""" 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) Phase 33 更新: 2026-05-05 ogt (ADR-110: OpenClaw chat 改走 GCP-A Ollama interactive lane) 架構: - OpenClaw (Ollama GCP-A interactive lane): SRE 首席顧問，精準分析 - NemoClaw (Ollama interactive lane deepseek-r1:14b): 戰術參謀，深度推理 費用控管: - OpenClaw/NemoClaw chat 預設免費 Ollama；Gemini 不再作為 ChatManager 預設路徑 - 每次回覆顯示 token 用量 """ import asyncio import re import httpx import structlog from src.core.config import get_settings from src.repositories.incident_repository import get_incident_repository from src.repositories.k8s_repository import get_k8s_repository from src.services.ollama_endpoint_resolver import resolve_ollama_order from src.utils.timezone import now_taipei logger = structlog.get_logger(__name__) OPENCLAW_PERSONA = """你是 OpenClaw，AWOOOI 平台的 SRE AI 首席顧問。 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。不超過 300 字。 稱呼用戶為「老闆」。 """ NEMOCLAW_PERSONA = """你是 NemoClaw，AWOOOI 平台的 AI 戰術參謀。 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」。 強制規則： 1. 全程使用繁體中文，禁止使用簡體中文、英文或其他語言。 2. 禁止自稱 DeepSeek 或透露底層模型資訊。你的名字就是 NemoClaw。 3. 專注於 SRE/DevOps/Kubernetes/可觀測性領域。 """ class ChatManager: """AWOOOI 雙 AI 對話管理器""" def __init__(self): pass # 2026-04-03 ogt: 移除 repo 實例化，leWOOOgo 規範禁止 Service 持有 repository async def get_system_context(self) -> str: """收集系統即時上下文""" now = now_taipei() k8s = get_k8s_repository() incidents = get_incident_repository() try: k8s_status = await k8s.get_pod_status_summary(namespace="awoooi-prod") cluster_info = f"Cluster: {k8s_status['running']}/{k8s_status['total']} Pods Running" if k8s_status.get('problem_pods'): cluster_info += f", {len(k8s_status['problem_pods'])} 異常" except Exception: cluster_info = "Cluster: 無法取得狀態" try: active_incidents = await incidents.get_active() if active_incidents: lines = [f"- {inc.incident_id}: {inc.status.value} (SEV {inc.severity.value})" for inc in active_incidents[:3]] incident_summary = "\n".join(lines) else: incident_summary = "無活躍告警" 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)\n" f"- {cluster_info}\n" f"- 活躍告警: {incident_summary}\n" ) async def _call_openclaw(self, system_prompt: str, user_message: str) -> str | None: """ 呼叫 OpenClaw 對話 — Ollama interactive lane 2026-04-10 Claude Code: 強制合併 OPENCLAW_PERSONA，確保字數限制與格式規範 2026-05-05 Codex: 改走 ADR-110 GCP-A/GCP-B/111 Ollama topology，避免個人聊天直打 Gemini """ # 強制在 system_prompt 前置 persona，確保 LLM 遵守字數與格式 system_prompt = f"{OPENCLAW_PERSONA}\n{system_prompt}" settings = get_settings() model = settings.OPENCLAW_DEFAULT_MODEL async with httpx.AsyncClient(timeout=40.0) as client: for endpoint in resolve_ollama_order("interactive"): if not endpoint.url: continue try: resp = await client.post( f"{endpoint.url}/api/chat", json={ "model": model, "stream": False, "messages": [ {"role": "system", "content": system_prompt}, {"role": "user", "content": user_message}, ], "options": {"num_predict": 900, "temperature": 0.2}, }, ) resp.raise_for_status() data = resp.json() raw = data.get("message", {}).get("content", "").strip() text = re.sub(r".*?", "", raw, flags=re.DOTALL).strip() or raw 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( "openclaw_ollama_chat_usage", model=model, endpoint=endpoint.url, provider=endpoint.provider_name, prompt_tokens=prompt_eval_count, output_tokens=eval_count, ) return f"{text}\n\n🦙 {model} | {total_tokens} tokens | 免費" except Exception as e: logger.warning( "openclaw_chat_endpoint_failed", provider=endpoint.provider_name, endpoint=endpoint.url, error=str(e), ) logger.warning("openclaw_chat_failed_all_endpoints", model=model) return None async def _call_nemotron(self, system_prompt: str, user_message: str) -> str | None: """ 呼叫 NemoClaw 對話 — Ollama 111 deepseek-r1:14b 2026-04-09 ogt: 改接 192.168.0.111 Ollama deepseek-r1:14b，SRE 推理能力最強 deepseek-r1 含 標籤，需過濾後才回傳 2026-04-10 Claude Code: 強制合併 NEMOCLAW_PERSONA，確保字數限制與格式規範 """ # 強制在 system_prompt 前置 persona system_prompt = f"{NEMOCLAW_PERSONA}\n{system_prompt}" # 2026-05-05 Codex: ADR-110 interactive lane，由 resolver 管理 GCP-A/GCP-B/111 拓撲 MODEL = "deepseek-r1:14b" async with httpx.AsyncClient(timeout=120.0) as client: for endpoint in resolve_ollama_order("interactive"): if not endpoint.url: continue try: resp = await client.post( f"{endpoint.url}/api/chat", json={ "model": MODEL, "stream": False, # Ollama 0.24 separates deepseek-r1 thinking from final text. # Chat callers expect message.content to contain the answer. "think": False, "messages": [ {"role": "system", "content": system_prompt}, {"role": "user", "content": user_message}, ], "options": {"num_predict": 1200}, }, ) resp.raise_for_status() data = resp.json() raw = data.get("message", {}).get("content", "").strip() # 過濾 deepseek-r1 的 ... 推理區塊 text = re.sub(r".*?", "", raw, flags=re.DOTALL).strip() if not text: text = raw # 萬一全是 think block，直接回傳原文 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_ollama_usage", model=MODEL, provider=endpoint.provider_name, prompt_tokens=prompt_eval_count, output_tokens=eval_count, ) return f"{text}\n\n🦙 {MODEL} | {total_tokens} tokens | 免費" except Exception as e: logger.warning( "nemotron_chat_endpoint_failed", model=MODEL, provider=endpoint.provider_name, endpoint=endpoint.url, error=str(e), ) logger.warning("nemotron_chat_failed_all_endpoints", model=MODEL) return None async def generate_response( self, user_id: int, # noqa: ARG002 username: str, # noqa: ARG002 message_text: str, ) -> str: """ 根據訊息決定回應模式: @openclaw → 只有 OpenClaw 回應 @nemo → 只有 NemoClaw 回應 其他 → OpenClaw 先回，NemoClaw 異步補充 """ context = await self.get_system_context() text = message_text.strip() # 模式 1: 指定 OpenClaw if text.lower().startswith("@openclaw"): msg = text[9:].strip() or text result = await self._call_openclaw(f"{OPENCLAW_PERSONA}\n{context}", msg) return f"🦞 OpenClaw:\n{result or '🔴 OpenClaw 無響應'}" # 模式 2: 指定 NemoClaw if text.lower().startswith("@nemo"): msg = text[5:].strip() or text result = await self._call_nemotron(f"{NEMOCLAW_PERSONA}\n{context}", msg) return f"🤖 NemoClaw:\n{result or '🔴 NemoClaw 無響應 (NIM 超時)'}" # 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"老闆問了: {text}\n\n請從 NemoClaw 角度補充或評論。", ) ) # OpenClaw 最多等 40s（含 context 取得時間），NemoClaw 最多等 60s # 2026-04-03 ogt: 移除 asyncio.shield — shield 會在超時後讓 task 繼續跑但無人等待，造成 silent leak try: openclaw_raw = await asyncio.wait_for(openclaw_task, timeout=40.0) except TimeoutError: openclaw_raw = None openclaw_block = f"🦞 OpenClaw:\n{openclaw_raw or '🔴 無響應'}" try: nemo_raw = await asyncio.wait_for(nemo_task, timeout=60.0) except TimeoutError: nemo_raw = None if nemo_raw: return f"{openclaw_block}\n\n🤖 NemoClaw:\n{nemo_raw}" return openclaw_block # Singleton _chat_manager: ChatManager | None = None def get_chat_manager() -> ChatManager: global _chat_manager if _chat_manager is None: _chat_manager = ChatManager() return _chat_manager