fix

tests/entrypoints/openai/test_render_no_inference.py

@@ -13,7 +13,7 @@
 
 from ...utils import RemoteOpenAIServer
 
-MODEL_NAME = "hmellor/tiny-random-LlamaForCausalLM"
+MODEL_NAME = "openai/gpt-oss-20b"
 
 
 @pytest.fixture(scope="module")
@@ -60,9 +60,25 @@ async def test_chat_completion_render_works_without_inference(client):
 
     conversation, engine_prompts = data
 
-    # Verify conversation is preserved
-    assert conversation[0]["role"] == "user"
-    assert "Hello" in conversation[0]["content"]
+    # Verify conversation contains messages
+    assert len(conversation) > 0
+
+    # Find the user message (Harmony models use 'author' dict instead of 'role')
+    user_msg = None
+    for msg in conversation:
+        # Check for direct 'role' key (standard format)
+        if msg.get("role") == "user":
+            user_msg = msg
+            break
+        # Check for nested 'author' dict (Harmony format)
+        author = msg.get("author")
+        if isinstance(author, dict) and author.get("role") == "user":
+            user_msg = msg
+            break
+
+    assert user_msg is not None, (
+        f"User message not found in conversation: {conversation}"
+    )
 
     # Verify tokenization occurred
     assert len(engine_prompts) > 0

vllm/entrypoints/openai/chat_completion/api_router.py

@@ -101,7 +101,12 @@ async def render_chat_completion(request: ChatCompletionRequest, raw_request: Re
     if isinstance(result, ErrorResponse):
         return JSONResponse(content=result.model_dump(), status_code=result.error.code)
 
-    return JSONResponse(content=result)
+    conversation, engine_prompts = result
+    # Serialize Pydantic models (Harmony) or pass through TypedDicts as-is
+    serialized_conversation = [
+        getattr(msg, "model_dump", lambda m=msg: m)() for msg in conversation
+    ]
+    return JSONResponse(content=[serialized_conversation, engine_prompts])
 
 
 def attach_router(app: FastAPI):