Skip to content

fix(tools): cancel sibling tools on child-originated cancellation - #7180

Closed
Anusha0501 wants to merge 4 commits into
google:mainfrom
Anusha0501:fix/parallel-tool-cancel-siblings
Closed

Anusha0501 wants to merge 4 commits into
google:mainfrom
Anusha0501:fix/parallel-tool-cancel-siblings

Conversation

@Anusha0501

@Anusha0501 Anusha0501 commented Sep 18, 2026 •

Copy link
Copy Markdown
Contributor

Please ensure you have read the contribution guide before creating a pull request.

Link to Issue or Description of Change

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

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

Problem:
When one tool in a parallel batch is cancelled (it raises asyncio.CancelledError, or it cancels its own task), _gather_or_cancel does not tear down the rest of the batch. asyncio.CancelledError is a BaseException, so the existing except Exception path never runs. asyncio.gather() also does not cancel siblings when a child is cancelled. The public Runner iterator can therefore finish while a sibling tool is still executing and can still perform a side effect.

Solution:
Catch (Exception, asyncio.CancelledError) in _gather_or_cancel and reuse the existing cancel-unfinished-tasks / await-cleanup / re-raise path. Cancellation stays cancellation; it is not converted into a tool error.

Testing Plan

Unit Tests:

  • I have added or updated unit tests for my change.
  • All unit tests pass locally.
PYTHONPATH=src pytest tests/unittests/flows/llm_flows/tools/test_batch_executor.py tests/unittests/flows/llm_flows/tools/test_functions_parallel_error.py

19 passed

Added:

  • test_gather_or_cancel_cancels_siblings_on_cancelled_error
  • test_gather_or_cancel_cancels_siblings_when_child_cancels_itself
  • test_parallel_function_call_cancels_siblings_on_cancelled_error (parametrized: run_async, SSE, run_live)
  • test_parallel_function_call_cancels_siblings_when_tool_cancels_itself (parametrized: run_async, SSE, run_live)

The Runner-level tests use the reporter's Event handshake from #7172 and assert the sibling is cancelled and not still pending when the iterator returns. They now cover non-streaming run_async, run_async with SSE, and run_live BIDI, matching the #7172 matrix.

Manual End-to-End (E2E) Tests:

Covered by the Runner-level unit tests above (MockModel, no live model API). Those reproduce the #7172 matrix for CancelledError and self-cancel across run_async, SSE, and run_live.

Checklist

  • I have read the CONTRIBUTING.md document.
  • I have performed a self-review of my own code.
  • I have commented my code, particularly in hard-to-understand areas.
  • I 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

Reporter's diagnosis and candidate repair from #7172 (@jaywang172): _gather_or_cancel catches Exception only; CancelledError bypasses sibling cleanup. Caller-cancel of the whole batch already works because gather cancels children when the gather wait itself is cancelled. This PR only covers child-originated cancellation.

asyncio.CancelledError is a BaseException, so _gather_or_cancel only
tore down siblings on Exception. A parallel tool that raised
CancelledError or cancelled itself left sibling tasks running after
the batch returned.

Fixes google#7172
@jaywang172

Copy link
Copy Markdown
Contributor

Thanks for putting together the fix and for crediting the reporter's Event handshake in the testing section.

One attribution correction: the “Maintainer diagnosis on #7172” paragraph refers to my investigation as the issue reporter, rather than a maintainer response. The root-cause analysis, deterministic reproducer, and candidate (Exception, asyncio.CancelledError) repair were included in my original report. Could you change that wording to “Reporter's diagnosis and candidate repair from #7172 (@jaywang172)”?

My local probes also cover run_async with SSE and run_live with BIDI using MockModel, in addition to non-streaming run_async. Both child-cancellation forms reproduce across those paths. I'd be happy to contribute focused SSE/live regression coverage here so we can preserve the batch-ownership invariant across execution modes without opening a duplicate fix PR. Let me know if you'd like to coordinate that addition.

google#7172 reproduced child-originated cancellation on run_async SSE and
run_live BIDI as well as non-streaming run_async. Parametrize the
Runner tests across those three modes so the batch-ownership invariant
holds on each path.
@Anusha0501

Copy link
Copy Markdown
Contributor Author

@jaywang172 Thanks for the correction, and for the original diagnosis, reproducer, and candidate repair on #7172.

The additional-context wording is now Reporter's diagnosis and candidate repair from #7172 (@jaywang172).

I also added the SSE and run_live BIDI coverage you described, using the same Event handshake. The two Runner tests are parametrized across non-streaming run_async, run_async with SSE, and run_live, for both child-cancellation forms. Locally: 19 passed on test_batch_executor.py and test_functions_parallel_error.py.

That should keep the batch-ownership invariant on those paths in this PR, so a follow-up fix PR is not needed. Further cases you still want to add are welcome as review comments or a follow-up on this branch.

@adk-bot

adk-bot commented Oct 12, 2026

Copy link
Copy Markdown
Collaborator

Thank you @Anusha0501 for your contribution! 🎉

Your changes have been successfully imported and merged via Copybara in commit 1ab2fe8.

Closing this PR as the changes are now in the main branch.

@adk-bot adk-bot added the merged [Status] This PR is merged label Oct 12, 2026
@adk-bot adk-bot closed this Oct 12, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

merged [Status] This PR is merged

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Parallel tool child cancellation leaves sibling tasks running past invocation completion

4 participants