Mobilerun device control — you are the agent

mobilerun is your hands and eyes on a real phone. You are the brain. You read the screen with mobilerun device ui / device screenshot, decide the next move with your own reasoning, and act with deterministic mobilerun device commands. No cloud, no Gemini, no mobilerun run, no API key — those are explicitly out of scope here.

Scripts live in scripts/ next to this file; reference docs in references/. Set two variables once per session, before the loop — the skill is invoked with the project cwd, not this folder, so a bare scripts/mr-preflight.sh will No such file or directory. And because each Bash call is a fresh shell (one command per call), no cd persists between calls to rescue a relative path — so address scripts absolutely:

export MR_DEVICE=<serial>                       # find it with: mobilerun devices
export SKILL=<absolute path to this skill dir>  # the folder holding this SKILL.md

Then call every script as bash "$SKILL/scripts/<name>" — that's how the runnable blocks below are written, so they work from any cwd. (Prose and tables name a script by its bare filename, e.g. mr-tap.sh, purely for brevity.)

The mental model

Six ideas make every other instruction in this skill obvious. Hold them and the right move usually follows on its own.

1. You are the agent. The cloud path (mobilerun run / Gemini) needs an API key this setup doesn't have, so the intelligence is you: you read the screen, choose the next move, and issue one deterministic mobilerun device command at a time. There's nothing else deciding — every tap is yours.
2. Act on the snapshot you just took. The element list is rebuilt by a pre-order walk and renumbered on every UI change, so index 9 in an old capture is a different element now. The tap you're about to make has to come from the capture you're looking at: snap → act on that snap → snap again to see what happened.
3. One command per call, and read its printed line. Each Bash call runs a single mobilerun command so you can read its result — Tapped (x,y), Typed:, the tree. Piping (cmd | tail) hands you tail's exit code instead of mobilerun's, so a real error reads as success; trust the printed words, or ${PIPESTATUS[0]} if you must pipe.
4. Move like a person. Pause between actions and aim a few pixels off dead-center. This reads human, and it's also more correct: the dwell lets animations and lazy lists settle before your next snap, and the offset avoids landing on an overlapping child element. mr-wait.sh and mr-tap.sh's jitter do this for you.
5. Compose outward steps, show them, then commit. Anything other people or the account will see — posting, sending, commenting, buying, installing/uninstalling, signing out, a factory reset — you build up to and then stop, show the user, and submit only on their yes. A reversible setting you're just testing is different: change it, verify, and restore it yourself.
6. Begin ready. mr-preflight.sh resolves the device, rides out the USB settle, and re-asserts accessibility in the form the version check expects — the things that otherwise make the first command of a session fail. Run it once up front and the loop starts on solid ground.

The through-line: perception before action. Hold the six ideas above and one truth explains them all — almost every failure on this device is a perception failure, not an action failure. You rarely fail because you tapped the wrong pixel; you fail because you acted on a belief about the screen that a fresh snap would have corrected. That belief takes four shapes, and naming the one you're in hands you the fix:

• Stale snapshot — you act on a capture the screen already moved past; any scroll renumbers the tree. Snap → act on that snap → snap again.
• Assumed location — you act on where you think you are, not what the header says: start resumes a leftover screen, and a stray back can drop you into another app entirely. Read the phone-state activityName before you act.
• Unconfirmed blank — you read "no output" as an empty tree without checking the snap itself was empty; a swallowed pipe looks identical to a wedged a11y service. Confirm the missing Clickable UI elements line before you believe it.
• Trusted doc — you act on a documented shape, flag, or schema you never ran; docs drift, the device does not. Run it once and read the real bytes.

So when something doesn't work, your first move is not a different tap — it's a fresh snap, because a surprising result is data about your perception before it is data about the device. The lone exception is a tool-contract gap: when the command itself cannot express what you need (clearing a field to empty), you fix the tool, not your looking — see references/text-input.md.

The loop

0. SETUP (once):      export MR_DEVICE=<serial> ; export SKILL=<abs path to this skill dir>
   PREFLIGHT (once):  bash "$SKILL/scripts/mr-preflight.sh"     # ready + heals a11y/version/USB
   ── then repeat ──
1. CAPTURE:           bash "$SKILL/scripts/mr-snap.sh" [--boxes]   # indexed UI tree (+ screenshot, +numbered boxes)
2. DECIDE (you):      read the tree/screenshot; pick the target element by its TEXT/resourceId;
                      copy its bounds (x1,y1,x2,y2). When unsure of layout, Read the SHOT= PNG.
3. ACT:               bash "$SKILL/scripts/mr-tap.sh" --bounds "x1,y1,x2,y2"   # jittered box-center tap
                      bash "$SKILL/scripts/mr-type.sh" "text" [--clear|--human]   (focus the field first)
                      mobilerun device swipe … / press back|home|enter / start <pkg>
                      bash "$SKILL/scripts/mr-wait.sh" [read|think|MIN MAX]   # human pause between steps
4. VERIFY:            bash "$SKILL/scripts/mr-snap.sh" ; confirm the expected change appeared.
                      If nothing changed, retry the tap once; if it's still unchanged, re-plan from
                      the new snap rather than repeating the same tap.

Give the loop a budget: if the goal state isn't reached in ~15–25 capture→act cycles, stop and report what you saw — a stalled plan is information, so re-plan instead of looping on.

The box-select click system (the core mental model)

mobilerun device ui prints every on-screen element as:

Current Clickable UI elements:
'index. className: resourceId; checkedState, text - bounds(x1,y1,x2,y2)':
9. TextView: "android:id/title", "Network & internet" - (221,462,566,519)
34. Switch: "…/switchWidget", "Dark theme" ; isChecked=False - (859,1912,996,2038)

• The bounds (x1,y1,x2,y2) are absolute device pixels: left,top,right,bottom.
• The Portal can draw these as numbered boxes on screen (mr-snap.sh --boxes, or adb shell content insert --uri content://com.mobilerun.portal/overlay_visible --bind visible:b:true). The number in each box's top-right corner matches the device ui index — this is your set-of-marks. Use it when you (or the user) want to see what's clickable.
• To click element N: center = ((x1+x2)//2, (y1+y2)//2), then tap (with jitter). mr-tap.sh --bounds "…" does the math + jitter — the same center mobilerun computes internally.
• Identify a target by its text / resourceId, not by a remembered number: the index is only meaningful in the snap it came from, so re-snap and re-find the element each time before you tap.

Full mechanics, overlap handling, raw JSON: references/box-select-and-clicking.md.

Capture: which signal to use

How to choose between them, and how to handle the temp-path screenshot: references/capture-state.md.

Acting: primitives

Gestures, scrolling a list to find an item, paging: references/navigation-and-gestures.md. Text entry, how the keyboard works on this device, atomic paste vs human typing: references/text-input.md.

Stealth & looking human

Default to human cadence: mr-wait.sh between actions, mr-wait.sh read while "reading" a post, jittered taps, natural scroll distances, and --human typing for anything a person authored (comments, messages). See references/stealth-and-human-pacing.md. When you compose text a human will be seen to have written, also read references/writing-like-a-human.md so it doesn't read like an LLM.

Planning any task

Decompose the goal into the loop above, expressed as concrete CLI steps before you touch the device. Worked end-to-end plans (Reddit "read N posts then comment", Settings toggles, search-and- open, multi-app flows) are in references/task-planning-playbook.md. Start there for any non-trivial request.

Reference routing