diff --git a/backend/internal/service/gemini_messages_compat_service.go b/backend/internal/service/gemini_messages_compat_service.go
index 335e1f81..792c8f4b 100644
--- a/backend/internal/service/gemini_messages_compat_service.go
+++ b/backend/internal/service/gemini_messages_compat_service.go
@@ -770,6 +770,14 @@ func (s *GeminiMessagesCompatService) Forward(ctx context.Context, c *gin.Contex
 			break
 		}
 
+		// 错误策略优先：匹配则跳过重试直接处理。
+		if matched, rebuilt := s.checkErrorPolicyInLoop(ctx, account, resp); matched {
+			resp = rebuilt
+			break
+		} else {
+			resp = rebuilt
+		}
+
 		if resp.StatusCode >= 400 && s.shouldRetryGeminiUpstreamError(account, resp.StatusCode) {
 			respBody, _ := io.ReadAll(io.LimitReader(resp.Body, 2<<20))
 			_ = resp.Body.Close()
@@ -1176,6 +1184,14 @@ func (s *GeminiMessagesCompatService) ForwardNative(ctx context.Context, c *gin.
 			return nil, s.writeGoogleError(c, http.StatusBadGateway, "Upstream request failed after retries: "+safeErr)
 		}
 
+		// 错误策略优先：匹配则跳过重试直接处理。
+		if matched, rebuilt := s.checkErrorPolicyInLoop(ctx, account, resp); matched {
+			resp = rebuilt
+			break
+		} else {
+			resp = rebuilt
+		}
+
 		if resp.StatusCode >= 400 && s.shouldRetryGeminiUpstreamError(account, resp.StatusCode) {
 			respBody, _ := io.ReadAll(io.LimitReader(resp.Body, 2<<20))
 			_ = resp.Body.Close()
@@ -1425,6 +1441,26 @@ func (s *GeminiMessagesCompatService) ForwardNative(ctx context.Context, c *gin.
 	}, nil
 }
 
+// checkErrorPolicyInLoop 在重试循环内预检查错误策略。
+// 返回 true 表示策略已匹配（调用者应 break），resp 已重建可直接使用。
+// 返回 false 表示 ErrorPolicyNone，resp 已重建，调用者继续走重试逻辑。
+func (s *GeminiMessagesCompatService) checkErrorPolicyInLoop(
+	ctx context.Context, account *Account, resp *http.Response,
+) (matched bool, rebuilt *http.Response) {
+	if resp.StatusCode < 400 || s.rateLimitService == nil {
+		return false, resp
+	}
+	body, _ := io.ReadAll(io.LimitReader(resp.Body, 2<<20))
+	_ = resp.Body.Close()
+	rebuilt = &http.Response{
+		StatusCode: resp.StatusCode,
+		Header:     resp.Header.Clone(),
+		Body:       io.NopCloser(bytes.NewReader(body)),
+	}
+	policy := s.rateLimitService.CheckErrorPolicy(ctx, account, resp.StatusCode, body)
+	return policy != ErrorPolicyNone, rebuilt
+}
+
 func (s *GeminiMessagesCompatService) shouldRetryGeminiUpstreamError(account *Account, statusCode int) bool {
 	switch statusCode {
 	case 429, 500, 502, 503, 504, 529: