platforms.md

Every screen is a device

What can join this fleet, what it takes, and — for each one — whether that has actually been done or is only believed to work. The last column is the point of the page. A list of platforms a project "supports" is worth very little; a list that says which ones somebody has watched register is worth something.

The shelf today

Platform	Runner	How it joins	Verified
Android phone, tablet	runner-android	adb install, or sideload	yes — the original shelf
iPhone, iPad	runner-ios	Xcode, TestFlight internal	yes
iOS simulator	runner-ios	simctl install	yes
macOS, Linux, Windows	runner-machine	fleet up, or npm start, or a LaunchAgent / systemd unit	macOS yes, on both architectures — an Apple M1 Pro and a 2016 Intel i7 both registered, passed npm run conformance, and produce comparable synthetic numbers with matching digests; both run under fleet service. Linux and Windows run their suite in CI on x64 and arm64. CI also installs fleet service on all three on every pull request that touches it -- a systemd user unit on Ubuntu, a LaunchAgent on macos-latest, a scheduled task on windows-latest -- waits for the collector it starts and, on macOS and Windows, for the agent to register, then uninstalls it. The Windows task's first run is the first time a Windows agent registered against a real collector; it found the task had never started anything, now fixed. On Linux the agent also registers inside the Docker image in CI, on x64 and arm64. No service has been watched to come back after a reboot, and nobody runs a Linux or Windows machine in a fleet — see what each one reports
Android emulator	runner-android	adb install to the emulator	yes

What Wave 5 added

Platform	Runner	How it joins	Verified
Any browser	collector/runner-web/	open http://<collector>/runner	yes — driven end to end by a real Chromium against a real collector, both backends, digest checked
Android TV, Google TV, Fire TV	runner-android	adb connect, then adb install	partly — the built APK declares a leanback launcher and a banner, checked with aapt2 dump badging. No TV has run it. (Exploratory QA has driven an Android TV emulator from outside, over adb; that is the executor, not this runner — see what it can drive)
Meta Quest	runner-android	sideload via adb	no — Quest is Android 12 and reports UI_MODE_TYPE_NORMAL, so the runner identifies it by the VR feature flag. Never run on one
Wear OS	runner-android	adb install over Wi-Fi debugging	no — the watch feature flag is read at run time; the manifest deliberately does not declare it (that would make the APK a Wear-only app in a store). Never run on one
Apple TV (tvOS)	runner-ios, FleetRunnerTV target	devicectl or simctl	builds — xcodebuild -scheme FleetRunnerTV succeeds. No Apple TV has run it. (Exploratory QA drives tvOS 26.5 simulators through FleetDriver, a separate UI-test bundle)
Vision Pro (visionOS)	runner-ios, FleetRunnerVision target	devicectl or simctl	no — the target is written and has never been compiled; the visionOS platform is not installed on the machine it was written on
Raspberry Pi, Jetson, Steam Deck	runner-machine	install-agent.sh, or the Docker image	no — but arm64 Linux in CI reports a nearly complete descriptor, so a board should register cleanly. No board has
Anything with Docker	fleet/Dockerfile	docker run ghcr.io/addisdev/fleet	in CI — every release publishes linux/amd64 and linux/arm64. CI builds and starts both on every pull request that touches them -- arm64 natively on GitHub's arm64 runner -- and waits for each one's own machine agent to register. No board or NAS has run it

What Wave 6 added

Platform	Runner	How it joins	Verified
Roku player, Roku TV	runner-roku	build.sh --install <ip>, then an ECP launch	no -- and less verified than anything above it. No Roku was available and no Roku emulator exists, so the channel has never run. CI compiles it with brighterscript's bsc on every pull request that touches it; the firmware's own compiler has never seen it. See Roku: the one that has never run
Roku, from the executor	collector/src/drivers/roku.ts	SSDP discovery on the LAN	partly -- the SSDP and device-info parsers are checked against recorded response shapes in drivers.test.ts; the UDP socket path is exercised by nothing, and no Roku has answered it

The desktop app is not a runner, and it has never run on the platform it is for. On every pull request that touches it, desktop / cargo check compiles its Rust on macOS, and desktop / launch smoke builds and launches it on Ubuntu under a virtual display: its fleet sidecar comes up and registers an agent, its settings window answers over WebDriver, and SIGTERM stops both. The tray, macOS, and the local-network bet it exists for are untested; see its README.

Android: one APK, every shape

The Android agent is a foreground service and a text field. Nothing in it needs a phone, and it was never a code problem — it was a packaging one. Three declarations in the manifest do the work:

• uses-feature android.hardware.touchscreen required="false" (and faketouch). A television has no touchscreen, and installers and TV launchers treat that feature as required-by-default and filter the app out entirely.
• category android.intent.category.LEANBACK_LAUNCHER on the main activity. Without it the APK installs onto a Fire TV perfectly well and then cannot be started from the remote — which reads as "it does not work on TV" and is really "there is no icon".
• android:banner. A TV launcher draws a 320×180 banner, not an icon.

The phones are unaffected: declaring a feature not-required never removes it from a device that has one.

What differs per shape is only the kind the runner declares, read from UiModeManager and the VR and watch feature flags. A headset is asked about before a TV, because Quest reports UI_MODE_TYPE_NORMAL and only the VR flag tells the truth. Phone versus tablet has no system answer at all, so it uses the 600 dp rule the resource system itself uses for sw600dp.

Apple: one source, several products

platform is resolved at compile time, not from UIDevice, because it names the SDK the binary was built against and that is a fact about the binary. A tvOS build and an iOS build are different products from one source, and a benchmark table has to be able to tell them apart.

tvOS needed exactly two accommodations, both in the sources rather than in a fork of them:

• No battery API. Not a battery reading -1 — isBatteryMonitoringEnabled does not exist, because an Apple TV is mains-powered. The runner reports 100% and charging, which is what a permanently plugged-in device is.
• No WebKit. The web-shots:webkit route is compiled out under #if canImport(WebKit), so an Apple TV never declares the capability — and since the capability list is derived from the route list, the declaration and the code stay one act. A television advertising a screenshot workload it cannot run would take those jobs off the queue from the phones that can.

WebShotsManifest.swift stays in the tvOS target. Its header says it is kept free of UIKit and WebKit so its parsing can be checked without a device, and Protocol.swift refers to the type it defines — excluding it took the whole protocol down with it.

watchOS is not done. The agent, the protocol and the Core ML backend would compile, but watchOS has no UIKit, so UIDevice and UIApplication need guards throughout rather than the two tvOS needed — and the watchOS platform is not installed here, so none of it could be compiled anyway. It is the next Apple target, not a missing one.

The browser: the one with no install

http://<collector>/runner enrols the browser that opens it. A smart TV's built-in browser, a games console, a Chromebook, a Kindle, an iPad too old for the App Store, a friend's phone via a QR code — none of them will run a signed native agent, and all of them have a browser.

Read the backend name before reading a number. The browser runner does not declare benchmark:synthetic, and that is deliberate rather than an omission. The fleet's synthetic backend measures SHA-256 throughput; a browser's only native hash is crypto.subtle.digest, which is asynchronous, so a thousand sequential folds is a thousand promise dispatches and the scheduler is a large part of what gets measured. crypto.subtle is also a secure-context API, so on http://fleet-host.local:8788 — the address a TV browser would use — it is not there at all.

So there are two backends, each named for what it is:

Backend	What it measures	Available
jssha	a synchronous SHA-256 written out in the page; no dispatch overhead, so the rate is the JS engine's throughput and is comparable between browsers	always, including plain HTTP
webcrypto	the platform's native SHA-256, paying async dispatch per round	secure contexts only

Neither is comparable with a native runner's synthetic rate and neither pretends to be. What is comparable is correctness: both fold the identical block through identical rounds and both report synthetic_digest, so the conformance suite proves the browser is doing the fleet's arithmetic even though its clock means something different.

A hidden tab is throttled by every browser, so the agent stops asking for work when the page is hidden — and aborts an in-flight claim rather than waiting out the collector's 25-second long poll, because a job handed to a tab that went hidden in that window would be measured through the throttle. It registers with a TTL, so a tab that is closed leaves the shelf instead of sitting there forever as an offline device nobody can find.

Roku: the one that has never run

A Roku is the cheapest always-on ARM box in most houses, it is mains-powered, and it never moves. On paper it is an ideal shelf device. runner-roku/ is a SceneGraph channel that speaks the protocol: it registers, long-polls, runs the fold, beacons and reports.

None of it has been run. Roku has never shipped an emulator and BrightScript runs only on the device. What CI does have is brighterscript's bsc, the community compiler: runner-roku / brighterscript parses and validates every .brs and SceneGraph .xml file in the channel on each pull request that touches it, so syntax errors and calls to functions that do not exist are caught off the television. Whether the firmware's own compiler agrees, and whether any of it behaves as written, is still for the first sideload to find out. That is worth stating plainly here rather than only in its README, because this page exists so that a row's last column can be trusted.

It names its backend roevp, for the browser's reason

BrightScript has no hash of its own. The only one on the platform is roEVPDigest -- a native OpenSSL object -- called once per round from an interpreted for loop, and at 4 KiB a round the native hash is a small part of the work. The rest is the interpreter dispatching a method call, allocating the returned hex string, parsing it back into a byte array, and copying 32 bytes with another interpreted loop.

So the rate measures BrightScript's dispatch as much as it measures the SoC, which is the same shape of problem as crypto.subtle being asynchronous, and it gets the same answer:

Backend	What it measures	Comparable with
synthetic	SHA-256 throughput, natively	every native runner
jssha	a browser JS engine's throughput	other browsers
roevp	one native hash per interpreted round, so mostly the BrightScript interpreter	other Rokus

All three fold the identical block through identical rounds and all three report synthetic_digest, so correctness is shared across every one of them and only the clock differs. A job asking a Roku for backend: "synthetic" is refused with a sentence rather than served a roevp number under that name.

Three descriptor fields are null, and that is the interesting part

soc, arch and ram_mb are all null. Roku exposes no CPU model and no core count; it has shipped both MIPS and ARM devices and no API says which one you are on; and its only memory API is GetGeneralMemoryLevel(), which answers "normal", "low" or "critical".

That last one is the refusal worth naming. It would have been easy to map three words onto a number, and ram_mb is one of the two fields targets.match is most often written against -- so an invented 1024 would not sit harmlessly in a descriptor, it would decide which jobs reach the device and then appear in tables beside numbers that were measured. The pressure signal instead rides on every beacon as mem_pressure, which is a free-form string in the result schema and the honest home for it.

Battery reads 100 and charging reads true, the same accommodation tvOS needed and for the same reason: a Roku is mains-powered, so require_charging is genuinely satisfied rather than being unmeasurable.

Two things a Roku owner should know before trusting a number

It only runs in the foreground. Roku suspends a channel when the user presses Home -- there is no background execution model for a channel and no way to ask for one. A Roku on this fleet is a Roku dedicated to it, or one that joins between programmes. It registers without a ttl_s for that reason: a suspended channel is an offline device, not a departed one.

Whether the screensaver suspends it is not known. Roku's documentation does not say whether a channel with no video playing keeps executing behind the screensaver, is suspended, or is left running and throttled. All three are plausible, and the third would quietly produce slow benchmark numbers with nothing anywhere to say why. Nobody has watched a Roku long enough to find out. The test, for whoever has one, is whether the beacons keep arriving after the screensaver appears.

Discovery is the first driver that is not a cable

collector/src/drivers/roku.ts is the first driver in the registry that reaches devices over the network: an SSDP M-SEARCH for roku:ecp, then GET http://<ip>:8060/query/device-info. adb, simctl and devicectl all answer "what is plugged into this Mac?"; this one answers "what Rokus are on this LAN?"

The difference shows up in the shape. SSDP has no end-of-list, so discovery is a fixed listening window rather than a command that returns -- which means a Roku that was asleep or dropped a datagram is simply absent from one pass and present in the next, and there is no way to tell that from there being no more Rokus.

It deliberately does not install. That needs HTTP digest auth, which Node's fetch does not do, and hand-writing RFC 7616 against no hardware would produce something that looks finished and fails the first time somebody has a Roku. runner-roku/build.sh installs with curl --digest instead, reading the developer password from the executor host's Keychain -- never from a job spec, which anyone who can enqueue can read, and which the dashboard renders.

What a cloud runner actually reports

The machine agent's CI matrix prints the descriptor each platform would send rather than asserting on it, because a cloud runner is allowed to be mostly nulls and pinning a field would fail the day a hosted image changes. What it printed on its first run is worth writing down, because the two Linux and Windows answers are nothing like each other.

arm64 Linux — near enough complete:

{ "model": "Virtual Machine", "soc": null, "ram_mb": 15947,
  "os": "linux-ubuntu-24.04", "platform": "linux", "kind": "ci",
  "arch": "arm64", "cpu_cores": 4 }

Only soc is missing: /proc/cpuinfo on arm64 has no model name line, which is the field the probe reads. A Raspberry Pi should register cleanly, and would also fill model from the device tree.

Windows — almost nothing:

{ "model": null, "soc": null, "ram_mb": null, "os": null,
  "platform": "windows", "kind": "ci", "arch": "x64",
  "gpu": null, "vram_mb": null, "cpu_cores": 4 }

Every field there came from wmic, and wmic is a removed feature on current Windows, not merely deprecated. Only the fields Node answers by itself survived.

The agent still registered and still ran work — the probes degrade to nulls exactly as designed, and nothing crashed. But os being null meant a targets.match expression could never select a Windows machine by its OS, and ram_mb being null meant it could not be selected by memory either, which left device_id, platform and arch as the only handles on a Windows agent.

That is the finding the platform matrix was added to produce, and it is the reason it prints the descriptor rather than asserting on it.

What was done about it

The Windows probes now lead with PowerShell Get-CimInstance and keep wmic as the fallback for older installs, rather than the other way round — one PowerShell process running six queries, because its startup is the expensive part and this sits on the path to registration:

Field	Class
model, ram_mb	Win32_ComputerSystem
soc	Win32_Processor
os	Win32_OperatingSystem
gpu, vram_mb	Win32_VideoController
kind	Win32_SystemEnclosure, then Win32_Battery

The same runner now reports:

{ "model": "Virtual Machine", "soc": "AMD EPYC 9V74 80-Core Processor",
  "ram_mb": 16379, "os": "windows-10.0.26100",
  "platform": "windows", "kind": "ci", "arch": "x64",
  "gpu": "Microsoft Hyper-V Video", "vram_mb": null, "cpu_cores": 4 }

os and ram_mb answer, so a Windows machine can be selected by its OS and by its memory. A hosted runner is a VM, so model reads Virtual Machine — that is a real answer, not a failure.

Two fields are null on purpose. AdapterRAM is a uint32, so every card with 4 GB or more reports the same saturated ceiling; reporting that as 4095 would let a match expression asking for 16000 skip the 24 GB machine that could have run the job, so the field says nothing instead — and the Hyper-V synthetic adapter above has no dedicated memory to report either way. And kind still reads ci rather than the chassis answer: a container's hardware belongs to somebody else, and "this will be gone in four minutes" is the more useful fact.

Adding one

1. Write the agent. It registers, long-polls, runs, reports. There is no SDK and no shared library, on purpose — five implementations in five languages keep the protocol honest in a way one library never could.
2. Declare platform, kind and capabilities at registration. Declare only what the agent can actually run: a capability it cannot honour takes a job off the queue from a device that could have run it.
3. Run the conformance suite: npm run conformance -- --device <id>. Nine clauses, each of them something that has actually gone wrong here. A skip is fine; a FAIL is a bug in the agent. See Writing a runner.
4. Add a row to the table at the top of this page, and put the honest answer in the last column.