refactor: clear fresh tech debt from the last 24 hours (2026-09-05, 2026-09-06)

Drop the TID251 cast import and both cast-ok casts from the refusal
message_delta rebuild by narrowing the TypedDict union on its type literal,
drop the redundant Mapping cast after the isinstance check in
_mapping_field, and type the Lyria predict read-only helpers as
Mapping[str, object] instead of a bare dict with mutable-ok.

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

litellm/llms/anthropic/experimental_pass_through/adapters/streaming_iterator.py

@@ -11,7 +11,6 @@ from typing import (
     Final,
     Literal,
     Protocol,
-    cast,  # noqa: TID251  # rebuilt message_delta dict spans the ContentBlockDelta/MessageBlockDelta union
     get_args,
 )
 
@@ -25,6 +24,7 @@ from litellm.types.llms.anthropic import (
     ContentBlockDelta,
     ContextManagementResponse,
     MessageBlockDelta,
+    MessageDelta,
     StreamingContentBlockDeltaType,
     UsageDelta,
     UsageIteration,
@@ -1006,26 +1006,28 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
         self,
         processed_chunk: ContentBlockDelta | MessageBlockDelta,
     ) -> ContentBlockDelta | MessageBlockDelta:
-        if processed_chunk.get("type") != "message_delta" or not self._refusal_text:
+        if processed_chunk["type"] != "message_delta" or not self._refusal_text:
             return processed_chunk
-        delta: Final = cast(Mapping[str, object], processed_chunk["delta"])  # cast-ok: keys checked before use
+        delta: Final = processed_chunk["delta"]
         if delta.get("stop_reason") == "max_tokens":
             return processed_chunk
         from litellm.llms.anthropic.experimental_pass_through.messages.utils import (
             refusal_stop_details,
         )
 
-        return cast(  # cast-ok: rebuilt dict matches the message_delta TypedDict shape for this branch
-            ContentBlockDelta | MessageBlockDelta,
-            {  # mutable-ok: fresh translation payload; never mutated after construction
-                **processed_chunk,
-                "delta": {  # mutable-ok: fresh message_delta payload; never mutated after construction
-                    **delta,
-                    "stop_reason": "refusal",
-                    "stop_details": refusal_stop_details(self._refusal_text),
-                },
-            },
-        )
+        refusal_delta: Final[MessageDelta] = {  # mutable-ok: fresh message_delta payload
+            **delta,
+            "stop_reason": "refusal",
+            "stop_details": refusal_stop_details(self._refusal_text),
+        }
+        if "context_management" in processed_chunk:
+            return MessageBlockDelta(
+                type="message_delta",
+                delta=refusal_delta,
+                usage=processed_chunk["usage"],
+                context_management=processed_chunk["context_management"],
+            )
+        return MessageBlockDelta(type="message_delta", delta=refusal_delta, usage=processed_chunk["usage"])
 
     @staticmethod
     def _delta_has_content(processed_chunk: Mapping[str, object]) -> bool:

litellm/llms/anthropic/experimental_pass_through/messages/utils.py

@@ -37,7 +37,7 @@ def _mapping_field(container: object, key: str) -> object | None:
     """One key of a raw provider payload, or None when the payload is not a mapping."""
     if not isinstance(container, Mapping):
         return None
-    return cast(Mapping[str, object], container).get(key)  # cast-ok: raw payload, callers re-check every value
+    return container.get(key)
 
 
 def _mapping_str_field(container: object, key: str) -> str | None:

litellm/proxy/pass_through_endpoints/llm_provider_handlers/vertex_passthrough_logging_handler.py

@@ -1,5 +1,6 @@
 import asyncio
 import re
+from collections.abc import Mapping
 from datetime import datetime
 from typing import TYPE_CHECKING, Any, Final, cast
 from urllib.parse import urlparse
@@ -380,7 +381,7 @@ class VertexPassthroughLoggingHandler:
     @staticmethod
     def _is_audio_predict_response(
         model: str,
-        json_response: dict,  # mutable-ok: predicate inspects the decoded provider response dictionary without mutation
+        json_response: Mapping[str, object],
     ) -> bool:
         return (
             VertexPassthroughLoggingHandler._get_audio_prediction_count(json_response=json_response) > 0
@@ -389,7 +390,7 @@ class VertexPassthroughLoggingHandler:
 
     @staticmethod
     def _get_audio_prediction_count(
-        json_response: dict,  # mutable-ok: counter inspects the decoded provider response dictionary without mutation
+        json_response: Mapping[str, object],
     ) -> int:
         predictions: Final = json_response.get("predictions")
         if not isinstance(predictions, list):