MCP SSE client hangs indefinitely on dropped connection

Description

When using an MCP server over SSE transport (e.g., via Tailscale to a remote host), if the network connection drops mid-session, Claude Code hangs indefinitely when it next tries to call an MCP tool. The user must press Esc to cancel and manually reconnect the MCP server.

Steps to Reproduce

1. Configure an MCP server with SSE transport (e.g., claude mcp add mem0-local --transport sse http://remote-host:8050/sse)
2. Start a conversation and use MCP tools successfully
3. Cause a network interruption (Tailscale hiccup, Wi-Fi drop, server restart, etc.)
4. Try to use an MCP tool again — Claude Code hangs with no timeout or error

Expected Behavior

• MCP tool call should timeout after a reasonable period (e.g., 30-60s)
• Claude Code should display an error like "MCP server unreachable" and offer to reconnect
• Ideally: automatic reconnection attempt before failing

Actual Behavior

• Claude Code blocks indefinitely waiting for a response on the dead SSE connection
• No timeout, no error message
• User must press Esc to cancel, then manually reconnect the MCP (/mcp → reconnect)

Root Cause Analysis

• TCP keepalive defaults on macOS are ~2 hours, so the OS doesn't detect the dead connection quickly
• The MCP SSE client doesn't implement any request-level timeout
• No SSE heartbeat/keepalive detection on the client side
• Server-side SSE keepalive pings can help the server detect dead clients, but don't fix the client hanging

Environment

• Claude Code CLI (latest)
• macOS (Sequoia 15.3)
• MCP server: SSE transport over Tailscale VPN
• Reproducible with any SSE MCP server when network drops

Suggested Fix

1. Add a configurable timeout to MCP tool calls over SSE (default ~30-60s)
2. Detect stale SSE connections (e.g., expect periodic keepalive events)
3. Auto-reconnect on connection failure before returning error to user