fix(ai): normalize SDK objects before truncating messages

sentry_sdk/ai/utils.py

@@ -572,7 +572,7 @@ def _truncate_single_message_content_if_present(
     return message
 
 
-def _find_truncation_index(messages: "List[Dict[str, Any]]", max_bytes: int) -> int:
+def _find_truncation_index(messages: "List[Any]", max_bytes: int) -> int:
     """
     Find the index of the first message that would exceed the max bytes limit.
     Compute the individual message sizes, and return the index of the first message from the back
@@ -668,10 +668,10 @@ def redact_blob_message_parts(
 
 
 def truncate_messages_by_size(
-    messages: "List[Dict[str, Any]]",
+    messages: "List[Any]",
     max_bytes: int = MAX_GEN_AI_MESSAGE_BYTES,
     max_single_message_chars: int = MAX_SINGLE_MESSAGE_CONTENT_CHARS,
-) -> "Tuple[List[Dict[str, Any]], int]":
+) -> "Tuple[List[Any], int]":
     """
     Returns a truncated messages list, consisting of
     - the last message, with its content truncated to `max_single_message_chars` characters,
@@ -682,18 +682,22 @@ def truncate_messages_by_size(
     In the single message case, the serialized message size may exceed `max_bytes`, because
     truncation is based only on character count in that case.
     """
-    serialized_json = json.dumps(messages, separators=(",", ":"))
+    normalized_messages = _normalize_data(messages, unpack=False)
+    if not isinstance(normalized_messages, list):
+        normalized_messages = [normalized_messages]
+
+    serialized_json = json.dumps(normalized_messages, separators=(",", ":"))
     current_size = len(serialized_json.encode("utf-8"))
 
     if current_size <= max_bytes:
-        return messages, 0
+        return normalized_messages, 0
 
-    truncation_index = _find_truncation_index(messages, max_bytes)
-    if truncation_index < len(messages):
-        truncated_messages = messages[truncation_index:]
+    truncation_index = _find_truncation_index(normalized_messages, max_bytes)
+    if truncation_index < len(normalized_messages):
+        truncated_messages = normalized_messages[truncation_index:]
     else:
-        truncation_index = len(messages) - 1
-        truncated_messages = messages[-1:]
+        truncation_index = len(normalized_messages) - 1
+        truncated_messages = normalized_messages[-1:]
 
     if len(truncated_messages) == 1:
         truncated_messages[0] = _truncate_single_message_content_if_present(

tests/test_ai_monitoring.py

@@ -403,6 +403,27 @@ def test_single_message_truncation_list_content_multiple_text_parts(self):
         # Second part gets truncated to 0 chars + ellipsis
         assert parts[1]["text"] == "..."
 
+    def test_truncate_handles_pydantic_like_objects(self):
+        """Pydantic-like SDK objects should be normalized before JSON sizing."""
+
+        class PydanticLike:
+            def __init__(self, payload):
+                self.payload = payload
+
+            def model_dump(self):
+                return self.payload
+
+        messages = [
+            {"role": "user", "content": "hello"},
+            PydanticLike({"type": "function_call", "name": "notify"}),
+        ]
+
+        result, truncation_index = truncate_messages_by_size(messages, max_bytes=10_000)
+
+        assert truncation_index == 0
+        assert len(result) == 2
+        assert result[1] == {"type": "function_call", "name": "notify"}
+
     @pytest.mark.parametrize("content", [None, 42, 3.14, True])
     def test_single_message_truncation_non_str_non_list_content(self, content):
         messages = [{"role": "user", "content": content}]