v1.0.0

First tagged release of Verbatim, a real-time, vendor-agnostic macOS dictation tool with
visible corrections. Supersedes the earlier v0.1.0 tag.

Note on scope. Per the project's own milestone plan
(docs/architecture/git-and-release.md, docs/product/roadmap.md), a v1.0.0 tag is
defined as landing at milestone M6 — public release, Windows support, signed/notarized
builds. As of this tag, the project is at M4 (functionally complete) / M5 (in
progress): macOS-only, source-only, not code-signed or notarized. This tag reflects
the maintainer's explicit decision to version the current state as 1.0.0; it does not
assert M6 exit criteria are met. See Known limitations below.

Highlights

• Real-time dictation with a two-layer pipeline — a live transcript streams as you speak
  (never blocked on the correction model), and a correction pass runs when a segment
  finalizes.
• Visible corrections — removed and replaced spans are shown, not silently rewritten,
  before the cleaned text is inserted.
• Vendor-agnostic core — speech-to-text and correction are independent, swappable roles
  behind two interfaces in packages/core, with one adapter file per vendor.
• Bring-your-own-key, local-first — keys are stored in the OS keychain and sent only to
  the vendor you choose. No audio or transcript content telemetry.

macOS desktop app (milestone M4 — functionally complete)

• Menu-bar app with a floating, non-activating overlay that never steals keyboard focus
  from the app underneath it.
• Focusable Settings window and a first-run onboarding flow for entering provider
  keys.
• Focused-field text injection via the Accessibility API with a clipboard-paste fallback;
  secure/password fields are refused (copied to clipboard instead).
• ⌥Space toggle with both tap-to-toggle and hold-to-talk.
• Rust config store + OS keychain; the app owns and supervises its own backend
  ("sidecar") and injects keys from the Keychain — no key ever crosses the renderer.
• Release sidecar packaging so the backend ships bundled inside the app.

Providers

• PyAI (default) — speech-to-text and correction.
• Deepgram — streaming speech-to-text.
• OpenAI — speech-to-text and correction.
• Anthropic — correction.
• Speech-to-text and correction vendors can be mixed (for example Deepgram STT +
  Anthropic correction).
• Multilingual: non-English dictation routes STT through Deepgram or OpenAI, since PyAI's
  speech-to-text is English-only.

Quality & polish (milestone M5 — in progress; shipped in this tag)

• Reliability: retry-with-backoff, streaming keepalive, and auto-reconnect wiring.
• Custom vocabulary: a user term list injected into correction and used to boost STT.
• Snippet text-expansion.
• Formatting modes: prose / message / code / raw.
• Revert-to-raw undo and a paste-last-result hotkey (both configurable).
• Concurrency guard for overlapping dictations.
• Opt-in, metadata-only telemetry with latency capture (content is never sent; transport
  is not yet enabled).

Web demo

• Vite browser app with a Demo (no mic) mode that replays a real captured sample — the
  full live-transcript, correction-diff, and formatted-output flow with no microphone or key.

Developer experience

• npm-workspaces monorepo: packages/core, apps/widget, apps/backend, apps/web.
• Headless core CLI pipeline (npm run pipeline) with an offline fixture + mock path.
• Vitest unit/integration suite (adapters tested against faithful mock servers, no live
  network) and a Playwright end-to-end demo test.
• CI gates on every pull request: lint, type-check, tests, secret scanning (gitleaks),
  dependency audit (npm audit), and CodeQL.

Known limitations {#known-limitations-v1}

• macOS only. Windows support is planned for a later release.
• No published binary. Install from source; builds are not yet code-signed or notarized.
• Long-input live transcript can occasionally show stitch artifacts from the streaming
  provider; the final (batch-transcribed) output is unaffected.
• Experimental, not verified end-to-end: Meetings ("Granola mode") and command-mode /
  wake-word features are prototypes. Meetings requires a macOS loopback audio device.
• Provider availability and exact model behavior depend on your vendor account and keys.

Upgrade notes

• First release — no migration required. If you had checked out the short-lived v0.1.0 tag,
  re-tag/pull to v1.0.0; v0.1.0 has been removed.
• Copy .env.example to .env for standalone/local development, or enter keys in the
  desktop app's Settings window (stored in the OS keychain).