fix(adk): address review feedback on guardrail fallback paths

- Preserve usage_metadata when choices is empty
- Keep MAX_TOKENS finish reason when an empty stream was truncated by
  length rather than blocked by a content filter
- Centralize the blocked-content placeholder in a module constant

Signed-off-by: Combrink van der Vyver <combrink@gmail.com>

Author: combrink

python/packages/kagent-adk/src/kagent/adk/models/_openai.py

@@ -38,6 +38,9 @@
 if TYPE_CHECKING:
     from google.adk.models.llm_request import LlmRequest
 
+# Emitted when a guardrail or content filter blocks a response, leaving no content to surface.
+_CONTENT_BLOCKED_PLACEHOLDER = "Response blocked by content policy."
+
 
 def _convert_role_to_openai(role: Optional[str]) -> str:
     """Convert google.genai role to OpenAI role."""
@@ -316,12 +319,22 @@ def _convert_tools_to_openai(tools: list[types.Tool]) -> list[ChatCompletionTool
 
 def _convert_openai_response_to_llm_response(response: ChatCompletion) -> LlmResponse:
     """Convert OpenAI response to LlmResponse."""
+    # Handle usage metadata
+    usage_metadata = None
+    if hasattr(response, "usage") and response.usage:
+        usage_metadata = types.GenerateContentResponseUsageMetadata(
+            prompt_token_count=response.usage.prompt_tokens,
+            candidates_token_count=response.usage.completion_tokens,
+            total_token_count=response.usage.total_tokens,
+        )
+
     if not response.choices:
         return LlmResponse(
             content=types.Content(
                 role="model",
-                parts=[types.Part.from_text(text="Response blocked by content policy.")],
+                parts=[types.Part.from_text(text=_CONTENT_BLOCKED_PLACEHOLDER)],
             ),
+            usage_metadata=usage_metadata,
             finish_reason=types.FinishReason.SAFETY,
         )
     choice = response.choices[0]
@@ -354,15 +367,6 @@ def _convert_openai_response_to_llm_response(response: ChatCompletion) -> LlmRes
 
     content = types.Content(role="model", parts=parts)
 
-    # Handle usage metadata
-    usage_metadata = None
-    if hasattr(response, "usage") and response.usage:
-        usage_metadata = types.GenerateContentResponseUsageMetadata(
-            prompt_token_count=response.usage.prompt_tokens,
-            candidates_token_count=response.usage.completion_tokens,
-            total_token_count=response.usage.total_tokens,
-        )
-
     # Handle finish reason
     finish_reason = types.FinishReason.STOP
     if choice.finish_reason == "length":
@@ -593,8 +597,12 @@ async def generate_content_async(
                 # Guardrail or content filter can produce zero content/tool chunks.
                 # An empty parts list causes downstream IndexError; emit a placeholder.
                 if not final_parts:
-                    final_parts.append(types.Part.from_text(text="Response blocked by content policy."))
-                    final_reason = types.FinishReason.SAFETY
+                    if final_reason == types.FinishReason.MAX_TOKENS:
+                        # Truncated by length before any content; not a safety block.
+                        final_parts.append(types.Part.from_text(text=""))
+                    else:
+                        final_parts.append(types.Part.from_text(text=_CONTENT_BLOCKED_PLACEHOLDER))
+                        final_reason = types.FinishReason.SAFETY
 
                 # Always yield final response to signal completion and valid metadata
                 final_content = types.Content(role="model", parts=final_parts)

python/packages/kagent-adk/tests/unittests/models/test_openai.py

@@ -567,6 +567,23 @@ async def gen():
         assert final_response.finish_reason == types.FinishReason.SAFETY
         assert final_response.content.parts[0].text == "Response blocked by content policy."
 
+    # An empty stream truncated by length keeps MAX_TOKENS instead of being marked SAFETY.
+    with mock.patch.object(openai_llm, "_client") as mock_client:
+
+        async def mock_length_stream_gen_func(*args, **kwargs):
+            async def gen():
+                yield MockChunk(finish_reason="length")
+
+            return gen()
+
+        mock_client.chat.completions.create.side_effect = mock_length_stream_gen_func
+
+        stream_results = [resp async for resp in openai_llm.generate_content_async(llm_request, stream=True)]
+
+        final_response = stream_results[-1]
+        assert final_response.finish_reason == types.FinishReason.MAX_TOKENS
+        assert final_response.content.parts[0].text == ""
+
 
 # ============================================================================
 # SSL/TLS Configuration Tests
@@ -988,6 +1005,8 @@ def test_empty_choices_returns_safety_finish_reason(self):
 
         assert llm_response.finish_reason == types.FinishReason.SAFETY
         assert llm_response.content.parts[0].text == "Response blocked by content policy."
+        assert llm_response.usage_metadata is not None
+        assert llm_response.usage_metadata.total_token_count == 8
 
     def test_preserves_thought_signature_from_openai_tool_call_response(self):
         response = self._MockResponse(