The suite mocks the generated client with MagicMock, which answers every
attribute whatever the real response looks like, so a client that stops
building nested models passes unit tests and only fails against the live
backend. Build the responses through the installed client's own models
instead: Session.toolkits() and mcp.create() then fail locally when
item.meta, item.connected_account.auth_config, or commands arrive as
bare mappings.
Claude-Session: https://claude.ai/code/session_018wNe8qK1kGJFoe8s1iDtpF