Stop Calling the Provider That Is Down

When a model provider starts timing out, retrying is the worst thing you can do. Every request waits its full 30-second timeout before failing, your workers fill with calls that are going nowhere, and an outage in one dependency becomes an outage in your whole application. A circuit breaker fails fast instead, and, crucially, knows how to find out when the provider is back.

Write circuit_breaker(timestamps, results, failure_threshold=3, cooldown=10.0).

results[i] is what the call would return: "ok" or "fail". Return what actually happened for each request: "ok", "fail", or "rejected" if the breaker refused to make the call at all.

• Closed: calls go through. failure_threshold consecutive failures opens it. A success resets the count.
• Open: every call is "rejected" without being attempted, until cooldown seconds have passed since it opened.
• Half-open: the next call after the cooldown is allowed through as a probe. Success closes the breaker; failure opens it again and restarts the cooldown.