Return last text parts on empty message (#1408)

Author: samuelcolvin

pydantic_ai_slim/pydantic_ai/_agent_graph.py

@@ -427,6 +427,18 @@ async def _run_stream() -> AsyncIterator[_messages.HandleResponseEvent]:
                     # No events are emitted during the handling of text responses, so we don't need to yield anything
                     self._next_node = await self._handle_text_response(ctx, texts)
                 else:
+                    # we've got an empty response, this sometimes happens with anthropic (and perhaps other models)
+                    # when the model has already returned text along side tool calls
+                    # in this scenario, if text responses are allowed, we return text from the most recent model
+                    # response, if any
+                    if allow_text_output(ctx.deps.output_schema):
+                        for message in reversed(ctx.state.message_history):
+                            if isinstance(message, _messages.ModelResponse):
+                                last_texts = [p.content for p in message.parts if isinstance(p, _messages.TextPart)]
+                                if last_texts:
+                                    self._next_node = await self._handle_text_response(ctx, last_texts)
+                                    return
+
                     raise exceptions.UnexpectedModelBehavior('Received empty model response')
 
             self._events_iterator = _run_stream()
@@ -530,14 +542,14 @@ async def _handle_text_response(
 
         text = '\n\n'.join(texts)
         if allow_text_output(output_schema):
+            # The following cast is safe because we know `str` is an allowed result type
             result_data_input = cast(NodeRunEndT, text)
             try:
                 result_data = await _validate_output(result_data_input, ctx, None)
             except _output.ToolRetryError as e:
                 ctx.state.increment_retries(ctx.deps.max_result_retries)
                 return ModelRequestNode[DepsT, NodeRunEndT](_messages.ModelRequest(parts=[e.tool_retry]))
             else:
-                # The following cast is safe because we know `str` is an allowed result type
                 return self._handle_final_result(ctx, result.FinalResult(result_data, None, None), [])
         else:
             ctx.state.increment_retries(ctx.deps.max_result_retries)

tests/test_agent.py

@@ -1813,3 +1813,59 @@ def test_instructions_with_message_history():
             ),
         ]
     )
+
+
+def test_empty_final_response():
+    def llm(messages: list[ModelMessage], info: AgentInfo) -> ModelResponse:
+        if len(messages) == 1:
+            return ModelResponse(parts=[TextPart('foo'), ToolCallPart('my_tool', {'x': 1})])
+        elif len(messages) == 3:
+            return ModelResponse(parts=[TextPart('bar'), ToolCallPart('my_tool', {'x': 2})])
+        else:
+            return ModelResponse(parts=[])
+
+    agent = Agent(FunctionModel(llm))
+
+    @agent.tool_plain
+    def my_tool(x: int) -> int:
+        return x * 2
+
+    result = agent.run_sync('Hello')
+    assert result.output == 'bar'
+
+    assert result.new_messages() == snapshot(
+        [
+            ModelRequest(parts=[UserPromptPart(content='Hello', timestamp=IsNow(tz=timezone.utc))]),
+            ModelResponse(
+                parts=[
+                    TextPart(content='foo'),
+                    ToolCallPart(tool_name='my_tool', args={'x': 1}, tool_call_id=IsStr()),
+                ],
+                model_name='function:llm:',
+                timestamp=IsNow(tz=timezone.utc),
+            ),
+            ModelRequest(
+                parts=[
+                    ToolReturnPart(
+                        tool_name='my_tool', content=2, tool_call_id=IsStr(), timestamp=IsNow(tz=timezone.utc)
+                    )
+                ]
+            ),
+            ModelResponse(
+                parts=[
+                    TextPart(content='bar'),
+                    ToolCallPart(tool_name='my_tool', args={'x': 2}, tool_call_id=IsStr()),
+                ],
+                model_name='function:llm:',
+                timestamp=IsNow(tz=timezone.utc),
+            ),
+            ModelRequest(
+                parts=[
+                    ToolReturnPart(
+                        tool_name='my_tool', content=4, tool_call_id=IsStr(), timestamp=IsNow(tz=timezone.utc)
+                    )
+                ]
+            ),
+            ModelResponse(parts=[], model_name='function:llm:', timestamp=IsNow(tz=timezone.utc)),
+        ]
+    )