dsh-android-agent — Android phone testing plugin for DSH

中文 | English

A mobile-app testing plugin for DeepSeek Harness: it connects over WebSocket to an Android phone
running dsh-android-agent on the same LAN and
lets you open apps, tap at coordinates, input text, swipe, send key events, take screenshots and
dump the UI tree from the dsh-web「Phone」panel, with the same phone_* tools exposed to agents.
Built entirely on official NPM SDK packages — no dsh source changes.

Features

Security model

• All /api/dsh-phone/* routes are loopback-only (with same-origin checks) — the execution
  surface that drives a real phone is never exposed to the LAN.
• Tokens are stored in plaintext in ~/.dsh/dsh-phone.json (0700 dir / 0600 file, atomic writes) —
  same trust model as dsh-ssh.
• Agents can only use devices configured by the user in the GUI first.
• Screenshots/actions consume real phone resources and battery; confirm before operating.

Phone prerequisites

1. Install and start the Agent app's foreground service (enable accessibility / screen capture on the phone, same LAN).
2. Note the Agent's WebSocket endpoint ws://phone-ip:8080/ws and its token.
3. Add the device in the dsh-web「Phone」panel: alias, WebSocket URL, token.

Install

This plugin is a standalone package (not part of the @linxin666/dsh-web-ui-all aggregate), installed via a local link:

### npm (not yet published — available once published)
dsh plugin --profile web add dsh-android-agent

### local standalone package (development, at D:\code\dsh\dsh-android-agent)
cd D:\code\dsh\dsh-android-agent
pnpm install && pnpm build
dsh plugin --profile web add link:D:\code\dsh\dsh-android-agent

Windows users can install with one command (build + link, PowerShell):

powershell -ExecutionPolicy Bypass -File D:\code\dsh\dsh-android-agent\scripts\install.ps1

Full install & verification checklist (build → tests → install → GUI/device checks) lives in
docs/VERIFY.md.

Restart dsh web afterwards: the「Phone」entry appears in the sidebar and the agent prompt picks
up the plugin announcement automatically.

Data

• Device config: ~/.dsh/dsh-phone.json
• Agent screenshots: ~/.dsh/phone-screenshots/

Development

pnpm install
pnpm typecheck
pnpm build

Known limitations

• inputText relies on accessibility injection (ACTION_SET_TEXT / clipboard paste); some input
  fields may reject it — under root/Shizuku you can fall back to adb-style input via shell.
• Screenshots travel as base64 inside JSON-RPC; panel preview defaults to JPEG (PNG supported).
• Auto-reconnect can replay unfinished operations (non-idempotent) — mind side effects in long flows.
• The plugin only operates devices the user configured in the GUI.