- The Frame says it has caught up as soon as an aimed event lands after a
stale one, so the page stops re-reading the panels.
- After a device pause it lets go of everything it holds (releases that
arrived while paused were dropped), and waits for the device once per
batch, not once per event.
- The quick focus check no longer freshens the panel geometry's age.
- Each request carries at most about 100 characters of text.
- Turning Control off while it connects doesn't report an error.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- The Frame side tracks keys as well as buttons and lets go of both when the
session ends.
- A stale tap tells the page, which re-reads the panels at once.
- A paused input device waits instead of ending the session; only a
disconnect does. An OS error on one event skips it.
- Presses check focus with two property reads and do the full lookup only
when it changed.
- Writes to an agent's stdin are serialized, so two devices sending at once
can't tear a line (the keyboard agent too).
- Connecting gives up with a message after 15 s instead of hanging on
"Connecting…"; text goes in 100-character pieces so releases don't wait
behind a long paste; a cancelled mouse gesture releases what's held.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- The focused panel's display comes from GAMESCOPE_FOCUS_DISPLAY (gamescope
packs ":1" into the first value), so a window id repeated across :0 and :1
can't be mistaken; the pid is only the fallback.
- Presses, keys, text and scrolls read focus afresh on the Frame; only moves
use a reading up to a second old.
- A gesture remembers the panel it started on and does nothing more if that
stops being the one in use; a press with no panel to aim at isn't sent.
- Opening a screenshot clears the panel Control would act on; switching to
another app releases held keys and buttons.
- Trimming keeps a click with its position; the error backoff holds for new
input too.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- Keys held on the Frame are released with buttons when Control stops or
the view loses focus; keys for the Frame no longer trigger Frame Control's
own shortcuts.
- Taps only act when the picture on screen is the panel in use; positions,
presses, keys, text and scrolls name their panel (display and window: ids
repeat across :0 and :1, told apart by pid), and the Frame drops them if
focus has moved on. Releases always go.
- While connecting, a tap keeps its position; on an error only releases wait
and retries back off; trimming a long queue never drops a release.
- Lifting one of two scrolling fingers ends the scroll; a cancelled touch
isn't a tap; clicks and holds on the bars around the picture do nothing.
- A capture loop from before a Live restart can't stop the new video.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The live view gets a second source and a way to use the Frame from it:
- Desktop: the app panel in use in the headset, streamed from its own window
(x11grab of gamescope's redirected window), so it doesn't move as the
wearer looks around. A picker shows any other panel, view only.
- Control: on the Desktop view a tap or click lands exactly where you put it;
drag is a mouse drag, press and hold right-clicks, two fingers scroll, and
on a computer the mouse, wheel and keyboard work directly. On the headset
view the view is a trackpad. A text field and key row type from a phone.
Input goes through gamescope's own EIS socket (the way Steam feeds Remote
Play input) with the libei already on the image: ui/frame_touch.py, over
the same long-lived ssh machinery as the keyboard agent, nothing to
install. It reaches the panel that has focus on either X display, which
the KDE Connect route can't. Verified on the Frame and from the iPhone app
in the Simulator.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Any unexpected error while launching clears the launch and reports it, so
the pad can always be started again; the temporary stderr file is made
inside the handled path and a failure reading it is tolerated.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
discard() swallows OSError as well as Failure, so a launch that fails and
can't remove its copy still reports the error and can be started again.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The agent holds its copy from its first status line on and tidies it up
however it ends. Before that (a launch error, or stopped before the agent
ran) the server removes the copy itself. discard() only ever removes
incoming copies, never the iPhone bundle's own. The retry race test waits
for both contenders' decisions instead of sleeping.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- A start turned off while copying removes the copy instead of starting an
agent that would be killed before it could tidy up.
- A local read error while copying removes the partial copy too.
- The retry race test holds the new start until the retry has decided, so
it fails every time without the fix (checked 3/3), not by luck.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- Each start keeps its copy (holding a lock on it) until it ends, however
it ends, then removes it; a restart can still unpack it.
- The server removes a partial copy when copying fails.
- Other copies are removed only when unlocked and over an hour old.
- Tests: a start racing the need-packages retry (one agent, not two), a
restart that must unpack its copy, a held copy surviving the sweep.
Both regression tests fail without their fix.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- The need-packages retry only runs if no start() has taken over, so two
agents can't end up running for one session.
- Each copy goes to its own incoming folder; the agent removes its own once
connected (and any a cancelled start left over an hour ago), so a stopped
start can't delete files another is copying or still needs.
- A missing package folder means need-packages, not an error.
- Tests keep fake agents running, so they check ready and which agent owns
the session, plus a bounded retry and the tidy rule.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- iPhone build fails if the bundle can't be made, instead of shipping a
stale archive with an empty version.
- A different build that another device is using right now is left running
and replaced once nobody is, rather than stopped under them.
- If what's installed changed after the server looked, the agent asks for
the packages (need-packages) and the server copies them and starts again.
- The installed-build check sends bytes, so Windows' CRLF can't break it.
- Starting again while a stopped start is still copying launches anew;
concurrent copies use their own temporary names.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The keyboard and trackpad no longer fetch KDE Connect from Valve's package
repository on the Frame. The desktop apps and the iPhone app's Frame bundle
carry Valve's arm64 build of kdeconnect 24.02.2-1 and the five libraries it
links (kcontacts, kpeople, modemmanager-qt, pulseaudio-qt, libfakekey),
pinned by SHA-256 in frame/kdeconnect/packages.json and downloaded at build
time from the kdeconnect-frame-24.02.2-1 release, which also holds Valve's
complete source package for each.
On first use the computer copies them over its SSH connection (the iPhone
bundle already has them on the Frame); the agent checks each SHA-256,
unpacks them and stamps which build it is, so later starts copy nothing.
No internet on the Frame, 3.6 MB instead of 8 MB, 18 MB unpacked instead of
82 MB (ModemManager and friends were packaging-only dependencies).
GPL/LGPL compliance: frame/kdeconnect/NOTICE.md names each exact version,
licence and source; per-project licence texts in frame/kdeconnect/LICENSES;
THIRD_PARTY_NOTICES.md; an About and licences dialog in the app.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Home → Keyboard and trackpad, in every version of Frame Control (Mac,
Windows, Linux, iPhone, iPad) with nothing to install on the device in
your hand. On a phone: a trackpad (drag, tap, two-finger scroll and
right-click) and a field that types on the Frame. On a computer: click
the pad to pass the mouse and keyboard through; Esc to stop.
First-party route: ui/frame_input_agent.py runs on the Frame and talks
KDE Connect's LAN protocol (v7) to kdeconnectd as if it were a phone.
KDE Connect isn't on the image, but Valve's package repository has it;
the agent fetches it and four libraries into ~ (no root, survives
updates), starts it, pairs by itself (accepting over D-Bus), and stops
it again when the last device disconnects. Each device has its own
identity; a stuck KDE Connect is restarted once.
Verified against the real Frame (SteamOS 0.4.1): first-time install,
pairing, pointer moves from the Mac's server and the iPhone app
(Simulator), Mac and iPhone at once, two installs at once, a frozen
daemon replaced, and the daemon stopping when the app quits.
Reviewed by GPT-6 Astra (xhigh) over seven rounds; all findings fixed
except per-event delivery acknowledgement (documented known limit).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Main now runs Android installs as background jobs and maps SSH failures to
one offline message. Alternative-version installs go through the same job,
the alternatives dialog waits on it with runJob, and send_error_json keeps
the apk blocker that opens the dialog.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
One malformed or unreachable repo no longer hides the others; skip bad
index entries; a refreshed raw index outdates its reduced copy; style the
dialog like the others; validate package ids with PKG_RE.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Reducing an index no longer deletes apk-catalog/data/index-v2.json, which
the catalogue build reads. Drop the measurement log from docs.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The menci/archlinuxarm base failed `pacman -Syu` on GitHub's arm64 runner.
Valve's Holo Core aarch64 preview is the base the Frame's SteamOS is built on,
the iPhone app's frame-container already uses it, and its repos carry every
package the fake needs. A failed image build now reruns with the full log.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The tabs UI (#4) runs Flatpak installs as server.start_job jobs, so a failed
install answers 200 with a job id and reports the error on /api/job. The test
now waits for the job instead of expecting an immediate 502.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
On the Frame, compat_log.txt has binary bytes in it, so plain grep only said
"binary file matches" and the x86-64 launch check missed Steam's "is not
installed" line. The fake now writes that line to a compat_log.txt that starts
with a NUL, and the end-to-end test finds it the way the smoke test does.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A launch that needs the program running kept looking after Steam's
"started" line, rather than failing on it before the process showed up.
Cleanup reads steamos-delete's log, which reports a failed sync but exits
0, and runs it twice to check Steam no longer lists our titles; until then
they stay tracked and the cleanup step fails.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Headset smoke test: drop the paired key from authorized_keys with a
same-mode copy swapped in, so a failed write can't truncate it; check the
throwaway key with no ssh_config or agent; clean up idempotently (tracked
and leftover titles, their json files, Steam's shortcuts via steamos-delete,
and ~/devkit-utils if it wasn't there before), with a failed cleanup a
failed step; count a launch only with fresh evidence (the process, Steam's
log, or the known missing-runtime line), matching with [d]evkit-game so
pgrep doesn't find its own shell. The test programs sleep 10 s.
Fake Frame: log a launch before its reaper can look for it. e2e: kill a
pairing client's process group when a test ends; accept an aarch64 program
running under QEMU on x86 hosts.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
tests/fakeframe: a container that stands in for the Frame (Arch Linux, or
Arch Linux ARM on arm64) with sshd, rsync, Valve's steamos-devkit-service
and hooks (vendored unmodified), a fake Steam client for the devkit pipe and
the DevTools port (the app's JavaScript runs in Node against stand-in
SteamClient/appStore objects), stubs for steam, wpctl, flatpak, podman,
Lepton and friends, battery and thermal files under /sys, and fault
switches (fakeframe-ctl): pairing mode, approve/deny/timeout, Steam not
running, headset asleep, sshd off, disk full, runtimes missing. A second
container is the computer running Frame Control.
tests/e2e: 29 tests driving the real ui/server.py, frame_connect.py and
frame_titles.py against it; skipped unless FRAME_E2E=1. scripts/e2e.sh
builds, runs and tears down; CI runs it on ubuntu-24.04-arm.
tests/smoke + scripts/frame-smoke.sh: the core cases against a real Frame,
recorded with its BUILD_ID, cleaning up after itself; --pair to pair a
throwaway key. docs/testing.md describes the layers.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
On the Frame, Steam refused to register titles whose id had a hyphen
(fc-smoke-exe) with "missing/invalid arguments", and registered the same
program as FCSmokeProbe (headset smoke test, 2026-09-27, BUILD_ID
20260922.6101926). Valve's client only allows ^[A-Za-z_][A-Za-z0-9_.]+$.
title_id now makes ids of letters, digits and _, not starting with a
digit, 2 to 64 long; new installs are checked against that, while titles
already on the Frame are still listed, launched and removed. Steam's error
text is trimmed before it's quoted, and the "install it again with Steam
running" hint only follows a Steam-not-running error.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
tests/frame-container/frame-image.sh extracts rootfs-A from Valve's Frame
recovery image (steamdeck-images.steamos.cloud/recovery), mounts it read-only
with a throwaway writable layer and starts the image's own sshd, so the iPhone
app can pair with and run its server on the real SteamOS for Frame userland.
Verified: password pairing then key login in the image's sshd log, the power
password check through the image's sudo, status reading SteamOS 0.3.0.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Valve publishes no Steam Frame OS image, so tests/frame-container builds the
Frame's SSH surface on Valve and Collabora's Holo Core aarch64 base: a
steamos user with a password and sudo, OpenSSH with keys and passwords,
Python, and a systemctl that only records requests.
Against it from the Simulator: password pairing, the host-key pin, the power
password check. Found and fixed: a changed host key or a refused login said
"Can't reach the Frame" and retried forever; they now say "Pair with the
Frame again" and offer that. The server's key rejection now says the header
may be wrong, not only missing.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
An iPhone can't run Python or ssh, but the Frame can. The app (ios/, SwiftUI)
connects with its own SSH key (Citadel), copies the server and helpers to
~/.cache/frame-control/<version> on the Frame once per version, starts
ui/server.py there with FRAME_LOCAL=1 on the Frame's 127.0.0.1, and shows the
page through an SSH tunnel. The server exits when the phone disconnects.
Server: FRAME_LOCAL=1 puts ui/local-bin on PATH, whose ssh stand-in runs each
`ssh frame COMMAND` locally (and serves as rsync's transport), so desktop and
phone share one code path. Android display goes through podman exec there, as
the Frame has no adb. FRAME_UI_KEY replaces the fixed X-Frame-UI value with a
per-session key. Power actions take the Developer Mode password via sudo -S.
--port 0 now prints the port it took.
Page: a bottom tab bar and safe areas on phones, Play buttons visible on touch
screens, saving through the share sheet, SSH/SFTP/Steam Link/remote desktop
opening in their iOS apps, and a password dialog for power.
App: pairing with the Developer Mode password once (never stored) or with a
key the user adds; host key pinned on first use; plain-language connection
errors with quiet retries; frame-control://install links; alerts and confirms.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The page was one 6,800px scroll with nine nav links (hidden below 1150px).
It is now four tabs, Home, Games, Android and Tools, switched with 1-4; old
section links still land on the right tab.
When the Frame can't be reached, the server turns ssh's connection errors into
one plain message (503, offline: true), the page shows a single banner with
Retry and Set Up Connection, retries every 8 s, and reloads every panel when
the Frame answers. Panels say "Waiting for the Frame" instead of raw ssh text.
Flatpak and Android catalogue installs run as background jobs the page polls,
so a slow install no longer holds a request for up to 15 minutes or reports a
false failure; the bottom bar counts running installs.
Files can be dropped anywhere in the window, as the README already said.
Recent reports show the newest five, with Show all. Android display explains
an empty or failed read. A topped-up headset on a charger reads as not
charging rather than "still draining, using 0.0 W".
Fixes a race where the catalogue and reports loads wrote the compat-db
mirror's .tmp file at once and one failed with a 500.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Waking a recv blocked in another thread needs the handle closed on
Windows, which isn't safe under a TLS read (the previous commit, reverted
after review). A stalled download there holds the quit for the 4 s grace
period; its partial file is swept on the next start.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- Only stored and deflated entries are read: Python 3.9's bzip2 and lzma
readers inflate without bound before trimming.
- loadTitles drops a response that a newer request has overtaken, so the
install dialog's replace warning uses the fresh list.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- APK reader: read members with a bounded read, since ZipFile.read inflates
a member fully before trimming it to a forged declared size; the reference
walk counts every entry it examines, dead ends and cycles included.
- Titles: a path that exists under the root wins over stripping the archive
prefix; the prefix is taken before a linked folder is staged elsewhere
(another drive on Windows); the install dialog loads a fresh title list
first and says so if it couldn't check.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- APK reader: cap AndroidManifest.xml, resources.arsc and icon sizes before
inflating them (APKs can come from install links), and follow resource
references without cycles and with a result budget.
- Sideloading: reserve devkit-steam (SteamOS's sideloaded-client trampoline);
the install dialog warns when a name replaces an installed title; a
manifest's exe may name the program as it is in the archive, above the
folder the installer steps into.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Tested on a Frame (BUILD_ID 20260922.6101926):
- Devkit pairing: the service runs and answers properties.json, but /register
refuses with "please put the Steam client in pairing mode" unless Steam is on
Settings > Developer > Pair new host. Both setup paths now say so and keep
asking for 2 minutes before falling back to the password.
- Sideloading: registering, launching, quoted start paths and Remove work; a
Windows exe runs under Proton 11 through FEX. An aarch64 build starts but
outside the runtime container, and an x86-64 Linux build doesn't start
because the x86-64 Steam Linux Runtime 4.0 isn't installed. The inspect note
for x86-64 Linux builds now says so.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- Links to localhost need FRAME_CONTROL_LOCAL_LINKS=1: otherwise any website's
link could make the app fetch from services on this computer.
- Title staging folders (unzipped titles) carry the server's PID and are swept
on the next start like download folders, so quitting mid-install doesn't leave
gigabytes behind.
- An install from a link refreshes Sideloaded titles.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- frame_apk: android: attributes win over same-named attributes in
other namespaces; string attributes that keep only a typed value (no
raw string) still resolve; a failed icon read leaves the icon out
instead of failing the install.
- The clipboard IPC origin check can't throw on odd frame URLs.
- fetch-deps.js: 60 s download timeout, at most 5 redirects, a SystemRoot
fallback for tar.exe, and prunes pydoc_data, venv and the static
libpython.
- Docs keep the clipboard-tool note for running the UI in a browser.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Combines the three feature branches on bundle-deps. Conflicts in server.py,
index.html, preload.js, README and test_server.py keep both sides. The
web-install downloader now uses urllib's default HTTPS context, so the
bundled CA list from 770f26c applies to it on Windows too.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Dropping a game's .zip, folder or .exe on Send to Frame now adds it to the
headset's Steam library through Valve's SteamOS Devkit title path, with the
runtime picked from the program's header: Windows PE -> Proton Experimental
(steam_play=1), aarch64 ELF -> SteamLinuxRuntime_4-arm64, x86-64 ELF ->
SteamLinuxRuntime_4 (through FEX). Other architectures are refused.
- frame/devkit-utils: Valve's devkit-utils vendored unmodified (MIT,
steamos-devkit v0.20260925.1), synced to ~/devkit-utils by stamp, bundled in
the app and compiled in CI.
- ui/frame_titles.py: inspect (safe unzip, ELF/PE classification, launch
target ranking), install(path, name=None, exe=None, runtime=None,
progress=None), list, launch, remove, plus a CLI.
- ui/server.py: /api/titles (inspect/install/discard/launch/remove),
/api/titles/job progress, and an upload mode 'title'.
- ui/index.html: confirm dialog (name, launch target, runtime), install
progress, and a Sideloaded titles list with Launch and Remove. The app's
preload passes a dropped folder's path.
- tests and docs/sideloading.md. Device-side behaviour is inferred from
Valve's source; the headset was offline, so none of it has been checked on
a Frame yet.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A site can link to frame-control://install?manifest=URL (or ?url=URL) to
install a title with Frame Control. Manifests use FrameDrop's format, so
framedrop.install/v1 is accepted as well as frame-control.install/v1.
- app/install-link.js parses links; main.js registers the scheme (plus
electron-builder protocols for Info.plist and the .desktop file), takes
links from open-url, second-instance argv and the first argv, and holds
them until the page asks for them through preload's onInstallLink.
- ui/frame_webinstall.py checks the URLs (HTTPS only; localhost over http
only when the link itself is local; no userinfo; every address public,
rechecked on redirects and pinned for the connection), reads the
manifest, downloads with a size cap and sha256 check, and dispatch()
sends .apk to frame_android and .zip/.exe to frame_titles when present.
- server.py adds /api/webinstall/check, start, job and cancel behind the
existing Host and X-Frame-UI guards; a start needs a one-time id from
check. Downloads stop on cancel and on shutdown, and leftovers from a
killed server are swept by PID.
- index.html asks before anything downloads (name, source host, file,
type, size, whether a sha256 was given) and shows progress.
- docs/web-install.md, docs/install.html (landing page, unpublished).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
connect.sh and frame_connect.py now try Valve's steamos-devkit-service
first: GET /properties.json for the login user, then POST /register with
a new RSA key (~/.ssh/id_rsa_frame_devkit, the only type it accepts), so
the user approves a prompt in the headset instead of typing a password.
Port 32000 closed, a timeout or a 403 falls back to the existing
password copy. The Host frame block lists both keys; a host counts as
found if port 22 or 32000 answers; with no host given, dns-sd or
avahi-browse look for _steamos-devkit._tcp. Inferred from Valve's
source, not yet verified on a Frame.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- app/build/fetch-deps.js downloads a standalone Python 3.12
(python-build-standalone) for every build and adb from Google's
platform-tools, pinned by SHA-256, and prunes what the server never
uses. It replaces the Windows-only embeddable Python.
- The app runs the bundled Python with -I -u -B -X utf8, so a PYTHONHOME
or PYTHONPATH meant for another Python can't break it and nothing is
written inside the signed macOS bundle.
- An adb you already have still goes first, so two adb versions don't
keep restarting each other's server. arm64 Linux has no official
platform-tools and keeps using the system adb.
- ui/frame_apk.py reads APK badging (package, label, version, min SDK,
ABIs, icon) from the binary manifest and resources.arsc, replacing
aapt2. It matches aapt2 on nine F-Droid APKs and finds launcher icons
the old path missed with adaptive icons.
- The app reads the computer's clipboard through Electron, so Linux no
longer needs wl-clipboard or xclip.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- Without the maintainer's key, Android compatibility reports stay on the
Mac and the UI says so; the shared database is never contacted.
- Remove personal infrastructure details from scripts and docs: the Drive
folder and gog wrapper now come from the environment, and the Chromium
build host is required instead of defaulted.
- Add an MIT license, tester instructions in the README, and bump to 0.2.0.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Live in the headset view now streams video instead of polling stereo
screenshots (~2 fps). SteamVR's steamvr-v4l2cam.service mirrors the headset
view into /dev/video99; ffmpeg on the Frame encodes it with x264 (720p30 by
default, AUD + repeated SPS/PPS), /api/stream relays the raw H.264 over SSH,
and the page splits it on access unit delimiters and decodes it with
WebCodecs into the existing viewer. Capture still takes a stereo still; the
desktop panel keeps capture polling, and the page falls back to it if the
video can't start.
The remote ffmpeg runs under a shell that kills it when the SSH channel
closes, stderr goes to a temp file, and a 10 s stall ends the stream. One
stream at a time; a new one supersedes the last.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
