""" OpenClaw AI Decision Engine - True LLM + SignOz Integration ============================================================ Phase 5: OpenClaw 實體化升級 (2026-03-21) 統帥校正: SignOz 為唯一全能視力中心 Features: - 真實 LLM SDK 整合 (告警預設 Ollama GCP-A → GCP-B → 111 → Gemini) - SignOz Gold Metrics 即時擷取 (P99/Error/RPS) - AIOps Agent 專業人格 (K8s 維運 + SRE RCA 專精) - 強制結構化 JSON 輸出 (符合 API 契約) - 動態告警上下文注入 + SignOz 數據 - Shadow Mode 調優指令生成 (日誌輸出，不執行) 防禦性工程鐵律: - Zero Trust: 預設不信任 LLM 輸出，必須通過 Pydantic 驗證 - Edge Case: 網路失敗、解析失敗、超時處理 - SignOz 失敗時優雅降級 (不阻塞主流程) """ import hashlib import json import random import re import time from datetime import datetime import httpx import structlog from src.core.config import settings from src.core.prompts import NEMOTRON_SYSTEM_PROMPT, OPENCLAW_SYSTEM_PROMPT from src.core.redis_client import get_redis from src.models.ai import ( AIRiskLevel, OpenClawDecision, SuggestedAction, ) from src.services.langfuse_client import langfuse_trace from src.services.model_registry import get_model_registry from src.services.ollama_failover_manager import get_ollama_failover_manager from src.services.signoz_client import GoldMetrics, get_signoz_client from src.utils.k8s_naming import normalize_resource_name from src.utils.timezone import now_taipei_iso logger = structlog.get_logger(__name__) # ============================================================================= # AIOps Agent System Prompt (專業人格 + 仲裁邏輯 + SignOz 數據) # ============================================================================= # 責任矩陣定義 RESPONSIBILITY_MATRIX = { "FE": "前端團隊 (Frontend)", "BE": "後端團隊 (Backend)", "INFRA": "基礎設施團隊 (Infrastructure/SRE)", "DB": "資料庫團隊 (Database/DBA)", "COLLAB": "協同處理 (需多團隊會診)", } # 信心度閾值 CONFIDENCE_THRESHOLD_COLLAB = 0.70 # 低於此閾值自動標記為 COLLAB # OPENCLAW_SYSTEM_PROMPT 已移至 src/core/prompts.py (Phase 17 P2 改進) # ============================================================================= # LLM Analysis Result - Using Pydantic for Schema Enforcement # ============================================================================= # We use OpenClawDecision from models/ai.py for Pydantic validation # This alias is for backwards compatibility LLMAnalysisResult = OpenClawDecision # ============================================================================= # kubectl_command 回填 helper # 2026-04-09 Claude Sonnet 4.6: I2 架構Review修復 — 補齊所有 tool 類型，消除兩處重複邏輯 (M3) # ============================================================================= def _backfill_kubectl_command(proposal: dict, tools: list) -> None: """將 AI tool call 結果回填為可執行的 kubectl_command。 proposal["kubectl_command"] 若已有值則不覆蓋（LLM 直接填的優先）。 """ if not tools or proposal.get("kubectl_command"): return _t = tools[0] _tool_name = _t.get("tool", "") _args = _t.get("args", {}) _ns = _args.get("namespace", proposal.get("namespace", "awoooi-prod")) if _tool_name == "restart_deployment": _deploy = _args.get("deployment_name", proposal.get("target_resource", "")) if _deploy: proposal["kubectl_command"] = f"kubectl rollout restart deployment/{_deploy} -n {_ns}" elif _tool_name == "delete_pod": _pod = _args.get("pod_name", "") if _pod: proposal["kubectl_command"] = f"kubectl delete pod {_pod} -n {_ns}" elif _tool_name == "scale_deployment": _deploy = _args.get("deployment_name", "") _replicas = _args.get("replicas", 2) if _deploy: proposal["kubectl_command"] = f"kubectl scale deployment/{_deploy} --replicas={_replicas} -n {_ns}" elif _tool_name == "delete_deployment": _deploy = _args.get("deployment_name", "") if _deploy: proposal["kubectl_command"] = f"kubectl delete deployment/{_deploy} -n {_ns}" elif _tool_name == "drain_node": _node = _args.get("node_name", "") if _node: proposal["kubectl_command"] = f"kubectl drain {_node} --ignore-daemonsets --delete-emptydir-data" elif _tool_name == "cordon_node": _node = _args.get("node_name", "") if _node: proposal["kubectl_command"] = f"kubectl cordon {_node}" elif _tool_name == "delete_service": _svc = _args.get("service_name", "") if _svc: proposal["kubectl_command"] = f"kubectl delete service/{_svc} -n {_ns}" # ============================================================================= # OpenClaw Service # ============================================================================= def _build_alert_cache_context_hash(alert_context: dict | None) -> str: """Build a stable LLM cache scope for repeat alerts without dynamic annotations.""" if not alert_context: return "" alertname = alert_context.get("alertname") or alert_context.get("alert_type", "") category = alert_context.get("alert_category", "") namespace = alert_context.get("namespace", "") target = alert_context.get("target_resource", "") severity = alert_context.get("severity", "") fingerprint = alert_context.get("fingerprint", "") return f"{alertname}:{category}:{namespace}:{target}:{severity}:{fingerprint}" class OpenClawService: """ OpenClaw AI 決策服務 - True LLM + SignOz Integration 實作 AI_FALLBACK_ORDER 備援機制。 告警/incident 上下文預設套用成本防線，只允許 Ollama GCP-A → GCP-B → 111。 新增 SignOz 整合: - 自動擷取 Gold Metrics - 數據驅動的 RCA 分析 - 動態 Trace URL 生成 """ def __init__(self): self._http_client: httpx.AsyncClient | None = None self._signoz = get_signoz_client() async def _get_client(self) -> httpx.AsyncClient: """取得 HTTP 客戶端 2026-04-14 Claude Sonnet 4.6: 從硬編 120s 改用 OPENCLAW_TIMEOUT 設定 (30s) 對齊 ADR-052 GAP-B4 的 25s + 5s buffer 設計。原 120s 違反 defense-in-depth。 """ if self._http_client is None or self._http_client.is_closed: _t = float(settings.OPENCLAW_TIMEOUT) self._http_client = httpx.AsyncClient( timeout=httpx.Timeout(_t, connect=10.0), ) return self._http_client async def close(self) -> None: """關閉連線""" if self._http_client: await self._http_client.aclose() self._http_client = None def _is_incident_alert_context(self, alert_context: dict | None) -> bool: """Return true when a request came from the alert/incident automation path.""" if not alert_context: return False alert_keys = { "alert_type", "alertname", "alert_name", "fingerprint", "incident_id", "severity", "signals", "target_resource", } return any(key in alert_context for key in alert_keys) def _cloud_fallback_allowed_for_alert(self, alert_context: dict | None) -> bool: """Cloud fallback is allowed after the ordered Ollama lane for alerts.""" if alert_context and alert_context.get("allow_cloud_fallback") is False: return False if not self._is_incident_alert_context(alert_context): return True return bool(getattr(settings, "ALERT_AI_ALLOW_CLOUD_FALLBACK", True)) def _alert_enforces_ollama_first(self, alert_context: dict | None) -> bool: """Alert and AI-governance lanes must try GCP-A/GCP-B/111 before Gemini backup.""" return ( bool(alert_context) and ( bool(alert_context.get("enforce_ollama_first")) or self._is_incident_alert_context(alert_context) ) and bool(getattr(settings, "ALERT_AI_ENFORCE_OLLAMA_FIRST", True)) ) async def _resolve_alert_provider_order( self, task_type: str = "diagnose", alert_context: dict | None = None, cloud_provider_order: list[str] | None = None, ) -> list[str]: """Resolve GCP-A/GCP-B/111, then OpenClaw/Nemo before Gemini backup.""" provider_order: list[str] = [] try: route = await get_ollama_failover_manager().select_provider(task_type=task_type) provider_order = [ endpoint.provider_name for endpoint in route.all_endpoints_in_order() if endpoint.provider_name.startswith("ollama") ] except Exception as route_error: logger.warning( "alert_ollama_route_lookup_failed", error=str(route_error), task_type=task_type, ) if not provider_order: provider_order = ["ollama_gcp_a", "ollama_gcp_b", "ollama_local"] deduped: list[str] = [] for provider_name in provider_order: if provider_name and provider_name not in deduped: deduped.append(provider_name) if not self._alert_enforces_ollama_first(alert_context): return deduped ollama_order = {"ollama_gcp_a": 0, "ollama_gcp_b": 1, "ollama_local": 2} ordered_ollama = [ provider_name for provider_name in deduped if provider_name in ollama_order ] ordered_ollama.sort(key=lambda provider_name: ollama_order[provider_name]) if not ordered_ollama: ordered_ollama = ["ollama_gcp_a", "ollama_gcp_b", "ollama_local"] if not self._cloud_fallback_allowed_for_alert(alert_context): return ordered_ollama cloud_aliases = {"nvidia": "openclaw_nemo"} cloud_candidates = { cloud_aliases.get(provider_name, provider_name) for provider_name in (cloud_provider_order or []) } try: from src.services.ai_control import is_provider_disabled if await is_provider_disabled("openclaw_nemo"): cloud_candidates.discard("openclaw_nemo") else: cloud_candidates.add("openclaw_nemo") except Exception as control_error: logger.warning("alert_openclaw_nemo_control_check_failed", error=str(control_error)) # Gemini remains the final paid backup, but alert traffic should use # OpenClaw/Nemo first whenever the router control plane has not disabled it. cloud_candidates.add("gemini") cloud_backup = [ provider_name for provider_name in ("openclaw_nemo", "gemini") if provider_name in cloud_candidates ] return ordered_ollama + cloud_backup # ========================================================================= # SignOz Integration # ========================================================================= async def get_signoz_context( self, service_name: str, namespace: str = "default", alert_timestamp: datetime | None = None, ) -> tuple[GoldMetrics | None, str]: """ 擷取 SignOz 上下文數據 Returns: (GoldMetrics, trace_url) or (None, fallback_url) """ try: metrics = await self._signoz.get_gold_metrics( service_name=service_name, namespace=namespace, time_window_minutes=10, ) trace_url = self._signoz.generate_trace_url( service_name=service_name, alert_timestamp=alert_timestamp, window_minutes=5, ) logger.info( "signoz_context_fetched", service=service_name, rps=metrics.rps, error_rate=metrics.error_rate, p99_latency=metrics.p99_latency_ms, ) return metrics, trace_url except Exception as e: logger.warning( "signoz_context_fetch_failed", service=service_name, error=str(e), ) # 降級: 返回 None 和靜態 URL fallback_url = f"{settings.SIGNOZ_URL}/traces?service={service_name}" return None, fallback_url def generate_auto_tuning_command( self, alert_type: str, target_resource: str, namespace: str, metrics: GoldMetrics | None = None, ) -> dict: """ 根據告警類型和 SignOz 數據生成調優指令 Shadow Mode: 僅生成指令，不執行 Phase 18.1.6: 整合 K8s 資源名稱驗證 (ADR-016) Returns: {command: str, description: str, type: str} """ # Phase 18.1.6: 先正規化資源名稱 normalized = normalize_resource_name(target_resource, namespace) if not normalized.is_k8s_resource: # 非 K8s 資源，返回提示訊息 logger.info( "non_k8s_resource_detected", original=target_resource, note=normalized.note, ) return { "type": "MANUAL", "command": f"# 非 K8s 資源: {target_resource}", "description": f"此資源不在 K8s 中，需人工處理。{normalized.note or ''}", } # 使用正規化後的名稱 resolved_name = normalized.normalized or target_resource resolved_ns = normalized.namespace or namespace if normalized.confidence < 0.8: logger.warning( "low_confidence_resource_name", original=target_resource, resolved=resolved_name, confidence=normalized.confidence, ) # 根據告警類型選擇調優策略 (使用正規化後的名稱) if "cpu" in alert_type.lower() or "high_cpu" in alert_type.lower(): # CPU 高 → 擴容或調整 limit if metrics and metrics.rps > 100: # 高流量場景 → HPA return { "type": "HPA", "command": f"kubectl autoscale deployment {resolved_name} --cpu-percent=70 --min=2 --max=10 -n {resolved_ns}", "description": f"SignOz RPS={metrics.rps:.0f}，配置 HPA 應對流量波動", } else: # 低流量但 CPU 高 → 調整資源 return { "type": "RESOURCE_LIMIT", "command": f"kubectl set resources deployment/{resolved_name} --limits=cpu=2000m -n {resolved_ns}", "description": "增加 CPU limit 緩解資源競爭", } elif "memory" in alert_type.lower() or "oom" in alert_type.lower(): return { "type": "RESOURCE_LIMIT", "command": f"kubectl set resources deployment/{resolved_name} --limits=memory=1Gi -n {resolved_ns}", "description": "增加 Memory limit 防止 OOM", } elif "pod_crash" in alert_type.lower() or "crash" in alert_type.lower(): return { "type": "RESTART", "command": f"kubectl rollout restart deployment/{resolved_name} -n {resolved_ns}", "description": "滾動重啟清除異常狀態", } elif "latency" in alert_type.lower() or "slow" in alert_type.lower(): if metrics and metrics.p99_latency_ms > 500: return { "type": "SCALE", "command": f"kubectl scale deployment {resolved_name} --replicas=+2 -n {resolved_ns}", "description": f"SignOz P99={metrics.p99_latency_ms:.0f}ms，擴容分散負載", } else: return { "type": "CACHE", "command": "# 檢查 Redis 連線池配置", "description": "建議增加緩存層減少後端壓力", } else: # 通用: 滾動重啟 return { "type": "RESTART", "command": f"kubectl rollout restart deployment/{resolved_name} -n {resolved_ns}", "description": "滾動重啟恢復服務", } # ========================================================================= # ========================================================================= # OpenClaw (Nemo) 委派仲裁 — 架構鐵律 # 2026-04-01 ogt: AWOOOI API 不直接打 LLM，委派給 OpenClaw (192.168.0.188:8089) # ========================================================================= async def _call_openclaw_analyze( self, incident_id: str, severity: str, signals: list[dict], affected_services: list[str], expert_context: dict | None = None, ) -> dict | None: """ 委派 Incident RCA 給 OpenClaw (Nemo) API Returns: proposal_dict if success, None if failed (fallback to direct LLM) """ try: client = await self._get_client() import json as _json def _to_serializable(obj): if isinstance(obj, dict): return {k: _to_serializable(v) for k, v in obj.items()} if isinstance(obj, list): return [_to_serializable(i) for i in obj] try: _json.dumps(obj) return obj except (TypeError, ValueError): return str(obj) payload = { "incident_id": incident_id, "severity": severity, "signals": _to_serializable(signals[:5]), "affected_services": affected_services, "expert_context": _to_serializable(expert_context) if expert_context else None, } # 2026-04-14 Claude Sonnet 4.6: 從硬編 130s 改用 OPENCLAW_TIMEOUT # 原 130s 讓 LLM 能卡 2m10s，超過 Ollama 真實返回時間（P95 54s） resp = await client.post( f"{settings.OPENCLAW_URL}/api/v1/analyze/incident", json=payload, timeout=httpx.Timeout(float(settings.OPENCLAW_TIMEOUT), connect=5.0), ) resp.raise_for_status() data = resp.json() # 驗證必要欄位 if not data.get("action_title") or not data.get("risk_level"): logger.warning("openclaw_analyze_invalid_response", incident_id=incident_id) return None try: confidence_value = float(data.get("confidence", 0.0)) except (TypeError, ValueError): confidence_value = 0.0 if data.get("degraded") or data.get("provider") == "openclaw_degraded" or confidence_value < 0.3: logger.warning( "openclaw_analyze_degraded_response", incident_id=incident_id, provider=data.get("provider"), confidence=data.get("confidence"), ) return None logger.info( "openclaw_analyze_success", incident_id=incident_id, confidence=data.get("confidence", 0), risk_level=data.get("risk_level"), ) # 轉換為 AWOOOI API proposal dict 格式 return { "action": data.get("action_title", "AI 分析"), "description": data.get("description", ""), "kubectl_command": data.get("kubectl_command") or "", "target_resource": data.get("target_resource", "unknown"), "namespace": data.get("namespace", "awoooi-prod"), "risk_level": data.get("risk_level", "medium"), "reasoning": data.get("reasoning", ""), "confidence": data.get("confidence", 0.8), "primary_responsibility": data.get("primary_responsibility", "INFRA"), "optimization_suggestions": data.get("optimization_suggestions", []), "signoz_correlation": data.get("signoz_correlation", ""), "from_cache": False, "provider": "openclaw_nemo", "ai_tokens": 0, "ai_cost": 0.0, } except Exception as e: logger.warning( "openclaw_analyze_failed", incident_id=incident_id, error=str(e), reason="fallback to direct LLM", ) return None # ========================================================================= # [ARCHIVED Phase 24 B4 — 2026-04-03 ogt] # 以下三個方法 (_call_ollama/_call_gemini/_call_claude) 為舊版 fallback chain # 新路徑: USE_AI_ROUTER=true → _call_with_fallback → AIRouterExecutor (ai_router.py) # 新 Provider 實作: apps/api/src/services/ai_providers/ (OllamaProvider/GeminiProvider/ClaudeProvider) # 回滾保留: USE_AI_ROUTER=false 時仍由 _call_with_fallback (line ~993) 呼叫此區塊 # 完整移除時機: Phase 24 完整驗收後 (ADR-052 D11) # ========================================================================= async def _call_ollama(self, prompt: str, *, ollama_only: bool = False) -> tuple[str, bool]: """ 呼叫 Ollama (支援 JSON Mode)。 USE_AI_ROUTER=true 正常會走 AIRouterExecutor；這裡是 legacy safety-net。 2026-05-05 Codex: safety-net 也必須遵守 ADR-110 三層 Ollama 路由，告警路徑預設只允許 GCP-A/GCP-B/111，不能只打 OLLAMA_URL 後直接掉 Gemini。 """ try: client = await self._get_client() # 從 ModelRegistry 取得模型配置 registry = get_model_registry() model_name = registry.get_model("ollama", "rca") if ollama_only: model_name = getattr(settings, "ALERT_OLLAMA_MODEL", "qwen3:14b") options = registry.get_provider_options("ollama") timeout_seconds = max( float(settings.OPENCLAW_TIMEOUT), float(getattr(settings, "OLLAMA_DIAGNOSE_TIMEOUT_SECONDS", settings.OPENCLAW_TIMEOUT)), ) endpoints: list[tuple[str, str]] = [] try: route = await get_ollama_failover_manager().select_provider() endpoints = [ (endpoint.provider_name, endpoint.url) for endpoint in route.all_endpoints_in_order() if endpoint.provider_name.startswith("ollama") and endpoint.url ] except Exception as route_error: logger.warning( "legacy_ollama_route_lookup_failed", error=str(route_error), ) if not endpoints: configured_endpoints = [ ("ollama_gcp_a", settings.OLLAMA_URL), ("ollama_gcp_b", getattr(settings, "OLLAMA_SECONDARY_URL", "")), ("ollama_local", getattr(settings, "OLLAMA_FALLBACK_URL", "")), ] seen_urls: set[str] = set() endpoints = [] for provider_name, endpoint_url in configured_endpoints: if endpoint_url and endpoint_url not in seen_urls: endpoints.append((provider_name, endpoint_url)) seen_urls.add(endpoint_url) if ollama_only: allowed_provider_order = {"ollama_gcp_a": 0, "ollama_gcp_b": 1, "ollama_local": 2} endpoints = [ (provider_name, endpoint_url) for provider_name, endpoint_url in endpoints if provider_name in allowed_provider_order ] endpoints.sort(key=lambda item: allowed_provider_order[item[0]]) if not endpoints: endpoints = [ ("ollama_gcp_a", settings.OLLAMA_URL), ("ollama_gcp_b", getattr(settings, "OLLAMA_SECONDARY_URL", "")), ("ollama_local", getattr(settings, "OLLAMA_FALLBACK_URL", "")), ] endpoints = [ (provider_name, endpoint_url) for provider_name, endpoint_url in endpoints if endpoint_url ] last_error = "" for provider_name, endpoint_url in endpoints: try: logger.info( "ollama_request_start", provider=provider_name, url=f"{endpoint_url}/api/generate", prompt_length=len(prompt), ) response = await client.post( f"{endpoint_url}/api/generate", json={ "model": model_name, "prompt": prompt, "stream": False, "format": "json", # 強制 JSON 輸出 "options": { "num_predict": options.get("num_predict", 1024), "temperature": options.get("temperature", 0.1), "top_p": options.get("top_p", 0.9), }, }, timeout=httpx.Timeout(timeout_seconds, connect=10.0), ) logger.info( "ollama_response_received", provider=provider_name, status_code=response.status_code, ) response.raise_for_status() data = response.json() result = data.get("response", "") logger.info( "ollama_response_parsed", provider=provider_name, response_length=len(result), ) return result, True except httpx.TimeoutException as e: last_error = f"{provider_name} timeout: {e}" logger.warning("ollama_timeout", provider=provider_name, error=str(e)) except Exception as e: last_error = f"{provider_name} failed: {e}" logger.warning( "ollama_call_failed", provider=provider_name, error=str(e), error_type=type(e).__name__, ) return last_error or "all Ollama endpoints failed", False except httpx.TimeoutException as e: logger.warning("ollama_timeout", error=str(e)) return f"Timeout: {e}", False except Exception as e: logger.warning( "ollama_call_failed", error=str(e), error_type=type(e).__name__, ) return str(e), False async def _call_gemini(self, prompt: str) -> tuple[str, bool, int, float]: """ 呼叫 Google Gemini (支援 JSON Mode) Returns: tuple: (response_text, success, total_tokens, cost_usd) - response_text: LLM 回應文本 - success: 是否成功 - total_tokens: 使用的 Token 總數 - cost_usd: 預估成本 (USD) 2026-03-29 ogt: 加入 Token/Cost 追蹤 """ if not settings.GEMINI_API_KEY: return "GEMINI_API_KEY not configured", False, 0, 0.0 try: client = await self._get_client() # 從 ModelRegistry 取得模型配置 registry = get_model_registry() model_name = registry.get_model("gemini", "rca") response = await client.post( f"https://generativelanguage.googleapis.com/v1beta/models/{model_name}:generateContent", headers={"x-goog-api-key": settings.GEMINI_API_KEY}, json={ "contents": [{"parts": [{"text": prompt}]}], "generationConfig": { "temperature": 0.1, "maxOutputTokens": 2048, "responseMimeType": "application/json", # 強制 JSON 輸出 }, }, timeout=30.0, ) response.raise_for_status() data = response.json() text = data["candidates"][0]["content"]["parts"][0]["text"] # 2026-03-29 ogt: 擷取 Token 使用量 usage_metadata = data.get("usageMetadata", {}) prompt_tokens = usage_metadata.get("promptTokenCount", 0) completion_tokens = usage_metadata.get("candidatesTokenCount", 0) total_tokens = usage_metadata.get("totalTokenCount", prompt_tokens + completion_tokens) # Gemini 1.5 Flash 定價 (per 1M tokens) # Input: $0.075 / 1M, Output: $0.30 / 1M cost_usd = (prompt_tokens * 0.000000075) + (completion_tokens * 0.0000003) logger.info( "gemini_response_received", response_length=len(text), prompt_tokens=prompt_tokens, completion_tokens=completion_tokens, total_tokens=total_tokens, cost_usd=f"${cost_usd:.6f}", ) return text, True, total_tokens, cost_usd except Exception as e: logger.warning("gemini_call_failed", error=str(e)) return str(e), False, 0, 0.0 async def _call_claude(self, prompt: str) -> tuple[str, bool]: """ 呼叫 Anthropic Claude (使用 Tool Use 強制 JSON) """ if not settings.CLAUDE_API_KEY: return "CLAUDE_API_KEY not configured", False try: client = await self._get_client() # Claude 使用 Tool Use 強制結構化輸出 response = await client.post( "https://api.anthropic.com/v1/messages", headers={ "x-api-key": settings.CLAUDE_API_KEY, "anthropic-version": "2023-06-01", "content-type": "application/json", }, json={ "model": get_model_registry().get_model("claude", "rca"), "max_tokens": 2048, "messages": [{"role": "user", "content": prompt}], "tools": [{ "name": "submit_analysis", "description": "Submit the RCA analysis result in structured format", "input_schema": { "type": "object", "properties": { "action_title": {"type": "string"}, "description": {"type": "string"}, "suggested_action": {"type": "string", "enum": ["RESTART_DEPLOYMENT", "DELETE_POD", "SCALE_DEPLOYMENT", "APPLY_HPA", "TUNE_RESOURCES", "INVESTIGATE", "OBSERVE", "NO_ACTION"]}, "kubectl_command": {"type": "string"}, "target_resource": {"type": "string"}, "namespace": {"type": "string"}, "risk_level": {"type": "string", "enum": ["low", "medium", "critical"]}, "blast_radius": { "type": "object", "properties": { "affected_pods": {"type": "integer"}, "estimated_downtime": {"type": "string"}, "related_services": {"type": "array", "items": {"type": "string"}}, "data_impact": {"type": "string", "enum": ["NONE", "READ_ONLY", "WRITE", "DESTRUCTIVE"]} }, "required": ["affected_pods", "estimated_downtime", "related_services", "data_impact"] }, "reasoning": {"type": "string"}, "deviation_analysis": {"type": "string"}, "confidence": {"type": "number"}, "affected_services": {"type": "array", "items": {"type": "string"}} }, "required": ["action_title", "description", "suggested_action", "kubectl_command", "target_resource", "namespace", "risk_level", "blast_radius", "reasoning", "confidence"] } }], "tool_choice": {"type": "tool", "name": "submit_analysis"}, }, timeout=30.0, ) response.raise_for_status() data = response.json() # 從 Tool Use 回應中提取 JSON for block in data.get("content", []): if block.get("type") == "tool_use" and block.get("name") == "submit_analysis": tool_input = block.get("input", {}) logger.info("claude_tool_use_response", input_keys=list(tool_input.keys())) return json.dumps(tool_input), True # Fallback: 嘗試從 text 內容提取 for block in data.get("content", []): if block.get("type") == "text": return block.get("text", ""), True return "No valid response from Claude", False except Exception as e: logger.warning("claude_call_failed", error=str(e)) return str(e), False # 2026-03-29 ogt: _call_nvidia 已移至 nvidia_provider.py (ARCHIVED) # 符合模組化規範 - 所有 NVIDIA API 呼叫統一由 NvidiaProvider / OpenClawNemoProvider 處理 # ========================================================================= # [END ARCHIVED Phase 24 B4] # ========================================================================= # ========================================================================= # Mock LLM - Intelligent Fallback with SignOz Data # ========================================================================= def _generate_mock_response( self, alert_context: dict, signoz_metrics: GoldMetrics | None = None, ) -> str: """ 規則引擎降級回應 (v8.0) — 生產用途，不是假數據 從 alert_rules.yaml 載入規則進行匹配，AI 分析失敗時的正式降級路徑。 命中 generic_fallback 時會回傳 rule_id="generic_fallback"， 由上層 async 方法（_call_with_fallback）觸發 auto_generate_rule() 學習新規則。 Returns: (json_str, rule_id) tuple 2026-04-09 ogt: 重構為規則引擎，移除 if/elif 硬編碼 2026-04-09 ogt: S2-4 架構師審查 — 修正 Mock 語意混淆，澄清為規則引擎生產路徑 """ from src.services.alert_rule_engine import match_rule time.sleep(random.uniform(0.3, 0.8)) # 模擬思考延遲 # SignOz 數據整合 signoz_correlation = "SignOz 數據擷取中..." if signoz_metrics: signoz_correlation = ( f"RPS={signoz_metrics.rps:.1f} ({signoz_metrics.rps_trend}), " f"Error={signoz_metrics.error_rate:.2f}%, " f"P99={signoz_metrics.p99_latency_ms:.0f}ms" ) mock_response = match_rule(alert_context) if mock_response is None: # match_rule 不應該回傳 None（有通用兜底），但防禦性處理 alert_type = alert_context.get("alert_type", "custom") target = alert_context.get("target_resource", "unknown") namespace = alert_context.get("namespace", "awoooi-prod") mock_response = { "action_title": f"重新啟動 {target} 服務", "description": f"⚙️ 規則匹配: {target} 發生異常，需進一步診斷確認根因。", "suggested_action": "RESTART_DEPLOYMENT", "kubectl_command": f"kubectl rollout restart deployment/{target} -n {namespace}", "target_resource": target, "namespace": namespace, "risk_level": "medium", "blast_radius": { "affected_pods": 1, "estimated_downtime": "5-15 min", "related_services": [target], "data_impact": "NONE", }, "primary_responsibility": "COLLAB", "responsibility_reasoning": "告警資訊不足，建議多團隊協同排查", "secondary_teams": ["BE", "INFRA"], "optimization_suggestions": [], "reasoning": f"[規則匹配] 根據告警 {alert_type} 先重啟恢復服務。", "deviation_analysis": "監控指標顯示異常偏離基準線", "confidence": 0.0, "affected_services": [target], "signoz_correlation": signoz_correlation, } # 補充 SignOz 關聯資訊（規則引擎不持有 signoz_metrics） mock_response["signoz_correlation"] = signoz_correlation if signoz_metrics: mock_response["description"] += f" {signoz_metrics.to_summary()}" rule_id = mock_response.get("rule_id", "unknown") logger.info( "mock_llm_response_generated", rule_id=rule_id, action_title=mock_response["action_title"], risk_level=mock_response["risk_level"], primary_responsibility=mock_response["primary_responsibility"], confidence=mock_response["confidence"], signoz_integrated=signoz_metrics is not None, is_mock=True, ) # 2026-04-09 ogt: rule_id 回傳給上層 async 方法觸發自動規則生成 # 不在此 sync 方法中呼叫 asyncio，避免 event loop 混用問題 (S1-1 架構師審查) return json.dumps(mock_response), rule_id # ========================================================================= # LLM Cache Layer (憲法要求: 嚴禁無快取裸奔) # ========================================================================= def _generate_cache_key(self, prompt: str, context_hash: str = "") -> str: """ 生成 LLM 快取鍵 有告警上下文時，使用 prompt family + 穩定告警維度，避免 annotations / SignOz 即時數值讓同一告警每 20 秒打穿快取；沒有上下文時仍用完整 prompt。 """ if context_hash: prompt_family_source = ( "openclaw_alert_analysis" if "## Alert Data:" in prompt else prompt[:512] ) prompt_family = hashlib.sha256(prompt_family_source.encode()).hexdigest()[:8] content = f"{prompt_family}:{context_hash}" else: content = prompt hash_digest = hashlib.sha256(content.encode()).hexdigest()[:16] return f"llm_cache:{hash_digest}" async def _call_with_cache( self, prompt: str, alert_context: dict | None = None, signoz_metrics: GoldMetrics | None = None, cache_ttl: int = 3600, # 1 hour default ) -> tuple[str, str, bool, bool, int, float]: """ 帶快取的 LLM 呼叫包裝器 憲法條款: 必須使用快取保護算力資源 Args: prompt: LLM prompt alert_context: 告警上下文 signoz_metrics: SignOz 指標 cache_ttl: 快取存活時間 (秒) Returns: (response, provider, success, from_cache, total_tokens, cost_usd) 2026-03-29 ogt: 加入 Token/Cost 追蹤 """ # 生成快取鍵 (基於 prompt + alert_context hash) # 2026-04-16 ogt + Claude Sonnet 4.6: 修復 — alertname 才是主要識別符 # 舊版用 alert_type:target_resource → 不同告警 (e.g. PostgreSQLDiskGrowth vs PodCrashLoop) # 在 alert_type="custom" 時共用同一快取鍵 → 全部回傳相同 LLM 結果 context_hash = _build_alert_cache_context_hash(alert_context) cache_key = self._generate_cache_key(prompt, context_hash) # 1. 嘗試從快取讀取 try: redis_client = get_redis() cached = await redis_client.get(cache_key) if cached: cached_data = json.loads(cached) logger.info( "llm_cache_hit", cache_key=cache_key[:20], provider=cached_data.get("provider", "cached"), ) return ( cached_data["response"], f"{cached_data['provider']}_cached", True, True, # from_cache 0, # tokens (cache hit, no new tokens) 0.0, # cost (cache hit, no cost) ) except Exception as e: logger.warning("llm_cache_read_failed", error=str(e)) # 2. Cache Miss - 呼叫 LLM logger.info("llm_cache_miss", cache_key=cache_key[:20]) response, provider, success, total_tokens, cost_usd = await self._call_with_fallback( prompt, alert_context, signoz_metrics ) # 3. 成功則寫入快取 if success: try: redis_client = get_redis() cache_data = { "response": response, "provider": provider, "cached_at": now_taipei_iso(), } await redis_client.set( cache_key, json.dumps(cache_data, ensure_ascii=False), ex=cache_ttl, ) logger.info( "llm_cache_write", cache_key=cache_key[:20], provider=provider, ttl=cache_ttl, ) except Exception as e: logger.warning("llm_cache_write_failed", error=str(e)) return response, provider, success, False, total_tokens, cost_usd # from_cache=False # ========================================================================= # Public LLM Interface (ILLMProvider Protocol) # ========================================================================= async def call( self, prompt: str, alert_context: dict | None = None, ) -> tuple[str, str, bool]: """ 呼叫 LLM (ILLMProvider Protocol 實作) #39 Error Analyzer Agent 使用此方法 Args: prompt: 完整的 prompt alert_context: 可選的告警上下文（含 incident_id/signals， 供 OPENCLAW_NEMO provider 使用結構化 API） Returns: (response, provider, success) 2026-04-16 ogt + Claude Sonnet 4.6: 修復 — _call_with_fallback 回傳 5 值， call() 回傳 3 值，避免呼叫端 (diagnostician_agent:107) ValueError 2026-04-16 ogt + Claude Sonnet 4.6: 加入 alert_context — 讓 diagnostician 能傳 snapshot 結構化資料給 OPENCLAW_NEMO，避免 prompt 被截斷為 garbage """ response, provider, success, _tokens, _cost = await self._call_with_fallback( prompt, alert_context=alert_context ) return response, provider, success # ========================================================================= # Fallback Chain # ========================================================================= async def _call_with_fallback( self, prompt: str, alert_context: dict | None = None, signoz_metrics: GoldMetrics | None = None, ) -> tuple[str, str, bool, int, float]: """ 依 AI_FALLBACK_ORDER 順序呼叫 AI 若 MOCK_MODE=True，直接回傳模擬結果。 若所有 Provider 失敗，fallback 到 Mock。 Returns: tuple: (response, provider, success, total_tokens, cost_usd) Phase 15.1: 整合 Langfuse LLMOps 追蹤 2026-03-29 ogt: 加入 Token/Cost 追蹤 2026-04-02 ogt: Phase 24 ADR-052 絞殺者包裝 — USE_AI_ROUTER 新舊並存 """ # ================================================================= # Phase 24 ADR-052: 絞殺者分支 (Strangler Fig) # USE_AI_ROUTER=true → 新 AIRouterExecutor 路由 # USE_AI_ROUTER=false → 舊 if/else fallback chain (現狀) # 回滾: kubectl set env deployment/awoooi-api USE_AI_ROUTER=false # Phase 24 C: Redis 狀態覆蓋 env var (/ai router on/off) # ================================================================= # Redis 狀態優先 (Phase 24 C — 2026-04-03 ogt) _use_ai_router = settings.USE_AI_ROUTER try: from src.services.ai_control import get_ai_router_enabled _redis_override = await get_ai_router_enabled() if _redis_override is not None: _use_ai_router = _redis_override except Exception: pass if _use_ai_router: try: # 2026-04-02 ogt: C2 修復 — 呼叫 AIRouter.route() 智慧路由 (非靜態 order) # D1 意圖分類路由、D7 隱私保護 (DIAGNOSE/CODE_REVIEW 強制 local) 生效 from src.services.ai_router import ( IntentType, get_ai_executor, get_ai_router, ) router = get_ai_router() executor = get_ai_executor() # Step 1: 取得路由決策 (含意圖分類 + 複雜度評分) decision = await router.route(prompt, alert_context) # Step 2: 從 RoutingDecision 建立 provider_order (主 + fallback) # Phase 24 C: Redis primary_provider 覆蓋路由決策 provider_order = [decision.selected_provider.value] + [ p.value for p, _ in decision.fallback_chain if p.value != decision.selected_provider.value ] try: from src.services.ai_control import ( get_primary_provider, is_provider_disabled, ) _primary = await get_primary_provider() if _primary and _primary != decision.selected_provider.value: # 把 primary 移到首位 (保留原始 fallback) provider_order = [_primary] + [p for p in provider_order if p != _primary] # 過濾被停用的 Provider # C2 修復 (2026-04-03 首席架構師審查): Python 3.11 不支援 list comprehension 中 await _filtered = [] for _p in provider_order: if not await is_provider_disabled(_p): _filtered.append(_p) if _filtered: provider_order = _filtered except Exception as _e: logger.warning("ai_control_override_failed", error=str(_e)) if self._alert_enforces_ollama_first(alert_context): original_provider_order = list(provider_order) provider_order = await self._resolve_alert_provider_order( task_type=decision.intent.value if decision.intent else "diagnose", alert_context=alert_context, cloud_provider_order=original_provider_order, ) logger.info( "alert_ollama_first_provider_order", original_provider_order=original_provider_order, provider_order=provider_order, cloud_fallback_allowed=self._cloud_fallback_allowed_for_alert(alert_context), ) # Step 3: D7 隱私 — CODE_REVIEW 強制 local # 2026-04-15 ogt: DIAGNOSE 移除 require_local（v4.3 決策：NIM 為主力，無隱私問題） # ai_router.py v4.3 已明確：「NIM 從 Phase 22 起就是主力，無隱私問題」 # require_local=True 對 DIAGNOSE 只會讓所有 provider 被 privacy_skip → 永遠失敗 require_local = decision.intent in (IntentType.CODE_REVIEW,) # 2026-04-29 ogt + Claude Code: 注入 task_type 讓 Ollama 用正確 timeout # 根因: ai_providers/ollama.py:77 讀 context["task_type"] 決定 timeout # - "diagnose"/"force_local" → OLLAMA_DIAGNOSE_TIMEOUT_SECONDS # - 其他/未注入 → OPENCLAW_TIMEOUT=30s（不夠 qwen2.5:7b 推理） # webhooks alert_context 從未注入 task_type → Ollama fallback 永遠 30s timeout # 對齊 decision.intent 後 Ollama fallback 真正能跑完 exec_context = dict(alert_context) if alert_context else {} if decision.intent == IntentType.DIAGNOSE: exec_context["task_type"] = "diagnose" if self._alert_enforces_ollama_first(alert_context): exec_context.setdefault("task_type", "diagnose") exec_context.setdefault("ollama_model", getattr(settings, "ALERT_OLLAMA_MODEL", "qwen3:14b")) exec_context["allow_gcp_heavy_model"] = bool( exec_context.get("allow_gcp_heavy_model", True) ) exec_context["alert_requires_ollama_before_cloud"] = True result = await executor.execute( prompt=prompt, provider_order=provider_order, context=exec_context, cache_ttl=3600, require_local=require_local, ) logger.info( "phase24_ai_router_used", provider=result.provider, success=result.success, latency_ms=round(result.latency_ms, 1), intent=decision.intent.value, routing_reason=decision.routing_reason, ) return result.raw_response, result.provider, result.success, result.tokens, result.cost_usd except Exception as e: # AIRouter 失敗時 fallback 到舊路徑 (安全網) logger.warning("phase24_ai_router_fallback_to_legacy", error=str(e)) # Mock Mode: 開發測試用 if settings.MOCK_MODE: logger.info("mock_mode_enabled", using="mock_llm") _mock_json, _rule_id = self._generate_mock_response(alert_context or {}, signoz_metrics) if _rule_id == "generic_fallback": import asyncio from src.services.alert_rule_engine import auto_generate_rule try: asyncio.create_task(auto_generate_rule( alert_context or {}, ollama_url=settings.OLLAMA_URL, model=settings.OPENCLAW_DEFAULT_MODEL, gemini_api_key=( getattr(settings, "GEMINI_API_KEY", "") if self._cloud_fallback_allowed_for_alert(alert_context) else "" ), )) except Exception as _e: logger.warning("auto_rule_trigger_failed", error=str(_e)) return _mock_json, "mock", True, 0, 0.0 # Phase 15.1 + 15.3: Langfuse 追蹤整合 + SignOz Deep Linking with langfuse_trace( "openclaw_fallback_chain", metadata={ "prompt_length": len(prompt), "fallback_order": settings.AI_FALLBACK_ORDER, "alert_fingerprint": (alert_context or {}).get("fingerprint", "unknown"), }, ) as trace: # Phase 15.3: SignOz → Langfuse 反向連結 # 在當前 OTEL span 中記錄 Langfuse trace_id if trace.langfuse_trace_id: from opentelemetry import trace as otel_trace from src.core.deep_linking import DeepLinking current_span = otel_trace.get_current_span() if current_span: current_span.set_attribute("langfuse.trace_id", trace.langfuse_trace_id) current_span.set_attribute( "langfuse.trace_url", DeepLinking.langfuse_trace_url(trace.langfuse_trace_id), ) # Phase 13.2: Rate Limiter 整合 (2026-03-26) # 防止雲端 API 用量暴衝，超限自動降級 from src.services.ai_rate_limiter import get_ai_rate_limiter rate_limiter = get_ai_rate_limiter() legacy_provider_order = list(settings.AI_FALLBACK_ORDER) if self._alert_enforces_ollama_first(alert_context): legacy_provider_order = ["ollama"] if self._cloud_fallback_allowed_for_alert(alert_context): legacy_provider_order.append("gemini") logger.info( "legacy_alert_ollama_first_provider_order", provider_order=legacy_provider_order, cloud_fallback_allowed=self._cloud_fallback_allowed_for_alert(alert_context), ) for provider in legacy_provider_order: # Rate Limit 檢查 (nvidia/gemini/claude 需檢查，ollama 不限) # 2026-03-30 ogt: 加入 nvidia (RPM=5 限制) if provider in ("nvidia", "gemini", "claude"): allowed, reason = await rate_limiter.check_and_increment(provider) if not allowed: logger.warning( "ai_rate_limit_skip", provider=provider, reason=reason, ) continue # 跳過此 provider，嘗試下一個 logger.info("ai_provider_attempt", provider=provider) start_time = time.time() model_name = self._get_model_name(provider) # 2026-03-29 ogt: Gemini 回傳 4 值 (含 token/cost)，其他 Provider 補 0 total_tokens = 0 cost_usd = 0.0 if provider == "ollama": response, success = await self._call_ollama( prompt, ollama_only=self._alert_enforces_ollama_first(alert_context), ) elif provider == "gemini": response, success, total_tokens, cost_usd = await self._call_gemini(prompt) elif provider == "nvidia": # 2026-03-29 ogt: 使用 NvidiaProvider.chat() (模組化規範) from src.services.nvidia_provider import get_nvidia_provider nvidia_provider = get_nvidia_provider() response, success, total_tokens, cost_usd = await nvidia_provider.chat(prompt, use_json_mode=True) elif provider == "claude": response, success = await self._call_claude(prompt) else: logger.warning("unknown_ai_provider", provider=provider) continue latency_ms = (time.time() - start_time) * 1000 # Langfuse: 記錄每次 LLM 呼叫 trace.generation( name=f"{provider}_call", model=model_name, input=prompt[:500], # 截斷避免過長 output=response[:500] if success else f"ERROR: {response[:200]}", metadata={ "success": success, "latency_ms": round(latency_ms, 2), "provider": provider, "total_tokens": total_tokens, "cost_usd": cost_usd, }, ) if success: logger.info( "ai_provider_success", provider=provider, latency_ms=latency_ms, total_tokens=total_tokens, cost_usd=f"${cost_usd:.6f}", ) # Langfuse: 記錄成功評分 trace.score(name="provider_success", value=1.0, comment=f"Success via {provider}") # 2026-03-29 ogt: 記錄累積成本 (Gemini/Claude) if cost_usd > 0: await rate_limiter.record_cost(provider, cost_usd) return response, provider, True, total_tokens, cost_usd logger.warning("ai_provider_failed_fallback", provider=provider, latency_ms=latency_ms) # 所有 Provider 失敗時，fallback 到 Mock (優雅降級) logger.warning("all_providers_failed_using_mock", fallback="mock_llm") trace.score(name="provider_success", value=0.0, comment="All providers failed, using mock") _mock_json, _rule_id = self._generate_mock_response(alert_context or {}, signoz_metrics) if _rule_id == "generic_fallback": import asyncio from src.services.alert_rule_engine import auto_generate_rule try: asyncio.create_task(auto_generate_rule( alert_context or {}, ollama_url=settings.OLLAMA_URL, model=settings.OPENCLAW_DEFAULT_MODEL, gemini_api_key=( getattr(settings, "GEMINI_API_KEY", "") if self._cloud_fallback_allowed_for_alert(alert_context) else "" ), )) except Exception as _e: logger.warning("auto_rule_trigger_failed", error=str(_e)) return _mock_json, "mock_fallback", True, 0, 0.0 def _get_model_name(self, provider: str) -> str: """取得 provider 對應的模型名稱 (從 ModelRegistry)""" registry = get_model_registry() return registry.get_model(provider, "rca") # ========================================================================= # Response Parsing (防禦性解析) # ========================================================================= def _extract_json_from_response(self, text: str) -> str | None: """從 LLM 回應中提取 JSON (含啟發式修補)""" # 0. 清理開頭結尾空白 text = text.strip() if not text: return None # 1. 嘗試直接解析 try: json.loads(text) return text except json.JSONDecodeError: pass # 2. 嘗試從 markdown code block 提取 patterns = [ r"```json\s*([\s\S]*?)\s*```", r"```\s*([\s\S]*?)\s*```", r"(\{[\s\S]*\})", # 貪婪匹配最大括號對 ] for pattern in patterns: match = re.search(pattern, text) if match: candidate = match.group(1) if "(" in pattern else match.group(0) candidate = candidate.strip() try: json.loads(candidate) return candidate except json.JSONDecodeError: # 3. 啟發式修補: 如果結尾缺少括號，嘗試補齊 if candidate.startswith("{") and not candidate.endswith("}"): for i in range(1, 5): # 嘗試補 1-5 個括號/引號 try: repaired = candidate + '"' * (i - 1) + "}" * i json.loads(repaired) logger.info("json_repaired_heuristically", level=i) return repaired except json.JSONDecodeError: continue continue # 4. 極端情況: 找出最後一個有效 key if "{" in text: start_idx = text.find("{") candidate = text[start_idx:] # 暴力去除非法尾綴 (如 \t\t...) candidate = re.sub(r"[ \t\r\n]+$", "", candidate) if not candidate.endswith("}"): candidate += '"}' # 嘗試最簡單的閉合 try: json.loads(candidate) return candidate except json.JSONDecodeError: pass return None def _validate_deployment_inventory( self, result: "OpenClawDecision | None", k8s_inventory: str, k8s_ns: str, ) -> None: """ 2026-04-19 ogt + Claude Opus 4.7 (抽取自 analyze_alert): 幻覺 deployment 名偵測與降級。雙路徑共用(analyze_alert + generate_incident_proposal)。 根因: NEMOTRON 即使 prompt 含 inventory 仍會拿 namespace 當 deployment 名 → 執行 kubectl rollout restart deployment/awoooi-prod → "not found" 修復: 正則抽出 kubectl 指令的 deployment 名,對照 inventory 白名單; 不在白名單 → 降級為 NO_ACTION + 轉純調查 get deploy + 信心 0。 """ if not result or not k8s_inventory: return _inventory_names = {n.strip() for n in k8s_inventory.split(",") if n.strip()} if not _inventory_names: return _kcmd = (result.kubectl_command or "").lower() import re as _re _m = _re.search(r"deployment[/\s]+([a-z0-9][a-z0-9-]*)", _kcmd) if not _m: return _deploy_guess = _m.group(1) if _deploy_guess in _inventory_names: return logger.warning( "openclaw_deployment_hallucination_detected", hallucinated=_deploy_guess, inventory=sorted(_inventory_names), original_kubectl_cmd=result.kubectl_command, original_action=( result.suggested_action.value if hasattr(result.suggested_action, "value") else str(result.suggested_action) ), namespace=k8s_ns, ) # 降級為安全調查動作,不執行破壞性操作 try: result.kubectl_command = f"kubectl get deploy -n {k8s_ns}" except Exception: pass try: result.target_resource = "unknown(hallucinated)" except Exception: pass try: result.suggested_action = SuggestedAction.NO_ACTION except Exception: pass try: result.action_title = f"[安全降級] 調查 {k8s_ns} 真實資源狀態" except Exception: pass try: result.description = ( f"[安全降級] 原 LLM 建議的 deployment '{_deploy_guess}' 不在叢集 inventory " f"({', '.join(sorted(_inventory_names))})。" f"已降級為純調查動作(kubectl get deploy),請手動確認實際問題資源。" ) except Exception: pass try: result.confidence = 0.0 except Exception: pass # 2026-04-21 ogt: 降級後風險強制 LOW，避免 NO_ACTION 因原 HIGH/CRITICAL risk # 留在 PENDING 等 Telegram 批准（已無執行內容，等待毫無意義） try: result.risk_level = AIRiskLevel.LOW except Exception: pass def _parse_analysis_result(self, raw_response: str) -> OpenClawDecision | None: """ 解析 LLM 分析結果 - 使用 Pydantic Schema Enforcement 關鍵：blast_radius 為 REQUIRED，使用 AIBlastRadius Pydantic 模型驗證 """ json_str = self._extract_json_from_response(raw_response) if not json_str: logger.error("json_extraction_failed", raw_response=raw_response[:200]) return None try: data = json.loads(json_str) # Step 1: 確保 blast_radius 存在且為正確格式 if "blast_radius" not in data or not isinstance(data["blast_radius"], dict): data["blast_radius"] = { "affected_pods": 1, "estimated_downtime": "~30s", "related_services": data.get("affected_services", []), "data_impact": "NONE" } else: # 確保 blast_radius 內的必填欄位存在 br = data["blast_radius"] if "affected_pods" not in br: br["affected_pods"] = 1 if "estimated_downtime" not in br: br["estimated_downtime"] = "~30s" if "related_services" not in br: br["related_services"] = data.get("affected_services", []) if "data_impact" not in br: br["data_impact"] = "NONE" # Step 2: 填補其他可選欄位 if "action_title" not in data: data["action_title"] = data.get("action", "未知操作") if "target_resource" not in data: data["target_resource"] = "unknown" if "suggested_action" not in data: data["suggested_action"] = "NO_ACTION" # Step 2.5: 2026-04-01 Claude Code - 斷片補全 (信心度必須誠實) # 🔴 禁止填入假信心度！截斷 = 0.0，讓 auto-approve 正確判斷 # 2026-04-12 ogt: NIM/Ollama 有時回傳字串 "0.85"，先嘗試 parse 再判斷 if "confidence" in data and isinstance(data["confidence"], str): try: data["confidence"] = float(data["confidence"]) except (ValueError, TypeError): data.pop("confidence", None) if "confidence" not in data or not isinstance(data["confidence"], int | float): data["confidence"] = 0.0 # 截斷/缺失 → 0.0，不可偽造 if "risk_level" not in data: data["risk_level"] = "low" # 2026-04-19 ogt + Claude Opus 4.7 修 AP-3: # primary_responsibility 有時 LLM 填空字串/None → resp_display 顯示「❓ 未知」 # 強制正規化: 空/None/不在白名單 → 用 kubectl 有無推 INFRA 或 BE (非「未知」) _valid_resp = {"FE", "BE", "INFRA", "DB", "COLLAB"} _cur_resp = str(data.get("primary_responsibility") or "").strip().upper() if _cur_resp not in _valid_resp: data["primary_responsibility"] = "INFRA" if "kubectl" in str(data) else "BE" if "suggested_action" not in data: data["suggested_action"] = "RESTART_DEPLOYMENT" if "restart" in str(data).lower() else "NO_ACTION" if "reasoning" not in data: data["reasoning"] = "AI 產出欄位缺失，系統自動補全以維持運作。" # Step 3: 使用 Pydantic 驗證 (會自動正規化 risk_level, data_impact 等) decision = OpenClawDecision(**data) logger.info( "pydantic_validation_success", action_title=decision.action_title, risk_level=decision.risk_level.value, blast_radius_pods=decision.blast_radius.affected_pods, ) return decision except Exception as e: logger.error( "pydantic_validation_failed", error=str(e), json_str=json_str[:300], ) return None # ========================================================================= # Main Analysis Methods # ========================================================================= async def analyze_alert( self, alert_context: dict, ) -> tuple[LLMAnalysisResult | None, str, str, GoldMetrics | None, str, int, float]: """ 分析告警並產生 RCA 結果 (含 SignOz 整合) Args: alert_context: 告警上下文 (alert_type, severity, target_resource, etc.) Returns: (analysis_result, ai_provider, raw_response, signoz_metrics, signoz_trace_url, total_tokens, cost_usd) 2026-03-29 ogt: 加入 Token/Cost 追蹤 """ # Step 0: 擷取 SignOz 上下文 service_name = alert_context.get("target_resource", "unknown") namespace = alert_context.get("namespace", "default") signoz_metrics, signoz_trace_url = await self.get_signoz_context( service_name=service_name, namespace=namespace, ) # 將 SignOz 數據加入 prompt signoz_context = "" if signoz_metrics: signoz_context = f""" ## 📊 SignOz Real-time Metrics (Last 10 min) {signoz_metrics.to_summary()} Trace URL: {signoz_trace_url} """ # Step 0.5: 擷取 K8s 叢集真實資源清單（Checkpoint-2 webhook path） # 2026-04-17 ogt + Claude Sonnet 4.6: 防止 NemoTron 幻覺 deployment/awoooi-service # 根因：webhook path 沒有叢集上下文 → LLM 盲猜資源名稱 → kubectl not found → trust 0 永遠 # 修復：每次分析前先拉真實 Deployment 清單，注入 prompt 強制 LLM 對齊 _k8s_ns = alert_context.get("namespace", "awoooi-prod") _k8s_inventory = await _fetch_k8s_inventory_for_openclaw(namespace=_k8s_ns) k8s_section = ( f"\n\n## 🔒 叢集實際資源清單（{_k8s_ns}）\n" f"kubectl_command 與 target_resource **必須**從以下名稱選擇，不可自行編造：\n" f"{_k8s_inventory}\n" if _k8s_inventory else "\n\n## ⚠️ 無法取得叢集清單，target_resource 請依 alertname 推斷，勿編造。\n" ) # P2.1 fix 2026-04-24 ogt + Claude Sonnet 4.6: 提取 MCP evidence_summary 注入 prompt # diagnosis_context 由 decision_manager 在呼叫前填入（pre_decision_investigator 產出） _raw_evidence = alert_context.get("diagnosis_context", "") or "" if _raw_evidence and not _raw_evidence.startswith(""): _raw_evidence = f"\n{_raw_evidence}\n" evidence_section = f"\n\n## 🔬 MCP 實時環境證據\n{_raw_evidence}\n" if _raw_evidence else "" # 格式化告警為 Prompt (2026-03-31 ogt: 強力截斷以符合 NVIDIA 4K 限制) # 優先保留 System Prompt，截斷 Alert Data available_len = 3500 - len(OPENCLAW_SYSTEM_PROMPT) - len(signoz_context) - len(k8s_section) - len(evidence_section) if available_len < 500: # 如果 SignOz 太長，也截斷它 signoz_context = signoz_context[:500] + "... (truncated)" available_len = 3500 - len(OPENCLAW_SYSTEM_PROMPT) - len(signoz_context) - len(k8s_section) - len(evidence_section) alert_json = json.dumps(alert_context, ensure_ascii=False, indent=2) if len(alert_json) > available_len: alert_json = alert_json[:available_len] + "... (truncated)" full_prompt = OPENCLAW_SYSTEM_PROMPT + signoz_context + k8s_section + evidence_section + "\n\n## Alert Data:\n" + alert_json logger.info( "openclaw_alert_analysis_start", alert_type=alert_context.get("alert_type"), target=alert_context.get("target_resource"), signoz_available=signoz_metrics is not None, ) # 呼叫 LLM (使用快取層保護算力) raw_response, provider, success, from_cache, total_tokens, cost_usd = await self._call_with_cache( full_prompt, alert_context, signoz_metrics, cache_ttl=1800, # 30 min for alert analysis ) if not success: logger.error("openclaw_all_providers_failed") return None, provider, raw_response, signoz_metrics, signoz_trace_url, 0, 0.0 if from_cache: logger.info("openclaw_using_cached_response", provider=provider) logger.info( "openclaw_llm_response_received", provider=provider, response_length=len(raw_response), ) # 解析結果 result = self._parse_analysis_result(raw_response) # 2026-04-18 → 2026-04-19: 幻覺 deployment 名偵測與降級 (共用 helper) self._validate_deployment_inventory(result, _k8s_inventory, _k8s_ns) if result: logger.info( "openclaw_analysis_complete", action_title=result.action_title, risk_level=result.risk_level, confidence=result.confidence, provider=provider, signoz_integrated=signoz_metrics is not None, ) else: logger.warning( "openclaw_analysis_parse_failed", raw_response=raw_response[:300], ) return result, provider, raw_response, signoz_metrics, signoz_trace_url, total_tokens, cost_usd # ========================================================================= # Phase 6.4: LLM Proposal Generation # ========================================================================= async def generate_incident_proposal( self, incident_id: str, severity: str, signals: list[dict], affected_services: list[str], expert_context: dict | None = None, ) -> tuple[dict | None, str, bool]: """ 為 Incident 生成 LLM-based 修復提案 Phase 6.4: 賦予大腦「生成解決方案」的思考能力 2026-03-27: 整合 Expert System 診斷上下文 Args: incident_id: Incident ID severity: 嚴重度 (P0/P1/P2/P3) signals: 關聯的告警訊號 affected_services: 受影響服務 expert_context: Expert System 初步診斷 (可選) - initial_diagnosis: 規則匹配結果 - diagnosis_description: 診斷描述 - suggested_diagnosis_commands: 建議診斷指令 - expert_confidence: 信心分數 - requires_human_review: 是否需人工介入 Returns: (proposal_dict, provider, success) proposal_dict 包含: - action: 建議動作 - description: 動作描述 - kubectl_command: kubectl 指令 - risk_level: 風險等級 - reasoning: LLM 推理過程 """ # 建構 prompt (2026-03-31 ogt: Nemotron-mini context 較小，限制數量與長度) signal_summary = "\n".join([ f"- {s.get('alert_name', 'unknown')}: {str(s.get('description', 'N/A'))[:100]}..." for s in signals[:3] # 最多 3 筆，每筆最多 100 字元 ]) target = affected_services[0] if affected_services else "unknown-service" # 擷取 SignOz 指標 signoz_metrics, signoz_trace_url = await self.get_signoz_context( service_name=target, namespace="awoooi-prod", ) signoz_context = "" if signoz_metrics: signoz_context = f""" ## 📊 SignOz Real-time Metrics {signoz_metrics.to_summary()} """ # 2026-03-27: 整合 Expert System 診斷上下文 # 2026-03-26: ADR-030 Phase 2 - 加入 K8s/SignOz 診斷上下文 expert_diagnosis_context = "" if expert_context: diagnosis_cmds = expert_context.get("suggested_diagnosis_commands", []) diagnosis_cmds_str = "\n".join([f" - `{cmd}`" for cmd in diagnosis_cmds]) if diagnosis_cmds else " - (無)" # ADR-030: 加入完整診斷上下文 (如果有)，並限制長度以符合 4K Context full_diagnosis = str(expert_context.get("diagnosis_context", ""))[:800] if len(str(expert_context.get("diagnosis_context", ""))) > 800: full_diagnosis += "... (truncated)" diagnosis_signals = expert_context.get("diagnosis_signals", []) signals_summary = "" if diagnosis_signals: signals_summary = "\n".join([ f" - [{s.get('severity', 'info').upper()}] {s.get('source', 'unknown')}: {s.get('message', 'N/A')[:100]}" for s in diagnosis_signals[:5] ]) expert_diagnosis_context = f""" ## 🔍 Expert System Initial Diagnosis - **Matched Rule**: {expert_context.get('initial_diagnosis', 'unknown')} - **Diagnosis**: {expert_context.get('diagnosis_description', 'N/A')} - **Confidence**: {expert_context.get('expert_confidence', 0.0):.0%} - **Requires Human Review**: {'Yes' if expert_context.get('requires_human_review') else 'No'} - **Suggested Diagnosis Commands**: {diagnosis_cmds_str} {f'''## 🩺 K8s/SignOz Deep Diagnosis (ADR-030) {full_diagnosis} ### Diagnosis Signals {signals_summary if signals_summary else " - (No signals detected)"} ''' if full_diagnosis else ''} **IMPORTANT**: The Expert System and Diagnostic Aggregator have provided context. Consider this data but apply your own analysis. If Expert says "human review required", provide diagnostic guidance rather than automated fixes. """ # 2026-03-31 ogt: 針對 NVIDIA Nemo-4B 使用超精簡 Prompt registry = get_model_registry() is_nemo = "nvidia" in (registry.get_model("nvidia", "rca") or "").lower() base_prompt = NEMOTRON_SYSTEM_PROMPT if is_nemo else OPENCLAW_SYSTEM_PROMPT proposal_prompt = f"""{base_prompt} {signoz_context} {expert_diagnosis_context} ## 🚨 Incident Context - **Incident ID**: {incident_id} - **Severity**: {severity} - **Affected Services**: {', '.join(affected_services)} - **Signal Count**: {len(signals)} ## 📋 Alert Signals {signal_summary} ## 🎯 Your Task Based on the above incident, signals, and Expert System diagnosis, generate a remediation proposal. You MUST respond with ONLY valid JSON following the schema above. Focus on: 1. Root cause analysis based on signals, SignOz data, and Expert diagnosis 2. Specific kubectl command to remediate (or diagnostic command if root cause unclear) 3. Risk assessment for the proposed action 4. Preventive recommendations 5. If Expert System flagged "human review required", prioritize diagnostic commands over fixes """ logger.info( "proposal_generation_start", incident_id=incident_id, severity=severity, signal_count=len(signals), signoz_available=signoz_metrics is not None, ) # 2026-04-01 ogt: 架構鐵律 — OpenClaw (Nemo) 是 AI 大腦，優先委派仲裁 # AWOOOI K8s Pod 不直接打 Ollama/NVIDIA，避免並發逾時 openclaw_result = await self._call_openclaw_analyze( incident_id, severity, signals, affected_services, expert_context ) if openclaw_result is not None: return openclaw_result, "openclaw_nemo", True # 使用快取呼叫 LLM alert_context = { "incident_id": incident_id, "alert_type": signals[0].get("alert_name", "incident") if signals else "incident", "target_resource": target, "severity": severity, } # 2026-03-29 ogt: 修復 tuple unpacking (Token/Cost 追蹤) raw_response, provider, success, from_cache, ai_tokens, ai_cost = await self._call_with_cache( proposal_prompt, alert_context, signoz_metrics, cache_ttl=3600, # 1 hour for proposals ) if not success: logger.error( "proposal_generation_failed", incident_id=incident_id, provider=provider, ) return None, provider, False # 解析 LLM 結果 result = self._parse_analysis_result(raw_response) # 2026-04-19 ogt + Claude Opus 4.7: 同 analyze_alert 也需幻覺驗證 # 此路徑沒有 inventory 預抓,動態抓 _k8s_ns_for_validate = alert_context.get("namespace", "awoooi-prod") if "alert_context" in dir() else "awoooi-prod" try: _k8s_inv = await _fetch_k8s_inventory_for_openclaw(namespace=_k8s_ns_for_validate) except Exception: _k8s_inv = "" self._validate_deployment_inventory(result, _k8s_inv, _k8s_ns_for_validate) if result: logger.info( "proposal_generation_complete", incident_id=incident_id, action_title=result.action_title, risk_level=result.risk_level, provider=provider, from_cache=from_cache, ) # 轉換為 proposal dict (optimization_suggestions 是 list[dict]) # 2026-03-29 ogt: 加入 ai_tokens/ai_cost 追蹤 proposal_dict = { "action": result.action_title, "description": result.description, "kubectl_command": result.kubectl_command, "target_resource": result.target_resource, "namespace": result.namespace, "risk_level": result.risk_level, "reasoning": result.reasoning, "confidence": result.confidence, "primary_responsibility": result.primary_responsibility, "optimization_suggestions": [ { "type": s.get("type", "UNKNOWN"), "description": s.get("description", ""), "kubectl_or_config": s.get("kubectl_or_config", ""), } for s in result.optimization_suggestions ], "signoz_correlation": result.signoz_correlation, "from_cache": from_cache, "provider": provider, "model": self._get_model_name(provider), # 2026-04-04 ogt: 底層模型名稱 "ai_tokens": ai_tokens, "ai_cost": ai_cost, } return proposal_dict, provider, True logger.warning( "proposal_parse_failed", incident_id=incident_id, raw_response=raw_response[:300], ) return None, provider, False # ========================================================================= # Phase 22: OpenClaw + Nemotron 協作 (ADR-044) # 2026-03-31 Claude Code: 統帥批准實作 # ========================================================================= async def _maybe_run_openclaw_agent_loop_shadow( self, *, proposal: dict, incident_id: str, severity: str, signals: list[dict], affected_services: list[str], expert_context: dict | None = None, ) -> None: """ ADR-105 P1: read-only Agent Loop shadow investigation. This is intentionally non-decisive: it proves MCP tool_use/audit wiring with local models and read-only tools, then falls back silently to the regular proposal when disabled or unavailable. """ if not settings.ENABLE_OPENCLAW_AGENT_LOOP_SHADOW: return try: from src.plugins.mcp.registry import get_provider_registry from src.services.ai_providers.agent_loop import AgentToolExecutor from src.services.ai_providers.permissions import is_read_only_tool from src.services.ai_router import get_ai_registry ai_registry = get_ai_registry() provider = ai_registry.get("ollama") or ai_registry.get("ollama_local") if provider is None or not hasattr(provider, "analyze_with_tools"): logger.warning( "openclaw_agent_loop_shadow_skipped", incident_id=incident_id, reason="no_local_tool_provider", ) return mcp_registry = get_provider_registry() providers = {p.name: p for p in mcp_registry.all()} allowed_servers = {"kubernetes", "prometheus", "signoz", "database", "rag", "grafana"} available_tools = [] for mcp_provider in providers.values(): if mcp_provider.name not in allowed_servers: continue try: provider_tools = await mcp_provider.list_tools() except Exception as exc: logger.warning( "openclaw_agent_loop_tool_list_failed", incident_id=incident_id, provider=mcp_provider.name, error=str(exc), ) continue available_tools.extend( tool for tool in provider_tools if tool.server_name in allowed_servers and is_read_only_tool(tool) ) if not available_tools: logger.warning( "openclaw_agent_loop_shadow_skipped", incident_id=incident_id, reason="no_readonly_tools", ) return executor = AgentToolExecutor( available_tools=available_tools, providers=providers, agent_role="openclaw", incident_id=incident_id, flywheel_node="reason", ) shadow_prompt = self._build_agent_loop_shadow_prompt( proposal=proposal, incident_id=incident_id, severity=severity, signals=signals, affected_services=affected_services, expert_context=expert_context, ) result = await provider.analyze_with_tools( prompt=shadow_prompt, available_tools=available_tools, tool_executor=executor.execute, max_iterations=settings.OPENCLAW_AGENT_LOOP_MAX_ITERATIONS, agent_role="openclaw", context={ "incident_id": incident_id, "severity": severity, "task_type": "diagnose", }, ) structured_shadow = self._parse_agent_loop_shadow_response(result.raw_response or "") proposal["agent_loop_shadow"] = { "enabled": True, "success": result.success, "provider": result.provider, "tokens": result.tokens, "latency_ms": round(result.latency_ms, 1), "error": result.error, "decision_impact": "none", "structured": structured_shadow, "confidence_delta": structured_shadow.get("confidence_delta", 0.0), "preview": (result.raw_response or "")[:700], } logger.info( "openclaw_agent_loop_shadow_complete", incident_id=incident_id, provider=result.provider, success=result.success, tools_available=len(available_tools), latency_ms=round(result.latency_ms, 1), confidence_delta=structured_shadow.get("confidence_delta", 0.0), parse_status=structured_shadow.get("parse_status"), ) except Exception as exc: logger.warning( "openclaw_agent_loop_shadow_failed", incident_id=incident_id, error=str(exc), ) @classmethod def _parse_agent_loop_shadow_response(cls, raw_response: str) -> dict: """ Normalize read-only Agent Loop output into durable metadata. The shadow result is intentionally non-decisive. Downstream code can inspect this structure for quality review, but it must not override the main proposal until ADR-105 canary graduation. """ text = (raw_response or "").strip() if not text: return { "parse_status": "empty", "root_cause_check": "", "evidence_used": [], "confidence_delta": 0.0, "missing_evidence": [], "human_or_ai_next_step": "", } payload = cls._extract_json_object(text) if not isinstance(payload, dict): return { "parse_status": "unparsed", "root_cause_check": "", "evidence_used": [], "confidence_delta": 0.0, "missing_evidence": [], "human_or_ai_next_step": "", "raw_preview": text[:700], } return { "parse_status": "ok", "root_cause_check": cls._clip_shadow_text(payload.get("root_cause_check"), max_chars=500), "evidence_used": cls._coerce_shadow_list(payload.get("evidence_used"), max_items=5), "confidence_delta": cls._coerce_agent_loop_confidence_delta( payload.get("confidence_delta", 0.0) ), "missing_evidence": cls._coerce_shadow_list(payload.get("missing_evidence"), max_items=5), "human_or_ai_next_step": cls._clip_shadow_text( payload.get("human_or_ai_next_step"), max_chars=500 ), } @staticmethod def _extract_json_object(text: str) -> dict | None: """Extract the first JSON object from plain or fenced LLM output.""" candidates = [text] fenced = re.search(r"```(?:json)?\s*(\{.*?\})\s*```", text, flags=re.DOTALL | re.IGNORECASE) if fenced: candidates.insert(0, fenced.group(1)) object_match = re.search(r"\{.*\}", text, flags=re.DOTALL) if object_match: candidates.append(object_match.group(0)) for candidate in candidates: try: parsed = json.loads(candidate) except (TypeError, json.JSONDecodeError): continue if isinstance(parsed, dict): return parsed return None @staticmethod def _clip_shadow_text(value: object, *, max_chars: int) -> str: if value is None: return "" return str(value).strip()[:max_chars] @classmethod def _coerce_shadow_list(cls, value: object, *, max_items: int) -> list[str]: if value is None: return [] if isinstance(value, list): items = value else: items = [value] normalized = [] for item in items: clipped = cls._clip_shadow_text(item, max_chars=240) if clipped: normalized.append(clipped) if len(normalized) >= max_items: break return normalized @staticmethod def _coerce_agent_loop_confidence_delta(value: object) -> float: """ Keep canary deltas conservative: metadata may lower confidence later, but positive boosts are recorded as 0 until the shadow path graduates. """ try: delta = float(value) except (TypeError, ValueError): return 0.0 return round(max(min(delta, 0.0), -0.15), 3) def _build_agent_loop_shadow_prompt( self, *, proposal: dict, incident_id: str, severity: str, signals: list[dict], affected_services: list[str], expert_context: dict | None = None, ) -> str: """Build a compact read-only investigation prompt for Agent Loop shadow mode.""" return f"""你是 OpenClaw 的唯讀 shadow investigator。你可以使用 MCP 工具查證，但不得要求任何寫入、重啟、刪除或通知動作。 請只回傳 JSON： {{ "root_cause_check": "你對目前根因的查證結論", "evidence_used": ["最多 5 條具體證據"], "confidence_delta": -0.1, "missing_evidence": ["還缺什麼證據"], "human_or_ai_next_step": "下一個安全步驟" }} Incident: {incident_id} Severity: {severity} Affected services: {json.dumps(affected_services, ensure_ascii=False)} Signals: {json.dumps(signals[:5], ensure_ascii=False, default=str)} Current proposal: {json.dumps(proposal, ensure_ascii=False, default=str)} Expert context: {json.dumps(expert_context or {}, ensure_ascii=False, default=str)} """ async def generate_incident_proposal_with_tools( self, incident_id: str, severity: str, signals: list[dict], affected_services: list[str], expert_context: dict | None = None, ) -> tuple[dict | None, str, bool]: """ Phase 22: OpenClaw + Nemotron 協作生成修復提案 架構: - OpenClaw = 仲裁者 (Arbitrator) - 決定「為什麼」和「風險等級」 - Nemotron = 執行者 (Executor) - 決定「怎麼做」和「具體指令」 觸發條件: - LOW 風險 → 僅 OpenClaw，跳過 Nemotron - MEDIUM/HIGH/CRITICAL → OpenClaw + Nemotron 雙軌 Args: incident_id: Incident ID severity: 嚴重度 (P0/P1/P2/P3) signals: 關聯的告警訊號 affected_services: 受影響服務 expert_context: Expert System 初步診斷 (可選) Returns: (proposal_dict, provider, success) proposal_dict 新增: - nemotron_enabled: bool - nemotron_tools: list[dict] (如果啟用) - nemotron_validation: str - nemotron_latency_ms: float """ # Feature Flag 檢查 if not settings.ENABLE_NEMOTRON_COLLABORATION: logger.info( "nemotron_collaboration_disabled", incident_id=incident_id, reason="Feature flag disabled", ) return await self.generate_incident_proposal( incident_id, severity, signals, affected_services, expert_context ) # Step 1: OpenClaw 仲裁 proposal, provider, success = await self.generate_incident_proposal( incident_id, severity, signals, affected_services, expert_context ) if not success or proposal is None: return proposal, provider, success await self._maybe_run_openclaw_agent_loop_shadow( proposal=proposal, incident_id=incident_id, severity=severity, signals=signals, affected_services=affected_services, expert_context=expert_context, ) # Step 2: 判斷是否需要 Nemotron risk_level = proposal.get("risk_level", "low").lower() if risk_level == "low": proposal["nemotron_enabled"] = False logger.info( "nemotron_skipped_low_risk", incident_id=incident_id, risk_level=risk_level, ) return proposal, provider, True # Step 3: 呼叫 Nemotron Tool Calling — 🔴 必須等到有結果，不可跳過 # 2026-04-07 ogt: 統帥指示 Nemotron 不能跳過，必須等到處理完成 logger.info( "nemotron_collaboration_start", incident_id=incident_id, risk_level=risk_level, ) max_retries = 2 last_error = None for attempt in range(1, max_retries + 1): try: nemotron_result = await self._call_nemotron_tools( incident_id=incident_id, reasoning=proposal.get("reasoning", ""), target_resource=proposal.get("target_resource", ""), suggested_action=proposal.get("action", ""), namespace=proposal.get("namespace", "awoooi-prod"), ) proposal["nemotron_enabled"] = True proposal["nemotron_tools"] = nemotron_result.get("tools", []) proposal["nemotron_validation"] = nemotron_result.get("validation", "⏳ 驗證中") proposal["nemotron_latency_ms"] = nemotron_result.get("latency_ms", 0.0) proposal["nemotron_tool_model"] = nemotron_result.get("tool_model", "") proposal["nemotron_tool_backend"] = nemotron_result.get("tool_backend", "") # 2026-04-09 Claude Sonnet 4.6: 將 Nemotron tool call 回填為 kubectl_command # 根本問題修復：approval_records.action 需要可執行指令才能被 parse_operation_from_action 解析 _backfill_kubectl_command(proposal, proposal["nemotron_tools"]) logger.info( "nemotron_collaboration_complete", incident_id=incident_id, tools_count=len(proposal["nemotron_tools"]), validation=proposal["nemotron_validation"], latency_ms=proposal["nemotron_latency_ms"], attempt=attempt, ) last_error = None break # 成功，跳出重試迴圈 except Exception as e: last_error = e logger.warning( "nemotron_collaboration_retry", incident_id=incident_id, error=str(e), attempt=attempt, max_retries=max_retries, ) if attempt < max_retries: import asyncio await asyncio.sleep(2) # 重試前等 2 秒 # 重試全部失敗 — fallback 到 Gemini 模擬 tool calling # 2026-04-08 ogt: NIM 完全不可用時，改用 Gemini 產生執行方案（不可跳過） if last_error is not None: logger.error( "nemotron_collaboration_exhausted", incident_id=incident_id, error=str(last_error), retries=max_retries, ) logger.info("nemotron_fallback_gemini_start", incident_id=incident_id) gemini_fallback_result = await self._call_nemotron_tools_via_gemini( incident_id=incident_id, reasoning=proposal.get("reasoning", ""), target_resource=proposal.get("target_resource", ""), suggested_action=proposal.get("action", ""), namespace=proposal.get("namespace", "awoooi-prod"), ) proposal["nemotron_enabled"] = True proposal["nemotron_tools"] = gemini_fallback_result.get("tools", []) proposal["nemotron_validation"] = gemini_fallback_result.get("validation", "⚠️ Gemini 代理") proposal["nemotron_latency_ms"] = gemini_fallback_result.get("latency_ms", 0.0) proposal["nemotron_tool_model"] = "gemini-fallback" proposal["nemotron_tool_backend"] = "Gemini 雲端" # 2026-04-09 Claude Sonnet 4.6: Gemini fallback 同樣回填 kubectl_command _backfill_kubectl_command(proposal, proposal["nemotron_tools"]) return proposal, provider, True async def _call_nemotron_tools( self, incident_id: str, reasoning: str, target_resource: str, suggested_action: str, namespace: str = "awoooi-prod", ) -> dict: """ 呼叫 Nemotron 執行 Tool Calling Args: incident_id: Incident ID reasoning: OpenClaw 推理結果 target_resource: 目標資源名稱 suggested_action: OpenClaw 建議的操作 namespace: K8s namespace Returns: { "tools": [{"tool": str, "args": dict, "valid": bool}], "validation": str, "latency_ms": float } """ import asyncio from src.services.nvidia_provider import get_nvidia_provider nvidia = get_nvidia_provider() start_time = time.time() # 建構 Tool Calling prompt # 2026-04-09 Claude Code: 明確指示 deployment_name，防止 Nemotron 填 placeholder tool_prompt = f"""根據以下 AI 分析結果，生成對應的 kubectl 操作指令： ## Incident 上下文 - Incident ID: {incident_id} - 目標資源 (deployment_name): {target_resource} - Namespace: {namespace} ## OpenClaw 分析 - 建議操作: {suggested_action} - 推理過程: {reasoning[:500]} ## 你的任務 使用提供的工具生成 kubectl 操作。 **重要**: deployment_name 必須填入 "{target_resource}"，不可使用 placeholder。 """ # 定義可用 Tools (K8s 操作) k8s_tools = [ { "type": "function", "function": { "name": "restart_deployment", "description": "重啟 Deployment (rollout restart)", "parameters": { "type": "object", "properties": { "deployment_name": {"type": "string"}, "namespace": {"type": "string", "default": "awoooi-prod"}, }, "required": ["deployment_name"], }, }, }, { "type": "function", "function": { "name": "scale_deployment", "description": "調整 Deployment 副本數", "parameters": { "type": "object", "properties": { "deployment_name": {"type": "string"}, "replicas": {"type": "integer"}, "namespace": {"type": "string", "default": "awoooi-prod"}, }, "required": ["deployment_name", "replicas"], }, }, }, { "type": "function", "function": { "name": "delete_pod", "description": "刪除 Pod (強制重建)", "parameters": { "type": "object", "properties": { "pod_name": {"type": "string"}, "namespace": {"type": "string", "default": "awoooi-prod"}, }, "required": ["pod_name"], }, }, }, ] try: # 2026-04-07 ogt: 統帥指示不可跳過 Nemotron，用 120 秒寬裕超時 timeout = 120 result = await asyncio.wait_for( nvidia.tool_call( messages=[{"role": "user", "content": tool_prompt}], tools=k8s_tools, ), timeout=timeout, ) latency_ms = (time.time() - start_time) * 1000 # 解析 Tool Calling 結果 tools = [] validation_passed = True if result and hasattr(result, "tool_calls") and result.tool_calls: for tc in result.tool_calls: tool_entry = { "tool": tc.tool_name if hasattr(tc, "tool_name") else str(tc.get("name", "unknown")), "args": tc.arguments if hasattr(tc, "arguments") else tc.get("arguments", {}), "valid": tc.valid if hasattr(tc, "valid") else True, } tools.append(tool_entry) if not tool_entry["valid"]: validation_passed = False elif result and isinstance(result, dict) and result.get("tool_calls"): for tc in result["tool_calls"]: tool_entry = { "tool": tc.get("name", "unknown"), "args": tc.get("arguments", {}), "valid": True, } tools.append(tool_entry) # 2026-04-09 Claude Code: 修正 Nemotron 回傳 placeholder 問題 # 若 deployment_name 是 placeholder 或空值，用 target_resource 覆蓋 _PLACEHOLDERS = {"", "", "", "unknown", "null", "None"} for t in tools: args = t.get("args", {}) if isinstance(args, dict): dn = args.get("deployment_name", "") if not dn or str(dn).strip("<>") in _PLACEHOLDERS or dn.startswith("<"): args["deployment_name"] = target_resource logger.info( "nemotron_placeholder_corrected", incident_id=incident_id, original=dn, corrected=target_resource, ) validation_status = "✅ 驗證通過" if validation_passed and tools else "❌ 驗證失敗" return { "tools": tools, "validation": validation_status, "latency_ms": latency_ms, } except TimeoutError: latency_ms = (time.time() - start_time) * 1000 logger.error( "nemotron_tool_call_timeout", incident_id=incident_id, timeout_seconds=timeout, ) # 超時也拋出，讓外層重試 raise except Exception as e: latency_ms = (time.time() - start_time) * 1000 logger.error( "nemotron_tool_call_error", incident_id=incident_id, error=str(e), ) raise async def _call_nemotron_tools_via_gemini( self, incident_id: str, reasoning: str, target_resource: str, suggested_action: str, namespace: str = "awoooi-prod", ) -> dict: """ NIM 完全不可用時，由 Gemini 代理產生 tool calling 執行方案。 2026-04-08 ogt: NIM timeout 後的唯一 fallback，不可跳過。 Returns: {"tools": [...], "validation": str, "latency_ms": float} """ import time as _time start_time = _time.time() prompt = f"""你是 K8s SRE 專家。根據以下分析，輸出對應的 kubectl 操作指令（JSON 格式）。 Incident ID: {incident_id} 目標資源: {target_resource} Namespace: {namespace} 建議操作: {suggested_action} 分析摘要: {reasoning[:300]} 請輸出以下 JSON 格式（只輸出 JSON，不要其他文字）： {{ "tool_name": "restart_deployment 或 scale_deployment 或 no_action", "deployment_name": "部署名稱", "namespace": "{namespace}", "reason": "一句話說明原因" }}""" try: text, success, _, _ = await self._call_gemini(prompt) latency_ms = (_time.time() - start_time) * 1000 if not success: logger.warning("nemotron_gemini_fallback_failed", incident_id=incident_id, error=text) return {"tools": [], "validation": "❌ NIM + Gemini 均不可用", "latency_ms": latency_ms} import json as _json data = _json.loads(text) tool_name = data.get("tool_name", "no_action") tools = [] if tool_name != "no_action": tools = [{ "tool": tool_name, "args": { "deployment_name": data.get("deployment_name", target_resource), "namespace": data.get("namespace", namespace), }, "valid": True, }] logger.info( "nemotron_gemini_fallback_success", incident_id=incident_id, tool=tool_name, latency_ms=latency_ms, ) return { "tools": tools, "validation": "✅ Gemini 代理驗證通過", "latency_ms": latency_ms, } except Exception as e: latency_ms = (_time.time() - start_time) * 1000 logger.error("nemotron_gemini_fallback_error", incident_id=incident_id, error=str(e)) return {"tools": [], "validation": f"❌ Gemini 代理失敗: {str(e)[:50]}", "latency_ms": latency_ms} # ========================================================================= # Shadow Mode Auto-Tuning # ========================================================================= async def execute_auto_tuning( self, approval_id: str, kubectl_command: str, description: str, ) -> dict: """ 執行自動調優 (Shadow Mode: 僅日誌輸出) 統帥鐵律: Shadow Mode 下嚴禁實際執行 K8s 命令 Args: approval_id: 簽核單 ID kubectl_command: kubectl 指令 description: 操作描述 Returns: {executed: bool, shadow_mode: bool, command: str, log: str} """ if settings.SHADOW_MODE_ENABLED: # Shadow Mode: 僅記錄，不執行 log_message = f"[SHADOW MODE] AI 生成的調優指令：{kubectl_command}" logger.info( "shadow_mode_auto_tuning", approval_id=approval_id, command=kubectl_command, description=description, executed=False, ) print(f"\n{'='*60}") print(log_message) print(f"描述: {description}") print(f"簽核單: {approval_id}") print(f"{'='*60}\n") return { "executed": False, "shadow_mode": True, "command": kubectl_command, "description": description, "log": log_message, } else: # 生產模式: 實際執行 (需要額外安全檢查) logger.warning( "auto_tuning_execution_attempted", approval_id=approval_id, command=kubectl_command, message="Production execution not yet implemented - requires multi-sig approval", ) return { "executed": False, "shadow_mode": False, "command": kubectl_command, "description": description, "log": "Production execution requires multi-sig approval", } # ============================================================================= # Singleton # ============================================================================= _openclaw: OpenClawService | None = None def get_openclaw() -> OpenClawService: """取得全域 OpenClaw 實例""" global _openclaw if _openclaw is None: _openclaw = OpenClawService() return _openclaw async def close_openclaw() -> None: """關閉 OpenClaw 連線""" global _openclaw if _openclaw: await _openclaw.close() _openclaw = None async def _fetch_k8s_inventory_for_openclaw( namespace: str = "awoooi-prod", timeout_sec: float = 3.0, ) -> str: """ 取得 K8s 叢集實際 Deployment/StatefulSet 清單，注入 analyze_alert prompt。 2026-04-17 ogt + Claude Sonnet 4.6 (Checkpoint-2 webhook path): - 根因：NemoTron 在 webhook path 收不到叢集清單 → 幻覺 deployment/awoooi-service - 修復：analyze_alert 前拉取真實資源名，注入 prompt，強制 LLM 從清單選擇 - 超時/失敗 → 返回 ""（prompt 仍正常但無鎖定效果，不中斷主流程） - 只執行唯讀 get 指令，不修改叢集 Returns: "awoooi-api, awoooi-web, ..." 格式字串，失敗時返回 "" """ import asyncio as _asyncio import structlog as _structlog _logger = _structlog.get_logger(__name__) try: cmd = ( f"kubectl get deployments,statefulsets -n {namespace} " "-o jsonpath='{.items[*].metadata.name}' 2>/dev/null" ) proc = await _asyncio.create_subprocess_shell( cmd, stdout=_asyncio.subprocess.PIPE, stderr=_asyncio.subprocess.PIPE, ) try: stdout, _ = await _asyncio.wait_for(proc.communicate(), timeout=timeout_sec) except TimeoutError: proc.kill() _logger.warning("k8s_inventory_timeout_openclaw", namespace=namespace) return "" raw = (stdout or b"").decode("utf-8", errors="replace").strip() if not raw: return "" names = [n.strip() for n in raw.split() if n.strip()] return ", ".join(names) except Exception as _e: _logger.warning("k8s_inventory_failed_openclaw", namespace=namespace, error=str(_e)) return "" # ============================================================================= # Phase 5 + SignOz Integration Complete # =============================================================================