feat(guardrails): support structured_messages in generic guardrail API response

Guardrails that modify full message structure (e.g. prompt compression)
can now return structured_messages in GUARDRAIL_INTERVENED responses.
LiteLLM passes it back through GenericGuardrailAPIInputs so the
compressed message array reaches the LLM provider instead of only
flat text replacements.

litellm/proxy/guardrails/guardrail_hooks/generic_guardrail_api/generic_guardrail_api.py

@@ -329,8 +329,9 @@ class GenericGuardrailAPI(CustomGuardrail):
         tools: Any,
         guardrail_response: GenericGuardrailAPIResponse,
     ) -> GenericGuardrailAPIInputs:
-        # Action is NONE or no modifications needed
         return_inputs = GenericGuardrailAPIInputs(texts=texts)
+        if guardrail_response.structured_messages:
+            return_inputs["structured_messages"] = guardrail_response.structured_messages
         if guardrail_response.texts:
             return_inputs["texts"] = guardrail_response.texts
         if guardrail_response.images:

litellm/types/proxy/guardrails/guardrail_hooks/generic_guardrail_api.py

@@ -89,6 +89,7 @@ class GenericGuardrailAPIResponse:
     texts: Optional[List[str]]
     images: Optional[List[str]]
     tools: Optional[List[ChatCompletionToolParam]]
+    structured_messages: Optional[List[AllMessageValues]]
     action: str
     blocked_reason: Optional[str]
 
@@ -99,12 +100,14 @@ class GenericGuardrailAPIResponse:
         blocked_reason: Optional[str] = None,
         images: Optional[List[str]] = None,
         tools: Optional[List[ChatCompletionToolParam]] = None,
+        structured_messages: Optional[List[AllMessageValues]] = None,
     ):
         self.action = action
         self.blocked_reason = blocked_reason
         self.texts = texts
         self.images = images
         self.tools = tools
+        self.structured_messages = structured_messages
 
     @classmethod
     def from_dict(cls, data: dict) -> "GenericGuardrailAPIResponse":
@@ -114,4 +117,5 @@ class GenericGuardrailAPIResponse:
             texts=data.get("texts"),
             images=data.get("images"),
             tools=data.get("tools"),
+            structured_messages=data.get("structured_messages"),
         )

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

@@ -581,6 +581,34 @@ class TestGuardrailActions:
             assert result_texts == ["[REDACTED]"]
             assert result_images is None
 
+    @pytest.mark.asyncio
+    async def test_action_intervened_with_structured_messages(
+        self, generic_guardrail, mock_request_data_input
+    ):
+        compressed_messages = [
+            {"role": "user", "content": "Analyze this."},
+            {"role": "user", "content": [{"type": "tool_result", "tool_use_id": "t1", "content": "compressed json"}]},
+        ]
+        mock_response = MagicMock()
+        mock_response.json.return_value = {
+            "action": "GUARDRAIL_INTERVENED",
+            "structured_messages": compressed_messages,
+        }
+        mock_response.raise_for_status = MagicMock()
+
+        with patch.object(
+            generic_guardrail.async_handler, "post", return_value=mock_response
+        ):
+            guardrailed_inputs = await generic_guardrail.apply_guardrail(
+                inputs={"texts": ["Analyze this."], "structured_messages": [
+                    {"role": "user", "content": "Analyze this."},
+                    {"role": "user", "content": [{"type": "tool_result", "tool_use_id": "t1", "content": "very long uncompressed json " * 100}]},
+                ]},
+                request_data=mock_request_data_input,
+                input_type="request",
+            )
+            assert guardrailed_inputs.get("structured_messages") == compressed_messages
+
 
 class TestImageSupport:
     """Test image handling in guardrail requests"""