v1.1.0

Highlights

• --debug flag logs one line per STT request (model, language, prompt, audio duration) so you can see exactly what's being sent. Output to stderr via the scribe logger.
• Streaming recipes documented: Balanced (default) and Patient profiles, with the cross-chunk context trade-off explained.

New features

• --words auto-formatting for backends that fold words into the prompt (whisper-futo, openai, groq). A bare wordlist Tierney Comet is now rendered as "Tierney, Comet." — the trailing period biases Whisper toward punctuated output. faster-whisper's hotwords channel is unchanged.
• --stream-first-chunk-min (default 3.0 s) raises the floor for the very first chunk of a streaming thread, giving Whisper enough audio for a punctuated bootstrap transcript whose tail seeds the rolling prompt. Inactive in Patient mode (--stream-context-length 0).
• Options → Prompt submenu in the tray with file pickers for prompt.txt / words.txt and a Reload action so editing those files in a text editor is a one-click refresh.
• Menu presets: 15 s added to Chunk max picker, 8× added to Context reset picker.

Fixes

• Auto-discovered prompt.txt / words.txt paths now persist into the argparse namespace at startup, so the tray menu correctly shows the loaded file instead of (none).

Docs

• docs/backends.md — new Prompt style biases output style section (why your transcripts may come back unpunctuated and what to do about it), plus Streaming recipes — two profiles.
• docs/cli.md — new --stream-first-chunk-min row.

v1.0.1

Fix: insert space between concatenated chunks when neither side has whitespace (768aa6b).

v1.0.0 — First-class Stream mode, restructured menus, language flags

v1.0.0 — First-class Stream mode, restructured menus, language flags

Major milestone. Stream mode is now a first-class recording mode, the
parameter surface is split and renamed (Stream / Clip vocabulary
throughout), the tray menu is restructured for clarity, and output
sinks are pluggable with live mid-recording switching.

Headline changes since v0.18.0:

• Mode toggle Stream / Clip at the top level. CLI gets --stream /
  --clip (--realtime, --pseudo-streaming kept as hidden aliases).

• Streaming-params refactor — --silence-duration / --duration split
  into purpose-specific flags:
  --stream-chunk-min / --stream-chunk-max / --stream-chunk-silence-break
  --stream-context-reset-silence (multiplier of silence-break)
  --stream-context-length (rolling-context char cap, 0 = OFF)
  --realtime-commit-silence (OpenAI realtime backend)
  --stream-timeout / --clip-timeout (mode-specific auto-stop)
  Old flags stay as hidden back-compat aliases.

• Auto + Max chunk-cut strategies. silence-break=0 → Auto (longest
  silence in window, respecting chunk-min); silence-break=None → Max
  (force-cut only, no silence-based cuts).

• Output / Keyboard / File restructure. Output radio with four
  mutually-exclusive destinations (Keyboard / Clipboard / Terminal /
  File). Keyboard advanced submenu (visible iff Output=Keyboard)
  groups the typer Backend and a new Input mode radio (keystroke vs
  paste). File destination has a Choose path… picker (tkinter).

• Output class refactor (scribe/output.py) — pluggable sinks
  (KeyboardOutput / ClipboardOutput / TerminalOutput / FileOutput).
  start_recording rebuilds the sink at every chunk if the user
  changed Output / Typer / Input mode mid-recording, so menu changes
  take effect live. FileOutput drops the trailing newline that made
  the realtime backend produce one-word-per-line files.

• Vosk as a leaf in the Model menu. Model resolves from the active
  language via autoselect_language(); Auto on vosk displays as
  'Auto (🇬🇧 en)' without mutating o.language.

• Language menu with flags — ISO 639-1 short codes (en / fr / de /
  it) + origin-country flags via desktop-ai-core v0.3.0's new
  default_country() / flag_for() helpers.

• 🏠/☁️ prefix on model labels distinguishes on-device from cloud;
  vendor·model uses a middle-dot separator.

• About submenu with app identity, version, license, GitHub link.

• Backend × mode smoke-test matrix with --dry-run plumbing in every
  backend so the recording pipeline is exercised without API keys or
  local models on disk. Adds a regression guard for the same class of
  AttributeError that hit silence_duration after the rename.

• pyproject discoverability: 23 classifiers (Topic, Audience,
  Environment, Natural Language, License, Dev Status) and a tighter
  23-keyword list. project.urls grows Source, Issues, Changelog,
  Funding.

• Menu / docs alignment audit, terminal-frontend readability fix
  (every picker parent Item now carries a static help= so TUI users
  see 'Chunk min' instead of 'min').

Cross-repo dep: requires desktop-ai-core ≥ v0.3.0.