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

internal/context/adapter_summary.go

@@ -0,0 +1,79 @@
+package context
+
+import (
+	"context"
+	"fmt"
+	"strings"
+
+	"github.com/edgeai/gateway/internal/adapter"
+	"github.com/edgeai/gateway/pkg/api"
+)
+
+// AdapterSummaryGenerator 使用 adapter 调用 LLM 生成对话摘要。
+type AdapterSummaryGenerator struct {
+	adapter     adapter.ModelAdapter
+	model       string
+	maxTokens   int
+	temperature float64
+}
+
+// NewAdapterSummaryGenerator 创建基于 adapter 的摘要生成器。
+// model 为实际模型名（如 qwen2:latest），maxTokens 限制摘要长度。
+func NewAdapterSummaryGenerator(adp adapter.ModelAdapter, model string, maxTokens int) *AdapterSummaryGenerator {
+	temp := 0.3
+	return &AdapterSummaryGenerator{
+		adapter:     adp,
+		model:       model,
+		maxTokens:   maxTokens,
+		temperature: temp,
+	}
+}
+
+// GenerateSummary 调用 LLM 对历史消息生成摘要。
+func (g *AdapterSummaryGenerator) GenerateSummary(ctx context.Context, messages []api.Message) (string, error) {
+	if len(messages) == 0 {
+		return "", nil
+	}
+
+	// 构建摘要请求的 prompt
+	prompt := buildSummaryPrompt(messages)
+
+	req := &adapter.ChatRequest{
+		RequestID:   "summary-" + fmt.Sprintf("%d", len(messages)),
+		Model:       g.model,
+		Messages:    prompt,
+		MaxTokens:   g.maxTokens,
+		Temperature: &g.temperature,
+		Stream:      false,
+	}
+
+	resp, err := g.adapter.ChatCompletion(ctx, req)
+	if err != nil {
+		return "", fmt.Errorf("summary generation failed: %w", err)
+	}
+
+	return resp.Content, nil
+}
+
+// buildSummaryPrompt 构建用于生成摘要的消息列表。
+func buildSummaryPrompt(messages []api.Message) []api.Message {
+	// 将历史消息拼接为文本
+	var sb strings.Builder
+	for _, m := range messages {
+		role := m.Role
+		content, _ := m.Content.(string)
+		sb.WriteString(fmt.Sprintf("%s: %s\n", role, content))
+	}
+
+	return []api.Message{
+		{
+			Role: "system",
+			Content: "你是一个对话摘要助手。请将以下对话历史压缩为简洁的摘要，" +
+				"保留关键信息、用户意图和重要结论。摘要不超过200字。",
+		},
+		{
+			Role:    "user",
+			Content: fmt.Sprintf("请总结以下对话历史：\n\n%s", sb.String()),
+		},
+	}
+}

internal/context/assembler.go

@@ -7,8 +7,9 @@ import (
 
 // Assembler assembles context messages for a chat request.
 type Assembler struct {
-	estimator *TokenEstimator
-	cfg       *config.ContextConfig
+	estimator  *TokenEstimator
+	cfg        *config.ContextConfig
+	summarizer *Summarizer
 }
 
 // NewAssembler creates a new context assembler.
@@ -19,6 +20,11 @@ func NewAssembler(cfg *config.ContextConfig) *Assembler {
 	}
 }
 
+// SetSummarizer 注入摘要器，启用后 trimSummaryAndRecent 策略将生成真实 LLM 摘要。
+func (a *Assembler) SetSummarizer(s *Summarizer) {
+	a.summarizer = s
+}
+
 // AssembleResult contains the assembled messages and metadata.
 type AssembleResult struct {
 	Messages     []api.Message
@@ -143,13 +149,21 @@ func (a *Assembler) trimSummaryAndRecent(messages []api.Message, budget int) tri
 		recentTokens += msgTokens
 	}
 
-	// Add summary placeholder if we trimmed anything
+	// Add summary if we trimmed anything
 	result := make([]api.Message, 0, len(systemMsgs)+1+len(recentMsgs))
 	result = append(result, systemMsgs...)
 	if len(recentMsgs) < len(rest) {
+		// 被裁剪的老消息
+		trimmedMsgs := rest[:len(rest)-len(recentMsgs)]
+		var summaryText string
+		if a.summarizer != nil {
+			summaryText = a.summarizer.Summarize(trimmedMsgs)
+		} else {
+			summaryText = "[Earlier conversation history has been summarized and omitted.]"
+		}
 		result = append(result, api.Message{
 			Role:    "system",
-			Content: "[Earlier conversation history has been summarized and omitted.]",
+			Content: summaryText,
 		})
 	}
 	result = append(result, recentMsgs...)

internal/context/summarizer.go

@@ -0,0 +1,48 @@
+package context
+
+import (
+	"context"
+	"fmt"
+	"strings"
+	"time"
+
+	"github.com/edgeai/gateway/pkg/api"
+)
+
+// SummaryGenerator 定义摘要生成接口，允许注入不同的实现。
+type SummaryGenerator interface {
+	// GenerateSummary 对给定消息生成对话摘要，返回摘要文本。
+	GenerateSummary(ctx context.Context, messages []api.Message) (string, error)
+}
+
+// Summarizer 使用 LLM 对历史消息生成对话摘要。
+type Summarizer struct {
+	generator SummaryGenerator
+	timeout   time.Duration
+}
+
+// NewSummarizer 创建一个新的摘要器。
+func NewSummarizer(generator SummaryGenerator, timeout time.Duration) *Summarizer {
+	return &Summarizer{
+		generator: generator,
+		timeout:   timeout,
+	}
+}
+
+// Summarize 对被裁剪的老消息生成摘要。
+// 如果摘要生成失败，回退到占位符文本。
+func (s *Summarizer) Summarize(messages []api.Message) string {
+	if s == nil || s.generator == nil || len(messages) == 0 {
+		return "[Earlier conversation history has been summarized and omitted.]"
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), s.timeout)
+	defer cancel()
+
+	summary, err := s.generator.GenerateSummary(ctx, messages)
+	if err != nil || strings.TrimSpace(summary) == "" {
+		return "[Earlier conversation history has been summarized and omitted.]"
+	}
+
+	return fmt.Sprintf("[Earlier conversation summary: %s]", strings.TrimSpace(summary))
+}