LiteLLM OpenTelemetry callback: tool-call responses recorded with empty parts; single-chunk streamed tool call recorded as finish reason stop

Finding: In a reported case (v1.104.2, LiteLLM Python SDK with the built-in OpenTelemetry callback), tool-call responses are traced with empty message parts, and a streamed tool call sent in a single chunk is traced with finish reason stop. This is an external report and was not reproduced by KFA. Source: https://github.com/BerriAI/litellm/issues/45796 Reported symptom 1: with OTEL_SEMCONV_STABILITY_OPT_IN=gen_ai_latest_experimental, when the model returns a tool call (streamed read to the end, or non-streamed), the span attribute gen_ai.output.messages is an assistant message with empty parts and finish_reason tool_calls. The tool name, call id and arguments do not appear, although the application receives the tool call correctly (https://github.com/BerriAI/litellm/issues/45796). Reported symptom 2: when a streamed response delivers the tool call and finish_reason tool_calls in the same chunk, the application sees tool_calls per chunk, but the span records gen_ai.response.finish_reasons as ["stop"] with empty parts (same source). Reporter's suspected cause (HYPOTHESIS from code reading, not maintainer-confirmed): in _transform_messages_to_otel_semantic_conventions only string or list content becomes parts, and content is None for tool-call messages. A tool_calls branch copies the calls only for plain dicts, so Message objects from a ModelResponse lose them (same source). Environment: reporter states v1.104.2 and says the code path is unchanged on main at commit 7ff5010. The reproducers use a mocked OpenAI transport with an in-memory span exporter; behavior with real providers or other versions is unknown. Status: no maintainer response in the supplied thread. A commenter without maintainer association said they would open a PR; no PR, commit or release was supplied, so there is no confirmed fix or shipped release. No workaround was reported. Practical implication: if you rely on gen_ai.output.messages to audit tool calls on this version, the traces may omit tool name and arguments, and finish reasons for single-chunk streamed tool calls may read stop. Treat this as source-reported, unverified by KFA, and check the issue for fix updates.