v2.17.0 — what real use found in a day¶
Released 2026-08-09. Upgrade:
v2.16.0 made the snap a first-class channel. Within a day of people actually using it, three defects surfaced that no amount of unit testing had caught, because each needed a real desktop, a real microphone and a long sentence.
Streaming dictation deleted text it had never typed¶
Reported as: a long dictation ends in a flood of repeated characters, then the correction pass removes text you wanted to keep. Three separate defects, each able to cause it alone.
A fixed timeout on the X11 injector. XdotoolInjector used timeout=10 on every method. At --delay 12 that expires around 833 characters — and it expires after xdotool has already typed part of the text. The caller sees a failure for work that partly happened, LinuxInjector reads that as a broken backend and falls back to the clipboard, and the text lands twice. YdotoolInjector has scaled its timeout since the Wayland flood fix, and its comment says exactly why; the X11 path simply never got the same fix. Timeouts now scale with the keystrokes actually requested.
The correction pass hit the same wall. Committing a streamed burst sends one shift+Left per typed character, so a long dictation crossed the timeout twice — once selecting, once retyping. A run of one repeated key now goes out as a single --repeat spec, the way inject_backspaces always has.
A genuine race. inject_partial runs on the daemon's poll thread while commit runs on hold-release, and the daemon's join(timeout=1.0) cannot outwait an injection allowed ten times that. A partial could land after commit had read the character count, leaving shift+Left × N selecting a span that no longer matched the screen — so the correction ate whatever happened to sit next to it. Every mutation now happens under one lock held across the injection itself, and a committed injector is sealed: a late partial is dropped rather than typed behind the final text.
Verified red/green — the new behavioural tests fail on the previous code and pass on this.
A fresh snap install never mentioned raw-input¶
snap install yazses leaves both audio-record and raw-input unconnected. snapd auto-connects neither, and a snap cannot connect its own. That produces the two most visible failures available — it cannot hear you, and it cannot see the hotkey — and the daemon starts and reports healthy in both cases.
yazses setup / start already surfaced the microphone. It now surfaces raw-input too, so the tool says what is wrong instead of leaving it to documentation nobody re-reads after installing. The Snap row in the install guide lists every required step, including yazses setup — which it told pipx users to run but never snap users.
sudo snap install yazses
sudo snap connect yazses:audio-record # microphone
sudo snap connect yazses:raw-input # hold-to-talk key
yazses setup
yazses start now survives a reboot¶
install_autostart() existed so that a pipx / uv tool / pip install gets a login service like the packaged ones do — its own docstring says "a daemon you must remember to launch is not a daemon." It had exactly one caller: yazses autostart enable, a command you only find if you already know it exists. The ordinary path was install → yazses start → dictate happily → reboot → the daemon is gone, with nothing anywhere saying why.
yazses start now installs the login service once the daemon is actually up, and says so the single time it does. Best-effort by design: a daemon that died during startup is never scheduled to run at every login. yazses start --no-autostart opts out.
Five default filler words were load-bearing content¶
like, right, sort of, kind of and actually shipped in the default [filters.disfluency] filler_words. Each is a hesitation in some positions and meaning in others, and the filter cannot tell them apart — so a hedge became a fact (she is sort of stable → she is stable), a negated assessment lost its predicate (that dose is not right → that dose is not), and a correction marker was erased. YazSes is used for clinical and legal notes, where a confident falsehood is worse than a visible disfluency. All five are removed from the defaults and remain available to anyone who wants aggressive filler removal. Contract 5.1.0 → 6.0.0.
Also¶
- Style-Consistency Enforcer now reads a config-driven rules source (#36)
yazses gitvoice— the Voice Git Choreographer is wired up (#37)- Diagrams-as-Code by voice parses a whole graph, not just one edge
- Tray → Settings… opens the graphical settings window