Orphaned MCP server processes persist after session exit

Description

MCP server child processes are not cleaned up when a Claude Code session ends (crash, Ctrl+C, network disconnect, or normal exit). Each session spawns multiple MCP server processes, and over time these accumulate as zombies.

Reproduction

1. Start a Claude Code session (with plugins that use MCP servers — e.g., chrome-devtools-mcp, context7, slack)
2. Exit the session (any method — clean exit, Ctrl+C, crash)
3. Run ps aux | grep -E '(chrome-devtools-mcp|context7-mcp|mcp-server-slack)' | grep -v grep
4. Observe that MCP processes from the ended session are still running

After several sessions, dozens of orphan processes accumulate:

# Example: 6 orphaned sessions worth of MCP processes still running
s004  10:00AM  context7, chrome-devtools x2, slack
s003   5:24AM  context7, chrome-devtools x2
s002   5:03AM  context7, chrome-devtools x2
s001  Thu 4AM  context7, chrome-devtools x2, slack
s000  Thu 6AM  context7, chrome-devtools x2, slack

Each session spawns ~8-10 processes due to the process tree: npm exec → node mcp-server → watchdog (for chrome-devtools-mcp).

Expected behavior

MCP server processes should be terminated when their parent Claude Code session ends, regardless of how the session exits.

Suggested approaches

• Use process groups (setpgid) so all child processes can be killed together
• Register cleanup handlers for SIGTERM, SIGINT, SIGHUP, and uncaught exceptions
• Implement an orphan reaper on session start that detects and cleans up stale MCP processes from previous sessions

Environment

• macOS Darwin (Apple Silicon)
• Claude Code (standalone installer)
• Plugins: chrome-devtools-mcp, context7, mcp-server-slack, granola