[Python][Live] transfer_to_agent is ignored when its function response is not content.parts[0]

[haydeem2]

Bug Description:

In Live mode (BaseLlmFlow.run_live()), agent transfer is gated on whether
transfer_to_agent is the first function response in event.content.parts[0].

When the model issues parallel function calls and a non-transfer tool's
response is merged first, ADK:

1. Executes both tools successfully
2. Merges both function responses into one event
3. Correctly sets event.actions.transfer_to_agent on the merged event
4. Does not start the child agent, because the transfer gate only inspects
   parts[0]

Reversing the merged response order (transfer first) makes transfer work. A lone
transfer_to_agent call also works. This is an order-dependent Live transfer
bug, not a failure to execute or merge parallel tools.

The non-Live async path uses function_response_event.actions.transfer_to_agent
directly and does not exhibit this parts[0] dependency.

Steps to Reproduce:

1. Install ADK 1.34.0 and download the reproducer (see Minimal Reproduction Code
   below).
2. Run python adk_live_transfer_order_repro.py.
3. Observe Layer 1: merged event has parts[0]=set_state,
   parts[1]=transfer_to_agent, and actions.transfer_to_agent set — but the
   stock predicate misses the transfer.
4. Observe Layer 2: with [set_state, transfer_to_agent] parallel order, the
   child agent does not start (XFAIL on stock ADK 1.34.0).
5. Re-run mentally / via script with reversed order [transfer_to_agent, set_state] — child starts (PASS).

Expected Behavior:

Transfer should be order-independent: if the merged event contains a
transfer_to_agent function response and event.actions.transfer_to_agent
names a target, the child agent should start exactly once.

Merged response order	Expected
[set_state, transfer_to_agent]	Child starts once
[transfer_to_agent, set_state]	Child starts once
lone transfer_to_agent	Child starts once
parallel non-transfer tools	No child starts

Observed Behavior:

Merged response order	Stock ADK 1.34.0
[set_state, transfer_to_agent]	Child does not start
[transfer_to_agent, set_state]	Child starts
lone transfer_to_agent	Child starts

Example output from the reproducer on stock google-adk==1.34.0:

Layer 1: 15 passed
Layer 2: 7 passed, 1 xfail, 1 skipped
  XFAIL [set_state, transfer] child starts in run_live  (known stock ADK bug)
RESULT: OK on stock ADK — bug reproduced

Root cause in BaseLlmFlow.run_live() (v1.34.0,
base_llm_flow.py#L578-L614):

if (
    event.content
    and event.content.parts
    and event.content.parts[0].function_response
    and event.content.parts[0].function_response.name == "transfer_to_agent"
):
    transfer_to_agent = event.actions.transfer_to_agent
    if transfer_to_agent:
        ...  # start child run_live()

The same parts[0] predicate remains in v2.6.0
(base_llm_flow.py#L681-L717).

Environment Details:

• ADK Library Version (pip show google-adk): 1.34.0 (also present in 2.6.0
  and current main as of 2026-07-31)
• Desktop OS: macOS (reproducer also runs on Linux; not Windows-tested)
• Python Version (python -V): 3.12.8 (reproducer requires 3.11+)

Model Information:

• Are you using LiteLLM: No
• Which model is being used: gemini-live-2.5-flash-native-audio (Gemini 2.5
  Live native audio) in production via direct ADK Live + Vertex AI Agent Engine.
  The attached reproducer uses a MockModel (no GCP / API key) to force
  parallel response order deterministically.

---

Regression:

Did this work in a previous version of ADK? If so, which one?

Not confirmed on older ADK versions for this specific parts[0] predicate. The
issue is present on 1.34.0, 2.6.0, and main (checked 2026-07-31).
Related Live transfer fixes (#5238,
#5536) addressed different
mechanisms (history attribution and resumption-handle inheritance) and do not
remove the parts[0] check.

Logs:

Please attach relevant logs. Wrap them in code blocks (```) or attach a
text file.

ADK Live transfer_to_agent response-order reproducer
  google-adk: 1.34.0
  Python:     3.12.8

--- Layer 1: event-level probe (merge + predicates) ---
  PASS  [set_state, transfer] parts[0] is set_state
  PASS  [set_state, transfer] parts[1] is transfer_to_agent
  PASS  [set_state, transfer] action names specialist
  PASS  [set_state, transfer] stock predicate misses transfer
  PASS  [set_state, transfer] patched predicate detects transfer
  → 15 passed, 0 failed, 0 xfail, 0 skipped

--- Layer 2: run_live() regression (MockModel) ---
  XFAIL [set_state, transfer] child starts in run_live  (known stock ADK bug)
  PASS  [transfer, set_state] child starts
  PASS  lone transfer: child starts
  → 7 passed, 0 failed, 1 xfail, 1 skipped

RESULT: OK on stock ADK — bug reproduced

Screenshots / Video:

N/A

Additional Context:

Parallel Live tool execution merges responses correctly in
merge_parallel_function_response_events()
— all parts are appended and actions (including transfer_to_agent) are merged.
The bug is solely in the Live transfer gate inside run_live().

Proposed fix (scan all function responses and require the transfer action):

has_transfer_response = any(
    response.name == "transfer_to_agent"
    for response in event.get_function_responses()
)
if has_transfer_response and event.actions.transfer_to_agent:
    ...  # existing child run_live() logic unchanged

Useful references:

• ADK Live event guide
• ADK tool performance guide (parallel tools)
• ADK events guide (transfer_to_agent action)
• Live transfer lifecycle PR #2588
• Related action-merge issue #1938

Full repo (optional): https://github.com/haydeem2/adk-live-transfer-order-repro

Minimal Reproduction Code:

Gist: https://gist.github.com/haydeem2/4608a4a4e306a20ae9b5dfaa54338a34

python3 -m venv .venv
source .venv/bin/activate
pip install "google-adk==1.34.0"
curl -fsSL -o adk_live_transfer_order_repro.py \
  "https://gist.githubusercontent.com/haydeem2/4608a4a4e306a20ae9b5dfaa54338a34/raw/adk_live_transfer_order_repro.py"
python adk_live_transfer_order_repro.py

Minimal merged-event proof (Layer 1 only — no run_live()):

from google.adk.events.event import Event
from google.adk.events.event_actions import EventActions
from google.adk.flows.llm_flows.functions import merge_parallel_function_response_events
from google.genai import types

def response_event(name, *, transfer_target=None):
    actions = EventActions()
    if transfer_target:
        actions.transfer_to_agent = transfer_target
    return Event(
        author="coordinator_agent",
        content=types.Content(
            role="user",
            parts=[types.Part(function_response=types.FunctionResponse(
                id=f"{name}-id", name=name, response={"result": "ok"},
            ))],
        ),
        actions=actions,
    )

merged = merge_parallel_function_response_events([
    response_event("set_state"),
    response_event("transfer_to_agent", transfer_target="specialist_agent"),
])

assert merged.content.parts[0].function_response.name == "set_state"
assert merged.content.parts[1].function_response.name == "transfer_to_agent"
assert merged.actions.transfer_to_agent == "specialist_agent"
# Stock run_live() gate checks parts[0] only → transfer missed

How often has this issue occurred?:

• Rare

In production the symptom is rare: the model must batch transfer_to_agent with
another tool and have the non-transfer response land in parts[0]. Prompting
for a lone transfer reduces frequency further. When those conditions align, the
bug is deterministic (100% in the attached reproducer).

[kinKingen]

Hi @haydeem2 — I reproduced the order-dependent transfer gate on current main at f4e72334 with an offline event-level probe.

The merged event contains:

• function responses: ['set_state', 'transfer_to_agent']
• event.actions.transfer_to_agent == 'specialist_agent'
• current parts[0] predicate: False
• scanning event.get_function_responses(): True

I also searched for an implementation linked to this issue and did not find a competing PR.

If maintainers agree, I’d like to work on a focused fix:

• gate Live transfer on both an existing transfer_to_agent function response and event.actions.transfer_to_agent;
• keep the existing connection close, resumption-handle reset, and child run_live() behavior unchanged;
• add regression coverage for [set_state, transfer], [transfer, set_state], a lone transfer, and parallel non-transfer responses, asserting the child starts exactly once or not at all as appropriate;
• validate with focused pytest, the required tox suite, and a mock Runner/Live E2E log.

Per the contribution guide, I’ll wait for confirmation or assignment before opening a PR.

AI assistance disclosure: I used OpenAI Codex for repository navigation, reproduction support, and drafting this proposal. I reviewed the code path, reproduction output, and proposed scope.