Function connectVoiceStream

src/services/voiceStreamSTT.ts:111–544

(
  callbacks: VoiceStreamCallbacks,
  options?: { language?: string; keyterms?: string[] },
)

// ─── Connection ────────────────────────────────────────────────────────

export async function connectVoiceStream(
  callbacks: VoiceStreamCallbacks,
  options?: { language?: string; keyterms?: string[] },
): Promise<VoiceStreamConnection | null> {
  // Ensure OAuth token is fresh before connecting
  await checkAndRefreshOAuthTokenIfNeeded()

  const tokens = getClaudeAIOAuthTokens()
  if (!tokens?.accessToken) {
    logForDebugging('[voice_stream] No OAuth token available')
    return null
  }

  // voice_stream is a private_api route, but /api/ws/ is also exposed on
  // the api.anthropic.com listener (service_definitions.yaml private-api:
  // visibility.external: true). We target that host instead of claude.ai
  // because the claude.ai CF zone uses TLS fingerprinting and challenges
  // non-browser clients (anthropics/claude-code#34094). Same private-api
  // pod, same OAuth Bearer auth — just a CF zone that doesn't block us.
  // Desktop dictation still uses claude.ai (Swift URLSession has a
  // browser-class JA3 fingerprint, so CF lets it through).
  const wsBaseUrl =
    process.env.VOICE_STREAM_BASE_URL ||
    getOauthConfig()
      .BASE_API_URL.replace('https://', 'wss://')
      .replace('http://', 'ws://')

  if (process.env.VOICE_STREAM_BASE_URL) {
    logForDebugging(
      `[voice_stream] Using VOICE_STREAM_BASE_URL override: ${process.env.VOICE_STREAM_BASE_URL}`,
    )
  }

  const params = new URLSearchParams({
    encoding: 'linear16',
    sample_rate: '16000',
    channels: '1',
    endpointing_ms: '300',
    utterance_end_ms: '1000',
    language: options?.language ?? 'en',
  })

  // Route through conversation-engine with Deepgram Nova 3 (bypassing
  // the server's project_bell_v2_config GrowthBook gate). The server
  // side is anthropics/anthropic#278327 + #281372; this lets us ramp
  // clients independently.
  const isNova3 = getFeatureValue_CACHED_MAY_BE_STALE(
    'tengu_cobalt_frost',
    false,
  )
  if (isNova3) {
    params.set('use_conversation_engine', 'true')
    params.set('stt_provider', 'deepgram-nova3')
    logForDebugging('[voice_stream] Nova 3 gate enabled (tengu_cobalt_frost)')
  }

  // Append keyterms as query params — the voice_stream proxy forwards
  // these to the STT service which applies appropriate boosting.