MCP remote server permanently stuck as 'failed/unavailable' even with valid OAuth token #33012

Description

Remote MCP servers (Supabase in this case) get permanently stuck in a status=failed / server unavailable state after a single connection hiccup. The client never retries the connection, and all subsequent tool calls return a misleading -32600 error.

Log evidence from ~/.local/share/opencode/log/opencode.log:

level=WARN run=adb4f769 message="server unavailable" key=supabase type=remote status=failed

Key observations

1. Irreversible failed state — Once the server is marked failed, it never recovers. No retry mechanism.
2. Token was valid the whole time — Manual curl to mcp.supabase.com with the stored OAuth token (sbp_oauth_... from mcp-auth.json) worked perfectly. The token issue was not expired or revoked.
3. Misleading error message — Tool calls return MCP error -32600: You do not have permission to perform this action, which suggests an auth problem. The real issue is the MCP session was never established.
4. Only workaround: nuke everything — Had to delete the Supabase entry from mcp-auth.json, run opencode mcp logout supabase and opencode mcp login supabase again to force a fresh session.

Steps to reproduce

1. Configure a remote MCP server (e.g., Supabase) with OAuth
2. Let the MCP connection fail once (network blip, server restart, rate limit, etc.)
3. Try calling any tool — it fails with -32600
4. The server remains unavailable for the entire session, across tool calls, regardless of token validity

Expected behavior

The MCP client should:

• Retry failed server connections with exponential backoff
• Re-initialize the MCP session when detecting a stale/broken session (Mcp-Session-Id expired/missing)
• Return an accurate error message when the session is not established (not "permission denied")

OpenCode version

1.17.8

OS

Linux (Arch-based)

Terminal

Ghostty