fix(serialization): strip base64 image data from streamed values events (#3631)

The SSE stream's `values` snapshots serialize the full state, including the
hide_from_ui human messages that ViewImageMiddleware fills with base64 image
payloads. #3535 stripped those from the REST wait/history/state endpoints via
serialize_channel_values_for_api, but the streaming path (worker publishes
serialize(chunk, mode="values")) still went through serialize_channel_values,
so the same base64 leaked to the frontend over SSE. Route values-mode
serialization through serialize_channel_values_for_api as well. Non-hidden
messages and https image URLs are left untouched.

backend/packages/harness/deerflow/runtime/serialization.py

@@ -133,11 +133,14 @@ def serialize(obj: Any, *, mode: str = "") -> Any:
     """Serialize LangChain objects with mode-specific handling.
 
     * ``messages`` — obj is ``(message_chunk, metadata_dict)``
-    * ``values`` — obj is the full state dict; ``__pregel_*`` keys stripped
+    * ``values`` — obj is the full state dict; ``__pregel_*`` keys stripped and
+      base64 ``data:`` image blocks dropped from hide_from_ui messages
     * everything else — recursive ``model_dump()`` / ``dict()`` fallback
     """
     if mode == "messages":
         return serialize_messages_tuple(obj)
     if mode == "values":
-        return serialize_channel_values(obj) if isinstance(obj, dict) else serialize_lc_object(obj)
+        # ``values`` snapshots stream the full state to the frontend, so they
+        # must drop base64 image payloads the same way the REST endpoints do.
+        return serialize_channel_values_for_api(obj) if isinstance(obj, dict) else serialize_lc_object(obj)
     return serialize_lc_object(obj)

backend/tests/test_serialization.py

@@ -328,3 +328,32 @@ def test_serialize_channel_values_for_api_no_messages():
 
     result = serialize_channel_values_for_api({"title": "empty"})
     assert result == {"title": "empty"}
+
+
+def test_serialize_values_mode_strips_base64_from_hidden_messages():
+    """The SSE stream emits ``values`` snapshots of the full state, so it must
+    strip base64 image data from hide_from_ui messages just like the REST
+    endpoints do — otherwise the same payload leaks over the stream."""
+    import json
+
+    from deerflow.runtime.serialization import serialize
+
+    state = {
+        "messages": [
+            _make_msg(
+                [
+                    {"type": "text", "text": "context"},
+                    {
+                        "type": "image_url",
+                        "image_url": {"url": "data:image/png;base64,iVBOR..."},
+                    },
+                ],
+                hide_from_ui=True,
+            ),
+        ],
+    }
+    result = serialize(state, mode="values")
+    # the hidden message survives (count/order preserved) but the data: block is gone
+    assert len(result["messages"]) == 1
+    assert "data:image/png;base64" not in json.dumps(result)
+    assert result["messages"][0]["content"] == [{"type": "text", "text": "context"}]