fix(workflow): preserve non-ASCII characters in LLM agent node input

src/google/adk/flows/llm_flows/contents.py

@@ -360,7 +360,7 @@ def _build_task_input_user_content(
         try:
           import json as _json
 
-          text = _json.dumps(dict(fc.args))
+          text = _json.dumps(dict(fc.args), ensure_ascii=False)
         except (TypeError, ValueError):
           text = str(fc.args)
         parts = [types.Part(text=text)]

src/google/adk/workflow/_llm_agent_wrapper.py

@@ -194,7 +194,7 @@ def _node_input_to_content(node_input: Any) -> types.Content:
   elif isinstance(node_input, BaseModel):
     text = node_input.model_dump_json()
   elif isinstance(node_input, (dict, list)):
-    text = json.dumps(node_input)
+    text = json.dumps(node_input, ensure_ascii=False)
   else:
     text = str(node_input)
   return types.Content(role='user', parts=[types.Part(text=text)])

tests/unittests/flows/llm_flows/test_contents.py

@@ -1598,3 +1598,41 @@ def test_rearrange_async_function_responses_early_returns_when_no_responses():
       events
   )
   assert result is events
+
+
+def test_task_input_user_content_preserves_non_ascii():
+  """Delegated task input must not escape non-ASCII FC args (issue #6279).
+
+  A chat coordinator delegates to a task sub-agent via a function call; the
+  task agent's first user turn is rebuilt from the FC args. Escaping non-Latin
+  characters to ``\\uXXXX`` there bloats prompt tokens and degrades responses.
+  """
+  fc_id = "fc_task_1"
+  events = [
+      Event(
+          invocation_id="inv1",
+          author="coordinator",
+          content=types.Content(
+              role="model",
+              parts=[
+                  types.Part(
+                      function_call=types.FunctionCall(
+                          id=fc_id,
+                          name="delegate",
+                          args={"query": "שלום עולם", "city": "北京"},
+                      )
+                  )
+              ],
+          ),
+      ),
+  ]
+
+  content = contents._build_task_input_user_content(  # pylint: disable=protected-access
+      events, isolation_scope=fc_id
+  )
+
+  assert content is not None and content.parts
+  text = content.parts[0].text
+  assert "שלום עולם" in text
+  assert "北京" in text
+  assert "\\u" not in text

tests/unittests/workflow/test_llm_agent_as_node.py

@@ -234,6 +234,45 @@ async def test_single_turn_input_event_inherits_branch_and_scope(
   assert event.isolation_scope == 'scope-1'
 
 
+# --- _node_input_to_content non-ASCII handling ---
+
+
+class TestNodeInputNonAscii:
+  """Node input must reach the LLM without escaping non-ASCII characters.
+
+  Escaping (e.g. json.dumps' default ensure_ascii=True) turns each non-Latin
+  character into a ``\\uXXXX`` sequence, which bloats prompt tokens and
+  degrades model responses for non-English inputs (issue #6279).
+  """
+
+  def _text_of(self, node_input: Any) -> str:
+    from google.adk.workflow._llm_agent_wrapper import _node_input_to_content
+
+    content = _node_input_to_content(node_input)
+    assert content.parts and content.parts[0].text is not None
+    return content.parts[0].text
+
+  def test_dict_input_preserves_non_ascii(self):
+    text = self._text_of({'query': 'שלום עולם', 'city': '北京'})
+    assert 'שלום עולם' in text
+    assert '北京' in text
+    assert '\\u' not in text
+
+  def test_list_input_preserves_non_ascii(self):
+    text = self._text_of(['שלום', '你好'])
+    assert 'שלום' in text
+    assert '你好' in text
+    assert '\\u' not in text
+
+  def test_base_model_input_preserves_non_ascii(self):
+    class _Payload(BaseModel):
+      topic: str
+
+    text = self._text_of(_Payload(topic='עברית 中文'))
+    assert 'עברית 中文' in text
+    assert '\\u' not in text
+
+
 # --- build_node auto-wrapping ---