fix tool call count

Author: benjibc

eval_protocol/pytest/default_single_turn_rollout_process.py

@@ -103,7 +103,10 @@ async def process_row(row: EvaluationRow) -> EvaluationRow:
             assert isinstance(response, ModelResponse), "Response should be ModelResponse"
             assert isinstance(response.choices[0], Choices), "Response choice should be a Choices"
 
+            print(f"DEBUG: Full Response: {response}")
             assistant_message = response.choices[0].message
+            print(f"DEBUG: Assistant Message: {assistant_message}")
+            print(f"DEBUG: Assistant Content: {assistant_message.content}")
             finish_reason = getattr(response.choices[0], "finish_reason", None)
 
             # Extract content
@@ -177,6 +180,10 @@ async def process_row(row: EvaluationRow) -> EvaluationRow:
                     total_tokens=total_tokens,
                 )
 
+            row.execution_metadata.finish_reason = finish_reason
+            if converted_tool_calls:
+                row.execution_metadata.tool_call_count = len(converted_tool_calls)
+
             row.execution_metadata.duration_seconds = time.perf_counter() - start_time
 
             default_logger.log(row)