fix(task): parse pydantic output before guardrail invocation

[giulio-leone]

Problem

When a task has output_pydantic and a guardrail configured, TaskOutput.pydantic is None on the first guardrail invocation but correctly parsed on retry attempts. This makes it impossible to write guardrail functions that inspect structured Pydantic output.

Root Cause

In both execute_task() (async) and execute_sync(), when the result is a string and guardrails are present, _export_output() was intentionally skipped:

elif not self._guardrails and not self._guardrail:
    raw = result
    pydantic_output, json_output = self._export_output(result)
else:
    raw = result
    pydantic_output, json_output = None, None  # <-- BUG: always None with guardrails

On retries (line ~1155), _export_output() was always called, creating inconsistent behavior.

Fix

Remove the conditional skip so _export_output() is always called regardless of guardrail presence:

else:
    raw = result
    pydantic_output, json_output = self._export_output(result)

Applied to both sync and async execution paths.

Test

Added test_guardrail_pydantic_output_available_on_first_attempt that verifies:

• Guardrail receives non-None pydantic on the first (and only) invocation
• The parsed Pydantic model matches the expected output

All 20 existing guardrail tests pass (2 pre-existing failures unrelated to this change).

Fixes #4369

---

Note

Medium Risk
Touches core task execution output handling (sync and async), which could subtly change when/what structured parsing occurs, but failures are swallowed with a debug log to preserve raw output behavior.

Overview
Guardrails now receive structured outputs on the first pass. In Task sync + async execution paths, string results are now always run through _export_output() before any guardrail invocation, so TaskOutput.pydantic/json_dict are populated consistently (instead of only after a retry).

Parsing failures are handled defensively (debug log + continue with raw output), and tests were updated/expanded with a regression case (test_guardrail_pydantic_output_available_on_first_attempt) to lock in the new behavior. A tiny formatting-only change was made in crew_agent_executor.py.

^{Written by Cursor Bugbot for commit 47680d2. This will update automatically on new commits. Configure here.}

[giulio-leone]

Friendly ping — CI is green and this is ready for review. Happy to address any feedback. Thanks!

[Copilot]

Pull request overview

Fixes an inconsistency in Task execution where structured outputs (output_pydantic / output_json) were not exported before guardrails ran on the first attempt, making TaskOutput.pydantic/json_dict unavailable to guardrails until retries.

Changes:

• Update both async (_aexecute_core) and sync (_execute_core) paths to call _export_output() even when guardrails are configured.
• Add a regression test ensuring TaskOutput.pydantic is populated for guardrails on the first invocation.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

File	Description
lib/crewai/src/crewai/task.py	Always exports structured output before invoking guardrails (sync + async).
lib/crewai/tests/test_task_guardrails.py	Adds regression test asserting guardrails receive non-None Pydantic output on first attempt.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

[cursor]

Broad try/except silences errors for non-guardrail tasks

Medium Severity

The old code called _export_output() without try/except for tasks without guardrails, allowing errors like ValidationError to propagate to the caller. The new code merges all non-BaseModel results into one else branch with a blanket except Exception, so tasks without guardrails now silently swallow _export_output() failures and produce pydantic=None / json_dict=None instead of raising. This masks legitimate conversion errors (e.g., ValidationError, TypeError from missing agent) that previously surfaced to users. The try/except guard is only needed when guardrails are present.

Additional Locations (1)

• lib/crewai/src/crewai/task.py#L718-L722