openwhispr/CHANGELOG.md

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

Fixed

• Idle OpenWhispr no longer keeps a vector database and an inference process running. The Qdrant sidecar and the embedding model started on every launch and stayed up for the whole session, and the meeting detectors ran from boot even with meeting notifications turned off. Semantic search now starts on the first assistant search, serves keyword results while it warms up, and releases Qdrant and the embedding session after five minutes without a search. Note edits made while it is asleep are journaled in the notes database and indexed on the next activation, so search results stay complete. The microphone and meeting-app detectors start only once your saved notification preferences are loaded, stop when meeting notifications are off, and a detector stopped mid-scan can no longer come back reporting a meeting app. (#2143)
• Assistant selection capture handles dormant macOS accessibility trees. When accessibility exposes only a browser or Electron window, selection capture now tries a synthetic copy instead of assuming nothing is selected. The caret probe also waits long enough for its native retries to finish. Automatic answer delivery still requires a verified writable field at capture and delivery time: an unknown target keeps the answer in the panel and, with Auto-Paste enabled, copies it for manual paste. Dormant inputs still require manual paste because an empty copy cannot distinguish them from an integrated terminal, a page without an input, or some live selections. (#1952)

[1.10.2] - 2026-09-15

A hotfix for 1.10.1. Generate AI Summary with your own API key or an enterprise provider no longer times out on a real transcript, a recording with nothing substantive in it gets a one-line summary instead of a blank page and an error, and OpenRouter models can read screenshots again.

Notes

• Generate AI Summary with your own API key or an enterprise provider no longer times out on a real transcript. Every bring-your-own-key request shared the 30-second deadline meant for dictation cleanup (enterprise providers had 60 seconds), and a timeout was treated like a dropped connection and retried three more times. A 2,000-word transcript through a reasoning model such as GPT-5.6 Terra needs longer than that to think before it writes, so the summary spun for about a minute and a half, four requests were billed, and it ended with a timeout error and no note. Note formatting now waits up to ten minutes, matching OpenAI's own client default, and a request that does hit its deadline fails once instead of being re-sent. Dictation cleanup keeps its 30-second deadline on every provider. (#2204)
• A recording with nothing in it gets a one-line summary instead of an error. Recordings that held only greetings, filler or a sound check could end Generate AI Summary with an enhancement error: Detailed Notes strips content that carries nothing, so the model omitted every section and returned a blank reply, which the blank-response safeguard from 1.10.1 correctly refused to save. Generate Notes and Detailed Notes now answer with one factual sentence, in the requested output language, saying the transcript and notes contained nothing substantive, and that instruction overrides the section and bullet-count rules. Untouched default prompts move to the new text automatically; an edited prompt is left alone, and the blank-response safeguard stays on. (#2201, thanks @Chadpiha)

Assistant

• OpenRouter models can read screenshots again. OpenRouter returns model IDs with a vendor prefix, such as google/gemini-3.5-flash-lite, while the cloud model registry stores the upstream ID gemini-3.5-flash-lite, so the vision-capability lookup found no model and screenshot commands were rejected as unsupported. The lookup now knows which provider it is asking for and, for OpenRouter only, retries without the vendor prefix when the exact ID misses; every other provider is unchanged. (#2203, thanks @fthozdemir)

[1.10.1] - 2026-09-14

A follow-up to 1.10.0 built from everything that landed since. Notes open straight onto their AI Summary when one exists and the raw tab becomes Your notes; the "Enhanced" label is gone in favour of AI Summary everywhere, a meeting that ends on its own still offers the summary, and the post-meeting notification is retired. Team spaces get one Settings dialog with a built-in emoji picker and a single flat member roster, Teams are called Groups in workspace copy, the sidebar gains Invite your team, and the tray and dictation pill pick up quick actions. Uploads accept video and nearly every audio container, Windows finds the right microphone again and starts listening sooner, local dictation skips a decode step, and bring-your-own-key Deepgram and AssemblyAI work again. Cleanup no longer pastes a reply the model cut short, whisper.cpp stops dropping words at window boundaries, local models size their context to the request, assistant answers finally render tables, and Reset app data restarts the app so history keeps working.

Notes

• "Enhanced" is now "AI Summary". The tab, the toggle and every mention of the generated note now match the Generate AI Summary action that produces it, in all 11 locales. (#2163)
• Generate AI Summary no longer saves your transcript back as the "Enhanced" note. With an OpenAI reasoning model such as GPT-5.6 Terra, the summary could finish with no error and an Enhanced tab that was just the raw transcript run together. Whenever the model returned no text, the OpenAI provider fell back to handing its input straight back, and the app saved that as the enhanced note. The fallback now only applies to plain dictation cleanup, which its callers already guard; a summary or any other prompted task reports an error instead, including a refusal or a response that ran out of tokens before writing anything. Every other reasoning path, bring-your-own-key, local, enterprise, and OpenWhispr Cloud alike, also reports an error on a blank result rather than saving a blank page. OpenAI's reasoning models additionally get at least 25k output tokens so hidden reasoning cannot crowd out the answer.
• Notes open on their AI Summary when one exists. Clicking a note in the sidebar used to land on the raw notes tab every time, even when a summary had already been generated. A note with an AI Summary now opens on that tab; a note without one still opens on your notes, and the Transcript tab is never the default. The "Notes" tab is also renamed "Your notes" so it reads clearly next to "AI Summary" and "Transcript". (#2196)
• Detailed Notes gets a shorter, sharper prompt. The 1.10.0 prompt was three times longer than it needed to be, described the transcript layout incorrectly, ended with a self-check list that added reasoning time, and asked for action items in a shape the editor's owner tagging could never match, so nobody got an owner chip. The rewrite keeps the rules that matter, uses the same - [ ] Action — Owner shape as Generate Notes, drops the "Owner not specified" filler, corrects transcription misspellings of names from the participant list and custom dictionary, only lists actions someone actually took on, and says what to do with plain notes that have no transcript. Tested side by side against the old prompt on real meetings of 30, 60, and 90 minutes. Anyone still on the shipped prompt moves to the new one automatically; an edited prompt is left alone.
• New note is a split button in the top bar. + New note creates a note in the current space or folder and starts recording, as before; its chevron adds Assistant chat, which opens a fresh conversation on the Chat tab, and Team space for anyone who can create one. The button sits right of the search bar next to the window controls on Windows and Linux, and at the far right on macOS. It is hidden during notes onboarding and in the side-panel layout. (#2163)
• No more duplicate date headers. The conversation and overview lists grouped items into Today, Yesterday, Previous 7 days and Older by comparing each item with the one before it, so a list whose timestamps were not perfectly sorted, which happens whenever something updates in the background, showed the same header twice and rendered duplicate keys. Items now land in a fixed set of buckets that come out in order. (#2000, thanks @hsusul)

Meetings

• The recording timer survives switching notes. The elapsed time lived inside the header control, and the editor is rebuilt for each note, so leaving a live recording and coming back restarted the clock at 00:00. The session's start time now lives in the meeting store and the header derives the elapsed time from it, which also stops the timer drifting while the window is throttled. (#2162)
• A meeting that ends on its own still offers Generate AI Summary. The offer was gated on the Transcript tab, and every detected or quick-action meeting opens on your notes, so the one time a meeting ended without you the button was hidden. The offer no longer looks at the open tab. (#2162)
• The post-meeting notification is gone. The card that announced the meeting had ended and offered thirty seconds to restart the recording has been removed. Recordings still stop automatically when the call does and the transcript is saved exactly as before; if a stop fails to persist, the "Recording stopped automatically" toast remains. (#2192, #2197)
• Windows system-audio capture ignores a stopped helper's last words. When a capture session was stopped or replaced, stderr output from the old helper process was still fed into the new session, corrupting its buffer and raising errors against the wrong recording. Late output from an inactive helper is now dropped, matching what the macOS capture manager already did. (#2017, thanks @hsusul)

Team spaces

• One Settings dialog per space. A space's menu now opens a single Settings dialog with two tabs. General holds the name-and-icon field, Leave space for members who hold access through a group, and Delete space for admins behind the existing type-the-name confirmation. Members holds the roster described below. Opening members from a note's header lands in the same dialog. Names save on blur, Enter or close, and a draft you never touched can no longer overwrite a teammate's rename. The old inline rename, the separate members dialog and the Change emoji item are gone. (#2186)
• Members is one flat list of people. Space settings → Members now shows everyone with access in a single roster whether they were added directly or through a group, with rows marked via Design where a group is what grants it. Each row carries one quiet menu — Admin and Member with a line explaining each, then Remove for direct members; a role inherited from a group cannot be lowered here and the menu says so. Add people searches the workspace roster and adds someone directly, and typing an unknown email invites them to the workspace and the space in one step. Groups move to a collapsed Groups with access section keeping the existing assign, access-cap and unassign controls. Creating a space no longer requires a group: name and icon, then people and groups if you want them, and whoever creates it becomes its admin. Leaving drops only your direct membership and tells you when a group still grants access; removing a direct member says the same. Teams are called Groups in workspace copy across all 11 locales — "Teamspaces" and copy about the human team are unchanged. (#2194)
• A built-in emoji picker replaces the OS panel. The system emoji panel behaved differently on every platform and never worked on Linux, which only ever saw a fallback grid. The new picker is searchable and categorised, with keyboard navigation, recents and a Remove action, and its labels and search terms are translated into ten languages from emojibase data. Emoji the OS font cannot draw are hidden after a one-time check; country flags are hidden on Windows, which has none. (#2186)
• Invite your team from the sidebar. Workspace owners and admins get an Invite your team row above Get a free month that opens the existing invitation dialog for their workspace. Members, signed-out users and users with no workspace do not see it, since the invitation call requires the admin role. (#2181)
• An existing Screen Recording grant counts as the screen-context opt-in. Resetting onboarding on macOS no longer shows the row as disabled while the OS permission is already granted. Windows keeps its explicit Enable. (#2186)

Dictation

• Windows pins the real microphone again. Chromium lists the Windows default input twice, once as default and once as the communications alias, both with the device's own label. The selection code required exactly one label match, saw two, and fell back to Chromium's generic default, which it never caches, so every hotkey press re-resolved the device. Aliases are now ignored when matching, and the device is resolved once when the dictation window loads instead of on the first press. (#2180, thanks @Chadpiha)
• Dictation starts listening sooner on Windows. Pressing the hotkey could take two seconds before the microphone opened. Since 1.9.0 the app pins the system-default microphone to a real device so its health checks can run, and on Windows finding that device means starting a PowerShell process that compiles C# on the fly — a lookup that ran on the first dictation after launch and again whenever an audio device came or went, and expired after 30 seconds either way. The answer is now resolved once at launch, kept until the OS reports a device change, and shared between windows, so the hotkey path only has to open the microphone. (#2178)
• Learned corrections skip everyday words. When you fixed a transcript by swapping one common word for another, such as "why" for "what", the app treated the replacement as a mishearing and saved it to your dictionary. Those entries taught the cleanup model nothing and padded every request. Corrections that resolve to a common English word are no longer learned; names and jargon are unaffected. (#2178)
• The dictation pill stops jumping while you talk. Every live-transcript chunk asked the window to show itself and re-enforce always-on-top, which on Windows and Linux meant a ShowWindow and two SetWindowPos calls against an already visible pill, so the preview flickered and redrew as you spoke. Both calls now run only when the pill is actually hidden, the idiom the other panels already used. (#2185, ported from #1995 by @hsusul)
• The Linux dictation pill stays hoverable. The cancel X appeared on the first hover and never again, so a recording could no longer be cancelled from the pill. Returning the window to click-through asks Electron to forward hover events, which only macOS honours; on Linux it left a 1×1 input region that nothing could hover back into. Linux now keeps normal hit-testing, as Windows already did, and the pill also stays visible after Retry. (#2161, thanks @boseq)
• A cleanup reply the model cut short is no longer pasted. When cleanup hit its output-token limit, the truncated text was pasted as if it were complete, so a long dictation could end mid-sentence with no warning. Cleanup now requires a complete reply: a truncated or empty one keeps your raw transcript, pastes that instead, and raises the Cleanup failed toast with a translated reason. Retrying cleanup from history reports the failure rather than saving the partial text, and Gemini's output cap is removed so a heavy-thinking model cannot trip it. (#2130)
• Gemini cleanup no longer pastes a half-finished dictation. When the cleanup response stopped early — Gemini counts its hidden thinking against the same output budget, so a long dictation could exhaust it — the truncated text was pasted as though it were the whole transcript, with nothing to say a sentence had gone missing. An incomplete or empty response is now rejected so the original transcript survives through the existing fallback. Output headroom was raised, thinking controls are set per model, Gemini 3 uses Google's recommended default temperature unless you override it, and the cleanup instructions are sent separately from your words. Other providers and local models are unchanged. (#2092, thanks @IdrisGit)
• Snippets whose trigger starts with the assistant's name expand instead of waking it. A trigger such as openwhispr review was read as an address to the assistant, which stripped the name and sent "review" off as a command. Any match for the assistant's name that falls inside a snippet trigger is now ignored, in both detection and the address stripping, so the two can never disagree. A trigger that is exactly the bare name still shadows the wake word. (#2153)
• Quick actions in the tray and the pill. The tray menu now leads with Start listening / Stop listening, Ask assistant and Start meeting recording, and the dictation pill's right-click menu gains Start meeting recording while idle. Each action runs where it belongs: listening goes to the existing dictation start and stop, a meeting goes through the same entry the meeting hotkey uses and brings a live recording forward instead of opening an empty note over it, and only Ask assistant asks the renderer, since only it can open the panel. If organisation policy forbids the assistant, or policy is still loading, the pill surfaces and says why. (#2163)
• Windows: a dictation started from the tray pastes where you were working. Starting from the tray captured OpenWhispr's own menu-owner window as the paste target, so the fast-paste helper pulled the foreground onto OpenWhispr, dropped the keystroke and restored the clipboard over your text. A target that is ours but is not one of our real windows is now dropped; the note editor remains a valid target. (#2163)

Transcription

• Local transcription now uses sherpa-onnx 1.13.8. Cohere Transcribe now returns an empty result for silent audio instead of hallucinating a sentence; the bundled ONNX Runtime also moves to 1.28.2. (#1951, thanks @emanuelet)
• Long local dictations transcribe faster. Parakeet and Orukeet decode audio in 15-second segments, and the app sent them to the local server one after another even though the server decodes on three threads. Segments now decode side by side, one per server thread and sized to the machine's cores, so a one-minute dictation no longer waits on four serial passes. Dictations under 15 seconds are unchanged. (#2178)
• Local dictation skips a decode step after you stop. The app records in compressed WebM/Opus and used to hand that to FFmpeg to unpack before Parakeet, Orukeet, Cohere Transcribe or whisper.cpp could start, about half a second on Windows and a few milliseconds on macOS. The microphone now also keeps a raw 16 kHz copy from the moment it opens, pre-roll included, and the local engine decodes that directly. The WebM is still saved to history and remains what cloud providers, very long recordings and any fallback receive. Transcripts are unchanged: on the real engine, raw and Opus-decoded audio differed in about one word in a hundred, in the raw copy's favour. whisper.cpp also no longer needs FFmpeg present for a ready WAV. (#2179)
• Your own Deepgram key works again. Every Deepgram connection failed with Unexpected server response: 401, even though the same key tested fine in Settings and worked against Deepgram directly. Deepgram accepts a raw API key only under its Token authorization scheme and reserves Bearer for the short-lived tokens OpenWhispr Cloud mints on your behalf — and the app was presenting your key as a Bearer. This broke bring-your-own-key dictation as well as Note Recording. (#2140, thanks @nikhilmaddirala)
• Your own AssemblyAI key works in Note Recording again. Recordings died seconds in with Input Duration Violation: 40.0 ms, followed by a cascade of reconnect errors. AssemblyAI accepts audio in 50-1000 ms frames, and the app measured those frames against dictation's 16 kHz — but Note Recording streams at 24 kHz, where the same frame lasts only 33-40 ms. The frame size now follows the rate each session actually opened at, and an oversized burst — which a busy moment can produce on the system-audio channel — is split instead of being sent as one frame that trips the same limit from the other end. Dictation was never affected. (#2140, thanks @nikhilmaddirala)
• Local transcription no longer silently drops parts of a recording. whisper.cpp decodes audio in 30-second windows, and the app was starting whisper-server with --no-timestamps. Without timestamp tokens the decoder cannot report where it stopped inside a window, so whisper.cpp skipped to the next 30-second mark regardless and the speech in between was discarded — with no gap or warning, and the surviving text still reading as a complete thought. Dictation was the most exposed, but notes and meeting recordings lost the same words whenever a passage ran up to a window boundary. The 60-character segment wrap that --no-timestamps was there to suppress is now switched off directly, so words still never split mid-word. (#2150)
• Upload takes video and almost any audio container. The Upload view refused MPEG audio, video with a speech track (.mp4, .mov, .mkv, .3gp) and lossless formats like .aiff, .wma and .caf, even though the bundled FFmpeg decodes every one of them, and dropping such a file did nothing at all. One shared extension list now drives the drop zone and the native file dialog; local engines already decode through FFmpeg, and for OpenWhispr Cloud and bring-your-own-key uploads a container the provider would reject is re-encoded once to mono 16 kHz MP3 on the way out, with the temp file removed afterwards. The nine formats that worked before still upload byte for byte. Files the app cannot read are listed as skipped, with the reason, instead of being ignored. (#2193, fixes #2188)
• Uploads and meetings use the local engine you chose in onboarding. When the upload or meeting scope had no engine of its own it fell back to Whisper, so a fresh install that picked Parakeet or Cohere Transcribe transcribed uploads with Whisper base. Unset scopes now inherit the dictation engine, and switching every scope to cloud writes the same provider into all of them. Explicitly choosing Whisper for uploads still sticks. (#2033, thanks @hsusul)
• The download tray says Installing once extraction starts. A local model that had finished transferring sat behind a full bar and a 100% label under a header that still said Download in progress, so it read as stalled while it unpacked. The header now says Installing with an animated ellipsis, only once every row in the tray has reached that stage, and the animation is skipped when the system asks for reduced motion. (#2137)
• Orukeet installs count on Hugging Face. Orukeet ships as a single archive, which Hugging Face's download counter does not recognise, so the model's real usage was invisible. After a fresh, successful install the app fetches the model's small JSON manifest from the same pinned revision once, in the background, with a three-second timeout and no retries, and saves the published provenance beside the model. Already-installed models, reused archives, cancelled or failed installs and every other model make no request; installation and transcription never wait for it, and the archive's checksum is verified against what the manifest publishes. (#2157, thanks @Nathan-Roll1)

AI models

• Local models size their context to the request. Generate Notes on any meeting over roughly 80 minutes failed with a raw llama.cpp error when Note Formatting used a bundled local model, because every llama-server start was capped at 16384 tokens of context even though all 35 bundled models declare 32768 or more. The cap existed to stop an unbounded context freezing a 16 GB Mac, so it is not raised but derived: the KV cost per token is read from the GGUF itself and the ceiling comes from the memory actually available. A request that still does not fit is refused with a plain sentence naming the model, the tokens it needs and the tokens that fit, rather than a JSON body. (#2155, fixes #2142)
• OpenCode Go works as a custom OpenAI-compatible provider. OpenCode Go routes each conversation by an x-opencode-session id and rejected every request without one. A base URL on opencode.ai is now recognised by host, the way Mistral, DeepSeek and Cerebras are, and each cleanup or formatting call carries one session id across its endpoint and parameter fallbacks and retries. Other endpoints are unaffected. (#2146, thanks @L4XB, fixes #2145)

Settings

• Reset app data restarts the app. The reset closed the local database and stopped local services, then only reloaded the window, so dictations still pasted but were never saved to history and semantic search stayed down until you quit and relaunched. Reset, and Delete account with device erase, now relaunch the app through the normal quit path. The relaunch drops --hidden so a login-started process comes back with a window, drops a stale sign-in deep link so the session the reset just cleared is not restored, defers a downloaded update so it cannot replace the executable mid-restart, and on AppImage hands off to a shell that waits for the mount to release before starting the file again. (#2134, fixes #2132)
• The organisation-managed notice sits where it belongs. The "managed by your organization" banner was inset with no gap before the first section; it now spans the full content width just under the close button, with a shield icon. Opening Settings no longer shows a focused close button: the dialog takes the initial focus itself and the ring appears only for keyboard focus. (#2163)

CLI

• List and search endpoints reject bad limit and folder_id values. The local CLI bridge converted query parameters with Number(), so ?limit=abc or ?folder_id=-1 became NaN or a negative number that flowed straight into SQL and returned an empty 200 rather than an error. Those parameters must now be positive integers and anything else is a 400 validation_error, in line with the bridge's error contract. (#2015, thanks @hsusul)

Assistant

• Voice Assistant on OpenWhispr Cloud with a stray "Separate vision model" switch. On a managed workspace, an override that was switched on but never given a model was shown with an invented provider and model, and every screenshot command went to that provider without a key, failing with "OpenAI API key not configured". Workspace policy no longer invents a target for an override the user never configured, and a chosen override that policy moves to another provider is cleared rather than repointed, so it asks to be picked again. A failed assistant request is now written to the debug log. (#2139)
• Assistant answers render tables. A table in an assistant reply collapsed into a single paragraph of literal pipes, and always had: the renderer follows CommonMark, which has no table syntax. Headings, bold and lists were never affected. (#2131)
• Long assistant answers stream more smoothly. Every streamed token re-parsed the entire answer as markdown, so a long reply got slower the further it ran. Streamed text now reaches the screen at most every 32 ms, flushed immediately on completion, cancel or close, so the answer you end up with is unchanged. (#2171)

[1.10.0] - 2026-09-11

The desktop app gets a new look: an inset content container with a top bar and ⌘K search, a redesigned Home and note editor, the Yowza brand font and Nucleo icons, and one brand-blue glass surface for every primary action, onboarding included. Notes start recording the moment you create them, offer an AI summary when a recording ends, and gain Detailed Notes and Follow-up email as built-in actions. Orukeet arrives as the recommended local speech-to-text model and local models run up to three times faster on Apple silicon. Managed enterprise speech-to-text lands for workspaces on Azure OpenAI or AI Foundry, transcription gains streaming from Deepgram, AssemblyAI and Gemini Live plus Cohere Transcribe as a local engine, gpt-transcribe becomes the OpenAI default, onboarding was rebuilt and reordered, Insights gains an opt-in leaderboard and backfilled history, updates install themselves, Arabic joins as the eleventh UI language, and Astra and Fable 5.1 join the reasoning models.

Look and feel

• A new shell. The main window's content sits in an inset, rounded container on a sunken backdrop, with a top bar carrying the sidebar toggle, the page title, and a ⌘K search for notes and transcripts. The sidebar drops its own search box and border, scrollbars are hidden on the content area and the notes tree, and the toggle glides with the sidebar when it collapses instead of jumping. Back to notes now also works with a note open in a narrow window; it only handled meeting mode before.
• Home redesigned. Transcriptions are grouped by day in one rounded container with hairline dividers, time on top, and Copy plus a kebab menu revealed on hover. Upcoming events show up to five day cards with weekday headers, today in brand blue, attendee stacks, and a hover-revealed Join & transcribe that fades over long titles. Events with nobody but you are hidden. Empty states are proper cards with guidance and one action.
• Yowza and Nucleo. The UI is set in the Yowza brand font with Yowza Soft for headings, and the lucide icons are replaced by the vendored Nucleo outline set behind the same names. The font files are licensed and fetched at build time from the private brand-assets release; builds without access fall back to Noto Sans. (#2117)
• Brand glass. Primary buttons, the mic and send circles, the today header, and the onboarding brand tile share one blue glass surface derived from the app icon, and the primary colour tokens follow it. Onboarding speaks the same blue as the rest of the app in both themes.
• A contrast floor. Labels, icons, borders, and placeholders across the control panel were raised to a minimum contrast so nothing sits below legibility.
• Arabic. The eleventh UI language, with full right-to-left layout, bidi isolation for user text, and localized prompts. (#1869)

Updates

• Automatic updates replace the update popup. A new Automatic updates toggle under Settings → System downloads updates in the background and installs them when the app next quits. It is on for new installs; existing installs keep today's flow until they opt in: the app still checks for updates, and the sidebar Update Available button and Settings → System let you download and install by hand. The update-available popup window and the App updates notification toggle are gone. On Linux the toggle and the check only appear for AppImage installs, since deb, rpm and tar.gz packages are updated by the package manager. (#2120)
• No more update error popups. A failed background update check used to raise a modal the next time Settings opened, and a destructive toast in the control panel. Background failures are now silent and retried on the next check; only a manual Check for Updates reports its own failure inline.

Enterprise

• Managed transcription. Workspaces on Azure OpenAI or AI Foundry can distribute a speech-to-text deployment to every member, with no API keys on user machines. Dictation, retry from history, and file upload use it; meeting recording keeps its existing lanes, and Enterprise cloud is now a selectable transcription mode when workspace policy allows it. Members need 1.10.0 or later — an older build is not offered the mode at all — so set Minimum app version to 1.10.0 before making enterprise cloud the only allowed mode. (#1964)
• Managed enterprise providers no longer require SSO. Require SSO stays an independent workspace control. Managed Azure text processing also accepts AI Foundry and AI Services endpoints, not only Azure OpenAI. (#1964)
• AWS Bedrock failures explain themselves. Invalid credentials are rejected before a request goes out instead of surfacing as a generic mid-request error, cleanup requests cancel without leaving the runtime wedged, and the underlying credential diagnostics survive into the message you actually see. (#1981)

Transcription

• Orukeet is the recommended local model. A multilingual Parakeet fine-tune from Oruk, about 672 MB, served by the same sherpa-onnx runtime as Parakeet, so there is no new sidecar. Oruk is the first tab wherever local models are listed, and a fresh install that picks the local route starts on Orukeet instead of the Whisper fallback. (#2085, #2114)
• Local models are up to three times faster on Apple silicon. sherpa-onnx now ships the arm64-native ONNX Runtime build instead of the universal binary, whose arm64 slice ran INT8 models far slower. On an M-series Mac an 11-second clip drops from 605 ms to 173 ms on Orukeet and from 665 ms to 230 ms on Parakeet, with identical transcripts. (#2116)
• gpt-transcribe is the new OpenAI default. OpenAI deprecated whisper-1, gpt-4o-transcribe and gpt-4o-mini-transcribe on 2026-08-26, with removal set for 2027-02-26. New bring-your-own-key selections land on gpt-transcribe instead. Nothing stored is rewritten — the three deprecated models stay selectable and reach the API untouched — so an existing setup keeps working until you change it. On gpt-transcribe your custom dictionary rides the model's dedicated keywords channel, one term per field, rather than being flattened into the prompt. (#2076)
• Deepgram and AssemblyAI join as streaming providers. Both are bring-your-own-key and streaming-only. AssemblyAI offers Universal-3.5 Pro and warns when a request would silently fall back to a lesser model. (#2063)
• Gemini Live streams dictation. Google's live speech model is available as a streaming bring-your-own-key provider, with cold-start audio flushed on readiness so the opening of a dictation is not clipped. (#2062)
• Groundwork for gpt-live-transcribe. The realtime client can now run OpenAI's realtime successor to gpt-4o-mini-transcribe when the server mints such a session. Nothing switches yet: no picker change and no default change. Note recording cannot route onto it. (#2077)
• Cohere Transcribe runs locally. A third local engine at 5.42% average WER across 14 languages — the most accurate offline model OpenWhispr ships — served by the same sherpa-onnx server as Parakeet, so there is no new sidecar to install. It has no automatic language detection, so the dictation language is part of the server's identity and changing it restarts the engine. Roughly 1.7GB to download and 2.7GB on disk. (#1949)
• Cohere Transcribe rejected some perfectly ordinary language codes. A locale written with an underscore rather than a hyphen, or carrying stray whitespace, failed to resolve and the engine fell back to English. (#1967, thanks @hsusul)
• Long custom dictionaries were being cut roughly in half. Every direct-API provider had its dictionary prompt capped at 900 characters, a limit only Groq (896) and the Whisper family (~224 prompt tokens) actually impose. On gpt-4o-transcribe and gpt-4o-mini-transcribe, which accept a 16k-token context, a 1,633-character dictionary silently became 899 — about half the words dropped, with no indication anywhere. The cap now applies only where the provider requires it, the 4o family gets an 8,000-character guard, and the dictionary view warns when your list outgrows the Groq/Whisper window so the top-wins ordering is no longer undocumented. Local whisper.cpp keeps the last 224 tokens rather than the first, so that path is trimmed to match. (#1632)
• Local dictations that came back as pieces of your own dictionary. whisper.cpp can echo its prompt instead of transcribing, so a dictation could return dictionary terms you never said. Output that looks like a prompt fragment is now retried without the dictionary and without VAD, and the retry is kept only when it carries more real content than the original. Detection was tightened so speech that genuinely contains dictionary terms — a comma-separated list you actually dictated, or a word you happened to say twice — is never rewritten. (#1960)
• The dictionary-echo check ran even where no dictionary was sent. A transcript from a provider that never received your custom dictionary could still be judged an echo of it and thrown away. The check now runs only where the dictionary was actually part of the request. (#1929, thanks @AdityaPainuli)

Meetings

• Recording works again after a reload. A renderer reload mid-recording left the main process holding the session, so every later start failed with "Operation in progress" until the app restarted. The session now stops when its window navigates away, and a restart from the same window retires the stale session first.
• Smoother notification swipe. The meeting notification card follows your pointer, fades as it goes, and leaves through the edge you dragged it toward. The window no longer shows scrollbars mid-animation, and on macOS it now sits above every other window, including OpenWhispr's own floating panels.
• Meeting recordings that went silent partway through on macOS. The macOS audio tap is built once against the devices present when recording starts and never follows the machine afterwards, so a route change — headphones connecting, a device going away — could stop capture while the helper process stayed alive. Nothing noticed: no exit, no error, and silence from a dead tap is indistinguishable from a quiet call, so the rest of the meeting was lost. A watchdog now restarts the helper when no audio arrives for 6 seconds or the pinned device reports that it changed, up to three attempts, and tells you either way. A genuinely quiet call only warns, since restarting a working capture would cost real audio. (#2040, thanks @alekc)
• Playback no longer raises a meeting prompt. Combined input/output devices report audio activity while merely playing sound, so on macOS watching a video could prompt you to start a meeting note. Aggregate-device activity now prompts only while a known meeting app is running, background speech services are excluded from microphone ownership, and a detection listener that crashes respawns with backoff instead of staying down for the rest of the session. (#2071)

Dictation

• Push-to-talk no longer loses a dictation that hits the safety ceiling. When a hold reaches its time limit with the keys still down, the stop used to paste into the held modifiers, land nowhere, and discard the transcript. The text is now kept on the clipboard behind the usual dictation-error surface. Linux's watchdog matches the five-minute ceiling of the other platforms. (#2102)
• The cancel button pours out of the recording pill. Hovering the pill grows a single fused outline rather than placing a separate control beside it, tweened between pill states so it never pops mid-animation. The compact pill and its window widened to match, fixing a cancel button that could clip outside the window edge on hover. Inside the Live Transcript panel the button keeps its bordered look with the same motion. (#1886)
• Copying part of an Assistant response. Selecting a passage and pressing Cmd/Ctrl+C copies just that selection. Previously a partial copy also flipped the footer button to "Copied" although the full response had never been copied; the two are now separate. (#2035, thanks @dajiaohuang)
• A failed paste no longer costs you the transcript. When clipboard delivery failed, the dictation was dropped instead of written to history, so the text was gone with nowhere to recover it. The transcript is now saved either way, and a clipboard-only result is reported as clipboard-only rather than as a completed paste. (#1979)
• Windows and Linux media pause/resume no longer block the main process. A slow or stuck PowerShell, nircmd, playerctl or dbus-send helper previously froze hotkeys, IPC and the dictation window until it exited, so a dictation could no longer be stopped and its recording was lost. Pause, resume and toggle now also run one at a time, so a quick tap can no longer strand media paused — previously possible on macOS. (#2073)

Notes

• The note editor was redesigned. A larger title, a combined date and attendees capsule, folder and space capsules, and one toolbar row: Transcript, Notes, and AI Summary on the left, the record control and a split Share on the right. Recording widens the record circle into a stop, waveform, and timer pill, and the Transcript tab shows a live wave while recording. Export moved into the Share dialog. The chat panel no longer opens by itself on notes with history, and starting a recording keeps you on the Notes tab.
• The Transcript tab is always reachable. Notes without a transcript show an empty state with a Start recording action instead of a disabled tab.
• New notes start recording. Creating a note begins a meeting recording for it right away, unless workspace policy blocks recording or another recording is live.
• Generate AI Summary after a recording. Once a recording ends, the Transcript view offers a Generate AI Summary pill until a summary exists. It runs the new Detailed Notes action.
• Two new built-in actions. Detailed Notes turns a transcript and your manual notes into accurate, comprehensive meeting notes that keep every client, project, number, and date, separate decisions from proposals, and give action items an owner and a due date. Follow-up email drafts the email you would send afterwards, and is the ask bar's default action, labelled Draft a follow-up email. Generate Notes is unchanged. Every built-in prompt stays editable in Custom Actions, and an edited prompt is never overwritten by an update.
• Folder search in the note header is legible again.
• Notes created offline could take a long time to reach mobile. Note creation sent the note's local last-edit time as updated_at and the server stored it verbatim. Mobile's delta cursor is the newest updated_at it has already seen, so a note uploaded well after it was written — offline work, or backup switched on after a stretch of local use — landed behind that cursor and stayed invisible on mobile until some later edit happened to push it forward. The server now stamps the creation time itself. (#2065)
• Granola imports silently dropped notes that shared a title and date. Two notes with the same title on the same day collided on the same generated id, and the later one was discarded without a word. Colliding rows now get versioned ids allocated deterministically across sorted files, so a reimport produces the same result, and the identity no longer includes note content — editing a note cannot change which id it owns. (#1955, thanks @hsusul)
• Transcript scroll stopped following the live meeting. Scrolling up inside a speaker picker — which renders in a portal, so its wheel events bubbled to the transcript even though the pointer was never over it — latched the transcript out of follow mode, and a wheel over non-scrolling chrome such as the recording header wedged it off with no way back. The scroller now only counts gestures actually aimed at it, and a recording that opens with partial text before its first final segment is followed from the start. (#1896, thanks @boseq)

Uploads

• Uploads warned that speaker detection "couldn't be applied" without trying. A run could have speaker detection enabled while the local speaker models were still downloading or had never been downloaded, and it reported failure without invoking the diarizer at all — on every file in a batch. Uploads now fetch the models they need and wait on an in-flight download rather than giving up. When labeling genuinely cannot be applied the transcript is still kept, the warning is localized, and the active model label is a link into the upload model settings. (#1991)

Onboarding & accounts

• Onboarding runs dictation, then Notes, then a live assistant demo. The demo is answered by the real assistant through the same pipeline the app uses, instead of echoing your own transcript back. (#2122)
• A new welcome screen. A brand-blue hero with the app icon in a glass tile, glass primary buttons, a borderless email field, and the tagline "The power of your voice. On your terms."
• Voice Assistant wording and scope. Onboarding consistently says Voice Assistant, spoken commands run on the model chosen under Settings → Voice Assistant rather than the Chat model, and sending a screenshot as context is an explicit opt-in. (#2106)
• Onboarding was rebuilt and now keeps your progress. The setup flow was reworked end to end, and leaving partway through no longer starts you over. Provider secrets entered during setup stay out of the persisted session, and a cancelled model download no longer leaves Proceed disabled on a model you had already downloaded. (#1984, thanks @boseq)
• Plus-addressed email accounts could not sign in. The desktop rejected any address containing a + before the request ever left the app, so an account the website had happily created with a plus alias was locked out of every in-app sign-in surface, re-authentication included. Neither the website nor the API had that restriction, so it blocked legitimate logins without stopping anything. (#2088, thanks @boseq)

AI models & routing

• The voice assistant can read and edit your dictionary and snippets. Three new tools fetch a snippet by its trigger, add or remove dictionary words, and create, replace, or delete snippets, with a confirmation step before any bulk cleanup. The CLI bridge gains matching snippet routes and the API key picker exposes dictionary and snippet scopes. (#2119)
• The separate vision model override is bring-your-own-key only. OpenWhispr Cloud already routes screenshot commands to a vision model, so offering it there was a no-op. Cloud reasoning requests now declare their purpose, so translation and note formatting can use their own fallback chains. (#2115)
• Astra and Fable 5.1 are available as reasoning models. (#2107)
• Gemini replies could come back empty. Responses split across multiple parts, or carrying the model's thinking alongside its answer, were not reassembled — so a perfectly good reply could surface as nothing at all. (#1969, thanks @hsusul)
• Dictation cleanup was sampling at temperature 0.7. Cleanup has been specified as deterministic since it was introduced, but neither cleanup config actually passed a temperature, so the three IPC-bridged providers fell back to their own defaults — 0.7 for a local llama-server model, 0.3 for Anthropic and enterprise. Both cleanup routes now pass 0 explicitly, and only where the model registry says the model accepts the parameter. The agent, translation, and selection-edit steps are unchanged. (#2007, thanks @xlurie)
• Tinfoil selections landed on the wrong model. Tinfoil retires GLM-5.2 on 2026-09-10 and its live model list already omits it, so the default lookup missed and fell through to whatever the enclave happened to list first. Providers now name their default model in the registry instead of relying on list position, and a persisted GLM-5.2 selection is repaired at startup rather than only after a successful catalog fetch — which missed offline launches and lost the race against the first request of a session. (#2018)

Models & GPU

• Local models download in parallel, and progress survives closing the window. Download progress is now owned by the main process rather than the window that started it, so closing the control panel no longer loses track of a download in flight, and several models can download at once with independent progress. Failures are tracked per model and wait for an explicit retry instead of being retried silently. (#1267, thanks @IdrisGit)
• The GPU acceleration retry is easier to find. After a GPU crash, retrying activation is reachable from the fallback notification itself rather than only from deep in settings. (#1913, thanks @IdrisGit)

Insights

• History before Insights existed is backfilled. Content-free counters are reconstructed from your retained local transcription history, so Insights is not empty for dictation that predates the feature. Reconstructed rows sort behind live events and never block current analytics. (#2075)
• A usage Insights view. A new tab summarizing your dictation activity over time, computed locally by default. Syncing insights to your account is opt-in and off unless you turn it on, activity is never attributed to an account other than the one signed in when it was recorded, and clearing history or deleting your account removes the underlying rows. A workspace that enforces local history off writes nothing at all. (#1959, thanks @boseq)
• An opt-in leaderboard. Compare your dictation activity with colleagues who have also joined. Joining is an explicit action that is never inherited from the Insights sync toggle, declining the sync prompt never joins you, and leaving is one button on the leaderboard header. The share card shows a name or an email local part, never a colleague's full address. (#2011, thanks @boseq)

Transcription routing

• Signing in no longer silently reroutes dictation to OpenWhispr Cloud. Restarting onboarding from Settings (Create Free Account, the plan cards, or picking OpenWhispr Cloud while signed out) armed a post-sign-in "cloud migration" that rewrote how audio was routed but not the mode the Settings picker reads. Whether you had chosen Local, your own API keys, or your own endpoint, Settings kept showing that choice while your audio went to OpenWhispr Cloud — and clicking the mode already shown did nothing. That migration is gone, and a profile already stuck this way is repaired on the next launch from the mode its picker shows, for dictation and audio upload. The repair only ever moves away from our servers, so it can never start sending audio that was staying on your device or going to your own provider. Note Recording was not affected by this particular bug: it reads the same setting its picker shows. It had a separate problem of its own, below. (#2086)
• One consequence of that repair worth knowing: the picker is the only record of what you asked for, so a profile is moved back to what it shows even if it has been transcribing elsewhere for a while. A picker reading "Local" returns to Local, and one reading "API Keys" or a self-hosted endpoint returns to your own credential. If the local model was never downloaded, or the API key has since been removed or expired, dictation now says so instead of quietly using OpenWhispr Cloud — pick a mode again in Settings → Speech to Text to settle it. (#2086)
• Note recordings could go to OpenWhispr Cloud for a user who is Local everywhere else. Profiles that upgraded straight from 1.6.7 or earlier to 1.6.10 or later copied their Note Recording settings from dictation before the modes those settings are read from existed, so Note Recording fell back to OpenWhispr Cloud while every other setting said Local. Those profiles now have the mode re-derived from what was copied. Running 1.6.8 or 1.6.9 in between avoided it — a six-day window, 14 to 20 April 2026. Where the copied settings name an endpoint or provider that cannot stream, Note Recording now says so instead of quietly using OpenWhispr Cloud. (#2097)
• Note formatting could follow dictation cleanup to OpenWhispr Cloud after a local setup was skipped by the same copy. Note formatting falls back to the dictation cleanup model when its own mode was never written, so those profiles kept working locally until they switched cleanup to OpenWhispr Cloud, at which point their note text followed. A skipped local setup is now restored to Local. A skipped cloud setup is deliberately left alone so it keeps following cleanup, rather than being pinned to a provider the user may have moved away from. (#2097)

CLI

• Transcribe files from the terminal with your local models. The CLI bridge gains POST /v1/transcribe, which takes a file path and runs whichever local model the app is set to use (whisper.cpp, Parakeet, Nemotron, Cohere, or Orukeet), and GET /v1/transcribe/models, which lists downloaded models. Audio never crosses the bridge; the app reads the file itself. Powers openwhispr transcribe <file> in @openwhispr/cli 0.3.0. (#2121)

macOS

• Accessibility features wait for their permission. Startup no longer touches accessibility APIs before the permission is granted, which removes the premature system prompt and the races it caused during first launch and onboarding. (#2101)

Windows

• winget manifests are published from CI with tag-pinned URLs. Every manifest since 1.7.5 pointed at the releases/latest download path, which broke as soon as the next release shipped. Manifests are now submitted on publish with permanent asset URLs, minutes after a release rather than days. (#2089)
• Local transcription was dead on arrival without the VC++ redistributable. The bundled whisper-server.exe links the dynamic MSVC runtime but shipped as a bare executable, so on any machine missing the VC++ 2015–2022 redistributable the Windows loader killed every spawn before it started — with empty output, so nothing explained the failure. The runtime DLLs now ship alongside the binary. (#1687)
• Parakeet failed to start on machines with an older system ONNX Runtime. Windows 11 ships ONNX Runtime 1.17 in System32, and on some machines the loader resolved the bundled sherpa-onnx binaries' runtime to that copy rather than the newer one sitting beside the executable. The process then died before it could listen, reported only as "parakeet-ws process died during startup". The bundled runtime now ships under a private name that no system loader rule can shadow. (#2074)

Linux

• Hyprland hotkeys using punctuation. Control+, and other unshifted punctuation accelerators failed validation and were never registered, because a literal comma is Hyprland's own bind delimiter. Those keys are now converted to their XKB names. (#1877, thanks @hsusul)
• Auto-paste dropped keystrokes on GNOME Wayland. The bundled paste helper creates a throwaway input device and sends keys 50ms later, but GNOME can take 150–500ms to open a hotplugged device — so the keystrokes arrived partly or not at all while the helper still reported success, which meant the reliable fallback was never tried. Where the ydotool daemon is running, paste now goes through its persistent device first. (#2078)

Security

• Two high-severity advisories in bundled dependencies were closed. The note editor's Tiptap packages moved from 3.30.4 to 3.31.3, clearing a quadratic ReDoS in Markdown attribute parsing (GHSA-j95f-988m-3j2f) and, by way of prosemirror-view 1.42.3, a cross-site scripting vector reachable from pasted content (GHSA-c8x8-7fp4-3x9w). js-yaml moved to 4.3.2, clearing unbounded CPU use on empty merge sources (GHSA-2883-xcg3-v3hh). Neither change alters behavior. (#2079)
• One advisory remains open. adm-zip's symlink-following issue (GHSA-vwc7-r8mq-g2x9, moderate) reaches OpenWhispr as a transitive dependency of onnxruntime-node, which still requires it as of 1.29.0, and it has no patched release to move to. It stays visible in the audit output rather than being suppressed. (#2079)

[1.9.2] - 2026-08-29

A repair release for two 1.9.1 regressions. Windows desktop sign-in works again — every provider button had gone dead — and the three transcription paths that only failed in packaged builds are back on all platforms. Meetings get three fixes of their own: recordings that captured only your voice on Windows, prompts that stopped appearing after the first call, and swipe-to-dismiss on the prompt cards.

Windows

• Desktop sign-in works again. Since 1.9.1, clicking Google, Microsoft, Apple, or SSO on Windows showed a spinner and then nothing. Browser links were routed through explorer.exe, which silently opens a File Explorer window instead of the browser for any URL carrying a query string — and every sign-in URL carries one. Connecting a calendar, opening Stripe checkout, and following tagged outbound links broke the same way. Those links now open directly; the ones explorer.exe can deliver still go through it, so meeting recordings keep capturing the audio of a browser launched from a meeting link. (#1932)

Transcription

• Three transcription paths failed in packaged builds. A model-registry import added in 1.9.1 was missing the attribute Node's module loader requires for JSON, so xAI dictation, re-transcribing a saved recording, and bring-your-own-key file uploads threw needs an import attribute of "type: json" on Windows, macOS, and Linux while ordinary dictation kept working. Development builds inline that import, which is why it reached release. (#1908)

Meetings

• Windows meetings that recorded only your own voice. On some machines the system-audio helper starts up reporting itself healthy and then captures nothing but digital silence, so meeting notes contained your microphone and none of the other participants — with no error anywhere in the app. The check that picked the capture method only asked Windows whether loopback capture was available, never whether any audio was actually arriving, so the working fallback was never reached. The helper now compares its own output against what your speakers or headphones are really playing, and a recording that hits this switches to the fallback a few seconds in and keeps the rest of the call. (#1935, thanks @KishenG)
• Meeting prompts kept appearing after the first call. Closing a meeting prompt any way other than its own buttons — a compositor killing the window on Hyprland, a failed load, onboarding taking over the screen — left the detection marked as still on screen, so no later meeting ever prompted again for the rest of the session. On Linux the prompt also fired on OpenWhispr's own microphone use, and it now waits until the audio server has said which application owns the capture. A second call raises a fresh prompt once the first one's microphone has actually gone quiet, while an ongoing call still prompts only once. Explicitly dismissing keeps its five-minute pause, and letting a card expire still costs nothing. (#1918, thanks @IdrisGit)
• Meeting prompts swipe away like desktop notifications. A horizontal pointer swipe of 80px in either direction dismisses a meeting card — detection, calendar reminder, or the auto-end restart offer — taking the same path as its close button. Swipes that start on an action button are ignored. (#1906)

[1.9.1] - 2026-08-27

Meeting recordings now stop the moment the call does, with a short window to pick the same note back up. Gemini joins the bring-your-own-key transcription providers, Granola users can import their back catalogue, and the dictation pill gets a brand-blue processing glow plus a companion surface for Agent mode. Account deletion no longer reaches across accounts, and 1.9.0's recording-time CPU regression — along with the GPU packs 1.8.3 quietly deleted — is repaired.

Meetings

• Meeting recordings stop as soon as the call ends. The microphone, silence, and meeting-app signals no longer open a 60-second Keep recording countdown first, so recordings now end about a minute sooner than before. A microphone release still waits 5 seconds of continued quiet before it counts, so a headset reconnecting or an app that drops the mic on mute cannot cut a live call short. A 30-second card then offers to restart the same note — keeping its transcript, folder, and speaker settings — and restarting holds auto-end off for five minutes. (#1898)
• Meetings that recorded you but nobody else. On Windows, the system-audio helper excludes OpenWhispr's own process tree from loopback capture — and a browser launched by clicking the meeting link from inside the app landed in that tree, so the other participants were never captured at all. The meeting link now opens outside the tree, and both capture helpers report enough detail for a debug log to tell real silence from a dead transcription socket. (#1864)
• A malformed speaker segment no longer throws. Nullish segments and non-object patches are ignored instead of crashing speaker assignment. (#1798, thanks @hsusul)

Transcription

• Gemini is a bring-your-own-key transcription provider. Google's dedicated speech model (gemini-3.5-transcribe) joins OpenAI, Groq, and Mistral for dictation and uploads, reusing the Gemini key you may already have saved for cleanup. Batch only, so it stays out of the streaming-only meeting picker. (#1899)

Dictation

• A Signal glow replaces the rainbow processing beam. The thinking state on the dictation pill was nine hues orbiting a 40px circle, which read as faint chromatic noise in light theme. It's now a cyan-to-indigo band with a white-hot comet head and a breathing halo that's legible from across the room. Agent mode keeps its ocean beam, and prefers-reduced-motion disables the motion entirely. (#1826)
• Agent mode gets its own dictation pill. A focusless companion surface appears on the opposite edge of the same display, with the real microphone waveform and the production entrance timing, and it hosts Live Transcript when preview mode is on. Agent and translation hotkeys can't activate it, and plain dictation no longer animates the Agent footer pill. (#1867)
• Voice Assistant answers paste into the field you were typing in. With Auto-Paste on, the assistant's completed response goes back to the writable input you had focused, revalidated at paste time. If the target moved, the field is protected, or nothing writable was focused, the response goes to the clipboard and stays visible in the panel instead of being lost. (#1859)
• Recording starts without waiting on the cloud transcription config. Local-mode and signed-out users sat through a 4–5 second IPC round trip before the microphone opened, for a config that can only change the decision for signed-in Cloud users. (#1710, thanks @AdityaPainuli)
• No trailing space after Chinese or Japanese text. Dictation appended an ASCII space after every paste, so consecutive Chinese dictations accumulated gaps that violate East Asian typography. Korean deliberately keeps its space, since modern Hangul separates words the same way English does. (#1823)

Notes

• Import your Granola history. Settings → General takes Granola's official CSV export and brings meeting notes across with their original dates, transcripts as native segments, and speaker names locked where the export labels them. Notes land in an "Imported" folder, and re-running the import skips what's already there rather than duplicating it. (#1813)
• Uploaded transcript timestamps stay in bounds. Provider names are matched regardless of casing or stray whitespace, and negative or inverted segment ranges are clamped instead of reaching the transcript. (#1800, thanks @hsusul)

Calendar

• The assistant can find you a free slot. A read-only availability tool unions your selected Google, Microsoft, and Apple calendars into busy intervals and open slots, honoring all-day events, self-declined invitations, buffers, and a minimum duration. No titles, attendees, links, or account identifiers reach the model. (#1821)
• Availability looks 31 days ahead instead of 7. "Suggest times Wednesday or Thursday next week" can land 13 days out and used to fail outright; each provider's sync window widened to match. (#1858)
• Recurring meetings on shared Microsoft calendars have names again. 1.9.0 backfilled series masters only from your own mailbox, so occurrences on a shared calendar 404'd on every sync and stuck as "Untitled Event" until the delta token expired a week later. The lookup is now calendar-scoped, and a failed backfill retries within ten minutes. (#1835)

Onboarding & accounts

• Screen Context is offered during setup. An optional macOS permission row alongside the existing ones, gated by organization policy. Microphone remains the only permission that blocks you. (#1856)
• Compact onboarding windows have real window controls. Minimize, maximize, and close on Windows and Linux; native traffic lights on macOS; and the window can be resized. (#1857)
• Sign-up checks whether your email already has an account before sending you down the wrong path, through the auth service rather than the optional cloud API, and handles a duplicate-account race at submit. (#1854)

Data safety

• Deleting one account no longer erases another account's local content. Account deletion now removes only that account's private notes and folders, preserves workspace-owned spaces and content with former-user attribution cleared, and leaves downloaded models and device settings intact. Deleting a folder likewise releases other accounts' notes to the space root instead of deleting them. A separate unchecked option erases all app data from the device when you're leaving OpenWhispr permanently. (#1812)

Enterprise

• Buy Enterprise without talking to sales. Signed-in users get an Upgrade to Enterprise button on the pricing tile that opens a checkout dialog — pick the workspace, seat count, and billing cycle, then pay through Stripe. Owners and admins of an active Enterprise workspace also get a one-click Enterprise console entry from Plans & Billing and Workspace settings. (#1623)
• Organizations can require local models. Managed users missing a required model hit a blocking onboarding step that downloads them with per-row progress and retry; already-onboarded users get a non-dismissable banner with an inline download instead. (#1836)

AI models & routing

• Thinking tags stop leaking, whatever their case. <Think> blocks from LAN fine-tunes and orphan </think> closers from R1-style templates survived the filter on both the streamed and non-streamed paths, landing reasoning text in titles and notes. (#1802, #1862, thanks @hsusul)
• An Anthropic reply with no text block no longer crashes cleanup. A response that hit its token limit before producing text threw a TypeError; it now reports what happened and your raw dictation still pastes. (#1709, thanks @AdityaPainuli)

Performance & GPU

• 1.9.0's recording-time CPU spike is fixed, and 1.8.3's deleted GPU packs heal themselves. The notes bottom bar animated a layout property inside a blurred surface sitting over the live transcript, so every partial forced a re-blur and CPU scaled with participant count. Separately, 1.8.3's pack migration deleted GPU packs while leaving acceleration flagged as enabled, silently dropping whisper-large onto the CPU binary — orphaned packs are now detected at launch and repaired instead of waiting for a manual retry. (#1863)
• The GPU banner refreshes when you close Settings. Installing a pack left the home screen showing the previous result until something else changed or the app restarted. (#1804, thanks @zhongwater123)

Elsewhere

• Settings stays open when you dismiss a dropdown. An open dropdown made the panel inert, so the next click anywhere inside it landed on the modal backdrop and closed the whole thing. (#1820)
• Debug logs record which release produced them. The System Info header carried platform, Node, and Electron versions but not the app's own. (#1688)
• Packaged builds block untrusted file navigation. The control-panel guard trusted every file:// URL. (#1776, thanks @hsusul)
• Self-hosted servers across the full IPv6 link-local range are reachable. The check matched fe80 literally rather than fe80::/10, rejecting fe90::, fea0::, and febf::. (#1733, thanks @hsusul)
• The CLI bridge handles UTF-8 correctly. A multibyte character split across two network chunks was corrupted, and the 1 MiB request limit counted characters instead of bytes. (#1386, #1777, thanks @hsusul)
• Date formatting survives null and invalid input instead of throwing on note lists and transcripts. (#1784, thanks @hsusul)
• Renderer stores import cleanly under partial browser environments, closing a class of import-time crash that briefly took CI down. (#1841, #1842)
• Two CI checks no longer share a name. A failing gate rendered as the twin of a passing one, so contributors had to dig through the logs to tell which suite had broken. (#1745, thanks @AdityaPainuli)

[1.9.0] - 2026-08-24

The chat agent and the voice agent become one Voice Assistant behind a redesigned dictation pill, guided setup is rebuilt end to end, and the notes surface gets a design pass with @mention tagging and real speaker identity in meeting notes. Meetings you forget to stop now end themselves. Linux gains push-to-talk on Wayland, Windows pastes back into the window you dictated into, and local transcription stops thanking you for watching.

Voice Assistant

• One assistant, one entry point. The separate always-on-top chat window is gone. The dictation pill is now the only surface: speak a standalone command and the answer streams into a floating panel with the chat's full toolset — notes search and editing, calendar, web search, clipboard, and memory when you're signed in — with conversations saved and resumable. Highlighted text is still edited in place. (#1597)
• Answers return to the focused field. With auto-paste enabled, standalone Voice Assistant answers now paste at a verified writable cursor. Without a writable cursor—or if the target changes while the answer is generated—the completed answer stays safely in the floating panel and is copied to the clipboard for manual paste; the existing Copy button confirms the automatic copy for six seconds. Empty or cancelled responses do not change the clipboard. (#1859)
• A pill that shows what it's doing. An idle orb that grows into a recording capsule with a live, level-driven waveform and an in-capsule cancel, then a processing state that never changes shape mid-cycle. Dragging the pill no longer walks it across the screen as it resizes. (#1597)
• Your dictionary reaches the assistant. Custom-dictionary words are injected into every assistant conversation, so replies come back using your names and jargon. (#1597)
• Follow-ups without reopening. Type into the panel or press the hotkey again to speak into the same conversation; Esc collapses it and cancels anything in flight; Copy takes the answer in one click. (#1597)
• Wake words work in your language. "Ehi Jarvis", "Привет, Jarvis", "ねぇ Jarvis" — localized vocative cues now count as addressing your agent, not just the English ones. (#1604, thanks @xAlcahest)
• Selection edits no longer fail on a missing completion marker. (#1587, thanks @hsusul)
• Assistant conversations survive the cloud migration. (#1772, thanks @hsusul)

Onboarding

• Guided setup, rebuilt. A new versioned flow walks through sign-in, permissions, language, use cases, hotkey capture, live dictation and assistant demos, and calendar connection — then branches into guided OpenWhispr Cloud, bring-your-own-key, local-model, or enterprise setup, with provider validation and models downloading in the background. Demo recordings are isolated from the rest of the app, so the floating pill and global hotkeys stay out of the way while you're being shown around. (#1670)
• Signing back in looks like signing up. The returning-user screen uses the same compact authentication shell as onboarding instead of a form embedded in a full-size control panel. (#1763)

Notes

• A design pass over the notes surface. A shared gradient send and mic identity across every input, a voice-reactive recording pill with an elapsed timer, and a translucent "liquid glass" bottom bar that note content scrolls under. The ask input stays mounted and animates as the chat panel opens instead of blinking out. (#1651)
• Voice notes straight into chat. The chat input's mic records with a full-width live waveform, timer, and cancel; the recording runs through the normal transcription pipeline (any provider, no cleanup pass) and lands in the input for you to review before sending. (#1651)
• @mention the people in your notes. Typing @ opens a picker fed by your account, meeting attendees, and known speakers; chips render with an avatar and round-trip through Markdown as readable links. Generated action items tag their owner automatically. (#1741)
• Create a note inside a folder. Folder rows get the same hover "+" that space rows already had, instead of making you create the note elsewhere and move it. (#1650)
• Timestamped transcripts and SRT export for uploads. Audio uploaded or ingested from a URL and transcribed through your own OpenAI, Groq, or Mistral key now keeps segment-level timestamps, so the Transcript tab and SRT/TXT/JSON/MD export light up for upload notes. (#1727)
• Cancelling an upload actually stops it. Cancel used to reset only the UI while local transcription and diarization kept burning CPU to the end of the file. (#1728)
• The chat panel no longer covers what you're reading. Opening it reserves scroll space and reveals the bottom of the note instead of forcing a manual scroll. (#1769, thanks @IdrisGit)
• Uploaded notes remember their speaker detection. A note created through Upload ran speaker detection but stored none of it — it now records that diarization ran, the speaker count you chose, and the audio duration, so it behaves like a meeting note when you record into it or resolve participants. Present since upload speaker detection shipped in 1.7.6. (#1610)
• Generated titles lose their quotes. Curly quotes, guillemets, and stacked wrappers are peeled off; apostrophes inside the title stay. (#1640, thanks @hsusul)
• Mirrored Markdown files stay inside your mirror folder. A folder named with .. could write note files outside it. (#1773, thanks @hsusul)
• A line break in a title no longer destroys a note's metadata. Control characters are escaped in mirrored frontmatter instead of invalidating the whole block. (#1646, thanks @hsusul)
• One undeletable file no longer strands the rest. Deleting a note whose mirror file is open in another editor now cleans up everything else. (#1649, thanks @hsusul)
• Future timestamps show a date instead of "now". (#1768, thanks @hsusul)
• The share dialog reads email domains correctly. Padded addresses and Name <addr> forms no longer make a personal Gmail look like a company domain. (#1683, thanks @hsusul)
• The unused workspace short name is gone from settings. (#1660)

Meetings & speakers

• Forgotten recordings end themselves. When the call is over — the meeting app releases your microphone, or both audio channels go quiet, or the app exits — a 60-second countdown starts, cancellable with Keep recording, and the card tells you which signal fired. Remote voices still playing defer the countdown, so a quiet-but-live meeting is never cut short. (#1494)
• Meeting notes know who is who. Every transcript line carries its resolved speaker name and the mic track carries your own name, a Meeting Context block names the note owner and invited participants, and the prompt is forbidden from inventing identities — so notes stop assuming you're whoever got named in conversation. (#1741)
• In-person meetings get speaker labels. Recordings with everyone in the room and no call ended with zero diarized segments, because only the (silent) system-audio channel was ever diarized. Those sessions now diarize the microphone track. (#1726)
• Windows system-audio capture is re-checked at meeting start. One early probe failure used to pin the whole session to the browser fallback. (#1474, thanks @stantheman0128)

Calendar

• Recurring Outlook meetings have names again. Microsoft's delta sync can return occurrences as bare stubs, which showed up as "Untitled Event"; they're now backfilled from the series master. (#1665)
• Personal time blocks stop prompting you to record. Focus time, reminders, and "lunch" — events with no attendees and no meeting link — no longer fire the meeting overlay, on Google, Microsoft, and Apple Calendar alike. (#1615, #1696, thanks @IdrisGit)
• Join recognizes more meeting links. Zoom webinars and personal links, and Teams /meet/ URLs. (#1692, thanks @hsusul)
• All-day events sit on the right day. West of UTC they landed on yesterday's card in Coming up. (#1742)
• A mismatched sign-in fails immediately instead of leaving Connect spinning for two minutes. (#1753, thanks @hsusul)
• Coming up, redesigned. Day cards with a today marker, per-event accent, an attendee avatar stack with RSVP state, and a "Join & take notes" button. (#1741)

Speech to text

• Dictation works again on the default realtime model. Since 1.8.2, every dictation configured for OpenAI realtime transcription failed with Unsupported realtime token provider: undefined. Provider routing now has a single source of truth. (#1631)
• Local transcription stops thanking you for watching. whisper.cpp's anti-hallucination thresholds are now sent on every request — the values the reporter measured cut hallucinated tails from 2.25% to 0.06% across 4,814 real dictations. (#1723)
• Your phone is no longer mistaken for your microphone. A phone offered through Continuity was classified as a built-in mic and auto-selected over the real one, producing empty transcripts. (#1515, thanks @xpipko)
• Parakeet is offered only where it can run. Its bundled ONNX runtime needs macOS 15.5; older Macs now fall back safely instead of loading an incompatible native library. (#1720)
• The preview closes when you didn't say anything. It used to stick in the cleanup state. (#1667, thanks @IdrisGit)

GPU acceleration

• 1.8.3's multi-GPU regression is fixed. Vulkan pins to the discrete GPU instead of defaulting to integrated graphics, NVIDIA detection finds nvidia-smi off PATH, and packs cleared during the upgrade are announced rather than silently deleted. (#1609)
• A downloaded pack is used, with or without the env flag. A pack on disk now implies intent; the flag became an explicit opt-out instead of a required opt-in that could silently fall back to CPU forever. (#1724)
• The GPU banner only appears when cleanup actually runs locally. Cloud-cleanup users saw "GPU acceleration available" on every launch, forever. (#1591)

AI models & routing

• Retired cloud models are gone. Groq's shut-down Qwen and Llama models and Tinfoil's replaced Kimi model no longer sit in the picker returning model_not_found, and the canary suites that should have caught it can now actually fail. (#1722)
• A working Gemini Flash Lite again. Google retired gemini-2.5-flash-lite for new API keys; Gemini 3.5 and 3.1 Flash Lite are in the registry, so cleanup doesn't have to run on a heavier model. (#1702, thanks @xAlcahest)
• gpt-oss on Tinfoil works. Every request on all five LLM surfaces returned a 400 over an unsupported reasoning-effort value. (#1611)
• Request parameters have one source of truth. Token-limit names, temperature support, and reasoning-effort values are now declared per provider and model family instead of being duplicated across five transports — the recurring cause of a new provider-and-model combination breaking in production. Local model parameters and their canaries got the same treatment. (#1620, #1714)
• A blank model reply never eats your dictation. Whitespace-only cleanup, agent, or chain output preserves the spoken text on every route. (#1618, #1645, thanks @hsusul)
• Thinking tags stop leaking into your text. Nested <think> blocks are stripped correctly, streamed or not, so stray tags stop landing in titles and notes. (#1619, #1644, thanks @hsusul)
• Browsing provider tabs no longer switches your model. The provider-and-model pair commits only when you click a model. (#1288, thanks @IdrisGit)
• Transient failures back off instead of failing. HTTP 408 joins 429 in the retry strategy, and flaky DNS and broken pipes are reported as the network errors they are. (#1734, #1682, thanks @hsusul)
• Snippets and dictionary hints can't crash a dictation. An empty or missing list threw and dropped the paste. (#1672, #1708, thanks @hsusul)
• Your agent's name is matched case-insensitively in the dictionary. (#1639, thanks @hsusul)

Linux

• Push-to-talk on Wayland. Hold-to-dictate now works on Hyprland, KDE, and GNOME 48+, which previously only delivered a single toggle with no press or release. (#1738, thanks @IdrisGit)
• Hyprland's Lua config is supported. On Hyprland 0.55+ (and Omarchy), shortcuts persisted only to the deprecated hyprland.conf and vanished on reload. (#1664, thanks @IdrisGit)
• Auto-paste works on non-QWERTY layouts. Compositor-native symbolic shortcuts and keysyms replace hardcoded keycodes, and dictated text stays on the clipboard when automatic paste can't land. (#1525, thanks @IdrisGit)
• Punctuation hotkeys bind on GNOME and KDE. Control+, and friends silently failed to register. (#1658, #1752, thanks @hsusul)
• Ptyxis and GNOME Console paste correctly, like the other terminals. (#1659, thanks @hsusul)
• Paste stops stalling on KDE. A stale desktop-portal session cost up to 19 seconds on every paste until restart; failures are now remembered and skipped. (#1629)
• Sway overlays keep your text field focused, so auto-paste lands. (#1718)

Windows

• Dictation pastes into the window you recorded from. If another window took the foreground while transcription ran, the text went there instead; the captured target is now restored first. (#1725)
• Packaged builds stop printing logs to the terminal. Use --console-logs if you want them. (#1719)
• Whisper models work under non-ASCII and redirected profiles. A CJK or Cyrillic user name crashed the local transcription server; a redirected USERPROFILE hid downloaded models. (#1514, #1721, thanks @stantheman0128)

Elsewhere

• Turning off App updates turns off update checks. The toggle used to gate only the popup, so firewalled machines still got network attempts on every launch and error dialogs when Settings opened. (#1662, thanks @AdityaPainuli)
• A wedged background service can't spin forever. The reaper now escalates to SIGKILL and verifies the process actually died, instead of clearing its record after one ignored signal and spawning another alongside it. (#1626)
• Chinese UI for zh-Hans and zh-Hant system locales. Those tags fell back to English. (#1691, thanks @hsusul)

[1.8.3] - 2026-08-12

GPU acceleration for local transcription gets an overhaul: the status you see is now the truth, enabling it works without a restart, older NVIDIA cards are routed to a backend that actually works on them, and a GPU failure can never cost you a dictation. LLM routing gains the same fail-closed treatment speech-to-text received in 1.8.2. Launch at login arrives on Linux, and transcription errors stop blaming your microphone.

GPU acceleration

• "GPU acceleration active" now means it. The indicator reflects what the transcription server is actually running on — ready, activating, active, or "GPU could not be activated" with a Retry — instead of turning green whenever a file finished downloading. (#1578)
• Enable GPU applies immediately. Downloading or removing a GPU pack reloads the transcription engine in place; previously nothing changed until the app was restarted, with no hint. (#1578)
• A GPU crash never costs you a dictation — and is remembered. If the GPU engine fails, the same recording is retried on CPU and pasted, and the failed backend isn't silently re-attempted (and its model reload re-paid) on every launch. Retry or re-download clears the memory. (#1578)
• GTX 10-series (Pascal) cards get CUDA. The CUDA pack now ships Pascal kernels and is offered to those cards; a Pascal machine already running the Vulkan pack keeps it instead of being re-prompted to download a second backend. (#1585)
• NVIDIA cards below the CUDA build's floor are offered a backend that works. Cards the build carries no kernels for (Maxwell and older) were offered a CUDA build that loaded the model, then crashed at first use while reporting success. They now get the Vulkan pack, like AMD and Intel GPUs. (#1576)
• GPU packs can no longer corrupt each other. The whisper CUDA and llama Vulkan packs shipped identically-named runtime libraries into one shared folder, so installing one could silently break the other — and removing one deleted the other's files. Each pack now owns its own directory, and old installs are healed on startup. (#1577)
• An interrupted GPU install can no longer fake success. Installs are staged and swapped in atomically, so a crash or power cut mid-install can't leave a truncated binary the app reports as installed forever. (#1577)
• GPU binaries updated (whisper.cpp pack 0.0.9). Same engine and source as 0.0.8, rebuilt with the Pascal CUDA kernels included (+17.5 MB per CUDA pack). (#1584)

Speech to text

• Action required for some Custom endpoints. If you selected the Custom speech-to-text provider but never changed its pre-filled URL (https://api.openai.com/v1), your audio and your custom key were being sent to OpenAI. That now fails with a clear error instead, so set a real endpoint under Settings → Speech to Text — or switch to the OpenAI provider, which the URL field does automatically once you re-enter it. (#1556)
• "No Audio Detected" means silence again. A broken transcription engine used to blame your microphone; it now reports the real error, keeps the recording for retry, and still falls back to your cloud provider when enabled. (#1575)
• Re-transcribing a Corti recording reaches Corti. History → Re-transcribe sent Corti recordings to OpenAI. (#1556)
• A leftover provider can no longer hijack self-hosted transcription. With self-hosted selected, a stale Mistral, xAI, or Corti selection could still divert your audio to that provider. (#1556)
• Each provider remembers its own model. Switching speech-to-text providers and back restores the model you had chosen instead of resetting it to the default — including custom model names on your own endpoint. (#1556)
• Mistral file uploads authenticate correctly. Upload was the last path still sending a Bearer token instead of Mistral's x-api-key. (#1556)
• Azure OpenAI endpoints work on every path. Dictation, re-transcribe, and upload all build deployment-style URLs, whether the endpoint is configured as Custom or self-hosted. (#1556)
• Uploads detect their own language again. A file you upload is no longer forced into your dictation language. (#1556)
• Realtime streaming waits for the transcript tail. Stopping a streamed dictation no longer races a fixed delay against the last words arriving. (#1573)
• Recordings discarded as dictionary echo are kept and surfaced. (#1559)
• Parakeet accepts non-PCM16 WAV uploads. (#1376)

AI models & routing

• Custom LLM endpoints fail closed. An empty, unparseable, or non-OpenAI custom URL no longer falls back to sending your prompt — and your custom API key — to OpenAI; it fails with a clear, localized error. Unknown providers are rejected instead of routed to OpenAI. (#1583)
• Your cleanup key stays on the cleanup endpoint. The dictation-cleanup custom key no longer rides along to other scopes' custom endpoints. (#1583)
• Every LLM scope remembers its model per provider. Switching provider tabs and back restores your previous choice across all six scopes, and retired cloud models are repointed to the provider's current default instead of failing forever. (#1583)
• The voice-agent vision override works. Screenshot-carrying commands can actually route to the configured vision model, and are enforced under the correct policy scope. (#1583)
• Custom-endpoint API keys move to the OS secure store. Five scopes stored their keys in plaintext; existing keys are migrated automatically. (#1583)

Voice agent

• Screen context, selection edits, and calendar answers are more reliable. Screenshot capture failures retry text-only without losing the command, selection edits recover from a rejected screenshot, and the agent's calendar tool and accessibility reads are fixed. (#1566)
• Selection edits work in apps with dormant accessibility trees. Commands dictated into Dia, Arc, Chrome, Claude Desktop, Slack, or VS Code no longer die before reaching the model when the selected text can't be read natively. (#1593)

Meetings & speakers

• Your own dictation no longer triggers meeting detection. (#1570)
• Manual speaker reassignments win. A segment you reassign keeps your label instead of being pulled back to its diarization cluster. (#1569)
• Gap segments follow the nearest speaker. (#1423)
• Realtime meeting streams stop cleanly. (#1553)
• Expected speaker counts are validated in one place. (#1555, #1522)

Calendar

• Large Google Calendars sync completely. Sync now follows all result pages instead of stopping after the first. (#1572)
• Meeting reminders re-arm after a provider reset. (#1486)
• Calendar API failures are reported with their cause. (#1553)
• Meeting join links tolerate malformed calendar data. Whitespace-only or missing join URLs no longer break the join flow. (#1580)

Startup & Linux

• Launch at login on Linux. Via an XDG autostart entry, with correct behavior across GNOME and KDE autostart editors — plus start-hidden fixes on Windows and a hardened Linux launcher probe. (#1518, #1574, thanks @edwin-luu)
• Linux text monitor builds link AT-SPI2 correctly. (#1544, thanks @iSparsh)
• Meeting notifications stay clickable after the first hover on Linux. (#1562)

macOS

• The Globe hotkey no longer also triggers the system Globe action. (#1567)

Notes & interface

• Meeting transcript timestamps export correctly to Markdown. (#1560)
• Retired default prompts are cleared from persisted settings. (#1561)
• Empty states close their layout gaps. (#1565)
• The "Coming Soon" badge translates again. A wrong-prefix key rendered raw text in every language; a new lint-style test guards all referenced keys. (#1592)
• Hotkey parsing handles left/right modifiers and trailing + correctly. (#1437, #1433)
• Retention cleanup waits for the first renderer sync. A fresh install can no longer sweep history before settings arrive. (#1558)

Security & enterprise

• IPv6 private and metadata addresses are blocked for enterprise endpoints. (#1440)

CLI

• Validation errors return HTTP 400 with a structured body. (#1521)

[1.8.2] - 2026-08-11

Meetings get more reliable speaker identity — labels you set now stick, Windows loopback gains live identification, and Intel Mac meetings no longer fail when the optional ONNX binding is unavailable. The voice agent can edit highlighted text in place or use an opt-in screenshot as context. Collaboration is now free for signed-in users, Microsoft and Apple Calendar join Google Calendar, and organizations gain server-enforced policy plus centrally managed Bedrock and Azure OpenAI access.

Voice agent

• Edit highlighted text by voice. Select text anywhere, trigger the voice agent, and say what you want changed — the agent rewrites the selection in place instead of appending a new block. (#1264)
• Optional screen context. With "Share screen context" enabled (Settings → AI Models → Voice Agent, off by default), pressing the voice agent hotkey captures the display your cursor is on and sends it with your command, so you can say things like "reply to this email" or "explain the error on screen". You can route screenshot-carrying commands to a separate vision-capable model, and a screenshot that can't be sent never costs you the command — it silently reruns text-only and tells you.
• Agent failures are no longer silent. If the agent can't process a command, you get an "Agent Unavailable" notice instead of your raw words appearing in whatever app you were in.

Meetings & speaker identity

• Speaker labels stay put. Manual speaker labels persist, identities stay stable across a meeting, and reconciliation no longer deletes a mapping you set by hand when live and offline speaker ids happen to match. (#1501)
• Live speaker identification on Windows. Windows loopback capture now identifies speakers live, matching macOS. (#1502)
• Intel Mac meetings no longer fail at startup. ONNX Runtime no longer ships a macOS x86_64 binding, so live speaker identification and voice fingerprinting remain unavailable on Intel Macs; meeting recording, transcription, offline diarization, and keyword note search continue normally. (#1538)
• Diarization writes land on the right note. Delayed diarization is always persisted to the note it belongs to, and completions are serialized so labels don't get crossed between notes. (#1539, #1495)
• Tinfoil realtime meeting transcription. Tinfoil joins the realtime meeting providers.
• Live speaker identification degrades gracefully. A missing onnxruntime binding disables live identification instead of breaking the meeting.
• Steadier meeting prompts. Meeting detection re-evaluates gated microphone state, prompts no longer expire early, and the countdown is scoped to its own window. (#1532)
• Speaker timestamps anchored correctly. Live speaker timestamps anchor to the first system chunk, and the roster-driven speaker cap only ever raises.

Organization policy

• Every policy field is enforced. Organization policy is now applied across the whole app rather than just the settings UI: restricted options are hidden with safe fallbacks, enforcement covers modes, providers, features, sharing, and retention, and a malformed policy response fails closed. (#1074, #1506)
• Managed Enterprise AI. Enterprise workspaces can centrally configure Amazon Bedrock or Azure OpenAI for cleanup, voice agent, note formatting, note chat, and translation. Employees sign in with company SSO instead of entering cloud keys; short-lived credentials stay in the desktop main process and prompts go directly to the organization's cloud account. (#1530)

Dictation

• No more clipped first words. Cold microphone opens could swallow the beginning of a recording; the mic is warmed so your first words are captured. (#845, #1493)
• Dropped transcript segments are retried and surfaced. Silently dropped segments are retried, and a genuine loss is reported rather than quietly shortening your transcript. (#1462)
• VAD is now opt-in. Voice activity detection is off by default, with dictionary-echo decodes rescued rather than discarded.
• Transcripts are no longer replaced unexpectedly. (#1461)

Notes & sync

• Collaboration is free. Any signed-in account can create and join team spaces, share individual notes, and sync shared content even when personal cloud backup is off. New users who missed an invitation email can join when invited or verified through company SSO; otherwise they can request access from workspace admins. (#1549)
• Workspace invites no longer stop at prepaid capacity. Free workspaces can add members without a subscription. Paid workspaces show the estimated prorated seat cost before an invitation is sent and increase Stripe capacity only if it is accepted. (#1549)
• Idle collaboration sync backs off. Accounts with no shared activity gradually reduce background checks to once per hour, while manual syncs and active collaboration remain immediate. (#1549)
• Tombstones survive. Note pulls and deletes are guarded so a tombstone can't be lost and a deleted note resurrected. (#1354, thanks @xAlcahest)
• Edits during a push are not lost. A note edited while its save is being acknowledged stays pending instead of being marked clean.

Account & billing

• Edit your profile. Change your display name in the app; email-and-password accounts can also change their password and optionally sign out other devices. Email address changes still go through support. (#1162)
• Correct plan shown for covered workspaces. Members covered by a workspace plan no longer see "Free", and an unknown entitlement is never reported as a free plan. (#1540, #1424)
• Current Business pricing. Plan cards now show $16 per user monthly or $160 per user annually. (#1551)

Calendars

• Apple Calendar support. Native EventKit integration joins Google Calendar for meeting detection and reminders. (#1237)
• Microsoft Calendar support. Connect Outlook.com or Microsoft 365 on macOS, Windows, or Linux. Microsoft Graph sync adds scheduled-meeting prompts, titles, attendees, and join links alongside Google and Apple Calendar events. (#1251)

Models

• Claude Opus 5. Added to the Anthropic provider. (#1452)
• Retired model ids removed from the registry. (#1482)
• Anthropic temperature compatibility. Models that reject a temperature parameter no longer fail. (#1475)

Translations & internationalization

• Chinese locales corrected. Spanish and untranslated strings that leaked into the Chinese locales are fixed. (#1504)
• Simplified vs Traditional Chinese respected for Chinese speech-to-text. (#1226)
• Speaker labels are translated, including "unknown speaker" and your own voice in exported transcripts.

Other fixes

• Security updates. Better Auth and affected packaged dependencies were updated to patched releases; the release dependency audit now reports no known vulnerabilities.
• More reliable cloud audio uploads. A poisoned TLS connection is discarded before retrying, and upload diarization never requests more speaker clusters than the local model supports. (#1496)
• Custom STT endpoint URLs are preserved when switching provider tabs. (#1463)
• Hotkey slots release their accelerators when unregistered, so a rebound hotkey no longer leaves the old one dead. (#1425, #1420, thanks @hsusul)
• Tailscale MagicDNS works over HTTP, and self-hosted mode is allowed for uploads.
• Local translation models with an empty provider route through llama.cpp correctly.
• The dictation agent's inference mode is authoritative, with providers normalized for local and self-hosted modes.

[1.8.1] - 2026-07-30

A critical fix on top of 1.8.0. If you installed 1.8.0, update to 1.8.1 right away.

Fixed

• Note content no longer leaks between notes. In 1.8.0, switching from one note to another could copy the second note's content into the first — even when you edited neither — because the editor re-mounting on a note switch registered a phantom edit against the note you'd just left. Editor edits and pending saves are now tied to the specific note they came from, and switching notes no longer emits a spurious change. (#1406)

[1.8.0] - 2026-07-30

Team spaces and web note sharing headline this release: create shared spaces inside your workspace, invite teammates with roles and one-click deep links, and publish notes to the web with link, domain, or invite-only visibility — all local-first, with access enforced server-side and revocation handled gracefully. Alongside it: workspace management lands in Settings, new Gemma 4 local models, configurable privacy retention, custom-dictionary management over the CLI, and a broad reliability pass across transcription, notes sync, translation, exports, and per-platform polish.

Team spaces & sharing

• Team spaces in the notes sidebar. A new TEAM SPACES section sits alongside your private space: create a space with a name and emoji, organize per-space folders and notes, drag notes between containers, and drive the whole tree from the keyboard. The editor shows where a note lives — a breadcrumb chip that jumps to and reveals its space, and an audience pill showing how many people can see it. Content is local-first: your notes stay in the on-device database as the source of truth, with each space syncing to the cloud in the background. (#925)
• Membership, roles, and invitations. Invite teammates by email with team spaces pre-selected; the invitation email lands on notes.openwhispr.com with a one-click "Open in OpenWhispr" deep link straight into the accept flow (and an honest copy-the-link path on phones). Space members are admins or members, workspace owners and admins hold implicit admin access to every space, and rosters are managed from the space's Members dialog or Settings → Workspace → Team spaces — with type-the-name confirmation before a space is deleted, everywhere it can be deleted from. (#925)
• Server-enforced access with graceful revocation. Membership checks live on the server — non-members get existence privacy, and revoked or archived spaces return typed errors the app understands. Losing access to a space removes its synced notes locally and relocates anything you never managed to sync into your private space, with a toast explaining what happened. If a teammate's edit arrives while you have unpushed changes, a conflict banner names who edited it and lets you take their copy or keep yours. (#925)
• Share notes on the web. Personal notes can be published to notes.openwhispr.com under three visibilities: anyone with the link, anyone signed in under your email domain, or invited people only. Rich link previews (title and snippet) are emitted only for link-visibility shares — restricted notes unfurl generically so nothing leaks into chat preview caches. Team notes deliberately have no public link; their audience is the space. (#925)
• Changes land live, everywhere. Space changes now broadcast across windows the way notes and folders already did: a teammate's rename, a role or member-count change, or a freshly accepted invitation appears in an open sidebar without remounting, skeleton rows clear as a new space's content finishes backfilling, and a deleted space vanishes from every window at once. (#925)
• Workspace management in Settings. The new Settings → Workspace section covers members (invite, remove, roles), team spaces, billing with per-seat counts, and developer settings — with workspace creation and switching built in. (#925)

Transcription & reliability

• Local transcription no longer splits words mid-word. Segment boundaries could cut a word in half and drop a fragment on either side; segments now break on natural gaps so words stay intact. (#1348, #1369)
• Local transcription timeouts scale with audio length. A fixed timeout could kill long recordings before they finished decoding — the limit now grows with the duration of the audio. (#1104)
• Sturdier chunked cloud uploads. Chunked cloud transcription gained per-request timeouts, a fresh-connection retry, proper cancellation, and a global cap so a stuck upload can no longer hang the pipeline. (#1330)
• Your transcript survives an empty cleanup. If the cloud cleanup pass returned nothing, the whole transcription used to come back blank; the raw transcript is now kept when cleanup yields no text. (#938)
• Recordings are recovered when the mic disconnects. Unplugging or losing the microphone mid-recording no longer discards what you already said. (#1261)
• Steadier NVIDIA Parakeet streaming. Streaming dictation falls back to a full decode if the live stream loses audio, and it keeps the newer partial result instead of an older one — so the committed text matches what you said. (#1283, #1285)
• Cleaner AssemblyAI streaming. Premature stream closes from AssemblyAI are now rejected rather than treated as a finished transcript. (#1240)
• Pinned, verified CUDA Whisper binaries. The CUDA whisper.cpp binaries are pinned to known versions and checked against their digests before use. (#1372)

Local models

• Gemma 4 QAT local models. Added the quantization-aware-trained Gemma 4 models, including MTP drafters for faster local generation. (#1268, #1269)
• Local llama-server pre-warming restored. The local LLM server warms up ahead of time again, so the first local reasoning request isn't slow. (#1363)
• Stale local model selections are cleared. Removing a downloaded model no longer leaves it selected as your active model. (#1312)
• Download state persists across the picker. A model download keeps its progress when the model picker remounts, instead of appearing to reset. (#1052)
• Correct model download filename. A local model download filename now matches its HuggingFace link so the download succeeds. (#1292)
• Canceled downloads clean up properly. Canceling a model download waits for the file to finish closing, so it no longer leaves a partial file behind. (#1252)

Translation

• Local translation routed through llama.cpp. Dictation translation with a local model now runs through the llama.cpp path like the rest of local reasoning. (#1327)
• Empty translation results are handled. An empty result from the translation chain is guarded instead of surfacing as a broken response. (#1258)

Notes

• Chat text selection survives live transcription. With meeting transcription running in Notes, every incoming transcript update could yank the text selection out of the embedded chat; selections in chat are now left alone while the transcript streams. (#1027)
• Cloud sync won't wipe locally-edited notes. A stale, empty copy arriving from the cloud can no longer overwrite a note you've edited locally. (#1291)
• Mirrored notes are distinguished from transcripts. Notes mirrored from elsewhere are no longer confused with transcription records. (#1276)
• Fixed a note race condition. (#1277)

History & export

• See and copy the full raw transcript. History shows the complete raw transcript with a dedicated "copy raw transcript" action. (#1289)
• History groups by your local date. Entries are grouped by local calendar day instead of UTC, so late-evening items land on the right day. (#1272)
• Meeting participants included in exports. Exported meeting transcripts now carry their participant list. (#1274)
• Cleaner exported segments. Consecutive speaker-less segments within the gap threshold are merged, and SRT millisecond overflow rolls correctly into seconds. (#1335, #1294)

Custom dictionary

• Manage your dictionary from the CLI. The CLI bridge can now list dictionary entries and apply bulk updates. (#1366)
• Fresh dictionary on startup. The dictionary is read from SQLite rather than a stale in-memory cache at launch. (#1304)

Privacy & security

• Configurable retention, actually enforced. Added a 1-day option to audio/transcript retention and made the configured retention periods actually take effect, so saved recordings and transcripts are cleaned up on schedule. (#1368)
• Stricter private-endpoint validation. Private HTTP endpoints (self-hosted / LAN) must use IPv4 literals, closing an SSRF-style hostname bypass. (#1318)

AI, providers & settings

• Prompt Studio tests the right scope. The agent-prompt test now runs against the cleanup scope and echoes its input, so what you test matches what runs. (#1359)
• Clearer API error messages. Top-level and string detail fields are extracted from provider error bodies instead of showing an opaque error. (#1339)
• Pasted base URLs are parsed correctly. Paths are joined before query strings, so a base URL with a query string no longer breaks the request. (#1310)
• Inherited endpoints keep their key. Selecting an inherited endpoint now carries the associated API key. (#1364)

Audio upload & import

• Opus uploads transcribe correctly. YouTube imports often arrive as .opus files, which were uploaded with a generic MPEG content type and rejected by cloud transcription as an unsupported format. Opus (and a filename-first media-type check) is now part of the upload pipeline, so short YouTube links and direct Opus files transcribe instead of erroring. (#1249)

Interface

• The macOS Dock icon follows the control panel. The Dock icon appears when the control panel opens and disappears when it closes to the tray — and nothing else can resurrect it, which fixes the icon reappearing after dictation. The separate "Show Dock Icon" setting is gone; the Dock now simply reflects whether the panel is open. (#1202)
• The tray icon toggles the control panel. Clicking the tray icon hides the panel if it's visible and shows-and-focuses it otherwise (previously it could only open it), with a matching toggle item in the tray menu. Especially useful on Linux, where a window parked on another workspace no longer gets stuck out of sight. (#1133)

Windows

• No more console window flashes. Native helper launches across dictation, meetings, GPU probes, and media control — where.exe lookups, nvidia-smi GPU detection, media pause/resume and paste helpers, the meeting AEC, text-edit monitor, and mic listener — now spawn hidden instead of flashing a console window mid-flow. (#1228, #1232, #1233, #1234)

Linux

• Wayland dev/runtime parity. The Wayland development server is aligned with packaged behavior and runs under XWayland. (#1323)
• More reliable portal paste. Modifier keys settle before the key press in portal-based paste, fixing dropped or garbled pastes. (#1257)
• Meeting detection degrades gracefully. Detection falls back to polling when pactl or a native mic listener isn't available. (#1350)

macOS

• Accessibility guidance when the app isn't listed. The permissions flow now handles the case where OpenWhispr is missing from the Accessibility list entirely. (#1302)

[1.7.6] - 2026-07-18

A feature release across the entire app on top of 1.7.5: dictation translation with its own hotkey and dedicated model, audio import that takes YouTube and direct URLs plus batch uploads with on-device speaker detection, one-click Vulkan GPU acceleration bringing local Whisper GPU support to AMD and Intel, NVIDIA Nemotron streaming models that decode dictation live and commit the moment you stop, Liquid AI LFM2/LFM2.5 local reasoning models, a collapsible sidebar, and meeting prompts unified into the in-app overlay — plus a deep reliability pass on self-hosted routing (uploads, retries, note formatting, chat), Mistral and Groq cleanup, local LLM memory safety, media pause/resume on macOS 15.4+, and Windows/Linux packaging.

Dictation

• Translation mode with a dedicated hotkey. A new Translation Hotkey (Settings → General → Translation Hotkey) starts a dictation that gets cleaned up and then translated before it's pasted — dictate in any language and the text comes out in your chosen one. Configure it under the new Settings → AI Models → Translation tab: an "Enable Dictation Translation" toggle, a "Spoken language" selector (Automatic also handles mixed-language dictation), up to 5 target languages with one active target you can switch between, a dedicated model/provider just for translation, and a customizable Translation Prompt with its own translate-aware test tab in the prompt studio. The hotkey works across every binding type (including the macOS Globe key, mouse buttons, and right-modifiers), translated entries carry a marker in History and can be re-run through the translation chain from the retry menu, and if translation is unavailable or fails your dictation is pasted untranslated with a visible toast instead of being lost. (#1189)
• Auto-learn no longer double-initializes at launch. Both renderer windows sync the auto-learn preference to the main process on mount, so the handler received the same value twice per startup, redoing its setup each time. Repeated same-value syncs are now recognized and ignored, while a real enable or disable behaves exactly as before. (#1109)

Transcription

• Vulkan GPU acceleration for AMD and Intel GPUs. Local Whisper's one-click "Enable GPU" flow, previously NVIDIA/CUDA-only, now covers AMD Radeon and Intel Arc/integrated GPUs on Windows and Linux via Vulkan: the GPU card in the transcription model picker and the "GPU acceleration available" banner automatically offer CUDA on NVIDIA machines and Vulkan otherwise. The Vulkan whisper-server binary is downloaded on demand with pinned, SHA-256-verified checksums, and any Vulkan startup failure — early exit, out-of-VRAM, or hang — falls back to CPU transcription automatically with an in-app notice. GPU-accelerated Whisper also re-warms correctly after sleep on both backends now; macOS is unaffected (it already uses Metal). (#1185)
• NVIDIA Nemotron streaming transcription models. Two new local models join the NVIDIA Parakeet family in the transcription model picker: Nemotron Speech Streaming EN 0.6B (632 MB, English) and Nemotron 3.5 ASR Streaming 0.6B (650 MB, multilingual with automatic language detection across 15 languages including Spanish, German, Japanese, Korean, Arabic, and Hindi). These are true streaming models, so the Live Transcription Preview now holds one persistent connection for the whole recording and updates partial text as you speak, instead of re-transcribing 1.5-second chunks — with automatic fallback to the chunked path if the stream can't start. The bundled sherpa-onnx runtime was upgraded to 1.13.4, which Nemotron requires for accurate decoding. (#1131)
• Streaming dictation now commits in a single pass. With a Nemotron streaming model selected, dictation is decoded live during capture — with or without the preview open — and the streamed text is committed the moment you stop, replacing the full re-decode that used to run after every dictation. End-of-dictation latency drops to one tail flush and per-dictation CPU is roughly halved. The stop flush is truncation-aware, and anything short of a clean flush — a dropped connection, truncated results — falls back to the proven record-then-transcribe path (surfacing the partial-transcription warning where applicable). The online sherpa-onnx server is also properly tuned now (ONNX threads sized to CPU cores instead of one, tighter decode loop, tail padding), server startup is race-free under concurrent callers, sidecars start on IPv6-disabled machines, corrupt model archives are deleted and re-downloaded instead of re-extracting forever, and non-16 kHz/mono WAV input is resampled instead of decoded as-is. (#1238)
• Local transcription survives a GPU crash instead of failing the dictation. On some machines the CUDA (or Vulkan) whisper-server aborts at its first kernel launch — an unsupported GPU, for instance — dropping the connection mid-request and failing the dictation outright. A connection error from a local GPU server is now checked against the server process actually having died; on a real crash, whisper-server is restarted on CPU, the same request is retried, and a "using CPU instead" notice appears once the CPU server is up. Intentional stops, remote servers, and model switches are left alone. (#1192)
• Retrying a transcription from History now uses your self-hosted server. Retranscribing audio from History ignored self-hosted transcription settings and routed through stale cloud-provider state and credentials. Retries now normalize your configured URL to an OpenAI-compatible /audio/transcriptions endpoint (base URLs, /v1 URLs, and full paths all accepted), fail closed with a clear configuration error when the URL is missing or malformed instead of silently falling back to the cloud, and never fetch unrelated cloud API keys — with the same fail-closed, key-forwarding behavior extended to chat and dictation streaming. (#1154)

Audio upload & import

• Transcribe from URLs, batch uploads, and speaker detection. The Upload view now takes links, not just files: paste a YouTube link (including youtube.com/live) or a direct HTTPS audio/video URL, or switch to batch mode to queue up to 50 URLs (one per line) alongside multiple dragged-in files — the queue processes sequentially with per-item progress, "Cancel all"/"Clear" controls, and keeps running even if you navigate away. URL downloads land in a new default "Videos" folder and are capped at 500 MB; playlists aren't supported, and a blocked YouTube download tells you to retry later instead of failing cryptically. A new "Speaker detection" toggle labels who said what: it auto-downloads the local diarization models on first use and runs on-device, with an optional speaker count (auto-detect by default); with your own API keys, OpenAI switches to gpt-4o-transcribe-diarize (costs may differ) and Mistral's Voxtral includes diarization at no extra cost, while Groq falls back to local detection. If diarization fails you still get a plain transcript rather than a failed upload. (#754)
• Uploading audio files now respects self-hosted transcription mode. With transcription set to self-hosted, uploaded audio either failed with 401s or was sent to a cloud provider you had switched away from. The upload view only routed cloud, local, and BYOK modes, so self-hosted fell into the BYOK branch carrying stale cloud provider settings. Uploads now resolve the self-hosted route first — failing closed with a clear error if the server URL is missing or invalid — post to your server without the 25 MB third-party size cap, and the upload UI reflects self-hosted readiness and labels. (#1180)
• Exported transcripts keep correct start times for merged segments. When back-to-back segments from the same speaker were merged for export, each addition overwrote the group's timestamp, so TXT, Markdown, SRT, and JSON exports stamped every merged block with the time of its last sentence instead of when the speaker started talking. Merged groups now track their start and end separately: timestamps and SRT cue starts reflect the beginning of speech, while the final cue's end time and the JSON duration still anchor to the end of the last merged segment. (#1194)
• Failed or cancelled URL downloads no longer leave partial files behind. When an audio download from a URL was cancelled, stalled, or produced no output, cleanup destroyed the write stream and deleted the temp file immediately — but destroy() returns before the stream's async open finishes, so a still-pending open could recreate the file right after it was removed, orphaning a partial download on disk. Cleanup now waits for the stream's close event before unlinking, so the temp file is reliably gone before the error is reported. (#1212)

Reasoning & models

• Liquid AI LFM2/LFM2.5 local reasoning models. Liquid AI is now a local model provider for on-device AI processing (cleanup, formatting, and other LLM features), with five models downloaded from the official LiquidAI Hugging Face repos: LFM2.5 1.2B (0.73 GB, recommended for its instruction following), LFM2.5 8B MoE (4.8 GB, high quality with only 1.5B active parameters), LFM2 2.6B (1.5 GB), LFM2.5 350M (0.38 GB), and LFM2.5 230M (0.25 GB, instant cleanup on modest hardware). All run on the existing local GGUF runtime with a 32k context window, under the LFM Open License v1.0 (free for commercial use below $10M annual revenue). (#1176)
• Local AI cleanup no longer freezes the whole machine. Starting local LLM cleanup or note formatting could hard-freeze the Mac — on 16 GB machines it escalated to a watchdog kernel panic, even with a small 2 GB model. The bundled llama-server was spawned with no --ctx-size while auto-fit was disabled, so it allocated the model's full trained context (131K tokens for Llama-3.2-3B) — a ~15 GB KV cache wired into unified memory on top of the weights, which starved the OS until it panicked. Every server start now uses one bounded context policy (16,384 tokens, capped at the model's trained context) across inference, prewarm, and GPU-change restarts, and a model whose weights genuinely don't fit now fails at startup with a visible error instead of taking down the machine. (#1236)
• Self-hosted model lists now work when the server URL is missing /v1. Pointing OpenWhispr at LM Studio, Ollama, or vLLM with a bare origin like http://127.0.0.1:1234 showed no models, and LM Studio's native /api/v1 base "succeeded" but returned entries with no usable model ids. Those servers serve the OpenAI-compatible API under /v1, and while inference already normalized URLs to /v1, the model-list UI never did. When the entered base yields no models, the panel now tries the sibling /v1 base and adopts the one that works — updating the saved URL and the input field so inference targets the same endpoint. (#1235)
• Mistral now works as a custom cleanup provider instead of failing with a 422. Adding Mistral as a custom OpenAI-compatible provider made every cleanup request fail with an error that surfaced as "[object Object]". Mistral validates requests strictly and rejected two fields OpenWhispr always sent — max_completion_tokens and chat_template_kwargs — and the error parser assumed a string message where Mistral returns a structured detail array. Mistral endpoints are now recognized by host and get their own request dialect (max_tokens, temperature kept, reasoning_effort: "none" as the thinking switch, and nothing at all for legacy Magistral models, which reason natively) on both the cleanup and agent/chat streaming paths, and API error bodies are parsed properly so failures name the rejected fields. (#1213)
• Groq dictation cleanup no longer fails silently after a pointless retry wait. With the default "disable thinking" setting on, every Groq cleanup request failed with a 400, sat through a 1+2+4 s retry backoff as if it were a network blip, then quietly pasted the raw transcript. The suppression code sent reasoning_effort: "none" plus chat_template_kwargs to Groq — gpt-oss models only accept low/medium/high, qwen3 only none/default, and chat_template_kwargs isn't supported at all — and thrown API errors carried no HTTP status, so deterministic 4xx errors were retried like network faults. Groq now gets a model-aware dialect, retries are reserved for network faults, 429s, and 5xx, and if cleanup still fails a toast tells you the dictation was pasted unpolished instead of leaving you guessing. (#1211)
• "Disable thinking output" now actually works on self-hosted Ollama. Qwen and similar reasoning models on self-hosted endpoints kept generating hidden thinking tokens on every dictation (~570 per cleanup request) even with the toggle on. OpenWhispr sent Ollama's native think: false field to the OpenAI-compatible /v1/chat/completions endpoint, which silently drops unknown fields, so the toggle did nothing. It now sends reasoning: {effort: "none"}, which Ollama v0.12.4+ maps to disabling thinking, while llama.cpp, vLLM, and LM Studio ignore the object and keep their existing mechanism; older Ollama releases that reject the field get one automatic retry without it. (#1103)

Chat

• Chat with custom BYOK providers no longer 404s. Chat agents pointed at custom OpenAI-compatible endpoints (LiteLLM, vLLM, DeepSeek, llama-server proxies) returned 404s or empty responses. The AI SDK's default OpenAI factory targets the Responses API (POST /responses), which those endpoints don't implement. Custom providers now go through the Chat Completions endpoint, the same way the OpenRouter, Corti, and local routes already did. (#1181)
• Chat no longer borrows Dictation Cleanup's endpoint, and self-hosted Chat keys are kept. With Chat set to a BYOK custom provider or an authenticated self-hosted server, conversations either inherited the unrelated Dictation Cleanup self-hosted endpoint or dropped the Chat server's API key mid-request. The Chat route is now resolved exclusively from Chat-owned settings — an explicit self-hosted URL wins, then enterprise, provider, or local — and the self-hosted API key is preserved in both the direct Chat Completions stream and the tool-enabled AI SDK stream. (#1146)

Notes

• Generate Notes with self-hosted formatting no longer silently calls OpenAI — and note names finally regenerate. With Note Formatting set to self-hosted, Generate Notes quietly sent note content to api.openai.com because the background action only forwarded endpoint settings for cloud and custom modes; self-hosted fell through with no provider, so the model name resolved to OpenAI and the configured server URL and key were never passed. Formatting and title generation now route to your self-hosted server (failing fast with a clear error if no URL is configured), <think> reasoning blocks are stripped from non-streaming self-hosted responses, and auto-assigned note names ("Untitled Note", "New note", unedited calendar-event titles) are regenerated while titles you typed yourself are preserved. (#1156)
• Note search now works in every language. Searching notes in Cyrillic, Chinese, Japanese, Arabic — or even accented Latin words like "café" — returned nothing. The search sanitizer stripped "special" characters with JavaScript's ASCII-only \w class, deleting non-ASCII text before the query ever reached the database. Queries are now tokenized with Unicode-aware matching, and each token is safely quoted and prefix-matched in SQLite FTS5 — so "Прив" finds "Привет мир" and "東京" finds "東京駅" — while punctuation-only input and raw FTS5 operators no longer produce invalid or match-all queries. (#1178)

Meeting notes

• Meeting reminders join the in-app overlay. Calendar reminders no longer arrive as native macOS notifications — which Focus/Do Not Disturb silenced and screen sharing hid — they now appear in the same always-on-top overlay as microphone-based meeting detection, one minute before the scheduled start. When the event has a meeting link, the primary button becomes "Join," which opens the meeting and starts the note in one click; otherwise it's the usual "Start Recording." The prompt shows the event title with context-aware copy, mic detection keeps working during scheduled meetings instead of being suppressed by them, and the overlay is content-protected so it never shows up in your screen shares or recordings. Your existing notification preferences for calendar reminders and meeting detection still apply, and the copy is localized in all ten languages. (#1182)
• Ignoring a calendar meeting reminder no longer suppresses the "take notes" prompt for that call. A reminder that expired unanswered went through the same path as clicking X, starting the mic detector's 5-minute cooldown — and since the mic stays hot for the whole call, joining a couple of minutes after the reminder meant no mic-activity event could ever prompt you again. Now only an expired audio-detection prompt cools the detector; a lapsed calendar reminder leaves mic detection armed, while explicitly dismissing any prompt still triggers the cooldown. The prompt buttons were also clarified in all 10 languages: "Join & transcribe" on calendar reminders and "Take notes" on detection prompts. (#1196)

Interface

• Collapsible sidebar with hover-to-peek. The main window's sidebar can now be collapsed to give notes and history the full width: a panel toggle button sits in the top-left corner (next to the traffic lights on macOS), and the collapsed state persists across launches. While collapsed, hovering the left edge or the toggle slides the sidebar in as a floating overlay so you can jump between Home, Chat, Notes, Dictionary, Upload, and Integrations without expanding it — it tucks itself away again shortly after the pointer leaves. The toggle is hidden in the compact meeting side-panel layout, and the button is labeled in all ten languages. (#1011)

Hotkeys

• Fn navigation shortcuts no longer turn push-to-talk into stray dictation on macOS. With bare Fn as push-to-talk, pressing Fn+Arrow (Home/End/Page Up/Page Down) or Fn+Backspace kept recording for the whole Fn hold and then transcribed the captured noise into the focused app — the globe listener only watched modifier-flag changes, so it never knew another key had been pressed. The Swift listener now also monitors keyDown and emits an interrupt once per Fn hold, which cancels the in-flight bare-Fn push-to-talk through the existing cancel pipeline; compound Fn hotkeys and tap mode are unaffected. (#826)

Audio & media

• Pausing your music during dictation works again on macOS 15.4+. macOS 15.4 closed the MediaRemote daemon to ordinary processes, and the old fallback typed an F8 keystroke that isn't a media key on modern Macs, so every player silently ignored pause/resume. OpenWhispr now bundles the mediaremote-adapter framework (v0.7.6, BSD-3) and drives it through /usr/bin/perl — the one path Apple still trusts to reach MediaRemote — with pause/resume running asynchronously off the main thread and a rewritten Swift helper that posts a genuine media-key event as the last resort; the macOS release pipeline also verifies the packed framework and its perl entry point in every build. (#806)
• A muted microphone no longer produces silent recordings. If your preferred mic never delivered audio (hardware mute switch, dead device after idle), the single re-acquire retry usually reopened the same silent device and the recording proceeded, capturing nothing. The mic health check now hops to the OS default input when the retry is still silent, remembers the rejected device for the rest of the session so it isn't pinned again, and if no input delivers audio at all the recording fails fast with a "Microphone Muted" error instead of transcribing silence. (#1204)
• Your microphone selection survives Chromium device-ID rotation. The chosen mic was saved by device ID alone, but Chromium rotates media device IDs over time, so the saved ID eventually matched nothing and recordings quietly fell back to the system default until you re-picked the device. The mic's label is now stored alongside its ID, and when the ID no longer resolves, the app remaps to the unique device carrying that label and re-persists the fresh ID — for dictation, meeting recordings, and the Settings picker alike. (#1172)

Windows

• Parakeet installation no longer hangs behind an old PATH tar. Model extraction now invokes Windows' built-in tar.exe directly and times out a stuck native extractor so the existing JavaScript fallback can finish the installation. The same hardened extraction path is used for diarization model setup and for CUDA whisper / Vulkan llama.cpp runtime archive extraction. (#1173)
• The Windows mic-activity listener finally ships. The native helper that watches when other apps start using the microphone (which drives meeting detection) never compiled under MSVC — ten unresolved externals, because user32 wasn't linked and the WASAPI COM GUIDs are declared by the Windows SDK for C++ __uuidof only — so every Windows build silently fell back to mic polling. The GUIDs are now defined in-source and user32 is linked, so Windows builds include the event-driven listener. (#1179)

Linux

• .deb installs, upgrades, and removals no longer fail or eat your models. The package's maintainer scripts ran under set -e, so any hiccup — no udev or systemd (containers, chroots), a GUI installer with no SUDO_USER or tty — could fail the whole package operation, and the post-remove script also ran during upgrades, deleting your downloaded Whisper models. The scripts are now strictly best-effort, environment-specific steps are skipped where udev or systemd don't exist, and cache cleanup runs only on a genuine remove or purge — never on upgrade. (#1160)

Updates

• Update notifications reach windows opened after launch. The updater captured window references once at startup, so a control panel created later (start minimized) or recreated after being closed held stale references and never received "update available", download progress, or "ready to install" events. The updater now reads the live window references from the window manager each time it sends an event, so every open window hears about updates regardless of when it was created. (#1175)

Dependencies

• Dependency updates. Routine patch bumps: vite 8.1.3 → 8.1.4, tinfoil 1.1.7 → 1.1.8, and i18next 26.3.4 → 26.3.6. (#1147, #1150, #1151)

[1.7.5] - 2026-07-10

A model-breadth and reliability release on top of 1.7.4: the latest cloud reasoning models (OpenAI GPT-5.6 and Anthropic Claude Fable 5 / Sonnet 5), Corti as a private clinical reasoning provider with in-region routing that keeps transcribed medical text off third-party LLMs, Tinfoil confidential transcription extended to uploaded audio and every batch path, OpenRouter as a first-class LLM provider with a searchable model picker, enterprise Agent Mode chat with region-aware Bedrock and a live model catalog, multiple hotkeys per action, and a broad stack of meeting-notes, notes, transcription, and platform fixes.

Reasoning & models

• New cloud reasoning models. OpenAI's GPT-5.6 family — Sol (flagship, 1M context), Terra (balanced), and Luna (fastest and lowest-cost) — and Anthropic's Claude Fable 5 (Mythos-class flagship, 1M context) and Claude Sonnet 5 are now selectable in their model pickers. Claude Fable 5 is also offered as an AWS Bedrock enterprise inference profile, and the Claude Opus 4.8 description was re-toned now that Fable 5 is the most capable Claude model. (#1130)
• Corti — clinical-grade AI reasoning (BYOK). Corti, added for medical transcription in 1.7.3, is now also a bring-your-own-key reasoning provider built on the Corti Models gateway. When a healthcare user accepts Corti during onboarding, dictation cleanup and every LLM scope route to Corti so transcribed clinical text never reaches a general-purpose model, and Corti's availability now counts toward having a reasoning provider (cleanup was previously skipped silently for Corti-only setups). The gateway is EU-only and speaks Chat Completions, noted at the top of the Corti reasoning tab. (#1111, #1127)
• Clinical reasoning stays in-region. The Corti Models gateway is EU-only and needs its own key, so a US clinical user used to get Corti transcription while every LLM scope stayed on the default provider (OpenAI) — sending transcribed clinical text to a third party. Onboarding now routes reasoning to Corti only in the EU region with a key, and to the HIPAA-compliant OpenWhispr Cloud everywhere else, never a third-party LLM. The region selector drives which fields appear, and the reasoning model selector flags the EU-only requirement in all ten locales. (#1128)
• OpenRouter — first-class LLM provider. OpenRouter is now its own cloud provider tab with its own encrypted API key, instead of requiring the generic Custom OpenAI-compatible field (which leaked the shared custom key on a tab switch). Its 300–400 models render in a new searchable, grouped model picker — filter, grouping by provider with per-vendor icons, virtualized rows, and a pinned "Selected" group — which also kicks in for any OpenAI-compatible provider above 12 models. Outbound partner links (BYOK key pages, HuggingFace model pages) now carry OpenWhispr UTM attribution. (#1002)
• Enterprise: region-aware Bedrock, a live model catalog, and Agent Mode chat. AWS Bedrock inference profiles are geo-scoped, but suggested model cards always used us.-prefixed IDs, so non-US regions rejected them with "the provided model identifier is invalid"; model IDs are now region-aware and rewritten when you change region. A new "Browse all models" button loads the live Bedrock catalog resolved against your own credentials and region, so a picked model is always invocable. Agent Mode chat — the tool-using agent overlay — now works with AWS Bedrock through an IPC-proxied model, and the stop button and window unmount actually cancel enterprise streams now instead of billing to completion. Suggested Bedrock models were refreshed to Claude Sonnet 5 and Claude Opus 4.8 and gained on-demand GPT-OSS 120B, DeepSeek V3.2, and Qwen3 Next 80B; enterprise inference now uses the calling scope's provider instead of always reading the cleanup provider. (#1118)

Transcription

• Tinfoil — confidential transcription for uploaded audio and retries. Tinfoil was wired only into the realtime dictation socket, so audio upload, retries, and the streaming fallback all fell through to api.openai.com — sending the Tinfoil key, or a user's audio under a valid OpenAI key. Every batch path now runs through the attested Tinfoil client (Voxtral), and the streaming fallback routes to the user's configured provider instead of always calling OpenWhispr Cloud. One Voxtral model now covers both realtime and batch, error codes survive the proxy boundary (so INVALID_KEY drives the Settings CTA and 429/5xx map to localized messages), and the custom dictionary is forwarded as a prompt. (#1120)
• Tinfoil model list is now dynamic. The Tinfoil catalog is fetched from /v1/models (verified against enclave attestation) and refreshed in the background, with three bundled models as an offline fallback. A refresh that retires a model you're using switches you off it with a toast rather than silently 404-ing, the fetched list is cached across launches, and new users start on a named default instead of whatever the list happens to return first. (#1115)
• Cloud recordings retry transient chunk failures. Long cloud recordings are split into ~4-minute chunks, and a single transient failure (Vercel timeout, 5xx) used to permanently drop a chunk — a silent hole in the transcript, or a failed job. Failed chunks now retry up to twice with backoff (skipping permanent errors: auth, word limit, no speech, other 4xx), non-JSON platform error bodies are classified as coded server errors instead of throwing "Invalid JSON response", and a partial-transcription warning surfaces as a dictation toast and an upload completion notice. (#1094)
• Local transcription no longer breaks from binaries stranded inside the app archive. fs.existsSync returns true for paths inside an asar, so a missing unpacked ffmpeg-static binary was reported as spawnable and every local transcription failed with a cryptic spawn ENOENT; resolution now falls through to the system/PATH scan with an actionable error. On Windows the ps-list vendor executable was likewise left inside the asar, breaking meeting-app process detection on every poll. Both binaries are now unpacked, and packaging fails loudly if either is missing rather than shipping broken transcription. (#1124)

Meeting notes

• Diarization no longer collapses every remote speaker into one. A meeting with no stored speaker count and no participants fell back to a default of two, capping "other" speakers at one — so every remote voice merged into a single speaker, and naming it locked that whole side of the call to one person. The live speaker cap is now seeded at meeting start from the note and its calendar participants (held in memory for the session so a later edit or sync can't clobber it mid-call), offline diarization can refine a user-locked cluster that it splits into multiple speakers, and a rolled-back start no longer leaks a stale cap. Local meeting edits (participants, calendar event, transcript) are also re-flagged as sync-pending and preserved on cloud pulls, so a later last-writer-wins pull can't wipe them. (#1126)
• Meeting streams survive the 60-minute session limit. OpenAI Realtime sessions expire after 60 minutes; meeting streams now auto-reconnect across the expiry with a pre-connect buffer so no audio is lost, and a failed reconnect tears down the half-open streams and restores the working ones instead of leaking sockets or stopping transcription. Dictation keeps its existing auto-stop-with-accumulated-transcript behavior at the limit. The ONNX worker is now also verified into production builds (packaging fails if it's missing), fixing a utility-process crash-loop. (#830)
• Local speech is no longer silently dropped from meeting transcripts. Held-back meeting mic segments were discarded on audio-only echo evidence even when no system transcript matched them — field logs showed genuine local speech during double-talk being deleted. A transcript text match is now the only condition that drops a held-back segment; audio evidence can delay a segment but never discard it, and late-arriving system transcripts can still retract released echo within the full duplicate window. (#1093)

Notes & history

• Note enhancement stops borrowing another note's transcript and renaming titled notes. The meeting-recording transcript is global and persists after a recording stops, and the enhance action used it unguarded — enhancing any note after a meeting fed it the last recording's transcript, producing identical enhanced content across notes and leaking one meeting into another. Enhancement is now gated on the transcript belonging to the active note, auto-titling is opt-in per run and only renames empty- or default-titled notes, and the transcript is folded into the staleness hash so edits to it re-trigger enhancement. (#1119)
• Line breaks show in the history view. History entries now render with whitespace-pre-wrap, matching how transcripts display everywhere else. (#1015)

Hotkeys

• Multiple hotkeys per action. Each hotkey slot — dictation, agent, voice agent, and meeting — can now be bound to several hotkeys, so the same action fires from different keyboards. A new "Add another hotkey" row stacks bindings with per-slot duplicate guarding and atomic updates (a partial registration failure rolls the slot back rather than leaving a dead entry). Full support across globalShortcut and the native listeners (macOS Globe/mouse/right-modifier, Windows/Linux low-level hook, push-to-talk); GNOME, KDE, and Hyprland apply the primary hotkey. Comma-key hotkeys (e.g. Control+,) are preserved instead of being split on the list separator, and bindings persist as a backward-compatible comma-separated list. (#1017)

Clipboard & paste

• Dictating into Claude Desktop and claude.ai no longer breaks after the first paste. The auto-learn correction monitor used to force the AXEnhancedUserInterface accessibility flag onto the app it pasted into. On some Chromium apps (Claude Desktop, claude.ai in any browser) that flag permanently blurs the message composer, so every dictation after the first pasted into a field that no longer had keyboard focus. Monitoring is now strictly read-only: apps that expose their accessibility tree keep auto-learn working as before, and apps that don't simply skip correction learning for that paste. Restart the affected app once after updating to clear the stuck flag. (#1116)

Audio & media

• No more false "No audio detected" on Windows. The speech gate's AudioContext can stay suspended or stall on Windows (Bluetooth headsets, wedged output devices), so its analyser read flat silence and rejected recordings that actually captured speech. A suspended context is now resumed, gate windows are skipped while the context isn't running (the recording proceeds straight to transcription and the gate reports "unavailable"), and cancelling a recording tears the gate down instead of leaking its interval and AudioContext. (#1125)

Windows

• No more Windows Firewall prompt for local Parakeet transcription. The bundled sherpa-onnx server only serves OpenWhispr itself over 127.0.0.1, but the upstream binary has no loopback-only bind option, so Windows raised an "allow public and private networks" prompt when it started. All-users installs now register a scoped inbound block rule for the server binary: the prompt is gone, the port is closed to the network, and transcription is unaffected because Windows never filters loopback traffic. The rule is removed on uninstall. (#1090)

Linux

• Wayland global hotkeys no longer crash on Node 24. The D-Bus integration for GNOME, KDE, and Hyprland global shortcuts moved off the unmaintained dbus-next (last updated in 2021, built on Node APIs removed in Node 24) to @homebridge/dbus-native. The switch also fixes the KDE pre-registration conflict check (the invoke callback delivers the unwrapped return value, not a message object, so the check never fired) and attaches an error listener to each session bus, so a stale DBUS_SESSION_BUS_ADDRESS can no longer take down the process with an unhandled error. (#1101)

[1.7.4] - 2026-07-07

A feature-and-hardening release on top of 1.7.3: Tinfoil confidential inference for both transcription and AI reasoning, Azure AI Foundry / Azure OpenAI speech-to-text, native Windows system-audio capture for meeting notes, enterprise SSO in onboarding, cross-device Snippets sync, ambient sync that now runs even in tray-only sessions, and a broad stack of fixes across cleanup routing, clipboard, media playback, hotkeys, local GPU transcription, macOS paste, updates, and Linux packaging — plus SOC 2 dependency remediation.

Transcription

• Tinfoil — confidential cloud transcription (BYOK). New bring-your-own-key provider built on Tinfoil's attested secure enclaves for private cloud speech-to-text in both dictation and uploaded audio. The client verifies enclave attestation before every request and connects over an enclave host assigned dynamically at runtime. (#944)
• Azure AI Foundry / Azure OpenAI speech-to-text. The custom transcription provider now recognizes Azure endpoints (*.cognitiveservices.azure.com, *.openai.azure.com, *.services.ai.azure.com) and builds the deployment-style URL Azure requires (/openai/deployments/{model}/audio/transcriptions?api-version=...) instead of the plain OpenAI {base}/audio/transcriptions shape, which returned 404 DeploymentNotFound. Auth now uses Azure's api-key header on Azure hosts. Enter your resource endpoint as the URL and your exact deployment name in the Model field; api-version defaults to a transcribe-capable preview and can be overridden by appending ?api-version=... to the endpoint. (#997)
• Configurable self-hosted model. Self-hosted / local transcription endpoints can now specify which model to request instead of being pinned to a fixed default. (#1043)
• Hardened realtime dictation streaming. The realtime streaming connection is more resilient, and 16 kHz capture is now linearly upsampled to the 24 kHz OpenAI's realtime API requires — fixing outright BYOK connection rejections (invalid_request_error: integer_below_min_value) and a 1.5× speed mismatch on cloud sessions pinned to 24 kHz. (#1044)
• Local GPU model reloads after sleep. Waking the machine no longer leaves the on-device GPU transcription model in a broken state; it's reloaded automatically. (#1032)
• Multi-GPU device selected by UUID. The local transcription server now pins its GPU by stable UUID instead of a numeric index, so it keeps using the intended device across reboots and driver reordering on multi-GPU machines. (#1018)
• Empty recordings no longer crash transcription. A zero-length recording is handled gracefully instead of taking down the transcription pipeline. (#891)
• Bundled VAD model in packaged builds. The ggml-silero voice-activity-detection model is now copied into the packaged app, so production builds run local transcription with VAD instead of logging "model not found" and running without it. (#1000)

Reasoning & models

• Tinfoil — confidential AI inference (BYOK). Tinfoil is also available as a private reasoning/agent provider, with six chat models (reasoning-capable ones expose a disable-thinking toggle, matching Groq) verified against enclave attestation before each request. (#875)
• Hardened dictation cleanup and wake-word routing. The cleanup prompt was rewritten to reliably transform — not reply to — transcripts: transcripts are framed in <transcript> tags with a trailing output anchor, cleanup runs deterministically at temperature 0, and cloud cleanup requests now pin promptMode: "cleanup" so the server can't flip them to the action prompt when the dictation agent is off. Agent routing now requires the agent to be genuinely addressed (at the start of a dictation, after a greeting cue, or opening a new sentence) instead of firing on any mention of the agent's name anywhere in the transcript. (#1073)
• Bundled llama.cpp updated to b9763 for local LLM inference. (#995)

Meeting notes

• Native Windows system-audio capture. Windows previously captured meeting audio only through Chromium's display-media loopback, which hears just the default output device — so a meeting playing to a non-default device produced silent notes with no error. A new native WASAPI process-loopback helper hears every application on every output device while excluding OpenWhispr's own process tree, and transparently falls back to the Chromium loopback path on Windows versions without process-loopback support (< 10 2004). (#960)

Sync

• Ambient sync now runs in tray-only sessions. Auto-sync scheduling (initial pass, window focus/visibility, network reconnect, and a 5-minute interval) moved out of the Control Panel and into SyncService, so a start-minimized session that never opens the panel still pushes dictations up and pulls changes from other devices down. Passes are serialized across windows with a Web Lock and share a single throttle window; manual syncs wait for the lock instead of being dropped. (#1070, #1072)
• Snippets sync across devices. Your spoken-trigger Snippets now sync across signed-in devices, matching the cross-device custom-dictionary sync added in 1.7.3. (#1037)
• Default folders link to existing cloud folders instead of registering duplicates, so signing in on a new device no longer creates a second copy of your built-in folders. (#1086)

Dictation & snippets

• Snippets match triggers containing Turkish İ and ı. Trigger matching used toLowerCase(), whose lowercase form of İ (U+0130) is "i" plus a combining dot, so a snippet like "İmza" never matched the transcript — and the regex /i flag case-folds neither İ nor dotless ı (U+0131). Triggers and matches are now folded to a canonical key, the pattern matches İ/ı explicitly, and the transcript is NFC-normalized, so "imza", "İmza", "İMZA" — and "Işık"/"IŞIK" for an "ışık" trigger — all expand the same snippet. (#1050)

Onboarding & sign-in

• Enterprise SSO sign-in. Onboarding now offers "Sign in with SSO" alongside social and email/password: it reuses your typed work email, opens the external browser to the SSO flow, and returns via the openwhispr:// callback. (#1034)
• API-key drafts are saved on click-outside instead of discarded. Typing a key and clicking the next field used to silently revert it to empty — worst for Corti BYOK users pasting a client ID and then clicking the client-secret field during onboarding. Click-outside now commits the draft; Escape and the ✕ button still cancel. (#1039)

Clipboard & paste

• Rich clipboard formats preserved on restore. After auto-paste restores your previous clipboard, RTF and other rich formats survive instead of being flattened to plain text. (#1020)
• Faster macOS clipboard restore. Removed an unnecessary delay when restoring the clipboard after pasting on macOS. (#1038)
• Reliable target-app focus before pasting on macOS. The paste keystroke is delivered session-wide, so it only lands in the right field when the captured target app is frontmost. The target is now located by scanning running applications (the previous lookup returned nil under JXA, so activation silently no-op'd and #668's fix never ran), an already-frontmost Chromium/Electron app is left alone to avoid dropping its text field's focus, and the voice-agent hotkeys now capture the target PID so their paste can refocus too — all with no added latency. (#1000)

Audio & media

• Mic retries on the default device when the pinned one is stale. If a previously selected deviceId no longer resolves, recording falls back to the system default microphone instead of failing. (#978)
• Media playback resumes when recording stops, not after transcription — so paused music or video comes back the moment you finish speaking rather than after the transcript is processed. (#1030)
• No media pause for a recording that already ended. A quick tap during streaming-recording startup could pause media with nothing left to resume it (and play cues out of order); post-start side effects are now gated on the recording still being active. (#1061)

Hotkeys

• Modifier-only hotkeys work simultaneously on Windows/Linux. The native low-level key listener was a singleton hardwired to the dictation slot, so only one modifier-only hotkey (dictation, voice agent, agent, or meeting) worked per session. It's now a multiplexer that watches one hook process per key and routes key-tagged events to the right slot, mirroring macOS — and it skips the listener entirely on GNOME/KDE/Hyprland, where shortcuts arrive via D-Bus. (#1001)

Linux

• Fall back to --no-sandbox when user namespaces are restricted, so the app still launches on hardened kernels that disable unprivileged user namespaces. (#1042)
• RPM installable on openSUSE. Fixed packaging so the .rpm installs there. (#1014)

Updates

• No crash on the first "Install & Restart" click after an update. The manual app.emit("before-quit") passed no event object, so event.preventDefault() threw and the first click did nothing (the second only "worked" because the crash had left the app in a shutting-down state). Pre-quit cleanup now hooks the correct before-quit-for-update event on Electron's native autoUpdater — fixing stalled macOS installs — and shuts sidecars down on the update path. (#1012)

Security & dependencies

• Cleared all critical/high dependency alerts (form-data, tar, undici, ws) as part of SOC 2 Type 2 secure-code remediation — 1 critical + 10 high Dependabot alerts resolved, no source changes. (#1051)
• Restored lockfile integrity hashes and added a CI guard. Regenerated package-lock.json on Node 24 so all 967 package entries carry resolved + integrity fields (up from 53), restoring tamper verification for npm ci, and added a lockfile-lint workflow to keep it that way. (#1069)

[1.7.3] - 2026-06-23

A big release: two new transcription providers (Corti for clinical-grade medical dictation and xAI), a reworked onboarding flow built around what you'll use OpenWhispr for, spoken Snippets, a dedicated Voice Agent hotkey, a redesigned dictionary with cross-device sync, dedicated Audio Upload transcription settings, discarded-dictation history, OS-level notification controls, Linux PipeWire system-audio capture, new AI models, and a stack of fixes across paste, audio, settings, and Linux window managers.

Transcription

• Corti — clinical-grade medical transcription (BYOK). New bring-your-own-key cloud provider built on Corti (corti.ai) for HIPAA-compliant, clinical-grade speech-to-text in dictation and uploaded-audio notes. The main-process client mints OAuth2 client-credentials tokens and stores credentials in the encrypted keychain like every other key. (#929)
• xAI speech-to-text. Added xAI as a cloud transcription provider. (#942)
• Corti onboarding polish. Added the Corti provider icon, and the "Get a key" link plus the onboarding Corti links now point to the corti.ai homepage with referral UTM tracking instead of the bare console.
• Self-hosted servers skip the API-key check so local / self-hosted transcription endpoints work without a key. (#835)
• Dedicated Audio Upload transcription settings. Uploaded audio files now have their own Speech-to-Text context (Settings → Speech-to-Text → Audio Upload) with an independent provider and model, split out from dictation the same way Note Recording was. Existing users' dictation preference is migrated over; new users default to OpenWhispr Cloud.
• Cancel an in-progress audio-file transcription from the upload screen — cancelling returns to the upload view and discards the result so nothing is saved.

Reasoning & models

• New models: Claude Opus 4.8 (#884), Gemini 3.5 Flash (#837), and Gemma 4 (#892) are now selectable in their respective model pickers.
• Correct limit-error handling for BYOK and cloud reasoning. (#941)

Dictation & notes

• Snippets — spoken trigger-word expansion. Save trigger → replacement pairs (e.g. "cal link" → "cal.com/anna/30min"); when a trigger is spoken during dictation it's replaced before pasting. (#934)
• Voice Agent hotkey. A dedicated global hotkey that sends a dictation straight to the dictation agent as a command — no wake word — and always bypasses the cleanup model, separate from the chat-agent overlay hotkey. (#932)
• Smart spacing around dictated text so inserted text spaces correctly against surrounding content. (#856, #868)
• Redesigned dictionary page — list view with hover-revealed inline edit/remove, an agent header card, and bulk import/export. (#933)
• Dictionary prompt-echo fix so dictionary terms no longer leak into transcripts. (#852)
• Cross-device custom dictionary sync. Your custom dictionary now syncs across signed-in devices, with last-writer-wins conflict resolution so edits and deletions converge cleanly. (#966)
• Discarded dictations are preserved in history. Cancelled, too-short, or failed dictations are now kept and surfaced behind a "Show Discarded" toggle in History instead of vanishing. (#964)

Onboarding

• Intent capture up front. A new "About you" step lets you multi-select what you'll use OpenWhispr for — dictation, meetings, healthcare, translation, AI commands, or uploading audio — and the rest of onboarding adapts to your choices.
• Inline Corti setup for healthcare. Picking healthcare surfaces Corti on the finish step — enter Corti credentials right there or open Settings with the Corti provider preselected, with a "Skip for now" escape.
• Skippable optional steps, onboarding progress moved into the macOS title bar, the quit button removed from the title bar, a back-to-sign-in escape on email verification, and ghost-variant styling for subtle auth actions.
• Provider chips styling fix on the notes onboarding screen. (#916)

Settings & notifications

• OS-level notification controls. New settings to scope which OS-level interruptions OpenWhispr raises. (#781)
• Remove button for the agent hotkey. (#824)
• Simpler meeting detection. Audio-based meeting detection is now driven by the notification toggle (notificationsEnabled && notifyMeetingDetection) instead of a separate Audio Detection setting; the standalone setting, its UI section, and the now-dead detection translations were removed so a detector can't burn CPU while notifications are off.
• Settings no longer grabs the microphone. Opening Settings used to call getUserMedia to read device labels, which started a mic session and interrupted other audio (e.g. paused music on macOS). It now enumerates devices first and only falls back to getUserMedia when labels are missing because permission hasn't been granted.
• Custom cleanup API key persists across restarts. (#893)
• Download progress is preserved across tab switches. (#735)
• Select trigger background aligned with surrounding controls. (#825)

Linux

• PipeWire system-audio capture for transcriptions via a direct PipeWire loopback (replacing the ScreenCast portal path). (#904)
• Nix flake for one-command install, plus a GitHub Action. (#886)
• Paste reliability: Shift+Insert paste for Electron app windows (#873) and when the window context is unknown (#827); the ydotool socket is now propagated to spawned clients (#962).
• Dictation hotkey no longer dies in Hyprland on config reload. (#919)
• Fall back to the system journal when the user journal has no KWin entries. (#776)

Fixes

• Auto-paste in Chromium-based apps on macOS. (#668, #823)
• Mic re-acquired when it goes silent after idle. (#922)
• Serialize .env writes to prevent an ENOENT rename race. (#940)
• Fall back to JS extraction when system unzip is unavailable. (#775)
• Localized Language Models settings across all locales (#887); fixed zh-CN note-files description mojibake (#848).
• Local semantic search restored in packaged builds. A regression that broke on-device semantic note search in packaged (production) builds is fixed. (#981)
• Faster local transcription: the bundled Whisper server now auto-tunes its thread count to the machine. (#994)
• Accurate live speaker counts in note recordings — per-segment "Speaker N" labels no longer climb past the expected speaker count, and the recorder panel now follows the cursor to the active monitor. (#967)
• Cloud users no longer need to manually pick a model for the Voice Agent hotkey or note formatting — both now reach the OpenWhispr cloud agent without an explicitly selected model.
• No stale clipboard restores during paste, and cloud requests that hit a stale auth token now recover via the session cookie (fixes onboarding intent silently failing to save after email/password sign-in).

[1.7.2] - 2026-05-20

A small patch on top of 1.7.1: zero unnecessary macOS Keychain prompts on first launch, working cloud transcription on Electron's net.fetch, the Note Formatting selector now actually controls model routing, Wave Terminal pastes via the terminal path, and a notes view stability fix.

Desktop & permissions

• No more spurious macOS Keychain prompts on first launch. Two changes drop first-launch prompts from ~3 to 0: the environment manager no longer eagerly probes the secret-crypto backend before any window appears (it now defers Keychain access until the user actually saves their first secret), and the safeStorage key-loss backup is written only when a new master key is generated, not on every launch.

Transcription

• Cloud transcription works again on Electron's net.fetch. The 1.7.0 migration from https.request to net.fetch carried over a manual Content-Length header, which Electron rejects as a forbidden Fetch header — net::ERR_INVALID_ARGUMENT before any bytes hit the wire. All five upload callers (cloud-transcribe, chunked cloud transcribe, retry, file upload, BYOK whisper-compatible) now let net.fetch set Content-Length itself.

Notes

• Note Formatting selector now actually controls model routing. The Note Formatting tab exposed model / mode / provider keys that no runtime path read — actionProcessingStore was overriding any caller-supplied model with getEffectiveCleanupModel(), so switching the selector was a no-op end-to-end. Generate Notes now resolves through the same per-scope plumbing as Cleanup, with a dictationCleanup fallback so users who never touched Note Formatting keep their existing behavior. Closes #784.
• Notes view state management. Fixed a stale-ref issue where switching between notes could lose unsaved enhanced-content edits.

Linux

• Wave Terminal paste. Added Wave Terminal to the terminal allowlist so auto-paste uses Ctrl+Shift+V instead of the default Ctrl+V. Closes #814.

[1.7.1] - 2026-05-20

A follow-up to 1.7.0 with a stronger encryption key for stored secrets, end-to-end voice activity detection on local Whisper, macOS mouse-button hotkeys, the full OpenAI Realtime GA migration, proxy-aware network paths, and a stack of platform fixes across Linux WMs, Intel Macs, and Windows.

Security

• Stronger encryption for stored secrets. Secrets are now encrypted with AES-256-GCM using a 32-byte master key stored in the OS keychain via @napi-rs/keyring (Keychain on macOS, Credential Manager on Windows, libsecret on Linux). This replaces Chromium's safeStorage saltysalt/peanuts fallback as the trust anchor. Existing safeStorage-encrypted blobs are migrated transparently on first read — no prompts, no re-entry.
• Linux without a keyring daemon continues to fall back to safeStorage, matching today's behavior.
• Key-loss protection. When the keyring backend loads successfully, the master key is now backed up via safeStorage so a future native-module break can still decrypt existing blobs instead of losing them silently.

Local Whisper: voice activity detection

• Silero VAD is now wired end-to-end. The 1.7.0 release exposed VAD tuning UI, but whisper-server was never actually started with --vad and the Silero model wasn't bundled. This release ships ggml-silero-v5.1.2.bin (~864 KB) via a new download-whisper-vad-model.js build step, resolves it through WhisperManager.getVadModelPath(), and emits the VAD flags only when both vadEnabled and the model path are present. Server restart now keys on the VAD model path so toggling the UI takes effect immediately.

Hotkeys

• macOS mouse-button hotkeys. Bind dictation to mouse buttons 4, 5, etc. directly. Capture, validate, and re-bind are wired through the existing globe-listener with a hardened race fix so re-binding doesn't stomp the newly-spawned child process. Compound mouse hotkeys (e.g. Cmd+MouseButton4) are explicitly rejected — the native tap only handles bare buttons.

Notes & dictation

• Automatic note renaming is now configurable. A new autoGenerateNoteTitle setting in Cleanup turns the auto-title behavior off; the setting is properly registered as a boolean so cross-window storage events don't leave it as the string "false" (which silently defeats the toggle). Action processing is now async so the UI doesn't block while a long action runs.
• Voice Agent rename. The Dictation Agent tab and toggle are now labeled "Voice Agent" in all 10 locales; the internal dictationAgent identifiers are unchanged. The agent system prompt was compressed ~2700 → ~1200 chars now that name detection (detectAgentName) handles routing.
• Custom prompts sync across windows without restart. Storage events for customPrompt.* keys now propagate Control Panel edits to the Main Window immediately.
• Side-panel layout flip is scoped to the notes view with an active note. It no longer leaks onto Home, Chat, or Upload when the window is narrow; the layout now also flips correctly when you drag the window narrow on an open note and reverts on widen.
• Dictation overlay can't steal focus on Linux WMs. The floating icon is now focusable: false, the cross-platform equivalent of the no_focus [instance="open-whispr"] workaround Sway/i3/wlroots users were applying by hand. Mouse clicks on mic and cancel still work; auto-paste no longer breaks because the original text field stays focused.
• Meeting-detected "Start Recording" no longer races on a fresh control panel. The main process used to fire the navigate-to-meeting-note IPC after did-finish-load but before React mounted its listener, so the click was silently lost on a newly-created window. Now uses the same store-and-drain handshake the notification window already follows.

Reasoning & local LLMs

• Voice Agent only triggers when explicitly addressed. With useDictationAgent on and useCleanupModel off, every utterance was previously routed to the agent. Now detectAgentName must match; otherwise the request falls back to cleanup if reachable, or skips reasoning. Closes #768.
• Self-hosted dictation agent uses its own credentials instead of silently falling back to the cleanup model's URL and API key.
• Qwen thinking is suppressed on local llama.cpp. Both the note-formatting and streaming chat-agent paths now send chat_template_kwargs.enable_thinking: false (the llama.cpp-compatible flag) so Qwen3.x doesn't exhaust its token budget inside <think> and return empty content.
• Stop sending Ollama-only think field to Groq. Groq strictly validates request bodies and rejects unknown fields, so the suppression helper now picks the right dialect per provider instead of attaching both.
• disableThinking actually reaches every reasoning route. resolveReasoningRoute previously returned a bare { kind: "cleanup" } and the agent config omitted the flag, so cleanupDisableThinking and dictationAgentDisableThinking were silently dropped before reaching processWithReasoningModel. Now propagated through cleanup, BYOK cleanup, and streaming BYOK cleanup — Qwen finally respects the toggle on every STT path.
• Thinking-model responses parse correctly. When a local llama-server response only includes reasoning_content (no content field — common for Qwen3 / DeepSeek-R1 in single-shot mode), the result is now read from that field instead of returning an empty string.
• Idle local LLM unloads from VRAM after a timeout, freeing the GPU for other workloads.
• LAN / custom-cloud URL handling. Accepts OpenAI-compatible URLs with or without a /v1 suffix, and no longer produces /v1/v1/chat/completions when the stored endpoint already includes the suffix. Both LAN and custom-cloud paths now mirror the SDK's existing v1-suffix fallback.

Transcription

• Self-hosted STT endpoint resolution fixed. getTranscriptionEndpoint was only reading cloudTranscriptionBaseUrl, which defaults to the OpenAI URL. When transcriptionMode is self-hosted, the resolver now reads remoteTranscriptionUrl and stops silently falling back to OpenAI (which produced a 401).
• Language preference is preserved on retry, in meetings, and in the dictation preview. preferredLanguage is now threaded through every transcription path; non-English users were previously dropped to auto-detect on those branches.
• Custom dictionary now reaches the meeting note cleanup prompt. The meeting branch bypassed the dictation cleanup helper and built its own system prompt inline, dropping the user's dictionary. The substitution logic is now in a reusable helper and called from both branches.
• Groq Whisper prompt cap fixed. Custom STT routing to api.groq.com is now detected by endpoint URL (not provider name) and the prompt budget is capped at 890 chars (down from 900) to leave margin on UTF-16 codepoint drift.
• No more boot-time BYOK auto-default override. A one-shot migration was switching cloudTranscriptionMode to byok on every cold boot whenever any BYOK key existed, overriding subscribed users' UI selection. Both the persistence bug and the "any key" signal are gone.
• Parakeet health check no longer stalls. Removed a transcribing flag that gated the watchdog interval but had no other reader — if any error path failed to clear it, the watchdog would skip every tick forever and miss a dead sidecar.
• Transcription sync survives empty-text cloud rows. The local transcriptions table enforces text NOT NULL, but the cloud allowed null/empty rows from earlier failed transcriptions. Pulling one of those would abort the entire sync loop with a SqliteError. Empty rows are now filtered on push, skipped on pull, and a defensive coalesce in upsertTranscriptionFromCloud guards against future bad inputs.

Streaming & cloud

• OpenAI Realtime GA migration complete. OpenAI removed the Realtime Beta API on 2026-05-12. 1.7.0 dropped the OpenAI-Beta header but kept the Beta wire format — transcription still broke because the GA server emits session.created / session.updated (not transcription_session.*) and rejects transcription_session.update in favor of session.update with session.type=transcription and a nested audio.input.* schema. The two server-event handlers and the session configuration payload are now on the GA shape. Closes #805.
• Graceful fallback when the OpenWhispr API URL is unconfigured. postServerToken now attaches a NO_API code so startStreamingRecording can fall back to batch recording instead of surfacing an unhandled error.

Networking & proxies

• All GitHub, cloud, and calendar fetches now go through Electron net.fetch. PR #687 swapped 23 fetch sites onto proxyFetch but missed pre-existing private helpers using raw http/https. Behind corporate proxies these failed with ENOTFOUND / ETIMEDOUT before the proxy-aware downloader was ever reached (e.g. "Enable GPU" hitting connect ETIMEDOUT <IP>:443). Covers downloadUtils (new shared fetchJson), llamaVulkanManager / whisperCudaManager (GitHub release metadata, preserving GITHUB_TOKEN), ipcHandlers.postMultipart (chunked cloud transcribe, BYOK whisper uploads, 5 callers total), Google Calendar OAuth (token exchange/refresh/revoke) and _apiGet with a 10s AbortSignal.timeout, and the legacy signed-token bearer exchange in main.js.

Calendar

• Primary-only sync toggle. A new "Sync primary calendar only" switch on the Google Calendar integration card (defaults on for new connections) ignores events from calendars shared with you end-to-end: fetch captures Google's primary flag, selection is filtered, stale events from deselected calendars are purged, and the notification cache and next-meeting timer are reset. Existing connected users are fixed-forward on next launch.

Platform fixes

• Meeting recording falls back to mic-only on Intel Macs when the native CoreAudio process tap fails (ScreenCaptureKit / audio HAL quirks on AMD GPU configurations). The error no longer aborts the entire session — closes #744.
• macos-media-remote works on macOS 15.4+. Three independent regressions (main-thread deadlock, etc.) caused the binary to always return NOT_PLAYING on macOS 26.4.x. Music pause/resume during dictation now works again on the latest macOS.
• Pop!OS COSMIC is forced to XWayland. COSMIC's XDG_CURRENT_DESKTOP=COSMIC fell outside the relaunch allowlist, so the app ran as a native Wayland client — breaking the orb's initial placement and drag. Now handled like GNOME and KDE.
• Terminal detection on GNOME Wayland now uses AT-SPI2 instead of the X11 selection heuristic, restoring Ctrl+Shift+V auto-paste for native Wayland terminal emulators. Fixes #725.
• Konsole on X11 auto-paste fixed. Konsole intermittently reports no WM_CLASS via xdotool, so terminal detection went blind and fell through to Ctrl+V — which AI terminal agents like Codex and Claude Code interpret as image-paste, producing "No image found in clipboard" instead of pasting text. Detection now reads /proc/<pid>/comm as a complementary signal, and Konsole + X11 specifically routes through xdotool windowactivate --sync key shift+Insert (the XTest Ctrl+Shift+V path was being silently dropped by a long-standing focus/grab quirk). Closes #184; original patch and root-cause by @JGKle.
• Linux launcher symlinks (e.g. /usr/bin/open-whispr → /opt/OpenWhispr/open-whispr from deb/rpm packages) no longer fail with "No such file or directory" — the wrapper now resolves the symlink target before sourcing.
• Windows: llama.cpp pinned to b8857 to keep whisper-server.exe (frozen at OpenWhispr/whisper.cpp 0.0.6) loading correctly. A llama.cpp release between b8861 and b9020 bumped ggml's ABI, leaving local Whisper users on 1.7.0 unable to transcribe; the download script now requests b8857 explicitly.
• Port availability check probes the wildcard address (0.0.0.0 / ::) so sidecars don't false-positive on a free port when something is already bound on all interfaces. Resolved with a consolidated serverUtils.isPortAvailable with IPv6 probe. Closes #748.

Docs & contributor experience

• .github/CONTRIBUTING.md now points to the canonical docs site so GitHub surfaces the link in the PR-creation UI.
• Arch Linux ydotool install guide added to the Linux platform docs.

[1.7.0] - 2026-04-30

A big release: new sign-in options, smoother meeting recording, faster cross-device sync, a more configurable AI setup, and the long-planned move to our new legal entity (Gizmo Labs Inc.) on macOS and Windows.

Sign in your way

• Sign in with Microsoft — new on this release, alongside Google and email/password.
• Sign in with Apple on macOS — native Apple ID flow.
• Self-hosted authentication. Sign-in now runs entirely on OpenWhispr infrastructure (we replaced Neon Auth with Better Auth at auth.openwhispr.com). No third-party vendor lock-in. Self-hosters can point at their own server via VITE_AUTH_URL.
• Bearer-token sessions stored in your OS keychain replace the old browser-style cookie jar, so signed-in state survives renderer crashes and Electron session resets. Existing 1.7.x users transition silently on first launch.
• Forgot password opens a browser tab to openwhispr.com/reset-password instead of an in-app form, matching how every other reset flow works on the web.
• Sign-in buttons disable with a tooltip when the OS hasn't registered the openwhispr:// callback handler, so OAuth never gets stuck without a return path.

Security

• API keys are now encrypted at rest. All 12 secrets — every BYOK API key (OpenAI, Anthropic, Gemini, Groq, AssemblyAI, Deepgram, Mistral) and every enterprise cloud credential (AWS, Azure, Vertex) — moved from plaintext .env and localStorage to per-key files encrypted with the OS keychain via Electron safeStorage (Keychain on macOS, DPAPI on Windows, libsecret on Linux). A one-time silent migration runs on first launch with round-trip verification before any plaintext is removed; a sentinel makes it idempotent and re-tryable on partial failure. Closes #532.
• Non-secret preferences (regions, endpoints, hotkeys, flags) continue to live in .env so power users can keep editing them by hand.
• Linux without a keyring falls back to plaintext rather than locking you out, matching Electron's default behavior.

Meeting recording

• Background recording. Meeting capture now lives in a global store with a side-effect-only mount, so the audio pipeline survives navigating to other notes, opening Settings, or any view unmount that previously killed it.
• Floating recording pill. When you record one note and navigate to another, a pill appears top-center showing live mic-activity bars, the recording note title, click-to-jump-back, and a stop button.
• Side-panel layout is opt-in. A new hotkey-only layout setting (full-width or side-panel) controls whether hotkey-triggered recordings snap the window to a 1/3 panel. Manual record-from-note and calendar joins always open full-width and auto-flip to side-panel only when the window narrows below 1024px.
• Per-note diarization preferences persist. Mid-session toggles for "label speakers" and the "others in call" stepper are saved against the note, so a stop/resume keeps your choices instead of falling back to the global default.
• Meeting metadata syncs across devices. Participants, calendar event ID, diarization toggle, and expected speaker count now travel through cloud sync alongside note content. Older clients that don't send these fields keep working — the columns are nullable and treated as optional server-side.
• Three interchangeable streaming providers: OpenAI Realtime, AssemblyAI Universal-3 Pro, and Deepgram. Which providers are available is set on the OpenWhispr server, so there's no desktop-side toggle to keep in sync.
• Cleaner mic capture. A new acoustic gate prevents system audio from leaking into your mic during meetings. Speech onsets are protected so your voice isn't clipped at the start of a sentence.
• Better echo cancellation with built-in noise suppression in the same pass.
• Speaker labels capped to attendee count — no more phantom "Speaker 3+" labels in 1-on-1s and small groups.
• Live diarization stays scoped to the current note's attendees instead of pulling in profiles from unrelated meetings.
• Stable AssemblyAI / Deepgram turns: fixed a crash on turn-end and a frame-size mismatch with AssemblyAI v3.
• Fewer dropped turns: speech-start timestamps now flow through the echo-leak detector for AssemblyAI and Deepgram (was OpenAI-only).
• Music pause/resume on Windows is reliable again. Switching to a windows-media-control Python sidecar with a hardened WinRT async bridge fixed a class of GSMTC failures that left playback paused after a recording ended; if GSMTC ever fails, the app falls back to a media-key tap.

AI configuration

• Per-scope language model setup: pick different providers and models for dictation cleanup, the agent, note formatting, and chat. An empty agent setting links back to your cleanup model with one click.
• Self-hosted reasoning got upgraded. The Self-Hosted card now exposes URL + API key + model picker (was URL-only), bringing it to parity with Cloud → Custom. Distinct help text on each: Cloud → Custom points at OpenRouter / Together; Self-Hosted explains it's for OpenAI-compatible servers on your local network (Ollama, LM Studio, vLLM, llama-server). Closes #661.
• Per-scope thinking-mode toggle. Models that support visible reasoning (e.g. GPT-5/o-series, DeepSeek-R1, Qwen-think) now expose a "show thinking" switch per scope. Defaults to suppressed so the dictation pipeline stays snappy; turn it on per scope when you want to see the model reason.
• NVIDIA Parakeet parakeet-unified-en-0.6b — a new English-only Parakeet model with state-of-the-art offline accuracy (5.91% avg WER on the HF Open ASR Leaderboard, vs 6.34% for v3) at a slightly smaller ~631MB.
• Switching agent mode between Cloud and Local no longer leaves stale provider state behind.
• More reliable ONNX inference (speaker embeddings, semantic search): long meetings no longer crash the app from a memory allocation failure deep in the speaker model.
• Local helpers shut down cleanly on Quit. Local Whisper, Parakeet, llama-server, Qdrant, and the diarization helper now stop properly when you Quit. If a previous session ever leaves one stuck, the app catches and cleans it up on the next launch — so dictation and transcription work right away without manual cleanup.
• Local LLM startup is more patient. llama-server with Vulkan acceleration now gets a longer startup window before the app gives up; a stuck server is also fully stopped before any re-download attempt, so partial files don't get clobbered mid-write.
• Model downloads work behind redirects. Whisper, Parakeet, Qdrant, MiniLM, and local LLM (GGUF) downloads from Hugging Face and GitHub Releases now follow 3xx redirects correctly — a regression where the manual redirect handler aborted the request before the follow could land has been fixed.
• Local LLM downloads share the same plumbing as Whisper and Parakeet (proxy-aware, resume on stall, retry with backoff), removing ~50 lines of duplicated download code.
• ONNX worker failures are now visible. When the speaker-embedding / semantic-search worker crashes, stderr and onnx-worker.log capture the cause; the parent caps respawn at 5 attempts and degrades to FTS5 keyword search instead of restarting in a tight loop.

Sync

• Cross-device delete propagates to other signed-in devices for all object types.
• Folder cascade delete with a confirmation dialog showing how many notes will be removed.
• Folders sync immediately on rename and create (was previously only on update).
• Transcriptions sync on save instead of waiting for the next app launch.
• No more duplicate transcriptions in the cloud. Each transcription is tagged with a client-generated UUID before upload so the cloud row and local row stay in lockstep — sync upserts the existing row instead of creating a second copy.
• Folder pull no longer overwrites a freshly-renamed folder with a stale cloud copy.
• Server stops overwriting updated_at on every note push, eliminating spurious sync loops.

CLI

• Local HTTP bridge for the openwhispr CLI: when the desktop is running, CLI commands hit it directly for note/folder/transcription operations and only fall back to the cloud API if the desktop is closed. Bearer-token auth, 127.0.0.1-only.

Network

• Proxy-aware fetches: Node-side requests now honor the system proxy.
• Trusts your OS CA store: corporate TLS interception with a trusted root no longer breaks Node-side requests.
• Helpful error messages: connectivity failures explain whether it's an auth-host issue, DNS, port 443, or a TLS certificate, instead of a generic "Network error".

Other notable improvements

• Auto-learn corrections now works for Cyrillic, CJK, Arabic, Devanagari, and other non-Latin scripts.
• Windows text-field detection falls back to UIA TextPattern when ValuePattern isn't available — restoring rich-edit support in apps like RichEdit, Monaco, Qt, and Electron-hosted controls.
• Chat — first message saved the moment the conversation is created, eliminating a race that could orphan it.
• Note move keeps the active note selected and removes the source-folder entry immediately.
• Local semantic search (Qdrant) writes to the user data directory, not the read-only app bundle.
• No more hotkey re-register storm when dismissing Settings with a hotkey field focused — the IPC handler short-circuits duplicate listening-mode changes (one call per slot, not four).
• macOS native helpers bundled correctly. macos-audio-tap, macos-globe-listener, macos-fast-paste, macos-mic-listener, macos-text-monitor, macos-media-remote, and the linux-* helpers now land under Resources/bin/ in the packaged app — matching where the runtime resolvers and CI verify step look for them.

macOS / Windows: bundle identifier change

The app's bundle ID changed from com.herotools.openwispr to com.gizmolabs.openwhispr to match our new legal entity (Gizmo Labs Inc.) and fix the long-standing typo. Your notes, settings, API keys, and downloaded models carry over automatically on first launch.

Auto-update from 1.6.x cannot reach this release. Please download the new build manually from openwhispr.com/download.

A one-time onboarding modal walks you through re-granting Microphone, Accessibility, and System Audio permissions on macOS.

Upgrade notes

• 1.6.x users: download manually from openwhispr.com/download.
• You'll be signed out and need to sign in again.
• macOS: re-grant Microphone in-app and Accessibility + Screen Recording in System Settings.
• Self-hosters: rename VITE_NEON_AUTH_URL → VITE_AUTH_URL; rename the legacy reasoning-model env vars (e.g. REASONING_MODEL → CLEANUP_MODEL); rename AGENT_KEY → CHAT_AGENT_KEY. Both old and new names work for two releases.

[1.6.10] - 2026-04-20

Added

• Speaker Diarization Controls: Global on/off toggle in Settings plus a session-scoped pill in the recording view with its own switch and a "1 other in call / 2 others in call" stepper. Unscoped recordings cap at a sensible default to prevent phantom speakers; calendar attendees or the stepper value override. When labeling is off, transcripts fall back to "You"/"Others" labels derived from audio source
• Auto-Label 1-on-1 Speakers: Automatically label system audio speakers in 1-on-1 meetings when exactly two participants are detected (user + one other), creating voice fingerprint profiles and triggering retroactive mapping across past notes
• Integrations View: New top-level Integrations surface hosting API key management (relocated from Settings) and a new MCP integration card with a copyable server URL chip for paid users
• Dedicated Meetings Settings: Separate Speech-to-Text and Language Model selectors for meeting recording, independent from dictation
• Streaming-Only Engine Filter: Note Recording picker now filters to streaming-capable engines (OpenAI Cloud, gpt-4o-transcribe, on-device, streaming LAN servers); self-hosted stays available with a caption warning

Changed

• Settings Reorganization: "AI Models" collapsed into Speech-to-Text and Language Models; "Speech & AI" sidebar group renamed to "AI Models"; Meetings section added with its own sub-tabs; multiple redundant headers removed (sidebar "Settings", Agent Mode, enterprise provider-tabs wrapper, system-prompt textarea wrapper)
• Agent Hotkey Relocated: Moved into the Hotkeys section where it belongs, no longer orphaned under Agent Mode
• MCP Pro Gating: Free users see an upgrade message on the MCP card instead of operational-looking setup steps; paid users get the full flow
• README Reframed: Positions OpenWhispr as an open-source alternative to WisprFlow (dictation) and Granola (meetings)
• Meeting Sub-tabs Simplified: Engine selectors now shown directly; "follow main settings" toggles dropped. A one-time migration preserves every existing user's behavior with zero breaking changes

Fixed

• Stop Binding Random Notes to 1-on-1 Attendees: Removed over-eager calendar-event adoption that was auto-linking unassigned notes to the first active 1-on-1 calendar event, stamping unrelated recordings with that attendee's speaker profile
• Echo Leak Detector Pre-AEC: Detector was receiving the AEC-cleaned mic buffer so correlation against system reference was always ~0; now runs on the raw mic buffer where the leak actually exists
• Cloud Sync — Transcriptions Reappear on Restart: Clearing history hard-deleted locally while cloud rows stayed intact; now soft-deletes with tombstones pushed to cloud before hard-delete
• Cloud Sync — Folders Reappear on Restart: Same delete-sync bug as transcriptions; mirrored the notes/conversations pattern with deleted_at, pushFolderDeletes, and a pull-side tombstone guard
• Folder Name Collision After Delete: UNIQUE(name) blocked recreating a folder with the same name while its tombstone sat unsynced; tombstoned rows now get a mangled internal name that frees the slot without leaking to cloud
• View Plans Deep-Link: SettingsModal was initializing activeSection to "account" regardless of initialSection because prevOpen equalled open on first mount; moved to lazy state initializers so the resolution branch fires correctly
• MCP i18n: Setup steps corrected from OAuth to API key auth
• Missing Integrations i18n Keys: Added to en, es, fr locales
• Legacy Prompts Deep-Link: Now routes to the Dictation Cleanup sub-tab where PromptStudio lives, instead of falling back to the default tab
• Japanese / Chinese Sidebar Descriptions: Rewritten to use the same vocabulary as the actual sub-tabs
• Spanish Enterprise Strings: Aligned on "en la nube" for consistency
• Parakeet Model Cache: "Open cache folder" now opens at the cache root so downloaded models are actually visible

[1.6.9] - 2026-04-16

Added

• Transcript Export: Export transcripts to disk as TXT, SRT, or JSON files
• Disable Auto-Paste Toggle: New setting to disable automatic pasting after dictation
• Cloud Sync: Bidirectional cloud sync for notes, folders, conversations, and transcriptions
• Linux Push-to-Talk: Native push-to-talk support on Linux via evdev with permission UX and setup guide
• Agent Folder Tools: Agent can list, create, and auto-create folders with semantic folder matching
• API Keys Management UI: Manage API keys from within Settings
• Documentation Link: Added documentation link to the support dropdown menu
• Notification Dismiss Circle: Hover-reveal dismiss circle on notification overlays

Changed

• Electron 41 & Node 24: Upgraded from Electron 39 to 41 and Node 22 to 24
• Dependency Upgrades: Bumped TipTap packages, Tailwind CSS/Vite, and other major dependencies
• Provider Tab Redesign: Redesigned provider tabs as compact pill buttons
• Local Model Spec Links: Replaced local model descriptions with clickable spec links
• llama.cpp Endpoint Detection: Detect llama.cpp via /v1/models and prefer /chat/completions over /v1/responses
• README Overhaul: Simplified README from 907 to ~130 lines with improved keywords, privacy messaging, and feature parity highlights

Fixed

• Windows Hotkey After Lock/Unlock: Prevent false hotkey activation after Win+L lock/unlock on Windows
• Windows Push-to-Talk Stuck Recording: Prevent stuck recording when push-to-talk key-up is missed; remove timeout cleanup
• Speech Gate Thresholds: Relaxed speech gate thresholds and added no-audio toast for failed transcriptions
• Transcription Retry Button: Show retry button on failed transcriptions
• Note Editor Sync: Sync editor content when a note is updated externally
• Speaker Diarization Thresholds: Tuned thresholds to prevent excessive speaker creation
• Live Transcription Preview: Wired up live transcription preview toggle and restored missing IPC handlers
• Speaker Reclustering: Added periodic speaker reclustering, unified recording modes, and autosave transcript
• Floating Icon Position: Persist floating icon position setting across restarts
• TLS/Certificate Errors: Surface TLS/certificate errors in model download UI
• llama.cpp Probe: Made llama.cpp probe one-shot and removed fragile hasMeta heuristic
• Agent Folder Navigation: Refetch notes when opening a note in a different folder
• Notification Window Sizing: Fixed notification window sizing and dismiss circle visibility
• Japanese Translations: Added missing Japanese translations and reordered handler for consistency
• German Locale: Fixed invalid JSON in German locale file
• i18n Spec Links: Fixed i18n spec link text and renamed to "Learn more"

[1.6.8] - 2026-04-14

Added

• Speaker Diarization: Live speaker identification during meeting recording with post-processing refinement when the call ends (auto-downloaded sherpa-onnx pyannote + voxceleb models)
• Speaker Reassignment UI: Click any bubble to assign it to a known speaker, calendar attendee, or contact — with attendee-aware picker and bulk-select reassignment
• Voice Fingerprint Linking: Attach voice profiles to contact emails from the speaker picker
• Meeting AEC Helper: Native WebRTC AEC3 sidecar for mic echo cancellation when system audio is captured, with graceful fallback to the JS echo leak detector
• Transcript-Level Dedupe: Retract events drop mic duplicates once system audio confirms the same speech, cleaning both the live view and the saved transcript
• Live Accuracy Hint: Subtle in-view hint during recording indicating that speaker labels will sharpen once the call ends

Changed

• Meeting AEC Helper is Prebuilt: Binaries are built on CI and downloaded like whisper-cpp / qdrant — contributors no longer need cmake, Python 3, or a C++ toolchain for a normal build

Fixed

• Speaker Reassignment for Own Bubbles: Reassigning a left-side (mic) bubble now correctly flips side, name, and color instead of staying locked as "You"
• Live Speaker Lock Persistence: Live-assigned speaker names survive across the session and through diarization merge
• Meeting System Audio Handling: Restore system audio handling after transcription path refactor
• Local Whisper Speech Gate: Stricter silence gate with peak-amplitude fallback to prevent dropped chunks on quiet but non-silent audio
• Transcript Merge: Preserve prior transcript when diarization merge arrives

Security

• CMake Quoter Escape: Single-pass backslash + quote escape in quoteCmake resolves a CodeQL incomplete-escape warning

[1.6.7] - 2026-04-02

Added

• Calendar Participants on Meeting Notes: Automatically link Google Calendar attendees to meeting notes when recording starts from a calendar event, with domain-grouped display and Gravatar avatars
• Save Notes as Files: Export notes to the local filesystem as Markdown files, mirroring folder hierarchy
• Responsive Settings Dialog: Settings dialog adapts to narrow windows — sidebar collapses to icon rail, rows stack vertically, plan grid reflows
• Chat Sidebar: Full sidebar chat tab with conversation history, cloud sync, and semantic search
• Chat UX Polish: Empty state with illustration, shimmer thinking/streaming indicator, stop button, action buttons and search dialog
• Local Semantic Search: Always-on Qdrant vector DB sidecar for offline semantic search across notes — hybrid FTS5 + vector with Reciprocal Rank Fusion
• Agent Tool Calling: Agentic tool-calling system with note management tools (get, create, update, search), cloud agent support with NDJSON streaming, and local model tool calling with RAG context injection
• Embedded Chat in Notes: Embedded chat panel in the note editor with floating and sidebar modes
• Per-GPU Device Selector: Choose a specific GPU for transcription and intelligence processing (#539)
• Settings Keyboard Shortcut: Cmd+, / Ctrl+, keyboard shortcut to open Settings
• Notes Actions Button: Actions sidebar button with redesigned action editor dialog
• Notes Folder Picker: Folder picker in the note metadata row with cleaned-up input styles
• Notes Sidebar Buttons: New note and search notes buttons in the sidebar
• Meeting Echo Cancellation: Echo cancellation on mic input and note metadata chips in meeting view
• Linux Wrapper Script: Wrapper script to force XWayland and support user flags (#507)

Changed

• Vercel AI SDK Migration: Agent mode migrated from raw API calls to Vercel AI SDK
• Notes Bottom Bar Redesign: Redesigned bottom bar with compact action picker
• Dialog Design System Alignment: All dialogs aligned with design system guidelines
• Removed Note Word Count: Removed word count from note editor
• Cloud Agent Streaming: Stream cloud agent responses directly from the renderer via IPC

Fixed

• Meeting Auto-Detection: Fix auto-detection not firing for browser meetings
• Meeting Transcription Provider: Use local transcription provider for notes/meeting recording (#530)
• Meeting Partial Transcript Spam: Prevent partial transcript spam and duplicate final segments
• Meeting Notification Timing: Resolve notification popup timing and detection lifecycle bugs
• Folder/Note Race Conditions: Resolve race conditions when switching folders quickly, prevent meeting view from exiting when changing folder, fix rapid delete/switch state management
• Clipboard Preservation: Preserve images and HTML in clipboard during paste-and-restore (#381)
• Transcription Retry Provider: Retry transcription uses configured provider instead of forcing Parakeet
• JSON Parse Validation: Validate JSON.parse result type before calling .replace() in prompts (#541)
• GPU Selector Polish: Address code review feedback, rename Intelligence GPU label, fix dropdown chevron padding (#539)
• Meeting Participant Saves: Fix calendar attendees not syncing to store and manual participant adds overwriting calendar data
• Chat Duplicate Conversations: Fix duplicate conversations — includeArchived filter returned all instead of only archived
• Linux Wayland Fixes: Force XWayland on KDE/GNOME Wayland, fix hotkey startup race; use uinput before portal on GNOME Wayland (#468, #494)
• Mic Permission Gate: Remove mic permission gate, fix system audio detection
• Windows Build Signing: Fix Windows build signing on PRs, add missing mic-listener download, add missing publisherName to Azure signing config
• Dead optimizeAudio Crash: Remove dead optimizeAudio call that crashes on recordings over 90 seconds (#524)
• Download URL Logging: Remove URL truncation from download log and add failure logging (#540)

Security

• Google Calendar Scopes: Narrow OAuth scope from calendar.readonly to calendar.events.readonly + calendar.calendarlist.readonly for minimal privilege
• picomatch: Bump to 4.0.4
• brace-expansion: Bump to 1.1.13 (security backport)
• yaml: Bump to 2.8.3
• tar: Bump to 7.5.13

[1.6.6] - 2026-03-19

Added

• Native macOS System Audio Tap: CoreAudio Tap API for direct system audio capture — eliminates the need for screen recording permission on macOS 14.2+
• TipTap Rich Text Editor: Migrated notes editor from plain Markdown to TipTap with Obsidian-style live preview — hides Markdown syntax except on the cursor line, with rich text rendering for enhanced and transcript views
• Dual-Channel Meeting Transcription: Separate mic and system audio channels with chat bubble UI for speaker-differentiated meeting transcripts
• Meeting Segment Timestamps: Persist segment timestamps in saved meeting transcripts with chronological ordering
• Meeting-Specific AI Prompts: Meeting notes generation now uses speaker-aware prompts for better context in generated summaries
• KDE Wayland Native Shortcuts: Native global shortcut support for KDE Plasma on Wayland using D-Bus, matching the existing GNOME and Hyprland approach (#486)
• Mistral Nemo 12B and Gemma 3 12B: Added to local model registry for on-device inference (#483)
• Post-Login Permissions Gate: Returning users now see a permissions check after login to ensure mic and system audio access

Changed

• Unified Notes Recording: All notes now use dual-stream transcription with simplified recording UX — always saves to transcript
• Notes Tab Rename: Renamed "Raw" tab to "Notes" and default to it during meetings
• Shared Note Title Generation: Extracted generateNoteTitle utility for consistent auto-titling across meeting and regular notes
• Simplified Permission Buttons: Consolidated permission prompts to a single "Grant Access" action (#490)
• screenRecording → systemAudio Rename: Renamed screenRecording references to systemAudio across the codebase for clarity
• macOS 15+ System Audio Consent: Trigger the native system audio consent dialog on macOS 15+ instead of the legacy screen recording prompt
• Improved Notes Output: Better generate notes output format and auto-title generation
• Update Notification Polish: Improved update notification transparency, icon, and copy
• Permission Re-validation: Re-validate mic and system audio permissions against the OS on component mount

Fixed

• Gemini Agent Streaming: Route Gemini agent streaming to the correct API endpoint
• Windows Mic Volume Mutation: Disable browser AGC to prevent Windows mic volume being permanently altered (#476)
• Linux Mono Transcription: Request stereo recording to prevent mono transcription failure on Linux
• Meeting Bluetooth Audio: Detach meeting AudioContexts from output device for Bluetooth compatibility; fix system audio loopback silence
• Meeting Detection Suppression: Suppress meeting detection notifications when meeting mode is already active
• Windows Paste Modifier Keys: Release held modifier keys before SendInput paste on Windows
• Meeting Session Reset: Reset meeting audio send counts between sessions
• Meeting Hotkey Behavior: Meeting hotkey always opens a new meeting regardless of current view
• STT Config Auth Timing: Retry STT config fetch before recording when auth isn't ready on mount
• Hotkey Restore on Failure: Restore previous hotkey on registration failure
• KDE Wayland Hotkeys: Force XWayland on KDE Wayland to fix hotkey registration
• Streaming Dictation Commands: Use TipTap editor commands for streaming dictation input
• Google OAuth Onboarding: Fix Google OAuth users skipping onboarding flow
• Realtime Dictation Default: Default streaming provider to openai-realtime for dictation; respect sttConfig dictation mode for realtime models
• KDE Plasma Overlay: Fix KDE Plasma hotkey and overlay window behavior — scoped window type changes to KDE only, preserving GNOME behavior (#491)
• Cleanup Prompt Refusal: Fix cleanup prompt refusing to output command-like transcriptions (#478)
• KDE Wayland Clipboard Paste: Replaced busy-wait with sleep and clean up temp file for KDE Wayland paste (#455)
• GNOME Agent Hotkey: Register agent hotkey as independent GNOME Wayland keybinding slot (#436)
• Agent Hotkey Conflict Warning: Show conflict warning when agent hotkey duplicates another mode
• Meeting Hotkey Registration: Await async registerSlot for meeting hotkey registration
• Media Pause During Dictation: Prevent paused media from being unpaused during dictation (#419)
• Meeting Chat Scroll Overlap: Fix meeting system audio transcription and chat scroll overlap
• macOS Media Remote Bundle: Include macos-media-remote in extraResources (#487)
• NSAudioCaptureUsageDescription: Restore plist entry and increase audio probe timeout

Security

• undici CVE-2026-1526: Bump undici to 6.24.1 to fix request smuggling vulnerability

[1.6.5] - 2026-03-17

Added

• Data Retention Toggle: New privacy setting to control whether transcription text is retained in history (Privacy & Data settings)

Fixed

• Meeting Detection Reset: Fix meeting detection not properly resetting after a meeting ends

[1.6.4] - 2026-03-15

Added

• Meeting Mode Hotkey: Dedicated hotkey to start/stop meeting transcription directly from the keyboard, independent of the dictation hotkey
• Account Deletion: Users can now delete their account from within the app
• Qwen3.5 Local Models: Added Qwen3.5 local models to the model registry; removed sub-1B models that were too small for practical use
• Model Descriptions in Picker: Local model picker now shows model descriptions to help users choose the right model
• Meeting Detection Toggle: New setting to enable/disable automatic meeting detection
• Dependabot: Automated weekly npm dependency updates via Dependabot
• CodeQL Static Analysis: GitHub Actions workflow for automated security scanning
• Zod Dependency: Added Zod for input validation and sanitization

Changed

• Multi-Monitor Floating Icon: The dictation floating icon now appears on the monitor where the cursor is, instead of always on the primary display
• Persistent Panel Position: Panel start position now persists across app restarts
• Compact Hotkey Tooltip: Overlay tooltip uses compact modifier symbols (e.g., ⌘⇧K instead of Cmd+Shift+K), wraps for long combos, and aligns to window edge based on panel position
• Cross-Window Settings Sync: Settings changes now sync across all open windows in real time
• Agent Chat Title: Renamed agent mode window title from "Agent Mode" to "Agent Chat"
• Windows Model Preservation: Local LLM models are now preserved during Windows app updates instead of being deleted

Fixed

• Meeting Hotkey Overwrite: Fixed meeting hotkey accidentally overwriting the dictation hotkey on save
• Meeting Snap Timing (macOS): Fixed meeting mode snap timing on macOS causing incorrect window positioning
• Meeting Detection False Positives: Reduced false-positive meeting detection notifications
• Hotkey Tooltip Display: Fixed hotkey tooltip not updating after changing the hotkey in settings
• Silence Detection Threshold: Lowered silence detection threshold to avoid rejecting valid speech that was previously considered too quiet (#411)

[1.6.3] - 2026-03-12

Changed

• System Audio Permission Clarity: Renamed "Screen Recording" to "System Audio" across all permission prompts, onboarding, and settings — makes it clear that OpenWhispr captures other participants' audio, not your screen
• Improved Permission Copy: Microphone permission now reads "Captures your voice for transcription"; System Audio reads "Captures other participants' audio from calls and meetings. We never record your screen."
• Electron 39: Upgraded from Electron 36 to 39, which uses the CoreAudio Tap API by default on macOS 14.2+ — eliminates the purple "screen recording" indicator, the "Your screen is being observed" lock screen message, and the misleading "Screen & System Audio Recording" permission prompt. Users now see "System Audio Recording Only" instead
• NSAudioCaptureUsageDescription: Added the new macOS 14.2+ audio capture usage description to Info.plist, enabling the separate system audio permission dialog
• better-sqlite3 12: Upgraded from v11 to v12 for Electron 39 V8 compatibility
• Localized in all 10 languages: All permission copy changes translated across en, pt, de, es, fr, it, ru, ja, zh-CN, zh-TW

Added

• Hyprland Wayland Support: Native global shortcut support for Hyprland using hyprctl keybindings + D-Bus, matching the existing GNOME Wayland approach (#416)

Fixed

• Soft Voice Recognition: Enabled Auto Gain Control (AGC) for dictation microphone input to automatically boost quiet speech — previously disabled, now matches meeting mode behavior
• OpenAI Realtime VAD Sensitivity: Lowered voice activity detection threshold from 0.5 to 0.3 (both client and API) so soft-spoken audio is no longer missed
• Speech Onset Clipping: Increased VAD prefix padding from 300ms to 500ms to capture the quiet beginning of soft speech that was previously cut off
• Wayland Clipboard Paste: Fixed wl-copy failing silently due to 1ms spawnSync timeout killing the fork before it completed — increased to 50ms (#416)
• Streaming Media Resume: Fixed media staying paused after recording silence with "Pause media on dictation" enabled — streaming path now fires the completion callback even when no speech is detected (#429)

[1.6.2] - 2026-03-11

Added

• System Audio for Notes: Mix system audio (via getDisplayMedia loopback) with microphone input for note recordings, enabling capture of meeting audio, YouTube lectures, and other system sounds
• Event-Driven Meeting Detection: Replaced polling-based meeting detection with native OS event APIs (CoreAudio on macOS, WASAPI on Windows, pactl on Linux) — reduces background CPU from 5–9% to near-zero (#404)
• Notes Onboarding: Added screen recording permission step to the notes onboarding wizard (macOS) so users can grant permission before their first recording

Changed

• Auto-Enable System Audio: System audio is now automatically enabled when screen recording permission is granted — removed the separate toggle button for a simpler recording experience
• Deferred Transcript Display: Recording transcript is no longer shown live during notes dictation; it appears after recording stops, matching the meeting notes flow for a cleaner experience

Fixed

• Windows Hotkey Stability: Track modifier state in native keyboard hook so modifier-only shortcuts (e.g. Control+Super) are detected reliably on Windows 11; keep floating recorder interactive; prefer compiling current key-listener source over downloaded binaries
• macOS Accessibility Permission Prompt: Detect missing accessibility trust after startup and notify users with auto-opened Privacy settings and toast guidance — fixes silent Globe key failures on fresh installs
• Realtime Streaming Warmup: Fix warmup gating so initial audio is no longer silently dropped; skip redundant session config in cloud mode; handle empty-buffer commit on disconnect gracefully
• Custom Dictionary Prompt Truncation: Truncate custom dictionary to respect Groq's 896-char limit, preventing 400 errors on large dictionaries (#405)
• Parakeet bzip2 on Windows 10: Add JS fallback for bzip2 extraction when native tar fails (#406)
• Business Plan Past-Due Check: Include business plan in past-due subscription check

Removed

• Removed the Monitor toggle button from the dictation widget (system audio mode is now automatic)

[1.6.1] - 2026-03-08

Added

• WebSocket Streaming for BYOK Dictation: OpenAI Realtime API streaming now works for standard dictation mode (not just meetings), enabling real-time transcription for Bring Your Own Key users
• Unified Streaming Path: Extended OpenAI Realtime WebSocket streaming to normal dictation, sharing the same streaming infrastructure as meeting transcription

Fixed

• Transcript Loss on Disconnect: Commit audio buffer before closing WebSocket and wait for final transcript before closing, preventing lost transcriptions during disconnects
• Dictation IPC Callbacks: Send plain strings from streaming IPC callbacks instead of objects, fixing downstream consumers
• Accessibility Permission Detection (macOS): Fix onboarding flow not detecting macOS accessibility permission correctly (#394)
• Custom Cloud Provider Classification: Treat Custom Cloud endpoints as self-hosted rather than third-party (#384)
• Blocking execSync in Meeting Detection: Replaced synchronous process detection with async alternative to prevent UI freezes on Windows
• BYOK Onboarding Override: Guard BYOK override for signed-in users and fix missing deps during onboarding (#397)
• Windows Media Pause Toggle: Check audio state before sending media key on Windows (#402)
• Linux Wayland Portal Permissions: Set desktop name on Linux for Wayland portal permissions (#389)
• Chrome Sandbox Permissions (Linux): Set SUID bit on chrome-sandbox during deb/rpm install

Changed

• Eliminated duplication and fixed style inconsistencies in dictation streaming helpers
• Cleaned up meeting detection code after the Windows input fix

[1.6.0] - 2026-03-06

Added

• Agent Mode: Glassmorphism chat overlay with real-time AI streaming — resizable window (8 edge/corner handles), dedicated hotkey, conversation history stored in SQLite, customizable system prompt, and support for all cloud/local AI providers
• Google Calendar Integration: Connect multiple Google accounts via OAuth 2.0 (PKCE), view upcoming meetings in the sidebar, and receive notifications when meetings are detected
• Meeting Recording & Live Transcription: Automatic meeting detection via process monitoring (Zoom, Teams, FaceTime) and sustained audio activity, with live transcription powered by OpenAI Realtime API over WebSocket
• Cloud Notes with Sync: Local-first note storage with FTS5 full-text search, folder organization, cloud sync, and semantic search — all notes are instantly searchable across title, content, and enhanced content
• Audio Retention & Retry: Transcription audio is now saved locally with configurable retention (default 30 days), enabling playback from history and one-click retry of failed transcriptions through the full pipeline
• Cmd+K Command Search: Global command palette to search across notes, transcripts, and folders with real-time results, keyboard navigation, and type-grouped display
• Auto-Pause Media Playback: Automatically pauses media (Spotify, Apple Music, etc.) during dictation and resumes afterward — uses MediaRemote framework on macOS, GSMTC on Windows, and MPRIS2 on Linux
• Screen Recording Permission Flow (macOS): Optional onboarding step and in-app prompts for screen recording permission, required for meeting audio capture on macOS
• Configurable Recorder Position: Choose where the voice recorder panel appears on screen (top, bottom, center)
• Auto-Paste Toggle: New toggle in clipboard settings to enable/disable automatic pasting after transcription
• Prompt Architecture Overhaul: Centralized prompt definitions in src/config/prompts.ts with customizable agent system prompts
• Dynamic Agent Window: Agent overlay starts at full screen height with drag-to-resize support, persisted window bounds across sessions
• Save Failed Transcriptions: Failed transcriptions are now saved with their audio for later retry instead of being lost
• Cloud Backup Toggle: Unified cloud backup into a single toggle for simpler settings

Changed

• Removed Input Monitoring Requirement (macOS): Replaced CGEvent tap with NSEvent monitor for Globe/Fn key detection, eliminating the need for Input Monitoring privacy permission
• Unified Screen Recording Permission UX: Consolidated screen recording permission prompts across onboarding, meetings, and integrations into a consistent experience

Fixed

• Agent Panel Readability: Made agent panel fully opaque for better text readability
• Local Model Streaming: Fixed local model support in agent streaming and resolved Metal OOM crash on macOS
• Mic Auto-Gain: Enabled microphone auto-gain and skip silent system audio chunks during meeting recording
• Meeting Audio: Fixed simultaneous system and mic audio capture for meetings
• KDE Wayland Paste: Fixed portal exit code 0 with no token being treated as success on KDE Wayland
• Meeting Detection: Suppressed false meeting detection when no active calendar meeting exists
• OpenAI Realtime Session: Fixed session configuration timing — now sends config after session created event with pcm16 format and VAD
• Agent Hotkey Persistence: Agent hotkey now properly persists to .env file across restarts
• Sidebar Height: Fixed sidebar not extending full window height
• Empty Transcription Handling: Silent return on empty transcription instead of pasting fallback string
• Command Search Styling: Fixed input styling, note type icons, sidebar spacing, and added deleted_at column support
• Onboarding Accessibility UX: Show device name in mic settings and improve accessibility permission guidance
• Orphaned Trial Note: Removed orphaned trialNote reference from free plan pricing
• Portal-Based Tooltips: Fixed tooltip positioning and replaced download action with reveal-in-folder
• State-Aware Media Pause: Don't unpause media that was already paused before dictation started
• WebSocket Audio Buffering: Parallelized WebSocket connection and audio capture, buffer early audio to prevent data loss at meeting start
• Video Track Loopback: Keep video tracks alive for loopback audio capture, remove invalid dispose call

[1.5.5] - 2026-03-01

Added

• Mode-Aware File Size Validation: Upload UI now enforces file size limits per transcription mode — local is unlimited, BYOK and Cloud free are capped at 25 MB, Cloud pro at 500 MB — with contextual messaging and CTA buttons (Create Account, Upgrade, Switch to Cloud)
• Large File Chunking: Files over 25 MB are automatically split via FFmpeg and transcribed in parallel with per-chunk progress reporting
• Gemma 3 Local Models: Added Gemma 3 (1B, 4B, 12B, 27B) to the local model registry with provider icon
• Groq Model Updates: Added new Groq models and removed deprecated ones (Maverick, Kimi K2 Instruct)
• Notes Editor Formatting Shortcuts: Cmd+B (bold), Cmd+I (italic), Cmd+E (code) keyboard shortcuts in the notes editor
• Linux Wayland Paste Improvements: Added ydotool support and improved wl-copy reliability for Wayland paste
• Granular Build Scripts: Added individual build target scripts for more flexible CI/CD

Fixed

• Fn/Globe Hotkey: Fn key now correctly treated as equivalent to Globe key on macOS
• GPU Activation: Fixed GPU activation flow and Vulkan fallback behavior
• Groq i18n: Updated Groq model descriptions and added missing translations across all locales

[1.5.4] - 2026-02-25

Added

• Auto-Learn Correction Monitoring: Detects user edits after paste and automatically updates the custom dictionary with learned corrections; native text monitor binaries for macOS (AXObserver with PID-based AX targeting), Windows, and Linux (with download-first strategy and CI workflow for prebuilt binaries); undo button on auto-learned dictionary toast; dictionary settings UI with translations across all locales
• Config-Driven STT Routing: STT mode (batch vs streaming) now driven by /api/stt-config per context (dictation vs notes); streaming provider adapter map supports Deepgram and AssemblyAI, replacing hardcoded Deepgram IPC calls with a generic interface
• Live Toggle in Notes: "Live" toggle in NoteEditor lets users override between streaming and batch transcription for notes

Fixed

• STT Metadata Forwarding: Forward complete STT metadata (sttWordCount, sttLanguage, actual Deepgram model, audio bytes, stt_processing_ms) and client end-to-end latency (client_total_ms) to API logging
• BYOK Transcription Logging: Fixed BYOK reasoning incorrectly suppressing transcribe logs

[1.5.3] - 2026-02-24

Added

• Unified GPU Banners: Replaced dual CUDA/Vulkan banners on the home screen with a single GPU acceleration banner; added GPU banners to Transcription Settings and AI Text Enhancement Settings
• GpuStatusBadge Redesign: Auto-retry flow (download → activating → GPU active) with 15s timeout, replacing confusing "CPU Only" and "Re-detect GPU" states; swapped hardcoded hex colors for bg-success/bg-warning design tokens
• Streaming Usage Tracking: Wired up the previously-uncalled /api/streaming-usage endpoint so Deepgram streaming transcriptions report word counts to the server
• Cloud API Telemetry: Forward STT metadata (sttProvider, sttModel, processing time, audio duration/size/format) and clientVersion/clientType/appVersion to all cloud API requests
• Internationalization: Added 15 missing i18n keys (app.mic.*, app.commandMenu.*, app.toasts.*, app.oauth.*, notes.enhance.title) across all 10 locale files

Fixed

• Windows Blank Screen: Fixed blank screen on return from sleep/minimize by adding render-process-gone handler, isCrashed() health checks on show/tray/second-instance paths, backgroundColor and backgroundThrottling to window config, and disable-gpu-compositing for win32
• IPC Echo Loop: Broke infinite IPC bounce in floating icon auto-hide toggle by guarding the setter with an early return when the value hasn't changed
• GPU Banner Navigation: GPU banner "Enable GPU" button now navigates to the correct "intelligence" settings section instead of invalid "reasoning" ID
• AI CTA Deep Link: Replaced legacy "aiModels" alias with canonical "intelligence" section ID in the AI enhancement CTA button
• Custom Endpoint Routing (#311): Moved reasoningProvider === "custom" check to the top of getModelProvider() so custom endpoint models are never misrouted through built-in providers; custom models now show a neutral Globe icon
• KDE Wayland Terminal Detection: Detect Konsole via kdotool (fast path) or KWin supportInformation via qdbus (zero-install fallback) so terminals receive Ctrl+Shift+V instead of Ctrl+V
• RAM Leak on Provider Switch: Whisper, Parakeet, and llama-server processes now stop when switching to cloud providers, freeing loaded models from RAM
• Streaming Usage Session Refresh: Wrapped cloudStreamingUsage in withSessionRefresh so expired sessions auto-refresh instead of silently dropping word counts
• Duplicate Transcription Logs: Skip telemetry logging in streaming-usage and transcribe endpoints when reasoning is enabled (the /api/reason endpoint already creates the combined row)
• Usage Cache Invalidation: useUsage hook now listens for usage-changed events to invalidate its cache and refetch immediately after transcription
• macOS Binary Architecture: Added Mach-O header verification to globe-listener and fast-paste build scripts; force rebuild when architecture-specific hash file is missing; runtime architecture check before spawning binary
• Globe Key Listener Resilience: Auto-restart globe key listener on unexpected exit code 0 (sleep/wake invalidation); reset restart counter after sustained uptime; only treat "Failed to create event tap" as fatal
• Parakeet Long Recordings: Lowered max segment duration from 30s to 15s for more reliable chunked transcription; downgraded reasoning failure log from error to warn

[1.5.2] - 2026-02-24

Fixed

• Reasoning Output: Resolved empty output for Qwen3/GPT-OSS models by raising local inference minimum tokens from 100 to 512; fixed custom endpoint models misrouting by checking reasoningProvider setting before name heuristics
• Google OAuth: Added newUserCallbackURL to desktop Google OAuth flow for proper new user registration
• Linux KDE Taskbar: Prevented dictation panel from appearing in KDE taskbar
• Intel Mac CI Builds: Fixed binary architecture mismatch by installing x64 ffmpeg-static binary and preventing prebuild hooks from deleting x64 binaries on arm64 CI runners (#196)

[1.5.1] - 2026-02-23

Added

• GPU-Accelerated Local Inference: Vulkan (Windows/Linux) and Metal (macOS) support for llama-server with automatic CPU fallback and GPU status badge in the reasoning model selector
• CUDA GPU Acceleration for Whisper: NVIDIA GPU acceleration for local Whisper transcription with automatic GPU detection, upgrade banner for existing users, and shared download progress UI
• On-Demand Vulkan Download: Vulkan llama-server binary downloads on-demand when the user opts in, saving 40-46MB from the app installer

Changed

• Vulkan Llama-Server Architecture: Switched from bundling the Vulkan binary to on-demand download into userData, mirroring the Whisper CUDA download pattern

Fixed

• macOS Paste Failure: Replaced osascript-based accessibility check with Electron's native isTrustedAccessibilityClient() and fixed focus transfer using hide()+showInactive() instead of blur() on NSPanel (#313)
• Windows Sherpa-onnx Extraction: Fixed tar extraction failing on Windows due to GNU tar interpreting drive letter colons as remote host separators — now uses relative paths (#284)
• macOS Auto-Update Architecture: Detect Rosetta translation via sysctl.proc_translated so Apple Silicon users stuck on an x64 build from older releases self-heal to the native arm64 build on next update

[1.5.0] - 2026-02-23

Added

• Notes System: Full-featured note-taking built into the control panel
  • Create, edit, and organize notes with a rich Markdown editor
  • Organize notes into custom folders with a default Personal folder
  • Upload audio files for transcription directly into notes
  • Real-time dictation widget for transcribing directly into a note
  • Drag-and-drop to reorder notes and move between folders
  • Guided onboarding flow for first-time notes users
• AI Actions on Notes: Apply AI-powered actions to note content
  • Action picker with customizable processing prompts
  • Action manager dialog for creating and editing action templates
  • Processing overlay with live progress feedback
• Sidebar Navigation: Redesigned control panel with persistent sidebar
  • New ControlPanelSidebar replaces the old tab-based layout
  • Dedicated views for History, Notes, Dictionary, and Settings
  • Collapsible sidebar for more content space
• Referral Program: Invite friends to earn free Pro months
  • Referral dashboard with invite tracking and status badges
  • Email invitation flow
  • Animated spectrogram share card with unique referral code
• New AI Models: Added Claude 4.6 (Opus), Gemini 3 Flash, and Gemini 3.1 Pro to the model registry
• Settings Store: Migrated settings state management to Zustand store for better performance and shared access across components
• Note Store & Action Store: New Zustand stores for notes and AI action state

Changed

• Control Panel Architecture: Extracted History, Dictionary, and Settings into standalone views, reducing ControlPanel complexity
• Settings Refactor: Extracted bulk of useSettings hook logic into settingsStore.ts for cleaner separation of concerns
• UI Polish: Updated numerous components with improved dark mode support, consistent spacing, and refined typography
• Locale Updates: Extended all 10 language files with notes, referral, and sidebar translation keys

Fixed

• macOS Auto-Update Architecture: Detect Rosetta translation via sysctl.proc_translated so Apple Silicon users stuck on an x64 build from older releases self-heal to the native arm64 build on next update
• Linux GTK Crash: Force GTK3 on Linux startup to avoid GTK symbol crash on systems with GTK4 installed (#291)
• CI Pipeline: Added Windows paste binary and key listener download steps to the build workflow (#298)
• Buy Me a Coffee: Updated funding link username

[1.4.11] - 2026-02-13

Added

• Japanese Locale: Full Japanese UI and prompt translations
• Windows Paste Terminal Detection: Added kitty to the Windows fast paste binary's terminal class list

Changed

• Windows Push-to-Talk Refactor: Moved PTT state management (hold timing, recording tracking, cooldown) from main process into windowManager for cleaner separation and consistency with macOS PTT patterns
• Audio Recording Reentrancy Guards: Added lock refs to useAudioRecording start/stop to prevent concurrent calls from rapid key presses
• Synchronous Activation Mode: getActivationMode() is now synchronous (reads from cache), removing unnecessary async overhead in all PTT and hotkey handlers
• Default Agent Name: Set default agent name to OpenWhispr

Fixed

• Hide vs Minimize: Dictation panel now consistently hides (rather than minimizing on Windows/Linux) for uniform cross-platform behavior
• Minimized Window Restore: Dictation panel restores from minimized state before showing, preventing invisible panel on Windows

[1.4.10] - 2026-02-13

Added

• Deepgram Streaming Liveness Check: Detects unresponsive warm connections within 2.5s and transparently reconnects with audio replay
• Batch Transcription Fallback: If streaming produces no text, automatically falls back to batch transcription via OpenWhispr Cloud
• Full Locale Codes: Pass full locale codes (e.g. en-US, zh-CN) to Deepgram instead of stripping to base codes, preserving dialect precision

Fixed

• Deepgram Token Expiry: Fixed token expiry clock resetting on every re-warm cycle, which prevented detection of expired tokens and caused persistent 401 errors
• Deepgram 401 Recovery: Invalidate cached tokens on authentication failures so subsequent attempts fetch fresh tokens instead of retrying stale ones

[1.4.9] - 2026-02-12

Fixed

• Deepgram Nova-3 Language Fallback: Automatically fall back to Nova-2 for languages not yet supported by Nova-3 (e.g., Chinese, Thai), preventing 400 Bad Request errors. Also switches from keyterm to keywords parameter when using Nova-2.

[1.4.8] - 2026-02-12

Added

• Referral Program: Invite friends to earn free Pro months with referral dashboard, email invitations, invite tracking with status badges, and animated spectrogram share card with unique referral code
• Notes System: Added sidebar navigation with notes system and dictionary view for organizing transcriptions
• Folder Organization: Notes can be organized into custom folders with a default Personal folder, folder management UI, and folder-aware note filtering. Upload flow now includes folder selection
• Internationalization v1: Full desktop localization across auth, settings, hooks, and UI with centralized renderer locale resources (#258)
• Chinese Language Split: Split Chinese into Simplified (zh-CN) and Traditional (zh-TW) with tailored AI instructions and one-time migration for existing users (#267)
• Russian Interface Language: Added Russian to interface language options
• Deepgram Token Refresh & Keyterms: Proactive token rotation for warm connections before expiry and keyterms pass-through for improved transcription accuracy

Fixed

• macOS Non-English Keyboard Paste: Fixed paste not working on non-English keyboard layouts (Russian, Ukrainian, etc.) by using physical key code instead of character-based keystroke in AppleScript fallback
• Whisper Language Auto-Detection: Pass --language auto to whisper.cpp explicitly so non-English audio isn't forced to English (#260)
• Model Download Pipeline: Inline redirect handling, deferred write stream creation, indeterminate progress bar for unknown sizes, and Parakeet ONNX file validation after extraction
• Sherpa-onnx Shared Libraries: Always overwrite shared libraries during download to prevent stale architecture-mismatched binaries, with --force support
• Chinese Translation Fixes: Minor translation corrections for Chinese interface strings
• Neon Auth Build Config: Fixed auth build configuration

[1.4.7] - 2026-02-11

Added

• Deepgram Streaming Transcription: Migrated real-time streaming transcription from AssemblyAI to Deepgram for improved reliability and accuracy (#249)

Fixed

• BYOK After Upgrade: Prefer localStorage API keys over process.env so Bring Your Own Key mode works correctly after upgrading (#263)
• PTT Double-Fire Prevention: Applied post-stop cooldown and press-identity checks to both macOS and Windows push-to-talk handlers
• Archive Extraction Retry: Reuse existing archive on extraction retry with improved error handling
• Email Verification Polling: Pass email param in verification polling and stop on 401 responses
• Auth Build Bundling: Added @neondatabase/auth packages to rollup externals for correct production bundling (#256)
• Neon Auth Build Config: Fixed Vite build configuration for Neon Auth packages (#266)

Changed

• Build System: Bumped Node version in build files

[1.4.6] - 2026-02-10

Added

• Robust Model Downloads: Hardened download pipeline with stall detection, disk space checks, and file validation for more reliable model installs
• Prompt Handling Improvements: Improved agent name resolution, prompt studio enhancements, and smarter prompt context assembly
• Past-Due Subscription Handling: Users with past-due subscriptions now see clear messaging and recovery options

Fixed

• Parakeet Long Audio: Fixed empty transcriptions for long audio by segmenting input before sending to Parakeet
• Plus-Addressed Emails: Reject plus-addressed emails (e.g., user+tag@example.com) during authentication
• Double-Click Prevention: Prevent duplicate requests when double-clicking checkout and billing buttons
• Auth Initialization Race: Await init-user before completing auth flow and fix missing user dependency

Changed

• Startup Performance: Preload lazy chunks during auth initialization for faster page transitions
• Code Cleanup: Removed excess comments and simplified window management logic

[1.4.5] - 2026-02-09

Added

• Dictation Sound Effects Toggle: New setting to enable/disable dictation audio cues with refined tones (warmer, softer frequencies, gentler attack, distinct start/stop)
• Toast Notification Redesign: Redesigned toast notifications as dark HUD surfaces for a more polished look
• Floating Icon Auto-Hide: New setting to auto-hide the floating dictation icon
• Loading Screen Redesign: Branded loading screen with logo and spinner
• Discord Support Link: Added Discord link to the support menu
• Auth-Aware Routing: Returning signed-out users now see a re-authentication screen instead of a broken state

Fixed

• Dropdown Dark Mode: Fixed dropdown styling in dark mode
• Toast Dark Mode: Fixed toast colouring in dark mode
• Globe Key Persistence: Globe key now persists to .env and dictation key syncs to localStorage
• Globe Listener Cross-Compilation: Cross-compiled globe listener for x64

Changed

• Startup Performance: Deferred non-critical manager initialization after window creation, lazy-loaded ControlPanel/OnboardingFlow/SettingsModal, converted env file writes to async, extracted SettingsProvider context, and split Radix/lucide into separate vendor chunks
• Scrollbar Styling: Subtle transparent-track scrollbar with thinner floating thumb

[1.4.4] - 2026-02-08

Fixed

• AI Enhancement CTA Persistence: Dismissing the "Enable AI Enhancement" banner now persists to localStorage so it stays hidden across sessions

Changed

• Code Cleanup: Removed excess comments and section dividers in ControlPanel

[1.4.3] - 2026-02-08

Added

• Mistral Voxtral Transcription: Added Mistral as a cloud transcription provider with Voxtral Mini model and custom dictionary support via context_bias
• TypeScript Compilation: Added TypeScript as an explicit dev dependency with project-level tsconfig.json

Fixed

• Linux Wayland Clipboard: Persistent clipboard ownership on Wayland so Ctrl+V works reliably after transcription
• Linux Window Flickering: Fixed transparent window flickering on Wayland and X11 compositors
• Windows Modifier-Only Hotkeys: Support modifier-only hotkeys on Windows via native keyboard hook
• Update Installation: Resolved quitAndInstall hang by removing close listeners that block window shutdown during updates
• Custom System Prompts: Pass custom system prompt to local and Anthropic BYOK reasoning
• Audio Cue Audibility: Improved dictation start/stop audio cue volume
• Language Selector: Fixed dropdown positioning and sizing inside settings modal
• Type Safety: Tightened Electron IPC callback return types, model picker styles, toast variant types, and event handler signatures across the codebase

Changed

• Code Cleanup: Removed excess comments, section dividers, and redundant JSDoc across components, hooks, and utilities

[1.4.2] - 2026-02-07

Fixed

• AssemblyAI Streaming Reliability: Fixed real-time WebSocket going silent after idle periods by adding keep-alive pings, readyState validation, re-warm recovery, and connection death handling

[1.4.1] - 2026-02-07

Added

• Runtime .env Configuration: Environment variables now reload at runtime without requiring app restart
• Settings Retention on Pro: Pro subscribers retain their settings when managing their subscription

Fixed

• macOS Microphone Permission: Resolved hardened-runtime mic permission prompt by routing through main-process IPC and unifying API key cache invalidation with event-based AudioManager sync
• AudioWorklet ASAR Loading: Inlined AudioWorklet as blob URL to fix module loading failure in packaged ASAR builds
• Google OAuth Flow: OAuth now opens in the system browser with deep link callback instead of navigating the Electron window
• Auth Security Hardening: Safe JSON parsing, guarded URL constructor, and fixed error information leaks in auth code
• Deep Link Focus: Control panel now correctly receives focus when opened via deep link
• Neon Auth Electron Compatibility: Routed auth flows through API proxy and fixed Origin header rejection for desktop app
• Billing Error Visibility: Checkout and billing errors now surface as toast notifications instead of failing silently
• Hotkey Persistence: Added file-based hotkey storage for reliable startup persistence (#181)
• Email Verification: Disabled Neon Auth email verification step for smoother onboarding

Changed

• Build Optimization: Binary dependencies are now cached during build for faster CI
• UI Polish: Fixed scrollbar styling, provider button styling, and voice recorder icon fill

[1.4.0] - 2026-02-06

Added

• OpenWhispr Cloud: Cloud-native transcription service — sign in and transcribe without managing API keys
  • Google OAuth and email/password authentication via Neon Auth
  • Email verification flow with polling and resend support
  • Password reset via email magic links
• Subscription & Billing: Free and Pro plans with Stripe-powered payments
  • Free plan with rolling weekly word limits (2,000 words/week)
  • Pro plan with unlimited transcriptions
  • 7-day free trial for new accounts with countdown display
  • In-app upgrade prompts when approaching or reaching usage limits
  • Stripe billing portal access for Pro subscribers
• Usage Tracking: Real-time usage display with progress bar, color-coded thresholds, and next billing date
• Account Section in Settings: Profile display, plan status badge, usage bar, billing management, and sign out
• Upgrade Prompt Dialog: When usage limit is reached, offers three paths — upgrade to Pro, bring your own key, or switch to local
• Cancel Processing Button: Cancel ongoing transcription processing mid-flight
• Dynamic Window Resizing: Window automatically resizes based on command menu and toast visibility
• Dark Mode Icon Inversion: Monochrome provider icons now automatically invert in dark mode for better visibility

Changed

• Onboarding Redesign: Auth-first onboarding flow
  • Signed-in users get a streamlined 3-step flow (Welcome → Setup → Activation)
  • Non-signed-in users get a 4-step flow with transcription mode selection
  • Permissions merged into Setup step for signed-in users
• Transcription Mode Architecture: Unified mode selection across OpenWhispr Cloud, Bring Your Own Key (BYOK), and Local
  • Signed-in users default to OpenWhispr Cloud
  • Non-signed-in users choose between BYOK and Local
• Design System Overhaul: Complete refactor of styling to use design tokens throughout the codebase
  • Button component now uses text-foreground, bg-muted, border-border instead of hardcoded hex values
  • Removed hardcoded classes and inline styles across components
  • Improved button and badge consistency
• Settings UI Redesign: Overhauled all settings pages with unified panel system, redesigned sidebar, and extracted permissions section
• Dark Mode Polish: Premium button styling, glass morphism toasts, and streamlined visuals
• App Channel Isolation: Development, staging, and production channels now use isolated user data directories

Fixed

• Light Mode UI Visibility: Fixed multiple UI elements that were invisible or hard to see in light mode:
  • Settings gear icon in permission cards now uses text-foreground
  • Troubleshoot button uses proper foreground color
  • Reset button in developer settings now correctly shows destructive color
  • Settings and Help icons in the toolbar are now properly visible
  • Check for Updates button now renders correctly in light mode
• Provider Tab Flashing: Resolved TranscriptionModelPicker tab flashing by extracting ModeToggle component and syncing internal state with props
• Local Reasoning Model Persistence: Fixed local reasoning model selection not persisting correctly
• Parakeet Model Status: Added dedicated IPC channel for Parakeet model status checks
• Groq Qwen3 Models: Removed thinking tokens from Qwen3 models on Groq provider
• OAuth Session Grace Period: Automatic session refresh with exponential backoff retry during initial OAuth establishment

[1.3.3] - 2026-01-28

Added

• ONNX Warm-up Inference: Parakeet server now runs warm-up inference on start to eliminate first-request latency from JIT compilation
• Startup Preferences Sync: Renderer startup preferences are now synced to .env for server pre-warming on restart

Changed

• macOS Tray Behavior: Hide to tray on macOS for consistent cross-platform behavior

Fixed

• macOS Launch Crash: Added disable-library-validation entitlement to resolve macOS launch crash (#120)
• Reasoning Model Default: Fixed useReasoningModel not correctly defaulting to enabled by persisting useLocalStorage defaults and aligning direct reads
• Windows Non-ASCII Usernames: Resolved whisper-server crash on Windows with non-ASCII usernames by pre-converting audio to WAV and routing temp files through ASCII-safe directory
• Windows Paths with Spaces: Fixed temp directory fallback to also detect paths with spaces on Windows

[1.3.2] - 2026-01-27

Changed

• Linux Paste Tools: Prefer xdotool over ydotool for better compatibility

Fixed

• Windows Zip Extraction: Use tar instead of PowerShell Expand-Archive for zip extraction on Windows to avoid issues with special characters

[1.3.1] - 2026-01-27

Changed

• Download System Refactor: Consolidated model download logic into shared utilities with resume support, retry logic, abort signals, and improved installing state UI
• Throttled Progress Display: Whisper model download progress updates are now throttled for smoother UI

[1.3.0] - 2026-01-26

Added

• NVIDIA Parakeet Support: Fast local transcription via sherpa-onnx runtime with INT8 quantized models
  • parakeet-tdt-0.6b-v3: Multilingual (25 languages), ~680MB
• Windows Push-to-Talk: Native Windows key listener with low-level keyboard hook for true push-to-talk functionality
  • Supports compound hotkeys like Ctrl+Shift+F11 or CommandOrControl+Space
  • Prebuilt binary automatically downloaded from GitHub releases
  • Fallback to tap mode if binary unavailable
• Custom Dictionary: Improve transcription accuracy for specific words, names, and technical terms
  • Add custom words through Settings → Custom Dictionary
  • Words are passed as hints to Whisper for better recognition
  • Works with both local and cloud transcription
• GitHub Actions Workflow: Automated CI workflow to build and release Windows key listener binary
• Shared Download Utilities: New scripts/lib/download-utils.js module with reusable download, extraction, and GitHub release fetching functions

Changed

• Download Scripts Refactored: All download scripts now use shared utilities for consistency
• GitHub API Authentication: Download scripts support GITHUB_TOKEN to avoid API rate limits in CI
• Debug Logging Cleanup: Extracted common window loading code and cleaned up debug logging

Fixed

• GNOME Wayland Hotkey Improvements: Improved hotkey handling on GNOME Wayland
• Hotkey Persistence: Fixed hotkey selection not persisting correctly
• Custom Endpoint API Keys: Fixed custom endpoint API keys not persisting to .env file
• Custom Endpoint State: Fixed custom endpoint using shared state instead of its own
• Linux Stale Hotkey Registrations: Clear stale hotkey registrations on startup on Linux
• Wayland XWayland Paste: Try xdotool on Wayland when XWayland is available
• llama-server Libraries: Bundle llama-server shared libraries and search from extract root for varying archive structures
• STT/Reasoning Debug Logging: Added missing debug logging for STT and reasoning pipelines

[1.2.16] - 2026-01-24

Fixed

• App Startup Hang: Fixed app initialization timing issues with Electron 36+
• Manager Initialization: Deferred manager initialization until after app.whenReady() to prevent hangs
• Debug Logger Initialization: Deferred debugLogger file initialization until app.whenReady()
• Config Bundling: Fixed missing config files in production builds
• whisper.cpp Binary Version: Updated whisper.cpp release names and bumped binary version

[1.2.15] - 2026-01-22

Added

• ydotool Fallback for Linux: Added ydotool as additional fallback option for clipboard paste operations on Linux systems

Changed

• Unified Prompt System: Refactored to single intelligent prompt system for improved consistency and maintainability
• whisper.cpp Remote: Refactored remote whisper.cpp integration for better reliability

[1.2.14] - 2026-01-22

Added

• Troubleshooting Mode: New debug logging section in settings with toggle for detailed diagnostic logs, log file path display, and direct folder access for easier support
• Custom Transcription Endpoint: Support for custom OpenAI-compatible transcription endpoints with configurable base URLs
• Enhanced Clipboard Debugging: Detailed clipboard operation logging for diagnosing paste issues across platforms

Changed

• API Key Management: Consolidated and refactored API key persistence with improved .env file handling and recovery mechanisms
• Local Network Detection: Refactored URL detection into reusable utility for better code organization
• Electron Builder: Updated to latest version for improved build performance

Fixed

• Windows/Linux Taskbar: Prevented dual taskbar entries on Windows and Linux by properly configuring window behavior
• Single Instance Lock: Enforced single instance lock with cleaner window state checks
• Model Provider Consistency: Removed redundant fallbacks and ensured consistent use of getModelProvider()
• Cross-env Support: Fixed Windows compatibility in pack script using cross-env
• Linux X11 Paste: Improved paste reliability by capturing target window ID upfront with windowactivate --sync, added xdotool type fallback for terminals
• Tray Minimize: Fixed minimize to tray functionality

[1.2.12] - 2026-01-20

Added

• LLM Download Cancellation: Added ability to cancel in-progress local LLM model downloads with throttled progress updates to prevent UI flashing

Changed

• Gemini Model Updates: Updated Gemini models to latest versions
• Linux Wayland Improvements: Improved Wayland paste detection with GNOME-specific handling and XWayland fallback support
• whisper.cpp CUDA Support: Updated whisper.cpp download script to include CUDA-enabled binaries

Fixed

• Windows Paste Delay: Adjusted paste delay timing on Windows for more reliable text insertion
• Blank Audio Prevention: Fixed issue where blank/silent audio recordings would paste empty text
• Newline Handling: Fixed newline formatting issues in transcribed text

[1.2.11] - 2026-01-18

Fixed

• ASAR Path Resolution: Fixed path resolution issues for bundled resources in packaged builds
• Update Checker: Fixed auto-update checker initialization
• Build Includes: Ensured services and models are properly included in production builds
• OS Module Import: Fixed OS module import ordering

[1.2.10] - 2026-01-17

Fixed

• Streaming Backpressure: Fixed proper streaming backpressure handling in audio processing
• Quit and Install: Fixed update installation on app quit

[1.2.9] - 2026-01-17

Fixed

• Path Resolution: Improved path resolution for better cross-platform compatibility

[1.2.8] - 2026-01-16

Added

• Microphone Input Selection: Choose your preferred microphone input device in settings, with built-in mic preference to prevent Bluetooth audio interruptions
• Push to Talk Mode: New recording mode option alongside the existing toggle mode
• Hotkey Listening Mode: Prevents conflicts when capturing new hotkeys by temporarily disabling the global hotkey
• Hotkey Fallback System: Automatic fallback with user notifications when preferred hotkey is unavailable
• Cross-Platform Accessibility Settings: Quick access to system accessibility settings on macOS

Changed

• Streamlined Onboarding: Removed redundant "How it Works" section, success dialogs, and manual save buttons for a smoother setup experience
• Improved Select Styling: Enhanced dropdown select component appearance

Fixed

• FFmpeg Availability Types: Corrected type definitions and optimized whisper-cpp download process
• Whisper Models Path: Fixed model storage path resolution
• Better Path Resolution: Improved error handling for file paths
• Open Mic Settings: Fixed system settings link for microphone configuration

[1.2.7] - 2026-01-13

Added

• Whisper Server HTTP Mode: Added persistent whisper-server for faster repeated transcriptions with automatic CLI fallback
• Pipeline Timing Instrumentation: Added detailed timing logs for each stage of the transcription pipeline
• Whisper Server Pre-warming: Server pre-warms on startup for faster first transcription

Changed

• Windows Clipboard: Reduced clipboard delays for faster text pasting on Windows

Fixed

• Windows Update Install: Simplified Windows update installation by using silent mode and removing redundant before-quit handling
• Mac Build Workflows: Fixed CI/CD to run separate workflows for Mac builds
• Mac DMG Build Race Condition: Fixed release workflow DMG build failure caused by concurrent arm64/x64 builds mounting same volume
• Windows Download Script: Fixed PowerShell Expand-Archive failure with bracket characters in directory names

[1.2.6] - 2026-01-13

Changed

• Settings Layout: Moved settings navigation to left side on Windows and Linux for improved consistency

Fixed

• Linux Whisper Detection: Fixed issue where Python-based Whisper could be used instead of whisper.cpp on Linux systems

[1.2.5] - 2026-01-13

Added

• Model Validation: Added validation when deleting or loading Whisper models to ensure model integrity
• Download Cancellation: Added ability to cancel in-progress model downloads in whisper pickers
• Windows Paste Performance: Added nircmd for faster text pasting on Windows

Fixed

• EventEmitter Memory Leak: Fixed memory leak caused by duplicate listener registration in useUpdater hook across ControlPanel and SettingsPage components
• FFmpeg Path Resolution: Fixed FFmpeg path resolution in unpacked ASAR for local whisper.cpp transcription

Changed

• UI Cleanup: Removed redundant UI elements for a cleaner interface

[1.2.4] - 2026-01-13

Changed

• whisper.cpp Packaging: Moved whisper.cpp binaries from ASAR to extraResources for improved reliability and faster startup

Fixed

• Package Lock Sync: Fixed package-lock.json synchronization with package.json dependencies

[1.2.3] - 2026-01-13

Added

• Extended Hotkey Support: Added numpad keys, media keys, and additional special keys (Pause, ScrollLock, PrintScreen, NumLock) for hotkey selection
• Improved Hotkey Error Messages: Registration failures now include helpful suggestions for alternative hotkeys

Changed

• Linux Paste Tools: Only show paste tools installation prompt on Linux when tools are not available

Fixed

• Hotkey Debugging: Added comprehensive debug logging to hotkey manager for troubleshooting registration issues

[1.2.2] - 2026-01-13

Fixed

• React Version Mismatch: Fixed blank screen caused by incompatible React and React-DOM versions in package-lock.json

[1.2.1] - 2026-01-13

Fixed

• Blank Screen on Upgrade: Fixed white screen issue for users upgrading from older versions with different onboarding step counts. The onboarding step index is now properly clamped to valid range.

[1.2.0] - 2026-01-13

Added

• Delete All Whisper Models: New option to delete all downloaded Whisper models at once
• Model Deletion Confirmation: Added confirmation dialog when deleting models in settings

Changed

• Migrated to whisper.cpp: Replaced Python-based Whisper with native whisper.cpp for faster, more reliable transcription
  • No longer requires Python installation
  • WebM-to-WAV audio conversion built-in
  • Significantly improved startup and transcription speed
• Streamlined Onboarding: Simplified setup flow with fewer steps now that Python is not required
• Download Cancellation: Added ability to cancel in-progress model downloads
• CI/CD Updates: Updated build and release workflows

Fixed

• IPC Handler: Fixed broken IPC handler for model operations
• Logging: Standardized logging across the application
• React Hook Dependencies: Improved React hook dependency arrays for better performance
• Button Styling: Fixed button styling consistency across the application

Removed

• Python Dependency: Removed Python requirement and all related installation code
• whisper_bridge.py: Removed Python-based Whisper bridge in favor of native whisper.cpp

[1.1.2] - 2026-01-12

Added

• Linux Package Dependencies: Recommended xdotool, wtype, and python3 packages for Linux users

Fixed

• Python Installation Race Condition: Fixed race condition in Python installation check that could cause installation to fail or hang

[1.1.1] - 2026-01-12

Added

• Cross-Platform Paste Tools Detection: Onboarding now detects and guides users through installing paste tools on Linux and Windows with auto-grant accessibility

Changed

• Qwen Model Compatibility: Disabled thinking mode for Qwen models on Groq to prevent compatibility issues
• Model Registry Refactor: disableThinking flag now uses the centralized model registry
• Consolidated ColorScheme Types: Removed redundant default exports and cleaned up inline font styles
• Provider Icons: Use static imports for provider icons to fix Vite bundling issues

Fixed

• Recording Cancellation: Restored cancel recording functionality that was accidentally removed
• Model Downloads: Implemented atomic downloads with temp file pattern and robust cleanup handling for cross-platform reliability
• Incomplete Download Prevention: Model file size validation now prevents incomplete downloads from showing as complete
• Windows PowerShell Performance: Optimized paste startup time on Windows

[1.1.0] - 2026-01-10

Added

• Compound Hotkey Support: Use multi-key combinations like Cmd+Shift+K or Ctrl+Alt+D for dictation
• Groq API Integration: Ultra-fast AI inference with Groq's cloud API
• Auto-Update UI: Download progress bars and install button in settings
• Recording Cancellation: Cancel an in-progress recording without transcribing
• Release Notes Viewer: Markdown-rendered release notes in settings

Changed

• Major Hotkey Refactor: Complete rewrite of hotkey selection with improved reliability and validation
• Consolidated Model Registry: Single source of truth for all AI models (modelRegistryData.json)
• Unified Model Picker: Reusable component for both transcription and reasoning model selection
• Improved Latency Logging: Numbered stage logs for recording, transcription, reasoning, and paste timing
• Reduced Paste Delay: Lowered from 100ms to 50ms for faster text insertion
• Code Quality: Added ESLint, Prettier for JS/TS, and Ruff for Python

Fixed

• Windows 11 Compatibility: Fixed PATH separator, cache directories, and process termination
• Python Virtual Environment: Fixed race condition and added Arch Linux venv support
• Microphone Detection: Improved onboarding flow for missing inputs with deep-linking to system settings
• Recording State Alignment: Recording now aligns to MediaRecorder's actual start/stop events
• Caching Optimizations: Cached accessibility, paste tool, and FFmpeg checks to reduce process spawns
• Window Titles: Electron window titles now set correctly after page load

[1.0.15] - 2026-01-05

Added

• Button to fully quit OpenWhispr processes from the application
• Linux terminal detection with automatic paste key switching (Ctrl+Shift+V for terminals)

Changed

• Standardized logging on log levels with renderer IPC and .env refresh for consistent debug output

Fixed

• Use kdotool for Wayland terminal detection, improving clipboard paste reliability
• Increased delay before restoring clipboard to avoid race conditions during paste operations
• Persist OpenAI key before onboarding test to prevent key loss during setup
• Windows Python discovery now correctly handles output parsing
• Keep FFmpeg debug schema as boolean type
• Fixed OpenWhispr documentation paths
• Windows: Resolved issue #16 with WAV validation, registry-based Python detection, and normalized FFmpeg paths

[1.0.13] - 2025-12-24

Added

• Enhanced Linux support with Wayland compatibility, multiple package formats (AppImage, deb, rpm, Flatpak), and native window controls
• Auto-detect existing Python during onboarding and gate the installer with a recheck option
• "Use Existing Python" skip flow to onboarding with confirmation dialog

Changed

• Reuse audio manager and stabilize dictation toggle callback to fix recording latency
• Add cleanup functions to IPC listeners to prevent memory leaks
• Make Flatpak opt-in for local builds only

Fixed

• Optimized transcription pipeline with caching, batched reads, and non-blocking operations for improved performance
• Reference error in settings page
• Removed redundant audio listener causing unnecessary processing
• Added IPC listener cleanup to prevent memory leaks
• Performance improvements: removed duplicate useEffect, fixed blur causing re-renders

CI/CD

• Add caching for Electron and Flatpak downloads
• Add Flatpak runtime installation to workflow
• Add Linux packaging dependencies to GitHub Actions workflow

[1.0.12] - 2025-11-13

Added

• Added scripts/complete-uninstall.sh plus a new TROUBLESHOOTING guide so you can collect arch diagnostics, clean caches, and reset permissions before reinstalling stubborn builds.
• Control Panel history now auto-refreshes through a shared store and IPC events, so new, deleted, or cleared transcripts sync instantly without a manual refresh.
• Distribution artifacts now include both Apple Silicon and Intel macOS DMG/ZIP outputs, and the README documents Debian/Ubuntu packaging along with optional xdotool support.

Changed

• The onboarding flow now validates dictation hotkeys before letting you continue, remembers whether cloud auth was skipped, and only persists sanitized API keys once supplied.
• History entries normalize timestamps and no longer run the removed legacy text cleanup helper, so the UI shows the exact Whisper output that was saved.

Fixed

• Local Whisper now finds Python on Windows more reliably by scanning typical install paths, honoring OPENWHISPR_PYTHON, and surfacing actionable ENOENT guidance.
• Whisper installs automatically retry pip operations that hit PEP‑668, TOML, or permission errors, sanitizing the output and falling back to --user + legacy resolver when needed.

[1.0.11] - 2025-10-13

Added

• Settings, onboarding, and the AI model selector now accept OpenAI-compatible custom base URLs for both transcription and reasoning providers, complete with validation and reset helpers.
• Windows now gets full tray behavior: closing the control panel hides it to the tray, left-click reopens it, and the UI adds a native close button.

Changed

• ReasoningService sends both input and messages payloads and automatically falls back between /responses and /chat/completions so older OpenAI-compatible endpoints keep working.

Fixed

• Successful endpoint detection is cached per base URL, so the app remembers whether to call /responses or /chat/completions instead of retrying the wrong path forever.
• Custom endpoint fields now enforce HTTPS (with localhost as the lone exception) across the UI and services, preventing API keys from ever leaving over plain HTTP.

[1.0.10] - 2025-10-07

Added

• Added a compile:globe build step that emits a macOS Globe listener binary into resources/bin before every dev, pack, or dist command so the hotkey ships with all builds.

Fixed

• Globe key failures now raise a macOS dialog, verify the bundled binary is executable, and kill/restart the listener cleanly so the shortcut survives packaging.

[1.0.9] - 2025-10-07

Changed

• Simplified the release workflow by removing the bespoke GitHub release job and letting electron-builder upload draft releases directly.

[1.0.8] - 2025-10-03

Fixed

• Globe/Fn hotkey reliability improved by showing the dictation panel before toggling, making focus optional, and surfacing listener spawn errors instead of failing silently.

[1.0.7] - 2025-10-03

Added

• Settings update controls now show download progress bars, install countdowns, and clearer messaging while fetching or installing new builds.

Changed

• Auto-update internals now track listeners, cache the last release metadata, and keep auto-download/auto-install disabled until the user explicitly triggers an update, eliminating the previous memory leaks.

Fixed

• Install & Restart now emits before-quit, enables autoInstallOnAppQuit, logs progress, and calls quitAndInstall(false, true) so updates actually apply when quitting or pressing the button.

[1.0.6] - 2025-09-11

Added

• Dictation Panel Command Menu: Clicking the floating panel reveals quick actions, including a one-click "Hide this for now" option.
• macOS Globe Key Support: Added a lightweight Swift listener so the Globe/Fn key can toggle dictation across the system.
• Globe Key Selection UI: Settings and onboarding keyboards now include a dedicated Globe key option.
• Hotkey Validation: Settings and onboarding now verify shortcut registration immediately, alerting users when a key can’t be bound.
• Model Cache Cleanup: Added an in-app command (and installer/uninstaller hooks) to delete all cached Whisper models.
• Tray Controls: macOS tray menu gained quick actions to show or hide the dictation panel.

Changed

• Dictation Overlay Placement: Window now anchors to the active workspace's bottom-right corner with a safety margin, preventing it from sliding off-screen on multi-monitor setups.
• Dictation Overlay Canvas: Enlarged the floating window so tooltips, menus, and error states render without being clipped while keeping click-through behaviour outside interactive elements.
• Keyboard UX: Virtual keyboard hides macOS-exclusive keys on Windows/Linux and standardises hotkey labels.

Fixed

• macOS Window Lifecycle: Ensured the dictation panel keeps the app visible in Dock and Command-Tab while retaining floating behaviour across spaces.
• Control Panel Stability: Reworked close/minimize handling so the panel stays interactive when switching apps and reopens cleanly without spawning duplicate windows.
• Always-On-Top Enforcement: Centralised the logic that reapplies floating window levels, eliminating redundant timers and focus quirks.
• Menu Labelling: macOS application menu items now display the correct OpenWhispr casing instead of "open-whispr".
• Non-mac Hotkey Guard: Prevented the mac-only Globe shortcut from being saved on Windows/Linux.

[1.0.5] - 2025-09-10

Fixed

• Build System: Fixed native module signing conflicts on macOS
  • Added npmRebuild: true to force rebuild of native modules during packaging
  • Added buildDependenciesFromSource: true to compile native dependencies from source
  • Added better-sqlite3 to asarUnpack array to properly unpack SQLite3 native module
  • Resolves "different Team IDs" error when launching notarized macOS apps
• CI/CD Pipeline: Fixed automated release workflow issues
  • Removed automatic version update step from release workflow (version should be set before tagging)
  • Added contents: write permission to allow workflow to create GitHub releases
  • Fixes "Resource not accessible by integration" error during releases

Technical Details

• This is a maintenance release focusing on build reliability and deployment infrastructure
• No feature changes or user-facing functionality updates
• All changes related to packaging, signing, and automated release processes

[1.0.4] - 2025-09-09

Added

• Multi-Provider AI Support: Integrated three major AI providers for text processing
  • OpenAI: Complete model suite including:
    • GPT-5 Series (Nano/Mini/Full) - Latest generation with deep reasoning
    • GPT-4.1 Series (Nano/Mini/Full) - Enhanced coding, 1M token context, June 2024 knowledge
    • o-series (o3/o3-pro/o4-mini) - Advanced reasoning models with extended thinking time
    • GPT-4o/4o-mini - Multimodal models with vision support
  • Anthropic: Claude Opus 4.1, Sonnet 4, and 3.5 variants for frontier intelligence
  • Google: Gemini 2.5 Pro/Flash/Flash-Lite and 2.0 Flash for advanced processing
• OpenAI Responses API Integration: Migrated from Chat Completions to the new Responses API
  • Simplified request format with input array instead of messages
  • New response parsing for output items with typed content
  • Automatic handling of model-specific requirements
  • Better support for GPT-5 and o-series reasoning models
• Enhanced Reasoning Service: Complete TypeScript rewrite with provider abstraction
  • Automatic provider detection based on selected model
  • Secure API key caching with TTL
  • Unified retry strategies across all providers
  • Provider-specific token optimization (up to 8192 for Gemini)
• Comprehensive Debug Logging: Enhanced reasoning pipeline with stage-by-stage logging
  • Provider selection and routing logs
  • API key retrieval and validation logs
  • Request/response details for all providers
  • Error tracking with full stack traces
• Improved Settings UI: Comprehensive API key management for all providers
  • Color-coded provider sections (OpenAI=green, Anthropic=purple, Gemini=blue)
  • Inline API key validation and secure storage
  • Provider-specific model selection with descriptions

Changed

• Default AI Model: Updated from GPT-3.5 Turbo to GPT-4o Mini for cost-efficient multimodal support
• Model Updates: Refreshed all AI models to their latest 2025 versions
  • OpenAI: Added GPT-5 family (released August 2025), migrated to Responses API
  • Anthropic: Updated to Claude Opus 4.1 and Sonnet 4, fixed model naming
  • Gemini: Added latest 2.5 series models, increased token limits
• ReasoningService: Migrated from JavaScript to TypeScript for better type safety
• API Endpoint Updates:
  • OpenAI: Migrated from /v1/chat/completions to /v1/responses
  • Request format simplified for better performance
  • Response parsing updated for new output structure
• Model Configuration Improvements:
  • Fixed Anthropic model names (using hyphens instead of dots)
  • Increased Gemini 2.5 Pro token limits (2000 minimum)
  • Removed temperature parameter for GPT-5 and o-series models
• Documentation: Updated CLAUDE.md, README.md with comprehensive provider information

Fixed

• API Key Persistence: All provider keys now properly save to .env file
  • Added saveAllKeysToEnvFile() method for consistent persistence
  • Keys reload automatically on app restart
  • Fixed Gemini and Anthropic key storage issues
• CORS Issues: Anthropic API calls now route through IPC handler
  • Avoids browser CORS restrictions in renderer process
  • Proper error handling in main process
• Empty Response Handling: Fixed "No text transcribed" error when AI returns empty
  • Falls back to original text when API returns nothing
  • Properly handles edge cases in response parsing
• Parameter Compatibility: Fixed OpenAI API parameter errors
  • GPT-5 models use simplified parameters (no max_tokens)
  • o-series models configured without temperature
  • Older models retain full parameter support

Technical Improvements

• Added Gemini API integration with proper authentication flow
• Implemented SecureCache utility for API key management
• Enhanced IPC handlers for multi-provider support
• Updated environment manager with Gemini key storage
• Improved error handling with provider-specific messages
• Added comprehensive retry logic with exponential backoff
• Enhanced error messages with detailed logging
• Better fallback strategies for API failures
• Improved response validation and parsing
• Centralized API configuration in constants file
• Unified debugging system across all providers

[1.0.3] - 2024-12-20

Added

• Local AI Models: Integration with community models for complete privacy
  • Support for Llama, Mistral, and other open-source models
  • Local model management UI with download progress
  • Automatic model validation and testing
• Enhanced Security: Improved API key storage and management
  • System keychain integration where available
  • Encrypted localStorage fallback
  • Automatic key rotation support

Fixed

• Resolved issues with Whisper model downloads on slow connections
• Fixed clipboard pasting reliability on Windows 11
• Improved error messages for better debugging
• Fixed memory leaks in long-running sessions

Changed

• Optimized audio processing pipeline for 30% faster transcription
• Reduced app bundle size by 15MB through dependency optimization
• Improved startup time by lazy-loading heavy components

[1.0.2] - 2024-12-19

Added

• Automatic Python Installation: The app now detects and offers to install Python automatically
  • macOS: Uses Homebrew if available, falls back to official installer
  • Windows: Downloads and installs official Python with proper PATH configuration
  • Linux: Uses system package manager (apt, yum, or pacman)
• Enhanced Developer Experience:
  • Added MIT LICENSE file
  • Improved documentation for personal vs distribution builds
  • Added FAQ section to README
  • Added security information section
  • Clearer prerequisites and setup instructions
  • Added comprehensive CLAUDE.md technical reference
• Dock Icon Support: App now appears in the dock with activity indicator
  • Changed LSUIElement from true to false in electron-builder.json
  • App shows in dock on macOS with the standard dot indicator when running

Changed

• Updated supported language count from 90+ to 58 (actual count in codebase)
• Improved README structure for better open source experience

[1.0.1] - 2024-XX-XX

Added

• Agent Naming System: Personalize your AI assistant with a custom name for more natural interactions
  • Name your agent during onboarding (step 6 of 8)
  • Address your agent directly: "Hey [AgentName], make this more professional"
  • Update agent name anytime through settings
  • Smart AI processing distinguishes between commands and regular dictation
  • Clean output automatically removes agent name references
• Draggable Interface: Click and drag the dictation panel to any position on screen
• Dynamic Hotkey Display: Tooltip shows your actual hotkey setting instead of generic text
• Flexible Hotkey System: Fixed hardcoded hotkey limitation - now fully respects user settings

Changed

• [BREAKING] Removed click-to-record functionality to prevent conflicts with dragging
• UI Behavior: Recording is now exclusively controlled via hotkey (no accidental triggering)
• Tooltip Text: Shows "Press {your-hotkey} to speak" with actual configured hotkey
• Cursor Styles: Changed to grab/grabbing cursors to indicate draggable interface

Fixed

• Hotkey Bug: Fixed issue where hotkey setting was stored but not actually used by global shortcut
• Documentation: Updated all docs to reflect current UI behavior and hotkey system
• User Experience: Eliminated confusion between drag and click actions

Technical Details

• Agent Naming Implementation:
  • Added centralized agent name utility (src/utils/agentName.ts)
  • Enhanced onboarding flow with agent naming step
  • Updated ReasoningService with context-aware AI processing
  • Added agent name settings section with comprehensive UI
  • Implemented smart prompt generation for agent-addressed vs regular text
• Added IPC handlers for dynamic hotkey updates (update-hotkey)
• Implemented window-level dragging using screen cursor tracking
• Added real-time hotkey loading from localStorage in main dictation component
• Updated WindowManager to support runtime hotkey changes
• Added proper drag state management with smooth 60fps window positioning
• Code Organization: Extracted functionality into dedicated managers and React hooks:
  • HotkeyManager, DragManager, AudioManager, MenuManager, DevServerManager
  • useAudioRecording, useWindowDrag, useHotkey React hooks
  • WindowConfig utility for centralized window configuration
  • Reduced WindowManager from 465 to 190 lines through composition pattern

[0.1.0] - 2024-XX-XX

Added

• Initial release of OpenWhispr (formerly OpenWispr)
• Desktop dictation application using OpenAI Whisper
• Local and cloud-based speech-to-text transcription
• Real-time audio recording and processing
• Automatic text pasting via accessibility features
• SQLite database for transcription history
• macOS tray icon integration
• Global hotkey support (backtick key)
• Control panel for settings and configuration
• Local Whisper model management
• OpenAI API integration
• Cross-platform support (macOS, Windows, Linux)

Features

• Speech-to-Text: Convert voice to text using OpenAI Whisper
• Dual Processing: Choose between local processing (private) or cloud processing (fast)
• Model Management: Download and manage local Whisper models (tiny, base, small, medium, large)
• Transcription History: View, copy, and delete past transcriptions
• Accessibility Integration: Automatic text pasting with proper permission handling
• API Key Management: Secure storage and management of OpenAI API keys
• Real-time UI: Live feedback during recording and processing
• Global Hotkey: Quick access via customizable keyboard shortcut
• Database Storage: Persistent storage of transcriptions with SQLite
• Permission Management: Streamlined macOS accessibility permission setup

Technical Stack

• Frontend: React 19, Vite, TailwindCSS, Shadcn/UI components
• Backend: Electron 36, Node.js
• Database: better-sqlite3 for local storage
• AI Processing: OpenAI Whisper (local and API)
• Build System: Electron Builder for cross-platform packaging

Security

• Local-first approach with optional cloud processing
• Secure API key storage and management
• Sandboxed renderer processes with context isolation
• Proper clipboard and accessibility permission handling