Saral v1.2.5

Saral 1.2.5

A quieter workspace, focused settings, and more useful voice workflows.

What changed

• Settings shows only its categories, with one return arrow beside Settings. Return restores the previous workspace page; selecting Settings again opens General.
• Restored the sidebar logo, including packaged version labels. Light, dark, and System appearance now also reach the native recording indicator and macOS Dock icon.
• Refined vocabulary forms, writing-style preferences, provider details, and Diagnostics. Home stays minimal.
• Saved API keys remain masked. Home offers Unlock API keys when the OS vault needs approval after an update.
• Voice Answers and selected-text actions default on while respecting saved choices. Fn + Control asks a question on Mac; Ctrl + Alt + Q is the Windows answer default. Ordinary dictation shortcuts stay unchanged.
• Selected-text requests use the meaning of the whole request: contextual edits replace the selection, questions preserve it and append an answer, and literal requests such as “write the word rewrite” insert those words exactly. Unclear requests ask for clarification; malformed model responses never replace the original. Secure fields, changed selections, and unsupported destinations fail safely. Remote selected-text context through Apple Screen Sharing is not supported.
• A three-step optional quick-start guide shows an animated typing example, speech/writing provider setup with automatic fallback, and privacy. Next, Back, Skip, Close, and Escape are supported. Skipping preserves preferences; reopen the guide from Help & about. Motion follows accessibility preferences.
• Live meetings detect speech on-device and transcribe bounded speech sections rather than sending silence. Record-first mode, multi-file import, and ordered transcript combination remain available.
• Support offers Ko-fi and matching QR codes for realswapnil@ptyes and realswapnil@ptaxis. Feedback prepares a user-reviewed email draft.
• Shortcut help reflects current settings. Privacy controls are last in Settings and the final setup screen.
• Owner dashboard source includes aggregate usage, performance, release and integration views. Its deployment status is documented separately; local preview data is synthetic.

Installation and limits

Use the installer for Apple Silicon, Intel Mac, or Windows x64. Quit Saral before replacing the Mac app. Settings, saved-key references, meetings, and downloaded models live outside the app. OS credential approval may still be necessary after an update.

Mac builds retain the existing community signing identity, but are not Apple Developer ID signed/notarized. Windows builds are unsigned. First-install OS warnings may occur; do not disable security protections.

Automatic fallback tries the next available configured provider for speech or writing when a request fails, within bounded retry limits. It cannot guarantee service if all configured providers fail, and provider charges still apply. Dictation needs a configured speech provider or a downloaded on-device model; voice answers also need a writing model. The workspace does not require completing the guide.

Automated privacy, routing, API, update, UI, capture, and native packaging checks gate publication. Physical Windows microphone/shortcut/editor testing remains assigned to the collaborator. Native selected-text append/replacement was verified with disposable TextEdit fixtures; arbitrary-editor compatibility and live-model intent accuracy are not universal guarantees. Voice answers are not live web search and can be wrong; cloud processing may incur provider charges.

Saral v1.2.4

Saral desktop prerelease

Saral is the existing FlowClone dictation app, now packaged for macOS Apple Silicon,
macOS Intel, and Windows x64. Your existing ~/.flowclone preferences are preserved.

Version 1.2.4

• Import up to 20 recordings at once, with sequential processing and per-clip cancellation. Select and reorder completed transcripts to save a combined copy with continuous timestamps, source labels, and text/Markdown/SRT export. Original recordings and transcripts remain intact.

• Add a separate, configurable hold-to-ask voice-answer shortcut (Fn + Control on Mac). Choose Answer only or Question + answer in Settings → General. Uses configured AI models, not live web search; failed answers keep the question in History without pasting it. Normal dictation stays separate.

• Improve safe rich-text paste with headings, emphasis, nested lists, numbered steps, comparison tables, blockquotes, and literal code blocks. Compatible editors receive HTML (plus native RTF on Mac), including through Apple Screen Sharing. Plain fields receive clean text without presentation markers; code editors and literal dictation preserve their original text.

• Fix literal v insertion through Apple Screen Sharing with native Command modifier transitions and physical left-Command flags. Explicit Send Clipboard transfer addresses stale revisions; any temporary automatic-sharing change is restored. Unsupported menu layouts or focus/clipboard changes stop safely and retain the result in History. The integration currently uses English Screen Sharing menu labels. Local paste remains unchanged; other remote clients still need validation.

• Fix Fn becoming unresponsive after a mismatched or missed key release. Track physical keys, reconcile missed macOS releases, re-enable disabled event taps, and reconnect stopped listeners without restarting the workspace. Readiness now includes listener health.

• Dispatch frozen speech-model helper processes before app startup so they cannot open a second workspace or leave the instance lock held after quitting.

Changes

• Settings → General now offers System, Light, and Dark appearance. The saved choice applies before the window appears, follows system changes when selected, and keeps summary cards, controls, dialogs, and shortcut text consistent.

• Dictation readiness reflects current permissions. Home shows only missing access; the full permission checklist remains in Help and setup.

• The microphone is closed while idle. Dictation activates it only when recording starts and releases it before transcription, on cancel, after failed activation, and on quit. The always-open rolling audio buffer has been removed.

• The window remembers its size and position, recovers from disconnected monitors, follows system dark mode, respects reduced motion, and provides visible keyboard focus.

• Setup allows imports and model configuration before dictation permissions are granted. Help offers a safe restart action and update-specific permission guidance.

• Recording start is guarded against repeated clicks. Quit and restart protect unfinished recordings, transcription and model downloads. Key/history/meeting deletion uses clear app dialogs with cancellation as the default.

• Local downloads check free storage before starting; already cached models can still be verified offline with limited space.

• Local Whisper is included in desktop packages, with a private On-device settings panel and explicit downloads, verification, cached weight reuse and offline recognition.

• Model catalogs display immediately from a local cache and refresh in the background on launch, every six hours, and after connections change. Newly discovered models appear without changing selected models or priorities. Failed/offline providers keep their previous catalog.

• Saved history and its metrics load at startup while recording permissions or first-run setup are pending.

• Mac API keys are consolidated into one native vault. The app reads one credential item at startup; explicit unlocking reconnects all providers. Key changes update the item in place and preserve its access control. No keys appear in settings JSON or migration temporary files. Windows retains native per-key items because its credential manager has a smaller value limit and no per-key password prompt.

• Version 1.2.0 introduces a persistent free community signing identity for Mac releases. The app and bundled engine keep stable certificate-pinned designated requirements across code changes. Every bundle is signed before archiving and verified with codesign --verify --deep --strict. A CI probe verifies that different executables satisfy the same identity. Trusted automatic Mac installation still requires Apple notarization.

• App startup no longer rewrites settings or opens Keychain password dialogs. Unavailable keys stay saved and can be restored explicitly in Settings → AI & models → provider → Restore saved keys. Denying access stops recovery; unrelated settings saves reuse already loaded keys.

• Failed startup closes the dictation child process before displaying an error. Missing engine executables are handled without an unhandled crash.

• Saral opens its own Electron desktop window with native menus, file import/export dialogs, Mac window controls, and a consistent set of vector icons. Settings stay inside the application.

• Navigation is reduced to Dictation, Meetings, Library, History, Settings, and Help. Vocabulary, snippets, and writing styles share the Library; advanced evaluation lives in Diagnostics.

• Meetings can record your microphone or supported computer audio plus microphone, and import audio/video with a bundled FFmpeg decoder. Transcripts, summaries, decisions, action items, transcript questions, and text/Markdown/SRT export stay together. Temporary audio is removed after processing; imported originals are preserved.

• The Mac Fn / Globe shortcut is recognized through native modifier events and is the new-install default. Shortcut recording works outside browser keyboard events; existing shortcuts are preserved.

• Updates check quietly after launch and every four hours, display an in-app notice, and download only after a click. Community builds verify checksums (and the pinned Mac certificate), then offer Open installer; replacing the free Mac app remains manual. Trusted native feeds additionally support restart to install. Recording/processing blocks restart. Cloudflare R2 hosting and public GitHub release variables are connected; publishing/signing secrets are required to activate the first stable feed. See UPDATES.md.

• AI setup is unified in Settings → AI & models. The overview contains compact provider priorities; provider details contain keys, selected models, and expandable catalogs. Drag or use arrows to reorder providers and models. Search, select all, and custom model IDs are supported.

• Speech and AI polish have independent fallback priorities. Reordering turns off the corresponding rotation mode. Missing keys, disabled providers, and empty model selections remain visible and recoverable without losing saved preferences.

• General settings use expandable groups for less common controls. Provider connection settings and rotation are tucked under Advanced.

• Desktop packages include a branded icon, a drag-to-Applications macOS DMG, and a per-user Windows setup EXE with shortcuts and an uninstaller. Portable ZIPs and checksums are also produced.

• Windows gets a system tray, a non-activating recording HUD, foreground process/window
  tracking, password-field detection, bounded caret context, and HTML clipboard insertion.

• Explicit provider routing survives restart. Cloud-only and offline-only profiles remain selected.

• Focus restoration is verified again before insertion. Moving to a different field or caret
  in the same window preserves the transcript in History instead of overwriting another field.

• A usable transcript can still be pasted when AI polish fails; the warning remains in History.

• Keys migrate from legacy config to the native OS credential store after a successful save.
  A locked/unavailable store preserves the existing config and reports an error.

• Private atomic UTF-8 writes prevent Windows encoding failures with Hindi or other Unicode text.

• Settings reject foreign origins, require the launch token, suppress referrers, and prevent framing.

• Only one app instance uses a given config, and quitting closes the microphone and listener.

• Stale clipboard timers cannot overwrite a newer paste, and clipboard restoration still
  runs if keyboard injection fails.

• Source scans and native executable smoke tests run before any release is published.

Install

The free community prerelease is version 1.2.4. Windows physical-device acceptance is pending collaborator testing. If macOS blocks a downloaded copy, use System Settings → Privacy & Security → Open Anyway for Saral. This is Apple’s per-app approval route; do not disable Gatekeeper or remove quarantine. All Mac provider keys now share one native Keychain vault. Use Settings → AI & models → Unlock saved keys if access is needed after an update; one unlock restores every provider. Legacy separate-item keys migrate once after successful recovery. For a local source installation that can still read old keys, tools/migrate_credentials.py transfers them into the installed app through a private pipe without per-key prompts.

Download the package for your operating system. On macOS, open the DMG and drag Saral.app onto Applications. On Windows, run Saral-1.2.4-windows-x64-Setup.exe and launch Saral from the Start menu. Neither requires installing Python, Node.js, FFmpeg, or dependencies. For the portable Windows ZIP, extract the entire archive and keep resources alongside Saral.exe.

These builds include cloud dictation and the local speech runtime. Cloud providers require your own key. Apple Silicon uses MLX Whisper; Intel/Windows ...

Saral v1.2.2

Saral desktop prerelease

Saral is the existing FlowClone dictation app, now packaged for macOS Apple Silicon,
macOS Intel, and Windows x64. Your existing ~/.flowclone preferences are preserved.

Changes

• Settings → General now offers System, Light, and Dark appearance. The saved choice applies before the window appears, follows system changes when selected, and keeps summary cards, controls, dialogs, and shortcut text consistent.

• Dictation readiness reflects current permissions. Home shows only missing access; the full permission checklist remains in Help and setup.

• The microphone is closed while idle. Dictation activates it only when recording starts and releases it before transcription, on cancel, after failed activation, and on quit. The always-open rolling audio buffer has been removed.

• The window remembers its size and position, recovers from disconnected monitors, follows system dark mode, respects reduced motion, and provides visible keyboard focus.

• Setup allows imports and model configuration before dictation permissions are granted. Help offers a safe restart action and update-specific permission guidance.

• Recording start is guarded against repeated clicks. Quit and restart protect unfinished recordings, transcription and model downloads. Key/history/meeting deletion uses clear app dialogs with cancellation as the default.

• Local downloads check free storage before starting; already cached models can still be verified offline with limited space.

• Local Whisper is included in desktop packages, with a private On-device settings panel and explicit downloads, verification, cached weight reuse and offline recognition.

• Model catalogs display immediately from a local cache and refresh in the background on launch, every six hours, and after connections change. Newly discovered models appear without changing selected models or priorities. Failed/offline providers keep their previous catalog.

• Saved history and its metrics load at startup while recording permissions or first-run setup are pending.

• Mac API keys are consolidated into one native vault. The app reads one credential item at startup; explicit unlocking reconnects all providers. Key changes update the item in place and preserve its access control. No keys appear in settings JSON or migration temporary files. Windows retains native per-key items because its credential manager has a smaller value limit and no per-key password prompt.

• Version 1.2.0 introduces a persistent free community signing identity for Mac releases. The app and bundled engine keep stable certificate-pinned designated requirements across code changes. Every bundle is signed before archiving and verified with codesign --verify --deep --strict. A CI probe verifies that different executables satisfy the same identity. Trusted automatic Mac installation still requires Apple notarization.

• App startup no longer rewrites settings or opens Keychain password dialogs. Unavailable keys stay saved and can be restored explicitly in Settings → AI & models → provider → Restore saved keys. Denying access stops recovery; unrelated settings saves reuse already loaded keys.

• Failed startup closes the dictation child process before displaying an error. Missing engine executables are handled without an unhandled crash.

• Saral opens its own Electron desktop window with native menus, file import/export dialogs, Mac window controls, and a consistent set of vector icons. Settings stay inside the application.

• Navigation is reduced to Dictation, Meetings, Library, History, Settings, and Help. Vocabulary, snippets, and writing styles share the Library; advanced evaluation lives in Diagnostics.

• Meetings can record your microphone or supported computer audio plus microphone, and import audio/video with a bundled FFmpeg decoder. Transcripts, summaries, decisions, action items, transcript questions, and text/Markdown/SRT export stay together. Temporary audio is removed after processing; imported originals are preserved.

• The Mac Fn / Globe shortcut is recognized through native modifier events and is the new-install default. Shortcut recording works outside browser keyboard events; existing shortcuts are preserved.

• Updates check quietly after launch and every four hours, display an in-app notice, and download only after a click. Community builds verify checksums (and the pinned Mac certificate), then offer Open installer; replacing the free Mac app remains manual. Trusted native feeds additionally support restart to install. Recording/processing blocks restart. Cloudflare R2 hosting and public GitHub release variables are connected; publishing/signing secrets are required to activate the first stable feed. See UPDATES.md.

• AI setup is unified in Settings → AI & models. The overview contains compact provider priorities; provider details contain keys, selected models, and expandable catalogs. Drag or use arrows to reorder providers and models. Search, select all, and custom model IDs are supported.

• Speech and AI polish have independent fallback priorities. Reordering turns off the corresponding rotation mode. Missing keys, disabled providers, and empty model selections remain visible and recoverable without losing saved preferences.

• General settings use expandable groups for less common controls. Provider connection settings and rotation are tucked under Advanced.

• Desktop packages include a branded icon, a drag-to-Applications macOS DMG, and a per-user Windows setup EXE with shortcuts and an uninstaller. Portable ZIPs and checksums are also produced.

• Windows gets a system tray, a non-activating recording HUD, foreground process/window
  tracking, password-field detection, bounded caret context, and HTML clipboard insertion.

• Explicit provider routing survives restart. Cloud-only and offline-only profiles remain selected.

• Focus restoration is verified again before insertion. Moving to a different field or caret
  in the same window preserves the transcript in History instead of overwriting another field.

• A usable transcript can still be pasted when AI polish fails; the warning remains in History.

• Keys migrate from legacy config to the native OS credential store after a successful save.
  A locked/unavailable store preserves the existing config and reports an error.

• Private atomic UTF-8 writes prevent Windows encoding failures with Hindi or other Unicode text.

• Settings reject foreign origins, require the launch token, suppress referrers, and prevent framing.

• Only one app instance uses a given config, and quitting closes the microphone and listener.

• Stale clipboard timers cannot overwrite a newer paste, and clipboard restoration still
  runs if keyboard injection fails.

• Source scans and native executable smoke tests run before any release is published.

Install

The free Mac community release is version 1.2.2. If macOS blocks a downloaded copy, use System Settings → Privacy & Security → Open Anyway for Saral. This is Apple’s per-app approval route; do not disable Gatekeeper or remove quarantine. All Mac provider keys now share one native Keychain vault. Use Settings → AI & models → Unlock saved keys if access is needed after an update; one unlock restores every provider. Legacy separate-item keys migrate once after successful recovery. For a local source installation that can still read old keys, tools/migrate_credentials.py transfers them into the installed app through a private pipe without per-key prompts.

Download the package for your operating system. On macOS, open the DMG and drag Saral.app onto Applications. On Windows, run Saral-1.2.2-windows-x64-Setup.exe and launch Saral from the Start menu. Neither requires installing Python, Node.js, FFmpeg, or dependencies. For the portable Windows ZIP, extract the entire archive and keep resources alongside Saral.exe.

These builds include cloud dictation and the local speech runtime. Cloud providers require your own key. Apple Silicon uses MLX Whisper; Intel/Windows builds use CPU Whisper. Download or verify weights inside Settings → AI & models → On-device, then select the model and priority. Downloaded models are reused and work offline. Weights remain separate from the app installer; startup never downloads them.

On macOS grant Microphone, Accessibility, and Input Monitoring permissions to Saral when using dictation. You can finish setup and import recordings before granting these. Updating from an older ad-hoc build requires one approval for the new identity. If an old Saral entry appears enabled but access remains off, turn that entry off/on or replace it with the Applications copy, then use Help → Restart Saral. Future community releases reuse this identity; actual permission retention still needs a user-granted permission and a subsequent update check.
On Windows enable desktop microphone access and use destination apps at the same
privilege level. A normal process cannot paste into an elevated administrator app.
The Windows shortcut for a new installation is ctrl+alt+space; macOS uses fn (Fn / Globe). In macOS Keyboard settings, set “Press Fn / Globe key to” to “Do Nothing” to avoid the system action. Computer audio recording needs macOS 14.2+ or Windows, plus the applicable capture permissions.

Validation limits

CI Mac community releases have valid persistent certificate signatures but are not Apple Developer ID signed or notarized. Local development builds remain ad-hoc. Both can still trigger first-install Gatekeeper warnings. Apple Developer signing/notarization and Windows signing credentials must be supplied separately for trusted distribution. Quarantine removal is not part of the installation or release process.
Automated tests cover routing, provider contracts, privacy, security, Windows clipboard
encoding and context behavior, packaged dependency loading, the actual native window,
media decoding and recording, meeting notes, exports, and...