Skip to content

Features in detail ​

How each part of Tetsu OS was built, measured and tested, moved out of CLAUDE.md on 2026-11-08 to keep that file short. Read the part you work on. Notes on who confirmed what are as of the date given; CLAUDE.md keeps the current list of what still needs checking.

Working on the test unit, confirmed by the user:

  • Boots to the launcher

  • Power button starts the device, and a short press shuts it down

  • USB-C root console

  • Boot log saved at about 15 and 60 seconds after init starts

  • The button test shows each control as it is pressed: filled while held, outlined once seen, with the last press named below. An unlisted key shows as KEY <code>.

  • Audio: speakers and headphones play, and outputs switch automatically with the jack, including during playback and when booting with headphones plugged in

  • USB console connects without Login incorrect, using getty -w, so the login prompt appears only after pressing Enter

  • GPU: OpenGL ES 3.1 through Panfrost on the Mali-G31. kmscube ran at 59.94 fps with no visible tearing. The GPU runs at a fixed 432 MHz, because there is no clock scaling table yet

  • The launcher lists games from /mnt/data/roms and the folders inside it, in any arrangement, as the file extension decides the console; hidden files and folders are left out. It has the button test as a menu entry and working volume buttons

  • A libretro frontend in the same app, in libretro/, runs the mGBA core from package/libretro-mgba. Pokemon FireRed runs at 59.7 fps with clean sound and no tearing, at about 50 C. Holding Menu + Start leaves a game. /var/log/launcher.log has timings

  • Games and saves are in roms and saves on the TETSU-DATA partition. They survive leaving a game, rebooting and reflashing with flash.sh

  • CPU speed scaling, up to 1512 MHz on this chip. It was fixed at 1008 MHz before, which gave only 43 fps in the game

  • Tamagotchi P1 emulation through package/libretro-tamalib, using tama.b in roms, at 60 fps with the CPU at its lowest 480 MHz. The launcher picks a core by file extension. The pet's whole state is saved once a minute, on leaving and at power-off, in saves/tama.state, and it pauses while the game is off

  • Lid sleep, using s2idle because TF-A has no deep sleep, from the menu and in games. Sound stops in sleep (patch 0011) and resumes on opening. A power button press with the lid closed doesn't shut down (patch 0010), and the device goes back to sleep. The Tamagotchi carries on correctly after sleep

  • Power-off after a long lid sleep (2026-11-10): after 2 hours of lid sleep on battery (Settings > SYSTEM > POWER OFF AFTER SLEEP: 1, 2 (default), 4 or 8 HOURS or NEVER, saved as sleep_power_off=), an RTC alarm (/sys/class/rtc/rtc0/wakealarm, set to the time left before each freeze) wakes the device, and it powers off, so s2idle doesn't drain the battery. A game in progress saved its resume state when the lid closed, and saves again on the shutdown's SIGTERM; playing.txt stays, so it resumes at the next start. On USB power (status other than Discharging) it keeps sleeping for that long again instead, as the battery doesn't drain and USB power may start the device again. The time counts with CLOCK_BOOTTIME, as Instant stops in s2idle. Waking for a power button press doesn't reset it. NEVER sets no alarm. LAUNCHER_LONG_SLEEP overrides the setting in seconds, for testing. Claude checked on the device that a +40 alarm wakes s2idle, that the boot clock advanced 41.7 s across it and that the alarm clears once it fires. The user confirmed on 2026-11-10, with LAUNCHER_LONG_SLEEP=120 and the cable unplugged, that the device powered off with the lid closed during a game and that the game resumed at the next start. The setting is the ninth row of SYSTEM, the only group with that many (MAX_GROUP_SETTINGS), which leaves room for one note line: its note about the power button. Claude checked on the device that the row shows 2 HOURS, Right saves 4h, and the note fits under the rows with SLEEP WHEN IDLE on; a sleep with a setting other than 2 hours hasn't been tried

  • Shutdown screen (2026-11-10): a power symbol in the accent colour above POWERING OFF (SCHALTET AUS), or REBOOTING or BATTERY EMPTY, stays on the screen until the device is off (screens::shutdown_screen, preview with --preview-shutdown). The power button's acpid script (/etc/acpi/PWRF/00000080) sends SIGUSR1 to every launcher process, waits 0.2 s and runs poweroff as before. The menus' process and a game's both listen (app::show_shutdown_on_signal, a signal_hook thread), as either may be drawing; each draws the screen with Framebuffer::draw_last, after which fb.rs leaves out every write of that process, so a game's next frame or a menu redraw can't replace it. A game stops running frames until the shutdown's SIGTERM, which saves its resume state as before, so it resumes at the next start. The framebuffer keeps its picture after the launcher stops: Claude checked on the device that the menu still showed 2 s after S90launcher stop, that killall -USR1 launcher showed the screen, and that a press afterwards left it as it was. Not yet tried: a real power button press, one during a game, and the menu's POWER OFF and REBOOT with the new screen

  • The battery level in the menu

  • Updating over USB. The gadget is serial plus CDC Ethernet, set up through configfs by /usr/sbin/usb-gadget start from S41usbnet, with the fixed MACs and the USB ID (0525:a4aa) that the g_cdc gadget used before. S41usbnet gives the device 10.35.35.1, and udhcpd gives the PC 10.35.35.2 with no router option. Over USB, dropbear allows root's empty password (-B); Settings > ADVANCED > SSH chooses where it listens (see below). scripts/push.sh {app|rootfs|boot|all} copies output over SSH with rsync

  • USB STORAGE, in Settings > DATA, shows the card's TETSU-DATA partition to the PC as a USB drive, so games can be copied without taking the card out. A drive is built in on Windows, macOS and Linux, unlike MTP, and needs no program on the device. usb-gadget storage syncs and unmounts /mnt/data, gives the partition to the kernel's mass storage function as a removable drive, and binds Ethernet plus the drive under another ID (0525:a4a5), because Windows remembers a device's interfaces by its ID. The serial port is left out for the endpoints, so SSH keeps working. The PC sees the partition as a disk with no partition table, which all three read. Ejecting the drive on the PC empties the function's lun.0/file, which the screen checks twice a second, and leaves the screen. B also leaves, and asks again while the PC still has the drive. Leaving runs usb-gadget network: it closes the drive with forced_eject, binds serial plus Ethernet again, repairs the partition with fsck.fat -a (for a PC unplugged without ejecting), mounts it and the launcher reads the settings and games again. Meanwhile the device doesn't sleep: closing the lid only turns the screen off. Volume and brightness don't change, as they are saved on the card. An empty battery leaves the screen, mounts the card and powers off. Writes need patch 0014 (see the MUSB fact below). Mounting, copying and deleting games and ejecting were checked by the user on Linux on 2026-11-08; Windows and macOS haven't been tried

  • TETSU-DATA is repaired at boot after an unclean shutdown (2026-11-12). S15data reads the FAT dirty bit (boot sector byte 65 on FAT32, 37 on FAT16) and runs fsck.fat -a -w only when it is set, logging to /var/log/fsck-data.log: 1.2 s on the 64 GB card, nothing on a clean boot. Without it, one power cut left directory entries pointing at clusters the FAT still had free (the card kept an older FAT sector, though the launcher fsyncs before renaming), and the kernel, on reading such a file, made the partition read-only: from then on settings, resume points and play time were silently not saved, which looked like resume and BOOT JINGLE OFF breaking. The kernel never clears a dirty bit it found at mount, so the warning repeats until a repair. A clean reboot during a game was checked to leave the partition clean and resume. Should TETSU-DATA still go read-only, the menus say so (CARD ERROR): storage::data_read_only reads /proc/mounts with the battery each second, and before a game resumes at start. A restarts, which repairs it; B goes on, and the warning comes back after each game. Checked on the device with mount -o remount,ro /mnt/data: shown within a second, B back to the menus; A (a plain reboot) and the German wording weren't tried there

  • A second card in TF2 (2026-11-11) adds games: those in its roms folder join the lists with TETSU-DATA's, unmarked, and a game at the same path on both cards shows once, from TETSU-DATA (library::find_all_roms). Their saves, states, covers, play time and settings stay on TETSU-DATA, as they are named after the game, so taking the card out loses nothing and a game moved between the cards keeps them. The card can be FAT32 or exFAT (CONFIG_EXFAT_FS), as most cards over 32 GB come; a Tetsu OS card's TETSU-DATA is used if it has one, or else its first FAT32 or exFAT file system, or the whole card without a partition table. The kernel side is patch 0018: mmc2, card detect on PE22 and a 3.3 V supply on PE4 that stays on, from ROCKNIX, plus mmc aliases that keep TF1 at mmcblk0 and TF2 at mmcblk2. /usr/sbin/cards knows a slot by its controller (4020000.mmc, 4022000.mmc) in the card's sysfs path, not its number. S15data finds TETSU-DATA on TF1 only, so a second Tetsu card is never taken for it, and runs cards mount, which mounts TF2's games at /mnt/tf2 with umask=022, waiting up to 2 s for the partitions. The launcher checks the slot once a second in its menus' loop, by the card's CID in /sys/block/mmcblk*/device/cid, which also tells a card swapped during sleep: a card going in is mounted, one coming out is unmounted (lazily if a file was open), and the lists are read again, with a sign at the top for 3 s: TF2: 12 GAMES, TF2: CAN'T READ THE CARD or TF2: CARD TAKEN OUT. A card that wouldn't mount is tried again only once taken out. The check also runs once before the games are first read, for a card the kernel found after S15data. Taking the card out during a game isn't guarded against. System info's GAME STORAGE shows both cards' free space while one is in, without the totals, for room. USB STORAGE gives the card in TF2 to the PC as a second drive, the whole card, so the PC can also format it: lun.1 of the same mass storage function, so no endpoint more. With no card there, the PC sees an empty card reader. The screen leaves once both are ejected and says EJECT THE DRIVES; going back runs partprobe on the card, in case the PC partitioned it anew, cards check (fsck.fat -a or exfatprogs' fsck.exfat -y) and cards mount. WI-FI TRANSFER lists both cards' games; a game on TF2 is named TF2: and its path, which no FAT32 or exFAT file name can start with. With a card in TF2, the page shows both cards' free space and a SEND TO choice (TF1 or TF2, TF2 by default, kept in the browser), and uploads send card=tf2; a card gone since is refused as no-card. Checked on the PC: cargo test (the merge, the slot by controller and CID, the mount lookup), clippy, the device tree compiled with dtc, the patch with apply-patches.sh, and previews of the USB STORAGE screen and the sign in both languages

  • Charging screen (2026-11-11): plugging in USB power starts the device, and when that alone started it, a battery with the charge level shows instead of the menus. The AXP717 keeps what turned it on in REG 0x20 (PWRON status, AXP717 datasheet V1.2, 6.15.2.19): bit 0 the power button, bit 2 VBUS insert, bit 1 the IRQ pin, bit 3 a charged battery, bit 4 a battery insert; a software reboot clears it. REG 0x21 keeps what turned it off (bit 1 software, as poweroff does). Measured on the device: 04 after plugging in, 01 after the power button, 00 after a reboot. Whether VBUS starts the PMIC is set in its efuse, so it can't be turned off. S05charging mounts debugfs, reads the register from /sys/kernel/debug/regmap/0-0034/registers (it isn't volatile, and nothing reads it before, so the first read reaches the chip) and writes /run/charging when bit 2 is set without bit 0. Then S21jingle stays quiet, and the launcher (charging.rs), after setting the charge limit and before anything else, shows screens::charging_screen until the file is gone: the battery outline with an accent fill to the level, the percentage, CHARGING or CHARGED (status other than Charging, as at CHARGE TO's limit) and PRESS POWER TO START. The backlight turns off after 10 s (bl_power); any button or opening the lid lights it for another 10 s. The power button's script (/etc/acpi/PWRF/00000080) removes the file instead of powering off, also during the boot logo before the launcher runs, and the launcher, checking every 100 ms, goes on to the menus as after a power-button start (the setup, resuming a game), without the jingle. If axp20x-usb/online reads 0 for 2 s, it shows POWERING OFF and powers off. The rest of the boot (Wi-Fi, SSH, USB networking) runs as usual; the launcher idles at 0% CPU on the screen. launcher --preview-charging FILE.ppm renders it. Claude checked on the device with a hand-made /run/charging that it draws, darkens after 10 s, lights again on B, and that the power button's script starts the menus. The user confirmed on 2026-11-11 a real start from the cable with no jingle, the battery screen, the screen going dark and lighting on a button, the power button starting the menus, and that unplugging powered it off and it stayed off. A power button press during the boot logo of a cable start went straight to the menus (the log shows the screen skipped). Powering off with the cable in stays off, as VBUS starts the PMIC only when it rises

  • The mini-HDMI port (2026-11-11) mirrors the screen at 640x480: a TV or monitor shows the same picture as the panel, menus, boot logo and games alike, and scales it itself. Patch 0001 already had the HDMI controller (sun8i-dw-hdmi, DesignWare HDMI 2.0 TX), its H616 PHY, TCON TV0 and the connector, and the launcher draws only into /dev/fb0, so the kernel's fbdev emulation does the mirroring: on a hotplug, drm_fb_helper_hotplug_event probes the connector with the 640x480 buffer as the size limit and drives HDMI from crtc-1 (mixer 1, TCON TV0) with the same framebuffer the panel's crtc-0 scans. The only change is video=HDMI-A-1:640x480@60 in extlinux.conf: with a TV in at boot, the helper otherwise sizes the buffer for the largest mode, such as 1920x1080, and the launcher would fill its top-left corner. The mode is CEA 640x480p59.94 (25.175 MHz), which every HDMI sink must take; if the EDID lacks it, the kernel adds it. FBIO_WAITFORVSYNC waits on CRTC 0, the panel, so games keep the panel's timing, and the TV, at the same nominal rate but its own clock, may tear. Sound stays on the speaker, headphones or Bluetooth: HDMI audio needs the H616's audio hub (AHUB), which mainline doesn't have. Closing the lid still sleeps, which turns HDMI off with the display driver. Turning the panel off (sleep, USB STORAGE, WI-FI TRANSFER) is only its backlight, so the TV keeps showing those screens. A risk to check with a real TV: the HDMI tmds clock comes from pll-video2 at 297 MHz, so for 640x480 it is 24.75 MHz (297 / 12), 1.7% under the 25.17 MHz TCON TV0 gives from pll-video1. Modes at 74.25 and 148.5 MHz divide exactly; this one doesn't. If the TV shows nothing or a broken picture, look there first. Checked by Claude over SSH, without a TV: echo on > /sys/class/drm/card1-HDMI-A-1/status forced a hotplug while the launcher ran, and the DRM state (/sys/kernel/debug/dri/1/state) showed crtc-1 at 640x480 with plane-4 on the panel's framebuffer, and no kernel errors; detect put it back. A boot with video=HDMI-A-1:640x480@60e (forced on) kept fb0 at 640x480 with both CRTCs on it, and the launcher drew at 3.95 s, as before

  • WI-FI TRANSFER, in Settings > DATA and as SEND GAMES OVER WI-FI in the setup's games step (2026-11-10): a web page for a phone or PC browser on the same network, for getting games on without a cable or card reader. Off unless its screen is open: transfer::wifi_transfer starts transfer_server::Server (the tiny_http crate) on port 80 of every address, so also over USB at 10.35.35.1, and stopping it stops uploads under way and waits for each request's thread. The screen shows the address, a six-digit code that is new each time (/dev/urandom, bytes from 250 up skipped), and a QR code of http://<address>/?code=<code> (the qrcode crate, version 3 at 5 pixels a module, dark on a white square with the 4-module border scanners need), then what is arriving with its percentage and what came of the last request. Not connected, it says so and checks again every 2 s.

    • Access: the code, from the QR address or the page's form, gets a random 128-bit session cookie (HttpOnly, SameSite=Lax) and a redirect that drops the code from the address; sessions end with the screen. Wrong codes are answered one at a time, each after a second, so guessing would take days on average. Every change also needs an X-Tetsu header, which a page on another site can't send without a preflight the server never answers. Plain HTTP: someone watching the network could read the cookie
    • The page (transfer_page.html, built in with include_str!) follows the device's language (English or German, its own table) and builds everything with DOM calls, so no file name becomes HTML. Games upload one at a time with progress: a game goes to its console by extension (.bin, .img, .iso, .wav, .sub and .sbi with the PlayStation, for a .cue), into the top-level folder in roms with most of that console's games, or a new one named after its key (GBA). A ZIP waits in roms until the game inside names its console. Uploads go to a hidden .upload-<name> beside the destination, are synced and renamed into place, and leftovers are removed when the server starts. The card keeps 16 MiB free, and FAT32's 4 GiB is the largest file. Per game: COVER takes any image the browser opens, shrunk with fit_size's rule to at most 200 by 300 and sent as the PPM the launcher reads; covers show on the page as BMP. SAVE replaces the game's save at save_path (at most 64 MiB) and removes its .resume, which would load the old save over it; the page asks first if there is one. DELETE GAME runs delete_game, as Y does. Buttons that send say so with ⤒ and change with what is there (ADD COVER or REPLACE COVER, UPLOAD SAVE or REPLACE SAVE); downloads are links with ⤓ and the file's name. The header and the browser's tab show the logo, which /logo.svg draws from icons::LOGO, a square per pixel. Saves and screenshots download (only single names in saves and screenshots), and a search filters the games. Leaving reads the games again if anything changed
    • B leaves, and asks again while an upload runs; that upload is then refused as incomplete and its file removed. The lid only turns the screen off, so uploads go on, as in USB STORAGE; an empty battery leaves and powers off
    • Claude checked on the device on 2026-11-10, from the PC over the device's Wi-Fi: the QR code decoded from a screenshot, the login and wrong codes (1.0 s), refusals (no cookie, no header, a .txt, a path as a name, a 400x400 cover, a PNG as a cover, .. and other folders for downloads), a .gb into GB, a ZIP's .gba into GBA, replacing, a cover sent and read back pixel for pixel, a save downloaded unchanged and removing the .resume, deleting with save and cover, 100 MB at 6.5 MB/s with the same MD5 on the card, the screen during an upload, B twice stopping one, and port 80 closed after leaving. In a headless Chrome, the page's own code: the QR address login, the wrong-code page, the queue with a refused file, a 600x900 JPEG cover shrunk to 200x300, search and deleting. Not checked: a real phone, the German page, a PlayStation .cue with its .bin, a file over 4 GiB, the lid during an upload
  • USB stays connected through sleep, so SSH works after opening the lid (patch 0012). Before, the device dropped off the bus until the cable was re-plugged

  • The in-game menu (pause.rs, drawn by canvas::pause_screen). Tapping Menu alone pauses, and Menu + Start still quits directly. Resume, 8 save state slots per game (saves/<game>.state1 to .state8), volume, brightness and quit. A game with battery memory also saves saves/<game>.resume on leaving and loads it at the next start, so in-game saves made after a slot are never rolled back. Volume and brightness are saved in /mnt/data/settings.txt and restored by the launcher. Claude tested every item on the device over SSH with injected presses and screenshots: saving, loading, levels, quitting, resuming and Menu + Start. The user tried it with FireRed on 2026-11-08 and confirmed it feels right

  • QUIT's START OVER (2026-11-11). Left and Right on QUIT choose KEEP PLACE (as before) or START OVER, which quits without the .resume: the save file is still written, then the .resume is removed (save::forget_place), so the next start begins from the game's own save, as a real console does after power-off. State slots stay. Each visit to the menu starts on KEEP PLACE, so A twice by accident still keeps the place. A game whose save is its whole state (Persistence::FullState, Tamagotchi) has no .resume, so its QUIT shows no choice. Menu + Start, lid sleep, power-off and SWITCH GAME keep the place. Tested on the PC (pause.rs tests and --preview-pause in both languages); the user checked it on the device on 2026-11-11

  • A stale .resume is skipped (2026-11-12). Leaving writes the save file first and the .resume after, so a .resume older than the save file (save::resume_is_stale, by modification time) means the game ended without leaving: the reset button (hardware, unseen by Linux), a power cut or a crash, after the game saved in the meantime. A state holds the game's save memory, so loading it would bring the older save back and the 2-second autosave would write that over the newer save file. The .resume is removed and the game starts from its save file instead. A save sent by USB STORAGE is covered too. Checked on the device: a .resume dated before the .srm was skipped and removed; a normal stop and start still resumed

  • The in-game menu's lists (2026-11-11), as fifteen rows had filled the panel to its edges. The main list has RESUME, SAVE STATE, LOAD STATE, CHANGE DISC, SCREENSHOT, DISPLAY, OPTIONS, ACHIEVEMENTS and QUIT, at most 9 rows (pause::MOST_ROWS) of 32 lines instead of 22. DISPLAY opens a list with BRIGHTNESS, SCALING, SCREEN EFFECT, LCD GHOSTING and BORDER, and OPTIONS one with CONTROLS, CORE OPTIONS and CHEATS (pause::List); the title names the list. Rows that open a list or page show >. B goes back from a list to the row that opened it, and from the main list resumes; Menu resumes from any list. Opening the menu again starts on RESUME in the main list. A page that CONTROLS, CORE OPTIONS or CHEATS opens goes back to OPTIONS. The hint stays below where the longest list ends, so it doesn't move from list to list. VOLUME left the menu: the volume buttons do the same and show their sign over it. Checked with cargo test and --preview-pause in English and German; not yet on the device

  • The game switcher (2026-11-11, switcher.rs): SWITCH GAME in the in-game menu's main list, just above QUIT, opens the RECENT games other than the one running, newest first (up to 7), leaving out any whose file is gone. It is left out of the menu when there is none, so the main list has at most 10 rows (pause::MOST_ROWS). A switches: the game saves as when quitting (resume state, play time, achievements), shows SWITCHING TO with the next game's console and name, writes that game to playing.txt and exits with code 4 (switcher::SWITCH_EXIT). The launcher's play_game reads it back as Ending::Switch, and act starts that game straight away, which resumes from its own .resume and moves to the front of RECENT, so the game just left is first in its switcher and A, A goes back. A power-off between the two resumes the chosen game at boot, and a chosen game that can't start is forgotten there. After the last game quits, the list selects the game played last, if it is in that list. The menu passes the game's path as the lists have it (GameSpec::listed), which a ZIP's unpacked file isn't. Settings > IN GAME > GAME SWITCHER (switcher=, settings::SwitcherStyle) chooses how it looks: LIST (the default), rows like the menu's with the console on the right, Up and Down wrapping; or COVERS, the selected cover large in a frame of the theme's colour (fit to 200x230) with two smaller ones (88x116) each side, the name, console and place (2 / 7) below, Left and Right stopping at the ends. Covers are read and shrunk in the game's process when the page opens (BoxArt, so a screenshot stands in as in the lists, and a game without either shows its console's pictogram on a card). B or Menu goes back to SWITCH GAME. Menu + Select during play (pause::SWITCHER_BUTTON, pause::Opening) opens the page straight from the game, where B or Menu resumes and the hint says so; with no other game it does nothing. Like Menu + Start, the game doesn't see the Select: the page opens before the next frame, and the game continues only once every button is let go. --preview-switcher and --preview-switcher-covers render both. Claude checked on the device with injected presses and screenshots on 2026-11-11: both styles, Super Mario Bros. to Super Mario World and back several times, each continuing exactly where it was left (1-2 with the same score and coins), playing.txt and RECENT after each switch and quit, B on the page, and the selection after quitting. Menu + Select, injected in one write: it opened the covers page over Super Mario Bros., B went straight back into the game, and it switched to Super Mario World; a Menu tap still opens the menu. One switched start of Super Mario World logged a single 451 ms underrun, a normal start and two more switches none. Not checked: the feel, the SWITCHING TO screen by eye (the switch is over before a screenshot over SSH), a PlayStation game, a ZIP, a game on TF2, hardcore, a power-off mid-switch, Menu + Select on the real buttons, the German wording on the device

  • The volume sign: a press of a volume button shows VOL, a bar and the percentage, centred at the top of the screen for 1.5 s, in the menus, during a game and in the in-game menu (screens::volume_sign, app::show_volume). It is 20 pixels tall and 6 from the top, so it stays in the border above a GBA picture, and covers the top of taller pictures. fb::Framebuffer holds it: draw stamps it on each screen, and blit writes around its area, so a game's picture never overwrites it and it doesn't flicker, with no cost per frame beyond splitting those rows. When its time is up, the menus draw again, and a game fills the area with its picture and black around it (libretro::picture_in_area), as it only redraws its own area. During a game the launcher process still changes the volume, as amixer there doesn't hold up a frame, and writes the new percent as one byte to the game's standard input (libretro::volume_levels). So the game shows the level that was set, not one read too early. Only the buttons show it: VOLUME in the settings has its own bar. Claude checked it on the device with injected presses and screenshots on 2026-11-09: a game list, Wi-Fi, the title keyboard, Super Mario Land during play and in its menu, and closing the menu with the sign up. A press every 1.4 s or so during play kept 59.6-59.75 fps with no underruns, the same as without. The user confirmed on 2026-11-09 that it feels good on real presses, and that the lowest step is quiet but audible

  • Volume moves in steps of 5%, by the buttons and by the VOLUME row in Settings (settings::VOLUME_STEP). The hardware has levels 0 to 31 at 1.5 dB each, with 0 and 1 muted, so each step uses its nearest level (percent_level: 0, 2, 3, 5, 6, 8 ... 29, 31), which makes only 0% silent. settings.txt still saves the level, and one between steps, from an older version, snaps to the nearest step at start (level_percent). Before, a press moved two levels, so going down from the top showed different numbers from going up. Checked on the device on 2026-11-09: a walk from 0 to 100 and back hit the same levels, stopped at each end, and a saved level of 1 came back as 2

  • During a game two processes save settings.txt, and saving writes every setting, so each reads the file again just before: the launcher before a volume button's save, or it undoes the game menu's display changes, and the game before saving those, or it brings back the old volume. Reproduced and checked on the device on 2026-11-09

  • settings-backup.txt (2026-11-12) holds the settings from before their last change. settings::save copies the file it replaces there first, if that file is whole (some lines, each key=value: settings::looks_whole). Loading uses the backup when settings.txt is missing, unreadable or not whole, and writes it back as settings.txt, also for S50dropbear, which reads ssh= itself. A reset right after a change once lost settings.txt's FAT entry; the boot repair emptied it and every setting went back to its default. Now at most the last change is lost. Checked on the device: an emptied settings.txt loaded the backup (one volume step behind) and was written back

  • NES through package/libretro-fceumm, SNES through package/libretro-snes9x (free for non-commercial use only), and Game Boy and Game Boy Color through mGBA. Consoles appear in release order, set by the order of SYSTEMS in systems.rs. New systems name saves with the full file name, such as Tetris.gb.srm; GBA and Tamagotchi keep <name>.srm and tama.state, so existing saves still load. Cores that make less sound per panel refresh than their sample rate, like NES and SNES at 60.10 fps, play it at a matching lower rate, or they would underrun every few seconds. The user confirmed all of their games play well. Headless, Yoshi's Island, Chrono Trigger and A Link to the Past emulate at 134 to 143 fps

  • The launcher redesign. The home screen is a grid of tiles with pixel-art pictograms (icons.rs, drawn for this project): RECENT (the last 8 games, in /mnt/data/recent.txt), each console with games, and SETTINGS. A console opens its list of games, with short names from names.rs. Holding a direction repeats (repeat.rs), Left and Right page, B goes back, and Start opens settings anywhere. Screens and button handling are in launcher.rs and screens.rs. Claude checked every screen on the device with injected presses and screenshots, and the user confirmed on 2026-11-08 that it feels very good

  • Game Boy games fill the screen height. CORE_OPTIONS in libretro/ turns off mGBA's Super Game Boy border, which shrank SGB-enhanced games to half the screen. Not yet seen by the user

  • PlayStation through package/libretro-pcsx-rearmed (PCSX ReARMed, ari64 recompiler and NEON renderer, built with the makefile's h5 platform, the same Cortex-A53). Games are .cue files with their .bin beside them in roms (also .chd and .pbp); the launcher lists only the .cue. No BIOS file is needed, because the core has its own replacement. The first memory card is the save memory, in saves/<game>.cue.srm; the second is off. The frontend skips reading a game into memory when a core sets need_fullpath, and for this core alone answers the option defaults it declares through SET_VARIABLES (CoreSetup::declared_defaults): unanswered, it kept no memory card and read nothing ahead from the disc, which caused 33 ms underruns while FF7 loaded. Pictures are shown at 4:3 (CoreSetup::aspect), since its modes run from 256 to 640 pixels wide. Frames of 128 lines or more are scaled in two bands on two cores (Band in libretro/scale.rs), the same output as one pass, which took FF7 from 8.7 to 4.1 ms per frame of the 20 ms that 50 Hz European games have. Claude ran Final Fantasy VII, Metal Slug X and Star Wars Episode I (all European) on the device: 50 fps, Metal Slug X in play at the lowest 480 MHz and 44 C with no underruns, FF7 sees a formatted card, and save states (4.4 MB) save and load. 50 Hz games set the panel to their own rate since 2026-11-14 (below). The user hasn't tried it yet, so sound and feel are unchecked

  • Sorting. Settings > THEME > SORT GAMES BY chooses NAME, RECENTLY PLAYED or MOST PLAYED (settings::GameSort, sort= in settings.txt, under THEME since GENERAL has no room). library::sort_library applies it to consoles' lists and favorites; RECENT stays in the order played and search results stay by name. Recently played goes by a game's newest save file and most played by its play time, as the details screen shows them. Games that tie, such as those never played, keep their order by name at the end. L1 and R1 only jump in a list by name (Library::jumps_by_letter), and the hint leaves out L R ABC otherwise, on RECENT as well. When the games are read again, the selection follows its game into the new order (Launcher::selected_path, select_path), in any sort and on RECENT.

    Claude tested on the device on 2026-11-09 with dummy games with chosen dates and times: each order was as expected in the Game Boy list, R1 did nothing and the hint changed outside NAME, and a game started from row 4 moved to the top under RECENTLY PLAYED with the selection on it. Sorting 1000 extra empty games took 0.41 s, which happens at each start and after each game, and not at all under NAME. settings.txt and recent.txt were put back, and the card's files were as before. FAVORITES sorted is covered by a unit test only

  • Play time (play_time.rs). Each game keeps its time in whole seconds beside its saves, as saves/<save file>.playtime, so it follows the save name through being zipped or moved and is deleted with the saves. The game process counts the time between frames and leaves out any gap of a second or more, which covers the in-game menu, sleep and anything else that stops it without any of them telling it. Fast-forward counts as the time it took. It writes on opening the menu, on the lid closing, on leaving (with a power-off's signal too) and every 5 minutes, so a core crash costs at most that. Details show it as UNDER A MINUTE, 12 MIN or 2 H 5 MIN, adding what a grouped game's discs had as games of their own. A game with saves from before this says UNKNOWN until it is next played; the file isn't counted among SAVE FILES.

    Claude tested on the device on 2026-11-09 with a dummy copy of a game started from its list: 20 s of play, 10 s frozen with kill -STOP as sleep would be, then 5 s read 25 at the menu, a 12 s menu and 10 s more read 36, and 8 s more ended by SIGTERM read 43. Details showed NOT YET, UNDER A MINUTE and, with a written 7500, 2 H 5 MIN, and deleting the game removed the file. The 5-minute write and a real lid closing haven't been seen on the device

  • Play statistics (play_log.rs, Settings > STATISTICS). Each time the game process writes a game's play time, it adds the whole seconds gained to /mnt/data/play-log.txt under the local date and the console's key, one line per day and console (2026-11-11 gb 1234), merged and written atomically, so a year is a few kilobytes. A game's own .playtime has no dates, so the user chose to count only from the log, which starts empty: older play isn't in it, and the total can be less than the games' times added up. Deleting a game leaves its time in the log. A session over midnight goes to the day of each write, at most 5 minutes off.

    The page is the twelfth tile on the settings grid. It shows the total since the log's first day, the last 8 weeks (Monday to Sunday) as bars with their H:MM above and the Monday below, this week in the accent colour, and the consoles by time, five at a time, Up and Down scrolling. With no log it says NO PLAY TIME YET. The library reads the log, so it is current after each game. --preview-statistics renders it with made-up weeks.

  • Display settings per console. Changing scaling, effect or ghosting in the in-game menu saves all three for that console, as display.gb=fit,crt,smart in settings.txt (Settings::console_displays, read with display_for). A console's key comes from SYSTEMS, so Game Boy and Color are separate. Settings > IN GAME is the default for consoles with no such line, with a note saying so and how many have their own, and USE FOR ALL CONSOLES, which takes A twice, removes those lines. RESET SETTINGS does too. A line with an unknown value is ignored as a whole. The game process finds its console from the game (game::run_game), through to Host::console. An existing setting carries on as the default, and a console takes it as its starting point the first time its menu changes.

    Claude tested on the device on 2026-11-09 with saves in /tmp: an effect changed in a Game Boy game wrote only display.gb, an NES game still showed the default, the Game Boy game showed its choice after a restart, the note counted one console, and a first press of USE FOR ALL CONSOLES changed nothing while a second removed the line. settings.txt was put back afterwards. How the effects look on each console is the user's to judge

  • Borders around small screens (libretro/border.rs). Settings > IN GAME > BORDER and the in-game menu's BORDER choose what fills the screen around the picture: AUTO (default), OFF (black, as before), BEZEL or THEME. It is the fourth part of a console's display line, as in display.gb=fit,none,smart,bezel, and border= for the default; lines from before it have three parts and keep AUTO. AUTO gives a bezel to the handhelds (Game Boy, Color and Advance, Game Gear, Neo Geo Pocket and Color, WonderSwan and Color, Lynx), whose small screens leave the most around them, and black to the others; the in-game menu shows what it gives, as AUTO (BEZEL). Tamagotchi is left out, as its picture hasn't been looked at. BEZEL draws a handheld's body in a colour that console came in (a plain dark grey for the others), lit from above, with rounded glass reaching 18 px past the picture, a light rim, a dark edge by the picture, a shadow on the body and a red power light where there is room left of the glass. Where the margin is too thin, the glass runs off the screen's edge so its corners don't show. No copied artwork or logos. THEME is the theme's background, darker downwards, with a 2 px line in the accent colour 4 px from the picture.

    The border is made for the whole screen once per picture size and choice (Host::update_border), so it costs nothing per frame; a size change, as the Mega Drive's modes make, draws it all again, which also clears what the old picture left. It is drawn at the start, after the menu (play::draw_whole_screen, in place of clearing to black), behind a sign whose time is up and where the fast-forward sign was (picture_in_area takes it), and the in-game menu dims it with the picture. The battery warning keeps its black box. LCD GHOSTING now shows only on cores with its option (mGBA), which makes room for BORDER: only PlayStation games have CHANGE DISC, so no game shows all 16 items. --preview-borders renders six of them as a sheet.

    Claude tested on the device on 2026-11-11 with a copy of Link's Awakening and its saves in /tmp: the bezel at FIT, 59.45 fps with 2.1 ms on the screen as before, the dimmed bezel behind the menu with AUTO (BEZEL), THEME chosen there showing at once on resuming and saved as display.gb=fit,none,smart,theme, and the fast-forward corner given back to the border after R2. settings.txt was put back. How it looks to the user, the other consoles, INTEGER, the German wording and a Mega Drive mode change are unchecked

  • A game's details. Y in a game list opens them (Screen::Details, screens::details_screen): the file's full name, console, size, last played, how many save files, and the cover, with PLAY, STAR or REMOVE STAR (COLLECTIONS since 2026-11-13), and DELETE GAME AND SAVES. The screen holds the game's path, not its place in the list, as playing or starring can move it. --preview-details renders it. The size is what the game takes on the card, so a ZIP counts as zipped, and a .cue or .m3u counts all its tracks and discs (game_files::GameFiles). Last played is the date of its newest save file, since leaving a game writes one, so it needs no file of its own and knows games played before this screen; a game that never saved says NOT YET. It trusts the clock at the time, so a device that had no network time can show an odd date.

    Delete needs A twice in a row, and any other press calls it off; the screen always opens on PLAY. It removes the game first, so a failure there leaves the saves, then its saves (the memory card, resume point, 8 slots and any .unreadable), its cover, its star and its RECENT entry, and any folder that left empty, up to but not including roms. For a grouped or .m3u game that includes each disc, its tracks and the saves those discs had by themselves. Only what is under roms counts as the game's: a line in a .cue or .m3u that leads elsewhere is left alone. Anything another game on the card also uses stays, such as the save a cartridge shares with its ZIP or a cover of the same name, until the last game using it goes. There is no undo.

    Claude tested on the device on 2026-11-09 on dummy games, first listing the card's files: a cartridge with all its kinds of save and a cover; a cartridge and a ZIP sharing a save and cover, which stayed after the first and went with the second; a starred .cue with two tracks in its own folder; and a game grouped from two discs in a folder each, with its list and a card from before grouping. A first press followed by another button deleted nothing. Afterwards the card's files, recent.txt and settings.txt were as before. Failing to delete hasn't been seen on the device, and neither has a real game been deleted

  • Jumping by letter. L1 and R1 in a game list go to the first game of the previous or next letter, and L1 inside a letter goes to that letter's start first. They wrap at the ends and repeat while held (repeat.rs). The letter is the first character of the name as shown, ignoring ASCII case, as lists sort; so a name with The in front counts as T. RECENT doesn't jump, being in the order played. Left and Right still page, though the hint no longer has room to say so, nor for B BACK. The hint calls this L R ABC.

  • Search. X on the home screen or in a list opens the keyboard (app::search_games), showing how many games match while typing. OK shows them as a list like FAVORITES, with each console's icon, and does nothing at 0 matches. A game matches when all the typed words are in its shown name or file name, in any order, ignoring case and common accents (library::find_games). It looks through all consoles, also those with their tile hidden. List::Search reads Library::search, and reload_library runs the search again after a game or a star, so results stay. B leaves to where the search began, and X in the results edits the same text. They last until the launcher stops.

    Claude tested on the device on 2026-11-09 with injected presses and screenshots, without starting a game: "mario" counted and listed 7 games from four consoles, adding a letter counted 0 and OK stayed, B left, and starring in the results kept them. In the Game Boy list, R1 and L1 moved as described and holding R1 for a second kept jumping and wrapped. German was checked in previews. The buttons' feel on the device is the user's to judge

  • ZIP games (rom_zip.rs, with the zip crate, deflate only). A .zip in roms lists as a game of the console its file inside belongs to, named after the ZIP, and an unknown, damaged or password-protected ZIP is left out with a line in the log. With several games in one, the first by name is used; other files and macOS's __MACOSX and ._ files are ignored, as are names with ... What a ZIP holds is read once per scan and cached, since system_for runs often. PlayStation games don't count, as a disc is several files, and a game over 128 MiB unpacked is refused, as /tmp is in memory.

    The game process unpacks to /tmp/unpacked-game/<zip's name>.<game's extension> before it opens the screen, gives that file to the core and removes it on leaving, so cores need nothing special. A power-off empties /tmp, and the next game replaces what a crash leaves. save_path names saves as for the unzipped game with the ZIP's name, so Tetris.gb zipped as Tetris.zip keeps its saves, states and cover, while RECENT and favorites keep paths and don't carry over. scripts/boxart.py doesn't know ZIPs; the on-device download does.

    Claude tested on the device on 2026-11-09 with ZIPs of 10 games, each also holding a readme.txt and a __MACOSX file. NES, Master System, PC Engine, Mega Drive, Game Boy, Color, Game Gear, SNES, Advance and Tamagotchi all started on their cores; a 16 MiB GBA game unpacked in 0.21 s. A PC Engine ZIP started from its list, saved a slot and quit, which removed the unpacked file, then resumed. 500 more ZIPs added about 0.06 s to the launcher's start, with the files already in the cache. Neo Geo Pocket, WonderSwan and Lynx use the same code but had no games to try, and the user hasn't played from a ZIP

  • Multi-disc PlayStation games. An .m3u is a text file with one disc per line, such as Final Fantasy VII (Europe) (Disc 1).cue; names start from the .m3u's folder, and lines starting with # are comments. library::find_roms reads every .m3u and leaves out the discs it names, ignoring case as FAT does, so the game shows once under the .m3u's name. The core reads the file itself and holds up to 8 discs. All discs share one memory card and one set of states, in saves/<game>.m3u.*. A cover comes from the collection's first disc (closest_cover).

    Discs group by themselves without an .m3u (library::group_discs). Names that differ only by a disc number count as one game, across all folders, as some sets give each disc its own: (Disc 2), [Disk 2 of 3] or (CD2) anywhere in brackets, or Disc 2 or CD2 at the end of the name. Tags must match, so (Europe) (Disc 1) and (USA) (Disc 2) stay two games. Of two copies of a disc, the first by path joins and the other stays its own game. The launcher writes the list to roms/.discs/<game>.m3u with full paths, only when it changed, which leaves the game folders as they are; a dot folder isn't scanned. If it can't be written, the discs stay separate. An old list of a game that no longer groups stays in .discs unused. Only PlayStation games group.

    Going from separate discs to one game changes the save name. So when an .m3u game has no memory card, game::adopt_disc_save copies the newest card of its discs, leaving theirs. States aren't copied, because each records its own disc number. A starred or recent disc no longer shows, because that path isn't a game any more.

    CHANGE DISC in the in-game menu (Item::Disc) appears when the core reports more than one disc, so multi-disc .pbp files should get it too; no such file has been tried. Left and Right choose a disc and A inserts it and resumes. The frontend answers the core's SET_DISK_CONTROL_EXT_INTERFACE (DiskControl in libretro/host.rs): it opens the lid, sets the disc and closes it, and PCSX ReARMed reports the lid as open for two more seconds, which is how a game notices. A disc whose file is missing is refused first, because the core would otherwise be left with no disc. If the core fails anyway, the previous disc goes back. The disc in use is part of a save state, so resuming and loading a slot bring back that state's disc, and the menu follows.

    Claude tested on the device on 2026-11-09 with generated discs of 300 sectors holding only a boot file and a looping program, as no PlayStation game was on the device. The list showed one game; the same disc and a missing third disc gave their messages; changing to disc 2 and back opened that disc's file (/proc/<pid>/fd); loading a slot and quitting and restarting both returned to disc 2; and starting from the list used saves/Disc Test.m3u.*. Grouping was tested the same way: two discs showed as one game, and the new card matched the newer disc's with both originals kept. With the user's real Final Fantasy VII (USA) discs 2 and 3, grouped from a folder each and run with a save in /tmp, changing disc during the intro opened disc 3 at about 60 fps with no underruns, and the intro carried on. A dummy disc's run logged one 15 s underrun after its change, which the real game didn't. No real game has asked for its next disc yet, so whether a game accepts the change at that prompt is unchecked

  • Mega Drive, Master System and Game Gear through package/libretro-genesis-plus-gx, and PC Engine cards through package/libretro-beetle-pce-fast. Files are .md, .gen and .smd; .sms; .gg; and .pce. Mega Drive .bin isn't listed, because the launcher picks a system by extension alone and would show every PlayStation disc's .bin as a game, so rename those to .md. CD games aren't offered, as they need a BIOS and .cue is PlayStation's. The Mega Drive, Master System and PC Engine show at 4:3, since their pixels aren't square; the Game Gear keeps its 160x144 shape, 533x480 with side borders. Both cores answer their declared option defaults. Genesis Plus GX reports its save size differently once running: the full 64 KiB until the first frame, then only up to the last changed byte. So saves load fully, the first autosave rewrites the file at its trimmed size, and a game that has never saved gets a 0-byte .srm, which is harmless. Its licence bans selling it or using it commercially, like Snes9x's, which matters for a release. Claude tested with homebrew from /tmp on the device: Cave Story (Mega Drive) emulates at 206 fps headless and plays at 59.8 fps with 9.3 ms of work per frame at 1008 MHz, Astroswab (Master System) at 378 fps, and Jazz Jackrabbit (Game Gear) at 503 fps. Save and load state and resume were proved on Jazz's level select, a Mega Drive save survived a reload, and states are about 1 MB. No PC Engine ROM was at hand, so a hand-written 8 KiB card checked that the core boots, draws correct colours, runs at 373 fps, keeps its 2 KiB save and writes an 80 KB state. The user tried all four consoles on 2026-11-09 and confirmed they work

  • Neo Geo Pocket and Color through package/libretro-beetle-ngp (.ngp; .ngc, .ngpc), WonderSwan and Color through package/libretro-beetle-wswan (.ws; .wsc), and Atari Lynx through package/libretro-handy (.lnx, .lyx). Handy was chosen over Beetle Lynx because it has a built-in boot replacement, so no BIOS file is needed. Its retro_init reads a log callback from an uninitialised struct when the frontend has no log interface, and crashed on its first message, so 0001-initialise-log-callback.patch starts it at NULL. All show in their own shape, with no CoreSetup::aspect. WonderSwan's native 75 Hz would play 20% slow when paced by the panel, but the core's default 60 Hz mode runs 5 frames per 4 refreshes and reports 60.38 fps. WonderSwan has no Select button, so Select rotates the picture for vertical games: the core does it in software when SET_ROTATION fails, and has no automatic choice. Lynx rotates by the cartridge header, and Select overrides it. Two of these save outside the frontend's save memory: Neo Geo Pocket writes saves/<game>.flash when a game unloads, which every normal exit does, and resume states carry the flash too; Lynx writes saves/<game>.eeprom. WonderSwan uses the .srm as usual. Claude tested on the device from /tmp with homebrew: Asteroids Neo (Neo Geo Pocket) emulates at 154 fps headless and played at about 59.5 fps with 8 ms of emulation at 1008 MHz, a little under the panel's 59.94, so its feel is worth watching; Daybreak Ledger (WonderSwan) at 337 fps, playing at 59.8 fps; and the cc65 template (Lynx) played at 59.7 fps with 2.4 ms of emulation at 480 MHz. Save states and resume files were written on all three with no errors, restarting showed Neo Geo Pocket and Lynx resuming, and Select rotated WonderSwan. WonderSwan wasn't restarted to see it resume, and loading a state wasn't checked by its picture on any of the three. Only a WonderSwan ROM renamed .wsc tested that tile, and no game that actually saves to flash has been tried. Nobody has heard these systems' sound or tried their controls

  • Doom and Cave Story (2026-11-12), the first games that aren't emulated: engines rebuilt as libretro cores, package/libretro-prboom (.wad) and package/libretro-nxengine. Both come with the system: NXEngine's source carries Cave Story's freeware data, and package/freedoom installs Freedoom: Phase 1 and Phase 2 (0.13.0, BSD, checked against the release's signed checksum). They go to /usr/share/tetsu/games, and S15data copies each folder to roms once, noting it in bundled-games.txt on TETSU-DATA, so a deleted game stays deleted and a folder of the same name is left alone; the first boot after the update took 3.5 s longer for the 60 MB. Since 2026-11-15 the copy runs in the background once the boot jingle has ended (S15data keeps the PID that S21jingle writes to /run/jingle.pid, which the launcher removes about 200 ms later, and waits 15 s at most), and a .copying folder left by a cut-off copy is removed first. Copying during the jingle kept the card writing, and the launcher's code, read from the card as it ran, took 1.5 s per animation frame: 2 frames instead of about 50, and the setup at 13.6 s. On the device, with the games and bundled-games.txt removed, twice: 51 and 52 frames at 3 ms, the launcher at 6.65 s, both games copied 7.3 s after the kernel started. The games appear when the launcher reads the games again, as it does after the setup. System::titles names known files (freedoom1 shows as "Freedoom: Phase 1", doom2 as "Doom II", Doukutsu.exe as "Cave Story") in the lists, RECENT, SWITCH GAME, the details screen and the transfer page, with either GAME NAMES. Cave Story's .exe only counts when it is named Doukutsu.exe (System::only_titled), so DoConfig.exe and other programs aren't games, nothing can send one over Wi-Fi, and it can't be zipped, as it needs its data folder; the details screen counts and deletes that whole folder, unless the program sits loose in roms. Neither core saves a state the frontend can use (NXEngine has none; PrBoom's size changes as a level runs, which load_state refuses), so they use Persistence::OwnFiles: no save file, resume point, SAVE STATE or LOAD STATE, and nothing kept on leaving or at a lid close. Each saves from its own menu: Doom in saves/<wad>/prbmsav*.dsg (with prboom.cfg), Cave Story at its save points in saves/profile.dat, writing settings.dat in its folder. R2 doesn't fast-forward them, as Doom changes weapons with it. The frontend now answers ENVIRONMENT_SHUTDOWN, so Doom's own QUIT GAME goes back to the launcher. PrBoom calls its log callback unchecked in many places, and with no log interface every game crashed at address 0 in M_LoadDefaults (found with gdb on qemu-aarch64-static), so 0001-fall-back-to-a-log-function.patch gives it one that prints warnings and errors; its search for optional MP3 music logs six FindFileInDir: not found lines per track. Its prboom.wad is built in. Measured on the device: PrBoom renders Freedoom at 192 fps headless and plays at 59.85 fps in 320x200, shown at 4:3 over the whole screen, with 7.8 ms per frame at 1008 MHz; NXEngine at 441 fps headless and 59.9 fps at 480 MHz with 1.8 ms, its 320x240 doubled exactly. Claude checked the tiles and titles, a new Freedoom game saved from Doom's menu, quit, restarted and loaded at the same spot, Doom's own QUIT GAME, the in-game menu without the state rows, Cave Story's intro, and deleting Cave Story from its details screen removing its folder and play time. Nobody has heard either game, played them on the buttons, or reached a Cave Story save point

  • Nine more consoles (2026-11-13). SG-1000 (.sg) runs on Genesis Plus GX, which already ships. Eight new cores, each its own package, built from a pinned commit: libretro-stella (Atari 2600, .a26; current Stella from stella-emu/stella, built in src/os/libretro, rather than Stella2014), libretro-prosystem (Atari 7800, .a78), libretro-vecx (Vectrex, .vec, built with HAS_GPU=0, as the frontend has no GL), libretro-beetle-vb (Virtual Boy, .vb, .vboy), libretro-pokemini (Pokemon Mini, .min), libretro-potator (Watara Supervision, .sv), libretro-arduous (Arduboy, .hex, a CMake package) and libretro-picodrive (32X, .32x). PicoDrive and Arduous have git submodules, so they download with SITE_METHOD = git. .bin, which several of these cores also take, stays unlisted, as for the Mega Drive. None of them needs a BIOS: vecx has the Vectrex's built in (an empty cartridge runs Mine Storm), PokeMini uses its FreeBIOS, ProSystem skips the 7800's, and PicoDrive emulates the 32X's own. All use Persistence::BatteryRam with declared defaults. Virtual Boy, 32X and Arduboy keep save memory in the .srm. PokeMini writes saves/<game>.eep itself when the game unloads, as the Lynx does. The 2600, 7800, Vectrex and Supervision have no save memory, so their games keep only the resume point and state slots. Only PicoDrive applies cheat codes, so only the 32X has a cheats folder. Box art comes from the libretro-thumbnails folders of the same names, and RetroAchievements uses consoles 33 (SG-1000), 25, 51, 46, 28, 24, 63, 71 and 10 (32X). The Supervision, Pokemon Mini and Arduboy are handhelds with a bezel. The tiles go in release order, from the Atari 2600 (1977) to the Arduboy (2016), after Cave Story.

    PicoDrive forces the libretro file interface (USE_LIBRETRO_VFS := 1), with which dr_mp3.h assigns stdio's FILE * to libretro's RFILE *, an error in GCC 15. The frontend doesn't offer that interface anyway, so it is built with USE_LIBRETRO_VFS=0. It builds for platform=aarch64, which uses the arm64 SH-2 dynarec for the 32X. PokeMini runs at the Pokemon Mini's 72 Hz unless pokemini_60hz_mode is on, which drops every sixth frame and is the core's default; it reaches the core because the frontend now takes the first options interface (see the cores' options below). FRONTEND_DEFAULTS sets pokemini_video_scale to 1x, so the frontend scales the 96x64 picture and draws its LCD effect and bezel, instead of the core's 4x dot matrix. Beetle VB shows the plain red and black picture by default (anaglyph mode with the preset disabled). Vectrex and Virtual Boy are 50 Hz, so they set the panel to 50 Hz, as PAL games do. PokeMini aborts on a file of 0x2100 bytes or less, which it refuses, but a real cartridge is never that small, and a game runs in its own process.

    Tested by Claude on the device, with free homebrew and test ROMs: Warcraft 32X (black, as it needs the user's Warcraft demo data), the 32X 240p Test Suite (its menu, at 256 fps headless and 59.63 fps playing with 4.2 ms of emulation at 720 MHz), the Capitan Sevilla 3D demo for the Virtual Boy (124 fps headless, 50.41 fps playing, 6.6 ms at 1008 MHz), Catacombs of the Damned for the Arduboy (190 fps headless), Mine Storm on the Vectrex (166 fps headless, 50.03 fps), colour bars written for the 2600 (157 fps headless, 59.84 fps after the first seconds), PokeMini's FreeBIOS padded to 16 KiB ("NO CART", 600 fps) and idle loops for the 7800, Supervision and SG-1000 (over 300 fps). The tiles and icons, the Virtual Boy game from the menus with SAVE STATE, CORE OPTIONS, QUIT and a resume at the same screen, and RetroAchievements identifying it, were checked through the screen. The Arduboy plays at about 57 fps, because schedutil keeps the CPU at 720 MHz while it needs 7.6 ms per frame there with uneven frames; at 1512 MHz it played at 59.84 fps. Nobody has heard these consoles, played them on the buttons, or tried a real 7800, Supervision, SG-1000 or Pokemon Mini game

  • With more consoles a full home screen reached four rows, too short for a tile's text, so GRID SIZE's NORMAL now shows three rows at a time and pages beyond, as LARGE does

Built but not yet tested on the device:

  • Bluetooth headphones (stage 2), through bluez-alsa 4.3.1 with SBC, which every headset has (AAC, aptX and the others need libraries that aren't in the image). /usr/sbin/bluetooth starts bluealsa -p a2dp-source after bluetoothd. bluealsa's 20-bluealsa.conf makes the ALSA device bluealsa, a plug that converts the sample rate. audio_out.rs follows whether a headset is connected, from a thread that asks bluealsa (its ObjectManager at /org/bluealsa, an A2DP-source PCM in sink mode) once a second, and only while the system bus runs. Menu sounds and game audio then play with aplay -D bluealsa, the volume buttons set the PCM's Volume (0 to 127 per channel, which bluealsa sends to headphones that take it or applies itself), and the codec's level is kept for the speaker. A game checks every 60 frames and starts aplay again on the other device when a headset connects or goes, or when aplay ended, as it does with No such device when the headset is turned off. The boot jingle stays on the speaker, since Bluetooth starts after it.

    BlueZ only had its audio plugins after a clean rebuild: turning BR2_PACKAGE_BLUEZ5_UTILS_PLUGINS_AUDIO on with bluez5_utils-reconfigure kept the old src/builtin.h, so a2dp and avrcp were compiled but never loaded, there was no org.bluez.Media1, and a headset's connection failed with BREDR.ProfileUnavailable. A BlueZ option change needs bluez5_utils-dirclean.

    Tested by the user on 2026-11-11 with a Bose QC35 II: pairing, playing a game's sound, and the sound moving to the speaker and back as the headphones went off and on. BlueZ names it "LE-Bose QC35 II", from its Low Energy advertising; the same device also shows as "Bose QC35 II" before pairing. Each move to the headphones starts with underruns of 0.5 to 1.1 s while the A2DP stream starts, and the game ran at 55 fps for those seconds, then 59.7. One reconnect failed once with Couldn't change BlueALSA PCM configuration: Input/output error, and the next check a second later started it. Not yet tried: the volume buttons, how the delay feels.

  • Bluetooth, stage 1 of 2 (controllers; headphones are stage 2, above), from Settings > BLUETOOTH, off by default. The RTL8821CS's Bluetooth half is on UART1 with the Realtek 3-wire (H5) protocol, already in the shared device tree (realtek,rtl8821cs-bt, enable PG19, device wake PG17, host wake PG16). The kernel has CONFIG_BT, BT_HCIUART_3WIRE and BT_HCIUART_RTL, UHID for BlueZ's input plugin, and the PlayStation, Sony, Nintendo and Microsoft HID drivers with their vibration. rtl8821cs_fw.bin and rtl8821cs_config.bin are built in like Wi-Fi's; linux-firmware's WHENCE makes the config a link to rtl8761b_config.bin (25 bytes), the one ROCKNIX uses on the H700, so external.mk installs those two files only. BlueZ 5.86 with D-Bus is in the image, but post-build.sh takes the S off their init scripts; /usr/sbin/bluetooth on|off|status starts and stops both, and unblocks or blocks the radio through rfkill. The launcher runs it after its first frame when the setting is on, and from the screen's ON/OFF row. It costs about 5 MB of RAM (bluetoothd 3.9 MB, dbus-daemon 1.7 MB).

    The screen (bluetooth.rs talks to BlueZ through zbus, bluetooth_menu.rs has the list and buttons, bluetooth_screen.rs runs it) searches while open and lists connected devices, then paired ones, then the rest by signal, leaving out those that never said their name. A pairs a new device, trusts it so it can reconnect by itself, and connects it; on a paired device A connects or disconnects, and X twice forgets it. BlueZ needs an agent even for Just Works pairing, and without one refused the Backbone Pro with AuthenticationFailed (No agent available for request type 2 in bluetoothd -d). While the screen is open, the launcher is the default agent (bluetooth::Agent, NoInputNoOutput): it accepts every confirmation and answers a PIN request with 0000, since there is no screen for a passkey.

    No BlueZ call runs on the screen's thread. Starting a search while a classic pairing (a Samsung TV, chosen by mistake) kept the radio busy waited 15 s for BlueZ's answer, and the screen froze that long, then carried out the presses made meanwhile. A reader thread now connects, registers the agent, searches and sends the devices every second; pairing, connecting, disconnecting and forgetting each run in a thread, one at a time. Pairings are in /var/lib/bluetooth on the root filesystem: they survive reboots but not a flash, because TETSU-DATA is FAT and BlueZ's folder names have colons. Most Bluetooth controllers report the D-pad as a hat, so input::Axes turns ABS_HAT0X/Y and the left stick (on past half way, off under a third) into D-pad presses, for gamepads only (a south button and no accelerometer, as a DualShock 4's motion sensors are a device of their own). A controller then acts as the device's own buttons, in menus and games.

    Measured on 2026-11-10:

    • Boot: the Realtek driver held the chip in reset for 100 ms and waited 500 ms in probe, which the kernel waits for before mounting the root filesystem, so init started at 0.91 s instead of 0.53 s and the launcher was drawn 0.35 s later (6.70 s against 6.36 s with the driver kept out by initcall_blacklist=hci_uart_init). Patch 0017 moves the wait to a delayed work. The firmware upload (0.75 s over the UART) then still overlapped the first init scripts, 0.2 s, so hci_uart.rtl_boot_sync_delay_ms=5000 holds back the first H5 link sync at boot: it now loads 6.1 s after the kernel starts. Six boots drew the launcher at 6.33 to 6.46 s, one at 6.76 s, unexplained. Delaying the reset itself instead broke Wi-Fi on every boot: rtw88 timed out on SDIO (-110) loading its firmware at 0.84 s while the Bluetooth half was held in reset.
    • Wi-Fi with Bluetooth on: with rtw88's deep power save, the Wi-Fi firmware logged firmware failed to leave lps state and once stopped sending, connected, until Wi-Fi reconnected; Bluetooth off alone didn't bring it back. With rtw88_core.disable_lps_deep set, 170 s of searching lost no ping (85 of 85), and 50 such errors came back when it was cleared again. Since 2026-11-13 it is set for good on the kernel command line (rtw88_core.disable_lps_deep=Y in extlinux.conf), see "Wi-Fi hang with Bluetooth off" below.
    • Wi-Fi hang with Bluetooth off (2026-11-09 and 2026-11-13): connected, with an address, but nothing sent (no ping to the gateway, no DNS, no DHCP answer), while receiving still worked; rtw88 logged failed to get tx report from firmware from 312 s after boot, at first once a minute, then several a second, and a WARNING ran rtw_watch_dog_work > rtw_enter_lps > rtw_sdio_deep_ps > rtw_power_mode_change. Restarting S42wifi didn't help, a reboot did. So deep power save hangs the firmware with Bluetooth off too. The driver reads disable_lps_deep only in rtw_core_start, when the interface comes up, so the runtime switch /usr/sbin/bluetooth used to make took effect only at the next Wi-Fi restart, and turning Bluetooth off switched deep power save back on. Now it is set from the start and the script leaves it alone. Plain power save (POWER SAVING, iw ... power_save) stays on. Not yet measured: what it costs in battery, and whether the hang stays away over many hours.
    • Sleep: the Realtek driver puts the Bluetooth half in reset while asleep and probes it again on waking. With an RTC-timed s2idle (Wi-Fi disconnected first, as the launcher does), Wi-Fi reconnected and pinged after waking, with Bluetooth off and on, and the adapter was powered again after it.
    • On the device, Claude saw the BLUETOOTH tile, the screen turning on and off, its search listing four named devices nearby, and D-Bus, bluetoothd and the rfkill block following the setting. On 2026-11-11 the user paired a Backbone Pro (LE only, HID over GATT, driven by hid-microsoft as 045E:0B13) and its buttons worked in the menus, until the freeze above. Its first pairing then failed every reconnect: the Backbone reconnected by itself every 2 s, the kernel started encryption with the stored Secure Connections key, the Backbone never answered, and the link hit its 1 s supervision timeout (btmon). Not Wi-Fi (the same with it disconnected) and not the stored connection parameters (the Backbone asks for the same, 15 ms, latency 4, 1 s). Forgotten and paired again (one attempt got no SMP Pairing Response in 30 s, the next worked), it failed its reconnects the same way, so the pairing wasn't the cause; nor were the connection parameters (the kernel's defaults, latency 0, failed too). The traces showed the difference: while pairing, encryption started about 1 s in, after the Backbone's connection parameter update and the data length change; on a reconnect BlueZ started it 26 ms in, and the Backbone went on sending unencrypted L2CAP data as if the request never came. The BlueZ patch bluez5_utils/0001-device-raise-le-security-on-demand.patch leaves the level low on connect (device_attach_att(), and device_connect_le() for Device1.Connect) and lets the first read the Backbone refuses with Insufficient Encryption raise it, which happens 0.45 to 0.7 s in. With it the Backbone reconnected, encrypted and worked, also after being turned off and on (user, 2026-11-11). The Backbone takes a new static address with each pairing. btmon is in the image (BR2_PACKAGE_BLUEZ5_UTILS_MONITOR) for traces like these.
    • Its buttons then did nothing after a reconnect: it came back as event6, the name it had before, and input::spawn_readers never opened a name twice. It now opens a name again once the reader of the device that had it has ended.

    The search retries until the adapter is powered: at first it failed with NotReady and never ran, while the screen said SEARCHING.

    push.sh rootfs no longer copies /etc/passwd, /etc/group and /etc/shadow: Buildroot adds package users such as dbus only to the image, so the build tree's copies lack them, and pushing them removed the user D-Bus runs as. The unit was given the dbus user and the /var/lib/dbus link by hand, as in the image.

  • Faster boot with a logo. The launcher logs when it is drawn after the kernel started and after power on, from the system counter (since_power_on in system_info.rs), which gives the time before the kernel exactly. U-Boot's CONFIG_BOOTDELAY is 0 instead of 2, and its diagnostic writes of bootcount.txt and uboot-env.txt are gone, which cost 0.30 s. The launcher now draws 3.61 s after power on, 2.53 of them before the kernel. U-Boot reads the card at about 24 MB/s, measured with a kernel padded by 10 MiB, so the 21 MB kernel takes 0.87 s to load. An LZ4 kernel (10.6 MB, in the frame format booti unpacks, made with Buildroot's host-lz4) was tried on 2026-11-08 and was 0.11 s slower over three boots each (3.73 s against 3.62 s), so unpacking costs about 0.55 s; the plain Image stays. Kernel messages go to ttynull (CONFIG_NULL_TTY) at loglevel=5: the boot logo needs a level above CONFIG_CONSOLE_LOGLEVEL_QUIET (4), and at 3 it didn't show. With console=tty3, warnings such as 4021000.mmc: card claims to support voltages below defined range still printed over the logo, as the user saw, because tty3 isn't allocated that early and the kernel falls back to the visible tty1. The logo is board/rg35xxsp/logo.ppm, a plain PPM made with launcher --preview-splash, copied over the kernel's by a hook in external.mk, and shown centred from about 0.44 s after the kernel starts. The init scripts take 0.01 to 0.04 s each. scripts/push.sh uboot writes U-Boot at 8 KiB on the card over SSH and checks it by reading it back. The user saw the logo with a kernel warning over it; with ttynull it is clean. On 2026-11-08 the logo became TETSU over a smaller OS, and the user saw it on the device with no kernel text over it. On 2026-11-09 the user chose the project's logo, a tetsubin (iron kettle, as tetsu means iron) with a gamepad's controls on its side: icons::LOGO, on the launcher's 32 by 24 grid, drawn above TETSU OS at 5 times its size. Its handle, lid and body mirror around x = 17, and the spout balances the body sitting a pixel right of the grid's centre

  • Scrolling names: a selected name too long for its row pauses 1.2 s, scrolls at 48 pixels a second to its end, pauses, and starts over (screens::scroll_offset). Only that row is redrawn, 20 times a second (fb::draw_area), which kept the launcher at 2% CPU. The battery is now read once a second rather than on every redraw

  • Favorites: Select stars or unstars the selected game in any list, kept in /mnt/data/favorites.txt (favorites.rs). Starred games show a yellow * before their name, and a FAVORITES tile with a star icon follows RECENT on the home screen, listing them by name with their console's icon. A list that empties, as favorites do when the last is unstarred, goes home at once. Claude tested starring, the tile, the list and unstarring on the device

  • Collections and hidden games (2026-11-13): the user's own lists beside FAVORITES, and games kept out of every list.

    • A game's details (Y) have a COLLECTIONS row, which replaced STAR (the screen has room for six rows), with how many of the favorites and collections the game is in. It opens a checklist (Screen::GameCollections, library::collections_menu): FAVORITES, each collection, then NEW COLLECTION, with ADDED on those the game is in. A puts it in or takes it out; NEW COLLECTION types a name on the keyboard and puts the game in it; Y on a collection renames it. Select still stars a game anywhere.
    • Kept in /mnt/data/collections.txt (library::collections): each name in brackets on its own line, then its games' paths in the order added. A collection goes once its last game leaves it, so none is ever empty in the file. Names are at most 11 characters, which fit a NORMAL tile (longer labels are cut with ..), without [, ] or , (the settings' hidden tiles are a comma list), and two can't differ only in case, as the screen shows capitals. A name already taken isn't accepted, and the keyboard's line says so.
    • Each collection with games has a tile after FAVORITES (Tile::Collection, the stacked boxes of icons::COLLECTION), with its count, and its games joined on the home screen's header line. Its list is like FAVORITES: by name or the chosen sort, with each console's icon, letter jumps when by name. One whose games are all on a TF2 card that is out keeps no tile but stays in the checklist and the settings list. HOME SCREEN TILES lists every collection, saved as collection:<name> in hidden_tiles=; a rename carries that along and a delete removes it.
    • Settings > DATA > COLLECTIONS lists them with their game counts: A renames, X twice deletes (any other press calls it off). Games stay on the card. They are in DATA, not GENERAL, as GENERAL has 8 rows and the settings screen fits 9 with no note.
    • HIDE GAME on the details hides a game (/mnt/data/hidden-games.txt, one path per line, library::hidden) and goes back to the list. A hidden game leaves its console's shelf before every other list is picked from it, so it is out of the console lists, RECENT, FAVORITES, collections, search, the tiles' counts and SWITCH GAME (switcher::others); a console with only hidden games has no tile. Names are made before hiding, so the others' short names don't change. Its files, saves and play time stay; it still counts when deleting another game's shared files, a power cut during it still resumes it, and the transfer page still lists and deletes it. Settings > DATA > HIDDEN GAMES lists them with their console's icon, and A shows one again.
    • Deleting a game takes it out of the collections and the hidden list too. A backup includes collections.txt and hidden-games.txt.
    • Previews: --preview-collection-tiles, --preview-collection, --preview-collections-details, --preview-game-collections, --preview-collections and --preview-hidden, with sample collections and hidden games
    • Claude checked on the device: a collection made from the details with the keyboard, a second game added, its tile and list, a game hidden from that list and shown again in HIDDEN GAMES, a rename and a delete with X twice; the files were removed after. Unit tests cover the file, the lists built from it, the presses and the screens
  • Box art. scripts/boxart.py lists the games on the device over SSH, fetches covers from the libretro-thumbnails collections on GitHub (following their symlink stubs, and falling back to the closest name, such as one with (GB Compatible)), shrinks them to at most 200x300 with Pillow, and copies them as PPM files to /mnt/data/boxart. Downloads are cached in ~/.cache/rg35xxsp-boxart, never in the project, because the covers are copyrighted. All 35 games got a cover. Game lists show the selected game's cover in a panel on the right (boxart.rs, screens::list_screen), or the console's icon without one, which leaves about 31 characters for names. Claude checked the GBA and SNES lists in screenshots, and the user said the covers look fine

  • Network time. BusyBox ntpd (S43ntpd, from pool.ntp.org) runs from boot and keeps retrying, so it syncs whenever Wi-Fi is up, and steps the clock on its first answer. /usr/libexec/ntp-synced writes the RTC with hwclock -w -u on each sync and logs it to /var/log/ntp.log; the kernel reads the RTC at boot (CONFIG_RTC_HCTOSYS). The clock had been 32.6 hours fast. Claude checked on the device: the first sync stepped it by -117520 s to match the PC, and after a reboot it was right from the RTC (0.7 s off) and synced again 16 s into boot, though ntpd started before Wi-Fi. The system clock stays UTC

  • Theme settings, a THEME tile in Settings: COLOR (9 accent colours, all bright enough for the dark text on a selected row), TITLE (SHOWN or HIDDEN) and TITLE TEXT, typed on the on-screen keyboard (screens::KeyboardView, shared with Wi-Fi passwords), up to 24 characters with case kept; saving it empty brings back TETSU OS. The accent is part of the Theme each canvas carries, so a game's menu uses it too. Saved as accent=, title_hidden= and title=. Claude checked the colour on the menus and a game's menu (from --preview-pause on the device), editing, hiding and resetting the title; the test unit is back on the defaults. The user then chose blue with the title hidden. Without the title, the home grid starts at 36 instead of 84 pixels (GRID_TOP_UNTITLED, just below the clock), so 3 rows of tiles grow from 108 to 124 pixels, and each tile's icon and labels are centred in it

  • The home screen's header (2026-11-11, chosen by the user from PC mock-ups): a band 74 pixels high, 10% lighter than the background, with a 2-pixel accent line under it (screens::home_header). On it are the boot logo's kettle (icons::LOGO) at twice its size, the title right of it, and the clock and battery top right. The default title shows as on the boot screen, TETSU in the accent and a smaller grey OS. A typed title is in the accent at the same size, or smaller when it would reach the clock. TITLE HIDDEN removes the whole header, as before. Under the title, the selected tile's name in white, then a grey line about it: RECENT the last game and when, FAVORITES their names, a console LAST: <game> and when, or NOT PLAYED YET, and SETTINGS what is in it. The line is cut with dots to fit, but never its time. The last played game of a console is the first of its games in the recent list, which was played after any other, or else the one with the newest save (library::last_played_game). It is kept in the library, which is read again after a game, so the home screen reads no files after the first look at each tile. clock::ago_text gives JUST NOW, minutes, hours (also up to 6 hours back past midnight), YESTERDAY, days, and after a week the date, by local days. Tests cover the band, the logo, the title sizes, cutting, the last played game and the times. Checked only in PC previews, in English and German

  • Menu sounds: Settings > GENERAL > MENU SOUNDS, ON or OFF (default). Directions and volume keys tick, A, Start and Select make a rising blip, B a falling one, in the launcher's own screens (menu_sound in app.rs) and on the on-screen keyboard, for Wi-Fi passwords and the title, where X and Y blip too (keyboard_sound). sounds.rs makes them as tones, 14 to 52 ms at 48 kHz, quiet next to games, and plays each through its own aplay, which holds the card about 70 ms; a sound is skipped while the last still plays, and play_game waits for it (sounds::wait), since the card takes one stream. Groups now have room for 6 settings. Claude checked on the device that presses play without errors, and the user confirmed on 2026-11-08 that they work and are satisfying, in the menus and on the keyboard

  • Boot jingle: Settings > GENERAL > BOOT JINGLE, ON (default, also for a settings file from before it) or OFF. Since 2026-11-10 it is matsukaze, the rush of a tetsubin before it boils, answered by two strikes of an iron wind chime (a Nambu furin, from the same ironware tradition), 1.7 s long, with a boot animation. The user chose it from 32 candidates heard as WAV files (chip tunes, slower chip tunes, iron, furin), made by scripts/jingles/make_jingles*.py (numpy): chip tones with band-limited pulses, struck metal as decaying inharmonic modes with beating pairs, filtered noise and a small Schroeder reverb. The launcher keeps the chosen one as samples, src/launcher/sounds/boot-jingle.s16 (48 kHz mono, 160 KB, peak 11000 of 32767), because its noise and reverb would not come out the same made again in Rust; scripts/jingles/boot_jingle.py writes it, and Claude checked it matched the chosen WAV to the sample. The first jingle, from 2026-11-09, was a spoon tapping an iron kettle, 1.3 s, made as tones in sounds.rs. S21jingle runs launcher --jingle in the background just after S20audio, logging to /var/log/jingle.log, and writes its process to /run/jingle.pid. The launcher loads as usual, then waits for that process before it draws anything (sounds::wait_for_boot_jingle, at most 4 s), so a resumed game finds the card free. Turning the setting on plays the sound once, without the animation. It doesn't play when waking from sleep. The animation (boot_animation.rs) is drawn by the jingle's process over the kernel's logo, starting and ending on it exactly and drawing only where it is black (and the title): steam puffs from the kettle's spout as loud as the rush, a ring of teal blocks leaves the kettle at each strike (0.50 and 0.64 s), the title flashes, and a shine sweeps across it. It is in the logo's 5-pixel blocks, fading in steps of its colours, with a fixed random sequence so each boot looks alike. Its time comes from the samples the card has played (hw_ptr in /proc/asound/card0/pcm0p/sub0/status, read each frame while the state is RUNNING), so it waits out the amp's 700 ms and keeps to the speaker; with no sound by 1.5 s it runs on the clock. It draws every other refresh, about 30 frames a second, with blit_canvas: Framebuffer::draw took 31 ms a frame there, blit_canvas 7 ms, plus 4 ms to make the frame. launcher --preview-boot-animation FILE.ppm renders eight moments of it on the PC. Measured on the device on 2026-11-10: 50 frames during boot, the jingle's process ending 6.29 s after power on and the launcher drawn at 6.35 s (5.84 s with the first jingle, 3.83 s with it off). Claude saw steam, rings and the logo at the end in captures of the framebuffer

  • Sleep when idle: Settings > SYSTEM > SLEEP WHEN IDLE: OFF (default), 2, 5, 10 or 30 minutes without a press in the launcher's own screens (home, lists, settings), not in games, the Wi-Fi screen, the keyboard or a download (Wake::Idle, sleep_until_woken). It turns the backlight off through bl_power, leaves Wi-Fi as lid sleep does, and sleeps with freeze. The power button wakes it: only it and the lid are wakeup sources, and patch 0010 drops the waking press so it doesn't shut down. If the lid closes during it, that wakes it too, and it carries on as lid sleep. Saved as idle=. Claude checked it on the device at 2 minutes with an RTC alarm as the waker: it slept 2 minutes after the last press (PM: suspend entry (s2idle)), woke 50 s later, turned the backlight back on and rejoined Wi-Fi. Not yet checked by the user: that the screen is dark while asleep, and waking with the power button. The test unit is back on OFF. It also powers off after POWER OFF AFTER SLEEP's time asleep on battery, like lid sleep (2026-11-10), with the same RTC alarm and timer; a power button press before then wakes it as before, and if the lid closes, the lid sleep keeps counting the same 2 hours. Claude checked it on the device with LAUNCHER_LONG_SLEEP=60, a fake Discharging battery and a fake LAUNCHER_POWEROFF: it slept at 2 idle minutes, the alarm woke it 61 s later and the power-off command ran. A real power-off from idle sleep hasn't been tried

  • Wi-Fi and Bluetooth in the corner (2026-11-12), left of the clock and battery level on every screen with them (corner::add_corner, Canvas::links). Wi-Fi shows while wlan0's operstate is up, that is connected; the kernel has no /proc/net/wireless, so there is no signal strength without asking wpa_supplicant. Bluetooth shows while it is on in the settings: dimmer with nothing connected, and as bright as the clock while a device is, which the kernel shows as an hci0:<handle> folder in /sys/class/bluetooth. Both are read from sysfs with the battery each second, and the menus draw again when they change. The Bluetooth rune is 9 pixel rows, centred on the 7-row text, since at 7 it reads as an asterisk. The home header's CORNER_W leaves room for both. Checked on the device: Wi-Fi connected, Bluetooth on with nothing connected, and turned off; a connected device wasn't tried

  • Battery display: Settings > BATTERY > SHOW IN CORNER (in SYSTEM until 2026-11-13): PERCENT, TIME LEFT (see BATTERY below), WHEN LOW (only at LOW BATTERY AT or below while not charging, in red) or HIDDEN, for the menus' top-right corner; the clock takes its place. Games' low-battery warning and shutdown don't change. Like the accent, it is part of the Theme. Saved as battery=. Groups now have room for 5 settings (MAX_GROUP_SETTINGS). Claude checked all three on the device, WHEN LOW with a fake 9% battery through LAUNCHER_BATTERY_DIR; the test unit is back on PERCENT

  • Covers in lists: Settings > BOX ART > COVERS IN LISTS, ON or OFF. Off leaves out the art panel, so names have the whole width, about 49 characters instead of 31. Saved as covers=. Claude checked the GBA list both ways; the test unit is back on ON

  • Game names: Settings > THEME > GAME NAMES: SHORT (names::display_names, as before) or FULL, the file name without the extension, with its tags. Saved as names=; build_library takes the choice. Claude checked a list in FULL on the device; the test unit is back on SHORT

  • Start screen: Settings > THEME > START SCREEN: HOME, RECENT, FAVORITES or LAST GAME, the list of the last game's console with that game selected, from the first line of recent.txt (Launcher::start_in, open_start_screen in library.rs). An empty list stays home, and a game resumed at boot returns to it. Saved as start=. Claude checked LAST GAME and RECENT on the device by restarting the launcher; the test unit is back on HOME

  • Background shade: Settings > THEME > BACKGROUND: DARK (the first blue-grey), BLACK or TINT, an eighth of the theme colour over near black, which follows the colour (settings::Background::rgb). Like the accent, it is part of the Theme; it is also the dark text on selected rows. Saved as background=. Claude checked BLACK and TINT on the device; the test unit is back on DARK

  • Grid size: Settings > THEME > GRID SIZE: NORMAL (4 to a row, every row shown) or LARGE (3 to a row with icons at 3x, two rows at a time, paging like the lists, with small arrows for rows above or below). settings::GridSize gives the columns and rows, the home screen and grid_move take them, and the settings grid stays at 4. Saved as grid=. Claude checked both pages of LARGE on the device; the test unit is back on NORMAL

  • Hiding home screen tiles: Settings > THEME > HOME SCREEN TILES lists RECENT, FAVORITES, every console (with games or not) and SETTINGS, each SHOWN or HIDDEN, toggled with A (Library::every_tile, screens::home_tiles_screen). Saved as hidden_tiles=recent,tamagotchi,... in settings.txt, using each system's key in SYSTEMS. Games of a hidden console stay in RECENT and FAVORITES. If every tile is hidden, SETTINGS shows anyway; hiding it shows a reminder that Start opens settings. Fewer tiles can mean fewer rows, which makes the tiles larger. Claude hid and showed tiles on the device; the test unit has none hidden

  • Time settings, in Settings > GENERAL (a TIME tile until 2026-11-13): TIME ZONE (58 cities, west to east, Left and Right) and CLOCK (24 HOUR, 12 HOUR or OFF) for a clock left of the battery level on every screen that shows it. The note under them shows the date and when the network last set the clock (from /var/log/ntp.log). Saved as timezone=<IANA name> and clock= in settings.txt; UTC until chosen. clock.rs holds each zone as a POSIX TZ rule, taken from the PC's tzdata file footers, so the device needs no zoneinfo: the launcher reads it with the jiff crate, and games get it as TZ, which glibc reads, so cores' clocks (Pokemon's RTC) show local time. A test checks every rule against the IANA database hourly for two years, and on the device glibc gave the same local times as the PC for all 58, in winter and summer. Claude checked the screens, the corner clock in both formats and off, and TZ in a game's environment. The zone is set to Berlin on the test unit. A game's clock was not checked inside a game

  • Setting the time by hand (2026-11-12), Settings > GENERAL > SET TIME (set_time.rs), for a device without Wi-Fi. Day, month, year, hour and minute in boxes, Left and Right moving between them and Up and Down changing one, wrapping within it like a watch (a minute past 59 stays in the hour), the day kept within the month and the year between 2025 and 2099. The hour follows CLOCK (4 PM in 12 HOUR), and the full date with its weekday shows below. It is local time in the chosen zone (clock::zoned: a time skipped when the clocks go forward counts as after the change, a repeated one as the first). A saves it with clock_settime and writes the RTC with hwclock -w -u, so it holds through a power-off; B leaves unchanged. Then it restarts ntpd (S43ntpd restart): a running BusyBox ntpd doesn't notice the jump and went on logging tiny offsets with the clock 2 hours off, while a new one steps it on its first answer. So the network still wins, at once with Wi-Fi up or once it connects, which the screen says. Claude set 23:22 instead of 21:22 on the device: the system clock, the RTC and the corner followed, and with the restart ntpd stepped it back within a second (-7175 s)

  • Vibration (2026-11-10). The motor is on PWM channel 2, pin PH2, as in ROCKNIX's 0150-add-forcefeedback; patch 0015 makes it a mainline pwm-vibrator at 1 kHz (CONFIG_INPUT_PWM_VIBRA), an input device with FF_RUMBLE. Before the patch, Claude set PH2 to pwm2 with devmem 0x0300B0FC 32 0x77777422 (PH config register, PH2 in bits 11:8) and ran the channel from /sys/class/pwm: the user felt pulses at 20, 35, 50, 70 and 100 % duty. rumble.rs uploads one rumble effect through the evdev crate and changes its magnitude when a game asks (RETRO_ENVIRONMENT_GET_RUMBLE_INTERFACE, offered only when the motor opened). One motor stands for both: the weak one counts half, as pwm-vibrator does it, and any rumble runs at 20 % or more, up to the strongest for Settings > IN GAME > VIBRATION (OFF, LOW 40 %, MEDIUM 70 %, HIGH 100 %, vibration= in settings.txt; the note below it says it is one setting for every game). Changing it buzzes for 0.3 s at the new level. The motor is held still while the in-game menu is open, from a lid close until it opens, at power-off and when the game ends; closing the device file, also in a crash, makes the kernel stop it. mGBA rumbles for Game Boy rumble cartridges and the Game Boy Player, and changes the level almost every frame. In Pokemon Pinball (Rumble Version) that left the motor running for good, also after leaving the game: ff-memless clears an effect's PLAYING flag when new values are uploaded until its timer runs, and a stop in that window cleared the effect without telling the driver. The kernel still showed the channel enabled at 26.6 % in /sys/kernel/debug/pwm with no game running. Patch 0016 aborts every started effect instead, and the user checked Pinball's strong rumble and leaving the game with it. PCSX ReARMed rumbles only with a DualShock, which retro_set_controller_port_device would select. Tried (2026-11-10): the core then switches the pad to analog by itself after a game's 16th configuration read (pad.c, "Auto-enabling dualshock analog mode", no option), and Metal Slug X took no input at all, while the plain pad works. So PlayStation games keep the plain pad and don't rumble. The user felt the setting's buzzes and told the levels apart. Adding the motor renumbered /dev/input, so press.py finds devices by name

  • Settings is a grid of group tiles, like the home screen (settings_grid). Since 2026-11-13 each group is one topic, so a setting is found where a user would look: GENERAL (volume, brightness, language, menu sounds, boot jingle, time zone, clock, set time), IN GAME (a DISPLAY page with scaling, effect, ghosting and border, vibration, the game switcher, and an ACHIEVEMENTS page), WI-FI and BLUETOOTH (their own screens), BOX ART, THEME (the menus' and home screen's look: colour, background, title, grid, tiles, start screen, game names, sorting), DATA (USB STORAGE, WI-FI TRANSFER, SAVE BACKUP: getting things on and off the device), SYSTEM (system info, button test, sleep, power-off after sleep, system update, reboot, power off), ADVANCED (SSH, root password, SSH keys, performance, boot messages, setup, reset), CORES, BATTERY and STATISTICS. Later on 2026-11-13 the TROPHIES tile became IN GAME's ACHIEVEMENTS page, the display rows IN GAME's DISPLAY page (SUBGROUPS in launcher.rs, whose B goes back to the row they came from, Group::parent), and BATTERY took its tile, with the battery display from SYSTEM and CHARGE TO from ADVANCED. Before, START SCREEN, GAME NAMES and BATTERY were in GENERAL, GAME SWITCHER in THEME, the sleep rows, USB STORAGE and WI-FI TRANSFER in SYSTEM, and the time rows in a TIME tile, where DATA is now. A opens a group's list, B goes back to its tile, and Start closes settings from anywhere in them. Groups and their settings are GROUPS and Group::settings in launcher.rs, so a new option is one more entry; MAX_GROUP_SETTINGS (9) is what the screen has room for, with no note; a group with a note has at most 8. Settings that can't be undone (delete covers, reboot, power off) need A twice. Group icons are in icons.rs. Claude checked each screen on the device; not yet seen by the user

  • Settings > ADVANCED (2026-11-08):

    • SSH: OFF, USB ONLY (default) or USB + WI-FI, in settings.txt as ssh=. S50dropbear (in the overlay, replacing Buildroot's) reads that line itself at boot and when the launcher runs restart after a change: USB ONLY listens on 10.35.35.1:22 with -B; USB + WI-FI on every address without -B. dropbear can't allow an empty password on one network and not another, so USB + WI-FI needs an SSH key or a root password. Left and Right pass over it without one (ssh::next_access), the screen says why, and the script falls back to USB ONLY too. Restarting stops only the listener, so open sessions stay connected. Until this, dropbear listened on Wi-Fi as well, with root's empty password
    • ROOT PASSWORD: typed on the on-screen keyboard; Start on an empty field removes it (B on an empty field leaves without a change, which the user first took for removing it, so the line above the field says so). R1 fills in a made-up one (ssh::random_password): three groups of four from 27 lowercase letters and digits without lookalikes (no 0, o, 1, i, l, and no g, p, q or y, which the font draws without descenders, like 9, P and Y), about 57 bits, from /dev/urandom. It shows in the field to note down or edit, and Start saves it. BusyBox mkpasswd -m sha512 hashes it from standard input. SSH KEYS: the count, and A twice removes them all. scripts/push.sh key [FILE.pub] adds this PC's key (id_ed25519, id_ecdsa or id_rsa) once. Both live in /mnt/data/ssh (root-password, authorized_keys), so updates and new images keep them: S50dropbear copies them to /etc/shadow and /root/.ssh at every start, which makes the password apply on the serial console too and replaces one set with passwd. push.sh rootfs overwrites /etc/shadow, so it restarts S50dropbear afterwards. Removing the last key or the password drops USB + WI-FI back to USB ONLY (login_changed in app.rs)
    • A release build (2026-11-14) allows no empty password anywhere. post-build.sh writes VARIANT_ID=release into os-release when built with TETSU_RELEASE=1 (the release workflow does for a tag) and locks root's entry in /etc/shadow (*). ssh::release() and S50dropbear read that line. Then SSH defaults to OFF, the server never gets -B, every access needs a key or a password (ssh::effective falls back to OFF, and S50dropbear starts nothing), and without a password root stays *, so the serial console can't log in either. Right or Left on SSH while it is OFF with nothing set opens ROOT PASSWORD with a made-up one already in the field ("NOTE IT DOWN - START TURNS SSH ON"); Start saves it and turns SSH on one step from OFF (Right: USB ONLY), and B, which deletes first, leaves it off once the field is empty. The note under the group says TURNING SSH ON MAKES A ROOT PASSWORD, and THE USB CONSOLE STILL WORKS only with a password set. Removing the password and keys turns SSH off again
    • Claude checked on the device (2026-11-14), with the device's os-release marked as a release and the new launcher and S50dropbear: with a key, SSH ran on USB without -B, the key logged in and the empty password was refused; without the key the server stopped and SSH showed OFF with the note; Right showed a made-up password, Start turned SSH to USB ONLY (the first try went to USB + WI-FI, from the old usb setting: fixed by stepping from what was in effect), the shown password logged in and the empty one was refused; removing the password turned SSH off and locked root again, and the serial console refused root ("bad salt"). Then put back to a development build. A whole release image on the device is not yet checked
    • PERFORMANCE (performance.rs, performance=): BATTERY is schedutil up to 1104 MHz, BALANCED (default) schedutil up to 1512 MHz, as before, and PERFORMANCE the performance governor at 1512 MHz. The launcher sets them at start and on a change; one cpufreq policy covers all four cores. The in-game menu's 1008 MHz hold stays below every cap
    • Claude checked on the device: OFF stops the server and USB ONLY brings it back; USB
      • WI-FI is skipped with nothing set; with a key it listens on 0.0.0.0:22, the key works over Wi-Fi and USB, and an empty password is refused on both; a password typed on the keyboard logs in over both and a wrong one is refused; after a reboot the hash, keys and Wi-Fi mode came back; removing keys kept Wi-Fi (password still set), and removing the password then fell back to USB ONLY with the empty password working again; each PERFORMANCE level set the governor and cap. Not checked: a game at BATTERY or PERFORMANCE, and how either feels or lasts
    • CHARGE TO (charge=, in BATTERY since 2026-11-13): FULL (4.2 V), ABOUT 85% (4.1 V) or ABOUT 70% (4.0 V), the voltage the AXP717 stops charging at, written to the battery's voltage_max (power::apply_charge_limit). The driver takes only these three. The PMIC keeps the value until the battery is disconnected, so the launcher writes it at every start; this unit's PMIC was at 4.2 V, so FULL changes nothing. The percentages are rough, for a typical cell. Charging current stays at the device tree's 1024 mA, which is also the driver's maximum. A software stop at a percentage (turning charging off through status) was left out: a crash could leave charging off. Claude checked that each value reached voltage_max and that FULL started charging again; the user saw the charge level behave so far, but not yet over a whole charge. The orange light can blink after choosing 85% or 70% with the battery above that limit. While it blinked at 85% (4.16 V, USB in), 0x01 read 0x95 and 0x4a had bit 0 set (battery over-voltage), with health Good and 0x02 clear. Writing 4.2 V and then 4.1 V gave a few seconds of charging, then 0x94 and no over-voltage bit, and the blinking was found stopped. Unplugging and replugging USB after that gave one blink only, with the same readings. The cause is inferred: no AXP717 charger or LED documentation was found (the AXP717D datasheet has neither), and what starts the blinking wasn't reproduced. No harm was seen. current_now has an unknown offset, so about 425 mA here isn't real
    • BOOT MESSAGES (boot_partition.rs): LOGO is console=ttynull loglevel=5, as before; TEXT is console=tty1 loglevel=7, so kernel and init messages show until the launcher draws. The launcher changes only those arguments on the APPEND line of /boot/extlinux/extlinux.conf in the root filesystem (on TETSU-BOOT until 2026-11-12), writes it as extlinux.new, reads it back and renames it over the old one. The value is read when ADVANCED is drawn. push.sh boot and flashing put back LOGO. Claude checked that TEXT and back gave the original file byte for byte, and that the device boots with TEXT (launcher drawn 1.48 s after the kernel started, against 1.38 s); the user saw the logo with the messages below it, as it should be
    • RESET SETTINGS (A twice): Settings::reset keeps volume and brightness and sets everything else to its default, so SSH goes back to USB ONLY, the time zone to UTC and the charge limit to FULL; the CPU, charge limit and SSH server are applied again. Games, saves, covers, Wi-Fi networks, keys, the root password and the boot messages stay. Checked on the device, then the user's settings were put back
  • SAVE LOGS TO CARD, X on Settings > SYSTEM > SYSTEM INFO (boot_partition::save_logs): writes launcher.log, dmesg.txt and system.txt (date, device, kernel, build, uptime, settings, storage, RAM type and voltage, memory, processes, network) to logs on TETSU-BOOT, replacing the last ones, and says where. Wi-Fi passwords, keys and the root password hash are in other files and left out. Checked on the device

  • System info (system_info.rs) shows 11 lines, as many as the screen has room for: battery with voltage, game storage, game count (a .cue and its .bin count once), the RAM as sold with its type and what is free (1 GB LPDDR3 - 856 MB FREE; free of the kernel's total if the type is unknown), CPU speed and temperature, the Wi-Fi network and address (from wpa_supplicant's STATUS), the USB address, time since boot, kernel release and build date, the system build and the app version. The system build is the project's git revision, which Buildroot writes to VERSION in /etc/os-release; MODIFIED means it was built with uncommitted changes. It reads once per redraw (a key press, a new minute or a battery change). Claude checked the screen on the device

  • Box art over Wi-Fi, from Settings > BOX ART (boxart_fetch.rs), with the same sources and matching as the script, which still works over USB. DOWNLOAD MISSING COVERS fetches games without one, DOWNLOAD ALL COVERS AGAIN replaces each cover once its new one is saved, and DELETE ALL COVERS removes every .ppm, including those of removed games. The group shows how many games have a cover. Each console's collection is thumbnails in SYSTEMS. The app downloads over HTTPS itself (boxart_fetch::Https: ureq with rustls, HTTPS only, checking certificates against the Mozilla roots built in by webpki-roots, so the device has no curl or certificate files since 2026-11-08); serde_json reads GitHub's file listing; the png crate decodes (png.rs); covers are shrunk by area averaging. After the switch from curl, four covers downloaded again were byte-identical to curl's. The screen shows progress, B stops after the current game, the lid stops the download before sleeping (sleep leaves the network), and games with no cover are listed. The PNG tests use images from src/launcher/testdata/make.py. Claude checked all 38 cached covers decode exactly as Pillow does, then on the device: all 38 downloaded in about a minute, at the script's sizes with a mean difference of at most 4.2 of 255 (Lanczos against area averaging), the closest-name fallback, a game with no cover, B while running, and not connected. Then again after the settings rework, with the png crate: delete all, download missing (38) and download all again (38). The lid during a download is untested

  • Display options, in Settings and the in-game menu, saved in settings.txt, after research into emulator shader techniques. Scaling is sharp-bilinear (axis_weights in libretro/scale.rs): a screen line across a game pixel boundary blends the two by coverage, so pixels look even at non-whole FIT scales. SCALING INTEGER uses whole multiples: NES 512x480, SNES 512x448, Game Boy 480x432, GBA 480x320. SCREEN EFFECT CRT darkens the gap between game rows by up to 75% for dark pixels and 25% for bright ones, brightened 1.25 times; LCD is a soft grid, brightened 1.12 times. Both use coverage of the pixel gap (gap_coverage), so they stay smooth at any scale. The gain doesn't clip (Highlights, 2026-11-11): once a pixel's brightest channel passes 216, it bends to reach 255 at the gain's top, and the other channels follow it, so highlights keep their detail and colours their hue (before, Super Mario Bros.' sky turned grey-cyan as its blue clipped). Colours with a channel near full are brightened less than before, and their scanlines are fainter. The effect is laid over each line as it is scaled, the tables are kept in a Scaler until the sizes or effect change, and the last result is kept along a line. On the device, a benchmark scaling real NES (512x480) and Game Boy (533x480) screenshots took 3.35 to 3.48 ms with CRT against 4.21 to 4.23 before, and 2.79 to 2.83 ms with LCD against 3.12 to 3.19; with no effect, about the same (1.35 to 1.57 ms). Random noise, the worst case, is slower: CRT 5.7 against 4.9 ms. LCD GHOSTING sets mGBA's interframe blending: OFF, SMART (default, only flickering pixels) or FULL, applied live through GET_VARIABLE_UPDATE. Always on: mGBA colour correction (Auto) and FCEUmm's measured smooth-v2-fbx NES palette. The old scanlines and grid names load as CRT and LCD. FireRed with CRT: scaling 4.6 ms per frame at 59.4 to 59.6 fps. Moving drawing to the GPU would mean replacing fbdev with DRM, so it wasn't done. Claude checked each effect in screenshots of FireRed run from copies of its saves, and the user confirmed they look much better

  • The sound probe never lowers the samples per frame below the core's timing. Starting games make less sound: LeafGreen measured 1065 against 1097, and lowering the rate made it wait on the audio at 58.1 fps. FireRed now plays at its full 65536 Hz

  • 50 Hz games at their own rate (2026-11-14, ui/panel_rate.rs). A game whose core runs at 45 to 55 fps (PAL games, Vectrex, Virtual Boy) sets the panel to the nearest rate it can make, instead of 59.94 Hz, where some frames showed twice. Settings > IN GAME > DISPLAY > 50 HZ GAMES AT 50 HZ turns it off (50hz=off, fifty_hz_off; on by default, also for an older settings file); it is the last row, after USE FOR ALL CONSOLES, since it is one setting for every console, and its note says so. The --play process opens /dev/dri/card1, becomes DRM master and sets a mode with SETCRTC on the first CRTC with fb_id -1, which keeps the framebuffer fbdev made, so drawing into /dev/fb0 doesn't change. The pixel clock stays 24 MHz (pll-video0 at 72 MHz, which HDMI also uses); only the porches grow, from the panel's 770 x 520 up to at most 800 x 700, before each sync pulse. The panel runs in DE mode (0x23 = 0xa0), so it takes them. Found rates: PAL SNES and NES 50.007 fps on 792 x 606 (50.005 Hz), Mega Drive 49.7015 on 789 x 612 (49.703), PlayStation 49.7607 on 783 x 616 (49.759), 50.000 on 800 x 600. The mode change turns the panel off and on, 162 ms, at the game's start and end. Closing the device, also when the process ends or crashes, makes the kernel restore its own mode (lastclose); the menus process holds only /dev/fb0, which this needs. While a master exists, the kernel refuses FBIO_WAITFORVSYNC with EBUSY (drm_client_modeset_wait_for_vblank), so the game waits with DRM_IOCTL_WAIT_VBLANK on the same device. The sound's rate uses the panel's real rate (playback_rate), and the TCON registers (0x651104c, 0x6511050: totals, the lines doubled) show the mode with devmem. Checked on the device with Pac-Man (Europe): 50.005 Hz set, the TCON at 792 x 606, 59.94 Hz again after QUIT, and 49.92 fps steady, the sound's 0.1% margin making the game wait on the sound about once every 12 s (a repeated frame). With the setting OFF, the same game left the panel at 520 lines and ran as before. A test program measured 50.033, 49.953 and 49.715 Hz from vblank times for 623, 624 and 627 lines. While the game holds the display, the kernel doesn't light a newly plugged HDMI TV until the game ends; one plugged in before keeps 59.94 Hz

  • Fast-forward: holding R2 runs extra frames in each panel refresh, within 6 ms and up to 8 frames, keeping only the last frame's picture and sound, so sound keeps flowing. A >> shows in the top-right corner, and the CPU is held at 1008 MHz meanwhile. A Game Boy game reached about 5x with no underruns; at an 8 ms budget it missed refreshes and the sound ran out. The launcher log's statistics show the speed. The user confirmed it works; the sound is very choppy, especially on NES, which the user accepts

  • Sound rate from measured samples. The first 30 frames run silent while the frontend counts samples per frame, then aplay starts at the matching rate (playback_rate). mGBA reports 61.17 fps for Super Game Boy games but makes sound as at 59.73, so the rate from its timing made them wait on the audio at 58.48 fps. They now run at 59.5

  • Resuming at boot. Starting a game writes its path to /mnt/data/playing.txt; quitting or a crash removes it, but shutting down mid-game stops the launcher before it can, and an empty battery keeps it on purpose. At start, the launcher shows RESUMING with the game for 1.5 seconds, where B cancels, then starts it, and its resume state continues it. Claude tested quitting, cancelling, and a real reboot mid-game, after which the game came back by itself. The user confirmed it works with the power button

  • Low-battery protection (power.rs). Running on battery, a game shows the charge in red in the top-left corner, which is border for every system, at Settings > BATTERY > LOW BATTERY AT or below (10% until 2026-11-13, now 15% by default). At 3% or below, or under 3.35 V, read twice 5 seconds apart, the game saves as when quitting, shows BATTERY EMPTY and exits with code 3, and the launcher powers off. The launcher also powers off when it reads empty. Closing the lid in a game saves its resume state, and the launcher waits 500 ms before sleeping. LAUNCHER_BATTERY_DIR and LAUNCHER_POWEROFF point the battery reading at test files and replace poweroff. Claude tested the warning, the empty path and the lid save on the device with those; a real empty battery hasn't been tried

  • Settings > BATTERY (2026-11-13): CHARGE HISTORY, HEALTH AND USE, SHOW IN CORNER, LOW BATTERY AT, CHARGED NOTICE and CHARGE TO. The AXP717 driver gives the level, voltage and status, but no capacity, cycle count or temperature, and its current_now has an offset the driver itself calls unknown (it read 429 mA at 100% not charging), so all of this comes from a log of the level over time.

    • The log (device/battery_log.rs): the menus' process, which stays up through games and puts the device to sleep, samples every 60 s, and at once on plugging in, unplugging or charging ending (a thread looks every 10 s). A line is time percent mV state activity in /mnt/data/battery-log.txt: state D discharging, C charging, K charging at constant current with a good supply, F on USB power and not charging; the activity is what the device did since the sample before: menu, sleep, a console's key, or off (powered off, or the launcher not running). K reads the charge phase in REG 0x01 bits 0-2 (2 constant current, 3 constant voltage, 4 done) from the regmap's debugfs listing, only while charging, and needs USB at 4.5 V or more. Samples stay in memory and are appended every 15 minutes, at a game's start and end, before sleep and at shutdown (also the power button's and the sleep timer's), only while TETSU-DATA is mounted read-write (storage::data_writable) and not while USB STORAGE has it (battery_log::hold). sleep::freeze wraps every sleep, so the time asleep is one sleep sample taken on waking. Samples older than 30 days go at start, once a day past that, into /mnt/data/battery-stats.txt (since, charged percentage points, capacity measurements), so cycles and capacity outlast the log. The log isn't part of SAVE BACKUP: it belongs to this battery
    • Estimates (device/battery_estimates.rs): each activity's drain is the percentage points lost over the time between discharging samples (gaps over 3 minutes awake and off count for nothing; a sleep counts whole), given once there is half an hour and 2 points (2 hours for sleep). Time left is the level above 3% over the current activity's rate, or all awake time's; the corner's TIME LEFT shows it in the menus ("80% 4H05"), and only the percentage until there is enough. Time to full adds each remaining point at the speed earlier charges had in that tenth of the range. Charge cycles add every rise while on USB power, 100 points a cycle. Capacity: a run of K samples of at least 20 minutes and 20 points gives the set charge current (constant_charge_current_max, 1024 mA) times its time over the points gained; the capacity shown is the median of the newest 5, against the profile's rated 3300 mAh. This assumes the gauge's percentage follows the charge linearly and the charger really delivers its set current in that phase; neither is checked yet
    • CHARGE HISTORY (device/battery_screen.rs): the status line (level, state, voltage), the time left or to full, and the level over 24 hours or 7 days (Left and Right), in the accent on the battery, green plugged in, grey asleep, and a gap while off, with ticks every 6 hours or each midnight in local time. HEALTH AND USE: capacity, cycles since the log began, when charging last ended, and BATTERY USE: the menus, each console and sleep, with points per hour and how long a full charge lasts at it (Up and Down scroll). launcher --preview-battery day|week|health FILE.ppm renders them from a made-up log
    • LOW BATTERY AT (low_battery=): 5, 10, 15 (default) or 20%, where the corner turns red, WHEN LOW shows, and a game warns (power::level)
    • CHARGED NOTICE (charged_notice=, on by default): BATTERY CHARGED as a sign at the top in the menus and in a game when the status goes from Charging to anything but Discharging (power::ChargeWatch)
    • Checked by Claude on the device: the log starting and sampling, its first write to the card, the BATTERY tile, its group and both pages (with only minutes of log), IN GAME's DISPLAY and ACHIEVEMENTS pages and B from them. Not checked: anything on battery (time left, drain rates), a real charge (time to full, K samples, the capacity measurement and whether it is near the truth), sleep samples, the CHARGED sign, the 30-day trim, USB STORAGE with the log, the German wording
  • The menu checks the battery every second and redraws when it changes. Before, charging status only showed after moving in the menu, or after 30 seconds

  • Wi-Fi, set up from Settings > WI-FI. The chip is a Realtek RTL8821CS on SDIO (024C:C821, mmc1), driven by rtw88 with rtw8821c_fw.bin and the regulatory database built into the kernel, because the driver probes at 0.5 s, before the root filesystem. external.mk points CONFIG_EXTRA_FIRMWARE_DIR at the target, so post-build.sh keeps that one firmware file there. S42wifi starts wpa_supplicant and wpa_cli -a /usr/libexec/wifi-event in the background, with networks in /mnt/data/wpa_supplicant.conf; the hook gives each connection its address (below). The launcher talks to wpa_supplicant's control socket (wifi.rs): the screen (wifi_menu.rs) lists networks in range by signal, with saved ones out of range last, scans every 10 s and on Y, joins with A, and opens a saved network's page with X (below). A password is typed on an on-screen keyboard (keyboard.rs) with lowercase and every printable symbol (font::glyph_cased, used only where case matters). A new network is saved only once it connects. WPA3-only, enterprise and WEP networks show as NOT SUPPORTED, since WPA3 would need OpenSSL. Claude saw the list of 9 real networks and the keyboard on the device. Sleep works with Wi-Fi (patch 0013), checked with an RTC wake alarm (echo +8 > /sys/class/rtc/rtc0/wakealarm; echo freeze > /sys/power/state). The user joined their network on 2026-11-08, and it came back after a reboot. Lid sleep while connected hung the device until the launcher started disconnecting first; with that, an RTC-timed sleep woke and rejoined within 8 s, and the user confirmed lid sleep works and Wi-Fi reconnects

  • More Wi-Fi settings (2026-11-12), all under Settings > WI-FI:

    • An ON/OFF row first (wifi= in settings.txt). Off runs S42wifi stop: no wpa_supplicant and wlan0 down, also at boot, and the corner shows no Wi-Fi.
    • A saved network's page (wifi_pages::NetworkPage), from X, or A on the connected one: CONNECTION DETAILS (address and prefix from ip, gateway from /proc/net/route, DNS from resolv.conf, MAC, signal, band and channel and speed from SIGNAL_POLL, security), IP ADDRESS AUTOMATIC or MANUAL with ADDRESS (192.168.1.50/24, or /24 without a prefix) and GATEWAY, DNS (up to 3, empty for the router's), AUTO-JOIN, MAC ADDRESS DEVICE or PRIVATE, and FORGET (A twice). MANUAL without an address opens the keyboard with the address and gateway DHCP gave. Values are checked as typed (network.rs), and each change is saved at once.
    • Where they are kept: wpa_supplicant's own network block, so forgetting removes them. The address settings are its id_str (ip=static;address=..;gateway=..;dns=..;join=manual, network::NetworkConfig), which wpa_cli passes on: /usr/libexec/wifi-event sets a static address, route and DNS (the gateway's without servers of its own), or runs udhcpc with /usr/libexec/wifi-dhcp, which puts the network's servers in place of the router's. It also applies POWER SAVING, and logs to /var/log/wifi-event.log. Leaving the page applies changed address settings with wifi-event wlan0 REFRESH, without reconnecting.
    • AUTO-JOIN off is disabled=1. Choosing a network turns the others off in wpa_supplicant, so instead of ENABLE_NETWORK all each save turns each network on or off by its own setting (Control::save_config); turning off the connected one doesn't leave it. Lid sleep rejoins the network it left even with AUTO-JOIN off.
    • PRIVATE is mac_addr=3 with a mac_value made up when it is chosen (local, unicast); wpa_supplicant shows it only as *, and 2.12 didn't save it, so patch wpa_supplicant/0001. rtw88 changing its address while Wi-Fi ran left the link associated but sending nothing two times in seven (only S42wifi restart brought it back), so a MAC change restarts Wi-Fi instead, six of six clean, and a new address on each connection (mac_addr=1) isn't offered.
    • Start opens WI-FI OPTIONS (wifi_pages::WifiOptions): HOSTNAME, COUNTRY, POWER SAVING and ADD A HIDDEN NETWORK (name, then password, empty for open; scan_ssid=1). The hostname (hostname=, lowercase letters, digits and inner hyphens) is set at boot by S16hostname after TETSU-DATA is mounted, with /etc/hosts; changing it runs S16hostname reload, which renews DHCP with the new name and restarts mdnsd. mdnsd (Buildroot's, with its SSH service) answers to <hostname>.local on Wi-Fi and USB; from the PC it resolved over both, the router passing it on over the LAN. COUNTRY (wifi_country=, network::COUNTRIES, 55 by name in the language in use, AUTOMATIC first) is set by S42wifi with iw reg set, 00 for AUTOMATIC, after which the router's own country still applies. While connected, a string of changes left cfg80211 with an intersection of them (country 98), so a new country applies on leaving the page by restarting Wi-Fi. POWER SAVING (wifi_power_save=) is iw dev wlan0 set power_save, at once and on each connect. RESET SETTINGS puts all four back and applies them.
    • Init scripts read these with /usr/libexec/setting KEY DEFAULT, which also falls back to settings-backup.txt.
    • Checked on the device by Claude: on and off, the page and details, a static address (pinged from the PC), DHCP with own DNS, AUTO-JOIN off saved as disabled=1, PRIVATE and back with a restart each, a hidden open network that failed after 30 s and was removed, forgetting, a hostname change (DHCP and .local), Germany and Vietnam and back to AUTOMATIC, power saving off, and all of it after a reboot. Not checked: lid sleep with AUTO-JOIN off, a real hidden network, the German wording on the device

Not working yet:

  • A faint high-pitched noise with USB connected. VBUS read 4.42 V at a 2 A input limit on a PC port, so it may be overloading the port. Not heard as of 2026-11-08, so not diagnosed. The user will say if it returns

  • USB host mode is disabled, so USB accessories don't work

  • The wrong-password message over Wi-Fi is untested

  • An occasional single audio underrun during play, sometimes at the start, seen in /var/log/launcher.log. It was there without the menu. Not yet investigated

The kernel is trimmed to the Allwinner platform with modules off, and the user confirmed that it boots and that everything still works, including power-off. The image went from 40.6 MB to 19.8 MB, and the root filesystem contents from 94 MB to 7.9 MB. The kernel now reached init at 2.16 seconds, down from 2.86. Removing console=ttyS0 from extlinux.conf then brought that to 0.52 seconds, because sending messages to that port at 115200 baud took most of the time. With a serial adapter, add it back to see kernel messages on that port.

The root filesystem contents went from 255 MB to 40 MB. Buildroot's Panfrost option forces LLVM, and through Mesa's OpenCL option Clang and libclc, but Panfrost's shaders are precompiled on the PC. external.mk builds Mesa with -Dllvm=disabled, and post-build.sh deletes the unused libraries and tools, then fails the build if anything links to a missing library. Claude checked over SSH that kmscube still runs at 59.94 fps on Panfrost. The rootfs partition is still a fixed 1 GiB.

  • Translations (2026-11-09): Settings > GENERAL > LANGUAGE switches the launcher and the in-game menu between English and German at once, with no restart. It is kept in settings.txt as language=en or de, and RESET SETTINGS leaves it, so the menus stay readable to whoever is using the device. i18n::tr("SETTINGS") looks the English text up in a language file, and i18n::trf("{} OF {}", &[&a, &b]) also fills in values, in order; a text with no translation stays English. Screens ask as they draw, which is why the switch is immediate. Each language is a plain text file built into the launcher with include_str!, src/launcher/lang/de.txt, in lines of ENGLISH = TRANSLATION, so there is nothing to go missing on the device. To add a language: add the file, a variant in settings::Language and a line in i18n::table. Console names, TETSU OS and units stay as they are, and the language names are never translated. The font has one fixed width and everything is in capitals, so [max N] sections in the file give the room each kind of text has: 11 characters on a tile, 53 on a full line. That is why German says OPTIONEN rather than EINSTELLUNGEN, as the user chose short wording over changing the layout. The tests in i18n.rs fail on a text on screen with no translation, an entry nothing uses, a lost {}, a character the font can't draw and a text too long for its section. They find the English keys by reading the source for tr("..."), plus the lists of labels in keys_elsewhere, so text in a new list of labels has to be added there. Deleting a line, an overlong tile and a character without a glyph each failed as they should. They can't notice a string that was never wrapped in tr, which needs a look at the screen. The font gained Ä, Ö, Ü and ß, which also fixes these letters showing blank in game and network names. Other accented letters are still blank. Ä, Ö and Ü are the plain capital with two dots drawn two rows above the line (font::dots_above): squeezed into the font's 7 rows with their dots, they looked lowercase beside other capitals, as in ZURüCK. Lowercase ä, ö and ü, for network names, fit their dots in the glyph. LAUNCHER_PREVIEW_LANGUAGE=de renders the --preview* screens in German on the PC. Tested on the device on 2026-11-09 by switching with the buttons and screenshots of each settings group, the settings grid, Wi-Fi, system info, the home tiles list, a game list and a confirmation, then a launcher restart; launcher --preview-pause on the device showed the in-game menu reading the language from the file. The translation was written by Claude, and the user checked the wording and approved all of it on 2026-11-09. Not seen on the device by Claude: USB STORAGE's screens, the box art download, the keyboard, full-screen messages such as powering off, and the in-game menu inside a running game. A full-screen message drops from the large text to a smaller size when it wouldn't fit (screens::message_screen), which some English ones already needed: THE CARD IS IN USE - TRY AGAIN ran off both sides. Checked by a unit test, not seen on the device.

  • Setup on a new install (2026-11-09): before the first home screen, four steps on one screen each (setup.rs, drawn by screens::setup_screen): LANGUAGE, WI-FI, TIME and GAMES. Up and Down on the language step switch the language at once. WI-FI opens the normal Wi-Fi screen and GAMES opens USB STORAGE, and coming back from either selects the way on. TIME has the time zone and the clock format, and shows the time as it will read. A goes on, B goes back, and Start leaves from any step, because everything chosen here is also in Settings. Choosing German preselects Berlin when no zone is set yet; English stays on UTC. Settings > ADVANCED > RUN SETUP AGAIN starts it again. It runs when settings.txt doesn't exist (settings::from_file) and neither does settings-backup.txt, which only a new install has. A file from before the setup has no setup line, so existing devices don't get it after an update, and an unreadable file doesn't ask either. While it runs the file has setup=pending, removed on finishing, so powering off partway starts it over, with the choices so far kept. RESET SETTINGS doesn't start it. If TETSU-DATA can't be mounted the file is also missing, so the setup shows on each boot until the card is fixed. The screen has no clock in the corner, as the time may be wrong until the network sets it. It doesn't sleep when idle, like the Wi-Fi screen; the lid, the volume buttons and powering off on an empty battery work as elsewhere. launcher --preview-setup 0..3 FILE.ppm renders a step on the PC. Tested on the device on 2026-11-09 by removing settings.txt with the launcher stopped and restarting it, with screenshots: language switching, the Wi-Fi screen and its way back, Berlin preselected, changing the zone and clock, a restart partway coming back to the first step with setup=pending and the choices kept, Start finishing it, no setup after the next restart, and RUN SETUP AGAIN. The user then went through it on a card with an empty TETSU-DATA, the same day, and the logs showed: setup on screen 5.77 s after power on, a Wi-Fi network joined and saved, the clock set by ntpd, USB STORAGE ejected by the PC with a clean check of the card, and the setup finished. The user confirmed that the game count went up after USB STORAGE, the menus were silent with menu sounds off, the setup stayed away after a reboot, and its German wording reads well. Not tried: the lid during it.

  • Screenshots. SCREENSHOT in the in-game menu saves the paused picture as a PNG in /mnt/data/screenshots (screenshot.rs, png::encode), and the menu stays open, so it can be followed by anything else. The file holds the game's own pixels at the core's size, without scaling, screen effects, the menu or any sign: video_refresh keeps a copy of each frame it shows (Host::source), as a core's pointer is only valid during the call. That copy runs on every frame; its cost hasn't been measured, and is a memory copy of at most about 600 KB (PlayStation at 640x480). Consoles with non-square pixels, such as the SNES and PlayStation, are saved as their pixels are, so a PC shows them a little off the 4:3 shape they have on the device.

    Files are named after the game's save file and a number: Tetris.gb 001.png. The save name already tells Tetris.gb from Tetris.nes, stays the same when a game is zipped, and suits grouped discs (G.m3u 001.png). A number and not a time, because the clock can be wrong before the first network time. The next number follows the highest there, so an older picture is never replaced.

    A game with no cover shows its newest screenshot in lists and on its details screen (BoxArt::get), unless Settings > BOX ART > SCREENSHOTS AS COVERS is OFF (on by default, saved as screenshot_covers=), which shows the console's icon instead. It's shrunk to the 200x300 of a cover by the same averaging and given the console's shape. It is read at display time and never copied to boxart, so BOX ART still counts that game as having no cover, DOWNLOAD MISSING COVERS still fetches one, and a real cover then takes over. DELETE ALL COVERS leaves screenshots. So does deleting a game, as DELETE GAME AND SAVES says nothing of them and they are the player's own pictures; a game put back later shows its old screenshot as a cover. Screenshots can be taken from the card with USB STORAGE or copied over SSH.

    The menu gained a row, so its panel grew to 450 lines (PAUSE_PANEL), keeping the row height. Claude tested on the device on 2026-11-09 with Super Mario Bros. Deluxe (160x144) and Super Mario World (256x224), with their covers moved aside. Saving took 0.02 to 0.03 s and files of 3 to 10 KB. Two shots in a row got 001 and 002. The World shot had no CRT lines with the effect on, and both looked right on the PC. Each showed as the cover right after quitting, World at 200x150; putting its real cover back replaced it on returning from the home screen. A file in the folder's place gave COULD NOT SAVE THE SCREENSHOT with the game still running. German was checked in the --preview-pause picture. Saves, RECENT and covers were put back and the test pictures removed. SCREENSHOTS AS COVERS was then flipped on the device with a game whose cover was moved aside: OFF showed the console's icon and ON the screenshot, each after going through the home screen. The other consoles, including PlayStation's wide modes, a ZIP and a full card, are covered by unit tests at most, or not at all

  • The screenshot gallery (library/gallery.rs). A SCREENSHOTS tile on the home screen, before SETTINGS, opens every picture in /mnt/data/screenshots; it is there only while there are any, and HOME SCREEN TILES can hide it like the others (screenshots in hidden_tiles=). It has no count under it, as the tiles count games. A game's details screen has a SCREENSHOTS row with that game's count, only while it has some, which opens the gallery with just its pictures. The details screen's facts and rows moved closer together (22 and 26 lines) so its five rows and the delete question still fit above the hint.

    The gallery shows thumbnails 4 by 3, newest first by the file's time, a game's own by their number when the times are the same (FAT keeps them to 2 seconds). Each is drawn in its console's shape, as covers are, in a 136x102 cell; a file that can't be read shows its console's icon. Below the grid are the game's icon, its name as the library has it and the number from the file (SUPER MARIO BROS. 001), and the place among them (3 OF 14). A file not named the menu's way shows its own name and a camera icon, and a game no longer on the card its file's name. The arrows move, L1 and R1 go a page; A shows the selected one large, up to 600x322 and enlarged with the same averaging as covers, where Left and Right step through all of them and B goes back to the grid. Y and then A deletes the selected one (PRESS A TO DELETE THIS SCREENSHOT; anything else calls it off); the next one is selected, and deleting the last leaves the gallery. Thumbnails are read when their page shows and kept while the gallery is open. Leaving after a delete reads the folder again and forgets the covers, so a game whose cover was that picture shows its next newest or its icon. The gallery is its own loop, like SET TIME, opened by run_launcher because it names the games from the library. Matching a picture to its game only asks games whose file is named like it, as a ZIP's save name means opening it; each game's name is remembered in Library::screenshot_bases.

    Claude tested on the device on 2026-11-13 with injected presses and screenshots, with copies (number 900) of the user's two pictures: the tile and its line, the grid with the copies first, the large view of a Game Boy picture, Y with its question and A deleting it, then Super Mario Bros.' details showing 2, its gallery with only its two, deleting one there and the row showing 1 on the way back. The user's pictures were left as they were. Not tried: more than a page of pictures, a broken file, the lid, the German wording on the device (checked in --preview-gallery with LAUNCHER_PREVIEW_LANGUAGE=de). The previews are --preview-gallery and --preview-gallery-full.

  • Button remapping and turbo. CONTROLS in the in-game menu (controls.rs, canvas::controls_screen) opens a page with a row for each of A, B, X, Y, L1, R1, L2, R2, SELECT and START. Left and Right choose what the button presses in the game: any of those, a d-pad direction or NONE. A turns turbo on or off for that button, so any number of buttons can have it. TURBO SPEED is FAST, NORMAL or SLOW, for the whole layout: 2, 3 or 5 frames pressed and as many released, so about 15, 10 or 6 presses a second at 60 fps. One frame each way isn't offered, as a game that reads the buttons every other frame could see no change. The d-pad itself and Menu can't be changed, so no layout leaves a game unplayable or the menu out of reach. The button names are the device's, whatever a console calls them, and a target the console doesn't have does nothing.

    SET FOR says whom changes are for. THIS CONSOLE keeps a line like controls.gb=normal,x:right*,r2:none in settings.txt, holding the speed and what differs from the default (* is turbo), and the default has no line. THIS GAME copies what the game had been using into Tetris.gb.srm.controls beside its saves, which then wins over the console's. It follows the save's name through zipping, and DELETE GAME AND SAVES removes it, while it counts neither as a save nor as last played. Going back to THIS CONSOLE removes the file, though the page remembers it until it closes, in case that was a slip. RESET THESE CONTROLS empties the layout shown. RESET SETTINGS removes console lines and all games' files. A line or file with anything unknown is ignored as a whole. Changes are saved and in use as they're made.

    Pad::update works out what the game sees once per frame, in Core::run_frame, and input_state only reads that, so all a core's reads in a frame agree and fast-forward's extra frames count as frames. A turbo button's cycle starts over with each press, so a quick tap always reaches the game. Host::controls stays the real buttons, for Menu + Start and the menu. R2 fast-forwards only while its row is unchanged; a changed R2, including R2 with turbo, is for the game alone, which the page notes on that row. A PlayStation game with R2 unchanged still sees it and fast-forwards, as before. There is no global default, and nothing in the launcher's Settings: controls are set from inside a game.

    The menu gained a row, so PAUSE_ROW went from 30 to 27 lines to stay on its panel. --preview-controls renders the page. Claude tested on the device on 2026-11-09 with Tetris and a save in /tmp, by injected presses and the framebuffer: X set to START passed the title screen, and Start still worked. On the level grid, holding X as RIGHT for about 0.6 s moved one cell, and with turbo at NORMAL six. Held R2 showed the fast-forward sign, and set to SELECT didn't. THIS GAME wrote its file and left the console's line alone, a restart opened on it, THIS CONSOLE removed the file, and RESET removed the line, bringing fast-forward back. German was checked in the previews. settings.txt was put back afterwards. How turbo feels in a real game, other consoles' cores and Menu + Start with a layout (press.py can't hold both) are not checked

  • Core options. Each core's own settings, such as FCEUmm's sprite limit and palette or mGBA's Game Boy model, are on a CORE OPTIONS page in the in-game menu and under Settings > CORES (core_options.rs, core_options_screen.rs, screens::option_rows). Every option the core declares is listed under the core's own categories, with its own names, choices and help text, which are in English whatever the language. Left and Right go through the choices, A opens a category or steps a choice, B goes back, and RESET ALL TO DEFAULTS empties the choices shown. The selected option's help text shows under the rows, cut with .. after three lines; long names and choices are cut the same way.

    The frontend now answers GET_CORE_OPTIONS_VERSION with 2 and takes SET_CORE_OPTIONS_V2 and _V2_INTL (reading us; the frontend reports no language, so local is empty). Cores built with libretro_core_options.h send only _V2_INTL and don't fall back if it is refused. SET_VARIABLES is read too, for Tamalib, which declares nothing. Since 2026-11-13 the first interface, SET_CORE_OPTIONS and SET_CORE_OPTIONS_INTL (definitions without categories), is read too, for PokeMini and Beetle VB, which send it after a version of 2 and fall back to nothing when it is refused. Answering their options showed that Beetle VB calls its log callback unchecked, so the frontend now answers GET_LOG_INTERFACE with launcher_core_log from core_log.c, in C as it takes variable arguments. It prints warnings and errors to the launcher's log, and stops after 300. Handy's and PrBoom's log patches are older and still harmless. Stella logs one File open/read error as it starts, and the game runs. SET_CORE_OPTIONS_DISPLAY hides options the core says don't apply, and the UPDATE_DISPLAY_CALLBACK is called after each change, so the page follows. Options with one choice, and mGBA's mgba_interframe_blending, which IN GAME > LCD GHOSTING sets, are left out. FRONTEND_DEFAULTS (SGB borders off, colour correction Auto, FCEUmm's Smooth FBX palette, no second PlayStation memory card) are the defaults the page shows and RESET goes back to.

    What a core gets for an option: its ghosting setting, then the choice if the core still offers that value, then the frontend's default, then the core's declared default for cores with declared_defaults. Cores without it (mGBA, FCEUmm, Snes9x, Tamalib) still get nothing for options nobody chose, as before, except an option chosen earlier in the same run, so going back to the default reaches the core. Values the core is given are kept for the whole run (Host::answers), as a core may hold the pointer.

    SET FOR works as on CONTROLS: THIS CONSOLE keeps key=value lines in /mnt/data/core-options/<console>.txt, only for values that differ from the default, and removes the file when none are left. THIS GAME copies what the game had been using into Tetris.gb.srm.options beside its saves, where an empty file means every default. Going back to THIS CONSOLE removes it. DELETE GAME AND SAVES removes the game's file. It counts neither as a save nor as last played. RESET SETTINGS removes the folder and every game's file. Consoles that share a core (Game Boy, Color and Advance on mGBA) keep their own choices. Changes are saved and handed to the core as they are made. The core reads them through GET_VARIABLE_UPDATE on the next frame, but some only apply on the next start, which the page says and which mGBA marks with (RESTART) in the name. Some options do nothing here, such as FCEUmm's aspect ratio, as the frontend sets the shape itself.

    Settings > CORES lists the 16 consoles, and each opens its core's options for the console, without SET FOR. The launcher never loads a core: launcher --core-options CORE calls only retro_set_environment, where every core here declares its options, and prints them as tab-separated lines with what the core hid. The launcher keeps the result while it runs. FCEUmm declares its options again on loading a game, with its dipswitches. The tile is labelled CORES, as CORE OPTIONS is too long for a tile; it is the ninth, on a row of its own. The in-game menu gained a row, so PAUSE_ROW went from 27 to 25 lines. The in-game page uses a wider panel than the menu (600 lines), as names come from the cores. --preview-core-options renders it with sample options.

    Claude tested on the device on 2026-11-09. --core-options counts: FCEUmm 44 options in 6 categories, Genesis Plus GX 62 in 6, Beetle PCE Fast 32 in 5, PCSX ReARMed 60 in 7, Snes9x 43 in 3, mGBA 17 in 5, Beetle Cygne 16, Handy 7, Beetle NeoPop 1, Tamalib none. Tetris was run with its save in /tmp. The page showed mGBA's categories. Changing the palette to DMG Green coloured the game on resuming and wrote gb.txt, and a restart started green. THIS GAME with RESET gave an empty game file and grayscale, while gb.txt stayed. THIS CONSOLE removed the game's file and brought green back, and RESET removed gb.txt. Super Mario Bros. showed Smooth FBX as its palette, and NES Classic took effect on resuming. Settings > CORES listed the consoles and NES's categories. Everything was put back, and the folder removed. Not checked: German wording on the device (the previews were), a core's hidden options changing during a game, PCSX ReARMed and Genesis Plus GX in a game, and options that only apply on a restart for each core

  • Cheats. CHEATS in the in-game menu lists the codes of the game's cheat file, each ON or OFF (cheats.rs, canvas::cheats_screen). A, Left or Right turns the selected code on or off, L1 and R1 move 12 rows, TURN ALL OFF does that, and DOWNLOAD CHEATS fetches the game's file. Under the rows is the selected code's whole description, which its row may cut; with no codes, where to copy a file. The page uses the CORE OPTIONS panel and its row drawing, now shared through screens::RowList. Both wide pages' titles moved up 8 lines, so the arrow for rows above the page no longer overlaps them.

    Files are RetroArch .cht files, in /mnt/data/cheats/<console key>/ on TETSU-DATA (cheats/gb, cheats/snes, ...). A game's file is the one named after it (its file name without the extension, so a ZIP's name), or else the one whose title matches, compared on letters and digits only, with the most tags in common, as box art is chosen. Quoted values may span lines, as some descriptions in libretro-database do; they are shown on one line. Codes without a code, and RetroArch's own memory-search cheats (handler = 1, such as the (Rumbles) files), are left out, as this frontend has no cheat engine of its own: the core applies each code itself through retro_cheat_set, in whatever formats it reads (Game Genie, Pro Action Replay, GameShark, Code Breaker or raw address:value). The CHEATS item shows only for consoles whose core does that (System::cheats): not the Lynx, Neo Geo Pocket or Tamagotchi, whose cores ignore codes.

    Turning a code on or off calls retro_cheat_reset and then gives the core the codes that are on, numbered from 0, as RetroArch does: mGBA takes every code it gets as on. The codes that are on are saved by their code in Tetris.gb.srm.cheats beside the saves, so a new download keeps those still in the file, and the file is removed when none are left. A start gives the core those codes again after the resume state is loaded, leaving out any no longer in the cheat file. The enable lines of a .cht file are ignored: codes only come on from the page. Like a game's controls, the file counts neither as a save nor as last played, DELETE GAME AND SAVES removes it, and RESET SETTINGS removes every game's list; the cheat files themselves stay.

    DOWNLOAD CHEATS runs on its own thread while the menu stays usable, and the row reads DOWNLOADING... meanwhile. It first asks raw.githubusercontent.com for libretro-database/master/cht/<console folder>/<game>.cht, which needs no listing, then GitHub's listing of the folder (git/trees/master:cht/<folder>, one API call, of the 60 an hour allowed) for the closest name. A file without usable codes counts as none. It is saved under the game's own name, so it is found first from then on. The result shows in the accent colour: 37 CHEATS DOWNLOADED, NO CHEATS FOUND FOR THIS GAME, or the box art download's messages, such as NO CONNECTION TO GITHUB. The folders hold 1,500 to 2,300 files per console, named in several ways: No-Intro names, bare titles, and tags for the device, such as (Game Genie) or (GameShark).

    mGBA's retro_cheat_reset freed its codes without undoing their changes, so a Game Genie or other ROM patch code, and a GBA hook code's breakpoint, stayed after being turned off. package/libretro-mgba/0001-undo-cheats-on-cheat-reset.patch disables, removes and frees each set instead. Snes9x and Genesis Plus GX restore the ROM themselves; FCEUmm's codes are read through its cheat list, which reset empties.

    Claude tested on the device on 2026-11-09, running Tetris with its save in /tmp. Real downloads from the PC: Super Mario Bros. (World) found by name (711 codes), Tetris (World) (Rev 1) (37), Pokemon Emerald and Crash Bandicoot by the closest name. On the device the menu showed CHEATS, the empty page its note, and a download with the Wi-Fi not passing traffic NO CONNECTION TO GITHUB; after a reboot it saved 37 codes and listed them. Game Genie "Glitched Graphics" scrambled the next screen. GameShark "No Piece Display Box" emptied the next-piece box, and "Always Get Square Pieces" dropped squares and showed a piece in the level select, which TURN ALL OFF ended. Both came back after quitting and starting again, with a code no longer in the file left out (cheats: 1 on). "Glitched Graphics" on from a fresh start scrambled the level select after the title screen. Turned off on the copyright screen, the unpatched core still scrambled it, and the patched core showed the game type screen as normal. Not checked: the other cores, a PlayStation code, a ZIP, the German wording on the device (the previews were)

  • RetroAchievements (achievements.rs), through rcheevos v12.5.0, the library RetroArch uses, copied into src/launcher/rcheevos/ (MIT; README.md there says what was left out and how to update) and compiled into the launcher by build.rs with the cc crate. Buildroot's TARGET_CONFIGURE_OPTS gives it the cross compiler. The FFI is written by hand, mirroring only the structs and fields read. rcheevos/libretro/libretro.h is the part of libretro.h that rc_libretro.c needs: the memory constants and the memory map.

    Settings > ACHIEVEMENTS (the tile says TROPHIES, as ACHIEVEMENTS is one character too long for it) has ACCOUNT and HARDCORE. A on ACCOUNT types the user name and then the password on the on-screen keyboard and logs in on a thread with rc_client's password login. A failed login goes back to the password keyboard with the name and password still typed and the reason, such as WRONG NAME OR PASSWORD, above the field; B on an empty password goes back to the name, also still typed, and B on an empty name leaves. Only the name and the token the server returns are kept, in /mnt/data/retroachievements.txt, mode 600, outside settings.txt, so RESET SETTINGS keeps the account. Logged in, the row shows the name, and A asks PRESS A AGAIN TO LOG OUT, which removes the file. HARDCORE is a setting, off by default (hardcore=on).

    ACHIEVEMENTS, the group's first row, turns them off for every game (achievements=off), keeping the account. A game's details screen (Y in a list) has an ACHIEVEMENTS row, ON or OFF like a setting, which A, Left or Right switches, for games of consoles RetroAchievements covers (Library::achievement_systems), which leaves an empty Tetris.gb.srm.achievements-off beside the saves. Both apply from the game's next start: a game with achievements off starts no session, so it has no notices, no ACHIEVEMENTS item, and none of hardcore's rules. The file counts as a setting, not a save, DELETE GAME AND SAVES removes it, and RESET SETTINGS removes every game's. The details rows moved up 4 lines to fit a fourth.

    A game process with an account, on a console RetroAchievements covers (console_id; all but the Tamagotchi), starts a Session after retro_load_game: it logs in with the token and calls rc_client_begin_identify_and_load_game with the game's path, which rcheevos hashes itself: the file unpacked from a ZIP, an .m3u's first disc, a .cue's data track. rcheevos' own disc reader handles .cue/.bin and .iso, not .chd, so a CHD game is NOT RECOGNISED. Changing a disc of an .m3u game calls rc_client_begin_identify_and_change_media with the disc's path.

    Memory: RETRO_ENVIRONMENT_SET_MEMORY_MAPS (36) is answered, and the map is copied with its address space names, as a core's own may not outlive the call. rc_libretro_memory_init then lays the console's address space over the map, or over retro_get_memory_data's regions for a core without one; it knows each console's layout. A new map lays the memory out again, as RetroArch does: mGBA gives its map only in its first retro_run (its deferred setup), and before that only its work RAM, so high RAM at $FF80 was missing and rcheevos disabled every Super Mario Land achievement at load (Invalid address 00FF9F); a_memory_map_given_after_the_start_is_used tests it. A core that shows no memory yet is asked again every 60 frames while the game loads. rc_client_do_frame runs after every retro_run, fast-forward's extra frames too, and rc_client_idle every 100 ms while the menu is open.

    Requests: rc_client hands each request to server_call, which runs it on a thread with ureq (HTTPS only, 15 s to connect, 30 s in all) and the user agent TetsuOS/0.1.0 (Linux) rcheevos/12.5.0. The answer goes back through a channel, and Session::pump, called every frame, gives it to rcheevos on the game's thread, so rcheevos and every callback run on one thread. A request that never reached the server answers RC_API_SERVER_RESPONSE_RETRYABLE_CLIENT_ERROR, so rcheevos retries unlocks itself. Leaving the game waits up to 5 s for requests still on their way.

    Offline (2026-11-13, achievements_offline.rs): every request goes through offline::answer. The answers a game needs to start, the token login (login2 with a token, never a password), the game's achievements (achievementsets, and gameid for a disc change) and the session with the user's unlocks (startsession), are kept as /mnt/data/achievements/offline/<user>-<request>-<game hash>.json whenever the server gives them. When the server can't be reached (no answer, or a 5xx), the kept answer is given to rcheevos instead, so a game played once online has its achievements offline, with OFFLINE - UNLOCKS WAIT FOR A CONNECTION after the loaded notice. A game never played online still fails with NO CONNECTION TO RETROACHIEVEMENTS. An unlock that can't reach the server is added to waiting.txt there (user, ID, hardcore, game hash, Unix time; no token) and rcheevos is told it went through, so it doesn't retry it; it is added to the kept session too, so the game starts with it unlocked next time offline, and an unlock that reaches the server is added there as well. The notice UNLOCKS WAIT FOR A CONNECTION shows once, for a connection lost during a game. Leaving a game (Session::finish, 2026-11-13) waits at most 0.5 s (FINISH_TIMEOUT), and only while an unlock is being sent (offline::unsent); those still unanswered then go to waiting.txt too (keep_unsent; one the server took meanwhile is answered User already has when sent). A login, achievementsets, startsession or ping still on its way is left behind, as the next game asks again. It used to wait up to 5 s for any request: with Wi-Fi hung (see Known issues in CLAUDE.md) each DNS lookup takes about 12 s to fail, so nearly every quit waited the full 5 s, while 0.36 s was measured from QUIT to the menu drawn with working Wi-Fi. With DNS pointed at an address that never answers and the game's achievements still pending, the game process now ends 0.17 s after the press. sync sends the waiting unlocks, rebuilt with rcheevos' own rc_api_init_award_achievement_request and the token of the account logged in, with o= the seconds since the unlock (signed into v=), so the server dates them when they happened: before each startsession (WAITING UNLOCKS SENT then shows), and every minute from a thread in the launcher's menus (WAITING_RETRY). Taken, or User already has, removes one; no connection, 401 or 403 stops and keeps the rest; any other refusal drops that one with a log line. waiting.lock (flock) keeps the launcher and a game from writing the file at once. Unlocks of another account wait until it logs in again; logging out removes the kept logins, which hold the token. Not offline: GAME COMPLETED and MASTERED (rcheevos takes them from the server's AchievementsRemaining), leaderboard entries (retried only while the game runs), and the player's score, which reads 0 after an unlock offline. An unlock dated by a wrong clock (no network time yet) is sent with that age. Tests: a_game_played_before_unlocks_offline_and_sends_it_later (a whole session online, offline with an unlock, offline again, online) and achievements_offline's own.

    Notices show at the bottom of the screen, centred, for 4 to 6 s, through the framebuffer's sign, like the volume, and one after another: the game loaded (3 OF 40 ACHIEVEMENTS UNLOCKED - HARDCORE over its title), an unlock with its badge and points, progress towards one (2 s, a newer one replacing any waiting, and only as it reaches 25, 50 and 75 % of the way, as rcheevos reports every step: Super Mario Bros.' 180 jumps showed a notice on each jump), GAME COMPLETED or GAME MASTERED, leaderboards started, failed and submitted (only in hardcore, where rcheevos runs them), a game the server doesn't know or that has no achievements, a failed login (WRONG NAME OR PASSWORD, LOG IN AGAIN IN SETTINGS for an expired token, NO CONNECTION TO RETROACHIEVEMENTS), and unlocks waiting for a connection and sent later. Leaderboard trackers and challenge indicators aren't shown.

    Each kind of notice has its own place: UNLOCKS (NEUE ERFOLGE), PROGRESS (FORTSCHRITT) and OTHER MESSAGES (SONSTIGES), everything else, from the loaded game to HARDCORE OFF FOR THIS GAME. Each is BOTTOM (the default) or TOP, 8 pixels from the edge either way, or HIDDEN, which leaves that kind out (Notice::kind). Hidden achievements still unlock, and the in-game menu's list still shows them. Settings > ACHIEVEMENTS has the three for every game (notices_unlocks=, notices_progress=, notices_other= in settings.txt). In a game, the in-game menu's ACHIEVEMENTS opens a list of LIST (the achievements page), then the same three for that game, each DEFAULT (STANDARD), which follows Settings, or a place of its own. The full Settings value didn't fit beside the German labels, so DEFAULT shows alone. A game's choices are kept as Tetris.gb.srm.notices beside its saves (progress=hidden lines, removed once all are DEFAULT), which count as game settings like .achievements-off: DELETE GAME AND SAVES and RESET SETTINGS remove them.

    Badges are downloaded once the game has loaded, on a thread, one after another, as /mnt/data/achievements/<badge name>.png, stopping at the first failure to connect. The page shows them at 32 pixels, averaged down from 64, and grey and darker while locked, made from the same file; a notice shows them at 64.

    ACHIEVEMENTS in the in-game menu, after OPTIONS, shows for a game with a session, and opens a list whose LIST is the achievements page. The page, on the CORE OPTIONS panel, has 2 OF 6 UNLOCKED - 6/101 POINTS - HARDCORE and the achievements in rcheevos' groups (ALMOST THERE, LOCKED, UNLOCKED, ...), each with its badge, title and points, or its progress such as 73/100; Up and Down move, wrapping, L1 and R1 (or Left and Right) a page, and the selected one's description is below. While loading, or for a game without achievements, it says so. The in-game menu's rows went from 23 to 22 lines, so the fifteenth item fits.

    The game's count is saved as Tetris.gb.srm.achievements beside its saves when it loads and after each unlock; a game's details screen shows ACHIEVEMENTS 3 OF 40 from it. It counts neither as a save nor as last played, and DELETE GAME AND SAVES removes it.

    Hardcore (Session::restricts): unlocks count as hardcore, the game starts from its save, not its resume point, LOAD STATE says HARDCORE ALLOWS NO LOADING, the CHEATS item and saved codes are left out, and R2 doesn't fast-forward. Saving states still works. The restrictions end once the game turns out to have no achievements or isn't recognised, but the resume point was already skipped. Core options that RetroArch disallows in hardcore (rc_libretro_get_disallowed_settings: a PAL region for FCEUmm, Snes9x, Genesis Plus GX and PCSX ReARMed, Game Genie for FCEUmm, Snes9x's layer and clipping options, PCSX ReARMed's lower clock) switch the game to softcore at start, or when changed on the CORE OPTIONS page, with HARDCORE OFF FOR THIS GAME: CORE OPTION and the option's name. Loading a state in softcore calls rc_client_deserialize_progress_sized with nothing, which restarts the achievements from the loaded state, so its memory jump isn't counted.

    Tests: cargo test achievements runs a whole session offline against canned server answers taken from rcheevos' own tests (a test-only FAKE_SERVER): password login and a refused one, token login, hashing the file, loading the game, reading a fake 2 KiB NES RAM through rc_libretro, an unlock with its notice, and the saved count. Previews: --preview-achievements and --preview-notice. --render also prints how much memory rcheevos finds at load and after the frames (achievement memory at load: N bytes in M regions); only mGBA's Game Boy games differ, 5 regions at load and 11 after.

    The server adds a WARNING: UNKNOWN EMULATOR achievement for this client, as it does for one it doesn't know by its user agent (TetsuOS/0.1.0 (Linux) rcheevos/12.5.0), and rcheevos unlocks it at once and leaves it out of counts. Being known takes registering the client with RetroAchievements; borrowing another emulator's user agent isn't an option.

    Claude tested on the device on 2026-11-09: the settings grid with the TROPHIES tile, the ACHIEVEMENTS group logged out with its notes, and the user name keyboard, left with B. --render of 300 frames, with saves in /tmp, found memory for every core with a game on the card, each the size of rcheevos' layout for the console: FCEUmm 65536 bytes in 8 regions (from 14 descriptors), Genesis Plus GX 40960 for the Master System and Game Gear and 131072 for the Mega Drive (no memory map: retro_get_memory_data), Beetle PCE Fast 8192, mGBA 212992 for the Game Boy and Color and 360448 for the Advance, Snes9x 657408, PCSX ReARMed 2098176. The user logged in on the device; unlocks in softcore work, and hardcore is refused while the server doesn't know the client. Super Mario Land loaded with all 34 achievements disabled until the late memory map was used, then with 34 of 35 active (the other is the warning). Not checked: an unlock reaching the server, notices and the page with an unlock, hardcore, the Neo Geo Pocket, WonderSwan and Lynx cores (no games on the card), a ZIP, a multi-disc game, the German wording on the device (the previews were), and anything over a real connection to the server

  • Card layout for A/B updates (2026-11-12): genimage.cfg has TETSU-BOOT (FAT, 128 MiB at 1 MiB), root slot A (TETSU-ROOT, 1 GiB, partition 2, bootable flag), an empty 1 GiB root slot B (partition 3) and, added by flash.sh, TETSU-DATA at 4 GiB as partition 4, which uses up MBR's four. The kernel and device tree are in each root slot's /boot (BR2_LINUX_KERNEL_INSTALL_TARGET), with extlinux.conf in /boot/extlinux/ (installed by post-build.sh). U-Boot's bootstd scans only the partitions with the bootable flag when any has it, and reads extlinux.conf from / or /boot/, so the flag picks the slot; extlinux.conf paths are absolute (/boot/Image), which U-Boot's extlinux bootmeth allows. TETSU-BOOT keeps README.txt, bootlog.txt and logs/. Slot B has no filesystem and isn't in sdcard.img (genimage stops the file at slot A's end), so flashing takes as long as before. For updates later: U-Boot 2026.10 has CONFIG_BOOTMETH_RAUC (off), which keeps BOOT_ORDER and tries per slot in its environment (CONFIG_ENV_IS_IN_FAT, on TETSU-BOOT) and runs a boot.scr from each slot's boot partition, which can be the root partition itself (CONFIG_BOOTMETH_RAUC_PARTITIONS="2,2 3,3"); an update writing slot B must change root=PARTUUID=54455453-02 in its extlinux.conf or script to -03. Checked in the image: the partition table, /boot in the root filesystem, the FAT's contents, and data-partition.sh on an 8 GiB card image (TETSU-DATA created as partition 4, then kept on a second run). On the device (2026-11-13), after the user flashed over the old layout: it boots from slot A (root=PARTUUID=54455453-02), TETSU-DATA mounts as mmcblk0p4 with its games and settings kept, and the kernel starts 2.59 to 2.66 s after power on, as before (2.53); the launcher drew at 6.58 and 6.60 s with the jingle. BOOT MESSAGES TEXT and back gave the original file byte for byte, and a reboot with TEXT had console=tty1 loglevel=7 on the kernel's command line, so U-Boot reads the file the device wrote. SAVE LOGS TO CARD and bootlog.txt reach TETSU-BOOT. push.sh boot copied, verified, rebooted and put LOGO back.

  • TETSU-DATA on the first start (2026-11-14): a card written with an image writer (balenaEtcher, Raspberry Pi Imager, dd) has only the image's partitions, so S15data, finding no TETSU-DATA, runs /usr/sbin/data-partition. It only acts on TF1's card with no fourth partition, at least 1 GiB past 4 GiB, partitions 1 to 3 ending before 4 GiB and the MBR's fourth entry empty. It writes that entry with dd (type 0x0c, 4 GiB to the card's end, as scripts/data-partition.sh does) and tells the kernel with util-linux's addpart, as BusyBox's partprobe re-reads the whole table, which fails while root is mounted. A FAT32 file system labelled TETSU-DATA found there is kept (a card that had Tetsu OS and was written again); anything else is erased (its first MiB) and formatted with dosfstools' mkfs.vfat -F 32 -n TETSU-DATA, as writing the image meant erasing the card. Each step goes to the kernel log, so bootlog.txt shows it. The image gained mkfs.fat (59 KiB), addpart (30 KiB) and libblkid, libuuid and libsmartcols; blkid stays BusyBox's. Checked on the device with a sparse 5.5 GiB loop disk holding the card's MBR (losetup -P): an empty area formatted in 0.26 s, a second run left the fourth partition alone, the entry removed again kept TETSU-DATA with a file on it, and a FAT labelled OTHER was formatted; a 64 GiB loop disk gave a 60.0 GiB TETSU-DATA. On the real card, with its fourth MBR entry cleared and a reboot: added /dev/mmcblk0p4 with the TETSU-DATA file system already there 1.15 s into the kernel, mounted with the games and settings, the launcher drawn at 4.67 s after power on against 4.43 s on the next, normal, boot. Not done on the real card: formatting, as it would erase the device's games. Its time is estimated from the FATs, 15 MB for a 60 GB card, which the card writes in 0.79 s.

  • The boot card identified again (2026-11-15): the v0.1 release card stopped 3 times in its first 4 cold starts at "Waiting for root device PARTUUID=54455453-02", and a freshly wiped and written SanDisk Ultra did on most cold first starts. The screen showed why: mmc0: new SD card at address d555 and mmcblk0: mmc0:d555 SD032 30.6 MiB with no partitions, against new high speed SDHC card at address aaaa, SD32G 29.7 GiB on a good start. SanDisk cards give aaaa as their first address after power up and d555 as the second, so the card was still in U-Boot's session: Linux's reset hadn't taken. TF1's supply is the always-on vcc-io, so Linux can't power cycle it. Taking the card out and putting it back, which does, made the boot carry on every time. Ruled out: the release build, Etcher, the data partition, U-Boot's environment file (written into the image beforehand, it still failed), the battery, a busy card (a kernel waiting on DATA0 found it never busy), and RAM to keep the log (it doesn't survive a reset here). A 2 s post-power-on-delay-ms booted once, but costs every start. Patch 0019 instead makes wait_for_root() identify the SD memory cards again when the root device is missing 1 s after the drivers are done, then every 2 s, 3 times at most: each card is dropped, its host powered down and up and rescanned. A good start finds its root before it ever waits (no "Waiting for root device" line), so it is unchanged: root mounted at 0.537 s, launcher at 6.65 s. Checked by Claude on the device with mmc_reidentify_test, which forces one round: the card was removed at 0.493 s, found again at 0.608 s with its four partitions, the root mounted at 0.695 s and the launcher drew at 6.99 s. Not yet seen: a real failing start recovering by itself. The user saw it only with SanDisk Ultra cards, never with a 64 GB SanDisk Extreme, and a reset always starts the device, so it is left as it is (2026-11-16), with a line in the release notes

  • Updates over Wi-Fi (2026-11-13): Settings > SYSTEM > SYSTEM UPDATE (update/) installs a new root filesystem, which holds the kernel too, into the root slot that isn't running, and keeps the running one to go back to. U-Boot is not updated this way.

    • Making one: scripts/make-update.sh gzips output/images/rootfs.ext4 (62 MB for the 1 GiB image; zstd would be 50 MB but needs another crate, and gzip is in flate2 already) into output/images/update/ with manifest.txt (format, device from the device tree, build, revision, file, file_size, image_size, image_sha256) and manifest.sig, the manifest's Ed25519 signature in hex. post-build.sh adds BUILD_ID=YYYYMMDD-HHMMSS (UTC) to os-release, the same as the manifest's build; the script checks with debugfs that the image is that build. The private key is outside the project (~/.config/tetsu/update-key.pem, made once with make-update.sh --new-key, or $TETSU_UPDATE_KEY); its public half is src/launcher/update-key.pub, compiled in, so the launcher must be built after the key exists. scripts/serve-update.sh serves the folder on port 8035 with Python; over the cable that is http://10.35.35.2:8035.
    • The screen: rows for this system's build, revision, root slot and the source, and two lines of status. A looks for an update (manifest.txt, manifest.sig next to it), which must have the project's signature, this device's compatible and a newer build than the running BUILD_ID (a system without one is older than any). A again installs it, unless the battery is under 30 % and not charging. X types the source on the keyboard (update_source= in the settings; empty brings back the one built in with TETSU_UPDATE_SOURCE, none so far). Plain HTTP is allowed, since the signature is what is trusted; redirects are followed, so GitHub release assets would work. During an install a bar shows the download, then the read-back; B twice stops it, and the lid only turns the screen off. At the end A reboots, B leaves it for the next start. SAVE LOGS TO CARD moved from SYSTEM's rows to X on SYSTEM INFO, to make room.
    • Installing (install.rs): the boot order first puts the running slot first and the other at 0 tries, so a half-written slot is never booted, even after a power cut. The download is unpacked and written to the other slot's partition (/dev/mmcblk0p3 or p2, found from the root device's disk) in 1 MiB chunks, synced every 64 MiB, with its SHA-256 taken; then the page cache is dropped (POSIX_FADV_DONTNEED) and the partition read back and hashed again. Both must match the manifest. The slot is then mounted to check boot/boot.scr, boot/Image and extlinux.conf, and its extlinux.conf gets root=PARTUUID=54455453-03 (or -02) and the running system's BOOT MESSAGES. Last, the boot order puts the new slot first with 3 tries and the running one behind it, and /mnt/data/update-pending.txt notes the slot and build.
    • Booting and falling back: uboot.fragment turns on U-Boot's RAUC bootmeth (BOOT_ORDER, BOOT_A_LEFT, BOOT_B_LEFT, 3 tries, slot A partition 2 and B 3, each its own boot partition) and bootflow scan; run distro_bootcmd as the boot command. Without CONFIG_BOOTSTD_FULL, bootflow scan refuses flags and a device ("Flags not supported") and returns at once, which a first try with -bG mmc0 did, so the old script booted and nothing was saved; plain, it boots the first bootflow, with the devices in boot_targets' order (mmc0 first). On the whole card the RAUC bootmeth comes before the extlinux bootmeth reaches the partitions. U-Boot takes a try off the first slot in BOOT_ORDER with any left, saves the environment and runs that slot's /boot/boot.scr (boot.cmd, made by mkimage from host-uboot-tools in post-build.sh), which sysboots the slot's own extlinux.conf, so BOOT MESSAGES still works per slot. A slot whose script fails falls through to the extlinux bootmeth, which boots the bootable partition, slot A. At the launcher's start (slots::mark_good_in_background, once per boot through /run/launcher-slot-marked, in the charging screen too) the running slot goes first with its 3 tries back, so only boots that never reach the launcher count; after 3 of them U-Boot boots the other slot. extlinux.conf now has panic=10 (was panic=0, which waits forever), so a kernel that panics restarts and counts as a failed try; a hang without a panic needs the power button held. If update-pending.txt names another slot than the running one, the update failed: SYSTEM UPDATE then says <build> DID NOT START, THIS SYSTEM WAS KEPT, and otherwise UPDATED TO <build>.
    • U-Boot's environment (boot_env.rs) is in two copies on TETSU-BOOT, uboot.env and uboot-redund.env (CONFIG_ENV_REDUNDANT), 64 KiB each: CRC32, a counter byte, then sorted name=value entries ending in NUL. The copy with a valid CRC and the higher counter (0 after 255) is current, and a save writes the other one, so a power cut mid-save leaves the current copy. Linux reads and writes them the way U-Boot does, changing only BOOT_* and keeping every other entry. The default CONFIG_ENV_FAT_DEVICE_AND_PART=":auto" meant the bootable partition, the ext4 root, where saving failed, so nothing was ever saved before; it is now ":1". U-Boot saves its whole environment, boot command included, and a saved boot command would outlive a later U-Boot's, so push.sh uboot removes both files when running slot A. A U-Boot without A/B never writes BOOT_ORDER, and then an install stops before writing anything: THIS CARD NEEDS ONE FLASH FOR UPDATES.
    • Tested on the device (2026-11-13), over the USB cable from serve-update.sh (the PC's firewall lets the device reach it there):
      • With the old U-Boot, the new system booted as before; the launcher logged that U-Boot keeps no boot order, found and verified the signed update, and its install stopped with THIS CARD NEEDS ONE FLASH FOR UPDATES before writing.
      • With the new U-Boot (push.sh uboot), U-Boot saved BOOT_A_LEFT=2 in uboot-redund.env and the launcher wrote uboot.env with 3 back and every other entry kept. The launcher drew 6.73 s after power on, against 6.60 before.
      • Installing a newer build into slot B took about 1 min 45 s for the 62 MB download, writing 1 GiB and reading it back; the copies showed B at 0 tries while writing, then BOOT_ORDER=B A. A restarted into slot B (root=PARTUUID=54455453-03, the new BUILD_ID), marked good, Updated logged, drawn at 6.61 s.
      • Fallback: from B, slot A's extlinux.conf got init=/sbin/tetsu-broken and the order A B with 3 tries and a pending note for A, as an install leaves them. A's kernel panicked and restarted three times, and B was back after 59 s with BOOT_A_LEFT=0; the screen said <build> DID NOT START, THIS SYSTEM WAS KEPT. A's extlinux.conf was put back afterwards, so the card holds the newer build in B, running, and the older one in A.
      • Not tested: a download over Wi-Fi rather than the cable (the same code), the lid during an install, B to stop one, a low battery, typing a source on the keyboard, a power cut while writing, a U-Boot environment save cut short, and the German wording on the device.
    • push.sh and slot B: push.sh boot copies boot.scr too, and on slot B points the copied extlinux.conf at partition 3. push.sh app and rootfs write the running slot, whichever it is. Flashing writes slot A and a fresh TETSU-BOOT, so the card starts again from A.
  • Build setup (2026-11-12): the Buildroot configuration is configs/common.config (what every device shares: system, packages, cores, ccache) followed by board/<board>/buildroot.config (CPU, toolchain, kernel, bootloaders, GPU and Wi-Fi firmware, the board's paths); make config concatenates them into output/tetsu_defconfig and runs defconfig with it, and BOARD= picks the board (only rg35xxsp). The cross toolchain is built once by make toolchain (also run by make when missing) in output-toolchain/ from configs/toolchain_a53_defconfig (GCC 15, binutils 2.45, glibc, headers 7.1), and boards use its host/ as a preinstalled external toolchain, so make clean keeps it. glibc has no RPC, so BR2_TOOLCHAIN_EXTERNAL_INET_RPC is off. ccache/ holds the compiler cache (BR2_CCACHE); Rust doesn't use it. Compared with the old single defconfig, the merged configuration differed only in toolchain and ccache options. On this PC (16 threads) the toolchain took 11 minutes and the clean image build after it 80, with an empty cache (1 GiB filled); output-toolchain/ takes 7.5 GiB, mostly its build folder.

  • Device profile (2026-11-12): device/profile.rs picks a Profile from /proc/device-tree/compatible (the device's own name, not the chip's): its name, and the battery, USB power and backlight folders that power.rs, charging.rs, settings.rs and sleep.rs used as constants. An unknown device logs a warning and gets the RG35XX SP's; a PC, with no device tree, too. The log has device: after launcher started, and SAVE LOGS TO CARD a device line. Features some devices lack (lid, vibration motor, second slot, Wi-Fi, an analog stick) join it with the first such device, as nothing would read them yet.

  • Save backup (2026-11-13): the saves kept off the card, so a dead card loses nothing. A backup is the same set both ways (backup::FOLDERS and FILES): saves (saves, states, resume points, play time, per-game settings), core-options, cheats, screenshots, and settings.txt, favorites.txt, collections.txt, hidden-games.txt, recent.txt and play-log.txt. Hidden files and the .tmp files of a write under way are left out, and so are games, covers, wpa_supplicant.conf and retroachievements.txt (passwords and a token).

    • Syncthing, Settings > DATA > SAVE BACKUP (backup/, save_backup=, off by default). The official static arm64 build of Syncthing 2.1.7 (package/syncthing, 32 MB in /usr/bin) runs from /usr/sbin/syncthing-service, with its keys, config and database in /mnt/data/.syncthing, so the device ID outlives updates. Its web page and REST API listen on 127.0.0.1:8384 only; the launcher reads the API key from config.xml. Each start sets it up (syncthing::Api::configure): one folder, tetsu-data, which is TETSU-DATA itself with .stignore taking in only the backup, send-only, ignorePerms and a 2 s time window for FAT32; no default folder, usage reports or upgrades; the device named by its hostname. The launcher starts it after the first frame while SAVE BACKUP is on, holds it with SIGSTOP during each game (it then takes no time from frames) and asks it to look for changes once the game ends, and stops it for USB STORAGE, which unmounts TETSU-DATA. S44syncthing stops it at power off, before S15data.
    • Pairing: the screen shows this device's ID at the bottom. Adding it in Syncthing on a PC, phone or server makes that device ask to pair, and it shows here as WANTS TO PAIR. A, then A again, pairs and shares the folder with it, which the other device then accepts (FOLDER NOT ACCEPTED until then; then SENDING n% and UP TO DATE). X dismisses a request; X twice unpairs. BACK UP NOW looks for changes at once. On the other device, file versioning on the folder keeps what the handheld deletes (a deleted game's saves), as send-only passes deletions on.
    • Restoring is A, then X, on a device asking to pair, as on a new card after the old one died: the new card has a new ID, which is added on the PC to the same folder. The card's own backup files are moved to .restore-set-aside and the folder made again, receive-only, with that device, so Syncthing knows nothing of the new card's files (whose newer times would otherwise win over the backup's). Once the device has accepted the folder and everything has come, local changes are reverted and the folder turns send-only; the launcher then starts again (it execs itself), as its settings and lists changed under it. A device with no files, 15 minutes without progress, or B twice stop it: the device is unpaired again and the files set aside are put back. A restore a power cut stopped goes on when the screen opens (the folder is still receive-only).
    • WI-FI TRANSFER's page has a Backup card: Download backup streams the set as one ZIP (tetsu-backup-<hostname>-<date>.zip, GET /api/backup, written into a pipe as it is sent, PNGs stored, the rest deflated, file times kept in UTC), and Restore from a backup takes such a ZIP (PUT /api/restore): received into a hidden file on TETSU-DATA, then each backup file in it replaces the one on the card, with its time; other entries, paths outside the set and files over 256 MB are skipped, and files the ZIP lacks stay. The screen says BACKUP RESTORED, and leaving it starts the launcher again.
    • Tested on the PC: cargo test covers the set, the ZIP both ways, refused entries, setting files aside and back, the ignore lines, the API key, reading devices and completion, the menu and the screen's lines; --preview-backup renders the screen and --preview-settings-grid the grid with DATA.
    • Checked by Claude on the device (2026-11-13), with a second Syncthing on the handheld itself (/tmp, API on 8386 through an SSH tunnel) standing in for the PC: turning it on set up the folder and .stignore and showed the ID; the stand-in's request showed as WANTS TO PAIR, A A paired, and once the folder was accepted it held exactly the set (69 files, 12.9 MB, no games) and the screen said UP TO DATE. Syncthing used about 80 MB of RAM at most for the whole system. pause stopped both processes (state T); a save written meanwhile went out within 20 s of resume. X X unpaired, the stand-in asked again, and A X restored: the card's files were set aside, a file only on the stand-in arrived, settings and saves came back byte for byte with their times, the launcher started again, and the folder was send-only and paired after. The page's ZIP (72 files, 3.3 MB, 1.6 s over USB) unpacked without errors; a non-ZIP was refused (415), a restore put 73 files in place and left out a roms entry, and B restarted the launcher. Turning it off stopped Syncthing within 4 s. Changing a folder blocks Syncthing's API until its rescan ends (over 10 s on the device), so calls wait up to a minute and the screen keeps its last state for up to 5 failed readings. Not checked: a real PC or phone, Wi-Fi rather than the USB link, a real new card, USB STORAGE with Syncthing running, a game's end sending its save, the lid during a restore, global discovery and relays (the image has no CA certificates, so the relay pool's HTTPS may fail; on a home network local discovery is enough), and the German wording on the device.

Hardware facts in full ​

  • TF2 is mmc2, with its own 3.3 V switch on PE4, patch 0018 (values from ROCKNIX's 0126 and 0159). With that supply off, the card is powered back through its pulled-up I/O pins, so it stays on, as the vendor firmware leaves it. The MMC hosts probe asynchronously, so without aliases the card numbers could change; root is found by PARTUUID either way.

  • The vibration motor is on PH2, PWM channel 2, patch 0015. It spins from 20 % duty. Its input device comes before the others, so /dev/input/event* numbers aren't fixed.

  • Don't trust the model name for hardware details. This unit has LPDDR3 at 1.2V, while mainline U-Boot's config and the kernel device tree assume LPDDR4 at 1.1V, and another documented SP has LPDDR4. Read the working boot0 instead. It starts 8 KiB into the card. At these offsets: 0x38 clock, 0x3c type (7 LPDDR3, 8 LPDDR4), 0x40 DX_ODT, 0x44 DX_DRI, 0x48 CA_DRI, 0x4c ODT_EN, 0xa0 TPR6, 0xa4 TPR10, 0xa8 TPR11, 0xac TPR12.

  • The SPL detects the RAM type (U-Boot patch 0001). It tries LPDDR4 at 1.1V first, counting it only if it trains on the full 32-bit bus and passes the size check, then LPDDR3 at 1.2V, so LPDDR4 chips never get 1.2V. The settings for both are in board/rg35xxsp/uboot.fragment. Patch 0006 gives reg_dcdc3 a range of 1.1V to 1.2V, and Linux leaves a voltage inside its range alone, so it keeps the SPL's. U-Boot proper needs the same range (U-Boot patch 0002): its regulator uclass sets an always-on regulator whose minimum equals its maximum, and set 1.1V again after the SPL. Tested on the LPDDR3 test unit (2026-11-18): the controller came up as LPDDR3 on the full bus with one rank (MSTR 0xC1040008), 1 GiB, vdd-dram at 1.2V through 3 reboots and a lid sleep. The time before the kernel was 2.74 to 3.03 s over four boots, against 2.53 to 2.66 s recorded before, so the failed LPDDR4 try costs about 0.1 to 0.2 s. Not tested: any LPDDR4 unit, a cold start from power-off with this SPL.

  • The RAM type is reported for testers (device/ram.rs): the launcher reads the DRAM controller's MSTR register (0x047FB000 on the H616, H618 and H700, found by the chip's name in the device tree's compatible) through /dev/mem once, and its device type bits give the type. SYSTEM INFO's MEMORY row shows it, the launcher's log starts with ram: LPDDR3, vdd-dram 1.20 V, SAVE LOGS TO CARD's system.txt has it, and bootlog.txt has a == ram section from S99bootlog (type, bus width, ranks, MSTR, the voltage and MemTotal), written even if the launcher doesn't start. Checked on the LPDDR3 unit (2026-11-18): all four show LPDDR3 at 1.20 V; the German row wasn't seen on the device, but 1 GB LPDDR3 - 856 MB FREI is 25 of the 28 characters it has.

  • A RAM setup failure resets the chip. The first boot stage turns the LED on, then panics and resets, which shows as a 1 Hz green blink with nothing on the card.

  • EHCI0 and OHCI0 share PHY 0 with the MUSB device controller. If they run, they switch the port to host mode and the USB console never connects, so patch 0007 disables them. USB host mode needs that reverted, plus a dual-role setup like ROCKNIX's 0151 and 0152 patches. Those also manage a 5V output on the port, so review them before use while the port is connected to a PC.

  • The USB controller has 4 IN and 4 OUT endpoints, besides the control endpoint. Serial and Ethernet take 2 IN and 1 OUT each, so together 4 IN and 2 OUT, and any further USB function needs one of them dropped: USB STORAGE drops the serial port.

  • Mainline's MUSB gadget driver throws away USB packets that arrive early. Since commit 3fc137386c46, musb_ep_restart flushes an RX packet that came before the gadget queued its request. Mass storage lost commands that way, and the PC reset the device every 30 seconds. Patch 0014 reads the packet instead.

  • Save states keep mGBA's Super Game Boy border mode. A state saved with the border on loads into the borderless frame offset and cut off, so such states can't be used since the border was turned off. Red's were deleted on 2026-11-08.

  • The CPU governor raises the speed slowly from idle while a game runs, because a game waits for the panel between frames. After a pause, games ran at 30 fps with audio underruns for about a second, so the menu holds scaling_min_freq at 1008 MHz until two seconds after resuming.

  • Plugging in USB power starts the device (VBUS rising, set in the AXP717's efuse; powering off with the cable in stays off). REG 0x20 says what started it; the charging screen uses it. Test shutdown with the cable unplugged.

  • Kernel power-off: axp20x_power_off writes register 0x32, but the AXP717 turns off through 0x27. Shutdown works because it falls through to PSCI, and TF-A then writes the correct register.

  • X and Y pins: the face buttons are X top, Y left, A right, B bottom, as confirmed by the user. The shared device tree has X on PA3 and Y on PA2, but this unit has them the other way round, so patch 0008 swaps the pins. Another SP may differ.

  • Kernel modules are off, and nothing could load them anyway, since there is no mdev or eudev. Add new drivers to linux.fragment as =y. Don't turn CONFIG_MODULES off in a fragment over a config that has =m entries: each one becomes built in, which added 1,377 drivers in a trial.

  • Audio levels: amixer percentages are raw steps, not loudness. DAC at 50% and Line Out at 50% total about -58 dB, which sounded silent. If the DAC switch is off, the DAC never runs and playback fails with error -5.

  • Headphones are silent while the Speaker path is on, so the two can't play at once.

  • Opening the sound card after about 5 s idle takes 700 ms longer: msleep(700) in sun4i_codec_spk_event in the kernel's sun4i-codec.c, waiting for the amp. Measured with 1.0 s of silence through aplay: 1.80 s after 12 s idle, 1.08 s straight after another sound or after 2 s idle.

  • The amp stays on for 5 s after a sound (pmdown_time; PI5 measured high for 5.29 s), while BusyBox init pauses only 2 s between the shutdown scripts and power-off. A shutdown right after a menu sound or a game could therefore cut the power with the amp on, which made the speaker crackle on some shutdowns. S20audio stop turns off Line Out, then Speaker, after the launcher has exited, which drops PI5 at once (0.02 s measured), and the user no longer hears the crackle. The controls are back on after a reboot (checked).

  • The display is card1, because Panfrost now takes card0. Name it explicitly, as in kmscube -D /dev/dri/card1, and stop the launcher first with /etc/init.d/S90launcher stop.

  • kmscube quits on any waiting input, printing user interrupted!. Paste its test as one line, because lines pasted after it count as input. Over SSH, closed input also counts, so give it open input: sleep 30 | kmscube -D /dev/dri/card1 -c 600.

  • The GPU needs CONFIG_SUN50I_H6_PRCM_PPU. It has no default, and without it Panfrost fails silently, with only ignoring dependency in the log.

  • TETSU-DATA starts at a fixed 4 GiB, past the image, so writing the image can't reach it. Writing the image does replace the partition table, so flash.sh adds the entry back, as partition 4 since 2026-11-12 (3 before); the system finds it by its label. After an image writer or a plain dd, the device adds it at its next start (/usr/sbin/data-partition, see the card layout).

  • The SP has one of two panels, found at each start (kernel patch 0020). ROCKNIX ships a device tree for each and leaves the choice to the user. Here the NV3052C driver resets the panel when it probes and reads its display ID (0x04) over the 3-wire SPI bus. An NV3052C reports 30 52 01 after a reset and 40 00 03 once set up. Any other ID, still after a second read 130 ms later, gets the second panel's setup and timings, from ROCKNIX's anbernic,rg35xx-sp-v2-panel.panel; an ID that can't be read keeps the NV3052C's. Only boards with the new anbernic,rg35xx-sp-panel compatible do this. The log says what it found, panel_newvision_nv3052c.panel=nv3052c or =v2 on the command line overrides it, and the launcher reads the same parameter in sysfs: games are paced at the second panel's 59.52 Hz, and 50 HZ GAMES AT 50 HZ does nothing on it, as nobody has seen that panel take the longer blanking. Tested on the NV3052C unit (2026-11-18): found on 8 of 8 starts, the launcher starting 1.45 to 1.75 s after the kernel as before. With =v2 forced there, the kernel set the second panel's mode (640 664 672 720 / 480 512 516 560) and blanking and unblanking gave no errors. The user saw the picture normal and back after a lid close. Not tested: any real second panel, its picture or its ID.

  • Wi-Fi blocks sleep without keep-power-in-suspend. rtw88 fails its SDIO suspend with -EINVAL ("Failed to host PM flag MMC_PM_KEEP_POWER"), which aborts the whole suspend, so lid sleep stops working. Patch 0013 adds the property to mmc1.

  • Sleeping while connected to Wi-Fi hangs the device for good. The kernel logs PM: suspend entry and never finishes suspending: programs stay frozen, so SSH and the console don't answer, but ping still does, and only holding the power button helps. Not connected, it sleeps fine. sleep_until_lid_opens therefore sends DISCONNECT before sleeping and RECONNECT after. The hanging driver call wasn't found: programs are frozen by then, and kernel messages go nowhere (ttynull), so finding it would need netconsole over usb0.

  • This wpa_supplicant has no syslog support: -s prints the usage and exits with status 0, so a script can't tell it failed.

  • The linux-sunxi wiki returns 403 to automated fetches. Use U-Boot's doc/board/allwinner/sunxi.rst, the ROCKNIX tree and the source code instead.

Tetsu OS is released under the MIT license. Emulator cores keep their own licences.