fix(ai): 移除假信心分數 + 顯示 AI 模型來源

問題: AI 仲裁顯示硬編碼信心分數 (0.75/0.88/0.92/0.70) 修復: - decision_manager: 預設 confidence 0.75 → 0.0 - decision_manager: Expert System confidence=0.0 + is_rule_based - openclaw: 所有 Mock Response confidence → 0.0 - telegram_gateway: 新增 ai_provider 欄位 - telegram_gateway: 動態來源標籤 (Ollama/Gemini/Claude/規則匹配) Telegram 卡片顯示: - confidence > 0 + provider=ollama → 🤖 Ollama 仲裁 - confidence > 0 + provider=gemini → 🤖 Gemini 仲裁 - confidence > 0 + provider=claude → 🤖 Claude 仲裁 - confidence == 0 → ⚙️ 規則匹配 Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
2026-03-29 15:19:51 +08:00
parent 12e49d844a
commit 5f9a6a7e55
4 changed files with 119 additions and 23 deletions
--- a/apps/api/src/services/decision_manager.py
+++ b/apps/api/src/services/decision_manager.py
@@ -102,8 +102,9 @@ async def _push_decision_to_telegram(
        action = proposal_data.get("action", proposal_data.get("kubectl_command", ""))
        description = proposal_data.get("description", "")
        reasoning = proposal_data.get("reasoning", "")
-        confidence = proposal_data.get("confidence", 0.75)
+        confidence = proposal_data.get("confidence", 0.0)  # 🔴 預設 0.0 表示未經 AI 分析
        source = proposal_data.get("source", "unknown")
+        ai_provider = proposal_data.get("provider", "")  # 2026-03-29 ogt: AI 模型來源

        # 建立 approval_id (使用 incident_id 作為追蹤)
        # 2026-03-27 ogt: 修復 INC-INC-INC- 重複前綴 bug
@@ -119,6 +120,7 @@ async def _push_decision_to_telegram(
            primary_responsibility="INFRA",
            confidence=confidence,
            namespace=incident.signals[0].labels.get("namespace", "default") if incident.signals else "default",
+            ai_provider=ai_provider,  # 2026-03-29 ogt: 顯示 AI 模型來源
        )

        # 🔴 發送成功後設置去重 key (TTL 10 分鐘)
@@ -222,16 +224,19 @@ def expert_analyze(incident: Incident) -> dict[str, Any]:

    rule = EXPERT_RULES[matched_rule]

+    # 2026-03-29 ogt: Expert System 不應該假裝有高信心分數
+    # 設為 0.0 強制標記為規則匹配，而非 AI 仲裁
    return {
        "source": "expert_system",
        "action": rule["action"].format(target=target),
        "description": rule["description"],
        "risk_level": rule["risk_level"],
-        "reasoning": rule["reasoning"],
-        "confidence": 0.75,  # Expert System 固定信心分數
+        "reasoning": f"[規則匹配] {rule['reasoning']}",  # 明確標示來源
+        "confidence": 0.0,  # 🔴 規則匹配不是 AI 仲裁，信心度設 0
        "kubectl_command": rule["action"].format(target=target),
        "matched_rule": matched_rule,
        "from_cache": False,
+        "is_rule_based": True,  # 新增標記
    }


--- a/apps/api/src/services/openclaw.py
+++ b/apps/api/src/services/openclaw.py
@@ -520,10 +520,11 @@ class OpenClawService:
        )

        # 根據告警類型生成專業 RCA + 仲裁
+        # 🔴 2026-03-29 ogt: Mock 響應必須標示為規則匹配，不是 AI 仲裁
        if "oom" in message.lower() or "memory" in alert_type.lower():
            mock_response = {
                "action_title": f"刪除異常 Pod {target} (OOMKilled)",
-                "description": f"🤖 AI 仲裁: {target} 發生 OOMKilled，根因為 JVM Heap 配置與 K8s memory limit 不匹配或存在記憶體洩漏。{signoz_summary}",
+                "description": f"⚙️ 規則匹配: {target} 發生 OOMKilled，根因為 JVM Heap 配置與 K8s memory limit 不匹配或存在記憶體洩漏。{signoz_summary}",
                "suggested_action": "DELETE_POD",
                "kubectl_command": f"kubectl delete pod {target} -n {namespace}",
                "target_resource": target,
@@ -550,9 +551,9 @@ class OpenClawService:
                        "kubectl_or_config": f"kubectl autoscale deployment {target.rsplit('-', 2)[0]} --memory-percent=80 --min=2 --max=5 -n {namespace}"
                    }
                ],
-                "reasoning": f"🤖 Pod OOMKilled 後 ReplicaSet 將自動重建，但需同步修正資源配置防止復發。{signoz_correlation}",
+                "reasoning": f"⚙️ Pod OOMKilled 後 ReplicaSet 將自動重建，但需同步修正資源配置防止復發。{signoz_correlation}",
                "deviation_analysis": f"Memory 使用率 {metrics.get('memory_percent', 99)}%，超出基準線 60% 達 +6.5σ",
-                "confidence": 0.88,
+                "confidence": 0.0,  # 🔴 規則匹配，非 AI 仲裁
                "affected_services": [target, "api-gateway"],
                "signoz_correlation": signoz_correlation,
            }
@@ -564,7 +565,7 @@ class OpenClawService:

            mock_response = {
                "action_title": f"擴展 {target} 副本數 + 啟用 HPA",
-                "description": f"🤖 AI 仲裁: {target} CPU 使用率過高，根因為流量突增或計算密集任務未配置自動擴展。{rps_context}",
+                "description": f"⚙️ 規則匹配: {target} CPU 使用率過高，根因為流量突增或計算密集任務未配置自動擴展。{rps_context}",
                "suggested_action": "SCALE_DEPLOYMENT",
                "kubectl_command": tuning["command"],
                "target_resource": target,
@@ -591,16 +592,16 @@ class OpenClawService:
                        "kubectl_or_config": f"kubectl set resources deployment/{target} --requests=cpu=500m --limits=cpu=2000m -n {namespace}"
                    }
                ],
-                "reasoning": f"🤖 水平擴展可即時分散負載，同時建議配置 HPA 防止復發。{signoz_correlation}",
+                "reasoning": f"[規則匹配] 水平擴展可即時分散負載，同時建議配置 HPA 防止復發。{signoz_correlation}",
                "deviation_analysis": f"CPU 使用率 {metrics.get('cpu_percent', 95)}%，超出基準線 50% 達 +4.5σ",
-                "confidence": 0.92,
+                "confidence": 0.0,  # 🔴 規則匹配不是 AI 仲裁，信心度設 0
                "affected_services": [target],
                "signoz_correlation": signoz_correlation,
            }
        elif "http" in alert_type.lower() or "5xx" in message.lower() or "502" in message.lower():
            mock_response = {
                "action_title": f"重啟 {target} + 檢查上游服務",
-                "description": f"🤖 AI 仲裁: {target} 產生 HTTP 5xx 錯誤，可能為應用程式例外或上游服務不可達。{signoz_summary}",
+                "description": f"⚙️ 規則匹配: {target} 產生 HTTP 5xx 錯誤，可能為應用程式例外或上游服務不可達。{signoz_summary}",
                "suggested_action": "RESTART_DEPLOYMENT",
                "kubectl_command": f"kubectl rollout restart deployment/{target} -n {namespace}",
                "target_resource": target,
@@ -627,9 +628,9 @@ class OpenClawService:
                        "kubectl_or_config": "# 檢查 Redis 連線池配置，建議 maxTotal=50"
                    }
                ],
-                "reasoning": f"🤖 HTTP 錯誤需協同排查，先重啟恢復服務同時通知相關團隊。{signoz_correlation}",
+                "reasoning": f"[規則匹配] HTTP 錯誤需協同排查，先重啟恢復服務同時通知相關團隊。{signoz_correlation}",
                "deviation_analysis": "錯誤率 5%，超出基準線 0.1% 達 +50σ",
-                "confidence": 0.65,
+                "confidence": 0.0,  # 🔴 規則匹配不是 AI 仲裁，信心度設 0
                "affected_services": [target, "nginx-ingress", "upstream-api"],
                "signoz_correlation": signoz_correlation,
            }
@@ -637,7 +638,7 @@ class OpenClawService:
            # 通用異常處理
            mock_response = {
                "action_title": f"重新啟動 {target} 服務",
-                "description": f"🤖 AI 仲裁: {target} 發生異常: {message[:80]}。需進一步診斷確認根因。{signoz_summary}",
+                "description": f"⚙️ 規則匹配: {target} 發生異常: {message[:80]}。需進一步診斷確認根因。{signoz_summary}",
                "suggested_action": "RESTART_DEPLOYMENT",
                "kubectl_command": f"kubectl rollout restart deployment/{target} -n {namespace}",
                "target_resource": target,
@@ -659,9 +660,9 @@ class OpenClawService:
                        "kubectl_or_config": tuning["command"],
                    }
                ],
-                "reasoning": f"🤖 根據告警 {alert_type} 先重啟恢復服務，同時安排深入診斷。{signoz_correlation}",
+                "reasoning": f"[規則匹配] 根據告警 {alert_type} 先重啟恢復服務，同時安排深入診斷。{signoz_correlation}",
                "deviation_analysis": "監控指標顯示異常偏離基準線",
-                "confidence": 0.70,
+                "confidence": 0.0,  # 🔴 規則匹配不是 AI 仲裁，信心度設 0
                "affected_services": [target],
                "signoz_correlation": signoz_correlation,
            }
--- a/apps/api/src/services/telegram_gateway.py
+++ b/apps/api/src/services/telegram_gateway.py
@@ -142,6 +142,10 @@ class TelegramMessage:
    # 2026-03-29 ogt: AI Token/Cost 追蹤
    ai_tokens: int = 0          # LLM Token 使用量
    ai_cost: float = 0.0        # LLM 成本 (USD)
+    # 2026-03-29 ogt: ADR-037 異常頻率統計
+    anomaly_frequency: dict | None = None  # AnomalyCounter 統計
+    # 2026-03-29 ogt: AI Provider 來源顯示
+    ai_provider: str = ""  # ollama/gemini/claude/expert_system/mock

    def format(self) -> str:
        """
@@ -198,6 +202,43 @@ class TelegramMessage:
        if self.ai_tokens > 0 or self.ai_cost > 0:
            ai_cost_display = f"💰 Tokens: {self.ai_tokens:,} / ${self.ai_cost:.4f}\n"

+        # 2026-03-29 ogt: ADR-037 異常頻率顯示
+        frequency_block = ""
+        if self.anomaly_frequency and self.anomaly_frequency.get("count_24h", 0) > 1:
+            freq = self.anomaly_frequency
+            escalation_emoji = {
+                None: "",
+                "REPEAT": "⚠️",
+                "ESCALATE": "🔴",
+                "PERMANENT_FIX": "🚨",
+            }.get(freq.get("escalation_level"), "")
+
+            frequency_block = (
+                f"━━━━━━━━━━━━━━━━━━━\n"
+                f"📊 <b>頻率統計</b> {escalation_emoji}\n"
+                f"├ 1h: <code>{freq.get('count_1h', 0)}</code> 次\n"
+                f"├ 24h: <code>{freq.get('count_24h', 0)}</code> 次\n"
+                f"└ 修復: <code>{freq.get('auto_repair_count', 0)}</code> 次\n"
+            )
+            if freq.get("escalation_level"):
+                frequency_block += f"🔺 升級: <b>{freq['escalation_level']}</b>\n"
+
+        # 2026-03-29 ogt: 根據 confidence + ai_provider 動態顯示來源
+        # confidence > 0 = 真實 AI 分析, confidence == 0 = 規則匹配/降級
+        if self.confidence > 0 and self.ai_provider:
+            # 顯示具體 AI 模型
+            provider_names = {
+                "ollama": "Ollama",
+                "gemini": "Gemini",
+                "claude": "Claude",
+            }
+            provider_display = provider_names.get(self.ai_provider.lower(), self.ai_provider.upper())
+            source_label = f"🤖 <b>{provider_display} 仲裁</b>"
+        elif self.confidence > 0:
+            source_label = "🤖 <b>AI 仲裁判定</b>"
+        else:
+            source_label = "⚙️ <b>規則匹配</b>"
+
        # 組裝訊息
        message = (
            f"═══════════════════════════\n"
@@ -206,11 +247,12 @@ class TelegramMessage:
            f"📋 <code>{html.escape(incident_id)}</code>\n"
            f"🎯 資源: <code>{safe_resource}</code>\n"
            f"━━━━━━━━━━━━━━━━━━━\n"
-            f"🤖 <b>AI 仲裁判定</b>\n"
+            f"{source_label}\n"
            f"👥 責任: {resp_display}\n"
            f"📊 信心: {conf_emoji} {confidence_pct}%\n"
            f"{ai_cost_display}"
            f"💡 原因: {safe_root_cause}\n"
+            f"{frequency_block}"
            f"{signoz_block}"
            f"━━━━━━━━━━━━━━━━━━━\n"
            f"🔧 建議: {safe_action}\n"
@@ -437,11 +479,16 @@ class TelegramGateway:
        # 2026-03-29 ogt: AI Token/Cost 追蹤
        ai_tokens: int = 0,
        ai_cost: float = 0.0,
+        # 2026-03-29 ogt: ADR-037 異常頻率統計
+        anomaly_frequency: dict | None = None,
+        # 2026-03-29 ogt: AI Provider 來源顯示
+        ai_provider: str = "",
    ) -> dict:
        """
        推送待簽核卡片到 Telegram (v7.0 含 SignOz 整合)

        SOUL.md 4.1 + AI 仲裁 + SignOz 訊息格式
+        Phase 21 (ADR-037): 含異常頻率統計

        Args:
            approval_id: 簽核單 ID
@@ -456,6 +503,7 @@ class TelegramGateway:
            signoz_*: SignOz Gold Metrics
            signoz_trace_url: 動態時間參數的 Trace URL
            auto_tuning_command: kubectl 調優指令
+            anomaly_frequency: 異常頻率統計 (ADR-037)

        Returns:
            dict: Telegram API 回應
@@ -475,7 +523,7 @@ class TelegramGateway:
                trace_url=signoz_trace_url,
            )

-        # 建立訊息結構 (含 AI 仲裁 + SignOz + Token/Cost)
+        # 建立訊息結構 (含 AI 仲裁 + SignOz + Token/Cost + 頻率統計)
        message = TelegramMessage(
            status_emoji=emoji,
            risk_level=risk_level.upper(),
@@ -493,6 +541,10 @@ class TelegramGateway:
            # 2026-03-29 ogt: AI Token/Cost 追蹤
            ai_tokens=ai_tokens,
            ai_cost=ai_cost,
+            # 2026-03-29 ogt: ADR-037 異常頻率統計
+            anomaly_frequency=anomaly_frequency,
+            # 2026-03-29 ogt: AI Provider 來源顯示
+            ai_provider=ai_provider,
        )

        # 格式化訊息