From a215b78d99ea2f853e9bfdf07911f1f94df50ca6 Mon Sep 17 00:00:00 2001
From: "ZOU Yi (BD/SWD-WDE1)" <Yi.ZOU2@cn.bosch.com>
Date: Mon, 14 Sep 2026 11:05:04 +0800
Subject: [PATCH] fix(sap): address review feedback on stream chunk validation

Type the chunk payload as dict[str, object] and narrow choices with an
isinstance guard before iterating, fix the in-place mutation suppression
that claimed the loop never mutated, build the litellm Usage with
model_validate instead of spreading a dict[str, Any], and make the usage
regression test go through model_dump_json, the path that actually raised
MockValSer. model_dump() self-heals via __getattr__, so the old test passed
on unfixed code.

Carry the same typing into to_openai_chunk and the payload literals it
validates, so no basedpyright rule counts higher than before the change.
---
 litellm/llms/sap/chat/handler.py              | 50 ++++++++++---------
 .../llms/sap/chat/test_sap_chat_calls.py      |  5 +-
 2 files changed, 30 insertions(+), 25 deletions(-)

diff --git a/litellm/llms/sap/chat/handler.py b/litellm/llms/sap/chat/handler.py
index 2867b64057b..d0be0086214 100755
--- a/litellm/llms/sap/chat/handler.py
+++ b/litellm/llms/sap/chat/handler.py
@@ -49,14 +49,16 @@ class _StreamParser:
 
     @staticmethod
     def _validate_chunk(
-        payload: dict,  # mutable-ok: normalized in place (pops empty logprobs) before validation
+        payload: dict[str, object],  # mutable-ok: normalized in place (pops empty logprobs) before validation
     ) -> OpenAIChatCompletionChunk:
-        for choice in payload.get("choices") or []:  # mutable-ok: only iterated, never mutated
-            if isinstance(choice, dict) and not choice.get("logprobs"):
-                choice.pop("logprobs", None)
+        choices: Final = payload.get("choices")
+        if isinstance(choices, list):
+            for choice in choices:
+                if isinstance(choice, dict) and not choice.get("logprobs"):
+                    choice.pop("logprobs", None)  # mutable-ok: pops the logprobs key in-place before model_validate
         chunk = OpenAIChatCompletionChunk.model_validate(payload)
         if chunk.usage is not None:
-            chunk.usage = Usage(**chunk.usage.model_dump())
+            chunk.usage = Usage.model_validate(chunk.usage.model_dump())
         return chunk
 
     @staticmethod
@@ -68,25 +70,24 @@ class _StreamParser:
         if not orc:
             return None
 
-        return _StreamParser._validate_chunk(
-            {
-                "id": orc.get("id") or evt.get("request_id") or "stream-chunk",
-                "object": orc.get("object") or "chat.completion.chunk",
-                "created": orc.get("created") or evt.get("created") or _now_ts(),
-                "model": orc.get("model") or "unknown",
-                "choices": [
-                    {
-                        "index": c.get("index", 0),
-                        "delta": c.get("delta") or {},
-                        "finish_reason": c.get("finish_reason"),
-                    }
-                    for c in (orc.get("choices") or [])
-                ],
-            }
-        )
+        payload: Final[dict[str, object]] = {
+            "id": orc.get("id") or evt.get("request_id") or "stream-chunk",
+            "object": orc.get("object") or "chat.completion.chunk",
+            "created": orc.get("created") or evt.get("created") or _now_ts(),
+            "model": orc.get("model") or "unknown",
+            "choices": [
+                {
+                    "index": c.get("index", 0),
+                    "delta": c.get("delta") or {},
+                    "finish_reason": c.get("finish_reason"),
+                }
+                for c in (orc.get("choices") or [])
+            ],
+        }
+        return _StreamParser._validate_chunk(payload)
 
     @staticmethod
-    def to_openai_chunk(event_obj: dict) -> OpenAIChatCompletionChunk | None:
+    def to_openai_chunk(event_obj: dict[str, object]) -> OpenAIChatCompletionChunk | None:
         """
         Accepts:
           - {"final_result": <openai-style CHUNK>}   (IMPORTANT: this is just another chunk, NOT terminal)
@@ -102,7 +103,10 @@ class _StreamParser:
 
         # FINAL RESULT IS *NOT* TERMINAL: treat it as the next chunk
         if "final_result" in event_obj:
-            fr: Final = event_obj["final_result"] or {}
+            final_result: Final = event_obj["final_result"]
+            if not isinstance(final_result, dict):
+                return None
+            fr: Final[dict[str, object]] = final_result
             # ensure it looks like an OpenAI chunk
             if "object" not in fr:
                 fr["object"] = "chat.completion.chunk"
diff --git a/tests/test_litellm/llms/sap/chat/test_sap_chat_calls.py b/tests/test_litellm/llms/sap/chat/test_sap_chat_calls.py
index 170354dae7e..90b4b39582b 100644
--- a/tests/test_litellm/llms/sap/chat/test_sap_chat_calls.py
+++ b/tests/test_litellm/llms/sap/chat/test_sap_chat_calls.py
@@ -1,3 +1,4 @@
+import json
 import httpx
 from unittest.mock import patch, PropertyMock
 
@@ -267,7 +268,7 @@ def test_validate_chunk_without_usage_keeps_none():
     assert chunk.usage is None
 
 
-def test_validated_usage_survives_nested_model_dump():
+def test_validated_usage_survives_nested_model_dump_json():
     from litellm.llms.sap.chat.handler import _StreamParser
     from litellm.types.utils import ModelResponseStream
 
@@ -276,7 +277,7 @@ def test_validated_usage_survives_nested_model_dump():
     model_response = ModelResponseStream()
     setattr(model_response, "usage", chunk.usage)
 
-    dumped = model_response.model_dump()
+    dumped = json.loads(model_response.model_dump_json())
     assert dumped["usage"]["total_tokens"] == 62528