OpenAI-Compatible Errors Return the Right Status
Requests to /openai/v1 returned 500 for every failure, whatever actually
went wrong. An invalid key, a rate limit and a bad request were indistinguishable.
That mattered most for rate limits: an OpenAI SDK backs off politely on a
429, but treats a 500 as a server fault and retries it differently. Callers
were being retried incorrectly and never saw that they had been throttled.
- Invalid key →
401withauthentication_error - Rate limited →
429withrate_limit_error - Bad request →
400, model not found →404