fix: save post-guardrail output to output_file instead of stale pre-guardrail result

[hztBUAA]

Summary

Fixes a bug where output_file always receives the raw pre-guardrail result instead of the final post-guardrail output when guardrails are present.

Root cause: In both _execute_core (sync) and _aexecute_core (async), the output_file save logic used stale local variables (json_output, pydantic_output, result) that were explicitly set to None/raw-string when guardrails are present (lines 687-688). After guardrails process and update task_output, the file save still referenced these stale locals, writing the original raw result instead of the guardrail-transformed output.

Fix: Replace local variable references with task_output attributes (task_output.json_dict, task_output.pydantic, task_output.raw) in both sync and async code paths. This ensures the file always reflects the final post-guardrail result, consistent with what task.output and the TaskCompletedEvent receive.

Changes

• lib/crewai/src/crewai/task.py: Use task_output.json_dict, task_output.pydantic, and task_output.raw for output_file save in both _execute_core and _aexecute_core
• lib/crewai/tests/test_task_guardrails.py: Add 3 tests verifying output_file content correctness with and without guardrails

Test plan

• test_output_file_saves_guardrail_result_not_stale_raw - verifies guardrail-transformed text is saved (single guardrail)
• test_output_file_saves_guardrail_result_with_guardrails_list - verifies with guardrails list
• test_output_file_saves_without_guardrails - verifies no regression without guardrails
• All 21 existing guardrail tests pass
• Existing output_file tests in test_task.py pass

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.