fix: patch bug w/ extended thinking mode involving text leaking into reasoning (#4341)

* fix: patch for bad native reasoning behavior w/ sonnet

* fix: cleanup

* fix: cleanup

* fix: another prompt tune for less flaking

letta/interfaces/anthropic_streaming_interface.py

@@ -289,6 +289,13 @@ class AnthropicStreamingInterface:
                 if not self.anthropic_mode == EventMode.TEXT:
                     raise RuntimeError(f"Streaming integrity failed - received BetaTextDelta object while not in TEXT EventMode: {delta}")
 
+                # Weird bug happens with native thinking where a single response can contain:
+                # [reasoning, text, tool_call]
+                # In these cases, we should pipe text out to null / ignore it
+                # TODO this will have to be redone to support non-tool calling message sending
+                if not self.put_inner_thoughts_in_kwarg:
+                    return
+
                 # Combine buffer with current text to handle tags split across chunks
                 combined_text = self.partial_tag_buffer + delta.text
 

tests/integration_test_send_message.py

@@ -136,6 +136,7 @@ USER_MESSAGE_ROLL_DICE_LONG_THINKING: List[MessageCreate] = [
             "Explain the concept of randomness and how true random number generation works. "
             "End with some interesting facts about polyhedral dice and their history in gaming. "
             "Remember, make your response detailed and at least 800 characters long."
+            "Absolutely do NOT violate this order of operations: (1) Think / reason, (2) Roll die, (3) Think / reason, (4) Call send_message tool."
         ),
         otid=USER_MESSAGE_OTID,
     )