[Cross-Ref] MCP Timeout & Reconnection — Reliability Gap in All CLIs

[howardpen9]

Pattern

MCP (Model Context Protocol) server connections are unreliable across all CLIs — timeouts, disconnections, and failed reconnections frustrate users.

Occurrences

Claude Code

• #31425 — MCP tool search fails (regression in 2.1.70)
• #31412 — OAuth token not persisted between sessions
• #31427 — Feature request: async/background MCP calls

OpenCode

• #16338 — UnsupportedFunctionalityError with MCP
• #16318 — Kimi K2.5 reasoning_content missing via MCP

Kimi CLI

• #1316 — MCP timeout causes entire CLI to be unavailable
• #1296 — Intermittent error from disconnected MCP

Analysis

MCP is still a young protocol. Common failure modes:

1. Stdio servers crash silently → no reconnection attempt
2. HTTP servers timeout → blocks the entire tool call pipeline
3. OAuth tokens expire → no automatic refresh
4. Server initialization races with tool discovery

Recommended Solution

• Implement MCP health checks with exponential backoff reconnection
• Add per-server timeout configuration (some servers are inherently slow)
• Queue tool calls during reconnection instead of failing immediately
• Persist auth tokens and implement refresh flows

Contribution Target

Kimi CLI — Fix #1316 with timeout configuration and graceful reconnection.