Performance targets are measured on Android tablets, and several sessions share them. This guide lists the devices, how to reserve one, and how to test on it without disturbing anyone else. Shared machine resources are in environment.
| Name | Model | Role | GPU |
|---|---|---|---|
tcl |
TCL TAB 11 Gen 2 (9465X) | Low tier reference | Mali-G52 MC2 |
movinkpad11 |
Wacom MovinkPad 11 (DTHA116) | Mid tier reference | Mali-G57 MC2 |
movinkpad14 |
Wacom MovinkPad Pro 14 (DTHA140) | Top tier reference | Adreno 735 |
huion |
Huion Kamvas Pad 12 (KP1202) | Pen input testing | Mali-G57 |
xppen |
XP-Pen Magic Note Pad (MNP1095) | Pen input testing | Mali-G57 |
sm_t878u |
Samsung Galaxy Tab S7 (SM-T878U) | Android 13 without shared presentation | Adreno 650 |
Tier hardware has the full specifications. Device quirks that affect tests:
- The MovinkPad 11 panel presents at 60 Hz unless the app is switched on in
Settings › Display › Adaptive; only listed apps get 90 Hz. Its EGL has no
buffer_age, so the Android UI redraws every frame in full. - The TCL needs about 18 s to start from a cold shader cache; waits shorter than that time out on the first run after an install.
- The XP-Pen's
libguithrottles FIFO presents on an older fence than AOSP and crashes when shared and FIFO presentation alternate. The Android host adds one frame of navigation latency on MediaTek Android 14 devices to avoid it; a navigation crash seen only on the XP-Pen points there first. - The Huion launcher force-stops an app it sees updated; wait a few seconds after
adb installbefore starting instrumentation. - The Galaxy Tab S7 runs Android 13, whose Vulkan loader has no swapchain maintenance, so ink uses FIFO presentation. Its factory Chrome 111 has no WebGPU.
- An install can stall on a Play Protect dialog. These are test devices:
adb shell settings put global verifier_verify_adb_installs 0turns the check off.
A tablet serves one session at a time. tools/devices/devices.py
records reservations and serializes commands:
tools/devices/devices.py list
tools/devices/devices.py reserve --tier mid --note "canvas bar frame cost"
tools/devices/devices.py reserve movinkpad14 --hours 8
tools/devices/devices.py run movinkpad11 -- adb install -r app-debug.apk
tools/devices/devices.py release- Reserve before use and release when done. A reservation names its owner,
the worktree directory by default, and expires after
--hours(4 by default). Subagents working in the same worktree share it. Release a disconnected tablet by its serial. - Pick by tier, not by name, unless you were assigned a tablet. If the user
assigns one, use only that one. If they say a tablet belongs to someone else,
leave it alone even when
listshows it free. - Run every device command through
run. It holds the tablet's lock (/tmp/capy-<serial>.lock, the same fileflockusers take), waits while another command runs, and requires a current reservation by this owner. Inside the command,adbtargets the tablet throughANDROID_SERIAL, andCAPY_ANDROID_SERIAL,CAPY_APPLICATION_ID,CAPY_WEB_PORTandCAPY_CDP_PORTare set. - Build outside
run; hold the lock only while the device is in use.
For user-authorized file collection, run --read-recordings <device> -- list
lists saved .capystrokes files in Download and Documents. Use
run --read-recordings <device> -- pull <remote.capystrokes> <local> to copy one.
This restricted mode accepts no arbitrary commands, leaves reservations and
apps alone, and requires no reservation because it only reads shared files.
-
Install experiments and test builds under your own application ID:
-PcapyApplicationId=$(tools/devices/devices.py appid) -PcapyAppLabel=<label>. The test package is<id>.test, anddevices.py runexports the same ID as$CAPY_APPLICATION_ID.apps/layer-android/run.shinstallsart.capycanvas.devand is for single-user work on an emulator or your own device;run.sh testuses your own ID. -
Never uninstall
art.capycanvasorart.capycanvas.editoror clear their data; they hold an artist's drawings. Never touch another session's application IDs. -
A
connectedDebugAndroidTestrun uninstalls the tested app afterwards, deleting its data. Always give it your own-PcapyApplicationId, asrun.sh testdoes. -
Before each run, check that none of your app's processes are still running. At the end, uninstall your own IDs and remove your own port forwards (
adb forward --remove tcp:$CAPY_CDP_PORT,adb reverse --remove tcp:$CAPY_WEB_PORT). -
Chrome on a tablet uses the device's real profile. Use your own origin and port, open your own tab, and never drive a tab you did not open:
LAYER_WEB_PORT=$CAPY_WEB_PORT ./apps/layer-web/run.sh tools/devices/devices.py run movinkpad11 -- sh -c 'adb reverse tcp:$CAPY_WEB_PORT tcp:$CAPY_WEB_PORT && adb forward tcp:$CAPY_CDP_PORT localabstract:chrome_devtools_remote && adb shell am start -a android.intent.action.VIEW -d http://127.0.0.1:$CAPY_WEB_PORT/ -p com.android.chrome'
curl --max-time 5 http://127.0.0.1:$CAPY_CDP_PORT/json/listlists the tabs. The Web guide runs the device journeys from there.
- VMs.
tools/windows-vm/windows-vm.pyruns one Windows 11 VM per worktree, named after the worktree. Commands that use a VM lock it; pass--vmto use another. Default VMs render with WARP. An optional hardware GPU pool reserves one card per VM for renderer tests and hardware UI journeys. Performance qualification still requires the specified reference hardware and measurement rules. - Physical PCs. The Windows 11 Home laptop has an Intel Core i7-1255U, Iris Xe and a 60 Hz panel. The Surface Pro, 11th Edition has a Snapdragon X Plus X1P64100, Adreno X1-85 and 16 GB RAM for native ARM64/D3D12 checks. Windows Home has no Sandbox for clean package installs. These PCs are not reachable from the Linux machine; work that needs them runs on the assigned PC, using private test desktops and profiles. Desktop diagnostics do not replace reference-tier or physical pen measurements.
Apple builds and tests need an Apple Silicon Mac with Xcode; there is no remote
Mac. The reference hardware is an M2 Pro Mac mini, whose display presents at
90 Hz, and a 13-inch M4 iPad Pro at 120 Hz. On Linux, only the layer-apple
Rust bridge tests run.
Use --platform apple with the same reservation tool. It discovers paired
physical iPads through Xcode's devicectl and shares the reservation and command
locks used by the Android runner. When several iPads are paired, select one by
its serial. Partial hardware metadata can omit reality; discovery accepts that
case only when isProductionFused is true. Explicit simulated devices remain
excluded.
DEVELOPER_DIR=/Applications/Xcode.app/Contents/Developer tools/devices/devices.py --platform apple list
CAPY_TEST_IPAD=serial-from-list
DEVELOPER_DIR=/Applications/Xcode.app/Contents/Developer tools/devices/devices.py --platform apple reserve "$CAPY_TEST_IPAD"
DEVELOPER_DIR=/Applications/Xcode.app/Contents/Developer tools/devices/devices.py --platform apple run "$CAPY_TEST_IPAD" -- \
sh -c 'xcrun devicectl device info details --device "$CAPY_APPLE_DEVICE_ID"'
DEVELOPER_DIR=/Applications/Xcode.app/Contents/Developer tools/devices/devices.py --platform apple release "$CAPY_TEST_IPAD"Apple commands require a current reservation. The runner exports
CAPY_APPLE_DEVICE_ID and the owner's isolated CAPY_APPLE_BUNDLE_ID; pass the
latter as an xcodebuild build setting when building the app.
- The iPad must be awake and unlocked, with Settings > Developer > Enable UI Automation on.
- Never replace or clear an installed
art.capycanvas.CapyCanvas,art.capycanvas.CapyCanvas.devor olderart.capycanvas.apple.ipad/.macapp; they may hold an artist's drawings. Build test and benchmark apps under your ownCAPY_APPLE_BUNDLE_IDand give test launches their ownCAPY_STORAGE_DIR(Apple guide). A free Personal Team profile limits how many apps a device may hold, and the XCTest runner counts as one. - There is no iPad hardware keyboard, second Mac display, 120 Hz Mac display or working iCloud Drive account. Leave those checks unverified; don't ask for them.
- Keep device UDIDs, team IDs and signing details out of commits.