feat: Add configurable stuck/loop detection (#11799)

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

openhands/controller/agent_controller.py

@@ -895,7 +895,7 @@ class AgentController:
 
         # Synchronize spend across all llm services with the budget flag
         self.state_tracker.sync_budget_flag_with_metrics()
-        if self._is_stuck():
+        if self.agent.config.enable_stuck_detection and self._is_stuck():
             await self._react_to_exception(
                 AgentStuckInLoopError('Agent got stuck in a loop')
             )

openhands/core/config/README.md

@@ -32,6 +32,7 @@ The `load_from_env` function in the config package is responsible for loading co
 export LLM_API_KEY='your_api_key_here'
 export LLM_MODEL='gpt-4'
 export AGENT_MEMORY_ENABLED='true'
+export AGENT_ENABLE_STUCK_DETECTION='false'  # Disable loop detection
 export SANDBOX_TIMEOUT='300'
 ```
 

openhands/core/config/agent_config.py

@@ -51,6 +51,8 @@ class AgentConfig(BaseModel):
     """Whether to enable SoM (Set of Marks) visual browsing."""
     enable_plan_mode: bool = Field(default=True)
     """Whether to enable plan mode, which uses the long horizon system message and add the new tool - task_tracker - for planning, tracking and executing complex tasks."""
+    enable_stuck_detection: bool = Field(default=True)
+    """Whether to enable stuck/loop detection. When disabled, the agent will not automatically detect and recover from loops."""
     condenser: CondenserConfig = Field(
         # The default condenser is set to the conversation window condenser -- if
         # we use NoOp and the conversation hits the LLM context length limit,