feat: 十轮网关优化 - 安全加固/可观测性/性能/可靠性

- SSE Keepalive Ping (15s心跳防止代理断连)
- Timing HTTP 头 (X-Timing-Queue/Inference/Total-Ms)
- Adapter Request-ID 传播到后端
- Session 清理日志回调
- Server 安全加固 (ReadHeaderTimeout/MaxHeaderBytes 防 slowloris)
- Usage Tracker 数据保留清理 (retentionDays + 定期清理)
- Config Reload 后 Adapter Registry 更新 (RegisterIfAbsent + RWMutex)
- Rate Limiter 空闲 Bucket 清理 (30分钟过期)
- Shutdown Drain 超时可配置 (ShutdownDrainSeconds)
- Config 模型字段校验增强 (provider/endpoint/actual_model)
- Auth 过期 Key 自动清理 (5分钟扫描)
- Admin API Rate Limiting
- Adapter Health Check 独立超时 (每个 adapter 3s)
- TCP 连接阶段超时 (DialContext 5s + KeepAlive 30s)
- 幂等键缓存、审计日志、Gzip 中间件、CORS Expose Headers
- Backpressure 响应头、熔断器 Prometheus 指标
- 连接池优化、Trace-ID 全链路传播

configs/config.yaml

@@ -1,14 +1,18 @@
 server:
   host: "0.0.0.0"
-  port: 8080
+  port: 39000
   admin_port: 8081
   max_request_body_mb: 20
+  cors_allowed_origins: ["*"]
 
 auth:
   enabled: true
   methods: [api_key]
   jwt_issuer: edge-ai-gateway
   jwt_secret_env: EDGEAI_JWT_SECRET
+  rate_limit_per_minute: 60
+  rate_limit_burst: 10
+  usage_window_minutes: 60
 
 scheduler:
   max_running_tasks: 8
@@ -21,9 +25,9 @@ timeouts:
   default_connect_ms: 5000
   default_queue_ms: 5000
   default_first_token_ms: 10000
-  default_inference_ms: 60000
+  default_inference_ms: 120000
   default_idle_ms: 15000
-  default_total_ms: 90000
+  default_total_ms: 180000
   cancel_grace_period_ms: 3000
 
 context:
@@ -32,11 +36,14 @@ context:
   max_session_messages: 200
   session_idle_ttl_minutes: 60
   enable_prompt_persistence: false
+  enable_llm_summary: true
+  summary_max_tokens: 256
+  summary_timeout_seconds: 15
 
 models:
   general-chat:
     provider: ollama
-    actual_model: deepseek-r1:1.5b
+    actual_model: qwen2:latest
     endpoint: http://127.0.0.1:11434
     context_window: 32768
     max_output_tokens: 4096
@@ -46,7 +53,7 @@ models:
 
   fast-chat:
     provider: ollama
-    actual_model: deepseek-r1:1.5b
+    actual_model: qwen2:latest
     endpoint: http://127.0.0.1:11434
     context_window: 16384
     max_output_tokens: 2048
@@ -54,16 +61,6 @@ models:
     residency: on_demand
     idle_unload_seconds: 600
 
-  vllm-chat:
-    provider: vllm
-    actual_model: deepseek-r1:1.5b
-    endpoint: http://127.0.0.1:8000
-    context_window: 32768
-    max_output_tokens: 4096
-    max_concurrency: 4
-    residency: always
-    cancel_supported: true
-
 routing:
   sensitive_data_local_only: true
   allow_cloud_fallback_by_default: false