Fix ContextWindowExceededError after 3 retries in react loop (#9110)

* Initial plan

* Fix ContextWindowExceededError after 3 retries in react loop

When _call_with_potential_trajectory_truncation exhausts all retry
attempts, it now raises a clear ValueError instead of returning None.
This prevents the AttributeError: 'NoneType' object has no attribute
'next_thought' that occurred when accessing properties on the None
return value.

The ValueError is caught in forward/aforward and causes the loop to
break gracefully, allowing the extract phase to proceed with whatever
trajectory was collected.

Added tests for both sync and async versions of this scenario.

Co-authored-by: TomeHirata <33407409+TomeHirata@users.noreply.github.com>

* Address code review feedback: Fix comments and error messages

Co-authored-by: TomeHirata <33407409+TomeHirata@users.noreply.github.com>

* Remove unnecessary try-catch around truncate_trajectory

Co-authored-by: TomeHirata <33407409+TomeHirata@users.noreply.github.com>

* Remove docstrings from test functions

Co-authored-by: TomeHirata <33407409+TomeHirata@users.noreply.github.com>

* Combine sync and async context window tests into one test

Co-authored-by: TomeHirata <33407409+TomeHirata@users.noreply.github.com>

* Verify inputs passed to extract in test

Co-authored-by: TomeHirata <33407409+TomeHirata@users.noreply.github.com>

---------

Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: TomeHirata <33407409+TomeHirata@users.noreply.github.com>

Author: Copilot

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 +204,54 @@ def mock_react(**kwargs):
     assert result.output_text == "Final output"
 
 
+@pytest.mark.asyncio
+async 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])
+
+    def mock_react(**kwargs):
+        raise litellm.ContextWindowExceededError("Context window exceeded", "dummy_model", "dummy_provider")
+
+    # Test sync version
+    extract_calls = []
+
+    def mock_extract(**kwargs):
+        extract_calls.append(kwargs)
+        return dspy.Prediction(output_text="Fallback output")
+
+    react.react = mock_react
+    react.extract = mock_extract
+
+    result = react(input_text="test input")
+    assert result.trajectory == {}
+    assert result.output_text == "Fallback output"
+    assert len(extract_calls) == 1
+    assert extract_calls[0]["input_text"] == "test input"
+    assert "trajectory" in extract_calls[0]
+
+    # Test async version
+    async_extract_calls = []
+
+    async def mock_react_async(**kwargs):
+        raise litellm.ContextWindowExceededError("Context window exceeded", "dummy_model", "dummy_provider")
+
+    async def mock_extract_async(**kwargs):
+        async_extract_calls.append(kwargs)
+        return dspy.Prediction(output_text="Fallback output")
+
+    react.react.acall = mock_react_async
+    react.extract.acall = mock_extract_async
+
+    result = await react.acall(input_text="test input")
+    assert result.trajectory == {}
+    assert result.output_text == "Fallback output"
+    assert len(async_extract_calls) == 1
+    assert async_extract_calls[0]["input_text"] == "test input"
+    assert "trajectory" in async_extract_calls[0]
+
+
 def test_error_retry():
     # --- a tiny tool that always fails -------------------------------------
     def foo(a, b):