[FrankYang0529]

Why

• Return an approved agent answer as output_type after human review #73904 (in apache-airflow-providers-common-ai 0.11.0rc1) made AgentOperator convert an answer approved in HITL review with a pydantic TypeAdapter built from output_type. The following cases work on 0.10.0 but break on 0.11.0rc1:
  • A list of output types, such as [Verdict, Other], raises PydanticSchemaGenerationError, and the task fails.
  • An output marker, such as ToolOutput(Verdict), raises AttributeError, and the task fails. NativeOutput, PromptedOutput and TextOutput fail the same way.
  • For an output function, TypeAdapter validates the approved text as the function's arguments and calls the function a second time.

How

• An AgentOperator reviewer can approve, reject or ask for a new answer, but cannot edit the answer. So AgentOperator now keeps the output of the last run or regeneration and returns it when the reviewer approves, instead of converting the approved text. The operator no longer calls rehydrate_pydantic_output.
• When the approved answer is a Pydantic model whose class the worker would not register for XCom deserialization, the operator returns the model as a dict.

---

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

• Yes - Claude Code

[kaxil]

Thanks, the approved answer now comes back as the agent returned it, so [str, int] keeps "42" a string and the list, marker and function shapes no longer fail. LGTM.

One thing for a follow-up: the path without review (_complete_run, the if self._serialize_model_output and isinstance(output, BaseModel) branch) still dumps a model only when serialize_output is set. With output_type=[Summary, int] and review off, a 3.3+ worker pushes a Summary instance that the downstream pull rejects with the allow-list ImportError, while the review path now pushes a dict. That was already true before this PR. A small _to_xcom helper holding the condition you added here, called from both returns, would make the two paths agree.

Also minor: the tests that set run_hitl_review's return value to the agent's own output (test_execute_with_enable_hitl_review_delegates_to_run_hitl_review, test_execute_with_hitl_returns_string_unchanged, test_hitl_branch_success_also_deletes_the_budget_key) can no longer tell which value execute returns. One of them could return different text and assert the result is still the agent's, and test_execute_with_hitl_rehydrates_base_model still has the name and comment from the parse-back.

[kaxil]

Merged but can you handle those minor comments in a follow-up PR @FrankYang0529

[FrankYang0529]

@kaxil Thanks for the reivew. Create a follow-up PR #74729.