AI polishing of a recorded or pasted transcript was a generic rewrite: it cleaned fillers but left names and in-house terms mangled even when the app already knew them. Polishing now corrects dictionary pronunciations and typed identifiers deterministically before and after the model, and hands the model the learned term list plus the user's own expressions with explicit rules: rewrite a misheard name, product or in-house term to the learned spelling, and never swap an ordinary noun or ending on similarity alone. Dictation output is re-checked the same way after the model runs.
64 KiB
Changelog
All notable changes to D3RO Voice will be documented in this file.
The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.
[Unreleased]
Planned
- macOS / Linux support
- Additional Whisper model management UI
- Plugin system for custom pipelines
[1.16.4] - 2026-11-06
AI correction of transcripts is now grounded in the vocabulary the app learned from typing.
Changed
- AI polishing (recorded or pasted transcripts) treats learned vocabulary as the correction source: names, product names and in-house terms the transcript misheard are rewritten to the learned spelling, while ordinary nouns and endings are left alone. The learned term list and the user expressions are given to the model, and dictionary pronunciations plus typed identifiers are fixed deterministically before and after the model runs.
- Dictation output is checked against the learned vocabulary once more after the language model finishes, so a term the model rewrote is restored.
Fixed
- Hangul-layout typing artifacts (dkssudgktpdy) no longer appear in the term list, the correction hints, or the style examples handed to the model.
[1.16.3] - 2026-11-06
Supersedes 1.16.2, which could rewrite ordinary Korean words in transcripts.
Fixed
- Stop the learned-vocabulary pass from rewriting ordinary Korean. Measured against real meeting transcripts it changed 286 of 3,005 words (
하면→라면,이제→이게,매일→메일,정보→정도). Deterministic correction is now limited to your dictionary's pronunciations and the Latin identifiers you type; Korean terms reach recognition through the prompt hint and the polish prompt instead. Re-measured: 0 of 3,005 words changed. - Drop Hangul-layout typing artifacts (for example
dkssudgktpdy) from the recognition hint so they no longer bias transcription, and keep the original casing of learned identifiers (VSCode).
Changed
- Korean terms are no longer auto-replaced word by word. They are used as a recognition hint and as reference for the local model, which decides with sentence context.
[1.16.2] - 2026-11-05
Published from an annotated tag. Installer and update metadata are served by the canonical Forgejo feed; no binaries are committed to this repository.
Added
- Meeting and live-caption transcription now use the vocabulary the app learned from your typing: dictionary words, pronunciations and frequently typed terms bias speech recognition before it runs, fix mistyped words as soon as each line is finalized (no model required), and are given to the local model when it polishes a line.
- Polished caption lines now reach the meeting transcript and the meeting screen instead of stopping inside the caption service, so the saved transcript carries the corrected text.
Fixed
- Learned vocabulary used to feed typing suggestions only, which is why transcripts kept mangling names and jargon the app already knew.
[1.16.1] - 2026-11-04
Published from an annotated tag. Installer and update metadata are served by the canonical Forgejo feed; no binaries are committed to this repository.
Fixed
- prevent fast correction input races
[1.16.0] - 2026-11-01
Added
- Import existing audio or video recordings into a meeting, retaining timestamped transcripts and playable audio for minutes generation.
Fixed
- Cancel meeting imports when the account changes, preventing transcripts from being saved to another account.
- Deliver the updated local speech engine with CPU thread limits and faster terminal UIA handling. Existing engines upgrade before use.
Changed
- Release the required speech engine before publishing the desktop update, and verify the packaged engine starts successfully.
[1.15.4] - 2026-10-31
Published from an annotated tag. Installer and update metadata are served by the canonical Forgejo feed; no binaries are committed to this repository.
Changed
- Suggestion generation pipeline modularization (
SuggestionService).- Separates model-request construction, first-candidate retries, generation-session lifecycle, stale-result validation, failure fallback, and filled-candidate ranking/publication into named boundaries while preserving cancellation, timeout, personal-memory fallback, and candidate ordering behavior.
Fixed
- Desktop test reliability and offline release gate.
- Restores the 35 previously failing offline tests by preserving real STT exports in partial mocks, consolidating popup renderer recovery coverage, and aligning built-in instruction and synthetic Unicode input expectations with current behavior.
- Splits Vitest into explicit unit, service, integration, evaluation, and all scopes; the default desktop test gate now runs 2,764 deterministic offline tests with zero failures.
[1.15.3] - 2026-10-31
Published from an annotated tag. Installer and update metadata are served by the canonical Forgejo feed; no binaries are committed to this repository.
Fixed
Ctrl + Alt + Entersuggestion acceptance reliability (@d3ro/core/keybinding-runtime,KeyBindingService,SuggestionService,bootstrap).- Fixes
ChordStateMachine.keyUpso releasing a shared modifier (CtrlorRightAlt) releases all held chords in_isDownthat depend on that modifier instead of stopping after the first match (commandCtrl+RightAlt), preventingsuggestion-accept(Ctrl+Alt+Enter) from remaining stuck in_isDownand silently dropping subsequent presses as auto-repeat. - Maps
UiohookKey.NumpadEnterand numpad navigation keys inUIOHOOK_TO_VKand adds a 25 ms OSglobalShortcutfallback trigger inKeyBindingService._syncSuggestionCapturesoCtrl+Alt+Enterstill fires even ifuiohookmisses a key event. - Keeps OS suggestion shortcut capture active while the
"생성 중..."pending indicator or an accept is in flight (isOverlayOrCaptureActive), and whenCtrl+Alt+Enteris pressed while candidate #1 is still streaming (_inFlight/_generating/_pending), waits up to 2.5 s for candidate #1 to finish and immediately accepts it instead of ignoring the key press.
- Fixes
- Duplicate typo-model check elimination (
TypoCorrectionService.propose).- Skips re-running
_askModelinTypoCorrectionService.proposewhenhandleTypingContexthas already verified the same region (_isRegionChecked(region)), preventing a redundant 336-token LLM call during suggestion candidate filling from evicting Ollama's suggestion KV cache.
- Skips re-running
[1.15.2] - 2026-10-31
Published from an annotated tag. Installer and update metadata are served by the canonical Forgejo feed; no binaries are committed to this repository.
Fixed
- Modifier key (
Ctrl/Alt/Shift/Win) classification & terminal buffer preservation (@d3ro/core/input-intelligence,@d3ro/core/terminal-prompt,InputTelemetryService).- Fixes
classifyKeyStrokeso modifier virtual keycodes (0x10..0x12,0x5b..0x5c,0xa0..0xa5) are classified as'modifier'before checkingmodifiers.ctrl || modifiers.alt || modifiers.meta(whichuiohook-napisets totrueon the modifier keydown itself). Previously, pressing or holdingCtrl/Altwas misclassified as'shortcut'(key=shortcut+C+A), wiping outTerminalKeystrokeTracker's active line buffer (prefixLen=0), clearing_typedInCurrentFocus, and triggering repeated UIA snapshots on every key-repeat tick. - Ignores standalone modifier keydowns and D3RO Voice's own
Ctrl+Altsuggestion navigation/accept chords (Ctrl+Alt+Up/Down/Left/Right/Enter/Space/Backspace) inInputTelemetryService._handleKeyDownandTerminalKeystrokeTracker.onKeyDownwithout advancing_inputRevision/_lastKeyAtor resetting the tracked terminal line.
- Fixes
- Terminal suggestion acceptance without clipboard race (
SuggestionService._insertAccepted).- Uses direct Unicode keyboard injection (
insertMethod = 'keyboard'viaSendInput(KEYEVENTF_UNICODE)) instead of clipboardCtrl+Vwhen accepting a suggestion in terminal mode (terminalMode === trueorisTerminalApp(appName), includingxterm-helper-textarea,WindowsTerminal, andKitedesk), preventing delayed terminal paste processing from pasting the restored previous clipboard content (such as copied URLs) and synchronizingTerminalKeystrokeTrackerviaapplyProgrammaticReplacement. - Stops the background candidate fill loop (
_fillRemainingCandidates) immediately whenSuggestionService.accept()is triggered so new candidates do not continue streaming while waiting for modifier release.
- Uses direct Unicode keyboard injection (
- UIA snapshot timer optimization (
InputTelemetryService._scheduleTextSnapshot,_readTextSnapshot).- Skips scheduling the 780 ms
_settleTimerUIA snapshot on non-typing keys (Enter,Arrow,Shortcut,Navigation,Function,Tab) and automatically cancels any pending_settleTimeronce the 380 ms_textTimersnapshot determines the field cannot trigger a suggestion (prefixLen < minPrefixChars, non-editable, or password), cutting redundant UIA bridge calls in half during everyday typing and navigation.
- Skips scheduling the 780 ms
[1.15.1] - 2026-10-31
Published from an annotated tag. Installer and update metadata are served by the canonical Forgejo feed; no binaries are committed to this repository.
Fixed
- Terminal &
xterm.jsKorean IME character-swallowing fix (터미널·에이전트 창 한글 타이핑 시 1글자씩 씹히는 현상 근본 해결).- Removes synchronous cross-process
isForegroundKoreanIme(SendMessageTimeoutW(WM_IME_CONTROL)) calls from the low-level keydown hook (InputTelemetryService._handleKeyDown) that were disrupting Windows TSF/DefIME composition state on every keystroke in terminal windows, replacing them with a one-time idle-snapshot IME sync (TerminalKeystrokeTracker.syncKoreanIme). - Adds a dedicated
xterm-helper-textareafast path insidecar/uia_bridge.pythat skipsGetTopLevelControl()DOM tree walks andTextPattern.GetSelection()/ExpandToEnclosingUnit(Character)/TextEditPattern.GetActiveComposition()COM calls (which stalled Chromium's renderer thread up to 1,500 ms during Korean IME composition and leftisComposingstucktrue), and extends_WAKE_TTL_SECONDSfrom 60 s to 86,400 s per HWND. - Suppresses 40 ms word-boundary UIA snapshots on normal Korean words (
_shouldRunFastTerminalBoundarySnapshot), removes the duplicate 680 ms_typoTimer, and stops 800 ms_sampleWhileTypingpolling on terminal surfaces so UIA never queriesxterm-helper-textareawhile you press the first consonant of the next word. - Prevents
TypoCorrectionServicefrom injectingcommitComposition(VK_SPACE+VK_BACK) on trailing composed Hangul syllables (/[가-힣]$/u) in terminals beforeSpaceis pressed.
- Removes synchronous cross-process
[1.15.0] - 2026-10-31
Published from an annotated tag. Installer and update metadata are served by the canonical Forgejo feed; no binaries are committed to this repository.
Added
- Multi-command ordered pipeline selection & reordering in Commands (
CommandsPage.tsx,@d3ro/core/instruction-template,VoiceModeService).- Replaces single-select radio buttons in the Commands screen with an ordered multi-select pipeline: check multiple commands in sequence (e.g.,
1단계: 구어·추임새·중복 정리→2단계: 격식체 변환or1단계: 구어·추임새·중복 정리→2단계: 텍스트 요약→3단계: 번역) and reorder steps anytime with↑/↓controls. - Synthesizes selected pipeline steps into a single ordered pipeline prompt (
composeInstructionPipelinePrompt) where later steps determine the final output tone/format/language without adding multi-roundtrip LLM latency, and supports testing the full multi-step pipeline directly in the Commands test runner.
- Replaces single-select radio buttons in the Commands screen with an ordered multi-select pipeline: check multiple commands in sequence (e.g.,
[1.14.0] - 2026-10-29
Published from an annotated tag. Installer and update metadata are served by the canonical Forgejo feed; no binaries are committed to this repository.
Added
- First-word-after-focus (
pc/ㅔㅊ/ㅖㅊ/내ㅔㅊ) instant Han/Eng conversion (검색창·새 입력칸 포커스 직후 첫 단어 및PC계열 한영 자동 전환).- Fixes the race where clicking into a search bar or switching windows and immediately typing
pc(ㅔㅊ/ㅖㅊ) before the first UIA snapshot absorbed the typed word into_textAtFocus(editedSinceFocus = false), blocking auto-correction. - Expands 2-letter loose-jamo acronyms (
ㅔㅊ/ㅖㅊ→PC,ㅐㄴ/ㅒㄴ→OS,ㅔㄱ/ㅖㄱ→PR,ㅊㅑ→CI,ㅍㅡ→VM,ㅔㅡ/ㅖㅡ→PM,ㅔㅐ/ㅖㅐ→PO,ㅗㄱ→HR,ㅑㄱ→IR,ㅇㅡ→DM,ㅡㅇ→MD,ㅔㅇ/ㅖㅇ→PD,ㅍㅔ/ㅍㅖ→VP,ㅠㅡ→BM,ㅆㅍ→TV,ㅑㅖ→IP) and Korean-prefix + loose-jamo English compounds (내ㅔㅊ→내PC,회사ㅔㅊ에서→회사PC에서,한글로ㅔㅊ라고→한글로PC라고,윈도우ㅐㄴ→윈도우OS,새ㅁㅔㅑ를→새API를).
- Fixes the race where clicking into a search bar or switching windows and immediately typing
- Speech disfluency, self-correction, redundancy & STT homophone post-processing (
@d3ro/core/dictation-cleaner&builtin-clean-speech).- Adds deterministic zero-latency dictation cleanup (
cleanDictatedText) right after STT transcription to remove standalone spoken fillers (어,음,그 쓸데없는— while preserving quoted mentions like어라든가,아라든가and demonstratives), mid-sentence self-corrections (검색 아니 검색이 아니라 타이핑을→타이핑을), word stutters (내가 내가→내가), redundant expressions (쓸데없는 의미없는→불필요한,정리해 가지고 해 주는 거→정리해 주는 기능), colloquial suffixes (예를 들자면은→예를 들어,보면은→보면), and clear contextual STT homophones (스시가 영어로 변경되지→PC가 영어로 변경되지,그 길을 쓸 이유가→그 기능을 쓸 이유가). - Upgrades
refinepost-processing (llm-prompts.ts), meeting transcript auto-polish (MeetingModeService.polishTranscript), and adds the built-in command preset구어·추임새·중복 정리(builtin-clean-speech, voice trigger정리해줘/clean this up) so spoken transcripts are cleaned up while strictly preserving the speaker's original tone (반말 / 해요체 / 합쇼체).
- Adds deterministic zero-latency dictation cleanup (
[1.13.0] - 2026-10-29
Published from an annotated tag. Installer and update metadata are served by the canonical Forgejo feed; no binaries are committed to this repository.
Added
- Graph-relational extreme personalization & sentence suggestion upgrades (그래프 관계성 기반 극한 개인화 및 문장 제시 고도화).
- Chains consecutive short messages within the same app (
_lastIndexedByScope, 10-minute session window) viafollowsedges and restores post-send chat context fromrecentWriteswhen Enter clears the input box. - Adds particle-invariant stemmed continuations (
stemContinuationsFrom) so phrases like"회의록 정리해서 내일 오전까지 공유드리겠습니다"continue seamlessly when you type"오늘 회의록을 정리해서", alongside 2-hop graph path scoring (rankTwoHopGraphTargetsacrossshares_terms → followsandfollows → follows). - Weights frequent
personalPhrases(count >= 2) and adds a Korean stemmed-word bigram fallback (_stemBigram) inPersonalNgramModel, while reranking generated LLM candidates byscoreCandidatePersonalization(graph overlap + Korean register match) without reducing suggestion frequency.
- Chains consecutive short messages within the same app (
- Stronger English detection via derivational morphology & phrase lookahead (파생 형태론 및 전방 구문 호응 기반 영어 감지 강화).
- Expands English stem lexicon (
EXTENDED_ENGLISH_STEMS:kubernetes,microsoft,supabase,postgres,typescript,refactor,endpoint,workflow,dashboard, etc.) with productive derivational prefixes (un-,re-,pre-,non-,sub-,over-,under-,auto-,multi-,micro-,inter-) and suffixes (-tion,-sion,-ation,-ment,-ness,-able,-ible,-ful,-less,-ive,-ous,-al,-ic,-ize,-or,-ist,-ism). - Detects clause-initial and post-Korean Hangul-typed English function words via lookahead (
"새 ㅠㄷ "→"to be ","새 솓 "→"to the ","새 애 this "→"to do this ","랙 솓 "→"for the ","뭉 솓 "→"and the ") while keeping Korean determiners intact ("오늘 새 API를 만들었어 ").
- Expands English stem lexicon (
Changed
- Zero-sleep single-syscall atomic replacement & in-flight delta rebasing (단일
SendInput원자적 치환 및 타이핑 중 델타 리베이싱).- Packs
(optional VK_SPACE) + N × VK_BACK + M × KEYEVENTF_UNICODEinto a single WindowsSendInputcall (sendWin32UnicodeReplacement,< 0.5 ms,0 mssleep) so physical keystrokes never interleave with auto-correction. - Rebases replacement plans (
rebaseTypoReplacementPlan) when you keep typing during the 40 ms UIA verification or background LLM check, preserving newly typed suffix characters (<= 12chars) in one atomic swap.
- Packs
[1.12.0] - 2026-10-28
Published from an annotated tag. Installer and update metadata are served by the canonical Forgejo feed; no binaries are committed to this repository.
Added
- Hangul-typed English contractions, acronyms, function words & fuzzy typo recovery (한영 축약형·약어·오타 자동 교정 고도화).
- Converts Hangul-typed English contractions and contraction+particle tokens (
ㅊ무ㅅ/ㅊ무'ㅅ/ㅊ무ㅜ'ㅅ→can't,ㅊ무'ㅅ는→can't는,애ㅜ'ㅅ→don't,ㅑ'ㅡ/ㅑㅡ→I'm,재ㅜ'ㅅ→won't), standalone sentence-initialㅑ→I, and English-clause function words (새 애→to do,뭉→and,랙→for,소됴→they,생묘→today,갸홋→right). - General Araso batchim-harmony and rare-stem acronym detection (
츤를/츤는→CMS를/CMS는,눈를→SNS를,운를→DNS를,긴를→RLS를,틔을→XML을,먀를→AI를,뗘를→EU를,쓰은→TM은,꾸은→RN은,띠이→EL이,꺼을→RJ을,ㅏㅅㅌ를→KTX를,ㅎㅅㅌ를→GTX를). - Recovers 1-edit Hangul-typed English typos (
ㅗ디ㅐ→hello,ㅔㅐㄴ냐ㅠㅣ→possible,세갸ㅜㅎ→spring,ㅡㅁ가애주→markdown), trailing bracket mistypes (help[→help), and split reflexive pronouns (my self→myself).
- Converts Hangul-typed English contractions and contraction+particle tokens (
Changed
- Cursor-tail-only replacement without racing your typing (타이핑 경합 없는 꼬리 구간 즉시 교정).
- Auto-correction now replaces only the contiguous token chain at the cursor tail (up to 22 characters) instead of backspacing and retyping entire lines, and skips background LLM typo calls during active typing bursts so your keystrokes are never interrupted.
- Design System v4 ("조용한 계기") across Desktop, Web, Admin, and Landing site.
- Pretendard Variable + JetBrains Mono typography,
#08090Cmatte obsidian surfaces,#E04F3Esignal-red recording accent, and WCAG AA contrast tokens across all surfaces.
- Pretendard Variable + JetBrains Mono typography,
[1.11.0] - 2026-10-27
Published from an annotated tag. Installer and update metadata are served by the canonical Forgejo feed; no binaries are committed to this repository.
Added
- Built-in instant Korean/English keyboard layout auto-switching (한영 자판 자동 전환 기본 탑재).
Words typed while the IME is in the wrong layout are converted in place as soon as you
press Space or pause — even when input telemetry collection and sentence suggestions
are turned off:
- English keys → Korean (
dkssudgktpdy→안녕하세요,rkatkgkqslek→감사합니다). - Korean jamo → English (
ㅗ디ㅣㅐ 재깅→hello world,내ㅡㄷ샤ㅡㄷ→sometime,내ㅡㄷ유ㅐ요→somebody,ㅑㄹ ㅑ 쇼ㅔㄷ ㅗ디ㅣㅐ→if I type hello). - Mixed English/Acronym/CamelCase + Korean particles & compound nouns (
githubdp→github에,ㅎ샤ㅗㅕㅠ에→github에,ㅠㅕㅎ가→bug가,PDFvkdlf/ㅖㅇㄹ파일→PDF파일,Reactzjavhsjsxm→React컴포넌트,dhsmfdmsgoogledptj→오늘은google에서,1:1ansdml→1:1문의, batchim-aware 2-letter acronymsTMdms→TM은vsTMsms→쓰는, single-letter + particleㄴ가→s가,ㄴ은→s은). - Lingering Shift on double consonants (
djEJs→어떤,qkRNjTdj→바꿨어) and sentence-context Hangul conversion (Stay hungry, stay foolishsms wkqtmdml akfdlek.→Stay hungry, stay foolish는 잡스의 말이다.).
- English keys → Korean (
- Terminal, Vim / Neovim, and xterm.js support (터미널·Vim·xterm.js 한영 전환 지원).
- Converts wrong-layout shell commands (
퍄ㅡ→vim,ㅜ퍄ㅡ→nvim,ㅎ샤→git,ㅣㄴ→ls,채ㅜㄴㅅ→const) and Vim ex-commands (:ㅈㅂ→:wq,:ㅈㅂ!→:wq!,:ㅂ!→:q!,:ㅈ→:w,:ㅂ→:q) across Windows Terminal, PowerShell, cmd, Neovide, Kitedesk, WezTerm, Alacritty, and xterm.js terminals (Agent Switchboard, VS Code, Cursor). - Suppresses auto-correction in Vim Normal mode (
vimMode === 'normal') so motions (hjkl,dd,ciw) are never altered, and tracks Korean IME conversion state on Windows TSF (IMC_GETCONVERSIONMODE).
- Converts wrong-layout shell commands (
- Word-by-word suggestion acceptance & instant personal n-gram candidate.
- Press
Ctrl+Alt+Spaceto insert only the first word of the active suggestion and keep the remainder visible.
- Press
Changed
- Direct Unicode typing (
SendInputKEYEVENTF_UNICODE) instead of clipboard paste. Typo replacements and text insertion on Windows now type Unicode characters directly without touching yourWin+Vclipboard history or colliding with background clipboard locks (ClipboardSvcGroup). - No mid-word interruption while typing. Boundary keys (Space, punctuation) trigger
layout conversion immediately (
25–40 ms), while mid-word pauses wait650 msso compound words (some→sometime/somebody) are never split or swallowed.
Fixed
- Characters are no longer dropped, duplicated, or preceded by phantom spaces during fast typing.
- Synthetic
KEYEVENTF_UNICODEkey events (keycode=0) are ignored by the global input hook, and programmatic key suppression only matches syntheticBackspace/Space/Ctrl+Vso physical letter keys pressed during async replacement are never swallowed. commitComposition(SpacebeforeBackspace) is sent only when the character before the caret is a composed Hangul syllable ([가-힣]), never on a single compatibility jamo (ㅎinㅠㅕㅎ,ㄷin내ㅡㄷ,ㅂin:ㅈㅂ).TerminalKeystrokeTracker.applyReplacementpreserves keys typed after replacement started (baseLength) and reconcilesxterm-helper-textareabuffers without re-correcting previously replaced tokens.
- Synthetic
[1.10.1] - 2026-10-24
Urgent fix release. Installs older than 1.10.1 receive this update automatically.
Fixed
- Accepting a suggestion with Ctrl+Alt+Enter now actually inserts it. When the keyboard hook missed a key release, the paste went out while Ctrl/Alt still counted as held and nothing was inserted. D3RO now asks Windows for the real key state.
- The suggestion popup no longer reappears while an accepted suggestion is being inserted.
- Typing is learned as whole sentences instead of fragments. Text was saved every fraction of a second, often cut mid-word. It is now saved when you pause, finish a sentence or send the message. Very short pieces, links, half-typed Hangul and broken emoji are no longer learned.
- Many stability fixes across recording, meetings, sync, file transcription, popups and updates.
[1.10.0] - 2026-10-18
Published from an annotated tag. Installer and update metadata are served by the canonical Forgejo feed; no binaries are committed to this repository.
Added
- Typo auto-correct and sentence prediction while you type. D3RO suggests how your sentence continues based on the surrounding document or conversation. With suggestions on, typo fixes appear in the same popup and are applied only when you accept the suggestion.
- A redesigned app. Home, records and meetings come first, and the whole interface uses a calmer, flatter look.
- Unified search across your records, meetings and memos.
- Esc cancels dictation, and the dictation popups now show what is really happening.
- Updates are verified before they install. The update file is checked against a signature, so a tampered download is refused.
Changed
- Suggestions appear right after you pause instead of mid-typing, stay put while you pick one, and also work in terminals.
- Pasting is faster: the 300 ms key delay is gone and every clipboard format is restored after a paste.
- Local models are no longer pinned in memory, so the app uses less RAM when idle.
Fixed
- Speech recognition falls back to the CPU when the GPU runtime cannot be used.
- File transcription no longer fails on clean installs because ffmpeg was missing.
- Clicking the recording tip cancels the recording; the Sound switch applies at once.
- Meeting transcript edits are kept, and the live recording view comes back after switching screens.
- Clearing all history now asks for confirmation.
- Keyboard shortcuts no longer steal key combinations from other apps on Windows.
- Offline license keys survive restarts and cloud sign-out.
- Many sync, quota, billing and security fixes from a third security review.
[1.9.0] - 2026-09-27
Published from an annotated tag. Installer and update metadata are served by the canonical Forgejo feed; no binaries are committed to this repository.
Added
- Preset commands you edit on the desktop are used on the phone. Changing the prompt of Translate, Summarize, Professional rewrite or Explain code updates the same command on your phone; restoring the default restores it there too.
Fixed
- Meeting transcripts edited on the desktop now show on the phone. Edits, auto-polish and speaker labels were kept on the desktop only because the phone shows a meeting from its transcript segments; those are now updated with every change.
[1.8.0] - 2026-09-27
Published from an annotated tag. Installer and update metadata are served by the canonical Forgejo feed; no binaries are committed to this repository.
Added
- Knowledge base syncs with your phone. Documents added on the desktop can be searched on the phone and web, and documents added there appear on the desktop and are indexed with the local model. Deleting one removes it everywhere.
- Recordings sync and play anywhere. Desktop recordings are uploaded (up to 50 MB each, can be turned off in Settings > Cloud) so the phone can play them, and each history card has a play button that also plays recordings made on the phone.
- Shared settings. Language (Korean/English), light/dark/system theme, auto-polish on or off and the active command follow the phone.
Fixed
- A document's text is kept even when the local embedding model is unavailable, so it can be indexed later without the original file.
[1.7.0] - 2026-09-27
Published from an annotated tag. Installer and update metadata are served by the canonical Forgejo feed; no binaries are committed to this repository.
Added
- Your desktop and phone now stay in step. Signing in keeps history (with titles and favorites), the dictionary, meetings with their memos and documents, memo tags, your own commands and dictation/meeting templates the same on the desktop, the phone and the web. Changes made offline wait and upload later; what you delete on one device is deleted on the others; edits from the phone appear within seconds.
- The desktop appears in the phone's device list. Disconnecting it from the phone signs the desktop out; signing out on the desktop removes it from the list.
- Records made before signing in come with you. The first account that signs in on this computer receives the history, dictionary and meetings kept in local mode.
- Favorites on the desktop, shared with the phone.
- Settings > Cloud shows how many changes are waiting to upload or were refused.
- Live captions stream as you listen. Words appear about a second after they are spoken and keep updating while the sentence is in progress; the part that is no longer changing is shown solid and the rest dimmed. A pause finishes the line, and long unbroken speech is split at natural boundaries instead of waiting six seconds.
- Captions are polished in context. Each finished line is corrected by the local model using the lines before it (spacing, punctuation, misheard words); rewrites that change too much are ignored. It can be turned off in Settings.
- Live captions can use their own speech model. Settings > STT now has a separate live-caption model, loaded next to the dictation model, so captions can run on a faster or more accurate model without changing dictation. The local speech engine is updated for this and is downloaded again once.
Changed
- Suggestions page with Up/Down only. Moving past the third candidate turns the page by itself, and at the last candidate the selection waits while more are still being generated. The separate Ctrl+Alt+Left/Right page shortcuts are gone; they also collided with the display-rotation shortcut of Intel graphics drivers.
- The live-caption model is chosen next to the other caption settings in the General tab.
Fixed
- Accepting a suggestion with Ctrl+Alt+Enter did not paste. The paste was sent while Ctrl+Alt were still held, so the app received Ctrl+Alt+V. The panel now closes at once and the text is pasted as soon as the keys are released.
- Clicking a suggestion did nothing when a new candidate arrived between pressing and releasing the mouse button; a candidate is now accepted on press.
- Suggestions rarely appeared in chat boxes that grow as you type. The growing box was taken for a different field on every keystroke, which kept blocking suggestions as "not typing". Fields are now told apart by window and control, and a mouse click re-checks whether you have typed since.
- Cloud sync was mostly one-way. Data created on the phone never reached the desktop after signing in, only the first meeting's memos and documents were uploaded, deletions were never shared, team meetings leaked into personal data, and live updates never connected. All fixed.
- Exporting account data failed for accounts with meetings; the export keeps its original format again.
[1.6.0] - 2026-09-24
Published from an annotated tag. Installer and update metadata are served by the canonical Forgejo feed; no binaries are committed to this repository.
Added
- Up to 12 next-sentence suggestions, three per page. The first candidate appears
as soon as it is ready and the rest are generated one at a time in the background.
Ctrl+Alt+Up/Downmove between candidates,Ctrl+Alt+Left/Rightflip pages,Ctrl+Alt+Enteraccepts, andEsccloses the panel while it is showing. Accepting, typing on, or closing stops generation. Existing shortcuts that were still on the 1.5.0 defaults move to this layout automatically; customised ones are kept. - A key guide inside the panel built from your current shortcuts, so it follows whatever you rebind.
- The live caption window can be moved. Hover it to reveal a small handle, drag it anywhere, and it reopens there next time; double-click the handle to return it to the bottom of the screen. The rest of the window still lets clicks through.
- Live captions say they are getting ready from the moment you start them until the first line arrives, instead of showing an empty screen for several seconds.
Changed
- The suggestion panel is redesigned: numbered candidates with a clear selection bar, two-line items that no longer break Korean words mid-word, no empty header row, and a two-line footer for progress and keys. The panel stays where it first appeared while candidates stream in, and sits outside the input box when the app cannot report a caret position.
- The suggestion model stays loaded for 10 minutes and warms up on launch, and on Ollama reconnect, so the first suggestion no longer times out on a cold reload.
- The personal phrase memory only learns everyday writing. Terminals, code editors and the Agent Switchboard coding hub are excluded, and lines that are mostly symbols or an empty field's placeholder text are ignored. Phrases learned under the old rules that break these rules are removed on the next launch.
- Suggestions are never offered in terminals.
Fixed
- Suggestion shortcuts started dictation. The voice service treated every shortcut it did not know as dictation, so the new suggestion keys switched on recording.
- Suggestions never appeared after updating to 1.5.0 on installs whose local speech engine was downloaded by an earlier version: the old engine lacked the focus-reading endpoint. An engine older than the app requires is now downloaded again automatically.
- Suggestions were cancelled while typing Korean because the syllable still being composed changed the text; cancelled requests also no longer use up the rate limit.
- The panel appeared after a mouse click without typing, could not be closed with X while generating, and showed a spinner next to finished candidates.
- Model runner windows flashed while typing when D3RO had started Ollama itself. It now starts Ollama through Ollama's own tray app.
- The speech engine could start twice when two parts of the app needed it at the same moment.
[1.5.0] - 2026-09-23
Published from an annotated tag. Installer and update metadata are served by the canonical Forgejo feed; no binaries are committed to this repository.
Added
- Next-sentence suggestions while you type. With the feature turned on, a small
ghost-text panel sits next to the caret in whatever app has focus and offers up to
three ways to continue the sentence.
Ctrl+Alt+Rightaccepts,Down/Upmove between candidates,Leftor the panel's X dismisses, and all four are ordinary rebindable shortcuts. The panel never takes focus and is click-through until you turn interaction on. - Typed-text learning from the field you are in, read through UI Automation and diffed as a snapshot, so committed Korean and Japanese IME text is counted correctly (key codes cannot reconstruct it). Password fields are refused before any read, and IME composition suppresses learning and suggestions until the character is committed.
- Weekly input insights (
Settings > Inputand a dashboard card): keystroke, click and scroll totals, mouse travel converted to meters, top hours and apps, flow windows, edit friction (edits per 100 characters) and per-app suggestion quality. - A personal phrase memory built from your own typed sentences and voice history, ranked by frequency, recency (30-day half-life) and same-app usage, and offered as prompt hints. Phrases can be deleted individually or all at once.
- A privacy receipt that reports the real number of retained rows and the 30-day retention rule, plus smart exclusion: only the app you are in can be suggested for the exclusion list, and only after repeated unreadable fields, never automatically.
- A shortcut safety audit in Settings that flags invalid or conflicting bindings across every action.
- A local personal graph that links your sentences by "follows" and shared-terms relationships, so related context can surface even when the current prefix differs. It is local SQLite only, with no embeddings or network.
Changed
- Local model requests are now bounded and individually cancellable.
LocalLLMServicegives every request its own abort controller, relays external cancellation, requires adoneframe before a stream counts as successful, and clears incomplete streams. Generation and streaming are capped at 2048 tokens / 120 seconds and chat at 512 / 60 seconds. Text suggestions usekeep_alive: 2mand no longer warm the model at startup. - Voice conversations are single-flight and pass their cancellation through to the active local chat request, so ending a conversation aborts its own inference instead of only changing what the UI shows.
- Console windows are suppressed for the app's Windows helper processes (text to speech, voice actions, audio-device and active-window enumeration, and every ffmpeg path).
Fixed
- Custom instructions ran the instruction instead of your text. Every built-in
instruction (translate, summarize, rephrase, explain code, free prompt) inserted the
instruction's own wording rather than the processed result, and only
{{text}}was ever substituted, which none of the presets use. The instruction is now the system prompt, the transcript is the text, and{{text}},{{userPrompt}}and{{targetLanguage}}are resolved in one shared place. This path had never worked in any shipped release. - The suggestion panel could refuse to close, revive itself, or appear on a bare
click. Holding the modifier that Windows turns into
Ctrl+Altno longer fires the voice shortcut underneath, closing the panel during generation discards that request instead of re-showing it, a click without typing no longer triggers a suggestion, and pressing Enter stops the same text from being suggested again. - The global input hook is reference-counted, so the shortcut service and input telemetry can attach at the same time without one disabling the other.
[1.4.0] - 2026-09-21
Added
- Several shortcuts per action. Every action now holds a list of bindings
instead of a single one, so a second (or third) shortcut can be added with the
+button and removed individually. - Mouse buttons as shortcuts. Side buttons (back / forward) can be bound on their own; right and middle click require a modifier. Left click cannot be bound because it drives every on-screen interaction. Mouse events cannot be swallowed, so a bound button still performs its original action — the settings screen says so before the binding is saved.
- Pick a key from a searchable list. Besides pressing the key, a shortcut can now be chosen from a grouped dropdown with a search box inside it, which also makes keys reachable that are awkward to press while recording.
- The history and command popups are rebindable.
Ctrl+Shift+VandCtrl+Shift+Cwere fixed in code and had no settings entry; they are now ordinary actions. The voice command shortcut, which had never been exposed in settings, is visible as well.
Fixed
- Double-press hands-free never triggered. Dictation and hands-free ship with the same default key and are told apart by the double press, but shortcut lookup stopped at the first match, so hands-free was unreachable in installed builds. Both actions now receive the event and are split by press type.
- Reserved system shortcuts could be assigned anyway. The check compared
joined key names, so pressing the modifiers in a different order let
Ctrl+Cand friends through. Combinations are now compared structurally. - Turning shortcuts off also killed the popup shortcuts permanently. Disabling global shortcuts released every registration in the process, including ones it did not own, and re-enabling never restored them.
- Enabling shortcuts after starting with them off left nothing bound.
- Shortcut labels were frozen in the language they were recorded in and did not follow the app language or platform conventions. They are now derived on display.
- The dashboard caption indicator never reflected the real state because the event payload was stored instead of the state inside it.
Changed
- Shortcuts are stored as one list per action. Existing shortcuts are migrated automatically on first launch; no reconfiguration is needed.
[1.3.7] - 2026-09-19
Fixed
- Recording and live caption overlays never worked in installed builds. Popup
pages loaded their scripts as classic
<script src>tags, which the renderer build does not bundle, so a packaged app rendered only the static markup: the recording tip froze at 0:00 with no wave bars and captions showed nothing. Popup scripts are now module scripts, a packaging check fails when a renderer page references an asset that was never emitted, and popups hold IPC until their renderer is ready and re-assert visibility on every show.
[1.3.6] - 2026-09-18
Fixed
- Installing the local speech engine always failed with a hash mismatch, leaving local transcription unusable. The downloaded parts were checked against bytes counted from the network stream, while the joined archive was checked against what was actually written to disk, so a truncated write passed part verification and only surfaced later as an archive mismatch with no usable diagnostic. Every check now reads the file on disk, the joined archive is size-checked before hashing, mismatch errors report the actual and expected values, and a failed part is discarded and retried up to three times.
[1.3.5] - 2026-09-18
Fixed
- The Ollama guide still listed the retired models as its recommendation while the app already defaults to . The guide now leads with the model the app actually uses and lists current lightweight alternatives.
[1.3.4] - 2026-09-18
Fixed
- Automatic updates could not start on 1.3.3: its installer was missing the updater configuration file, because the packaging path used to guarantee the native module build does not create it. The file is now written from the single feed source and its presence in the packaged app is verified before publishing, so an installer that cannot update is never shipped.
[1.3.3] - 2026-09-18
Fixed
- The released installer could not start: it carried a
better-sqlite3build for the host Node runtime instead of Electron, so the app failed immediately with aNODE_MODULE_VERSIONmismatch when opening its database. Packaging now guarantees and verifies the Electron build of every runtime-sensitive native module before an installer or archive is produced, and publishes only from that verified tree. - The local language model default still pointed at a retired model: the stored default
was a
*.gguffile name that Ollama cannot serve, and the settings, onboarding, and Ollama guide still recommendedgemma2:2bwhile the service code already preferredgemma4:e4b. All surfaces now default togemma4:e4b.
[1.3.2] - 2026-09-18
Changed
- Auto-update works again: the installer no longer carries the local speech engine and ffmpeg. Bundling them pushed the installer to 189 MB, and the download feed sits behind Cloudflare, which rejects uploads over ~100 MiB with HTTP 413 — so update metadata could not be published at all. The installer is now 90.5 MiB and the update feed is published again.
- The app fetches its speech engine and ffmpeg when they are first needed, verifying every part and the joined archive by SHA-256 before installing them under the app data folder. Updates stay small (the engine is not re-downloaded on every release), and the Settings > STT tab shows the runtime status with a manual download action.
[1.3.1] - 2026-09-18
Fixed
- Manual install no longer needs 7-Zip: the signing-free channel now also publishes the
app as byte-split
zipparts. The install script joins them and extracts with the built-in WindowsExpand-Archive, so a user with nothing but Windows can install (the 7z volumes remain the smaller Scoop path). Split volumes from a different build are never mixed: every artifact of a release comes from one build, and a published version is not overwritten.
[1.3.0] - 2026-09-18
Published from an annotated tag through CI. Installer and update metadata are served by the canonical Forgejo feed; no binaries are committed to this repository.
Added
- Live partial transcript while dictating: while the hotkey is held, the app now transcribes the recent window of audio every 1.5 seconds and shows it in the recording tip, so the text can be seen forming before the key is released. Partials never reach the clipboard or the result popup; only the final transcription is inserted.
- Instant first dictation: the local speech engine (sidecar process + Whisper model) is warmed up in the background at app start, so the first press does not wait for the model to load.
- Packaged local speech engine: desktop installers now ship the faster-whisper
sidecar (
sidecar.exeplus runtime data, including the Silero VAD model) and ffmpeg, so local transcription works on a fresh install without Python on the machine. - Sidecar build tooling:
npm --prefix apps/desktop run sidecar:setupandsidecar:build, plus a packaging-time bundle verifier that fails the build when the engine or its VAD data is missing.
Added
- Signing-free install path (
portablechannel): the canonical feed sits behind Cloudflare, which rejects any upload body over ~100 MiB with HTTP 413 (measured: 60 MiB accepted, 110 MiB rejected). The sidecar-carrying app exceeds that, so releases are now also published as 95 MiB 7z split volumes (688 MB app → 162 MiB) with a Scoop bucket manifest and a verifiable manual installer script. This channel never touches the auto-update feed and needs no Authenticode certificate, so users can install while the signing certificate is still being procured.
Changed
- Local engine connections now target the IPv4 loopback (
127.0.0.1) instead oflocalhost. On machines wherelocalhostresolves only to IPv6, every local request (Ollama and the speech sidecar) was refused and local AI silently did nothing. - Whisper decoding is tuned for dictation: previous-text conditioning is disabled so repeated hallucinations cannot compound, silence is trimmed more aggressively, and low-confidence fallbacks are bounded. Same transcript quality, roughly five times faster on the same machine.
- The sidecar reuses an already-loaded model instead of reloading it, and reports the load time it measured.
- Local engine logs stream to the app log as complete UTF-8 lines instead of mangled fragments, so failures are diagnosable.
- Failed local-engine startup now fails immediately with an actionable message (missing bundled engine, damaged virtualenv, or missing SoX) instead of waiting for a 30 second health check and reporting a generic error.
Fixed
- Local transcription never worked in packaged builds: the sidecar was not part of the packaged resources and no pipeline step built it, so the app always fell back to a system Python that had no faster-whisper installed.
- Local transcription and local LLM never worked in development: the sidecar and SoX
paths were resolved against the Vite output directory (
out/main) instead of the app root, so recording failed with a SoXENOENTand the sidecar fell back to a Python without the runtime dependencies. - Silero VAD data was missing from the bundled engine, which would have made silence-trimmed transcription fail at runtime even with the engine bundled.
- Audio capture and the speech sidecar no longer flash a console window on Windows, and
SoX/spawn failures name the fix (
npm --prefix apps/desktop run setup:sox).
[1.2.0] - 2026-09-16
Published from an annotated tag through CI. Installer and update metadata are served by the canonical Forgejo feed; no binaries are committed to this repository.
Added
- Canonical desktop release channel: desktop auto-update now reads one Forgejo Generic Package Registry feed, published by a version-agnostic publisher (
scripts/ci/publish-forgejo-release.mjs) from tag-triggered Forgejo Actions, GitLab, and GitHub pipelines. The GitLab registry stays a legacy mirror for pre-Forgejo installs. - Update policy SSOT (
release/update-policy.json): channels (latest/beta/alpha), minimum supported version, forced install, full-installer thresholds, staged rollout percentage, and a remote kill switch, enforced at runtime byapps/desktop/src/main/update-policy.ts. - Dictionary import/export on desktop and web: round-trip import with per-entry conflict reporting, a web dictionary client, and an expanded knowledge add form.
- Multi-transport push delivery: Web Push (VAPID) and token-based Apple Push (APNs) transports join Firebase Cloud Messaging, with a Cloudflare Worker cron drain over an outbox table.
- Team activity feed: team activity events, migration, and the web feed component.
- Shared entitlement gating in
@d3ro/corefor free/paid feature boundaries. - Infrastructure map (
docs/map/) documenting the infrastructure and feature status per platform, with an update protocol so feature work and the map stay in step. - 21 unit tests for update policy decisions and feed helpers, plus tests for dictionary I/O, entitlements, push drain, Web Push, and APNs payloads.
Changed
- Desktop runtime updater feed moved from GitLab project 1172 to the canonical Forgejo registry;
electron-builder.yml, the metadata verifier, and both CI publishers now enforce the canonical/mirror split. - Release metadata verifier self-test expanded to negative cases covering the feed contract, policy schema, and Forgejo publisher invariants.
- The canonical publisher refuses to re-publish a version whose registry assets already hold different bytes, so a same-version re-release fails closed instead of overwriting a shipped installer.
- Landing site and web console download centers now link the canonical Forgejo feed instead of repository-local installer paths, which are not part of any deploy artifact.
- Admin console data views (models, pipelines, users, subscriptions, audit log) read live back-office data, with a unified sidebar and console theme.
- Desktop settings, ad surfaces, license, and meeting-export UI aligned on the shared theme tokens; meeting export filenames now go through one sanitizer.
- Developer-only automation, scratch captures, .NET build output, and Playwright run output left the release checkout.
- Product version moved to
1.2.0with Android version code and iOS build number1020001.
Fixed
- Desktop configuration writes fail explicitly when the config store is unavailable instead of falling back to a throwaway in-memory store.
- Speech provider and model failures reach the UI as errors instead of surfacing as empty transcriptions.
- Removed the Forgejo release-publishing prohibition; the legacy hardcoded-version script is replaced by a version-gated publisher.
- Dropped a stale admin bundle from the API server web root.
- Mobile team, meeting, memo, template, command, and dictionary screens follow the server contract, and report submission no longer hangs without confirming to the user.
[1.1.0] - 2026-08-29
Added
- Android/iOS product mobile app: implemented account bootstrap and lifecycle, onboarding, recording and audio import, transcription, history, meetings, templates, memos, teams and invitations, notifications, knowledge, AI Talk, generated documents, data portability, and role-aware administration flows.
- Mobile-native recording and intake: added Android foreground recording with persistent controls and recovery, external
ACTION_SENDaudio/video intake, durable upload processing, and supported-device on-device Whisper transcription. - Server-verified mobile monetization: added Google Play purchase and restore verification, entitlement reconciliation, AdMob rewarded SSV validation, and replay-safe reward receipts.
- Generated-content safety controls: added generation receipts, shared generative-AI safety instructions, and an authenticated report flow for owned AI-generated meeting documents.
- Release verification tooling: added Android artifact, App Links, production Firebase/AdMob configuration, Play asset, secret scanning, signed provenance, and publication-boundary checks.
- Complete mobile icon set: added canonical Android legacy/adaptive/monochrome launchers, the 512px Play icon, and all required iPhone, iPad, and App Store marketing icon slots.
Changed
- Unified mobile authentication and invitation links on the canonical
d3ro-voiceapp scheme and added fail-closed verification for the HTTPS App Links contract. - Hardened Supabase and mobile production configuration to fail closed when required public configuration is absent or does not match the release identity.
- Expanded CI coverage across Node.js 24, .NET 10, Deno Edge Functions, React Native tests, Android emulator gates, 16 KB page-size compatibility, and reproducible release evidence.
- Aligned account deletion guidance with external subscription cancellation requirements and centralized local sensitive-data purge on logout, account deletion, and session loss.
- Centralized product version
1.1.0, Android version code1010001, release notes, updater metadata, and app-store changelogs under release SSOT checks. - Finalized benchmarked
ko-KR/en-USPlay listing copy and a Console-previewed 1024×500 canonical feature graphic with preserved generation prompt, source, output, and hashes.
Security
- Added atomic authorization and replay protection for teams, invitations, push delivery, transcription quotas, billing, ad rewards, administrative actions, data portability, and content reports.
- Removed synthetic success fallbacks from protected STT and AI paths; provider and configuration failures now return explicit failures without creating fabricated user data.
- Strengthened generated-document reporting so only the owner can report an existing document backed by immutable generation audit evidence.
- Removed the repository-exposed desktop license private key and prefix-only paid-tier activation; production now accepts only Ed25519 licenses signed by the rotated external key.
- Made Windows stable packaging fail closed unless an external public-trust code-signing PFX is provided and both the NSIS installer and unpacked app pass exact-version, updater SHA-512, Authenticode
Valid, and expected signer-subject verification; local self-signed development certificates are rejected.
Fixed
- Corrected realtime Edge Function model routing and type checks for team and enterprise tiers.
- Removed the legacy
d3ro://deep-link surface to prevent divergent authentication callback identities.
[0.2.1-alpha] - 2026-07-22
Fixed
- v0.2.0-alpha 치명적 버그 (앱 실행 불가): refactor-wave WS2(AppConfig SSOT)에서 7개 키(
customInstructions,llmChains,voiceCommandRules,voiceCommandsEnabled,activeInstructionId,activeChainId,captionAudioSource)를 AppConfig 인터페이스에 추가하고as never16건을 제거했으나CONFIG_DEFAULTS(ConfigService.ts) 기본값을 누락 → 기존 사용자 config(0.1.x) 마이그레이션 시configGet()이undefined반환 → main 프로세스.map()크래시.CONFIG_DEFAULTS에 기본값([],[],[],false,'',null,'mic') 추가. - 교훈: typecheck/lint/unit test가 GREEN이어도 런타임을 보장하지 않음. 배포 전
npm run dev실행 + 산물 설치 e2e 필수 — 이번에 e2e를 건너뛰어 사용자가 발견.
[0.2.0-alpha] - 2026-07-22
다운로드: Windows
D3RO-Voice-Setup-0.2.0-alpha-x64.exe· macOSD3RO-Voice-0.2.0-alpha-arm64.dmg(Release assets 첨부) — 기존 Windows 사용자는 자동 업데이트(4시간 주기)로 수신됩니다. macOS는 ad-hoc 서명 — Gatekeeper 차단 시 우클릭→열기.
Midnight Glass v2 전면 재설계 + refactor-wave 대규모 코드 품질 정리. 기능 추가보다 시각·구조 개선 중심.
Added
- Midnight Glass v2 전면 리디자인: 미드나이트 네이비 + 글래스모피즘 카드(헤어라인 그라디언트 + 앰비언트 글로우) + 시안→블루→퍼플→마젠타 그라디언트. 6종 테마(dark/light/nord/solarized/catppuccin/dracula) 전환. 보더리스 UI + 커스텀 TitleBar. Pretendard Variable 폰트 오프라인 번들.
- TiltCard DS: 커서 추종 3D 틸트 + 스페큘러 빛 반사 (reduced-motion 완전 존중)
- GradientWave / StatRing DS 신규 (기존 CrtDisplay WebGL 오실로스코프를 그라디언트 스펙트럼 웨이브로 교체)
- 실시간 부분 전사: 받아쓰기 홀드 중 RecordingTip에 interim 전사 표시
- Windows 자동 업데이트: electron-updater + GitLab Generic Registry feed (4시간 주기, 재시작 다이얼로그)
- 멀티플랫폼 배포: macOS arm64(ad-hoc 서명) + Windows 동시 릴리스. GitLab Package Registry(버전별+latest) + Release 자동 생성
- 번들 Ollama v0.32.1: 온보딩 gemma4:e4b pull 412 실패 해소, GPU 런타임 가지치기로 ~1GB→119MB 슬림화
- gpt-realtime-2.1 라이브 음성 대화 (Premium 백엔드, OpenAI Realtime) — 로컬 파이프라인 자동 fallback
- Whisper large-v3-turbo 기본 STT: 6배 빠름, 사이드카 백그라운드 다운로드 API
Fixed
- 한국어 날짜 "TODAY"/"YESTERDAY" 표시 버그 —
date.today/yesterday중복 정의(JSON last-wins) 해소 - 팝업/렌더러 색상 충돌 — 팝업이 v1 주황(
#f25b29)으로 렌더되던 것을 파랑(#3b82f6)으로 SSOT 정합 - 받아쓰기 파이프라인 4버그: press 액션이 STT 초기화를 await하며 큐 점유→release 지연/유령 세션; 사이드카 crash 루프; RecordingTip hide 타이머 leak; 빈 STT 가드
- 온보딩 무진행 구간 UX: verifying(모델 검증) 단계 피드백 + 중복 실행 가드
- 마이크 테스트 조기 종료: 초기
level:0을 종료 신호로 오인하던 버그 - 트레이 아이콘 빈 이미지 버그
Changed
accent.amber→accent.main전면 통일 (162건, deprecated 별칭 2곳 제거) — desktop/web/admin/mobile/ui-native 전역. 6 테마 일관.- IPC 채널 SSOT 일원화:
ipc-channels.ts에 9 그룹/키 추가(INSTRUCTION, POPUP_*, CLIPBOARD, APP 등), 하드코딩 채널 문자열 제거,ipcSuccess/ipcError헬퍼 통일,catch(error)패턴 32건 - AppConfig 타입 강화: 7개 누락 키 추가,
as never16건 제거, dynamicrequire→ static import - 팝업 i18n: 4종 팝업(command/history/result/recording-tip) 한국어/영어 하드코딩 →
t()키 주입 (popup.*17키) - 매직 리터럴 토큰화: 인라인 fontSize/fontWeight/borderRadius →
d3roTypo/d3roRadius토큰 (렌더러 전역)
Removed
- DS 미사용 3종: MetalDial, ButtonGroup, CrtDisplay (barrel export 포함)
- 미사용 의존성 8종:
@mui/icons-material,@d3ro/api-client,fluent-ffmpeg(+@types),pdf-parse,node-record-lpcm16(+.d.ts),electron-audio-loopback,@electron-toolkit/preload - i18n 고아 키 967건 (12 locale) — camelCase license.feature 15키 등 동적 참조는 P1 재검증으로 보존
services/index.ts데드 레지스트리,electron-builder.ymlLinux 섹션
Internal
- refactor-wave Wave 1+2: 정책(
docs/REFACTOR_POLICY.md) 수립 → 공격적 감사(106건) → 9 워크스트림 병렬 → 분리 커밋 9개 + 메타 2. 공유 패키지 토큰 SSOT, IPC/타입 정합, 데드코드 대량 제거. 보고서docs/REFACTOR_WAVE1_REPORT.md·WAVE2_REPORT.md.
[1.0.0] - 2026-08-20
Added
Core Pipeline (Phase 1-3)
- Electron 33 + React 19 + MUI 7 + TypeScript strict mode
- AudioCaptureService: SoX-based PCM16 16kHz mono microphone capture
- LocalSTTService: faster-whisper Python sidecar (dual-condition flush pattern)
- VoiceModeService: 9-state RecognitionState + 4-state AudioState dual state machine
- HotkeyService: uiohook-napi global keyboard hooking (hold/toggle/double-press)
- TextInsertService: clipboard save/set/Ctrl+V/restore via @nut-tree-fork/nut-js
- RecordingTip popup: 9 wave bars with cosine distribution, thinking progress bar
- ResultPopup: auto-close with mouse hover pause, copy button
- Settings modal: General/Audio/STT/LLM tabs
- System tray with close-to-tray support
Cursor History Popup (Phase 3.5)
- Ctrl+Shift+V global shortcut opens recent transcription history at cursor position
- Arrow key navigation, Enter to paste, 1-9 direct select, ESC dismiss
- focusable: false maintains active app focus
LLM Integration (Phase 4)
- LocalLLMService: Ollama REST API with NDJSON streaming
- 6 system prompts: refine/translate/summarize/grammar/expand/custom
- VoiceModeService LLM integration with original text fallback
Database & History (Phase 5)
- better-sqlite3 + drizzle-orm (WAL mode)
- HistoryService: CRUD + search + stats + 30-day retention policy
- DictionaryService: custom words + STT initialPrompt injection
- Dashboard with real statistics
Custom Instructions & i18n (Phase 6)
- CustomInstructionService: 5 presets + user-defined CRUD
- i18n system: ko/en with t() function and React context
Build & Test (Phase 7)
- vitest: 41 unit tests (HistoryService, DictionaryService, VoiceModeService, etc.)
- electron-builder: NSIS installer for Windows x64
- GitLab CI/CD: lint, typecheck, test, build, release pipeline
- SoX download script, PyInstaller sidecar build script
Speakly Pattern Enhancement (Phase 7.5)
- SoundEffectService: WAV preload, fire-and-forget playback
- AutoLaunchService: app.setLoginItemSettings
- TextInsertService: basic insertion verification
- VoiceModeService: sound effect integration on session events
UI Redesign (Phase 8)
- Industrial instrument aesthetic (dark theme, amber #f25b29 accent)
- Design system: 9 DS components (CrtDisplay, InstrumentPanel, Led, PhysicalButton, MetalCard, PhosphorText, MetalDial, ScreenPanel, ButtonGroup)
- d3roPalette SSOT: all colors via CSS custom properties
- HotkeyRecordModal: key combination recording UI
- DashboardPage: feature-focused layout with stats + history + system status
Quality & UX (Phase 9)
- 4-step onboarding wizard (welcome, mic, hotkey, Ollama)
- Microphone test UI with level meter
- Audio recording WAV file saving
- MetalDial component with drag rotation
- Ctrl+Shift+C command popup at cursor position
- Dictionary inline editing
- Ollama installation guide with nudging bubble
- Real-time UI refresh on data changes
Killer Features (Phase 10)
- CaptionService: 3-second chunk continuous transcription with overlay popup
- VoiceCommandService: keyword-to-command matching engine (4 preset keywords)
- ScreenContextService: active window detection + selected text capture
- ChainService: sequential LLM command pipeline with step-by-step events
- MemoService: memo tags DB table + tag CRUD + markdown export
- Shared components: EmptyStateCard, SearchInput, PageHeader, HistoryEntryCard
- Shared utilities: formatters.ts (formatDuration, getDateKey, formatNumber)
Monetization (Phase 11)
- LicenseService: Free/Pro/Pro+ 3-tier system
- Feature gating with 16 feature enums and tier-based access mapping
- Daily usage quota tracking (Free: 20 dictation/day, 10 LLM/day)
- LemonSqueezy API integration for license key activation
- UpgradePromptModal, ProBadge, useProFeature hook
- Settings License tab with tier comparison table
Pro Features (Phase 12)
- FileTranscriptionService: ffmpeg PCM conversion, 30s chunk sequential STT
- MeetingSummaryService: caption session LLM summarization
- DictationTemplateService: CRUD + session state machine, 3 presets
Pro+ Premium (Phase 13)
- VoiceConversationService: STT-LLM-TTS loop for voice conversations
- TTSPlaybackService: Windows SAPI via PowerShell
- RAGService: Ollama nomic-embed-text embeddings, SQLite vector storage, cosine similarity
- KnowledgeBasePage: document management, indexing progress, Q&A UI
- VoiceActionService: LLM JSON action plans, 6 presets, dangerous command blocking
Landing Page
- Promotional website at
site/directory - Vite + React 19 + Tailwind CSS
- 4 design references integrated into D3RO brand identity
- CRT screen effects, crosshair decorations, noise textures (CSS only)
- 10-language i18n: en/ko/ja/zh/es/fr/de/pt/ru/vi
- GitHub Pages auto-deployment workflow
Technical Details
- Total IPC channels: ~200+
- Total services: 20+
- Total error codes: 0-878
- Design system components: 9
- Supported languages (app): ko, en
- Supported languages (landing): 10
- Test count: 41 unit tests
- Build target: Windows x64 NSIS installer