* fix(input): stop controller rumble when the game that started it is gone
A crashed game never sends the stop for a held rumble, so the keepalive re-armed it every 1.5 s until the session ended. A held effect now stops on the next keepalive once the process that sent it or the Steam game it started in has exited. The fake controller also stops its rumble when it's closed or the playing effect is erased, like the kernel does.
* test: read the test's own pid from /proc/self
* fix(input): track who plays rumble on each slot, check the game when a hold starts
The process that plays an effect now owns it, and only the handle currently playing on a slot sends the stop when it closes or erases its effect. A held effect records whether a game is running when it starts, and keeps checking that its sender still exists even when its start time couldn't be read.
* fix(input): only cache a running game, not its absence
A new hold claimed right after a no-game result reused it, so a game that started within the cache window was never watched.
* fix(input): treat a vanished sender of our own user as gone
Guest processes run as the app's user, so the app can always see them in /proc. If one is missing at the first lookup it already exited; only senders of another user (root-spawned guests) are left unchecked.
* fix(input): a stop only ends rumble the same process started
The fake controller's playing-handle tracking is per process, so another process closing or erasing its older effect still sent a slot-wide stop. The app now ignores a stop whose sender isn't the one behind the slot's current effect; stops from an unknown sender work as before.
---------
Co-authored-by: Tapuuk <177422350+Tapuuk@users.noreply.github.com>
* fix(ci): a test build's files outlive the catalog that names them, and a merged PR's build lasts until Preview carries it
- A new push's PR release keeps the previous push's files until the catalog names the new
ones, then five minutes more, so a download started from the old catalog still finds its APK.
The app also reloads the catalog and retries once when a test build's download fails.
- Closing a PR without merging still deletes its test build. A merged PR's build stays, and
stays listed, until the newest Preview build was built from a main holding the merge, or
three days pass (docs-only merges build no Preview). Main's build, the PR cleanup and the
sweep share one script for this.
* fix(ci): name the current push's APK in the catalog and prune nothing it still names
* fix(ci): only a Preview with every APK carries a merge or is offered
* fix(ci): publish no preview while none has every APK
* test(ci): build_catalog publishes no preview while its APKs are uploading
* fix(venus): DXVK runs on Mali GPUs without dual-source blending
DXVK requires dualSrcBlend for D3D11 and skips any GPU without it, so on
Mali through Venus no adapter was found. The Mali compatibility layer now
reports the feature and maxFragmentDualSrcAttachments 1, keeps it out of
the driver's device, turns fragment outputs decorated Index 1 into Private
variables (shader modules and inline stage code), and maps SRC1 blend
factors to source alpha in pipelines and vkCmdSetColorBlendEquationEXT.
Only active when the driver lacks the feature.
Adds a mock loader and driver test for the layer, and fetches the
layer's pinned headers for it in CI.
* build(venus): pin the Mali component with dual-source blending
* fix(agent): one-state composer with model and usage on top; reading back stays detached while the agent streams
- The composer no longer has a compact reading bar that a first tap expands: the editor is
always the editor, and the model picker and context ring move to the top beside the menu.
- The pet chat drops its DeeDee title so that top bar gets the whole width.
- Scrolling back stops following new output; new entries and streamed text only scroll the
list while it is at the bottom, and sending a message resumes following.
- When the editor slides back in or the keyboard rises, a list at the bottom keeps its last
lines above the editor, even mid-drag.
* fix(agent): a revise while reading back shows the editor it focuses
* feat(agent): dictate a message, and attach what the running game shows
- With nothing typed, the editor's send button dictates instead, through Android's speech
recognizer; what is heard follows the text already there. It asks for the microphone the
first time.
- While a game session is open, a camera button attaches the game's current frame. It is
copied from the game's own surface, as the bottom screen's screenshots are, so the chat or
drawer over it is not in the picture.
* feat(agent): a notification when a turn ends unseen, and replies that end in choices offer them as buttons
- A turn that ends while neither the Agent page nor the pet is in view posts a notification:
finished, failed or asking, with the reply's first line. Opening the Agent page clears it.
- Claude Code's question tool is not available in print mode, so the brief asks the agent to
end a reply with a choices block instead. The chat shows its options as buttons, and the one
picked is sent as the next message.
* feat(agent): a subagent's work shows under the call that started it
Claude Code tags a subagent's messages with the call that started it. Its text, thinking and
streamed deltas used to show as the main agent's reply, and its tool calls mixed into the
turn. Now only its tool calls are kept, stored with their parent, and shown inside the
parent's card: a live line with the step count and the latest call, and every call when
the card is opened. Its token counts no longer overwrite the context reading.
* feat(agent): copy a reply or a tool's text, and open a turn's changes in one place
- A turn's footer has a copy button for everything the agent said in it, and tool input and
output boxes have one each.
- The files-changed chip opens every edit the turn made, file by file.
* feat(agent): search conversations in History
A field above the list keeps the conversations whose title, messages or replies hold the
text, ignoring case. Chat files are read off the main thread, a quarter second after the
last keystroke.
* fix(agent): sending while reading back jumps to the latest
* feat(agent): manage Claude Code's MCP servers and skills from the Agent menu
A new panel lists the MCP servers claude mcp list reports, adds one for every project from a
URL (over HTTP) or a start command (over stdio), and removes it; connectors from the
claude.ai account are listed but managed there. It also lists the skills in
~/.claude/skills with their descriptions, and deletes one after a confirming second tap.
* fix(agent): review follow-ups for the Agent tab refinements
- Choices only come from one fenced block that ends the reply.
- The Agent page counts as on screen only while its activity is started, so a turn ending
during a game still notifies.
- Picking a choice or Go jumps to the latest like sending does.
- Dictation stops when the chat is hidden or the keyboard sends; stopping it by hand keeps
the recognizer until its final result.
- A failed JPEG encode is not attached.
- The copied reply is read from the turn each time.
- MCP start commands keep their quoted arguments; the CLI is read beside its 90 s deadline;
deleting a skill runs off the main thread and removes links without following them.
* fix(agent): scroll the shrinking list after layout, not during measure
* fix(agent): fence-aware choices parsing; dictation stops overwriting edited drafts
* fix(agent): ignore choices blocks nested in a code fence
* fix(agent): treat tilde fences like backtick fences when finding choices
* test(agent): cover a tilde-fenced choices block
* fix(epic): install the Epic Games Launcher from its Linux .deb
Epic no longer serves the Linux AppImage: the installer URL answers 404
errors.com.epicgames.launcher.app_direct_url_not_found (30003), so Install
failed for everyone. Epic now ships EpicGamesLauncher-x86_64.deb
(0.113.0+g2164614).
- The installer downloads the .deb, checks its ar magic and the CDN ETag,
and droiddeck-epic unpacks its data.tar.* member straight from the
archive.
- The .deb links WebKitGTK 4.1, JavaScriptCoreGTK 4.1 and GTK 3, which the
Steam Runtime 4 image does not ship. 32 Debian 13 amd64 packages, pinned
by SHA-256 and size (60.2 MB download, 201 MB unpacked), are unpacked
into the launcher's desktop/ folder (shared libraries only) and put on
its LD_LIBRARY_PATH. Debian 13 matches the runtime's own base, so only
sonames the runtime lacks are added. Each package falls back to
snapshot.debian.org when the deb.debian.org pool has moved on.
- WebKit starts its helper processes from its built-in
/usr/lib/x86_64-linux-gnu/webkit2gtk-4.1 path, so each launch links that
folder in the runtime to the layer's copy.
- An installed AppImage launcher keeps starting through its AppRun; the
next update check moves it to the .deb.
Closure checked offline: every ELF in the .deb and the layer resolves with
the runtime's own ld-linux-x86-64.so.2 inside a sandbox holding only the
runtime, the layer and libGLESv2 (0 missing, 0 version or undefined
symbol errors with LD_BIND_NOW).
Based on zooluis's research in #2268.
Co-authored-by: zooluis <150430329+zooluis@users.noreply.github.com>
* fix(epic): provide desktop GLib data paths
* fix(epic): keep archive symlinks from redirecting the desktop layer and icon
---------
Co-authored-by: zooluis <150430329+zooluis@users.noreply.github.com>
* feat(updates): install any kept Preview build, link builds to their PR and commit, flag stale test builds
* fix(updates): address review of the Updates tab history and test-build watch
* fix(updates): notify on switch to a followed PR, keep higher-version rebuilds installable
* fix(ci): let full reruns replace every artifact, note only expired PR builds
* fix(ci): leave the rebuild note only when the PR build is too old to rerun
* fix(ci): poll unfinished PR builds together and queue version-bump runs
* fix(ci): also rebuild PR builds that were running when the versionCode changed
* fix(ci): rebuild fork and just-finished PR builds, recheck the PR head before a rerun
* fix(ci): fail the rebuild sweep when finding candidates fails
* fix(ci): time the bump from the rebuild run's creation and scan unfinished builds of any age
* fix(ci): check the latest build of every open PR instead of scanning run history
* fix(ci): count PR builds that finish between the version push and the rebuild run
* fix(ci): use the push time from the event as the rebuild boundary
* fix(runtime): record why a runtime install failed in the session events
* fix(runtime): report empty or invalid runtime catalogs in the failure event
* fix(runtime): keep each install failure reason on its own job
* fix(runtime): name the busy reason for rejected installs and drop the owner's listener
* fix(runtime): keep each install call's failure reason on its calling thread
* fix(runtime): keep the interrupted reason when a joined install finishes during the wait
With a single GPU heap DXVK reports the whole heap as shared memory too, so games still saw ~8 GB as shared even with the cap on. Set dxgi.maxSharedMemory to the same value, like BattleNet already does.
Co-authored-by: Tapuuk <177422350+Tapuuk@users.noreply.github.com>
* feat: every HUD reading has its own switch and a row order
* feat: readings get a reorder dialog of their own
* refactor: drag alone moves a reading
* fix: the readings list is one column with a switch per row
* fix: one scrolling readings list, titled HUD info
* refactor: derive the readings order from the defaults instead of a second list
* fix: the readings dialog loses its doubled title and follows the saved order
* refactor: no title in the readings dialog, its actions sit beside the list
* refactor: readings dialog actions take the free side on landscape, the foot on portrait
* fix: the readings list follows the size it draws and drags with a live gap
* fix: the readings dialog fills the window and keeps scrolling under a held drag
* fix: reset restores every size's order, and the new HUD strings are translated
* refactor: one table for each size's rows and one for its text sizes
* refactor: the readings list swaps rows as the drag passes them, and the list animates the rest
* refactor: give the readings rows test tags for the local UI test
* refactor: HUD readings follow one configurable row table
* refactor: the Strip HUD draws its rows from the same table
* refactor: each size orders its rows by its own table, and the frame graph is a row
* fix: the readings dialog keeps the row order a drag leaves
* fix: each size keeps its own row order, and reset clears only the order
* refactor: the graph's size lives in one place, and the section markers are gone
* fix: a size change and the first save of a legacy config both keep the order
* chore: retrigger the build
* feat: a reading can be grabbed with Y and shuffled with the d-pad
* feat: controller hints float beside the dialog, and a moved row stays in view
* feat: Y grabs on a hold and drops on release, and a touch no longer flips the dialog
* fix: keep the key hints off the rows on narrow windows, and let Enter toggle a row
* fix(battlenet): a Battle.net download that stops moving ends as retryable, and a killed update comes back as an update
A job reply without a state or byte counts no longer counts as a started
download, and an Agent job that reports no change for ten minutes is
stopped with its Agent and fails with a Retry instead of waiting forever.
Updates remember their game folder in the saved queue, so one restored
after the app was closed runs agent-update on that folder instead of a
fresh install.
* fix(battlenet): a stalled job stops only the Agent it started
A game or Battle.net client running in the prefix keeps running when a
stalled Agent job is ended, and the progress test keeps the job alive
across polls with a controlled clock.
* fix(battlenet): only real progress holds off the stall timeout
The download rate and time estimate can change while no bytes move, so
the stall timer now resets only on byte counts, progress, state, phase,
pause or error.
* test(battlenet): bound the stalled-rate fake child's polls
* fix(session): keep the session awake while a browser hand-off is pending
When a program in the session opens a link in the Android browser, the
session view goes to the background and auto or manual suspend froze the
session within a couple of seconds. Programs that wait for the browser to
call back to a local listener (external sign-in flows) were frozen before
the callback arrived, so the sign-in never completed, and the session then
needed a tap to continue.
An open-url request now starts a browser hand-off that counts as a visible
surface until the session view comes back to the front or ten minutes pass.
Ordinary backgrounding still suspends as before.
* fix(session): hold through a display power change and only after the browser opens
* l10n: translate the strings added since the last pass
* l10n: add the input library and input strings in every locale
* l10n: keep the input hints unformatted and drop the untranslatable keys
aapt2 rejects input_axis_action_hint and input_trigger_action_hint without
formatted="false", because their translations hold two bare percent signs.
The default resource sets it, the 32 locales did not.
The 40 translatable="false" keys are left out of every locale again; lint's
ExtraTranslation check is fatal in release builds and the entries were added
here by mistake.
Also fixes wording the review caught: the Danish and Norwegian labels for
sticks and neighboring directions, the Bulgarian and Vietnamese D-pad names,
the Greek release and remap wording, the page names Hungarian, Indonesian and
Russian point users to, the Arabic log folder path, and the Finnish library
removal hint that read as deleting the folder.
* l10n: translate the strings the newer upstream commits added
* l10n: drop the two keys upstream no longer ships
* Start a launcher-run game without scanning its folder and keep Play on Starting until it launches
* Create the shared added-games cache under one lock
* Read a rewritten game manifest again instead of reusing the saved ranking
* Tie a saved ranking to the declared executables that exist
* fix(native): drive links stored as rootfs paths reach the game library mount
* fix(native): relative drive links reach the game library mount
* fix(native): keep the path after a relative drive link as written
* fix(native): relative links that climb past the rootfs root stop at the guest root
Proton copies default_pfx/dosdevices/z: into each prefix with readlink and
symlink. The relinked target climbs as many folders as default_pfx is deep,
which is one more than a Steam prefix, so natively the kernel left the rootfs
and Z: pointed at the app's files folder. Proot stops ".." at the guest root;
the hook now does the same for links it resolves.
* fix(native): resolve drive links in prefixes stored in bind sources outside the rootfs
* fix(native): detect equal-length mount conversion for bound prefixes
* feat(agent): GitHub sign-in floats its device code over the browser
* fix(agent): the device code falls back to the browser when nothing could open
* fix(agent): dd-gh-login waits for DroidDeck to report whether the page opened
* feat(storage): name, select and remove every prefix and store game
* feat(storage): bulk remove picks prefixes, shader caches or both
* fix(storage): count and clear a game's shader cache in every library
* fix(storage): guard shared prefixes, batches and active downloads
* fix(storage): count each shader cache once across games
* fix(storage): hold store uninstalls against downloads and keep every shortcut's options
* fix(stores): keep download cleanup and uninstall from overlapping
* fix(stores): wait for an installer's file discard before uninstalling
* fix(stores): a reinstall waits for the cancelled download's file cleanup
* fix(storage): fast first scan, readable 4:3 layout and every shader cache counted
* fix(storage): keep store sizes current and track discards from failing jobs
* fix(stores): uninstall waits for a cancelled update's cleanup
* test(stores): a reinstall waits for a failed job's discard
* fix(session): the Steam UI on the Mali bridge stops flickering
On the Mali bridge Steam's own web helper layout makes the Deck UI footer
and overlays flicker (Mali-G78 MP14). When the browser UI is on the GPU,
default to ANGLE on Vulkan without Chromium's Vulkan compositing. Values
set by the user still win, and the software UI path is unchanged.
* fix(session): keep empty Mali bridge CEF settings as set
With "Run without proot" on, the Steam client hung about a second after
login and aborted three minutes later with "fatal stalled cross-thread
pipe". The directory cache holds its lock while it adds an inotify watch.
Without proot that call reaches the native userland hook, whose
inotify_add_watch creates a missing directory with mkdir() by name. That
mkdir resolves to the cache's own wrapper, which waits for the same lock
on the same thread. Steam's IPC threads then queue behind it.
The cache now counts the locks each thread holds. While a thread holds
one, calls that come back into the cache (opendir, the change wrappers,
created files, missing-name checks, closedir of a stream it did not open)
go straight to the next library. The same guard covers the hook's
remove(), which calls unlink() by name while the cache's remove wrapper
holds its operation lock.
* fix(proot): fchdir takes the open directory's path instead of walking it again
* fix(proot): fchdir keeps trailing-period names and rechecks the binding's parent
The descriptor path now carries the same final marker the walk produced, so a directory whose name ends with a period is not truncated. The parent of the matching binding is resolved again, so a binding whose ancestor was replaced by a symlink takes the full walk as before. The check script gains a trailing-period directory.
* fix(proot): fchdir requires the binding's parent to be a directory
* fix(ui): keep the header height when search opens
The search field in the page header reserved a line for its empty label, which made the open field taller than the header row and pushed the page down.
* fix(ui): return focus to the search button when search closes
Closing the header search with Back removed the field while it held focus, so the highlight was lost. The search button takes focus once it is back on screen.
Down from the row above a short last row landed on a tile further left, because the grid's own search only sees the tiles laid out so far. Down now goes to the tile directly below, or to the last tile when there is none below, and scrolls the next row in first when it is not laid out yet. Widgets record focus too, so the grid knows when focus is not on a tile.