Removed context_managed_run

app.py

@@ -578,7 +578,7 @@ original_call = model.__call__
 model.__call__ = try_model_call_with_fallbacks
 
 # Reduce the model's output tokens immediately to improve chances of success
-model.max_tokens = 1000  # Reduce from 2096 to 1000 to stay under token limits
+model.max_tokens = 750  # Reduce from 2096 to 750 for all outputs by default
 
 # Import tool from Hub
 image_generation_tool = load_tool("agents-course/text-to-image", trust_remote_code=True)
@@ -586,59 +586,7 @@ image_generation_tool = load_tool("agents-course/text-to-image", trust_remote_co
 with open("prompts.yaml", 'r') as stream:
     prompt_templates = yaml.safe_load(stream)
 
-# Override CodeAgent run to automatically clear context periodically
-original_code_agent_run = CodeAgent.run
-
-def context_managed_run(self, task, max_steps=None, persist_conversation=None, stream=False, reset=False, additional_args=None):
-    """Override to periodically clear conversation context after steps."""
-    # Initialize step counter if not present
-    if not hasattr(self, '_context_step_counter'):
-        self._context_step_counter = 0
-    
-    # Reset counter if reset is requested
-    if reset:
-        self._context_step_counter = 0
-    else:
-        # Increment counter
-        self._context_step_counter += 1
-    
-    # Every 5 steps, perform aggressive context management
-    if self._context_step_counter >= 5:
-        print("Performing periodic conversation cleanup")
-        self._context_step_counter = 0
-        
-        # Clear most of the conversation history while keeping essential parts
-        if hasattr(self, 'conversation') and len(self.conversation) > 8:
-            original_length = len(self.conversation)
-            
-            # Find system messages
-            system_messages = [msg for msg in self.conversation if msg.get('role') == 'system']
-            
-            # Get the most recent exchanges (last 4 pairs of messages)
-            recent_messages = self.conversation[-8:]
-            
-            # Rebuild conversation with system messages + note + recent messages
-            self.conversation = system_messages + [
-                {'role': 'system', 'content': '[NOTICE: Previous conversation history has been trimmed to manage context size]'}
-            ] + recent_messages
-            
-            print(f"Conversation history trimmed from {original_length} to {len(self.conversation)} messages")
-    
-    # Call the original run method with all parameters (as they were received)
-    return original_code_agent_run(
-        self, 
-        task, 
-        max_steps=max_steps,
-        persist_conversation=persist_conversation,
-        stream=stream,
-        reset=reset,
-        additional_args=additional_args
-    )
-
-# Apply the monkey patch to CodeAgent class
-CodeAgent.run = context_managed_run
-    
-# Initialize the agent using standard smolagents patterns
+# Update the agent to use more conservative settings
 agent = CodeAgent(
     model=model,
     tools=[
@@ -651,16 +599,13 @@ agent = CodeAgent(
         Check_Dataset_Validity,
         visit_webpage_tool,  # This is correctly initialized as VisitWebpageTool()
     ],
-    max_steps=12,
-    verbosity_level=1,
+    max_steps=8,  # Reduce from 12 to 8
+    verbosity_level=0,  # Reduce from 1 to 0 to minimize internal conversation
     grammar=None,
     planning_interval=2,
     name="Research Assistant",
-    description="""An AI assistant that can search the web, create datasets, and answer questions # Note about working within token limits
-                # When using with queries that might exceed token limits, consider:
-                # 1. Breaking tasks into smaller sub-tasks
-                # 2. Limiting the amount of data returned by search tools
-                # 3. Using the planning_interval to enable more effective reasoning""",
+    description="""An AI assistant that can search the web, create datasets, and answer questions.
+                This assistant automatically manages token limits for better stability.""",
     prompt_templates=prompt_templates
 )
 
@@ -677,7 +622,7 @@ print(f"Available tools: final_answer, Sonar_Web_Search_Tool, {search_tool_name}
 # To fix the TypeError in Gradio_UI.py, you would need to modify that file
 # For now, we'll just use the agent directly
 try:
-    GradioUI(agent).launch()
+    GradioUI(agent).launch(share=True)
 except TypeError as e:
     if "unsupported operand type(s) for +=" in str(e):
         print("Error: Token counting issue in Gradio UI")