From 7793917aaedd64cbb91c122b102c380f82497007 Mon Sep 17 00:00:00 2001
From: TemoSulava <sulava.temo@yahoo.com>
Date: Thu, 22 Oct 2026 21:43:49 +0400
Subject: [PATCH] fix(proxy): set model on non-streaming RejectedRequestError
 responses

When a pre-call hook returns a string, the chat and text completion
routes answer with it through RejectedRequestError. The non-streaming
response was a bare ModelResponse/TextCompletionResponse with
model=None, so OpenAI SDK clients failed to parse the chat response.
Take the model from the request data, falling back to the exception's,
the same way the streaming branch does.
---
 litellm/proxy/proxy_server.py                 |  2 +
 .../test_rejected_request_response_model.py   | 81 +++++++++++++++++++
 2 files changed, 83 insertions(+)
 create mode 100644 tests/unit/proxy/test_rejected_request_response_model.py

diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py
index df948b69a67..bd96548a078 100644
--- a/litellm/proxy/proxy_server.py
+++ b/litellm/proxy/proxy_server.py
@@ -12824,6 +12824,7 @@ async def chat_completion(
             request_data=_data,
         )
         _chat_response = litellm.ModelResponse()
+        _chat_response.model = _data.get("model") or e.model or ""
         _chat_response.choices[0].message.content = e.message
 
         if data.get("stream", None) is not None and data["stream"] is True:
@@ -13009,6 +13010,7 @@ async def completion(
         else:
             _response = litellm.TextCompletionResponse()
             _response.choices[0].text = e.message
+            _response.model = _data.get("model") or e.model or ""
             return _response
     except Exception as e:
         await proxy_logging_obj.post_call_failure_hook(
diff --git a/tests/unit/proxy/test_rejected_request_response_model.py b/tests/unit/proxy/test_rejected_request_response_model.py
new file mode 100644
index 00000000000..398fef243d7
--- /dev/null
+++ b/tests/unit/proxy/test_rejected_request_response_model.py
@@ -0,0 +1,81 @@
+"""Regression test for the non-streaming RejectedRequestError response.
+
+A pre-call hook that returns a string raises ``RejectedRequestError``, and the
+route answers with that string instead of calling the LLM. The non-streaming
+response used to be a bare ``ModelResponse()`` with ``model=None``, which the
+OpenAI SDK rejects (``model: Input should be a valid string``).
+"""
+
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import pytest
+from fastapi import Request, Response
+from openai.types.chat import ChatCompletion
+
+from litellm.exceptions import RejectedRequestError
+from litellm.proxy._types import UserAPIKeyAuth
+from litellm.proxy.proxy_server import chat_completion, completion
+
+
+async def _run_rejected_request(route, body, exception_model="my-model"):
+    exception = RejectedRequestError(
+        message="cached answer",
+        model=exception_model,
+        llm_provider="",
+        request_data=dict(body),
+    )
+    with (
+        patch(
+            "litellm.proxy.proxy_server.read_request_body",
+            new_callable=AsyncMock,
+            return_value=dict(body),
+        ),
+        patch(
+            "litellm.proxy.proxy_server.ProxyBaseLLMRequestProcessing.base_process_llm_request",
+            new_callable=AsyncMock,
+            side_effect=exception,
+        ),
+        patch("litellm.proxy.proxy_server.proxy_logging_obj") as mock_proxy_logging,
+    ):
+        mock_proxy_logging.post_call_failure_hook = AsyncMock()
+
+        return await route(
+            request=MagicMock(spec=Request),
+            fastapi_response=MagicMock(spec=Response),
+            user_api_key_dict=UserAPIKeyAuth(),
+        )
+
+
+@pytest.mark.asyncio
+async def test_rejected_chat_completion_response_has_model():
+    response = await _run_rejected_request(
+        chat_completion,
+        {"model": "my-model", "messages": [{"role": "user", "content": "hi"}]},
+    )
+
+    assert response.model == "my-model"
+    ChatCompletion.model_validate(response.model_dump())
+
+
+@pytest.mark.asyncio
+async def test_rejected_text_completion_response_has_model():
+    response = await _run_rejected_request(completion, {"model": "my-model", "prompt": "hi"})
+
+    # finish_reason is still None here, so this is not validated against
+    # openai.types.Completion the way the chat response is.
+    assert response.model == "my-model"
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "route, body",
+    [
+        (chat_completion, {"model": "my-model", "messages": [{"role": "user", "content": "hi"}]}),
+        (completion, {"model": "my-model", "prompt": "hi"}),
+    ],
+)
+async def test_rejected_response_falls_back_to_request_model(route, body):
+    # RejectedRequestError is public, so a custom hook can raise it without a model.
+    response = await _run_rejected_request(route, body, exception_model=None)
+
+    assert response.model == "my-model"