feat: add hermes gateway llm adapter

custom_agent.py

@@ -9,6 +9,7 @@ from dataclasses import dataclass
 from pathlib import Path
 
 from dotenv import load_dotenv
+from hermes_gateway import GatewaySessionState, HermesGatewayLLM
 from memory import MemoryRecallClient
 from tts import BlackboxTTS
 
@@ -638,12 +639,20 @@ async def entrypoint(ctx: JobContext) -> None:
     LLM_BASE_URL = os.getenv("CUSTOM_LLM_BASE_URL")
     LLM_MODEL = os.getenv("CUSTOM_LLM_MODEL", "qwen-max")
     LLM_API_KEY = os.getenv("CUSTOM_LLM_API_KEY")
+    LLM_PROVIDER = os.getenv("CUSTOM_LLM_PROVIDER", "openai").strip().lower()
     TEXT_LLM_MODEL = os.getenv("CUSTOM_TEXT_LLM_MODEL", LLM_MODEL)
     VISION_LLM_MODEL = os.getenv("CUSTOM_VISION_LLM_MODEL", LLM_MODEL)
     INPUT_MODE = _normalize_input_mode(os.getenv("CUSTOM_AGENT_INPUT_MODE"))
-    if not LLM_API_KEY:
+    if LLM_PROVIDER not in {"openai", "openai-compatible", "hermes", "hermes_gateway", "openclaw"}:
+        raise RuntimeError(f"Unsupported CUSTOM_LLM_PROVIDER={LLM_PROVIDER!r}")
+    if LLM_PROVIDER in {"openai", "openai-compatible"} and not LLM_API_KEY:
         raise RuntimeError(f"CUSTOM_LLM_API_KEY is not set in {CUSTOM_ENV_PATH}")
-    logger.info("Using LLM model=%s base_url=%s", LLM_MODEL, LLM_BASE_URL or "OpenAI default")
+    logger.info(
+        "Using LLM provider=%s model=%s base_url=%s",
+        LLM_PROVIDER,
+        LLM_MODEL,
+        LLM_BASE_URL or "OpenAI default",
+    )
 
     TTS_URL = os.getenv("CUSTOM_TTS_URL") or os.getenv(
         "VOXCPM_TTS_URL", "http://localhost:5000/tts-blackbox"
@@ -668,38 +677,75 @@ async def entrypoint(ctx: JobContext) -> None:
     )
     stt_stream = stt.StreamAdapter(stt=blackbox_stt, vad=ctx.proc.userdata["vad"])
 
-    import httpx
-    from openai import AsyncClient as OpenAIAsyncClient
+    if LLM_PROVIDER in {"hermes", "hermes_gateway", "openclaw"}:
+        gateway_url = os.getenv("CUSTOM_HERMES_GATEWAY_URL", "").strip()
+        if not gateway_url:
+            raise RuntimeError(f"CUSTOM_HERMES_GATEWAY_URL is not set in {CUSTOM_ENV_PATH}")
 
-    # OpenAI-compatible endpoints can be used by setting CUSTOM_LLM_BASE_URL.
-    http_client = httpx.AsyncClient(verify=_env_bool("CUSTOM_LLM_VERIFY_SSL", False))
-
-    if LLM_BASE_URL:
-        openai_client = OpenAIAsyncClient(
-            api_key=LLM_API_KEY,
-            base_url=LLM_BASE_URL,
-            http_client=http_client,
+        hermes_agent_id = os.getenv("CUSTOM_HERMES_AGENT_ID") or None
+        hermes_session_mode = os.getenv("CUSTOM_HERMES_SESSION_MODE", "per_room").strip().lower()
+        if hermes_session_mode != "per_room":
+            raise RuntimeError("CUSTOM_HERMES_SESSION_MODE must be per_room")
+        hermes_token = (
+            os.getenv("CUSTOM_HERMES_API_KEY")
+            or os.getenv("CUSTOM_HERMES_TOKEN")
+            or LLM_API_KEY
+            or None
+        )
+        hermes_state = GatewaySessionState(
+            room_name=ctx.room.name,
+            agent_id=hermes_agent_id,
+            session_mode=hermes_session_mode,
+        )
+        base_llm = HermesGatewayLLM(
+            url=gateway_url,
+            token=hermes_token,
+            state=hermes_state,
+            agent_id=hermes_agent_id,
+            model_name=os.getenv("CUSTOM_HERMES_MODEL", "hermes-agent"),
+            request_timeout=_env_float("CUSTOM_HERMES_REQUEST_TIMEOUT", 30.0),
+        )
+        text_llm = base_llm
+        vision_llm = base_llm
+        logger.info(
+            "Using Hermes/OpenClaw gateway url=%s agent_id=%s session_key=%s",
+            gateway_url,
+            hermes_agent_id or "default",
+            hermes_state.session_key,
         )
     else:
-        openai_client = OpenAIAsyncClient(
-            api_key=LLM_API_KEY,
-            http_client=http_client,
-        )
+        import httpx
+        from openai import AsyncClient as OpenAIAsyncClient
 
-    base_llm = openai.LLM(
-        model=LLM_MODEL,
-        client=openai_client,
-    )
-    text_llm = (
-        openai.LLM(model=TEXT_LLM_MODEL, client=openai_client)
-        if TEXT_LLM_MODEL != LLM_MODEL
-        else base_llm
-    )
-    vision_llm = (
-        openai.LLM(model=VISION_LLM_MODEL, client=openai_client)
-        if VISION_LLM_MODEL != LLM_MODEL
-        else base_llm
-    )
+        # OpenAI-compatible endpoints can be used by setting CUSTOM_LLM_BASE_URL.
+        http_client = httpx.AsyncClient(verify=_env_bool("CUSTOM_LLM_VERIFY_SSL", False))
+
+        if LLM_BASE_URL:
+            openai_client = OpenAIAsyncClient(
+                api_key=LLM_API_KEY,
+                base_url=LLM_BASE_URL,
+                http_client=http_client,
+            )
+        else:
+            openai_client = OpenAIAsyncClient(
+                api_key=LLM_API_KEY,
+                http_client=http_client,
+            )
+
+        base_llm = openai.LLM(
+            model=LLM_MODEL,
+            client=openai_client,
+        )
+        text_llm = (
+            openai.LLM(model=TEXT_LLM_MODEL, client=openai_client)
+            if TEXT_LLM_MODEL != LLM_MODEL
+            else base_llm
+        )
+        vision_llm = (
+            openai.LLM(model=VISION_LLM_MODEL, client=openai_client)
+            if VISION_LLM_MODEL != LLM_MODEL
+            else base_llm
+        )
     vision_store = VisionFrameStore(
         max_age_seconds=_env_float("CUSTOM_VISION_FRAME_MAX_AGE_SECONDS", 8.0)
     )
@@ -707,7 +753,7 @@ async def entrypoint(ctx: JobContext) -> None:
     session: AgentSession = AgentSession(
         # 1. Custom ASR blackbox with StreamAdapter
         stt=stt_stream,
-        # 2. OpenAI-compatible LLM, e.g. MiniMax, Qwen, or OpenAI.
+        # 2. LLM backend, OpenAI-compatible or Hermes/OpenClaw gateway.
         llm=base_llm,
         # 3. TTS blackbox
         tts=BlackboxTTS(