diff --git a/backend/services/ai_chat_budget_test.go b/backend/services/ai_chat_budget_test.go
new file mode 100644
index 0000000..8dae92f
--- /dev/null
+++ b/backend/services/ai_chat_budget_test.go
@@ -0,0 +1,34 @@
+package services
+
+import (
+	"strings"
+	"testing"
+)
+
+// TestChatHistoryBudgetTrim 历史按字符预算裁剪:超预算时最旧被丢、标记注入。
+// 通过构造超长历史走 RunAIChat 会因未启用而提前返回,这里直接测裁剪语义的独立函数化验证——
+// 为可测性把裁剪抽成了 trimHistory。
+func TestChatHistoryBudgetTrim(t *testing.T) {
+	long := strings.Repeat("a", 15000)
+	history := []ChatMessage{
+		{Role: "user", Content: long},
+		{Role: "assistant", Content: long},
+		{Role: "user", Content: "最后一句"},
+	}
+	kept, truncated := trimHistory(history, 24000)
+	if !truncated {
+		t.Fatal("超出预算应标记 truncated")
+	}
+	if len(kept) != 2 {
+		t.Fatalf("应保留最近 2 条(丢弃最旧的 15000 字条目),got %d", len(kept))
+	}
+	if kept[0].Content != long || kept[len(kept)-1].Content != "最后一句" {
+		t.Fatal("应保留最近的消息")
+	}
+	// 不超预算时全保留
+	small := []ChatMessage{{Role: "user", Content: "hi"}, {Role: "assistant", Content: "ok"}}
+	kept2, truncated2 := trimHistory(small, 24000)
+	if truncated2 || len(kept2) != 2 {
+		t.Fatalf("小历史不应裁剪: %v %d", truncated2, len(kept2))
+	}
+}
diff --git a/backend/services/ai_provider.go b/backend/services/ai_provider.go
index 09c4deb..f4cc9b4 100644
--- a/backend/services/ai_provider.go
+++ b/backend/services/ai_provider.go
@@ -162,6 +162,37 @@ type ChatMessage struct {
 	Content string `json:"content" binding:"required"`
 }
 
+// trimHistory 按字符预算从最旧开始裁剪对话历史(指南 07 TrimMessages);
+// 返回保留的历史与是否发生了裁剪(调用方在 system 里标注省略)
+func trimHistory(history []ChatMessage, budget int) ([]ChatMessage, bool) {
+	var kept []ChatMessage
+	total := 0
+	truncated := false
+	for i := len(history) - 1; i >= 0; i-- {
+		size := len([]rune(history[i].Content))
+		if total+size > budget && len(kept) > 0 {
+			truncated = true
+			break
+		}
+		kept = append([]ChatMessage{{Role: history[i].Role, Content: history[i].Content}}, kept...)
+		total += size
+	}
+	return kept, truncated
+}
+
+// isRetryableAIError 网络抖动/限流/上游临时故障才值得重试;鉴权与参数错重试无意义
+func isRetryableAIError(err error) bool {
+	if err == nil {
+		return false
+	}
+	msg := err.Error()
+	return strings.Contains(msg, "429") ||
+		strings.Contains(msg, "500") || strings.Contains(msg, "502") ||
+		strings.Contains(msg, "503") || strings.Contains(msg, "504") ||
+		strings.Contains(msg, "connection refused") || strings.Contains(msg, "context deadline exceeded") ||
+		strings.Contains(msg, "EOF") || strings.Contains(msg, "timeout")
+}
+
 // RunAIChat 全站 AI 助手浮窗:多轮对话 + 当前页面上下文注入(轻量信号摘要)。
 // 与整页分析共享配置/配额/脱敏/Pro 闸门;context 里的 sourceId 可选,带了就注入日志摘要。
 func RunAIChat(ctx context.Context, cfg models.AIConfig, userID uint, history []ChatMessage, page string, pageContext string) (string, error) {
@@ -169,7 +200,8 @@ func RunAIChat(ctx context.Context, cfg models.AIConfig, userID uint, history []
 		return "", fmt.Errorf("AI 未启用:请先在「智能运维 → 模型配置」中开启并保存")
 	}
 	if cfg.DailyTokenCap > 0 && AITokenUsedToday(userID) >= cfg.DailyTokenCap {
-		return "", fmt.Errorf("今日 AI 用量已达上限,明天再试或调整上限")
+		// 指南 05.3:超限不伪装成功 —— 结构化 notice 而非报错,对话上下文在前端保留
+		return "\u2014\u2014 今日 AI 配额已用完(" + fmt.Sprintf("%d tokens),明天再试或在「AI 模型接入」调整上限 \u2014\u2014", cfg.DailyTokenCap), nil
 	}
 
 	// 提示词设计:身份固定(不猜底层模型)、风格收敛、话题回引导、上下文数据
@@ -187,8 +219,13 @@ func RunAIChat(ctx context.Context, cfg models.AIConfig, userID uint, history []
 		}
 		system += "\n\n" + RedactTextOnly(trimmed, cfg.RedactExtra)
 	}
+	// 上下文预算(指南 07 TrimMessages):按字符预算从最旧开始丢弃,优先保留近期对话
+	trimmed, truncated := trimHistory(history, 24000)
+	if truncated {
+		system += "\n(更早的对话已省略)"
+	}
 	msgs := []aiMessage{{Role: "system", Content: system}}
-	for _, m := range history {
+	for _, m := range trimmed {
 		msgs = append(msgs, aiMessage{Role: m.Role, Content: m.Content})
 	}
 	// 脱敏:用户消息里的敏感串同样替换
@@ -197,10 +234,27 @@ func RunAIChat(ctx context.Context, cfg models.AIConfig, userID uint, history []
 			msgs[i].Content = RedactTextOnly(msgs[i].Content, cfg.RedactExtra)
 		}
 	}
-	reply, promptTokens, completionTokens, err := aiChat(ctx, cfg, msgs)
+	// 指南 05.2:429/5xx/网络错误指数退避重试(1s→2s,最多 2 次);鉴权/参数错不重试
+	var reply string
+	var promptTokens, completionTokens int
+	var err error
+	for attempt := 0; ; attempt++ {
+		reply, promptTokens, completionTokens, err = aiChat(ctx, cfg, msgs)
+		if err == nil || attempt >= 2 || !isRetryableAIError(err) {
+			break
+		}
+		backoff := time.Duration(1<<attempt) * time.Second
+		log.Printf("[ai-chat] 可重试错误(%v),%s 后第 %d 次重试", err, backoff, attempt+1)
+		select {
+		case <-time.After(backoff):
+		case <-ctx.Done():
+			return "", ctx.Err()
+		}
+	}
 	if err != nil {
 		return "", err
 	}
+	reply = RedactTextOnly(reply, cfg.RedactExtra)
 	// 计入每日配额:落一条 chat 类型用量记录
 	db := repository.GetDB()
 	if db != nil {
diff --git a/frontend/src/components/AiAssistant.tsx b/frontend/src/components/AiAssistant.tsx
index 86041e2..0dfe529 100644
--- a/frontend/src/components/AiAssistant.tsx
+++ b/frontend/src/components/AiAssistant.tsx
@@ -85,17 +85,22 @@ export const AiAssistant = memo(function AiAssistant() {
             {messages.length === 0 && (
               <p className="py-8 text-center text-xs text-muted-foreground">{t('assistant.emptyHint')}</p>
             )}
-            {messages.map((m, i) => (
-              <div key={i} className={m.role === 'user' ? 'flex justify-end' : 'flex justify-start'}>
-                <div className={
-                  m.role === 'user'
-                    ? 'max-w-[85%] rounded-lg bg-primary px-3 py-2 text-primary-foreground'
-                    : 'max-w-[85%] whitespace-pre-wrap rounded-lg bg-muted px-3 py-2'
-                }>
-                  {m.content}
+            {messages.map((m, i) => {
+              const isQuota = m.role === 'assistant' && m.content.includes('今日 AI 配额已用完')
+              return (
+                <div key={i} className={m.role === 'user' ? 'flex justify-end' : 'flex justify-start'}>
+                  <div className={
+                    isQuota
+                      ? 'max-w-[85%] whitespace-pre-wrap rounded-lg border border-warning/40 bg-warning/10 px-3 py-2 text-warning'
+                      : m.role === 'user'
+                        ? 'max-w-[85%] rounded-lg bg-primary px-3 py-2 text-primary-foreground'
+                        : 'max-w-[85%] whitespace-pre-wrap rounded-lg bg-muted px-3 py-2'
+                  }>
+                    {m.content}
+                  </div>
                 </div>
-              </div>
-            ))}
+              )
+            })}
             {sending && (
               <div className="flex justify-start">
                 <div className="flex items-center gap-2 rounded-lg bg-muted px-3 py-2 text-xs text-muted-foreground">