fix(caching): answer 400 for a non-list /cache/delete keys and accept null Stability lists

Return pydantic errors without the caller's input on a bad /cache/delete body, let the Bedrock Stability
image edit schema take null finish_reasons and images like the old reads did, declare user_id and user_role
on the plugin session claim, wrap the retyped Vertex fields in ReadOnly, and drop unused Guardrails AI types

litellm/llms/bedrock/image_edit/stability_transformation.py

@@ -50,8 +50,8 @@ else:
 
 @with_config(ConfigDict(extra="allow", strict=True))
 class _StabilityResponse(TypedDict):
-    finish_reasons: ReadOnly[NotRequired[Sequence[object]]]
-    images: ReadOnly[NotRequired[Sequence[str | None]]]
+    finish_reasons: ReadOnly[NotRequired[Sequence[object] | None]]
+    images: ReadOnly[NotRequired[Sequence[str | None] | None]]
 
 
 _STABILITY_RESPONSE: Final = TypeAdapter(_StabilityResponse)

litellm/proxy/caching_routes.py

@@ -2,7 +2,7 @@ from collections.abc import Mapping
 from typing import Final
 
 from fastapi import APIRouter, Depends, HTTPException, Request
-from pydantic import ConfigDict, TypeAdapter, with_config
+from pydantic import ConfigDict, TypeAdapter, ValidationError, with_config
 from typing_extensions import NotRequired, ReadOnly, TypedDict
 
 import litellm
@@ -172,6 +172,8 @@ async def cache_delete(request: Request):
                 status_code=500,
                 detail=f"Cache type {litellm.cache.type} does not support deleting a key. only `redis` is supported",
             )
+    except ValidationError as e:
+        raise HTTPException(status_code=400, detail=e.errors(include_url=False, include_input=False))
     except Exception as e:
         raise HTTPException(
             status_code=500,

litellm/proxy/guardrails/guardrail_hooks/guardrails_ai/guardrails_ai.py

@@ -8,12 +8,11 @@
 import json
 import os
 from collections.abc import Sequence
-from typing import TYPE_CHECKING, Final, Literal, TypedDict
+from typing import TYPE_CHECKING, Final, Literal
 
 from fastapi import HTTPException
 from pydantic import ConfigDict, TypeAdapter, with_config
-from typing_extensions import NotRequired, ReadOnly
-from typing_extensions import TypedDict as _ValidatedTypedDict
+from typing_extensions import NotRequired, ReadOnly, TypedDict
 
 import litellm
 from litellm._logging import verbose_proxy_logger
@@ -38,26 +37,13 @@ class GuardrailsAIResponse(TypedDict):
     validationPassed: bool
 
 
-class InferenceData(TypedDict):
-    name: str
-    shape: list[int]
-    data: list
-    datatype: str
-
-
-class GuardrailsAIResponsePreCall(TypedDict):
-    modelname: str
-    modelversion: str
-    outputs: list[InferenceData]
-
-
 @with_config(ConfigDict(extra="allow", strict=True))
-class _PreCallOutput(_ValidatedTypedDict):
+class _PreCallOutput(TypedDict):
     data: ReadOnly[NotRequired[Sequence[str]]]
 
 
 @with_config(ConfigDict(extra="allow", strict=True))
-class _PreCallResponse(_ValidatedTypedDict):
+class _PreCallResponse(TypedDict):
     outputs: ReadOnly[NotRequired[Sequence[_PreCallOutput]]]
 
 

litellm/proxy/plugin_routes.py

@@ -45,6 +45,8 @@ router: Final = APIRouter()
 @with_config(ConfigDict(extra="allow", strict=True))
 class _PluginSessionClaim(TypedDict):
     plugin: NotRequired[ReadOnly[object]]
+    user_id: NotRequired[ReadOnly[str]]
+    user_role: NotRequired[ReadOnly[str]]
     exp: NotRequired[ReadOnly[int | float | str]]
 
 

litellm/types/llms/vertex_ai.py

@@ -140,7 +140,7 @@ class Schema(TypedDict, total=False):
     title: str
     description: str
     nullable: bool
-    default: object
+    default: ReadOnly[object]
     items: "Schema"
     minItems: str
     maxItems: str
@@ -155,7 +155,7 @@ class Schema(TypedDict, total=False):
     minLength: str
     maxLength: str
     pattern: str
-    example: object
+    example: ReadOnly[object]
     anyOf: list["Schema"]
 
 
@@ -794,9 +794,9 @@ class VertexVideoGenerationResponse(TypedDict, total=False):
 
     name: str
     done: bool
-    response: dict[str, object]
-    metadata: dict[str, object]
-    error: dict[str, object]
+    response: ReadOnly[dict[str, object]]
+    metadata: ReadOnly[dict[str, object]]
+    error: ReadOnly[dict[str, object]]
 
 
 VERTEX_CREDENTIALS_TYPES = str | dict[str, str]

tests/unit/llms/bedrock/image_edit/test_stability_transformation.py

@@ -1,3 +1,8 @@
+from unittest.mock import MagicMock
+
+import httpx
+import pytest
+
 from litellm.llms.bedrock.image_edit.stability_transformation import BedrockStabilityImageEditConfig
 
 
@@ -12,3 +17,31 @@ def test_transform_image_edit_request_returns_the_json_body_and_no_files():
     )
 
     assert request == ({"output_format": "png", "prompt": "add a red hat", "image": "iVBORy1ieXRlcw=="}, {})
+
+
+@pytest.mark.parametrize(
+    ("body", "expected_images"),
+    [
+        ({"images": ["aW1hZ2U="], "finish_reasons": [None], "seeds": [42]}, ["aW1hZ2U="]),
+        ({"images": ["aW1hZ2U="], "finish_reasons": None}, ["aW1hZ2U="]),
+        ({"images": None, "finish_reasons": None}, []),
+        ({}, []),
+    ],
+)
+def test_transform_image_edit_response_returns_each_image_and_tolerates_null_lists(body, expected_images):
+    response = BedrockStabilityImageEditConfig().transform_image_edit_response(
+        model="stability.stable-image-inpaint-v1:0",
+        raw_response=httpx.Response(200, json=body),
+        logging_obj=MagicMock(),
+    )
+
+    assert [image.b64_json for image in response.data or []] == expected_images
+
+
+def test_transform_image_edit_response_raises_the_first_finish_reason():
+    with pytest.raises(Exception, match="Bedrock Stability error: Filter reason: prompt"):
+        BedrockStabilityImageEditConfig().transform_image_edit_response(
+            model="stability.stable-image-inpaint-v1:0",
+            raw_response=httpx.Response(200, json={"images": [None], "finish_reasons": ["Filter reason: prompt"]}),
+            logging_obj=MagicMock(),
+        )

tests/unit/proxy/test_caching_routes.py

@@ -70,10 +70,12 @@ def test_cache_delete_passes_string_and_numeric_keys_to_the_cache(mock_redis_del
     mock_redis_delete.delete_cache_keys.assert_awaited_once_with(keys=keys)
 
 
-def test_cache_delete_rejects_keys_that_are_not_a_list(mock_redis_delete):
-    response = client.post("/cache/delete", headers={"Authorization": "Bearer sk-9876"}, json={"keys": "abc"})
+def test_cache_delete_rejects_keys_that_are_not_a_list_with_a_400_that_hides_them(mock_redis_delete):
+    response = client.post("/cache/delete", headers={"Authorization": "Bearer sk-9876"}, json={"keys": "sk-secret-key"})
 
-    assert response.status_code == 500
+    assert response.status_code == 400
+    assert response.json()["detail"][0]["loc"] == ["keys"]
+    assert "sk-secret-key" not in response.text
     mock_redis_delete.delete_cache_keys.assert_not_awaited()