fix: Fix infinite loop when using code_executor

[hiroakis]

Link to Issue or Description of Change

1. Link to an existing issue (if applicable):

• Closes: [BUG]: Inifinite loop when using code_executor #3921

2. Or, if no issue exists, describe the change:

Problem:
When using code_executor (e.g., UnsafeLocalCodeExecutor), the agent enters an infinite loop. The LLM generates code, the code executor runs it and returns a result, but then the LLM generates the same code again instead of providing a final response.

Solution:
The infinite loop was caused by two issues in _get_contents() flow:

1. _contains_empty_content() was not checking for executable_code and code_execution_result parts. This caused CODE_EXECUTION_RESULT events to be filtered out as "empty content", so the LLM never saw the execution results.

2. Consecutive model contents were sent as separate Content objects. The Gemini API expects executable_code and code_execution_result to be in the same Content with multiple parts. When sent separately, the LLM doesn't understand that the code was already executed.

This fix:

• Adds checks for p.executable_code and p.code_execution_result in _contains_empty_content()
• Merges consecutive model contents into a single Content object in _get_contents()

Testing Plan

Unit Tests:

• I have added or updated unit tests for my change.
• All unit tests pass locally.

$ uv run pytest tests/unittests --tb=no -q

......................................................................................................................................... [  3%]
......................................................................................................................................... [  7%]
......................................................................................................................................... [ 11%]
......................................................................................................................................... [ 15%]
......................................................................................................................................... [ 18%]
......................................................................................................................................... [ 22%]
......................................................................................................................................... [ 26%]
......................................................................................................................................... [ 30%]
......................................................................................................................................... [ 34%]
......................................................................................................................................... [ 37%]
......................................................................................................................................... [ 41%]
......................................................................................................................................... [ 45%]
......................................................................................................................................... [ 49%]
......................................................................................................................................... [ 53%]
......................................................................................................................................... [ 56%]
......................................................................................................................................... [ 60%]
......................................................................................................................................... [ 64%]
......................................................................................................................................... [ 68%]
......................................................................................................................................... [ 72%]
......................................................................................................................................... [ 75%]
......................................................................................................................................... [ 79%]
......................................................................................................................................... [ 83%]
......................................................................................................................................... [ 87%]
......................................................................................................................................... [ 90%]
......................................................................................................................................... [ 94%]
......................................................................................................................................... [ 98%]
.....................................................                                                                                     [100%]
=============================================================== warnings summary ================================================================

...

3615 passed, 1 skipped, 2116 warnings in 35.38s

Manual End-to-End (E2E) Tests:

• Reproduction Code

Based on the code from issue #3921:

import asyncio
from google.adk.agents import Agent
from google.adk.artifacts import InMemoryArtifactService
from google.adk.code_executors import UnsafeLocalCodeExecutor
from google.adk.runners import Runner
from google.adk.sessions import InMemorySessionService
from google.genai import types

agent = Agent(
    model="gemini-2.0-flash",
    name="code_executor_agent",
    instruction="You are a helpful assistant that can execute Python code.",
    code_executor=UnsafeLocalCodeExecutor(),
)

async def main():
    session_service = InMemorySessionService()
    artifact_service = InMemoryArtifactService()

    runner = Runner(
        app_name="test_app",
        agent=agent,
        session_service=session_service,
        artifact_service=artifact_service,
    )

    session = await session_service.create_session(
        app_name="test_app",
        user_id="user1",
    )

    content = types.Content(
        role="user",
        parts=[types.Part.from_text(text="write python code that computes 2+2")]
    )

    print("Starting agent run...")
    count = 0
    async for event in runner.run_async(
        user_id="user1",
        session_id=session.id,
        new_message=content,
    ):
        count += 1
        print(f"\n=== Event {count} ===")
        print(f"Author: {event.author}")
        print(f"Is final: {event.is_final_response()}")

        if event.content and event.content.parts:
            for i, part in enumerate(event.content.parts):
                if hasattr(part, 'code_execution_result') and part.code_execution_result:
                    print(f"Part {i}: CODE_EXECUTION_RESULT")
                    print(f"  Outcome: {part.code_execution_result.outcome}")
                elif hasattr(part, 'executable_code') and part.executable_code:
                    print(f"Part {i}: EXECUTABLE_CODE")
                    print(f"  Code: {part.executable_code.code[:50]}...")
                elif hasattr(part, 'text') and part.text:
                    print(f"Part {i}: TEXT = {part.text[:100]}...")
                elif hasattr(part, 'function_call') and part.function_call:
                    print(f"Part {i}: FUNCTION_CALL")
                elif hasattr(part, 'function_response') and part.function_response:
                    print(f"Part {i}: FUNCTION_RESPONSE")
                else:
                    print(f"Part {i}: OTHER")

        if count > 10:
            print("\nStopping after 10 events...")
            break

if __name__ == "__main__":
    asyncio.run(main())

• Bug Reproduction (google-adk==1.21.0):

uv init -p 3.10 issue3921
cd issue3921/
uv add google-adk==1.21.0
export GOOGLE_CLOUD_PROJECT=$(gcloud config get-value project)
export GOOGLE_CLOUD_LOCATION=us-central1
export GOOGLE_GENAI_USE_VERTEXAI=true
uv run agent.py # Reproduction Code above

• Result (infinite loop - stopped after 10 events):

Starting agent run...

=== Event 1 ===
Author: code_executor_agent
Is final: True
Part 0: EXECUTABLE_CODE
  Code: result = 2 + 2
print(result)...

=== Event 2 ===
Author: code_executor_agent
Is final: False
Part 0: CODE_EXECUTION_RESULT
  Outcome: OUTCOME_OK

=== Event 3 ===
Author: code_executor_agent
Is final: True
Part 0: EXECUTABLE_CODE
  Code: result = 2 + 2
print(result)...

... (repeats infinitely)

=== Event 11 ===
Author: code_executor_agent
Is final: True
Part 0: EXECUTABLE_CODE
  Code: result = 2 + 2
print(result)...

Stopping after 10 events...

• After applying this fix::

mkdir contributing/samples/issue3921-fix
uv run contributing/samples/issue3921-fix/agent.py # Reproduction Code above

• Result (completes normally with 3 events):

Starting agent run...

=== Event 1 ===
Author: code_executor_agent
Is final: True
Part 0: EXECUTABLE_CODE
  Code: result = 2 + 2
print(result)...

=== Event 2 ===
Author: code_executor_agent
Is final: False
Part 0: CODE_EXECUTION_RESULT
  Outcome: OUTCOME_OK

=== Event 3 ===
Author: code_executor_agent
Is final: True
Part 0: TEXT = The previous request asked to compute 2+2 in Python, and the code returned 4. No further instruction...

Checklist

• have read the https://github.com/google/adk-python/blob/main/CONTRIBUTING.md document.
• have performed a self-review of my own code.
• have commented my code, particularly in hard-to-understand areas.
• have added tests that prove my fix is effective or that my feature works.
• New and existing unit tests pass locally with my changes.
• I have manually tested my changes end-to-end.
• Any dependent changes have been merged and published in downstream modules.

Additional context

The new test test_events_with_code_execution_are_included verifies that:

1. Events with executable_code and code_execution_result are NOT filtered out
2. Consecutive model contents are merged into a single Content with multiple parts

[gemini-code-assist]

Summary of Changes

Hello @hiroakis, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request addresses a critical bug that caused agents utilizing the code_executor to enter an infinite loop. The fix ensures that the agent correctly processes and understands code execution results by preventing essential content from being filtered out and by properly merging related content parts for the LLM. This allows for a smooth and complete conversational flow when code execution is involved, ultimately improving the reliability and functionality of the agent system.

Highlights

• Infinite Loop Fix: Resolves an infinite loop issue occurring when the code_executor is used, where the LLM repeatedly generates the same code without progressing to a final response.
• Content Filtering Correction: The _contains_empty_content() function was updated to correctly identify and not filter out executable_code and code_execution_result parts, ensuring the LLM receives crucial execution feedback.
• Consecutive Content Merging: Implemented logic in _get_contents() to merge consecutive model contents, specifically for executable_code and code_execution_result, into a single Content object as expected by the Gemini API for proper understanding of code execution flow.
• New Unit Test: A new unit test, test_events_with_code_execution_are_included, was added to verify that code execution events are correctly processed and merged, preventing regressions.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request effectively resolves a critical infinite loop issue that occurred when using a code_executor. The problem was well-diagnosed, stemming from two distinct issues: code_execution_result events being incorrectly filtered as empty, and consecutive model contents not being merged as the Gemini API expects. The fix is clean and directly addresses both root causes by updating _contains_empty_content to recognize code execution parts and by adding logic to _get_contents to merge consecutive model contents. The addition of the test_events_with_code_execution_are_included unit test is excellent, as it thoroughly validates both aspects of the fix. The changes are correct, well-implemented, and improve the robustness of the code execution flow.

[ryanaiagent]

Hi @hiroakis , Thank you for your contribution! We appreciate you taking the time to submit this pull request. Can you fix lint errors by running autoformat.sh

[hiroakis]

@ryanaiagent Thank you for your review! I've addressed the feedback in ea4515d and also resolved the merge conflicts in 13f8063.