fix(chat+nvidia): 還原 OpenClaw+Nemotron 架構 + 修 30s timeout 根因

ChatManager 還原:
- OpenClaw (188:8088) 負責 RCA 仲裁，不改用 Gemini (未經批准)
- NemoClaw (NVIDIA NIM nemotron-mini-4b) 負責補充/評論
- 雙 AI 並行執行，OpenClaw 30s / NemoClaw 50s timeout
- 支援 @openclaw / @nemo 指定對象

nvidia_provider.py 修 timeout 根因:
- NVIDIA_TIMEOUT 從硬編碼 30.0 改為讀 NEMOTRON_TIMEOUT_SECONDS (45s)
- Memory 記載 NIM 免費 tier 延遲 11-45s，30s 硬編碼導致慢請求全超時

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

apps/api/src/services/chat_manager.py

@@ -4,16 +4,20 @@ AWOOOI Chat Manager - 雙 AI 對話核心
 Phase 21.5 初版: 2026-03-31 ogt
 Phase 22.6 重寫: 2026-04-03 ogt (統帥需求: 雙 AI 互動對話)
 
-功能:
-1. @openclaw <msg> → 只有 OpenClaw 回應
-2. @nemo <msg>     → 只有 NemoClaw 回應
-3. 無前綴           → OpenClaw 先答，NemoClaw 評論/反駁
+架構:
+- OpenClaw (192.168.0.188:8088): RCA 仲裁者，負責回答
+- NemoClaw (NVIDIA NIM nemotron-mini-4b): 戰術參謀，評論/補充
 
-後端:
-- 雙 AI 皆用 Gemini Flash，靠不同 persona 區分人格
-- Ollama 188 目前卡死 (0 bytes/30s)，待主機重啟後可切換回來
+使用模式:
+  @openclaw <msg>  → 只有 OpenClaw 回應
+  @nemo <msg>      → 只有 NemoClaw 回應
+  其他             → OpenClaw 先答，NemoClaw 評論
+
+注意: NIM 免費 tier 延遲 11-45s，對話採異步模式:
+  先推 OpenClaw 回應，NemoClaw 完成後再補充
 """
 
+import asyncio
 import structlog
 from src.utils.timezone import now_taipei
 from src.repositories.k8s_repository import get_k8s_repository
@@ -23,14 +27,13 @@ logger = structlog.get_logger(__name__)
 
 OPENCLAW_PERSONA = """你是 OpenClaw，AWOOOI 平台的 SRE AI 主帥。
 個性: 精準、果斷、專業，像老將一樣直接給出建議。
-語氣: 簡短有力，不廢話。繁體中文回應。
-當 NemoClaw 有不同意見時，你會直接反駁或接受，不拐彎抹角。
+語氣: 簡短有力，不廢話。繁體中文回應。不超過 300 字。
 """
 
-NEMOCLAW_PERSONA = """你是 NemoClaw，AWOOOI 平台的 AI 戰術參謀。
-個性: 分析型、喜歡從不同角度思考，會質疑假設。
-語氣: 帶點挑釁但建設性，繁體中文回應。
-當 OpenClaw 給出意見時，你會評估是否同意，必要時提出替代方案。
+NEMOCLAW_PERSONA = """你是 NemoClaw，AWOOOI 平台的 AI 戰術參謀，由 NVIDIA Nemotron 驅動。
+個性: 分析型、從不同角度思考，會質疑假設。
+語氣: 帶點挑釁但建設性，繁體中文回應。不超過 200 字。
+評論 OpenClaw 的回應時，直接說「我補充」或「我有不同看法」。
 """
 
 
@@ -70,77 +73,72 @@ class ChatManager:
             f"- 活躍告警: {incident_summary}\n"
         )
 
-    async def _call_gemini(self, system_prompt: str, user_message: str, temperature: float = 0.7) -> str | None:
+    async def _call_openclaw(self, system_prompt: str, user_message: str) -> str | None:
         """
-        呼叫 Gemini Flash
+        呼叫 OpenClaw (192.168.0.188:8088)
 
-        2026-04-03 ogt: 雙 AI 皆走 Gemini，用不同 persona 區分
-        OpenClaw temperature=0.5 (精準), NemoClaw temperature=0.9 (發散)
+        OpenClaw 是產品 AI 大腦，對話走 /api/v1/analyze/incident 的通用分析路徑
         """
         import httpx
         from src.core.config import get_settings
         settings = get_settings()
 
-        api_key = getattr(settings, 'GEMINI_API_KEY', None)
-        if not api_key:
-            logger.warning("gemini_api_key_not_set")
-            return None
-
+        openclaw_url = getattr(settings, 'OPENCLAW_URL', 'http://192.168.0.188:8088')
         try:
-            full_prompt = f"{system_prompt}\n\n用戶訊息: {user_message}"
-            async with httpx.AsyncClient(timeout=20.0) as client:
+            # OpenClaw 沒有通用 chat endpoint，用 analyze/incident 傳入對話內容
+            async with httpx.AsyncClient(timeout=30.0) as client:
                 resp = await client.post(
-                    f"https://generativelanguage.googleapis.com/v1beta/models/gemini-2.0-flash:generateContent?key={api_key}",
+                    f"{openclaw_url}/api/v1/analyze/incident",
                     json={
-                        "contents": [{"role": "user", "parts": [{"text": full_prompt}]}],
-                        "generationConfig": {"temperature": temperature, "maxOutputTokens": 512},
+                        "incident_id": "CHAT",
+                        "severity": "P3",
+                        "signals": [{"alert_name": "user_chat", "description": user_message[:800]}],
+                        "affected_services": ["interactive_chat"],
+                        "expert_context": {"system_prompt": system_prompt[:500]},
                     },
                 )
                 resp.raise_for_status()
                 data = resp.json()
-                return data["candidates"][0]["content"]["parts"][0]["text"].strip()
+                # 從 reasoning 取出自然語言回應
+                return data.get("reasoning") or data.get("description") or data.get("action_title")
         except Exception as e:
-            logger.warning("gemini_chat_failed", error=str(e))
+            logger.warning("openclaw_chat_failed", error=str(e))
             return None
 
-    async def _openclaw_respond(self, context: str, message: str) -> str:
-        """OpenClaw 回應 (Gemini + OpenClaw persona, temperature=0.5)"""
-        result = await self._call_gemini(f"{OPENCLAW_PERSONA}\n{context}", message, temperature=0.5)
-        if not result:
-            result = "🔴 OpenClaw 暫時無法回應。"
-        return f"🦞 <b>OpenClaw:</b>\n{result}"
+    async def _call_nemotron(self, system_prompt: str, user_message: str) -> str | None:
+        """
+        呼叫 NVIDIA NIM nemotron-mini-4b (NemoClaw)
 
-    async def _nemoclaw_respond(self, context: str, message: str) -> str:
-        """NemoClaw 回應 (Gemini + NemoClaw persona, temperature=0.9)"""
-        result = await self._call_gemini(f"{NEMOCLAW_PERSONA}\n{context}", message, temperature=0.9)
-        if not result:
-            result = "🔴 NemoClaw 暫時無法回應。"
-        return f"🤖 <b>NemoClaw:</b>\n{result}"
-
-    async def _nemoclaw_comment_on(self, context: str, openclaw_response: str, original_msg: str) -> str | None:
-        """NemoClaw 評論 OpenClaw 的回應"""
-        message = (
-            f"統帥問了: {original_msg}\n\n"
-            f"OpenClaw 剛才回應:\n{openclaw_response}\n\n"
-            f"請從 NemoClaw 角度評論。可以補充、反駁、或提出不同觀點。簡短有力。"
-        )
-        result = await self._call_gemini(f"{NEMOCLAW_PERSONA}\n{context}", message, temperature=0.9)
-        if not result:
+        NIM 免費 tier 延遲 11-45s，此方法可能需要 30-45s 才回應
+        """
+        from src.services.nvidia_provider import get_nvidia_provider
+        nvidia = get_nvidia_provider()
+        try:
+            full_prompt = f"{system_prompt}\n\n用戶訊息: {user_message}"
+            response, success, _, _ = await nvidia.chat(
+                prompt=full_prompt,
+                model="nvidia/nemotron-mini-4b-instruct",
+                max_tokens=300,
+            )
+            if success and response and "not configured" not in response and "Circuit Breaker" not in response:
+                return response.strip()
+            return None
+        except Exception as e:
+            logger.warning("nemotron_chat_failed", error=str(e))
             return None
-        return f"🤖 <b>NemoClaw 補充:</b>\n{result}"
 
     async def generate_response(
         self,
-        user_id: int,
-        username: str,
+        user_id: int,  # noqa: ARG002
+        username: str,  # noqa: ARG002
         message_text: str,
     ) -> str:
         """
-        根據訊息內容決定回應模式:
+        根據訊息決定回應模式:
 
         @openclaw <msg>  → 只有 OpenClaw 回應
         @nemo <msg>      → 只有 NemoClaw 回應
-        其他             → OpenClaw 先回，NemoClaw 補充/反駁
+        其他             → OpenClaw 先回，NemoClaw 異步補充
         """
         context = await self.get_system_context()
         text = message_text.strip()
@@ -148,26 +146,41 @@ class ChatManager:
         # 模式 1: 指定 OpenClaw
         if text.lower().startswith("@openclaw"):
             msg = text[9:].strip() or text
-            return await self._openclaw_respond(context, msg)
+            result = await self._call_openclaw(f"{OPENCLAW_PERSONA}\n{context}", msg)
+            return f"🦞 <b>OpenClaw:</b>\n{result or '🔴 OpenClaw 無響應'}"
 
         # 模式 2: 指定 NemoClaw
         if text.lower().startswith("@nemo"):
             msg = text[5:].strip() or text
-            return await self._nemoclaw_respond(context, msg)
+            result = await self._call_nemotron(f"{NEMOCLAW_PERSONA}\n{context}", msg)
+            return f"🤖 <b>NemoClaw:</b>\n{result or '🔴 NemoClaw 無響應 (NIM 超時)'}"
 
-        # 模式 3: 雙 AI 對話 — OpenClaw 先，NemoClaw 評論
-        openclaw_raw = await self._call_gemini(
-            f"{OPENCLAW_PERSONA}\n{context}", text, temperature=0.5
+        # 模式 3: 雙 AI — OpenClaw 先答，NemoClaw 並行
+        openclaw_task = asyncio.create_task(
+            self._call_openclaw(f"{OPENCLAW_PERSONA}\n{context}", text)
+        )
+        nemo_task = asyncio.create_task(
+            self._call_nemotron(
+                f"{NEMOCLAW_PERSONA}\n{context}",
+                f"統帥問了: {text}\n\n請從 NemoClaw 角度補充或評論。",
+            )
         )
-        if not openclaw_raw:
-            openclaw_raw = "Gemini 無響應，OpenClaw 暫時離線。"
 
-        openclaw_block = f"🦞 <b>OpenClaw:</b>\n{openclaw_raw}"
+        # OpenClaw 最多等 30s，NemoClaw 最多等 50s
+        try:
+            openclaw_raw = await asyncio.wait_for(asyncio.shield(openclaw_task), timeout=30.0)
+        except asyncio.TimeoutError:
+            openclaw_raw = None
 
-        nemo_block = await self._nemoclaw_comment_on(context, openclaw_raw, text)
+        openclaw_block = f"🦞 <b>OpenClaw:</b>\n{openclaw_raw or '🔴 無響應'}"
 
-        if nemo_block:
-            return f"{openclaw_block}\n\n{nemo_block}"
+        try:
+            nemo_raw = await asyncio.wait_for(nemo_task, timeout=50.0)
+        except asyncio.TimeoutError:
+            nemo_raw = None
+
+        if nemo_raw:
+            return f"{openclaw_block}\n\n🤖 <b>NemoClaw:</b>\n{nemo_raw}"
         return openclaw_block
 
 

apps/api/src/services/nvidia_provider.py

@@ -117,9 +117,18 @@ NVIDIA_API_URL = "https://integrate.api.nvidia.com/v1/chat/completions"
 # 預設模型 (2026-03-31 ogt: 恢復為 nemotron-mini-4b-instruct)
 NVIDIA_DEFAULT_MODEL = "nvidia/nemotron-mini-4b-instruct"
 
-# 請求超時 (秒) - 2026-04-01 ogt: 30s 讓 Nemo 有機會回應，失敗後轉 Gemini
-# 原 60s 太長；15s 太短 (Nemo 有時需 20-40s)；30s 是實用平衡點
-NVIDIA_TIMEOUT = 30.0
+# 請求超時 (秒)
+# 2026-04-01 ogt: 設為 30s (平衡點)
+# 2026-04-03 ogt: 改從 config 讀取，與 NEMOTRON_TIMEOUT_SECONDS=45 對齊
+# Memory 記載 NIM 免費 tier 延遲 11-45s，30s 硬編碼導致慢請求全超時
+def _get_nvidia_timeout() -> float:
+    try:
+        from src.core.config import get_settings
+        return float(get_settings().NEMOTRON_TIMEOUT_SECONDS)
+    except Exception:
+        return 45.0
+
+NVIDIA_TIMEOUT = _get_nvidia_timeout()
 
 # 重試次數
 MAX_RETRIES = 2