Reliability in ClawAI: fallback, breakers and resumable streams

Automatic fallback to another model

Every routed request carries a list of candidate models. If the chosen model fails mid-request, ClawAI moves to the next candidate automatically, and the answer records which model actually responded, not only the one first chosen.

A shared breaker for exhausted providers

When a provider account runs out of credit, a breaker takes that provider out of rotation for ten minutes, then lets a single test call through to see whether it has recovered. The breaker’s state is shared in Redis across every chat server, so one failure is learned once rather than rediscovered by each server, and each server falls back to its own copy if Redis is unavailable.

Stop that always stops, streams that resume

Pressing Stop is broadcast to every chat server, so whichever one is running the model aborts it. If your connection drops while an answer is streaming, reconnecting replays the events you missed from a buffer instead of losing the rest of the answer.

Questions people ask

What happens if a model fails halfway through an answer?

ClawAI falls back to the next candidate model automatically, and the answer shows which model actually produced it.

What does the provider breaker protect against?

A provider account that has run out of credit. It is skipped for ten minutes and then tested with one call, instead of every request failing against it in the meantime.

Do I lose an answer if my connection drops?

No. When the stream reconnects, the events you missed are replayed from a server-side buffer and the answer continues.