Raw event.content assigned to child.output without transformation in _reconstruct_node_states during node re-run, for the nodes with message_as_input = True

[samarth1224]

Describe the Bug:
I encountered bug during the execution of following code:-

analyzer = Agent(
    name="analyzer_agent",
    model="gemini-3.1-flash-lite-preview",
    output_schema=Action,
    instruction=analyzer_prompt,
)
 @node
async def main_workflow(ctx: Context):
    report = await ctx.run_node(check_network_status_node)
    if not report:
        yield "network Issue"
        return

    agent_output = await ctx.run_node(analyzer, report) #Analyzer is LLMAgent
    agent_output = Action.model_validate(agent_output)
    
    while agent_output.output_type == 'request_approval':

        result = await ctx.run_node(handle_process, agent_output)
        # main_workflow node start re-running after here
        # And node subsequent to this do not run right after.
        if result:
            agent_output = await ctx.run_node(analyzer,result)
            agent_output = Action.model_validate(agent_output)
        else:
            print('No result')
     
    yield agent_output

root_agent = Workflow(
    name="network_analyzer",
    edges=[("START", main_workflow)]
)

The await ctx.run_node(analyzer, report) is producing types.Content object. instead of dictionary object expected, during the re-run/resuming of the node( main_workflow re-runs, therefore. all the child of it re-reruns)

The fresh execution of the nodes runs fine and produces the excepted output. The issue evolves when the parent node(main_workflow) is re-rerun/resume after yielding the RequestInput response, it re-runs the dynamic child nodes using ctx.run_node(child_node) , which as expected, do not actually re-rerun but rehydrates the node's state using the Session Events(SSE) present in invocation context. This Events have been altered in _consume_event_queue()(which consume''s event from the event queue) function in the runner before putting them into Session.
A Node with message_as_input == True (Specifically, LLMAgent node), will have its output nullified in the _consume_event_queue() function.

     # _consume_event_queue() function ....
        if event.node_info.message_as_output and event.content is not None:
          event = event.model_copy()
          event.output = None
       #rest of the  function......

Everything good till here since, LLMAgent produces a event with a content object, therefore, a specific output field in the event is not required in the event.
But the main issue lies in the fact that event.content is not being processed before assigning it to child.output in _reconstruct_node_state() function which reconstruct the node's prior state(the state when the node previously executed), and this child.output is propagated as the node's output up till ctx.run_node.

#...... _reconstruct_node_state() function

    if is_direct or is_delegated:
      if event.output is not None:
        child.output = event.output
        child.branch = event.branch
      elif use_message_as_output:
        child.output = event.content
  #.....rest of the function

So basically, the nodes with message_as_output == True, are producing the expected output during fresh execution but when re-rerun and subsequently re-hydrated they produce the raw event.content instead of the output.

In nutshell:

1)Fresh Run: run_llm_agent_as_node() populates event.output. _track_event_in_context() sees this and correctly assigns the string/schema to ctx.output.

2)Persistence: _consume_event_queue() sets event.output = None to optimize for message_as_output logic and saves the raw event.content to the Session.

3)Resumption: When the main_workflow resumes, _reconstruct_node_states() attempts to rebuild the child state. Because event.output is now None, it falls back to:

elif use_message_as_output:
    child.output = event.content # <--- Error: Assigns raw container/object 

4)Failure: This raw container is returned to ctx.run_node(analyzer, report), causing Action.model_validate(agent_output) to fail because it received types.Contentobject instead of a valid Action schema or dictionary.

Steps to Reproduce:
Please provide a numbered list of steps to reproduce the behavior:

1. Create a main function node with LLMagent node and a node which yields RequestInput object.
2. Make the LLMAgent node run before the RequestInput Node.
3. Have the LLMAgent's output validated if output schema is provided or check it against the string data type.
4. The error will occur when the main function node is rerun after the getting the input from user.

Expected Behavior:
When re-running the node. LLMAgent node's output should have have been the output returned during the fresh execution that is a dictionary.

Observed Behavior:
The raw types.Content is returned which failed to verify against the output_schema of the LLMagent node

Environment Details:

• ADK Library Version (pip show google-adk): 2.0.0b1
• Desktop OS:** Windows 10
• Python Version (python -V): Python 3.13.4

Model Information:

• Are you using LiteLLM: No
• Which model is being used: gemini-3.1-flash-lite-preview

---

🟡 Optional Information

From what I understand about the codebase, during a child node's fresh run the events are enqueued in the event queue, and event in this queue are consumed by the runner where as stated earlier they are altered.
And before the events are put in the event queue they are tracked in the context of the child node in the function _track_event_in_context() which puts event.output into child nodes ctx.output.

  def _track_event_in_context(self, event: Event, ctx: Context) -> None:
    """Write yielded event results to ctx (source of truth)."""
    if event.output is not None:
      ctx.output = event.output
    elif event.node_info and event.node_info.message_as_output:
      ctx.output = event.content
# ..... rest of the function 

But I suppose this function intends to make the

ctx.output = event.content

instead of

ctx.output = event.output

But since it can not , because in process_llm_agent_output() function it puts the event.output = output
hence, populating the event.output.

So there could additionally be one more fix by either:

1. Assign ctx.output = event.content (off course after processing the content, maybe using the process_llm_agent_output() function)

or

2. If we strictly want the output field to be None, even during the fresh execution of the node, remove process_llm_agent_output() function to process the output in the run_llm_agent_as_node() method.

Screenshots / Video:

Logs:

  File "D:\Projects\googleadk\.venv\Lib\site-packages\google\adk\workflow\_node_runner.py", line 127, in run
    await self._execute_node(ctx, node_input)
  File "D:\Projects\googleadk\.venv\Lib\site-packages\google\adk\workflow\_node_runner.py", line 250, in _execute_node
    await self._run_node_loop(ctx, node_input)
  File "D:\Projects\googleadk\.venv\Lib\site-packages\google\adk\workflow\_node_runner.py", line 264, in _run_node_loop
    async for event in agen:
      self._track_event_in_context(event, ctx)
      await self._enqueue_event(event, ctx)
  File "D:\Projects\googleadk\.venv\Lib\site-packages\google\adk\workflow\_base_node.py", line 218, in run
    async for item in agen:
    ...<12 lines>...
        yield Event(output=validated)
  File "D:\Projects\googleadk\.venv\Lib\site-packages\google\adk\workflow\_function_node.py", line 508, in _run_impl
    async for item in items:
    ...<2 lines>...
        yield event
  File "D:\Projects\googleadk\agent\agent.py", line 78, in main_workflow
    agent_output = Action.model_validate(agent_output)
  File "D:\Projects\googleadk\.venv\Lib\site-packages\pydantic\main.py", line 732, in model_validate
    return cls.__pydantic_validator__.validate_python(
           ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^
        obj,
        ^^^^
    ...<5 lines>...
        by_name=by_name,
        ^^^^^^^^^^^^^^^^
    )
    ^
pydantic_core._pydantic_core.ValidationError: 1 validation error for Action
  Input should be a valid dictionary or instance of Action [type=model_type, input_value=Content(
  parts=[
    Pa...),
  ],
  role='model'
), input_type=Content]

Screenshots / Video:

Additional Context:
'N/A

Minimal Reproduction Code:
Please provide a code snippet or a link to a Gist/repo that isolates the issue.

from google.adk import Workflow
from google.adk.events import RequestInput
from google.adk import Context
from google.adk.agents import Agent
from google.adk.workflow import node

from pydantic import BaseModel

from typing import Any
class MySchema(BaseModel):
    greeting: str

analyzer = Agent(
    name="my_agent",
    model="gemini-3.1-flash-lite-preview",
    output_schema=MySchema,
    instruction="Your job is to greet the user.",
)
@node(rerun_on_resume=False)
async def get_user_approval(ctx: Context, node_input: Any):
    """Yields a RequestInput to pause the workflow and wait for user input."""
    yield RequestInput(message=f'please approve or reject.',response_schema = str)

@node(rerun_on_resume=True)
async def handle_process(ctx: Context, node_input: Any):
    """The orchestrator calling the interactive step."""
    user_response = await ctx.run_node(get_user_approval,node_input)
    if user_response.lower() == "yes":
        yield 'approved'
    yield "Denied"
    return

@node(rerun_on_resume=True)
async def my_workflow(ctx:Context):
    agent_input = "my agent input"
    agent_output = await ctx.run_node(analyzer,agent_input)
    agent_output = MySchema.validate(agent_output)
    result = await ctx.run_node(handle_process)
    yield result

root_agent = Workflow(
    name="greet_user",
    edges=[("START", my_workflow)]
)

Screenshots / Video:

How often has this issue occurred?:

• Always (100%)