diff --git a/apps/api/src/services/openclaw.py b/apps/api/src/services/openclaw.py
index 7be8971f1..96687ece8 100644
--- a/apps/api/src/services/openclaw.py
+++ b/apps/api/src/services/openclaw.py
@@ -18,6 +18,7 @@ Features:
 - SignOz 失敗時優雅降級 (不阻塞主流程)
 """
 
+import hashlib
 import json
 import re
 import time
@@ -27,6 +28,7 @@ import httpx
 import structlog
 
 from src.core.config import settings
+from src.core.redis_client import get_redis
 from src.models.ai import (
     AIRiskLevel,
     AIBlastRadius,
@@ -714,6 +716,100 @@ class OpenClawService:
 
         return json.dumps(mock_response)
 
+    # =========================================================================
+    # 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
+
+        使用 prompt 內容的 SHA256 作為快取鍵，確保相同問題不重複呼叫 LLM
+        """
+        content = f"{prompt}:{context_hash}"
+        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]:
+        """
+        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)
+        """
+        # 生成快取鍵 (基於 prompt + alert_context hash)
+        context_hash = ""
+        if alert_context:
+            # 使用告警類型 + 目標資源作為上下文 hash
+            context_hash = f"{alert_context.get('alert_type', '')}:{alert_context.get('target_resource', '')}"
+
+        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
+                )
+        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 = 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": datetime.now().isoformat(),
+                }
+                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  # from_cache=False
+
     # =========================================================================
     # Fallback Chain
     # =========================================================================
@@ -899,17 +995,21 @@ Trace URL: {signoz_trace_url}
             signoz_available=signoz_metrics is not None,
         )
 
-        # 呼叫 LLM
-        raw_response, provider, success = await self._call_with_fallback(
+        # 呼叫 LLM (使用快取層保護算力)
+        raw_response, provider, success, from_cache = 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
 
+        if from_cache:
+            logger.info("openclaw_using_cached_response", provider=provider)
+
         logger.info(
             "openclaw_llm_response_received",
             provider=provider,
@@ -936,6 +1036,157 @@ Trace URL: {signoz_trace_url}
 
         return result, provider, raw_response, signoz_metrics, signoz_trace_url
 
+    # =========================================================================
+    # Phase 6.4: LLM Proposal Generation
+    # =========================================================================
+
+    async def generate_incident_proposal(
+        self,
+        incident_id: str,
+        severity: str,
+        signals: list[dict],
+        affected_services: list[str],
+    ) -> 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
+
+        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
+
+        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
+        signal_summary = "\n".join([
+            f"- {s.get('alert_name', 'unknown')}: {s.get('description', 'N/A')}"
+            for s in signals[:10]  # 最多 10 筆
+        ])
+
+        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()}
+"""
+
+        proposal_prompt = f"""{OPENCLAW_SYSTEM_PROMPT}
+
+{signoz_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 and signals, generate a remediation proposal.
+You MUST respond with ONLY valid JSON following the schema above.
+Focus on:
+1. Root cause analysis based on signals and SignOz data
+2. Specific kubectl command to remediate
+3. Risk assessment for the proposed action
+4. Preventive recommendations
+"""
+
+        logger.info(
+            "proposal_generation_start",
+            incident_id=incident_id,
+            severity=severity,
+            signal_count=len(signals),
+            signoz_available=signoz_metrics is not None,
+        )
+
+        # 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,
+        }
+
+        raw_response, provider, success, from_cache = 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)
+
+        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
+            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.type,
+                        "description": s.description,
+                        "kubectl_or_config": s.kubectl_or_config,
+                    }
+                    for s in result.optimization_suggestions
+                ],
+                "signoz_correlation": result.signoz_correlation,
+                "from_cache": from_cache,
+                "provider": provider,
+            }
+            return proposal_dict, provider, True
+
+        logger.warning(
+            "proposal_parse_failed",
+            incident_id=incident_id,
+            raw_response=raw_response[:300],
+        )
+        return None, provider, False
+
     # =========================================================================
     # Shadow Mode Auto-Tuning
     # =========================================================================
diff --git a/apps/api/src/services/proposal_service.py b/apps/api/src/services/proposal_service.py
index 8af5313a6..ebdafb3c2 100644
--- a/apps/api/src/services/proposal_service.py
+++ b/apps/api/src/services/proposal_service.py
@@ -19,7 +19,6 @@ Decision Proposal Service - Phase 6.4 決策輸出層
 """
 
 from datetime import datetime, timezone
-from typing import Any
 from uuid import UUID
 
 import structlog
@@ -43,6 +42,7 @@ from src.models.incident import (
 )
 from src.services.approval_db import get_approval_service
 from src.services.trust_engine import trust_engine, normalize_action_pattern, RiskLevel
+from src.services.openclaw import get_openclaw
 
 logger = structlog.get_logger(__name__)
 
@@ -95,14 +95,20 @@ class ProposalService:
     決策提案服務 - Phase 6.4
 
     職責:
-    1. 分析 Incident 生成修復建議
+    1. 分析 Incident 生成修復建議 (LLM-based)
     2. 評估風險等級
     3. 建立 ApprovalRequest (向下相容前端)
     4. 更新 Incident 狀態與關聯
+
+    Phase 6.4 升級:
+    - 整合 OpenClaw LLM 生成智能提案
+    - 使用 _call_with_cache 保護算力資源
+    - Fallback 到模板方案確保可用性
     """
 
     def __init__(self) -> None:
         self._approval_service = get_approval_service()
+        self._openclaw = get_openclaw()
 
     # =========================================================================
     # 核心方法: 從 Incident 生成 Proposal
@@ -147,12 +153,51 @@ class ProposalService:
                 signal_count=len(incident.signals),
             )
 
-            # 2. 分析 signals 決定修復動作
-            action_type, action, description = self._determine_action(incident)
-
-            # 3. 評估風險等級
-            base_risk = SEVERITY_TO_RISK.get(incident.severity, ApprovalRiskLevel.MEDIUM)
+            # 2. 呼叫 OpenClaw LLM 生成提案 (Phase 6.4 核心)
             target = incident.affected_services[0] if incident.affected_services else "unknown"
+            signals_dict = [s.model_dump() for s in incident.signals]
+
+            llm_proposal, provider, llm_success = await self._openclaw.generate_incident_proposal(
+                incident_id=incident_id,
+                severity=incident.severity.value,
+                signals=signals_dict,
+                affected_services=incident.affected_services,
+            )
+
+            # 使用 LLM 結果或 fallback 到模板
+            if llm_success and llm_proposal:
+                action = llm_proposal["action"]
+                description = f"{llm_proposal['description']}\n\n**AI 推理**: {llm_proposal['reasoning']}"
+                action_type = llm_proposal.get("primary_responsibility", "default").lower()
+
+                # LLM 提供的 risk_level 轉換
+                llm_risk = llm_proposal.get("risk_level", "medium")
+                risk_map = {
+                    "low": ApprovalRiskLevel.LOW,
+                    "medium": ApprovalRiskLevel.MEDIUM,
+                    "critical": ApprovalRiskLevel.CRITICAL,
+                }
+                base_risk = risk_map.get(llm_risk, ApprovalRiskLevel.MEDIUM)
+
+                logger.info(
+                    "llm_proposal_generated",
+                    incident_id=incident_id,
+                    provider=provider,
+                    action=action[:50],
+                    risk_level=llm_risk,
+                    confidence=llm_proposal.get("confidence", 0),
+                )
+            else:
+                # Fallback 到模板方案
+                logger.warning(
+                    "llm_proposal_fallback_to_template",
+                    incident_id=incident_id,
+                    provider=provider,
+                )
+                action_type, action, description = self._determine_action(incident)
+                base_risk = SEVERITY_TO_RISK.get(incident.severity, ApprovalRiskLevel.MEDIUM)
+
+            # 3. 評估風險等級 (TrustEngine 調整)
             action_pattern = normalize_action_pattern(action_type, {"resource": target})
 
             risk_adjustment = trust_engine.evaluate_adjusted_risk(
@@ -173,6 +218,24 @@ class ProposalService:
             blast_radius = self._build_blast_radius(incident)
             dry_run_checks = self._build_dry_run_checks(incident)
 
+            # 建立 metadata (含 LLM 資訊)
+            metadata = {
+                "incident_id": incident_id,
+                "severity": incident.severity.value,
+                "signal_count": len(incident.signals),
+                "affected_services": incident.affected_services,
+                "trust_adjustment": risk_adjustment.to_dict(),
+            }
+
+            # 加入 LLM 相關資訊 (Phase 6.4)
+            if llm_success and llm_proposal:
+                metadata["llm_provider"] = llm_proposal.get("provider", "unknown")
+                metadata["llm_confidence"] = llm_proposal.get("confidence", 0)
+                metadata["llm_from_cache"] = llm_proposal.get("from_cache", False)
+                metadata["kubectl_command"] = llm_proposal.get("kubectl_command", "")
+                metadata["signoz_correlation"] = llm_proposal.get("signoz_correlation", "")
+                metadata["optimization_suggestions"] = llm_proposal.get("optimization_suggestions", [])
+
             approval_create = ApprovalRequestCreate(
                 action=action,
                 description=description,
@@ -180,13 +243,7 @@ class ProposalService:
                 blast_radius=blast_radius,
                 dry_run_checks=dry_run_checks,
                 requested_by="OpenClaw AI",
-                metadata={
-                    "incident_id": incident_id,
-                    "severity": incident.severity.value,
-                    "signal_count": len(incident.signals),
-                    "affected_services": incident.affected_services,
-                    "trust_adjustment": risk_adjustment.to_dict(),
-                },
+                metadata=metadata,
             )
 
             approval = await self._approval_service.create_approval(approval_create)