refactor: VAD upgrade, retry, dead code cleanup, AGENT removal

- audio: switch VAD to webrtcvad with RMS gate + fallback to RMS
- audio: honor FOLLOWUP_TIMEOUT — short silence wait after bot response
- llm: retry with exponential backoff on network errors and 5xx
- llm: VOICE_MAX_TOKENS env (default 300) instead of hardcoded 150
- tts: optional VAD-based barge-in (BARGE_IN_ENABLED, off by default)
- tts: remove dead start_barge_in_listener / was_barge_in helpers
- config: drop AGENT/LUSYA_AGENT — routing happens via session_key
- modes: remove unused imports, pass FOLLOWUP_TIMEOUT to follow-up record()
- docs: full rewrite of README and CLAUDE.md to match current architecture

satellite/audio.py

@@ -2,22 +2,59 @@ import os
 import pyaudio
 import numpy as np
 
-from .config import SILENCE_THRESHOLD, SILENCE_DURATION, MAX_DURATION, log
+from .config import (
+    SILENCE_THRESHOLD, SILENCE_DURATION, MAX_DURATION,
+    FOLLOWUP_TIMEOUT, VAD_AGGRESSIVENESS, log,
+)
 from .stt import transcribe
 
 ECHO_WARMUP = float(os.getenv("ECHO_WARMUP", "0.5"))  # сек пропуска в начале — гасит эхо от TTS
 
+try:
+    import webrtcvad
+    _vad = webrtcvad.Vad(VAD_AGGRESSIVENESS)
+    _VAD_OK = True
+except Exception as e:
+    log.warning(f"webrtcvad недоступен, fallback на RMS: {e}")
+    _vad = None
+    _VAD_OK = False
+
+# webrtcvad требует фрейм 10/20/30 мс при 8/16/32/48 кГц
+SAMPLE_RATE = 16000
+FRAME_MS = 30
+FRAME_SAMPLES = int(SAMPLE_RATE * FRAME_MS / 1000)  # 480
+FRAME_BYTES = FRAME_SAMPLES * 2                      # int16
+
+
+def _is_speech(frame: bytes) -> bool:
+    """Единое решение по VAD: webrtcvad + RMS-гейт, чтобы не ловить шёпот и эхо."""
+    amplitude = float(np.abs(np.frombuffer(frame, dtype=np.int16)).mean())
+    if amplitude < SILENCE_THRESHOLD:
+        return False
+    if _VAD_OK:
+        try:
+            return _vad.is_speech(frame, SAMPLE_RATE)
+        except Exception:
+            pass
+    return True  # RMS уже прошёл — считаем речью
+
+
+def record(initial_silence_timeout: float | None = None) -> str:
+    """Запись до тишины + STT.
+    initial_silence_timeout — через сколько секунд выйти если пользователь вообще не начал говорить.
+    По умолчанию FOLLOWUP_TIMEOUT (короткое ожидание после ответа бота).
+    """
+    if initial_silence_timeout is None:
+        initial_silence_timeout = FOLLOWUP_TIMEOUT
 
-def record() -> str:
-    """Запись до тишины (VAD) + STT. Игнорирует ECHO_WARMUP в начале."""
     try:
         audio = pyaudio.PyAudio()
         stream = audio.open(
             format=pyaudio.paInt16,
             channels=1,
-            rate=16000,
+            rate=SAMPLE_RATE,
             input=True,
-            frames_per_buffer=1024,
+            frames_per_buffer=FRAME_SAMPLES,
         )
     except Exception as e:
         log.exception("Не удалось открыть микрофон")
@@ -25,30 +62,38 @@ def record() -> str:
         return ""
 
     print("🎙️  Говори...")
-    frames = []
-    silent_chunks = 0
+    frames: list[bytes] = []
     speaking_started = False
-    max_chunks = int(16000 / 1024 * MAX_DURATION)
-    silence_chunks_needed = int(16000 / 1024 * SILENCE_DURATION)
-    warmup_chunks = int(16000 / 1024 * ECHO_WARMUP)
+    trailing_silence = 0  # фреймы тишины после начала речи
+    initial_silence  = 0  # фреймы тишины до начала речи
+
+    max_frames              = int(MAX_DURATION * 1000 / FRAME_MS)
+    warmup_frames           = int(ECHO_WARMUP * 1000 / FRAME_MS)
+    silence_frames_needed   = int(SILENCE_DURATION * 1000 / FRAME_MS)
+    initial_silence_limit   = int(initial_silence_timeout * 1000 / FRAME_MS)
 
     try:
-        for i in range(max_chunks):
-            data = stream.read(1024, exception_on_overflow=False)
-            if i < warmup_chunks:
-                continue  # гасим эхо от TTS / звука активации
+        for i in range(max_frames):
+            data = stream.read(FRAME_SAMPLES, exception_on_overflow=False)
+            if i < warmup_frames:
+                continue
             frames.append(data)
 
-            amplitude = np.abs(np.frombuffer(data, dtype=np.int16)).mean()
-
-            if amplitude > SILENCE_THRESHOLD:
+            if _is_speech(data):
                 speaking_started = True
-                silent_chunks = 0
-            elif speaking_started:
-                silent_chunks += 1
-                if silent_chunks >= silence_chunks_needed:
-                    print("🔇 Конец речи")
-                    break
+                trailing_silence = 0
+            else:
+                if speaking_started:
+                    trailing_silence += 1
+                    if trailing_silence >= silence_frames_needed:
+                        print("🔇 Конец речи")
+                        break
+                else:
+                    initial_silence += 1
+                    if initial_silence >= initial_silence_limit:
+                        print("😴 Пользователь молчит, выхожу")
+                        speaking_started = False
+                        break
     except Exception as e:
         log.exception("Ошибка при записи аудио")
         print(f"⚠️  Ошибка записи: {e}")