Phone Harness

phone-harness

Direct control of a phone from Python scripts. The same helpers drive every kind of phone below; what differs is how each one sees and touches the screen, and that is what the per-phone sections below are for. Read Which phone, the shared Working method, and then only the section for the phone in front of you.

Which phone?

"My phone" means the personal one. The personal iPhone and the cloud iPhone are two different devices with different apps and logins: the first is in the user's pocket and mirrored onto the Mac, the second lives in a data centre and is reached over HTTPS. When the user says "my phone", "my iPhone", or names an app they use, drive the personal iPhone. Use a cloud phone only when the user names it ("the cloud phone", cloud start) or the task is clearly about it. If the personal phone cannot be reached, say so and ask; never fall through to a cloud session that happens to be running, even one the user started - it is theirs, and it is a different phone.

phone-harness config shows the default platform (ios on a Mac, android elsewhere) and every other setting. phone-harness cloud shows whether a cloud phone is attached; while one is, scripts drive it regardless of the config default. PHONE_HARNESS_PLATFORM=android overrides per call. On a cloud iPhone any explicit platform selects a local backend; see Cloud iPhone.

When not to use any of them: if the task is doable on the Mac or the web - a website, an API, an app with a web equivalent - do it there and leave the phone alone. Use a phone when the task genuinely needs one: phone-only apps, things tied to a phone number or 2FA, checking how something looks on a phone.

Working method (every phone)

phone-harness <<'PY'
# task: report the OS version and model name from Settings
# step: open Settings, find About, read the screen
open_app("Settings")
print([o["text"] for o in ocr()][:10])
PY

• Invoke as phone-harness with a heredoc. Helpers are pre-imported. Start every script with # task: (the user's request in one sentence, identical across the scripts of one request) and # step: (what this script does).
• Tell the user what you are doing as you go. A phone task is many short scripts and the user sees none of them: one line before each script (what you are about to do), one line after (what you saw). Never run two scripts in a row in silence. The # task:/# step: comments are not this - the user cannot see them.
• Read with ocr(), not by eyeballing screenshots. Every visible string comes back with a tap-ready centre: [{text, confidence, source, x, y, w, h}]. Filter in Python before printing. tap_text("Weather") taps by label and, on failure, raises with what IS visible - read the exception before retrying. screenshot() costs more but shows what OCR cannot: icons, images, state.
• Act, verify, adapt. There is no DOM to assert against and no return value that means "it worked", so:
  1. Name what should change before you act - a title, a row, a field's contents. Most phone failures are silent no-ops; if you cannot name the expected change you cannot tell success from one.
  2. Do one thing, then check that one thing: wait_for_text("Got it"), wait_for_app("com.android.chrome") or wait_stable(), then one ocr(). Do not wait(n) for a screen to load.
  3. Once a sequence is proven, batch it: a whole sub-task in one script is far faster than a call per turn. Keep one cheap check at the end.
  4. When a check fails, isolate: re-run that one action, look at the screen, form one guess, test it. Do not re-run the batch hoping it lands.
  5. Keep what you learn in the agent workspace's agent_helpers.py. A checkout uses agent-workspace/. A pip or uv install uses ~/.phone-harness/agent-workspace, or ~/.local/share/phone-harness/agent-workspace when ~/.phone-harness is a git checkout, so an upgrade does not wipe it. PH_AGENT_WORKSPACE overrides that path. The directory is created from the defaults only when it is missing. An existing workspace is never overwritten. To get a fresh default, delete the workspace folder; it is recreated on the next run.
• The harness reports, you decide. Helpers return observations, never a verdict on your intent. Diff two ocr() sets, watch one label, count rows, poll until something appears - whatever matches what you asked for.
• Navigation: home(), back(), app_switcher(), open_app("Notes"), type_text("..."), press("enter"), long_press(x, y), tap(x, y).
• scroll says what you want to SEE; swipe says which way the finger goes. English disagrees the same way - "scroll down the page" and "swipe up for the next video" describe the same motion. scroll("down") reveals what is further down; swipe("up") is a thumb flick, the phrasing everyone uses for "next". scroll, scroll_screen, scroll_until and scroll_collect take the content direction; only swipe takes finger motion. "left"/"right" work on both. Which one moves what differs per phone - see each section.
• Walking a list: scroll_until(done) stops when your predicate on the visible boxes is met; scroll_collect(extract, key=...) walks and de-dupes, returning {items, stop, scrolls} with stop of 'reached-end' or 'max-scrolls'. Both end on YOUR check, so an extractor that misses rows ends the walk early. scroll_screen() is the single step and returns {before, after, boxes} for you to judge. at= aims the gesture at an inner list or strip; only the scroll view under that point moves.
• type_text needs a focused field on every phone. Tap the field, wait for the keyboard, then type; verify with ocr(), because unfocused text goes elsewhere or nowhere.
• Consent. This is the user's real phone and real accounts. Stop and ask before anything outward-facing or hard to reverse: sending, posting, following, purchasing, deleting, changing settings. Never type a PIN, a password or a 2FA code; on a cloud phone, hand the controls over instead (phone-harness cloud open, below).
• The harness reconnects what it can; the rest is the user's. On a personal iPhone, start the task with ensure_mirroring(): it opens iPhone Mirroring if it is closed, presses the window's Connect once, and brings the window to the front so the user can watch. Pairing, unlocking, tapping Allow, locking an iPhone that says "iPhone in Use" or "Timed Out", approving a cloud sign-in in the browser: when the harness still says the phone is not reachable, relay its message and ask - never loop-poll. Retry once after the user says it is done; that retry presses Connect again.

Cloud phones

There are two things. A phone is a row in the account catalog (GET /me -> available): a label, a platform, an id, and whether it is a temporary phone. The service keeps no default phone; a preference is the chooser's. A session is one of those phones rented right now. cloud ls prints one row per catalog phone and writes any live session on that row (* is the session this process targets). A session lines up with the row that has the same device_id or profile_id; a session with neither lines up with the temporary row. A cloud iPhone session does not echo that id: its device is the word "iPhone", and its profile (iphone-...) matches no catalog row. The catalog row's state is running while a session holds it. A session whose ids match no row counts as unmatched. When exactly one row of that platform and kind is running and exactly one unmatched session has that same platform and kind, the session is written on that row. A session that still matches nothing is listed on its own.

cloud start with no name starts this machine's cloud.phone setting (config set cloud.phone "Desk phone") when one is set; otherwise it sends no platform or kind and gets the service's rule for an unnamed start (GET /me -> unnamed): a temporary Android. A name is a catalog id or label, or ios / android when only one entry has that platform. Several matches print the catalog and stop (a terminal may ask which one). The CLI turns the chosen row into platform and kind, plus device_id for an iPhone or profile_id for the saved Android. The temporary row sends neither id.

What the helpers do next follows the session's link, not the platform name. An adb link is the Android section. A control link ({url, token}) is Cloud iPhone.

phone-harness cloud                          # signed in? which session is attached? time left?
phone-harness cloud ls                       # catalog, live sessions inlined; --json for the rows
phone-harness cloud start                    # cloud.phone if set, else a temporary Android; opens the live view
phone-harness cloud start ios --for 15m      # one iPhone, fifteen minutes
phone-harness cloud start "Temporary phone" --for 90s
phone-harness --session SID <<'PY'           # this process only; or PHONE_HARNESS_SESSION
print(screen_info())
PY
phone-harness cloud stop                     # ends billing; returns at once
phone-harness cloud phone                    # the saved Android's stored state
phone-harness cloud watch                    # reopen the read-only live view (--print to share)
phone-harness cloud open                     # the dashboard: the user takes the controls

• Pin a session when more than one can be up. --session SID or PHONE_HARNESS_SESSION selects one phone for this process, and every cloud command and helper honours it. With neither, a person at a terminal gets the attached session: the last cloud start, or cloud use SID. cloud use rewrites that shared fallback. Parallel agents should pass --session and leave cloud use alone, or they will steal the fallback from each other.
• It bills by the minute while it is up. Start it once and keep it for the whole conversation. Stop it when the user is done, and say that you did. --for is the length: 90s, 15m / 15min, or a bare number of minutes. Default 15, cap 30 (cloud.minutes, cloud.max_minutes).
• Stop before the deadline; a session cannot be extended. One that simply runs out keeps a saved Android's data but not its running state. The harness warns on stderr under two minutes; phone-harness cloud shows the time left. cloud stop returns at once. Stopping an adb phone that has a profile saves it; cloud start of that phone waits out a save still in progress. A control link keeps its data on the phone, so there is nothing to wait for.
• Already running. Starting a phone this machine already holds for the same catalog choice reuses that session. If the service says it is held elsewhere (409 profile_running), the message names the SID and says to attach with --session SID. Do that; do not start a second one. 409 device_unavailable means the phone is granted but not on its host. 409 session_limit names how many sessions are in use. 503 phones_busy includes Retry-After. 403 account_blocked means rentals are blocked; a 403 about credit or balance means add credit. A 400 that names no such phone reprints the catalog.
• Not signed in means the user runs phone-harness cloud login and approves it in a browser. Relay that; you cannot do it for them.
• Watching versus controlling. cloud start opens the live view so the user can watch. cloud watch reopens it. For a password, a 2FA prompt, or the user taking over, run phone-harness cloud open and wait until they say they are done.
• cloud phone is the saved Android only (GET /me/profile). cloud phone reset --yes wipes that phone's stored state: apps, logins, and data. The CLI has no iPhone reset; that is the operator's.
• A cloud session that ends does not hand you another phone. If the session attached to this script expires, the error says session <sid> expired. If the service loses it, the error says the cloud phone failed and is unreachable. Either way the script stops. It does not drive the personal phone, and the CLI does not rent a replacement. Start the same catalog phone again with a longer duration and continue the task: phone-harness cloud start <same phone> --for <longer>. <same phone> is that phone's id or label (or ios / android when only one entry has that platform). <longer> is a duration longer than the session that just ended (90s, 15m, or a number of minutes). A --session pin whose session expired is the same error, naming that sid. A missing cloud state file still means no cloud phone was attached, and the personal phone is used as today. An unreadable state file is an error, not a cue to switch phones. cloud stop clears the failure; a new cloud start does too.
• The adb address the CLI shows is not a secret; the unlock code is. Do not look for it, print it, or ask the user for it.

Cloud iPhone

Start it from the catalog: phone-harness cloud start ios when that is unambiguous, or the entry's id or label (cloud ls). The create body is platform: "ios", kind: "device", and device_id set to that entry's id. A ready session carries control: {url, token, expires_at} and no adb. From then on the helpers drive it over HTTPS. This machine needs no adb, no shell, and no Mac. It is a phone granted to the user's account and kept between sessions - not the personal iPhone mirrored on the Mac, which is a different device with its own apps and logins. cloud stop ends billing and returns at once; the phone keeps its own data. There is no Android-style disk save, and nothing on this phone for cloud phone to wipe.

An Android session can stay up beside it. Target one with --session SID. cloud use SID only changes the terminal fallback.

Length, billing, ls, watch, and stop are the Cloud phones commands above. An iPhone session reports billing_mode: disabled. Run scripts as plain phone-harness (with --session when you must pin one). PHONE_HARNESS_PLATFORM=... selects a local backend and skips the attached cloud phone.

The mirroring section describes a Mac window. This phone is the HTTPS path.

The client posts {"op", "kw"} to <control url>/op and reads the host's list from GET <control url>/ops (remote.py). supports(op) is that list. send of an op the host did not list raises Unsupported before the POST (this cloud phone cannot '...'). The op names and the shapes below are the client vocabulary in transport.py. Result shapes are what this repo documents or decodes.

Trust GET /ops when the host's list is shorter than this table.

• Coordinates are screen.bounds. tap(x, y), find_window(), send("screen.require"), and OCR centres all use that space: {x, y, w, h, id}. A point read off the PNG goes through tap_image_point() once screen_info() returns img_px. Otherwise tap an OCR centre.
• screen_info(), image_point(), and tap_image_point() measure the PNG so a point you read off a screenshot can become a tap. If screen_info() raises ModuleNotFoundError: No module named 'Quartz', that helper imported the macOS Vision module to read the PNG's pixel size. ocr() sends screen.text to the host and still works - keep using tap_text(). If screen_info() returns img_px, those three helpers work on any OS. PR #108 ("screen_info() no longer needs macOS") is the change that reads the PNG header instead of importing Quartz; until screen_info() returns img_px off a Mac, treat the Quartz error as those three helpers only.
• type_text() types into the focused field. Tap the field first. Text sent before the field has focus goes nowhere, so check with ocr() afterwards.
• Act, then look. One action, then ocr() or screenshot(), then the next. Do not wait(n) for a screen to load. The client allows screen.capture 60s, apps.launch and apps.list 90s, and input.text the timeout in the table.
• Errors the client raises. HTTP 400 with unsupported: true raises Unsupported (retrying will not help). HTTP 404 raises RuntimeError: the session is gone or the grant expired. Any other HTTP status raises RuntimeError with the host's error string. Create-time device_unavailable, profile_running, and phones_busy happen on cloud start, before any op, and are described under Cloud phones.
• ui(), back(), and current_app() raise Unsupported when the host does not list tree, nav.back, and apps.current. home() is the Home button. shell() is Unsupported.
• Confirm open_app. It sends apps.launch and can return with the app still not in front. Check with ocr() or a screenshot. If it missed, home() and tap the icon. tap_icon() in agent-workspace/agent_helpers.py aims 35 points above the label for the mirroring window, so measure this phone's icon yourself.