From 4d952e7b8119fc5bdf75dbc52682f309496f4800 Mon Sep 17 00:00:00 2001
From: Rod Boev <rod.boev@gmail.com>
Date: Thu, 25 Jun 2026 07:34:19 -0400
Subject: [PATCH] fix(vllm_model): handle explicit null metadata in
 preprocessing

Signed-off-by: Rod Boev <rod.boev@gmail.com>
---
 responses_api_models/vllm_model/app.py        |  3 +++
 .../vllm_model/tests/test_app.py              | 24 +++++++++++++++++++
 2 files changed, 27 insertions(+)

diff --git a/responses_api_models/vllm_model/app.py b/responses_api_models/vllm_model/app.py
index 3e96dae44d..e6e86e1634 100644
--- a/responses_api_models/vllm_model/app.py
+++ b/responses_api_models/vllm_model/app.py
@@ -301,6 +301,9 @@ def _preprocess_chat_completion_create_params(self, request: Request, body_dict:
             chat_template_kwargs = deepcopy(self.config.chat_template_kwargs)
 
         metadata = body_dict.get("metadata", dict())
+        if metadata is None:
+            metadata = {}
+            body_dict["metadata"] = metadata
 
         # Merge global config chat_template_kwargs with per-request overrides in metadata (e.g. per-sample reasoning on/off)
         metadata_chat_template_kwargs_str = metadata.get("chat_template_kwargs", "{}")
diff --git a/responses_api_models/vllm_model/tests/test_app.py b/responses_api_models/vllm_model/tests/test_app.py
index b32ecd8c0e..6cc8b76325 100644
--- a/responses_api_models/vllm_model/tests/test_app.py
+++ b/responses_api_models/vllm_model/tests/test_app.py
@@ -3306,6 +3306,30 @@ async def mock_create_chat_completion(**kwargs):
         assert captured_kwargs["min_tokens"] == 20
         assert captured_kwargs["new_param"] == "value"
 
+    def test_metadata_none_normalizes_to_empty_dict(self) -> None:
+        config = VLLMModelConfig(
+            host="0.0.0.0",
+            port=8081,
+            base_url="http://api.openai.com/v1",
+            api_key="dummy_key",  # pragma: allowlist secret
+            model="dummy-model",
+            entrypoint="",
+            name="",
+            return_token_id_information=False,
+            uses_reasoning_parser=False,
+        )
+        model = VLLMModel(config=config, server_client=MagicMock(spec=ServerClient))
+
+        body = {
+            "model": "dummy-model",
+            "messages": [{"role": "user", "content": "hello"}],
+            "metadata": None,
+        }
+        result = model._preprocess_chat_completion_create_params(MagicMock(), body)
+
+        assert result["messages"][0]["content"] == "hello"
+        assert result["metadata"] == {}
+
 
 # ──────────────────────────────────────────────────────────────────────────────
 # Audio sidechannel splice (metadata.audio_data → user-message content block)