Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add skills/bearlike/assistant/device-controlnpx skills add bearlike/Assistant --skill device-controlgit clone --depth 1 https://github.com/bearlike/AssistantWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00041 | $0.01625 |
| Opus 5 | $0.00020 | $0.00813 |
| Sonnet 5 | $0.00008 | $0.00325 |
| Haiku 4.5 | $0.00004 | $0.00162 |
Grade A, and why
device-control scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 144 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Driving the device
You are operating the user's real phone. Every tap lands on a live screen, and the user is watching.
Ask first, then work, then hand it back
Control is granted, not assumed. device_ui, device_action and device_shell
refuse with device_control_not_started until you hold a grant.
device_control_start— take control.- Do the work.
device_control_stop— give it back, as soon as you are done.
The grant lasts for this turn only. If the user replies and you need the screen
again, start again — it is cheap, and already_active is a normal answer.
While it is held, the phone shows a persistent notification saying you can control it, with a Stop the user can press at any moment. Leaving a grant open after you have finished leaves that notification sitting there, which is why step 3 is not optional.
When start refuses
Every refusal names something the user can do. Relay it and stop — none of these clears by retrying.
| Outcome | What it means | What to tell the user |
|---|---|---|
granted |
you have control | nothing; get on with it |
already_active |
you already had it | nothing |
shizuku_not_installed |
the Shizuku app is missing | install Shizuku, then start it |
shizuku_not_running |
installed, service down — normal after a restart | open Shizuku and start the service |
permission_denied |
running, but Aura is not authorised in it | open Aura's Settings and tap the screen-control row |
permission_denied is the one worth reading carefully: the authorisation lives
inside Shizuku, not in Android's permission screen, so "check app permissions"
is the wrong advice.
The loop
Observe, act, read what came back, decide. The action's own result already carries the settled element list, so you rarely need a separate observation between steps.
device_ui(action="elements")— see what is on screen.device_action(action="tap", index=N)— act on an element by its index.- Read the element list in that result. It is the new screen.
- Repeat until the task is done, or until you are stuck and should say so.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 2d ago First seen · 144 lines · 41 tokens per session scan A fc6734034b6c
device-control is a skill published in the GitHub repository bearlike/Assistant (41 stars, last pushed 8d ago), licensed MIT. It adds 41 tokens to every session and 1,625 once invoked, about $0.0002 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
largest-files-finder
List the top 3 largest files in a given directory.
extract-action-items
Extract a list of action items with owners and deadlines from a meeting transcript.
cross-border-customer-service-brief
Cluster refund, return, complaint, and bad-review notes into a daily support brief with response guidance and escalation hints.
cross-border-listing-image-layout-audit
Audit listing image and layout notes — first-impression hero, info hierarchy, lifestyle vs studio, copy density, trust signals, compliance — into a follow-up checklist.
cross-border-price-match-review
Cluster competitor price, coupon, shipping, and bundle observations into a price-gap review with explicit human-approval guidance.
cross-border-product-scout
Turn detected market signals (or an upstream spike review) into screened follow-up directions and explicit human-review checkpoints.