With a session open, you write in the internal widget just like in any chat: the assistant replies with access to whatever corresponds to your real role (Manual 15, section 4; Agent control, Manual 52) — the same audience and per-medium permission narrowings that apply to the rest of the Platform.
The session stays open while there’s activity, up to the limit set on the instance (section 3). After that much time without using it, it asks you to identify yourself again — your previous conversation isn’t lost from the Platform (your tenant’s conversations stay just like any other), but you’ll have to log back in to keep using the widget.