feat(omni): 支持 GLM-4.6V 多模态模型（智谱 API）

backend/miloco/src/miloco/admin/router.py

@@ -933,6 +933,7 @@ def _full_omni_payload() -> dict:
             "api_key_masked": _mask_api_key(p.api_key),
             "has_key": bool(p.api_key),
             "active": p.label == active.label,
+            "extra_headers": dict(p.extra_headers or {}),
         }
         for p in m.omni_profiles
     ]
@@ -946,6 +947,7 @@ def _full_omni_payload() -> dict:
                 "api_key_masked": _mask_api_key(active.api_key),
                 "has_key": True,
                 "active": True,
+                "extra_headers": dict(active.extra_headers or {}),
             },
         )
     health = asdict(get_omni_circuit_breaker().snapshot())
@@ -957,6 +959,7 @@ def _full_omni_payload() -> dict:
             "api_key_masked": _mask_api_key(active.api_key),
             "has_key": bool(active.api_key),
             "health": health,
+            "extra_headers": dict(active.extra_headers or {}),
         },
         "profiles": profiles,
     }
@@ -969,6 +972,7 @@ def _profiles_as_dicts() -> list[dict]:
             "model": p.model,
             "base_url": p.base_url,
             "api_key": p.api_key,
+            "extra_headers": dict(p.extra_headers or {}),
         }
         for p in get_settings().model.omni_profiles
     ]
@@ -981,6 +985,13 @@ class OmniConfigBody(BaseModel):
     api_key: str | None = None  # 留空 = 沿用该档案原 key(不被打码值覆盖)
     original_label: str | None = None  # 正在编辑的档案原名(支持改名/定位);None=新增
     activate: bool = True  # True=同时设为当前生效;False=只入列表(激活由 /activate 负责)
+    extra_headers: dict[str, str] = Field(
+        default_factory=dict,
+        description=(
+            "附加请求头（可选）。部分网关要求携带自定义头才走特定通道或计费口径，"
+            "例如智谱 GLM Coding Plan 需要 X-Title 才计入 MCP 通道。"
+        ),
+    )
 
 
 class OmniSelectBody(BaseModel):
@@ -1033,15 +1044,23 @@ async def put_omni_config(
         raise HTTPException(status_code=409, detail=f"档案名「{label}」已存在")
     # 传 base_url 让 _key_by_label 校验"URL 未变才沿用旧 key",防跨 URL 复用凭证。
     key = _key_by_label(orig or label, body.api_key, base_url=base_url)
-    entry = {"label": label, "base_url": base_url, "model": model, "api_key": key}
+    entry = {
+        "label": label,
+        "base_url": base_url,
+        "model": model,
+        "api_key": key,
+        "extra_headers": dict(body.extra_headers or {}),
+    }
     tgt = orig or label
     will_activate = body.activate or _label_is_active(tgt)
     if will_activate:
         if not key:
             raise HTTPException(
                 status_code=400, detail={"code": "no_key", "message": "未配置 API Key"}
             )
-        result = await _probe.probe_omni(model, base_url, key)
+        result = await _probe.probe_omni(
+            model, base_url, key, extra_headers=dict(body.extra_headers or {})
+        )
         if not result.get("ok"):
             raise HTTPException(status_code=400, detail=result)
     if target:
@@ -1081,7 +1100,12 @@ async def activate_omni_config(
                     status_code=400,
                     detail={"code": "no_key", "message": "未配置 API Key"},
                 )
-            result = await _probe.probe_omni(p.model, p.base_url, p.api_key)
+            result = await _probe.probe_omni(
+                p.model,
+                p.base_url,
+                p.api_key,
+                extra_headers=dict(p.extra_headers or {}),
+            )
             if not result.get("ok"):
                 raise HTTPException(status_code=400, detail=result)
             update_shared_config(
@@ -1091,6 +1115,7 @@ async def activate_omni_config(
                         "model": p.model,
                         "base_url": p.base_url,
                         "api_key": p.api_key,
+                        "extra_headers": dict(p.extra_headers or {}),
                     }
                 }
             )
@@ -1188,7 +1213,7 @@ async def test_omni_config(
 ):
     """用表单值（缺省回退当前已保存配置）探测配置可用性。
 
-    OpenAI 兼容族（MiMo/Qwen）两阶段：先 GET /models 验鉴权/可达，再发一次 max_tokens=1 的
+    OpenAI 兼容族（MiMo/Qwen/GLM）两阶段：先 GET /models 验鉴权/可达，再发一次 max_tokens=1 的
     极简 chat 真正验证该模型可用；非 OpenAI 兼容族（Gemini 等原生协议）没有等价 GET /models
     预检语义，直接走 adapter 化的 chat 探测。消耗极少量 token，不计入 miloco 用量统计。
     返回 {ok, code, status, latency_ms, message}。"""
@@ -1208,7 +1233,19 @@ async def test_omni_config(
             message="ok",
             data={"ok": False, "code": "no_key", "message": "未配置 API Key"},
         )
-    result = await _probe.probe_omni(model, base_url, api_key)
+    # 取匹配档案（或当前 active）的 extra_headers，探测链路与推理链路保持同一配置。
+    label = (body.label or "").strip()
+    extra_headers: dict[str, str] = {}
+    if label:
+        for p in get_settings().model.omni_profiles:
+            if p.label == label:
+                extra_headers = dict(p.extra_headers or {})
+                break
+    if not extra_headers:
+        extra_headers = dict(omni.extra_headers or {})
+    result = await _probe.probe_omni(
+        model, base_url, api_key, extra_headers=extra_headers
+    )
     # 测通 + 三元组精确匹配当前 active + 熔断非 ok → 主动清熔断,与 put/activate/retry
     # 恢复路径对齐。护栏:测别的档案 / 未保存的新配置时不动状态。
     # OPEN_CONFIG 下 tick 不会自动探测(只探 OPEN_RECOVERABLE),不清则用户测通了红条仍不消失,
@@ -1275,8 +1312,22 @@ async def list_omni_models(
                 "message": "未配置 API Key",
             },
         )
+    # 取匹配档案（或当前 active）的 extra_headers，探测链路与推理链路保持同一配置。
+    label = (body.label or "").strip()
+    extra_headers: dict[str, str] = {}
+    if label:
+        for p in get_settings().model.omni_profiles:
+            if p.label == label:
+                extra_headers = dict(p.extra_headers or {})
+                break
+    if not extra_headers:
+        extra_headers = dict(get_settings().model.omni.extra_headers or {})
     return NormalResponse(
-        code=0, message="ok", data=await _probe.fetch_models(base_url, api_key)
+        code=0,
+        message="ok",
+        data=await _probe.fetch_models(
+            base_url, api_key, extra_headers=extra_headers
+        ),
     )
 
 
@@ -1401,7 +1452,12 @@ async def retry_omni_probe(current_user: str = Depends(verify_token)):
         return NormalResponse(code=0, message="ok", data=_full_omni_payload())
 
     try:
-        result = await _probe.probe_omni(omni.model, omni.base_url, omni.api_key)
+        result = await _probe.probe_omni(
+            omni.model,
+            omni.base_url,
+            omni.api_key,
+            extra_headers=dict(omni.extra_headers or {}),
+        )
     except asyncio.CancelledError:
         # 客户端断开 HTTP(用户切页/关 tab/网络抖动)时 FastAPI 抛 CancelledError。
         # 此前 retry_now() 已把 state 置 HALF_OPEN,若不复位则 before_call 永久短路、

backend/miloco/src/miloco/config/settings.py

@@ -161,6 +161,14 @@ class OmniModelSettings(BaseModel):
         default="",
         description="多模态模型 API Key；为空时视为未配置，插件与后端启动前校验",
     )
+    extra_headers: dict[str, str] = Field(
+        default_factory=dict,
+        description=(
+            "附加请求头（可选）。部分网关要求携带自定义头才走特定通道或计费口径，"
+            "例如智谱 GLM Coding Plan 需要 {\"X-Title\": \"...\"} 才计入 MCP 通道；"
+            "默认空，是否携带由用户自行决定。"
+        ),
+    )
 
 
 class ModelSettings(BaseModel):
@@ -711,6 +719,7 @@ def _propagate_model_omni_to_perception(self) -> "MilocoSettings":
             "model": self.model.omni.model,
             "base_url": self.model.omni.base_url,
             "api_key": self.model.omni.api_key,
+            "extra_headers": dict(self.model.omni.extra_headers),
         }
         if merged != existing:
             new_engine = {**self.perception.engine, "omni": merged}

backend/miloco/src/miloco/perception/engine/config.py

@@ -24,7 +24,7 @@ class InputConfig:
     # 在构造前设 config.input.video_short_edge 期望生效，那样会被静默忽略。
     video_short_edge: int = 512
     # media_resolution: 仅 Gemini 生效的「每帧视觉 token 预算」档位。""/"low" = 66 tok/帧
-    # (默认、最省);"high" = 264 tok/帧(小目标/文字更清但 4× token)。mimo/qwen 忽略此字段。
+    # (默认、最省);"high" = 264 tok/帧(小目标/文字更清但 4× token)。mimo/qwen/glm 忽略此字段。
     # 实测:小目标清晰度主要由输入像素分辨率(video_short_edge)决定,本档位只控每帧 token 预算,
     # 故默认 low;identity 等细节敏感场景可经 CLI 切 high。运行时由 GeminiAdapter 实时读 settings。
     media_resolution: str = ""
@@ -396,6 +396,9 @@ class OmniConfig:
     top_p: float = 0.95
     timeout: float = 30.0
     stream: bool = False
+    # 附加请求头（可选）。部分网关要求携带自定义头才走特定通道/计费口径
+    # （如智谱 GLM Coding Plan 的 X-Title MCP 流量标识）；默认空。
+    extra_headers: dict[str, str] = field(default_factory=dict)
 
 
 @dataclass

backend/miloco/src/miloco/perception/engine/omni/omni.py

@@ -93,7 +93,7 @@ async def run_omni(
     edge_packet: IdentityPacket, context: OmniContext, config: OmniConfig
 ) -> OmniOutput:
     """Run Omni layer: build prompt → call model → parse response."""
-    payload = build_prompt(edge_packet, context)
+    payload = build_prompt(edge_packet, context, adapter=get_adapter(config.model))
     raw_response = await call_omni(payload, config)
     output = parse_omni_response(raw_response, _rule_name_to_id(context))
     output.usage = extract_usage(raw_response)
@@ -104,7 +104,7 @@ async def run_omni_batch(
     edge_packets: list[IdentityPacket], context: OmniContext, config: OmniConfig
 ) -> OmniOutput:
     """Run Omni layer for multiple devices in the same room."""
-    payload = build_batch_prompt(edge_packets, context)
+    payload = build_batch_prompt(edge_packets, context, adapter=get_adapter(config.model))
     raw_response = await call_omni(payload, config)
     output = parse_omni_response(raw_response, _rule_name_to_id(context))
     output.usage = extract_usage(raw_response)
@@ -325,6 +325,7 @@ async def _call_omni_messages(
         "Content-Type": "application/json",
         **adapter.auth_headers(api_key),
         "User-Agent": MILOCO_USER_AGENT,
+        **(config.extra_headers or {}),
     }
     try:
         await cb.before_call()
@@ -476,7 +477,7 @@ async def run_omni_stream(
     on_early_suggestions: Callable[[list[Suggestion]], Awaitable[None]] | None = None,
 ) -> OmniOutput:
     """Run Omni layer with streaming — extracts actionable fields early via callbacks."""
-    payload = build_stream_prompt(edge_packet, context)
+    payload = build_stream_prompt(edge_packet, context, adapter=get_adapter(config.model))
     return await _stream_and_parse(
         payload,
         config,
@@ -497,7 +498,7 @@ async def run_omni_batch_stream(
     on_early_suggestions: Callable[[list[Suggestion]], Awaitable[None]] | None = None,
 ) -> OmniOutput:
     """Run Omni layer for multiple devices with streaming — extracts actionable fields early."""
-    payload = build_batch_stream_prompt(edge_packets, context)
+    payload = build_batch_stream_prompt(edge_packets, context, adapter=get_adapter(config.model))
     return await _stream_and_parse(
         payload,
         config,

backend/miloco/src/miloco/perception/engine/omni/omni_client.py

@@ -157,6 +157,7 @@ def resolve_live_omni_config(base: OmniConfig) -> OmniConfig:
         model=o.model,
         base_url=o.base_url,
         api_key=o.api_key or base.api_key,
+        extra_headers=dict(o.extra_headers),
     )
     _maybe_reset_breaker_on_config_change(resolved)
     return resolved
@@ -217,6 +218,7 @@ async def call_omni(
         "Content-Type": "application/json",
         **adapter.auth_headers(api_key),
         "User-Agent": MILOCO_USER_AGENT,
+        **(config.extra_headers or {}),
     }
     try:
         await cb.before_call()  # 熔断 OPEN → 直接抛 CircuitOpenError
@@ -362,7 +364,7 @@ def _build_messages(payload: dict, adapter: OmniProviderAdapter) -> list[dict]:
 
     if payload.get("video_base64"):
         content.append(adapter.build_video_block(payload["video_base64"], media_info))
-    elif payload.get("audio_base64"):
+    elif payload.get("audio_base64") and adapter.supports_audio_input:
         content.append(adapter.build_audio_block(payload["audio_base64"], media_info))
 
     # Crop images (from tracker)
@@ -435,6 +437,7 @@ async def call_omni_stream(
         "Content-Type": "application/json",
         **adapter.auth_headers(api_key),
         "User-Agent": MILOCO_USER_AGENT,
+        **(config.extra_headers or {}),
     }
     url = adapter.endpoint(config.base_url, config.model, stream=True)
 

backend/miloco/src/miloco/perception/engine/omni/probe.py

@@ -66,7 +66,9 @@
     return {"code": "http_error", "message": f"服务返回异常（HTTP {r.status_code}）"}
 
 
-async def fetch_models(base_url: str, api_key: str) -> dict[str, Any]:
+async def fetch_models(
+    base_url: str, api_key: str, extra_headers: dict[str, str] | None = None
+) -> dict[str, Any]:
     """拉取 provider 模型列表(GET /models)。
 
     模型下拉在「选定 model 之前」拉取,没有 model 可路由 adapter,故按 base_url 判 provider:
@@ -88,6 +90,7 @@
         if is_gemini
         else {"Authorization": f"Bearer {api_key}"}
     )
+    headers.update(extra_headers or {})
     try:
         async with httpx.AsyncClient(timeout=_TIMEOUT) as client:
             r = await client.get(f"{base}/models", headers=headers)
@@ -179,7 +182,12 @@
         return 500, 0, False, {}
 
 
-async def probe_chat(model: str, base_url: str, api_key: str) -> dict[str, Any]:
+async def probe_chat(
+    model: str,
+    base_url: str,
+    api_key: str,
+    extra_headers: dict[str, str] | None = None,
+) -> dict[str, Any]:
     """极简 chat 探测(max_tokens=1)真校验模型是否可用。
 
     走 provider adapter 生成 body,兼容不同 provider 的强制要求(Qwen 强制
@@ -210,6 +218,7 @@
     headers = {
         **adapter.auth_headers(api_key),
         "Content-Type": "application/json",
+        **(extra_headers or {}),
     }
     t0 = time.monotonic()
     try:
@@ -330,7 +339,12 @@
     }
 
 
-async def probe_omni(model: str, base_url: str, api_key: str) -> dict[str, Any]:
+async def probe_omni(
+    model: str,
+    base_url: str,
+    api_key: str,
+    extra_headers: dict[str, str] | None = None,
+) -> dict[str, Any]:
     """两阶段探测:GET /models 预检 → 极简 chat 真校验。
 
     - GET /models 网络错 → unreachable
@@ -351,12 +365,16 @@
     )
 
     if not isinstance(get_adapter(model), OpenAICompatAdapter):
-        return await probe_chat(model, base, api_key)
+        return await probe_chat(model, base, api_key, extra_headers)
     try:
         async with httpx.AsyncClient(timeout=_TIMEOUT) as client:
             r = await client.get(
-                f"{base}/models", headers={"Authorization": f"Bearer {api_key}"}
+                f"{base}/models",
+                headers={
+                    "Authorization": f"Bearer {api_key}",
+                    **(extra_headers or {}),
+                },
             )
     except Exception as e:  # noqa: BLE001
         return {
             "ok": False,
@@ -377,4 +395,4 @@
             "status": r.status_code,
             "message": f"服务返回异常（HTTP {r.status_code}）",
         }
-    return await probe_chat(model, base, api_key)
+    return await probe_chat(model, base, api_key, extra_headers)