Merge branch 'main' into spelling-issue-template

Author: wuliang229

CHANGELOG.md

@@ -79,6 +79,8 @@
 * Fix issue with MCP tools throwing an error ([1a4261a](https://github.com/google/adk-python/commit/1a4261ad4b66cdeb39d39110a086bd6112b17516))
 * Remove redundant `format` field from LiteLLM content objects ([489c39d](https://github.com/google/adk-python/commit/489c39db01465e38ecbc2c7f32781c349b8cddc9))
 * Update the contribution analysis tool to use original write mode ([54db3d4](https://github.com/google/adk-python/commit/54db3d4434e0706b83a589fa2499d11d439a6e4e))
+* Fix agent evaluations detailed output rows wrapping issue([4284c61](https://github.com/google/adk-python/commit/4284c619010b8246c1ecaa011f14b6cc9de512dd))
+* Update dependency version constraints to be based on PyPI versions([0b1784e](https://github.com/google/adk-python/commit/0b1784e0e493a0e2df1edfe37e5ed5f4247e7d9d))
 
 ### Improvements
 
@@ -97,6 +99,7 @@
 * Add sample agent for VertexAiCodeExecutor ([edfe553](https://github.com/google/adk-python/commit/edfe5539421d196ca4da14d3a37fac7b598f8c8d))
 * Adds a new sample agent that demonstrates how to integrate PostgreSQL databases using the Model Context Protocol (MCP) ([45a2168](https://github.com/google/adk-python/commit/45a2168e0e6773e595ecfb825d7e4ab0a38c3a38))
 * Add example for using ADK with Fast MCP sampling ([d3796f9](https://github.com/google/adk-python/commit/d3796f9b33251d28d05e6701f11e80f02a2a49e1))
+* Refactor gepa sample code and clean-up user demo colab([63353b2](https://github.com/google/adk-python/commit/63353b2b74e23e97385892415c5a3f2a59c3504f))
 
 ## [1.17.0](https://github.com/google/adk-python/compare/v1.16.0...v1.17.0) (2025-10-22)
 

contributing/samples/gepa/adk_agent.py

@@ -263,30 +263,35 @@ def run_environment_loop(
           ),
       ],
   )
-  runner = runners.InMemoryRunner(
-      agent=agent,
-      app_name='eval_app',
-      plugins=plugins,
-  )
-  session = asyncio.run(
-      runner.session_service.create_session(
-          app_name='eval_app', user_id='eval_user'
-      )
-  )
-  env_reset_res = env.reset(task_index=task_index)
-  initial_message = types.Content(
-      role='user', parts=[types.Part(text=env_reset_res.observation)]
-  )
-  # The initial message is generated by the environment `reset` within the
-  # implementation of this function - as the first step of the trace.
-  # We yield this first step to ensure we provide a full trace to the user.
-  yield event_lib.Event(
-      author='user',
-      content=initial_message,
-  )
-  for event in runner.run(
-      user_id=session.user_id,
-      session_id=session.id,
-      new_message=initial_message,
-  ):
-    yield event
+
+  async def _async_run():
+    runner = runners.InMemoryRunner(
+        agent=agent,
+        app_name='eval_app',
+        plugins=plugins,
+    )
+    session = await runner.session_service.create_session(
+        app_name='eval_app', user_id='eval_user'
+    )
+    env_reset_res = env.reset(task_index=task_index)
+    initial_message = types.Content(
+        role='user', parts=[types.Part(text=env_reset_res.observation)]
+    )
+    # The initial message is generated by the environment `reset` within the
+    # implementation of this function - as the first step of the trace.
+    # We yield this first step to ensure we provide a full trace to the user.
+    events = [
+        event_lib.Event(
+            author='user',
+            content=initial_message,
+        )
+    ]
+    async for event in runner.run_async(
+        user_id=session.user_id,
+        session_id=session.id,
+        new_message=initial_message,
+    ):
+      events.append(event)
+    return events
+
+  return asyncio.run(_async_run())

contributing/samples/gepa/adk_agent_test.py

@@ -324,24 +324,25 @@ async def _run_async_impl(self, ctx):
   async def _mock_create_session(*args, **kwargs):
     del args, kwargs
     await asyncio.sleep(0.1)
-    return
+    mock_session = mock.Mock()
+    mock.user_id = "fake-user=id"
+    mock.id = "fake-session-id"
+    return mock_session
 
   with mock.patch.object(runners, "InMemoryRunner") as mock_runner_cls:
     mock_runner = mock_runner_cls.return_value
     mock_runner.session_service.create_session.side_effect = (
         _mock_create_session
     )
     mock_runner.run.return_value = []
-    next(
-        adk_agent.run_environment_loop(
-            instruction="some-instruction",
-            env=_TestEnv([]),
-            temperature=0.123,
-            tools=[],
-            task_index=0,
-            agent_model="some-test-model",
-            plugins=[_TestPlugin([])],
-        )
+    adk_agent.run_environment_loop(
+        instruction="some-instruction",
+        env=_TestEnv([]),
+        temperature=0.123,
+        tools=[],
+        task_index=0,
+        agent_model="some-test-model",
+        plugins=[_TestPlugin([])],
     )
     mock_runner_cls.assert_called_once()
     _, runner_kwargs = mock_runner_cls.call_args

contributing/samples/gepa/gepa_tau_bench.ipynb

@@ -31,21 +31,53 @@
       "execution_count": null,
       "metadata": {
         "id": "GqUHYdvRJ7pt",
-        "language": "python"
+        "language": "python",
+        "cellView": "form"
       },
       "outputs": [],
       "source": [
         "#@title Install Tau-bench and GEPA\n",
+        "!git clone https://github.com/google/adk-python.git\n",
         "!git clone https://github.com/sierra-research/tau-bench.git\n",
         "%cd tau-bench/\n",
         "!pip install -e . --quiet\n",
         "\n",
         "%cd ..\n",
         "!pip install gepa --quiet\n",
         "\n",
-        "!pip install retry --quiet"
+        "!pip install retry --quiet\n"
       ]
     },
+    {
+      "cell_type": "code",
+      "source": [
+        "#@title Configure python dependencies\n",
+        "import sys\n",
+        "\n",
+        "sys.path.append('/content/tau-bench')\n",
+        "sys.path.append('/content/adk-python/contributing/samples/gepa')"
+      ],
+      "metadata": {
+        "cellView": "form",
+        "id": "k0nrsIca0yXr"
+      },
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "#@title Authentication\n",
+        "from google.colab import auth\n",
+        "auth.authenticate_user()"
+      ],
+      "metadata": {
+        "cellView": "form",
+        "id": "NsXa217t03vL"
+      },
+      "execution_count": null,
+      "outputs": []
+    },
     {
       "cell_type": "code",
       "execution_count": null,
@@ -55,7 +87,7 @@
       },
       "outputs": [],
       "source": [
-        "#@title Setup and Authentication\n",
+        "#@title Setup\n",
         "from datetime import datetime\n",
         "import json\n",
         "import logging\n",

contributing/samples/gepa/rater_lib.py

@@ -120,7 +120,7 @@ def format_user_agent_conversation(conv: list[dict[str, Any]]) -> str:
         res = (
             res
             + 'The execution result from the agent of function'
-            f' {fc["name"]} is: \n{fc["args"]}\n'
+            f' {fc["name"]} is: \n{fc["response"]}\n'
         )
   return res
 

contributing/samples/live_bidi_streaming_single_agent/agent.py

@@ -65,8 +65,9 @@ async def check_prime(nums: list[int]) -> str:
 
 
 root_agent = Agent(
-    # model='gemini-2.0-flash-live-preview-04-09',  # for Vertex project
-    model='gemini-2.0-flash-live-001',  # for AI studio key
+    model='gemini-live-2.5-flash-preview-native-audio-09-2025',  # vertex
+    # model='gemini-2.5-flash-native-audio-preview-09-2025',  # for AI studio
+    # key
     name='roll_dice_agent',
     description=(
         'hello world agent that can roll a dice of 6 sides and check prime'

src/google/adk/agents/base_agent.py

@@ -164,19 +164,16 @@ def _load_agent_state(
       ctx: InvocationContext,
       state_type: Type[AgentState],
   ) -> Optional[AgentState]:
-    """Loads the agent state from the invocation context, handling resumption.
+    """Loads the agent state from the invocation context.
 
     Args:
       ctx: The invocation context.
       state_type: The type of the agent state.
 
     Returns:
-        The current state if resuming, otherwise None.
+        The current state if exists, otherwise None.
     """
-    if not ctx.is_resumable:
-      return None
-
-    if self.name not in ctx.agent_states:
+    if ctx.agent_states is None or self.name not in ctx.agent_states:
       return None
     else:
       return state_type.model_validate(ctx.agent_states.get(self.name))

src/google/adk/agents/llm_agent.py

@@ -85,6 +85,16 @@
     list[_SingleAfterModelCallback],
 ]
 
+_SingleOnModelErrorCallback: TypeAlias = Callable[
+    [CallbackContext, LlmRequest, Exception],
+    Union[Awaitable[Optional[LlmResponse]], Optional[LlmResponse]],
+]
+
+OnModelErrorCallback: TypeAlias = Union[
+    _SingleOnModelErrorCallback,
+    list[_SingleOnModelErrorCallback],
+]
+
 _SingleBeforeToolCallback: TypeAlias = Callable[
     [BaseTool, dict[str, Any], ToolContext],
     Union[Awaitable[Optional[dict]], Optional[dict]],
@@ -364,6 +374,21 @@ class LlmAgent(BaseAgent):
     The content to return to the user. When present, the actual model response
     will be ignored and the provided content will be returned to user.
   """
+  on_model_error_callback: Optional[OnModelErrorCallback] = None
+  """Callback or list of callbacks to be called when a model call encounters an error.
+
+  When a list of callbacks is provided, the callbacks will be called in the
+  order they are listed until a callback does not return None.
+
+  Args:
+    callback_context: CallbackContext,
+    llm_request: LlmRequest, The raw model request.
+    error: The error from the model call.
+
+  Returns:
+    The content to return to the user. When present, the error will be
+    ignored and the provided content will be returned to user.
+  """
   before_tool_callback: Optional[BeforeToolCallback] = None
   """Callback or list of callbacks to be called before calling the tool.
 
@@ -587,6 +612,20 @@ def canonical_after_model_callbacks(self) -> list[_SingleAfterModelCallback]:
       return self.after_model_callback
     return [self.after_model_callback]
 
+  @property
+  def canonical_on_model_error_callbacks(
+      self,
+  ) -> list[_SingleOnModelErrorCallback]:
+    """The resolved self.on_model_error_callback field as a list of _SingleOnModelErrorCallback.
+
+    This method is only for use by Agent Development Kit.
+    """
+    if not self.on_model_error_callback:
+      return []
+    if isinstance(self.on_model_error_callback, list):
+      return self.on_model_error_callback
+    return [self.on_model_error_callback]
+
   @property
   def canonical_before_tool_callbacks(
       self,

src/google/adk/agents/remote_a2a_agent.py

@@ -372,7 +372,7 @@ def _construct_message_parts_from_session(
       if _is_other_agent_reply(self.name, event):
         event = _present_other_agent_message(event)
 
-      if not event.content or not event.content.parts:
+      if not event or not event.content or not event.content.parts:
         continue
 
       for part in event.content.parts:

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

@@ -977,6 +977,44 @@ async def _run_and_handle_error(
     Yields:
       A generator of LlmResponse.
     """
+
+    from ...agents.llm_agent import LlmAgent
+
+    agent = invocation_context.agent
+    if not isinstance(agent, LlmAgent):
+      raise TypeError(
+          f'Expected agent to be an LlmAgent, but got {type(agent)}'
+      )
+
+    async def _run_on_model_error_callbacks(
+        *,
+        callback_context: CallbackContext,
+        llm_request: LlmRequest,
+        error: Exception,
+    ) -> Optional[LlmResponse]:
+      error_response = (
+          await invocation_context.plugin_manager.run_on_model_error_callback(
+              callback_context=callback_context,
+              llm_request=llm_request,
+              error=error,
+          )
+      )
+      if error_response is not None:
+        return error_response
+
+      for callback in agent.canonical_on_model_error_callbacks:
+        error_response = callback(
+            callback_context=callback_context,
+            llm_request=llm_request,
+            error=error,
+        )
+        if inspect.isawaitable(error_response):
+          error_response = await error_response
+        if error_response is not None:
+          return error_response
+
+      return None
+
     try:
       async with Aclosing(response_generator) as agen:
         async for response in agen:
@@ -985,12 +1023,10 @@ async def _run_and_handle_error(
       callback_context = CallbackContext(
           invocation_context, event_actions=model_response_event.actions
       )
-      error_response = (
-          await invocation_context.plugin_manager.run_on_model_error_callback(
-              callback_context=callback_context,
-              llm_request=llm_request,
-              error=model_error,
-          )
+      error_response = await _run_on_model_error_callbacks(
+          callback_context=callback_context,
+          llm_request=llm_request,
+          error=model_error,
       )
       if error_response is not None:
         yield error_response