VoiceVector v0.6.2

Windows delivery and polish, from user testing:

• Single-exe download — no zip. The Windows release is now one signed
  VoiceVector-windows-x64.exe (SmartScreen shows a verified publisher);
  no zip to extract, no .config sidecar, and the in-app updater swaps the
  exe directly.
• Real app icon everywhere — exe, title bar, and system tray (the tray
  previously showed the default application glyph).
• Fluent styling across the app — checkboxes, combo dropdowns, radio
  buttons, text fields and tooltips now match the violet/Fluent look instead
  of WPF's legacy default chrome.
• Optional user-land installer — first run offers to copy the app to
  %LOCALAPPDATA%\Programs\VoiceVector with a per-user Start Menu shortcut
  and optional start-with-Windows. Never needs admin; declining keeps the
  portable single file.

Full changelog: https://github.com/Sammons/voicevector/blob/main/CHANGELOG.md

VoiceVector v0.6.1

• Windows builds are now Authenticode-signed. Releases are signed via
  Azure Artifact Signing with a public-trust certificate issued to
  Sammons Software LLC; the signing key never leaves Microsoft's HSM and
  CI authenticates keylessly through GitHub OIDC. The exe in this and all
  future releases shows a verified publisher in SmartScreen instead of the
  "unknown publisher" warning. Past unsigned releases keep validating
  forever via their timestamp.
• No code changes since v0.6.0 — this release exists to ship the signature.

Full changelog: https://github.com/Sammons/voicevector/blob/main/CHANGELOG.md

VoiceVector v0.6.0

Big release: review-before-pasting, multi-machine peering, and AI routing.
Multi-machine and routing are labeled Experimental in Settings — new and
still rough; great to try, don't rely on them yet.

• Router picks a numbered destination, not an opaque window id. The router
  model now gets one flat, numbered list of destinations (0 = leave the cursor
  where it is) and replies with a single small number, which the app maps back
  to the real machine + window. Previously it had to echo a 5-digit window id
  verbatim, which small models frequently got wrong — so routing failed and
  fell back to a local paste more often than it should. Much more reliable.
• See where routed text is going, and pick the field. When the router
  targets a window, the staging card now shows a thumbnail of that window with
  the destination input field highlighted; press ⇥ (⇧⇥ to go back) to cycle
  through the window's input fields before you hit ⏎. The receiving machine
  focuses the exact field you chose. Works for a window on this machine or a
  paired one (macOS; the target must share its screen). New window peer
  request and a field index on deliver (docs/multi-machine.md).
• Routed text lands in the app's input field. After the receiving machine
  foregrounds the target app, its text input isn't always focused, so the
  paste went nowhere. macOS now walks the foregrounded window's Accessibility
  tree and focuses the first editable text field/area (the chat box, terminal,
  etc.) before pasting. Needs the receiving Mac's Accessibility permission.
  (Windows foregrounds and pastes into the app's focused control; Linux/Wayland
  pastes into the focused window — no per-field control there.)
• Cross-machine routing actually reaches paired machines. Pairing now
  records the peer's network address on both sides (the receiving side reads
  it from the connection), so you no longer have to type it in by hand. And a
  paired machine is offered to the router even when it isn't sharing its
  screens — "send this to " routes to its focused window (window 0);
  sharing screens still adds per-window targeting. (Delivery into a machine
  still requires that machine's "May paste into me" permission.)
• Auto-submit (press Enter after pasting) — a per-hotkey option, off by
  default. After the text is pasted the app presses Enter, so a dictated chat
  message or terminal command is sent without touching the keyboard. It also
  travels with a routed delivery: the receiving machine presses Enter, which
  is the point for the cross-machine case where no one is at that keyboard.
  Leave it off for editors, where Enter just inserts a line.
• Speak the destination. When you open a router dictation by naming where
  it should go — "Hey Slack, …", "Send this to the terminal —", "Tell Ben
  that …" — the router now honours that spoken destination (it's given your
  original words alongside the window lists), and the cleanup strips the
  addressing phrase out of the pasted text instead of leaving it in.
• Routed text only pastes once the target window is actually focused. On
  macOS and Windows the app now confirms the destination window came to the
  front before pasting; if the system refuses the focus change, the text is
  copied with a "click it and press ⌘V/Ctrl+V" notice instead of landing in
  whatever window happened to be focused (e.g. the terminal). (Linux/Wayland
  can't focus another app's window, so it pastes into the focused one as
  before.)
• The AI router's choice is validated and bounced back. The destination
  must be a machine and window that were actually offered (or window 0 for the
  current focus); an invalid or hallucinated pick is returned to the router
  with a specific error telling it what was wrong, up to three tries, before
  falling back to the normal local paste. No more silent routing to nothing.
• A spoken revision that transcribes to nothing now says so ("Didn't catch
  that…") instead of silently doing nothing.
• Multi-machine peering (Settings → Multi-machine, all three apps).
  VoiceVector instances on the same tailnet/LAN pair with a 6-digit code
  confirmed on both screens (Bluetooth-style numeric comparison over a
  commit-then-reveal exchange), then talk over TLS with pinned self-signed
  identities — no central authority. Per-peer permissions, both off by
  default: "may see my screens" and "may paste into me". Protocol:
  docs/multi-machine.md.
• Route with AI (per hotkey, needs Review before pasting). A router model
  looks at the draft, your windows and screens — and paired machines' too —
  and picks the window the text was meant for; the staging card shows its
  choice ("→ Slack on crankshaft") and ⏎ sends it there: the target window
  is raised before pasting locally, or the text is delivered to the paired
  machine, which pastes and saves it as a routed entry. Any router failure
  falls back to the normal paste. Prompt: shared/prompts/router.txt.
  Linux note: Wayland hides other apps' windows, so a Linux machine offers
  its screens and receives text into the focused window, without per-window
  targeting.

Full changelog: https://github.com/Sammons/voicevector/blob/main/CHANGELOG.md

VoiceVector v0.4.1

• Screenshot context covers every display. One image per display is
  attached (not just the frontmost window), each preceded by a caption saying
  which display is active — the one the text will be inserted into — and on
  macOS and Windows that window is outlined in red in its screenshot. Linux
  crops the portal's desktop capture per monitor; Wayland hides the focused
  window, so with several displays the caption says the active one is unknown.
• Screenshots are saved with the entry. <id>-screen-N.jpg beside the
  WAV and Markdown, described by screenshots / activeScreenshot /
  screenshotOutline front-matter keys (docs/storage-format.md). One set per
  dictation, taken when the hotkey fires; a retry from the library reuses it
  instead of capturing whatever is on screen now. Deleted with the entry.
• Fix (macOS): the app crashed at launch when "Keep the microphone always
  warm" was on.
• Fix (macOS): binding a hotkey to a regular (non-modifier) key crashed
  the app whenever Settings showed that hotkey's name.
• Fix (macOS): a newly added hotkey that was never set swallowed the
  letter A everywhere (its placeholder is key code 0, which is A).

Full changelog: https://github.com/Sammons/voicevector/blob/main/CHANGELOG.md

VoiceVector v0.4.0

• Linux app (apps/linux/, Ubuntu 26.04 / GNOME 50 reference; Wayland
  and X11). C on GTK 4 / libadwaita, dynamically linked against system
  libraries only — a ~230 KB binary, no runtime to install. Hotkeys through
  the XDG GlobalShortcuts portal by default (sudo-free, hold-to-talk works),
  or raw evdev input for modifier-only keys after a one-time input group
  step; paste through the RemoteDesktop portal with a clipboard fallback;
  API keys in the Secret Service. Same storage format, profiles, review mode,
  and prompts as the other two apps, checked by a 104-assertion self-test in
  CI. See docs/linux/README.md.
• Review before pasting (per hotkey). The cleaned text is staged in a
  card above the recording pill instead of being pasted. Press the hotkey and
  say a change — "make it shorter", "turn that into a list", "sign it Ben" —
  as many times as you like; a review model (the cleanup model by default, or
  one you pick) rewrites the draft each time. ⏎ pastes, Esc discards. Neither
  key reaches the app underneath while a draft is staged.
• Screenshot context (per hotkey). A screenshot of the frontmost window is
  captured when the hotkey fires and attached to the cleanup call and every
  revision, so the model knows what you're looking at. macOS asks for Screen
  Recording once; models without vision fall back to text automatically.
• Reviewer prompt lives in shared/prompts/review.txt, embedded and
  self-test-checked in both apps like the cleanup prompts.

Full changelog: https://github.com/Sammons/voicevector/blob/main/CHANGELOG.md

VoiceVector v0.3.2

Recording feel, especially with external audio interfaces.

• Instant HUD. The recording overlay appears the moment the hotkey fires;
  the microphone opens in the background. USB interfaces (Focusrite Scarlett
  and friends) take about half a second to open, and that used to block the
  whole UI.
• Warm microphone (Settings → General). Two independent options: keep the
  microphone open for 15 seconds after a recording (on by default — back-to-back
  takes start instantly) and always keep it open while VoiceVector runs (off by
  default). Either shows the OS mic-in-use indicator while the mic is open.
• Level meter that works at any gain. Conservatively gained interfaces
  produce speech around −40 dBFS — perfectly transcribable, but the old linear
  meter barely moved. The meter is now a fixed, quantized scale (−60 dBFS
  silence → −15 dBFS full, three steps), and handles Float32/Int16/Int32 and
  multichannel input, taking the loudest channel.
• Voice detection relative to the noise floor so silence-gap streaming
  (transcribe-while-you-pause) works on quiet inputs too.
• VoiceVector --probe-audio [seconds] (macOS) prints the default input's
  format, per-channel levels, and device open latency, for diagnosing this
  class of problem.

Full changelog: https://github.com/Sammons/voicevector/blob/main/CHANGELOG.md

VoiceVector v0.3.1

Full Changelog: https://github.com/Sammons/voicevector/commits/v0.3.1