# v2.40.1 — hold the key we actually listen for **Released 2026-09-25.** Upgrade: ```sh sudo snap refresh yazses # snap pipx upgrade yazses # or: uv tool upgrade yazses scoop update yazses # Windows ``` ## The first thing a new user was told to do did nothing The APT and pipx installers opened with this, before anything else: ``` Hold Space → speak → release → text appears anywhere ``` Holding Space does nothing. `[hotkey] key` defaults to `auto`, which resolves to `platform.default_hotkey` — **`right_alt`** on Linux — and first-run seeding never writes a hotkey, so `auto` is what a fresh install runs with. The modifier was chosen deliberately, so that dictation never collides with normal typing the way the space bar would. This is the first thing anyone does after installing. Following the instruction produces silence, with no error to explain it, and the reasonable conclusion is that YazSes does not work. The same sentence was in the `.deb`'s `Description:` field — what `apt show` and every software centre display *before* you install — in the Linux section of the GitHub release body, and in the "hold this key" tables on the Russian and Hindi home pages, whose Linux rows still read `Space` after the English ones had been corrected. The README and the docs home page were fixed in v2.27.0. That entry says the correction covered "four places". It was a list written by hand, which is why it missed nine more. The list is now a test: `tests/test_hotkey_naming_truth.py` reads each platform's `default_hotkey` out of the source and scans every user-facing file, failing the build if one tells a reader to hold a key no platform defaults to, or pairs an OS with another OS's key. Two of those places only reach a user at a tag, which is why this is a release rather than a documentation commit: the `.deb` description is baked into the package at build time, and the release body is written at publish time. Every `.deb` published before this one still says Space. ## Wayland's default injector no longer drops your accents On Wayland, `[injection] backend = "auto"` picks `ydotool`, and the text went straight to `ydotool type`. That presses keycodes against the active XKB layout, so a character the layout has no keycode for — `å`, `ä`, `ö`, French and German accents, any non-Latin script — pressed nothing at all. Nothing errored: the command exited 0, the daemon reported success, and the only symptom was the wrong words on screen ([#329](https://github.com/MSKazemi/yazses/issues/329)). The layout-independent path already existed as an opt-in `unicode` backend ([#364](https://github.com/MSKazemi/yazses/issues/364)), which resolves a character through `libxkbcommon` and presses it via a private `/dev/uinput` keyboard. The ydotool backend now splits the text into ASCII and non-ASCII runs, keeps sending the ASCII runs through `ydotool type` exactly as before, and routes the rest through that injector — so dictating in your own language works on the default `auto` path with no config change. It needs no new permission: `yazses setup`'s udev rule already grants `/dev/uinput` to the same `input`-group membership `ydotool` itself requires before `auto` will select it. ## Also in this release Everything else here is documentation and packaging truth — pages that described a YazSes different from the one in the tree: - **The Fedora package was four releases behind, again.** `packaging/fedora/yazses.spec` said `Version: 2.36.0` while the project shipped 2.40.0, so COPR built and published 2.36.0 as the current package. `dnf copr enable mskazemi/yazses` still installs 2.36.0 until a rebuild. The spec is right now, and — the part that matters — the release script rewrites it and the packaging-version guard derives which files it checks from the tree instead of a hand-written list, so the next one cannot slip through. - **Install instructions named channels that do not serve what they claimed.** Every channel claim was re-checked against each registry's own API. `yay -S yazses` cannot work — YazSes has never been published to the AUR — winget holds a build seven releases old rather than "a few days" behind, and COPR serves 2.36.0. The pages say so now. - **The support page denied two installers that had shipped for 23 releases.** Intel Mac and Windows-on-ARM users were told their native installer "lands at the next tagged release"; the `.dmg` and the `.exe` have been attached to every release since v2.22.0. - **Every contributor page promised a test suite that runs in seconds.** A full run takes minutes — 15713 tests measured at 4m53s on a recent laptop — not the "about 30 seconds" eleven pages quoted. They now describe the scale instead of a figure that varies by an order of magnitude with the machine. - **The documented contributor setup did not produce the clean mypy run it promised.** A correct fresh `uv sync` reported 12 `import-not-found` errors nobody had introduced, because `PySide6` moved into the `desktop` extra without an entry in the `[tool.mypy]` overrides that silence every other optional backend. - **The Store notes said the Face-Gesture Switch did not exist.** It has shipped since v2.40.0; the reason the MSIX does not declare `webcam` is the frozen bundle carrying no camera extra, which is what the page says now. The Store page also states plainly that YazSes is **not** listed yet, so nobody finishes it believing otherwise. - **The contribute links pointed at finished work.** Newcomer-facing pages recruited against closed issues, two of them saying outright that the work was "open right now". They now point at the directories and pages that take the pull request. No issue was opened, closed or commented on; only the links moved. - **The second paper is being written in the open**, under `design/publication/paper-v2/` — manuscript, claim ledger and the reconstruction of the August measurement campaign, every number resolving to a committed artifact under `paper/results/`. - **The Microsoft Store artwork is drawn by a script** rather than by a snippet pasted into a README, so it stopped being the retired blue logo; and the image drift guards that bind each asset to its generator now measure a tolerance instead of comparing pixels exactly, which was never portable across architectures.