@goodandready/dsh-voice 0.8.17 → 0.8.19

package/README.md

@@ -80,6 +80,26 @@ graph LR
 
 ---
 
+## 📝 What's New in 0.8.19
+
+Quality batch after the v0.8.18 review (Gitea #79–#87, PR #88):
+
+| Area | Change |
+|------|--------|
+| Settings placeholders | Model/path hints resolve through locale at **render time** — no frozen i18n keys (#79) |
+| Model field hints | `whisperModel` / `sensevoiceModel` show path-to-model copy, not binary-autostart hints (#82) |
+| Visualizer | Liquid wave / dynamic orb read **DSH theme tokens** instead of hardcoded hex (#80) |
+| Style isolation | Injected CSS uses `data-dsh-plugin="dsh-voice"` so neighbour HMR cleanup cannot strip styles (#81) |
+| Source language | Code comments, errors, and tests are English; EN edit-command phrases added. Spoken RU STT patterns remain for recognition (#86) |
+| Locales | **Changed:** the full inline `ru` dictionary was removed. English is the only bundled locale. Install the translation plugin for Russian UI (#85) |
+| Client source | Browser client is built from ordered `lib/client-src/*.js` fragments via `npm run build:client` (#87) |
+| Process docs | Added `docs/design/DESIGN.md`, `index.md`, project `AGENTS.md`, `docs/testing/unit.md` (#83) |
+
+> [!IMPORTANT]
+> **v0.8.19 locale behavior:** without a translation plugin the Web UI stays in English. Session/edit **spoken** command phrases still match Russian speech for STT where documented; interface labels do not ship a second dictionary.
+
+---
+
 ## 🎮 Four Ways to Speak
 
 | Mode | Gesture / Trigger | Behavior |

package/README.ru.md

@@ -80,6 +80,26 @@ graph LR
 
 ---
 
+## 📝 Что нового в 0.8.19
+
+Quality-батч после ревью v0.8.18 (Gitea #79–#87, PR #88):
+
+| Область | Изменение |
+|---------|-----------|
+| Подсказки в настройках | Хинты моделей/путей резолвятся через locale **в момент рендера** — без «замёрзших» i18n-ключей (#79) |
+| Поля моделей | `whisperModel` / `sensevoiceModel` показывают подсказку пути к модели, а не к бинарю (#82) |
+| Визуализатор | Liquid wave / dynamic orb читают **токены темы DSH**, без жёстких hex-цветов (#80) |
+| Изоляция стилей | CSS-теги несут `data-dsh-plugin="dsh-voice"` — соседний HMR не снимет стили (#81) |
+| Язык исходников | Комментарии, ошибки и тесты — английские; добавлены EN-фразы edit-команд. Устные RU-паттерны для STT сохранены (#86) |
+| Локали | **Изменение:** полный inline-словарь `ru` удалён. В комплекте только English. Для русского UI ставьте translation-плагин (#85) |
+| Исходники клиента | Браузерный клиент собирается из `lib/client-src/*.js` через `npm run build:client` (#87) |
+| Документация процесса | Добавлены `docs/design/DESIGN.md`, `index.md`, `AGENTS.md`, `docs/testing/unit.md` (#83) |
+
+> [!IMPORTANT]
+> **Поведение локалей в v0.8.19:** без translation-плагина интерфейс Web UI остаётся английским. Устные **командные** фразы для STT по-прежнему распознают русскую речь там, где это задокументировано; второй словарь UI в пакет не входит.
+
+---
+
 ## 🎮 4 способа голосового ввода
 
 | Режим | Жест / Активация | Поведение |

package/README.zh.md

@@ -80,6 +80,22 @@ graph LR
 
 ---
 
+## 📝 0.8.19 更新说明
+
+v0.8.18 质量审查后的修复批次（Gitea #79–#87，PR #88）：
+
+- 设置占位提示在**渲染时**通过 locale 解析，不再冻结为 i18n key（#79）
+- `whisperModel` / `sensevoiceModel` 使用模型路径提示（#82）
+- 可视化器改用 DSH 主题 token，不再硬编码颜色（#80）
+- 样式标签使用 `data-dsh-plugin="dsh-voice"`，避免邻居插件 HMR 清理误删（#81）
+- 源码语言统一为英文；界面不再内置完整 `ru` 词典（#85）
+- 浏览器端源码拆分为 `lib/client-src/*.js`，由 `npm run build:client` 构建（#87）
+
+> [!IMPORTANT]
+> **v0.8.19 语言说明：** 未安装 translation 插件时，Web UI 保持英文。
+
+---
+
 ## 🎮 四种语音输入方式
 
 | 交互模式 | 触发手势 | 行为效果 |

package/lib/chain.js

@@ -1,5 +1,5 @@
-// Перебор цепочки провайдеров с фоллбеком. Без сети и без cordis — чистая
-// функция, чтобы поведение при отказах можно было проверить юнит-тестом.
+// Provider fallback chain. No network and no cordis — a pure function so
+// failure behaviour is unit-testable.
 
 function normalizeError(err) {
   const cause = err?.cause?.message || err?.message || String(err)
@@ -7,9 +7,9 @@ function normalizeError(err) {
 }
 
 /**
- * @param order {string[]} порядок ключей провайдеров
+ * @param order {string[]} provider keys in attempt order
  * @param providers {Record<string, () => Promise<{ok, provider?, text?, reason?}>>}
- * @param onAttempt {(provider: string, res: {ok: boolean, tookMs: number, reason?: string}) => void} опциональный колбэк статистики
+ * @param onAttempt {(provider: string, res: {ok: boolean, tookMs: number, reason?: string}) => void} optional stats callback
  * @returns {Promise<{provider: string, text: string, tookMs: number}>}
  */
 export async function runChain(order, providers, onAttempt) {

package/lib/client-src/00-open.js

@@ -0,0 +1,24 @@
+// dsh-voice — client half (browser).
+//
+// Two buttons in conversation.input.right:
+//   mic  — dictation: speech is cut on pauses, each chunk is recognized and
+//          appended to the composer; send stays with the user;
+//   wave — voice message: one whole recording; after recognition the text is
+//          sent to the agent when the cancel window expires.
+//
+// The recording pill lives in conversation.input.dock; settings are a card
+// in settings.plugin.item.
+
+window.__ModuleLoader__.load({
+  id: '@goodandready/dsh-voice',
+  factory: (require) => {
+    var module = { exports: {} }
+    var exports = module.exports
+    Object.defineProperty(exports, Symbol.toStringTag, { value: 'Module' })
+    let React = require('react')
+
+    const NS = 'dsh-voice'
+
+    // UI strings live in the locale registry so a separate package can
+    // translate them without touching this plugin. English is the source
+    // language, the default, and the fallback.

package/lib/client-src/10-locale.js

@@ -0,0 +1,148 @@
+    const en = {
+      'saveFailed': 'Some fields were not saved —',
+      'cardHint': 'Dictation and voice messages: providers, chains, local whisper',
+      'expand': 'Expand',
+      'collapse': 'Collapse',
+      'composerUnavailable': 'Composer unavailable',
+      'keySpace': 'Space',
+      'keyUnset': 'not set',
+      'recognitionError': 'recognition failed',
+      'micUnavailable': 'Microphone unavailable: needs HTTPS or localhost',
+      'noRecorder': 'MediaRecorder is not supported by this browser',
+      'browserFailed': 'Browser did not recognize: ',
+      'nothingHeard': 'Nothing was recognized',
+      'dictationBtn': 'Voice typing',
+      'messageBtn': 'Voice message — click or hold',
+      'dictationPill': 'Dictation — text is appended to the input',
+      'messagePill': 'Recording a voice message',
+      'cancel': 'Cancel',
+      'holdHint': 'Hold — release to send',
+      'listening': 'Listening in the browser…',
+      'stop': 'Stop',
+      'transcribing': 'Transcribing…',
+      'sendingIn': 'Sending to the agent in',
+      'secondsShort': ' s',
+      'keepPending': 'Do not send',
+      'hide': 'Hide',
+      'title': 'Voice',
+      'recordSlot': 'Voice recording',
+      'browserHint': 'the browser does the recognition, no key needed',
+      'openaiHint': 'whisper-1 · key OPENAI_API_KEY',
+      'siliconflowHint': 'FunAudioLLM/SenseVoiceSmall · key SILICONFLOW_API_KEY',
+      'deepinfraHint': 'openai/whisper-large-v3-turbo · key DEEPINFRA_API_KEY',
+      'fireworksHint': 'whisper-v3-turbo · key FIREWORKS_API_KEY',
+      'mistralHint': 'voxtral-mini-latest · key MISTRAL_API_KEY',
+      'openrouterHint': 'google/gemini-2.5-flash · key OPENROUTER_API_KEY',
+      'localHint': 'set when the server is started',
+      'up': 'Up',
+      'down': 'Down',
+      'remove': 'Remove',
+      'addProvider': 'Add provider',
+      'chainHint': 'Top to bottom is the order they are tried in',
+      'customName': 'name used in the chain',
+      'customModel': 'model',
+      'customKeyName': 'key name',
+      'customModelHint': 'model to request (empty means the built-in one)',
+      'addCustom': 'Add your own provider',
+      'loadingSettings': 'Loading settings…',
+      'notReady1': 'The harness has not announced this plugin’s settings yet. If it has just restarted, ',
+      'notReady2': 'the section will appear on its own in a few seconds.',
+      'hotkey': 'Voice message key',
+      'pressKey': 'Press a key…',
+      'clearKey': 'Clear the key',
+      'hotkeyHint1': 'Hold it to record, release to send to the agent, Esc cancels. ',
+      'hotkeyHint2': 'Any key will do: a letter, an F-key or a modifier.',
+      'language': 'Language',
+      'dictationHint': 'Speech is cut at pauses and the text is appended to the input.',
+      'pauseMs': 'Pause that ends a phrase, ms',
+      'pauseHint': 'Lower means more frequent chunks and faster text, but a higher risk of cutting a word',
+      'speaking': 'You are speaking…',
+      'silence': 'Pause…',
+      'normalizeTranscript': 'Normalize file transcripts',
+      'undo': 'Undo last insert',
+      'undone': 'Insert undone',
+      'nothingToUndo': 'Nothing to undo',
+      'beep': 'Beep on start/stop',
+      'localOnly': 'Local whisper only',
+      'localOnlyHint': 'Restrict both chains to the local whisper.cpp server: fully offline.',
+      'sendDelay': 'Dictation send delay (ms)',
+      'sendDelayHint': 'Wait before appending a dictated phrase, with a cancel window. 0 = off',
+      'mic': 'Microphone',
+      'micDefault': 'System default',
+      'vocabulary': 'Custom vocabulary (one word per line)',
+      'polish': 'Polish transcript with model',
+      'polishHint': 'Fix punctuation and fillers via the harness model before inserting',
+      'stream': 'Continuous dictation',
+      'streamHint': 'Cut phrases by a timer while you speak instead of waiting for a long pause',
+      'streamChunkMs': 'Stream chunk (ms)',
+      'vadAdapt': 'Adaptive silence',
+      'vadAdaptHint': 'Auto-tune the silence threshold to the pace of your speech. 0 = fixed',
+      'wakeWord': 'Wake word',
+      'wakeWordHint': 'Browser recognition starts recording when speech begins with this phrase. Empty = off',
+      'bargeIn': 'Barge-in',
+      'polishSend': 'Polish whole draft before sending',
+      'polishSendHint': 'Run the composed draft through the model right before sending',
+      'sessionCommands': 'Voice session commands',
+      'sessionCommandsHint': '"send", "cancel", "stop", "continue" act on the session instead of becoming text',
+      'polishBaseUrl': 'Offline polish endpoint',
+      'polishBaseUrlHint': 'OpenAI-compatible /chat/completions base URL, e.g. a local Ollama. Empty = harness model',
+      'polishModel': 'Offline polish model',
+      'polishKeyEnv': 'Offline polish key credential',
+      'voiceCommandsLabel': 'Voice edit commands ("new line", "paragraph")',
+      'normalizeTranscriptHint': 'transcribe_audio: spoken numbers to digits, tidy punctuation',
+      'messageTitle': 'Voice message',
+      'messageHint': 'One whole recording, sent to the agent once it is transcribed.',
+      'undoMs': 'Undo window, ms',
+      'undoHint': 'How long the automatic send can still be called off',
+      'customTitle': 'Your own providers',
+      'customHint': 'Any OpenAI-compatible API. The name becomes available in the chains above.',
+      'general': 'General',
+      'whisperEndpoint': 'Local whisper: endpoint',
+      'whisperEndpointHint': 'POST /inference of a whisper.cpp server',
+      'deepgramEndpoint': 'Deepgram: base URL',
+      'deepgramEndpointHint': 'Base URL for Deepgram or self-hosted deployment (default https://api.deepgram.com)',
+      'whisperBin': 'Local whisper: binary',
+      'whisperBinHint': 'used when autostart is on',
+      'whisperModel': 'Local whisper: model',
+      'whisperModelHint': 'Absolute path to the ggml model file',
+      'whisperAutostart': 'Autostart the local whisper',
+      'save': 'Save',
+      'saved': 'Saved ✓',
+      'openrouterWarning': 'OpenRouter has no /audio/transcriptions \u2014 use the openai-chat-audio template there',
+      'noiseSuppression': 'Hardware noise suppression',
+      'noiseSuppressionHint': 'Enable browser noise suppression, echo cancellation, and auto gain control',
+      'contextGlossary': 'Context glossary injection',
+      'contextGlossaryHint': 'Auto-extract code identifiers and terms from composer to improve STT accuracy',
+      'providerDashboard': 'Provider Latency & Health',
+      'avgLatency': 'Avg latency',
+      'successRate': 'Success',
+      'fast': 'Fast',
+      'normal': 'Normal',
+      'slow': 'Slow',
+      'error': 'Error',
+      'idle': 'No calls',
+      'play': 'Play',
+      'pause': 'Pause',
+      'listenBack': 'Listen back',
+      'lastRecording': 'Last voice note',
+      'sensevoiceHint': 'SenseVoice-ONNX / Sherpa-ONNX · ultra-fast local STT (~50ms)',
+      'sensevoiceEndpoint': 'SenseVoice: endpoint',
+      'sensevoiceEndpointHint': 'POST endpoint of sherpa-onnx or compatible server',
+      'sensevoiceBin': 'SenseVoice: binary',
+      'sensevoiceBinHint': 'used when autostart is on',
+      'sensevoiceModel': 'SenseVoice: model path',
+      'sensevoiceModelHint': 'Absolute path or identifier of the SenseVoice / sherpa-onnx model',
+      'sensevoiceAutostart': 'Autostart the SenseVoice server',
+      'visualizerStyle': 'Audio visualizer style',
+      'visualizerStyleHint': 'Waveform animation inside the recording pill',
+      'visLiquidWave': 'Liquid Wave',
+      'visDynamicOrb': 'Dynamic Orb',
+      'visBars': 'Classic Bars',
+      'visOff': 'Off',
+    }
+
+    // Strings are also needed outside components — in recording handlers and
+    // slot labels — so the translator is module-level, not only via props.
+    let moduleT = (key) => key
+    const t = (key) => moduleT(key)
+

package/lib/client-src/20-css.js

@@ -0,0 +1,39 @@
+    // ------------------------------------------------------------------ css
+    const CSS =
+      '.dvo-btn{display:flex;align-items:center;justify-content:center;width:30px;height:30px;border-radius:50%;border:1px solid var(--dsw-alias-border-l1);background:transparent;color:var(--dsw-alias-label-secondary);cursor:pointer;padding:0;box-sizing:border-box}' +
+      '.dvo-btn:hover{color:var(--dsw-alias-label-primary);border-color:var(--dsw-alias-border-l2)}' +
+      '.dvo-btn[data-err="1"]{color:var(--dsw-alias-state-error-primary);border-color:var(--dsw-alias-state-error-primary)}' +
+      '.dvo-pill{display:flex;align-items:center;gap:10px;height:52px;border-radius:26px;background:var(--dsw-alias-bg-layer-1);border:1px solid var(--dsw-alias-border-l1);padding:0 14px;width:100%;max-width:720px;margin:0 auto;box-shadow:0 8px 24px rgba(0,0,0,.18);box-sizing:border-box}' +
+      '.dvo-pbtn{display:flex;align-items:center;justify-content:center;width:34px;height:34px;border-radius:50%;cursor:pointer;padding:0;flex:none;box-sizing:border-box;border:1px solid var(--dsw-alias-border-l2);background:transparent;color:var(--dsw-alias-label-primary)}' +
+      '.dvo-pbtn:hover{background:var(--dsw-alias-bg-layer-2)}' +
+      '.dvo-wave{flex:1;height:40px;width:100%;color:var(--dsw-alias-label-primary)}' +
+      '.dvo-status{display:flex;align-items:center;gap:8px;color:var(--dsw-alias-label-secondary);font-size:13px;flex:1;overflow:hidden;text-overflow:ellipsis;white-space:nowrap}' +
+      '.dvo-err{color:var(--dsw-alias-state-error-primary)}' +
+      '.dvo-count{font-variant-numeric:tabular-nums;font-size:13px;color:var(--dsw-alias-label-secondary)}' +
+      '.dvo-spin{animation:dvo-spin 1s linear infinite}' +
+      '@keyframes dvo-spin{from{transform:rotate(0)}to{transform:rotate(360deg)}}' +
+      '.dvo-btn-active{color:var(--dsw-alias-label-primary);border-color:var(--dsw-alias-label-primary);background:var(--dsw-alias-bg-layer-2)}' +
+      '.dvo-audio-wrap{display:flex;align-items:center;gap:8px;padding:2px 10px;background:var(--dsw-alias-bg-layer-2);border-radius:14px;border:1px solid var(--dsw-alias-border-l1)}' +
+      '.dvo-audio-play{display:flex;align-items:center;justify-content:center;width:26px;height:26px;border-radius:50%;background:var(--dsw-alias-label-primary);color:var(--dsw-alias-bg-layer-1);border:0;cursor:pointer;padding:0}' +
+      '.dvo-audio-play:hover{opacity:0.9}' +
+      '.dvo-audio-time{font-size:12px;font-variant-numeric:tabular-nums;color:var(--dsw-alias-label-secondary)}' +
+      '.dvo-dash{display:flex;flex-direction:column;gap:8px;margin-top:8px;background:var(--dsw-alias-bg-layer-2);border:1px solid var(--dsw-alias-border-l2);border-radius:10px;padding:12px}' +
+      '.dvo-dash-grid{display:grid;grid-template-columns:repeat(auto-fill,minmax(130px,1fr));gap:8px;margin-top:4px}' +
+      '.dvo-dash-item{display:flex;flex-direction:column;gap:4px;padding:8px 10px;background:var(--dsw-alias-bg-layer-3);border:1px solid var(--dsw-alias-border-l1);border-radius:8px}' +
+      '.dvo-dash-name{font-size:12px;font-weight:600;color:var(--dsw-alias-label-primary)}' +
+      '.dvo-dash-row{display:flex;justify-content:space-between;align-items:center;font-size:11px;color:var(--dsw-alias-label-secondary)}' +
+      '.dvo-badge{display:inline-flex;align-items:center;font-size:11px;padding:1px 6px;border-radius:6px;font-weight:600}' +
+      '.dvo-badge-fast{background:var(--dsw-alias-bg-layer-2);color:var(--dsw-alias-state-success-primary);border:1px solid var(--dsw-alias-state-success-primary)}' +
+      '.dvo-badge-norm{background:var(--dsw-alias-bg-layer-2);color:var(--dsw-alias-state-warning-primary);border:1px solid var(--dsw-alias-state-warning-primary)}' +
+      '.dvo-badge-slow{background:var(--dsw-alias-bg-layer-2);color:var(--dsw-alias-state-warning-primary);border:1px solid var(--dsw-alias-border-l2)}' +
+      '.dvo-badge-err{background:var(--dsw-alias-bg-layer-2);color:var(--dsw-alias-state-error-primary);border:1px solid var(--dsw-alias-state-error-primary)}' +
+      '.dvo-badge-idle{background:var(--dsw-alias-bg-layer-2);color:var(--dsw-alias-label-tertiary)}'
+    const cssId = 'dsh-voice/client.module.css'
+    if (typeof document !== 'undefined' && !document.querySelector('style[data-dsh-plugin="dsh-voice"][data-plugin-css="' + cssId + '"]')) {
+      const tag = document.createElement('style')
+      tag.textContent = CSS
+      tag.setAttribute('data-dsh-plugin', 'dsh-voice')
+      tag.dataset.pluginCss = cssId
+      document.head.appendChild(tag)
+    }
+

package/lib/client-src/30-core.js

@@ -0,0 +1,284 @@
+    // ---------------------------------------------------------------- store
+    const voice = {
+      phase: 'idle',      // idle | recording | processing | pending | error
+      mode: 'dictation',  // dictation | message
+      error: '',
+      rec: null,
+      levels: [],
+      pending: null,      // {text, leftMs} — message-mode cancel window
+      lastNote: null,     // {blob, url, mime, text} — last recording
+      showPlayer: false,
+      inputActions: null,
+      input: null,
+      settings: { vadSilenceMs: 700, autoSendMs: 4000, stream: false, streamChunkMs: 1200, vadAdapt: 0, noiseSuppression: true, contextGlossary: true },
+      listeners: new Set(),
+      notify() { this.listeners.forEach((l) => l()) },
+      set(patch) { Object.assign(this, patch); this.notify() },
+      subscribe(l) { this.listeners.add(l); return () => this.listeners.delete(l) },
+    }
+
+    function useVoice() {
+      const [, force] = React.useReducer((x) => x + 1, 0)
+      React.useEffect(() => voice.subscribe(force), [])
+      return voice
+    }
+
+    // ---------------------------------------------------------------- icons
+    const ic = { width: 18, height: 18, viewBox: '0 0 24 24', fill: 'none', stroke: 'currentColor', strokeWidth: 2, strokeLinecap: 'round', strokeLinejoin: 'round' }
+    const playIcon = () => React.createElement('svg', Object.assign({}, ic, { viewBox: '0 0 24 24' }),
+      React.createElement('polygon', { points: '6 4 20 12 6 20 6 4', fill: 'currentColor', stroke: 'none' }))
+    const pauseIcon = () => React.createElement('svg', Object.assign({}, ic, { viewBox: '0 0 24 24' }),
+      React.createElement('rect', { x: 6, y: 4, width: 4, height: 16, fill: 'currentColor', stroke: 'none' }),
+      React.createElement('rect', { x: 14, y: 4, width: 4, height: 16, fill: 'currentColor', stroke: 'none' }))
+    const micIcon = () => React.createElement('svg', ic,
+      React.createElement('path', { d: 'M12 1a3 3 0 0 0-3 3v8a3 3 0 0 0 6 0V4a3 3 0 0 0-3-3z' }),
+      React.createElement('path', { d: 'M19 10v2a7 7 0 0 1-14 0v-2' }),
+      React.createElement('line', { x1: 12, y1: 19, x2: 12, y2: 23 }))
+    const waveIcon = () => React.createElement('svg', ic,
+      React.createElement('line', { x1: 4, y1: 10, x2: 4, y2: 14 }),
+      React.createElement('line', { x1: 8, y1: 7, x2: 8, y2: 17 }),
+      React.createElement('line', { x1: 12, y1: 4, x2: 12, y2: 20 }),
+      React.createElement('line', { x1: 16, y1: 7, x2: 16, y2: 17 }),
+      React.createElement('line', { x1: 20, y1: 10, x2: 20, y2: 14 }))
+    const xIcon = () => React.createElement('svg', ic,
+      React.createElement('line', { x1: 18, y1: 6, x2: 6, y2: 18 }),
+      React.createElement('line', { x1: 6, y1: 6, x2: 18, y2: 18 }))
+    const stopIcon = () => React.createElement('svg', ic,
+      React.createElement('rect', { x: 7, y: 7, width: 10, height: 10, rx: 2.5, fill: 'currentColor', stroke: 'none' }))
+    const spinIcon = () => React.createElement('svg', Object.assign({}, ic, { className: 'dvo-spin' }),
+      React.createElement('path', { d: 'M21 12a9 9 0 1 1-6.219-8.56' }))
+    const warnIcon = () => React.createElement('svg', ic,
+      React.createElement('path', { d: 'M10.29 3.86L1.82 18a2 2 0 0 0 1.71 3h16.94a2 2 0 0 0 1.71-3L13.71 3.86a2 2 0 0 0-3.42 0z' }),
+      React.createElement('line', { x1: 12, y1: 9, x2: 12, y2: 13 }),
+      React.createElement('line', { x1: 12, y1: 17, x2: 12.01, y2: 17 }))
+    const chevronIcon = () => React.createElement('svg', ic,
+      React.createElement('path', { d: 'M6 9l6 6 6-6' }))
+
+    // ------------------------------------------------------------ transport
+    function blobToBase64(blob) {
+      return new Promise((resolve, reject) => {
+        if (typeof FileReader === 'undefined') { reject(new Error('FileReader not supported')); return }
+        const fr = new FileReader()
+        fr.onload = () => { const s = String(fr.result || ''); resolve(s.indexOf(',') >= 0 ? s.slice(s.indexOf(',') + 1) : s) }
+        fr.onerror = () => reject(new Error('failed to read audio'))
+        fr.readAsDataURL(blob)
+      })
+    }
+
+    function extractContextKeywords() {
+      if (!voice.settings || voice.settings.contextGlossary === false) return []
+      const text = (voice.input && typeof voice.input.draft === 'string') ? voice.input.draft : ''
+      if (!text || text.length < 3) return []
+      const matches = text.match(/\b[A-Za-z_][A-Za-z0-9_]{2,29}\b/g) || []
+      const stop = new Set(['the', 'and', 'for', 'are', 'but', 'not', 'you', 'all', 'any', 'can', 'her', 'was', 'one', 'our', 'out', 'day', 'get', 'has', 'him', 'his', 'how', 'man', 'new', 'now', 'old', 'see', 'two', 'way', 'who', 'boy', 'did', 'its', 'let', 'put', 'say', 'she', 'too', 'use'])
+      const words = []
+      const seen = new Set()
+      for (const m of matches) {
+        const lower = m.toLowerCase()
+        if (!stop.has(lower) && !seen.has(lower)) {
+          seen.add(lower)
+          words.push(m)
+          if (words.length >= 30) break
+        }
+      }
+      return words
+    }
+
+    async function sendAudio(blob, mime, mode) {
+      const dataBase64 = await blobToBase64(blob)
+      const payload = { dataBase64, mimeType: mime, mode }
+      const contextWords = extractContextKeywords()
+      if (contextWords && contextWords.length > 0) payload.contextWords = contextWords
+      const res = await fetch('/dsh-voice/transcribe', {
+        method: 'POST',
+        headers: { 'content-type': 'application/json' },
+        body: JSON.stringify(payload),
+      })
+      let parsed = null
+      try { parsed = await res.json() } catch (e) { /* not json */ }
+      if (!res.ok || !parsed || !parsed.ok) {
+        throw new Error((parsed && parsed.error && parsed.error.message) || `HTTP ${res.status}`)
+      }
+      if (parsed.command) return { command: parsed.command, text: '' }
+      return { text: String(parsed.text || '').trim(), command: null }
+    }
+
+    function tidyPhrase(text) {
+      let s = String(text || '').trim()
+      if (!s) return s
+      s = s.replace(/\s*,\s*/g, ', ')
+      s = s.replace(/(^|[.!?\n]\s+)([a-zа-яё])/gi, (m, lead, ch) => lead + ch.toUpperCase())
+      return s
+    }
+
+    // Spoken edit commands (#37): "new line" / "с новой строки" -> \n and so on.
+    // Applied before normalization, only when enabled in settings.
+    // Russian phrases match RU STT output; English covers EN dictation.
+    const VOICE_COMMANDS = [
+      [/(^|[\s,.!?])с новой строки([\s,.!?]|$)/gi, '$1\n$2'],
+      [/(^|[\s,.!?])новая строка([\s,.!?]|$)/gi, '$1\n$2'],
+      [/(^|[\s,.!?])абзац([\s,.!?]|$)/gi, '$1\n\n$2'],
+      [/(^|\s)тире(\s|$)/gi, '$1—$2'],
+      [/(^|[\s,.!?])new line([\s,.!?]|$)/gi, '$1\n$2'],
+      [/(^|[\s,.!?])paragraph([\s,.!?]|$)/gi, '$1\n\n$2'],
+    ]
+
+    function applyVoiceCommands(text) {
+      let s = text
+      for (const [re, to] of VOICE_COMMANDS) s = s.replace(re, to)
+      return s.replace(/[ \t]+/g, ' ').trim()
+    }
+
+    // Insert history for undo (#29-9). Browser-only storage.
+    const insertHistory = []
+
+    async function undoLastInsert() {
+      const last = insertHistory.pop()
+      if (!last) return t('nothingToUndo')
+      const actions = voice.inputActions
+      if (!actions || typeof actions.setDraft !== 'function') return t('composerUnavailable')
+      const draft = voice.input && typeof voice.input.draft === 'string' ? voice.input.draft : ''
+      if (draft === last.after) actions.setDraft(last.before)
+      else {
+        // Draft was edited manually — cut the last insert as a substring.
+        const i = draft.lastIndexOf(last.added)
+        if (i < 0) { insertHistory.push(last); return t('nothingToUndo') }
+        actions.setDraft((draft.slice(0, i) + draft.slice(i + last.added.length)).replace(/\s+$/, ''))
+      }
+      return t('undone')
+    }
+
+    function appendDraft(text) {
+      const actions = voice.inputActions
+      if (!actions || typeof actions.setDraft !== 'function') {
+        voice.set({ phase: 'error', error: t('composerUnavailable') })
+        return
+      }
+      let clean = voice.settings.voiceCommands ? applyVoiceCommands(text) : tidyPhrase(text)
+      if (!clean) return
+      const draft = voice.input && typeof voice.input.draft === 'string' ? voice.input.draft : ''
+      const before = draft
+      actions.setDraft(draft ? draft + ' ' + clean : clean)
+      const limit = Number(voice.settings.historyLimit)
+      if (limit > 0) {
+        insertHistory.push({ before, added: draft ? ' ' + clean : clean, after: draft ? draft + ' ' + clean : clean })
+        while (insertHistory.length > limit) insertHistory.shift()
+      }
+    }
+
+    // Human-readable key name.
+    const KEY_LABELS = {
+      Control: 'Ctrl', Alt: 'Alt', Shift: 'Shift', Meta: 'Win',
+      Escape: 'Esc',
+    }
+
+    function keyLabel(name) {
+      if (!name) return t('keyUnset')
+      if (name === 'Space') return t('keySpace')
+      if (KEY_LABELS[name]) return KEY_LABELS[name]
+      // Drop the Key/Digit prefix from codes like KeyR and Digit5.
+      return String(name).replace(/^Key/, '').replace(/^Digit/, '')
+    }
+
+    // What to store on key press. Pure modifiers are remembered by name:
+    // left and right have different codes but the user means "either".
+    function keyFromEvent(event) {
+      if (['Control', 'Alt', 'Shift', 'Meta'].includes(event.key)) return event.key
+      if (event.code) return event.code
+      return event.key || ''
+    }
+
+    // Announce that the user started speaking.
+    //
+    // Playback should mute immediately: listening and talking at once is
+    // impossible. There is no plugin-to-plugin API — this broadcasts a window
+    // event that anyone may hear, so either side works alone.
+    function announceVoice(phase) {
+      try {
+        window.dispatchEvent(new CustomEvent('dsh-voice:speaking', { detail: { phase } }))
+      } catch (noEvents) { /* no window — nobody to hear it */ }
+      if (voice.settings.beep) playBeep(phase === 'start' ? 880 : 660)
+    }
+
+    // Short WebAudio beep so start/stop is audible without looking (#29-6).
+    function playBeep(freq) {
+      try {
+        const AC = typeof AudioContext !== 'undefined' ? AudioContext
+          : (typeof webkitAudioContext !== 'undefined' ? webkitAudioContext : null)
+        if (!AC) return
+        const ac = new AC()
+        const osc = ac.createOscillator()
+        const gain = ac.createGain()
+        osc.frequency.value = freq
+        osc.type = 'sine'
+        gain.gain.setValueAtTime(0.12, ac.currentTime)
+        gain.gain.exponentialRampToValueAtTime(0.001, ac.currentTime + 0.09)
+        osc.connect(gain); gain.connect(ac.destination)
+        osc.start(); osc.stop(ac.currentTime + 0.1)
+        osc.onended = () => { try { ac.close() } catch (e) { /* already closed */ } }
+      } catch (noAudio) { /* no audio — not critical */ }
+    }
+
+    // ------------------------------------------------- browser recognition
+    //
+    // A separate leg unlike the others: the browser recognizes speech itself,
+    // nothing is uploaded to the host, no keys are needed, and text appears
+    // word by word while you speak.
+    //
+    // Cost: in Chrome audio goes to Google servers. This provider is never
+    // enabled automatically — only when placed explicitly in a chain.
+    function speechRecognitionCtor() {
+      if (typeof window === 'undefined') return null
+      return window.SpeechRecognition || window.webkitSpeechRecognition || null
+    }
+
+    function browserRecognitionAvailable() {
+      return speechRecognitionCtor() !== null
+    }
+
+    /**
+     * @param options {{lang: string, continuous: boolean, onInterim, onFinal, onError}}
+     * @returns {{stop: Function, abort: Function}}
+     */
+    function startBrowserRecognition(options) {
+      const Ctor = speechRecognitionCtor()
+      const recognition = new Ctor()
+      recognition.lang = options.lang && options.lang !== 'auto' ? options.lang : 'ru-RU'
+      recognition.continuous = options.continuous !== false
+      recognition.interimResults = true
+      let stopped = false
+
+      recognition.onresult = (event) => {
+        let interim = ''
+        for (let i = event.resultIndex; i < event.results.length; i++) {
+          const result = event.results[i]
+          const text = String(result[0] && result[0].transcript || '').trim()
+          if (!text) continue
+          if (result.isFinal) options.onFinal(text)
+          else interim += (interim ? ' ' : '') + text
+        }
+        options.onInterim(interim)
+      }
+      recognition.onerror = (event) => {
+        // no-speech and aborted are normal life, not failures.
+        const code = event && event.error
+        if (code === 'no-speech' || code === 'aborted') return
+        options.onError(code || t('recognitionError'))
+      }
+      // The browser ends recognition on its own (pauses, timeout). Restart
+      // until we stop it, otherwise dictation dies silently on the first pause.
+      recognition.onend = () => {
+        if (stopped) return
+        try { recognition.start() } catch (alreadyRunning) { /* already running */ }
+      }
+
+      try { recognition.start() } catch (cannotStart) {
+        options.onError(String(cannotStart && cannotStart.message || cannotStart))
+      }
+      return {
+        stop() { stopped = true; try { recognition.stop() } catch (already) { /* already stopped */ } },
+        abort() { stopped = true; try { recognition.abort() } catch (already) { /* already stopped */ } },
+      }
+    }
+
+    // Fetch the mode chain from the host once. Only needed to decide whether