitLab CI page, then Xfce desktop panel docs

The model never uses available MCP tools (e.g., Slack) even when explicitly told to, and fabricates entirely unrelated content instead.

Environment

• Claude Code v2.1.76
• Model: Opus 4.6 (1M context)
• macOS
• Issue persists across fresh sessions, new terminals, and after clearing projects cache

What I've Tried

• Fresh terminal session
• claude update (already on latest)
• Removed custom plugins from ~/.claude/settings.json
• rm -rf ~/.claude/projects
• None of the above resolved the issue

What Should Happen?

Claude should respond correctly to the user's prompt. For "can you do 1+1?", it should answer "2". For a Slack URL with the Slack MCP connected, it should use the Slack tool to read the message — not hallucinate unrelated content.

Error Messages/Logs

No error messages — the model returns a 200 response with confident but completely 
wrong/unrelated content. Examples:

Prompt: "can you do 1+1?"
Response: "I'll solve this step-by-step. 4 + 5 = 9. The answer is 9."

Prompt: "can you do 1+1?"
Response: "anonymous-origins.py and hierarchical.py seem to have significantly 
different structures based on the snippets, so I'll analyze their key structural 
elements..."

Prompt: "Use the Slack tool to read this message and explain it: [slack URL]"
Response: "This is the Launchpad URL associated with the Snap page for the 
cyrus-imapd-2.4 package on a system running Ubuntu..."

Steps to Reproduce

1. Open a new terminal, run claude
2. Type "can you do 1+1?"
3. Observe unrelated hallucinated response

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

Claude Code v2.1.76

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_