fix: prevent executor agent from getting stuck in action loop (#251)

- Add critical rules to executor system prompt: use wait as fallback when
  no viable action exists, always output a valid action, try different
  approaches after repeated failures
- Add repeated-action loop detection in handle_executor_result: when the
  same action is repeated 3+ times consecutively, set error_flag_plan to
  force the manager to re-plan

Author: adntgv

droidrun/agent/droid/droid_agent.py

@@ -7,6 +7,7 @@
 - When reasoning=True: Uses Manager (planning) + Executor (action) workflows
 """
 
+import json
 import logging
 from typing import TYPE_CHECKING, Type, Awaitable, Union
 
@@ -762,9 +763,23 @@ async def handle_executor_result(
         """
         Process Executor result and continue.
 
-        Checks for error escalation and loops back to Manager.
+        Checks for error escalation, repeated action loops, and loops back to Manager.
         Note: Max steps check is now done in run_manager pre-flight.
         """
+        # Check for repeated action loop (same action N times in a row)
+        repeat_thresh = 3
+        if len(self.shared_state.action_pool) >= repeat_thresh:
+            recent_actions = self.shared_state.action_pool[-repeat_thresh:]
+            try:
+                normalized = [json.dumps(a, sort_keys=True) for a in recent_actions]
+                if len(set(normalized)) == 1:
+                    logger.warning(
+                        f"⚠️ Executor stuck: same action repeated {repeat_thresh} times in a row: {normalized[0]}"
+                    )
+                    self.shared_state.error_flag_plan = True
+            except (TypeError, ValueError):
+                pass  # Non-serializable actions, skip detection
+
         # Check error escalation and reset flag when errors are resolved
         err_thresh = self.shared_state.err_to_manager_thresh
 

droidrun/config/prompts/executor/system.jinja2

@@ -93,8 +93,10 @@ No actions have been taken yet.
 Whatever the current subgoal says to do, do that EXACTLY. Do not substitute with what you think is better. Do not optimize. Do not consider screen state. Parse the subgoal text literally and execute the matching atomic action.
 
 IMPORTANT:
-1. Do NOT repeat previously failed actions multiple times. Try changing to another action.
+1. Do NOT repeat previously failed actions multiple times. If an action failed, try a DIFFERENT action or approach.
 2. Must do the current subgoal.
+3. If you have tried the same action 2+ times and it keeps failing, try a completely different approach. If truly stuck with no viable action, use `{"action": "wait", "duration": 1.0}` as a fallback and explain why in the Description (e.g., "No actionable element found for subgoal").
+4. ALWAYS output a valid action. There is no "skip" or "do nothing" option — use `wait` with duration 1.0 if uncertain.
 
 Provide your output in the following format, which contains three parts: