LiteLLM 1.105.0 makes several failed streams visible. Test the error path.
The 11 October release changes Responses handling to fall back on pre-output drops, fail truncated streams and honor request timeouts. It also turns an empty Bedrock converse stream into a 502 and surfaces Messages stream failures as error events. These fixes matter when an agent treats an empty reply as a completed turn. 1
Exercise the provider path your harness uses: a failure before the first output, an interruption after partial output, and a request timeout. Check what the client receives and whether the task is marked failed. Record each retry so a partial operation cannot silently run twice.
The same release gates SSO debug routes behind ENABLE_SSO_DEBUG, off by default, and adds several redaction fixes. Upgrade with your current configuration captured; confirm that operational logs remain useful without exposing the request body you intended to suppress. 1
An error event starts recovery. The caller still needs to preserve task state and decide whether retrying is safe. A provider switch starts another attempt; it needs its own recorded outcome.
Test timeout, partial-output and empty-stream cases through your client before rollout.