description

felixnic pointed out on #3949 that after set_model() mid-session, llm.model_name can still show the first model. i checked it with real AssistantMessage / ResultMessage types and it reproduces.

• turn 1 runs on sonnet, then set_model("haiku"), turn 2 runs on haiku
• the result message's model_usage covers the whole session, so it lists both models
• the instrumentor picked the model with the most output tokens from that list, so turn 2 got sonnet even though the assistant message said haiku

fix:

• if the span already has a model from an assistant message, model_usage is not used to replace it
• if no model was seen on the span, model_usage is still used as before, so the multi-model pick from [bug][claude-agent-sdk] _extract_model_name_from_usage returns first dict key, mis-attributes multi-model runs #3136 is unchanged

before and after, same script:

tests:

• 2 new unit tests in test_wrappers_unit.py, the first one fails on main and passes with the fix
• pytest tests in the package: 54 passed

separate from #3949, which only touches the per-turn token counts.

checklist:

• Follows OpenInference configuration to hide sensitive info
• Spans properly inherit from context attributes
• Properly respects suppress tracing context