[lelamp] Tuning voice agent

lelamp/config.py

@@ -332,6 +332,7 @@ def _lamp_cfg_get(key: str, default: str = "") -> str:
 # For Gemini: "off" disables automatic activity detection; any other value enables it.
 # For OpenAI: maps to turn_detection type in session config.
 REALTIME_TURN_DETECTION: str = os.environ.get("LELAMP_REALTIME_TURN_DETECTION", "off")
+REALTIME_CONNECT_MAX_RETRIES: int = int(os.environ.get("LELAMP_REALTIME_CONNECT_MAX_RETRIES", "3"))
 
 # --- Realtime: Gemini Live ---
 REALTIME_GEMINI_API_KEY: str = (

lelamp/service/realtime/config.py

@@ -17,15 +17,15 @@
 
 
 def _load_language() -> str | None:
-    """Load language from Lamp's config.json (stt_language field)."""
+    """Load language from device config.json (stt_language field)."""
     from lelamp.config import _lamp_cfg_get
 
     lang: str = _lamp_cfg_get("stt_language", "").strip()
     return lang if lang else None
 
 
 def _parse_turn_detection(value: str) -> OpenAITurnDetectionType | None:
-    """Parse LELAMP_REALTIME_TURN_DETECTION into an OpenAITurnDetectionType or None (off)."""
+    """Parse HAL_REALTIME_TURN_DETECTION into an OpenAITurnDetectionType or None (off)."""
     v = value.strip().lower()
     if v in ("off", "none", ""):
         return None
@@ -51,8 +51,6 @@ class OpenAIConfig(BaseModel):
     )
     truncation_type: OpenAITruncationType = OpenAITruncationType.RETENTION_RATIO
     truncation_retention_ratio: float = 0.5
-    max_retries: int = 3
-    reconnect_delay_s: float = 2.0
 
 
 class GeminiConfig(BaseModel):
@@ -73,9 +71,3 @@ class GeminiConfig(BaseModel):
         "",
     )
     context_window_compression: bool = True
-    max_retries: int = 3
-    reconnect_delay_s: float = 2.0
-    send_timeout_s: float = 10.0
-    recv_timeout_s: float = 300.0
-    queue_poll_s: float = 1.0
-    join_timeout_s: float = 5.0