fix: use async LLM calls in converter for async workflows (issue #5230)

[devin-ai-integration]

Summary

When using akickoff() (async workflow), the converter module was called synchronously via self._export_output() inside _aexecute_core and _ainvoke_guardrail_function, which calls llm.call() and blocks the event loop.

This PR adds async counterparts throughout the converter pipeline:

• Converter.ato_pydantic() / Converter.ato_json() — use llm.acall() instead of llm.call()
• aconvert_to_model(), ahandle_partial_json(), aconvert_with_instructions() — async versions of the module-level utility functions
• Task._aexport_output() — async version of _export_output wired into _aexecute_core (line 605) and both branches of _ainvoke_guardrail_function (lines 1286, 1331)

Sync methods are untouched; backward compatibility is preserved.

Review & Testing Checklist for Human

• ato_json still calls sync _create_instructor().to_json() when supports_function_calling() is True — this means the function-calling JSON path still blocks the event loop. The instructor library doesn't expose an async interface here. Verify this is acceptable or needs a run_in_executor wrapper.
• Verify all three _aexport_output call sites in task.py are correct (lines ~605, ~1286, ~1331). Confirm no other async paths still call the sync _export_output.
• ato_pydantic non-function-calling fallback calls sync handle_partial_json on ValidationError — this is safe (no LLM call, just regex/JSON parsing) but worth confirming.
• Run an end-to-end async workflow with output_pydantic or output_json set on a task, using crew.akickoff(), and confirm the event loop is no longer blocked during output conversion.

Notes

• All new async utility functions mirror the logic of their sync counterparts. The duplication is intentional to avoid changing the sync code path.
• 20 new async tests added covering: ato_pydantic/ato_json happy paths, retry logic, max-attempt errors, all three async utility functions, partial JSON handling, and explicit assertions that llm.call() is never invoked from the async path.

Link to Devin session: https://app.devin.ai/sessions/fec9a6f7345348bd99ecbc42d20e7876

---

Note

Medium Risk
Updates async task execution to use new async conversion helpers, which changes how structured outputs are parsed and may affect task results/guardrail retries in async workflows. Risk is contained to async paths; sync behavior is preserved.

Overview
Fixes async workflows (Task._aexecute_core / _ainvoke_guardrail_function) so structured output export no longer calls the synchronous converter and blocks the event loop.

Adds an async converter pipeline in utilities/converter.py (Converter.ato_pydantic/ato_json plus aconvert_to_model/ahandle_partial_json/aconvert_with_instructions) and wires it via a new Task._aexport_output, with expanded async test coverage ensuring llm.call() is never used on the async path.

^{Written by Cursor Bugbot for commit 415b894. This will update automatically on new commits. Configure here.}

[devin-ai-integration]

Prompt hidden (unlisted session)

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment and CI monitoring

[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.}