fix: preserve non-ASCII characters in agent input

Merge #6282

`input_schema` inputs containing non-Latin characters (Hebrew, Chinese, etc.) reach the LLM as `\uXXXX` escapes, which bloats prompt tokens (~6x for Hebrew) and degrades model responses, as reported in #6279.

The escaping comes from `json.dumps` being called with its default `ensure_ascii=True` on the LLM-bound input text. Two paths were affected:

- `utils/content_utils.py` - `to_user_content()` for `dict`/`list` node input (used by `workflow/_llm_agent_wrapper.py`)
- `flows/llm_flows/contents.py` – `_build_task_input_user_content()`, which rebuilds a delegated task's function-call args as the sub-agent's first user turn. This path takes priority over the wrapper fallback, so it affects the common chat/root → task sub-agent delegation case.

Both now serialize with `ensure_ascii=False`, matching how the output-schema path already serializes responses (`_output_schema_processor.py`), fixed earlier in #2936/#2937.

Fixes #6279
Closes #6282

Co-authored-by: Jason Zhang <jasoncz@google.com>
PiperOrigin-RevId: 947866027

Author: 2 people

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

@@ -387,7 +387,7 @@ 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/utils/content_utils.py

@@ -64,7 +64,7 @@ def to_user_content(value: Any) -> types.Content:
     deep-copied)
   - str -> single text part
   - BaseModel -> model_dump_json() text part
-  - dict/list -> json.dumps() text part
+  - dict/list -> json.dumps() text part (non-ASCII preserved, not escaped)
   - anything else -> str() text part
   """
   if isinstance(value, types.Content):
@@ -74,7 +74,7 @@ def to_user_content(value: Any) -> types.Content:
   elif isinstance(value, BaseModel):
     text = value.model_dump_json()
   elif isinstance(value, (dict, list)):
-    text = json.dumps(value)
+    text = json.dumps(value, ensure_ascii=False)
   else:
     text = str(value)
   return types.Content(role='user', parts=[types.Part(text=text)])

tests/unittests/flows/llm_flows/test_contents.py

@@ -1965,3 +1965,41 @@ def test_recover_compacted_parallel_call_reinjects_sibling_response():
       if part.function_response
   }
   assert response_ids == {"lr-1", "reg-1"}
+
+
+def test_task_input_user_content_preserves_non_ascii():
+  """Delegated task input must not escape non-ASCII FC args.
+
+  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/utils/test_content_utils.py

@@ -66,3 +66,25 @@ def test_to_user_content_other_input_is_str():
   content = to_user_content(42)
   assert content.role == 'user'
   assert content.parts[0].text == '42'
+
+
+def test_to_user_content_dict_input_preserves_non_ascii():
+  """Non-ASCII input must reach the LLM as-is, not as \\uXXXX escapes.
+
+  Escaping (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.
+  """
+  content = to_user_content({'query': 'שלום עולם', 'city': '北京'})
+  text = content.parts[0].text
+  assert 'שלום עולם' in text
+  assert '北京' in text
+  assert '\\u' not in text
+
+
+def test_to_user_content_list_input_preserves_non_ascii():
+  content = to_user_content(['שלום', '你好'])
+  text = content.parts[0].text
+  assert 'שלום' in text
+  assert '你好' in text
+  assert '\\u' not in text