agent-device（Agent 设备）

Mobile Automation with agent-device

For agent-driven exploration: use refs. For deterministic replay scripts: use selectors.

Quick start

agent-device open Settings --platform ios
agent-device snapshot -i
agent-device press @e3
agent-device wait text "Camera"
agent-device alert wait 10000
agent-device diff snapshot -i
agent-device fill @e5 "test"
agent-device close

If not installed, run:

npx -y agent-device

Core workflow

1. Open app or deep link: open [app|url] [url] (open handles target selection + boot/activation in the normal flow)
2. Snapshot: snapshot to get refs from accessibility tree
3. Interact using refs (press @ref, fill @ref "text"; click is an alias of press)
4. Re-snapshot after navigation/UI changes
5. Close session when done

Commands

Navigation

agent-device boot                 # Ensure target is booted/ready without opening app
agent-device boot --platform ios  # Boot iOS target
agent-device boot --platform android # Boot Android emulator/device target
agent-device open [app|url] [url] # Boot device/simulator; optionally launch app or deep link URL
agent-device open [app] --relaunch # Terminate app process first, then launch (fresh runtime)
agent-device open [app] --activity com.example/.MainActivity # Android: open specific activity (app targets only)
agent-device open "myapp://home" --platform android          # Android deep link
agent-device open "https://example.com" --platform ios       # iOS deep link (opens in browser)
agent-device open MyApp "myapp://screen/to" --platform ios   # iOS deep link in app context
agent-device close [app]          # Close app or just end session
agent-device reinstall <app> <path> # Uninstall + install app in one command
agent-device session list         # List active sessions

boot requires either an active session or an explicit selector (--platform, --device, --udid, or --serial). boot is a fallback, not a regular step: use it when starting a new session only if open cannot find/connect to an available target.

Snapshot (page analysis)

agent-device snapshot                  # Full XCTest accessibility tree snapshot
agent-device snapshot -i               # Interactive elements only (recommended)
agent-device snapshot -c               # Compact output
agent-device snapshot -d 3             # Limit depth
agent-device snapshot -s "Camera"      # Scope to label/identifier
agent-device snapshot --raw            # Raw node output
agent-device diff snapshot             # Structural diff against previous session baseline

XCTest is the iOS snapshot engine: fast, complete, and no Accessibility permission required.

Snapshot diff notes:

• First diff snapshot call initializes baseline for the current session.
• Subsequent diff snapshot calls compare current UI to prior baseline and then update baseline.
• Use this for compact change tracking between adjacent UI states.

Find (semantic)

agent-device find "Sign In" click
agent-device find text "Sign In" click
agent-device find label "Email" fill "user@example.com"
agent-device find value "Search" type "query"
agent-device find role button click
agent-device find id "com.example:id/login" click
agent-device find "Settings" wait 10000
agent-device find "Settings" exists

Settings helpers

agent-device settings wifi on
agent-device settings wifi off
agent-device settings airplane on
agent-device settings airplane off
agent-device settings location on
agent-device settings location off
agent-device settings faceid match
agent-device settings faceid nonmatch
agent-device settings faceid enroll
agent-device settings faceid unenroll

Note: iOS wifi/airplane toggles status bar indicators, not actual network state. Airplane off clears status bar overrides. iOS settings helpers are simulator-only. Use match/nonmatch as the canonical command values. Think of them as validate/invalidate outcomes when describing intent.

Logs (token-efficient debugging)

Use the detailed logs workflow reference: skills/agent-device/references/logs.md

Recommended minimum:

agent-device logs doctor
agent-device logs start
agent-device logs path

App state

agent-device appstate

• Android: appstate reports live foreground package/activity.
• iOS: appstate is session-scoped and reports the app tracked by the active session on the target device.
• For iOS appstate, ensure a matching session exists (for example open --session <name> --platform ios --device "<name>" <app>).

Interactions (use @refs from snapshot)

agent-device press @e1                # Canonical tap command (`click` is an alias)
agent-device focus @e2
agent-device fill @e2 "text"           # Clear then type (Android: verifies value and retries once on mismatch)
agent-device type "text"               # Type into focused field without clearing
agent-device press 300 500             # Tap by coordinates
agent-device press 300 500 --count 12 --interval-ms 45
agent-device press 300 500 --count 6 --hold-ms 120 --interval-ms 30 --jitter-px 2
agent-device press @e1 --count 5             # Repeat taps on the same target
agent-device press @e1 --count 5 --double-tap # Use double-tap gesture per iteration
agent-device swipe 540 1500 540 500 120
agent-device swipe 540 1500 540 500 120 --count 8 --pause-ms 30 --pattern ping-pong
agent-device longpress 300 500 800     # Long press on iOS and Android
agent-device scroll down 0.5
agent-device pinch 2.0              # Zoom in 2x (iOS simulator only)
agent-device pinch 0.5 200 400     # Zoom out at coordinates (iOS simulator only)
agent-device back
agent-device home
agent-device app-switcher
agent-device wait 1000
agent-device wait text "Settings"
agent-device is visible 'id="settings_anchor"'  # selector assertions for deterministic checks
agent-device is text 'id="header_title"' "Settings"
agent-device alert get

Get information

agent-device get text @e1
agent-device get attrs @e1
agent-device screenshot out.png

Deterministic replay and updating

agent-device open App --relaunch      # Fresh app process restart in the current session
agent-device open App --save-script   # Save session script (.ad) on close (default path)
agent-device open App --save-script ./workflows/app-flow.ad  # Save to custom file path
agent-device replay ./session.ad      # Run deterministic replay from .ad script
agent-device replay -u ./session.ad   # Update selector drift and rewrite .ad script in place

replay reads .ad recordings. --relaunch controls launch semantics; --save-script controls recording. Combine only when both are needed. --save-script path is a file path; parent directories are created automatically. For ambiguous bare values, use --save-script=workflow.ad or ./workflow.ad.

Fast batching (JSON steps)

Use batch when an agent already has a known short sequence and wants fewer orchestration round trips.

agent-device batch \
  --session sim \
  --platform ios \
  --udid 00008150-001849640CF8401C \
  --steps-file /tmp/batch-steps.json \
  --json

Inline JSON works for small payloads:

agent-device batch --steps '[{"command":"open","positionals":["settings"]},{"command":"wait","positionals":["100"]}]'

Step format:

[
  { "command": "open", "positionals": ["settings"], "flags": {} },
  { "command": "wait", "positionals": ["label=\"Privacy & Security\"", "3000"], "flags": {} },
  { "command": "click", "positionals": ["label=\"Privacy & Security\""], "flags": {} },
  { "command": "get", "positionals": ["text", "label=\"Tracking\""], "flags": {} }
]

Batch best practices:

• Batch one screen-local flow at a time.
• Add sync guards (wait, is exists) after mutating steps (open, click, fill, swipe).
• Treat prior refs/snapshot assumptions as stale after UI mutations.
• Prefer --steps-file over inline JSON.
• Keep batches moderate (about 5-20 steps).
• Use failure context (step, partialResults) to replan from the failed step.

Stale accessibility tree note:

• Rapid mutations can outrun accessibility tree updates.
• Mitigate with explicit waits and phase splitting (navigate, verify/extract, cleanup).

Trace logs (XCTest)

agent-device trace start               # Start trace capture
agent-device trace start ./trace.log   # Start trace capture to path
agent-device trace stop                # Stop trace capture
agent-device trace stop ./trace.log    # Stop and move trace log

Devices and apps

agent-device devices
agent-device apps --platform ios              # iOS simulator + iOS device, includes default/system apps
agent-device apps --platform ios --all        # explicit include-all (same as default)
agent-device apps --platform ios --user-installed
agent-device apps --platform android          # includes default/system apps
agent-device apps --platform android --all    # explicit include-all (same as default)
agent-device apps --platform android --user-installed

Best practices

• press is the canonical tap command; click is an alias with the same behavior.
• press (and click) accepts x y, @ref, and selector targets.
• press/click support gesture series controls: --count, --interval-ms, --hold-ms, --jitter-px, --double-tap.
• --double-tap cannot be combined with --hold-ms or --jitter-px.
• swipe supports coordinate + timing controls and repeat patterns: swipe x1 y1 x2 y2 [durationMs] --count --pause-ms --pattern.
• swipe timing is platform-safe: Android uses requested duration; iOS uses normalized safe timing to avoid longpress side effects.
• longpress is coordinate-based and supported on iOS and Android.
• Pinch (pinch <scale> [x y]) is iOS simulator-only; scale > 1 zooms in, < 1 zooms out.
• Snapshot refs are the core mechanism for interactive agent flows.
• Use selectors for deterministic replay artifacts and assertions (e.g. in e2e test workflows).
• Prefer snapshot -i to reduce output size.
• Prefer scoped snapshots (-s "<label>" or -s @ref) for screen-local tasks.
• Add -d <depth> when only upper tree levels matter; avoid full-tree snapshots by default.
• Use diff snapshot after mutations to detect structural changes with less output than full re-read.
• Refresh refs immediately after navigation/modal/list mutations before issuing next ref-targeted action.
• Use --raw only for debugging parser/tree edge-cases; avoid it for normal agent loops due to size.
• On iOS, snapshots use XCTest and do not require Accessibility permission.
• If XCTest returns 0 nodes (foreground app changed), treat it as an explicit failure and retry the flow/app state.
• open <app|url> [url] can be used within an existing session to switch apps or open deep links.
• open <app> updates session app bundle context; open <app> <url> opens a deep link on iOS.
• Use open <app> --relaunch during React Native/Fast Refresh debugging when you need a fresh app process without ending the session.
• Use --session <name> for parallel sessions; avoid device contention.
• Use --activity <component> on Android to launch a specific activity (e.g. TV apps with LEANBACK); do not combine with URL opens.
• On iOS devices, http(s):// URLs fall back to Safari automatically; custom scheme URLs require an active app in the session.
• iOS physical-device runner requires Xcode signing/provisioning; optional overrides: AGENT_DEVICE_IOS_TEAM_ID, AGENT_DEVICE_IOS_SIGNING_IDENTITY, AGENT_DEVICE_IOS_PROVISIONING_PROFILE.
• Default daemon request timeout is 45000ms. For slow physical-device setup/build, increase AGENT_DEVICE_DAEMON_TIMEOUT_MS (for example 120000).
• For daemon startup troubleshooting, follow stale metadata hints for ~/.agent-device/daemon.json / ~/.agent-device/daemon.lock.
• Use fill when you want clear-then-type semantics.
• Use type when you want to append/enter text without clearing.
• On Android, prefer fill for important fields; it verifies entered text and retries once when IME reorders characters.
• If using deterministic replay scripts, use replay -u during maintenance runs to update selector drift in replay scripts. Use plain replay in CI.

References

• references/snapshot-refs.md
• references/session-management.md
• references/permissions.md
• references/video-recording.md
• references/coordinate-system.md
• references/batching.md