agent-device

Pre-flight (auto)

These checks evaluate at skill load. If any line shows FAIL, stop and surface the fix before running any device command.

agent-device version: !R=0.18.0; V=$(agent-device --version 2>/dev/null); [ -n "$V" ] && [ "$(printf '%s\n%s\n' "$R" "$V" | sort -V | head -1)" = "$R" ] && echo "OK ($V)" || echo "FAIL (need v$R+, got: ${V:-not installed}). Fix: npm install -g agent-device@latest"

Bundled CLI skills dir: !D="$(npm root -g)/agent-device/skills/agent-device"; test -s "$D/SKILL.md" && echo "OK ($D)" || echo "FAIL (missing $D/SKILL.md). Fix: npm install -g agent-device@latest"

HybridApp mode: !M=$(scripts/is-hybrid-app.sh 2>/dev/null | tail -1); [ "$M" = "true" ] && echo "OK (HybridApp)" || echo "FAIL (got: ${M:-unknown}). This skill only supports the HybridApp build - ensure the Mobile-Expensify submodule is present."

Bring-up

Run this sequence the first time the user asks for device interaction in a session, before any open / snapshot / replay.

1. Platform

If the user prompt names ios or android explicitly, use it. Otherwise ask. Only iOS and Android are supported; reject other platforms.

2. Bundle ID

HybridApp dev builds only (the pre-flight gate enforces this).

3. Confirm dev build is installed

agent-device apps --platform <p> --json

If the resolved bundle ID is missing from the list, STOP and instruct the developer to run the matching build command from the App repository root. The build script detects HybridApp mode and builds the native app from Mobile-Expensify/.

4. Metro

agent-device metro prepare --public-base-url http://localhost:8081 --port 8081 --kind react-native

If metro prepare fails, STOP and surface the error verbatim.

5. Pick a target device

agent-device devices --platform <p> --json

• Prefer the first device with booted=true.
• If none are booted, choose the default target device (usually the first listed), then continue to step 6 to detect and clear any stale session bound to that device before opening.
• If multiple are booted, ask the user which.

Capture the device name and (for iOS) the simulator UDID, or (for Android) the serial.

6. Session reuse vs reset

agent-device session list --json

For each entry whose device_udid (iOS) or serial (Android) matches the chosen device:

• If the session was created earlier in the current Claude invocation, reuse it silently.
• Otherwise prompt: reuse (continue with the existing session) or reset (force-close it).
  • To reset: agent-device close --shutdown --session <name>. --shutdown also frees the simulator.

7. Open

agent-device open <bundle-id> --platform <p> --device "<name>"

If open errors with "app not installed", revisit step 3.

8. Sanity

agent-device snapshot -i

Confirm the app rendered. From here, follow the Agent decision loop for repeatable flows or drive interactively.

9. Interaction safety

After opening or relaunching the App, inspect agent-device snapshot -i for React Native development overlays before interacting.

If the snapshot reports a LogBox warning, run:

agent-device react-native dismiss-overlay
agent-device snapshot -i

Continue only when the fresh snapshot no longer reports the overlay. Multiple LogBox banners can require repeated dismiss-overlay and fresh snapshot -i calls. If the snapshot reports a RedBox fatal error, stop and surface the error instead of dismissing it.

Before pressing an action that may be covered by an overlay or system UI:

agent-device snapshot -i
agent-device screenshot --overlay-refs

Use a stable selector or a fresh @eN reference. Confirm the target is reported as hittable. Never use coordinates to bypass interactionBlocked: "covered", reason: "offscreen_ref", or targetHittable: false.

When a target is rejected, capture agent-device screenshot --overlay-refs. Dismiss a recoverable LogBox overlay when present, capture a fresh snapshot -i, and retry only through a selector or fresh reference. Otherwise stop and report the blocker.

After pressing the action, verify the expected destination or control state with wait, is, find, or a fresh snapshot.

Canonical skill references

Read these files directly for device automation guidance (bootstrap, exploration, verification, debugging): !echo "$(npm root -g)/agent-device/skills/agent-device"

Flows

Repeatable steps (sign-in, onboarding, etc.) are captured as composable .ad snippets under flows/. For interactive usage, propose and run only flows/macros/ helpers. flows/tests/ belongs to a separate QA workflow and must not be proposed by this skill; QA/perf runs execute them via agent-device test <path>.