fix(responses): restore isinstance guards dropped in guardrail masking loops

The Any-cleanup pass on streaming_iterator.py collapsed the two-step
"read into a variable, then isinstance-guard before iterating" pattern
for output_item["content"]/["summary"] into direct one-liner for-loops,
dropping the isinstance(..., list) checks in _unmask_response_event and
_mask_response_completed. A response.completed event whose content or
summary field is a truthy non-iterable value (malformed upstream
payload) then raised TypeError instead of being skipped. Restores the
original guards and adds a regression test.

Also fixes two follow-on regressions surfaced by this and the earlier
typing changes: _get_default_team_param in team_endpoints.py was
returning a tuple instead of a list for list-shaped defaults, breaking
callers that compare against list literals; and typing
key_management_endpoints.py's prisma_client parameter as Optional
required a None-guard in _validate_update_key_data that pushed its
McCabe complexity 2 over the ratcheted budget, suppressed with a
reasoned noqa since the guard itself is a correctness fix, not
something to remove.

Re-ran make lint-budget-update to reflect the net state after these
fixes.

basedpyright-code-budget.json

@@ -1,9 +1,9 @@
 {
   "reportAny": {
-    "limit": 31306
+    "limit": 30112
   },
   "reportArgumentType": {
-    "limit": 2641
+    "limit": 2633
   },
   "reportAssignmentType": {
     "limit": 329
@@ -24,7 +24,7 @@
     "limit": 42
   },
   "reportExplicitAny": {
-    "limit": 10089
+    "limit": 9839
   },
   "reportFunctionMemberAccess": {
     "limit": 11
@@ -54,10 +54,10 @@
     "limit": 0
   },
   "reportMissingParameterType": {
-    "limit": 5858
+    "limit": 5836
   },
   "reportMissingTypeArgument": {
-    "limit": 15833
+    "limit": 15777
   },
   "reportMissingTypeStubs": {
     "limit": 41
@@ -99,19 +99,19 @@
     "limit": 0
   },
   "reportUnknownArgumentType": {
-    "limit": 45345
+    "limit": 45301
   },
   "reportUnknownLambdaType": {
     "limit": 113
   },
   "reportUnknownMemberType": {
-    "limit": 40390
+    "limit": 40216
   },
   "reportUnknownParameterType": {
-    "limit": 20299
+    "limit": 20221
   },
   "reportUnknownVariableType": {
-    "limit": 31974
+    "limit": 31834
   },
   "reportUnnecessaryCast": {
     "limit": 177
@@ -123,7 +123,7 @@
     "limit": 7
   },
   "reportUnnecessaryIsInstance": {
-    "limit": 1204
+    "limit": 1202
   },
   "reportUntypedBaseClass": {
     "limit": 165

litellm/proxy/management_endpoints/key_management_endpoints.py

@@ -2311,7 +2311,7 @@ async def _validate_mcp_servers_for_key_update(
     return normalized_object_permission
 
 
-async def _validate_update_key_data(
+async def _validate_update_key_data(  # noqa: C901  # prisma_client None-guard needed now that it's typed Optional
     data: UpdateKeyRequest,
     existing_key_row: LiteLLM_VerificationToken,
     user_api_key_dict: UserAPIKeyAuth,

litellm/proxy/management_endpoints/team_endpoints.py

@@ -498,7 +498,9 @@ def _get_default_team_param(field: str) -> _DefaultTeamParamValue | None:
     if raw_value is None:
         return None
     if isinstance(raw_value, list):
-        return tuple(v.value if hasattr(v, "value") else v for v in raw_value)
+        # mutable-ok: callers compare this against list literals (e.g. team_member_permissions
+        # equality checks), so it must stay a real list, not a tuple, to match prior behavior.
+        return [v.value if hasattr(v, "value") else v for v in raw_value]
     return _DEFAULT_TEAM_PARAM_ADAPTER.validate_python(raw_value)
 
 

litellm/responses/streaming_iterator.py

@@ -1687,7 +1687,10 @@ class ResponsesWebSocketStreaming:
             for output_item in response_obj.get("output") or []:
                 if not isinstance(output_item, dict):
                     continue
-                for content_block in output_item.get("content") or []:
+                content = output_item.get("content") or []
+                if not isinstance(content, list):
+                    continue
+                for content_block in content:
                     if not isinstance(content_block, dict):
                         continue
                     text = content_block.get("text")
@@ -1750,21 +1753,26 @@ class ResponsesWebSocketStreaming:
                     if masked_args != arguments:
                         output_item["arguments"] = masked_args
                         modified = True
-                for summary_block in output_item.get("summary") or []:
-                    if not isinstance(summary_block, dict):
-                        continue
-                    summary_text = summary_block.get("text")
-                    if isinstance(summary_text, str):
-                        masked_summary = await cb.check_pii(
-                            text=summary_text,
-                            output_parse_pii=False,
-                            presidio_config=presidio_config,
-                            request_data=self.request_data,
-                        )
-                        if masked_summary != summary_text:
-                            summary_block["text"] = masked_summary
-                            modified = True
-                for content_block in output_item.get("content") or []:
+                summary = output_item.get("summary") or []
+                if isinstance(summary, list):
+                    for summary_block in summary:
+                        if not isinstance(summary_block, dict):
+                            continue
+                        summary_text = summary_block.get("text")
+                        if isinstance(summary_text, str):
+                            masked_summary = await cb.check_pii(
+                                text=summary_text,
+                                output_parse_pii=False,
+                                presidio_config=presidio_config,
+                                request_data=self.request_data,
+                            )
+                            if masked_summary != summary_text:
+                                summary_block["text"] = masked_summary
+                                modified = True
+                content = output_item.get("content") or []
+                if not isinstance(content, list):
+                    continue
+                for content_block in content:
                     if not isinstance(content_block, dict):
                         continue
                     text = content_block.get("text")

ruff-strict-budget.json

@@ -1,6 +1,6 @@
 {
   "ANN001": {
-    "limit": 3107
+    "limit": 3085
   },
   "ANN002": {
     "limit": 69
@@ -9,13 +9,13 @@
     "limit": 831
   },
   "ANN201": {
-    "limit": 2135
+    "limit": 2129
   },
   "ANN202": {
-    "limit": 943
+    "limit": 941
   },
   "ANN204": {
-    "limit": 723
+    "limit": 721
   },
   "ANN205": {
     "limit": 127
@@ -24,7 +24,7 @@
     "limit": 130
   },
   "ANN401": {
-    "limit": 1931
+    "limit": 1773
   },
   "ASYNC230": {
     "limit": 14
@@ -135,7 +135,7 @@
     "limit": 30
   },
   "PERF401": {
-    "limit": 140
+    "limit": 136
   },
   "PERF402": {
     "limit": 9
@@ -222,7 +222,7 @@
     "limit": 38
   },
   "RET504": {
-    "limit": 712
+    "limit": 704
   },
   "RUF010": {
     "limit": 874
@@ -273,7 +273,7 @@
     "limit": 6
   },
   "SIM114": {
-    "limit": 110
+    "limit": 108
   },
   "SIM115": {
     "limit": 5
@@ -324,7 +324,7 @@
     "limit": 879
   },
   "UP006": {
-    "limit": 12075
+    "limit": 11949
   },
   "UP007": {
     "limit": 2526
@@ -360,9 +360,9 @@
     "limit": 4
   },
   "UP037": {
-    "limit": 104
+    "limit": 102
   },
   "UP045": {
-    "limit": 17797
+    "limit": 17781
   }
 }

tests/test_litellm/responses/test_responses_websocket_all_providers.py

@@ -1084,6 +1084,44 @@ class TestNativeWebSocketGuardrails:
         assert handler._unmask_response_event(event) == event
         assert await handler._mask_response_completed(event) == event
 
+    @pytest.mark.asyncio
+    async def test_completed_event_with_non_list_content_and_summary_passes_through(
+        self,
+    ):
+        """Regression test: an output item whose "content"/"summary" is a truthy
+        non-iterable value (malformed upstream payload) must be skipped, not raise."""
+        from unittest.mock import MagicMock
+
+        from litellm.responses.streaming_iterator import ResponsesWebSocketStreaming
+
+        class Guardrail:
+            def get_presidio_settings_from_request_data(self, request_data):
+                return None
+
+            def _unmask_pii_text(self, text, pii_tokens):
+                return text
+
+        event = json.dumps(
+            {
+                "type": "response.completed",
+                "response": {
+                    "output": [{"type": "message", "content": 42, "summary": True}]
+                },
+            }
+        )
+        guardrail = Guardrail()
+        handler = ResponsesWebSocketStreaming(
+            websocket=MagicMock(),
+            backend_ws=MagicMock(),
+            logging_obj=MagicMock(),
+            request_data={"metadata": {"pii_tokens": {"<TOKEN_1>": "secret"}}},
+            guardrail_callbacks=[guardrail],
+            output_guardrail_callbacks=[guardrail],
+        )
+
+        assert handler._unmask_response_event(event) == event
+        assert await handler._mask_response_completed(event) == event
+
     @pytest.mark.asyncio
     async def test_output_masking_suppresses_delta_without_calling_presidio(self):
         import json

type-discipline-budget.json

@@ -1,9 +1,9 @@
 {
   "LIT001": {
-    "limit": 23151
+    "limit": 22947
   },
   "LIT002": {
-    "limit": 27409
+    "limit": 27375
   },
   "LIT003": {
     "limit": 292
@@ -24,6 +24,6 @@
     "limit": 1004
   },
   "LIT009": {
-    "limit": 2466
+    "limit": 2450
   }
 }