fix: fall back on OpenAI streaming overloads

[Mason Daugherty (mdrxy)]

OpenAI Responses streaming overloads can raise a plain APIError without status_code, causing runs to fail before trying the configured fallback.

• Retry only exact base-class instances (type(exc) is openai.APIError) when classify_exception(exc) == "provider_overloaded". This does not make every APIError retryable: unrelated plain APIErrors classified as provider_unavailable still propagate, and subclasses retain their existing handling.
• Reuse the existing overload classifier, which recognizes overloaded in the message or extracted provider code. This supports the reported status-less streaming error without assuming an HTTP status or structured error code; it is a text-based classification, not a strict structured-code match.
• Preserve retry timing, model selection, configuration, and exhaustion behavior. Behavioral coverage includes successful fallback for the exact overload message and propagation of unrelated API, authentication, invalid-request, and context-length errors.

Made by Open SWE · openai:gpt-6-astra (low)

[open-swe]

✅ Open SWE Review: No issues found

Open SWE reviewed this PR and found no potential bugs to report.

Open in Web • View Open SWE trace