fix(streaming): handle None chunks in _chunk_to_dict

app/routes/chat.py

@@ -54,7 +54,16 @@
 
 
 def _chunk_to_dict(chunk) -> dict:
-    """Normalize a litellm chunk (Pydantic model or dict) into a plain dict."""
+    """Normalize a litellm chunk (Pydantic model or dict) into a plain dict.
+
+    Returns an empty dict for None chunks — some LiteLLM stream wrappers
+    yield None as a heartbeat/keepalive signal, and crashing on those
+    would kill the stream for a non-event. Other unexpected types fall
+    through to dict() which will raise TypeError if the object is not
+    iterable, surfacing the bug rather than silently swallowing it.
+    """
+    if chunk is None:
+        return {}
     if isinstance(chunk, dict):
         return chunk
     if hasattr(chunk, "model_dump"):

tests/unit/test_chunk_to_dict_guard.py

@@ -0,0 +1,32 @@
+"""Regression: _chunk_to_dict must handle None chunks gracefully.
+
+Some LiteLLM stream wrappers yield None as a heartbeat/keepalive signal.
+Without a guard, dict(None) raises TypeError and kills the stream for a
+non-event. The function must return an empty dict for None so the stream
+continues uninterrupted."""
+
+from app.routes.chat import _chunk_to_dict
+
+
+def test_chunk_to_dict_none_returns_empty():
+    assert _chunk_to_dict(None) == {}
+
+
+def test_chunk_to_dict_dict_passthrough():
+    d = {"id": "1", "model": "gpt-4o"}
+    assert _chunk_to_dict(d) is d
+
+
+def test_chunk_to_dict_pydantic_model():
+    class FakeModel:
+        def model_dump(self, exclude_none=False):
+            return {"id": "1", "model": "gpt-4o"}
+
+    result = _chunk_to_dict(FakeModel())
+    assert result == {"id": "1", "model": "gpt-4o"}
+
+
+def test_chunk_to_dict_iterable_fallback():
+    # A tuple of pairs is iterable and dict()-able
+    result = _chunk_to_dict((("key", "value"),))
+    assert result == {"key": "value"}