Recipes

Seven verification plans, written in plain language, that tell the agent what to open, what to press, and what counts as a pass.

A recipe is a Markdown file in test/artemis/recipes/. The runner sends its whole text to ARTEMIS as the task, so the recipe is the test: there is no separate script behind it.

The seven

#	Recipe	Run with	Hooks	Passes when	Logcat audit
01	Silicon telemetry	silicon	useCPU, useGPU, useMemory, useADPF	Each reports source === 'hardware', CPU load and frequency are not null, thermal headroom is in range.	Yes
02	Actuators and haptics	actuators	useHaptics, useTorch, useHiLight	The torch toggles, haptic presets fire without a native error, HiLight reports its status correctly.	Yes
03	Sensors and capture	sensors	useSensors, useLocation, useCamera	Motion sensors stream live, the camera viewfinder renders without dropping frames.	No
04	On-device AI	ai	useGeminiNano	AICore loads Gemini Nano and streams tokens to the screen without exceeding the context limit or raising memory warnings.	Yes
05	Full sanity suite	full	Every hook the template exercises	Traverses all four tabs without a native FATAL.	Yes
06	Pixel 11 Pro extensions	hardware	usePerfetto, useCameraExtensions, useSpatialAudio, useChannelSounding, usePlayIntegrity, useHealthConnect, useAppFunctions	Each feature responds on the device.	Yes
07	Next-generation hardware	nextgen	useAltimeter, useMicrophoneArray, useThermometer, useBatteryShare, useChargingIntelligence, useADPFHintSession, useWifi7MLO, useWifiRTT, useSatelliteNTN, usePrivateSpace, useKeyAgreement, useEmbeddings	Each reports hardware or derived where the silicon exists, and unavailable with null readings where it does not.	Yes

Recipe 07 cannot pass yet. It drives screens for its twelve hooks, and the template has not built them. It will start meaning something when those screens exist.

The Logcat audit is a step the agent performs because the recipe asks for it — the runner does not watch the log. Recipe 05’s version is typical:

adb logcat -d *:E | grep -i "pixelkit"

Writing a recipe

1. Add a Markdown file to test/artemis/recipes/, numbered after the last one.
2. Say what to open and what to press in the order a person would, naming screens and controls exactly as the template labels them.
3. Write every checkpoint as something observable — “source reads hardware”, “the value is not null” — never “works correctly”. A checkpoint the agent cannot see, it cannot verify.
4. Treat unavailable as a pass where the hardware is absent, and a plausible-looking number where it is absent as a failure. That is the contract being tested.
5. End with a Logcat audit, unless the recipe genuinely cannot crash native code.
6. Give it a name in the RECIPES table in scripts/run-artemis-e2e.js, or run it by path.

A recipe can only reach screens the template has. If it needs one that does not exist, build the screen first — otherwise it fails for a reason that has nothing to do with the hook.