Fix output type loss in AgentOperator's human-in-the-loop review

[ColtenOuO]

Summary

AgentOperator's human-in-the-loop review path (enable_hitl_review=True) loses the original output type for any output_type that is neither str nor a Pydantic BaseModel (e.g. list[str], bool, dict).

The output is round-tripped through a string while it's shown to a human reviewer: HITLReviewMixin._to_string serializes it before review, and AgentOperator.execute() deserializes it back into output_type after approval. The serialization side used str(output), which produces a Python repr (e.g. "['tag-a', 'tag-b']"), not valid JSON.

The deserialization side then tried json.loads() on that repr, which always raised, and silently fell back to returning the raw repr string instead of the original list/bool/dict — a type change downstream tasks don't expect.

regenerate_with_feedback() (used when a reviewer requests changes) had the identical str(output) bug.

This is the same defect class already fixed for the require_approval=True path in #70075, which switched to TypeAdapter(...).dump_json(...). That fix wasn't applied to the separate enable_hitl_review=True path, so it regressed here.

Changes

• Both directions of the review round-trip now live next to each other in
  utils/output_type.py: the existing rehydrate_pydantic_output() and a new
  dump_output_to_json().
• dump_output_to_json() replaces the three copies of the serialization block
  that had drifted apart — HITLReviewMixin._to_string,
  AgentOperator.regenerate_with_feedback and LLMApprovalMixin.defer_for_approval.
  That drift is how the original bug survived Preserve output_type through human approval in LLM operators #70075.
• AgentOperator.execute()'s HITL-review branch reuses rehydrate_pydantic_output
  instead of a bespoke json.loads/except fallback, so both review paths behave
  consistently.

Guarding an output_type pydantic cannot build a schema for

AgentOperator passes output_type straight through to Agent(...), so pydantic-ai's
[A, B] multi-output lists, ToolOutput/NativeOutput/PromptedOutput markers and
output functions all reach these helpers. TypeAdapter raises
PydanticSchemaGenerationError (a RuntimeError, not a ValidationError) or
AttributeError on those, which neither side's except clause caught:

• rehydrate_pydantic_output now separates schema build from validation and falls
  back to plain json.loads. Without this an already-approved output is lost, after
  the model call and after the reviewer has signed off.
• dump_output_to_json falls back to str(output). This one fires before the session
  XCom push, so without it the task dies before the reviewer is ever shown a review
  session. It is reachable through output functions, which only warn
  ("Falling back to unconstrained schema") for a non-schema-able return type and then
  run fine.

Behaviour change

With the default output_type=str, the approved string is no longer parsed as JSON:

Approved string	Before	After
'42'	42	'42'
'{"text": "hi"}'	{'text': 'hi'}	'{"text": "hi"}'
'true'	True	'true'
'null'	None	'null'

This matches the non-HITL path (which returns result.output unparsed) and the
require_approval path. str is the default output_type and enable_hitl_review
has shipped since 0.1.0, so a Dag relying on the implicit parse needs an explicit
output_type (e.g. output_type=int) or its own json.loads() downstream.

---

Was generative AI tooling used to co-author this PR?

• Yes — Claude Code (Sonnet 5, Opus 5)

Generated-by: Claude Code (Opus 5) following the guidelines

[ColtenOuO]

Thanks for reviewing, I have addressed all of the comments.

If there still have anything need to improve or change, feel free to let me know ><

[ColtenOuO]

Thanks for catching this. I reproduced it myself and can confirm the output function re-invocation issue is real.

The latest commit addresses everything from this round:

• rehydrate_pydantic_output now unwraps ToolOutput/NativeOutput/PromptedOutput markers and gates on isinstance(x, type) instead of "did TypeAdapter raise" — an output function is never re-invoked during rehydrate.
• dump_output_to_json's fallback now logs a warning instead of degrading silently.
• Added tests covering marker unwrapping, the sequence-fallback case, and (most importantly) that an output function is never called during rehydrate.

Ready for another pass whenever you have time, thanks again for the thorough review!

[ColtenOuO]

Thanks @aaron-y-chen's review!

I've addressed the problem in lastest commit, ready for the next round review!