fix

Author: Dylan Huang

eval_protocol/models.py

@@ -122,7 +122,7 @@ class Code(int, Enum):
         UNAUTHENTICATED = 16
 
         # Custom codes for rollout states (using higher numbers to avoid conflicts)
-        FINISHED = 100  # Custom code for rollout finished
+        FINISHED = 100
 
     @classmethod
     def rollout_running(cls) -> "Status":

eval_protocol/pytest/utils.py

@@ -282,7 +282,7 @@ async def execute_row_with_backoff(task: asyncio.Task, row: EvaluationRow) -> Ev
             try:
                 # Try original task first
                 result = await task
-                result.rollout_status.status = RolloutStatus.Status.FINISHED
+                result.rollout_status = Status.rollout_finished()
                 return result
             except Exception as e:
                 # NOTE: we perform these checks because we don't put the backoff decorator on initial batch call. we don't want to retry whole batch if anything fails.
@@ -295,17 +295,15 @@ async def execute_row_with_backoff(task: asyncio.Task, row: EvaluationRow) -> Ev
                     # Use shared backoff function for retryable exceptions
                     try:
                         result = await execute_row_with_backoff_retry(row)
-                        result.rollout_status.status = RolloutStatus.Status.FINISHED
+                        result.rollout_status = Status.rollout_finished()
                         return result
                     except Exception as retry_error:
                         # Backoff gave up
-                        row.rollout_status.status = RolloutStatus.Status.ERROR
-                        # row.rollout_status.termination_reason = str(retry_error)
+                        row.rollout_status = Status.rollout_error(str(retry_error))
                         return row
                 else:
                     # Non-retryable exception - fail immediately
-                    row.rollout_status.status = RolloutStatus.Status.ERROR
-                    # row.rollout_status.termination_reason = str(e)
+                    row.rollout_status = Status.rollout_error(str(e))
                     return row
 
         # Process all tasks concurrently with backoff retry