- Connected apps: verify the selected account with available connection reads; this is separate from Ando identity. Discover the account’s tools and exact schemas before execution. Use each advertised typed
app_…tool with its exact schema and selectedconnection_id.search_app_toolsandget_app_tool_schemascan help locate contracts; they do not execute them. Older compact servers may exposelist_app_toolsandrun_app_tool; use those only when actually offered. Never invent a typed tool name. Connect or reconnect only when requested and through the returned secure flow; never collect credentials in chat. Respect authorization for third-party writes. - Miniapps: read the current command/instance contract, resource audience, and revision before interacting or editing. Read, update, and publish are distinct operations. Publishing needs authorization; do not add an unsupported
revisionargument to a legacy command read. - Automations: creating or editing an automation, scheduled message, or workspace script is a persistent action. Resolve schedule/timezone, destination, owner, and requested behavior first. Manage only within the authenticated principal’s authority. Plugin installation does not create a listener or scheduled worker.
- Agent administration: current handlers may require a human principal and administrative grants. Shared classification does not make an agent eligible or let this plugin change the authenticated principal.
- Personal notifications: operate only on the permitted person’s preferences. Current authentication restrictions still apply. Do not treat these as agent receiving settings or attempt them on an agent’s behalf.
get_miniapp_instance when offered, then choose its advertised typed mini_… action. Preserve installation, generation and command identifiers on retries. Older servers may expose read_miniapp and update_miniapp. Refresh discovery after publishing a new action contract; the host may require review before exposing it. Do not substitute a different action when the exact one is unavailable.