Fono

0.2.0 — 2026-04-27

Single-binary local stack: STT (whisper.cpp) and LLM cleanup
(llama.cpp) now ship together in one statically-linked fono binary,
out of the box, with hardware-accelerated CPU SIMD selected at runtime.

Added — single-binary local STT + LLM

• llama-local is now part of the default features set. The previous
  compile_error! guard in crates/fono/src/lib.rs is gone — both
  whisper-rs and llama-cpp-2 link into the same ELF.
• .cargo/config.toml adds -Wl,--allow-multiple-definition to
  deduplicate the otherwise-colliding ggml symbols vendored by both sys
  crates. Both copies originate from the same ggerganov upstream and
  are ABI-compatible; the linker keeps one set, no UB at runtime.
• New accel-cuda / accel-metal / accel-vulkan / accel-rocm /
  accel-coreml / accel-openblas features on crates/fono that
  forward to matching whisper-rs / llama-cpp-2 features for opt-in
  GPU acceleration.
• Startup banner prints a new hw accel : <accelerators> + CPU <SIMD>
  line (runtime SIMD probe: AVX512 / AVX2 / AVX / SSE4.2 + FMA + F16C on
  x86; NEON + DotProd + FP16 on aarch64).
• LlamaLocal::run_inference redirects llama.cpp / ggml's internal
  printf-style logging through tracing (matches the existing
  whisper_rs::install_logging_hooks pattern). Default verbosity now
  emits a single LLM ready: <model> (<MB>, <threads> threads, ctx=<n>) in <ms> line; cosmetic load-time warnings (control-token type,
  n_ctx_seq < n_ctx_train) are silenced. Re-enable on demand with
  FONO_LOG=llama-cpp-2=info.
• New smoke test crates/fono/tests/local_backends_coexist.rs boots
  WhisperLocal and LlamaLocal in the same process to lock in the
  no-collision contract.

Added — wizard local LLM path

• First-run wizard now offers Local LLM cleanup (qwen2.5, private, offline) as a top-level option in both the Local and Mixed paths, in
  addition to Skip and Cloud. New configure_local_llm helper picks
  a tier-aware model: qwen2.5-3b-instruct (HighEnd),
  qwen2.5-1.5b-instruct (Recommended/Comfortable),
  qwen2.5-0.5b-instruct (Minimum/Unsuitable). All Apache-2.0 per
  ADR 0004.
• The wizard's auto-download now fires for either local STT or local
  LLM (was STT-only).

Added — tray UX

• Tray STT and LLM submenus now show a ● marker beside the active
  backend (was missing — active_backends() returned the trait name()
  while the comparison logic expected the canonical config-string
  identifier).
• Switching to the local STT or LLM backend from the tray now ensures
  the corresponding model file is on disk first, with a "downloading…"
  notification, a "ready" notification on completion, and a clear error
  notification on failure (with the orchestrator reload skipped to keep
  the user on a working backend).

Changed — hotkey defaults

• toggle = "F9" (was Ctrl+Alt+Space). Single key, no default
  binding on any major desktop, easy to fire blind.
• hold = "F8" (was Ctrl+Alt+Grave). Adjacent to F9 for natural
  push-to-talk muscle memory.
• cancel = "Escape" unchanged (only grabbed while recording).
• paste_last hotkey removed. The tray's "Recent transcriptions"
  submenu and the fono paste-last CLI cover the same need with a
  better UX (re-paste any of the last 10, not just the newest).
  Request::PasteLast IPC and Cmd::PasteLast CLI are preserved and
  now route directly to orch.on_paste_last().

Changed — release profile size

• [profile.release] now sets strip = "symbols" and lto = "thin",
  trimming the dev cargo build --release artifact from ~23 MB → ~19 MB
  (no code removal — only .symtab / .strtab deduplication).
  release-slim (used by packaging CI) is unchanged at ~15 MB.

Documented

• docs/status.md — new entries for hotkey ergonomics and the
  single-binary local-stack resolution.
• docs/troubleshooting.md, docs/wayland.md, README.md updated for
  the new default hotkeys.
• New plans: plans/2026-04-27-shared-ggml-static-binary-v1.md (the
  shared-ggml strategy that informed the linker-dedupe shortcut),
  plans/2026-04-27-llama-dynamic-link-sota-v1.md,
  plans/2026-04-27-candle-backend-benchmark-v1.md,
  plans/2026-04-27-local-stt-llm-resolution-v1.md.

0.1.0 — 2026-04-25

First public release. Pipeline (audio → STT → LLM → inject) is fully wired
end-to-end; default release ships local whisper.cpp out of the box.

Added — pipeline

• SessionOrchestrator (crates/fono/src/session.rs) glues hotkey FSM →
  cpal capture → silence trim → STT → optional LLM cleanup → text injection
  → SQLite history. Hot-swappable backends behind RwLock<Arc<dyn …>>.
• fono record — one-shot CLI dictation (microphone → stdout / inject).
• fono transcribe <wav> — runs a WAV file through the same pipeline; useful
  for verifying API keys without a microphone.

Added — providers

• STT: local whisper.cpp (small / base / medium models), Groq cloud
  (whisper-large-v3-turbo), OpenAI cloud, optional Deepgram / AssemblyAI /
  Cartesia stubs.
• LLM cleanup: optional, off-by-default. OpenAI-compatible endpoints
  (Cerebras, Groq, OpenAI, OpenRouter, Ollama) and Anthropic.
• STT and TextFormatter traits with prewarm() so the first dictation
  after daemon start is not cold (latency plan L2/L3).
• fono use {stt,llm,cloud,local,show} — one-command provider switching;
  rewrites config atomically and hot-reloads the orchestrator (no restart).
• fono keys {list,add,remove,check} — multi-provider API-key vault with
  reachability probes.
• Per-call overrides: fono record --stt openai --llm anthropic.

Added — hardware-adaptive setup

• crates/fono-core/src/hwcheck.rs — pure-Rust probe of physical/logical
  cores, RAM, free disk, and CPU features (AVX2/NEON/FMA). Maps to a
  five-level LocalTier (Unsuitable, Minimum, Comfortable,
  Recommended, High-end).
• Wizard prints the live tier and steers the user toward local vs cloud
  based on what the machine can sustain.
• fono hwprobe [--json] exposes the snapshot for scripts.
• fono doctor shows the active hardware tier alongside provider
  reachability and the chosen injector.

Added — input / output

• Default key-injection backend Injector::XtestPaste — pure-Rust X11 XTEST
  paste via x11rb + xsel/wl-copy/xclip clipboard write. No system
  dependencies beyond a clipboard tool. Shift+Insert is the default paste
  shortcut (universal X11 binding).
• Override paste shortcut via [inject].paste_shortcut = "ctrl-v" in config
  or FONO_PASTE_SHORTCUT=ctrl-shift-v env var.
• Always-clipboard safety net: every successful dictation also writes to both
  CLIPBOARD and PRIMARY selections (general.also_copy_to_clipboard = true).
• Always-notify: notify-rust toast on every dictation
  (general.notify_on_dictation = true).
• fono test-inject "<text>" [--shortcut <variant>] — smoke-tests injection
  and clipboard delivery without speaking.

Added — tray

• Recent transcriptions ▸ submenu with the last 10 dictations; click to
  re-paste.
• STT: <active> ▸ and LLM: <active> ▸ submenus for live provider
  switching from the tray (same code path as fono use).
• Open history folder (was misrouted to Dolphin in pre-release; now opens
  the directory itself via xdg-open).

Added — safety + observability

• Per-stage tracing breadcrumbs at info: capture=…ms trim=…ms stt=…ms llm=…ms inject=…ms (raw_chars → cleaned_chars).
• Pipeline in-flight guard refuses concurrent recordings with a toast.
• Skip-LLM-when-short heuristic (configurable llm.skip_if_words_lt) saves
  150–800 ms per short dictation.
• Trim leading/trailing silence pre-STT (audio.trim_silence); ~30 % faster
  STT on 5 s utterances with 1.5 s of tail silence.

Added — benchmark harness

• New crates/fono-bench/ crate: 6-language LibriVox fixture set (en, es,
  fr, de, it, ro), Word Error Rate + per-stage latency report, criterion
  benchmark, regression gate. CI-fast (network-free) and full-stack modes.

Documented

• docs/plans/2026-04-25-fono-pipeline-wiring-v1.md (W1–W22, all landed).
• docs/plans/2026-04-25-fono-latency-v1.md (L1–L30, 17 landed, 13
  deferred-to-v0.2 with rationale).
• docs/plans/2026-04-25-fono-local-default-v1.md (H1–H25).
• docs/plans/2026-04-25-fono-provider-switching-v1.md (S1–S27).
• docs/plans/2026-04-25-fono-roadmap-v2.md (post-v0.1 roadmap).
• ADR docs/decisions/0007-local-models-build.md — glibc-linked default
  release vs musl-slim cloud-only artifact.

Models locked in v0.1.0

| Provider | Model | License | First-run download |
|---|---|---|---|
| Whisper local | ggml-small.bin (multilingual) | MIT | ~466 MB |
| Whisper local (light) | ggml-base.bin | MIT | ~142 MB |
| Groq cloud STT | whisper-large-v3-turbo | (cloud, no license) | n/a |
| OpenAI cloud STT | whisper-1 | (cloud) | n/a |
| Cerebras cloud LLM | llama-3.3-70b | (cloud) | n/a |
| Groq cloud LLM | llama-3.3-70b-versatile | (cloud) | n/a |

Local LLM (Qwen2.5 / SmolLM2) is opt-in behind the llama-local Cargo
feature and ships fully wired in v0.2.

Verification

• 86 unit + integration tests; 2 latency-smoke #[ignore] tests.
• cargo clippy --workspace --no-deps -- -D warnings clean (pedantic +
  nursery).
• DCO sign-off enforced on every commit.

Known limitations

• No streaming STT/LLM yet (latency plan L6/L7/L8 deferred to v0.2). Latency
  on cloud Groq+Cerebras is ~1 s end-to-end on a 5 s utterance.
• Wayland global hotkey requires compositor binding to fono toggle
  (org.freedesktop.portal.GlobalShortcuts not yet stable in upstream
  compositors).
• Local LLM cleanup (Qwen / SmolLM) is opt-in / preview.
• Real winit + softbuffer overlay window is a stub (event channel only).