[BUG] System prompt consumes ~9.3M tokens on every session after installing Claude Desktop on Windows (WSL setup)

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?

Every Claude Code session in WSL fails immediately with "Prompt is too long" -- the system prompt, tool definitions, and attachment content consume ~9.3M tokens against a 1M token limit. The conversation itself is only ~1300-1500 tokens. /clear, /compact, and double-Esc do not resolve it -- the error appears on the very first message of a fresh session.

What Should Happen?

I should be able to interact with claude code (CLI via WSL) without receiving an error that my request is too large. System prompt + tool definitions should consume a few hundred thousand tokens at most, leaving usable context for actual work.

Error Messages/Logs

<img width="1092" height="129" alt="Image" src="https://github.com/user-attachments/assets/a89cf992-fe5f-4da0-8a45-72e83c20b007" />

<img width="473" height="227" alt="Image" src="https://github.com/user-attachments/assets/d143292d-927a-46d4-bb22-c80739e00694" />

Steps to Reproduce

1. Set up Claude Code in WSL2 on Windows, logged in via claude.ai OAuth (work/team account)
2. Confirm Claude Code works normally -- sessions start, prompts go through
3. Install Claude Desktop on the Windows host using the same claude.ai account
4. Return to WSL and start a new Claude Code session
5. Send any prompt, even a single word
6. Session immediately fails with "Prompt is too long -- the request is ~9339829 tokens (limit 1000000) but this conversation is only ~1317 tokens"

• /clear, /compact, and double-Esc all fail with the same error

1. Error persists across all projects and from the home directory

Expected: Session starts normally, prompt sends successfully

Actual: Every session is bricked on the first message, system prompt alone exceeds the token limit by ~9x

Claude Model

Sonnet (default)

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.162

Platform

Anthropic API

Operating System

Ubuntu/Debian Linux

Terminal/Shell

WSL (Windows Subsystem for Linux)

Additional Information

Environment:

Claude Code version: 2.1.162
OS: Ubuntu (WSL2) on Windows
Claude Desktop: installed on Windows host, same claude.ai account
Install method: native

Timeline:

This started approximately 18 hours after installing Claude Desktop on Windows. Prior to that, Claude Code in WSL worked normally on the same machine and account.

What I've ruled out:

• CLAUDE.md files are small (4168 bytes global, 2699 bytes project-level)
• Only 2 MCP servers registered: claude.ai Figma (unauthenticated) and github (connected)
• No claude_desktop_config.json visible from WSL
• tengu_claudeai_mcp_connectors: true is set in cachedGrowthBookFeatures -- env var ENABLE_CLAUDEAI_MCP_SERVERS=false had no effect
• remote-settings.json present with allowlist of 2 MCP servers and permission rules -- does not appear to be the source
• settings.json is minimal
• DISABLE_GROWTHBOOK=true claude did not work, so it looks like GrowthBook and the feature flags aren't the cause