fix(models): pass NOT_GIVEN to Anthropic when no system_instruction

When LlmRequest has no system_instruction (e.g. during event compaction
via LlmEventSummarizer), AnthropicLlm passed system=None to the
Anthropic SDK, which serializes to JSON null and the API rejects with
'system: Input should be a valid list'. Use NOT_GIVEN to omit the
parameter entirely instead.

Close google#5318

Co-authored-by: George Weale <gweale@google.com>
PiperOrigin-RevId: 932664392

Author: GWeale

src/google/adk/models/anthropic_llm.py

@@ -553,11 +553,14 @@ async def generate_content_async(
         else NOT_GIVEN
     )
     thinking = _build_anthropic_thinking_param(llm_request.config)
+    system = NOT_GIVEN
+    if llm_request.config.system_instruction is not None:
+      system = llm_request.config.system_instruction
 
     if not stream:
       message = await self._anthropic_client.messages.create(
           model=model_to_use,
-          system=llm_request.config.system_instruction,
+          system=system,
           messages=messages,
           tools=tools,
           tool_choice=tool_choice,
@@ -567,14 +570,15 @@ async def generate_content_async(
       yield message_to_generate_content_response(message)
     else:
       async for response in self._generate_content_streaming(
-          llm_request, messages, tools, tool_choice, thinking
+          llm_request, messages, system, tools, tool_choice, thinking
       ):
         yield response
 
   async def _generate_content_streaming(
       self,
       llm_request: LlmRequest,
       messages: list[anthropic_types.MessageParam],
+      system: Union[str, types.Content, NotGiven],
       tools: Union[Iterable[anthropic_types.ToolUnionParam], NotGiven],
       tool_choice: Union[anthropic_types.ToolChoiceParam, NotGiven],
       thinking: Union[
@@ -591,7 +595,7 @@ async def _generate_content_streaming(
     model_to_use = self._resolve_model_name(llm_request.model)
     raw_stream = await self._anthropic_client.messages.create(
         model=model_to_use,
-        system=llm_request.config.system_instruction,
+        system=system,
         messages=messages,
         tools=tools,
         tool_choice=tool_choice,

tests/unittests/models/test_anthropic_llm.py

@@ -21,6 +21,7 @@
 from unittest.mock import AsyncMock
 from unittest.mock import MagicMock
 
+from anthropic import NOT_GIVEN
 from anthropic import types as anthropic_types
 from google.adk import version as adk_version
 from google.adk.models import anthropic_llm
@@ -2134,3 +2135,93 @@ async def test_generate_content_async_pairs_invalid_tool_ids(
   ]
   assert len(set(use_ids)) == expected_unique
   assert set(use_ids) == set(result_ids)
+
+
+@pytest.mark.asyncio
+async def test_non_streaming_no_system_instruction_passes_not_given():
+  """system=NOT_GIVEN when LlmRequest has no system_instruction."""
+  llm = AnthropicLlm(model="claude-sonnet-4-20250514")
+
+  mock_message = anthropic_types.Message(
+      id="msg_test",
+      content=[
+          anthropic_types.TextBlock(text="ok", type="text", citations=None)
+      ],
+      model="claude-sonnet-4-20250514",
+      role="assistant",
+      stop_reason="end_turn",
+      stop_sequence=None,
+      type="message",
+      usage=anthropic_types.Usage(
+          input_tokens=1,
+          output_tokens=1,
+          cache_creation_input_tokens=0,
+          cache_read_input_tokens=0,
+          server_tool_use=None,
+          service_tier=None,
+      ),
+  )
+
+  mock_client = MagicMock()
+  mock_client.messages.create = AsyncMock(return_value=mock_message)
+
+  request = LlmRequest(
+      model="claude-sonnet-4-20250514",
+      contents=[Content(role="user", parts=[Part.from_text(text="Hi")])],
+  )
+  assert request.config.system_instruction is None
+
+  with mock.patch.object(llm, "_anthropic_client", mock_client):
+    _ = [r async for r in llm.generate_content_async(request, stream=False)]
+
+  mock_client.messages.create.assert_called_once()
+  _, kwargs = mock_client.messages.create.call_args
+  assert kwargs["system"] is NOT_GIVEN
+
+
+@pytest.mark.asyncio
+async def test_streaming_no_system_instruction_passes_not_given():
+  """system=NOT_GIVEN on the streaming path when no system_instruction."""
+  llm = AnthropicLlm(model="claude-sonnet-4-20250514")
+
+  events = [
+      MagicMock(
+          type="message_start",
+          message=MagicMock(usage=MagicMock(input_tokens=1, output_tokens=0)),
+      ),
+      MagicMock(
+          type="content_block_start",
+          index=0,
+          content_block=anthropic_types.TextBlock(text="", type="text"),
+      ),
+      MagicMock(
+          type="content_block_delta",
+          index=0,
+          delta=anthropic_types.TextDelta(text="ok", type="text_delta"),
+      ),
+      MagicMock(type="content_block_stop", index=0),
+      MagicMock(
+          type="message_delta",
+          delta=MagicMock(stop_reason="end_turn"),
+          usage=MagicMock(output_tokens=1),
+      ),
+      MagicMock(type="message_stop"),
+  ]
+
+  mock_client = MagicMock()
+  mock_client.messages.create = AsyncMock(
+      return_value=_make_mock_stream_events(events)
+  )
+
+  request = LlmRequest(
+      model="claude-sonnet-4-20250514",
+      contents=[Content(role="user", parts=[Part.from_text(text="Hi")])],
+  )
+  assert request.config.system_instruction is None
+
+  with mock.patch.object(llm, "_anthropic_client", mock_client):
+    _ = [r async for r in llm.generate_content_async(request, stream=True)]
+
+  mock_client.messages.create.assert_called_once()
+  _, kwargs = mock_client.messages.create.call_args
+  assert kwargs["system"] is NOT_GIVEN