In new_stream, when agent_name from metadata is None but run_name exists, the span name correctly falls back to run_name (line 1022-1023), but GEN_AI_AGENT_NAME data attribute is never set (line 1032-1033 only checks agent_name is not None). This differs from new_invoke which sets GEN_AI_AGENT_NAME to run_name when truthy. Users calling stream() without explicit metadata but with a configured run_name will have inconsistent telemetry - the span name shows the agent name but the data attribute doesn't.

If langchain passes metadata=None explicitly in kwargs, kwargs.get("metadata", {}) returns None (not {}), causing .get("lc_agent_name") to raise AttributeError. While capture_internal_exceptions() suppresses this, the remaining span data (DATA_FIELDS, tools, prompts/messages) will not be recorded since they're in the same with block. The safer pattern is (kwargs.get("metadata") or {}).get("lc_agent_name").

In new_stream, when agent_name (from metadata) is None but run_name is available, the span name correctly uses run_name (line 1022-1023), but GEN_AI_AGENT_NAME is not set (lines 1032-1033 only check agent_name is not None). This is inconsistent with new_invoke which sets GEN_AI_AGENT_NAME when run_name is set (lines 963-964). Users expecting consistent agent name data between invoke and stream will not see the agent name in stream spans when using run_name configuration.