feat(eval): add evaluate_full_response option to rubric-based evaluation

.github/.release-please-manifest.json

@@ -1,3 +1,3 @@
 {
-  ".": "1.29.0"
+  ".": "1.30.0"
 }

.github/release-please-config.json

@@ -1,6 +1,6 @@
 {
   "$schema": "https://raw.githubusercontent.com/googleapis/release-please/main/schemas/config.json",
-  "last-release-sha": "6b1600fbf53bcf634c5fe4793f02921bc0b75125",
+  "last-release-sha": "80a7ecf4b31e4c6de4a1425b03422f384c1a032d",
   "packages": {
     ".": {
       "release-type": "python",

CHANGELOG.md

@@ -1,5 +1,32 @@
 # Changelog
 
+## [1.30.0](https://github.com/google/adk-python/compare/v1.29.0...v1.30.0) (2026-04-13)
+
+
+### Features
+
+* Add Auth Provider support to agent registry ([f2c68eb](https://github.com/google/adk-python/commit/f2c68eb1536f1c0018c2cf7ee3f4417ca442080c))
+* Add Parameter Manager integration to ADK ([b0715d7](https://github.com/google/adk-python/commit/b0715d77a2a433bb2ed07a2475cc4d1f2d662b6c))
+* Add support for Gemma 4 models in ADK ([9d4ecbe](https://github.com/google/adk-python/commit/9d4ecbe9fd1141693e4682cbfe4d542cc62b76ac)), closes [#5156](https://github.com/google/adk-python/issues/5156)
+* allow users to include artifacts from artifact_service in A2A events using provided interceptor ([e63d991](https://github.com/google/adk-python/commit/e63d991be84e373fd31be29d4b6b0e32fdbde557))
+* emit a `TaskStatusUpdateEvent` for ADK events with no output parts but with event.actions ([dcc485b](https://github.com/google/adk-python/commit/dcc485b23e3509e2e386636d841033b91c9a401c))
+* Live avatar support in ADK ([a64a8e4](https://github.com/google/adk-python/commit/a64a8e46480753439b91b9cfd41fd190b4dad493))
+* **live:** expose live_session_resumption_update as Event in BaseLlmFlow ([2626ad7](https://github.com/google/adk-python/commit/2626ad7c69fb64a88372225d5583085fc08b1fcd)), closes [#4357](https://github.com/google/adk-python/issues/4357)
+* Promote BigQuery tools to Stable ([abcf14c](https://github.com/google/adk-python/commit/abcf14c166baf4f8cc6e919b1eb4c063bf3a92af))
+* **samples:** add sample for skill activation via environment tools ([2cbb523](https://github.com/google/adk-python/commit/2cbb52306910fac994fe1d29bdfcfacb258703b4))
+
+
+### Bug Fixes
+
+* Add "gcloud config unset project" command to express mode flow ([e7d8160](https://github.com/google/adk-python/commit/e7d81604126cbdb4d9ee4624e1d1410b06585750))
+* avoid load all agents in adk web server ([cb4dd42](https://github.com/google/adk-python/commit/cb4dd42eff2df6d20c5e53211718ecb023f127fc))
+* Change express mode user flow so it's more clear that an express mode project is being created ([0fedb3b](https://github.com/google/adk-python/commit/0fedb3b5eb2074999d8ccdb839e054ea80da486f))
+* Custom pickling in McpToolset to exclude unpicklable objects like errlog ([d62558c](https://github.com/google/adk-python/commit/d62558cc2d7d6c0372e43c9f009c8c7a6863ff0a))
+* Fix credential leakage vulnerability in Agent Registry ([e3567a6](https://github.com/google/adk-python/commit/e3567a65196bb453cdac4a5ae42f7f079476d748))
+* Include a link to the deployed agent ([547766a](https://github.com/google/adk-python/commit/547766a47779915a8a47745237a46882a02dae9a))
+* preserve interaction ids for interactions SSE tool calls ([9a19304](https://github.com/google/adk-python/commit/9a1930407a4eff67093ea9f14292f1931631a661)), closes [#5169](https://github.com/google/adk-python/issues/5169)
+* validate user_id and session_id against path traversal ([cbcb5e6](https://github.com/google/adk-python/commit/cbcb5e6002b5bae89de5309caf7b9bb02d563cfc)), closes [#5110](https://github.com/google/adk-python/issues/5110)
+
 ## [1.29.0](https://github.com/google/adk-python/compare/v1.28.0...v1.29.0) (2026-04-09)
 
 

pyproject.toml

@@ -44,7 +44,7 @@ dependencies = [
   "google-cloud-spanner>=3.56.0, <4.0.0",                    # For Spanner database
   "google-cloud-speech>=2.30.0, <3.0.0",                     # For Audio Transcription
   "google-cloud-storage>=2.18.0, <4.0.0",                    # For GCS Artifact service
-  "google-genai>=1.64.0, <2.0.0",                            # Google GenAI SDK
+  "google-genai>=1.72.0, <2.0.0",                            # Google GenAI SDK
   "graphviz>=0.20.2, <1.0.0",                                # Graphviz for graph rendering
   "httpx>=0.27.0, <1.0.0",                                   # HTTP client library
   "jsonschema>=4.23.0, <5.0.0",                              # Agent Builder config validation

src/google/adk/agents/run_config.py

@@ -198,6 +198,9 @@ class RunConfig(BaseModel):
   response_modalities: Optional[list[str]] = None
   """The output modalities. If not set, it's default to AUDIO."""
 
+  avatar_config: Optional[types.AvatarConfig] = None
+  """Avatar configuration for the live agent."""
+
   save_input_blobs_as_artifacts: bool = Field(
       default=False,
       deprecated=True,

src/google/adk/evaluation/eval_metrics.py

@@ -144,6 +144,19 @@ class RubricsBasedCriterion(BaseCriterion):
       ),
   )
 
+  evaluate_full_response: bool = Field(
+      default=False,
+      description=(
+          "Whether to evaluate the full agent response including intermediate"
+          " natural language text (e.g. text emitted before tool calls) in"
+          " addition to the final response. By default, only the final"
+          " response text is sent to the judge. When True, text from all"
+          " intermediate invocation events is concatenated with the final"
+          " response before evaluation. This is useful for agents that emit"
+          " text both before and after tool calls within a single invocation."
+      ),
+  )
+
 
 class HallucinationsCriterion(BaseCriterion):
   """Criterion to use when evaluating agents response for hallucinations."""

src/google/adk/evaluation/rubric_based_final_response_quality_v1.py

@@ -274,7 +274,18 @@ def format_auto_rater_prompt(
     """Returns the autorater prompt."""
     self.create_effective_rubrics_list(actual_invocation.rubrics)
     user_input = get_text_from_content(actual_invocation.user_content)
-    final_response = get_text_from_content(actual_invocation.final_response)
+
+    # When evaluate_full_response is enabled, include text from intermediate
+    # invocation events (e.g. text emitted before tool calls) in addition to
+    # the final response. This is useful for agents that stream text, call
+    # tools, then stream more text within a single invocation.
+    criterion = self._eval_metric.criterion
+    evaluate_full = getattr(criterion, "evaluate_full_response", False)
+
+    if evaluate_full:
+      final_response = self._get_full_response_text(actual_invocation)
+    else:
+      final_response = get_text_from_content(actual_invocation.final_response)
 
     rubrics_text = "\n".join([
         f"*  {r.rubric_content.text_property}"
@@ -310,3 +321,25 @@ def format_auto_rater_prompt(
     )
 
     return auto_rater_prompt
+
+  @staticmethod
+  def _get_full_response_text(invocation: Invocation) -> str:
+    """Concatenates all NL text from invocation events and the final response.
+
+    When an agent emits text before a tool call (e.g. presenting a plan),
+    that text is stored in intermediate_data.invocation_events but not in
+    final_response. This method collects text from both sources to give the
+    judge a complete picture of the agent's output.
+    """
+    parts = []
+    if invocation.intermediate_data and isinstance(
+        invocation.intermediate_data, InvocationEvents
+    ):
+      for evt in invocation.intermediate_data.invocation_events:
+        text = get_text_from_content(evt.content)
+        if text:
+          parts.append(text)
+    final_text = get_text_from_content(invocation.final_response)
+    if final_text:
+      parts.append(final_text)
+    return "\n\n".join(parts)

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

@@ -90,6 +90,9 @@ def _build_basic_request(
   llm_request.live_connect_config.context_window_compression = (
       invocation_context.run_config.context_window_compression
   )
+  llm_request.live_connect_config.avatar_config = (
+      invocation_context.run_config.avatar_config
+  )
 
 
 class _BasicLlmRequestProcessor(BaseLlmRequestProcessor):

src/google/adk/models/gemini_llm_connection.py

@@ -115,16 +115,7 @@ async def send_content(self, content: types.Content):
       is_gemini_31 = model_name_utils.is_gemini_3_1_flash_live(
           self._model_version
       )
-      is_gemini_api = self._api_backend == GoogleLLMVariant.GEMINI_API
-
-      # As of now, Gemini 3.1 Flash Live is only available in Gemini API, not
-      # Vertex AI.
-      if (
-          is_gemini_31
-          and is_gemini_api
-          and len(content.parts) == 1
-          and content.parts[0].text
-      ):
+      if is_gemini_31 and len(content.parts) == 1 and content.parts[0].text:
         logger.debug('Using send_realtime_input for Gemini 3.1 text input')
         await self._gemini_session.send_realtime_input(
             text=content.parts[0].text
@@ -149,11 +140,7 @@ async def send_realtime(self, input: RealtimeInput):
       is_gemini_31 = model_name_utils.is_gemini_3_1_flash_live(
           self._model_version
       )
-      is_gemini_api = self._api_backend == GoogleLLMVariant.GEMINI_API
-
-      # As of now, Gemini 3.1 Flash Live is only available in Gemini API, not
-      # Vertex AI.
-      if is_gemini_31 and is_gemini_api:
+      if is_gemini_31:
         if input.mime_type and input.mime_type.startswith('audio/'):
           await self._gemini_session.send_realtime_input(audio=input)
         elif input.mime_type and input.mime_type.startswith('image/'):

src/google/adk/sessions/vertex_ai_session_service.py

@@ -270,6 +270,7 @@ async def append_event(self, session: Session, event: Event) -> Event:
 
     reasoning_engine_id = self._get_reasoning_engine_id(session.app_name)
 
+    # Build config (Monolithic approach)
     config = {}
     if event.content:
       config['content'] = event.content.model_dump(
@@ -286,9 +287,6 @@ async def append_event(self, session: Session, event: Event) -> Event:
               k: json.loads(v.model_dump_json(exclude_none=True, by_alias=True))
               for k, v in event.actions.requested_auth_configs.items()
           },
-          # TODO: add requested_tool_confirmations, agent_state once
-          # they are available in the API.
-          # Note: compaction is stored via event_metadata.custom_metadata.
       }
     if event.error_code:
       config['error_code'] = event.error_code
@@ -311,10 +309,8 @@ async def append_event(self, session: Session, event: Event) -> Event:
       metadata_dict['grounding_metadata'] = event.grounding_metadata.model_dump(
           exclude_none=True, mode='json'
       )
-    # Store compaction data in custom_metadata since the Vertex AI service
-    # does not yet support the compaction field.
-    # TODO: Stop writing to custom_metadata once the Vertex AI service
-    # supports the compaction field natively in EventActions.
+
+    # ALWAYS write to custom_metadata
     if event.actions and event.actions.compaction:
       compaction_dict = event.actions.compaction.model_dump(
           exclude_none=True, mode='json'
@@ -324,8 +320,6 @@ async def append_event(self, session: Session, event: Event) -> Event:
           key=_COMPACTION_CUSTOM_METADATA_KEY,
           value=compaction_dict,
       )
-    # Store usage_metadata in custom_metadata since the Vertex AI service
-    # does not persist it in EventMetadata.
     if event.usage_metadata:
       usage_dict = event.usage_metadata.model_dump(
           exclude_none=True, mode='json'
@@ -335,7 +329,12 @@ async def append_event(self, session: Session, event: Event) -> Event:
           key=_USAGE_METADATA_CUSTOM_METADATA_KEY,
           value=usage_dict,
       )
+
     config['event_metadata'] = metadata_dict
+
+    # Persist the full event state using raw_event. If the client-side SDK
+    # does not support this field, it will raise a ValidationError, and we
+    # will fall back to legacy field-based storage.
     config['raw_event'] = event.model_dump(
         exclude_none=True,
         mode='json',
@@ -345,7 +344,8 @@ async def append_event(self, session: Session, event: Event) -> Event:
     # Retry without raw_event if client side validation fails for older SDK
     # versions.
     async with self._get_api_client() as api_client:
-      try:
+
+      async def _do_append(cfg: dict[str, Any]):
         await api_client.agent_engines.sessions.events.append(
             name=(
                 f'reasoningEngines/{reasoning_engine_id}/sessions/{session.id}'
@@ -355,22 +355,16 @@ async def append_event(self, session: Session, event: Event) -> Event:
             timestamp=datetime.datetime.fromtimestamp(
                 event.timestamp, tz=datetime.timezone.utc
             ),
-            config=config,
+            config=cfg,
         )
+
+      try:
+        await _do_append(config)
       except pydantic.ValidationError:
+        logger.warning('Vertex SDK does not support raw_event, falling back.')
         if 'raw_event' in config:
           del config['raw_event']
-        await api_client.agent_engines.sessions.events.append(
-            name=(
-                f'reasoningEngines/{reasoning_engine_id}/sessions/{session.id}'
-            ),
-            author=event.author,
-            invocation_id=event.invocation_id,
-            timestamp=datetime.datetime.fromtimestamp(
-                event.timestamp, tz=datetime.timezone.utc
-            ),
-            config=config,
-        )
+        await _do_append(config)
     return event
 
   def _get_reasoning_engine_id(self, app_name: str):
@@ -429,8 +423,8 @@ def _get_raw_event(api_event_obj: Any) -> dict[str, Any] | None:
 
 def _from_api_event(api_event_obj: vertexai.types.SessionEvent) -> Event:
   """Converts an API event object to an Event object."""
-  # Read event data from raw_event first before falling back to top level
-  # fields.
+  # Prioritize reading from raw_event to restore full state. Fall back to
+  # top-level fields for older data that lacks raw_event.
   raw_event_dict = _get_raw_event(api_event_obj)
   if raw_event_dict:
     event_dict = copy.deepcopy(raw_event_dict)
@@ -439,8 +433,9 @@ def _from_api_event(api_event_obj: vertexai.types.SessionEvent) -> Event:
         'id': api_event_obj.name.split('/')[-1],
         'invocation_id': getattr(api_event_obj, 'invocation_id', None),
         'author': getattr(api_event_obj, 'author', None),
-        'timestamp': timestamp_obj.timestamp() if timestamp_obj else None,
     })
+    if timestamp_obj:
+      event_dict['timestamp'] = timestamp_obj.timestamp()
     return Event.model_validate(event_dict)
 
   actions = getattr(api_event_obj, 'actions', None)
@@ -514,6 +509,13 @@ def _from_api_event(api_event_obj: vertexai.types.SessionEvent) -> Event:
         usage_metadata_data
     )
 
+  timestamp_obj = getattr(api_event_obj, 'timestamp', None)
+  timestamp = (
+      timestamp_obj.timestamp()
+      if timestamp_obj
+      else datetime.datetime.now(datetime.timezone.utc).timestamp()
+  )
+
   return Event(
       id=api_event_obj.name.split('/')[-1],
       invocation_id=api_event_obj.invocation_id,
@@ -522,7 +524,7 @@ def _from_api_event(api_event_obj: vertexai.types.SessionEvent) -> Event:
       content=_session_util.decode_model(
           getattr(api_event_obj, 'content', None), types.Content
       ),
-      timestamp=api_event_obj.timestamp.timestamp(),
+      timestamp=timestamp,
       error_code=getattr(api_event_obj, 'error_code', None),
       error_message=getattr(api_event_obj, 'error_message', None),
       partial=partial,

src/google/adk/tools/_automatic_function_calling_util.py

@@ -368,6 +368,11 @@ def from_function_with_options(
           parameters_json_schema[name] = types.Schema.model_validate(
               json_schema_dict
           )
+          if param.default is not inspect.Parameter.empty:
+            if param.default is not None:
+              parameters_json_schema[name].default = param.default
+            else:
+              parameters_json_schema[name].nullable = True
         except Exception as e:
           _function_parameter_parse_util._raise_for_unsupported_param(
               param, func.__name__, e
@@ -392,6 +397,11 @@ def from_function_with_options(
         type='OBJECT',
         properties=parameters_json_schema,
     )
+    declaration.parameters.required = (
+        _function_parameter_parse_util._get_required_fields(
+            declaration.parameters
+        )
+    )
 
   if variant == GoogleLLMVariant.GEMINI_API:
     return declaration

src/google/adk/utils/model_name_utils.py

@@ -130,9 +130,6 @@ def is_gemini_2_or_above(model_string: Optional[str]) -> bool:
 def is_gemini_3_1_flash_live(model_string: Optional[str]) -> bool:
   """Check if the model is a Gemini 3.1 Flash Live model.
 
-  Note: This is a very specific model name for live bidi streaming, so we check
-  for exact match.
-
   Args:
     model_string: The model name
 
@@ -141,5 +138,4 @@ def is_gemini_3_1_flash_live(model_string: Optional[str]) -> bool:
   """
   if not model_string:
     return False
-
-  return model_string == 'gemini-3.1-flash-live-preview'
+  return model_string.startswith('gemini-3.1-flash-live')

src/google/adk/version.py

@@ -13,4 +13,4 @@
 # limitations under the License.
 
 # version: major.minor.patch
-__version__ = "1.29.0"
+__version__ = "1.30.0"

tests/unittests/agents/test_run_config.py

@@ -17,6 +17,7 @@
 from unittest.mock import patch
 
 from google.adk.agents.run_config import RunConfig
+from google.genai import types
 import pytest
 
 
@@ -64,3 +65,35 @@ def test_audio_transcription_configs_are_not_shared_between_instances():
   assert (
       config1.input_audio_transcription is not config2.input_audio_transcription
   )
+
+
+def test_avatar_config_initialization():
+  custom_avatar = types.CustomizedAvatar(
+      image_mime_type="image/jpeg", image_data=b"image_bytes"
+  )
+  avatar_config = types.AvatarConfig(
+      audio_bitrate_bps=128000,
+      video_bitrate_bps=1000000,
+      customized_avatar=custom_avatar,
+  )
+  run_config = RunConfig(avatar_config=avatar_config)
+
+  assert run_config.avatar_config == avatar_config
+  assert run_config.avatar_config.customized_avatar == custom_avatar
+  assert (
+      run_config.avatar_config.customized_avatar.image_mime_type == "image/jpeg"
+  )
+  assert run_config.avatar_config.customized_avatar.image_data == b"image_bytes"
+
+
+def test_avatar_config_with_name():
+  avatar_config = types.AvatarConfig(
+      audio_bitrate_bps=128000,
+      video_bitrate_bps=1000000,
+      avatar_name="test_avatar",
+  )
+  run_config = RunConfig(avatar_config=avatar_config)
+
+  assert run_config.avatar_config == avatar_config
+  assert run_config.avatar_config.avatar_name == "test_avatar"
+  assert run_config.avatar_config.customized_avatar is None