adb-android-automation

该 Skill 使用 ADB 连接 Android 设备，由 Gemini 或 OpenAI 模型驱动自主导航应用界面，执行点餐、购票、购物、叫车等操作；结束后返回结构化结果、步骤和截图，敏感操作需用户授权。

mohd-mursaleen

热度 23.00

最近更新 2026-03-24 14:11:48

分类：自动化依赖：代码执行 / 读写文件 / 第三方 API / 内部调用大模型应用场景：自动化脚本 / 工作流 / 购物 / 旅行

安装技能

复制下面的话术发给你的 Agent，它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill： npx secure-skills-cli add android-automation

Metadata

name: iota-sidequest
description: Run autonomous Android automations over ADB. Use this for any task inside an Android app — ordering food, booking tickets, shopping, rides, or anything else. OpenClaw gathers details from the user, fires the automation, and reports back.
metadata: {"openclaw":{"emoji":"🌀","requires":{"bins":["adb"],"anyBins":["python3","python"]},"primaryEnv":"GEMINI_API_KEY","optionalEnv":["OPENAI_API_KEY"]}}

Iota Sidequest

Use this skill when the user wants to do something inside an Android app through ADB.

Setup

Before the first run: {baseDir}/scripts/setup_runtime.sh. Ensure a device is visible in adb devices.

API key — set one of:

Provider	Env var	Default model
Gemini (default)	`GEMINI_API_KEY`	`gemini-3.1-pro-preview`
OpenAI	`OPENAI_API_KEY`	`gpt-5.2`

If only OPENAI_API_KEY is set, the runtime auto-switches to OpenAI. To force a provider when both keys exist, set PHONE_AGENT_PROVIDER=openai (or gemini).

How It Works

Gather details from the user. Read {baseDir}/references/planning.md for what to ask per domain.
Run the automation in background with exit notification enabled:
```
"{baseDir}/scripts/run_phase.sh" "<task description>"
```
Wait for the process to exit. Do not poll or babysit.
Read the result JSON printed to stdout.
Act on the result (see table below).

The runtime is autonomous — it opens the app, navigates, taps, types, and scrolls on its own. It stops when the task is done, when it needs a user decision, or when it gets stuck.

Writing the Task Description

One natural-language string. Include: what app, what to do, what you know, what NOT to do yet.

First run example:

Open Swiggy Instamart. Search for SuperYou chocolate protein bar. Add 1 to cart. Go to checkout. Default address. Do not select payment or confirm order.

Continuation after user input:

Continue in Swiggy from checkout screen. SuperYou Protein Bar x1 in cart at ₹99. Select Google Pay. User authorized payment. Complete the order.

Reading the Result

Key fields in the result JSON:

Field	Meaning
`status`	`done`, `needs_input`, `stuck`, `max_steps`, `timeout`, or `failed`
`why`	Human-readable reason the automation stopped
`what_i_did`	Narrative of what the automation accomplished
`where_i_am`	Description of the current screen
`steps`	Step-by-step log of actions and observations
`screenshot_path`	Path to the final screenshot

What To Do After Each Result

Status	Action
`done`	Tell the user the result. Task complete.
`needs_input`	Show user: `where_i_am` + attach the screenshot image + ask the question from `why`.
`stuck`	Show user: `what_i_did` + `where_i_am` + attach the screenshot image. Ask what to do.
`max_steps`	Show user where it got to + attach screenshot. Ask if it should continue.
`timeout`	Same as `max_steps`.
`failed`	Tell user what went wrong. Ask before retrying.

When asking the user for input, always attach the screenshot image from screenshot_path so the user can see the current screen.

Continue automatically only when you have all needed information. Never wait for user to say "continue" or "now" between runs — if you have enough context, just fire the next run.

Safety

Never guess through: payment method, OTP, final purchase confirmation, account switching, cancellation, or refund. If the user already authorized a sensitive action, include that authorization in the task description.

Reference

Read {baseDir}/references/planning.md before complex tasks for domain-specific checklists on what details to gather.