bug(llmo): fix anthropic tool call response (#297)

* bug(llmo): fix anthropic's tool call response

* bug(llmo): fix tool calls response handling for anthropic

* bug(llmo): run formatter

* bug(llmo): bump version

* bug(llmo): add date to changelog

Author: Radu-Raicea

CHANGELOG.md

@@ -1,3 +1,7 @@
+# 6.3.2 - 2025-07-31
+
+- fix: Anthropic's tool calls are now handled properly
+
 # 6.3.0 - 2025-07-22
 
 - feat: Enhanced `send_feature_flags` parameter to accept `SendFeatureFlagsOptions` object for declarative control over local/remote evaluation and custom properties

posthog/ai/utils.py

@@ -118,7 +118,12 @@ def format_response(response, provider: str):
 def format_response_anthropic(response):
     output = []
     for choice in response.content:
-        if choice.text:
+        if (
+            hasattr(choice, "type")
+            and choice.type == "text"
+            and hasattr(choice, "text")
+            and choice.text
+        ):
             output.append(
                 {
                     "role": "assistant",
@@ -225,8 +230,21 @@ def format_response_gemini(response):
 
 def format_tool_calls(response, provider: str):
     if provider == "anthropic":
-        if hasattr(response, "tools") and response.tools and len(response.tools) > 0:
-            return response.tools
+        if hasattr(response, "content") and response.content:
+            tool_calls = []
+
+            for content_item in response.content:
+                if hasattr(content_item, "type") and content_item.type == "tool_use":
+                    tool_calls.append(
+                        {
+                            "type": content_item.type,
+                            "id": content_item.id,
+                            "name": content_item.name,
+                            "input": content_item.input,
+                        }
+                    )
+
+            return tool_calls if tool_calls else None
     elif provider == "openai":
         # Handle both Chat Completions and Responses API
         if hasattr(response, "choices") and response.choices:
@@ -378,6 +396,7 @@ def call_llm_and_track_usage(
         }
 
         tool_calls = format_tool_calls(response, provider)
+
         if tool_calls:
             event_properties["$ai_tools"] = with_privacy_mode(
                 ph_client, posthog_privacy_mode, tool_calls

posthog/test/ai/anthropic/test_anthropic.py

@@ -88,6 +88,31 @@ def mock_anthropic_response_with_cached_tokens():
     )
 
 
+@pytest.fixture
+def mock_anthropic_response_with_tool_use():
+    return Message(
+        id="msg_123",
+        type="message",
+        role="assistant",
+        content=[
+            {"type": "text", "text": "I'll help you with that."},
+            {
+                "type": "tool_use",
+                "id": "tool_1",
+                "name": "get_weather",
+                "input": {"location": "New York"},
+            },
+        ],
+        model="claude-3-opus-20240229",
+        usage=Usage(
+            input_tokens=20,
+            output_tokens=10,
+        ),
+        stop_reason="end_turn",
+        stop_sequence=None,
+    )
+
+
 def test_basic_completion(mock_client, mock_anthropic_response):
     with patch(
         "anthropic.resources.Messages.create", return_value=mock_anthropic_response
@@ -434,3 +459,49 @@ def test_cached_tokens(mock_client, mock_anthropic_response_with_cached_tokens):
         assert props["$ai_http_status"] == 200
         assert props["foo"] == "bar"
         assert isinstance(props["$ai_latency"], float)
+
+
+def test_tool_use_response(mock_client, mock_anthropic_response_with_tool_use):
+    with patch(
+        "anthropic.resources.Messages.create",
+        return_value=mock_anthropic_response_with_tool_use,
+    ):
+        client = Anthropic(api_key="test-key", posthog_client=mock_client)
+        response = client.messages.create(
+            model="claude-3-opus-20240229",
+            messages=[{"role": "user", "content": "What's the weather like?"}],
+            posthog_distinct_id="test-id",
+            posthog_properties={"foo": "bar"},
+        )
+
+        assert response == mock_anthropic_response_with_tool_use
+        assert mock_client.capture.call_count == 1
+
+        call_args = mock_client.capture.call_args[1]
+        props = call_args["properties"]
+
+        assert call_args["distinct_id"] == "test-id"
+        assert call_args["event"] == "$ai_generation"
+        assert props["$ai_provider"] == "anthropic"
+        assert props["$ai_model"] == "claude-3-opus-20240229"
+        assert props["$ai_input"] == [
+            {"role": "user", "content": "What's the weather like?"}
+        ]
+        # Should only include text content, not tool_use content
+        assert props["$ai_output_choices"] == [
+            {"role": "assistant", "content": "I'll help you with that."}
+        ]
+        assert props["$ai_input_tokens"] == 20
+        assert props["$ai_output_tokens"] == 10
+        assert props["$ai_http_status"] == 200
+        assert props["foo"] == "bar"
+        assert isinstance(props["$ai_latency"], float)
+        # Verify that tools are captured separately
+        assert props["$ai_tools"] == [
+            {
+                "type": "tool_use",
+                "id": "tool_1",
+                "name": "get_weather",
+                "input": {"location": "New York"},
+            }
+        ]

posthog/version.py

@@ -1,4 +1,4 @@
-VERSION = "6.3.1"
+VERSION = "6.3.2"
 
 if __name__ == "__main__":
     print(VERSION, end="")  # noqa: T201