Repository navigation
Fix pulling AgentOperator output for list, marker and function types - #74729
Conversation
4635497 to
f1806a4
Compare
f5d341a to
edf86d0
Compare
edf86d0 to
ecdae8f
Compare
kaxil
left a comment
There was a problem hiding this comment.
LGTM, approving. A few nits you can fold in here or leave for later:
The first sentence of the changelog note reads like every model in every output becomes a dict. A plain output_type=Summary that the worker registers still pushes the instance, so I'd lead with the five output types like structured_output.rst does.
timedelta is missing from the HITL lists in structured_output.rst and the changelog note. Plain model_dump() kept it as a timedelta; JSON mode gives you "PT1M30S".
In test_execute_with_hitl_returns_the_model_instance the review text is the same model the agent returned, so the test can't tell which one came back. Having run_hitl_review return '{"text": "Edited"}' instead would catch that.
|
@FrankYang0529 Could you fix those so we can merge the PR, plz |
Signed-off-by: PoAn Yang <payang@apache.org>
Signed-off-by: PoAn Yang <payang@apache.org>
Signed-off-by: PoAn Yang <payang@apache.org>
Signed-off-by: PoAn Yang <payang@apache.org>
Signed-off-by: PoAn Yang <payang@apache.org>
ecdae8f to
8be607e
Compare
|
Thanks! I fixed all three nits:
|
Follow-up for #74537 (review).
_convert_output_for_xcom. Both returns of_complete_runnow call this method. Without review,AgentOperatornow also pushes a model as a dict when the worker would not register the model's class, for example withoutput_type=[Summary, int].test_execute_with_hitl_rehydrates_base_modeltotest_execute_with_hitl_returns_the_model_instanceand remove its outdated comment about the JSON string.Was generative AI tooling used to co-author this PR?