fix(tools): add fallback to code_execution_result.output in AgentTool.run_async when text parts are empty

[ilyesJammouci]

Fixes #5481

Summary

AgentTool currently extracts tool output from text parts only.
When an inner agent finishes with code_execution_result parts and no text part, AgentTool returns an empty result string to the parent agent.

This causes silent output loss and prevents parent-level recovery logic.

---

Changes

Updated agent_tool.py in AgentTool.run_async:

• Kept existing text merge behavior as the primary path
• Added a fallback path when merged text is empty:
  • Collect output from code_execution_result parts
  • Join collected outputs into the final merged text
• Preserve empty string behavior when neither text nor code execution output exists

Before

• Extraction logic only used text parts

After

• Text-first extraction with fallback to code_execution_result.output

Big thanks to @Number531 for the clear bug report and for proposing the fix

---

Motivation

In multi-agent workflows using AgentTool with code execution, inner model responses may contain:

• executable_code
• code_execution_result without a textual summary.

Returning an empty result in this case drops valid computation output and can derail long-running optimization or orchestration flows.

This change makes AgentTool robust to that response shape and avoids silent failures.

---

Test Plan

Manual E2E verification

Reproduced the issue with the local runner flow:

python -m google.adk.agent.my_agent.run_repro --offline --runs 30

Before the fix

• Parent tool result was empty when the inner response contained only code_execution_result parts

After the fix

• Parent tool result contains code_execution_result.output for the same response shape
• The empty-result failure mode is no longer observed in this path

Unit tests

• Added regression test in test_agent_tool.py:
  • test_run_async_uses_code_execution_result_output_when_no_text

[adk-bot]

Response from ADK Triaging Agent

Hello @ilyesJammouci, thank you for creating this PR!

To help reviewers verify the fix, could you please provide a manual end-to-end (E2E) test as described in our contribution guidelines? This could be a screenshot from adk web or logs from the runner tool demonstrating that the issue is resolved.

This information will help reviewers to review your PR more efficiently. Thanks!

[rohityan]

Hi @ilyesJammouci , Thank you for your contribution! We appreciate you taking the time to submit this pull request. Can you please fix the failing test before we can proceed with the review.

[ilyesJammouci]

Hi @rohityan, thank you for the feedback! I've fixed the failing test. Please let me know if anything else needs to be addressed.