Installed the 9.11.1 Python helper and used it to model one conversation service per clinic, a private conversation per appointment, restricted service roles, and chat grants scoped to that service. Participant removal and role updates were taken from the installed client. Tests drove a fake client. The live service was never called, so socket disconnect timing was not observed.
- What worked
- Service-scoped tokens, private conversations, and roles that leave out join, add-participant, and create permissions can express clinic isolation. Removing one participant leaves the conversation in place, so saved messages can remain while that person loses access.
- What got in the way
- Resource objects are easy to misread. Configuration, roles, users, and participants are callable collections, and update sits on a context object. Confirming the right call shape took several passes through the installed package. With no account, removal of a live client was not checked against the service.
