Code Executor Output Not Available in Next React Round

[liuzengh]

Problem Description

When using the built-in React multi-round flow in adk-python with the /*ACTION*/ - /*REASONING*/ pattern, the next round cannot access the output results from the code executor in the previous round.

Root Cause Analysis

The current code executor is designed as a processor. This means it functions similarly to an insertion point that:

1. Extracts code snippets from the LLM's message response
2. Executes the code
3. Appends the execution results to the end of the message response

In the React mode:

• /*ACTION*/ represents tool calls, where results are fed back to the LLM
• However, the current code executor is not implemented as a tool call
• Therefore, its results are not automatically fed back to the LLM

Impact

Since the code executor results are not passed back to the LLM, the subsequent round's /*REASONING*/ or /*FINAL_ANSWER*/ stages cannot access the execution output.

Proposed Solution

Either:

1. Refactor the code executor to work as a tool call so its results are automatically fed back to the LLM
2. Modify the processor logic to explicitly inject code executor results into the context for the next round

[liuzengh]

add PlanReActPlanner and UnsafeLocalCodeExecutor to adk-python/contributing/samples/fields_planner

root_agent = Agent(
    model='gemini-2.5-flash',
    # model='gemini-2.0-flash',
    name='data_processing_agent',
    description=(
        'hello world agent that can roll a dice of 8 sides and check prime'
        ' numbers.'
    ),
    tools=[
        roll_die,
        check_prime,
    ],
    planner=PlanReActPlanner(),
    code_executor=UnsafeLocalCodeExecutor(),
    generate_content_config=types.GenerateContentConfig(
        safety_settings=[
            types.SafetySetting(  # avoid false alarm about rolling dice.
                category=types.HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT,
                threshold=types.HarmBlockThreshold.OFF,
            ),
        ]
    ),
)

run results:

Enter your messages (type "quit" or "exit" to stop):
You: roll a random side die, and write a python code to print the result
** User says: {'parts': [{'text': 'roll a random side die, and write a python code to print the result'}], 'role': 'user'}
** data_processing_agent: /*PLANNING*/
1. Ask the user to specify the number of sides for the die to be rolled, as the request "roll a random side die" is ambiguous.
2. Once the user provides the number of sides, use the `default_api.roll_die` tool to roll the die.
3. Capture the result from the `roll_die` tool.
4. Construct a Python code snippet that prints the captured die roll result.
5. Provide the constructed Python code snippet as the final answer.

/*REASONING*/
The user asked to roll a "random side die", but the `roll_die` function requires a specific number of sides. Therefore, I need to ask the user for clarification on how many sides the die should have before I can roll it.
Please specify the number of sides for the die.
You: 8 side
** User says: {'parts': [{'text': '8 side'}], 'role': 'user'}
** data_processing_agent: /*PLANNING*/
1. Roll an 8-sided die using the `default_api.roll_die` tool.
2. Capture the result from the `roll_die` tool.
3. Construct a Python code snippet that prints the captured die roll result.
4. Provide the constructed Python code snippet as the final answer.

/*ACTION*/

** data_processing_agent: /*REASONING*/
The 8-sided die was rolled, and the result is 4. Now I need to generate the Python code to print this result.

/*FINAL_ANSWER*/



** data_processing_agent: I'm sorry, I cannot execute code. The previous response was the python code to print the result. Can I help you with anything else?
You: