fix(guardrails): inspect responses reasoning content and summary text

litellm/proxy/guardrails/_content_utils.py

@@ -8,7 +8,7 @@ skip the other shapes — these helpers normalise that so every hook sees
 every text fragment.
 """
 
-from collections.abc import Callable, Iterator
+from collections.abc import Callable, Iterator, Mapping
 from typing import Any, Final
 
 # Call types whose body carries free-form chat / prompt text that
@@ -33,7 +33,9 @@ def is_text_content_call_type(call_type: str) -> bool:
     return call_type in TEXT_CONTENT_CALL_TYPES
 
 
-TEXT_PART_TYPES: Final[frozenset[str]] = frozenset({"text", "input_text", "output_text"})
+TEXT_PART_TYPES: Final[frozenset[str]] = frozenset(
+    {"text", "input_text", "output_text", "summary_text", "reasoning_text"}
+)
 
 # Responses-API item types whose ``output`` field carries user/tool text
 # that guardrails should inspect.  ``function_call_output`` is the
@@ -42,6 +44,16 @@ TEXT_PART_TYPES: Final[frozenset[str]] = frozenset({"text", "input_text", "outpu
 _OUTPUT_ITEM_TYPES: Final[frozenset[str]] = frozenset({"function_call_output", "custom_tool_call_output"})
 
 
+def _part_text(part: Mapping[str, object]) -> str | None:
+    """Return non-empty plaintext from any content part that carries ``text``."""
+    if not isinstance(part, dict):
+        return None
+    text = part.get("text")
+    if isinstance(text, str) and text:
+        return text
+    return None
+
+
 def _iter_text_parts_in_content(content: Any) -> Iterator[str]:
     """Yield text fragments from a ``message.content`` value (string or
     multimodal list). Non-text parts (images, audio, …) are skipped."""
@@ -58,10 +70,9 @@ def _iter_text_parts_in_content(content: Any) -> Iterator[str]:
                 continue
             if not isinstance(part, dict):
                 continue
-            if part.get("type") in TEXT_PART_TYPES:
-                text = part.get("text")
-                if isinstance(text, str) and text:
-                    yield text
+            text = _part_text(part)
+            if text is not None:
+                yield text
 
 
 def _coerce_input_to_messages(input_value: Any) -> list[dict[str, Any]]:
@@ -75,8 +86,23 @@ def _coerce_input_to_messages(input_value: Any) -> list[dict[str, Any]]:
         if isinstance(item, str):
             messages.append({"role": "user", "content": item})
         elif isinstance(item, dict):
-            if item.get("type") in TEXT_PART_TYPES:
+            if _part_text(item) is not None:
                 messages.append({"role": item.get("role") or "user", "content": [item]})
+            elif item.get("type") == "reasoning":
+                if "content" in item:
+                    messages.append(
+                        {  # mutable-ok: append reasoning content
+                            "role": item.get("role") or "assistant",
+                            "content": item["content"],
+                        }
+                    )
+                if isinstance(item.get("summary"), list):
+                    messages.append(
+                        {  # mutable-ok: append reasoning summary
+                            "role": item.get("role") or "assistant",
+                            "content": item["summary"],
+                        }
+                    )
             elif "content" in item:
                 messages.append({"role": item.get("role") or "user", "content": item["content"]})
             elif item.get("type") in _OUTPUT_ITEM_TYPES and "output" in item:
@@ -126,12 +152,7 @@ def walk_user_text(data: dict[str, Any], visit: Callable[[str], str]) -> int:
                 if isinstance(part, str) and part:
                     visited += 1
                     new_parts.append(visit(part))
-                elif (
-                    isinstance(part, dict)
-                    and part.get("type") in TEXT_PART_TYPES
-                    and isinstance(part.get("text"), str)
-                    and part["text"]
-                ):
+                elif isinstance(part, dict) and _part_text(part) is not None:
                     visited += 1
                     new_parts.append({**part, "text": visit(part["text"])})
                 else:
@@ -158,10 +179,14 @@ def walk_user_text(data: dict[str, Any], visit: Callable[[str], str]) -> int:
                     visited += 1
                     input_value[idx] = visit(item)
             elif isinstance(item, dict):
-                if item.get("type") in TEXT_PART_TYPES:
-                    if isinstance(item.get("text"), str) and item["text"]:
-                        visited += 1
-                        input_value[idx] = {**item, "text": visit(item["text"])}
+                if _part_text(item) is not None:
+                    visited += 1
+                    input_value[idx] = {**item, "text": visit(item["text"])}  # mutable-ok: rewrite text part in place
+                elif item.get("type") == "reasoning":
+                    if "content" in item:
+                        item["content"] = _rewrite_content(item["content"])
+                    if isinstance(item.get("summary"), list):
+                        item["summary"] = _rewrite_content(item["summary"])
                 elif "content" in item:
                     item["content"] = _rewrite_content(item["content"])
                 elif item.get("type") in _OUTPUT_ITEM_TYPES and "output" in item:

litellm/responses/litellm_completion_transformation/transformation.py

@@ -638,7 +638,7 @@ class LiteLLMCompletionResponsesConfig:
                 # Not followed by an assistant message — keep the reasoning
                 # standalone instead of dropping it.
                 merged.extend(  # mutable-ok: append reasoning messages
-                    [
+                    [  # mutable-ok: append reasoning messages
                         ChatCompletionResponseMessage(
                             role="assistant",
                             content=None,
@@ -652,7 +652,7 @@ class LiteLLMCompletionResponsesConfig:
             merged.append(msg)
 
         merged.extend(  # mutable-ok: append trailing reasoning
-            [
+            [  # mutable-ok: append trailing reasoning
                 ChatCompletionResponseMessage(
                     role="assistant",
                     content=None,
@@ -1196,6 +1196,8 @@ class LiteLLMCompletionResponsesConfig:
             if text_parts:
                 return "\n".join(text_parts)
 
+        # Guardrail traversal in litellm/proxy/guardrails/_content_utils.py
+        # inspects and rewrites these summary blocks before they are forwarded.
         summary: Final[object] = input_item.get("summary")
         if isinstance(summary, list):
             text_parts = []  # mutable-ok: text accumulator  # rebind-ok: text accumulator

tests/test_litellm/proxy/guardrails/test_content_utils.py

@@ -149,6 +149,22 @@ def test_iter_message_text_responses_api_tool_call_taxonomy():
     assert list(iter_message_text(data)) == ["hello", "sunny"]
 
 
+def test_iter_message_text_inspects_reasoning_content_and_summary():
+    """VERIA: reasoning items forwarded as ``reasoning_content`` must be
+    inspected, including ``summary`` blocks the bridge reads as a fallback."""
+    data = {
+        "input": [
+            {
+                "type": "reasoning",
+                "id": "rs_1",
+                "content": [{"type": "summary_text", "text": "content secret"}],
+                "summary": [{"type": "summary_text", "text": "summary secret"}],
+            }
+        ]
+    }
+    assert list(iter_message_text(data)) == ["content secret", "summary secret"]
+
+
 # ── walk_user_text ────────────────────────────────────────────────────────────
 
 
@@ -308,6 +324,27 @@ def test_walk_user_text_redacts_mixed_list_input():
     assert data["input"][2] == {"type": "image_url", "image_url": {"url": "..."}}
 
 
+def test_walk_user_text_redacts_reasoning_content_and_summary():
+    """VERIA: in-place redaction must cover both plaintext shapes the bridge
+    forwards from a reasoning item."""
+    data = {
+        "input": [
+            {
+                "type": "reasoning",
+                "id": "rs_1",
+                "content": [{"type": "summary_text", "text": "AKIAEXAMPLE content"}],
+                "summary": [{"type": "summary_text", "text": "AKIAEXAMPLE summary"}],
+            }
+        ]
+    }
+    visited = walk_user_text(data, lambda s: s.replace("AKIAEXAMPLE", "[REDACTED]"))
+    assert visited == 2
+    item = data["input"][0]
+    assert item["content"][0]["text"] == "[REDACTED] content"
+    assert item["summary"][0]["text"] == "[REDACTED] summary"
+    assert item["id"] == "rs_1"
+
+
 # ── build_inspection_messages ─────────────────────────────────────────────────
 
 
@@ -462,6 +499,23 @@ def test_build_inspection_messages_empty_data():
     assert build_inspection_messages({"input": ""}) == []
 
 
+def test_build_inspection_messages_includes_reasoning_summary():
+    """VERIA: remote guardrail APIs must see reasoning summaries even when
+    the reasoning item has no ``content`` field."""
+    data = {
+        "input": [
+            {
+                "type": "reasoning",
+                "id": "rs_1",
+                "summary": [{"type": "summary_text", "text": "secret summary"}],
+            }
+        ]
+    }
+    assert build_inspection_messages(data) == [
+        {"role": "assistant", "content": "secret summary"}
+    ]
+
+
 # ── has_non_string_content ────────────────────────────────────────────────────