From f71bbfb41e0ed67e875a4b6abf88c28a3691407e Mon Sep 17 00:00:00 2001
From: Kevin Lin <klin5061@gmail.com>
Date: Mon, 25 Aug 2025 21:17:08 -0700
Subject: [PATCH] fix: patch anthropic send message tests (#4173)

Co-authored-by: Sarah Wooders <sarahwooders@gmail.com>
---
 letta/llm_api/anthropic_client.py      | 24 ++++++++++++++++--------
 letta/schemas/message.py               |  7 +++++++
 tests/integration_test_send_message.py | 14 ++++++++------
 3 files changed, 31 insertions(+), 14 deletions(-)

diff --git a/letta/llm_api/anthropic_client.py b/letta/llm_api/anthropic_client.py
index ef54ff62..6950213f 100644
--- a/letta/llm_api/anthropic_client.py
+++ b/letta/llm_api/anthropic_client.py
@@ -288,13 +288,22 @@ class AnthropicClient(LLMClientBase):
             anthropic_tools = None
 
         thinking_enabled = False
-        if messages:
-            for message in messages:
-                if isinstance(message.get("content"), list):
-                    for content_block in message["content"]:
-                        if content_block.get("type") == "thinking":
-                            thinking_enabled = True
-                            break
+        if messages and len(messages) > 0:
+            # Check if the last assistant message starts with a thinking block
+            # Find the last assistant message
+            last_assistant_message = None
+            for message in reversed(messages):
+                if message.get("role") == "assistant":
+                    last_assistant_message = message
+                    break
+
+            if (
+                last_assistant_message
+                and isinstance(last_assistant_message.get("content"), list)
+                and len(last_assistant_message["content"]) > 0
+                and last_assistant_message["content"][0].get("type") == "thinking"
+            ):
+                thinking_enabled = True
 
         try:
             count_params = {
@@ -305,7 +314,6 @@ class AnthropicClient(LLMClientBase):
 
             if thinking_enabled:
                 count_params["thinking"] = {"type": "enabled", "budget_tokens": 16000}
-
             result = await client.beta.messages.count_tokens(**count_params)
         except:
             raise
diff --git a/letta/schemas/message.py b/letta/schemas/message.py
index 5376cd5f..2b294dad 100644
--- a/letta/schemas/message.py
+++ b/letta/schemas/message.py
@@ -863,6 +863,13 @@ class Message(BaseMessage):
                                 "data": content_part.data,
                             }
                         )
+                    if isinstance(content_part, TextContent):
+                        content.append(
+                            {
+                                "type": "text",
+                                "text": content_part.text,
+                            }
+                        )
             elif text_content is not None:
                 content.append(
                     {
diff --git a/tests/integration_test_send_message.py b/tests/integration_test_send_message.py
index 344863fe..c8d794b3 100644
--- a/tests/integration_test_send_message.py
+++ b/tests/integration_test_send_message.py
@@ -399,7 +399,7 @@ def validate_openai_format_scrubbing(messages: List[Dict[str, Any]]) -> None:
             assert content is None
 
 
-def validate_anthropic_format_scrubbing(messages: List[Dict[str, Any]]) -> None:
+def validate_anthropic_format_scrubbing(messages: List[Dict[str, Any]], reasoning_enabled: bool) -> None:
     """
     Validate that Anthropic/Claude format assistant messages with tool_use have no <thinking> tags.
     Args:
@@ -433,10 +433,12 @@ def validate_anthropic_format_scrubbing(messages: List[Dict[str, Any]]) -> None:
         # Verify that the message only contains tool_use items
         tool_use_items = [item for item in content_list if item.get("type") == "tool_use"]
         assert len(tool_use_items) > 0, "Assistant message should have at least one tool_use item"
-        assert len(content_list) == len(tool_use_items), (
-            f"Assistant message should ONLY contain tool_use items when reasoning is disabled. "
-            f"Found {len(content_list)} total items but only {len(tool_use_items)} are tool_use items."
-        )
+
+        if not reasoning_enabled:
+            assert len(content_list) == len(tool_use_items), (
+                f"Assistant message should ONLY contain tool_use items when reasoning is disabled. "
+                f"Found {len(content_list)} total items but only {len(tool_use_items)} are tool_use items."
+            )
 
 
 def validate_google_format_scrubbing(contents: List[Dict[str, Any]]) -> None:
@@ -1983,7 +1985,7 @@ def test_inner_thoughts_toggle_interleaved(
         validate_openai_format_scrubbing(messages)
     elif llm_config.model_endpoint_type == "anthropic":
         messages = response["messages"]
-        validate_anthropic_format_scrubbing(messages)
+        validate_anthropic_format_scrubbing(messages, llm_config.enable_reasoner)
     elif llm_config.model_endpoint_type in ["google_ai", "google_vertex"]:
         # Google uses 'contents' instead of 'messages'
         contents = response.get("contents", response.get("messages", []))