Add voice state tags, SuperTTS configs, and demo WS log groups.

Parse leading <state> tags from LLM replies and emit response.state over the product websocket while stripping tags from TTS/text streams. Add FastGPT+Xfyun voice configs (including state-enabled preset), SuperTTS support, and context sync for interrupted turns. Refresh the voice demo with a state indicator and collapsible audio delta websocket log groups.

Co-authored-by: Cursor <cursoragent@cursor.com>

src/voice/config.py

@@ -26,6 +26,9 @@ def resolve_voice_config_path() -> Path:
 
 DEFAULT_VOICE_CONFIG = resolve_voice_config_path()
 
+SUPPORTED_LLM_PROVIDERS = frozenset({"openai", "fastgpt"})
+_LLM_PROVIDER_ALIASES = {"llm": "openai", "openai": "openai", "fastgpt": "fastgpt"}
+
 
 @dataclass(frozen=True)
 class ServerConfig:
@@ -93,11 +96,20 @@ class TurnConfig:
     )
 
 
+@dataclass(frozen=True)
+class ResponseStateConfig:
+    enabled: bool = False
+    tag: str = "state"
+    event_type: str = "response.state"
+    max_prefix_chars: int = 256
+
+
 @dataclass(frozen=True)
 class AgentConfig:
     system_prompt: str = "You are a helpful, friendly voice assistant."
     greeting: str | None = None
     greeting_mode: str = "generated"
+    response_state: ResponseStateConfig = field(default_factory=ResponseStateConfig)
 
 
 @dataclass(frozen=True)
@@ -106,6 +118,7 @@ class LLMConfig:
     api_key: str = ""
     base_url: str | None = None
     model: str = "gpt-4o-mini"
+    app_id: str | None = None
     temperature: float | None = 0.7
     chat_id: str | None = None
     variables: dict[str, str] = field(default_factory=dict)
@@ -113,6 +126,19 @@ class LLMConfig:
     timeout_sec: float = 60.0
     send_system_prompt: bool = False
 
+    @property
+    def is_fastgpt(self) -> bool:
+        return self.provider == "fastgpt"
+
+    @property
+    def is_openai(self) -> bool:
+        return self.provider == "openai"
+
+    @property
+    def uses_local_context_history(self) -> bool:
+        """Whether the pipeline should seed and maintain local LLM context history."""
+        return not self.is_fastgpt or self.send_system_prompt
+
 
 @dataclass(frozen=True)
 class STTConfig:
@@ -147,6 +173,8 @@ class TTSConfig:
     pitch: int = 50
     timeout_sec: float = 30.0
     source_sample_rate_hz: int | None = None
+    oral_level: str = "mid"
+    text_aggregation_mode: str | None = None
 
 
 @dataclass(frozen=True)
@@ -183,14 +211,24 @@ def config_from_dict(data: dict) -> EngineConfig:
         agent["greeting"] = None
     if agent.get("greeting_mode") not in (None, "generated", "fixed", "off"):
         raise ValueError("agent.greeting_mode must be one of: generated, fixed, off")
+    response_state = ResponseStateConfig(**_dict(agent.pop("response_state")))
+    if response_state.max_prefix_chars < 1:
+        raise ValueError("agent.response_state.max_prefix_chars must be greater than 0")
+    if not response_state.tag:
+        raise ValueError("agent.response_state.tag must not be empty")
+    if not response_state.event_type:
+        raise ValueError("agent.response_state.event_type must not be empty")
 
     stt = _dict(services.get("stt") or services.get("asr"))
     if stt.get("language") == "":
         stt["language"] = None
 
     llm = _dict(services.get("llm"))
+    llm["provider"] = _normalize_llm_provider(llm.get("provider", LLMConfig().provider))
     if llm.get("chat_id") == "":
         llm["chat_id"] = None
+    if llm.get("app_id") == "":
+        llm["app_id"] = None
     if not isinstance(llm.get("variables"), dict):
         llm["variables"] = {}
 
@@ -219,7 +257,7 @@ def config_from_dict(data: dict) -> EngineConfig:
                 )
             ),
         ),
-        agent=AgentConfig(**agent),
+        agent=AgentConfig(**agent, response_state=response_state),
         services=ServicesConfig(
             llm=LLMConfig(**llm),
             stt=STTConfig(**stt),
@@ -230,3 +268,14 @@ def config_from_dict(data: dict) -> EngineConfig:
 
 def _dict(value: object) -> dict:
     return dict(value) if isinstance(value, dict) else {}
+
+
+def _normalize_llm_provider(value: object) -> str:
+    provider = str(value or LLMConfig().provider).strip().lower()
+    normalized = _LLM_PROVIDER_ALIASES.get(provider)
+    if normalized is None:
+        supported = ", ".join(sorted(SUPPORTED_LLM_PROVIDERS | {"llm"}))
+        raise ValueError(
+            f"services.llm.provider must be one of: {supported}; got {value!r}"
+        )
+    return normalized