mcp: no reconnect after network change/socket reset; stale Error until server process is killed #52564

Description

MCP servers that lose their connections (e.g., a network change such as joining a VPN resets their sockets) are never reconnected: the long-lived opencode serve --service server performs zero reconnect attempts when new CLI/TUI clients attach, and the TUI shows Error for those servers indefinitely. The only recovery is killing the server process to force a cold start.

Environment

• opencode version: 2.0.21
• OS: Darwin 25.2.0 (macOS, arm64)
• Terminal: Apple_Terminal, TERM=xterm-256color
• Shell: /bin/zsh
• Install/channel: latest (https://opencode.ai/v2/install)
• Active plugins: one local file plugin auto-loaded from ~/.config/opencode/plugins/
• Architecture note: a long-lived server (opencode serve --service) holds all MCP state; TUI/CLI invocations attach to it (--auto, -s <session>).

Reproduction

1. Have a running opencode serve --service process with MCP servers connected at cold start (server log: mcp connected server=context7 tools=2, plus other remote and uvx-based local servers).
2. Change networks mid-session — in my case, join a VPN. Existing MCP sockets die. Server log shows (server name redacted):

   level=WARN message="mcp http request failed" errors=[{"message":"The socket connection was closed unexpectedly","code":"ECONNRESET"}] server=<remote-mcp-server>
   

3. Quit the TUI and relaunch (opencode -s <sessionID>), or attach new CLIs (opencode --auto). Observed three CLI attach events within one minute.
4. Check MCP status in the TUI.

Expected Behavior

After a connection loss or when a new client attaches, the server retries failed/disconnected MCP servers (with backoff), and the TUI status recovers without manual process kills.

Actual Behavior

• After the socket reset, the log contains no MCP reconnect attempts at all — no mcp connect failed, no retry — across multiple CLI reattach events.
• The TUI shows Error for every server whose connection predated the network change, indefinitely, even though all affected hosts are reachable from the same machine at that moment (verified with curl).
• New sessions attached to the same server do not trigger MCP re-init.
• The only recovery found: kill <serve --service PID>, then relaunch — a fresh server cold-starts all MCP connections successfully.
• Corollary confusion: because the server survives TUI restarts, "I restarted opencode" does not restart MCP state at all.

Additional Context

• Related auth issue: after upgrading v1 → v2.0.21, remote OAuth MCP servers went to needs_auth (requires a login before it can register a client) despite ~/.local/share/opencode/mcp-auth.json holding valid, unexpired access tokens for them (v1-era store). Full re-auth via opencode mcp auth <name> was required per server. A mcp-auth.json.bak exists alongside, suggesting a store migration that didn't carry the tokens forward.
• Servers that (re)connected after the network change (via mcp auth) were healthy; every connection established before the change stayed dead forever — consistent with missing retry/keepalive recovery.
• Reproducible consistently across the network-change scenario; 100% of pre-existing connections were affected.