stdio MCP servers should auto-reconnect like HTTP/SSE transports

Problem

When a stdio-based MCP server (e.g. Playwright MCP) disconnects, Claude Code does not attempt automatic reconnection. Users must manually run /mcp to reconnect. This happens frequently — roughly every 10-20 minutes on macOS — making stdio MCP servers unreliable for longer sessions.

HTTP/SSE/WebSocket transports have exponential backoff reconnection (5 attempts, 1s-16s), but stdio is explicitly excluded from this logic.

Root Cause

In the minified cli.js (v2.1.92), line 6078:

if(y!=="stdio"&&y!=="sdk")

This condition skips automatic reconnection for stdio transport. Reconnecting a stdio server is just re-spawning the child process — functionally identical to what /mcp does manually.

Proposed Fix

Change:

if(y!=="stdio"&&y!=="sdk")

To:

if(y!=="sdk")

This is a one-line change that gives stdio servers the same auto-reconnect behavior that HTTP/SSE/WebSocket servers already have.

Environment

• Claude Code 2.1.92 (Bun standalone binary, macOS arm64)
• MCP server: @playwright/mcp@0.0.70 (stdio transport)
• macOS 15.4 (Darwin 25.4.0)

Related

• #43177
• #37482 (stdin pipe loss on macOS)