diff --git a/backend/internal/service/openai_account_runtime_block_fastpath_test.go b/backend/internal/service/openai_account_runtime_block_fastpath_test.go
index 9a6475f23e..e013ba8831 100644
--- a/backend/internal/service/openai_account_runtime_block_fastpath_test.go
+++ b/backend/internal/service/openai_account_runtime_block_fastpath_test.go
@@ -123,6 +123,59 @@ func TestOpenAI429FastPath_SparkQuotaOnlyBlocksSparkModel(t *testing.T) {
 	require.Greater(t, time.Until(repo.lastModelRateLimitedUntil), 6*24*time.Hour)
 }
 
+func TestOpenAI429FastPath_SparkTransient429UsesShortFallback(t *testing.T) {
+	repo := &oauth429RateLimitRepo{}
+	rateLimits := NewRateLimitService(repo, nil, &config.Config{}, nil, nil)
+	svc := &OpenAIGatewayService{rateLimitService: rateLimits}
+	rateLimits.SetAccountRuntimeBlocker(svc)
+	account := &Account{ID: 428, Platform: PlatformOpenAI, Type: AccountTypeOAuth}
+	headers := http.Header{}
+	headers.Set("x-codex-primary-used-percent", "37")
+	headers.Set("x-codex-primary-reset-after-seconds", "604800")
+	headers.Set("x-codex-primary-window-minutes", "10080")
+	headers.Set("x-codex-secondary-used-percent", "20")
+	headers.Set("x-codex-secondary-reset-after-seconds", "3600")
+	headers.Set("x-codex-secondary-window-minutes", "300")
+
+	shouldDisable := svc.handleOpenAIAccountUpstreamError(
+		context.Background(), account, http.StatusTooManyRequests, headers,
+		[]byte(`{"error":{"type":"rate_limit_error","code":"rate_limit_exceeded"}}`),
+		"gpt-5.3-codex-spark",
+	)
+
+	require.False(t, shouldDisable)
+	require.Equal(t, 1, repo.setModelRateLimitCalls)
+	require.Less(t, time.Until(repo.lastModelRateLimitedUntil), time.Minute)
+	require.Greater(t, time.Until(repo.lastModelRateLimitedUntil), time.Second)
+}
+
+func TestOpenAIStream429_SparkQuotaUsesQuotaHeaders(t *testing.T) {
+	repo := &oauth429RateLimitRepo{}
+	rateLimits := NewRateLimitService(repo, nil, &config.Config{}, nil, nil)
+	svc := &OpenAIGatewayService{rateLimitService: rateLimits}
+	rateLimits.SetAccountRuntimeBlocker(svc)
+	account := &Account{ID: 429, Platform: PlatformOpenAI, Type: AccountTypeOAuth}
+	headers := http.Header{}
+	headers.Set("x-codex-primary-used-percent", "100")
+	headers.Set("x-codex-primary-reset-after-seconds", "604800")
+	headers.Set("x-codex-primary-window-minutes", "10080")
+	headers.Set("x-codex-secondary-used-percent", "20")
+	headers.Set("x-codex-secondary-reset-after-seconds", "3600")
+	headers.Set("x-codex-secondary-window-minutes", "300")
+	payload := []byte(`{"type":"error","error":{"type":"rate_limit_error","code":"rate_limit_exceeded"}}`)
+
+	status, shouldDisable := svc.handleOpenAIStreamTerminalAccountSideEffects(
+		nil, account, payload, "quota exhausted", headers, "gpt-5.3-codex-spark",
+	)
+
+	require.Equal(t, http.StatusTooManyRequests, status)
+	require.False(t, shouldDisable)
+	require.Equal(t, 1, repo.setModelRateLimitCalls)
+	require.Equal(t, "gpt-5.3-codex-spark", repo.lastModelRateLimitKey)
+	require.Greater(t, time.Until(repo.lastModelRateLimitedUntil), 6*24*time.Hour)
+	require.False(t, svc.isOpenAIAccountRuntimeBlocked(account))
+}
+
 func TestOpenAI429FastPath_SparkShadowQuotaStaysModelScoped(t *testing.T) {
 	repo := &oauth429RateLimitRepo{}
 	rateLimits := NewRateLimitService(repo, nil, &config.Config{}, nil, nil)
diff --git a/backend/internal/service/openai_gateway_passthrough.go b/backend/internal/service/openai_gateway_passthrough.go
index aaa74e4a7f..65df226dd9 100644
--- a/backend/internal/service/openai_gateway_passthrough.go
+++ b/backend/internal/service/openai_gateway_passthrough.go
@@ -1572,17 +1572,16 @@ func (s *OpenAIGatewayService) handleOpenAIStreamTerminalAccountSideEffects(
 		if c != nil && c.Request != nil {
 			ctx = c.Request.Context()
 		}
-		accountHeaders := headers
-		if statusCode == http.StatusTooManyRequests {
-			// The enclosing HTTP response succeeded. Its quota snapshot describes
-			// normal account state and must not become the reset for a semantic 429
-			// carried by a stream terminal event.
-			accountHeaders = nil
-		}
 		model := firstNonEmpty(canonicalModel...)
 		if model == "" {
 			model = firstNonEmpty(gjson.GetBytes(payload, "model").String(), gjson.GetBytes(payload, "response.model").String())
 		}
+		accountHeaders := headers
+		if statusCode == http.StatusTooManyRequests && !(isCodexSparkModel(model) && isOpenAIOAuthAccount(account)) {
+			// 普通模型的流式 429 不能继承外层 HTTP 200 的全局 quota 快照；
+			// 只有 OAuth/SetupToken 的 Spark 配额 429 才需要保留 headers 读取明确的 5h/7d reset。
+			accountHeaders = nil
+		}
 		return statusCode, s.handleOpenAIAccountUpstreamError(ctx, account, statusCode, accountHeaders, payload, model)
 	default:
 		return statusCode, false
diff --git a/backend/internal/service/ratelimit_service.go b/backend/internal/service/ratelimit_service.go
index e0263e8f3e..ec3c134e5a 100644
--- a/backend/internal/service/ratelimit_service.go
+++ b/backend/internal/service/ratelimit_service.go
@@ -2208,7 +2208,12 @@ func (s *RateLimitService) HandleOpenAICodexSparkRateLimit(ctx context.Context,
 		return false
 	}
 	now := time.Now()
-	_, resetAt := classifyOpenAIOAuth429(headers, responseBody)
+	disposition, resetAt := classifyOpenAIOAuth429(headers, responseBody)
+	// Spark 只有明确耗尽 5h/7d 窗口时才能使用上游长 reset；普通瞬时 429
+	// 即使携带全局 reset 头，也只能使用短时回避，避免错误冷却数天。
+	if disposition != openAIOAuth429Quota5h && disposition != openAIOAuth429Quota7d {
+		resetAt = nil
+	}
 	if resetAt == nil || !resetAt.After(now) {
 		cooldown, ok := s.get429FallbackCooldown(ctx, account)
 		if !ok || cooldown <= 0 {