26 - 运行监控面板

// BEFORE_AGENT_EXECUTION 创建 Span 时设置的标签:
Span span = tracer.spanBuilder("agent.execute")
    .setAttribute("agent.id", agentId)
    .setAttribute("task.id", taskId)
    .setAttribute("conversation.id", conversationId)
    .setAttribute("model.id", modelId)
    .startSpan();

// 跨生命周期 Span 追踪:
// Agent 执行是异步的，BEFORE 在一个线程创建 Span，
// AFTER/ERROR/TIMEOUT 可能在另一个线程关闭 Span
ConcurrentHashMap<String, Span> activeSpans = new ConcurrentHashMap<>();

// BEFORE: 存入
activeSpans.put(taskId, span);

// AFTER/ERROR/TIMEOUT: 取出并关闭
Span span = activeSpans.remove(taskId);
if (span != null) {
    span.end();
}

GET /api/debug/interceptor-stats

{
  "modelInterceptors": [
    {"name": "LoggingModelInterceptor", "order": 10}
  ],
  "toolInterceptors": [
    {"name": "CachingToolInterceptor", "order": 20},
    {"name": "RetryToolInterceptor", "order": 30}
  ],
  "cacheStats": {
    "cacheSize": 42,
    "cacheHits": 156,
    "cacheMisses": 89,
    "hitRate": 0.637,
    "ttlMs": 300000,
    "lastEviction": "2026-04-28T10:30:00Z"
  },
  "retryStats": {
    "maxRetries": 2,
    "baseDelayMs": 500,
    "totalRetries": 23,
    "retriesExhausted": 3,
    "avgRetryDelayMs": 750
  }
}

GET /api/debug/context-stats

{
  "totalInvocations": 1250,
  "strategyUsage": {
    "COMPOSITE": 820,
    "SLIDING_WINDOW": 215,
    "SUMMARIZE": 180,
    "TRIM": 35
  },
  "avgMessagesBeforeTrim": 28.5,
  "avgMessagesAfterTrim": 8.2,
  "avgReductionRate": 0.712,
  "summarizeFallbackCount": 3,
  "truncationTriggered": 12
}

GET /api/debug/memory-stats

{
  "totalManaged": 580,
  "strategyUsage": {
    "DELETE": 320,
    "SUMMARIZE": 150,
    "TRIM": 80,
    "FILTER": 30
  },
  "avgMessagesDeleted": 15.3,
  "summarizeSuccess": 145,
  "summarizeFailed": 5,
  "filterMatchRate": 0.42
}

GET /api/debug/orchestration-stats

{
  "totalOrchestrations": 340,
  "modeUsage": {
    "SEQUENTIAL": 180,
    "PARALLEL": 95,
    "LLM_ROUTING": 45,
    "LOOP": 20
  },
  "avgDurationMs": {
    "SEQUENTIAL": 12500,
    "PARALLEL": 4200,
    "LLM_ROUTING": 3800,
    "LOOP": 18900
  },
  "parallelStats": {
    "avgAgentsPerExecution": 3.2,
    "errorIsolationTriggered": 8,
    "timeoutCount": 2
  },
  "loopStats": {
    "avgIterations": 3.5,
    "maxIterationsReached": 1
  },
  "routingStats": {
    "fallbackTriggered": 4,
    "routingAccuracy": 0.911
  }
}

{
  "graphInfra": {
    "checkpointSaver": "memory",
    "observationEnabled": true,
    "diagramEnabled": true,
    "observationRegistryActive": false,
    "longTermMemoryStoreAvailable": true,
    "agentRuntimeModes": {
      "chat": "GRAPH_PARALLEL",
      "translate": "LEGACY",
      "copywriting": "LEGACY"
    }
  }
}

public interface MiddlewareMonitor {
    String getName();                  // 组件名称，如 "MySQL"
    String getType();                  // 类型标识，如 "relational_database"
    Map<String, Object> getStatus();   // 运行状态快照
}

// 连接池指标采集 (HikariCP)
HikariPoolMXBean poolBean = hds.getHikariPoolMXBean();
pool.put("activeConnections", poolBean.getActiveConnections());
pool.put("idleConnections", poolBean.getIdleConnections());
pool.put("totalConnections", poolBean.getTotalConnections());
pool.put("maximumPoolSize", hds.getMaximumPoolSize());

// 响应时间测试
SELECT 1  →  responseTimeMs

// 运行时长
SHOW STATUS LIKE 'Uptime'  →  "12h 35m"

// 主探测：临时 MilvusClientV2，3s 超时，最多重试 2 次
ConnectConfig config = ConnectConfig.builder()
    .uri("http://" + host + ":" + port)
    .connectTimeoutMs(3000).build();
MilvusClientV2 client = new MilvusClientV2(config);
ListCollectionsResp resp = client.listCollections();  // 验证连通

// 遍历所有集合采集向量总数
for (String collName : resp.getCollectionNames()) {
    GetCollectionStatsResp stats = client.getCollectionStats(
        GetCollectionStatsReq.builder().collectionName(collName).build()
    );
    totalVectors += stats.getNumOfEntities();
}

// 备用探测：Spring AI VectorStore.similaritySearch("ping")
// Fallback 检测：vectorStore.getClass().isAnonymousClass()

// 24h 快速概览
List<LlmCallStatsRecord> records = repository.findByStatHourBetween(
    LocalDateTime.now().minusHours(24), LocalDateTime.now()
);
int totalCalls = records.stream().mapToInt(r -> r.getCallCount()).sum();
int successCalls = records.stream().mapToInt(r -> r.getSuccessCount()).sum();
long totalTokens = inputTokens + outputTokens;
double successRate = (double) successCalls / totalCalls * 100;

// 趋势数据按小时分组
Map<LocalDateTime, List<LlmCallStatsRecord>> byHour =
    records.stream().collect(Collectors.groupingBy(r -> r.getStatHour()));
// 每小时: callCount, successCount, failureCount,
//         inputTokens, outputTokens, cacheTokens, reasoningTokens

// 模型测试
ChatResponse response = chatModel.call(new Prompt(prompt));
tokensPerSecond = completionTokens / (elapsedMs / 1000.0);

// CacheRegistry — 显式注册，不依赖 Spring Cache 抽象
@Component
public class CacheRegistry {
    private final Map<String, Cache<?, ?>> caches = new ConcurrentHashMap<>();
    public void register(String name, Cache<?, ?> cache) { caches.put(name, cache); }
}

// CacheMonitorService — 采集每个缓存的 Caffeine Stats
var stats = cache.stats();
stats.hitCount();        // 命中次数
stats.missCount();       // 未命中次数
stats.evictionCount();   // 淘汰次数
cache.estimatedSize();   // 当前条目数

// 整体命中率
overallHitRate = totalHits / (totalHits + totalMisses)

// 管理操作
cache.invalidate(key);   // 删除单条目
cache.invalidateAll();   // 清空全部

// Redis: 通过 RedisCallback 执行 PING + 解析 INFO 原始输出
@Service
@ConditionalOnClass(RedisTemplate.class)
public class RedisMonitorService implements MiddlewareMonitor {
    // PING 探测
    Object pong = redisTemplate.execute(
        (RedisCallback<Object>) RedisConnection::ping
    );
    // INFO 解析: version, mode, usedMemory, totalKeys, uptime, opsPerSec ...
    // 集群: clusterGetNodes() → host, port, flags
}

// RocketMQ: 检测 Producer 内部工厂链
@Service
@ConditionalOnClass(RocketMQTemplate.class)
public class RocketMQMonitorService implements MiddlewareMonitor {
    // Producer 可用性检测
    boolean available = producer.getDefaultMQProducerImpl()
        .getmQClientFactory() != null;
}