Fono

Context-aware dictation, Esc-to-cancel on Wayland, and smarter
first-run model picks. Fono now reads which window is focused when you
press the hotkey and silently adjusts how it transcribes and cleans up
your speech — terminal windows get a shell-vocabulary bias so ls,
git commit, and chmod 755 come out right; code editors get a
language-specific vocabulary hint. No configuration required.

Added

• Window-aware context injection. At hotkey-press time Fono reads
  the focused window class and title and silently passes a tailored
  initial_prompt to Whisper (local and cloud) and a matching cleanup
  suffix to the LLM. Built-in profiles cover terminal emulators (shell
  vocabulary: ls -la, grep -r, chmod 755, git commit, etc.),
  per-language code editors (Cursor, Zed, Kate; hints derived from
  the file extension in the window title), and private windows
  (KeePassXC, Bitwarden — history writes suppressed). Existing
  [[context_rules]] entries take precedence as usual. Detection
  covers X11, sway, Hyprland, and GNOME Wayland (with XWayland
  fallback for GNOME 46+ where the Shell introspect API is restricted).
  Visible at any time with FONO_LOG=fono::context=debug.
• Terminal project and agent detection. When a terminal emulator is
  focused, Fono walks /proc to find the CWD and detects the active
  project type (Rust, Python, Node, Go, Docker, K8s) and whether a
  coding agent (Forge, Claude Code, Codex, Aider, Goose, and others)
  is running. Project type refines the Whisper vocabulary hint; agent
  detection is stored for future prompt biasing and is currently no-op.
• Escape cancels recordings on Wayland. The portal hotkey
  backend (KDE / sway / Hyprland) opens a transient second
  GlobalShortcuts session while a recording is active; the
  GNOME-Wayland gsettings shim writes a temporary fono-cancel
  custom-keybinding for the same duration. Either way Esc is only
  grabbed while Fono needs it.
• fono cancel CLI verb. Aborts an active recording or
  assistant turn. Idempotent. Backs Request::Cancel and the
  Esc grab above.
• Native aarch64 release binary. fono-vX.Y.Z-aarch64 is now
  built on a hosted ubuntu-22.04-arm runner alongside the x86_64
  builds and is gated by the same size-budget check (same glibc 2.35
  floor; verified end-to-end on a Debian 13 aarch64 host).

Fixed

• Dictation on PipeWire-only Linux hosts. On stock Ubuntu 24.04
  (and similar systems without pulseaudio-utils installed) the
  pw-cat capture helper was missing --raw and emitted a
  containerized stream that Fono interpreted as PCM — recordings
  came out as noise. The capture path now passes --raw and
  produces clean audio on every PipeWire setup.
• Wizard recommendation accuracy on older iGPUs. The picker no
  longer credits CPU-only builds with a GPU multiplier they can't
  deliver, and Vulkan-capable integrated GPUs are split into two
  classes (Integrated at 1.3× for fp16-only parts like UHD 620,
  IntegratedTensor at 2.0× for fp16 + VK_KHR_cooperative_matrix
  parts like Lunar Lake Xe2 and Apple Silicon). The small.en
  registry anchor was also off by 2× — fixed against the matrix.
  Net effect: older laptops are recommended small or small.en
  instead of a turbo model that can't keep up; modern tensor-iGPU
  laptops correctly get large-v3-turbo. fono doctor now uses
  the same affordability walk as the wizard so the two never
  disagree.

Changed

• Wizard model selection is now data-driven. A new three-class
  HostGpu classifier (None / Integrated / Discrete), derived
  from the Vulkan probe's deviceType + shaderFloat16 bit, replaces
  the previous static accelerated() 4× heuristic with multipliers
  1× / 2× / 4× (no PCI tables, no runtime calibration, no maintained
  device lists; see ADR 0028). The live-mode Borderline affordability
  middle state and the two LIVE_REALTIME_MIN_* constants are gone —
  the wizard now applies a single BATCH_REALTIME_MIN = 2.0 gate and
  every shortlist entry is comfortable by construction. Quantization
  defaults are unified on q8_0 across the registry (per the ADR 0027
  2026-05-25 amendment): tiny, tiny.en, small, small.en, and
  large-v3-turbo all default to q8_0; the previous q5_1 defaults
  remain reachable via [stt.local].quantization. wer_by_lang
  English numbers are refreshed to Open-ASR-Leaderboard means (rounded
  up: tiny 12→16, tiny.en 9→13, small 6→10, small.en 5→9,
  large-v3-turbo 4→8) so the accuracy buckets the wizard surfaces
  match the public prior users will find elsewhere.

• Assistant history now survives dictation pivots. Previously a
  press of the dictation hotkey (F7) while an assistant conversation
  was in flight would wipe the rolling chat history; the next F8
  turn would start fresh and the assistant would tell you it had no
  memory of what you'd just said. Dictation and the assistant have
  always had fully separate histories — the dictation transcript log
  lives in SQLite, the assistant's chat turns live in memory — so
  the cross-wired auto-clear was surprising rather than protective.
  The pivot still stops any in-flight assistant playback so it
  doesn't talk over your dictation, but the chat history is
  preserved and you can resume the conversation on the next F8.

Removed

• fono assistant stop CLI verb — use fono cancel instead.
• "Stop assistant" tray entry — redundant with fono cancel.
• [assistant].auto_clear_on_dictation config key. No longer
  read; remove it from your config.toml if present (unknown keys
  are silently ignored, so existing configs keep working). The
  remaining knobs (history_window_minutes, history_max_turns)
  and the tray "Forget conversation" entry / fono assistant forget CLI still cover every legitimate need to drop history.

Full Changelog: https://github.com/bogdanr/fono/compare/v0.8.1...v0.8.2