@goodandready/dsh-voice 0.9.35 → 0.9.37

package/lib/host-routes.js

@@ -1,6 +1,6 @@
 // HTTP host routes registered on ctx.webServer.
 import os from 'node:os'
-import { readBody, isLoopback, isTrustedCaller } from './http-util.js'
+import { readBody, isLoopback, isTrustedCaller, getVoiceSessionToken } from './http-util.js'
 import { resolveSensevoiceProvider } from './sensevoice-installer.js'
 import { applyVoiceMacros } from './macros.js'
 import { createSyntheticWav } from './wav.js'
@@ -45,6 +45,7 @@ export function registerHostRoutes(ctx, deps) {
       const sensevoiceError = sensevoiceOk ? null : getSensevoiceError()
       writeJson(res, 200, {
         ok: true,
+        sessionToken: getVoiceSessionToken(),
         whisperRunning: whisperOk,
         whisperError: whisperOk ? null : whisperError,
         sensevoiceRunning: sensevoiceOk,
@@ -154,7 +155,7 @@ export function registerHostRoutes(ctx, deps) {
     path: '/dsh-voice/benchmark',
     handler: async (req, res) => {
       if (req.method !== 'POST') { writeJson(res, 405, { ok: false, error: { code: 'method', message: 'POST only' } }); return }
-      if (!isTrustedCaller(req)) { writeJson(res, 403, { ok: false, error: { code: 'forbidden', message: 'forbidden source origin' } }); return }
+      if (!isTrustedCaller(req, { loopbackOnly: true })) { writeJson(res, 403, { ok: false, error: { code: 'forbidden', message: 'forbidden source origin' } }); return }
 
       const cfg = live()
       const effectiveProvider = resolveSensevoiceProvider(cfg)

package/lib/http-util.js

@@ -1,4 +1,5 @@
 // Pure HTTP helpers shared by host routes. No cordis, no network.
+import { randomBytes, timingSafeEqual } from 'node:crypto'
 
 export class BodyTooLargeError extends Error {
   constructor(message = 'body too large') {
@@ -9,6 +10,29 @@ export class BodyTooLargeError extends Error {
   }
 }
 
+let sessionSecret = ''
+export function getVoiceSessionToken() {
+  if (!sessionSecret) {
+    sessionSecret = randomBytes(24).toString('hex')
+  }
+  return sessionSecret
+}
+
+export function setVoiceSessionToken(t) {
+  sessionSecret = String(t || '')
+}
+
+export function verifyVoiceSessionToken(token) {
+  if (!token || typeof token !== 'string') return false
+  const secret = getVoiceSessionToken()
+  if (token.length !== secret.length) return false
+  try {
+    return timingSafeEqual(Buffer.from(token), Buffer.from(secret))
+  } catch {
+    return false
+  }
+}
+
 export function writeJson(res, code, body) {
   try {
     const headers = { 'Content-Type': 'application/json', 'Cache-Control': 'no-store' }
@@ -77,12 +101,16 @@ export function isLoopback(addr) {
 }
 
 /**
- * Verify that the HTTP request originates from a trusted caller (loopback or same-origin).
- * Fail-closed security guard for state-changing or quota-consuming POST routes.
+ * Verify that the HTTP request originates from a trusted caller.
+ * Fail-closed security guard for state-changing or quota-consuming POST routes (#117, #173, #277).
+ * - Loopback callers (127.0.0.1, ::1) are trusted unless rejected by browser security headers.
+ * - Non-loopback LAN mutating/quota-consuming POST/PUT requests require valid session token (#277).
+ * - Routes marked loopbackOnly strictly reject non-loopback callers.
  * @param {import('node:http').IncomingMessage} req
+ * @param {{ loopbackOnly?: boolean }} [options]
  * @returns {boolean}
  */
-export function isTrustedCaller(req) {
+export function isTrustedCaller(req, options = {}) {
   if (!req) return false
 
   // 1. Browser security headers: immediately reject cross-site and same-site requests (#117, #173)
@@ -113,7 +141,6 @@ export function isTrustedCaller(req) {
   }
 
   // 4. For state-changing or data-carrying requests (POST/PUT/PATCH), require application/json or octet-stream
-  // Cross-origin simple form submissions cannot set application/json without a CORS preflight.
   const method = String(req.method || 'GET').toUpperCase()
   if (method === 'POST' || method === 'PUT' || method === 'PATCH') {
     const contentType = String(req.headers?.['content-type'] || '').toLowerCase()
@@ -128,7 +155,20 @@ export function isTrustedCaller(req) {
   const remote = req.socket?.remoteAddress || req.connection?.remoteAddress || ''
   if (isLoopback(remote)) return true
 
-  // 6. Non-loopback requests must be verified same-origin
+  // If route strictly requires loopback (e.g. host model installer or hardware benchmark)
+  if (options && options.loopbackOnly) return false
+
+  // 6. Non-loopback LAN requests: POST/PUT/PATCH quota-burning operations require session auth token (#277)
+  const isMutating = method === 'POST' || method === 'PUT' || method === 'PATCH'
+  if (isMutating) {
+    const incomingToken = String(req.headers?.['x-dsh-voice-token'] || req.headers?.['x-voice-token'] || '')
+    if (incomingToken && verifyVoiceSessionToken(incomingToken)) {
+      return true
+    }
+    return false
+  }
+
+  // 7. Non-state-changing GET requests from LAN: verify same-origin
   if (secSite === 'same-origin') return true
   if (origin) {
     try {
@@ -146,4 +186,4 @@ export function isTrustedCaller(req) {
   }
 
   return false
-}
+}

package/lib/index.js

@@ -17,6 +17,7 @@ import { registerConfigRoutes } from './config-routes.js'
 import { registerHostRoutes } from './host-routes.js'
 import { createConfigReader } from './config-compat.js'
 import { createStatsTracker, mergeContextVocabulary } from './stats.js'
+import { buildTranscriptionPrompt } from './transcribe-core.js'
 import { registerTranscribeAudioTool } from './tool.js'
 import { createPolishText } from './polish.js'
 import { buildProviderOrder, sessionCommand } from './transcribe-core.js'
@@ -114,7 +115,9 @@ export function apply(ctx, baseConfig) {
   ctx.inject(['settings'], (sctx) => {
     settingsService = sctx.settings
     try {
-      scope = sctx.settings.register(NS, Config, { base: baseConfig })
+      if (typeof sctx.settings?.register === 'function') {
+        scope = sctx.settings.register(NS, Config, { base: baseConfig })
+      }
       if (scope && typeof scope.get === 'function') {
         getConfig = () => scope.get() ?? describeRow()?.config ?? readEntryConfig()
       } else {
@@ -133,16 +136,10 @@ export function apply(ctx, baseConfig) {
     triggerAutostart()
   })
 
-  if (typeof ctx.on === 'function') {
-    if (typeof ctx.effect === 'function') {
-      ctx.effect(() => ctx.on('loader/volatile-update', () => triggerAutostart()), 'dsh-voice: volatile autostart')
-      ctx.effect(() => ctx.on('settings/document-updated', () => triggerAutostart()), 'dsh-voice: settings autostart')
-      ctx.effect(() => ctx.on('config', () => triggerAutostart()), 'dsh-voice: config autostart')
-    } else {
-      ctx.on('loader/volatile-update', () => triggerAutostart())
-      ctx.on('settings/document-updated', () => triggerAutostart())
-      ctx.on('config', () => triggerAutostart())
-    }
+  if (typeof ctx.effect === 'function' && typeof ctx.on === 'function') {
+    ctx.effect(() => ctx.on('loader/volatile-update', () => triggerAutostart()), 'dsh-voice: volatile autostart')
+    ctx.effect(() => ctx.on('settings/document-updated', () => triggerAutostart()), 'dsh-voice: settings autostart')
+    ctx.effect(() => ctx.on('config', () => triggerAutostart()), 'dsh-voice: config autostart')
   }
 
   triggerAutostart()
@@ -161,13 +158,17 @@ export function apply(ctx, baseConfig) {
       defaultModels: DEFAULT_MODELS,
     })
 
-    const vocab = cfg.contextGlossary
-      ? mergeContextVocabulary(cfg.vocabulary, contextWords)
-      : (Array.isArray(cfg.vocabulary) ? cfg.vocabulary : [])
+    const { promptText, keywords, vocabulary: vocab } = buildTranscriptionPrompt({
+      vocabulary: cfg.vocabulary,
+      contextWords: cfg.contextGlossary ? contextWords : [],
+      bilingualDeveloperPrompt: cfg.bilingualDeveloperPrompt !== false,
+      bilingualCustomPrompt: cfg.bilingualCustomPrompt || '',
+      language: modeCfg.language,
+    })
 
     const providers = makeProviders(
       { resolveKey, fetchImpl: fetch, cfg, toWav: (b, sig) => toWav16k(b, cfg.ffmpegBin, sig || signal) },
-      { bytes, mime, lang: modeCfg.language, signal, models, vocabulary: vocab },
+      { bytes, mime, lang: modeCfg.language, signal, models, vocabulary: vocab, promptText, keywords },
     )
     return runChain(order, providers, statsTracker.record)
   }

package/lib/local-daemon.js

@@ -35,8 +35,10 @@ export function createLocalDaemons({ live, ctx }) {
         try { child.kill() } catch { /* ignore */ }
         child = null
       }
+      const rawLang = cfg?.dictation?.language
+      const safeLang = isAutoLang(rawLang) ? 'auto' : String(rawLang || '').replace(/[^a-zA-Z0-9_-]/g, '') || 'auto'
       const command = `${JSON.stringify(cfg.whisperBin)} -m ${JSON.stringify(cfg.whisperModel)}`
-        + ` --host 127.0.0.1 --port 8001 -t 8 -p 1 -l ${isAutoLang(cfg.dictation.language) ? 'auto' : cfg.dictation.language}`
+        + ` --host 127.0.0.1 --port 8001 -t 8 -p 1 -l ${JSON.stringify(safeLang)}`
       const specOpts = { command, onExpiry: 'none', timeoutMs: 60000, stdoutMaxBytes: 4 * 1024 * 1024 }
       const spec = typeof ctx.shell?.resolve === 'function' ? ctx.shell.resolve(specOpts) : specOpts
 

package/lib/providers.js

@@ -87,10 +87,12 @@ function pickModel(models, key) {
 
 export function makeProviders(deps, req) {
   const { resolveKey, fetchImpl, cfg } = deps
-  const { bytes, mime, lang, signal, models } = req
+  const { bytes, mime, lang, signal, models, promptText, keywords } = req
   const vocab = Array.isArray(req.vocabulary)
     ? req.vocabulary.map((w) => String(w || '').trim()).filter(Boolean).join(', ')
     : ''
+  const effectivePrompt = (typeof promptText === 'string' && promptText.trim()) ? promptText.trim() : vocab
+  const effectiveKeywords = (typeof keywords === 'string' && keywords.trim()) ? keywords.trim() : vocab
 
   async function deepgram() {
     const key = await resolveKey(cfg.deepgramKeyEnv)
@@ -99,7 +101,7 @@ export function makeProviders(deps, req) {
     const base = String(cfg.deepgramBaseUrl || 'https://api.deepgram.com').replace(/\/+$/, '')
     const url = `${base}/v1/listen?model=${encodeURIComponent(model)}`
       + (!isAutoLang(lang) ? `&language=${encodeURIComponent(lang)}` : '')
-      + (vocab ? `&keywords=${encodeURIComponent(vocab)}` : '')
+      + (effectiveKeywords ? `&keywords=${encodeURIComponent(effectiveKeywords)}` : '')
       + '&smart_format=true'
     const res = await fetchImpl(url, {
       method: 'POST',
@@ -124,7 +126,7 @@ export function makeProviders(deps, req) {
     form.append('file', new Blob([bytes], { type: mime }), fileName(mime))
     form.append('model', pickModel(models, 'groq'))
     if (!isAutoLang(lang)) form.append('language', lang)
-    if (vocab) form.append('prompt', vocab)
+    if (effectivePrompt) form.append('prompt', effectivePrompt)
     form.append('response_format', 'json')
     const res = await fetchImpl('https://api.groq.com/openai/v1/audio/transcriptions', {
       method: 'POST',
@@ -187,7 +189,7 @@ export function makeProviders(deps, req) {
     // Also override a server started with --translate: this plugin always needs
     // a transcript in the spoken language, never Whisper's English translation.
     form.append('translate', 'false')
-    if (vocab) form.append('prompt', vocab)
+    if (effectivePrompt) form.append('prompt', effectivePrompt)
     form.append('response_format', 'json')
     const res = await fetchImpl(cfg.whisperUrl, { method: 'POST', body: form, signal })
     // whisper.cpp returns 400 with a JSON body on internal failures (e.g. decode
@@ -228,7 +230,7 @@ export function makeProviders(deps, req) {
     if (isOpenAI) {
       form.append('model', pickModel(models, 'sensevoice'))
       if (!isAutoLang(lang)) form.append('language', lang)
-      if (vocab) form.append('prompt', vocab)
+      if (effectivePrompt) form.append('prompt', effectivePrompt)
       form.append('response_format', 'json')
     }
 
@@ -282,7 +284,7 @@ export function makeProviders(deps, req) {
       form.append('file', new Blob([bytes], { type: mime }), fileName(mime))
       form.append('model', model)
       if (lang && lang !== 'auto') form.append('language', lang)
-      if (vocab) form.append('prompt', vocab)
+      if (effectivePrompt) form.append('prompt', effectivePrompt)
       form.append('response_format', 'json')
       const res = await fetchImpl(`${base}/audio/transcriptions`, {
         method: 'POST', headers, body: form, signal,
@@ -303,7 +305,7 @@ export function makeProviders(deps, req) {
         format = 'wav'
       }
       const ask = (spec.prompt || CHAT_AUDIO_PROMPT)
-        + (vocab ? ` Vocabulary hints (spell these correctly): ${vocab}.` : '')
+        + (effectivePrompt ? ` Vocabulary hints (spell these correctly): ${effectivePrompt}.` : '')
         + (lang && lang !== 'auto' ? ` The audio language is ${lang}.` : '')
       const res = await fetchImpl(`${base}/chat/completions`, {
         method: 'POST',

package/lib/schema.js

@@ -112,6 +112,10 @@ export const BaseConfig = z.object({
     .description('Custom words (names, terms) hinted to providers so they recognize them correctly.'),
   contextGlossary: z.boolean().default(true)
     .description('Automatically extract code words and terms from context and hint them to STT providers.'),
+  bilingualDeveloperPrompt: z.boolean().default(true)
+    .description('Inject bilingual developer primer (git, docker, npm, PR, etc.) into the STT prompt so mixed RU/EN technical speech is transcribed accurately.'),
+  bilingualCustomPrompt: z.string().default('')
+    .description('Optional custom terms or prompt primer for the bilingual developer prompt.'),
   noiseSuppression: z.boolean().default(true)
     .description('Enable hardware noise suppression, echo cancellation, and auto gain control.'),
   voiceCommands: z.boolean().default(false)

package/lib/sensevoice-installer.js

@@ -1,7 +1,8 @@
 // lib/sensevoice-installer.js — 1-Click installer for SenseVoice-Small & Whisper Large-v3-Turbo
 // Downloads pre-quantized ONNX models and configures autostart on the host via sherpa-onnx.
 
-import { createWriteStream, existsSync, readdirSync } from 'node:fs'
+import { createWriteStream, createReadStream, existsSync, readdirSync } from 'node:fs'
+import { createHash } from 'node:crypto'
 import { mkdir, unlink } from 'node:fs/promises'
 import { homedir } from 'node:os'
 import path from 'node:path'
@@ -20,6 +21,29 @@ export const WHISPER_TURBO_MODEL_URL =
 export const WHISPER_TURBO_MODEL_MIRROR =
   'https://huggingface.co/csukuangfj/sherpa-onnx-whisper-turbo/resolve/main/sherpa-onnx-whisper-turbo.tar.bz2'
 
+export const MODEL_SHA256 = {
+  'sensevoice': 'f6b2a72ebcb1ac7a764d4cfccd886e6bcb2a95c4657c2199d0ba95ed4b9ea71a',
+  'whisper-turbo': 'b11acbbcd660b44a8e0df33724feb5aaa709cf65668f2823d59f656312544f22',
+}
+
+export function verifyArchiveSha256(filePath, expectedSha256) {
+  if (!expectedSha256) return Promise.resolve(true)
+  return new Promise((resolve, reject) => {
+    const hash = createHash('sha256')
+    const stream = createReadStream(filePath)
+    stream.on('data', (chunk) => hash.update(chunk))
+    stream.on('end', () => {
+      const digest = hash.digest('hex').toLowerCase()
+      if (digest === expectedSha256.toLowerCase()) {
+        resolve(true)
+      } else {
+        reject(new Error(`SHA-256 verification failed: expected ${expectedSha256}, got ${digest}`))
+      }
+    })
+    stream.on('error', reject)
+  })
+}
+
 let installerState = {
   installing: false,
   progress: 0,
@@ -176,7 +200,7 @@ export async function downloadArchive(url, destPath, onProgress) {
 
 export function extractTarBz2(archivePath, targetDir) {
   return new Promise((resolve, reject) => {
-    const child = spawn('tar', ['-xjf', archivePath, '-C', targetDir], { windowsHide: true })
+    const child = spawn('tar', ['--no-same-owner', '-xjf', archivePath, '-C', targetDir], { windowsHide: true })
     let stderr = ''
     child.stderr.on('data', (d) => { stderr += d })
     child.on('close', (code) => {
@@ -233,6 +257,13 @@ export async function startSensevoiceInstall(ctx, cfg, updateConfigFn, startSens
 
     if (!downloadOk) throw new Error(`Could not download ${targetModel} model from primary or mirror URL`)
 
+    installerState.step = 'verifying checksum'
+    installerState.progress = 92
+    const expectedHash = MODEL_SHA256[targetModel]
+    if (expectedHash) {
+      await verifyArchiveSha256(archivePath, expectedHash)
+    }
+
     installerState.step = 'extracting'
     installerState.progress = 95
     await extractTarBz2(archivePath, modelDir)
@@ -281,7 +312,7 @@ export function registerSensevoiceInstaller(ctx, { liveConfig, isAlive, updateCo
     kind: 'exact',
     path: '/dsh-voice/sensevoice-installer',
     handler: async (req, res) => {
-      if (!isTrustedCaller(req)) {
+      if (!isTrustedCaller(req, { loopbackOnly: true })) {
         writeJson(res, 403, { ok: false, error: 'Forbidden: untrusted caller' })
         return
       }

package/lib/transcribe-core.js

@@ -1,3 +1,4 @@
+import { mergeContextVocabulary } from './stats.js'
 // Pure helpers for the transcribe path: chain selection and session commands.
 
 /**
@@ -47,4 +48,46 @@ export function sessionCommand(text) {
     if (re.test(t)) return cmd
   }
   return null
-}
+}
+
+export const DEFAULT_BILINGUAL_PROMPT = 'TypeScript, Python, JavaScript, Docker, Git commit, PR, push, merge, branch, npm, pnpm, API, JSON, async, await, bug, deploy, Linux, bash. Смешанная речь разработчика: русский язык с английскими терминами, командами и кодом.'
+
+export function buildTranscriptionPrompt({
+  vocabulary = [],
+  contextWords = [],
+  bilingualDeveloperPrompt = true,
+  bilingualCustomPrompt = '',
+  language = 'ru',
+  maxWords = 50,
+} = {}) {
+  const mergedVocab = mergeContextVocabulary(vocabulary, contextWords, maxWords)
+  const isRuOrAuto = !language || language === 'ru' || language === 'auto'
+  let primer = ''
+  if (bilingualDeveloperPrompt && isRuOrAuto) {
+    primer = (typeof bilingualCustomPrompt === 'string' && bilingualCustomPrompt.trim())
+      ? bilingualCustomPrompt.trim()
+      : DEFAULT_BILINGUAL_PROMPT
+  }
+  const vocabStr = mergedVocab.join(', ')
+  let promptText = ''
+  if (primer && vocabStr) {
+    promptText = `${primer} Термины: ${vocabStr}.`
+  } else if (primer) {
+    promptText = primer
+  } else {
+    promptText = vocabStr
+  }
+
+  const keywords = [...mergedVocab]
+  if (bilingualDeveloperPrompt && isRuOrAuto) {
+    for (const w of ['Docker', 'Git', 'commit', 'PR', 'merge', 'branch', 'pnpm', 'npm', 'API', 'JSON', 'TypeScript']) {
+      if (!keywords.includes(w)) keywords.push(w)
+    }
+  }
+
+  return {
+    vocabulary: mergedVocab,
+    promptText: promptText.trim(),
+    keywords: keywords.slice(0, maxWords).join(', '),
+  }
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@goodandready/dsh-voice",
-  "version": "0.9.35",
+  "version": "0.9.37",
   "description": "Voice input for DeepSeek Harness: dictation chunked by pauses and voice messages, each with its own provider fallback chain (Deepgram, Groq, HuggingFace, local whisper.cpp, plus any OpenAI-compatible API of your own).",
   "license": "MIT",
   "type": "module",