fix: keep concurrent tool results in request order

[he-yufeng]

Motivation

ConcurrentToolExecutor currently passes the same tool_results list into every running task. Each task appends when it finishes, so the follow-up toolResult message is ordered by scheduler completion instead of by the model's original toolUse order.

That makes prompts unstable for parallel tool calls: a fast second tool can appear before a slow first tool, breaking prompt-cache and replay determinism.

Closes #2112.

Changes

• Give each concurrent task its own result list.
• Merge those per-task results back into the shared tool_results list in tool_uses order after all tasks finish.
• Add a regression test where the first tool intentionally finishes after the second tool, but tool_results still follows request order.

Testing

• .\.venv\Scripts\python.exe -m pytest tests\strands\tools\executors\test_concurrent.py -q
• .\.venv\Scripts\python.exe -m ruff check src\strands\tools\executors\concurrent.py tests\strands\tools\executors\test_concurrent.py
• .\.venv\Scripts\python.exe -m ruff format --check src\strands\tools\executors\concurrent.py tests\strands\tools\executors\test_concurrent.py
• .\.venv\Scripts\python.exe -m mypy src\strands\tools\executors\concurrent.py
• .\.venv\Scripts\python.exe -m py_compile src\strands\tools\executors\concurrent.py tests\strands\tools\executors\test_concurrent.py
• git diff --check

---

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[github-actions]

Assessment: Approve

Clean, minimal fix that correctly addresses the ordering issue described in #2112. The approach of per-task result lists merged in tool_uses order is straightforward and maintains the existing contract.

Review Notes

• Correctness: The fix properly isolates results per-task and merges in deterministic order after all tasks complete. The error path (exception raised in the while loop) skips the merge, which produces empty tool_results on failure—consistent with the existing test expectation (test_concurrent_executor_reraises_exceptions asserts tool_results == []).
• Testing: The regression test is well-designed with a clear timing differential that validates ordering. Existing tests continue to pass with the new semantics.

Nice targeted fix! 👍

[opieter-aws]

@he-yufeng the implementation looks correct, thanks! Can you please rebase and fix failing CI?

[he-yufeng]

Rebased this branch onto current main and force-pushed 27dda6d.

Validation after rebase:

• uvx hatch test -n 0 --basetemp .tmp/pytest tests/strands/tools/executors/test_concurrent.py tests/strands/tools/test_caller.py::test_tool_caller_raises_reference_error_after_agent_collected
• uvx griffe check --search strands-py/src --format github strands --against upstream/main
• uvx ruff format --check strands-py/src/strands/tools/executors/concurrent.py strands-py/tests/strands/tools/executors/test_concurrent.py
• uvx ruff check strands-py/src/strands/tools/executors/concurrent.py strands-py/tests/strands/tools/executors/test_concurrent.py
• git diff --check

The earlier check-api warning was against the stale local main ref; it passes against current upstream/main after the rebase. Full format-check still reports an unrelated pre-existing formatting difference in tests/strands/models/test_gemini.py, so I did not include unrelated formatting churn in this PR.

[github-actions]

Assessment: Approve

Clean, minimal fix that correctly addresses the ordering issue described in #2112. The approach of per-task result lists merged in tool_uses order is straightforward and maintains the existing contract.

Review Notes

• Correctness: The fix properly isolates results per-task and merges in deterministic order after all tasks complete. The error path (exception raised in the while loop) skips the merge, which produces empty tool_results on failure—consistent with the existing test expectation (test_concurrent_executor_reraises_exceptions asserts tool_results == []).
• Testing: The regression test is well-designed with a clear timing differential that validates ordering. Existing tests continue to pass with the new semantics.

Nice targeted fix.

[chaynabors]

Hey @he-yufeng, sorry for the trouble, was this contribution made under the terms of Apache 2.0? Asking because we found another license in the codebase that has since been removed, and would otherwise have to revert this.