Add McpReconnect tool so the model can reconnect failed MCP servers

Problem

When an MCP server fails mid-session (Playwright is a frequent offender), the model detects the failure but has no way to fix it. It has to ask the user to manually run /mcp and reconnect, which breaks flow and is frustrating.

The user can reconnect via /mcp in the same session and it works fine. The model just has no way to trigger this itself.

Proposed solution

Add a tool (e.g. McpReconnect) that lets the model programmatically reconnect a failed MCP server by name. This would allow the model to transparently recover from MCP failures without interrupting the user.

Alternatively, an auto-reconnect opt-in setting per MCP server would also work, though the tool approach is more flexible.

Why the workarounds don't work

• Hooks (PostToolUseFailure): Can restart the underlying service (e.g. Chrome for Playwright) but can't reconnect the MCP protocol layer in Claude Code.
• Windows MCP / desktop automation: The model is occupying the terminal while responding, so it can't type /mcp into its own input. Even if it could, /mcp is an interactive menu.
• The model has no access to slash commands or UI actions: Only tools are available.

Context

This is especially painful for MCP servers that connect to external processes (Playwright via CDP to Chrome, etc.) where transient failures are common.