Assist degrades instead of stopping

Thinkr runs on Google's Gemini models, and like any upstream dependency, that connection can fail in ways that have nothing to do with your spec. Until now, one of those failure modes stopped work outright: if our primary model route refused a request, the call retried a few times against the same route, waited, and gave up. You got "Thinkr Assist is unavailable right now" and about fifteen seconds of your afternoon back.

What changed

Assist now treats a refused route as a routing problem instead of a dead end.

When the primary route won't serve a request, the call moves to a secondary one and completes there. The model may be a smaller one, so a critique or a draft can read slightly tighter than usual — but it runs, and your work doesn't stop mid-sentence. The failure modes that genuinely can't be rescued, like a prompt the model declines, still surface as errors rather than being quietly retried somewhere else.

We also stopped retrying failures that cannot succeed. Some refusals are permanent for the lifetime of the request; retrying them three times with backoff only spent your time before showing the same message. Those now fail immediately, which turns a fifteen-second wait into an instant one, and more often skips straight to the fallback that does work.

Why you may never notice this

That's the intent. The point of this change is that a problem on our side becomes a slightly different answer rather than no answer. The last time this failure mode fired, it ran for several days against a handful of workspaces before we caught it, because nothing in the product said anything was wrong. Now the fallback keeps the work moving and our own monitoring says so loudly on the first occurrence.

No action needed on your side. If a critique or draft ever reads thinner than you expect, retrying a few minutes later will usually pick up the primary route again.