Desktop app: Claude uses unrequested MCP fallback; claude.ai connectors drop and stay stuck on 'pending'

Environment: Claude desktop app (Code tab), Windows 11, model Opus 5. Working on a WordPress / WooCommerce / Dokan multi-vendor staging site through a claude.ai custom connector.

What went wrong

1. Out-of-scope commentary. While fixing an image bug, Claude called another vendor's media on the site "a mistake" and commented on content the user hadn't asked about.
2. Unrequested fallback to a different MCP server. When the site's claude.ai connector got stuck on pending, Claude tried to call an unrelated MCP server (list_clients on a separate user-configured server) as a workaround, without asking. The user rejected the call.
3. Connector reliability. In the middle of the session, every claude.ai connector (the site connector, Google Drive, Context7, Claude Docs, visualize) disappeared at once. After being re-enabled they came back but stayed pending indefinitely. reconnect_session_connector returned "This session has no MCP server by that name" even though the connector was listed by name and ID, so a simple footer edit was blocked.

Expected

• Stay within the requested task and don't editorialize about unrelated site data.
• Never switch to a different MCP server as a fallback without explicit user approval.
• claude.ai connectors shouldn't silently drop from a live session, and a pending connector should be recoverable (reconnect should work, or report a clear error).