humanintheloop

Google ADK human-in-the-loop (durable approval)

A human-in-the-loop Google ADK agent running durably on Temporal with the googleadk contrib integration. The agent has a sensitive delete_resource function tool that must not run without a human's approval.

How it works:

1. The model calls delete_resource. On its first invocation the tool sees ctx.ToolConfirmation() == nil, calls ctx.RequestConfirmation("Delete …?", nil), and returns without doing the delete — so ADK pauses the agent.
2. ApprovalWorkflow detects the pause via googleadk.PendingConfirmations and durably blocks on a Temporal signal (googleadk.ConfirmationSignalName, read with workflow.GetSignalChannel) carrying a googleadk.ConfirmationDecision.
3. When the decision arrives, the workflow resumes the run with googleadk.ConfirmationResponse(decision). ADK re-dispatches the original tool call, which now sees a confirmation and performs the delete (or is blocked if denied).

This is the differentiator: the wait for the human is durable. The workflow can sit blocked for minutes or days and survive worker restarts — no state is lost. When the approval signal finally arrives, the agent resumes exactly where it paused.

Notes

• delete_resource runs in-workflow and only simulates the delete (it returns a status map) to keep the demo deterministic. A real destructive operation does I/O and must not run in the workflow — expose it with googleadk.ActivityAsTool so it runs worker-side under Temporal's retry/timeout policy. The confirmation gate is identical either way: the tool still calls ctx.RequestConfirmation(...) before doing the work.
• The workflow handles one pending confirmation per resume pass. This is also the recommended pattern: as the googleadk.ConfirmationResponse docs note, resuming several decisions in one pass can re-dispatch the approved tool calls in an order that is not replay-stable, so when the confirmed tools dispatch Activities, answer one decision per Run pass (additional pending confirmations surface again on the next pass).

Prerequisites

• A running Temporal server (e.g. temporal server start-dev).
• A Gemini API key from https://aistudio.google.com/apikey, exported worker-side.

Steps to run this sample

1. Start a Temporal server (see prerequisites).

2. In a second terminal, start the worker:

export GEMINI_API_KEY=...
go run googleadk/humanintheloop/worker/main.go

3. In a third terminal, run the starter:

go run googleadk/humanintheloop/starter/main.go

The starter asks the agent to delete a resource; the workflow pauses awaiting approval, and the starter then sends an approval signal (via client.SignalWorkflow) to demonstrate the resume. In a real system the signal would come from an operator clicking "approve" in a UI, possibly much later.

Test without a live LLM

workflow_test.go scripts the model to call delete_resource, uses env.RegisterDelayedCallback to deliver the approval through the real Temporal signal, and asserts the delete completes only after approval (plus a denial case). No API key or network needed:

go test ./googleadk/humanintheloop/...