Fix
Removing a member now signs them out
Removing someone from your workspace ends their session instead of leaving them signed in until their token expired on its own. Separately, Thinkr Assist now falls through to a working model when one is unavailable, rather than failing the request.
Two fixes this release, both in places you only notice when they go wrong.
Removing a member ends their session
Removing someone from a workspace used to delete their membership and nothing else. Because Thinkr checks the signature on your session token locally, a removed member's existing token kept working, and could be renewed, so in practice access did not end when you clicked Remove.
Removing a member now signs that person out of Thinkr. Their session is ended at the identity provider, so they cannot renew it, and the next time their app asks for anything they land on the "access removed" screen. The audit entry for the removal records whether the sign-out completed, so there is evidence either way.
Two things worth knowing. The sign-out is global, because sessions are not per-workspace: someone who belongs to two workspaces and is removed from one gets signed out and will need to sign in again, after which they still have the other. And an invitation that was never accepted has no session to end, so revoking one is unchanged.
This bounds access to the token the person is already holding rather than ending it the same instant. Closing that last gap needs a membership check on every request, which is planned.
Thinkr Assist falls through to a working model
Thinkr tries a sequence of models and steps down to the next one when the first is busy. That step-down only happened for rate limits, so if the next model in the chain was not available to our project at all, the chain stopped there and the request failed, even though a model further down would have worked.
The step-down now also happens when a model is unavailable, not just when it is busy. In practice this means fewer outright failures on PRD generation, critiques, and decks when a model is being retired or is temporarily ungated. The match is deliberately narrow, so a genuine problem with your request still fails fast instead of being retried on another model.