Bug: `transfer_to_agent` not found in `canonical_tools()` during HITL `request_confirmation` resume

[settler-av]

Using tool_context.request_confirmation() inside a before_tool_callback to gate transfer_to_agent with human approval fails on the confirmation resume with:

ValueError: Tool 'transfer_to_agent' not found. Available tools:

This is a structural incompatibility between two ADK subsystems:

• _AgentTransferLlmRequestProcessor injects transfer_to_agent dynamically into llm_request.tools_dict at flow time. It is never added to agent.tools.
• On confirmation resume, _RequestConfirmationLlmRequestProcessor (Step 4) rebuilds tools_dict exclusively from await agent.canonical_tools(), which only reads from agent.tools.
• Since transfer_to_agent was never in agent.tools, tools_dict = {} → _get_tool("transfer_to_agent", {}) → ValueError.

Steps to Reproduce:

1. Create an LlmAgent orchestrator with sub_agents (e.g. RemoteA2aAgent or plain LlmAgent)
2. Add a before_tool_callback that calls tool_context.request_confirmation() when tool.name == "transfer_to_agent"
3. Send a message that causes the LLM to call transfer_to_agent
4. Respond to the adk_request_confirmation event with ToolConfirmation(confirmed=True)
5. Error is raised

Script that reproduce the issue

import asyncio
from google.adk.agents import LlmAgent
from google.adk.runners import Runner
from google.adk.sessions import InMemorySessionService
from google.adk.tools.tool_context import ToolContext
from google.adk.tools.tool_confirmation import ToolConfirmation
from google.adk.tools.base_tool import BaseTool
from google.genai import types
import os
from dotenv import load_dotenv
from google.adk.models.lite_llm import LiteLlm

load_dotenv()

# ── HITL callback ─────────────────────────────────────────────────────────────

async def human_approval_callback(tool: BaseTool, args: dict, tool_context: ToolContext):
    if tool.name != "transfer_to_agent":
        return None

    if tool_context.tool_confirmation is None:          # first call → suspend
        tool_context.request_confirmation(hint=f"Approve transfer to '{args.get('agent_name')}'?")
        return {"status": "pending"}

    if tool_context.tool_confirmation.confirmed:        # resume → approved
        return None
    return {"status": "denied"}                         # resume → denied


# ── Agents ────────────────────────────────────────────────────────────────────

math_agent = LlmAgent(
    name="math_assistant",
    model=LiteLlm(
        model=f"azure/{os.getenv('AZURE_MODEL_NAME')}",
        api_key=os.getenv("AZURE_API_KEY"),
        api_base=os.getenv("AZURE_API_BASE"),
        api_version=os.getenv("AZURE_API_VERSION"),
    ),
    description="Specialist for math questions.",
    instruction="You are a math expert. Answer math questions concisely.",
)

root_agent = LlmAgent(
    name="orchestrator",
    model=LiteLlm(
        model=f"azure/{os.getenv('AZURE_MODEL_NAME')}",
        api_key=os.getenv("AZURE_API_KEY"),
        api_base=os.getenv("AZURE_API_BASE"),
        api_version=os.getenv("AZURE_API_VERSION"),
    ),
    description="Routes questions to specialists.",
    instruction="If the user asks a math question, transfer to math_assistant immediately.",
    sub_agents=[math_agent],
    before_tool_callback=human_approval_callback,
)


# ── Runner ────────────────────────────────────────────────────────────────────

APP, USER, SESSION = "repro", "user1", "s1"
session_service = InMemorySessionService()
runner = Runner(agent=root_agent, app_name=APP, session_service=session_service)


async def main():
    await session_service.create_session(app_name=APP, user_id=USER, session_id=SESSION)

    # Turn 1 — trigger a transfer
    print("=== Turn 1: user asks a math question ===")
    confirmation_id = None
    async for event in runner.run_async(
        user_id=USER, session_id=SESSION,
        new_message=types.Content(role="user", parts=[types.Part(text="What is 12 * 12?")]),
    ):
        if event.content:
            for part in event.content.parts or []:
                fc = getattr(part, "function_call", None)
                if fc and fc.name == "adk_request_confirmation":
                    confirmation_id = fc.id
                    print(f"  → Approval requested (id={fc.id})")

    if not confirmation_id:
        print("No transfer triggered — try a clearer math prompt.")
        return

    # Turn 2 — approve the transfer
    # BUG: this raises ValueError: Tool 'transfer_to_agent' not found. Available tools:
    print("\n=== Turn 2: human approves the transfer ===")
    approval = ToolConfirmation(confirmed=True)
    async for event in runner.run_async(
        user_id=USER, session_id=SESSION,
        new_message=types.Content(role="user", parts=[types.Part(
            function_response=types.FunctionResponse(
                name="adk_request_confirmation",
                id=confirmation_id,
                response={"response": approval.model_dump_json()},
            )
        )]),
    ):
        if event.is_final_response() and event.content:
            for part in event.content.parts or []:
                if part.text:
                    print(f"  Final: {part.text}")


if __name__ == "__main__":
    asyncio.run(main())

Expected Behavior:

After human approval, the confirmation processor re-executes transfer_to_agent and the transfer to the sub-agent completes normally.

Observed Behavior:

ValueError: Tool 'transfer_to_agent' not found. Available tools: 

  File "google/adk/flows/llm_flows/functions.py", in _get_tool
    raise ValueError(error_msg)
  File "google/adk/flows/llm_flows/request_confirmation.py", in run_async
    {
        tool.name: tool
        for tool in await agent.canonical_tools(   # ← transfer_to_agent never here
            ReadonlyContext(invocation_context)
        )
    },

Environment Details:

• ADK Library Version: 2.0.0b1
• Desktop OS: Linux
• Python Version: 3.10+

Model Information:

• Are you using LiteLLM: Yes
• Which model is being used: gpt-4.1

---

🟡 Optional Information

Regression: Not a regression — pre-existing architectural gap. request_confirmation() was designed for regular tools in agent.tools; transfer_to_agent is a flow-level tool and never goes through that path.

Root Cause (potential):

agent_transfer.py ~line 51:

transfer_to_agent_tool = TransferToAgentTool(agent_names=[...])
await transfer_to_agent_tool.process_llm_request(...)  # adds to llm_request.tools_dict ONLY
# never added to agent.tools → invisible to canonical_tools()

request_confirmation.py ~line 168:

{
    tool.name: tool
    for tool in await agent.canonical_tools(   # only reads agent.tools
        ReadonlyContext(invocation_context)
    )
}
# tools_dict = {} → Tool 'transfer_to_agent' not found

How often has this issue occurred?: Always (100%)

[surajksharma07]

@settler-av Step 4 in request_confirmation.py rebuilds tools_dict exclusively from canonical_tools(), so transfer_to_agent (only registered dynamically into llm_request.tools_dict by agent_transfer.py:51-66) is always invisible at confirmation-resume time regardless of how sub_agents is configured.

A workaround worth trying: right after the canonical_tools() comprehension on line 174, also call _get_transfer_targets(agent) and inject a TransferToAgentTool(agent_names=[...]) when targets exist — this mirrors exactly what the agent-transfer preprocessor does and should close the gap without touching the broader API.

Could you apply that locally and verify it carries the full approve → transfer and reject → block cycle cleanly without the ValueError?

[settler-av]

@surajksharma07 Given solution worked for the script that I shared.

Let me know if I can raise a PR implementing stated solution for the same.