[BUG] v2.1.105: ESC during MCP tool call kills MCP server process (regression from v2.1.104)

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?

Pressing ESC to cancel a long-running MCP tool call in v2.1.105 kills the MCP server process entirely. All tools from
that server become unavailable and it shows as "failed" in the /mcp dialog.

What Should Happen?

ESC should cancel the tool call (mark as "rejected by user") while keeping the MCP server alive and all tools
available — which is exactly how v2.1.104 behaves

Error Messages/Logs

"The following MCP servers have disconnected: meshcode-meshcode-self-improve-backend"                                
  "The following deferred tools are no longer available (their MCP server disconnected)"

Steps to Reproduce

1. Have any MCP server with a long-running async tool (e.g., a tool that awaits for 60+ seconds)
2. Call that tool from Claude Code
3. While the tool is running, press ESC to cancel it
4. The MCP server process is killed — all tools become unavailable
5. /mcp dialog shows the server as failed

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

2.1.104

Claude Code Version

2.1.105

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

Verified with the exact same MCP server code on both versions:

• v2.1.104: ESC cancels tool, server stays alive ✅
• v2.1.105: ESC cancels tool, server killed ❌

Tested with multiple server code variations (with/without signal handlers, with/without atexit hooks, minimal
mcp.run() only) — all crash on v2.1.105, all survive on v2.1.104.

mcp Python package v1.27.0 (also tested v1.26.0 — same result on v2.1.105).