[vishal-bulbule]

Link to Issue or Description of Change

1. Link to an existing issue (if applicable):

• Closes: Workflow: LlmAgent with sub_agents/transfer loses ctx.output for downstream edges #7782

Problem:

A top-level LlmAgent node with sub_agents gets the single_turn mode from build_node. When it transfers, the transfer is handled twice:

1. run_llm_agent_as_node does not stop on transfer_to_agent in single_turn mode, so the function-call postprocessor runs the target inline inside the node.
2. NodeRunner also copies transfer_to_agent onto the node context, so DynamicNodeScheduler runs the target again as a separate hop. That hop runs in chat mode, sets no output, and its context is the one Workflow._handle_completion gets.

So for START -> lead(sub_agents) -> capture, capture receives None and every transfer calls the target model twice per turn. On resume (a tool confirmation or long-running tool inside the target) the workflow fast-forwards lead and resumes the target as the scheduler hop only, so the reply is lost there as well.

Solution:

Let the scheduler own the transfer in both the fresh and the resumed path:

• _llm_agent_wrapper.py: the single_turn branch stops after an event with transfer_to_agent, as the chat and task branches already do. The target no longer runs inline.
• _dynamic_node_scheduler.py: when the transfer chain starts at a single_turn node, each LlmAgent hop whose mode is None or chat runs as a single_turn copy, with rerun_on_resume=True unless the target set it (what build_node gives an LlmAgent node). Its reply goes through process_llm_agent_output, so the reply of the agent that ends the chain is the node output, and a hop paused on a tool call runs its model again on resume. A parent stamps chat on every LlmAgent sub-agent without a mode, so explicit and defaulted chat targets are both switched; task and single_turn sub-agents are tools rather than transfer targets and keep their mode.

Chains that start in chat or task mode are unchanged. I also tried giving these nodes chat mode in build_node instead, but a chat node waits for an output that never comes, so the downstream node never runs.

Testing Plan

Unit Tests:

• I have added or updated unit tests for my change.
• All unit tests pass locally.

New tests in tests/unittests/workflow/test_llm_agent_as_node.py (16 cases, most parametrized over is_resumable):

Test	main (1d81896)	this PR
test_single_turn_transfer_output_reaches_downstream (two turns, different targets, each target called once)	fail	pass
test_single_turn_transfer_terminal_node_replies_once	fail	pass
test_single_turn_transfer_back_to_router_reaches_downstream	fail	pass
test_single_turn_nested_transfer_reaches_downstream	fail	pass
test_single_turn_transfer_target_structured_output_reaches_downstream (output_schema plus output_key)	fail	pass
test_single_turn_transfer_target_resumes_after_confirmation	fail	pass
test_single_turn_transfer_target_resumes_after_long_running_tool	fail	pass
test_single_turn_transfer_explicit_chat_target_reaches_downstream	fail	pass
test_single_turn_transfer_keeps_task_mode_target	fail	pass
test_single_turn_transfer_respects_target_rerun_on_resume	fail	pass

tests/unittests/workflow, agents, flows, runners and test_runners.py: all pass (Python 3.11).
Full tests/unittests, each top-level directory in its own process (minimal venv, so tests that need optional extras fail the same way on main): 16,280 passed on this branch and 16,269 on main, and the failing set is identical except the new tests, which fail only on main.

Manual End-to-End (E2E) Tests:

The graph from #7782 with gemini-3.5-flash on Agent Platform, a before_model_callback counting model calls per agent:

Scenario	main	this PR
Two turns (translate, then summarize), not resumable	echo gets None both turns; lead 5, translate 2, summarize 7 calls	echo gets both replies; lead 2, translate 1, summarize 1 calls
Two turns, resumable	echo gets None both turns; lead 6, translate 9, summarize 3 calls	same as above
translate tool needs confirmation, resume, not resumable	echo gets None after the resume	echo gets the translation after the resume; translate 2 calls (before and after the pause)
Same, resumable	same as above	same as above

On main the call counts vary between runs because transfers bounce between agents; one earlier run on main ended with RecursionError in the postprocessor's inline transfer during the first turn.

Checklist

• I have read the CONTRIBUTING.md document.
• I have performed a self-review of my own code.
• I have commented my code, particularly in hard-to-understand areas.
• I have added tests that prove my fix is effective or that my feature works.
• New and existing unit tests pass locally with my changes.
• I have manually tested my changes end-to-end.
• Any dependent changes have been merged and published in downstream modules.

Additional context

The first version of this PR cleared the transfer after the inline run. That fixed fresh runs only: on resume the target runs as the scheduler hop, so the resume cases above still failed. This version removes the inline run for single_turn nodes instead, so there is one path.