From 3b0b1567c507b8ff322800cea057dd67229fef64 Mon Sep 17 00:00:00 2001
From: homelab <me@cxd.name>
Date: Fri, 31 Jul 2026 16:09:05 +0800
Subject: [PATCH 1/9] =?UTF-8?q?feat(omni):=20=E6=94=AF=E6=8C=81=20GLM-4.6V?=
 =?UTF-8?q?=20=E5=A4=9A=E6=A8=A1=E6=80=81=E6=A8=A1=E5=9E=8B=20(=E6=99=BA?=
 =?UTF-8?q?=E8=B0=B1=20API)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- 新增 GlmAdapter: 继承 MiMoAdapter, 视频块/请求体同构 (video_url + fps +
  media_resolution / thinking:disabled 实测兼容)
- get_adapter 按模型名子串匹配 ("glm" in model.lower()) 路由到 GlmAdapter,
  glm-4.6v / glm-4.6v-flash / zhipu/glm-4.6v 均命中
- GLM-4.6V 是纯视觉模型, 不支持 input_audio (音频降级见后续 commit);
  X-Title 由用户经 model.omni.extra_headers 配置 (见后续 commit)
---
 .../miloco/perception/engine/omni/provider.py | 20 +++++++++++
 .../perception/engine/omni/test_provider.py   | 35 ++++++++++++++++++-
 2 files changed, 54 insertions(+), 1 deletion(-)

diff --git a/backend/miloco/src/miloco/perception/engine/omni/provider.py b/backend/miloco/src/miloco/perception/engine/omni/provider.py
index 2f9dfac91..9f5e7079a 100644
--- a/backend/miloco/src/miloco/perception/engine/omni/provider.py
+++ b/backend/miloco/src/miloco/perception/engine/omni/provider.py
@@ -161,6 +161,20 @@ def build_request_body(
         return body
 
 
+class GlmAdapter(MiMoAdapter):
+    """GLM-4.6V API adapter（智谱大模型开放平台，OpenAI 兼容协议）。
+
+    GLM-4.6V 请求体与 MiMoAdapter 完全同构（video_url + fps + media_resolution、
+    input_audio、thinking:disabled 均接受，实测兼容），唯一差异是智谱网关要求
+    携带 ``X-Title`` 头（不带会被 429 余额不足拦截），故只覆写 auth_headers。
+    """
+
+    def auth_headers(self, api_key: str) -> dict[str, str]:
+        headers = super().auth_headers(api_key)
+        headers["X-Title"] = "4.5V MCP Local"
+        return headers
+
+
 class QwenOmniAdapter(OpenAICompatAdapter):
     """Qwen3.5-Omni 系列 API adapter（qwen3.5-omni-plus / qwen3.5-omni-flash）。
 
@@ -461,6 +475,7 @@ def adjust_fps_for_omni(fps: int, omni_fps: int) -> int:
 
 
 _DEFAULT_ADAPTER = MiMoAdapter()
+_GLM_ADAPTER = GlmAdapter()
 _QWEN_ADAPTER = QwenOmniAdapter()
 _GEMINI_ADAPTER = GeminiAdapter()
 
@@ -471,6 +486,9 @@ def get_adapter(model: str) -> OmniProviderAdapter:
     Qwen 侧仅支持 Qwen3.5-Omni 系列（qwen3.5-omni-plus / qwen3.5-omni-flash），
     旧版 qwen3-omni-flash 不支持多模态组合输入，无法满足 fused 模式需求。
 
+    GLM 走 OpenAI 兼容协议（glm-4.6v 等），请求体与 MiMo 同构，
+    仅鉴权头差异由 GlmAdapter 处理。
+
     Gemini 走原生 generateContent 协议（OpenAI 兼容端点不支持视频输入）。
     """
     name = model.lower()
@@ -478,4 +496,6 @@ def get_adapter(model: str) -> OmniProviderAdapter:
         return _QWEN_ADAPTER
     if "gemini" in name:
         return _GEMINI_ADAPTER
+    if "glm" in name:
+        return _GLM_ADAPTER
     return _DEFAULT_ADAPTER
diff --git a/backend/miloco/tests/perception/engine/omni/test_provider.py b/backend/miloco/tests/perception/engine/omni/test_provider.py
index f526019da..86dd7e9eb 100644
--- a/backend/miloco/tests/perception/engine/omni/test_provider.py
+++ b/backend/miloco/tests/perception/engine/omni/test_provider.py
@@ -3,6 +3,7 @@
 from miloco.perception.engine.omni import provider
 from miloco.perception.engine.omni.provider import (
     GeminiAdapter,
+    GlmAdapter,
     LocalMediaInfo,
     MiMoAdapter,
     OpenAICompatAdapter,
@@ -46,16 +47,24 @@ def test_gemini(self):
     def test_gemini_case_insensitive(self):
         assert isinstance(get_adapter("Gemini-3-Pro"), GeminiAdapter)
 
+    def test_glm(self):
+        assert isinstance(get_adapter("glm-4.6v"), GlmAdapter)
+
+    def test_glm_case_insensitive(self):
+        assert isinstance(get_adapter("GLM-4.6V"), GlmAdapter)
+
     def test_openai_compat_family(self):
-        # MiMo / Qwen 都归 OpenAI 兼容族；Gemini 不是。
+        # MiMo / Qwen / GLM 都归 OpenAI 兼容族；Gemini 不是。
         assert isinstance(get_adapter("xiaomi/mimo-v2.5"), OpenAICompatAdapter)
         assert isinstance(get_adapter("qwen3.5-omni-flash"), OpenAICompatAdapter)
+        assert isinstance(get_adapter("glm-4.6v"), OpenAICompatAdapter)
         assert not isinstance(get_adapter("gemini-3-flash-preview"), OpenAICompatAdapter)
 
     def test_singleton(self):
         assert get_adapter("xiaomi/mimo-v2.5") is get_adapter("xiaomi/mimo-v2.5")
         assert get_adapter("qwen3.5-omni-flash") is get_adapter("qwen3.5-omni-plus")
         assert get_adapter("gemini-3-flash") is get_adapter("gemini-3-pro")
+        assert get_adapter("glm-4.6v") is get_adapter("glm-4.6v")
 
 
 class TestMiMoAdapter:
@@ -92,6 +101,30 @@ def test_request_body_stream(self):
         assert body["thinking"] == {"type": "disabled"}
 
 
+class TestGlmAdapter:
+    adapter = GlmAdapter()
+
+    def test_video_block_same_as_mimo(self):
+        block = self.adapter.build_video_block("AAAA", _VIDEO_MEDIA)
+        assert block["type"] == "video_url"
+        assert block["fps"] == 1
+        assert block["media_resolution"] == "max"
+        assert block["video_url"]["url"].startswith("data:video/mp4;base64,")
+
+    def test_request_body_inherits_mimo(self):
+        body = self.adapter.build_request_body(
+            _MESSAGES, model="glm-4.6v",
+            max_tokens=512, temperature=0.1, top_p=0.95, stream=False,
+        )
+        assert body["thinking"] == {"type": "disabled"}
+        assert body["stream"] is False
+
+    def test_auth_headers_include_x_title(self):
+        headers = self.adapter.auth_headers("sk-test")
+        assert headers["Authorization"] == "Bearer sk-test"
+        assert headers["X-Title"] == "4.5V MCP Local"
+
+
 class TestQwenOmniAdapter:
     adapter = QwenOmniAdapter()
 

From 9ac54969475c8d0d79047be63f2e6c951b20610f Mon Sep 17 00:00:00 2001
From: homelab <me@cxd.name>
Date: Fri, 31 Jul 2026 16:11:42 +0800
Subject: [PATCH 2/9] =?UTF-8?q?docs(omni):=20=E6=B3=A8=E6=98=8E=20GlmAdapt?=
 =?UTF-8?q?er=20=E9=80=82=E9=85=8D=E7=9A=84=E6=98=AF=E6=99=BA=E8=B0=B1=20G?=
 =?UTF-8?q?LM=20Coding=20Plan=20=E7=9A=84=20MCP=20=E6=B5=81=E9=87=8F?=
 =?UTF-8?q?=E9=80=9A=E9=81=93?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

X-Title: 4.5V MCP Local 是 Coding Plan Key 必须携带的标识,
不带会按普通资源包计费并返回 429 余额不足。
---
 backend/miloco/src/miloco/perception/engine/omni/provider.py | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/backend/miloco/src/miloco/perception/engine/omni/provider.py b/backend/miloco/src/miloco/perception/engine/omni/provider.py
index 9f5e7079a..280ceb828 100644
--- a/backend/miloco/src/miloco/perception/engine/omni/provider.py
+++ b/backend/miloco/src/miloco/perception/engine/omni/provider.py
@@ -167,6 +167,10 @@ class GlmAdapter(MiMoAdapter):
     GLM-4.6V 请求体与 MiMoAdapter 完全同构（video_url + fps + media_resolution、
     input_audio、thinking:disabled 均接受，实测兼容），唯一差异是智谱网关要求
     携带 ``X-Title`` 头（不带会被 429 余额不足拦截），故只覆写 auth_headers。
+
+    注意：``X-Title: 4.5V MCP Local`` 是智谱 GLM Coding Plan 的 MCP 流量标识，
+    使用 Coding Plan 签发的 API Key 时必须携带，否则网关按普通资源包计费并
+    返回 429 余额不足。非 Coding Plan 的 GLM 视觉 Key 是否要求该头未验证。
     """
 
     def auth_headers(self, api_key: str) -> dict[str, str]:

From 6c1d07e4a3dea36d12cdbfcd35463b79d146175a Mon Sep 17 00:00:00 2001
From: homelab <me@cxd.name>
Date: Fri, 31 Jul 2026 16:53:24 +0800
Subject: [PATCH 3/9] =?UTF-8?q?fix(omni):=20GLM=20=E4=B8=8D=E6=94=AF?=
 =?UTF-8?q?=E6=8C=81=E9=9F=B3=E9=A2=91=E8=BE=93=E5=85=A5,=20audio-only=20?=
 =?UTF-8?q?=E7=AA=97=E5=8F=A3=E9=99=8D=E7=BA=A7=20text-only=20+=20probe=20?=
 =?UTF-8?q?=E8=A1=A5=20X-Title?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

pr-review 发现的问题修复:

🔴 GLM-4.6V 是纯视觉模型(文本/图片/视频/文件), input_audio 属于
   GLM-4-Voice/GLM-Realtime 线, 发给 4.6V 稳定 400 → 夜间"有声无画面
   变化"窗口会把熔断打进需人工干预的 CONFIG 态:
   - OmniProviderAdapter 新增 supports_audio_input 能力位(默认 True)
   - GlmAdapter 置 False
   - prompt_builder audio route / omni_client on_demand 路径按能力位
     降级为 text-only, 不发音频块

🟡 probe.fetch_models / probe_omni 预检硬编码 Bearer, 漏掉 GLM 的
   X-Title 头(429 → 模型下拉列表拉不出来): 按 hostname 判断补头,
   与 adapter.auth_headers 保持一致

🔵 注释/文档: 3 处 OpenAI 兼容族枚举补 GLM; singleton 测试改用
   glm-4.6v-flash 覆盖同族复用; PR 描述改子串匹配口径

实测: GLM-4.6V 拒 m4a(400 format 不支持), 尝试 wav 也报 base64 格式
错误, 音频输入不可用 → 降级是正确方案
---
 backend/miloco/src/miloco/admin/router.py     |  2 +-
 .../src/miloco/perception/engine/config.py    |  2 +-
 .../perception/engine/omni/omni_client.py     |  2 +-
 .../miloco/perception/engine/omni/probe.py    | 14 ++++++++++-
 .../perception/engine/omni/prompt_builder.py  |  8 +++++-
 .../miloco/perception/engine/omni/provider.py | 25 +++++++++++++------
 .../engine/omni/test_prompt_builder.py        | 15 +++++++++++
 .../perception/engine/omni/test_provider.py   |  7 +++++-
 8 files changed, 62 insertions(+), 13 deletions(-)

diff --git a/backend/miloco/src/miloco/admin/router.py b/backend/miloco/src/miloco/admin/router.py
index 47770ccb1..fd53035ea 100644
--- a/backend/miloco/src/miloco/admin/router.py
+++ b/backend/miloco/src/miloco/admin/router.py
@@ -777,7 +777,7 @@ 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}。"""
diff --git a/backend/miloco/src/miloco/perception/engine/config.py b/backend/miloco/src/miloco/perception/engine/config.py
index 017b904cb..6cb7c9c83 100644
--- a/backend/miloco/src/miloco/perception/engine/config.py
+++ b/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 = ""
diff --git a/backend/miloco/src/miloco/perception/engine/omni/omni_client.py b/backend/miloco/src/miloco/perception/engine/omni/omni_client.py
index 15a1e05eb..bc01bcabc 100644
--- a/backend/miloco/src/miloco/perception/engine/omni/omni_client.py
+++ b/backend/miloco/src/miloco/perception/engine/omni/omni_client.py
@@ -362,7 +362,7 @@ 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)
diff --git a/backend/miloco/src/miloco/perception/engine/omni/probe.py b/backend/miloco/src/miloco/perception/engine/omni/probe.py
index cccf40d2e..3b9117067 100644
--- a/backend/miloco/src/miloco/perception/engine/omni/probe.py
+++ b/backend/miloco/src/miloco/perception/engine/omni/probe.py
@@ -83,11 +83,19 @@ async def fetch_models(base_url: str, api_key: str) -> dict[str, Any]:
     is_gemini = (
         (urlparse(base).hostname or "").lower() == "generativelanguage.googleapis.com"
     )
+    # 智谱 GLM（open.bigmodel.cn / api.z.ai）需带 X-Title 头（Coding Plan MCP 流量
+    # 标识，不带被 429 拦截），与 adapter.auth_headers 保持一致。
+    is_glm = (urlparse(base).hostname or "").lower() in (
+        "open.bigmodel.cn",
+        "api.z.ai",
+    )
     headers = (
         {"x-goog-api-key": api_key}
         if is_gemini
         else {"Authorization": f"Bearer {api_key}"}
     )
+    if is_glm:
+        headers["X-Title"] = "4.5V MCP Local"
     try:
         async with httpx.AsyncClient(timeout=_TIMEOUT) as client:
             r = await client.get(f"{base}/models", headers=headers)
@@ -352,10 +360,14 @@ async def probe_omni(model: str, base_url: str, api_key: str) -> dict[str, Any]:
 
     if not isinstance(get_adapter(model), OpenAICompatAdapter):
         return await probe_chat(model, base, api_key)
+    headers = {"Authorization": f"Bearer {api_key}"}
+    # 智谱 GLM 需带 X-Title 头（Coding Plan MCP 流量标识），与 adapter 一致。
+    if (urlparse(base).hostname or "").lower() in ("open.bigmodel.cn", "api.z.ai"):
+        headers["X-Title"] = "4.5V MCP Local"
     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=headers
             )
     except Exception as e:  # noqa: BLE001
         return {
diff --git a/backend/miloco/src/miloco/perception/engine/omni/prompt_builder.py b/backend/miloco/src/miloco/perception/engine/omni/prompt_builder.py
index 91e1abed8..1f198a3e6 100644
--- a/backend/miloco/src/miloco/perception/engine/omni/prompt_builder.py
+++ b/backend/miloco/src/miloco/perception/engine/omni/prompt_builder.py
@@ -233,7 +233,13 @@ def build_fused_payload(
             user_content.append({"type": "text", "text": f"当前时间: {context.current_time}"})
         if context.room_name:
             user_content.append({"type": "text", "text": f"位置: {context.room_name}"})
-        if audio_b64 and len(audio_b64) >= _MIN_AUDIO_B64_LEN:
+        if not adapter.supports_audio_input:
+            logger.warning(
+                "event=fused_audio_unsupported adapter=%s, provider 不支持 input_audio, "
+                "本窗口降级为 text-only",
+                type(adapter).__name__,
+            )
+        elif audio_b64 and len(audio_b64) >= _MIN_AUDIO_B64_LEN:
             user_content.append(adapter.build_audio_block(audio_b64, _audio_only_media_info(ep.sample_rate)))
         elif audio_b64:
             logger.warning(
diff --git a/backend/miloco/src/miloco/perception/engine/omni/provider.py b/backend/miloco/src/miloco/perception/engine/omni/provider.py
index 280ceb828..782e0d863 100644
--- a/backend/miloco/src/miloco/perception/engine/omni/provider.py
+++ b/backend/miloco/src/miloco/perception/engine/omni/provider.py
@@ -12,7 +12,7 @@
   - parse_response / parse_stream_chunk —— 把 provider 响应反解析回 OpenAI 形态
     ``{choices:[{message:{content}}], usage:{...}}``。
 
-OpenAI 兼容族（MiMo / Qwen）继承 ``OpenAICompatAdapter``，协议方法走默认实现，只覆写各自
+OpenAI 兼容族（MiMo / Qwen / GLM）继承 ``OpenAICompatAdapter``，协议方法走默认实现，只覆写各自
 的 block / body 差异。Gemini 走原生 ``generateContent`` 协议（OpenAI 兼容端点不支持视频输入）。
 """
 
@@ -43,6 +43,12 @@ class LocalMediaInfo:
 
 
 class OmniProviderAdapter(ABC):
+    """provider 是否接受 ``input_audio`` 块。``False`` 时 audio-only 窗口降级为
+    text-only、不发音频块（见 prompt_builder 的 audio route / omni_client 的
+    on_demand 路径）。MiMo / Qwen / Gemini 默认支持，GLM 关闭。
+    """
+
+    supports_audio_input: bool = True
 
     @abstractmethod
     def build_video_block(self, video_base64: str, media: LocalMediaInfo) -> dict[str, Any]:
@@ -164,15 +170,20 @@ def build_request_body(
 class GlmAdapter(MiMoAdapter):
     """GLM-4.6V API adapter（智谱大模型开放平台，OpenAI 兼容协议）。
 
-    GLM-4.6V 请求体与 MiMoAdapter 完全同构（video_url + fps + media_resolution、
-    input_audio、thinking:disabled 均接受，实测兼容），唯一差异是智谱网关要求
-    携带 ``X-Title`` 头（不带会被 429 余额不足拦截），故只覆写 auth_headers。
+    GLM-4.6V 请求体与 MiMoAdapter 基本同构（video_url + fps + media_resolution
+    实测兼容），但**不支持音频输入**——GLM-4.6V 是纯视觉语言模型（输入模态为
+    文本/图片/视频/文件），``input_audio`` 属于 GLM-4-Voice / GLM-Realtime 线，
+    发给 4.6V 会被 400 拒。故 ``supports_audio_input = False``，audio-only 窗口
+    由 prompt_builder 降级为 text-only。
 
-    注意：``X-Title: 4.5V MCP Local`` 是智谱 GLM Coding Plan 的 MCP 流量标识，
-    使用 Coding Plan 签发的 API Key 时必须携带，否则网关按普通资源包计费并
-    返回 429 余额不足。非 Coding Plan 的 GLM 视觉 Key 是否要求该头未验证。
+    鉴权差异：智谱网关要求携带 ``X-Title`` 头（不带会被 429 余额不足拦截），
+    故覆写 auth_headers。该头是智谱 GLM Coding Plan 的 MCP 流量标识，使用
+    Coding Plan 签发的 API Key 时必须携带，否则网关按普通资源包计费并返回
+    429 余额不足。非 Coding Plan 的 GLM 视觉 Key 是否要求该头未验证。
     """
 
+    supports_audio_input = False
+
     def auth_headers(self, api_key: str) -> dict[str, str]:
         headers = super().auth_headers(api_key)
         headers["X-Title"] = "4.5V MCP Local"
diff --git a/backend/miloco/tests/perception/engine/omni/test_prompt_builder.py b/backend/miloco/tests/perception/engine/omni/test_prompt_builder.py
index 51a21bff1..e0a7fa6d3 100644
--- a/backend/miloco/tests/perception/engine/omni/test_prompt_builder.py
+++ b/backend/miloco/tests/perception/engine/omni/test_prompt_builder.py
@@ -552,6 +552,21 @@ def test_audio_route_emits_input_audio_block(self):
         audio_block = next(b for b in user_blocks if b["type"] == "input_audio")
         assert audio_block["input_audio"]["data"].startswith("data:audio/m4a;base64,")
 
+    def test_audio_route_glm_drops_audio_block(self):
+        """GLM 不支持 input_audio：audio-only 窗口降级为 text-only，不发音频块。"""
+        from miloco.perception.engine.omni.omni_client import _build_messages
+        from miloco.perception.engine.omni.provider import GlmAdapter
+
+        ep = _audio_only_packet()
+        payload = build_prompt(ep, OmniContext())
+        messages = _build_messages(payload, GlmAdapter())
+        user_blocks = messages[1]["content"]
+        types = [b["type"] for b in user_blocks]
+
+        assert "input_audio" not in types
+        assert "video_url" not in types
+        assert any(b["type"] == "text" for b in user_blocks)
+
     def test_video_route_emits_video_url_block(self):
         from miloco.perception.engine.omni.omni_client import _build_messages
         from miloco.perception.engine.omni.provider import MiMoAdapter
diff --git a/backend/miloco/tests/perception/engine/omni/test_provider.py b/backend/miloco/tests/perception/engine/omni/test_provider.py
index 86dd7e9eb..6300e3e87 100644
--- a/backend/miloco/tests/perception/engine/omni/test_provider.py
+++ b/backend/miloco/tests/perception/engine/omni/test_provider.py
@@ -64,7 +64,7 @@ def test_singleton(self):
         assert get_adapter("xiaomi/mimo-v2.5") is get_adapter("xiaomi/mimo-v2.5")
         assert get_adapter("qwen3.5-omni-flash") is get_adapter("qwen3.5-omni-plus")
         assert get_adapter("gemini-3-flash") is get_adapter("gemini-3-pro")
-        assert get_adapter("glm-4.6v") is get_adapter("glm-4.6v")
+        assert get_adapter("glm-4.6v") is get_adapter("glm-4.6v-flash")
 
 
 class TestMiMoAdapter:
@@ -124,6 +124,11 @@ def test_auth_headers_include_x_title(self):
         assert headers["Authorization"] == "Bearer sk-test"
         assert headers["X-Title"] == "4.5V MCP Local"
 
+    def test_supports_audio_input_false(self):
+        # GLM-4.6V 是纯视觉模型（文本/图片/视频/文件），input_audio 属于
+        # GLM-4-Voice / GLM-Realtime 线，audio-only 窗口必须降级 text-only。
+        assert self.adapter.supports_audio_input is False
+
 
 class TestQwenOmniAdapter:
     adapter = QwenOmniAdapter()

From 015c5dcb57d6f29728a4c02fab3c385aa00f8483 Mon Sep 17 00:00:00 2001
From: homelab <me@cxd.name>
Date: Fri, 31 Jul 2026 17:23:05 +0800
Subject: [PATCH 4/9] =?UTF-8?q?refactor(omni):=20X-Title=20=E6=94=B9?=
 =?UTF-8?q?=E4=B8=BA=E7=94=A8=E6=88=B7=E9=85=8D=E7=BD=AE=20extra=5Fheaders?=
 =?UTF-8?q?,=20=E4=BF=AE=E5=A4=8D=20prompt=20=E9=9F=B3=E9=A2=91=E5=A3=B0?=
 =?UTF-8?q?=E6=98=8E?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

pr-review 第二轮意见修复:

🔴 GLM Coding Plan 的 X-Title 头是计费口径/服务条款层面的选择,
   不应由 adapter 硬编码且无法关闭:
   - OmniModelSettings / OmniConfig 新增 extra_headers 字段(默认空)
   - GlmAdapter 移除 auth_headers 覆写, 退回纯继承(仅保留
     supports_audio_input = False)
   - omni.py / omni_client.py 三处请求头组装合并 config.extra_headers
   - probe.py 移除硬编码 X-Title, 还原纯净实现

🟡 supports_audio_input 只挡了音频块, 没挡 prompt 里"本轮有音频"
   的声明 — GLM 每个 audio-only 窗口会收到"请转录音频"的 system
   prompt 却没有音频(空烧钱 + 可能脑补 speeches):
   - 新增 _adapter_hears_audio() 助手
   - audio 路由对听不见音频的 provider 退回 video 路由(帧还在)
   - video 路由 has_audio/has_speech 按听力置 False, 既有机制剥掉
     speeches/env_sounds
   - build_prompt 系列 / run_omni 系列透传 adapter

🔵 基类 docstring 修正; 新增 extra_headers 生效/默认测试
   测试: omni + probe + config 398 个全过
---
 backend/miloco/src/miloco/config/settings.py  |  8 +++
 .../src/miloco/perception/engine/config.py    |  3 +
 .../src/miloco/perception/engine/omni/omni.py |  9 +--
 .../perception/engine/omni/omni_client.py     |  2 +
 .../miloco/perception/engine/omni/probe.py    | 14 +----
 .../perception/engine/omni/prompt_builder.py  | 53 ++++++++++++------
 .../miloco/perception/engine/omni/provider.py | 36 +++++-------
 .../engine/omni/test_omni_client_circuit.py   | 56 +++++++++++++++++++
 .../perception/engine/omni/test_provider.py   |  6 +-
 9 files changed, 131 insertions(+), 56 deletions(-)

diff --git a/backend/miloco/src/miloco/config/settings.py b/backend/miloco/src/miloco/config/settings.py
index c8abbe6c7..41a45685c 100644
--- a/backend/miloco/src/miloco/config/settings.py
+++ b/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):
diff --git a/backend/miloco/src/miloco/perception/engine/config.py b/backend/miloco/src/miloco/perception/engine/config.py
index 6cb7c9c83..c49514124 100644
--- a/backend/miloco/src/miloco/perception/engine/config.py
+++ b/backend/miloco/src/miloco/perception/engine/config.py
@@ -348,6 +348,9 @@ 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
diff --git a/backend/miloco/src/miloco/perception/engine/omni/omni.py b/backend/miloco/src/miloco/perception/engine/omni/omni.py
index 331e7bbe4..4e31779cb 100644
--- a/backend/miloco/src/miloco/perception/engine/omni/omni.py
+++ b/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,
diff --git a/backend/miloco/src/miloco/perception/engine/omni/omni_client.py b/backend/miloco/src/miloco/perception/engine/omni/omni_client.py
index bc01bcabc..7ac57e3f8 100644
--- a/backend/miloco/src/miloco/perception/engine/omni/omni_client.py
+++ b/backend/miloco/src/miloco/perception/engine/omni/omni_client.py
@@ -217,6 +217,7 @@ 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
@@ -435,6 +436,7 @@ 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)
 
diff --git a/backend/miloco/src/miloco/perception/engine/omni/probe.py b/backend/miloco/src/miloco/perception/engine/omni/probe.py
index 3b9117067..cccf40d2e 100644
--- a/backend/miloco/src/miloco/perception/engine/omni/probe.py
+++ b/backend/miloco/src/miloco/perception/engine/omni/probe.py
@@ -83,19 +83,11 @@ async def fetch_models(base_url: str, api_key: str) -> dict[str, Any]:
     is_gemini = (
         (urlparse(base).hostname or "").lower() == "generativelanguage.googleapis.com"
     )
-    # 智谱 GLM（open.bigmodel.cn / api.z.ai）需带 X-Title 头（Coding Plan MCP 流量
-    # 标识，不带被 429 拦截），与 adapter.auth_headers 保持一致。
-    is_glm = (urlparse(base).hostname or "").lower() in (
-        "open.bigmodel.cn",
-        "api.z.ai",
-    )
     headers = (
         {"x-goog-api-key": api_key}
         if is_gemini
         else {"Authorization": f"Bearer {api_key}"}
     )
-    if is_glm:
-        headers["X-Title"] = "4.5V MCP Local"
     try:
         async with httpx.AsyncClient(timeout=_TIMEOUT) as client:
             r = await client.get(f"{base}/models", headers=headers)
@@ -360,14 +352,10 @@ async def probe_omni(model: str, base_url: str, api_key: str) -> dict[str, Any]:
 
     if not isinstance(get_adapter(model), OpenAICompatAdapter):
         return await probe_chat(model, base, api_key)
-    headers = {"Authorization": f"Bearer {api_key}"}
-    # 智谱 GLM 需带 X-Title 头（Coding Plan MCP 流量标识），与 adapter 一致。
-    if (urlparse(base).hostname or "").lower() in ("open.bigmodel.cn", "api.z.ai"):
-        headers["X-Title"] = "4.5V MCP Local"
     try:
         async with httpx.AsyncClient(timeout=_TIMEOUT) as client:
             r = await client.get(
-                f"{base}/models", headers=headers
+                f"{base}/models", headers={"Authorization": f"Bearer {api_key}"}
             )
     except Exception as e:  # noqa: BLE001
         return {
diff --git a/backend/miloco/src/miloco/perception/engine/omni/prompt_builder.py b/backend/miloco/src/miloco/perception/engine/omni/prompt_builder.py
index 1f198a3e6..e932c6153 100644
--- a/backend/miloco/src/miloco/perception/engine/omni/prompt_builder.py
+++ b/backend/miloco/src/miloco/perception/engine/omni/prompt_builder.py
@@ -101,43 +101,49 @@ def build_prompt(
     identity_packet: IdentityPacket,
     context: OmniContext,
     label_lookup: "dict[str, str] | None" = None,
+    adapter: "OmniProviderAdapter | None" = None,
 ) -> dict:
     """Build the prompt payload for the omni model (single device).
 
     Args:
         label_lookup: person_id (UUID) → 姓名/标签 反查表，渲染 "已识别人物" 段时把
                       UUID 替换为人名。None 时直接渲染 person_id 字段值（与旧行为兼容）。
+        adapter: 感知 provider adapter；听不见音频的 provider（如 GLM）会把
+                 audio 路由退回 video 路由，且 prompt 不声称本轮有音频。
 
     Returns dict with keys: system_prompt, user_content, video_base64, media_info, crops.
     """
-    return _build_payload([identity_packet], context, stream=False, label_lookup=label_lookup)
+    return _build_payload([identity_packet], context, stream=False, label_lookup=label_lookup, adapter=adapter)
 
 
 def build_batch_prompt(
     identity_packets: list[IdentityPacket],
     context: OmniContext,
     label_lookup: "dict[str, str] | None" = None,
+    adapter: "OmniProviderAdapter | None" = None,
 ) -> dict:
     """Build the prompt payload for multi-device omni inference (same room)."""
-    return _build_payload(identity_packets, context, stream=False, label_lookup=label_lookup)
+    return _build_payload(identity_packets, context, stream=False, label_lookup=label_lookup, adapter=adapter)
 
 
 def build_stream_prompt(
     identity_packet: IdentityPacket,
     context: OmniContext,
     label_lookup: "dict[str, str] | None" = None,
+    adapter: "OmniProviderAdapter | None" = None,
 ) -> dict:
     """Build prompt payload for streaming omni call (single device, speeches first)."""
-    return _build_payload([identity_packet], context, stream=True, label_lookup=label_lookup)
+    return _build_payload([identity_packet], context, stream=True, label_lookup=label_lookup, adapter=adapter)
 
 
 def build_batch_stream_prompt(
     identity_packets: list[IdentityPacket],
     context: OmniContext,
     label_lookup: "dict[str, str] | None" = None,
+    adapter: "OmniProviderAdapter | None" = None,
 ) -> dict:
     """Build prompt payload for streaming omni call (multi-device, speeches first)."""
-    return _build_payload(identity_packets, context, stream=True, label_lookup=label_lookup)
+    return _build_payload(identity_packets, context, stream=True, label_lookup=label_lookup, adapter=adapter)
 
 
 def build_query_prompt(
@@ -223,7 +229,9 @@ def build_fused_payload(
     # audio route：无视觉信息，候选作废。与 video 同款 message 隔离（待判断规则/只读历史
     # 各自独立 user 消息）；本轮事实只放"当前时间 + 音频"——audio 无视频，不渲染名册/gallery/
     # 待识别 track（名册的 bbox 是为"把姓名对应到视频里的人"，audio 场景无意义）。
-    if _resolve_route(packets) == "audio":
+    # provider 听不见音频（如 GLM 纯视觉模型）时退回 video 路由：帧本来就在（只是没变化），
+    # 至少还有画面这份真证据，且 prompt 不会声称"本轮有音频"。
+    if _resolve_route(packets) == "audio" and _adapter_hears_audio(adapter):
         scene = SceneDescriptor(route="audio", has_identity=False, stream=False)
         system_prompt = build_system_prompt(scene, include_home_profile=False, camera_prompt=context.camera_prompt)
         ep = packets[0]
@@ -233,13 +241,7 @@ def build_fused_payload(
             user_content.append({"type": "text", "text": f"当前时间: {context.current_time}"})
         if context.room_name:
             user_content.append({"type": "text", "text": f"位置: {context.room_name}"})
-        if not adapter.supports_audio_input:
-            logger.warning(
-                "event=fused_audio_unsupported adapter=%s, provider 不支持 input_audio, "
-                "本窗口降级为 text-only",
-                type(adapter).__name__,
-            )
-        elif audio_b64 and len(audio_b64) >= _MIN_AUDIO_B64_LEN:
+        if audio_b64 and len(audio_b64) >= _MIN_AUDIO_B64_LEN:
             user_content.append(adapter.build_audio_block(audio_b64, _audio_only_media_info(ep.sample_rate)))
         elif audio_b64:
             logger.warning(
@@ -264,10 +266,13 @@ def build_fused_payload(
     # has_speech 只由本轮 VAD 决定：本轮真有人声（含 pending 的延续语音）→ VAD 自然过、
     # 保留 speeches、模型把 <pending_speech> 拼成完整句；本轮无人声 → 剥 speeches，挂着的
     # pending 半句不强行补全（否则模型会就着噪声脑补出一个完成句，正是要根除的幻觉）。
+    # provider 听不见音频时 has_audio/has_speech 一并置 False：prompt 不声称本轮有音频，
+    # 既有机制会把 speeches / env_sounds 从 schema 和任务清单里剥掉。
+    _hears = _adapter_hears_audio(adapter)
     scene = SceneDescriptor(
         route="video", has_identity=bool(candidates), stream=False,
-        has_audio=_batch_video_has_audio(packets),
-        has_speech=_batch_video_has_speech(packets),
+        has_audio=_batch_video_has_audio(packets) and _hears,
+        has_speech=_batch_video_has_speech(packets) and _hears,
         identity_match_disabled=matching_moot,
     )
     system_prompt = build_system_prompt(scene, include_home_profile=False, camera_prompt=context.camera_prompt)
@@ -370,16 +375,22 @@ def _build_payload(
     stream: bool,
     label_lookup: "dict[str, str] | None" = None,
     include_home_profile: bool = True,
+    adapter: "OmniProviderAdapter | None" = None,
 ) -> dict:
     route = _resolve_route(packets)
+    # provider 听不见音频（如 GLM 纯视觉模型）时，audio 路由退回 video 路由：
+    # 帧本来就在（只是没变化），至少还有画面这份真证据，且 prompt 不会声称"本轮有音频"。
+    if route == "audio" and not _adapter_hears_audio(adapter):
+        route = "video"
     # has_audio：video 路由下音频未过 gate 时为 False → schema 剥掉 speeches/env_sounds，
     # 避免模型就着画面脑补人声。audio 路由恒有音频。
     # has_speech：video 路由下 VAD 判无人声时为 False → 只剥 speeches、保留 env_sounds。
-    has_audio = True if route == "audio" else _batch_video_has_audio(packets)
+    _hears = _adapter_hears_audio(adapter)
+    has_audio = True if route == "audio" else (_batch_video_has_audio(packets) and _hears)
     # has_speech 只由本轮 VAD 决定：本轮真有人声（含 pending 的延续语音）→ VAD 自然过、
     # 拼接照常；本轮无人声 → 剥 speeches，挂着的 pending 半句不强行补全（否则模型会就着
     # 噪声脑补出完成句，正是要根除的幻觉）。
-    has_speech = True if route == "audio" else _batch_video_has_speech(packets)
+    has_speech = True if route == "audio" else (_batch_video_has_speech(packets) and _hears)
     scene = SceneDescriptor(
         route=route, has_identity=False, stream=stream,
         has_audio=has_audio, has_speech=has_speech,
@@ -1123,6 +1134,16 @@ def _audio_only_media_info(sample_rate: int) -> LocalMediaInfo:
     )
 
 
+def _adapter_hears_audio(adapter: "OmniProviderAdapter | None") -> bool:
+    """provider 能不能听见音频。
+
+    决定的是「prompt 敢不敢声称本轮有音频」，比 supports_audio_input（只管
+    input_audio 块发不发）覆盖面更广：video 路由的 mp4 音轨同样听不见。
+    adapter 为 None（旧调用路径）时按支持处理，保持行为不变。
+    """
+    return adapter is None or adapter.supports_audio_input
+
+
 def _get_video_short_edge() -> int:
     try:
         from miloco.config import get_settings
diff --git a/backend/miloco/src/miloco/perception/engine/omni/provider.py b/backend/miloco/src/miloco/perception/engine/omni/provider.py
index 782e0d863..b2ee4dd36 100644
--- a/backend/miloco/src/miloco/perception/engine/omni/provider.py
+++ b/backend/miloco/src/miloco/perception/engine/omni/provider.py
@@ -43,11 +43,11 @@ class LocalMediaInfo:
 
 
 class OmniProviderAdapter(ABC):
-    """provider 是否接受 ``input_audio`` 块。``False`` 时 audio-only 窗口降级为
-    text-only、不发音频块（见 prompt_builder 的 audio route / omni_client 的
-    on_demand 路径）。MiMo / Qwen / Gemini 默认支持，GLM 关闭。
-    """
+    """各 provider 协议适配的抽象基类（见模块 docstring 的 adapter 职责清单）。"""
 
+    # provider 是否接受 ``input_audio`` 块。``False`` 时 audio-only 窗口降级为
+    # text-only、不发音频块（见 prompt_builder 的 audio route / omni_client 的
+    # on_demand 路径）。MiMo / Qwen / Gemini 默认支持，GLM 关闭。
     supports_audio_input: bool = True
 
     @abstractmethod
@@ -168,27 +168,21 @@ def build_request_body(
 
 
 class GlmAdapter(MiMoAdapter):
-    """GLM-4.6V API adapter（智谱大模型开放平台，OpenAI 兼容协议）。
-
-    GLM-4.6V 请求体与 MiMoAdapter 基本同构（video_url + fps + media_resolution
-    实测兼容），但**不支持音频输入**——GLM-4.6V 是纯视觉语言模型（输入模态为
-    文本/图片/视频/文件），``input_audio`` 属于 GLM-4-Voice / GLM-Realtime 线，
-    发给 4.6V 会被 400 拒。故 ``supports_audio_input = False``，audio-only 窗口
-    由 prompt_builder 降级为 text-only。
-
-    鉴权差异：智谱网关要求携带 ``X-Title`` 头（不带会被 429 余额不足拦截），
-    故覆写 auth_headers。该头是智谱 GLM Coding Plan 的 MCP 流量标识，使用
-    Coding Plan 签发的 API Key 时必须携带，否则网关按普通资源包计费并返回
-    429 余额不足。非 Coding Plan 的 GLM 视觉 Key 是否要求该头未验证。
+    """GLM-4.6V API adapter（智谱开放平台，OpenAI 兼容协议）。
+
+    鉴权与请求体与 MiMoAdapter 一致（video_url + fps + media_resolution、
+    thinking:disabled 实测兼容）；**不支持音频输入**——GLM-4.6V 是纯视觉语言
+    模型（输入模态为文本/图片/视频/文件），``input_audio`` 属于 GLM-4-Voice /
+    GLM-Realtime 线，发给 4.6V 会被 400 拒。故 ``supports_audio_input = False``，
+    audio-only 窗口由 prompt_builder 降级为 text-only。
+
+    若使用智谱 GLM Coding Plan 签发的 Key，需要携带 ``X-Title`` MCP 流量标识头
+    （否则网关按普通资源包计费并返回 429 余额不足）——请通过配置的
+    ``model.omni.extra_headers`` 自行添加，本 adapter 不做硬编码。
     """
 
     supports_audio_input = False
 
-    def auth_headers(self, api_key: str) -> dict[str, str]:
-        headers = super().auth_headers(api_key)
-        headers["X-Title"] = "4.5V MCP Local"
-        return headers
-
 
 class QwenOmniAdapter(OpenAICompatAdapter):
     """Qwen3.5-Omni 系列 API adapter（qwen3.5-omni-plus / qwen3.5-omni-flash）。
diff --git a/backend/miloco/tests/perception/engine/omni/test_omni_client_circuit.py b/backend/miloco/tests/perception/engine/omni/test_omni_client_circuit.py
index 31d96787d..3203a8097 100644
--- a/backend/miloco/tests/perception/engine/omni/test_omni_client_circuit.py
+++ b/backend/miloco/tests/perception/engine/omni/test_omni_client_circuit.py
@@ -84,6 +84,62 @@ def _payload() -> dict:
     return {"system_prompt": "sys", "user_content": "u"}
 
 
+def _capture_client(captured: dict):
+    """捕获请求头/body 的 fake AsyncClient。"""
+
+    class _C:
+        def __init__(self, *a, **k):
+            pass
+
+        async def __aenter__(self):
+            return self
+
+        async def __aexit__(self, *a):
+            return False
+
+        async def post(self, url, headers=None, json=None, **k):
+            captured["url"] = url
+            captured["headers"] = headers or {}
+            captured["json"] = json or {}
+            return _FakeResp(200, {"choices": [], "usage": {}})
+
+    return _C
+
+
+async def test_call_omni_sends_extra_headers(monkeypatch):
+    """model.omni.extra_headers 应合并进每个请求头（GLM Coding Plan X-Title 用例）。"""
+    import miloco.perception.engine.omni.omni_client as oc
+
+    captured: dict = {}
+    monkeypatch.setattr(oc.httpx, "AsyncClient", _capture_client(captured))
+    cfg = OmniConfig(
+        model="glm-4.6v",
+        base_url="https://open.bigmodel.cn/api/paas/v4",
+        api_key="sk-glm",
+        temperature=0,
+        top_p=1,
+        max_completion_tokens=1,
+        timeout=1.0,
+        extra_headers={"X-Title": "4.5V MCP Local"},
+    )
+    await oc.call_omni(_payload(), cfg)
+
+    assert captured["headers"]["Authorization"] == "Bearer sk-glm"
+    assert captured["headers"]["X-Title"] == "4.5V MCP Local"
+
+
+async def test_call_omni_no_extra_headers_by_default(monkeypatch):
+    """未配置 extra_headers 时请求头不含额外自定义头。"""
+    import miloco.perception.engine.omni.omni_client as oc
+
+    captured: dict = {}
+    monkeypatch.setattr(oc.httpx, "AsyncClient", _capture_client(captured))
+    await oc.call_omni(_payload(), _cfg())
+
+    assert captured["headers"]["Authorization"] == "Bearer sk-1"
+    assert "X-Title" not in captured["headers"]
+
+
 # ─── call_omni × 熔断 ───────────────────────────────────────────────────────
 
 
diff --git a/backend/miloco/tests/perception/engine/omni/test_provider.py b/backend/miloco/tests/perception/engine/omni/test_provider.py
index 6300e3e87..c5b9f4bee 100644
--- a/backend/miloco/tests/perception/engine/omni/test_provider.py
+++ b/backend/miloco/tests/perception/engine/omni/test_provider.py
@@ -119,10 +119,12 @@ def test_request_body_inherits_mimo(self):
         assert body["thinking"] == {"type": "disabled"}
         assert body["stream"] is False
 
-    def test_auth_headers_include_x_title(self):
+    def test_auth_headers_plain_bearer(self):
+        # GLM 鉴权就是标准 Bearer；X-Title 这类额外头由用户经
+        # model.omni.extra_headers 配置，adapter 不硬编码。
         headers = self.adapter.auth_headers("sk-test")
         assert headers["Authorization"] == "Bearer sk-test"
-        assert headers["X-Title"] == "4.5V MCP Local"
+        assert "X-Title" not in headers
 
     def test_supports_audio_input_false(self):
         # GLM-4.6V 是纯视觉模型（文本/图片/视频/文件），input_audio 属于

From e4094d1a7ec9fa2cc84f34ca8a264b1be2e5a27f Mon Sep 17 00:00:00 2001
From: homelab <me@cxd.name>
Date: Fri, 31 Jul 2026 17:25:45 +0800
Subject: [PATCH 5/9] =?UTF-8?q?fix(config):=20model.omni.extra=5Fheaders?=
 =?UTF-8?q?=20=E4=B8=8B=E6=8E=A8=E5=88=B0=20perception.engine?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

_propagate_model_omni_to_perception 下推时漏了 extra_headers 字段,
OmniConfig 拿不到用户配置的附加请求头(X-Title 等)
---
 backend/miloco/src/miloco/config/settings.py | 1 +
 1 file changed, 1 insertion(+)

diff --git a/backend/miloco/src/miloco/config/settings.py b/backend/miloco/src/miloco/config/settings.py
index 41a45685c..a6d3213af 100644
--- a/backend/miloco/src/miloco/config/settings.py
+++ b/backend/miloco/src/miloco/config/settings.py
@@ -668,6 +668,7 @@ 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}

From 0c12793aca57d872afe815b53427f6278450e5c8 Mon Sep 17 00:00:00 2001
From: homelab <me@cxd.name>
Date: Fri, 31 Jul 2026 17:58:04 +0800
Subject: [PATCH 6/9] =?UTF-8?q?fix(omni):=20extra=5Fheaders=20=E7=A9=BF?=
 =?UTF-8?q?=E5=88=B0=E6=8E=A2=E6=B5=8B=E9=93=BE=E8=B7=AF=20+=20=E7=83=AD?=
 =?UTF-8?q?=E6=9B=B4=E6=96=B0=20+=20=E9=9F=B3=E8=BD=A8=E7=BC=96=E7=A0=81?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

pr-review 第三轮意见修复:

🔴 探测链路(probe.py)没接 extra_headers — Coding Plan key 的探测
   请求 100% 429, 导致: 配置存不进去 / 模型下拉为空 / 熔断后自愈
   永久失效 / 手动重试也救不回来:
   - fetch_models / probe_chat / probe_omni 各加 extra_headers 参数
   - router.py 5 处调用点传入(OmniConfigBody 新增 extra_headers 字段,
     按 label 查档案 extra_headers, 缺省用当前 active)

🟡 resolve_live_omni_config 热更新漏 extra_headers — 改配置后
   请求头不生效, 需重启进程:
   - replace() 补 extra_headers=dict(o.extra_headers)

🟡 音频路降级零测试覆盖 — 新增 TestGlmAudioFallback 4 用例:
   audio-only 退回 video 路由 / MiMo 保持音频路 / fused 路径降级 /
   video 路由 GLM 剥掉 speeches 字段

🔵 注释口径修正(降级 text-only → 退回 video 路由) + get_adapter
   docstring 修正(鉴权头相同, 唯一差异是 supports_audio_input)

🔵 mp4 音轨编码与 has_audio 同判据 — _packet_audio_included /
   _encode_video / _encode_batch_video 传 adapter, GLM 视频路不混
   音轨, 避免白传模型解码不了的 AAC

测试: omni + config 相关 395 个全过
---
 backend/miloco/src/miloco/admin/router.py     | 53 +++++++++++++++---
 .../perception/engine/omni/omni_client.py     |  1 +
 .../miloco/perception/engine/omni/probe.py    | 28 ++++++++--
 .../perception/engine/omni/prompt_builder.py  | 36 +++++++++----
 .../miloco/perception/engine/omni/provider.py | 17 +++---
 .../engine/omni/test_omni_client_circuit.py   |  1 +
 .../engine/omni/test_prompt_builder.py        | 54 +++++++++++++++++++
 7 files changed, 163 insertions(+), 27 deletions(-)

diff --git a/backend/miloco/src/miloco/admin/router.py b/backend/miloco/src/miloco/admin/router.py
index fd53035ea..226db6af2 100644
--- a/backend/miloco/src/miloco/admin/router.py
+++ b/backend/miloco/src/miloco/admin/router.py
@@ -570,6 +570,13 @@ 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):
@@ -622,7 +629,13 @@ 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:
@@ -630,7 +643,9 @@ async def put_omni_config(
             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:
@@ -670,7 +685,12 @@ 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(
@@ -797,7 +817,9 @@ 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)
+    result = await _probe.probe_omni(
+        model, base_url, api_key, extra_headers=dict(omni.extra_headers or {})
+    )
     # 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,
@@ -864,8 +886,22 @@ 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
+        ),
     )
 
 
@@ -930,7 +966,12 @@ 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、
diff --git a/backend/miloco/src/miloco/perception/engine/omni/omni_client.py b/backend/miloco/src/miloco/perception/engine/omni/omni_client.py
index 7ac57e3f8..37d0ef6ab 100644
--- a/backend/miloco/src/miloco/perception/engine/omni/omni_client.py
+++ b/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
diff --git a/backend/miloco/src/miloco/perception/engine/omni/probe.py b/backend/miloco/src/miloco/perception/engine/omni/probe.py
index cccf40d2e..87716f521 100644
--- a/backend/miloco/src/miloco/perception/engine/omni/probe.py
+++ b/backend/miloco/src/miloco/perception/engine/omni/probe.py
@@ -66,7 +66,9 @@ async def probe_reachable(base_url: str) -> dict | None:
     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 @@ async def fetch_models(base_url: str, api_key: str) -> dict[str, Any]:
         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 @@ async def _probe_stream_chat(
         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 @@ async def probe_chat(model: str, base_url: str, api_key: str) -> dict[str, Any]:
     headers = {
         **adapter.auth_headers(api_key),
         "Content-Type": "application/json",
+        **(extra_headers or {}),
     }
     t0 = time.monotonic()
     try:
@@ -330,7 +339,12 @@ async def probe_chat(model: str, base_url: str, api_key: str) -> dict[str, Any]:
     }
 
 
-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
@@ -351,11 +365,15 @@ async def probe_omni(model: str, base_url: str, api_key: str) -> dict[str, Any]:
     )
 
     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 {
diff --git a/backend/miloco/src/miloco/perception/engine/omni/prompt_builder.py b/backend/miloco/src/miloco/perception/engine/omni/prompt_builder.py
index e932c6153..f5cf2d92b 100644
--- a/backend/miloco/src/miloco/perception/engine/omni/prompt_builder.py
+++ b/backend/miloco/src/miloco/perception/engine/omni/prompt_builder.py
@@ -261,7 +261,9 @@ def build_fused_payload(
         }
 
     short_edge = _get_video_short_edge()
-    video_b64, media_info = _encode_batch_video(packets, short_edge=short_edge)
+    video_b64, media_info = _encode_batch_video(
+        packets, short_edge=short_edge, adapter=adapter
+    )
 
     # has_speech 只由本轮 VAD 决定：本轮真有人声（含 pending 的延续语音）→ VAD 自然过、
     # 保留 speeches、模型把 <pending_speech> 拼成完整句；本轮无人声 → 剥 speeches，挂着的
@@ -271,7 +273,7 @@ def build_fused_payload(
     _hears = _adapter_hears_audio(adapter)
     scene = SceneDescriptor(
         route="video", has_identity=bool(candidates), stream=False,
-        has_audio=_batch_video_has_audio(packets) and _hears,
+        has_audio=_batch_video_has_audio(packets, adapter) and _hears,
         has_speech=_batch_video_has_speech(packets) and _hears,
         identity_match_disabled=matching_moot,
     )
@@ -386,7 +388,7 @@ def _build_payload(
     # 避免模型就着画面脑补人声。audio 路由恒有音频。
     # has_speech：video 路由下 VAD 判无人声时为 False → 只剥 speeches、保留 env_sounds。
     _hears = _adapter_hears_audio(adapter)
-    has_audio = True if route == "audio" else (_batch_video_has_audio(packets) and _hears)
+    has_audio = True if route == "audio" else (_batch_video_has_audio(packets, adapter) and _hears)
     # has_speech 只由本轮 VAD 决定：本轮真有人声（含 pending 的延续语音）→ VAD 自然过、
     # 拼接照常；本轮无人声 → 剥 speeches，挂着的 pending 半句不强行补全（否则模型会就着
     # 噪声脑补出完成句，正是要根除的幻觉）。
@@ -409,7 +411,9 @@ def _build_payload(
         base["media_info"] = _audio_only_media_info(ep.sample_rate)
     else:
         short_edge = _get_video_short_edge()
-        video_b64, media_info = _encode_batch_video(packets, short_edge=short_edge)
+        video_b64, media_info = _encode_batch_video(
+            packets, short_edge=short_edge, adapter=adapter
+        )
         base["video_base64"] = video_b64
         base["media_info"] = media_info
     return base
@@ -1169,14 +1173,24 @@ def _get_video_short_edge() -> int:
 _AUDIO_ONLY_ENABLED = True
 
 
-def _packet_audio_included(ep: IdentityPacket) -> bool:
+def _packet_audio_included(
+    ep: IdentityPacket,
+    adapter: "OmniProviderAdapter | None" = None,
+) -> bool:
     """该 packet 的音频是否会被合成进 mp4：audio gate 通过即带（trigger=None 视为通过，
-    兼容主动查询 / 旧路径）。speeches / env_sounds 字段的取舍与此一致——没喂音频就别问。"""
+    兼容主动查询 / 旧路径）。speeches / env_sounds 字段的取舍与此一致——没喂音频就别问。
+    provider 听不见音频（如 GLM 纯视觉模型）时不合成音轨：mp4 里混一段模型解码不了的
+    AAC 是白花的带宽和 token 前处理开销，且与 prompt 里 has_audio=False 自相矛盾。"""
+    if not _adapter_hears_audio(adapter):
+        return False
     trig = ep.trigger
     return trig is None or trig.audio_active
 
 
-def _batch_video_has_audio(packets: list[IdentityPacket]) -> bool:
+def _batch_video_has_audio(
+    packets: list[IdentityPacket],
+    adapter: "OmniProviderAdapter | None" = None,
+) -> bool:
     """video 路由最终合进 mp4 的音频是否存在。
 
     与 ``_encode_batch_video`` 选设备口径一致（首个有 frames 的 device），据该 device 的
@@ -1185,7 +1199,7 @@ def _batch_video_has_audio(packets: list[IdentityPacket]) -> bool:
     """
     for ep in packets:
         if ep.all_frames:
-            return _packet_audio_included(ep)
+            return _packet_audio_included(ep, adapter)
     return False
 
 
@@ -1208,6 +1222,7 @@ def _batch_video_has_speech(packets: list[IdentityPacket]) -> bool:
 def _encode_video(
     identity_packet: IdentityPacket,
     short_edge: int = _VIDEO_SHORT_EDGE,
+    adapter: "OmniProviderAdapter | None" = None,
 ) -> tuple[str | None, LocalMediaInfo | None]:
     """Encode all frames + audio into mp4 video, return ``(base64, media_info)``。"""
     frames = identity_packet.all_frames
@@ -1216,7 +1231,7 @@ def _encode_video(
 
     audio = (
         identity_packet.audio_clip
-        if _packet_audio_included(identity_packet)
+        if _packet_audio_included(identity_packet, adapter)
         else np.empty(0, dtype=np.int16)
     )
     return _encode_video_mp4(
@@ -1440,6 +1455,7 @@ def _encode_audio_only_mp4(
 def _encode_batch_video(
     edge_packets: list[IdentityPacket],
     short_edge: int = _VIDEO_SHORT_EDGE,
+    adapter: "OmniProviderAdapter | None" = None,
 ) -> tuple[str | None, LocalMediaInfo | None]:
     """Encode video from the first device that has frames.
 
@@ -1447,7 +1463,7 @@ def _encode_batch_video(
     返回 ``(base64_str, media_info)``。
     """
     for ep in edge_packets:
-        b64, media_info = _encode_video(ep, short_edge=short_edge)
+        b64, media_info = _encode_video(ep, short_edge=short_edge, adapter=adapter)
         if b64 is not None:
             return b64, media_info
     return None, None
diff --git a/backend/miloco/src/miloco/perception/engine/omni/provider.py b/backend/miloco/src/miloco/perception/engine/omni/provider.py
index b2ee4dd36..48efcebe7 100644
--- a/backend/miloco/src/miloco/perception/engine/omni/provider.py
+++ b/backend/miloco/src/miloco/perception/engine/omni/provider.py
@@ -45,9 +45,13 @@ class LocalMediaInfo:
 class OmniProviderAdapter(ABC):
     """各 provider 协议适配的抽象基类（见模块 docstring 的 adapter 职责清单）。"""
 
-    # provider 是否接受 ``input_audio`` 块。``False`` 时 audio-only 窗口降级为
-    # text-only、不发音频块（见 prompt_builder 的 audio route / omni_client 的
-    # on_demand 路径）。MiMo / Qwen / Gemini 默认支持，GLM 关闭。
+    # provider 是否接受 ``input_audio`` 块。``False`` 时：
+    #   1. audio-only 窗口退回 video 路由——帧本来就在（只是没变化），改送画面
+    #      而非降级为 text-only（见 prompt_builder._build_payload / build_fused_payload）；
+    #   2. video 路由下 has_audio / has_speech 一并置 False，schema 剥掉
+    #      speeches / env_sounds，prompt 不声称本轮有音频；
+    #   3. messages 组装层兜底不拼 audio 块（omni_client._build_messages）。
+    # MiMo / Qwen / Gemini 默认支持，GLM 关闭。
     supports_audio_input: bool = True
 
     @abstractmethod
@@ -174,7 +178,7 @@ class GlmAdapter(MiMoAdapter):
     thinking:disabled 实测兼容）；**不支持音频输入**——GLM-4.6V 是纯视觉语言
     模型（输入模态为文本/图片/视频/文件），``input_audio`` 属于 GLM-4-Voice /
     GLM-Realtime 线，发给 4.6V 会被 400 拒。故 ``supports_audio_input = False``，
-    audio-only 窗口由 prompt_builder 降级为 text-only。
+    audio-only 窗口退回 video 路由（帧本来就在，改送画面而非降级为 text-only）。
 
     若使用智谱 GLM Coding Plan 签发的 Key，需要携带 ``X-Title`` MCP 流量标识头
     （否则网关按普通资源包计费并返回 429 余额不足）——请通过配置的
@@ -495,8 +499,9 @@ def get_adapter(model: str) -> OmniProviderAdapter:
     Qwen 侧仅支持 Qwen3.5-Omni 系列（qwen3.5-omni-plus / qwen3.5-omni-flash），
     旧版 qwen3-omni-flash 不支持多模态组合输入，无法满足 fused 模式需求。
 
-    GLM 走 OpenAI 兼容协议（glm-4.6v 等），请求体与 MiMo 同构，
-    仅鉴权头差异由 GlmAdapter 处理。
+    GLM 走 OpenAI 兼容协议（glm-4.6v 等），请求体与 MiMo 同构，鉴权头亦相同
+    （均为标准 Bearer）；唯一差异是 ``supports_audio_input = False``
+    （GLM-4.6V 是纯视觉模型，不接受音频输入）。
 
     Gemini 走原生 generateContent 协议（OpenAI 兼容端点不支持视频输入）。
     """
diff --git a/backend/miloco/tests/perception/engine/omni/test_omni_client_circuit.py b/backend/miloco/tests/perception/engine/omni/test_omni_client_circuit.py
index 3203a8097..a7f33aed3 100644
--- a/backend/miloco/tests/perception/engine/omni/test_omni_client_circuit.py
+++ b/backend/miloco/tests/perception/engine/omni/test_omni_client_circuit.py
@@ -288,6 +288,7 @@ class _Mo:
         model = "m1"
         base_url = "https://x/v1"
         api_key = "sk-NEW"
+        extra_headers: dict[str, str] = {}
 
     class _M:
         omni = _Mo()
diff --git a/backend/miloco/tests/perception/engine/omni/test_prompt_builder.py b/backend/miloco/tests/perception/engine/omni/test_prompt_builder.py
index e0a7fa6d3..c9a4206f2 100644
--- a/backend/miloco/tests/perception/engine/omni/test_prompt_builder.py
+++ b/backend/miloco/tests/perception/engine/omni/test_prompt_builder.py
@@ -1409,3 +1409,57 @@ def test_system_prompt_no_member_matching_leak_when_moot(self):
         assert self._TASK_MATCH_MARKER in full
         assert self._EXAMPLE_A_MARKER in full
         assert self._MATCH_ONLY_MARKER in full
+
+
+class TestGlmAudioFallback:
+    """GLM（纯视觉，不支持 input_audio）的音频路降级行为。
+
+    锁的是第三轮 review 指出的核心行为：GLM 在 audio-only 窗口应退回 video 路由
+    （送画面而非降级为 text-only），has_audio/has_speech 压 False，fused 路径同样降级。
+    """
+
+    def test_audio_only_falls_back_to_video_for_glm(self):
+        from miloco.perception.engine.omni.prompt_builder import _build_payload
+        from miloco.perception.engine.omni.provider import GlmAdapter
+
+        payload = _build_payload(
+            [_audio_only_packet()], OmniContext(), stream=False, adapter=GlmAdapter()
+        )
+        assert "audio_base64" not in payload  # 不再走音频路
+        assert payload["video_base64"]  # 改送视频帧
+        assert "speeches" not in payload["system_prompt"]  # 音频字段已剥掉
+
+    def test_audio_only_keeps_audio_route_for_mimo(self):
+        from miloco.perception.engine.omni.prompt_builder import _build_payload
+        from miloco.perception.engine.omni.provider import MiMoAdapter
+
+        payload = _build_payload(
+            [_audio_only_packet()], OmniContext(), stream=False, adapter=MiMoAdapter()
+        )
+        assert "audio_base64" in payload  # MiMo 照常走音频路
+        assert "video_base64" not in payload
+
+    def test_fused_audio_falls_back_to_video_for_glm(self):
+        from miloco.perception.engine.omni.prompt_builder import build_fused_payload
+        from miloco.perception.engine.omni.provider import GlmAdapter
+
+        payload = build_fused_payload(
+            packets=[_audio_only_packet()], context=OmniContext(),
+            candidates=[], gallery_snapshot={}, adapter=GlmAdapter(),
+        )
+        user_blocks = payload["messages"][-1]["content"]
+        assert all(b["type"] != "input_audio" for b in user_blocks)
+        # fused 降级后同样送视频帧
+        assert any(b["type"] == "video_url" for b in user_blocks)
+
+    def test_video_route_glm_strips_audio_fields(self):
+        from miloco.perception.engine.omni.prompt_builder import build_fused_payload
+        from miloco.perception.engine.omni.provider import GlmAdapter
+
+        payload = build_fused_payload(
+            packets=[_video_route_packet()], context=OmniContext(),
+            candidates=[], gallery_snapshot={}, adapter=GlmAdapter(),
+        )
+        # video 路由 + GLM：prompt 不声称本轮有音频
+        system_prompt = payload["messages"][0]["content"]
+        assert "speeches" not in system_prompt

From 43641535ee8d96aea480e73bc21ea0f6e39bddc6 Mon Sep 17 00:00:00 2001
From: homelab <me@cxd.name>
Date: Fri, 31 Jul 2026 20:59:56 +0800
Subject: [PATCH 7/9] =?UTF-8?q?fix(omni):=20extra=5Fheaders=20=E7=A9=BF?=
 =?UTF-8?q?=E9=80=8F=E6=94=B6=E5=B0=BE=20=E2=80=94=20probe=20=E5=9B=9E?=
 =?UTF-8?q?=E9=80=80/=E7=86=94=E6=96=AD=E8=87=AA=E6=84=88/=E6=A1=A3?=
 =?UTF-8?q?=E6=A1=88=E5=BA=8F=E5=88=97=E5=8C=96?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

pr-review 第四轮意见修复(同一字段的遗漏点):

🔴 probe_omni 内部 chat 回退路径漏传 extra_headers — GET /models
   通过后最后一跳 probe_chat 不带 X-Title → 429, 保存/激活/测试/
   重试全失败
🔴 processor._run_omni_probe 熔断自动探活漏传 — GLM 感知一次
   抖动后熔断半开探活永远 429, 永久停摆
🔴 activate_omni_config 写回 active 漏 extra_headers — 激活 GLM
   档案后推理立即 429

🟡 _profiles_as_dicts 不返回 extra_headers — DELETE 清零所有
   剩余档案的自定义头
🟡 _full_omni_payload 不返回 — 前端编辑无法回显, 保存即丢

测试: omni + admin 434 个全过
---
 backend/miloco/src/miloco/admin/router.py                 | 5 +++++
 backend/miloco/src/miloco/perception/engine/omni/probe.py | 2 +-
 backend/miloco/src/miloco/perception/processor.py         | 7 ++++++-
 3 files changed, 12 insertions(+), 2 deletions(-)

diff --git a/backend/miloco/src/miloco/admin/router.py b/backend/miloco/src/miloco/admin/router.py
index 226db6af2..2ced06af6 100644
--- a/backend/miloco/src/miloco/admin/router.py
+++ b/backend/miloco/src/miloco/admin/router.py
@@ -522,6 +522,7 @@ 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
     ]
@@ -535,6 +536,7 @@ 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())
@@ -546,6 +548,7 @@ 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,
     }
@@ -558,6 +561,7 @@ 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
     ]
@@ -700,6 +704,7 @@ 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 {}),
                     }
                 }
             )
diff --git a/backend/miloco/src/miloco/perception/engine/omni/probe.py b/backend/miloco/src/miloco/perception/engine/omni/probe.py
index 87716f521..be0d4804f 100644
--- a/backend/miloco/src/miloco/perception/engine/omni/probe.py
+++ b/backend/miloco/src/miloco/perception/engine/omni/probe.py
@@ -395,4 +395,4 @@ async def probe_omni(
             "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)
diff --git a/backend/miloco/src/miloco/perception/processor.py b/backend/miloco/src/miloco/perception/processor.py
index 0c23e4558..04b0511e8 100644
--- a/backend/miloco/src/miloco/perception/processor.py
+++ b/backend/miloco/src/miloco/perception/processor.py
@@ -104,7 +104,12 @@ async def _run_omni_probe() -> None:
                 ClassifiedError("no_key", "未配置 API Key", ErrorCategory.CONFIG),
             )
             return
-        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 {}),
+        )
         if result.get("ok"):
             await cb.record_probe_result(True, None)
             return

From 9bdf7bee388cb9e149d99ead84d971201b07773d Mon Sep 17 00:00:00 2001
From: homelab <me@cxd.name>
Date: Fri, 31 Jul 2026 21:11:00 +0800
Subject: [PATCH 8/9] =?UTF-8?q?test(omni):=20=E4=BF=AE=20=5FFakeOmni/=5Ffa?=
 =?UTF-8?q?ke=5Fprobe=20mock=20=E9=80=82=E9=85=8D=20extra=5Fheaders?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

processor._run_omni_probe 现在传 extra_headers, 测试 mock 需同步
---
 .../miloco/tests/perception/test_omni_probe_tick_drive.py    | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/backend/miloco/tests/perception/test_omni_probe_tick_drive.py b/backend/miloco/tests/perception/test_omni_probe_tick_drive.py
index 924346a5b..31f8995cc 100644
--- a/backend/miloco/tests/perception/test_omni_probe_tick_drive.py
+++ b/backend/miloco/tests/perception/test_omni_probe_tick_drive.py
@@ -36,6 +36,7 @@ class _FakeOmni:
         model = "m"
         base_url = "https://x/v1"
         api_key = "sk-x"
+        extra_headers: dict[str, str] = {}
 
     class _FakeModel:
         omni = _FakeOmni()
@@ -62,7 +63,7 @@ async def test_tick_drive_probe_success_recovers_to_closed(monkeypatch, _mock_om
     """probe 成功 → 熔断从 OPEN_RECOVERABLE 回 CLOSED,感知恢复。"""
     from miloco.perception import processor as _processor
 
-    async def _fake_probe(model, base_url, api_key):
+    async def _fake_probe(model, base_url, api_key, extra_headers=None):
         return {"ok": True, "code": "ok", "status": 200, "latency_ms": 10}
 
     monkeypatch.setattr(
@@ -99,7 +100,7 @@ async def test_tick_drive_probe_failure_grows_backoff(monkeypatch, _mock_omni_co
     """probe 继续失败 → 状态回 OPEN_RECOVERABLE,backoff 涨,in-flight 清位。"""
     from miloco.perception import processor as _processor
 
-    async def _fake_probe(model, base_url, api_key):
+    async def _fake_probe(model, base_url, api_key, extra_headers=None):
         return {"ok": False, "code": "unreachable", "message": "still down"}
 
     monkeypatch.setattr(

From 7d996925b2ceeaa5c809c31582d04c7999a673f0 Mon Sep 17 00:00:00 2001
From: homelab <me@cxd.name>
Date: Fri, 31 Jul 2026 21:11:48 +0800
Subject: [PATCH 9/9] =?UTF-8?q?fix(omni):=20test=5Fomni=5Fconfig=20?=
 =?UTF-8?q?=E6=8C=89=20label=20=E5=8F=96=20extra=5Fheaders=20+=20probe=20?=
 =?UTF-8?q?=E7=A9=BF=E9=80=8F=E6=B5=8B=E8=AF=95?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

pr-review 第五轮意见:

🟡 test_omni_config 的 extra_headers 取法与 list_omni_models 不一致
   — 测试非 active 的 GLM 档案时 429, 统一按 label 查档案/缺省 active

🔵 probe.py extra_headers 路径零测试 — 新增 3 用例锁住 5 处合并点
   (fetch_models / probe_chat / probe_omni 两跳), 防重构静默回归
---
 backend/miloco/src/miloco/admin/router.py     | 12 +++-
 .../perception/engine/omni/test_probe.py      | 68 +++++++++++++++++++
 2 files changed, 79 insertions(+), 1 deletion(-)

diff --git a/backend/miloco/src/miloco/admin/router.py b/backend/miloco/src/miloco/admin/router.py
index 2ced06af6..330980bc7 100644
--- a/backend/miloco/src/miloco/admin/router.py
+++ b/backend/miloco/src/miloco/admin/router.py
@@ -822,8 +822,18 @@ async def test_omni_config(
             message="ok",
             data={"ok": False, "code": "no_key", "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(omni.extra_headers or {})
     result = await _probe.probe_omni(
-        model, base_url, api_key, extra_headers=dict(omni.extra_headers or {})
+        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。
diff --git a/backend/miloco/tests/perception/engine/omni/test_probe.py b/backend/miloco/tests/perception/engine/omni/test_probe.py
index c43215eaa..bfbbdb1e5 100644
--- a/backend/miloco/tests/perception/engine/omni/test_probe.py
+++ b/backend/miloco/tests/perception/engine/omni/test_probe.py
@@ -429,3 +429,71 @@ async def test_probe_chat_stream_429_preserves_retry_after(monkeypatch):
     assert r["code"] == "rate_limited"
     # 关键:Retry-After 被解析出来传给上层 _grow_backoff_locked
     assert r["retry_after_seconds"] == 45.0
+
+
+# ─── extra_headers 穿透 ─────────────────────────────────────────────────────
+
+
+def _capture_client(captured: dict):
+    """捕获请求头/body 的 fake AsyncClient（GET + POST）。"""
+
+    class _C:
+        def __init__(self, *a, **k):
+            pass
+
+        async def __aenter__(self):
+            return self
+
+        async def __aexit__(self, *a):
+            return False
+
+        async def get(self, url, headers=None, **k):
+            captured["get_headers"] = headers or {}
+            captured["get_url"] = url
+            return _FakeResp(200, {"data": [{"id": "glm-4.6v"}]})
+
+        async def post(self, url, headers=None, json=None, **k):
+            captured["post_headers"] = headers or {}
+            captured["post_url"] = url
+            return _FakeResp(200, {"choices": [{"message": {"content": "pong"}}]})
+
+    return _C
+
+
+async def test_fetch_models_passes_extra_headers(monkeypatch):
+    captured: dict = {}
+    monkeypatch.setattr(probe.httpx, "AsyncClient", _capture_client(captured))
+    await probe.fetch_models(
+        "https://open.bigmodel.cn/api/paas/v4",
+        "sk-glm",
+        extra_headers={"X-Title": "4.5V MCP Local"},
+    )
+    assert captured["get_headers"]["Authorization"] == "Bearer sk-glm"
+    assert captured["get_headers"]["X-Title"] == "4.5V MCP Local"
+
+
+async def test_probe_chat_passes_extra_headers(monkeypatch):
+    captured: dict = {}
+    monkeypatch.setattr(probe.httpx, "AsyncClient", _capture_client(captured))
+    await probe.probe_chat(
+        "glm-4.6v",
+        "https://open.bigmodel.cn/api/paas/v4",
+        "sk-glm",
+        extra_headers={"X-Title": "4.5V MCP Local"},
+    )
+    assert captured["post_headers"]["X-Title"] == "4.5V MCP Local"
+
+
+async def test_probe_omni_passes_extra_headers_both_stages(monkeypatch):
+    """GET /models 与 chat 回退两跳都要带 extra_headers（Coding Plan 429 回归锁）。"""
+    captured: dict = {}
+    monkeypatch.setattr(probe.httpx, "AsyncClient", _capture_client(captured))
+    result = await probe.probe_omni(
+        "glm-4.6v",
+        "https://open.bigmodel.cn/api/paas/v4",
+        "sk-glm",
+        extra_headers={"X-Title": "4.5V MCP Local"},
+    )
+    assert result.get("ok")
+    assert captured["get_headers"]["X-Title"] == "4.5V MCP Local"
+    assert captured["post_headers"]["X-Title"] == "4.5V MCP Local"