Changelog

Every change, version by version.

Newest first: everything that changed in Freally Teleprompt, in every language this site speaks.

Unreleased

Documentation

  • The website is rebuilt in the Freally family's design, and it speaks all 18 of the app's languages — this changelog included. It offers no downloads.

v1.5.0

Anything in the room can drive the prompter.

Added

  • Foot pedals and presenter remotes. A pedal is how a prompter is actually driven on a shoot — your hands are on a camera, not a keyboard. Open Shortcuts, press the binding you want to change, then press the pedal: whatever it sends is what gets stored, with no driver and no setup. Most pedals and remotes send F13–F24 precisely because nothing else uses them, and those now bind like any other key.
  • Keys that work when Freally Teleprompt is not in front. The operator is usually in a camera app, a DAW or OBS, so every transport command can be given a second binding that the operating system delivers wherever you are.
  • Nothing is claimed until you ask. No system-wide key is registered out of the box — taking one away from every other program uninvited is not something an app should do on first run.
  • A refused key says so. If another program already owns the combination, the row tells you, with the reason. Under Wayland, which does not let applications claim keys system-wide at all, the dialog says that too rather than showing a binding that quietly does nothing.
  • A searchable, rebindable map of every command, with conflict detection — search by what a command is called, or by the key itself when you want to know what F13 is currently doing. Clear a binding, or restore the shipped defaults, at any time.
  • One table drives all of it. The preview, the projector window, a pedal and a system-wide hotkey all read the same bindings, so changing a key changes it everywhere at once — including the talent's projector, which follows a rebind without being reopened.
  • The space bar, arrows and Home still type normally while you are writing a script: a bare key runs its command only when you are not in the editor.

Fixed

Importing a document (1.4.0's Import), where the theme is: a converter that quietly changes your script is worse than one that refuses it.

  • A PDF no longer claims to know what it left behind. Every other format is read by walking a document model, so "nothing else was left behind" is something the importer can actually check. A PDF has no such structure — an empty list there means we could not tell — and saying otherwise was a false reassurance on the format most likely to be carrying figures and footnotes.
  • A Word text box is no longer reported as a missing picture. Its text was being imported all along, and counted as a picture that had been dropped at the same time. It now becomes its own paragraph rather than being spliced into whichever line happened to anchor it — "we cut ON CAMERA to the wide" is not what the document says.
  • Pictures and running heads are counted once, and only where they are used. A single photo was reported as two, and Word writes three headers and three footers into every document whether or not they are used — so one running head was reported as six. An operator reading "Pictures left out: 2" goes looking for a second picture that never existed.
  • Characters Word stores as elements are no longer lost silently — co‑operate imported as cooperate. A symbol-font glyph is now counted as something left out rather than guessed at, because its code point is an index into a font, not a character.
  • A file that claims to be Unicode and isn't is now reported. Some tools stamp the three UTF-8 marker bytes and then write Western European text anyway; that file imported with every accented character replaced by � and the report saying nothing had been dropped — the exact failure the encoding note exists to prevent, reached down the one path that skipped the check.
  • Normalising a dash still cannot invent a pause. A zero-width space or a stray marker byte sitting between two dashes — which web- and PDF-sourced text routinely carries — split the run in two, and a —​— b imported as a -- caesura nobody wrote. RTF took the same wrong turn by a different route.
  • RTF: an embedded picture or object no longer leaks into the script. The binary payload was being read as if it were text, so the first } byte inside it — which arbitrary binary contains about once every 256 bytes — ended the skip early and emitted the rest into the script.
  • RTF: emoji and other astral characters survive. They arrive as two halves of a surrogate pair, and neither half is a character on its own, so both were being dropped. Non-breaking spaces and hyphens now import as themselves too.
  • The app no longer freezes while a document is read. A large PDF could hold the window for up to ninety seconds.
  • A slow import can no longer land in the wrong dialog. Close it, reopen it, choose a different file, and the first document's result used to arrive and swap the report, the preview and the name — so pressing Import saved a script you had not chosen, under a name that had changed under your hand.
  • A failed save no longer leaves a stub behind. If the write failed — a read-only folder, a full disk, a file a sync client had locked — an empty script was left in the library under your chosen name, and importing again failed with "a script called that already exists". You had to delete it by hand. The rollback never hides the real error.

Find & replace (1.4.0's Find)

  • Replace now edits the script as it is, not as it was a moment ago. Both Replace and Replace all were computed against a copy that lags a round-trip behind the editor — so a dictated utterance arriving in that window was written over and lost, and a match's position could point at the wrong characters.
  • Replacing with text that contains what you searched for now walks forward. Replacing "the" with "there" turned one word into "therere" over three presses while the other two sat untouched, and the counter read "1 of 3" the whole time.

Rehearsal report (1.3.0's rehearsal mode)

  • One stray heading no longer collapses the report. A twelve-paragraph script carrying a single # Title from a Markdown import — or one shot-list line reading #3 CAMERA B — dropped from twelve rows to two, losing the per-paragraph timings the report exists for. One marker labels a script; it does not divide it into sections.

v1.4.0

Getting the script in, and finding your way around it.

Added

  • Open the file you were sent. Until now the only ways text reached the prompter were typing and pasting, which is not how any script arrives. Import on the toolbar reads Word (.docx), RTF, PDF, plain text and Markdown and turns them into clean prompter text: paragraphs kept, bold and fonts and colours flattened, curly quotes and em dashes and ellipses normalised, and a file saved in the wrong encoding read and reported rather than shown as mojibake.
  • It tells you what it could not bring across — pictures, footnotes, comments, headers and footers, link addresses, embedded objects, each with a count. A converter that dropped them quietly is one you find out about on a shoot.
  • Nothing is saved until you say so. You see the text and the report, name the script, and only then does a file appear in your library — and an import will never overwrite a script that already exists.
  • Normalising a dash can never invent a pause: a typeset — becomes an ordinary hyphen, and two of them in a row do not turn into this app's -- caesura token. A .txt you wrote with real caesuras keeps them.
  • Section markers, with a jump list. Any line beginning # — # Verse 2, # Q&A — is a landmark, as is any line that is one of your skip words ([Chorus], Bridge). They appear in a jump list beside the seek bar, are ruled onto the bar itself, and next / previous walk them. Nothing to configure, and an imported Markdown script arrives with its list already built.
  • Previous goes to the top of the section you are in, then past it — the way a previous-track button behaves.
  • Find and replace (Ctrl+F), scoped to the open script: a live match count, match case, whole words only, replace and replace-all. The current match is lit in the editor without taking your typing out of the search box, and a replace-all is a single press of Ctrl+Z.
  • Script statistics, under the read-time estimate: word count, character count, and a warning when a line is long enough to be a wrapped paragraph. The counts are of what is actually performed — skipped labels and pause tokens are left out, so they agree with the duration beside them.

Changed

  • The rehearsal report's sections now split on markers as well as skip labels, so a script written with # headings rehearses section by section with no keywords configured. A section is named without its decoration: [Verse 1] reports as Verse 1.

Security

  • Every document parser runs in a separate process. Three of the five formats are read by third-party crates over a file the app did not write, and release builds abort on panic — so a malformed PDF would otherwise have taken the prompter down mid-read. It also gives the work a time limit and keeps a runaway parser's memory out of the app.
  • The import's temporary file is created in a freshly-made directory with an unguessable name, and on Unix one only its owner can enter.

v1.3.0

Reading to time — and the words that shouldn't count.

Added

  • Tap a tempo, and the app learns yours. BPM has always been turned into a reading speed with one number: how many characters you get through in a beat. It used to be a guess — the same guess for a balladeer and a rapper, who are an order of magnitude apart. Settings → Timing now lets you tap along at the tempo you actually perform at, and measures that number against your own reading speed instead of assuming it.
  • A calibration that puts part of the musical range out of the engine's reach narrows the range the BPM box offers rather than showing you numbers it would quietly change underneath you.
  • A metronome. An optional click on every beat while the script scrolls, accented on the first beat of the bar, with the start countdown doubling as its count-in. The app synthesises it — there is no audio file in the installer and nothing is fetched. It is scheduled against the audio clock from each beat's index, so it does not drift against the scroll: measured over a ten-minute read at the fastest tempo the app offers, the last click is still exact.
  • A bar and beat counter, beside the reading surface, with bar lines ruled on the seek bar — so "sixteen bars" is something you can see rather than something you count in your head. Both appear only when you are working to a tempo.
  • Rehearsal mode, with a timing report. Read the script through and the app tells you, section by section, what it should have taken against what it did, and offers the pace you actually delivered it at. It records timings and nothing else — there is no microphone in it and no audio path at all, which is asserted by a test rather than only promised here.
  • A quiet pace warning while rehearsing, once the read has slipped more than a few seconds from its plan — so an over-running section is visible before the end rather than after it.
  • Words you read but do not perform. Put "Chorus", "Verse" or "Bridge" in Settings → Timing and a line that is nothing but one of them — Chorus, [Verse 1], ## Bridge — costs no time at all, so your lyrics still land on the bar you wrote them for. The same word inside a real line skips only itself, so "back to the chorus now" keeps its timing. They stay on screen, dimmed, and read-aloud never says them.

v1.2.0

The record button, and what happens to the words after you say them.

Changed

  • The record button is a button, not a sentence. It is now a circle with a glyph in it — ● to record, ■ to stop — with the instruction beside it rather than inside it, so the control stops changing width as you use it. Its colour follows your pointer as well as its state: hover the record button and it turns green, and while it is recording, hovering it turns it red. Moving away puts it back without changing what it will do.
  • Dictated words now go through the editor, which is where they always should have gone. Three things follow from that, and each of them was a real fault before:
  • Undo takes back one thing you said, not the whole session. Ctrl+Z used to jump past everything you had dictated in a single step.
  • Text you type while recording is kept. The next thing you said used to overwrite it.
  • Words land where your cursor is, not always at the end of the script.

Fixed

  • Dictation reported itself ready when the speech-model folder existed, even if the model inside it was missing or half-copied — which put a record button on screen that could only fail when pressed. It now checks the one file the recogniser cannot start without.
  • On an install under a drive-letterless volume (a folder-mounted drive), the model path handed to the recogniser was turned into one that named nothing at all. It now names the model. Whether the recogniser then loads it is untested on such a setup — but a failure there says which model it could not open, instead of quietly looking somewhere else entirely.
  • Text inserted with the cursor at the very top of the script could land at the bottom instead. The editor could not read a cursor position sitting on the script as a whole rather than inside a line of it, and answered with the length of everything — so a paste, or a dictated line, went to the end.
  • Dictation could report itself unavailable while a perfectly good model was installed: a stray or half-copied vosk-model-en in your data folder was preferred over the one inside the app, and then rejected for being incomplete. The app now passes over an unusable folder and uses the model it shipped with.

Added

  • npm run dev:vosk for contributors — a development run that can actually dictate. Dictation is release-only otherwise, which made it awkward to work on.

Documentation

  • The docs site's search finds dictation — the feature shipped in 1.1.0 and was not in the index, so searching for it returned nothing. In the other direction, search results that could only ever 404 are gone: the shared index listed five Freally apps that have no site yet, which was 39 of its 109 entries.
  • THIRD-PARTY.md records Vosk (engine, model, and the Windows MinGW runtime) alongside the fonts; NOTICE now carries the pointers to that runtime's corresponding source. Both files ship inside the installer.
  • The Settings note for dictation says what happens when the LAN mirror is on: dictated text reaches its viewers as it is written, the same as typing.

v1.1.0

One way to use your voice, and it works the moment you switch it on.

Added

  • Write your script by speaking (FT-33) — turn on Dictation in Settings → Voice and a record button appears above the script. Press it and what you say is written in; press it again to stop.

It needs no training and no setup: the speech model ships inside the installer, so it works the moment you switch it on. Nothing is downloaded at runtime and nothing is sent anywhere — the recogniser and its model are on your machine, and that stays true offline. The microphone is open only while recording, and nothing you say is written to a file but the text itself.

Removed

  • Voice commands are gone. They required you to record every command — play, pause, faster, slower, next pause, back to top — in your own voice before any of them did anything, and then to hold a button down while speaking. Until you had done that recording, holding the button opened the microphone and achieved nothing at all, with no way to tell that from a fault. Dictation covers the same ground with nothing to train.
  • "Follow my reading" is gone for the same reason: one clear way to use your voice is better than three that overlap. The alignment work behind it still exists as a library and may return in a different shape.

The installers are correspondingly larger than 1.0.0's — the speech model is about 40 MB, and it now ships whether or not you use it.

v1.0.0

The first published release: everything from Phase 3 (hands-free voice control) and Phase 5's polish pass, cut as signed installers for Windows, macOS and Linux.

Added

  • A one-minute welcome tour (FT-50) — the first time you open the app, four short steps introduce the editor and its pause marks, the pace controls, and the projector. Skip it at any point, and run it again whenever you like from Settings → General → Show the tour again.
  • A theme that follows your system (FT-50) — alongside Dark and Light there is now Same as my system, which switches with your OS and repaints immediately when your machine changes at sunset. Dark is still the default.
  • Voice commands (FT-30, FT-31) — train play, pause, faster, slower, next pause, back to top in your own voice (a few takes each) and drive the prompter hands-free. A visible ● Listening indicator shows whenever the microphone is open; choose push-to-talk (mic open only while you hold a button) or always-listening. Off by default. Model-free: matched by classic signal processing against your own recordings, on-device.
  • Voice-following (FT-32, FT-34, FT-35) — an opt-in mode that scrolls the script to keep up as you read it aloud, recognising the script's own words with a bundled Apache-2.0 Vosk model. It steps aside to manual the moment it loses your place (so it never scrolls somewhere wrong) and drives only the operator preview, never the projector. Off by default, and greyed out with a reason where the model isn't installed. The deterministic alignment layer that decides where in the script you are is fully owned and unit-tested.
  • Installers, signed and self-updating (FT-51, FT-52) — Windows (.exe and .msi), a macOS universal .dmg, and Linux AppImage and .deb, all built from one tagged commit. The in-app updater checks the releases page, and every update is verified against a signing key the app carries — an update that cannot be verified is refused rather than installed.

Changed

  • Keyboard and screen-reader accessibility pass (FT-50) — every focusable control now shows a clear focus ring; Tab stays inside an open dialog instead of wandering onto the window behind it; closing a dialog hands focus back to the button that opened it; and the Settings category list now reports the pane you are actually looking at, even while a search is filtering it. The app also honours reduce motion throughout. The script's own scroll is deliberately untouched by that setting — it is the thing the app is for, and you already control its speed directly.

Privacy

  • No network, ever. Voice commands never touch the network or disk — audio is matched in memory against your own recordings, and nothing you say is written to a file. Voice-following runs the recogniser entirely on-device; the model ships in the installer and downloads nothing at runtime.

_Voice-following is not usable in this build. The recogniser needs a speech model, and this release does not yet bundle one — the feature detects that and reports itself unavailable with a reason, rather than failing when you turn it on. Voice commands are unaffected: they need no model and work here. The alignment and grammar layers underneath are complete and tested._

v0.300.0

Offline autocomplete in the script editor: it suggests the rest of the word as you type, in any of the 18 languages, from word lists that ship inside the app.

Added

  • Ghost-text autocomplete (FT-20, FT-21) — start a word and the rest of it appears dimmed ahead of the cursor. Tab accepts it, Esc dismisses it, and typing straight past it ignores it. It completes whole common phrases too, once you finish a word and press space.
  • Autocomplete settings — a new Editor pane with an on/off switch and its own suggestion language, separate from the app's language, because writing a script in one language while running the app in another is normal.
  • Word and phrase tables for all 18 languages (FT-22) — roughly 464,000 words and 345,000 phrases in total, bundled in the installer. Only the language you are actually using is loaded, so this costs the app about 0.5 MB of memory at a time rather than all 12 MB.

Changed

  • The language picker now starts with English and then runs alphabetically by each language's own name — Deutsch, Español, Français, and so on — and stays in that order whichever language the app is in. It previously followed the English names, so the order shifted depending on what you had selected.

Privacy

  • Autocomplete never touches the network. It is a lookup against tables inside the installer — no model, no service, no telemetry. Nothing you type is sent anywhere, and the feature works with the machine offline.

v0.200.0

The standalone teleprompter: scripts you can keep, a projector for the talent, a real transport, and the prompter reading itself aloud.

Added

  • Script library (FT-10) — .ftscript files you can create, open, rename and delete, with a recent list and autosave while you type. They are plain UTF-8 text in one folder, so a script stays readable, diffable and rescuable in any text editor if this app ever goes away. The app reopens whatever you had open last.
  • Caesura chips in the editor (FT-11) — every inline -- / --2 renders as an atomic pill showing its pause length. The caret can only sit before or after one, Backspace/Delete removes the whole token in a single press, Shift+Arrow selects it as a unit, copy/cut carry the real -- text, and a pasted script has its sloppy spacing normalised. Typing the second dash expands a pause for you.
  • Projector window (FT-12) — open the reading surface on any connected display, filling it edge to edge, with a mirror flip for beam-splitter prompter glass. It runs off the same scroll state as the operator preview, so the two can never disagree about where the read is, and it carries its own transport, seek bar and keyboard shortcuts for the person actually reading.
  • LAN mirror (FT-12) — mirror the scrolling script to a phone, tablet or second laptop on your own network, from a link and QR code the app shows you. Each device sets its own font size without changing what anyone else sees. Off by default, loopback-only until you explicitly open it to the network, protected by a key generated fresh each launch, and read-only — nothing a browser sends can change the scroll, the script, or a setting.
  • A real transport (FT-13) — play / pause / stop / rewind / fast-forward as proper SVG buttons with hold-to-repeat, a YouTube-style seek bar with elapsed and total read time (caesura pauses counted) and hover-to-preview the words at any point, click-a-word-to-start, mousewheel scrubbing, and keyboard shortcuts on the projector.
  • BPM speed mode (FT-14) — read the pace as characters per second or as BPM (20–250) for musical delivery. The engine stays chars/sec throughout; BPM is a display and entry mode, and every BPM you can type lands inside the engine's own range rather than being silently clamped underneath you.
  • Appearance controls (FT-15) — typeface, weight, text colour, side margins, line spacing, and where the reading guide sits on screen. These live in the engine rather than in one window's state, so the preview, the projector and the LAN mirror always look alike.
  • Read aloud with per-OS speech synthesis (FT-16) — have the prompter read the script to you at the pace you have set, to audition timing before a take. The highlight follows the spoken word, caesuras become real silences (-- is never pronounced "dash"), and seeking jumps the speech. It is preview-only: it never moves the talent's projector. Uses the voices your OS already has (Windows OneCore/SAPI, macOS AVSpeechSynthesis, Linux Speech Dispatcher / espeak-ng) — no speech engine is bundled.
  • Every language now renders, on every machine. The app bundles the Noto fonts for all 18 languages it ships in — Latin, Greek, Cyrillic, Vietnamese, Arabic, Devanagari, Japanese, Korean and Simplified Chinese. Switching to Japanese on a machine with no Japanese font installed used to show boxes instead of the interface; now the app carries what it needs. Only the character subsets a chosen language actually uses are ever loaded, and the fonts back every reading typeface too, so picking "Serif" can never make a language unreadable. The fonts are under the SIL Open Font License; the notice ships with the app as THIRD-PARTY.md.

Changed

  • The app draws its own window. No OS title bar: a centred title, with Settings, About, and minimise / maximise-restore / close on the right — the same arrangement Freally Player uses, so the suite's windows read alike. The bar is draggable, double-click still maximises, and the window is still resizable from every edge and corner.
  • Settings is now a proper settings dialog, in the same shape as Freally Capture's: a category sidebar (General, Reading, Appearance, Projector, Network) with a search box, and an OK / Cancel / Apply footer. Apply is dead until something actually changes, and a dot marks the categories you have edited but not yet applied.
  • Minimize to the system tray (Settings → General). Off by default. With it on, minimising hides the window and puts an icon in the tray; restoring brings the window back and takes the icon away — the app is in the taskbar or in the tray, never both. The tray menu is translated like the rest of the app.
  • The app icon is now the full artwork at every size. 0.100.0 swapped in a cropped, glyph-only "F" for icons of 48px and below, on the grounds that the wordmark in the artwork is unreadable that small. One consistent mark turned out to matter more than a legible-but-different one, so the crop is gone — taskbar, tray and favicon all show the same icon now.

Fixed

Found by the review pass over this phase, before release:

  • A pause chip could appear where the prompter took no pause. The editor and read-aloud each carried their own copy of the -- grammar, without the bounds the scroll engine applies — so --2.5.3 was drawn as a pause the prompter would skip straight past. All three now share one scanner.
  • Right-clicking the tray icon restored the window instead of opening the menu, which made the tray's own "Quit" unreachable.
  • The first-run agreement's buttons could sit below the bottom of the window now that the app draws its own title bar.
  • Settings could show a value the app was not using. A number the backend clamped (a port below 1024, say) kept displaying as the one that was typed until the next restart; the dialog now shows what was actually stored.
  • Searching Settings could leave the wrong panel on screen, with no highlighted category to explain it.
  • Stop is now one action in the engine rather than a pause-and-rewind pair composed by each button, so anything else that stops the scroll behaves the same way.

Testing

  • Playwright now covers every Phase 1 feature it can reach — the script library's whole lifecycle, chip behaviour, the display picker, the transport, BPM conversion, the appearance pipeline, and the contract that read-aloud never touches the shared scroll. The mocked bridge records IPC calls, so the tests assert what the UI actually asked the backend to do.
  • The app is now launched and photographed on Windows, macOS and Linux every CI run (scripts/app-screenshot.mjs), failing if it dies on start or paints nothing. A compile that succeeds and a webview that renders are different things.
  • Every language is checked for legibility, not just for strings. A test switches the app into each of the 18 locales and proves every character of its own interface is covered by a bundled font. A missing script renders as boxes, which looks like working text to a screenshot and to every other check — so Arabic, Hindi, Japanese, Korean and Chinese could each have broken silently, and only for the people who read them.
  • A translation that quietly fell back to English now fails the build. Bulk translation passes leave English behind for anything they cannot translate, and the result parses and ships looking finished. Genuine exceptions — "OK", "Port", proper nouns — are listed with a reason.
  • Linux rendering is now proven, not assumed. The CI screenshot runs in a container with a software renderer, because GitHub's bare Linux runner has no working GL and produced an empty window whether or not the app was healthy — a check that could not fail was telling us nothing.
  • Everything that genuinely cannot be automated — a second monitor, prompter glass, audible speech, a phone on your Wi-Fi — now has step-by-step drills in Live-To-Do-List.md.

Notes

  • Still no AI, no accounts, no telemetry. The LAN mirror serves your script to devices you point at it, on your own network; nothing is uploaded or relayed. The only outbound request the app makes remains the update check.

v0.100.0

The scaffold phase: the app now builds, runs, and gates on the agreement, with the character-based scroll engine ported over from Freally Capture.

Added

  • Tauri v2 scaffold (FT-00) — a Rust workspace (src-tauri) plus a React/TypeScript UI (ui), sharing Freally Capture's tooling shape: ci.yml, scripts/ci-local.mjs, deny.toml, and a pinned rust-toolchain.toml, so npm run ci:local runs the full gate from day one.
  • Icon set (FT-01) — generated from the source artwork by scripts/make-icons.py, for Windows, macOS and Linux plus the website favicons. Sizes at or below 48px use a glyph-only "F" variant, because the wordmark baked into the artwork is illegible below 64px (measured, not guessed).
  • The character-based teleprompter engine (FT-02) — teleprompter.rs and its TypeScript twin ui/src/lib/caesura.ts, ported from Freally Capture: visible-character scroll offsets, inline -- caesura pauses ( --2 for a custom hold), seek/step/speed controls, and a start-countdown pre-roll. Preview and projector lay text on one fixed-width stage and CSS-scale it, so they wrap at the identical column.
  • Single-window app shell (FT-03) — toolbar, script input, and the live prompter surface, with Settings on the draft/apply pattern (nothing takes effect until Apply; Cancel restores exactly what you had) and an 18-locale Fluent i18n foundation.
  • Docs site + changelog (FT-04) — the site is served by GitHub Pages from /docs, matching the other Freally apps.
  • First-run EULA gate (FT-05) — the app is unusable until the agreement is accepted. Acceptance is versioned and preserved across every settings save (a bug that shipped in Freally Capture; there is now a regression test for it).
  • Problem reporter and update check — the Havoc-standard crash reporter (scrubbed local report, native "stopped unexpectedly" notice, restart, and GitHub/Gmail/mail submission you trigger yourself) and a launch-time update check that shows the version and real release notes before doing anything.

Notes

  • No AI, no accounts, no telemetry. The only outbound request is the update check; it sends nothing about you, your machine, or your scripts.
  • Download links on the site are placeholders until the release pipeline lands.

Known limitations

  • There are no installers yet. 0.100.0 is a source milestone; the signed, per-OS release pipeline is a later phase.
  • Automatic updates are not active. The updater ships without its signing public key, so it refuses every package rather than installing an unverified one — it fails safe, but it also cannot succeed. "Check for updates" will report an error until code signing lands.