Skip to content

v0.4.1 — the rest of the machine

Choose a tag to compare

@Brunux Brunux released this 23 Oct 09:35
· 252 commits to main since this release

v0.4.1 — the rest of the machine

A patch number, but the largest batch since the port first booted: this is the
release where the session caught up with the hardware around it. A second screen
on either Type-C port and the menu that arranges it; the camera that was a USB
device all along; the translation of the selected text; the screen reachable
read-only from another machine; the idle screen that finally ends on a mouse move
and stops holding the lock chain behind it; and the Developer menu, which turns
the account into one encrypted archive on a stick.

The kernel and the compositor are the ones v0.4.0 shipped — Linux
7.2.7 with the same configuration, swayfx 0.6 on scenefx 0.5 — with one
device-tree patch this release adds for the external display.

What changed since v0.4.0

  • A second screen, on both ports. The port shipped no external-display
    pipeline at all: dpi@14016000 was disabled and there was no ANX7625, so the
    kernel created one connector and card1-DP-1 did not exist.
    kernels/mediatek-mt81/patches/mt8192-asurada-external-display.patch adds the
    three nodes ChromeOS's asurada DTSI carries — the DPI output with its 16 pins
    at 6 mA, the bridge on i2c@11d00000 with its pp*_dpbrdg rails and GPIO
    8/9/7, and the DPI input endpoint — to the shared asurada dtsi. The bridge then
    has one DP transmitter and two downstream Type-C ports, and
    TCPC_SWITCH_0/1 (7-bit 0x2c on i2c-7) pick which one it reaches: 0x21
    is the left port, 0x12 the right, and the chip powers up on the left.
    Mainline registers a typec port and no mux, so nothing ever wrote them:
    measured, a monitor on the right reported HPD, answered EDID over AUX and got
    1920x1080@60 while the panel showed its no-signal pattern, and the same monitor
    on the left worked. /usr/local/sbin/lazyos-dp-mux, run by a udev rule when a
    port's DisplayPort alt mode appears, writes the registers with i2cset -f
    after the page sync ChromeOS's i2c_access_workaround() does. The timing is
    the feature: the connector's hotplug arrives before the alt-mode device (46 ms,
    measured) and the chip trains the link when it is enabled, so a write afterwards
    retrains nothing. A resume is the half udev cannot see — the bridge powers down
    with the pipeline and returns to its default crosspoint — so
    ~/.config/bin/lazyos-dp-resume runs from the idle chain's after-resume and
    does the crosspoint plus one 1280x720 modeset, which is what retrains the link
    (a disable/enable would move the workspace to the internal panel). Where it
    sits is Settings → Display arrangement (rofi-display): right, left, above,
    below, mirrored, resolution and scale, with --status as the single copy of the
    answer the Settings row reads. The panel is never moved, only the monitor, and
    every arrangement is also written to ~/.config/sway/config.d/40-display.conf,
    because a runtime position dies with the session. Only the monitor can leave
    a picture-less machine when the cable comes out, so the no-monitor branch brings
    the panel back and Alt+Ctrl+Shift+D runs rofi-display --panel-on blind for
    the case where every output is off. wdisplays is the manual row.
  • The camera needed no driver, only a window. The Hayato's camera is a USB
    UVC device on the internal hub (0408:30d5 Quanta USB2.0 HD UVC WebCam), so
    uvcvideo binds it and udev creates /dev/video0; the tablet's video2/video3
    are the SoC's encoder and decoder, which is why lazyos-camera selects the node
    by driver and never by number. The device tree carries no MIPI sensor, so
    the rear camera a detachable is expected to have is not a missing module — it
    has no sensor to drive. The window is guvcview (GTK3, so it inherits
    gtk-3.0/gtk.css, and it brings the camera's own exposure, gain and white
    balance controls); the wrapper keeps its photos out of $HOME (into
    ~/Pictures/Camera/~/Videos), forces -g gtk3 -r sdl because guvcview
    remembers a headless run and would reopen with no window and the device held,
    raises a card when the window dies at once, and focuses an existing instance
    instead of opening a second one. The camera portal stays closed (this image has
    no access backend, so the portal logs one line per session); native programs do
    not care: Firefox reads V4L2 directly (verified: a 640x480 stream at 30 fps
    with the portal missing), so a video call works while a Flatpak app sees no
    camera. Rows: the root menu's Camera, org.lazyos.camera.desktop and a
    camera-web icon in the LazyOS theme.
  • Translate is a key, not a right-click (Alt+Ctrl+T). The mouse selection —
    the primary selection, which a drag already writes — or the clipboard goes into
    one language (target=en unless ~/.config/lazyos/translate.conf says
    otherwise) and the answer is copied; a long answer opens in a read window. The
    engine was chosen by measurement, not taste: Google's translate_a/single
    answers this machine with a captcha page, while
    clients5.google.com/translate_a/t?client=dict-chrome-ex answers plain JSON,
    auto-detects the source and echoes it, so the card can say Spanish → English
    and an English selection can answer Already English. It takes GET only
    (POST 302s to /sorry), which is why the selection rides in the URL and is
    capped at 3000 characters; MyMemory is the fallback in 450-character pieces, and
    when neither answers the card says so instead of showing half a translation.
    Settings → Translate is the same list as translate config from a terminal.
    A busy card with a stop button, both endpoints, and nothing resident.
  • The screen can be looked at from another machine, read-only and only over
    ssh: wayvnc (312 kB + libneatvnc 218 kB + libaml 82 kB), started and stopped
    from rofi-vnc (Settings → Share this screen, and the root menu). It binds
    127.0.0.1 only — wayvnc 0.10 authenticates only through a certificate, a
    private key and a password file this image does not ship, and a server on the
    network address would be an unauthenticated desktop on every network the
    machine joins; the loopback address is reached through the ssh login the image
    already answers, which makes the tunnel the authentication — and it runs
    --disable-input, because remote input would arrive as a keyboard nobody
    can attribute. nohup so the server outlives the menu and $! is wayvnc's own
    pid, pid and log in $XDG_RUNTIME_DIR/lazyos-vnc, the pid trusted only after
    kill -0 and /proc/PID/comm says wayvnc, and an ss -ltnH "sport = :5900"
    read-back instead of believing the pid. The bar's custom/vnc eye appears while
    it runs and a right click stops it. No busy card on purpose: a card older than
    five minutes is dropped as stale, and a share can last an afternoon.
  • Agents (Alt+Ctrl+Shift+A): launch, install (npm install -g --prefix ~/.local, with --allow-scripts because npm 12 blocks the postinstall that
    picks OpenCode's platform binary), pick a default, and keep accounts. An account
    is a directory under ~/.local/share/lazyos/agents/<agent>/<name> used through
    the variable the CLI documents — CLAUDE_CONFIG_DIR for Claude Code,
    CODEX_HOME for Codex; Gemini and Copilot have no such variable and the menu
    says so instead of pretending. Every row carries a glyph and a display/key TSV
    maps the rendered row back to its action. The root menu's AI group now opens
    with Agent session (rofi-agents --default, in ~/Work), and with no CLI
    at all it raises a card instead of opening a terminal that closes at once.
    OpenCode is the one agent the image ships, at ~/.opencode/bin:
    opencode upgrade rewrites the binary in place only from there and answers
    installation method not found anywhere else (measured on 2.1.3), so
    tools/opencode/ pins the arm64 tarball of the npm platform package, verifies
    its SHA256, and rootfs-setup.sh extracts it into the user's home. The image is
    signed out — the credentials live in the data directory, which
    lib-dist.sh excludes from anything published.
  • The idle screen ends on the pointer, and the chain stops waiting for it.
    A pty hears the mouse only when the program inside asks for it, so the
    dismissal is \033[?1003h\033[?1006h (all-motion tracking, SGR encoding; foot
    implements both), printed with the raw-mode setup and turned off from an EXIT
    trap so a --run in a real shell does not leave escapes in the prompt. Without
    it the loop saw key bytes only — motion and clicks were the terminal's own
    business, which is exactly the report ("the screensaver does not end with mouse
    movement"). And swayidle -w waits for each timed command in its own event
    loop (main.c, cmd_exec → waitpid), so a 150 s command that stayed in the
    foreground held the 300 s lock, the 360 s DPMS, the resume lines and
    before-sleep for as long as the wordmark animated: a session left alone never
    locked and the panel never went dark. The launcher now starts its foot and
    returns (foot … </dev/null &) — measured on a headless sway, with -w a
    timeout 1 'sleep 100' kept a 2 s timeout from firing within 5 s, and without
    -w it fired at 2 s. swaylock is pointer-blind on purpose (seat.c answers
    motion, buttons and the wheel with // Who cares), so typing is the unlock and
    a mouse move at five minutes only turns the panel back on.
  • The Developer menu (Alt+Ctrl+D, the root menu's Developer row and
    org.lazyos.dev.desktop): export my profile, restore a profile, save
    a diagnostics bundle
    . The archive is tar --zstd encrypted with gpg
    symmetric AES256 — a tar because the sticks are exFAT, which drops modes and
    symlinks (a copied id_ed25519 without its 0600 is a key ssh refuses), and
    encrypted because it carries ~/.ssh, ~/.gnupg and the Firefox password
    store; the passphrase goes to gpg through file descriptor 3 with
    --pinentry-mode loopback, never on a command line, and the plaintext tar is
    removed on every exit path. What travels is the account: the keys, the shell,
    the configs of the tools this image ships, ~/.local/bin, the applications,
    the agents (the menu's account directories, OpenCode's config, and the
    CLIs' own homes ~/.claude, ~/.codex, ~/.gemini, ~/.copilot,
    ~/.config/github-copilot, minus conversation transcripts and caches), and the
    Firefox profile (~/.config/mozilla — the XDG path this Firefox uses — or
    ~/.mozilla with MOZ_LEGACY_HOME=1: bookmarks and history, logins.db +
    key4.db, cert9.db, cookies, add-ons, bookmark backups). On the card here
    that is 238 MB of profile in, 3.2 MB out: storage/ (202 MB of IndexedDB
    and localStorage), security_state, the session and the SQLite
    -wal/-shm checkpoints stay behind, and a copy of a live profile is the
    database at its last checkpoint. OpenCode's opencode.db stays out because it
    holds its sign-in and the session history, and the manual says to sign in again.
    Restore lists the archives it can see, backs the current profile up to
    ~/.cache/lazyos/backups first, refuses absolute and .. members before
    extracting anything, and puts Cancel first with a typed restore as the last
    gate. The diagnostics bundle is this boot's report, journal, dmesg, hardware and
    packages, unencrypted because none of it is a secret.
    The passphrase prompt is the one hardware bug this paid for: sway's exec
    inherits /dev/tty7, so [ -t 0 ] alone showed a busy card while read()
    waited on a VT nobody was looking at. on_terminal() now also requires the
    process to be the terminal's foreground group (tpgid == pgid) and falls back
    to a masked rofi prompt whenever WAYLAND_DISPLAY is there; a scripted run is
    refused before the tar is built.
  • The workspace pills say what is running in them. workstyle 0.8.2 is built
    here (tools/workstyle/, three patched lines), costs 1 MB of PSS and no CPU
    while idle, and renames each workspace to "<number>: <icons>" with
    deduplicate_icons. The glyphs are the session's UI_I_* alphabet; the bar
    draws the whole name ({value}), the bindings say workspace number N, and the
    persistent-workspaces that never worked is gone rather than duplicated.
  • The keyboard and the prompt. kitty ships configured as the session's second
    terminal — solarized, the same font chain with symbol_map to the Mono variant
    so eza --icons keeps its columns, hide_window_decorations (swayfx is the
    frame), confirm_os_window_close 0 (Alt+Q is a close request that cannot
    answer a dialog) and background_blur 0 (swayfx advertises neither blur
    protocol) — and the prompt is the powerline one in the session's palette, with
    the triangles U+E0B0/U+E0B2 because JetBrains Mono draws them at the text
    size while the rounded caps exist only in the symbol font, which foot draws at
    that font's size (both terminals were screenshotted before the choice), and no
    $os segment because twelve of the plane-15 MDI glyphs have no font on this
    image. The touchpad's four three-finger swipes: left/right move the session, and
    up/down type Ctrl+Tab/Ctrl+Shift+Tab into the focused Firefox
    through wtype, gated by [app_id="(?i)^firefox" con_id="__focused__"] —
    measured, criteria without con_id ran while a terminal was focused, and the
    prefix costs 4 ms inside the compositor while starting no process at all outside
    Firefox.
  • The small things the card asked for. The menus find /usr/sbin and /sbin
    (lib-rofi.sh appends them once; the session's PATH is greetd's, so Missing:
    parted mkswap swapon blkid
    came with all five installed), the disk guard names
    its fourth reason (a removable, unmounted card that carries a bootable kernel
    partition is the card that installed this system), the AI section opens with
    the action, the key view draws a picture on every row, the file-search check
    looks inside the drop-in directory, and the site's install section walks card to
    eMMC.
  • A new package, and no new service. gnupg joins packages/tools.list for
    the Developer menu and for the keyring a developer expects; gpg-agent is
    D-Bus-activated, idles out and is listed among the expected user units. Every
    other feature here is a one-shot script or a package this image already had.

Validation

The tree builds with scripts/build.sh on the card's own port: Linux 7.2.7 with
the display patch, scenefx and swayfx from source, workstyle from its crate,
debootstrap for forky, and the kpart signed with the vboot developer keys —
futility show answers Body verification succeeded. The image, its xz, the
root filesystem tarball and the kpart are sanitized and asserted clean by
scripts/lib-dist.sh, and SHA256SUMS covers the four.

scripts/check-system.sh is also the release's own to-do list. Run against the
v0.4.0 card with this tree it reports 244 OK, 7 WARN, 16 FAIL, and the
sixteen are exactly what this release adds and the card does not have yet:
iproute2/iputils-ping/zip/strace/ncdu, the idle screen's ?1003h and
its early return, the idle-inhibit drop-in, the volume overlay, the Search-key
bindings, the translate key and its manual line, the VNC menu and its bar module,
the welcome window's key, the agents menu's ~/.opencode/bin line and
--default, the AI session row, wayvnc and fonts-noto-core. A card written
from this image is where those sixteen become OK; the other 244 lines are the
ones the v0.4.0 card already passes.

The behaviours with numbers behind them were measured on this machine rather than
assumed: the DP crosspoint on both ports and across a resume (0x12 → 0x00 →
0x21), the camera's 640x480 stream in Firefox with no portal, the translation
endpoints answering JSON, the idle chain's 2 s versus 5 s, the swipe's 4 ms, and
the Developer archive's 238 MB → 3.2 MB.

Cutting the release also found one packaging bug of its own: scripts/lib-dist.sh
checked that the image carries no OpenCode state by running find on
~/.local/share/opencode — after dist_sanitize had just removed it. find
exits 1 for a missing path, its message went to /dev/null, and set -e ended
make-release.sh right after "Sanitizing" with no output at all. The check now
tolerates the absence it is looking for, which is the state it is supposed to
find.

Assets

asset what it is
google-asurada.img.xz 6 GiB disk image: the three ChromeOS GPT partitions (kernel, /boot, root), ready for dd
google-asurada-rootfs.tar.zst the same root filesystem as a tarball (with /boot inside), for writing your own card with scripts/write-sd.sh
google-asurada.vmlinuz.kpart the signed FIT kernel partition of that image; it embeds the image's root=UUID, so it belongs to this image and no other
SHA256SUMS SHA-256 of the three assets
sha256sum -c SHA256SUMS
xz -dc google-asurada.img.xz | sudo dd of=/dev/mmcblk1 bs=4M conv=fsync

On the Chromebook: developer mode, crossystem dev_boot_usb=1 dev_boot_signed_only=0, insert the
card and pick external disk in the firmware menu. It comes up to a greetd prompt — user
brunux, password debian, passwordless sudo. Change the password before using it for
anything real.
scripts/install-emmc.sh installs it to internal storage, and on the card
Alt+Space → Install LazyOS is the same call with the questions asked first.

What it does not carry

The artifact is sanitized and asserted clean by scripts/lib-dist.sh:
no Wi-Fi profiles (a released image knows no network and asks for one), no SSH host keys and no
machine-id (both generated on first boot, so every installation has its own identity), no
OpenCode sign-in and no personal files beyond the session configuration this repository ships.

Known limitations

  • The night light cannot reach the external monitor. MT8192's second display path
    (OVL1→RDMA1→DPI) has no gamma block bound, so modetest -M mediatek -p lists
    GAMMA_LUT on the 1366x912 CRTC and nothing on the 1920x1080 one; wlsunset logs
    gamma control of output … failed and the monitor keeps its own colour
    temperature.
  • The camera portal is not implemented (there is no access backend), so a
    Flatpak application sees no camera while Firefox and other native programs work.
    This is a limitation, not a bug to wait on.
  • The three-finger tab swipes do nothing when Firefox runs on Xwayland
    (MOZ_ENABLE_WAYLAND=0): the criteria need an app_id, and an X11 window has
    none.
  • The profile archive does not carry OpenCode's sign-in (it lives in
    ~/.local/share/opencode/opencode.db with the session history), and a Firefox
    profile copied while the browser is open is each database at its last
    checkpoint.
  • Wayvnc is loopback-only and read-only: the way in is an ssh -L tunnel
    (rofi-vnc has the command on a row), and the menu's input is deliberately
    off.
  • Install LazyOS refuses to run from the disk it would overwrite, which is the
    correct answer when the menu is opened on the eMMC install; the installation is
    made from the microSD, which can always be selected again in the firmware menu.
    The copy cannot be stopped: killing rsync or mkfs leaves a disk that looks
    written and is not.
  • The image is 6 GiB (IMAGE_SIZE_MIB=6144), which leaves about 2 GiB in the
    root partition; a larger card keeps the space after p3 unused unless the
    tarball is written with scripts/write-sd.sh, which sizes p3 to the whole
    card. The tarball's /etc/fstab is the unconfigured one the build leaves
    behind.
  • Artifacts are not byte-reproducible yet (FIT timestamp, filesystem UUIDs).
  • The battery gauge's chemistry and time-to-empty properties are upstream failures;
    waybar reads the capacity directly, which works.
  • There is no hardware video decoder, and Vulkan is not usable (PanVK answers
    Mali-G57 not supported); the session renders through EGL and Panfrost, and
    video is CPU work.

Upstream

Vendored from pmaports commit 0e458a2417d6f5ed61c7b70621d5632e30136091 (2026-10-15), mainline
Linux 7.2.7, swayfx 0.6 on scenefx 0.5, workstyle 0.8.2, OpenCode 2.1.3.