feat(chat): NemoClaw 改接 Ollama 111 deepseek-r1:14b

2026-04-09 ogt: 棄用 Claude Haiku，改用本地 deepseek-r1:14b
- 端點: http://192.168.0.111:11434
- 過濾 <think>...</think> 推理區塊，只回傳結論
- timeout 120s（14b 推理較慢）
- 完全免費，不計入 Claude API 費用

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

apps/api/src/services/chat_manager.py

@@ -4,13 +4,15 @@ 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 (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)
+Phase 22.8 更新: 2026-04-09 ogt (老闆指示: NemoClaw→Ollama 111 deepseek-r1:14b，SRE 推理更強)
 
 架構:
 - 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
+- NemoClaw (Ollama 192.168.0.111 deepseek-r1:14b): 戰術參謀，深度推理
 
 費用控管:
 - Gemini Flash: Input $0.075/1M tokens, Output $0.30/1M tokens
+- NemoClaw: 免費 (本地 Ollama)
 - 每次回覆顯示 token 用量與費用
 - 月上限 $10 USD (由 ai_rate_limiter 控管)
 """
@@ -29,7 +31,7 @@ OPENCLAW_PERSONA = """你是 OpenClaw，AWOOOI 平台的 SRE AI 首席顧問。
 稱呼用戶為「老闆」。
 """
 
-NEMOCLAW_PERSONA = """你是 NemoClaw，AWOOOI 平台的 AI 戰術參謀，由 NVIDIA Nemotron 驅動。
+NEMOCLAW_PERSONA = """你是 NemoClaw，AWOOOI 平台的 AI 戰術參謀，由 DeepSeek-R1 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。不超過 200 字。
 稱呼用戶為「老闆」。評論 OpenClaw 的回應時，直接說「我補充」或「我有不同看法」。
@@ -144,64 +146,50 @@ class ChatManager:
 
     async def _call_nemotron(self, system_prompt: str, user_message: str) -> str | None:
         """
-        呼叫 NemoClaw 對話 — Claude API (claude-haiku-4-5)
+        呼叫 NemoClaw 對話 — Ollama 111 deepseek-r1:14b
 
-        2026-04-03 ogt: 老闆指示改接 Claude API，快速且中文能力強
-        用 haiku 降低費用
+        2026-04-09 ogt: 改接 192.168.0.111 Ollama deepseek-r1:14b，SRE 推理能力最強
+        deepseek-r1 含 <think> 標籤，需過濾後才回傳
         """
         import httpx
-        from src.core.config import get_settings
-        settings = get_settings()
+        import re
 
-        api_key = getattr(settings, 'CLAUDE_API_KEY', None)
-        if not api_key:
-            logger.warning("nemotron_chat_failed", error="CLAUDE_API_KEY not configured")
-            return None
+        OLLAMA_URL = "http://192.168.0.111:11434"
+        MODEL = "deepseek-r1:14b"
 
         try:
-            async with httpx.AsyncClient(timeout=30.0) as client:
+            async with httpx.AsyncClient(timeout=120.0) as client:
                 resp = await client.post(
-                    "https://api.anthropic.com/v1/messages",
-                    headers={
-                        "x-api-key": api_key,
-                        "anthropic-version": "2023-06-01",
-                        "content-type": "application/json",
-                    },
+                    f"{OLLAMA_URL}/api/chat",
                     json={
-                        "model": "claude-haiku-4-5-20251001",
-                        "max_tokens": 300,
-                        "system": system_prompt,
-                        "messages": [{"role": "user", "content": user_message}],
+                        "model": MODEL,
+                        "stream": False,
+                        "messages": [
+                            {"role": "system", "content": system_prompt},
+                            {"role": "user", "content": user_message},
+                        ],
+                        "options": {"num_predict": 400},
                     },
                 )
                 resp.raise_for_status()
                 data = resp.json()
-                text = data["content"][0]["text"].strip()
+                raw = data.get("message", {}).get("content", "").strip()
 
-                # Token/費用統計 — Claude Haiku 4.5: Input $0.80/1M, Output $4.00/1M
-                usage = data.get("usage", {})
-                in_tok = usage.get("input_tokens", 0)
-                out_tok = usage.get("output_tokens", 0)
-                cost = (in_tok * 0.0000008) + (out_tok * 0.000004)
+                # 過濾 deepseek-r1 的 <think>...</think> 推理區塊
+                text = re.sub(r"<think>.*?</think>", "", raw, flags=re.DOTALL).strip()
+                if not text:
+                    text = raw  # 萬一全是 think block，直接回傳原文
 
-                # 月累計到 Redis
-                from src.core.redis_client import get_redis
-                from src.utils.timezone import now_taipei
-                redis = get_redis()
-                month_key = f"claude_cost:{now_taipei().strftime('%Y-%m')}"
-                try:
-                    current = float(await redis.get(month_key) or 0)
-                    new_total = current + cost
-                    await redis.set(month_key, str(round(new_total, 6)), ex=40 * 24 * 3600)
-                except Exception:
-                    new_total = cost
+                eval_count = data.get("eval_count", 0)
+                prompt_eval_count = data.get("prompt_eval_count", 0)
+                total_tokens = eval_count + prompt_eval_count
 
-                logger.info("nemotron_claude_usage", in_tokens=in_tok, out_tokens=out_tok,
-                            cost_usd=round(cost, 6), monthly_total_usd=round(new_total, 4))
+                logger.info("nemotron_ollama_usage", model=MODEL,
+                            prompt_tokens=prompt_eval_count, output_tokens=eval_count)
 
-                return f"{text}\n\n<i>📊 {in_tok+out_tok} tokens | ${cost:.4f} | 本月累計 ${new_total:.4f}</i>"
+                return f"{text}\n\n<i>🦙 {MODEL} | {total_tokens} tokens | 免費</i>"
         except Exception as e:
-            logger.warning("nemotron_chat_failed", error=str(e))
+            logger.warning("nemotron_chat_failed", model=MODEL, error=str(e))
             return None
 
     async def generate_response(