fix: handle cancelled single function tools as tool failures

[elainegan-openai]

Summary

• handle CancelledError from a function tool via the normal tool-failure path, even when it is the only function tool in the turn
• preserve nested Agent.as_tool() fallback output when a recovered nested run ends without a final message
• add regression coverage for execution, non-streamed runner, streamed runner, and nested agent-tool paths

Why

Single-tool turns were handling CancelledError differently from turns with multiple function tools. That asymmetry could cancel the entire run instead of surfacing a normal tool error back to the model.

The same path also exposed a nested agent-tool regression: when a recovered nested run ended without a final message, Agent.as_tool() returned an empty string instead of the nested tool failure output.

Validation

• ruff check src/agents/agent.py src/agents/run_internal/tool_execution.py tests/test_example_workflows.py tests/test_run_step_execution.py tests/test_agent_runner.py tests/test_agent_runner_streamed.py
• PYTHONPATH=src pytest -q tests/test_example_workflows.py
• PYTHONPATH=src pytest -q tests/test_run_step_execution.py -k 'single_tool_call_uses_default_failure_error_function_for_cancelled_tool or multiple_tool_calls_use_default_failure_error_function_for_manual_cancelled_tool or test_execute_function_tool_calls_parent_cancellation_skips_post_invoke_work'
• PYTHONPATH=src pytest -q tests/test_agent_runner.py -k 'single_tool_call_with_cancelled_tool_reaches_final_output or parallel_tool_call_with_cancelled_sibling_reaches_final_output'
• PYTHONPATH=src pytest -q tests/test_agent_runner_streamed.py -k 'streamed_single_tool_call_with_cancelled_tool_reaches_final_output or streamed_parallel_tool_call_with_cancelled_sibling_reaches_final_output'

[elainegan-openai]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: fe9f06e436

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".