475 Commits
Author SHA1 Message Date
Ryan Vogel 44be686fde Android: never suggest slurs, and swear words only when asked (#2081)
* Android: bundle a curated, attributed never-suggest list

Separate identity slurs from opt-in profanity for English, Spanish, German and Bokmål, reusing CC BY 4.0 LDNOOBW data with an explicit source revision, modifications and credit on the in-app licences screen. Include common inflections, keep neutral anatomy and ordinary local meanings, and match selected expressions as whole pairs.

Checked all 653 rows against the pinned four upstream lists, verified the attribution and licence, and compared all four HXTM files from Android, iOS and the debug APK byte-for-byte. No language model or probability is replaced.

* Android: filter keyboard candidates without rewriting typed words

Replace unconditional HXTM output with lookup-time flags for prefix and next-word suggestions, deterministic fixes, whole-word spatial readings, and both halves of missed spaces. Still use blocked words for probabilities, and keep a fully typed spelling (including out-of-model words) through space or a distinct literal chip. The Inter Keyboard switch enables profanity only; preference-aware caches refresh even in the active editor.

Verified 8 filter tests and 5 existing tests in each of debug and release, including every curated entry across the four layouts. Reproduced offensive suggestions and space replacements on the Studio API 35 IME before the change; after it, confirmed safe near-miss results, intentional literals, live toggle refresh, and light/dark settings. The 4,000-sentence phase-4 replay changed English WER from 7.4045% to 7.4117% and left Spanish at 10.3301%.
2026-11-20 18:02:48 +00:00
Dax 3fc14b5ec9 labs: neutral voice for chosen modes (#2043) (#2049)
* labs: neutral voice for chosen modes (#2043)

* labs: tighten neutral voice guard and reuse cleanup helpers

* labs: share neutral voice quote and fallback helpers

* labs: keep folder paths protected in neutral voice
2026-11-20 11:51:08 -05:00
Dax 90cca52748 windows: don't paste into the desktop or taskbar, keep the dictation on the clipboard (#2016)
* windows: don't paste into the desktop or taskbar, keep the dictation on the clipboard (#2009)

* windows: keep the shell fallback plain and unmarked, and accept a copy in practice (#2009)
2026-11-19 19:10:26 -05:00
Kit Langton ec5e7282ca fix(labs): Switch model with app matches a picked app under its Finder name too (#2018)
The app picker saves the name Finder shows for the bundle's file, but Switch model with app looked modes up by the running app's own name only, so a mode picked as "Rex Beta" (Rex Beta.app runs as Rex) never switched the model. The frontmost read used by the switcher now shares capture_frontmost's name read and passes Finder's name alongside, and the mode lookup accepts either, as Modes and YOLO mode already do. Windows keeps matching its process name.
2026-11-19 15:17:36 -08:00
Dax 4d28a692d8 add fluid speech pro to the speech model catalog (#2007)
* add fluid speech pro to the speech model catalog

FluidVoice's Fluid Speech Pro, a Parakeet TDT 0.6B v3 retrain that also
declares Moondream's Parakeet Ultra as a base, for the same 25 European
languages. TalkMate converted it to a Q8_0 transcribe.cpp GGUF at d59dc79
(741,819,456 bytes, sha256 c3848e98...); the file says
parakeet/tdt-0.6b-fluid-speech-pro and announces its extra pre-Conformer
block as stt.parakeet.encoder.pre_block_kernel, so it needs a transcribe.cpp
that reads enc.pre_block.

Like Parakeet v3 it takes no language hint, detects the language under Auto,
does not stream, and is never recommended or badged. Its card publishes no
LibriSpeech test-clean figure, so it carries no single word error rate and no
benchmark record; its per-language FLEURS figures and the CC BY 4.0 credit
are in its details.

* bump(transcribe.cpp): 295c7e6, fluid speech pro's pre-conformer block

Fluid Speech Pro's GGUF announces stt.parakeet.encoder.pre_block_kernel and
carries eight enc.pre_block.* tensors: a residual FrameCleaner block between
the subsampler and Conformer layer 0. The old pin's graph never read them, so
the model loaded as plain Parakeet v3 and its encoder was wrong.

Both transcribe-cpp crates move to 295c7e601 on hex/combined-fluid, a
fast-forward of bcb639f that runs the block on pre_encode before xscaling
when the KV is present, and fails the load rather than skipping tensors when
it is only partly there. Cargo.lock changes only those two sources
(`getrandom`'s tempfile pointer was reverted to what it was), and
nix/package.nix carries the hash from a failed `nix build .#hex`, then a full
one whose check phase passed.

`hex model install fluid-speech-pro` followed by `hex model transcribe` on
whisper.cpp's jfk.wav prints the expected sentence, as does Parakeet v3 on
the same clip.

* fluid speech pro: snake_case id, FLEURS-only figures, blurb, notices
2026-11-19 15:19:06 -05:00
Kit Langton 998131ef82 fix(modes): match a picked app under its Finder name too (#2001)
The app picker saves the name Finder shows for the bundle's file, but the frontmost read reports the running app's own name, so a picked app whose file is named differently (Rex Beta.app runs as Rex, Visual Studio Code.app as Code) never activated its Mode and Labs › YOLO mode never pressed Return in it. The macOS frontmost read now also keeps Finder's name for the bundle when it differs, and application matching accepts either, so existing saved lists start working without re-picking.
2026-11-19 09:34:33 -08:00
Dax 002b575c95 Add KB-Whisper Medium for Swedish (#1991) (#1992) 2026-11-19 05:44:56 -05:00
Dax a9f992004f linux: round a Wayland window's bottom corners (#1981)
* linux: round a Wayland window's bottom corners

Bump Rune to 431469b999268a56ef7fbc60ed55975753ee22df. Under GNOME Wayland
winit draws the client-side frame, whose headerbar rounds the top corners and
whose shadow cuts a 10-point rounded window, but the content's surface stayed
square: the bottom corners showed past the frame's border and the border's
straight bands crossed the shadow's cut-out (#1779).

Rune (#1188) rounds the content's bottom corners with the frame while the
window floats, and keeps them square when it is maximized, tiled or
fullscreen; a compositor that draws the decorations itself (KDE, sway,
Hyprland) keeps the surface it had. Patch sctk-adwaita from the same rev as
winit, since Rune's winit calls the vendored frame's
`AdwaitaFrame::content_corner_radius`.

The rev also carries Rune's macOS/iOS/Linux build fix after the crate
extractions (#1203), and its wheel fix (#1214): a wheel over a scroll bar now
scrolls the viewport it follows, so Hex's wheel-over-the-right-edge tests keep
their positions.

The bump brings the type-erased `window::Host` and window/ reorganisation,
the rune-core/rune-gpu/rune-image/rune-spell extraction, then rune-text,
rune-media and rune-edit, the non-generic `RuntimeCore`, windows-sys 0.61,
deep Taffy stack segments, one flatten and one detached-table prune, the new
wheel distance and chaining, and text, markdown and scroll fixes. No Hex call
sites changed.

Goldens are refreshed for Rune's visuals: the OpenCode icon catalog's marks
(terminal, search, keyboard, globe) in every pane and overlay, and the
shortcut-mode bars from Rune's border-over-fill change.

cargo check --all-targets, cargo test (2811 passed), cargo clippy
--all-targets --all-features, cargo fmt --check, code-shape.py,
check-nix-git-pins.py and its tests, nix build of the cargoDeps, and
cargo test rune_ui all pass.

* history: read Dictation stats from the window's settings

The pane closed its stats dialog and reloaded the rollup from the
process-wide labs switch. A test that applies another LabsSettings (as
several do) flips that mirror under it, so the dialog vanished mid-test.
Take the window's own `settings.labs.history_stats` instead, as the Stats
button and the dialog already do.

* nix: vendor Rune's fonts for rune-text, and run the suite serially

Rune's crates/rune-text includes crates/rune/assets/Inter-*.ttf by a path
relative to the workspace root; a vendored crate sits in git/ beside
rune-0.0.1, so link git/rune to it for the include.

The package check also runs plain cargo test, which shares Hex's
process-global lab switches and model-download state. nextest isolates
each test everywhere else, so run this one test at a time.

* goldens: refresh the meeting capsules on Rune's icons

Dev's capsule layout moved the time and close button above the title; with
Rune's icon catalog the copy-agent-prompt mark is the boxed terminal. Only
those four renders change.

* nix: run the package check one test at a time the hook's way

`preCheck` runs before cargoCheckHook, which then sets
RUST_TEST_THREADS=$NIX_BUILD_CORES, so the export there was a no-op and the
suite still raced on its process-global state. Set
dontUseCargoParallelTests instead, and take the rune vendor directory's name
from its cargoGitDependencies key, failing loudly if its assets are missing.
Drop the link once Rune moves the fonts into rune-text.
2026-11-19 05:18:55 -05:00
Dax 538d16f668 windows: sign the setup engine so smart app control can run setup (#1987)
* windows: sign the setup engine so smart app control can run setup

Setup extracts a copy of its engine to %TEMP% at launch and writes the
installed uninstaller from the same bytes. Signing only the outer
hex-setup-*.exe left both unsigned, and Windows 11 Smart App Control
blocked the extracted copy ("Unable to execute file in the temporary
directory", error 4551).

SignedUninstaller embeds the engine's signature, but the signature has to
exist before the compile that uses it, and the compile job holds no
credentials. So the release signs it in two passes: the new
windows-setup-engine.yml compiles once with SignedUninstaller=yes and keeps
the unsigned uninst-*.e32 ISCC asks to be signed (ISCC leaves it and aborts
as a compile failure, exit 2); sign-windows-engine signs and verifies it in
the release environment; windows-installers.yml then compiles with the
signed engine and asserts the installed uninstaller's signature.

Test builds and runs without the Azure values keep the unsigned single
compile. Inno Setup is pinned to 6.7.1, read from the compiler's own banner
(ISCC.exe's version resource is 0.0.0), so both passes use one compiler.

* windows: harden the setup engine passes

Review fixes for the signed setup engine:

- sign-windows-engine requires sign-windows to have succeeded, and its
  steps are unguarded like sign-windows-installers, so it can never upload
  an empty digest.
- setup-engine runs only when Windows is signed (it waits for
  sign-windows), so unsigned runs really do a single compile with no
  engine pass at all.
- the Inno re-pin uses choco upgrade --allow-downgrade, because choco
  install --version treats a preinstalled package as already installed.
- Invoke-Iscc collects native stderr safely under Windows PowerShell 5.1
  and pwsh with $PSNativeCommandUseErrorActionPreference on, and echoes
  the compiler's lines so build logs stay readable.
- the version probe compiles a minimal script from standard input into a
  throwaway OutputDir, so it never touches the caller's -OutDir.
- only .e32 exists for the engine; drop the unused .e64 handling.
- the uninstaller signature test also checks the release publisher, and
  the test-builds doc warns that Smart App Control cannot be turned back
  on without resetting Windows.

* windows: correct the setup engine comments
2026-11-19 02:38:03 -05:00
Dax 8d3ea95f91 show why speech engine fell back to the cpu after a gpu out-of-memory failure (#1984) (#1985) 2026-11-19 01:01:37 -05:00
Dax e4d85b44aa windows: join back-to-back dictations with a space (#1980)
* windows: join back-to-back dictations with a space

move the input-activity revision, continuation join and its tests into a
platform-neutral continuation module. windows now invalidates it on real
keys and clicks from the low-level hooks and on foreground-window changes,
and the paster joins dictation and paste last the way macos does.

* windows: end continuations on real input and skip hex's own trigger

injected input from other tools and the auxiliary mouse buttons end a paste
continuation as on macOS, while Hex's own synthetic events and a dictation
trigger that Labs › Share shortcut lets through do not. A standalone macOS
paste clears the continuation only where dev did: at the start of typing, or
after the clipboard is preserved and checked.

* windows: a bare trigger key press ends the paste continuation

Labs › Share shortcut lets the dictation trigger's key reach the app in
front, and the hook treated every event for that key as hex's own. A
plain Space while Ctrl+Space is bound then kept the continuation, so the
next dictation joined with a space for no reason. Mark the press as the
shortcut's only once it claims the capture, and drive the hook's own
state machine in a test.
2026-11-18 19:28:07 -05:00
Dax 84ed9a52c8 labs: fold yearly stats into dictation stats (#1976)
* labs: fold yearly stats into dictation stats

* labs: strip retired yearly_stats key and keep it under a lock
2026-11-18 12:22:38 -05:00
Dax 0ec0438c22 history: yearly stats heatmap, time saved and busiest hour (#1971)
* history: add yearly stats rollup, time saved, and busiest hour

* history: fix clearing and seeding the usage rollup
2026-11-18 07:08:05 -05:00
Dax 717932b905 fix: label an unset OpenCode rewrite model as automatic and log the model that answered (#1961)
* fix: label an unset OpenCode rewrite model as automatic and log the model that answered (#1955)

* fix: initialize the processing model on the macOS cleanup pass

* fix: gate the unranked model fixture to the Linux tests

* fix: don't log a failed automatic rewrite as opencode's default

* docs: describe where the rewrite model is recorded
2026-11-18 02:40:37 -05:00
Dax 10d103ff8c labs: bring the Modes editor back behind a lab (#1963)
* labs: bring the Modes editor back behind a lab

The read-only Output pane replaced the per-app Modes editor, so there was
no UI left to turn a mode's OpenCode rewrite on or off or pick its model
(#1960).

Labs › Modes editor (off by default, every OS) shows the old editor as a
main pane right after Output, adds it to the Command-number order, and
makes the Modes destination (the menu bar and `hex window show --pane
modes`) open it instead of Output. Off, the sidebar, `--pane modes`, and
every scene render exactly as before; turning it off while the editor
shows leaves that pane for Output.

The card, its translations, site/src/labs.json, the `hex settings` key,
and the Labs docs come with it.

* labs: flesh out the Modes editor lab's copy and docs

Say what the Modes editor actually edits (apps and websites, corrections,
OpenCode rewrite on or off and its model, transformations) in the CLI and
`hex settings` descriptions, keep the card line and its translations as
they are, note Modes in the sidebar and hex://settings/<pane> docs, align
the Labs table row, and exercise the off-while-Modes fallback by pulling
it out of poll into one tested helper.
2026-11-18 02:40:22 -05:00
Dax ebdd971ce1 fix(dictation): release the warm mic while the display sleeps or the screen is locked (#1956)
* fix(dictation): release the warm mic while the display sleeps or the screen is locked (#1953)

* fix(dictation): address review on warm mic away release

* docs: fix hex microphone resume wording
2026-11-18 02:34:47 -05:00
Dax b5b5a27158 linux: send portal paste chords as evdev keycodes (#1941)
* linux: send portal paste chords as evdev keycodes (#1939)

KWin looks a keysym up in the desktop's active layout group. With a
non-Latin layout (Russian) active it can't find Latin `v` there, and its
fallback to a scratch keymap lost the held Ctrl: the app received a bare
`v` and Ctrl stayed down in the compositor.

NotifyKeyboardKeycode takes Linux evdev keycodes (KEY_LEFTCTRL 29,
KEY_LEFTSHIFT 42, KEY_A 30, KEY_C 46, KEY_V 47, KEY_ENTER 28), which name
the physical key, so the chord is the same in every layout. The session
already asks for the KEYBOARD device type, which keycode notifications
require, so only the method and its numbers change.

Not observed on a real KDE session.

* linux: resolve paste keycodes from the desktop keymap (#1939)

A fixed evdev code is wrong on a non-QWERTY Latin layout: Dvorak puts `v`
on KEY_DOT, so Ctrl+V would arrive as Ctrl+K, and AZERTY puts `a` on
KEY_Q, so select-all would be Ctrl+Q, which quits apps.

Paste now reads the compositor's keymap from a one-shot wl_seat keyboard
(bounded at 250 ms, the fd read through a File with a size cap, no
unsafe), parses it with xkbcommon, and resolves `a`, `c`, and `v` at level
1 across every layout group. A chord goes out as NotifyKeyboardKeycode
only when every group that has the letter agrees on one evdev code (the
xkb keycode minus 8); disagreeing groups such as `us,us(dvorak)`, an
ambiguous group, a missing letter, or no keymap at all send the whole
chord as keysyms, as before. Ctrl (29), Shift (42), and Enter (28) are
fixed codes, and Return never reads the keymap.

KWin 6.8's fake input answered a Latin `v` under a Russian layout from a
scratch keymap that lost the held Ctrl: the app got a bare `v` and Ctrl
stayed stuck (#1939). Mutter's keysym path drops the `v` the same way.

The keymap is read per chord because a portal session can outlive the
keymap it started under, and a layout or variant change replaces it.

Not observed on a real KDE or GNOME session.

* linux: read the shared keymap safely and check every fixed key (#1939)

Compositors hand every client the same keymap fd, and an fd's offset is
shared, so the sequential read left it at EOF: every later paste's read
came back empty, silently falling back to keysyms, and other clients
reading the same fd saw nothing. Read positionally instead
(`FileExt::read_exact_at`, never moving the offset) and strip the NUL the
desktop ends the keymap with.

Ctrl, Shift, and Return are only where the keymap says they are. Options
move them (`ctrl:swapcaps` puts Caps_Lock on 29, `ctrl:nocaps` leaves
Control_L on two keys), so keycode mode now requires Control_L on 29,
Shift_L on 42 when the chord holds shift, and Return on 28, each in every
layout group; otherwise the whole chord is keysyms. Return without a
keymap is a keysym again, as it was before.

The wait no longer spins: `guard.read()` returns Ok(0) when nothing is
ready, so the old loop retried without ever yielding. It now asks for a
wl_display sync once the keyboard exists and stops when that answers
(the keymap comes before it), stops at once when the seat has no
keyboard, and polls the Wayland fd with a deadline like `linux/clipboard`
does. The registry roundtrip before it keeps the comment saying what it
is: Wayland's unbounded startup roundtrip, as in `linux/paste.rs`.

Tests parse a self-contained inline keymap through xkbcommon (no
xkeyboard-config), read it twice through one fd, and cover the moved
modifier and Return fallbacks. The Cyrillic fixtures use the legacy
keysyms xkbcommon reports, in `ru,us` order.

Not observed on a real KDE or GNOME session.

* docs(linux): correct the keymap comments (#1939)

KEY_CAPSLOCK is 58, where `ctrl:swapcaps` puts Control_L, not a key no
keyboard has. The map now says what each key actually requires: a letter
needs one code across every group that has it, while Ctrl, Shift, and
Enter must be on 29, 42, and 28 in every group.

Mutter loses a keysym outside the active group, not `v` specifically.
The wtype paragraph keeps its original wrapping, and a paste that cannot
read the keymap says so at debug level.
2026-11-17 13:23:38 -05:00
Dax 919f726add sort language menus alphabetically (#1937)
* sort language menus alphabetically

* open language menus on their chosen language

Whisper's list is longer than its menu shows, so sorting it by name can
push the saved language's check below the edge. A select's press now
reveals the chosen row through Rune's scroll request, in the kit's
Select and Menu alike, and the map names the menu's order where it
still described the catalog's.

* docs: note language menus open on the chosen language
2026-11-17 11:31:28 -05:00
Kit Langton 6be1f7ba1e fix(history): join adjacent selected dictations (#1931)
* fix(history): join adjacent selected dictations

* test(history): check selected ranges against Linux renders
2026-11-16 22:09:40 -08:00
Dax 9a79a25475 make the labs card copy plain (#1929)
* make the labs card copy plain

Rewrite the Labs card titles and lines in words a regular user understands:
YOLO mode says the message sends itself, Live text and Failure notices name
the small recording indicator, Vocabulary help says which speech models it
affects, and the key-side cards name Option, Command, Shift, and Control.
Audit every other card, updating the CLI descriptions, all nine translated
languages, the website lab catalog, the feature docs, and the render goldens.

* fix labs copy review findings

Restore the ISO qualifier on the section-key card, spell out that
double-tap and hold-through shortcuts contain no letter or number key,
cover all Global mode stages in the API text card, and name the print
command in the print-start copy. Restore the developer API card titles,
fix the Annotate capture wording, and retranslate the affected strings.
Remove every replaced translation key and stale lab-name reference.

* fix second labs copy review

Update the live typing test and YOLO goldens. Add a Windows variant of the
hold-through-keys copy naming Ctrl, Alt, Shift, and the Windows key. Remove
the last unused rewrite-by-voice translation key, keep Annotate untranslated
in the replay title, and clean the remaining old lab-name comments.

* use the Polish Annotate pane name in its replay title

Polish calls the Annotate pane Adnotacje, so the Annotate last 15 seconds
title says Ostatnie 15 sekund w Adnotacjach again. Other languages keep
Annotate because their pane name is Annotate.
2026-11-16 22:32:02 -05:00
Kit Langton 18e1051d0d feat(history): select several dictations without a Lab (#1924)
* feat(history): select several dictations without a Lab

* Adjust the Labs registry count

* Refresh Linux goldens for the smaller Labs list

* Update Labs counts in Linux goldens

* Refresh the History group counts against the current Labs list
2026-11-16 17:57:43 -08:00
Dax d7104ac925 use 24-hour meeting times outside english (#1927) 2026-11-16 20:16:08 -05:00
Kit Langton 51c9974aa5 feat(settings): find a setting or lab without leaving your pane (#1914)
* Update the Rune window and input dependency

* Find settings and Labs cards from the focused Hex window

* Focus an indexed slider without starting a drag

* Check shortcut capture, modal boundaries, and Annotate availability

* Refresh Labs goldens after graduating settings search

* Remove redundant row conversions after retiring the filter
2026-11-16 16:11:48 -08:00
Kit Langton fedab81d81 feat(voice filter): preview effects without changing your call (#1906)
* feat(voice filter): preview effects without changing the call (t1598)

* test(voice filter): check pass-through after the preset fade (t1598)

* test(voice filter): refresh the setup and preview goldens (t1598)

* chore(voice filter): finish the preview checks (t1598)
2026-11-16 15:37:31 -08:00
Dax 7dcd6dbf5d add azerbaijani to transcription languages (#1921) 2026-11-16 16:45:02 -05:00
Dax f070dcffc7 fix lower audio on outputs with partial channel volume controls (#1916)
* fix(audio): lower outputs that lack volume on every channel

macOS ducking needed all of an output's stream channels to have a settable
volume scalar, so a USB adapter reporting four channels with controls on only
the first two was left unchanged, silently. Ramp only the channels that can be
set, then fall back to the main volume and the virtual main volume Mute uses,
and warn when the output has no controllable volume at all.

fixes #1915

* fix(audio): keep a read-only channel's balance when lowering

A channel that can be read but not set used to be dropped from the ramp, which
left a settable channel beside it lopsided, and a plain main volume write could
flatten a balance that cannot be restored. Such a channel now sends the whole
output to the virtual main volume, which macOS keeps balanced; the main volume
is a fallback only when no channel reports a volume at all. The warning for an
output with nothing to lower now says which failure it saw and includes the
device id when there is one.

* fix(audio): keep an unreadable volume from making the ramp lopsided

AudioObjectHasProperty now decides whether a channel has a volume scalar at
all; a property that exists but cannot be read, or returns an invalid value,
counts as read-only, so lowering still goes to the virtual main volume instead
of skipping the channel and ramping only the ones beside it. The warning for an
output with nothing to lower now also covers an unreadable stream layout.
2026-11-16 15:12:32 -05:00
Dax 72ac81ad61 fix(meetings): finish recordings across quits and update restarts (#1900)
* fix(meetings): recover recordings lost to a quit or update restart

Rune ends a quitting app with `_exit` unless it is told the app has work
after its loop, so every in-app quit (Quit Hex, `hex quit`, Restart to
update) skipped Hex's drain since the Nov 14 Rune bump. A meeting
recording was killed mid-write: zeroed WAV headers, no `final.pending`,
and nothing to recover at the next launch.

- `rune::orderly_exit()` in the macOS, Windows, and Linux app loops, so
  the existing teardown runs on every quit; the AppKit terminate path
  still drains from its will-terminate notification.
- macOS joins the meeting controller before installing a staged update
  and before the other workers, and a recording gets 30 s to finalise
  rather than 5.
- `meeting::recording()` is a process-wide flag the updater reads.
  Restart to update refuses with a reason, re-checks before swapping,
  and keeps the release staged.
- Launch recovery selects every unlocked unfinished meeting, not just
  `final.pending`; repairs zeroed or wrong WAV RIFF/data sizes from the
  file length, fills sample counts and duration from the audio, and runs
  the existing final pass so the transcript, Markdown export, and notes
  are written. Failed and completed meetings are untouched.

Checks: `a_zeroed_wav_header_is_repaired_from_its_file_length`,
`a_stranded_recording_is_selected_for_recovery`,
`recovery_fills_the_manifest_from_the_repaired_audio`,
`the_flag_holds_until_the_recorder_drops_it`,
`restart_to_update_waits_while_a_meeting_records`.

* fix(meetings): settle recovery attempts and keep retention honest

Reviews of #1900:
- a recovery that dies during the final pass leaves recovery.attempt;
  the next launch marks the meeting Interrupted ("recovery did not
  finish") and never retries that audio. final.pending folders that
  already have a transcript are finished too.
- a repaired, unfinalised track overwrites list_all's relabel with the
  audio's real duration and ended_at.
- retention only waits for stranded meetings while recovery can run
  (Meetings on, model installed); the MCP query uses the same rule.
- repair only rewrites a zero or overlong data chunk, so a trailing
  chunk or pad byte is never read as audio.
- the macOS status item dims Restart to update while a meeting records.
- finish_restart keeps Ready only when an update is staged.
- macOS installs a staged update before joining the controller when
  nothing records; Windows arms a 20 s teardown deadline; Linux writes
  its reopen marker only after a restart is accepted.
- the recording flag is a counter, tested without assuming idleness.

* fix(meetings): bound recovery retries and harden quits

- recovery.attempt now counts attempts: a pass that dies is retried once,
  then the meeting is marked Interrupted ("recovery did not finish") and
  that audio is never handed to Hex again. The pass holds the meeting
  busy flag, so the quit deadline and Restart to update wait for it.
- a finished meeting with a stale attempt marker keeps its final.pending,
  which the ordinary final pass then exports; settling no longer deletes
  it.
- pending recoveries stop between meetings when a recording starts and
  are offered again once Hex is idle.
- an empty finalised track is not "unfinalised"; the audio's duration
  wins whenever repair changed the manifest's track counts.
- can_recover reads meetings.enabled and the model from settings.json in
  readers (`hex mcp`, `hex meetings`), matching the app.
- updater: State::Applied is never overwritten by a later check, and a
  restart check that races a quit never swaps a second time.
- macOS installs the staged bundle before joining the meeting
  controller, marks quitting before any drain, and extends the 5 s exit
  deadline to 30 s from the quit's start while meeting audio is in
  flight. `meeting::record` refuses once quitting.
- Windows marks quitting after the loop, bounds teardown at 20 s (30 s
  while meeting audio is in flight) and exits 1 when the window failed;
  Linux marks quitting before joining meetings and removes a reopen
  marker when a restart is refused.

* fix(updater): keep the staged update installable when a late download is dropped

- check() now returns the state the locks actually show when it discards a
  release during a quit: Applied only if one was installed, else Ready while
  an older update is staged, else Current. install_on_quit therefore still
  installs the staged update a plain quit is about to apply.
- after waiting for meeting audio, the macOS exit deadline gives the rest of
  the drain up to 5 s more, inside the 30 s bound from the quit's start.
  Windows polls meeting audio to 30 s instead of choosing a bound once.
- the Restart to update refusal reads "Hex is finishing a meeting; restart to
  update when it's done", true for a recording and a launch recovery.
- a recording refused because Hex is quitting removes the empty folder it
  created (and only when it holds nothing but the lock file).
- docs: the macOS drain installs before joining the controller, and the
  deadline/discarded-release wording matches the code.
2026-11-16 12:30:57 -05:00
Kit Langton 2e8e3f65b0 fix(hud): keep the dictation and meeting overlays dark (#1834)
* Trace a recording pill hover and reversal offscreen (t1571)

* Keep the recording meter anchored and show each audio track (t1571)

* Keep light and dark recording snapshots from Linux (t1571)

* Draw dictation’s listening field inside the meeting pill (t1571)

* Refresh meeting pill goldens with the listening field (t1571)

* Paint just the listening dots on the meeting surface (t1571)

* Keep dots-only recording goldens in light and dark (t1571)

* Keep recording controls and the dictation HUD dark on light desktops (t1571)

* Refresh the dark overlays on light-host goldens (t1571)

* Keep the finished meeting in its corner and let Close dismiss it (t1571)

* Refresh finished summary goldens with the close control (t1571)

* Open shipped features from What’s new (t1574)

* Keep release-specific links distinct from feature navigation (t1574)

* Document and translate the feature feed (t1574)

* Verify dark and light navigation to both controls (t1574)

* Show feature icons and calendar-day headings in What’s new (t1574)

* Verify local-day boundaries and refresh both feed goldens (t1574)

* Separate What's new navigation from build details (t1574)

* Check What’s new selection through accessibility (t1574)

* Refresh only the lower sidebar in the Linux goldens (t1574)

* Refresh the pill, Sounds, and What's new goldens after the stack (t1571)

* Hold the recording clock still in the dark-host test (t1571)
2026-11-16 00:22:48 -08:00
Kit Langton 5e272d4d2e feat(settings): open new features from the sidebar (#1866)
* Open shipped features from What’s new (t1574)

* Keep release-specific links distinct from feature navigation (t1574)

* Document and translate the feature feed (t1574)

* Verify dark and light navigation to both controls (t1574)

* Show feature icons and calendar-day headings in What’s new (t1574)

* Verify local-day boundaries and refresh both feed goldens (t1574)

* Separate What's new navigation from build details (t1574)

* Check What’s new selection through accessibility (t1574)

* Refresh only the lower sidebar in the Linux goldens (t1574)

* Refresh the What's new goldens with the newer labs (t1574)
2026-11-16 00:11:41 -08:00
Kit Langton db06244662 Play requested Sounds previews while Hex volume is Off (t1581) (#1880) 2026-11-15 22:29:41 -08:00
Dax f1afb3c06a fix: recognize ardia and zen as browsers (#1875)
* fix: recognize ardia and zen for pause browser media

Ardia (com.ardia.browser, audio in com.ardia.browser.helper) and Zen
(app.zen-browser.zen, helpers under app.zen-browser) were missing from the
browsers microphone activity knows, so Also pause the browser never heard
them and left them alone.

Add both to the browser list, Ardia's helper and Zen's whole app.zen-browser
prefix to the canonical mapping (Zen's helpers are app.zen-browser.*, never
org.mozilla.*), and their display names. The prefix keeps the existing dot
boundary, with negatives for app.zen-browserx and com.ardia.browserx.

* fix: recognize ardia and zen in call detection

The browser list microphone activity knows is shared with meeting
detection, so Ardia and Zen still read as unknown microphone apps there:
no browser-call offer, no Meet name from their titles, and Zen's
plugin-container never folded to its browser.

Add both to call detection's browser names, matching Zen's whole name
like Arc so Citizen and Zendesk stay out, strip their Chromium and
Firefox brand suffixes (Zen Browser and Zen Twilight on Windows and
Linux, no brand suffix on macOS), count Zen's plugin-container as a
helper, and cover each browser with a detection test plus Zen's
substring negative. Zen's Windows stem was already listed; Ardia has no
Windows build.

Docs: meetings' call detection and helper-process notes learn both.
2026-11-15 23:15:47 -05:00
Dax 9ea343f69d fix(updater): keep one .Hex-update folder, not one per staged dev build (#1867)
* fix(updater): keep one .Hex-update folder, not one per staged dev build

Every staged release unpacked its own `.Hex-update-<version>` folder beside
`Hex.app`, and the only removal was `finish_on_launch`. The Dev channel checks
every ten minutes, so a Hex that runs for days staged a full ~80 MB copy of the
app per published build and left every one of them in `/Applications`
(the folder it replaces was never removed); the launch sweep only ran when Hex
next started, which for a menu-bar app can be days.

Staging a newer release now removes the folder of the one it replaces, once
the replacement is the staged update, so a running install holds exactly one
folder and the app can never be caught without its staged update. The launch
sweep stays as the safety net for a swap or an interrupted stage, and now
skips symlinks (never followed) and the running bundle itself.

The sweep moved to `updater/staging.rs` so it can be tested off macOS:
`launch_cleanup_removes_every_unpacked_update_but_the_running_bundle`,
`a_newer_update_removes_the_folder_of_the_one_it_replaces`,
`a_bundle_running_from_a_staging_folder_keeps_that_folder`, and
`a_symlink_named_like_a_staged_update_is_never_followed`, plus the macOS-only
`a_newer_staged_release_removes_the_folder_of_the_one_it_replaces`.

Observed on macOS 26 with the published 0.0.0-dev-1046 build: the launch sweep
removed both a folder holding an older signed build and one holding a
swapped-in peer beside /Applications/Hex.app.

* fix(updater): never stage a release after a quit installed one

a quit installs the staged update itself (`install_on_quit`) while the updater
thread can still be in `check()`. two things went wrong in that window:

- the late `check` overwrote `shared.staged`, and its `remove_superseded` then
  deleted the folder holding the bundle the quit had just replaced (the one
  the draining process runs from), orphaning the release it had unpacked.
- `finish_restart` swapped again after `install_on_quit` had applied the
  update, exchanging the same two bundles back and losing the update.

`check` now takes the state and staged locks together, in the order
`install_on_quit` takes them, and refuses a release that lands while quitting or
after `State::Applied`: it removes the folder it unpacked and keeps the
installed update. `finish_restart` reports done without swapping once the state
is Applied. The macOS `discard` helper removes that unused folder.

`staging::remove_beside` now canonicalizes the bundle's folder and compares the
kept folder by name (only when its own folder resolves there), so a `..`, `./`,
or symlinked spelling still names the folder to keep, and a path outside it
protects nothing.

tests: `a_restart_after_a_quit_installed_the_update_does_not_swap_again` (fails
without the Applied guard), `a_release_that_finishes_staging_after_a_quit_installed_one_is_discarded`
(fails when the decision adopts everything), and
`the_folder_to_keep_is_recognized_through_a_non_canonical_spelling` (fails with
the old raw path comparison).
2026-11-15 18:20:36 -05:00
Dax 58f5cd5202 add microphone key alone lab (#1864)
* add microphone key alone lab (#181)

* gate every microphone key shortcut on its lab (#181)

* reset mic-key cancel and screenshot shortcuts too (#181)
2026-11-15 15:10:30 -05:00
Dax c4e361b214 labs: copy a dictation's kept audio as a file (#1860)
* labs: copy a kept dictation recording as a file

Adds the Copy dictation audio lab (Labs › History, off by default,
labs.copy_dictation_audio). It stores no audio: History entries that
Labs › Keep recordings gave a WAV gain Copy audio beside Play, and the
macOS menu bar and the Linux and Windows tray gain Copy last dictation
audio, which copies the newest entry that kept its audio. macOS uses
Scrawl's file pasteboard, Windows CF_HDROP, and Linux a text/uri-list
through the process's one LinuxPaster, which now lives in
linux::paste::with_paster so captures and History share it.

A missing recording or a refused clipboard says so under the row in
History, as a failed Play does; the menus log it, as Copy recent
dictations does. The newest-recording lookup skips entries whose file
is gone, with a unit test. Docs, translations, labs.json, and the Labs
goldens follow.

* labs: copy dictation audio off the UI thread and behind both switches

Review fixes for the Copy dictation audio lab:

- History's Copy audio now runs like Transcribe again: the file lookup on
  the window, then a worker named hex-copy-audio does the clipboard write
  (which can wait on another app, Annotate's paster, X11 or wl-copy) and
  reports Copied or the error back through the window's waker. The button
  reads Copying… while it runs.
- HistoryStore::newest_recording now takes now_ms and filters by the
  store's retention cutoff, like recent() and search(); the unit test
  covers a file deleted behind the store, the cutoff boundary, and an
  expired entry.
- The menu bar and tray item, and the History action, need both this lab
  and Keep recordings; the gate and every handler re-check both. A copy
  failure from the menu or tray stays logged, as the menu's other copy
  actions do, and the docs record that limit.
- Copy's text check and Copy audio's check clear each other.
- A UI test pins Copy audio absent with either switch off and present
  with both on.
- Card and site summary say menu bar or tray; the validateMenuItem doc
  names both dimmed rows; WindowsDesktopHost::history follows new(); the
  History feature map mentions the action.

* labs: hold text Copy behind a running Copy audio, and refresh the tray

Follow-ups from the PR review:

- A pending Copy audio keeps a text Copy from starting, since its worker
  still lands the WAV on the clipboard; text Copy only clears a settled
  audio state. The UI test pins it.
- The Linux and Windows trays compare Copy audio's availability with the
  last value, so a recording that appears or expires redraws the menu
  without a snapshot or publish change; the gate is read first and
  History only with a tray, keeping the check cheap.
- Move Copy audio's worker state into history/audio.rs so
  rune_ui/history.rs stays under the 1,500-line cap.
2026-11-15 12:08:47 -05:00
Dax 82ca73f945 reload personal config when imported .js files change (#1856)
* reload personal config when imported .js files change

* watch and load the same config file kinds

the watcher skipped foo.d.js, foo.d.mjs, and foo.d.json because it treated
every .d stem as a declaration file, while the loader ran any file imported
by its exact name as javascript even though edits to it never reloaded.

both now use one predicate: ts, js, mjs, mts, and json in any case, excluding
only .d.ts and .d.mts declarations. the loader refuses other extensions when
a config imports them.
2026-11-15 10:46:47 -05:00
Dax 369a303bfc add khmer to the whisper language list (#1852)
* add khmer to the whisper language list

* restore km in whisper's code list, tighten khmer docs
2026-11-15 08:15:08 -05:00
Dax 64234cc2e5 windows: keep opening the microphone when its default format can't be read (#1849)
* windows: keep opening the microphone when its default format can't be read (#1845)

A device that reports no default format was abandoned before the direct
WASAPI paths and the formats it lists were tried, which is the one failure
the Windows default-input handle hits and a named device does not: cpal
activates it through ActivateAudioInterfaceAsync rather than the endpoint.

Try those paths instead of returning, build the alternative formats from
the ranges alone (48 kHz, else 44.1 kHz, else the range's top) when there
is no default to match, and keep the read error, with the driver's reason,
for the message. The open thread, startup, and saved-microphone retries
render causes with {error:#}, and a released microphone that will not open
is logged where it used to fail silently.

* windows: name the ordered fallbacks for a microphone with no default format (#1845)

Review fixes: the Windows tree row lists Hex's own WASAPI paths before the
formats the device lists, matching the code; the module doc says a device
with no readable default format still tries them; and listing a device's
formats logs its cause like every other path.
2026-11-14 23:49:07 -05:00
Dax defa21337f Add Labs › Upcoming meetings in menu bar (#1846)
* Add Labs › Upcoming meetings in menu bar

The macOS status menu lists the next calendar meetings while the lab is on:
up to five meetings from the same filtered EventKit read the meeting
reminders already make, each its local start-end time and title with "Now · "
while in progress, and a Join row for a recognized call link that opens it
and never starts a recording. A meeting without a link is informational.

The reminder poll publishes its meeting list into a bounded snapshot
(calendar_menu) and the status item only formats it on the main thread, on
the same 500 ms tick as the meeting state; with the lab off the menu and the
controller are unchanged. Off by default, macOS only, in the Meetings Labs
group, with the labs catalog card, gate, hex settings key, translations, and
site/src/labs.json.

* Share the menu's local clock with the meeting query time

* Return objc2 Bool from the status menu validator's early exit

* Fix the calendar menu's read gate, titling side effect, and copy

The menu asked the controller to read EventKit whenever the lab was on,
even with no calendar connected, and a menu-only read overwrote the
titling/reminder list, so detected calls were named from the calendar
with reminders off. The read now needs the lab and a connected calendar,
and only feeds the titling list while reminders are active; the menu
snapshot owns the menu-only read. Turning reminders back on forces one
fresh read instead of waiting out the minute.

Menu click now goes through the reminder's ControllerCommand::JoinMeeting
so it opens the link and marks the occurrence answered, and the rows carry
the meeting rather than only its URL. Rows starting on a later local day
read "Tomorrow" or their short weekday, from iso_local's date part. The
Upcoming strings move to the menu translation table, with Now · {},
Tomorrow, and short weekdays translated.

Clearing an empty snapshot takes no write lock, and refresh_upcoming no
longer holds the links RefCell borrow across AppKit calls. The summary
copy no longer claims a Join item or that no recording starts at all.

* Make the menu's join suppress that occurrence's reminder

The menu reused the reminder capsule's Join command, whose answered()
only clears the reminder on screen. Joining from the menu before the lead
time still let the reminder pop, and joining while it showed forgot it
without withdrawing the capsule.

Add ControllerCommand::JoinFromMenu and ReminderScheduler::joined_elsewhere:
it records the occurrence as reminded so it never pops, and returns the
occurrence when its reminder is showing so the controller sends
ReminderWithdrawn. The reminder capsule's Join path is unchanged. Menus
open the link and still never record.

Also inject the date in the row tests so they pass in any time zone, force
the menu's first read when the lab comes on, read the lab atomic first in
section(), and name the watcher tests for what they cover.
2026-11-14 23:41:22 -05:00
Dax 687ea32d65 Bump Rune to 1899e64: overlays pace frames with the swap chain on Windows (#1832)
* Bump Rune to 1899e64: overlays pace frames with the swap chain on Windows

Windows HUD memory grew and the app could hang after dictation because the
overlay presented frames without waiting for swap chain room after Rune #844;
Rune #985 makes an overlay's frames wait for room in its swap chain, as a
window's do.

nix: vendor each Cargo source in its own directory. Rune's tiny-skia-path is a
path dependency inside Rune (a Git crate at 0.11.4 here), the same name and
version as resvg's from crates.io, and cargo's directory source keys a crate by
name and version alone.

Goldens: Rune's text drawing changed over the 152 commits since d6715352.

* Review: patch accesskit_macos, and make the nix vendor split name what breaks

Rune #975 put accesskit_macos in its [patch.crates-io], so Hex patches it too
(Rune's README asks consumers to); it is the fifth and last crate Rune patches.

The vendor script override now checks each fragment it rewrites and names the
missing one, its comment says crates.io goes to registry/ and Git sources to
git/, nixfmt formats the file, and check-nix-git-pins.py's docstring says the
hashes reach importCargoLock through the cargoDeps override.

* linux: keep Hex running when the window can't start

Rune r399 makes App::run print a startup error and exit 1 on native targets
instead of panicking, so a Linux Hex with no display died, the next launch
became a fresh instance, and Check Linux's single-instance exercise saw
`hex start` hang and `hex app --hidden` reset.

run_app now returns Rune's try_run error and run_window reports it and waits,
as it did for the panic (#979).

* macos, windows: take rune's startup error instead of exiting

Rune r399 prints a startup error and exits the process from `run` on
native targets, which skipped Hex's own stop: macOS's drain (workers,
event-log flush, a recording kept for the next launch) and Windows' GPU
draw-marker disarm, so the launch that failed before drawing was read as
a crash while drawing. Both apps take `try_run`'s error now, log it, and
stop the way Quit does; Windows also reports it from `hex app`. The
Linux window already took the error (#979).

* macos: a launch whose window never started exits 1

Taking Rune's try_run error lets the app stop through the drain, but
process_exit::finish ends with _exit(0), so a failed GUI launch reported
success. finish_with takes the code; the window's failure exits 1 while a
normal quit keeps 0.

* docs: overlays wait for swap-chain room on windows since rune #985

Rune #985 makes every overlay's frames wait for room in its swap chain, so
the notes that said only a window waits and that the HUD's 60 Hz timer worked
around a suspected cause were stale. The timer is a cap on top of the slot
wait now. The macOS and Windows startup-error comments cited #979, which is
the Linux window issue; they describe Rune r399 instead.

* windows: a failed start keeps the GPU draw marker

disarm() means a frame reached the screen or Hex quit cleanly; a start that
never drew is neither, so it keeps the marker and the next launch falls back to
the high-performance GPU.

* docs: say what was measured about overlay pacing

* docs: say what stays bounded on windows, not what shrinks

The meetings map and the dictation map described the fix as memory that no
longer grows, and as Rune holding a frame's slot and skipping it. Rune #985
defers the frame until its swap chain has room, then draws it, which bounds
the frames in flight; the growth itself hasn't been re-measured on Windows
with the fix.
2026-11-14 20:40:20 -05:00
Dax 86d033b8f5 clear raw and paste-last shortcuts that share the dictation shortcut (#1830) (#1831)
dictate without processing and paste last dictation are not filtered at
runtime the way the lab shortcuts' active_* accessors are, so a saved one on
the dictation chord started both. the base-aware save merge could also land
the dictation shortcut on a saved one, and a load-time repair could reset
dictation onto one.

clear those two on load and on the merged settings before a save. the load
clear is a repair, so the file keeps its bytes until the next save.
2026-11-14 14:15:57 -05:00
Dax a94583e8d9 fix(onboarding): give the launch-at-login ask the platform's words (#1819)
Windows and Linux asked "Open Hex when your Mac starts?" and had Hex
napping "in your menu bar". The ask's copy now follows `Home`: macOS
keeps its Mac and menu bar, Windows says computer and system tray, Linux
says computer and in the background. The new strings are translated in
every language, and a headless test checks all three platforms.
2026-11-14 00:36:16 -05:00
Dax 94eb637eef labs: clean up dictations with apple intelligence (#1814)
* labs: clean up dictations with apple intelligence (#1810)

Adds Labs › Clean up with Apple Intelligence, off by default and macOS only:
one short on-device Foundation Models pass over each dictation whose mode does
not already rewrite with OpenCode. The instructions and the guard are the iOS
cleaner's, so both apps accept and refuse the same answers. An unavailable
model, timeout, refusal, guardrail block, or an answer the guard refuses all
paste the corrected transcript; the stage's own observation records the
reason in History and the live log. Raw dictations never reach it.

* labs: address cleanup review

- read the lab switch (and the caller's cancellation flag) before Apple
  Intelligence's availability, so a lab-off or cancelled dictation never
  touches Foundation Models.
- keep the managed row the Labs registry builds, so an organization's lock
  still disables the card; a test forces the lock on its own thread.
- count CJK numerals as digits, matching Character.isNumber.
- keep dictated quotes and the transcript's own spacing when tidying an
  answer.
- docs: the real worker name, the local API's processed route, the 2 s
  budget, and the guard's scalar-versus-grapheme counting.

* labs: keep the cleanup lock test to linux

The panes tests render on Linux only, so the forced-policy seam has no
users on macOS and failed the macOS check's -D warnings.

* labs: regenerate the mac labs goldens with the cleanup row

* labs: keep the cleanup card inside the labs file cap
2026-11-13 23:51:47 -05:00
Dax 9d040bacb1 feat(labs): let the rewrite read text around the cursor (macOS) (#1815)
* feat(labs): let the rewrite read text around the cursor (macOS)

* fix(labs): bound returned cursor text, escape tags, re-check lab at finish
2026-11-13 23:36:24 -05:00
Kit Langton 4dda1cb6ca feat(labs): YOLO mode presses Return after long dictations (#1812)
* Add Labs › YOLO mode: press Return after long dictations (t1558)

* Add a Labs YOLO mode scene (t1558)

* Update Labs goldens for YOLO mode (t1558)

* Box the Windows dictation output job (t1558)
2026-11-13 20:02:34 -08:00
Kit Langton 214de68edc feat(settings): quiet Reset on shortcut rows, ⌫ turns a shortcut off (#1809)
* Shortcut rows: a quiet Reset, ⌫ turns a shortcut off, and Restore shortcut default graduates (t1548)

* Goldens: no Clear icon on shortcut rows, one fewer lab (t1548)

* Gate the icon button to macOS, its only remaining user (t1548)
2026-11-13 17:09:24 -08:00
Dax 8d62e7ec22 fix(linux): pass pw-cat --raw only where pipewire knows it (#1803)
* fix(meetings): pass pw-cat's --raw only where PipeWire knows it

`--raw` only exists from PipeWire 1.3.81 on. 1.2 and older reject it as an
unknown option and exit before doing anything, so meetings recorded the
microphone alone and the Hex Microphone source never started. Since 1.4
`-` also means a WAV file unless `--raw` is passed, so the flag cannot just
be dropped: probe `pw-cat --help` once per process and pass it only where
the usage lists it, keeping it when the probe cannot run.

Also stop resetting a meeting source's backoff on every successful open: a
source that dies right after each spawn now keeps backing off instead of
respawning and logging twice a second forever.

* fix(linux): test pw-cat's real help and trust only a successful probe

The fixtures are now the exact `pw-cat --help` output of 1.0.5 and 1.4.0
(show_usage with its format args filled in, from a build without
-Dpw-cat-ffmpeg as Arch and Ubuntu build it), not an excerpt, so the
option the probe looks for is checked against the text it really prints.

The probe also asks for the untranslated usage with LC_ALL=C, and keeps
--raw when the help process starts but exits non-success, as its comment
already claimed.
2026-11-13 16:30:28 -05:00
Kit Langton fb6c61b0d9 feat(cli): hex transcribe reads any common media file, with or without the app (#1799)
* Make hex transcribe work with any common media file, with or without the app (t1555)

* Regenerate the Labs goldens without Transcribe files (t1555)
2026-11-13 13:14:15 -08:00
Dax ba8b5698ac windows: keep per-language model choice when switching keyboard language (#1788)
* windows: follow the focused window's layout and ignore hex's own windows for keyboard model switching

* windows: skip a window that closed mid-read; log keyboard source changes at info
2026-11-13 11:18:21 -05:00
Dax b65f9880e0 Labs: smaller model files (K-quant imports, macOS) (#1786)
Adds Labs › Smaller model files (k-quant-models, labs.k_quant_models, off by
default, macOS only). While on, an imported speech model may use Q6_K, Q5_K,
and Q4_K weights beside F32, F16, and Q8_0: Metal has their mul_mat and
get_rows kernels, while the Linux and Windows Vulkan builds embed shaders
only for the catalog's types and would abort, so the lab is not listed there.

The tensor-type decision is passed into the header check
(gguf_header::TensorTypes) instead of read inside it, and the import's
manifest records `k_quants` (serde default false, so older manifests load).
That lets the selection gate a K-quant import exactly as Parakeet Ultra is
gated: its availability folds in the lab, the listener falls back with
FallbackReason::LabsOff and keeps the saved choice, and `hex model use`
refuses it by name.

The header check now also sizes the K-quant tensors from ggml's block sizes
(256 values in 144, 176, and 210 bytes), so a cut-short file is refused and
the GPU-memory estimate counts them, and a found file's compression reads
Q6_K/Q5_K/Q4_K instead of F32.

Off, every OS keeps the F32, F16, and Q8_0 rule and its message byte for
byte.
2026-11-13 10:11:35 -05:00
Dax 4dc3106df6 linux: restore the readable clipboard formats when one can't be read (#1787)
* fix(linux): restore the clipboard formats that can be read

* fix(linux): harden the partial clipboard restore
2026-11-13 10:06:06 -05:00