What CI cannot check. Everything device-free is in make ci, which the
CI workflow runs on every push — formatting,
vet, both modules' tests, the four non-Go clients' own tests, and six
cross-compile targets. This file is the rest, and the rest is where most
serious defects in this project have come from: 328 of 408 were found only by
running against a real device.
Work through it before tagging a release, on three substrates: an Android emulator, an iPhone simulator and a real iPhone — and a real Android phone for the items that say so. Most of it is now the scripts in checks/, which assert what they can and print what a person should read.
Emulator and simulator half, 2026-10-07: every check in
checks/ on a Pixel 7 AVD (Android 15) and an iPhone 17 Pro
simulator (iOS 26.5), side by side, about two hours. Android passed all 48
that apply to it, after the emulator stopped answering adb mid-run and was
shut down and booted again; the simulator found three defects and two
checks out of date (CHALLENGES 288–290), and fourteen checks that ran on the
other platform and failed for the wrong reason (291). icecubes-ios.sh
failed twice and passed once, a lead in ROADMAP. Not run: the phones and
Windows, and the manual items below.
The iPhone half, 2026-10-07: an iPhone 15 Plus, iOS 26.6.2. "A real
iPhone" below, all five, and then every check that runs on a phone: all
passed or refused correctly (battery, shake), except icecubes-ios.sh, the
lead in ROADMAP, now seen on the phone too. The phone locked mid-batch and
three checks failed for that alone, passing when run again; one of them
misread it as a finding (CHALLENGES 292). Still to run: the Pixel, Windows
and the manual items.
The Pixel half, 2026-10-07: a Pixel 8 Pro, Android 17: calculator.sh,
external-driver.sh, crashes.sh and clients.sh (all five clients), and
every check that runs on a phone, with the owner's settings compared before
and after — all fifteen read back as found, and no mock location left. Two
defects found and fixed (CHALLENGES 293, 294). chrome.sh and pwa.sh
stopped at Chrome's own notification promo, which Mobium refused to tap
through, rightly; answering it is the owner's. Not run on this phone:
clock-timer.sh, which force-stops Clock and so would cancel the owner's
alarms, and compose-app.sh and third-party-app.sh, whose apps are not on
it. Still to run: Windows and the manual items.
Last full run: 2026-09-25, on a Pixel 7 AVD (API 35), an iPhone 17 Pro
simulator (iOS 26.5) and an iPhone 15 Plus (iOS 26.6.2): 29 minutes by the
clock, from booting the emulator to clean-stop.sh reporting clean, with the
three devices' scripts running side by side, each under its own
MOBIUM_SESSION. The scripts took 16 minutes on the phone, 8 on the
simulator and 6½ on the emulator; the phone's are the long pole. Everything below passed except what is listed here, and
it found four defects before they shipped (CHALLENGES 91 and 92, and two
checks that could only pass alone), and the phone below found three more:
- A real Android phone, run the same afternoon on a Pixel 8 Pro
(Android 17):
calculator.sh,clock-timer.sh,external-driver.sh,crashes.shandclients.shpass. It found two defects the emulator could not (CHALLENGES 93, 94) and one in the client flows. - Not run: Windows — the laptop has not arrived; see that section.
- Skipped by design on the phone: setting a location and writing the clipboard, which a real iPhone refuses and the check asserts the refusal.
A hosted Android emulator job is possible and a hosted iOS simulator job is easy. Neither is here, for one reason: a device test that runs unattended tells you it failed, not what the screen looked like. Every defect in CHALLENGES.md was diagnosed by reading a hierarchy or comparing a computed point against a screenshot, which is a human activity.
That is a judgment, not a law. Revisit it when there is a suite worth running unattended — a regression pack of screens with known-correct maps would be the thing to build first.
It also runs against the advice in Emulator vs Simulator vs Real Device, which this project's author wrote. That piece puts virtual devices squarely in CI — roughly 90% of pre-commit validation, 70/30 virtual to real at multi-merge, real devices nightly — on the grounds that a bug costs more the later it is caught, so the cheap substrate belongs early.
The argument is right and the conclusion does not transfer, for a reason specific to what is being tested here. That pipeline assumes the app under test is the thing that might be wrong and the harness is trusted. Here the harness is the thing under test, and its defects are overwhelmingly misreadings of a device rather than crashes: a label attached to the wrong node, a tap computed 186 pixels high, a rotation read in the wrong unit, a password printed in plaintext. Each was caught by looking at a hierarchy or a screenshot. A green unattended run would have reported all of them as passing, because nothing threw.
So the split is the same and the line sits elsewhere: everything that can be
decided by an assertion is in make ci and runs on every push, and everything
that needs a person reading output is here. The day a regression pack can
assert "this screen maps to exactly these refs", most of this file moves left
and the advice applies unchanged.
SETUP lists what each substrate cannot do at all.
Setup is in SETUP.md, or WINDOWS.md for Windows.
Start the emulator, then mobium daemon stop so nothing serves a stale build.
-
mobium deviceslists the emulator -
mobium mapon the launcher returns app icons asbutton, notlink -
mobium launch com.android.settings && mobium tap "text=Network & internet"with no sleep between them — the implicit wait is what makes this work -
mobium wait "text=Network & internet"returns a ref thatmobium tapthen accepts without an interveningmap -
mobium scroll-to "text=About emulated device"reaches it, and reports a scroll count above zero -
mobium scroll-to "text=Definitely Not Here"stops at the end of the list in a few swipes, not fifteen -
mobium grant com.android.chrome allgrants every declared permission, confirmed withadb shell dumpsys package com.android.chrome | grep granted=false -
mobium --driver uiautomator mapworks on an ordinary screen and refuses the About page by name (see defect 25) -
mobium screenshot -o /tmp/x.pngwrites a PNG whose coordinates agree with whatmapreported -
mobium appearance darkthenlight, each confirmed by reading it back -
mobium appslists what is installed;--systemincludes the platform's - After a
--driver uiautomatorrun,adb shell ls /data/local/tmphas no mobium file in it -
mobium uninstall com.android.settingsfails, naming DELETE_FAILED_INTERNAL_ERROR rather than reporting success
-
mobium --driver wda mapon the home screen returns icons - Wait for something already on screen, and across a cold app launch
-
mobium scroll-toreaches a Settings row below the fold, and tapping it opens that screen -
mobium grant com.apple.Preferences photossucceeds and says it could not be verified -
mobium grant <anything> camerais refused, naming the missing service - Coordinates from
mapland where the screenshot says they should — iOS reports points, screenshots are pixels, and a 3x error is silent -
mobium appearance darkand a screenshot that is actually dark;autois refused -
mobium appslists WebDriverAgentRunner; uninstalling it works and the next command reinstalls it unaided -
./docs/checks/ios-webview.shpasses — WKWebViews over Remote Web Inspector, including that a tap in Safari's page, placed from text both sides report (CHALLENGES 248), opens the page its link points to. Stop the daemon first: a page's inspector target belongs to one debugger at a time, so a session left switched into a page blocks it - Switch into a WebView, leave, and switch back in on the same daemon
session. This is the one that breaks silently — a page's target is
announced once per connection, so a missing
_rpc_forwardDidClose:makes the second attach fail on a page that is plainly there
Paired, unlocked, and with Developer Mode on — see SETUP.md. The first command builds and signs WebDriverAgent, which takes a minute or two.
-
./docs/checks/ios-device.sh <udid>passes — app switches timed against the 61-second stall (defect 71), rows with no bounds, typing read back, and every refusal a phone owes -
./docs/checks/third-party-app-ios.sh <udid>passes — Wikipedia from the App Store, one of two third-party apps driven on iOS; the other, NetNewsWire, isnetnewswire-ios.sh, below -
MOBIUMAPP_BUNDLE=<path>/MobiumApp.app ./docs/checks/mobium-app.sh <udid>passes, WebViews included — the phone's reach them over lockdown, not the simulator's socket -
./docs/checks/crashes.sh <udid>passes with MobiumApp installed: its logged line is in the captured log, and its Crash Demo's report is listed, read in full, and its cause found in the log rather than the report -
mobium --device <udid> current, without--driver, names the wda backend rather than saying there is no such device (defect 88)
-
./docs/checks/calculator.sh <serial>passes on an emulator -
./docs/checks/clock-timer.sh <serial>passes on an emulator - Both pass on a real device, unchanged
-
./docs/checks/external-driver.sh <serial>passes — the third-party driver protocol, with the reference driver's map diffed against the built-in backend's on the same screen -
./docs/checks/device-state.sh <serial>passes — all four orientations read back, the hierarchy follows the rotation, a tap lands in landscape, and an app pinned toja-JPrenders in Japanese with no split characters -
./docs/checks/third-party-app.sh <serial> wikipedia.apkpasses — Wikipedia, the first app driven here that nobody at Google wrote; F-Droid and Aegis followed, and Seal iscompose-app.sh, below. If a labeling change is in this release, these are the checks that catch it -
./docs/checks/gestures.shand./docs/checks/zoom.shpass on the emulator, the simulator and the phone — each gesture asserted against the one it could be mistaken for, and the refusals where a platform cannot send one (defects 84–86) -
./docs/checks/mobium-app.shpasses on the emulator and the simulator -
./docs/checks/crashes.shpasses on the emulator and the simulator — device logs and crash reports, each against a crash caused on purpose
These checks are in checks/ but were not on this list for the last full run (2026-09-25), and most were written after it, so that run vouches for none of them. What each proves, and the devices it runs on, is in checks/README.md.
- Third-party apps:
netnewswire-ios.sh <udid>(NetNewsWire, iOS) andcompose-app.sh <serial>(Seal, the first Jetpack Compose app) - MobiumApp's demos:
login.sh,keyboard.sh,otp.sh,dialogs.sh,autowait.sh,obstruction.sh,hit-test.sh,wait-states.sh,map-diff.sh,source.sh,clear-data.sh,files.sh,battery.sh,biometric.sh,accessibility.shand, on Android,screen-reader.sh - Inside a WebView:
web-actionability.sh,web-type.shandweb-storage.sh; andios-webview-probe.sh <simulator-udid>, the platform assumption with no Mobium code - Back:
back.sh <android-serial | simulator-udid | iphone-udid> - Files at a path:
device-paths.sh <android-serial | simulator-udid | iphone-udid>, withMOBIUM_DEBUGGABLE_APPon an emulator - A PWA from the Play Store:
twa.sh <android-serial>, with OYO Lite installed from Play - Flutter:
flutter.sh <android-serial | simulator-udid | iphone-udid>, with MobiumApp'sflutter/demo installed - A mobile browser and a PWA:
chrome.sh <emulator-serial | android-serial>andpwa.sh <emulator-serial | android-serial | simulator-udid>, both with the network - Device state:
orientation.sh,network.sh <android-serial>,proxy.sh <android-serial>,shake.sh, and on iOSlocale-ios.sh,timezone-ios.sh,notifications-ios.shandaudit.sh - Recording:
record.sh,trace.shandcodegen.sh - Audio:
audio.sh <emulator-serial>andaudio-sim.sh <simulator-udid> - Devices started and stopped:
boot.sh <avd> [simulator-udid] - The test runner:
test-runner.sh,test-ui.sh(needs Vibium), andMOBIUM_GRID=<node> test-grid.sh
The extension point is a published protocol, so a break here is a break in somebody else's code that we cannot see.
-
examples/drivers/mobium-driver-adbstill imports nothing but the Python standard library. If a reference driver needs anything from this repository, the extension point is not open -
examples/drivers/PROTOCOL.md's method table matches whatinternal/mobiumdriveractually sends, field for field - Every capability in
mobiumdriver.KnownCapabilitieshas a matchingAs*helper, and nothing aboveinternal/mobiumdriverasserts a capability interface directly (grep 's\.driver\.(mobiumdriver\.'finds nothing) -
mobium doctorlists a driver placed on PATH, and says none are there when none are
Each client covers the whole tool surface, and make ci proves that by
name. ./docs/checks/clients.sh <serial> proves it by driving: the same flow
through each client's own API against a booted Android device — reading,
acting, waiting, scrolling, lifecycle, permissions, the device log and crash
reports, and a failure that must arrive as that client's own exception.
- Python
- JavaScript
- Go
- Java
- .NET
-
mobium doctoron a healthy machine reports no problems and exits 0 - With
PATHstripped of adb it names the fix and exits 1 - It still runs with
MOBIUM_HOMEset to something over the socket length limit — that is the case the first version could not report
-
mobium daemon stopfollowed immediately by any command works (defect 28) - One SIGTERM stops an idle daemon and removes its PID file; a stuck call
can no longer hold it up past half a minute (defect 89 — the case itself
is
TestShutdownIsBoundedWhenClosingHangs, since nothing on a device can be made to hang on purpose) - Quit the emulator mid-session; the next command recovers rather than failing until the daemon is restarted (defect 9)
-
--jsonoutput carries structured fields, not wrapped prose - The MCP server answers
tools/listand onetools/callover stdio
-
./docs/checks/clean-stop.shreports clean
- The Pinned agents workflow has
passed recently, or run
MOBIUM_NETWORK_TESTS=1 go test ./internal/device/ -run Network - If a pin was bumped, both checksums were regenerated from the download rather than edited by hand
Unsupported until it has run on Windows. The named-pipe transport and the other known gaps are written and cross-compile, and WINDOWS.md, "Doing the work", is what to run on a Windows machine. Windows becomes supported when those pass, and only then.
-
make crosscompilepasses (CI does this) - On Windows:
go test ./internal/daemon/ ./internal/paths/passes (CI does this, five times over, in thego-windowsjob), and the CLI auto-starts a daemon (CI does this too, and reads its status back) that survives its terminal closing (still by hand: no runner has a terminal to close)
The release workflow attaches mobium-<version>.mcpb — the MCP bundle for
macOS and Linux — and server.json, the registry entry pointing at it,
after checking both (packaging/mcp). Publishing
the entry is a person's step, under the name dev.mobium/mobium, which the
registry grants to whoever proves control of mobium.dev.
Not yet published (checked 2026-11-03): the registry has no
dev.mobium/mobium, and mobium.dev, on Spaceship's name servers, has no
TXT record of any kind. v0.1.0-rc1's release carries both files. Publish
first with v0.1.0, not a release candidate: an rc that is the only version
listed may be what the registry shows as the latest, and a published
version cannot be published again with other contents.
Once, before the first publish — these are the domain owner's, since the key is what proves control of the domain:
-
An Ed25519 key, made with OpenSSL 3 (macOS's own
opensslcannot), kept out of every repository:mkdir -p ~/.config/mcp-registry && cd ~/.config/mcp-registry /opt/homebrew/opt/openssl@3/bin/openssl genpkey -algorithm Ed25519 -out mcp-registry.pem chmod 600 mcp-registry.pem -
Its public key, for the DNS record:
/opt/homebrew/opt/openssl@3/bin/openssl pkey -in mcp-registry.pem -pubout -outform DER | tail -c 32 | base64 -
A TXT record on the apex of
mobium.dev(host@, not a selector or subdomain):v=MCPv1; k=ed25519; p=<public key> -
Seen from outside:
dig +short TXT mobium.dev @1.1.1.1prints it
Each release, after the release workflow has attached the assets:
-
mcp-publisherv1.8.1 or later, checked against the release's checksums before it is run:b=https://github.com/modelcontextprotocol/registry/releases/download/v1.8.1 curl -fsSLO "$b/mcp-publisher_darwin_arm64.tar.gz" && curl -fsSLO "$b/registry_1.8.1_checksums.txt" grep ' mcp-publisher_darwin_arm64.tar.gz$' registry_1.8.1_checksums.txt | shasum -a 256 -c tar -xzf mcp-publisher_darwin_arm64.tar.gz mcp-publisher -
Log in, the private key read from its file so it never reaches the shell's history:
./mcp-publisher login dns --domain mobium.dev --private-key \ "$(/opt/homebrew/opt/openssl@3/bin/openssl pkey -in ~/.config/mcp-registry/mcp-registry.pem -outform DER | tail -c 32 | xxd -p -c 64)" -
The release's own
server.json, never one built locally, published from its folder:gh release download v<version> -R mobiumdev/mobium -p server.json ./mcp-publisher publish -
The entry is there:
curl -s "https://registry.modelcontextprotocol.io/v0/servers?search=mobium" | grep -o '"dev.mobium/mobium"[^}]*"version":"[^"]*"'shows the new version. Search matches the part of a name after the slash, so searching fordev.mobiumfinds nothing, measured with the registry's ownio.modelcontextprotocol/everything