[FEATURE] Allow hooks to trigger MCP server reconnection

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

SessionStart hooks can start/restart MCP servers and output system messages, but they cannot programmatically trigger MCP reconnection. After a hook restarts an MCP server (e.g., as part of an environment switch workflow), the user must manually run /mcp → Reconnect before tools become available again.

This creates a multi-step manual workflow where it could be seamless:

1. A skill restarts an MCP server with new configuration
2. The user runs /clear (triggering the SessionStart hook)
3. The hook detects the restarted server
4. The user must manually run /mcp → Reconnect ← this step should be automatable

Proposed Solution

Allow the SessionStart hook's JSON response to include a directive that triggers MCP reconnection for specified servers. For example:

{
  "systemMessage": "MCP server restarted with new config.",
  "hookSpecificOutput": {
    "hookEventName": "SessionStart",
    "reconnectMcpServers": ["snowflake"]
  }
}

Claude Code would process this by automatically re-initializing the MCP session for the named servers, the same way /mcp → Reconnect does today.

Alternative Solutions

• Document the manual step clearly: This is the current workaround. The hook outputs a message telling the user to run /mcp. Works but adds friction.
• Auto-reconnect on connection failure: Claude Code could detect a broken MCP connection on the next tool call and automatically reconnect. This would help but is reactive rather than proactive, and the first tool call after a restart would still fail.

Priority

Low - Nice to have

Feature Category

MCP server integration

Use Case Example

Workflows that restart MCP servers mid-session — such as switching between environments, rotating credentials, or changing server configuration — currently require the user to remember and execute /mcp → Reconnect after every switch. This is error-prone (users forget and get confusing tool errors) and adds friction to what should be a single-action workflow.

Additional Context

_No response_