Interactions API in single_turn sub-agent result dropped as orphan ("Dropping function responses with no matching function call")

[Dumeng]

🔴 Required Information

Please ensure all items in this section are completed to allow for efficient
triaging. Requests without complete information may be rejected / deprioritized.
If an item is not applicable to you - please mark it as N/A

Describe the Bug:
With Gemini(use_interactions_api=True), when a chat agent delegates to a mode="single_turn" sub-agent, the parent's next request drops the delegation's function_response.

ADK logs:

WARNING google_adk...tools._rearranger: Dropping function responses with no matching function call: ['call_root_1']

The Interactions delta then carries only the sub-agent's quoted final text. The server-side interaction is left with an unanswered function_call, and anything an after_tool_callback attached to the result is lost.

Steps to Reproduce:

import asyncio
from collections.abc import AsyncGenerator

from google.adk.agents import Agent
from google.adk.models.google_llm import Gemini
from google.adk.models.llm_request import LlmRequest
from google.adk.models.llm_response import LlmResponse
from google.adk.runners import InMemoryRunner
from google.genai import types
from pydantic import PrivateAttr


class Scripted(Gemini):
    use_interactions_api: bool = True
    _script: list[LlmResponse] = PrivateAttr(default_factory=list)
    _requests: list[LlmRequest] = PrivateAttr(default_factory=list)

    async def generate_content_async(
        self, llm_request: LlmRequest, stream: bool = False
    ) -> AsyncGenerator[LlmResponse, None]:
        self._requests.append(llm_request.model_copy(deep=True))
        yield self._script.pop(0)


def turn(iid: str, part: types.Part) -> LlmResponse:
    return LlmResponse(content=types.Content(role="model", parts=[part]), interaction_id=iid)


async def main() -> None:
    child_model = Scripted(model="gemini-3-flash")
    child_model._script = [turn("c1", types.Part(text="child answer"))]
    child = Agent(name="child", description="d", mode="single_turn", model=child_model)

    root_model = Scripted(model="gemini-3-flash")
    root_model._script = [
        turn("r1", types.Part(function_call=types.FunctionCall(
            id="call_root_1", name="child", args={"request": "do it"}))),
        turn("r2", types.Part(text="done")),
    ]
    root = Agent(name="root", model=root_model, sub_agents=[child])

    runner = InMemoryRunner(agent=root, app_name="repro")
    session = await runner.session_service.create_session(app_name="repro", user_id="u")
    async for _ in runner.run_async(
        user_id="u", session_id=session.id,
        new_message=types.Content(role="user", parts=[types.Part(text="hi")]),
    ):
        pass

    # Expected: [..., function_call(call_root_1), function_response(call_root_1)]
    # Actual:   [user: "For context: ... [child] said: child answer"]
    for c in root_model._requests[1].contents:
        print(c.role, [p.function_response or p.function_call or p.text for p in c.parts])


asyncio.run(main())

Environment Details:

• ADK Library Version (pip show google-adk): 2.10.0
• Desktop OS: Linux
• Python Version (python -V): 3.14.5

Model Information:

• Are you using LiteLLM: No
• Which model is being used: gemini-3.8-flash

---

🟡 Optional Information

Providing this information greatly speeds up the resolution process.

Regression:
N/A

Logs:
See above

Screenshots / Video:
N/A

Additional Context:
Add any other context about the problem here.

Minimal Reproduction Code:
See above

How often has this issue occurred?:

Always (100%)

Potential root cause

1. _SingleTurnAgentTool.run_async runs the child with tool_context.run_node(..., override_branch="<agent>@<fc_id>") but no override_isolation_scope. So the child's events (including its synthetic author="user" input event) are unscoped and visible to the parent. They land in session.events between the parent's function_call and its function_response.
2. On the Interactions path, _get_current_turn_contents scans backward for the turn start. It only keeps scanning past a function_response whose author is user (_is_submitted_tool_result). The parent's result is authored by the parent agent, so the scan stops at the child's last event.
3. The parent's function_call falls outside the slice, so drop_orphaned_function_responses removes the result.

Task mode doesn't hit this, because _dispatch_task_fc passes override_isolation_scope=fc.id and _synthesize_task_fr_event authors the result as user.

Suggested fix (either)

• In _get_current_turn_contents, add every event's function_response ids to unmatched_response_ids, not only user-submitted ones. That's what the invariant in the comment above the loop requires ("the slice must reach back far enough to include the call it answers"). Verified locally: the slice becomes [user, function_call, function_response] and the Interactions delta is exactly the function_response.
• Or run _SingleTurnAgentTool children under override_isolation_scope=fc_id, as task mode does.

[charan-rathore]

I'd like to take this. I plan to keep agent-authored function responses paired with their calls in the current-turn context slice, with an offline regression test for the single-turn sub-agent case. Would a focused PR be welcome?

[surajksharma07]

The trace made this easy to follow @Dumeng. Reproduced on 2.10.0 and the second request ends up with only the quoted "[child] said" text, no function_response.

Going with your first option locally (tracking every function_response id in _get_current_turn_contents, not just user-authored ones, and keeping the anchor check as it is) gave me [user, function_call, function_response] and the flows/agents/models unit tests still pass apart from a few unrelated Anthropic failures I get without the change too.

Could you try that on your real setup ideally with an after_tool_callback that modifies the result and check the orphan warning is gone? Also worth checking include_contents='none'.

@charan-rathore happy for you to pick up the PR once that's confirmed. Please add the offline regression test and run the full suite before raising it.