claude-phone-local 2.0.0 → 2.0.2

package/.env.example

@@ -73,6 +73,13 @@ AUDIO_BASE_URL=http://voice-app:3000
 PIPER_VOICE=en_US-lessac-medium
 WHISPER_MODEL=small
 
+# ---- voice activity detection (VAD) ----
+# How long the caller must be silent before we finalize their utterance and
+# start transcribing. Every ms here is dead air the caller hears on every
+# single turn. Lower = snappier, but too low risks cutting off a mid-sentence
+# pause as if they'd finished talking.
+VAD_END_SILENCE_MS=700
+
 # ---- cloud mode settings (only needed if STT_MODE/TTS_MODE=cloud) ----
 # ELEVENLABS_API_KEY=your-elevenlabs-api-key
 # ELEVENLABS_VOICE_ID=your-default-voice-id

package/claude-api-server/server.js

@@ -210,11 +210,11 @@ function runClaudeOnce({ fullPrompt, callId, timestamp }) {
 const VOICE_CONTEXT = `[VOICE CALL CONTEXT]
 This query comes via voice call. You MUST include BOTH of these lines in your response:
 
-🗣️ VOICE_RESPONSE: [Your conversational answer in 40 words or less. This is what gets spoken aloud via TTS. Be natural and helpful, like talking to a friend.]
+🗣️ VOICE_RESPONSE: [Your conversational answer, spoken aloud via TTS. Be natural and helpful, like talking to a friend. Keep it as tight as the answer allows - a yes/no or a quick fact might be one sentence, but if the caller asked something that genuinely needs more (multiple steps, several items, an explanation), take the space to answer it properly instead of truncating. Don't pad it, but don't cut off partway through a real answer either - target roughly 2-4 sentences for anything non-trivial, more if the content actually requires it.]
 
 🎯 COMPLETED: [Status summary in 12 words or less. This is for logging only.]
 
-IMPORTANT: The VOICE_RESPONSE line is what the caller HEARS. Make it conversational and complete - don't just say "Done" or "Task completed". Actually answer their question or confirm what you did in a natural way.
+IMPORTANT: The VOICE_RESPONSE line is what the caller HEARS. Make it conversational and complete - don't just say "Done" or "Task completed". Actually answer their question or confirm what you did in a natural way. A short question deserves a short answer, but never sacrifice a complete answer just to hit a word count.
 
 SLACK DELIVERY: When the caller requests delivery to Slack (phrases like "send to Slack", "post to #channel", "message me when done"):
 1. Do the requested work (research, generate content, analyze, etc.)

package/cli/bin/cli-main.js

@@ -1,5 +1,6 @@
 import { Command } from 'commander';
 import chalk from 'chalk';
+import { createRequire } from 'node:module';
 import { setupCommand } from '../lib/commands/setup.js';
 import { startCommand } from '../lib/commands/start.js';
 import { stopCommand } from '../lib/commands/stop.js';
@@ -20,12 +21,15 @@ import { uninstallCommand } from '../lib/commands/uninstall.js';
 import { registerMcpServerWithOutput, mcpServerPath } from '../lib/mcp-register.js';
 import { loadConfig as _loadCfgForMcp } from '../lib/config.js';
 
+const require = createRequire(import.meta.url);
+const pkg = require('../../package.json');
+
 const program = new Command();
 
 program
   .name('claude-phone')
   .description('Voice interface for Claude Code via SIP - Call your AI, and your AI can call you')
-  .version('1.0.0');
+  .version(pkg.version);
 
 program
   .command('setup')

package/mcp-server/index.js

@@ -37,10 +37,19 @@ const TOOLS = [
       'hit a blocker, or finished something they asked to be told about.\n\n' +
       'RETURNS IMMEDIATELY - it does not wait for the call to finish, so keep working ' +
       'on the task while the phone rings.\n\n' +
-      'mode="conversation" keeps the line open so they can answer back and the voice ' +
-      'agent talks with them. To hear what they said, call call_status with the ' +
-      'returned callId once you reach a point where their answer matters. If they do ' +
-      'not pick up or say nothing, the call ends by itself and call_status reports it.',
+      'CHOOSING THE MODE - this is not optional, pick correctly:\n' +
+      '- mode="announce": speaks the message and hangs up. NO reply is ever collected, ' +
+      'not even if your message ends in a question. Only use this for pure FYI ' +
+      'notifications where you genuinely do not need anything back ("the build finished").\n' +
+      '- mode="conversation": keeps the line open so they can answer back and the voice ' +
+      'agent talks with them. REQUIRED whenever your message asks a question, requests a ' +
+      'decision, says "let me know", or the task you are calling about is not actually ' +
+      'finished until they respond.\n\n' +
+      'If you use mode="conversation", the call is NOT done when this tool returns - ' +
+      'you MUST poll call_status with the returned callId (a few times, spaced out, while ' +
+      'continuing other work) until state is COMPLETED, then read conversationHistory for ' +
+      'what they said before treating the task as finished. Calling call_me and moving on ' +
+      'without ever checking call_status silently discards their answer - never do that.',
     inputSchema: {
       type: 'object',
       properties: {
@@ -130,10 +139,13 @@ server.setRequestHandler(CallToolRequestSchema, async (req) => {
             `Calling ${to} now (callId=${callId}, status=${out.status || 'started'}). ` +
             'This returned immediately - carry on with the task while it rings.' +
             (mode === 'conversation'
-              ? `\n\nConversation mode: they can reply. When you need their answer, call ` +
-                `call_status with callId=${callId} and read conversationHistory. If they ` +
-                'do not answer or stay silent, the call ends on its own.'
-              : '\n\nAnnounce mode: it speaks the message and hangs up. No reply is collected.'),
+              ? `\n\nConversation mode - this task is NOT finished yet. You must poll ` +
+                `call_status with callId=${callId} until state is COMPLETED, then read ` +
+                'conversationHistory for their reply before considering this done. Do not ' +
+                'stop after this message - come back to call_status.'
+              : '\n\nAnnounce mode: it speaks the message and hangs up. No reply is collected. ' +
+                'If you actually needed an answer, you used the wrong mode - call call_me ' +
+                'again with mode="conversation".'),
         }],
       };
     }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-phone-local",
-  "version": "2.0.0",
+  "version": "2.0.2",
   "description": "Local/offline fork of NetworkChuck's claude-phone: talk to Claude Code over 3CX/SIP with faster-whisper STT + Piper TTS in one Docker container.",
   "type": "module",
   "bin": {

package/voice-app/lib/audio-fork.js

@@ -44,7 +44,11 @@ class AudioForkSession extends EventEmitter {
     ws,
     callUuid,
     sampleRate = 16000,
-    endSilenceMs = 1500,
+    // Silence after speech before we finalize the utterance and start
+    // transcribing. This is pure dead air on every single turn, so it's kept
+    // as short as the VAD noise-floor tracking reliably tolerates - too low
+    // and a mid-sentence breath gets mistaken for end-of-speech.
+    endSilenceMs = parseInt(process.env.VAD_END_SILENCE_MS || '700', 10),
     minSpeechMs = 350,
     maxUtteranceMs = 60000
   }) {

package/voice-app/lib/sip-handler.js

@@ -121,11 +121,15 @@ function isGoodbye(transcript) {
  * Priority: VOICE_RESPONSE > CUSTOM COMPLETED > COMPLETED > first sentence
  */
 function extractVoiceLine(response) {
-  // Priority 1: VOICE_RESPONSE (new format)
-  var voiceMatch = response.match(/🗣️\s*VOICE_RESPONSE:\s*([^\n]+)/im);
+  // Priority 1: VOICE_RESPONSE (new format). Stops at the next labeled line
+  // (COMPLETED, or another 🗣️/🎯 marker) rather than the first newline, so a
+  // longer answer that wraps onto multiple lines isn't truncated. The word cap
+  // is a sanity ceiling against a runaway response, not a target length - real
+  // answers that need more room than a one-liner are expected and fine.
+  var voiceMatch = response.match(/🗣️\s*VOICE_RESPONSE:\s*([\s\S]+?)(?=\n\s*🎯|\n\s*🗣️|$)/im);
   if (voiceMatch) {
     var text = voiceMatch[1].trim().replace(/\*+/g, '').replace(/\[.*?\]/g, '').trim();
-    if (text && text.split(/\s+/).length <= 60) {
+    if (text && text.split(/\s+/).length <= 200) {
       return text;
     }
   }
@@ -241,6 +245,11 @@ async function conversationLoop(endpoint, dialog, callUuid, options, deviceConfi
     session = await sessionPromise;
     console.log('[' + new Date().toISOString() + '] AUDIO Fork connected');
 
+    // Pre-render the goodbye clip in the background so hanging up doesn't
+    // wait on a fresh Piper round-trip right when the caller wants off the
+    // line. Regenerated per detected language below since turnVoice can change.
+    let goodbyeUrlPromise = ttsService.generateSpeech("Goodbye! Call again anytime.", turnVoice);
+
     // Main conversation loop
     let turnCount = 0;
     const MAX_TURNS = 20;
@@ -292,10 +301,16 @@ async function conversationLoop(endpoint, dialog, callUuid, options, deviceConfi
       const detectedLang = sttResult.language;
 
       // Answer in whatever language the caller just used.
+      const previousTurnVoice = turnVoice;
       turnVoice = voiceForLanguage(detectedLang, voiceId);
       console.log('[' + new Date().toISOString() + '] WHISPER [' + (detectedLang || '?') +
         ' -> voice ' + turnVoice + ']: "' + transcript + '"');
 
+      // Re-render the pre-warmed goodbye clip if the caller's language changed.
+      if (turnVoice !== previousTurnVoice) {
+        goodbyeUrlPromise = ttsService.generateSpeech("Goodbye! Call again anytime.", turnVoice);
+      }
+
       if (!transcript || transcript.trim().length < 2) {
         const clarifyUrl = await ttsService.generateSpeech("Sorry, I didn't catch that. Could you repeat?", turnVoice);
         await endpoint.play(clarifyUrl);
@@ -303,11 +318,21 @@ async function conversationLoop(endpoint, dialog, callUuid, options, deviceConfi
       }
 
       if (isGoodbye(transcript)) {
-        const byeUrl = await ttsService.generateSpeech("Goodbye! Call again anytime.", turnVoice);
+        const byeUrl = await goodbyeUrlPromise;
         await endpoint.play(byeUrl);
         break;
       }
 
+      // Fire the Claude query immediately - everything else in this block
+      // (thinking phrase, hold music/filler loop) runs concurrently with it
+      // instead of blocking it, since generating+playing the thinking phrase
+      // used to add its own TTS round-trip before the query even started.
+      console.log('[' + new Date().toISOString() + '] CLAUDE Querying (device: ' + deviceName + ')...');
+      const claudeQueryPromise = claudeBridge.query(
+        transcript,
+        { callId: callUuid, devicePrompt: devicePrompt }
+      );
+
       // THINKING FEEDBACK
       const thinkingPhrase = getRandomThinkingPhrase();
       console.log('[' + new Date().toISOString() + '] THINKING: "' + thinkingPhrase + '"');
@@ -342,14 +367,9 @@ async function conversationLoop(endpoint, dialog, callUuid, options, deviceConfi
         }
       })();
 
-      // Query Claude with device-specific prompt
-      console.log('[' + new Date().toISOString() + '] CLAUDE Querying (device: ' + deviceName + ')...');
       let claudeResponse;
       try {
-        claudeResponse = await claudeBridge.query(
-          transcript,
-          { callId: callUuid, devicePrompt: devicePrompt }
-        );
+        claudeResponse = await claudeQueryPromise;
       } finally {
         waiting = false;
         try { await keepAlive; } catch (e) {}
@@ -383,9 +403,11 @@ async function conversationLoop(endpoint, dialog, callUuid, options, deviceConfi
   } finally {
     console.log('[' + new Date().toISOString() + '] CONVERSATION Cleanup...');
 
-    try {
-      await claudeBridge.endSession(callUuid);
-    } catch (e) {}
+    // Fire-and-forget: this is a host-side HTTP bookkeeping call with its own
+    // multi-second timeout. Awaiting it here used to leave the call connected
+    // and silent for up to 5s after the caller said goodbye, before the SIP
+    // dialog was actually torn down.
+    claudeBridge.endSession(callUuid).catch(function () {});
 
     if (forkRunning) {
       try {