fix(chat): OpenClaw 改走 Gemini Flash，移除 Ollama 依賴

Ollama 188 完全卡死 (0 bytes/30s timeout)，無法作為對話後端。
雙 AI 皆使用 Gemini Flash，靠不同 persona 和 temperature 區分:
- OpenClaw: temperature=0.5 (精準果斷)
- NemoClaw: temperature=0.9 (分析發散)

同時 kubectl set env ENABLE_NEMOTRON_COLLABORATION=false
停止每個 incident 白白等待 30s Nemotron timeout

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

apps/api/src/services/chat_manager.py

@@ -5,13 +5,13 @@ 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 / @nemo 路由 — 指定 AI 回應
-2. 無前綴 — 兩個 AI 輪流回應，並互相評論
-3. AI 互相對話 — NemoClaw 看到 OpenClaw 的回應後可補充/反駁
+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: 用 Ollama qwen2.5:7b-instruct (本地, 快)
-- NemoClaw: 用 Gemini Flash (雲端, 快) — NIM nemotron-mini 太慢 (15s+)
+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
 """
 
 import structlog
@@ -21,7 +21,6 @@ from src.repositories.incident_repository import get_incident_repository
 
 logger = structlog.get_logger(__name__)
 
-# 人格設定
 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。
@@ -65,43 +64,26 @@ class ChatManager:
         except Exception:
             incident_summary = "無法取得告警"
 
-        return f"""## 系統狀態 ({now.strftime('%Y-%m-%d %H:%M')} 台北)
-- {cluster_info}
-- 活躍告警: {incident_summary}
-"""
+        return (
+            f"## 系統狀態 ({now.strftime('%Y-%m-%d %H:%M')} 台北)\n"
+            f"- {cluster_info}\n"
+            f"- 活躍告警: {incident_summary}\n"
+        )
 
-    async def _call_ollama(self, system_prompt: str, user_message: str) -> str:
-        """呼叫 Ollama (OpenClaw 用)"""
-        import httpx
-        try:
-            async with httpx.AsyncClient(timeout=30.0) as client:
-                resp = await client.post(
-                    "http://192.168.0.188:11434/api/chat",
-                    json={
-                        "model": "qwen2.5:7b-instruct",
-                        "messages": [
-                            {"role": "system", "content": system_prompt},
-                            {"role": "user", "content": user_message},
-                        ],
-                        "stream": False,
-                        "options": {"temperature": 0.7, "num_predict": 512},
-                    },
-                )
-                resp.raise_for_status()
-                data = resp.json()
-                return data.get("message", {}).get("content", "").strip()
-        except Exception as e:
-            logger.warning("ollama_chat_failed", error=str(e))
-            return None
+    async def _call_gemini(self, system_prompt: str, user_message: str, temperature: float = 0.7) -> str | None:
+        """
+        呼叫 Gemini Flash
 
-    async def _call_gemini(self, system_prompt: str, user_message: str) -> str:
-        """呼叫 Gemini Flash (NemoClaw 用)"""
+        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)
+        """
         import httpx
         from src.core.config import get_settings
         settings = get_settings()
 
-        api_key = settings.GEMINI_API_KEY if hasattr(settings, 'GEMINI_API_KEY') else None
+        api_key = getattr(settings, 'GEMINI_API_KEY', None)
         if not api_key:
+            logger.warning("gemini_api_key_not_set")
             return None
 
         try:
@@ -111,7 +93,7 @@ class ChatManager:
                     f"https://generativelanguage.googleapis.com/v1beta/models/gemini-2.0-flash:generateContent?key={api_key}",
                     json={
                         "contents": [{"role": "user", "parts": [{"text": full_prompt}]}],
-                        "generationConfig": {"temperature": 0.8, "maxOutputTokens": 512},
+                        "generationConfig": {"temperature": temperature, "maxOutputTokens": 512},
                     },
                 )
                 resp.raise_for_status()
@@ -122,37 +104,27 @@ class ChatManager:
             return None
 
     async def _openclaw_respond(self, context: str, message: str) -> str:
-        """OpenClaw 回應"""
-        system = f"{OPENCLAW_PERSONA}\n{context}"
-        result = await self._call_ollama(system, message)
+        """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 暫時離線，Ollama 無響應。"
+            result = "🔴 OpenClaw 暫時無法回應。"
         return f"🦞 <b>OpenClaw:</b>\n{result}"
 
     async def _nemoclaw_respond(self, context: str, message: str) -> str:
-        """NemoClaw 回應"""
-        system = f"{NEMOCLAW_PERSONA}\n{context}"
-        result = await self._call_gemini(system, message)
+        """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:
-            # Gemini 失敗時 fallback 到 Ollama
-            result = await self._call_ollama(system, message)
-        if not result:
-            result = "🔴 NemoClaw 暫時離線。"
+            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:
+    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}
-
-OpenClaw 的回應是:
-{openclaw_response}
-
-請你從 NemoClaw 的角度評論上面的回應。可以補充、反駁、或提出不同觀點。"""
-
-        system = f"{NEMOCLAW_PERSONA}\n{context}"
-        result = await self._call_gemini(system, message)
-        if not result:
-            result = await self._call_ollama(system, message)
+        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:
             return None
         return f"🤖 <b>NemoClaw 補充:</b>\n{result}"
@@ -168,7 +140,7 @@ OpenClaw 的回應是:
 
         @openclaw <msg>  → 只有 OpenClaw 回應
         @nemo <msg>      → 只有 NemoClaw 回應
-        其他             → OpenClaw 先回，NemoClaw 評論
+        其他             → OpenClaw 先回，NemoClaw 補充/反駁
         """
         context = await self.get_system_context()
         text = message_text.strip()
@@ -183,27 +155,24 @@ OpenClaw 的回應是:
             msg = text[5:].strip() or text
             return await self._nemoclaw_respond(context, msg)
 
-        # 模式 3: 雙 AI 對話
-        # Step 1: OpenClaw 先回
-        openclaw_raw = await self._call_ollama(
-            f"{OPENCLAW_PERSONA}\n{context}", text
+        # 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
         )
         if not openclaw_raw:
-            openclaw_raw = "Ollama 無響應，OpenClaw 暫時離線。"
+            openclaw_raw = "Gemini 無響應，OpenClaw 暫時離線。"
 
         openclaw_block = f"🦞 <b>OpenClaw:</b>\n{openclaw_raw}"
 
-        # Step 2: NemoClaw 評論 OpenClaw 的回應
         nemo_block = await self._nemoclaw_comment_on(context, openclaw_raw, text)
 
         if nemo_block:
             return f"{openclaw_block}\n\n{nemo_block}"
-        else:
-            return openclaw_block
+        return openclaw_block
 
 
 # Singleton
-_chat_manager = None
+_chat_manager: ChatManager | None = None
 
 
 def get_chat_manager() -> ChatManager: