voicesmith-mcp 1.0.13 → 1.0.15

package/config.json

@@ -4,7 +4,8 @@
     "voices_path": "~/.local/share/voicesmith-mcp/models/voices-v1.0.bin",
     "default_voice": "am_eric",
     "default_speed": 1.0,
-    "audio_player": "mpv"
+    "audio_player": "mpv",
+    "duck_media": true
   },
   "stt": {
     "model_size": "base",

package/config.py

@@ -26,6 +26,7 @@ class TTSConfig:
     default_voice: str = "am_eric"
     default_speed: float = 1.0
     audio_player: str = "mpv"
+    duck_media: bool = False
 
 
 @dataclass
@@ -99,6 +100,8 @@ def load_config(config_path: Optional[Path] = None) -> AppConfig:
                     config.tts.default_speed = float(tts["default_speed"])
                 if "audio_player" in tts:
                     config.tts.audio_player = tts["audio_player"]
+                if "duck_media" in tts:
+                    config.tts.duck_media = bool(tts["duck_media"])
 
             # STT config
             if "stt" in data:
@@ -179,6 +182,7 @@ def save_config(config: AppConfig, config_path: Optional[Path] = None) -> None:
             "default_voice": config.tts.default_voice,
             "default_speed": config.tts.default_speed,
             "audio_player": config.tts.audio_player,
+            "duck_media": config.tts.duck_media,
         },
         "stt": {
             "model_size": config.stt.model_size,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "voicesmith-mcp",
-  "version": "1.0.13",
+  "version": "1.0.15",
   "description": "Local AI voice for coding assistants — TTS & STT via MCP. Kokoro ONNX + faster-whisper, fully offline.",
   "bin": {
     "voicesmith-mcp": "bin/cli.js"

package/server.py

@@ -41,7 +41,7 @@ from shared import (
     get_logger,
 )
 from config import load_config, save_config, get_config_path, AppConfig
-from session_registry import register_session, unregister_session
+from session_registry import register_session, rename_session, unregister_session
 
 logger = get_logger("server")
 
@@ -82,7 +82,7 @@ def _init_tts(config: AppConfig):
 
     try:
         _tts_engine = KokoroEngine(config.tts.model_path, config.tts.voices_path)
-        _audio_player = AudioPlayer(config.tts.audio_player)
+        _audio_player = AudioPlayer(config.tts.audio_player, duck_media=config.tts.duck_media)
         _speech_queue = SpeechQueue(_tts_engine, _audio_player)
         logger.info("TTS subsystem initialized")
     except TTSEngineError as e:
@@ -454,14 +454,18 @@ async def listen(timeout: float = 15, prompt: str = "", silence_threshold: float
         logger.info(f"Listening (prompt: {prompt})")
 
     try:
+        loop = asyncio.get_running_loop()
+
         # Play ready sound so the user knows to start speaking
         # Skip for push-to-talk (HTTP) — it has its own beep
         if prompt != "push-to-talk":
-            loop = asyncio.get_event_loop()
             await loop.run_in_executor(None, _play_ready_sound)
 
         start = time.perf_counter()
 
+        # Reset VAD state from any prior recording (LSTM hidden state + context)
+        _vad.reset()
+
         # Record audio with VAD
         audio = await _mic_capture.record(
             vad=_vad,
@@ -479,7 +483,6 @@ async def listen(timeout: float = 15, prompt: str = "", silence_threshold: float
         recording_ms = (time.perf_counter() - start) * 1000
 
         # Transcribe
-        loop = asyncio.get_event_loop()
         result = await loop.run_in_executor(
             None, _stt_engine.transcribe, audio, STT_SAMPLE_RATE
         )
@@ -565,6 +568,10 @@ async def get_voice_registry() -> dict:
 async def set_voice(name: str, voice: str) -> dict:
     """Assign or reassign a voice to an agent name.
 
+    Also renames the session so name and voice always match.
+    The name is derived from the voice ID (e.g., "am_fenrir" -> "Fenrir").
+    If the derived name is taken by another session, returns name_occupied error.
+
     Args:
         name: Agent name to assign.
         voice: Kokoro voice ID (e.g., "am_eric"). Must be valid.
@@ -579,17 +586,41 @@ async def set_voice(name: str, voice: str) -> dict:
             "message": f"Voice '{voice}' not found. Use list_voices to see available options.",
         }
 
-    _registry.set_voice(name, voice)
+    # Derive canonical name from voice ID (e.g., "am_fenrir" -> "Fenrir")
+    # The voice ID format is {prefix}_{name}, so split on underscore and capitalize
+    parts = voice.split("_", 1)
+    new_name = parts[1].capitalize() if len(parts) == 2 else name
+
+    old_name = _session_info["name"] if _session_info else name
+
+    # Update sessions.json with conflict check
+    if _session_info:
+        try:
+            updated = rename_session(os.getpid(), new_name, voice)
+            if updated:
+                _session_info.update(updated)
+        except ValueError:
+            return {
+                "success": False,
+                "error": "name_occupied",
+                "message": f"'{new_name}' is occupied by another session.",
+            }
+
+    # Update voice registry (remove old entry, add new)
+    _registry.rename_voice(old_name, new_name, voice)
 
     # Persist last voice name so it survives session restart / resume
     if _config is not None:
-        _config.last_voice_name = name
+        _config.last_voice_name = new_name
         try:
             save_config(_config)
         except Exception as e:
             logger.warning(f"Failed to persist last_voice_name: {e}")
 
-    return {"success": True, "name": name, "voice": voice}
+    result = {"success": True, "name": new_name, "voice": voice}
+    if old_name != new_name:
+        result["previous_name"] = old_name
+    return result
 
 
 @mcp.tool()

package/session_registry.py

@@ -258,6 +258,52 @@ def register_session(
     return session
 
 
+def rename_session(pid: int, new_name: str, new_voice: str) -> Optional[dict]:
+    """Rename this server's session in the registry.
+
+    Updates the name and voice fields for the entry matching pid.
+    Returns the updated session dict, or None if PID not found.
+    Raises ValueError if new_name is taken by another active session.
+    """
+    path = _sessions_path()
+    if not path.exists():
+        return None
+
+    try:
+        with open(path, "r+") as f:
+            fcntl.flock(f, fcntl.LOCK_EX)
+            sessions = _read_sessions(path)
+            sessions = _clean_stale(sessions)
+
+            # Find our entry
+            our_entry = None
+            for s in sessions:
+                if s.get("pid") == pid:
+                    our_entry = s
+                    break
+
+            if our_entry is None:
+                return None
+
+            # Check if new_name is taken by another session
+            if new_name != our_entry["name"]:
+                for s in sessions:
+                    if s.get("name") == new_name and s.get("pid") != pid:
+                        raise ValueError(
+                            f"'{new_name}' is occupied by another session (pid {s.get('pid')})"
+                        )
+
+            our_entry["name"] = new_name
+            our_entry["voice"] = new_voice
+            _write_sessions(path, sessions)
+            return dict(our_entry)
+    except ValueError:
+        raise
+    except OSError as e:
+        logger.warning(f"Failed to rename session: {e}")
+        return None
+
+
 def unregister_session() -> None:
     """Remove this server's session from the registry."""
     path = _sessions_path()

package/stt/mic_capture.py

@@ -1,7 +1,12 @@
 """Microphone capture with VAD-controlled recording."""
 
 import asyncio
+import os
+import platform
 import queue
+import socket
+import subprocess
+import threading
 import time
 from typing import Optional
 
@@ -12,6 +17,65 @@ from stt.vad import VoiceActivityDetector
 
 logger = get_logger("stt.mic")
 
+_CHUNK_SAMPLES = 512        # Silero VAD requires exactly 512-sample chunks at 16kHz
+_CHUNK_BYTES   = _CHUNK_SAMPLES * 4   # float32 = 4 bytes/sample → 2048 bytes/chunk
+_ZERO_CHECK_CHUNKS = 10    # ~320ms of silence before detecting TCC denial
+
+_AUDIO_SERVICE_SOCKET  = "/tmp/voicesmith-audio.sock"
+_LAUNCHAGENT_LABEL     = "com.voicesmith-mcp.audio"
+_LAUNCHAGENT_PLIST     = os.path.expanduser(
+    f"~/Library/LaunchAgents/{_LAUNCHAGENT_LABEL}.plist"
+)
+
+
+def _find_app_binary(name: str) -> Optional[str]:
+    """Return path to a named binary inside VoiceSmithMCP.app, or None."""
+    install_dir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+    binary = os.path.join(install_dir, "VoiceSmithMCP.app", "Contents", "MacOS", name)
+    return binary if os.path.isfile(binary) and os.access(binary, os.X_OK) else None
+
+
+def _launchagent_available() -> bool:
+    """Return True if the VoiceSmithMCP audio LaunchAgent plist is installed."""
+    return os.path.isfile(_LAUNCHAGENT_PLIST)
+
+
+def _ensure_audio_service_running() -> None:
+    """Start the audio LaunchAgent if it is not already running.
+
+    The service is started via launchctl.  We then wait up to 3 seconds for
+    the Unix socket to appear, which signals the service is ready to accept
+    connections.
+    """
+    # If the socket exists and is connectable, service is already running.
+    if _socket_ready():
+        return
+
+    logger.info("Starting audio service via launchctl")
+    try:
+        subprocess.run(
+            ["launchctl", "start", _LAUNCHAGENT_LABEL],
+            capture_output=True,
+            timeout=5,
+        )
+    except Exception as e:
+        raise MicCaptureError(f"Failed to start audio service: {e}") from e
+
+    # Wait up to 3 s for the socket to appear.
+    for _ in range(30):
+        if _socket_ready():
+            return
+        time.sleep(0.1)
+    raise MicCaptureError(
+        "VoiceSmith audio service did not start in time.  "
+        f"Check {_LAUNCHAGENT_PLIST} and launchctl output."
+    )
+
+
+def _socket_ready() -> bool:
+    """Return True if the audio service socket file exists."""
+    return os.path.exists(_AUDIO_SERVICE_SOCKET)
+
 
 class MicCapture:
     """Microphone capture with voice activity detection."""
@@ -31,7 +95,11 @@ class MicCapture:
     ) -> Optional[np.ndarray]:
         """Record audio from the microphone until silence is detected.
 
-        Uses VAD to detect speech and stop recording after a period of silence.
+        On macOS, prefers the audio-service LaunchAgent backend which runs
+        under launchd (ppid=1), ensuring macOS TCC attributes mic permission
+        to VoiceSmithMCP.app rather than to the user's terminal app.
+        Falls back to the audio-capture subprocess if the LaunchAgent is not
+        installed, and to sounddevice on non-macOS systems.
 
         Args:
             vad: VoiceActivityDetector instance for speech detection.
@@ -48,6 +116,150 @@ class MicCapture:
         if self._recording:
             raise MicCaptureError("Another recording is already in progress")
 
+        # Reset VAD state between recordings.
+        vad.reset()
+
+        if platform.system() == "Darwin":
+            if _launchagent_available():
+                return await self._record_via_socket(
+                    vad, timeout, silence_threshold, cancel_event
+                )
+            # Legacy: subprocess fallback for installs without the LaunchAgent.
+            audio_capture_bin = _find_app_binary("audio-service") or _find_app_binary("audio-capture")
+            if audio_capture_bin:
+                return await self._record_via_subprocess(
+                    audio_capture_bin, vad, timeout, silence_threshold, cancel_event
+                )
+
+        return await self._record_via_sounddevice(
+            vad, timeout, silence_threshold, cancel_event
+        )
+
+    # ── LaunchAgent socket backend (macOS primary) ─────────────────────────────
+
+    async def _record_via_socket(
+        self,
+        vad: VoiceActivityDetector,
+        timeout: float,
+        silence_threshold: float,
+        cancel_event: Optional[asyncio.Event],
+    ) -> Optional[np.ndarray]:
+        """Record via the VoiceSmithMCP audio LaunchAgent (Unix socket).
+
+        The LaunchAgent runs under launchd so macOS TCC attributes mic access
+        to com.voicesmith-mcp.launcher, not to the parent terminal app.
+        """
+        loop = asyncio.get_running_loop()
+
+        # Ensure the service is up and the socket is ready.
+        try:
+            await loop.run_in_executor(None, _ensure_audio_service_running)
+        except MicCaptureError:
+            raise
+        except Exception as e:
+            raise MicCaptureError(f"Audio service error: {e}") from e
+
+        # Open socket connection.
+        sock = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
+        try:
+            sock.connect(_AUDIO_SERVICE_SOCKET)
+        except OSError as e:
+            sock.close()
+            raise MicCaptureError(f"Cannot connect to audio service: {e}") from e
+
+        self._recording = True
+        self._stop_flag = False
+        self._audio_queue = queue.Queue()
+
+        def _reader() -> None:
+            """Background thread: reads socket chunks → audio_queue."""
+            try:
+                while True:
+                    data = b""
+                    while len(data) < _CHUNK_BYTES:
+                        got = sock.recv(_CHUNK_BYTES - len(data))
+                        if not got:
+                            return  # service closed connection
+                        data += got
+                    self._audio_queue.put(np.frombuffer(data, dtype=np.float32).copy())
+            except Exception as exc:
+                logger.debug(f"socket reader thread exiting: {exc}")
+
+        reader_thread = threading.Thread(target=_reader, daemon=True)
+        reader_thread.start()
+        logger.info("Microphone recording started (audio-service socket)")
+
+        try:
+            self._flush_queue(int(0.2 * self._sample_rate / _CHUNK_SAMPLES))
+            return await self._run_vad_loop(vad, timeout, silence_threshold, cancel_event)
+        finally:
+            sock.close()  # signals service to stop sending for this session
+            reader_thread.join(timeout=1)
+            self._recording = False
+
+    # ── Subprocess backend (macOS legacy fallback) ─────────────────────────────
+
+    async def _record_via_subprocess(
+        self,
+        binary: str,
+        vad: VoiceActivityDetector,
+        timeout: float,
+        silence_threshold: float,
+        cancel_event: Optional[asyncio.Event],
+    ) -> Optional[np.ndarray]:
+        """Record using a CoreAudio binary inside VoiceSmithMCP.app (legacy)."""
+        self._recording = True
+        self._stop_flag = False
+        self._audio_queue = queue.Queue()
+
+        try:
+            proc = subprocess.Popen(
+                [binary],
+                stdout=subprocess.PIPE,
+                stderr=subprocess.PIPE,
+                close_fds=True,
+            )
+        except Exception as e:
+            self._recording = False
+            raise MicCaptureError(f"Failed to start audio binary: {e}") from e
+
+        logger.info("Microphone recording started (subprocess fallback)")
+
+        def _reader() -> None:
+            try:
+                while True:
+                    data = proc.stdout.read(_CHUNK_BYTES)
+                    if not data or len(data) < _CHUNK_BYTES:
+                        break
+                    self._audio_queue.put(np.frombuffer(data, dtype=np.float32).copy())
+            except Exception as exc:
+                logger.debug(f"subprocess reader thread exiting: {exc}")
+
+        reader_thread = threading.Thread(target=_reader, daemon=True)
+        reader_thread.start()
+
+        try:
+            self._flush_queue(int(0.2 * self._sample_rate / _CHUNK_SAMPLES))
+            return await self._run_vad_loop(vad, timeout, silence_threshold, cancel_event)
+        finally:
+            proc.terminate()
+            try:
+                proc.wait(timeout=1)
+            except Exception:
+                proc.kill()
+            reader_thread.join(timeout=1)
+            self._recording = False
+
+    # ── sounddevice backend (non-macOS fallback) ───────────────────────────────
+
+    async def _record_via_sounddevice(
+        self,
+        vad: VoiceActivityDetector,
+        timeout: float,
+        silence_threshold: float,
+        cancel_event: Optional[asyncio.Event],
+    ) -> Optional[np.ndarray]:
+        """Record using sounddevice / PortAudio (fallback for non-macOS)."""
         try:
             import sounddevice as sd
         except Exception as e:
@@ -56,15 +268,6 @@ class MicCapture:
         self._recording = True
         self._stop_flag = False
         self._audio_queue = queue.Queue()
-        chunks: list[np.ndarray] = []
-        speech_detected = False
-        silence_duration = 0.0
-        loop = asyncio.get_event_loop()
-
-        # Reset VAD state — the LSTM hidden state and context window must
-        # be cleared between recordings to avoid stale state from previous
-        # audio affecting speech detection.
-        vad.reset()
 
         stream = None
         try:
@@ -72,94 +275,138 @@ class MicCapture:
                 samplerate=self._sample_rate,
                 channels=1,
                 dtype="float32",
-                blocksize=512,  # Silero VAD expects 512-sample chunks at 16kHz
+                blocksize=_CHUNK_SAMPLES,
                 callback=self._audio_callback,
             )
             stream.start()
-            logger.info("Microphone recording started")
-
-            # Discard the first ~200ms of audio to avoid picking up residual
-            # speaker output (Tink sound or TTS playback that just finished).
-            # This prevents VAD from detecting speaker bleed as "speech" and
-            # then cutting off when the bleed stops.
-            flush_chunks = int(0.2 * self._sample_rate / 512)  # ~6 chunks
-            for _ in range(flush_chunks):
-                try:
-                    self._audio_queue.get(timeout=0.1)
-                except queue.Empty:
-                    break
-
-            start_time = asyncio.get_event_loop().time()
-
-            while not self._stop_flag:
-                # Check cancellation
-                if cancel_event and cancel_event.is_set():
-                    logger.info("Recording cancelled by event")
-                    break
-
-                # Check timeout
-                elapsed = asyncio.get_event_loop().time() - start_time
-                if elapsed >= timeout:
-                    if not speech_detected:
-                        logger.info("Recording timed out with no speech detected")
-                    else:
-                        logger.info("Recording timed out")
-                    break
-
-                # Get audio chunk from queue
-                try:
-                    chunk = await loop.run_in_executor(
-                        None, self._audio_queue.get, True, 0.1
-                    )
-                except queue.Empty:
-                    continue
-
-                chunks.append(chunk)
-                is_speech = vad.is_speech(chunk)
-
-                if is_speech:
-                    speech_detected = True
-                    silence_duration = 0.0
-                elif speech_detected:
-                    # Count silence after speech was detected
-                    chunk_duration = len(chunk) / self._sample_rate
-                    silence_duration += chunk_duration
-                    if silence_duration >= silence_threshold:
-                        logger.info(
-                            f"Silence threshold reached ({silence_threshold}s), stopping"
-                        )
-                        break
-
-            if not chunks or not speech_detected:
-                return None
-
-            return np.concatenate(chunks).flatten()
+            logger.info("Microphone recording started (sounddevice)")
 
+            self._flush_queue(int(0.2 * self._sample_rate / _CHUNK_SAMPLES), chunk_timeout=0.1)
+            return await self._run_vad_loop(vad, timeout, silence_threshold, cancel_event)
         except MicCaptureError:
             raise
         except Exception as e:
             raise MicCaptureError(f"Recording failed: {e}") from e
         finally:
-            # Safely tear down the audio stream. The CoreAudio IO thread may
-            # still be executing the callback when we call stop(). Wait briefly
-            # between stop() and close() to let the IO thread finish — this
-            # prevents the segfault in libffi/PortAudio where the callback
-            # dereferences freed memory.
             if stream is not None:
                 try:
                     stream.stop()
-                    time.sleep(0.05)  # Let CoreAudio IO thread finish
+                    time.sleep(0.05)
                     stream.close()
                 except Exception as e:
                     logger.debug(f"Stream teardown: {e}")
             self._recording = False
 
-    def _audio_callback(self, indata, frames, time, status) -> None:
+    # ── Shared helpers ─────────────────────────────────────────────────────────
+
+    def _flush_queue(self, n_chunks: int, chunk_timeout: float = 0.15) -> None:
+        """Discard the first n_chunks from the audio queue (drops speaker bleed)."""
+        for _ in range(n_chunks):
+            try:
+                self._audio_queue.get(timeout=chunk_timeout)
+            except queue.Empty:
+                break
+
+    # ── Shared VAD loop ────────────────────────────────────────────────────────
+
+    async def _run_vad_loop(
+        self,
+        vad: VoiceActivityDetector,
+        timeout: float,
+        silence_threshold: float,
+        cancel_event: Optional[asyncio.Event],
+    ) -> Optional[np.ndarray]:
+        """VAD recording loop — shared by all capture backends.
+
+        Reads 512-sample float32 chunks from self._audio_queue, runs Silero VAD
+        on each, and returns when silence_threshold is exceeded after speech,
+        timeout elapses, or cancel_event fires.
+
+        Raises:
+            MicCaptureError: If audio is all-zeros (TCC denial detected).
+        """
+        loop = asyncio.get_running_loop()
+        chunks: list[np.ndarray] = []
+        speech_detected = False
+        silence_duration = 0.0
+        zero_check_done = False
+        start_time = loop.time()
+
+        while not self._stop_flag:
+            if cancel_event and cancel_event.is_set():
+                logger.info("Recording cancelled by event")
+                break
+
+            elapsed = loop.time() - start_time
+            if elapsed >= timeout:
+                if not speech_detected:
+                    logger.info("Recording timed out with no speech detected")
+                else:
+                    logger.info("Recording timed out")
+                break
+
+            try:
+                chunk = await loop.run_in_executor(
+                    None, self._audio_queue.get, True, 0.1
+                )
+            except queue.Empty:
+                continue
+
+            chunks.append(chunk)
+
+            if not zero_check_done and len(chunks) >= _ZERO_CHECK_CHUNKS:
+                zero_check_done = True
+                if all(np.max(np.abs(c)) == 0.0 for c in chunks):
+                    raise MicCaptureError(self._zero_audio_message())
+
+            is_speech = vad.is_speech(chunk)
+
+            if is_speech:
+                speech_detected = True
+                silence_duration = 0.0
+            elif speech_detected:
+                silence_duration += len(chunk) / self._sample_rate
+                if silence_duration >= silence_threshold:
+                    logger.info(
+                        f"Silence threshold reached ({silence_threshold}s), stopping"
+                    )
+                    break
+
+        if not chunks or not speech_detected:
+            return None
+
+        return np.concatenate(chunks).flatten()
+
+    # ── sounddevice callback ───────────────────────────────────────────────────
+
+    def _audio_callback(self, indata, frames, time_info, status) -> None:
         """Sounddevice callback — pushes audio chunks to the queue."""
         if status:
             logger.warning(f"Audio callback status: {status}")
         self._audio_queue.put(indata.copy())
 
+    # ── Error message ──────────────────────────────────────────────────────────
+
+    @staticmethod
+    def _zero_audio_message() -> str:
+        """Build an error message for zero-amplitude mic input."""
+        msg = (
+            "Microphone is returning silent audio. "
+            "The audio stream opened successfully but every sample is zero."
+        )
+        if platform.system() == "Darwin":
+            msg += (
+                "\n\nmacOS is blocking mic access.  The VoiceSmithMCP audio service "
+                "may not have been granted Microphone permission yet.  "
+                "Check System Settings > Privacy & Security > Microphone and "
+                "ensure VoiceSmithMCP is enabled.\n\n"
+                "If VoiceSmithMCP is not listed, re-run the installer:\n"
+                "  ./install.sh"
+            )
+        return msg
+
+    # ── Properties / control ──────────────────────────────────────────────────
+
     @property
     def is_recording(self) -> bool:
         """Return whether the microphone is currently recording."""

package/tts/audio_player.py

@@ -10,6 +10,7 @@ import time
 import soundfile as sf
 
 from shared import PlaybackResult, AudioPlayerError, AUDIO_LOCK_PATH, get_logger
+from tts.media_duck import duck, unduck
 
 logger = get_logger("tts.audio_player")
 
@@ -17,8 +18,9 @@ logger = get_logger("tts.audio_player")
 class AudioPlayer:
     """Plays audio samples through an external player process."""
 
-    def __init__(self, player_command: str = "mpv") -> None:
+    def __init__(self, player_command: str = "mpv", duck_media: bool = False) -> None:
         self._player_command = player_command
+        self._duck_media = duck_media
         self._process: subprocess.Popen | None = None
 
         # Detect platform fallback if player_command is not available
@@ -82,19 +84,23 @@ class AudioPlayer:
 
             # Cross-session audio lock: prevents overlapping playback
             # flock is kernel-managed — auto-released on crash, no stale locks
-            with open(AUDIO_LOCK_PATH, "w") as lock_file:
-                fcntl.flock(lock_file, fcntl.LOCK_EX)
-
-                start = time.perf_counter()
-                self._process = subprocess.Popen(
-                    cmd,
-                    stdout=subprocess.DEVNULL,
-                    stderr=subprocess.DEVNULL,
-                )
-                self._process.wait()
-                duration_ms = (time.perf_counter() - start) * 1000
-
-            # Lock released when lock_file closes
+            paused_apps = duck() if self._duck_media else []
+            try:
+                with open(AUDIO_LOCK_PATH, "w") as lock_file:
+                    fcntl.flock(lock_file, fcntl.LOCK_EX)
+
+                    start = time.perf_counter()
+                    self._process = subprocess.Popen(
+                        cmd,
+                        stdout=subprocess.DEVNULL,
+                        stderr=subprocess.DEVNULL,
+                    )
+                    self._process.wait()
+                    duration_ms = (time.perf_counter() - start) * 1000
+
+                # Lock released when lock_file closes
+            finally:
+                unduck(paused_apps)
 
             if self._process.returncode != 0:
                 return PlaybackResult(

package/tts/media_duck.py

@@ -0,0 +1,146 @@
+"""macOS media ducking via osascript.
+
+Pauses media apps (Apple Music, Spotify) and browser tabs (Chrome, Brave,
+Edge, Safari) before VoiceSmith audio playback and resumes them afterward.
+No-ops on non-macOS systems.
+
+Browser ducking uses JavaScript injection via AppleScript.  The first time
+each browser is targeted, macOS will prompt for Automation permission — approve
+once and it is remembered.
+
+Usage:
+    paused = duck()        # pause everything playing; returns opaque token list
+    ...play audio...
+    unduck(paused)         # resume only what we paused
+"""
+
+import platform
+import subprocess
+
+from shared import get_logger
+
+logger = get_logger("tts.media_duck")
+
+# ── Native media apps ─────────────────────────────────────────────────────────
+
+# (display name, AppleScript target)
+_APPS = [
+    ("Apple Music", "Music"),
+    ("Spotify",     "Spotify"),
+]
+
+# ── Browsers ──────────────────────────────────────────────────────────────────
+
+# (display name, AppleScript target, family: "chrome" | "safari")
+_BROWSERS = [
+    ("Google Chrome",   "Google Chrome",   "chrome"),
+    ("Brave Browser",   "Brave Browser",   "chrome"),
+    ("Microsoft Edge",  "Microsoft Edge",  "chrome"),
+    ("Safari",          "Safari",          "safari"),
+]
+
+# JS injected into every tab on duck: pause playing media and mark it.
+_JS_PAUSE = (
+    "document.querySelectorAll('video,audio').forEach(function(v){"
+    "if(!v.paused){v.pause();v.dataset.voicesmithPaused='1'}"
+    "})"
+)
+
+# JS injected on unduck: resume only elements we marked, then clear the mark.
+_JS_RESUME = (
+    "document.querySelectorAll('video,audio').forEach(function(v){"
+    "if(v.dataset.voicesmithPaused){delete v.dataset.voicesmithPaused;v.play()}"
+    "})"
+)
+
+# ── Helpers ───────────────────────────────────────────────────────────────────
+
+def _osascript(script: str) -> str:
+    """Run an AppleScript (may be multi-line); return stdout stripped, or '' on error."""
+    try:
+        result = subprocess.run(
+            ["osascript"],
+            input=script,
+            capture_output=True,
+            text=True,
+            timeout=5,
+        )
+        return result.stdout.strip()
+    except Exception:
+        return ""
+
+
+def _browser_script(target: str, family: str, js: str) -> str:
+    """Build an AppleScript that runs js in every tab of target browser."""
+    if family == "safari":
+        exec_stmt = f'do JavaScript "{js}" in t'
+    else:  # chrome family
+        exec_stmt = f'execute t javascript "{js}"'
+
+    return f"""\
+tell application "{target}"
+    repeat with w in windows
+        repeat with t in tabs of w
+            try
+                {exec_stmt}
+            end try
+        end repeat
+    end repeat
+end tell"""
+
+
+# ── Public API ────────────────────────────────────────────────────────────────
+
+def duck() -> list[str]:
+    """Pause any playing media apps and browser tabs.
+
+    Returns:
+        Opaque list of tokens — pass unchanged to unduck().
+    """
+    if platform.system() != "Darwin":
+        return []
+
+    paused: list[str] = []
+
+    # Native apps (Music, Spotify)
+    for display_name, target in _APPS:
+        if _osascript(f'application "{target}" is running') != "true":
+            continue
+        if _osascript(f'tell application "{target}" to get player state') == "playing":
+            _osascript(f'tell application "{target}" to pause')
+            paused.append(target)
+            logger.debug(f"Ducked {display_name}")
+
+    # Browsers — inject pause JS into every tab
+    for display_name, target, family in _BROWSERS:
+        if _osascript(f'application "{target}" is running') != "true":
+            continue
+        _osascript(_browser_script(target, family, _JS_PAUSE))
+        paused.append(f"browser:{target}")
+        logger.debug(f"Ducked browser tabs in {display_name}")
+
+    return paused
+
+
+def unduck(paused: list[str]) -> None:
+    """Resume apps and browser tabs paused by duck().
+
+    Args:
+        paused: The list returned by a previous duck() call.
+    """
+    if platform.system() != "Darwin":
+        return
+
+    for token in paused:
+        if token.startswith("browser:"):
+            target = token[len("browser:"):]
+            # family lookup for resume script
+            family = next(
+                (f for _, t, f in _BROWSERS if t == target),
+                "chrome",
+            )
+            _osascript(_browser_script(target, family, _JS_RESUME))
+            logger.debug(f"Unducked browser tabs in {target}")
+        else:
+            _osascript(f'tell application "{token}" to play')
+            logger.debug(f"Unducked {token}")

package/tts/speech_queue.py

@@ -55,7 +55,7 @@ class SpeechQueue:
         speed: float,
     ) -> SpeakResult:
         """Internal: synthesize and play text, blocking until done."""
-        loop = asyncio.get_event_loop()
+        loop = asyncio.get_running_loop()
         self._speaking = True
         total_duration_ms = 0.0
         total_synthesis_ms = 0.0

package/voice_registry.py

@@ -87,6 +87,22 @@ class VoiceRegistry:
         logger.info(f"Set voice '{voice_id}' for '{name}'")
         return True
 
+    def rename_voice(self, old_name: str, new_name: str, voice_id: str) -> bool:
+        """Rename an agent's registry entry and set a new voice.
+
+        Removes the old name entry and creates a new one.
+        If old_name == new_name, just updates the voice in place.
+        Returns True if the voice_id is valid, False otherwise.
+        """
+        if voice_id not in ALL_VOICE_IDS:
+            logger.warning(f"Invalid voice ID '{voice_id}' for rename '{old_name}' -> '{new_name}'")
+            return False
+        if old_name != new_name and old_name in self._registry:
+            del self._registry[old_name]
+        self._registry[new_name] = voice_id
+        logger.info(f"Renamed '{old_name}' -> '{new_name}' with voice '{voice_id}'")
+        return True
+
     def get_registry(self) -> dict[str, str]:
         """Return a copy of the current registry."""
         return dict(self._registry)