hotfix(chat): Chat answer streaming and smooth markdown rendering

surfsense_backend/app/tasks/chat/streaming/flows/shared/llm_bundle.py

@@ -130,7 +130,9 @@ async def load_llm_bundle(
             billing_tier="free",
         )
         return (
-            SanitizedChatLiteLLM(model=model_string, **litellm_kwargs),
+            SanitizedChatLiteLLM(
+                model=model_string, **{**litellm_kwargs, "streaming": True}
+            ),
             agent_config,
             None,
         )
@@ -174,7 +176,9 @@ async def load_llm_bundle(
         billing_tier=str(global_model.get("billing_tier", "free")).lower(),
     )
     return (
-        SanitizedChatLiteLLM(model=model_string, **litellm_kwargs),
+        SanitizedChatLiteLLM(
+            model=model_string, **{**litellm_kwargs, "streaming": True}
+        ),
         agent_config,
         None,
     )

surfsense_backend/tests/unit/tasks/chat/streaming/test_llm_bundle.py

@@ -0,0 +1,154 @@
+"""Contracts for chat LLM construction in streaming flows.
+
+``stream_new_chat`` / ``stream_resume_chat`` depend on LangChain receiving
+token chunks from ``ChatLiteLLM``. ``langchain-litellm`` defaults
+``streaming`` to ``False``, so the shared bundle loader must opt in
+explicitly for both DB-backed and global model paths.
+"""
+
+from __future__ import annotations
+
+from types import SimpleNamespace
+from typing import Any
+
+import pytest
+
+import app.tasks.chat.streaming.flows.shared.llm_bundle as llm_bundle
+
+pytestmark = pytest.mark.unit
+
+
+class _CapturedChatLiteLLM:
+    calls: list[dict[str, Any]] = []
+
+    def __init__(self, **kwargs: Any) -> None:
+        self.kwargs = kwargs
+        self.__class__.calls.append(kwargs)
+
+
+@pytest.fixture(autouse=True)
+def _patch_common_bundle_dependencies(monkeypatch: pytest.MonkeyPatch):
+    """Keep these tests focused on the LLM constructor contract."""
+
+    _CapturedChatLiteLLM.calls = []
+
+    async def _fake_search_space(_session: Any, _search_space_id: int) -> SimpleNamespace:
+        return SimpleNamespace(id=42, user_id="user-1")
+
+    monkeypatch.setattr(llm_bundle, "_load_search_space", _fake_search_space)
+    monkeypatch.setattr(llm_bundle, "SanitizedChatLiteLLM", _CapturedChatLiteLLM)
+    monkeypatch.setattr(llm_bundle, "register_model_usage_metadata", lambda **_kw: None)
+    monkeypatch.setattr(
+        llm_bundle,
+        "has_capability",
+        lambda _model, capability: capability in {"chat", "vision"},
+    )
+
+    return None
+
+
+async def test_load_llm_bundle_enables_streaming_for_db_models(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    connection = SimpleNamespace(
+        provider="openai",
+        api_key="sk-test",
+        base_url=None,
+        extra={"litellm_params": {"temperature": 0.1}},
+    )
+    model = SimpleNamespace(
+        id=7,
+        model_id="gpt-4o-mini",
+        display_name="GPT 4o Mini",
+        connection=connection,
+    )
+
+    async def _fake_db_model(_session: Any, *, model_id: int, search_space: Any) -> Any:
+        assert model_id == 7
+        assert search_space.id == 42
+        return model
+
+    monkeypatch.setattr(llm_bundle, "_load_db_model", _fake_db_model)
+    monkeypatch.setattr(
+        llm_bundle,
+        "to_litellm",
+        lambda _conn, _model_id: (
+            "openai/gpt-4o-mini",
+            {"api_key": "sk-test", "temperature": 0.1},
+        ),
+    )
+
+    llm, agent_config, error = await llm_bundle.load_llm_bundle(
+        object(),
+        config_id=7,
+        search_space_id=42,
+    )
+
+    assert error is None
+    assert llm is not None
+    assert agent_config is not None
+    assert _CapturedChatLiteLLM.calls == [
+        {
+            "model": "openai/gpt-4o-mini",
+            "api_key": "sk-test",
+            "temperature": 0.1,
+            "streaming": True,
+        }
+    ]
+
+
+async def test_load_llm_bundle_enables_streaming_for_global_models(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    global_model = {
+        "id": -11,
+        "connection_id": -101,
+        "model_id": "claude-sonnet-4-5",
+        "display_name": "Claude Sonnet",
+        "billing_tier": "premium",
+    }
+    global_connection = {
+        "id": -101,
+        "provider": "anthropic",
+        "api_key": "sk-ant-test",
+        "base_url": None,
+        "extra": {"litellm_params": {"temperature": 0.2}},
+    }
+    monkeypatch.setattr(
+        llm_bundle.config,
+        "GLOBAL_MODELS",
+        [global_model],
+        raising=False,
+    )
+    monkeypatch.setattr(
+        llm_bundle.config,
+        "GLOBAL_CONNECTIONS",
+        [global_connection],
+        raising=False,
+    )
+    monkeypatch.setattr(
+        llm_bundle,
+        "to_litellm",
+        lambda _conn, _model_id: (
+            "anthropic/claude-sonnet-4-5",
+            {"api_key": "sk-ant-test", "temperature": 0.2},
+        ),
+    )
+
+    llm, agent_config, error = await llm_bundle.load_llm_bundle(
+        object(),
+        config_id=-11,
+        search_space_id=42,
+    )
+
+    assert error is None
+    assert llm is not None
+    assert agent_config is not None
+    assert _CapturedChatLiteLLM.calls == [
+        {
+            "model": "anthropic/claude-sonnet-4-5",
+            "api_key": "sk-ant-test",
+            "temperature": 0.2,
+            "streaming": True,
+        }
+    ]

surfsense_web/components/assistant-ui/chat-viewport.tsx

@@ -27,8 +27,8 @@ export interface ChatViewportProps {
 export const ChatViewport: FC<ChatViewportProps> = ({ children, footer }) => (
 	<ThreadPrimitive.Viewport
 		turnAnchor="top"
-		autoScroll={false}
-		scrollToBottomOnRunStart={false}
+		autoScroll
+		scrollToBottomOnRunStart
 		scrollToBottomOnInitialize
 		scrollToBottomOnThreadSwitch
 		className="aui-thread-viewport relative flex flex-1 min-h-0 flex-col overflow-y-auto px-4 scroll-smooth"

surfsense_web/components/assistant-ui/markdown-text.tsx

@@ -110,7 +110,7 @@ const MarkdownTextImpl = () => {
 	return (
 		<CitationUrlMapContext.Provider value={urlMapRef}>
 			<MarkdownTextPrimitive
-				smooth={false}
+				smooth
 				remarkPlugins={[remarkGfm, [remarkMath, { singleDollarTextMath: false }]]}
 				rehypePlugins={[rehypeKatex]}
 				className="aui-md"