Currently, when a rate limit (429) or server load status is received from the LLM provider, OpenAICompatibleProvider.chat attempts to extract the retry delay from the Retry-After header.
According to the HTTP specification, Retry-After can be formatted either as a numeric value in seconds (e.g., 120) or as an HTTP-date string (e.g., Wed, 21 Oct 2015 07:28:00 GMT).
When the LLM API gateway returns a date string, calling float(retry_after) raises a ValueError, crashing the entire scoring execution and preventing successful fallback retries.
Currently, when a rate limit (429) or server load status is received from the LLM provider, OpenAICompatibleProvider.chat attempts to extract the retry delay from the Retry-After header.
According to the HTTP specification, Retry-After can be formatted either as a numeric value in seconds (e.g., 120) or as an HTTP-date string (e.g., Wed, 21 Oct 2015 07:28:00 GMT).
When the LLM API gateway returns a date string, calling float(retry_after) raises a ValueError, crashing the entire scoring execution and preventing successful fallback retries.