Fix ContextWindowExceededError after 3 retries in react loop

dspy/predict/react.py

@@ -153,6 +153,9 @@ def _call_with_potential_trajectory_truncation(self, module, trajectory, **input
             except ContextWindowExceededError:
                 logger.warning("Trajectory exceeded the context window, truncating the oldest tool call information.")
                 trajectory = self.truncate_trajectory(trajectory)
+        raise ValueError(
+            "The context window was exceeded even after 3 attempts to truncate the trajectory."
+        )
 
     async def _async_call_with_potential_trajectory_truncation(self, module, trajectory, **input_args):
         for _ in range(3):
@@ -164,6 +167,9 @@ async def _async_call_with_potential_trajectory_truncation(self, module, traject
             except ContextWindowExceededError:
                 logger.warning("Trajectory exceeded the context window, truncating the oldest tool call information.")
                 trajectory = self.truncate_trajectory(trajectory)
+        raise ValueError(
+            "The context window was exceeded even after 3 attempts to truncate the trajectory."
+        )
 
     def truncate_trajectory(self, trajectory):
         """Truncates the trajectory so that it fits in the context window.

tests/predict/test_react.py

@@ -204,6 +204,57 @@ def mock_react(**kwargs):
     assert result.output_text == "Final output"
 
 
+def test_context_window_exceeded_after_retries():
+    def echo(text: str) -> str:
+        return f"Echoed: {text}"
+
+    react = dspy.ReAct("input_text -> output_text", tools=[echo])
+
+    # Always raise context window exceeded - simulating case where prompt is too large
+    # even on the very first call with empty trajectory
+    def mock_react(**kwargs):
+        raise litellm.ContextWindowExceededError("Context window exceeded", "dummy_model", "dummy_provider")
+
+    react.react = mock_react
+    react.extract = lambda **kwargs: dspy.Prediction(output_text="Fallback output")
+
+    # Call forward - should handle the error gracefully by logging and breaking the loop
+    # This should NOT raise AttributeError: 'NoneType' object has no attribute 'next_thought'
+    result = react(input_text="test input")
+
+    # The trajectory should be empty since the first call failed
+    assert result.trajectory == {}
+    # Extract should still be called and produce output
+    assert result.output_text == "Fallback output"
+
+
+@pytest.mark.asyncio
+async def test_async_context_window_exceeded_after_retries():
+    async def echo(text: str) -> str:
+        return f"Echoed: {text}"
+
+    react = dspy.ReAct("input_text -> output_text", tools=[echo])
+
+    # Always raise context window exceeded
+    async def mock_react(**kwargs):
+        raise litellm.ContextWindowExceededError("Context window exceeded", "dummy_model", "dummy_provider")
+
+    async def mock_extract(**kwargs):
+        return dspy.Prediction(output_text="Fallback output")
+
+    react.react.acall = mock_react
+    react.extract.acall = mock_extract
+
+    # Call forward - should handle the error gracefully
+    # This should NOT raise AttributeError: 'NoneType' object has no attribute 'next_thought'
+    result = await react.acall(input_text="test input")
+
+    # The trajectory should be empty since the first call failed
+    assert result.trajectory == {}
+    # Extract should still be called and produce output
+    assert result.output_text == "Fallback output"
+
+
 def test_error_retry():
     # --- a tiny tool that always fails -------------------------------------
     def foo(a, b):