fix(orchestrator): Enforce retry budget to prevent unbounded loops

backend/tests/test_review_fix_unbounded_loop.py

@@ -0,0 +1,85 @@
+import pytest
+from unittest.mock import MagicMock, patch
+from ea_chatbot.graph.workflow import create_workflow
+from ea_chatbot.graph.state import AgentState
+from langchain_core.messages import AIMessage, HumanMessage
+
+def test_orchestrator_loop_retry_budget():
+    """Verify that the orchestrator loop is bounded and terminates after max retries."""
+    
+    mock_analyzer = MagicMock()
+    mock_planner = MagicMock()
+    mock_delegate = MagicMock()
+    mock_worker = MagicMock()
+    mock_reflector = MagicMock()
+    mock_synthesizer = MagicMock()
+    
+    # 1. Analyzer: Proceed to planning
+    mock_analyzer.return_value = {"next_action": "plan"}
+    
+    # 2. Planner: One task
+    mock_planner.return_value = {
+        "checklist": [{"task": "Unsolvable Task", "worker": "data_analyst"}],
+        "current_step": 0,
+        "iterations": 0
+    }
+    
+    # We'll use the REAL delegate and reflector logic to verify the fix
+    # But we mock the LLM calls inside them if necessary.
+    # Actually, it's easier to just mock the node return values but follow the logic.
+    
+    from ea_chatbot.graph.nodes.delegate import delegate_node
+    from ea_chatbot.graph.nodes.reflector import reflector_node
+    
+    # Mocking the LLM inside reflector to always be unsatisfied
+    with patch("ea_chatbot.graph.nodes.reflector.get_llm_model") as mock_get_llm:
+        mock_llm = MagicMock()
+        # Mark as NOT satisfied
+        mock_llm.with_structured_output.return_value.invoke.return_value = MagicMock(satisfied=False, reasoning="Still bad.")
+        mock_get_llm.return_value = mock_llm
+        
+        app = create_workflow(
+            query_analyzer=mock_analyzer,
+            planner=mock_planner,
+            # delegate=delegate_node, # Use real
+            data_analyst_worker=mock_worker,
+            # reflector=reflector_node, # Use real
+            synthesizer=mock_synthesizer
+        )
+        
+        # Mock worker to return something
+        mock_worker.return_value = {"result": "Bad Output", "messages": [AIMessage(content="Bad")]}
+        mock_synthesizer.return_value = {"messages": [AIMessage(content="Failure Summary")], "next_action": "end"}
+        
+        initial_state = AgentState(
+            messages=[HumanMessage(content="test")],
+            question="test",
+            analysis={},
+            next_action="",
+            iterations=0,
+            checklist=[],
+            current_step=0,
+            vfs={},
+            plots=[],
+            dfs={}
+        )
+        
+        # Run the graph. If fix works, it should hit iterations=3 and route to synthesizer.
+        # We use a recursion_limit higher than our retry budget but low enough to fail fast if unbounded.
+        final_state = app.invoke(initial_state, config={"recursion_limit": 20})
+        
+        # Assertions
+        # 1. We tried 3 times (iterations 0, 1, 2) and failed on 3rd.
+        # Wait, delegate routes to summarize when iterations >= 3.
+        # Reflector increments iterations.
+        # Loop: 
+        # Start: it=0
+        # Delegate (it=0) -> Worker -> Reflector (fail, it=1) -> Delegate (it=1)
+        # Delegate (it=1) -> Worker -> Reflector (fail, it=2) -> Delegate (it=2)
+        # Delegate (it=2) -> Worker -> Reflector (fail, it=3) -> Delegate (it=3)
+        # Delegate (it=3) -> Summarize (it=0)
+        
+        assert final_state["iterations"] == 0 # Reset in delegate or handled in synthesizer
+        # Check if we hit the failure summary
+        assert "Failed to complete task" in final_state["summary"]
+        assert mock_worker.call_count == 3