Skip to content

Latest commit

 

History

History
420 lines (310 loc) · 19.4 KB

CHANGELOG.md

File metadata and controls

420 lines (310 loc) · 19.4 KB

Changelog

All notable changes to LocalFlow are documented here. Earlier entries describe the upstream FreeFlow releases this fork is based on.

This project uses semantic versioning for public releases. Use MAJOR.MINOR.PATCH, where:

  • MAJOR changes include breaking behavior or major compatibility changes.
  • MINOR changes add user-visible features and improvements.
  • PATCH changes fix bugs, polish existing behavior, or make small internal improvements.

[2.1.8] - 2026-10-26

Tenth LocalFlow macOS beta. Requires an Apple Silicon Mac running macOS 26 or later. Uses the same app identity and local data locations as 2.1.7.

Added

  • Granting Accessibility no longer means finding LocalFlow in System Settings yourself. Grant Access opens the Accessibility pane with a card beside it: drag LocalFlow from the card into the list. The card follows the Settings window and confirms once access is granted.

Fixed

  • Installing an in-app update is safer. LocalFlow checks that it can replace itself before downloading, and asks you to move it to Applications when it is running from the DMG or a folder it cannot write to. A failure while swapping in the new version no longer damages the new app, and LocalFlow always reopens. This applies to updates installed by 2.1.8 and later.
  • The menu-bar shortcut menus now disable every choice that would be rejected. Before, some choices stayed enabled and did nothing when picked.

[2.1.7] - 2026-10-21

Ninth LocalFlow macOS beta. Requires an Apple Silicon Mac running macOS 26 or later. Uses the same app identity and local data locations as 2.1.6.

Changed

  • LocalFlow keeps one version of its speech model in memory instead of four, so it uses much less memory while it runs. On Macs with limited memory the model is less likely to be pushed out of memory, which should keep transcription fast.
  • After the first launch, LocalFlow no longer spends several minutes preparing three extra model versions in the background.
  • Recordings longer than 15 seconds are transcribed in 15-second sections instead of 30-second ones. For dictations between 15 and 30 seconds, most of the audio is now transcribed while you are still speaking, so the text arrives sooner after you stop.

[2.1.6] - 2026-10-21

Eighth LocalFlow macOS beta. Requires an Apple Silicon Mac running macOS 26 or later. Uses the same app identity and local data locations as 2.1.5.

Fixed

  • Dictionary suggestions from corrections could fail to appear in browsers and Electron apps such as Chrome and T3 Code. These apps can update a text field a moment after the paste, so LocalFlow now tries again for up to two seconds before giving up on watching it.
  • Text fields that show spaces as no-break spaces, as many web editors do, no longer stop the correction watcher from finding the pasted text.

[2.1.5] - 2026-10-20

Seventh LocalFlow macOS beta. Requires an Apple Silicon Mac running macOS 26 or later. Uses the same app identity and local data locations as 2.1.4.

Added

  • Dictionary words can be limited to apps. In Settings → Dictionary, add an app from the open apps or from Applications, then choose for each word whether it is used in every app or only in chosen apps. A dictation uses the words for every app plus those of the app in front when recording starts. Correction suggestions in an app offer "Add to " and "Add everywhere". Removing an app deletes the words that only it used. The 1000-word limit covers all words together, and the ignore list still applies everywhere. LocalFlow saves the name and bundle ID of each app you add, and nothing else about the apps you dictate into.

Fixed

  • Pressing Escape during transcription and starting a new dictation right away could leave the new dictation unable to be cancelled.
  • A very large ignore list could make the dictionary unreadable at the next launch.

[2.1.4] - 2026-10-17

Sixth LocalFlow macOS beta. Requires an Apple Silicon Mac running macOS 26 or later. Uses the same app identity and local data locations as 2.1.3.

Changed

  • Long dictations finish almost as soon as you stop. Each full 30-second stretch is transcribed while you are still speaking, so the wait after you stop no longer grows with the length of the dictation.
  • Silence is shortened before transcription, so dictations with dead air or long pauses transcribe faster. Silence at the start and end keeps 0.3 seconds, and a pause longer than 0.8 seconds, such as while searching for a word, is shortened to 0.8 seconds. Shorter pauses are untouched. The saved recording is not changed.
  • The speech model is kept in a stable copy in ~/Library/Application Support/com.ajbarryiii.localflow.dev/Models, so later updates that keep the same model no longer prepare it again (about a minute and a half per model size). 2.1.4 itself prepares the model once more. The copy shares disk space with the app until an update replaces the app, after which it uses about 370 MB.

Fixed

  • The Add Word and Edit Word sheet in Settings → Dictionary no longer cuts off its sides.

[2.1.3] - 2026-10-15

Fifth LocalFlow macOS beta. Requires an Apple Silicon Mac running macOS 26 or later. Uses the same app identity and local data locations as 2.1.2.

Added

  • Spoken numbers are written as digits ("twenty three percent" becomes "23%", "fiscal twenty twenty one" becomes "Fiscal 2021"). Zero to nine stay as words except in lists, labels, units and ranges ("1, 2, 3", "step 3", "5 GB", "5 to 15"). Years, times, decimals, money, percentages, ordinals and dates are handled. Turn it off in Settings → Clipboard → "Write spoken numbers as digits".

Changed

  • Correction suggestions are on by default. Turning them off is remembered. They still read only the pasted text and a little around it, for up to a minute, never password fields, and keep that text in memory only.

[2.1.2] - 2026-10-13

Fourth LocalFlow macOS beta. Requires an Apple Silicon Mac running macOS 26 or later. Uses the same app identity and local data locations as 2.1.1.

Added

  • Custom dictionary (Settings → Dictionary). Add names and technical terms that are often misheard; each entry biases recognition toward its words with its own strength, and entries containing capitals are written exactly as entered.
  • Optional correction suggestions, off by default. After a dictation, when you correct a misheard word in the pasted text, LocalFlow re-decodes the same audio with that word boosted and offers to add it to the dictionary only if that reproduces your correction. It reads only the pasted text and a little around it, for up to a minute, never password fields, and keeps that text in memory only.
  • Optional double-tap prompt mode, off by default: tap Hold to Talk, then press and hold it again within 0.4 seconds to dictate with the prompt tag.

Improved

  • Recordings longer than one chunk (15 seconds, or 30 seconds once optimization completes) are transcribed in chunks that overlap by 2 seconds and are joined at a word both chunks agree on. This removes the invented periods, capitals and cut-off words that used to appear where chunks met.
  • Model files are verified with bounded reads instead of loading the whole encoder into memory.

Fixed

  • Temporary recordings from cancelled or failed dictations are deleted instead of being left behind.

[2.1.1] - 2026-10-09

Third LocalFlow macOS beta. Requires an Apple Silicon Mac running macOS 26 or later. Uses the same app identity and local data locations as 2.1.0.

Added

  • Optional Hold to Prompt shortcut that prefixes dictation with an editable tag (default [dictated]) for use with AI agents. Disabled by default; voice macros are never tagged.
  • Speaking pace in Settings: average and latest words per minute, total words, and dictation count, with a reset action. Statistics retain word and time totals rather than transcript text.

Improved

  • Dictation becomes ready after the 15-second model function prepares; the remaining functions optimize in the background while dictation stays available.
  • A refreshed model bundle supports 30-second transcription chunks once optimization completes. Until then, recordings use 15-second chunks.

Fixed

  • In-app updates verify the downloaded app's code signature, developer team, and app identity before installation. Ad hoc and unsigned builds must use a manual download.
  • Failed update downloads or installations now show an alert.

[2.1.0] - 2026-10-06

Second LocalFlow beta. Requires an Apple Silicon Mac running macOS 26 or later.

Upgrading from 2.0.0: the 2.0.0 beta still used FreeFlow's development identifier (com.zachlatta.freeflow.dev). This beta uses LocalFlow's own (com.ajbarryiii.localflow.dev), so macOS treats it as a new app. Grant Microphone and Accessibility again, re-enable Launch at Login if you use it, and re-create any voice macros. Preferences and history from 2.0.0 are left untouched in the old FreeFlow Dev location; they are not shared with or deleted by this version.

Added

  • Spoken quotes, brackets, and code marks. Say "quote … end quote", "paren … close paren", "open bracket … close bracket", "open brace … close brace", "backtick … end backtick", or "double asterisk … close double asterisk" to insert "…", (…), […], {…}, `…`, or **…**. Only matched pairs convert; a lone "quote" or "the paren" stays as spoken.
  • Spoken capitalization. "all caps on … all caps off" uppercases the words in between and "all lowercase … end lowercase" lowercases them. Starting a dictation with "all caps" or "all lowercase", without a closing command, changes the whole dictation.
  • A setting to turn these spoken commands off, under Settings → Clipboard. It is on by default.

Changed

  • LocalFlow uses its own bundle identifiers and stores its data under Application Support/LocalFlow (Application Support/LocalFlow Dev for this beta), so it never shares preferences, permissions, or history with an installed copy of FreeFlow.
  • Rebuilding the app no longer forces the several-minute model preparation when the model itself has not changed.

[2.0.0] - 2026-10-04

This is the first LocalFlow beta. Requires an Apple Silicon Mac running macOS 26 or later. The beta is signed with the maintainer's Developer ID and notarized by Apple for distribution outside the Mac App Store.

Added

  • Bundled LocalFlow English speech recognition, derived from NVIDIA Parakeet v2, with no API key, provider account, or model download required.
  • Model preparation at startup and reuse throughout the session. First-time preparation can take several minutes; Settings shows progress and offers a retry if preparation fails.
  • LocalFlow app and menu-bar icons, with attribution to FreeFlow and the model's original authors included in the application.

Changed

  • Speech recognition runs entirely on the Mac. Cloud transcription, LLM cleanup, translation, Edit Mode, and screenshot/app context capture have been removed.
  • Microphone and Accessibility are the only dictation permissions. Setup guidance reflects macOS 27's Device Control and Data Access permission pane.
  • Hold to Talk, Tap to Toggle, Paste Again, deterministic voice macros, and local history/retry remain available. Existing preferences and local history keep their compatible storage locations.
  • Update checks use the LocalFlow fork's releases. The updater may contact GitHub; audio and transcripts remain local.

Beta limitations

  • English dictation only; Intel Macs and macOS versions earlier than 26 are not supported by this model bundle.
  • Dictations longer than 15 seconds use independent chunks and can lose context at chunk boundaries. There is no LLM cleanup or cloud fallback.
  • Live microphone dictation and paste worked in the preceding local build; the packaged beta still needs tester verification of shortcuts, permissions, and installation on other Macs.

[1.2.1] - 2026-08-11

Improved

  • Model pickers now show only currently supported Groq options, with context models limited to those that can analyze screenshots.

Fixed

  • Fixed typing lag in custom vocabulary and prompt editors by saving changes when editing finishes instead of after every keystroke.

[1.2.0] - 2026-07-14

Added

  • A Preserve exact wording option that skips transcript cleanup while still supporting literal translation, voice macros, and Edit Mode.
  • An instruction guard that retries or falls back to the literal transcript when cleanup appears to answer a dictated prompt instead of preserving it.
  • An option to keep dictations in clipboard-manager history.
  • A menu bar action for adding copied words directly to the custom vocabulary while avoiding duplicates.

Improved

  • Post-processing now switches to the fallback model immediately when the primary model is rate-limited, remembers daily limits across restarts, and shows their reset times in Settings.
  • The default fallback model now uses Qwen 3.6 27B, and models scheduled for Groq shutdown in July or August 2026 have been removed from the picker.
  • Transcription failures now distinguish network outages from slow providers and display clearer errors in the recording overlay.
  • The default context model now uses Qwen 3.6 27B and strips reasoning tags from context summaries.
  • Local model requests now honor their configured timeouts, and transcription uses response formats compatible with a wider range of OpenAI-style models.
  • Permission and update checks use fewer unnecessary background wakeups.

Fixed

  • Fixed a recording overlay resource leak that could leave hidden animations running after the overlay closed.

[1.1.0] - 2026-06-03

Added

  • Model pickers in Settings for post-processing, fallback, context, and transcription models, including Qwen 3 32B and custom model entries.
  • A recording overlay display picker for choosing the active window, primary display, or a specific connected monitor.
  • In-pill error notifications so transient failures such as network or provider errors are visible without opening logs.
  • Advanced timeout overrides for local model and slow network setups.

Improved

  • Retried dictations now place the successful transcript on the clipboard and update Paste Again.
  • Paste Again now preserves the latest raw transcript earlier in the dictation flow, so it remains useful if later cleanup or pasting fails.
  • Post-processing handles reasoning-oriented model output more cleanly, including Qwen thinking tags and providerless model aliases.

Fixed

  • Fixed cases where transcription could hang indefinitely when a provider accepted a connection but never returned a response.
  • Fixed false screen-recording permission alerts from unrelated permission messages.
  • Fixed duplicate in-pill error notifications being dismissed by an older timer.

[1.0.0] - 2026-05-20

FreeFlow is now considered feature-complete and stable enough for a 1.0 release.

Added

  • Paste Again shortcut for re-pasting the most recent dictation.
  • Recent transcript history in the menu bar, with copy actions for quickly reusing previous dictations.
  • Run Log copy controls for both literal and cleaned transcript output.
  • Menu bar actions for opening the Run Log and checking for updates.
  • Debug settings for troubleshooting overlays and update prompts.
  • A polished drag-to-Applications DMG background for installer builds.

Improved

  • Recording feedback now uses a cleaner minimalist menu-bar overlay, with clearer command-mode state.
  • Transcribing and processing feedback appears sooner and more consistently after recording stops.
  • Shortcut labels now use friendlier modifier names alongside symbols.
  • Setup and recovery flows are more resilient when restoring app state.
  • Sentence-ending dictations now paste with trailing spacing that better matches normal writing.
  • Development builds and main-branch release automation are easier to identify and validate.

Fixed

  • Fixed shortcut collision checks for edit mode and manual modifier bindings.
  • Fixed cases where dictation could terminate automatically while still in progress.
  • Fixed clipboard restoration after dictation when the original clipboard content is unchanged.
  • Marked transient dictation clipboard contents so clipboard managers can avoid saving them.
  • Preserved spoken instructions verbatim during post-processing.
  • Simplified transcription submission errors into clearer one-line messages.

[0.3.3] - 2026-04-25

Added

  • Output Language setting for automatically translating dictated text before it is pasted.
  • Transcription Language setting for choosing the language FreeFlow listens for during dictation.
  • Recording state flag file for external tools that need to know when FreeFlow is actively recording.
  • Distinct FreeFlow Dev app and menu bar icons so development builds are easier to tell apart from release builds.

Improved

  • Permission prompts and setup screens now use the correct app name for the installed build.
  • Release notes in update prompts now render changelog formatting more clearly.
  • Development builds now have clearer bundle naming and icon handling.

Fixed

  • Fixed audio recording crashes caused by unexpected input formats, resampling, and upload-path conversion.
  • Fixed cases where FreeFlow could silently fall back when the selected microphone was unavailable.
  • Fixed paste shortcuts on Colemak-DH and other non-QWERTY keyboard layouts.
  • Fixed output language handling when custom system prompts are enabled.

[0.3.2] - 2026-04-23

Fixed

  • Removed the pause-based audio interruption mode that could misfire and resume playback unexpectedly; dictation now only mutes audio.

[0.3.1] - 2026-04-23

Added

  • Faster live dictation with realtime transcription support.
  • A setting for choosing the realtime transcription model.
  • Run log exports, so you can save a full dictation run for debugging or sharing.
  • A Copy Transcript action in the run log.
  • A voice command for submitting text: say "press enter" at the end of a dictation.
  • Audio controls that can mute or pause other audio while you dictate, then restore it when recording stops.
  • Build details in Settings for easier troubleshooting.
  • Direct shortcuts from FreeFlow to the right macOS permission settings.
  • A What’s New popup when an update is available.

Improved

  • Recording feedback now feels more responsive.
  • The run log is easier to scan and use.
  • Exported run logs include more useful context for reproducing issues.
  • Realtime transcription is more reliable when recordings are cancelled, retried, or finish with no text.
  • Provider settings are easier to edit without accidental whitespace or half-saved values.
  • FreeFlow now warns you if alert sounds may be hard to hear because system audio is muted or very low.
  • Update prompts now show the version, release date, and release notes more clearly.
  • FreeFlow now uses proper version numbers for updates instead of internal build names.

Fixed

  • Fixed cases where arrow or navigation keys could be mistaken for Fn shortcut input.
  • Fixed a clipboard timing issue that could paste the wrong content.
  • Fixed empty realtime transcriptions getting stuck instead of finishing cleanly.
  • Fixed waveform glitches caused by invalid audio levels.
  • Filtered out more common transcription artifacts.
  • Fixed alert sound hints staying visible after alert sounds are turned off.
  • Fixed update checks so users only see real app releases, not internal builds.
  • Fixed update checks so the app does not offer an older or already-installed version.