fix(model_armor): scan each responses field once, not twice

Separating delta fields stopped the terminal body from matching the delta text, so a turn
with two visible fields sent Model Armor both copies. Only the delta fields the body does not
already carry are appended now.

litellm/proxy/guardrails/guardrail_hooks/model_armor/model_armor.py

@@ -984,8 +984,8 @@ class ModelArmorGuardrail(CustomGuardrail, VertexBase):
         return tuple(str(getattr(chunk, attr, None)) for attr in _RESPONSES_DELTA_FIELD_ATTRS)
 
     @staticmethod
-    def _responses_delta_text(all_chunks: Sequence[object]) -> str:
-        """Text a ``/v1/responses`` stream has already spelled out in its delta events.
+    def _responses_delta_field_texts(all_chunks: Sequence[object]) -> tuple[str, ...]:
+        """Text each field of a ``/v1/responses`` turn has already spelled out in its delta events.
 
         One field's deltas are joined as they streamed, since a finding can be split across them,
         and separate fields stay apart, so a reasoning summary running into the visible answer
@@ -997,7 +997,7 @@ class ModelArmorGuardrail(CustomGuardrail, VertexBase):
             if getattr(chunk, "type", None) in _RESPONSES_DELTA_EVENT_TYPES
             and isinstance(delta := getattr(chunk, "delta", None), str)
         )
-        return "\n".join(
+        return tuple(
             "".join(delta for field, delta in deltas if field == streamed_field)
             for streamed_field in dict.fromkeys(field for field, _ in deltas)
         )
@@ -1011,17 +1011,14 @@ class ModelArmorGuardrail(CustomGuardrail, VertexBase):
         """Text to scan for a buffered stream, which is everything the client is about to receive.
 
         A ``/v1/responses`` stream also spells out reasoning summaries and tool-call arguments in
-        delta events that its terminal body never repeats, so body and deltas are scanned together.
+        delta events that its terminal body never repeats, so every delta field the body does not
+        already carry is scanned after it.
         """
         content: Final = self._extract_streaming_content(assembled_response)
         if surface is not _StreamSurface.RESPONSES:
             return content
-        delta_text: Final = self._responses_delta_text(all_chunks)
-        if delta_text in content:
-            return content
-        if content in delta_text:
-            return delta_text
-        return f"{content}\n{delta_text}"
+        unscanned: Final = tuple(text for text in self._responses_delta_field_texts(all_chunks) if text not in content)
+        return "\n".join(part for part in (content, *unscanned) if part)
 
     @staticmethod
     def _apply_sanitized_content(assembled_response: ModelResponse, sanitized_content: str) -> None:

tests/test_litellm/proxy/guardrails/guardrail_hooks/test_model_armor.py

@@ -4855,6 +4855,66 @@ async def test_streaming_responses_one_fields_deltas_still_join_into_a_single_fi
     assert "Streaming response blocked by Model Armor" in rendered
 
 
+@pytest.mark.asyncio
+async def test_streaming_responses_fields_the_body_repeats_are_not_scanned_a_second_time():
+    """A turn whose visible fields all reach the terminal body is scanned once, not twice.
+
+    Two output_text fields stream as deltas and come back in the completed body, so scanning the
+    deltas on top of the body would send Model Armor two copies of everything the client sees.
+    """
+    from litellm.types.llms.openai import (
+        OutputTextDeltaEvent,
+        ResponseCompletedEvent,
+        ResponsesAPIResponse,
+        ResponsesAPIStreamEvents,
+    )
+
+    paragraphs = ("the first thing to know", "a second and separate point")
+    text_deltas = tuple(
+        OutputTextDeltaEvent(
+            type=ResponsesAPIStreamEvents.OUTPUT_TEXT_DELTA,
+            item_id=f"msg_{index}",
+            output_index=index,
+            content_index=0,
+            delta=paragraph,
+        )
+        for index, paragraph in enumerate(paragraphs)
+    )
+    completed = ResponseCompletedEvent(
+        type=ResponsesAPIStreamEvents.RESPONSE_COMPLETED,
+        response=ResponsesAPIResponse(
+            id="resp_1",
+            created_at=0,
+            model="gpt-5-mini",
+            object="response",
+            output=[
+                {
+                    "type": "message",
+                    "id": f"msg_{index}",
+                    "role": "assistant",
+                    "status": "completed",
+                    "content": [{"type": "output_text", "text": paragraph, "annotations": []}],
+                }
+                for index, paragraph in enumerate(paragraphs)
+            ],
+            parallel_tool_calls=False,
+            tool_choice="auto",
+            tools=[],
+        ),
+    )
+    guardrail = _surface_guardrail()
+    post = _armor_post_mock(_MODEL_ARMOR_CLEAN)
+
+    with patch.object(guardrail.async_handler, "post", post):
+        delivered = await _drain_surface_hook(guardrail, (*text_deltas, completed))
+
+    post.assert_called_once()
+    scanned = post.call_args.kwargs["json"]["modelResponseData"]["text"]
+    assert [scanned.count(paragraph) for paragraph in paragraphs] == [1, 1]
+    rendered = "".join(str(item) for item in delivered)
+    assert all(paragraph in rendered for paragraph in paragraphs)
+
+
 def test_every_responses_delta_event_is_in_the_scanned_set():
     """Every ``.delta`` the Responses event enum defines is model output on its way to the client."""
     from litellm.proxy.guardrails.guardrail_hooks.model_armor.model_armor import (