dcdpr / jp / 25557841148

66%

push

github

fix(llm, openai): Classify overload errors as transient/retryable (#624)

The `classify_stream_error` function previously only treated
`server_error` and `api_error` types as transient. This meant that
overload-specific error signals were silently falling through to
`StreamError::Other`, making them non-retryable.

Two additional cases are now covered:

- `service_unavailable_error` type with `code=server_is_overloaded` is
the documented in-stream form of OpenAI's 503 "engine is currently
overloaded" condition. The discriminator lives in `code`, not `type`, so
both fields are now checked.

- `overloaded_error` type, used by some OpenAI-compatible providers that
reuse Anthropic's error taxonomy.

Either a matching `type` or a matching `code` is sufficient to classify
the error as transient, preserving any `retry-after` hint that may be
present on the error.

Signed-off-by: Jean Mertz <git@jeanmertz.com>

5 of 5 new or added lines in 1 file covered. (100.0%)

25771 of 38887 relevant lines covered (66.27%)

229.95 hits per line