feat(chat): OpenClaw→Gemini Flash + NemoClaw→Ollama llama3.2:3b

老闆指示 (2026-04-03):
- OpenClaw: Gemini 1.5 Flash API，每次回覆附 token+費用統計
- NemoClaw: Ollama llama3.2:3b，本地快速回應 (3-8s)
- 費用控管: Gemini 月上限 $10 USD

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

apps/api/src/services/chat_manager.py

@@ -2,19 +2,17 @@
 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)
+Phase 22.6 重寫: 2026-04-03 ogt (老闆需求: 雙 AI 互動對話)
+Phase 22.7 更新: 2026-04-03 ogt (老闆指示: OpenClaw→Gemini, NemoClaw→Ollama llama3.2:3b)
 
 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
+- OpenClaw (Gemini API): SRE 首席顧問，精準分析
+- NemoClaw (Ollama llama3.2:3b): 戰術參謀，快速補充
 
-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
+費用控管:
+- Gemini Flash: Input $0.075/1M tokens, Output $0.30/1M tokens
+- 每次回覆顯示 token 用量與費用
+- 月上限 $10 USD (由 ai_rate_limiter 控管)
 """
 
 import asyncio
@@ -78,46 +76,77 @@ class ChatManager:
 
     async def _call_openclaw(self, system_prompt: str, user_message: str) -> str | None:
         """
-        呼叫 OpenClaw 對話 — 走 NVIDIA NIM meta/llama-3.1-8b-instruct
+        呼叫 OpenClaw 對話 — Gemini Flash API
 
-        2026-04-03 ogt: OpenClaw 8088 的 analyze/incident 是告警分析 API，
-        回覆是告警格式，不適合自然語言對話。
-        改用 NIM llama-3.1-8b 做 chat，與 NemoClaw 同樣走免費 NIM cloud。
+        2026-04-03 ogt: 老闆指示改用 Gemini，費用控管月上限 $10 USD
+        每次回覆附帶 token 用量與費用統計
         """
-        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="meta/llama-3.1-8b-instruct",
-                max_tokens=300,
-            )
-            if success and response and "not configured" not in response and "Circuit Breaker" not in response:
-                return response.strip()
+        import httpx
+        from src.core.config import get_settings
+        settings = get_settings()
+
+        api_key = settings.GEMINI_API_KEY
+        if not api_key:
+            logger.warning("openclaw_chat_failed", error="GEMINI_API_KEY not configured")
             return None
+
+        # Gemini 1.5 Flash: 快速、便宜
+        model = "gemini-1.5-flash"
+        try:
+            async with httpx.AsyncClient(timeout=30.0) as client:
+                resp = await client.post(
+                    f"https://generativelanguage.googleapis.com/v1beta/models/{model}:generateContent?key={api_key}",
+                    json={
+                        "system_instruction": {"parts": [{"text": system_prompt}]},
+                        "contents": [{"parts": [{"text": user_message}]}],
+                        "generationConfig": {"maxOutputTokens": 300, "temperature": 0.7},
+                    },
+                )
+                resp.raise_for_status()
+                data = resp.json()
+                text = data["candidates"][0]["content"]["parts"][0]["text"].strip()
+
+                # Token/費用統計
+                usage = data.get("usageMetadata", {})
+                in_tok = usage.get("promptTokenCount", 0)
+                out_tok = usage.get("candidatesTokenCount", 0)
+                cost = (in_tok * 0.000000075) + (out_tok * 0.0000003)
+                logger.info("openclaw_gemini_usage", in_tokens=in_tok, out_tokens=out_tok, cost_usd=round(cost, 6))
+
+                return f"{text}\n\n<i>📊 {in_tok+out_tok} tokens | ${cost:.4f}</i>"
         except Exception as e:
             logger.warning("openclaw_chat_failed", error=str(e))
             return None
 
     async def _call_nemotron(self, system_prompt: str, user_message: str) -> str | None:
         """
-        呼叫 NVIDIA NIM nemotron-mini-4b (NemoClaw)
+        呼叫 NemoClaw 對話 — Ollama llama3.2:3b (本地，快速)
 
-        NIM 免費 tier 延遲 11-45s，此方法可能需要 30-120s 才回應
+        2026-04-03 ogt: 老闆指示改用 Ollama 小模型取代 NIM，加快回應速度
         """
-        from src.services.nvidia_provider import get_nvidia_provider
-        nvidia = get_nvidia_provider()
+        import httpx
+        from src.core.config import get_settings
+        settings = get_settings()
+
+        ollama_url = getattr(settings, 'OLLAMA_URL', 'http://192.168.0.188:11434')
         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
+            async with httpx.AsyncClient(timeout=60.0) as client:
+                resp = await client.post(
+                    f"{ollama_url}/api/chat",
+                    json={
+                        "model": "llama3.2:3b",
+                        "stream": False,
+                        "messages": [
+                            {"role": "system", "content": system_prompt},
+                            {"role": "user", "content": user_message},
+                        ],
+                        "options": {"num_predict": 250},
+                    },
+                )
+                resp.raise_for_status()
+                data = resp.json()
+                text = data.get("message", {}).get("content", "").strip()
+                return text or None
         except Exception as e:
             logger.warning("nemotron_chat_failed", error=str(e))
             return None