Fono

Added

• Voice assistant — F10 hold-to-talk, streaming chat, TTS playback.
  A second push-to-talk key (F10 by default) turns Fono into an
  offline-capable voice assistant. The pipeline diverges after STT:
  instead of cleaning the transcript and injecting it, Fono asks a
  chat-capable LLM, streams the reply sentence-by-sentence into a TTS
  backend, and plays the audio. First sentence starts speaking before
  the model finishes generating, so time-to-first-audio is bounded by
  one sentence's synth latency rather than the full reply.
• [assistant] and [tts] config blocks. Independent backend
  selection from the [llm] cleanup pipeline — pick a fast local 3B
  for cleanup and a bigger cloud model for the assistant, or any
  mix-and-match. Multi-turn rolling history with a configurable time
  window (default 5 minutes) and max-turn cap (default 12). Pressing
  the dictation key clears assistant context (configurable);
  pressing F10 again mid-reply barges in with history retained;
  Escape stops playback ("shut up") without forgetting.
• Cloud assistant backends. Anthropic (Claude Haiku 4.5) and the
  full OpenAI-compatible family — OpenAI (gpt-5.4-mini), Cerebras
  (qwen-3-235b-a22b-instruct-2507), Groq (openai/gpt-oss-120b),
  OpenRouter, Ollama. Each
  ships in the default binary; one feature flag per family lets slim
  builds drop unused providers.
• Cloud cleanup model defaults refreshed to match retired and
  newly-released models: Cerebras llama3.1-8b, Groq
  openai/gpt-oss-20b, OpenAI gpt-5.4-nano, Anthropic
  claude-haiku-4-5-20251001. The OpenAI-compat client now sends
  max_completion_tokens (the new field name newer OpenAI models
  require; older models still accept it).
• TTS backends. Wyoming protocol client (any
  wyoming-piper-style server on the LAN), the OpenAI
  /v1/audio/speech API (24 kHz PCM stream), and an in-process
  Piper stub that points users at Wyoming-piper for now (the
  static-musl ship build can't yet pull in onnxruntime). Audio
  playback uses paplay on the Linux release variant (no libasound
  link, matches the existing parec capture path) or cpal behind the
  cpal-backend feature.
• CLI surface. fono use assistant <backend>,
  fono use tts <backend> [--uri tcp://host:port],
  fono assistant {press,release,stop} for scripted end-to-end
  testing.
• Tray. New Stop assistant and Forget conversation entries;
  Assistant backend and TTS backend submenus mirror the existing
  STT/LLM submenus and switch backends live via Reload. Tray icon
  flips amber while the assistant is active.
• Wizard. fono setup ends with an opt-in assistant + TTS step;
  reuses any cloud key already entered earlier in the flow so a
  single OPENAI_API_KEY powers both chat and TTS without a second
  prompt.
• Doctor. fono doctor exercises both factories at startup so a
  missing API key or unreachable Wyoming server surfaces in one
  place; new Providers (assistant) and Providers (TTS) tables
  show key/URI status per backend with an active marker.
• Overlay feedback for the assistant flow. Recording paints
  green ("ASSISTANT") with the chosen waveform style; the post-
  release thinking + speaking phase paints amber ("THINKING") with
  per-style synthetic animations distinct from the real-audio
  recording shape:
  • FFT — Gaussian "scanner" (σ ≈ 8 bins out of 100) sweeps
    across the panel; per-bin breathing baseline blends in via
    a screen composite so the bell emerges smoothly.
  • Bars — symmetric centre-out, peak at midline rippling
    outward.
  • Oscilloscope — two interfering sine waves with edge taper
    pinning x = 0 / x = 1 to the centerline; central antinode
    reaches ±1.0 without clipping.
  • Heatmap — two anti-phased Gaussian "neural strands"
    crossing over the rolling 6 s window; transitions seamlessly
    from recording-FFT data without clearing the cache.
    Default [overlay].style flipped from Bars → FFT — most active
    visualisation across both phases.
• Runtime overlay style swap. Changing [overlay].style via
  fono use, the tray Waveform style submenu, or fono config edit now applies on the next frame instead of waiting for a
  daemon restart.
• Smoke-test binary (cargo run --release --example smoke_assistant -p fono) exercises each cloud assistant + the
  OpenAI TTS path end-to-end. The release CI's new
  cloud-assistant job runs the --ci subset (Groq + Cerebras,
  the providers whose API keys are stored as GitHub Secrets).

Fixed

• FSM stuck on a sub-300 ms F10 tap. Brief F10 taps released
  before MIN_RECORDING left the orchestrator's
  on_assistant_hold_release early-returning without firing
  ProcessingDone; the FSM sat in AssistantThinking forever and
  silently rejected subsequent F8/F9/F10 presses. Every early-return
  path now emits ProcessingDone; AssistantRecording also accepts
  ProcessingDone as a safety net.
• Audio playback worker dying after every cancel. pb.stop()
  used to send Cmd::Stop which made the worker break out of its
  loop; the next turn's enqueue then failed with "audio playback
  worker stopped". Cmd::Drain now drains queued items + clears the
  abort flag without exiting the worker, so multi-turn conversations
  keep working across barge-ins, Forget, and dictation pivots.
• Frozen overlay during the post-release phase. The level task
  was aborted in stop_and_drain the moment capture ended, leaving
  the waveform on its last pre-release frame for 4–5 s while STT +
  LLM ran. The overlay now switches into the synthetic thinking
  animation as soon as F10 is released, and the FFT thinking
  visualisation gets even-spaced inter-bar gaps via integer-aligned
  slot widths.

Full Changelog: https://github.com/bogdanr/fono/compare/v0.6.1...v0.7.0