Releases

Changed

• The leaderboard now labels its figure "time back" instead of "time saved", matching Home, onboarding, and the manual. The number is how long the same words would have taken to type at 40 wpm; nothing subtracts the time dictating them actually took, so calling it a saving overstated it. The number itself is unchanged, as are the minutesSaved field and minutes_saved column that already-shipped clients depend on.

Changed

• The leaderboard now draws a fixed seven rows whatever the standings return, so the height of the card no longer reveals how many people use OpenVoiceFlow. Rows the service doesn't name are masked — a shimmering placeholder with no rank, name, or total — and the tail fades out rather than ending on a countable last row. The card states no rule and shows no threshold: naming the bar would let anyone counting the visible names work out what population they were counting.

Fixed

• Public leaderboard standings now include only rows with at least 60 minutes saved and show no more than five entries, while still showing your own row separately.
• Legacy auto-generated leaderboard names now use the same exact 10–99 suffix range as the native app, preserving custom-looking names ending in 00–09.

Added

• App names across Home, History, and Personalize now use recognizable app or company logos, including Claude, Discord, Notion, and Outlook. Installed macOS app icons remain the first choice, with bundled brand marks as reliable fallbacks.

Changed

• Leaderboard nicknames now save when Return is pressed or the field loses focus, immediately publish that installation's aggregate totals, and refresh the standings after a successful update. Each installation keeps its own row even when multiple computers use the same nickname.
• Opening Leaderboard now republishes that installation's saved aggregate totals before fetching standings, restoring existing local usage without requiring another dictation or nickname change.
• The leaderboard service now uses Neon's supported serverless database client and accepts both current and legacy Vercel database environment-variable names.

Fixed

• History's Copy action now changes to “Copied” with a checkmark after a successful copy and resets automatically after 1.5 seconds.
• Leaderboard loading and nickname-sync failures now show an honest retry action instead of appearing as empty standings.

Fixed

• Personalized Styles now shows recognizable Discord and Gmail icons, using installed app icons first and bundled brand marks when those apps are unavailable.
• Clicking the dashboard outside the centered Feedback sheet now dismisses it, while clicks inside the sheet or in unrelated OpenVoiceFlow windows keep their normal behavior.

Fixed

• Restored Feedback as a centered native modal instead of a left-anchored popover, while retaining the always-visible optional email field and Cancel, Send, and Escape dismissal.

Changed

• Simplified feedback to one optional email field in a native macOS popover that dismisses with Cancel, Send, Escape, or an outside click.
• Moved Know Me into Personalize so related controls stay together.
• Changed generated leaderboard aliases to compact one-token handles while preserving custom names during legacy migration.

Changed

• Settings no longer exposes the anonymous analytics device ID or the long leaderboard-sharing explanation, and the Feedback sheet no longer shows its usage-snapshot explainer. The sharing toggle, leaderboard name, deletion action, and feedback behavior are unchanged.

Changed

• “Where you dictate” now shows each app’s real icon with a time-share ring, making the per-app breakdown easier to scan while preserving the existing local-only usage data and accessibility labels.

Changed

• The History pane now carries a subtitle: "Stored only on this Mac — never synced to a server." Scoped deliberately to what's true unconditionally (this list is local storage, full stop) rather than a blanket "nothing leaves this machine" claim, since cloud AI cleanup (opt-in, off by default) does send dictated text to the selected provider before it lands here — that's covered separately by the existing footer line and the privacy docs.

Changed

• Dictionary, Snippets, and Styles are now one Personalize pane in the dashboard sidebar, switched with tabs instead of three separate rows. Each tab keeps its item count, the content area now sits in a single card, and switching tabs animates instead of jumping. No data or behavior changes — same stores, same add/remove actions, same empty states. Docs updated to describe the new location (Dashboard ▸ Personalize).

Fixed

• The Feedback sheet's category picker ("Something's broken" / "Feature idea" / "Just saying thanks" / "Something else") overflowed a 440pt-wide sheet, truncating the leading and trailing segments. Segmented control now shows short labels (Bug/Idea/Thanks/Other); the full wording still goes into the email subject and body. Sheet widened to 480pt for margin.

Added

• Anonymous usage sharing and an in-app Leaderboard, ranked by time saved. On by default — a real, disclosed change from every prior release, which sent nothing. Settings ▸ Privacy ▸ "Share anonymous usage & leaderboard rank" turns it off, deletes nothing retroactively but stops every future request immediately; a "Delete my leaderboard data" button removes a past submission outright. What's sent: a random per-device ID, a display name you can change, aggregate word/time/streak/feature-usage counters already shown on Home, and a server-derived country (no IP stored). Never dictation text, snippets, dictionary entries, or the Know-Me profile. Full detail in PRIVACY.md §7 and the docs site's Privacy architecture ▸ Analytics & leaderboard section.

Changed

• Docs and PRIVACY.md updated to describe the above plainly, replacing the "no telemetry" claims that were accurate through 0.5.7 and are not anymore.

Added

• A Feedback item in the sidebar, below Settings. Opens a short form (bug / idea / praise / other, plus an optional reply email) and, only on Send, opens a mailto: with the message and a small aggregate usage snapshot (words dictated, time saved, streak, last 7 days, first-use date) — never dictation text, snippets, dictionary entries, or the Know-Me profile, and nothing is sent until Send is pressed.

Added

• The language picker now offers all 99 languages in Whisper's multilingual set (previously eleven), including Ukrainian. Tiny, Small, Medium, and Large turbo all share the same tokenizer, so the list applies to every engine offered in Settings.

Fixed

• The progress bar no longer parks at 100% with no way forward. WhisperKit compiles the model for your chip after the download — minutes on a first large-model run, with no progress signal — so the transfer now owns 0–90% and the compile creeps through the 90s under "Downloaded — now optimizing for your Mac. First run only." 100% arrives only with the Try It button.

Changed

• RECOMMENDED now adapts to the Mac: Large turbo on Apple Silicon with disk headroom, Small on Intel or a tight disk. No permissions involved.

Fixed

• The download meter showed "0 of 0 MB · 0.0 MB/s" over a moving bar: WhisperKit reports abstract progress units for multi-file model downloads, not bytes. The meter now shows percent (and the always-honest ETA) when units can't be bytes, instead of inventing megabytes.

Fixed

• Large turbo could never download. Its model id didn't exist in the WhisperKit repo; every user who chose it hit "no models found" disguised as a connection error. Correct id shipped, saved settings migrate silently.
• Browsing the engine chooser no longer starts a download per click: a 1.5 s grace window ("Starting in a moment — switch engines freely") starts the transfer only once the choice rests, and switching cancels the superseded download outright. Model loads are single-flight, so overlapping transfers can no longer corrupt each other's cache.
• Download failure copy points at Details instead of guessing "check your connection" — the guess was wrong exactly when it mattered.

Added

• Live transcript in the HUD while the key is held ("Show words as you speak", default on; respects the text-echo privacy setting).
• Speech-engine choice in onboarding — four models with size and benefit, nothing preselected, nothing downloading until you pick.
• Launch at login (SMAppService), default on, with a Settings toggle.
• Onboarding's "I live up there" now anchors a callout to the real menu-bar icon, with a mock menu-bar illustration when the icon is squeezed out.
• Dashboard banner explaining when macOS hides the menu-bar icon (full bar / notch) and the ⌘-drag fix.

Changed

• Onboarding ends in the dashboard instead of an empty desktop; the try-it card acknowledges key-down instantly.
• The "names I'd get wrong" question moved out of onboarding (Know-Me keeps it).

Fixed

• The hotkey now works from launch. Previously nothing started the key listener after first run — every relaunch came up deaf until Start Dictation was clicked in the menu. The Start/Stop menu item is gone (Pause remains); a Turn On Dictation button appears only if a permission was revoked.

Added

• PERMISSIONS card in dashboard Settings with live status per grant; re-granting revives dictation without a relaunch.

Fixed

• Dock presence obeys the setting; fn is the default hotkey; onboarding gives real feedback while permissions land.

Added

• Automatic background updates (Sparkle) with honest in-app update cues.
• Per-app dictation breakdown ("Where you dictate").

Fixed

• Failed pastes recover; the transcription model hot-swaps without restart.

Fixed — Fn / 🌐 Globe key no longer fails silently

• Fn was never a working hotkey. macOS doesn't deliver the Fn/Globe key to pynput, so selecting it left dictation permanently dead with no feedback. Choosing Fn now surfaces an immediate modal explaining the limitation and pointing at Right Command (the default), the menu-bar picker hides Fn unless it's the current (legacy) selection, and onboarding no longer offers it.

Fixed — production-hardening pass (end-to-end audit follow-up)

• Hotkey can't be wedged by a consent dialog. The microphone is now armed on the event-tap thread and the ~150 ms of osascript context capture runs on a worker, so a stalled Automation/Accessibility dialog can no longer block (and let macOS disable) the global hotkey.
• Runaway recordings self-terminate. A max-duration watchdog force-stops a dictation whose key-release was lost, instead of leaving the mic and whisper-stream running indefinitely.
• No PortAudio stream leak. recorder.stop() always releases the audio stream even when the input device is unplugged mid-recording; that failure is now surfaced instead of silently freezing the overlay.
• Clipboard data-loss fixed. Selected-text capture no longer overwrites a non-text clipboard (image/file) and never restores empty text over the selection.
• Security defense-in-depth. Secrets are written with O_EXCL|O_NOFOLLOW (symlink-swap safe); ~/.openvoiceflow/logs/ is 0700; LLM/Ollama responses are read with a 16 MB cap; ollama_url rejects non-HTTP schemes and warns when it points off-box. Hung pbcopy children are reaped, not orphaned.
• Upgrades actually reinstall dependencies. The DMG venv marker is now version-stamped, so a release that adds or bumps a dependency reinstalls for existing upgraders instead of breaking on import.
• Overlay polish. HUD timers run in the common run-loop mode (no freeze during menu tracking), fades honor Reduce Motion, and the window is ordered out after fading so it doesn't linger in every Space.
• Know-Me interview no longer self-destructs when Escape is pressed while a text field is focused.

Fixed — first launch could silently do nothing (no menu bar icon, no wizard)

• Stale bootstrap lock. ~/.openvoiceflow/.bootstrap.lock survives reboots, and its stored pid could be reused by an unrelated live process — the alive check then matched, so every launch exited silently forever. The lock now only counts when the pid is alive *and* is really our launcher.sh.
• Command Line Tools stub. command -v python3 always succeeds on macOS (Apple ships an installer stub), so a Mac without the Command Line Tools got past the check and failed later with no message. First launch now probes xcode-select, triggers the CLT installer, and shows instructions.
• Broken venv skipped as "installed". A macOS/CLT update can break the venv's interpreter or native wheels while the .ovf-deps-installed marker stays present, so the bootstrap was skipped and every launch died on import. The launcher now health-checks the venv (imports numpy/sounddevice/pynput/ rumps/objc) and rebuilds it when the check fails.
• Silent progress + silent fallbacks. First launch now shows an up-front dialog explaining the ~5-minute setup and where the menu bar icon appears (including the 14"/16" notch tip); the venv/pip steps fail with a visible dialog instead of exiting quietly; a skipped setup wizard and a menu-bar (rumps) failure now say so instead of dropping to an invisible background process.
• Input Monitoring permission (native launcher). The native launcher now proactively requests Input Monitoring (IOHIDRequestAccess) at first launch, in addition to Microphone and Accessibility, so the grant is in place before the dead-listener watchdog would ever need to fire.

Fixed

• Startup now checks the Input Monitoring permission — the one the global hotkey listener actually needs. It was never verified (only the doctor knew about it), so a missing grant produced a "Ready" menu with a hotkey that silently never fired. A denied probe warns loudly with a one-click System Settings link but never blocks startup, because Accessibility trust can also grant listen access.
• A new dead-listener watchdog catches the same failure by measuring reality: if no key events at all arrive within 15 seconds of the listener starting *and* macOS reports Input Monitoring denied, the menu bar shows a front-and-center alert with the fix. An idle user without the permission problem gets a one-time gentle tip instead. (Previously this self-check existed only for the fn/Globe key, never for the default Right Command hotkey.)
• Setup failures at launch now surface as a modal alert with a one-click settings link instead of only a Notification Center banner — notifications require a permission of their own and the status item can hide behind a MacBook notch, so the old failure path could be entirely invisible.
• A corrupt or wrong-shaped file in ~/.openvoiceflow/ (snippets, stats, profile, dictionary aliases, transcript logs, seen-tips) can no longer make every dictation fail with a generic error: all user-data loaders now validate shape, drop malformed entries, and degrade to safe defaults. --show-profile prints recovery guidance instead of a literal null when the profile file is corrupt.
• Every osascript/launchctl/open subprocess call now has a timeout, so a pending macOS consent dialog or a wedged launchd can no longer freeze the hotkey thread permanently. If auto-paste times out, a notification explains that the text is already on the clipboard, ready for ⌘V.
• Auto-learn no longer records ordinary content edits (e.g. "june" → "july") as permanent corrections: the word-similarity threshold was raised from 0.4 to 0.55, calibrated so genuine mishearings ("mir" → "Meer", "recieve" → "receive") are still learned.
• Setup wizard: an llm_backend of none no longer breaks the backend screen; switching backends clears the previous backend's prefilled API key instead of finishing without saving a key; a failed final config write now shows an error dialog instead of a Finish button that silently does nothing.
• Know Me interview: no longer clobbers a configured --style (the style radio preselects from config, and styles the interview can't express, like code or email, are preserved unless actively changed); Enter-key bindings no longer leak between screens; context-menu paste into multi-line fields is no longer dropped; a hand-edited profile can't crash the wizard.
• Menu bar: notifications fall back to the osascript path when rumps.notification is unavailable (non-framework Python installs), and Check for Updates… can no longer get stuck on unexpected network/response errors. The duplicate startup update check in menu-bar mode was removed.
• Overlay: the compact auto-learn pill no longer permanently shrinks the font of every subsequent HUD message.
• CLI: value-taking flags now reject empty strings (--search "" used to launch the app; --add-command "" corrupted every dictation); config setters combined with an action flag now print a warning instead of being silently ignored; --auto-learn, --update-check, and --log-transcripts now compose on one command line.
• Doctor: macOS version detection resolves the "10.16" Big Sur compatibility shim via sysctl kern.osproductversion and reports "unknown" (a WARN) rather than false-failing supported systems; bare-major versions ("12") no longer compare below the (12, 0) minimum.

Added

• CI now compiles the native launcher (clang -fsyntax-only) and runs a functional test suite (tests/test_launcher_flows.py) that renders the real launcher.sh and exercises the stale-lock, missing-CLT, broken-venv, and first-run-visibility paths with shimmed macOS commands.
• The menu-bar app now shows a Dock icon while running (toggleable via Advanced → Show in Dock, default on). A menu-bar-only icon can hide behind a MacBook notch, leaving no visible sign the app is running; clicking the Dock icon opens the native status summary with the active dictation shortcut.
• 17 regression tests pinning the defensive-loader, learner-threshold, updater, and macOS-version behaviors (tests/test_defensive_loaders.py), plus 12 pinning the Input Monitoring gate, dead-listener watchdog, and Dock default (tests/test_visibility_and_watchdog.py).

Changed

• CI lint now covers the entire repository (ruff check .), not just voiceflow/.

Added

• A polished native menu-bar experience led by Open OpenVoiceFlow, with grouped Dictation Shortcut, AI Cleanup, Writing Style, Personalization, and Advanced menus.
• A user-triggered Check for Updates… action that always reports whether a release is available, the app is current, or the service could not be reached.
• Direct menu links to macOS Microphone and Accessibility settings.

Changed

• The status item now uses dark-mode-safe SF Symbols with accessible labels: a waveform while ready, a pause symbol while paused, and a warning symbol when setup needs attention. A visible OpenVoiceFlow text label remains as the fallback if symbols are unavailable.
• Menu selections now use native macOS checkmarks, clean human-readable names, every supported dictation shortcut, standard separators, and ⌘Q for Quit OpenVoiceFlow instead of emoji-prefixed labels.
• The long-lived Python UI process now stays out of the Dock and uses the bundled OpenVoiceFlow icon for native dialogs. The Current App row refreshes as focus changes without mistaking the UI process for the user's app.

Fixed

• The waveform status item is created before backend validation begins, so a slow local backend no longer makes launch look unresponsive.
• Edit Profile now runs its Tk interview in an isolated process instead of risking a native Tk crash in the menu-bar process.
• The Download button in the update alert now recognizes AppKit's native button result and opens the release page correctly.

Added

• The menu-bar menu now includes a persistent How to Use guide, and the first-run hotkey tip also appears in the floating HUD instead of relying only on a potentially hidden macOS notification.
• voiceflow/platform_support.py: one place for OS, macOS-version, architecture (Apple Silicon / Intel / Rosetta), and permission detection.
• CLI platform gate: on Linux/Windows, openvoiceflow now prints a clear "macOS-only" explanation with uninstall guidance and exits cleanly instead of crashing with a traceback. --doctor and --show-config keep working so users can inspect state first.
• New doctor checks: operating system + version, architecture (warns on Rosetta-translated Python and Intel Homebrew on Apple Silicon), and the three macOS permissions dictation depends on (Microphone, Accessibility, Input Monitoring) with click-to-fix System Settings links.
• A launch below the supported macOS floor (12 Monterey) now prints an upgrade warning.
• CI: new non-macos-guard job on ubuntu-latest pins the "friendly message, never a traceback" guarantee and runs the full suite on Linux.
• Website: the download page now detects the visitor's OS. Windows, Linux, ChromeOS, Android, and iPhone/iPad visitors get an inert "Not available for <OS>" notice instead of a live DMG button; Safari-on-Mac visitors get an Apple Silicon / Intel recommendation via a WebGL renderer fallback (Chromium browsers already used architecture hints).

Fixed

• Short streaming dictations no longer lose the final transcript fragment when whisper-stream exits without a trailing newline.
• Fresh installs now use the reliable batch recorder by default. Experimental real-time streaming remains available through the menu bar or --streaming on. Existing v0.3.2 installs are reset to batch mode once and can opt back into streaming afterward.
• voiceflow.recorder no longer imports sounddevice at module load — the import crashed the whole app (OSError: PortAudio library not found) on machines without PortAudio before any error handling could run.
• find_whisper_cpp / the doctor's Homebrew check use shutil.which instead of spawning which, which crashed on Windows.
• The keyboard-listener (pynput) import is now guarded in CLI and menu-bar modes: a missing input backend produces a clear error instead of a crash.
• Clipboard/keystroke/sound helpers (pbcopy, pbpaste, osascript, afplay) no longer propagate FileNotFoundError when the binaries are missing; failures surface as user-visible notifications.
• --autostart refuses politely off-macOS instead of writing a ~/Library/LaunchAgents folder on Linux; on macOS it now prefers the supported launchctl bootstrap/bootout verbs, falling back to the deprecated load/unload.
• validate_setup reports a broken audio backend (OSError) instead of only a missing sounddevice package.

Fixed

• The DMG bootstrap now executes the virtual-environment Python interpreter at its original path, preventing a silent dynamic-library crash on launch.
• Native startup failures now show an alert with a direct link to the launcher log instead of making the app appear to do nothing.
• First-run Tk onboarding runs in an isolated process, so an incompatible system Tk build falls back to local transcription without killing the app.
• Menu-bar settings no longer clear uninitialized native submenus during startup.

Fixed

• The notarized macOS app now carries the hardened-runtime entitlements for microphone input and Apple Events, allowing recording and auto-paste to pass macOS privacy enforcement.
• DMGs now use a small native launcher so macOS attributes Microphone and Accessibility consent to OpenVoiceFlow instead of its Python bootstrap.
• First launch uses the native macOS consent prompts and no longer stacks two conflicting custom permission dialogs.

Added

• Pytest scaffold under tests/ with regression tests for every Wave-1 ship-stopper (test_python39_compat, test_install_sh, test_config_migration, test_onboarding, test_voice_commands_count, test_updater) plus privacy invariants (test_chmod_600, test_privacy_defaults) and package smoke tests. 60 tests total, green on Python 3.9, 3.10, and 3.11.
• [dev] extras in pyproject.toml (pytest, pytest-cov, ruff, build, twine).
• CI Python matrix: 3.9, 3.10, 3.11 on macos-latest. Lint is now a blocking step. A separate build job exercises python -m build and twine check.
• --update-check on/off CLI flag and matching update_check config key for opting out of the daily GitHub-Releases ping.
• --log-transcripts on/off CLI flag for explicit control of the ~/.openvoiceflow/logs/ daily files.
• "Privacy at a glance" panel in the README that maps every on-disk file and every network egress in one table.
• voiceflow/_secure_io.py — secure_write_json() and secure_chmod() helpers that all config / profile / dictionary / snippets / stats / daily-log writes now go through, enforcing mode 600.
• Community-health docs: SECURITY, PRIVACY, CONTRIBUTING, CODE_OF_CONDUCT, SUPPORT, VERSIONING, AGENTS, plus docs/COMPLIANCE, docs/COMPATIBILITY, docs/THREAT_MODEL, docs/ARCHITECTURE, docs/legal/DPA-template, docs/legal/THIRD_PARTY_NOTICES. ~14 files, ~2,100 lines.
• RELEASE.md — maintainer's three-command playbook plus when-things-go-wrong table.
• .github/workflows/release.yml rebuilt: tag-driven, Trusted Publisher PyPI publish on non-pre-release tags, split arm64/x86_64 DMG attach, verify-version gate that fails the release if tag ↔ pyproject ↔ __version__ disagree.

Changed

• Config-key migration: cleanup_prompt → llm_prompt. Existing ~/.openvoiceflow/config.json files are migrated transparently on load; the rename was the v0.2.0 schema change that previously dropped users' custom prompts on upgrade. (Fixes SS5.)
• Privacy defaults flipped to off for fresh installs. New users get log_transcripts: False and auto_learn: False until they opt in via the Know Me onboarding interview. Existing users keep their setting — the migration logic only applies the new defaults when no config exists.
• install.sh shim now execs the pip-installed openvoiceflow console script directly instead of python3 -m openvoiceflow (which never existed — the module is voiceflow). (Fixes SS2.)
• Lint is now blocking in CI. Previously ruff check was advisory; the 43-finding backlog meant noise floor only ever climbed.
• README hero rewritten so the privacy framing matches the actual data flow (audio is local; transcripts go to whatever LLM backend you pick; default Gemini is cloud, Ollama is local).

Fixed

• Code-review pass (2026-07) — full-codebase review; the notable fixes: - Streaming mode + snippet match no longer crashes with a NameError (the dictation was silently lost with only an error sound). - OpenAI, Anthropic, and Groq backends now honor per-app style and app context (previously silently dropped — the per-app style feature only worked on OpenRouter/Ollama). - Empty LLM responses fall back to the raw transcript instead of pasting nothing. - Custom voice commands containing backslashes no longer crash every dictation; command expansions are no longer re-processed by later commands (single-pass replacement). - Snippet triggers only match on word boundaries — a trigger like sig no longer swallows a dictation starting with "significant…". - Menu bar: LLM Backend / Hotkey / Style submenus are now populated at startup (previously empty until first use); Streaming/Auto-Style/ Auto-Learn toggles now apply to the running session; stopping or quitting aborts an in-flight recording instead of orphaning the whisper-stream process (hot mic). - Corrupt config.json no longer bricks every CLI entry point — the bad file is preserved as config.json.corrupt and defaults are restored. - Whisper model downloads use curl --fail with a temp file, so an HTTP error page or interrupted transfer can't be mistaken for a valid model forever after (fixed in transcriber.py, install.sh, and the DMG launcher). - Batch transcription timeout raised 30 s → 300 s so long dictations aren't discarded; Ollama cleanup timeout raised 30 s → 120 s for cold model loads. - Mic failures during recording start are surfaced (notification + error sound) instead of being silently swallowed with stale state. - Overlay no longer sticks on screen after too-short/no-speech/error aborts. - Selected-text capture no longer erases image/file clipboards when nothing was selected; pbpaste/pbcopy calls have timeouts. - --show-config no longer masks hotkey (over-broad secret matching); --streaming-step validates its value and 0 is no longer ignored; --language no longer crashes on a null whisper_model; env-var-only API-key setups aren't forced through onboarding on every launch. - Onboarding: re-running the wizard and switching backends no longer saves the old backend's key under the new backend's field; the Finish button survives a corrupt config; headless runs fail with instructions instead of a traceback (also fixed for --profile). - Interview: pressing Escape on the final screen no longer mislabels a saved profile as skipped. - --autostart on fails with a clear message when the executable can't be resolved (previously installed a LaunchAgent that silently never launched); the plist is generated with plistlib so paths containing & can't produce invalid XML. - install.sh works under curl | bash (wizard prompts read from /dev/tty), installs from the script's own directory instead of the CWD, creates ~/.zshrc when no shell rc exists, and uses set -euo pipefail. - Release workflow: DMG job sequenced after the PyPI job so the two release-upload steps can't race creating the same GitHub Release.

Changed (code-review pass)

• The typed 🎙 recording indicator is now opt-in ("recording_indicator": true): it edits the frontmost document and could delete a user character when focus changed mid-dictation; the overlay HUD remains the default recording feedback.
• paste_text no longer moves the caret to end-of-line before pasting — text is pasted at the cursor, as documented.
• --set-key BACKEND - reads the key from stdin so it stays out of shell history and ps output (used by install.sh, which also hides key input).

Security (code-review pass)

• Config/profile/dictionary writes are atomic and created with mode 600 from the first byte (previously created 644 then chmod'd — a crash mid-write could truncate the file or leave secrets world-readable).
• ~/.openvoiceflow/ is chmod 700; LaunchAgent stdout/stderr logs are pre-created with mode 600 (they capture dictated text via stdout).
• Update-notification strings from the GitHub API are escaped before AppleScript interpolation (a crafted release tag/URL could otherwise break out of the string literal).
• Internal docs/superpowers/ documents removed from the repository, and the website build excludes any such directory as defense-in-depth.
• SS2 — install.sh shim no longer crashes with ModuleNotFoundError: No module named 'openvoiceflow'.
• SS3 — from __future__ import annotations added to all 30 modules using X | None syntax. pip install . and openvoiceflow --version now work on Python 3.9 (macOS 12 default) in addition to 3.10 / 3.11.
• SS5 — Silent loss of custom cleanup_prompt on v0.1 → v0.2 upgrade. Migration now preserves the user's prompt under the new key.
• SS6 — Onboarding "Personalize OpenVoiceFlow" button no longer silently swallows every exception from interview.run_interview(). Errors surface with a logged traceback and an in-app message.
• The cargo-culted PyObjC import in overlay.py that broke the module for anyone installing without the [overlay] extra.
• 43 lint findings (37 unused imports, 4 placeholder-less f-strings, 1 multiple-statements, 1 unused variable). All resolved; lint now gates CI.

Security

• All ~/.openvoiceflow/*.json files (config, profile, dictionary, snippets, stats) and the daily transcript logs are written with mode 600 — owner-only — instead of inheriting umask 022. API keys and personal data are no longer world-readable on default macOS user accounts.
• Lint gating in CI closes the "style debt grew silently" loop.
• Twelve community-health files added so a procurement reviewer can find the answers they need without filing an issue: SECURITY.md, PRIVACY.md, THREAT_MODEL.md, COMPLIANCE.md, CODE_OF_CONDUCT.md, CONTRIBUTING.md, SUPPORT.md, CHANGELOG.md (this file), VERSIONING.md, COMPATIBILITY.md, legal/DPA-template.md, legal/THIRD_PARTY_NOTICES.md.

Added

• Know Me — tkinter onboarding interview. Captures name, occupation, industry, frequently-mentioned names/tools, communication style. Output is injected into every LLM cleanup prompt as system context.
• Auto-learn — watches the focused text field via the macOS Accessibility API for 30 seconds after every paste and learns word-level substitutions from your edits. Levenshtein-gated, substitutions only (never inserts/deletes), 5-sample minimum.
• Streaming transcription via whisper-stream with refinement-replacement and Jaccard-overlap deduplication. Falls back to non-streaming for the canonical text.
• Per-app context + per-app styles — frontmost-app detection via Apple events; the cleanup prompt picks up the right style automatically (e.g. terse in Slack, formal in Mail).
• Voice commands — say "new line", "period", "delete that", etc. Configurable via --add-command / --remove-command / --list-commands / --voice-commands on/off.
• Clipboard / selected-text context — whatever's selected when you start dictating gets handed to the LLM as context.
• History search — --search QUERY greps the daily transcript logs.
• Voice snippets — trigger-phrase → expansion. --add-snippet / --remove-snippet / --list-snippets.
• Personal dictionary — proper-noun and acronym table. --add-word / --remove-word / --list-words.
• Launch-at-login — --autostart on/off; writes a LaunchAgent plist.
• Statistics — --stats shows dictations, words, time saved.
• Auto-update notifier — once-a-day check against the GitHub Releases API; quiet on failure.
• Floating macOS overlay HUD — visual feedback during recording, streaming, and auto-learn moments.
• New CLI flags across the board: --version, --show-config, --set-prompt, --clear-prompt, --set-key, --app-style, --remove-app-style, --list-app-styles, --auto-style on/off, --streaming on/off, --auto-learn on/off.

Changed

• DMG installer split into separate arm64 and x86_64 artifacts — one-click install on any Mac, no Rosetta dance.
• pyproject.toml gained an [overlay] extra for the optional PyObjC dependencies (Cocoa, Quartz). rumps (the menubar lib) transitively installs Cocoa, so menubar users get the HUD by default.
• Codebase grew from ~13 modules to ~28. Backend cleanup() interface now accepts context, app_context, override_style; all 5 existing backends were updated consistently.
• Anthropic default model: claude-3-5-haiku-20241022.
• Groq default model: llama-3.1-8b-instant.

Fixed

• 22 bugs from a deep QA audit (#3): debounce on the recording hotkey, auto-paste timing window, error reporting in system.py, and a long tail of small UX papercuts.

Fixed

• Rosetta + macOS compatibility. The launcher now smart-detects architecture (Apple Silicon vs Intel) and bootstraps the right venv.
• First-launch dependency installation: brew, whisper-cpp, the ggml-*.bin model, and the Python venv now install on demand instead of failing if anything is missing.
• DMG install path on Intel Macs.
• Several broken GitHub URLs across the README and install scripts.

Added

• 13-module Python package (voiceflow/).
• Local transcription via whisper.cpp. Audio never leaves the Mac.
• Five LLM cleanup backends: Gemini, Groq, OpenAI, Anthropic, Ollama, plus a "none" pass-through.
• tkinter onboarding wizard for first-run setup.
• DMG installer (universal, with Rosetta fallback on Intel).
• Hotkey-driven dictation with auto-paste at the cursor.
• Configurable prompt, model, hotkey, language, and style.
• MIT licensed.