feat(agents): add Redis checkpoint for agent loop crash recovery

- New RedisClientService: optional ioredis wrapper, gracefully degrades without REDIS_URL
- New RedisModule: global NestJS module providing Redis connectivity
- AgentCheckpoint interface: captures turn, messages, cost, agents, timestamp
- Agent loop saves checkpoint after each tool execution batch (TTL=10min)
- On restart with same conversationId+requestId, loads checkpoint and resumes from saved state
- Checkpoint auto-deleted after load to prevent stale recovery
- Coordinator injects @Optional() RedisClientService, builds save/load callbacks
- Zero impact when Redis is not configured — checkpoint silently skipped

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

packages/services/conversation-service/src/infrastructure/agents/agents.module.ts

@@ -49,6 +49,9 @@ import { EVALUATION_RULE_REPOSITORY } from '../../domain/repositories/evaluation
 // Tool Hooks
 import { ToolHooksService } from './hooks/tool-hooks.service';
 
+// Redis (optional — checkpoint persistence)
+import { RedisModule } from '../cache/redis.module';
+
 /**
  * Anthropic Client Provider
  * 共享的 Anthropic SDK 实例，所有 Agent 共用
@@ -82,6 +85,7 @@ const AnthropicClientProvider = {
     TypeOrmModule.forFeature([TokenUsageORM, EvaluationRuleORM, ConversationORM]),
     McpModule,
     PaymentModule,
+    RedisModule,
   ],
   providers: [
     // Shared Anthropic client

packages/services/conversation-service/src/infrastructure/agents/coordinator/agent-loop.ts

@@ -125,14 +125,43 @@ export async function* agentLoop(
     traceId,
   } = params;
 
-  const currentTurn = params.currentTurnCount || 0;
+  let currentTurn = params.currentTurnCount || 0;
-  const currentCost = params.currentCostUsd || 0;
+  let currentCost = params.currentCostUsd || 0;
   /** 日志前缀 — 包含 traceId 用于端到端追踪 */
   const lp = traceId ? `[trace:${traceId}]` : `[conv:${conversationId}]`;
   let totalInputTokens = 0;
   let totalOutputTokens = 0;
   const agentsUsed: SpecialistAgentType[] = [];
 
+  // ---- Checkpoint Recovery (if available) ----
+  // 首轮（currentTurn=0）尝试从 Redis 恢复上次崩溃前的状态
+  if (currentTurn === 0 && params.checkpoint) {
+    try {
+      const saved = await params.checkpoint.load();
+      if (saved) {
+        logger.log(`${lp} Recovering from checkpoint at turn ${saved.turn} (${Date.now() - saved.timestamp}ms ago)`);
+        // 用 checkpoint 的 messages 替代当前 messages，跳到 saved turn 继续
+        yield* agentLoop(
+          anthropicClient,
+          {
+            ...params,
+            messages: saved.messages,
+            currentTurnCount: saved.turn,
+            currentCostUsd: saved.costSoFar,
+            checkpoint: undefined, // 恢复后不再尝试加载，只保存
+          },
+          toolExecutor,
+          additionalTools,
+          additionalConcurrencyMap,
+          progressEventSink,
+        );
+        return;
+      }
+    } catch (err) {
+      logger.warn(`${lp} Checkpoint load failed (starting fresh): ${err}`);
+    }
+  }
+
   // ---- Safety Checks ----
   if (currentTurn >= maxTurns) {
     yield {
@@ -545,6 +574,22 @@ export async function* agentLoop(
     { role: 'user', content: toolResultMessages },
   ];
 
+  // ---- Save Checkpoint (if configured) ----
+  if (params.checkpoint) {
+    try {
+      await params.checkpoint.save({
+        turn: currentTurn + 1,
+        messages: nextMessages,
+        costSoFar: currentCost + turnCost,
+        agentsUsed,
+        timestamp: Date.now(),
+      });
+      logger.debug(`${lp} Checkpoint saved at turn ${currentTurn + 1}`);
+    } catch (err) {
+      logger.warn(`${lp} Checkpoint save failed (non-fatal): ${err}`);
+    }
+  }
+
   // ---- Recurse ----
   yield* agentLoop(
     anthropicClient,

packages/services/conversation-service/src/infrastructure/agents/coordinator/coordinator-agent.service.ts

@@ -9,7 +9,7 @@
  * 4. 产出流式事件给 ConversationService
  */
 
-import { Injectable, OnModuleInit } from '@nestjs/common';
+import { Injectable, OnModuleInit, Optional } from '@nestjs/common';
 import { ConfigService } from '@nestjs/config';
 import Anthropic from '@anthropic-ai/sdk';
 import { Logger } from '@nestjs/common';
@@ -37,6 +37,7 @@ import {
   ClaudeMessage,
   SystemPromptBlock,
   SpecialistAgentType,
+  CheckpointCallbacks,
 } from '../types/agent.types';
 import { StreamEvent } from '../types/stream.types';
 import { ConversationContext } from '../types/context.types';
@@ -61,6 +62,9 @@ import { InputGateService } from './input-gate.service';
 import { ToolHooksService } from '../hooks/tool-hooks.service';
 import { ToolType } from '../hooks/tool-hooks.types';
 
+// Redis (optional checkpoint persistence)
+import { RedisClientService } from '../../cache/redis-client.service';
+
 // ============================================================
 // Compatibility Types (与 ClaudeAgentServiceV2 的 StreamChunk 兼容)
 // ============================================================
@@ -157,6 +161,8 @@ export class CoordinatorAgentService implements OnModuleInit {
     private readonly tenantContext: TenantContextService,
     // Tool hooks (PreToolUse/PostToolUse interception)
     private readonly toolHooks: ToolHooksService,
+    // Redis (optional — for checkpoint persistence)
+    @Optional() private readonly redisClient?: RedisClientService,
   ) {}
 
   onModuleInit() {
@@ -287,6 +293,9 @@ export class CoordinatorAgentService implements OnModuleInit {
         return gateResult;
       };
 
+      // 5.5. Build checkpoint callbacks (if Redis available)
+      const checkpoint = this.buildCheckpointCallbacks(context.conversationId, traceId);
+
       // 6. Build agent loop params
       const loopParams: AgentLoopParams = {
         messages: enrichedMessages,
@@ -301,6 +310,7 @@ export class CoordinatorAgentService implements OnModuleInit {
         currentTurnCount: 0,
         currentCostUsd: 0,
         evaluationGate: evaluationGateCallback,
+        checkpoint,
       };
 
       // 6. Create progress event sink (shared between tool executor and agent loop)
@@ -782,6 +792,42 @@ export class CoordinatorAgentService implements OnModuleInit {
     }
   }
 
+  // ============================================================
+  // Checkpoint Persistence
+  // ============================================================
+
+  /** Checkpoint TTL: 10 minutes */
+  private readonly CHECKPOINT_TTL_SECONDS = 600;
+
+  /**
+   * 构建 checkpoint 回调 — 如果 Redis 不可用则返回 undefined
+   */
+  private buildCheckpointCallbacks(
+    conversationId: string,
+    requestId: string,
+  ): CheckpointCallbacks | undefined {
+    if (!this.redisClient?.isAvailable()) return undefined;
+
+    const key = `agent:checkpoint:${conversationId}:${requestId}`;
+
+    return {
+      save: async (checkpoint) => {
+        await this.redisClient!.setex(
+          key,
+          this.CHECKPOINT_TTL_SECONDS,
+          JSON.stringify(checkpoint),
+        );
+      },
+      load: async () => {
+        const raw = await this.redisClient!.get(key);
+        if (!raw) return null;
+        // 加载后删除 — 避免过期 checkpoint 重复恢复
+        await this.redisClient!.del(key);
+        return JSON.parse(raw);
+      },
+    };
+  }
+
   // ============================================================
   // Event Mapping
   // ============================================================

packages/services/conversation-service/src/infrastructure/agents/types/agent.types.ts

@@ -262,6 +262,21 @@ export interface CoordinatorTurnStats {
 // Coordinator Loop Parameters
 // ============================================================
 
+/** Agent Loop 状态快照 — 用于崩溃恢复 */
+export interface AgentCheckpoint {
+  turn: number;
+  messages: ClaudeMessage[];
+  costSoFar: number;
+  agentsUsed: string[];
+  timestamp: number;
+}
+
+/** Checkpoint 回调 — save/load 由 coordinator 注入，agent-loop 调用 */
+export interface CheckpointCallbacks {
+  save: (checkpoint: AgentCheckpoint) => Promise<void>;
+  load: () => Promise<AgentCheckpoint | null>;
+}
+
 /** Agent Loop 参数 */
 export interface AgentLoopParams {
   messages: ClaudeMessage[];
@@ -284,6 +299,8 @@ export interface AgentLoopParams {
     turnCount: number,
     agentsUsed: string[],
   ) => Promise<import('../coordinator/evaluation-gate.service').GateResult>;
+  /** 可选 checkpoint 回调 — 提供时在每轮 tool 执行后保存状态快照 */
+  checkpoint?: CheckpointCallbacks;
 }
 
 /** Claude API 消息格式 */

packages/services/conversation-service/src/infrastructure/cache/redis-client.service.ts

@@ -0,0 +1,82 @@
+/**
+ * Redis Client Service
+ * 可选的 Redis 连接 — 如果 REDIS_URL 未配置则静默跳过
+ *
+ * 用途：Agent Loop 状态快照（checkpoint）
+ */
+
+import { Injectable, OnModuleInit, OnModuleDestroy, Logger, Optional } from '@nestjs/common';
+import { ConfigService } from '@nestjs/config';
+import Redis from 'ioredis';
+
+@Injectable()
+export class RedisClientService implements OnModuleInit, OnModuleDestroy {
+  private readonly logger = new Logger(RedisClientService.name);
+  private client: Redis | null = null;
+
+  constructor(@Optional() private readonly configService?: ConfigService) {}
+
+  async onModuleInit() {
+    const redisUrl = this.configService?.get<string>('REDIS_URL');
+    if (!redisUrl) {
+      this.logger.log('REDIS_URL not configured — checkpoint persistence disabled');
+      return;
+    }
+
+    try {
+      this.client = new Redis(redisUrl, {
+        maxRetriesPerRequest: 3,
+        lazyConnect: true,
+        connectTimeout: 5000,
+      });
+      await this.client.connect();
+      this.logger.log(`Connected to Redis: ${redisUrl.replace(/\/\/.*@/, '//*****@')}`);
+    } catch (error) {
+      this.logger.warn(`Redis connection failed (checkpoint disabled): ${error}`);
+      this.client = null;
+    }
+  }
+
+  async onModuleDestroy() {
+    if (this.client) {
+      await this.client.quit();
+      this.client = null;
+    }
+  }
+
+  /** Redis 是否可用 */
+  isAvailable(): boolean {
+    return this.client !== null && this.client.status === 'ready';
+  }
+
+  /** SET with TTL (seconds) */
+  async setex(key: string, ttlSeconds: number, value: string): Promise<void> {
+    if (!this.client) return;
+    try {
+      await this.client.setex(key, ttlSeconds, value);
+    } catch (error) {
+      this.logger.debug(`Redis setex failed: ${error}`);
+    }
+  }
+
+  /** GET */
+  async get(key: string): Promise<string | null> {
+    if (!this.client) return null;
+    try {
+      return await this.client.get(key);
+    } catch (error) {
+      this.logger.debug(`Redis get failed: ${error}`);
+      return null;
+    }
+  }
+
+  /** DEL */
+  async del(key: string): Promise<void> {
+    if (!this.client) return;
+    try {
+      await this.client.del(key);
+    } catch (error) {
+      this.logger.debug(`Redis del failed: ${error}`);
+    }
+  }
+}

packages/services/conversation-service/src/infrastructure/cache/redis.module.ts

@@ -0,0 +1,16 @@
+/**
+ * Redis Module
+ * 可选的 Redis 连接模块 — 提供 RedisClientService
+ */
+
+import { Module, Global } from '@nestjs/common';
+import { ConfigModule } from '@nestjs/config';
+import { RedisClientService } from './redis-client.service';
+
+@Global()
+@Module({
+  imports: [ConfigModule],
+  providers: [RedisClientService],
+  exports: [RedisClientService],
+})
+export class RedisModule {}