#16: with display, capture and encode all on an RTX 5070 the packed-RGB
input path is 19% slower in the probe and 25% end to end than CPU
swscale, while on an iGPU-composited desktop it was neutral to faster.
Keep it on by default only when no connected DRM connector belongs to an
nvidia-bound card; UU_NVENC_RGB=1 forces it, =0 still disables it.
Co-authored-by: VallierDev <[email protected]>
Co-authored-by: Claude Opus 5.5 <[email protected]>
* Restart X11 capture at the new size when the desktop is resized mid-stream
A RandR resize during a session made x11grab fail with BadMatch on the old
size. read_encoder treated that as an encoder failure and walked the whole
fallback chain (NVENC -> CPU, HEVC -> H.264) with the stale size, then the
stream stopped ("Encoder output ended") and the viewer froze. New connections
already re-probed the geometry (#18); live ones did not.
- When an X11 stream fails and the root size differs, skip the encoder
fallback and restart capture at the new size with the same settings.
- Absolute pointer mapping follows the refreshed X11 size.
- New connections compare the root size with their capture instead of the
engine's change flag, which a live stream may already have consumed.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
* Change the X11 desktop resolution when the viewer picks one
The official Windows host changes its display resolution for the viewer's
分辨率 choice; this host only scaled the encoded stream, so a smaller choice
made text smaller and blurrier instead of larger.
On X11 the capture/resolution RPC now switches the primary output's mode with
xrandr (an existing mode of that size, else a CVT reduced-blanking mode at
60 Hz, or 30 Hz above an HDMI 2.0 pixel clock), re-applying it when mutter
resets the mode after the list changes. resolution_type 3 (original) and the
end of the last desktop stream restore the mode in use before the first
change. The stream restarts capture at the new native size and reports it in
the screen sources and CaptureSettingSync. UU_DESKTOP_RESIZE=0 keeps the old
encoder scaling; Wayland is unchanged.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
---------
Co-authored-by: VallierDev <[email protected]>
Co-authored-by: Claude Opus 5.5 <[email protected]>
With --features inproc-av on X11 and UU_VIDEO_ENCODER=vaapi, VAAPI runs in process: a DRI3 DMA-BUF of the root
(XShm upload without DRI3) goes through VPP to NV12 and the hardware encoder. SPS/PPS(/VPS) and slice headers are
packed by the host because radeonsi (Mesa 26) does not write them. libva/libva-drm are dlopen'ed; release builds do
not enable inproc-av. The NVENC path is unchanged.
Accepted by the contributor on AMD HawkPoint1 (Mesa 26.2.3), 1920x1200@60, with the official Mac viewer
(HEVC, presets, fps, scaling). Intel iHD and 4K are not verified. Thanks @ferstar.
H.264 High 4:4:4 / HEVC FRExt with chromaFormatIDC 3 and AQ off when 真彩 is requested; GPUs without YUV444 encode fail the capability check and fall back. full_chroma_available() no longer excludes the in-process path. UU_PROBE_CHROMA=444 for the probe.
Co-authored-by: ferstar <[email protected]>
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Replaces the FFmpeg-based in-process encoder with a native NVENC engine (vendored nv-codec-headers n12.1.14.0, libcuda/libnvidia-encode dlopen'ed at runtime, primary context reuse). Fallback: native NVENC -> FFmpeg NVENC process -> libx264. inproc-av no longer links FFmpeg; VAAPI/CPU always use the process pipeline. Default builds are unchanged.
Co-authored-by: ferstar <[email protected]>
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The official viewer (libstreamer, Mac 4.43) treats a video_codec_capability entry
with width or height 0 as unsupported. The 4.37 Windows host we copied in #49 ran on
a GTX 660 without HEVC/4:4:4, so its 0x0 entries meant exactly that. Evidence by
@ferstar in #77.
Co-authored-by: VallierDev <[email protected]>
Co-authored-by: ferstar <[email protected]>
Co-authored-by: Claude Opus 5.5 <[email protected]>
uu_wl_move returns -2 only when the compositor has no wlr-virtual-pointer; other failures are returned as before. Document UU_WAYLAND_NO_VIRTUAL_KEYBOARD.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Compositors without wlr-virtual-pointer, or that restrict the virtual keyboard to their input method (Denial), now bring up capture and fall back to uinput for input. UU_WAYLAND_NO_VIRTUAL_KEYBOARD=1 skips the virtual keyboard.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
From the official Mac client (UURemote 4.42; QRCodeGenLoginResponse decodes only qrcode_jump_url
and status_query_ticket, the status/by_qrcode bodies carry that ticket under "token"):
- gen/login is sent without a body; the status ticket is status_query_ticket (older token as fallback).
- The status long poll re-asks right after each answer with a 60 s request timeout (was 12 s).
- Status 3 regenerates at once, status 0 after 0.35 s, 4 logs in (3 no longer calls by_qrcode).
- Errors such as HTTP 400 code 1190 and timeouts stop polling until the user clicks the new
刷新二维码 button (Enter in a terminal; one minute when neither exists). This replaces the
v0.2.19 backoff and the re-register hint.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- Never spin on immediate unchanged status answers or transport errors (2 s backoff); repeated
QR codes that expire within seconds back off exponentially and point to re-registering the
device identity, which recovered a flagged device in #68.
- Custom bitrate field 8 is bit/s (#67); values that can only be Mbit/s or kbit/s are scaled
instead of capping the encoder at 1 kbit/s.
- doctor: without DISPLAY the X11 engine still tries :0/:1, so the #66 warning says that.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Without DISPLAY and without WAYLAND_DISPLAY the desktop backend cannot
capture anything: the host logs 'Desktop unavailable ... x11_init failed'.
The report named X11 as the selected backend and nothing else, which reads
like a working setup. Name the missing display and note that the terminal,
file transfer and port mapping tools still work.
systemctl start returns 0 when the unit is skipped by a Condition or keeps crashing under
Restart=; show now reads ActiveState/SubState/ConditionResult while waiting and names the real
problem. ffmpeg-sys-next's build script runs before ours, so FFmpeg >= 7 cannot be caught in
build.rs; the README and AGENTS state the 6.x requirement next to the build commands.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Mirrors the official Windows host (GameViewerServer reportLanInfo / PushMessageHandler::wakeOnLan /
send_wol): report the LAN identity after sign-in and answer wake_on_lan pushes with a magic packet
to the target's subnet broadcast. Field names come from the official log strings; off by default
and unverified end to end. wake_on_lan push key names are always logged for evidence.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Session, selected backend (and the X11-inside-Wayland trap), engine plugin load results, uinput
and render-node access, FFmpeg encoders/version, the default codec, sign-in and service state.
Prints booleans, names and versions only; home directory and user name are masked. The issue
template and READMEs ask for it.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- KDE portal sessions hold an org.freedesktop.ScreenSaver inhibitor while a desktop connection
is active (GNOME keeps SessionManager.Inhibit), so the idle timeout no longer locks mid-session.
- Log a hint when the desktop locks: viewers only see black frames until it is unlocked.
- Warn when the X11 backend is selected while a Wayland session is running (XWayland root
capture is all black without any error).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The in-process encoder opened its context at the raw capture size, so an
odd logical desktop (e.g. two stacked 150% monitors, 2560x2507) made every
H.264/HEVC/NVENC encoder fail to open. The encoder size is now rounded up to
even, and a native capture is copied into a zeroed frame of that size so the
extra row/column stays black. Scaled output paths are unchanged.
Co-authored-by: ferstar <[email protected]>
libx265 is about ten times slower than libx264 (15 vs 155 fps at 1080p on a 4-core Ampere A1),
so without NVENC/VAAPI the host now picks H.264 even when both sides offer H.265;
UU_VIDEO_CODEC=h265 still forces it. The release workflow builds x86_64 and aarch64 packages
(each with both engine plugins) and publishes them from a single job.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The release tarball was X11 only, so Hyprland/niri and GNOME/KDE Wayland users had to build from
source (#43, #54). With --features engine-plugins the wayland and portal engines are built as
libuu-wayland.so / libuu-portal.so and dlopen'ed only when the session selects them, so X11 hosts
need none of their libraries. The wayland plugin is built on Ubuntu 22.04 with the main binary;
the portal plugin needs libei >= 1.1 and is built on 24.04 (glibc >= 2.39). Source builds without
engine-plugins keep linking the engines statically.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Release packages are X11 only; on GNOME/KDE Wayland auto selection picks the
portal backend, which then failed every 5 s with the opaque "portal
unavailable (code 2)".
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The portal backend reports the compositor's logical desktop size, which is
odd when a scaled monitor's logical extent is odd: two stacked monitors at
150% scale capture 2560x2507. H.264/HEVC 4:2:0 encoders (and NVENC's surface
allocation) reject odd dimensions, so with such a desktop every encoder
failed to open:
[libx264] height not divisible by 2 (2560x2507)
[h264_nvenc] InitializeEncoder failed: out of memory (10)
ERROR Live desktop stream stopped: Encoder output ended
and the official viewer stayed on 正在传输画面 while input still worked.
Pad the odd row/column in the filter chain when the native capture size is
known and no output scaling was requested (a no-op case omitted for even
sizes). The host display mode, the reported capture size and the explicit
分辨率 scaling path are unchanged; the fixed-input probe keeps its chain.
Co-authored-by: wen0531 <[email protected]>
A local HTTP/SOCKS proxy that intercepts TLS caused login to fail with
"tls handshake eof". Build all cloud API clients with no_proxy() unless
UU_USE_PROXY is explicitly truthy, so direct connections are the default
and environments that genuinely need a proxy can still enable it.
The mobile clients' input method sends `text_input {content}` over CONTROL. Upstream only
implements it for X11, so a Wayland host (Hyprland here) dropped the characters with a
warning. Wire it up: characters the session keymap already produces are pressed as
ordinary keycodes, and characters it cannot produce are carried by spare keycodes (the
keymap has ~33 named but unmapped ones) declared for the duration of one batch, in text
order, then removed again.
The focused client translates keycodes with the keymap the compositor last handed it, and
a keymap uploaded to a keyboard the seat already has is never handed on again: measured on
Hyprland 0.56.2, pressing the remapped keycodes then reproduces the previous message's
text (a space came out as the previous message's first character, and every message
repeated the one before it). Re-attaching the virtual keyboard is what makes the
compositor send the keymap to the client, so each batch does that, waits for the
compositor to install the keymap, and releases held keys first because the new keyboard
starts with none. No timing sleep is needed once the keyboard is re-attached, so there is
none. Plain latin text and spaces never touch the keymap at all. The pristine keymap is
restored after every call, so uu_wl_key keeps its own symbol map and the compositor is
never left with spare keys bound.
`wayland-text-probe` reports how many characters one upload can carry and types a sample;
UU_PROBE_SEQUENCE types one message per |-separated field, the way the phone's input
method sends them, which is what reproduced this locally.
Co-authored-by: Hey Claw <[email protected]>
Co-authored-by: Claude Code <[email protected]>
The iPhone client ignored our flat forward_setting device_capability and kept
quality and chroma locked. Send it nested under data.device_capability with
type beside it, list the formats the encoder really produces (4:4:4 when
available), and add sender_para_info, as the official Windows host does.
Verified with the UU iPhone app 4.42.0.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
* HEVC: split access units by picture boundaries, and stop passing -aud to hevc_vaapi
hevc_vaapi with `-aud 1` writes an AUD only at IRAPs and drops every other picture from
the bitstream: a 90-frame 640x360 clip came out as 3 IDRs and no P-frames. The reader
splits access units on AUDs, so each "frame" became one IDR plus a whole GOP, and the
IRAP gate then rejected all of them - the H.265 path sent no video at all (measured
~12 kbit/s outbound against 4-6 Mbit/s for H.264).
HEVC now leaves AUD off and hevc_au_offsets() splits on RFC 7798 picture boundaries (an
AUD, or the first NAL after a VCL NAL that is non-VCL or starts a picture), which also
keeps multi-slice pictures in one unit and covers the libx265/NVENC shapes. Parameter
sets still repeat in-band ahead of every IDR, so no out-of-band sprop is needed.
The fixed-input probe hardcoded H.264 in three places (settings, output muxer, first-unit
check), which is why this path was never measurable; it now follows UU_VIDEO_CODEC.
Verified with desktop-video-probe (90-frame fixed input): hevc_vaapi and libx265 both
emit 90 frames, first unit an IRAP with its parameter sets.
Assisted-by: DeepSeek V4.1 Flash-1017
* HEVC: keep suffix SEI, filler and end markers with the preceding picture
RFC 7798 / H.265 7.4.2.4.4 lets only AUD, VPS/SPS/PPS, prefix SEI and types
41-44/48-55 open an access unit; trailing NALs stay with the picture. Verified on
an RTX 3090 with the fixed-input probe: hevc_nvenc and libx265 both yield 90/90
decodable units with an IRAP first.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
---------
Co-authored-by: VallierDev <[email protected]>
Co-authored-by: Claude Opus 5.5 <[email protected]>
The 100-500 kbit default keeps NVENC keyframes inside the buffer, but VAAPI quantizes a
whole desktop keyframe when the buffer is that small: on a 2560x1440 H.264 VAAPI stream
capped at 30 Mbit/s the first IDR measured 59,910 bytes and the picture blurred once per
GOP before sharpening back.
UU_VAAPI_BUFFER_MS (10-1000 ms, VAAPI only) sizes the buffer as bitrate x ms/1000. With
250 ms on the same input the first IDR measured 142,397 bytes and the worst-frame Y-SSIM
over 180 static frames went from 0.9492 to 0.9933, with no rise in average bitrate.
Opt-in: unset, out of range or unparsable keeps the old buffer.
Assisted-by: DeepSeek V4.1 Flash-1017
Host: accept the official text_input {content} action on any channel and type it on
X11 via spare-keycode Unicode keysyms (verified with xev); log non-CONTROL JSON
action names. Viewer: Windows-type hosts get only the case-10 text change (a rich
format list overwrote it and was never rendered), Mac gets only the rich offer.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Log unsupported CONTROL input actions (names only) and the server NumLock
state for numpad keys; skip the SDL window test on macOS (main-thread SDL).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The official Mac host answers check_env on the BINARY terminal channel next
to TERM frames; the viewer parsed it as a TERM frame and closed. Verified
live: Mac host terminal opens and echoes input.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The Mac host offers pasteboard text as format 0 "public.utf8-plain-text"
(UTF-8), not CF_UNICODETEXT, so the viewer never requested it. Accept and
request that format, and offer it next to CF_UNICODETEXT from the viewer.
Host-side offers to official clients are unchanged. Mac -> viewer verified
live; viewer -> Mac not yet. Adds format-name/echo diagnostics (no data).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
FFmpeg 9 removed -vsync, so v0.2.2/v0.2.3 decoders exited at start and every
frame failed. Verified against the official Mac host on FFmpeg 9.0.2.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
POST /api/v1/device/{id}/unbind, the path the official Windows host and Mac
viewer use for 「删除设备」. Verified live on a disposable test device; the
shared-device path (/api/v1/device/share/{id}/unbind) comes from strings only.
CLI: uuremote-viewer unbind DEVICE [--shared]; launcher: 更多工具 → 删除设备
with a confirmation dialog.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
* Fix viewer Dock churn by reusing SDL clipboard window
* Handle clipboard text boundaries and read only after changes
* Keep SDL clipboard scoped to the focused desktop and validate CI call sites
The official Windows host only acknowledges the rich format-list offer
and never asks for the data. Each local change now also sends RpcRequest
case 10 ClipboardTextChangeRequest with the text (its
handle_text_change_request -> notifyReceiveText path); the host answers
case 6 err=1 (CB_SUCCEEDED), logged as a status only. This project's
host ignores case 10 and keeps using the rich offer. Note: on the tested
Windows host the acknowledged text still did not reach the desktop
clipboard; under investigation.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Live against an official Windows host (PowerShell via ConPTY):
- Its rpc_main.proto has no check_env (requests end at 25), so a missing
answer within 3 s no longer aborts; a negative answer still does.
- OPEN carries "option":"none"; the host rejects an OPEN without it
("terminal open: missing 'option' field", kError 1004). This
project's host ignores the key.
- The host answers OPEN with {code,message}, header session 0, and keeps
one terminal per connection as session 0; a nonzero header still
overrides the body for this project's host.
- Echo requests whose SimpleAction omits the default action 0 are
answered too.
Create/reattach on this project's host unchanged.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The official Windows host answers every PortMappingFrame with session_id
1 regardless of the viewer's value, so the random 64-bit session never
matched and every stream ended with a SYN timeout. One port-mapping peer
carries one session; use 1 (this project's host echoes it). Stream
failures and unmatched frames are now logged with reasons and wire types
only. Verified: HTTP through the official Windows host and this
project's host.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Live against an official Windows host:
- Uploads split Windows targets (C:\dir\name) on either separator and
keep the host's separator for the directory; '/x' keeps the root.
- The upload request carries FileInfo.modified_time, the zlib FileList in
list_data (field 6) and a task_unique_id, as official viewers with the
list-compress feature do; the Windows host rejected the inline-only
request with code 2. The inline copy stays for older hosts.
- Downloads accept blocks up to 256 KiB: the Windows host sent a 64 KiB
file as one block, which the old 32 KiB limit rejected with code 3.
64 KiB and 1 MiB round trips matched by sha256 on the Windows host and
on this project's host.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Offer H.265 (RFC 7798) next to H.264 and list H.265 and 4:4:4 decoder
caps; the host picks the codec. Depacketize single NAL, AP and FU units
with the existing reorder/loss handling and decode with the ffmpeg CLI
(-f hevc). UU_VIEWER_CHROMA=444 requests true color via
CaptureSettingRequest field 7 and follows CaptureSettingSync;
UU_VIEWER_CODEC=h264 keeps H.264 only.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Without polkit permission (for example no active local session)
systemctl reboot/poweroff waited for an authentication dialog on the
controlled screen. Pass --no-ask-password so the action fails and is
logged instead. README: the viewer's terminal session list, reboot and
shutdown, and its Wayland window/clipboard were tested in a VM.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
* Add the experimental native viewer (uuremote-viewer)
A Linux viewer/controller for UU Remote hosts, built separately with
--features native-viewer,viewer-sdl; the host build and behaviour are
unchanged.
- Cloud join by device and remote assistance by share ID + code (code
via stdin only), WebRTC offer side, SDL desktop window with ffmpeg CLI
H.264 decoding, Opus sound, keyboard/mouse, view-only mode, text
clipboard on TEXT_DATA_CHANNEL.
- Tools: file browser/transfer (zero- and one-based download blocks),
terminal and terminal session list, port mapping (127.0.0.1 only),
reboot/shutdown; tools reuse the desktop peer while one is open.
- GTK launcher with device list and tool windows.
- CI builds and tests the viewer; release tarballs include it.
Logs carry field numbers, sizes and states only.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
* ci: install the ffmpeg CLI for the viewer audio tests
Co-Authored-By: Claude Opus 5.5 <[email protected]>
---------
Co-authored-by: VallierDev <[email protected]>
Co-authored-by: Claude Opus 5.5 <[email protected]>
The X11 engine has one process-wide display connection, but every
ScreenCapture closed it on drop. During guest assistance the device
init's screen-size probe created and dropped a capture after the main
one, closing the display under it: the stream stopped with
"X11 display wake failed" and the viewer got no video. A capture
replaced after a resize had the same hazard. Count live X11 captures
and close only on the last drop.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
* fix(portal): bootstrap video for iPhone locked auto unlock
* docs: keep the #29 experiment record in the PR instead of the repository
Co-Authored-By: Claude Opus 5.5 <[email protected]>
---------
Co-authored-by: VallierDev <[email protected]>
Co-authored-by: Claude Opus 5.5 <[email protected]>
* fix(portal): establish locked desktop control before capture
* test(portal): preserve v4 pixel inventory after deferred capture settings
* fix(control): refresh lock state after viewer Echo handshake
* fix(portal): announce desktop target before locked capture wait
* ci: run the desktop_power native test under ASan/UBSan
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
* input: skip the Portal release sweep while no session is bound
The first locked connection creates its peer before any Portal exists; the sweep
then logged two failed injections and a 'reset incomplete' warning for nothing.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
* docs: describe locked cold connections, auto unlock limits and iPhone status
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
* fix(portal): treat a cached screen with a closed Portal session as not ready
begin() now marks the state as waiting when the cached session is dead, and
permission/video/clipboard consumers honor it instead of reporting Allow and
starting the stale video path before maintain() has rebuilt capture.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
* fix(portal): gate input on a ready desktop and share recovery backoff
Input dispatch kept only the cached screen and dropped State.waiting, so a closed
Portal passed the gate once the screen unlocked. The supervisor's recover() and the
encoder's stream-loss recovery now share failure backoff and a terminal-denial flag,
so a denied consent is not re-prompted by the other path.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
* fix(control): keep unknown lock state unlocked, stop waits on Closing, never drop capture settings
- QuerySystemState with no logind lock state (Xvfb/linger) reports unlocked as before.
- The Portal wait loop also exits while the channel is Closing.
- Capture-setting requests back-pressure instead of try_send dropping the latest setting.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
---------
Co-authored-by: Claude Sonnet 5.5 <[email protected]>
Co-authored-by: VallierDev <[email protected]>
The codec-negotiation and ICE/SCTP acceptance tests used real local sockets
and occasionally hit the 30 s connect timeout when the whole suite ran in
parallel. They now use webrtc-rs vnet (shared test_network helper), so they
are deterministic and the acceptance test no longer needs #[ignore].
Test-only change.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
UU_PROBE_INPUT=<video or image sequence> bypasses desktop capture and encodes
the file with the production FFmpeg encoder arguments, so encoder
measurements no longer depend on desktop content; UU_PROBE_LOOP=1 repeats the
input up to UU_PROBE_FRAMES. Missing/undecodable input exits non-zero. The
session path is unchanged. Verified with Y4M, MP4 and PNG sequences (repeat
runs byte-identical); NVENC/VAAPI not hardware-verified.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Follow-up to #26: retire_if_closed errors (poisoned lock, panicked task) are
logged and treated as not stale instead of returning from serve().
Co-Authored-By: Claude Opus 5.5 <[email protected]>
* fix(portal): recover capture after live scaling changes
* fix(portal): keep input available during session recovery
* fix(portal): serialize reconnect recovery and throttle lock probes