Sub-agent with sequential LRO tools fails to resume

[tony-lenski]

🔴 Required Information

Describe the Bug:

When a sub-agent has multiple LRO tools that are called sequentially, the first HITL pause/resume works correctly, but the second resumption fails — the runner early-exits with no LLM content. Two independent bugs combine to cause this.

Steps to Reproduce:

1. Create a resumable app with an orchestrator LlmAgent that delegates to a sub-agent
2. The sub-agent has two LongRunningFunctionTools (e.g., a picker and a confirmation step)
3. User triggers the first LRO tool → agent pauses → user responds → agent resumes correctly
4. Agent calls the second LRO tool → agent pauses → user responds → agent fails to resume

See Minimal Reproduction Code below for a copy-paste-runnable script.

Expected Behavior:

Both resumptions should work identically — the sub-agent receives the tool response, the LLM is called, and it generates text or further tool calls.

Observed Behavior:

The first resumption works. The second resumption produces no LLM-generated content — the agent silently stops.

Environment Details:

• ADK Library Version: 1.30.0 (bug present in 1.27.0–1.30.0, worked in 1.26.0)
• Desktop OS: macOS
• Python Version: 3.13+

Model Information:

• Are you using LiteLLM: No
• Which model is being used: gemini-2.0-flash (not model-dependent — bug is in the runner/agent framework)

---

🟡 Optional Information

Regression:

Yes — worked in ADK 1.26.0, broken since 1.27.0 (the _resolve_invocation_id rewrite).

In 1.26.0, if no invocation_id was passed, the runner always created a new invocation — no stale flags to trip over. In 1.27.0, _resolve_invocation_id (runners.py:356-383) auto-infers invocation_id from the FunctionResponse by searching session.events for the matching FunctionCall. This forces the resumed-invocation path, which replays stale flags from the previous pause.

Additional Context:

Root cause

The first resumption works because the orchestrator's initial run exits via should_pause (line 494), which does not set end_of_agent. The problems appear on the second resumption, when the orchestrator enters the sub-agent resume block (line 474-483) for the first time.

Bug 1 — should_pause_invocation ignores existing responses

should_pause_invocation (invocation_context.py) returns True if any FunctionCall ID is in long_running_tool_ids — without checking whether a FunctionResponse already exists. On the second resumption, the already-answered first LRO still triggers the pause guard in base_llm_flow.py:838-851, and the sub-agent's LLM flow exits immediately.

Bug 2 — Premature end_of_agent on the orchestrator

After the first resumption, the orchestrator enters the sub-agent resume block (line 474-483) and unconditionally sets end_of_agent=True — even when the sub-agent only paused for the second LRO (not truly completed). On the second resumption, populate_invocation_agent_states replays this stale flag and the runner early-exits at runners.py:597.

Trigger conditions

All three are required:

1. ResumabilityConfig(is_resumable=True)
2. Root LlmAgent delegates to a sub-agent that has LRO tools
3. The sub-agent calls LRO tools sequentially (two or more pause/resume cycles in the same invocation)

Suggested fix

Bug 1 is addressed by PR #5072 (has_unresolved_long_running_tool_calls), open since 2026-03-30. The fix below is only for Bug 2, which #5072 does not cover.

Bug 2 — The orchestrator should only set end_of_agent=True when the sub-agent truly completed, not when it paused for another LRO:

Diff for llm_agent.py (sub-agent resume block, ~line 474)

   if agent_state is not None and (
       agent_to_transfer := self._get_subagent_to_resume(ctx)
   ):
+    sub_agent_paused = False
     async with Aclosing(agent_to_transfer.run_async(ctx)) as agen:
       async for event in agen:
+        # Requires Bug 1 fix — should_pause_invocation must be response-aware
+        if ctx.should_pause_invocation(event):
+          sub_agent_paused = True
         yield event

-    ctx.set_agent_state(self.name, end_of_agent=True)
-    yield self._create_agent_state_event(ctx)
+    if not sub_agent_paused:
+      ctx.set_agent_state(self.name, end_of_agent=True)
+      yield self._create_agent_state_event(ctx)
     return

Backward-compatible — non-LRO sub-agent completions still set the flag exactly as before. The check is skipped entirely when is_resumable is False or when the root agent calls LRO tools directly (no sub-agent delegation), so existing setups are unaffected.

Verified: with both #5072 (Bug 1) and this fix (Bug 2) applied to ADK 1.30.0, all sequential resumptions work correctly.

Minimal Reproduction Code:

import asyncio
from google.adk.agents import LlmAgent
from google.adk.apps import App, ResumabilityConfig
from google.adk.runners import Runner
from google.adk.sessions import InMemorySessionService
from google.adk.tools import LongRunningFunctionTool
from google.genai import types


async def select_item(tool_context) -> dict:
    """Step 1: user picks an item (blocks for user input)."""
    return {"status": "pending"}


async def confirm_choice(tool_context) -> dict:
    """Step 2: user confirms the choice (blocks for user input)."""
    return {"status": "pending"}


sub_agent = LlmAgent(
    name="picker",
    model="gemini-2.0-flash",
    instruction=(
        "First call select_item to let the user pick. "
        "After they respond, call confirm_choice to confirm. "
        "After both are done, summarize what was picked and confirmed."
    ),
    tools=[
        LongRunningFunctionTool(func=select_item),
        LongRunningFunctionTool(func=confirm_choice),
    ],
)
root_agent = LlmAgent(
    name="orchestrator",
    model="gemini-2.0-flash",
    instruction="Delegate to the picker agent.",
    sub_agents=[sub_agent],
)
app = App(
    name="repro",
    root_agent=root_agent,
    resumability_config=ResumabilityConfig(is_resumable=True),
)
session_service = InMemorySessionService()
runner = Runner(app=app, session_service=session_service)


async def main():
    session = await session_service.create_session(app_name="repro", user_id="u")

    # Step 1: agent delegates → picker calls select_item → pauses for user input
    step1_events = []
    async for event in runner.run_async(
        user_id="u",
        session_id=session.id,
        new_message=types.Content(
            role="user", parts=[types.Part(text="Pick something")]
        ),
    ):
        step1_events.append(event)

    fc1_id = next(
        fc.id
        for e in step1_events
        for fc in e.get_function_calls()
        if fc.name == "select_item" and e.long_running_tool_ids
    )

    # Step 2: user responds → picker resumes → calls confirm_choice → pauses again
    step2_events = []
    async for event in runner.run_async(
        user_id="u",
        session_id=session.id,
        new_message=types.Content(
            role="user",
            parts=[
                types.Part(
                    function_response=types.FunctionResponse(
                        id=fc1_id,
                        name="select_item",
                        response={"result": "option_a"},
                    )
                )
            ],
        ),
    ):
        step2_events.append(event)

    fc2_id = next(
        fc.id
        for e in step2_events
        for fc in e.get_function_calls()
        if fc.name == "confirm_choice" and e.long_running_tool_ids
    )

    # Step 3: user confirms → BUG: no LLM content, agent silently stops
    step3_events = []
    async for event in runner.run_async(
        user_id="u",
        session_id=session.id,
        new_message=types.Content(
            role="user",
            parts=[
                types.Part(
                    function_response=types.FunctionResponse(
                        id=fc2_id,
                        name="confirm_choice",
                        response={"confirmed": True},
                    )
                )
            ],
        ),
    ):
        step3_events.append(event)

    has_llm_content = any(
        e.get_function_calls()
        or (
            e.content
            and e.content.parts
            and any(p.text for p in e.content.parts)
        )
        for e in step3_events
        if e.author != "user"
    )
    assert has_llm_content, (
        f"BUG: Step 3 produced {len(step3_events)} events but no LLM content"
    )


asyncio.run(main())

How often has this issue occurred?:

• Always (100%) — deterministic with sequential LRO tools in a sub-agent

Related Issues:

• LongRunningFunctionTool resume fails: unresolved pause check + streaming ID mismatch #5064 — Covers Bug 1. PR fix(flows): resume long-running tools after matching responses #5072 fixes it with has_unresolved_long_running_tool_calls(). Verified that fix(flows): resume long-running tools after matching responses #5072 alone is not sufficient — Bug 2 must also be fixed (different code path at ~line 474, the sub-agent resume block, not the early-return at line 496 that fix(flows): resume long-running tools after matching responses #5072 modifies).
• Long-Running Tool FunctionResponse Routing Bypasses Custom Agent Orchestration #3986 — Related: FunctionResponse routing bypasses custom agent orchestration.