Fono

Show your screen, dictate in any language. This release teaches the voice
assistant and your coding agents to look at what you're pointing at, fixes
AI cleanup so it stops dropping text and accents on non-English dictation, and
adds a few new looks for the recording overlay.

Added

• Point at your screen and ask. The F8 voice assistant and any
  connected coding agent can now see your screen when you reference
  something on it — "what does this error mean?", "read this dialog to
  me". Fono grabs the focused window automatically, or opens your
  desktop's region picker so you can frame exactly what to share, then
  hands the picture to the model. Private windows (KeePassXC, Bitwarden,
  1Password) are never captured. Works out of the box with whatever
  screenshot tool you already have (scrot, grim, maim, spectacle,
  gnome-screenshot, …) — no new required dependencies. fono doctor
  shows whether capture is ready.
• New looks for the recording overlay. Three fresh visualisation
  styles join the picker: Aurora Beziers (Siri-style glowing
  ribbons), System/360 (a retro mainframe console-lamp spectrum),
  and Terrain 3D (your voice as a flowing 3D landscape). Pick one
  from the tray's Visualization menu.

Changed

• The voice assistant is on by default. The pipeline that powers F8
  and the coding-agent voice loop now works without extra setup. If you
  had explicitly turned it off, that choice is respected.
• Voice mode talks more naturally. The built-in voice preset for
  coding agents was rewritten: agents now listen by default, only ask
  bounded A/B/C questions when it actually helps, never ask you to
  approve risky actions by voice, and open each spoken turn with a short
  cue so you have a moment to refocus before the answer.

Fixed

• Dictation cleanup no longer drops your words — or your accents.
  On non-English dictation, the AI cleanup step could silently come back
  empty and inject the raw, unpolished transcript instead; diacritics
  (ă, î, ș, ț, é, ñ, …) could also get lost on the way to the cursor.
  Both are fixed: cleanup now reliably tidies up non-English text and
  restores the correct accented characters. When a coding agent is in
  focus in a terminal, dictation is framed as prose (capitalisation and
  punctuation) rather than shell commands.
• The assistant now actually answers about your screen. Previously
  it captured the screen but spoke a placeholder instead of describing
  what it saw. It now sends the image to the model and reads back the
  real answer.
• Escape reliably cancels while the agent is listening, and Ctrl-C
  restores the tray icon cleanly when you stop a voice session.

Full Changelog: https://github.com/bogdanr/fono/compare/v0.9.0...v0.9.1