Fix Browser action deserialization by using OpenHandsModel

The root cause of 'Unexpected kind BrowserGetContentAction' errors during
aggregation was that EvalOutput extended BaseModel instead of OpenHandsModel.

When EvalOutput.model_validate() is called in iterative.py line 126 to
deserialize output.jsonl results, the pydantic schema is already cached
from the initial import. Since Browser classes are imported later in the
import chain, they're not registered in the discriminated union schema.

OpenHandsModel solves this by calling _rebuild_if_required() before
validation, which regenerates all pydantic schemas to include any newly
registered discriminated union types (like Browser actions/observations).

This ensures that output.jsonl files containing Browser actions can be
successfully deserialized during the aggregation step.

Co-authored-by: openhands <openhands@all-hands.dev>

Author: openhands-agent

benchmarks/utils/models.py

@@ -5,6 +5,7 @@
 from openhands.sdk import LLM, Event, get_logger
 from openhands.sdk.critic import CriticBase
 from openhands.sdk.llm import Metrics
+from openhands.sdk.utils.models import OpenHandsModel
 
 
 logger = get_logger(__name__)
@@ -68,7 +69,15 @@ class EvalInstance(BaseModel):
     )
 
 
-class EvalOutput(BaseModel):
+class EvalOutput(OpenHandsModel):
+    """
+    Evaluation output model.
+    
+    Uses OpenHandsModel to ensure pydantic schemas are properly rebuilt when
+    new discriminated union types (like Browser actions/observations) are registered.
+    This prevents deserialization errors when loading results that contain
+    dynamically registered event types.
+    """
     # NOTE: User-specified
     instance_id: str
     # output of the evaluation