fix(anthropic): warn when response is truncated by max_tokens limit

griptape/drivers/prompt/anthropic_prompt_driver.py

@@ -107,6 +107,13 @@ def try_run(self, prompt_stack: PromptStack) -> Message:
 
         logger.debug(response.model_dump())
 
+        if response.stop_reason == "max_tokens":
+            logger.warning(
+                "Response reached the max_tokens limit (%d). Output may be truncated. "
+                "Increase the max_tokens parameter to get longer responses.",
+                self.max_tokens,
+            )
+
         return Message(
             content=[self.__to_prompt_stack_message_content(content) for content in response.content],
             role=Message.ASSISTANT_ROLE,
@@ -126,6 +133,12 @@ def try_stream(self, prompt_stack: PromptStack) -> Iterator[DeltaMessage]:
             elif event.type == "message_start":
                 yield DeltaMessage(usage=DeltaMessage.Usage(input_tokens=event.message.usage.input_tokens))
             elif event.type == "message_delta":
+                if event.delta.stop_reason == "max_tokens":
+                    logger.warning(
+                        "Response reached the max_tokens limit (%d). Output may be truncated. "
+                        "Increase the max_tokens parameter to get longer responses.",
+                        self.max_tokens,
+                    )
                 yield DeltaMessage(usage=DeltaMessage.Usage(output_tokens=event.usage.output_tokens))
 
     def _base_params(self, prompt_stack: PromptStack) -> dict:

tests/unit/drivers/prompt/test_anthropic_prompt_driver.py

@@ -524,3 +524,63 @@ def test_try_run_omits_sampling_params_for_opus_4_7(self, mock_client, prompt_st
         assert "temperature" not in call_kwargs.kwargs
         assert "top_p" not in call_kwargs.kwargs
         assert "top_k" not in call_kwargs.kwargs
+
+    def test_try_run_warns_on_max_tokens(self, mocker, prompt_stack, caplog):
+        import logging
+
+        # Given
+        mock_client = mocker.patch("anthropic.Anthropic")
+        mock_client.return_value = Mock(
+            messages=Mock(
+                create=Mock(
+                    return_value=Mock(
+                        stop_reason="max_tokens",
+                        usage=Mock(input_tokens=5, output_tokens=10),
+                        content=[Mock(type="text", text="truncated-output")],
+                    )
+                )
+            )
+        )
+        driver = AnthropicPromptDriver(model="claude-3-haiku", api_key="api-key", max_tokens=10)
+
+        # When
+        with caplog.at_level(logging.WARNING, logger="griptape"):
+            driver.try_run(prompt_stack)
+
+        # Then
+        assert any("max_tokens" in r.message for r in caplog.records if r.levelno == logging.WARNING)
+
+    def test_try_stream_warns_on_max_tokens(self, mocker, prompt_stack, caplog):
+        import logging
+
+        # Given
+        mock_client = mocker.patch("anthropic.Anthropic")
+        mock_client.return_value = Mock(
+            messages=Mock(
+                create=Mock(
+                    return_value=iter(
+                        [
+                            Mock(type="message_start", message=Mock(usage=Mock(input_tokens=5))),
+                            Mock(
+                                type="content_block_start",
+                                index=0,
+                                content_block=Mock(type="text", text="truncated"),
+                            ),
+                            Mock(
+                                type="message_delta",
+                                usage=Mock(output_tokens=10),
+                                delta=Mock(stop_reason="max_tokens"),
+                            ),
+                        ]
+                    )
+                )
+            )
+        )
+        driver = AnthropicPromptDriver(model="claude-3-haiku", api_key="api-key", stream=True, max_tokens=10)
+
+        # When
+        with caplog.at_level(logging.WARNING, logger="griptape"):
+            list(driver.try_stream(prompt_stack))
+
+        # Then
+        assert any("max_tokens" in r.message for r in caplog.records if r.levelno == logging.WARNING)