Fix visual browsing (#7278)

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

openhands/agenthub/visualbrowsing_agent/visualbrowsing_agent.py

@@ -202,6 +202,7 @@ Note:
         tabs = ''
         last_obs = None
         last_action = None
+        set_of_marks = None  # Initialize set_of_marks to None
 
         if len(state.history) == 1:
             # for visualwebarena, webarena and miniwob++ eval, we need to retrieve the initial observation already in browser env
@@ -217,6 +218,9 @@ Note:
                 # agent has responded, task finished.
                 return AgentFinishAction(outputs={'content': event.content})
             elif isinstance(event, Observation):
+                # Only process BrowserOutputObservation and skip other observation types
+                if not isinstance(event, BrowserOutputObservation):
+                    continue
                 last_obs = event
 
         if len(prev_actions) >= 1:  # ignore noop()

openhands/memory/conversation_memory.py

@@ -52,7 +52,6 @@ class ConversationMemory:
         initial_messages: list[Message],
         max_message_chars: int | None = None,
         vision_is_active: bool = False,
-        enable_som_visual_browsing: bool = False,
     ) -> list[Message]:
         """Process state history into a list of messages for the LLM.
 
@@ -64,11 +63,13 @@ class ConversationMemory:
             max_message_chars: The maximum number of characters in the content of an event included
                 in the prompt to the LLM. Larger observations are truncated.
             vision_is_active: Whether vision is active in the LLM. If True, image URLs will be included.
-            enable_som_visual_browsing: Whether to enable visual browsing for the SOM model.
         """
 
         events = condensed_history
 
+        # log visual browsing status
+        logger.debug(f'Visual browsing: {self.agent_config.enable_som_visual_browsing}')
+
         # Process special events first (system prompts, etc.)
         messages = initial_messages