初始化导入 new-api 源码

service/audio.go

@@ -0,0 +1,48 @@
+package service
+
+import (
+	"encoding/base64"
+	"fmt"
+	"strings"
+)
+
+func parseAudio(audioBase64 string, format string) (duration float64, err error) {
+	audioData, err := base64.StdEncoding.DecodeString(audioBase64)
+	if err != nil {
+		return 0, fmt.Errorf("base64 decode error: %v", err)
+	}
+
+	var samplesCount int
+	var sampleRate int
+
+	switch format {
+	case "pcm16":
+		samplesCount = len(audioData) / 2 // 16位 = 2字节每样本
+		sampleRate = 24000                // 24kHz
+	case "g711_ulaw", "g711_alaw":
+		samplesCount = len(audioData) // 8位 = 1字节每样本
+		sampleRate = 8000             // 8kHz
+	default:
+		samplesCount = len(audioData) // 8位 = 1字节每样本
+		sampleRate = 8000             // 8kHz
+	}
+
+	duration = float64(samplesCount) / float64(sampleRate)
+	return duration, nil
+}
+
+func DecodeBase64AudioData(audioBase64 string) (string, error) {
+	// 检查并移除 data:audio/xxx;base64, 前缀
+	idx := strings.Index(audioBase64, ",")
+	if idx != -1 {
+		audioBase64 = audioBase64[idx+1:]
+	}
+
+	// 解码 Base64 数据
+	_, err := base64.StdEncoding.DecodeString(audioBase64)
+	if err != nil {
+		return "", fmt.Errorf("base64 decode error: %v", err)
+	}
+
+	return audioBase64, nil
+}

service/billing.go

@@ -0,0 +1,78 @@
+package service
+
+import (
+	"fmt"
+
+	"github.com/QuantumNous/new-api/logger"
+	relaycommon "github.com/QuantumNous/new-api/relay/common"
+	"github.com/QuantumNous/new-api/types"
+	"github.com/gin-gonic/gin"
+)
+
+const (
+	BillingSourceWallet       = "wallet"
+	BillingSourceSubscription = "subscription"
+)
+
+// PreConsumeBilling 根据用户计费偏好创建 BillingSession 并执行预扣费。
+// 会话存储在 relayInfo.Billing 上，供后续 Settle / Refund 使用。
+func PreConsumeBilling(c *gin.Context, preConsumedQuota int, relayInfo *relaycommon.RelayInfo) *types.NewAPIError {
+	session, apiErr := NewBillingSession(c, relayInfo, preConsumedQuota)
+	if apiErr != nil {
+		return apiErr
+	}
+	relayInfo.Billing = session
+	return nil
+}
+
+// ---------------------------------------------------------------------------
+// SettleBilling — 后结算辅助函数
+// ---------------------------------------------------------------------------
+
+// SettleBilling 执行计费结算。如果 RelayInfo 上有 BillingSession 则通过 session 结算，
+// 否则回退到旧的 PostConsumeQuota 路径（兼容按次计费等场景）。
+func SettleBilling(ctx *gin.Context, relayInfo *relaycommon.RelayInfo, actualQuota int) error {
+	if relayInfo.Billing != nil {
+		preConsumed := relayInfo.Billing.GetPreConsumedQuota()
+		delta := actualQuota - preConsumed
+
+		if delta > 0 {
+			logger.LogInfo(ctx, fmt.Sprintf("预扣费后补扣费：%s（实际消耗：%s，预扣费：%s）",
+				logger.FormatQuota(delta),
+				logger.FormatQuota(actualQuota),
+				logger.FormatQuota(preConsumed),
+			))
+		} else if delta < 0 {
+			logger.LogInfo(ctx, fmt.Sprintf("预扣费后返还扣费：%s（实际消耗：%s，预扣费：%s）",
+				logger.FormatQuota(-delta),
+				logger.FormatQuota(actualQuota),
+				logger.FormatQuota(preConsumed),
+			))
+		} else {
+			logger.LogInfo(ctx, fmt.Sprintf("预扣费与实际消耗一致，无需调整：%s（按次计费）",
+				logger.FormatQuota(actualQuota),
+			))
+		}
+
+		if err := relayInfo.Billing.Settle(actualQuota); err != nil {
+			return err
+		}
+
+		// 发送额度通知（订阅计费使用订阅剩余额度）
+		if actualQuota != 0 {
+			if relayInfo.BillingSource == BillingSourceSubscription {
+				checkAndSendSubscriptionQuotaNotify(relayInfo)
+			} else {
+				checkAndSendQuotaNotify(relayInfo, actualQuota-preConsumed, preConsumed)
+			}
+		}
+		return nil
+	}
+
+	// 回退：无 BillingSession 时使用旧路径
+	quotaDelta := actualQuota - relayInfo.FinalPreConsumedQuota
+	if quotaDelta != 0 {
+		return PostConsumeQuota(relayInfo, quotaDelta, relayInfo.FinalPreConsumedQuota, true)
+	}
+	return nil
+}

service/billing_session.go

@@ -0,0 +1,314 @@
+package service
+
+import (
+	"fmt"
+	"net/http"
+	"strings"
+	"sync"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/logger"
+	"github.com/QuantumNous/new-api/model"
+	relaycommon "github.com/QuantumNous/new-api/relay/common"
+	"github.com/QuantumNous/new-api/types"
+
+	"github.com/bytedance/gopkg/util/gopool"
+	"github.com/gin-gonic/gin"
+)
+
+// ---------------------------------------------------------------------------
+// BillingSession — 统一计费会话
+// ---------------------------------------------------------------------------
+
+// BillingSession 封装单次请求的预扣费/结算/退款生命周期。
+// 实现 relaycommon.BillingSettler 接口。
+type BillingSession struct {
+	relayInfo        *relaycommon.RelayInfo
+	funding          FundingSource
+	preConsumedQuota int  // 实际预扣额度（信任用户可能为 0）
+	tokenConsumed    int  // 令牌额度实际扣减量
+	fundingSettled   bool // funding.Settle 已成功，资金来源已提交
+	settled          bool // Settle 全部完成（资金 + 令牌）
+	refunded         bool // Refund 已调用
+	mu               sync.Mutex
+}
+
+// Settle 根据实际消耗额度进行结算。
+// 资金来源和令牌额度分两步提交：若资金来源已提交但令牌调整失败，
+// 会标记 fundingSettled 防止 Refund 对已提交的资金来源执行退款。
+func (s *BillingSession) Settle(actualQuota int) error {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if s.settled {
+		return nil
+	}
+	delta := actualQuota - s.preConsumedQuota
+	if delta == 0 {
+		s.settled = true
+		return nil
+	}
+	// 1) 调整资金来源（仅在尚未提交时执行，防止重复调用）
+	if !s.fundingSettled {
+		if err := s.funding.Settle(delta); err != nil {
+			return err
+		}
+		s.fundingSettled = true
+	}
+	// 2) 调整令牌额度
+	var tokenErr error
+	if !s.relayInfo.IsPlayground {
+		if delta > 0 {
+			tokenErr = model.DecreaseTokenQuota(s.relayInfo.TokenId, s.relayInfo.TokenKey, delta)
+		} else {
+			tokenErr = model.IncreaseTokenQuota(s.relayInfo.TokenId, s.relayInfo.TokenKey, -delta)
+		}
+		if tokenErr != nil {
+			common.SysLog(fmt.Sprintf("error adjusting token quota after funding settled (userId=%d, tokenId=%d, delta=%d): %s",
+				s.relayInfo.UserId, s.relayInfo.TokenId, delta, tokenErr.Error()))
+		}
+	}
+	if s.funding.Source() == BillingSourceSubscription {
+		s.relayInfo.SubscriptionPostDelta += int64(delta)
+	}
+	s.settled = true
+	return tokenErr
+}
+
+// Refund 退还所有预扣费，幂等安全，异步执行。
+func (s *BillingSession) Refund(c *gin.Context) {
+	s.mu.Lock()
+	if s.settled || s.refunded || !s.needsRefundLocked() {
+		s.mu.Unlock()
+		return
+	}
+	s.refunded = true
+	s.mu.Unlock()
+
+	logger.LogInfo(c, fmt.Sprintf("用户 %d 请求失败, 返还预扣费（token_quota=%s, funding=%s）",
+		s.relayInfo.UserId,
+		logger.FormatQuota(s.tokenConsumed),
+		s.funding.Source(),
+	))
+
+	tokenId := s.relayInfo.TokenId
+	tokenKey := s.relayInfo.TokenKey
+	isPlayground := s.relayInfo.IsPlayground
+	tokenConsumed := s.tokenConsumed
+	funding := s.funding
+
+	gopool.Go(func() {
+		if err := funding.Refund(); err != nil {
+			common.SysLog("error refunding billing source: " + err.Error())
+		}
+		if tokenConsumed > 0 && !isPlayground {
+			if err := model.IncreaseTokenQuota(tokenId, tokenKey, tokenConsumed); err != nil {
+				common.SysLog("error refunding token quota: " + err.Error())
+			}
+		}
+	})
+}
+
+func (s *BillingSession) NeedsRefund() bool {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	return s.needsRefundLocked()
+}
+
+func (s *BillingSession) needsRefundLocked() bool {
+	if s.settled || s.refunded || s.fundingSettled {
+		return false
+	}
+	if s.tokenConsumed > 0 {
+		return true
+	}
+	if sub, ok := s.funding.(*SubscriptionFunding); ok && sub.preConsumed > 0 {
+		return true
+	}
+	return false
+}
+
+func (s *BillingSession) GetPreConsumedQuota() int {
+	return s.preConsumedQuota
+}
+
+// ---------------------------------------------------------------------------
+// PreConsume — 统一预扣费入口（含信任额度旁路）
+// ---------------------------------------------------------------------------
+
+func (s *BillingSession) preConsume(c *gin.Context, quota int) *types.NewAPIError {
+	effectiveQuota := quota
+
+	if s.shouldTrust(c) {
+		effectiveQuota = 0
+		logger.LogInfo(c, fmt.Sprintf("用户 %d 额度充足, 信任且不需要预扣费 (funding=%s)", s.relayInfo.UserId, s.funding.Source()))
+	} else if effectiveQuota > 0 {
+		logger.LogInfo(c, fmt.Sprintf("用户 %d 需要预扣费 %s (funding=%s)", s.relayInfo.UserId, logger.FormatQuota(effectiveQuota), s.funding.Source()))
+	}
+
+	if effectiveQuota > 0 {
+		if err := PreConsumeTokenQuota(s.relayInfo, effectiveQuota); err != nil {
+			return types.NewErrorWithStatusCode(err, types.ErrorCodePreConsumeTokenQuotaFailed, http.StatusForbidden, types.ErrOptionWithSkipRetry(), types.ErrOptionWithNoRecordErrorLog())
+		}
+		s.tokenConsumed = effectiveQuota
+	}
+
+	if err := s.funding.PreConsume(effectiveQuota); err != nil {
+		if s.tokenConsumed > 0 && !s.relayInfo.IsPlayground {
+			if rollbackErr := model.IncreaseTokenQuota(s.relayInfo.TokenId, s.relayInfo.TokenKey, s.tokenConsumed); rollbackErr != nil {
+				common.SysLog(fmt.Sprintf("error rolling back token quota (userId=%d, tokenId=%d, amount=%d, fundingErr=%s): %s",
+					s.relayInfo.UserId, s.relayInfo.TokenId, s.tokenConsumed, err.Error(), rollbackErr.Error()))
+			}
+			s.tokenConsumed = 0
+		}
+		errMsg := err.Error()
+		if strings.Contains(errMsg, "no active subscription") || strings.Contains(errMsg, "subscription quota insufficient") {
+			return types.NewErrorWithStatusCode(fmt.Errorf("订阅额度不足或未配置订阅: %s", errMsg), types.ErrorCodeInsufficientUserQuota, http.StatusForbidden, types.ErrOptionWithSkipRetry(), types.ErrOptionWithNoRecordErrorLog())
+		}
+		return types.NewError(err, types.ErrorCodeUpdateDataError, types.ErrOptionWithSkipRetry())
+	}
+
+	s.preConsumedQuota = effectiveQuota
+	s.syncRelayInfo()
+	return nil
+}
+
+func (s *BillingSession) shouldTrust(c *gin.Context) bool {
+	if s.relayInfo.ForcePreConsume {
+		return false
+	}
+
+	trustQuota := common.GetTrustQuota()
+	if trustQuota <= 0 {
+		return false
+	}
+
+	tokenTrusted := s.relayInfo.TokenUnlimited
+	if !tokenTrusted {
+		tokenQuota := c.GetInt("token_quota")
+		tokenTrusted = tokenQuota > trustQuota
+	}
+	if !tokenTrusted {
+		return false
+	}
+
+	switch s.funding.Source() {
+	case BillingSourceWallet:
+		return s.relayInfo.UserQuota > trustQuota
+	case BillingSourceSubscription:
+		return false
+	default:
+		return false
+	}
+}
+
+func (s *BillingSession) syncRelayInfo() {
+	info := s.relayInfo
+	info.FinalPreConsumedQuota = s.preConsumedQuota
+	info.BillingSource = s.funding.Source()
+
+	if sub, ok := s.funding.(*SubscriptionFunding); ok {
+		info.SubscriptionId = sub.subscriptionId
+		info.SubscriptionPreConsumed = sub.preConsumed
+		info.SubscriptionPostDelta = 0
+		info.SubscriptionAmountTotal = sub.AmountTotal
+		info.SubscriptionAmountUsedAfterPreConsume = sub.AmountUsedAfter
+		info.SubscriptionPlanId = sub.PlanId
+		info.SubscriptionPlanTitle = sub.PlanTitle
+	} else {
+		info.SubscriptionId = 0
+		info.SubscriptionPreConsumed = 0
+	}
+}
+
+func NewBillingSession(c *gin.Context, relayInfo *relaycommon.RelayInfo, preConsumedQuota int) (*BillingSession, *types.NewAPIError) {
+	if relayInfo == nil {
+		return nil, types.NewError(fmt.Errorf("relayInfo is nil"), types.ErrorCodeInvalidRequest, types.ErrOptionWithSkipRetry())
+	}
+
+	pref := common.NormalizeBillingPreference(relayInfo.UserSetting.BillingPreference)
+	if relayInfo.TokenKey != "" {
+		if token, tokenErr := model.GetTokenByKey(relayInfo.TokenKey, false); tokenErr == nil && token != nil {
+			tokenConsumePref := model.NormalizeTokenConsumeGroup(token.ConsumeGroup)
+			if tokenConsumePref != "" {
+				pref = tokenConsumePref
+			}
+		}
+	}
+
+	isSubscriptionGroup, err := model.IsSubscriptionDedicatedGroup(relayInfo.UsingGroup)
+	if err != nil {
+		return nil, types.NewError(err, types.ErrorCodeQueryDataError, types.ErrOptionWithSkipRetry())
+	}
+
+	tryWallet := func() (*BillingSession, *types.NewAPIError) {
+		userQuota, err := model.GetUserQuota(relayInfo.UserId, false)
+		if err != nil {
+			return nil, types.NewError(err, types.ErrorCodeQueryDataError, types.ErrOptionWithSkipRetry())
+		}
+		if userQuota <= 0 {
+			return nil, types.NewErrorWithStatusCode(
+				fmt.Errorf("用户额度不足, 剩余额度: %s", logger.FormatQuota(userQuota)),
+				types.ErrorCodeInsufficientUserQuota, http.StatusForbidden,
+				types.ErrOptionWithSkipRetry(), types.ErrOptionWithNoRecordErrorLog())
+		}
+		if userQuota-preConsumedQuota < 0 {
+			return nil, types.NewErrorWithStatusCode(
+				fmt.Errorf("预扣费额度失败, 用户剩余额度: %s, 需要预扣费额度: %s", logger.FormatQuota(userQuota), logger.FormatQuota(preConsumedQuota)),
+				types.ErrorCodeInsufficientUserQuota, http.StatusForbidden,
+				types.ErrOptionWithSkipRetry(), types.ErrOptionWithNoRecordErrorLog())
+		}
+		relayInfo.UserQuota = userQuota
+		session := &BillingSession{relayInfo: relayInfo, funding: &WalletFunding{userId: relayInfo.UserId}}
+		if apiErr := session.preConsume(c, preConsumedQuota); apiErr != nil {
+			return nil, apiErr
+		}
+		return session, nil
+	}
+
+	trySubscription := func() (*BillingSession, *types.NewAPIError) {
+		subConsume := int64(preConsumedQuota)
+		if subConsume <= 0 {
+			subConsume = 1
+		}
+		session := &BillingSession{
+			relayInfo: relayInfo,
+			funding: &SubscriptionFunding{
+				requestId: relayInfo.RequestId,
+				userId:    relayInfo.UserId,
+				modelName: relayInfo.OriginModelName,
+				amount:    subConsume,
+				group:     relayInfo.UsingGroup,
+			},
+		}
+		if apiErr := session.preConsume(c, int(subConsume)); apiErr != nil {
+			return nil, apiErr
+		}
+		return session, nil
+	}
+
+	if isSubscriptionGroup {
+		switch pref {
+		case model.TokenConsumeGroupWalletOnly:
+			return nil, types.NewErrorWithStatusCode(fmt.Errorf("当前请求分组是订阅专属分组，仅允许订阅扣费"), types.ErrorCodeInsufficientUserQuota, http.StatusForbidden, types.ErrOptionWithSkipRetry(), types.ErrOptionWithNoRecordErrorLog())
+		default:
+			return trySubscription()
+		}
+	}
+
+	subAllowed, subAllowedErr := model.IsGroupAllowedByAnyActiveSubscription(relayInfo.UserId, relayInfo.UsingGroup)
+	if subAllowedErr != nil {
+		return nil, types.NewError(subAllowedErr, types.ErrorCodeQueryDataError, types.ErrOptionWithSkipRetry())
+	}
+
+	switch pref {
+	case model.TokenConsumeGroupSubscriptionOnly:
+		if !subAllowed {
+			return nil, types.NewErrorWithStatusCode(fmt.Errorf("当前请求分组不在订阅可消费分组内，不能使用订阅扣费"), types.ErrorCodeInsufficientUserQuota, http.StatusForbidden, types.ErrOptionWithSkipRetry(), types.ErrOptionWithNoRecordErrorLog())
+		}
+		return trySubscription()
+	case model.TokenConsumeGroupWalletOnly:
+		return tryWallet()
+	default:
+		return tryWallet()
+	}
+}

service/channel.go

@@ -0,0 +1,115 @@
+package service
+
+import (
+	"fmt"
+	"net/http"
+	"strings"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/constant"
+	"github.com/QuantumNous/new-api/dto"
+	"github.com/QuantumNous/new-api/model"
+	"github.com/QuantumNous/new-api/setting/operation_setting"
+	"github.com/QuantumNous/new-api/types"
+)
+
+func formatNotifyType(channelId int, status int) string {
+	return fmt.Sprintf("%s_%d_%d", dto.NotifyTypeChannelUpdate, channelId, status)
+}
+
+// disable & notify
+func DisableChannel(channelError types.ChannelError, reason string) {
+	common.SysLog(fmt.Sprintf("通道「%s」（#%d）发生错误，准备禁用，原因：%s", channelError.ChannelName, channelError.ChannelId, reason))
+
+	// 检查是否启用自动禁用功能
+	if !channelError.AutoBan {
+		common.SysLog(fmt.Sprintf("通道「%s」（#%d）未启用自动禁用功能，跳过禁用操作", channelError.ChannelName, channelError.ChannelId))
+		return
+	}
+
+	success := model.UpdateChannelStatus(channelError.ChannelId, channelError.UsingKey, common.ChannelStatusAutoDisabled, reason)
+	if success {
+		subject := fmt.Sprintf("通道「%s」（#%d）已被禁用", channelError.ChannelName, channelError.ChannelId)
+		content := fmt.Sprintf("通道「%s」（#%d）已被禁用，原因：%s", channelError.ChannelName, channelError.ChannelId, reason)
+		NotifyRootUser(formatNotifyType(channelError.ChannelId, common.ChannelStatusAutoDisabled), subject, content)
+	}
+}
+
+func EnableChannel(channelId int, usingKey string, channelName string) {
+	success := model.UpdateChannelStatus(channelId, usingKey, common.ChannelStatusEnabled, "")
+	if success {
+		subject := fmt.Sprintf("通道「%s」（#%d）已被启用", channelName, channelId)
+		content := fmt.Sprintf("通道「%s」（#%d）已被启用", channelName, channelId)
+		NotifyRootUser(formatNotifyType(channelId, common.ChannelStatusEnabled), subject, content)
+	}
+}
+
+func ShouldDisableChannel(channelType int, err *types.NewAPIError) bool {
+	if !common.AutomaticDisableChannelEnabled {
+		return false
+	}
+	if err == nil {
+		return false
+	}
+	if types.IsChannelError(err) {
+		return true
+	}
+	if types.IsSkipRetryError(err) {
+		return false
+	}
+	if operation_setting.ShouldDisableByStatusCode(err.StatusCode) {
+		return true
+	}
+	//if err.StatusCode == http.StatusUnauthorized {
+	//	return true
+	//}
+	if err.StatusCode == http.StatusForbidden {
+		switch channelType {
+		case constant.ChannelTypeGemini:
+			return true
+		}
+	}
+	oaiErr := err.ToOpenAIError()
+	switch oaiErr.Code {
+	case "invalid_api_key":
+		return true
+	case "account_deactivated":
+		return true
+	case "billing_not_active":
+		return true
+	case "pre_consume_token_quota_failed":
+		return true
+	case "Arrearage":
+		return true
+	}
+	switch oaiErr.Type {
+	case "insufficient_quota":
+		return true
+	case "insufficient_user_quota":
+		return true
+	// https://docs.anthropic.com/claude/reference/errors
+	case "authentication_error":
+		return true
+	case "permission_error":
+		return true
+	case "forbidden":
+		return true
+	}
+
+	lowerMessage := strings.ToLower(err.Error())
+	search, _ := AcSearch(lowerMessage, operation_setting.AutomaticDisableKeywords, true)
+	return search
+}
+
+func ShouldEnableChannel(newAPIError *types.NewAPIError, status int) bool {
+	if !common.AutomaticEnableChannelEnabled {
+		return false
+	}
+	if newAPIError != nil {
+		return false
+	}
+	if status != common.ChannelStatusAutoDisabled {
+		return false
+	}
+	return true
+}

service/channel_affinity.go

@@ -0,0 +1,953 @@
+package service
+
+import (
+	"fmt"
+	"hash/fnv"
+	"regexp"
+	"strconv"
+	"strings"
+	"sync"
+	"time"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/dto"
+	"github.com/QuantumNous/new-api/pkg/cachex"
+	"github.com/QuantumNous/new-api/setting/operation_setting"
+	"github.com/QuantumNous/new-api/types"
+	"github.com/gin-gonic/gin"
+	"github.com/samber/hot"
+	"github.com/tidwall/gjson"
+)
+
+const (
+	ginKeyChannelAffinityCacheKey   = "channel_affinity_cache_key"
+	ginKeyChannelAffinityTTLSeconds = "channel_affinity_ttl_seconds"
+	ginKeyChannelAffinityMeta       = "channel_affinity_meta"
+	ginKeyChannelAffinityLogInfo    = "channel_affinity_log_info"
+	ginKeyChannelAffinitySkipRetry  = "channel_affinity_skip_retry_on_failure"
+
+	channelAffinityCacheNamespace           = "new-api:channel_affinity:v1"
+	channelAffinityUsageCacheStatsNamespace = "new-api:channel_affinity_usage_cache_stats:v1"
+)
+
+var (
+	channelAffinityCacheOnce sync.Once
+	channelAffinityCache     *cachex.HybridCache[int]
+
+	channelAffinityUsageCacheStatsOnce  sync.Once
+	channelAffinityUsageCacheStatsCache *cachex.HybridCache[ChannelAffinityUsageCacheCounters]
+
+	channelAffinityRegexCache sync.Map // map[string]*regexp.Regexp
+)
+
+type channelAffinityMeta struct {
+	CacheKey       string
+	TTLSeconds     int
+	RuleName       string
+	SkipRetry      bool
+	ParamTemplate  map[string]interface{}
+	KeySourceType  string
+	KeySourceKey   string
+	KeySourcePath  string
+	KeyHint        string
+	KeyFingerprint string
+	UsingGroup     string
+	ModelName      string
+	RequestPath    string
+}
+
+type ChannelAffinityStatsContext struct {
+	RuleName       string
+	UsingGroup     string
+	KeyFingerprint string
+	TTLSeconds     int64
+}
+
+const (
+	cacheTokenRateModeCachedOverPrompt           = "cached_over_prompt"
+	cacheTokenRateModeCachedOverPromptPlusCached = "cached_over_prompt_plus_cached"
+	cacheTokenRateModeMixed                      = "mixed"
+)
+
+type ChannelAffinityCacheStats struct {
+	Enabled       bool           `json:"enabled"`
+	Total         int            `json:"total"`
+	Unknown       int            `json:"unknown"`
+	ByRuleName    map[string]int `json:"by_rule_name"`
+	CacheCapacity int            `json:"cache_capacity"`
+	CacheAlgo     string         `json:"cache_algo"`
+}
+
+func getChannelAffinityCache() *cachex.HybridCache[int] {
+	channelAffinityCacheOnce.Do(func() {
+		setting := operation_setting.GetChannelAffinitySetting()
+		capacity := setting.MaxEntries
+		if capacity <= 0 {
+			capacity = 100_000
+		}
+		defaultTTLSeconds := setting.DefaultTTLSeconds
+		if defaultTTLSeconds <= 0 {
+			defaultTTLSeconds = 3600
+		}
+
+		channelAffinityCache = cachex.NewHybridCache[int](cachex.HybridCacheConfig[int]{
+			Namespace: cachex.Namespace(channelAffinityCacheNamespace),
+			Redis:     common.RDB,
+			RedisEnabled: func() bool {
+				return common.RedisEnabled && common.RDB != nil
+			},
+			RedisCodec: cachex.IntCodec{},
+			Memory: func() *hot.HotCache[string, int] {
+				return hot.NewHotCache[string, int](hot.LRU, capacity).
+					WithTTL(time.Duration(defaultTTLSeconds) * time.Second).
+					WithJanitor().
+					Build()
+			},
+		})
+	})
+	return channelAffinityCache
+}
+
+func GetChannelAffinityCacheStats() ChannelAffinityCacheStats {
+	setting := operation_setting.GetChannelAffinitySetting()
+	if setting == nil {
+		return ChannelAffinityCacheStats{
+			Enabled:    false,
+			Total:      0,
+			Unknown:    0,
+			ByRuleName: map[string]int{},
+		}
+	}
+
+	cache := getChannelAffinityCache()
+	mainCap, _ := cache.Capacity()
+	mainAlgo, _ := cache.Algorithm()
+
+	rules := setting.Rules
+	ruleByName := make(map[string]operation_setting.ChannelAffinityRule, len(rules))
+	for _, r := range rules {
+		name := strings.TrimSpace(r.Name)
+		if name == "" {
+			continue
+		}
+		if !r.IncludeRuleName {
+			continue
+		}
+		ruleByName[name] = r
+	}
+
+	byRuleName := make(map[string]int, len(ruleByName))
+	for name := range ruleByName {
+		byRuleName[name] = 0
+	}
+
+	keys, err := cache.Keys()
+	if err != nil {
+		common.SysError(fmt.Sprintf("channel affinity cache list keys failed: err=%v", err))
+		keys = nil
+	}
+	total := len(keys)
+	unknown := 0
+	for _, k := range keys {
+		prefix := channelAffinityCacheNamespace + ":"
+		if !strings.HasPrefix(k, prefix) {
+			unknown++
+			continue
+		}
+		rest := strings.TrimPrefix(k, prefix)
+		parts := strings.Split(rest, ":")
+		if len(parts) < 2 {
+			unknown++
+			continue
+		}
+		ruleName := parts[0]
+		rule, ok := ruleByName[ruleName]
+		if !ok {
+			unknown++
+			continue
+		}
+		if rule.IncludeUsingGroup {
+			if len(parts) < 3 {
+				unknown++
+				continue
+			}
+		}
+		byRuleName[ruleName]++
+	}
+
+	return ChannelAffinityCacheStats{
+		Enabled:       setting.Enabled,
+		Total:         total,
+		Unknown:       unknown,
+		ByRuleName:    byRuleName,
+		CacheCapacity: mainCap,
+		CacheAlgo:     mainAlgo,
+	}
+}
+
+func ClearChannelAffinityCacheAll() int {
+	cache := getChannelAffinityCache()
+	keys, err := cache.Keys()
+	if err != nil {
+		common.SysError(fmt.Sprintf("channel affinity cache list keys failed: err=%v", err))
+		keys = nil
+	}
+	if len(keys) > 0 {
+		if _, err := cache.DeleteMany(keys); err != nil {
+			common.SysError(fmt.Sprintf("channel affinity cache delete many failed: err=%v", err))
+		}
+	}
+	return len(keys)
+}
+
+func ClearChannelAffinityCacheByRuleName(ruleName string) (int, error) {
+	ruleName = strings.TrimSpace(ruleName)
+	if ruleName == "" {
+		return 0, fmt.Errorf("rule_name 不能为空")
+	}
+
+	setting := operation_setting.GetChannelAffinitySetting()
+	if setting == nil {
+		return 0, fmt.Errorf("channel_affinity_setting 未初始化")
+	}
+
+	var matchedRule *operation_setting.ChannelAffinityRule
+	for i := range setting.Rules {
+		r := &setting.Rules[i]
+		if strings.TrimSpace(r.Name) != ruleName {
+			continue
+		}
+		matchedRule = r
+		break
+	}
+	if matchedRule == nil {
+		return 0, fmt.Errorf("未知规则名称")
+	}
+	if !matchedRule.IncludeRuleName {
+		return 0, fmt.Errorf("该规则未启用 include_rule_name，无法按规则清空缓存")
+	}
+
+	cache := getChannelAffinityCache()
+	deleted, err := cache.DeleteByPrefix(ruleName)
+	if err != nil {
+		return 0, err
+	}
+	return deleted, nil
+}
+
+func matchAnyRegexCached(patterns []string, s string) bool {
+	if len(patterns) == 0 || s == "" {
+		return false
+	}
+	for _, pattern := range patterns {
+		if pattern == "" {
+			continue
+		}
+		re, ok := channelAffinityRegexCache.Load(pattern)
+		if !ok {
+			compiled, err := regexp.Compile(pattern)
+			if err != nil {
+				continue
+			}
+			re = compiled
+			channelAffinityRegexCache.Store(pattern, re)
+		}
+		if re.(*regexp.Regexp).MatchString(s) {
+			return true
+		}
+	}
+	return false
+}
+
+func matchAnyIncludeFold(patterns []string, s string) bool {
+	if len(patterns) == 0 || s == "" {
+		return false
+	}
+	sLower := strings.ToLower(s)
+	for _, p := range patterns {
+		p = strings.TrimSpace(p)
+		if p == "" {
+			continue
+		}
+		if strings.Contains(sLower, strings.ToLower(p)) {
+			return true
+		}
+	}
+	return false
+}
+
+func extractChannelAffinityValue(c *gin.Context, src operation_setting.ChannelAffinityKeySource) string {
+	switch src.Type {
+	case "context_int":
+		if src.Key == "" {
+			return ""
+		}
+		v := c.GetInt(src.Key)
+		if v <= 0 {
+			return ""
+		}
+		return strconv.Itoa(v)
+	case "context_string":
+		if src.Key == "" {
+			return ""
+		}
+		return strings.TrimSpace(c.GetString(src.Key))
+	case "gjson":
+		if src.Path == "" {
+			return ""
+		}
+		storage, err := common.GetBodyStorage(c)
+		if err != nil {
+			return ""
+		}
+		body, err := storage.Bytes()
+		if err != nil || len(body) == 0 {
+			return ""
+		}
+		res := gjson.GetBytes(body, src.Path)
+		if !res.Exists() {
+			return ""
+		}
+		switch res.Type {
+		case gjson.String, gjson.Number, gjson.True, gjson.False:
+			return strings.TrimSpace(res.String())
+		default:
+			return strings.TrimSpace(res.Raw)
+		}
+	default:
+		return ""
+	}
+}
+
+func buildChannelAffinityCacheKeySuffix(rule operation_setting.ChannelAffinityRule, usingGroup string, affinityValue string) string {
+	parts := make([]string, 0, 3)
+	if rule.IncludeRuleName && rule.Name != "" {
+		parts = append(parts, rule.Name)
+	}
+	if rule.IncludeUsingGroup && usingGroup != "" {
+		parts = append(parts, usingGroup)
+	}
+	parts = append(parts, affinityValue)
+	return strings.Join(parts, ":")
+}
+
+func setChannelAffinityContext(c *gin.Context, meta channelAffinityMeta) {
+	c.Set(ginKeyChannelAffinityCacheKey, meta.CacheKey)
+	c.Set(ginKeyChannelAffinityTTLSeconds, meta.TTLSeconds)
+	c.Set(ginKeyChannelAffinityMeta, meta)
+}
+
+func getChannelAffinityContext(c *gin.Context) (string, int, bool) {
+	keyAny, ok := c.Get(ginKeyChannelAffinityCacheKey)
+	if !ok {
+		return "", 0, false
+	}
+	key, ok := keyAny.(string)
+	if !ok || key == "" {
+		return "", 0, false
+	}
+	ttlAny, ok := c.Get(ginKeyChannelAffinityTTLSeconds)
+	if !ok {
+		return key, 0, true
+	}
+	ttlSeconds, _ := ttlAny.(int)
+	return key, ttlSeconds, true
+}
+
+func getChannelAffinityMeta(c *gin.Context) (channelAffinityMeta, bool) {
+	anyMeta, ok := c.Get(ginKeyChannelAffinityMeta)
+	if !ok {
+		return channelAffinityMeta{}, false
+	}
+	meta, ok := anyMeta.(channelAffinityMeta)
+	if !ok {
+		return channelAffinityMeta{}, false
+	}
+	return meta, true
+}
+
+func GetChannelAffinityStatsContext(c *gin.Context) (ChannelAffinityStatsContext, bool) {
+	if c == nil {
+		return ChannelAffinityStatsContext{}, false
+	}
+	meta, ok := getChannelAffinityMeta(c)
+	if !ok {
+		return ChannelAffinityStatsContext{}, false
+	}
+	ruleName := strings.TrimSpace(meta.RuleName)
+	keyFp := strings.TrimSpace(meta.KeyFingerprint)
+	usingGroup := strings.TrimSpace(meta.UsingGroup)
+	if ruleName == "" || keyFp == "" {
+		return ChannelAffinityStatsContext{}, false
+	}
+	ttlSeconds := int64(meta.TTLSeconds)
+	if ttlSeconds <= 0 {
+		return ChannelAffinityStatsContext{}, false
+	}
+	return ChannelAffinityStatsContext{
+		RuleName:       ruleName,
+		UsingGroup:     usingGroup,
+		KeyFingerprint: keyFp,
+		TTLSeconds:     ttlSeconds,
+	}, true
+}
+
+func affinityFingerprint(s string) string {
+	if s == "" {
+		return ""
+	}
+	hex := common.Sha1([]byte(s))
+	if len(hex) >= 8 {
+		return hex[:8]
+	}
+	return hex
+}
+
+func buildChannelAffinityKeyHint(s string) string {
+	s = strings.TrimSpace(s)
+	if s == "" {
+		return ""
+	}
+	s = strings.ReplaceAll(s, "\n", " ")
+	s = strings.ReplaceAll(s, "\r", " ")
+	if len(s) <= 12 {
+		return s
+	}
+	return s[:4] + "..." + s[len(s)-4:]
+}
+
+func cloneStringAnyMap(src map[string]interface{}) map[string]interface{} {
+	if len(src) == 0 {
+		return map[string]interface{}{}
+	}
+	dst := make(map[string]interface{}, len(src))
+	for k, v := range src {
+		dst[k] = v
+	}
+	return dst
+}
+
+func mergeChannelOverride(base map[string]interface{}, tpl map[string]interface{}) map[string]interface{} {
+	if len(base) == 0 && len(tpl) == 0 {
+		return map[string]interface{}{}
+	}
+	if len(tpl) == 0 {
+		return base
+	}
+	out := cloneStringAnyMap(base)
+	for k, v := range tpl {
+		if strings.EqualFold(strings.TrimSpace(k), "operations") {
+			baseOps, hasBaseOps := extractParamOperations(out[k])
+			tplOps, hasTplOps := extractParamOperations(v)
+			if hasTplOps {
+				if hasBaseOps {
+					out[k] = append(tplOps, baseOps...)
+				} else {
+					out[k] = tplOps
+				}
+				continue
+			}
+		}
+		if _, exists := out[k]; exists {
+			continue
+		}
+		out[k] = v
+	}
+	return out
+}
+
+func extractParamOperations(value interface{}) ([]interface{}, bool) {
+	switch ops := value.(type) {
+	case []interface{}:
+		if len(ops) == 0 {
+			return []interface{}{}, true
+		}
+		cloned := make([]interface{}, 0, len(ops))
+		cloned = append(cloned, ops...)
+		return cloned, true
+	case []map[string]interface{}:
+		cloned := make([]interface{}, 0, len(ops))
+		for _, op := range ops {
+			cloned = append(cloned, op)
+		}
+		return cloned, true
+	default:
+		return nil, false
+	}
+}
+
+func appendChannelAffinityTemplateAdminInfo(c *gin.Context, meta channelAffinityMeta) {
+	if c == nil {
+		return
+	}
+	if len(meta.ParamTemplate) == 0 {
+		return
+	}
+
+	templateInfo := map[string]interface{}{
+		"applied":             true,
+		"rule_name":           meta.RuleName,
+		"param_override_keys": len(meta.ParamTemplate),
+	}
+	if anyInfo, ok := c.Get(ginKeyChannelAffinityLogInfo); ok {
+		if info, ok := anyInfo.(map[string]interface{}); ok {
+			info["override_template"] = templateInfo
+			c.Set(ginKeyChannelAffinityLogInfo, info)
+			return
+		}
+	}
+	c.Set(ginKeyChannelAffinityLogInfo, map[string]interface{}{
+		"reason":            meta.RuleName,
+		"rule_name":         meta.RuleName,
+		"using_group":       meta.UsingGroup,
+		"model":             meta.ModelName,
+		"request_path":      meta.RequestPath,
+		"key_source":        meta.KeySourceType,
+		"key_key":           meta.KeySourceKey,
+		"key_path":          meta.KeySourcePath,
+		"key_hint":          meta.KeyHint,
+		"key_fp":            meta.KeyFingerprint,
+		"override_template": templateInfo,
+	})
+}
+
+// ApplyChannelAffinityOverrideTemplate merges per-rule channel override templates onto the selected channel override config.
+func ApplyChannelAffinityOverrideTemplate(c *gin.Context, paramOverride map[string]interface{}) (map[string]interface{}, bool) {
+	if c == nil {
+		return paramOverride, false
+	}
+	meta, ok := getChannelAffinityMeta(c)
+	if !ok {
+		return paramOverride, false
+	}
+	if len(meta.ParamTemplate) == 0 {
+		return paramOverride, false
+	}
+
+	mergedParam := mergeChannelOverride(paramOverride, meta.ParamTemplate)
+	appendChannelAffinityTemplateAdminInfo(c, meta)
+	return mergedParam, true
+}
+
+func GetPreferredChannelByAffinity(c *gin.Context, modelName string, usingGroup string) (int, bool) {
+	setting := operation_setting.GetChannelAffinitySetting()
+	if setting == nil || !setting.Enabled {
+		return 0, false
+	}
+	path := ""
+	if c != nil && c.Request != nil && c.Request.URL != nil {
+		path = c.Request.URL.Path
+	}
+	userAgent := ""
+	if c != nil && c.Request != nil {
+		userAgent = c.Request.UserAgent()
+	}
+
+	for _, rule := range setting.Rules {
+		if !matchAnyRegexCached(rule.ModelRegex, modelName) {
+			continue
+		}
+		if len(rule.PathRegex) > 0 && !matchAnyRegexCached(rule.PathRegex, path) {
+			continue
+		}
+		if len(rule.UserAgentInclude) > 0 && !matchAnyIncludeFold(rule.UserAgentInclude, userAgent) {
+			continue
+		}
+		var affinityValue string
+		var usedSource operation_setting.ChannelAffinityKeySource
+		for _, src := range rule.KeySources {
+			affinityValue = extractChannelAffinityValue(c, src)
+			if affinityValue != "" {
+				usedSource = src
+				break
+			}
+		}
+		if affinityValue == "" {
+			continue
+		}
+		if rule.ValueRegex != "" && !matchAnyRegexCached([]string{rule.ValueRegex}, affinityValue) {
+			continue
+		}
+
+		ttlSeconds := rule.TTLSeconds
+		if ttlSeconds <= 0 {
+			ttlSeconds = setting.DefaultTTLSeconds
+		}
+		cacheKeySuffix := buildChannelAffinityCacheKeySuffix(rule, usingGroup, affinityValue)
+		cacheKeyFull := channelAffinityCacheNamespace + ":" + cacheKeySuffix
+		setChannelAffinityContext(c, channelAffinityMeta{
+			CacheKey:       cacheKeyFull,
+			TTLSeconds:     ttlSeconds,
+			RuleName:       rule.Name,
+			SkipRetry:      rule.SkipRetryOnFailure,
+			ParamTemplate:  cloneStringAnyMap(rule.ParamOverrideTemplate),
+			KeySourceType:  strings.TrimSpace(usedSource.Type),
+			KeySourceKey:   strings.TrimSpace(usedSource.Key),
+			KeySourcePath:  strings.TrimSpace(usedSource.Path),
+			KeyHint:        buildChannelAffinityKeyHint(affinityValue),
+			KeyFingerprint: affinityFingerprint(affinityValue),
+			UsingGroup:     usingGroup,
+			ModelName:      modelName,
+			RequestPath:    path,
+		})
+
+		cache := getChannelAffinityCache()
+		channelID, found, err := cache.Get(cacheKeySuffix)
+		if err != nil {
+			common.SysError(fmt.Sprintf("channel affinity cache get failed: key=%s, err=%v", cacheKeyFull, err))
+			return 0, false
+		}
+		if found {
+			return channelID, true
+		}
+		return 0, false
+	}
+	return 0, false
+}
+
+func ShouldSkipRetryAfterChannelAffinityFailure(c *gin.Context) bool {
+	if c == nil {
+		return false
+	}
+	v, ok := c.Get(ginKeyChannelAffinitySkipRetry)
+	if ok {
+		b, ok := v.(bool)
+		if ok {
+			return b
+		}
+	}
+	meta, ok := getChannelAffinityMeta(c)
+	if !ok {
+		return false
+	}
+	return meta.SkipRetry
+}
+
+func MarkChannelAffinityUsed(c *gin.Context, selectedGroup string, channelID int) {
+	if c == nil || channelID <= 0 {
+		return
+	}
+	meta, ok := getChannelAffinityMeta(c)
+	if !ok {
+		return
+	}
+	c.Set(ginKeyChannelAffinitySkipRetry, meta.SkipRetry)
+	info := map[string]interface{}{
+		"reason":         meta.RuleName,
+		"rule_name":      meta.RuleName,
+		"using_group":    meta.UsingGroup,
+		"selected_group": selectedGroup,
+		"model":          meta.ModelName,
+		"request_path":   meta.RequestPath,
+		"channel_id":     channelID,
+		"key_source":     meta.KeySourceType,
+		"key_key":        meta.KeySourceKey,
+		"key_path":       meta.KeySourcePath,
+		"key_hint":       meta.KeyHint,
+		"key_fp":         meta.KeyFingerprint,
+	}
+	c.Set(ginKeyChannelAffinityLogInfo, info)
+}
+
+func AppendChannelAffinityAdminInfo(c *gin.Context, adminInfo map[string]interface{}) {
+	if c == nil || adminInfo == nil {
+		return
+	}
+	anyInfo, ok := c.Get(ginKeyChannelAffinityLogInfo)
+	if !ok || anyInfo == nil {
+		return
+	}
+	adminInfo["channel_affinity"] = anyInfo
+}
+
+func RecordChannelAffinity(c *gin.Context, channelID int) {
+	if channelID <= 0 {
+		return
+	}
+	setting := operation_setting.GetChannelAffinitySetting()
+	if setting == nil || !setting.Enabled {
+		return
+	}
+	if setting.SwitchOnSuccess && c != nil {
+		if successChannelID := c.GetInt("channel_id"); successChannelID > 0 {
+			channelID = successChannelID
+		}
+	}
+	cacheKey, ttlSeconds, ok := getChannelAffinityContext(c)
+	if !ok {
+		return
+	}
+	if ttlSeconds <= 0 {
+		ttlSeconds = setting.DefaultTTLSeconds
+	}
+	if ttlSeconds <= 0 {
+		ttlSeconds = 3600
+	}
+	cache := getChannelAffinityCache()
+	if err := cache.SetWithTTL(cacheKey, channelID, time.Duration(ttlSeconds)*time.Second); err != nil {
+		common.SysError(fmt.Sprintf("channel affinity cache set failed: key=%s, err=%v", cacheKey, err))
+	}
+}
+
+type ChannelAffinityUsageCacheStats struct {
+	RuleName            string `json:"rule_name"`
+	UsingGroup          string `json:"using_group"`
+	KeyFingerprint      string `json:"key_fp"`
+	CachedTokenRateMode string `json:"cached_token_rate_mode"`
+
+	Hit           int64 `json:"hit"`
+	Total         int64 `json:"total"`
+	WindowSeconds int64 `json:"window_seconds"`
+
+	PromptTokens         int64 `json:"prompt_tokens"`
+	CompletionTokens     int64 `json:"completion_tokens"`
+	TotalTokens          int64 `json:"total_tokens"`
+	CachedTokens         int64 `json:"cached_tokens"`
+	PromptCacheHitTokens int64 `json:"prompt_cache_hit_tokens"`
+	LastSeenAt           int64 `json:"last_seen_at"`
+}
+
+type ChannelAffinityUsageCacheCounters struct {
+	CachedTokenRateMode string `json:"cached_token_rate_mode"`
+
+	Hit           int64 `json:"hit"`
+	Total         int64 `json:"total"`
+	WindowSeconds int64 `json:"window_seconds"`
+
+	PromptTokens         int64 `json:"prompt_tokens"`
+	CompletionTokens     int64 `json:"completion_tokens"`
+	TotalTokens          int64 `json:"total_tokens"`
+	CachedTokens         int64 `json:"cached_tokens"`
+	PromptCacheHitTokens int64 `json:"prompt_cache_hit_tokens"`
+	LastSeenAt           int64 `json:"last_seen_at"`
+}
+
+var channelAffinityUsageCacheStatsLocks [64]sync.Mutex
+
+// ObserveChannelAffinityUsageCacheByRelayFormat records usage cache stats with a stable rate mode derived from relay format.
+func ObserveChannelAffinityUsageCacheByRelayFormat(c *gin.Context, usage *dto.Usage, relayFormat types.RelayFormat) {
+	ObserveChannelAffinityUsageCacheFromContext(c, usage, cachedTokenRateModeByRelayFormat(relayFormat))
+}
+
+func ObserveChannelAffinityUsageCacheFromContext(c *gin.Context, usage *dto.Usage, cachedTokenRateMode string) {
+	statsCtx, ok := GetChannelAffinityStatsContext(c)
+	if !ok {
+		return
+	}
+	observeChannelAffinityUsageCache(statsCtx, usage, cachedTokenRateMode)
+}
+
+func GetChannelAffinityUsageCacheStats(ruleName, usingGroup, keyFp string) ChannelAffinityUsageCacheStats {
+	ruleName = strings.TrimSpace(ruleName)
+	usingGroup = strings.TrimSpace(usingGroup)
+	keyFp = strings.TrimSpace(keyFp)
+
+	entryKey := channelAffinityUsageCacheEntryKey(ruleName, usingGroup, keyFp)
+	if entryKey == "" {
+		return ChannelAffinityUsageCacheStats{
+			RuleName:       ruleName,
+			UsingGroup:     usingGroup,
+			KeyFingerprint: keyFp,
+		}
+	}
+
+	cache := getChannelAffinityUsageCacheStatsCache()
+	v, found, err := cache.Get(entryKey)
+	if err != nil || !found {
+		return ChannelAffinityUsageCacheStats{
+			RuleName:       ruleName,
+			UsingGroup:     usingGroup,
+			KeyFingerprint: keyFp,
+		}
+	}
+	return ChannelAffinityUsageCacheStats{
+		CachedTokenRateMode:  v.CachedTokenRateMode,
+		RuleName:             ruleName,
+		UsingGroup:           usingGroup,
+		KeyFingerprint:       keyFp,
+		Hit:                  v.Hit,
+		Total:                v.Total,
+		WindowSeconds:        v.WindowSeconds,
+		PromptTokens:         v.PromptTokens,
+		CompletionTokens:     v.CompletionTokens,
+		TotalTokens:          v.TotalTokens,
+		CachedTokens:         v.CachedTokens,
+		PromptCacheHitTokens: v.PromptCacheHitTokens,
+		LastSeenAt:           v.LastSeenAt,
+	}
+}
+
+func observeChannelAffinityUsageCache(statsCtx ChannelAffinityStatsContext, usage *dto.Usage, cachedTokenRateMode string) {
+	entryKey := channelAffinityUsageCacheEntryKey(statsCtx.RuleName, statsCtx.UsingGroup, statsCtx.KeyFingerprint)
+	if entryKey == "" {
+		return
+	}
+
+	windowSeconds := statsCtx.TTLSeconds
+	if windowSeconds <= 0 {
+		return
+	}
+
+	cache := getChannelAffinityUsageCacheStatsCache()
+	ttl := time.Duration(windowSeconds) * time.Second
+
+	lock := channelAffinityUsageCacheStatsLock(entryKey)
+	lock.Lock()
+	defer lock.Unlock()
+
+	prev, found, err := cache.Get(entryKey)
+	if err != nil {
+		return
+	}
+	next := prev
+	if !found {
+		next = ChannelAffinityUsageCacheCounters{}
+	}
+	currentMode := normalizeCachedTokenRateMode(cachedTokenRateMode)
+	if currentMode != "" {
+		if next.CachedTokenRateMode == "" {
+			next.CachedTokenRateMode = currentMode
+		} else if next.CachedTokenRateMode != currentMode && next.CachedTokenRateMode != cacheTokenRateModeMixed {
+			next.CachedTokenRateMode = cacheTokenRateModeMixed
+		}
+	}
+	next.Total++
+	hit, cachedTokens, promptCacheHitTokens := usageCacheSignals(usage)
+	if hit {
+		next.Hit++
+	}
+	next.WindowSeconds = windowSeconds
+	next.LastSeenAt = time.Now().Unix()
+	next.CachedTokens += cachedTokens
+	next.PromptCacheHitTokens += promptCacheHitTokens
+	next.PromptTokens += int64(usagePromptTokens(usage))
+	next.CompletionTokens += int64(usageCompletionTokens(usage))
+	next.TotalTokens += int64(usageTotalTokens(usage))
+	_ = cache.SetWithTTL(entryKey, next, ttl)
+}
+
+func normalizeCachedTokenRateMode(mode string) string {
+	switch mode {
+	case cacheTokenRateModeCachedOverPrompt:
+		return cacheTokenRateModeCachedOverPrompt
+	case cacheTokenRateModeCachedOverPromptPlusCached:
+		return cacheTokenRateModeCachedOverPromptPlusCached
+	case cacheTokenRateModeMixed:
+		return cacheTokenRateModeMixed
+	default:
+		return ""
+	}
+}
+
+func cachedTokenRateModeByRelayFormat(relayFormat types.RelayFormat) string {
+	switch relayFormat {
+	case types.RelayFormatOpenAI, types.RelayFormatOpenAIResponses, types.RelayFormatOpenAIResponsesCompaction:
+		return cacheTokenRateModeCachedOverPrompt
+	case types.RelayFormatClaude:
+		return cacheTokenRateModeCachedOverPromptPlusCached
+	default:
+		return ""
+	}
+}
+
+func channelAffinityUsageCacheEntryKey(ruleName, usingGroup, keyFp string) string {
+	ruleName = strings.TrimSpace(ruleName)
+	usingGroup = strings.TrimSpace(usingGroup)
+	keyFp = strings.TrimSpace(keyFp)
+	if ruleName == "" || keyFp == "" {
+		return ""
+	}
+	return ruleName + "\n" + usingGroup + "\n" + keyFp
+}
+
+func usageCacheSignals(usage *dto.Usage) (hit bool, cachedTokens int64, promptCacheHitTokens int64) {
+	if usage == nil {
+		return false, 0, 0
+	}
+
+	cached := int64(0)
+	if usage.PromptTokensDetails.CachedTokens > 0 {
+		cached = int64(usage.PromptTokensDetails.CachedTokens)
+	} else if usage.InputTokensDetails != nil && usage.InputTokensDetails.CachedTokens > 0 {
+		cached = int64(usage.InputTokensDetails.CachedTokens)
+	}
+	pcht := int64(0)
+	if usage.PromptCacheHitTokens > 0 {
+		pcht = int64(usage.PromptCacheHitTokens)
+	}
+	return cached > 0 || pcht > 0, cached, pcht
+}
+
+func usagePromptTokens(usage *dto.Usage) int {
+	if usage == nil {
+		return 0
+	}
+	if usage.PromptTokens > 0 {
+		return usage.PromptTokens
+	}
+	return usage.InputTokens
+}
+
+func usageCompletionTokens(usage *dto.Usage) int {
+	if usage == nil {
+		return 0
+	}
+	if usage.CompletionTokens > 0 {
+		return usage.CompletionTokens
+	}
+	return usage.OutputTokens
+}
+
+func usageTotalTokens(usage *dto.Usage) int {
+	if usage == nil {
+		return 0
+	}
+	if usage.TotalTokens > 0 {
+		return usage.TotalTokens
+	}
+	pt := usagePromptTokens(usage)
+	ct := usageCompletionTokens(usage)
+	if pt > 0 || ct > 0 {
+		return pt + ct
+	}
+	return 0
+}
+
+func getChannelAffinityUsageCacheStatsCache() *cachex.HybridCache[ChannelAffinityUsageCacheCounters] {
+	channelAffinityUsageCacheStatsOnce.Do(func() {
+		setting := operation_setting.GetChannelAffinitySetting()
+		capacity := 100_000
+		defaultTTLSeconds := 3600
+		if setting != nil {
+			if setting.MaxEntries > 0 {
+				capacity = setting.MaxEntries
+			}
+			if setting.DefaultTTLSeconds > 0 {
+				defaultTTLSeconds = setting.DefaultTTLSeconds
+			}
+		}
+
+		channelAffinityUsageCacheStatsCache = cachex.NewHybridCache[ChannelAffinityUsageCacheCounters](cachex.HybridCacheConfig[ChannelAffinityUsageCacheCounters]{
+			Namespace: cachex.Namespace(channelAffinityUsageCacheStatsNamespace),
+			Redis:     common.RDB,
+			RedisEnabled: func() bool {
+				return common.RedisEnabled && common.RDB != nil
+			},
+			RedisCodec: cachex.JSONCodec[ChannelAffinityUsageCacheCounters]{},
+			Memory: func() *hot.HotCache[string, ChannelAffinityUsageCacheCounters] {
+				return hot.NewHotCache[string, ChannelAffinityUsageCacheCounters](hot.LRU, capacity).
+					WithTTL(time.Duration(defaultTTLSeconds) * time.Second).
+					WithJanitor().
+					Build()
+			},
+		})
+	})
+	return channelAffinityUsageCacheStatsCache
+}
+
+func channelAffinityUsageCacheStatsLock(key string) *sync.Mutex {
+	h := fnv.New32a()
+	_, _ = h.Write([]byte(key))
+	idx := h.Sum32() % uint32(len(channelAffinityUsageCacheStatsLocks))
+	return &channelAffinityUsageCacheStatsLocks[idx]
+}

service/channel_affinity_template_test.go

@@ -0,0 +1,247 @@
+package service
+
+import (
+	"fmt"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+	"time"
+
+	relaycommon "github.com/QuantumNous/new-api/relay/common"
+	"github.com/QuantumNous/new-api/setting/operation_setting"
+	"github.com/gin-gonic/gin"
+	"github.com/stretchr/testify/require"
+)
+
+func buildChannelAffinityTemplateContextForTest(meta channelAffinityMeta) *gin.Context {
+	rec := httptest.NewRecorder()
+	ctx, _ := gin.CreateTestContext(rec)
+	setChannelAffinityContext(ctx, meta)
+	return ctx
+}
+
+func TestApplyChannelAffinityOverrideTemplate_NoTemplate(t *testing.T) {
+	ctx := buildChannelAffinityTemplateContextForTest(channelAffinityMeta{
+		RuleName: "rule-no-template",
+	})
+	base := map[string]interface{}{
+		"temperature": 0.7,
+	}
+
+	merged, applied := ApplyChannelAffinityOverrideTemplate(ctx, base)
+	require.False(t, applied)
+	require.Equal(t, base, merged)
+}
+
+func TestApplyChannelAffinityOverrideTemplate_MergeTemplate(t *testing.T) {
+	ctx := buildChannelAffinityTemplateContextForTest(channelAffinityMeta{
+		RuleName: "rule-with-template",
+		ParamTemplate: map[string]interface{}{
+			"temperature": 0.2,
+			"top_p":       0.95,
+		},
+		UsingGroup:     "default",
+		ModelName:      "gpt-4.1",
+		RequestPath:    "/v1/responses",
+		KeySourceType:  "gjson",
+		KeySourcePath:  "prompt_cache_key",
+		KeyHint:        "abcd...wxyz",
+		KeyFingerprint: "abcd1234",
+	})
+	base := map[string]interface{}{
+		"temperature": 0.7,
+		"max_tokens":  2000,
+	}
+
+	merged, applied := ApplyChannelAffinityOverrideTemplate(ctx, base)
+	require.True(t, applied)
+	require.Equal(t, 0.7, merged["temperature"])
+	require.Equal(t, 0.95, merged["top_p"])
+	require.Equal(t, 2000, merged["max_tokens"])
+	require.Equal(t, 0.7, base["temperature"])
+
+	anyInfo, ok := ctx.Get(ginKeyChannelAffinityLogInfo)
+	require.True(t, ok)
+	info, ok := anyInfo.(map[string]interface{})
+	require.True(t, ok)
+	overrideInfoAny, ok := info["override_template"]
+	require.True(t, ok)
+	overrideInfo, ok := overrideInfoAny.(map[string]interface{})
+	require.True(t, ok)
+	require.Equal(t, true, overrideInfo["applied"])
+	require.Equal(t, "rule-with-template", overrideInfo["rule_name"])
+	require.EqualValues(t, 2, overrideInfo["param_override_keys"])
+}
+
+func TestApplyChannelAffinityOverrideTemplate_MergeOperations(t *testing.T) {
+	ctx := buildChannelAffinityTemplateContextForTest(channelAffinityMeta{
+		RuleName: "rule-with-ops-template",
+		ParamTemplate: map[string]interface{}{
+			"operations": []map[string]interface{}{
+				{
+					"mode":  "pass_headers",
+					"value": []string{"Originator"},
+				},
+			},
+		},
+	})
+	base := map[string]interface{}{
+		"temperature": 0.7,
+		"operations": []map[string]interface{}{
+			{
+				"path":  "model",
+				"mode":  "trim_prefix",
+				"value": "openai/",
+			},
+		},
+	}
+
+	merged, applied := ApplyChannelAffinityOverrideTemplate(ctx, base)
+	require.True(t, applied)
+	require.Equal(t, 0.7, merged["temperature"])
+
+	opsAny, ok := merged["operations"]
+	require.True(t, ok)
+	ops, ok := opsAny.([]interface{})
+	require.True(t, ok)
+	require.Len(t, ops, 2)
+
+	firstOp, ok := ops[0].(map[string]interface{})
+	require.True(t, ok)
+	require.Equal(t, "pass_headers", firstOp["mode"])
+
+	secondOp, ok := ops[1].(map[string]interface{})
+	require.True(t, ok)
+	require.Equal(t, "trim_prefix", secondOp["mode"])
+}
+
+func TestShouldSkipRetryAfterChannelAffinityFailure(t *testing.T) {
+	tests := []struct {
+		name string
+		ctx  func() *gin.Context
+		want bool
+	}{
+		{
+			name: "nil context",
+			ctx: func() *gin.Context {
+				return nil
+			},
+			want: false,
+		},
+		{
+			name: "explicit skip retry flag in context",
+			ctx: func() *gin.Context {
+				ctx := buildChannelAffinityTemplateContextForTest(channelAffinityMeta{
+					RuleName:   "rule-explicit-flag",
+					SkipRetry:  false,
+					UsingGroup: "default",
+					ModelName:  "gpt-5",
+				})
+				ctx.Set(ginKeyChannelAffinitySkipRetry, true)
+				return ctx
+			},
+			want: true,
+		},
+		{
+			name: "fallback to matched rule meta",
+			ctx: func() *gin.Context {
+				return buildChannelAffinityTemplateContextForTest(channelAffinityMeta{
+					RuleName:   "rule-skip-retry",
+					SkipRetry:  true,
+					UsingGroup: "default",
+					ModelName:  "gpt-5",
+				})
+			},
+			want: true,
+		},
+		{
+			name: "no flag and no skip retry meta",
+			ctx: func() *gin.Context {
+				return buildChannelAffinityTemplateContextForTest(channelAffinityMeta{
+					RuleName:   "rule-no-skip-retry",
+					SkipRetry:  false,
+					UsingGroup: "default",
+					ModelName:  "gpt-5",
+				})
+			},
+			want: false,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			require.Equal(t, tt.want, ShouldSkipRetryAfterChannelAffinityFailure(tt.ctx()))
+		})
+	}
+}
+
+func TestChannelAffinityHitCodexTemplatePassHeadersEffective(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	setting := operation_setting.GetChannelAffinitySetting()
+	require.NotNil(t, setting)
+
+	var codexRule *operation_setting.ChannelAffinityRule
+	for i := range setting.Rules {
+		rule := &setting.Rules[i]
+		if strings.EqualFold(strings.TrimSpace(rule.Name), "codex cli trace") {
+			codexRule = rule
+			break
+		}
+	}
+	require.NotNil(t, codexRule)
+
+	affinityValue := fmt.Sprintf("pc-hit-%d", time.Now().UnixNano())
+	cacheKeySuffix := buildChannelAffinityCacheKeySuffix(*codexRule, "default", affinityValue)
+
+	cache := getChannelAffinityCache()
+	require.NoError(t, cache.SetWithTTL(cacheKeySuffix, 9527, time.Minute))
+	t.Cleanup(func() {
+		_, _ = cache.DeleteMany([]string{cacheKeySuffix})
+	})
+
+	rec := httptest.NewRecorder()
+	ctx, _ := gin.CreateTestContext(rec)
+	ctx.Request = httptest.NewRequest(http.MethodPost, "/v1/responses", strings.NewReader(fmt.Sprintf(`{"prompt_cache_key":"%s"}`, affinityValue)))
+	ctx.Request.Header.Set("Content-Type", "application/json")
+
+	channelID, found := GetPreferredChannelByAffinity(ctx, "gpt-5", "default")
+	require.True(t, found)
+	require.Equal(t, 9527, channelID)
+
+	baseOverride := map[string]interface{}{
+		"temperature": 0.2,
+	}
+	mergedOverride, applied := ApplyChannelAffinityOverrideTemplate(ctx, baseOverride)
+	require.True(t, applied)
+	require.Equal(t, 0.2, mergedOverride["temperature"])
+
+	info := &relaycommon.RelayInfo{
+		RequestHeaders: map[string]string{
+			"Originator": "Codex CLI",
+			"Session_id": "sess-123",
+			"User-Agent": "codex-cli-test",
+		},
+		ChannelMeta: &relaycommon.ChannelMeta{
+			ParamOverride: mergedOverride,
+			HeadersOverride: map[string]interface{}{
+				"X-Static": "legacy-static",
+			},
+		},
+	}
+
+	_, err := relaycommon.ApplyParamOverrideWithRelayInfo([]byte(`{"model":"gpt-5"}`), info)
+	require.NoError(t, err)
+	require.True(t, info.UseRuntimeHeadersOverride)
+
+	require.Equal(t, "legacy-static", info.RuntimeHeadersOverride["x-static"])
+	require.Equal(t, "Codex CLI", info.RuntimeHeadersOverride["originator"])
+	require.Equal(t, "sess-123", info.RuntimeHeadersOverride["session_id"])
+	require.Equal(t, "codex-cli-test", info.RuntimeHeadersOverride["user-agent"])
+
+	_, exists := info.RuntimeHeadersOverride["x-codex-beta-features"]
+	require.False(t, exists)
+	_, exists = info.RuntimeHeadersOverride["x-codex-turn-metadata"]
+	require.False(t, exists)
+}

service/channel_affinity_usage_cache_test.go

@@ -0,0 +1,105 @@
+package service
+
+import (
+	"fmt"
+	"net/http/httptest"
+	"testing"
+	"time"
+
+	"github.com/QuantumNous/new-api/dto"
+	"github.com/QuantumNous/new-api/types"
+	"github.com/gin-gonic/gin"
+	"github.com/stretchr/testify/require"
+)
+
+func buildChannelAffinityStatsContextForTest(ruleName, usingGroup, keyFP string) *gin.Context {
+	rec := httptest.NewRecorder()
+	ctx, _ := gin.CreateTestContext(rec)
+	setChannelAffinityContext(ctx, channelAffinityMeta{
+		CacheKey:       fmt.Sprintf("test:%s:%s:%s", ruleName, usingGroup, keyFP),
+		TTLSeconds:     600,
+		RuleName:       ruleName,
+		UsingGroup:     usingGroup,
+		KeyFingerprint: keyFP,
+	})
+	return ctx
+}
+
+func TestObserveChannelAffinityUsageCacheByRelayFormat_ClaudeMode(t *testing.T) {
+	ruleName := fmt.Sprintf("rule_%d", time.Now().UnixNano())
+	usingGroup := "default"
+	keyFP := fmt.Sprintf("fp_%d", time.Now().UnixNano())
+	ctx := buildChannelAffinityStatsContextForTest(ruleName, usingGroup, keyFP)
+
+	usage := &dto.Usage{
+		PromptTokens:     100,
+		CompletionTokens: 40,
+		TotalTokens:      140,
+		PromptTokensDetails: dto.InputTokenDetails{
+			CachedTokens: 30,
+		},
+	}
+
+	ObserveChannelAffinityUsageCacheByRelayFormat(ctx, usage, types.RelayFormatClaude)
+	stats := GetChannelAffinityUsageCacheStats(ruleName, usingGroup, keyFP)
+
+	require.EqualValues(t, 1, stats.Total)
+	require.EqualValues(t, 1, stats.Hit)
+	require.EqualValues(t, 100, stats.PromptTokens)
+	require.EqualValues(t, 40, stats.CompletionTokens)
+	require.EqualValues(t, 140, stats.TotalTokens)
+	require.EqualValues(t, 30, stats.CachedTokens)
+	require.Equal(t, cacheTokenRateModeCachedOverPromptPlusCached, stats.CachedTokenRateMode)
+}
+
+func TestObserveChannelAffinityUsageCacheByRelayFormat_MixedMode(t *testing.T) {
+	ruleName := fmt.Sprintf("rule_%d", time.Now().UnixNano())
+	usingGroup := "default"
+	keyFP := fmt.Sprintf("fp_%d", time.Now().UnixNano())
+	ctx := buildChannelAffinityStatsContextForTest(ruleName, usingGroup, keyFP)
+
+	openAIUsage := &dto.Usage{
+		PromptTokens: 100,
+		PromptTokensDetails: dto.InputTokenDetails{
+			CachedTokens: 10,
+		},
+	}
+	claudeUsage := &dto.Usage{
+		PromptTokens: 80,
+		PromptTokensDetails: dto.InputTokenDetails{
+			CachedTokens: 20,
+		},
+	}
+
+	ObserveChannelAffinityUsageCacheByRelayFormat(ctx, openAIUsage, types.RelayFormatOpenAI)
+	ObserveChannelAffinityUsageCacheByRelayFormat(ctx, claudeUsage, types.RelayFormatClaude)
+	stats := GetChannelAffinityUsageCacheStats(ruleName, usingGroup, keyFP)
+
+	require.EqualValues(t, 2, stats.Total)
+	require.EqualValues(t, 2, stats.Hit)
+	require.EqualValues(t, 180, stats.PromptTokens)
+	require.EqualValues(t, 30, stats.CachedTokens)
+	require.Equal(t, cacheTokenRateModeMixed, stats.CachedTokenRateMode)
+}
+
+func TestObserveChannelAffinityUsageCacheByRelayFormat_UnsupportedModeKeepsEmpty(t *testing.T) {
+	ruleName := fmt.Sprintf("rule_%d", time.Now().UnixNano())
+	usingGroup := "default"
+	keyFP := fmt.Sprintf("fp_%d", time.Now().UnixNano())
+	ctx := buildChannelAffinityStatsContextForTest(ruleName, usingGroup, keyFP)
+
+	usage := &dto.Usage{
+		PromptTokens: 100,
+		PromptTokensDetails: dto.InputTokenDetails{
+			CachedTokens: 25,
+		},
+	}
+
+	ObserveChannelAffinityUsageCacheByRelayFormat(ctx, usage, types.RelayFormatGemini)
+	stats := GetChannelAffinityUsageCacheStats(ruleName, usingGroup, keyFP)
+
+	require.EqualValues(t, 1, stats.Total)
+	require.EqualValues(t, 1, stats.Hit)
+	require.EqualValues(t, 25, stats.CachedTokens)
+	require.Equal(t, "", stats.CachedTokenRateMode)
+}

service/channel_select.go

@@ -0,0 +1,162 @@
+package service
+
+import (
+	"errors"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/constant"
+	"github.com/QuantumNous/new-api/logger"
+	"github.com/QuantumNous/new-api/model"
+	"github.com/QuantumNous/new-api/setting"
+	"github.com/gin-gonic/gin"
+)
+
+type RetryParam struct {
+	Ctx          *gin.Context
+	TokenGroup   string
+	ModelName    string
+	Retry        *int
+	resetNextTry bool
+}
+
+func (p *RetryParam) GetRetry() int {
+	if p.Retry == nil {
+		return 0
+	}
+	return *p.Retry
+}
+
+func (p *RetryParam) SetRetry(retry int) {
+	p.Retry = &retry
+}
+
+func (p *RetryParam) IncreaseRetry() {
+	if p.resetNextTry {
+		p.resetNextTry = false
+		return
+	}
+	if p.Retry == nil {
+		p.Retry = new(int)
+	}
+	*p.Retry++
+}
+
+func (p *RetryParam) ResetRetryNextTry() {
+	p.resetNextTry = true
+}
+
+// CacheGetRandomSatisfiedChannel tries to get a random channel that satisfies the requirements.
+// 尝试获取一个满足要求的随机渠道。
+//
+// For "auto" tokenGroup with cross-group Retry enabled:
+// 对于启用了跨分组重试的 "auto" tokenGroup：
+//
+//   - Each group will exhaust all its priorities before moving to the next group.
+//     每个分组会用完所有优先级后才会切换到下一个分组。
+//
+//   - Uses ContextKeyAutoGroupIndex to track current group index.
+//     使用 ContextKeyAutoGroupIndex 跟踪当前分组索引。
+//
+//   - Uses ContextKeyAutoGroupRetryIndex to track the global Retry count when current group started.
+//     使用 ContextKeyAutoGroupRetryIndex 跟踪当前分组开始时的全局重试次数。
+//
+//   - priorityRetry = Retry - startRetryIndex, represents the priority level within current group.
+//     priorityRetry = Retry - startRetryIndex，表示当前分组内的优先级级别。
+//
+//   - When GetRandomSatisfiedChannel returns nil (priorities exhausted), moves to next group.
+//     当 GetRandomSatisfiedChannel 返回 nil（优先级用完）时，切换到下一个分组。
+//
+// Example flow (2 groups, each with 2 priorities, RetryTimes=3):
+// 示例流程（2个分组，每个有2个优先级，RetryTimes=3）：
+//
+//	Retry=0: GroupA, priority0 (startRetryIndex=0, priorityRetry=0)
+//	         分组A, 优先级0
+//
+//	Retry=1: GroupA, priority1 (startRetryIndex=0, priorityRetry=1)
+//	         分组A, 优先级1
+//
+//	Retry=2: GroupA exhausted → GroupB, priority0 (startRetryIndex=2, priorityRetry=0)
+//	         分组A用完 → 分组B, 优先级0
+//
+//	Retry=3: GroupB, priority1 (startRetryIndex=2, priorityRetry=1)
+//	         分组B, 优先级1
+func CacheGetRandomSatisfiedChannel(param *RetryParam) (*model.Channel, string, error) {
+	var channel *model.Channel
+	var err error
+	selectGroup := param.TokenGroup
+	userGroup := common.GetContextKeyString(param.Ctx, constant.ContextKeyUserGroup)
+
+	if param.TokenGroup == "auto" {
+		if len(setting.GetAutoGroups()) == 0 {
+			return nil, selectGroup, errors.New("auto groups is not enabled")
+		}
+		autoGroups := GetUserAutoGroup(userGroup)
+
+		// startGroupIndex: the group index to start searching from
+		// startGroupIndex: 开始搜索的分组索引
+		startGroupIndex := 0
+		crossGroupRetry := common.GetContextKeyBool(param.Ctx, constant.ContextKeyTokenCrossGroupRetry)
+
+		if lastGroupIndex, exists := common.GetContextKey(param.Ctx, constant.ContextKeyAutoGroupIndex); exists {
+			if idx, ok := lastGroupIndex.(int); ok {
+				startGroupIndex = idx
+			}
+		}
+
+		for i := startGroupIndex; i < len(autoGroups); i++ {
+			autoGroup := autoGroups[i]
+			// Calculate priorityRetry for current group
+			// 计算当前分组的 priorityRetry
+			priorityRetry := param.GetRetry()
+			// If moved to a new group, reset priorityRetry and update startRetryIndex
+			// 如果切换到新分组，重置 priorityRetry 并更新 startRetryIndex
+			if i > startGroupIndex {
+				priorityRetry = 0
+			}
+			logger.LogDebug(param.Ctx, "Auto selecting group: %s, priorityRetry: %d", autoGroup, priorityRetry)
+
+			channel, _ = model.GetRandomSatisfiedChannel(autoGroup, param.ModelName, priorityRetry)
+			if channel == nil {
+				// Current group has no available channel for this model, try next group
+				// 当前分组没有该模型的可用渠道，尝试下一个分组
+				logger.LogDebug(param.Ctx, "No available channel in group %s for model %s at priorityRetry %d, trying next group", autoGroup, param.ModelName, priorityRetry)
+				// 重置状态以尝试下一个分组
+				common.SetContextKey(param.Ctx, constant.ContextKeyAutoGroupIndex, i+1)
+				common.SetContextKey(param.Ctx, constant.ContextKeyAutoGroupRetryIndex, 0)
+				// Reset retry counter so outer loop can continue for next group
+				// 重置重试计数器，以便外层循环可以为下一个分组继续
+				param.SetRetry(0)
+				continue
+			}
+			common.SetContextKey(param.Ctx, constant.ContextKeyAutoGroup, autoGroup)
+			selectGroup = autoGroup
+			logger.LogDebug(param.Ctx, "Auto selected group: %s", autoGroup)
+
+			// Prepare state for next retry
+			// 为下一次重试准备状态
+			if crossGroupRetry && priorityRetry >= common.RetryTimes {
+				// Current group has exhausted all retries, prepare to switch to next group
+				// This request still uses current group, but next retry will use next group
+				// 当前分组已用完所有重试次数，准备切换到下一个分组
+				// 本次请求仍使用当前分组，但下次重试将使用下一个分组
+				logger.LogDebug(param.Ctx, "Current group %s retries exhausted (priorityRetry=%d >= RetryTimes=%d), preparing switch to next group for next retry", autoGroup, priorityRetry, common.RetryTimes)
+				common.SetContextKey(param.Ctx, constant.ContextKeyAutoGroupIndex, i+1)
+				// Reset retry counter so outer loop can continue for next group
+				// 重置重试计数器，以便外层循环可以为下一个分组继续
+				param.SetRetry(0)
+				param.ResetRetryNextTry()
+			} else {
+				// Stay in current group, save current state
+				// 保持在当前分组，保存当前状态
+				common.SetContextKey(param.Ctx, constant.ContextKeyAutoGroupIndex, i)
+			}
+			break
+		}
+	} else {
+		channel, err = model.GetRandomSatisfiedChannel(param.TokenGroup, param.ModelName, param.GetRetry())
+		if err != nil {
+			return nil, param.TokenGroup, err
+		}
+	}
+	return channel, selectGroup, nil
+}

service/codex_credential_refresh.go

@@ -0,0 +1,104 @@
+package service
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"strings"
+	"time"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/constant"
+	"github.com/QuantumNous/new-api/model"
+)
+
+type CodexCredentialRefreshOptions struct {
+	ResetCaches bool
+}
+
+type CodexOAuthKey struct {
+	IDToken      string `json:"id_token,omitempty"`
+	AccessToken  string `json:"access_token,omitempty"`
+	RefreshToken string `json:"refresh_token,omitempty"`
+
+	AccountID   string `json:"account_id,omitempty"`
+	LastRefresh string `json:"last_refresh,omitempty"`
+	Email       string `json:"email,omitempty"`
+	Type        string `json:"type,omitempty"`
+	Expired     string `json:"expired,omitempty"`
+}
+
+func parseCodexOAuthKey(raw string) (*CodexOAuthKey, error) {
+	if strings.TrimSpace(raw) == "" {
+		return nil, errors.New("codex channel: empty oauth key")
+	}
+	var key CodexOAuthKey
+	if err := common.Unmarshal([]byte(raw), &key); err != nil {
+		return nil, errors.New("codex channel: invalid oauth key json")
+	}
+	return &key, nil
+}
+
+func RefreshCodexChannelCredential(ctx context.Context, channelID int, opts CodexCredentialRefreshOptions) (*CodexOAuthKey, *model.Channel, error) {
+	ch, err := model.GetChannelById(channelID, true)
+	if err != nil {
+		return nil, nil, err
+	}
+	if ch == nil {
+		return nil, nil, fmt.Errorf("channel not found")
+	}
+	if ch.Type != constant.ChannelTypeCodex {
+		return nil, nil, fmt.Errorf("channel type is not Codex")
+	}
+
+	oauthKey, err := parseCodexOAuthKey(strings.TrimSpace(ch.Key))
+	if err != nil {
+		return nil, nil, err
+	}
+	if strings.TrimSpace(oauthKey.RefreshToken) == "" {
+		return nil, nil, fmt.Errorf("codex channel: refresh_token is required to refresh credential")
+	}
+
+	refreshCtx, cancel := context.WithTimeout(ctx, 10*time.Second)
+	defer cancel()
+
+	res, err := RefreshCodexOAuthTokenWithProxy(refreshCtx, oauthKey.RefreshToken, ch.GetSetting().Proxy)
+	if err != nil {
+		return nil, nil, err
+	}
+
+	oauthKey.AccessToken = res.AccessToken
+	oauthKey.RefreshToken = res.RefreshToken
+	oauthKey.LastRefresh = time.Now().Format(time.RFC3339)
+	oauthKey.Expired = res.ExpiresAt.Format(time.RFC3339)
+	if strings.TrimSpace(oauthKey.Type) == "" {
+		oauthKey.Type = "codex"
+	}
+
+	if strings.TrimSpace(oauthKey.AccountID) == "" {
+		if accountID, ok := ExtractCodexAccountIDFromJWT(oauthKey.AccessToken); ok {
+			oauthKey.AccountID = accountID
+		}
+	}
+	if strings.TrimSpace(oauthKey.Email) == "" {
+		if email, ok := ExtractEmailFromJWT(oauthKey.AccessToken); ok {
+			oauthKey.Email = email
+		}
+	}
+
+	encoded, err := common.Marshal(oauthKey)
+	if err != nil {
+		return nil, nil, err
+	}
+
+	if err := model.DB.Model(&model.Channel{}).Where("id = ?", ch.Id).Update("key", string(encoded)).Error; err != nil {
+		return nil, nil, err
+	}
+
+	if opts.ResetCaches {
+		model.InitChannelCache()
+		ResetProxyClientCache()
+	}
+
+	return oauthKey, ch, nil
+}

service/codex_credential_refresh_task.go

@@ -0,0 +1,140 @@
+package service
+
+import (
+	"context"
+	"fmt"
+	"strings"
+	"sync"
+	"sync/atomic"
+	"time"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/constant"
+	"github.com/QuantumNous/new-api/logger"
+	"github.com/QuantumNous/new-api/model"
+
+	"github.com/bytedance/gopkg/util/gopool"
+)
+
+const (
+	codexCredentialRefreshTickInterval = 10 * time.Minute
+	codexCredentialRefreshThreshold    = 24 * time.Hour
+	codexCredentialRefreshBatchSize    = 200
+	codexCredentialRefreshTimeout      = 15 * time.Second
+)
+
+var (
+	codexCredentialRefreshOnce    sync.Once
+	codexCredentialRefreshRunning atomic.Bool
+)
+
+func StartCodexCredentialAutoRefreshTask() {
+	codexCredentialRefreshOnce.Do(func() {
+		if !common.IsMasterNode {
+			return
+		}
+
+		gopool.Go(func() {
+			logger.LogInfo(context.Background(), fmt.Sprintf("codex credential auto-refresh task started: tick=%s threshold=%s", codexCredentialRefreshTickInterval, codexCredentialRefreshThreshold))
+
+			ticker := time.NewTicker(codexCredentialRefreshTickInterval)
+			defer ticker.Stop()
+
+			runCodexCredentialAutoRefreshOnce()
+			for range ticker.C {
+				runCodexCredentialAutoRefreshOnce()
+			}
+		})
+	})
+}
+
+func runCodexCredentialAutoRefreshOnce() {
+	if !codexCredentialRefreshRunning.CompareAndSwap(false, true) {
+		return
+	}
+	defer codexCredentialRefreshRunning.Store(false)
+
+	ctx := context.Background()
+	now := time.Now()
+
+	var refreshed int
+	var scanned int
+
+	offset := 0
+	for {
+		var channels []*model.Channel
+		err := model.DB.
+			Select("id", "name", "key", "status", "channel_info").
+			Where("type = ? AND status = 1", constant.ChannelTypeCodex).
+			Order("id asc").
+			Limit(codexCredentialRefreshBatchSize).
+			Offset(offset).
+			Find(&channels).Error
+		if err != nil {
+			logger.LogError(ctx, fmt.Sprintf("codex credential auto-refresh: query channels failed: %v", err))
+			return
+		}
+		if len(channels) == 0 {
+			break
+		}
+		offset += codexCredentialRefreshBatchSize
+
+		for _, ch := range channels {
+			if ch == nil {
+				continue
+			}
+			scanned++
+			if ch.ChannelInfo.IsMultiKey {
+				continue
+			}
+
+			rawKey := strings.TrimSpace(ch.Key)
+			if rawKey == "" {
+				continue
+			}
+
+			oauthKey, err := parseCodexOAuthKey(rawKey)
+			if err != nil {
+				continue
+			}
+
+			refreshToken := strings.TrimSpace(oauthKey.RefreshToken)
+			if refreshToken == "" {
+				continue
+			}
+
+			expiredAtRaw := strings.TrimSpace(oauthKey.Expired)
+			expiredAt, err := time.Parse(time.RFC3339, expiredAtRaw)
+			if err == nil && !expiredAt.IsZero() && expiredAt.Sub(now) > codexCredentialRefreshThreshold {
+				continue
+			}
+
+			refreshCtx, cancel := context.WithTimeout(ctx, codexCredentialRefreshTimeout)
+			newKey, _, err := RefreshCodexChannelCredential(refreshCtx, ch.Id, CodexCredentialRefreshOptions{ResetCaches: false})
+			cancel()
+			if err != nil {
+				logger.LogWarn(ctx, fmt.Sprintf("codex credential auto-refresh: channel_id=%d name=%s refresh failed: %v", ch.Id, ch.Name, err))
+				continue
+			}
+
+			refreshed++
+			logger.LogInfo(ctx, fmt.Sprintf("codex credential auto-refresh: channel_id=%d name=%s refreshed, expires_at=%s", ch.Id, ch.Name, newKey.Expired))
+		}
+	}
+
+	if refreshed > 0 {
+		func() {
+			defer func() {
+				if r := recover(); r != nil {
+					logger.LogWarn(ctx, fmt.Sprintf("codex credential auto-refresh: InitChannelCache panic: %v", r))
+				}
+			}()
+			model.InitChannelCache()
+		}()
+		ResetProxyClientCache()
+	}
+
+	if common.DebugEnabled {
+		logger.LogDebug(ctx, "codex credential auto-refresh: scanned=%d refreshed=%d", scanned, refreshed)
+	}
+}

service/codex_oauth.go

@@ -0,0 +1,317 @@
+package service
+
+import (
+	"context"
+	"crypto/rand"
+	"crypto/sha256"
+	"encoding/base64"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"net/http"
+	"net/url"
+	"strings"
+	"time"
+
+	"github.com/QuantumNous/new-api/common"
+)
+
+const (
+	codexOAuthClientID     = "app_EMoamEEZ73f0CkXaXp7hrann"
+	codexOAuthAuthorizeURL = "https://auth.openai.com/oauth/authorize"
+	codexOAuthTokenURL     = "https://auth.openai.com/oauth/token"
+	codexOAuthRedirectURI  = "http://localhost:1455/auth/callback"
+	codexOAuthScope        = "openid profile email offline_access"
+	codexJWTClaimPath      = "https://api.openai.com/auth"
+	defaultHTTPTimeout     = 20 * time.Second
+)
+
+type CodexOAuthTokenResult struct {
+	AccessToken  string
+	RefreshToken string
+	ExpiresAt    time.Time
+}
+
+type CodexOAuthAuthorizationFlow struct {
+	State        string
+	Verifier     string
+	Challenge    string
+	AuthorizeURL string
+}
+
+func RefreshCodexOAuthToken(ctx context.Context, refreshToken string) (*CodexOAuthTokenResult, error) {
+	return RefreshCodexOAuthTokenWithProxy(ctx, refreshToken, "")
+}
+
+func RefreshCodexOAuthTokenWithProxy(ctx context.Context, refreshToken string, proxyURL string) (*CodexOAuthTokenResult, error) {
+	client, err := getCodexOAuthHTTPClient(proxyURL)
+	if err != nil {
+		return nil, err
+	}
+	return refreshCodexOAuthToken(ctx, client, codexOAuthTokenURL, codexOAuthClientID, refreshToken)
+}
+
+func ExchangeCodexAuthorizationCode(ctx context.Context, code string, verifier string) (*CodexOAuthTokenResult, error) {
+	return ExchangeCodexAuthorizationCodeWithProxy(ctx, code, verifier, "")
+}
+
+func ExchangeCodexAuthorizationCodeWithProxy(ctx context.Context, code string, verifier string, proxyURL string) (*CodexOAuthTokenResult, error) {
+	client, err := getCodexOAuthHTTPClient(proxyURL)
+	if err != nil {
+		return nil, err
+	}
+	return exchangeCodexAuthorizationCode(ctx, client, codexOAuthTokenURL, codexOAuthClientID, code, verifier, codexOAuthRedirectURI)
+}
+
+func CreateCodexOAuthAuthorizationFlow() (*CodexOAuthAuthorizationFlow, error) {
+	state, err := createStateHex(16)
+	if err != nil {
+		return nil, err
+	}
+	verifier, challenge, err := generatePKCEPair()
+	if err != nil {
+		return nil, err
+	}
+	u, err := buildCodexAuthorizeURL(state, challenge)
+	if err != nil {
+		return nil, err
+	}
+	return &CodexOAuthAuthorizationFlow{
+		State:        state,
+		Verifier:     verifier,
+		Challenge:    challenge,
+		AuthorizeURL: u,
+	}, nil
+}
+
+func refreshCodexOAuthToken(
+	ctx context.Context,
+	client *http.Client,
+	tokenURL string,
+	clientID string,
+	refreshToken string,
+) (*CodexOAuthTokenResult, error) {
+	rt := strings.TrimSpace(refreshToken)
+	if rt == "" {
+		return nil, errors.New("empty refresh_token")
+	}
+
+	form := url.Values{}
+	form.Set("grant_type", "refresh_token")
+	form.Set("refresh_token", rt)
+	form.Set("client_id", clientID)
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, tokenURL, strings.NewReader(form.Encode()))
+	if err != nil {
+		return nil, err
+	}
+	req.Header.Set("Content-Type", "application/x-www-form-urlencoded")
+	req.Header.Set("Accept", "application/json")
+
+	resp, err := client.Do(req)
+	if err != nil {
+		return nil, err
+	}
+	defer resp.Body.Close()
+
+	var payload struct {
+		AccessToken  string `json:"access_token"`
+		RefreshToken string `json:"refresh_token"`
+		ExpiresIn    int    `json:"expires_in"`
+	}
+
+	if err := common.DecodeJson(resp.Body, &payload); err != nil {
+		return nil, err
+	}
+	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+		return nil, fmt.Errorf("codex oauth refresh failed: status=%d", resp.StatusCode)
+	}
+
+	if strings.TrimSpace(payload.AccessToken) == "" || strings.TrimSpace(payload.RefreshToken) == "" || payload.ExpiresIn <= 0 {
+		return nil, errors.New("codex oauth refresh response missing fields")
+	}
+
+	return &CodexOAuthTokenResult{
+		AccessToken:  strings.TrimSpace(payload.AccessToken),
+		RefreshToken: strings.TrimSpace(payload.RefreshToken),
+		ExpiresAt:    time.Now().Add(time.Duration(payload.ExpiresIn) * time.Second),
+	}, nil
+}
+
+func exchangeCodexAuthorizationCode(
+	ctx context.Context,
+	client *http.Client,
+	tokenURL string,
+	clientID string,
+	code string,
+	verifier string,
+	redirectURI string,
+) (*CodexOAuthTokenResult, error) {
+	c := strings.TrimSpace(code)
+	v := strings.TrimSpace(verifier)
+	if c == "" {
+		return nil, errors.New("empty authorization code")
+	}
+	if v == "" {
+		return nil, errors.New("empty code_verifier")
+	}
+
+	form := url.Values{}
+	form.Set("grant_type", "authorization_code")
+	form.Set("client_id", clientID)
+	form.Set("code", c)
+	form.Set("code_verifier", v)
+	form.Set("redirect_uri", redirectURI)
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, tokenURL, strings.NewReader(form.Encode()))
+	if err != nil {
+		return nil, err
+	}
+	req.Header.Set("Content-Type", "application/x-www-form-urlencoded")
+	req.Header.Set("Accept", "application/json")
+
+	resp, err := client.Do(req)
+	if err != nil {
+		return nil, err
+	}
+	defer resp.Body.Close()
+
+	var payload struct {
+		AccessToken  string `json:"access_token"`
+		RefreshToken string `json:"refresh_token"`
+		ExpiresIn    int    `json:"expires_in"`
+	}
+	if err := common.DecodeJson(resp.Body, &payload); err != nil {
+		return nil, err
+	}
+	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+		return nil, fmt.Errorf("codex oauth code exchange failed: status=%d", resp.StatusCode)
+	}
+	if strings.TrimSpace(payload.AccessToken) == "" || strings.TrimSpace(payload.RefreshToken) == "" || payload.ExpiresIn <= 0 {
+		return nil, errors.New("codex oauth token response missing fields")
+	}
+	return &CodexOAuthTokenResult{
+		AccessToken:  strings.TrimSpace(payload.AccessToken),
+		RefreshToken: strings.TrimSpace(payload.RefreshToken),
+		ExpiresAt:    time.Now().Add(time.Duration(payload.ExpiresIn) * time.Second),
+	}, nil
+}
+
+func getCodexOAuthHTTPClient(proxyURL string) (*http.Client, error) {
+	baseClient, err := GetHttpClientWithProxy(strings.TrimSpace(proxyURL))
+	if err != nil {
+		return nil, err
+	}
+	if baseClient == nil {
+		return &http.Client{Timeout: defaultHTTPTimeout}, nil
+	}
+	clientCopy := *baseClient
+	clientCopy.Timeout = defaultHTTPTimeout
+	return &clientCopy, nil
+}
+
+func buildCodexAuthorizeURL(state string, challenge string) (string, error) {
+	u, err := url.Parse(codexOAuthAuthorizeURL)
+	if err != nil {
+		return "", err
+	}
+	q := u.Query()
+	q.Set("response_type", "code")
+	q.Set("client_id", codexOAuthClientID)
+	q.Set("redirect_uri", codexOAuthRedirectURI)
+	q.Set("scope", codexOAuthScope)
+	q.Set("code_challenge", challenge)
+	q.Set("code_challenge_method", "S256")
+	q.Set("state", state)
+	q.Set("id_token_add_organizations", "true")
+	q.Set("codex_cli_simplified_flow", "true")
+	q.Set("originator", "codex_cli_rs")
+	u.RawQuery = q.Encode()
+	return u.String(), nil
+}
+
+func createStateHex(nBytes int) (string, error) {
+	if nBytes <= 0 {
+		return "", errors.New("invalid state bytes length")
+	}
+	b := make([]byte, nBytes)
+	if _, err := rand.Read(b); err != nil {
+		return "", err
+	}
+	return fmt.Sprintf("%x", b), nil
+}
+
+func generatePKCEPair() (verifier string, challenge string, err error) {
+	b := make([]byte, 32)
+	if _, err := rand.Read(b); err != nil {
+		return "", "", err
+	}
+	verifier = base64.RawURLEncoding.EncodeToString(b)
+	sum := sha256.Sum256([]byte(verifier))
+	challenge = base64.RawURLEncoding.EncodeToString(sum[:])
+	return verifier, challenge, nil
+}
+
+func ExtractCodexAccountIDFromJWT(token string) (string, bool) {
+	claims, ok := decodeJWTClaims(token)
+	if !ok {
+		return "", false
+	}
+	raw, ok := claims[codexJWTClaimPath]
+	if !ok {
+		return "", false
+	}
+	obj, ok := raw.(map[string]any)
+	if !ok {
+		return "", false
+	}
+	v, ok := obj["chatgpt_account_id"]
+	if !ok {
+		return "", false
+	}
+	s, ok := v.(string)
+	if !ok {
+		return "", false
+	}
+	s = strings.TrimSpace(s)
+	if s == "" {
+		return "", false
+	}
+	return s, true
+}
+
+func ExtractEmailFromJWT(token string) (string, bool) {
+	claims, ok := decodeJWTClaims(token)
+	if !ok {
+		return "", false
+	}
+	v, ok := claims["email"]
+	if !ok {
+		return "", false
+	}
+	s, ok := v.(string)
+	if !ok {
+		return "", false
+	}
+	s = strings.TrimSpace(s)
+	if s == "" {
+		return "", false
+	}
+	return s, true
+}
+
+func decodeJWTClaims(token string) (map[string]any, bool) {
+	parts := strings.Split(token, ".")
+	if len(parts) != 3 {
+		return nil, false
+	}
+	payloadRaw, err := base64.RawURLEncoding.DecodeString(parts[1])
+	if err != nil {
+		return nil, false
+	}
+	var claims map[string]any
+	if err := json.Unmarshal(payloadRaw, &claims); err != nil {
+		return nil, false
+	}
+	return claims, true
+}

service/codex_wham_usage.go

@@ -0,0 +1,56 @@
+package service
+
+import (
+	"context"
+	"fmt"
+	"io"
+	"net/http"
+	"strings"
+)
+
+func FetchCodexWhamUsage(
+	ctx context.Context,
+	client *http.Client,
+	baseURL string,
+	accessToken string,
+	accountID string,
+) (statusCode int, body []byte, err error) {
+	if client == nil {
+		return 0, nil, fmt.Errorf("nil http client")
+	}
+	bu := strings.TrimRight(strings.TrimSpace(baseURL), "/")
+	if bu == "" {
+		return 0, nil, fmt.Errorf("empty baseURL")
+	}
+	at := strings.TrimSpace(accessToken)
+	aid := strings.TrimSpace(accountID)
+	if at == "" {
+		return 0, nil, fmt.Errorf("empty accessToken")
+	}
+	if aid == "" {
+		return 0, nil, fmt.Errorf("empty accountID")
+	}
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, bu+"/backend-api/wham/usage", nil)
+	if err != nil {
+		return 0, nil, err
+	}
+	req.Header.Set("Authorization", "Bearer "+at)
+	req.Header.Set("chatgpt-account-id", aid)
+	req.Header.Set("Accept", "application/json")
+	if req.Header.Get("originator") == "" {
+		req.Header.Set("originator", "codex_cli_rs")
+	}
+
+	resp, err := client.Do(req)
+	if err != nil {
+		return 0, nil, err
+	}
+	defer resp.Body.Close()
+
+	body, err = io.ReadAll(resp.Body)
+	if err != nil {
+		return resp.StatusCode, nil, err
+	}
+	return resp.StatusCode, body, nil
+}

service/convert.go

@@ -0,0 +1,988 @@
+package service
+
+import (
+	"encoding/json"
+	"fmt"
+	"strings"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/constant"
+	"github.com/QuantumNous/new-api/dto"
+	"github.com/QuantumNous/new-api/relay/channel/openrouter"
+	relaycommon "github.com/QuantumNous/new-api/relay/common"
+	"github.com/QuantumNous/new-api/relay/reasonmap"
+	"github.com/samber/lo"
+)
+
+func ClaudeToOpenAIRequest(claudeRequest dto.ClaudeRequest, info *relaycommon.RelayInfo) (*dto.GeneralOpenAIRequest, error) {
+	openAIRequest := dto.GeneralOpenAIRequest{
+		Model:       claudeRequest.Model,
+		Temperature: claudeRequest.Temperature,
+	}
+	if claudeRequest.MaxTokens != nil {
+		openAIRequest.MaxTokens = lo.ToPtr(lo.FromPtr(claudeRequest.MaxTokens))
+	}
+	if claudeRequest.TopP != nil {
+		openAIRequest.TopP = lo.ToPtr(lo.FromPtr(claudeRequest.TopP))
+	}
+	if claudeRequest.TopK != nil {
+		openAIRequest.TopK = lo.ToPtr(lo.FromPtr(claudeRequest.TopK))
+	}
+	if claudeRequest.Stream != nil {
+		openAIRequest.Stream = lo.ToPtr(lo.FromPtr(claudeRequest.Stream))
+	}
+
+	isOpenRouter := info.ChannelType == constant.ChannelTypeOpenRouter
+
+	if isOpenRouter {
+		if effort := claudeRequest.GetEfforts(); effort != "" {
+			effortBytes, _ := json.Marshal(effort)
+			openAIRequest.Verbosity = effortBytes
+		}
+		if claudeRequest.Thinking != nil {
+			var reasoning openrouter.RequestReasoning
+			if claudeRequest.Thinking.Type == "enabled" {
+				reasoning = openrouter.RequestReasoning{
+					Enabled:   true,
+					MaxTokens: claudeRequest.Thinking.GetBudgetTokens(),
+				}
+			} else if claudeRequest.Thinking.Type == "adaptive" {
+				reasoning = openrouter.RequestReasoning{
+					Enabled: true,
+				}
+			}
+			reasoningJSON, err := json.Marshal(reasoning)
+			if err != nil {
+				return nil, fmt.Errorf("failed to marshal reasoning: %w", err)
+			}
+			openAIRequest.Reasoning = reasoningJSON
+		}
+	} else {
+		thinkingSuffix := "-thinking"
+		if strings.HasSuffix(info.OriginModelName, thinkingSuffix) &&
+			!strings.HasSuffix(openAIRequest.Model, thinkingSuffix) {
+			openAIRequest.Model = openAIRequest.Model + thinkingSuffix
+		}
+	}
+
+	// Convert stop sequences
+	if len(claudeRequest.StopSequences) == 1 {
+		openAIRequest.Stop = claudeRequest.StopSequences[0]
+	} else if len(claudeRequest.StopSequences) > 1 {
+		openAIRequest.Stop = claudeRequest.StopSequences
+	}
+
+	// Convert tools
+	tools, _ := common.Any2Type[[]dto.Tool](claudeRequest.Tools)
+	openAITools := make([]dto.ToolCallRequest, 0)
+	for _, claudeTool := range tools {
+		openAITool := dto.ToolCallRequest{
+			Type: "function",
+			Function: dto.FunctionRequest{
+				Name:        claudeTool.Name,
+				Description: claudeTool.Description,
+				Parameters:  claudeTool.InputSchema,
+			},
+		}
+		openAITools = append(openAITools, openAITool)
+	}
+	openAIRequest.Tools = openAITools
+
+	// Convert messages
+	openAIMessages := make([]dto.Message, 0)
+
+	// Add system message if present
+	if claudeRequest.System != nil {
+		if claudeRequest.IsStringSystem() && claudeRequest.GetStringSystem() != "" {
+			openAIMessage := dto.Message{
+				Role: "system",
+			}
+			openAIMessage.SetStringContent(claudeRequest.GetStringSystem())
+			openAIMessages = append(openAIMessages, openAIMessage)
+		} else {
+			systems := claudeRequest.ParseSystem()
+			if len(systems) > 0 {
+				openAIMessage := dto.Message{
+					Role: "system",
+				}
+				isOpenRouterClaude := isOpenRouter && strings.HasPrefix(info.UpstreamModelName, "anthropic/claude")
+				if isOpenRouterClaude {
+					systemMediaMessages := make([]dto.MediaContent, 0, len(systems))
+					for _, system := range systems {
+						message := dto.MediaContent{
+							Type:         "text",
+							Text:         system.GetText(),
+							CacheControl: system.CacheControl,
+						}
+						systemMediaMessages = append(systemMediaMessages, message)
+					}
+					openAIMessage.SetMediaContent(systemMediaMessages)
+				} else {
+					systemStr := ""
+					for _, system := range systems {
+						if system.Text != nil {
+							systemStr += *system.Text
+						}
+					}
+					openAIMessage.SetStringContent(systemStr)
+				}
+				openAIMessages = append(openAIMessages, openAIMessage)
+			}
+		}
+	}
+	for _, claudeMessage := range claudeRequest.Messages {
+		openAIMessage := dto.Message{
+			Role: claudeMessage.Role,
+		}
+
+		//log.Printf("claudeMessage.Content: %v", claudeMessage.Content)
+		if claudeMessage.IsStringContent() {
+			openAIMessage.SetStringContent(claudeMessage.GetStringContent())
+		} else {
+			content, err := claudeMessage.ParseContent()
+			if err != nil {
+				return nil, err
+			}
+			contents := content
+			var toolCalls []dto.ToolCallRequest
+			mediaMessages := make([]dto.MediaContent, 0, len(contents))
+
+			for _, mediaMsg := range contents {
+				switch mediaMsg.Type {
+				case "text", "input_text":
+					message := dto.MediaContent{
+						Type:         "text",
+						Text:         mediaMsg.GetText(),
+						CacheControl: mediaMsg.CacheControl,
+					}
+					mediaMessages = append(mediaMessages, message)
+				case "image":
+					// Handle image conversion (base64 to URL or keep as is)
+					imageData := fmt.Sprintf("data:%s;base64,%s", mediaMsg.Source.MediaType, mediaMsg.Source.Data)
+					//textContent += fmt.Sprintf("[Image: %s]", imageData)
+					mediaMessage := dto.MediaContent{
+						Type:     "image_url",
+						ImageUrl: &dto.MessageImageUrl{Url: imageData},
+					}
+					mediaMessages = append(mediaMessages, mediaMessage)
+				case "tool_use":
+					toolCall := dto.ToolCallRequest{
+						ID:   mediaMsg.Id,
+						Type: "function",
+						Function: dto.FunctionRequest{
+							Name:      mediaMsg.Name,
+							Arguments: toJSONString(mediaMsg.Input),
+						},
+					}
+					toolCalls = append(toolCalls, toolCall)
+				case "tool_result":
+					// Add tool result as a separate message
+					toolName := mediaMsg.Name
+					if toolName == "" {
+						toolName = claudeRequest.SearchToolNameByToolCallId(mediaMsg.ToolUseId)
+					}
+					oaiToolMessage := dto.Message{
+						Role:       "tool",
+						Name:       &toolName,
+						ToolCallId: mediaMsg.ToolUseId,
+					}
+					//oaiToolMessage.SetStringContent(*mediaMsg.GetMediaContent().Text)
+					if mediaMsg.IsStringContent() {
+						oaiToolMessage.SetStringContent(mediaMsg.GetStringContent())
+					} else {
+						mediaContents := mediaMsg.ParseMediaContent()
+						encodeJson, _ := common.Marshal(mediaContents)
+						oaiToolMessage.SetStringContent(string(encodeJson))
+					}
+					openAIMessages = append(openAIMessages, oaiToolMessage)
+				}
+			}
+
+			if len(toolCalls) > 0 {
+				openAIMessage.SetToolCalls(toolCalls)
+			}
+
+			if len(mediaMessages) > 0 && len(toolCalls) == 0 {
+				openAIMessage.SetMediaContent(mediaMessages)
+			}
+		}
+		if len(openAIMessage.ParseContent()) > 0 || len(openAIMessage.ToolCalls) > 0 {
+			openAIMessages = append(openAIMessages, openAIMessage)
+		}
+	}
+
+	openAIRequest.Messages = openAIMessages
+
+	return &openAIRequest, nil
+}
+
+func generateStopBlock(index int) *dto.ClaudeResponse {
+	return &dto.ClaudeResponse{
+		Type:  "content_block_stop",
+		Index: common.GetPointer[int](index),
+	}
+}
+
+func buildClaudeUsageFromOpenAIUsage(oaiUsage *dto.Usage) *dto.ClaudeUsage {
+	if oaiUsage == nil {
+		return nil
+	}
+	usage := &dto.ClaudeUsage{
+		InputTokens:              oaiUsage.PromptTokens,
+		OutputTokens:             oaiUsage.CompletionTokens,
+		CacheCreationInputTokens: oaiUsage.PromptTokensDetails.CachedCreationTokens,
+		CacheReadInputTokens:     oaiUsage.PromptTokensDetails.CachedTokens,
+	}
+	if oaiUsage.ClaudeCacheCreation5mTokens > 0 || oaiUsage.ClaudeCacheCreation1hTokens > 0 {
+		usage.CacheCreation = &dto.ClaudeCacheCreationUsage{
+			Ephemeral5mInputTokens: oaiUsage.ClaudeCacheCreation5mTokens,
+			Ephemeral1hInputTokens: oaiUsage.ClaudeCacheCreation1hTokens,
+		}
+	}
+	return usage
+}
+
+func StreamResponseOpenAI2Claude(openAIResponse *dto.ChatCompletionsStreamResponse, info *relaycommon.RelayInfo) []*dto.ClaudeResponse {
+	if info.ClaudeConvertInfo.Done {
+		return nil
+	}
+
+	var claudeResponses []*dto.ClaudeResponse
+	// stopOpenBlocks emits the required content_block_stop event(s) for the currently open block(s)
+	// according to Anthropic's SSE streaming state machine:
+	// content_block_start -> content_block_delta* -> content_block_stop (per index).
+	//
+	// For text/thinking, there is at most one open block at info.ClaudeConvertInfo.Index.
+	// For tools, OpenAI tool_calls can stream multiple parallel tool_use blocks (indexed from 0),
+	// so we may have multiple open blocks and must stop each one explicitly.
+	stopOpenBlocks := func() {
+		switch info.ClaudeConvertInfo.LastMessagesType {
+		case relaycommon.LastMessageTypeText, relaycommon.LastMessageTypeThinking:
+			claudeResponses = append(claudeResponses, generateStopBlock(info.ClaudeConvertInfo.Index))
+		case relaycommon.LastMessageTypeTools:
+			base := info.ClaudeConvertInfo.ToolCallBaseIndex
+			for offset := 0; offset <= info.ClaudeConvertInfo.ToolCallMaxIndexOffset; offset++ {
+				claudeResponses = append(claudeResponses, generateStopBlock(base+offset))
+			}
+		}
+	}
+	// stopOpenBlocksAndAdvance closes the currently open block(s) and advances the content block index
+	// to the next available slot for subsequent content_block_start events.
+	//
+	// This prevents invalid streams where a content_block_delta (e.g. thinking_delta) is emitted for an
+	// index whose active content_block type is different (the typical cause of "Mismatched content block type").
+	stopOpenBlocksAndAdvance := func() {
+		if info.ClaudeConvertInfo.LastMessagesType == relaycommon.LastMessageTypeNone {
+			return
+		}
+		stopOpenBlocks()
+		switch info.ClaudeConvertInfo.LastMessagesType {
+		case relaycommon.LastMessageTypeTools:
+			info.ClaudeConvertInfo.Index = info.ClaudeConvertInfo.ToolCallBaseIndex + info.ClaudeConvertInfo.ToolCallMaxIndexOffset + 1
+			info.ClaudeConvertInfo.ToolCallBaseIndex = 0
+			info.ClaudeConvertInfo.ToolCallMaxIndexOffset = 0
+		default:
+			info.ClaudeConvertInfo.Index++
+		}
+		info.ClaudeConvertInfo.LastMessagesType = relaycommon.LastMessageTypeNone
+	}
+	if info.SendResponseCount == 1 {
+		msg := &dto.ClaudeMediaMessage{
+			Id:    openAIResponse.Id,
+			Model: openAIResponse.Model,
+			Type:  "message",
+			Role:  "assistant",
+			Usage: &dto.ClaudeUsage{
+				InputTokens:  info.GetEstimatePromptTokens(),
+				OutputTokens: 0,
+			},
+		}
+		msg.SetContent(make([]any, 0))
+		claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+			Type:    "message_start",
+			Message: msg,
+		})
+		//claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+		//	Type: "ping",
+		//})
+		if openAIResponse.IsToolCall() {
+			info.ClaudeConvertInfo.LastMessagesType = relaycommon.LastMessageTypeTools
+			info.ClaudeConvertInfo.ToolCallBaseIndex = 0
+			info.ClaudeConvertInfo.ToolCallMaxIndexOffset = 0
+			var toolCall dto.ToolCallResponse
+			if len(openAIResponse.Choices) > 0 && len(openAIResponse.Choices[0].Delta.ToolCalls) > 0 {
+				toolCall = openAIResponse.Choices[0].Delta.ToolCalls[0]
+			} else {
+				first := openAIResponse.GetFirstToolCall()
+				if first != nil {
+					toolCall = *first
+				} else {
+					toolCall = dto.ToolCallResponse{}
+				}
+			}
+			resp := &dto.ClaudeResponse{
+				Type: "content_block_start",
+				ContentBlock: &dto.ClaudeMediaMessage{
+					Id:    toolCall.ID,
+					Type:  "tool_use",
+					Name:  toolCall.Function.Name,
+					Input: map[string]interface{}{},
+				},
+			}
+			resp.SetIndex(0)
+			claudeResponses = append(claudeResponses, resp)
+			// 首块包含工具 delta，则追加 input_json_delta
+			if toolCall.Function.Arguments != "" {
+				idx := 0
+				claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+					Index: &idx,
+					Type:  "content_block_delta",
+					Delta: &dto.ClaudeMediaMessage{
+						Type:        "input_json_delta",
+						PartialJson: &toolCall.Function.Arguments,
+					},
+				})
+			}
+		} else {
+
+		}
+		// 判断首个响应是否存在内容（非标准的 OpenAI 响应）
+		if len(openAIResponse.Choices) > 0 {
+			reasoning := openAIResponse.Choices[0].Delta.GetReasoningContent()
+			content := openAIResponse.Choices[0].Delta.GetContentString()
+
+			if reasoning != "" {
+				if info.ClaudeConvertInfo.LastMessagesType != relaycommon.LastMessageTypeThinking {
+					stopOpenBlocksAndAdvance()
+				}
+				idx := info.ClaudeConvertInfo.Index
+				claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+					Index: &idx,
+					Type:  "content_block_start",
+					ContentBlock: &dto.ClaudeMediaMessage{
+						Type:     "thinking",
+						Thinking: common.GetPointer[string](""),
+					},
+				})
+				idx2 := idx
+				claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+					Index: &idx2,
+					Type:  "content_block_delta",
+					Delta: &dto.ClaudeMediaMessage{
+						Type:     "thinking_delta",
+						Thinking: &reasoning,
+					},
+				})
+				info.ClaudeConvertInfo.LastMessagesType = relaycommon.LastMessageTypeThinking
+			} else if content != "" {
+				if info.ClaudeConvertInfo.LastMessagesType != relaycommon.LastMessageTypeText {
+					stopOpenBlocksAndAdvance()
+				}
+				idx := info.ClaudeConvertInfo.Index
+				claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+					Index: &idx,
+					Type:  "content_block_start",
+					ContentBlock: &dto.ClaudeMediaMessage{
+						Type: "text",
+						Text: common.GetPointer[string](""),
+					},
+				})
+				idx2 := idx
+				claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+					Index: &idx2,
+					Type:  "content_block_delta",
+					Delta: &dto.ClaudeMediaMessage{
+						Type: "text_delta",
+						Text: common.GetPointer[string](content),
+					},
+				})
+				info.ClaudeConvertInfo.LastMessagesType = relaycommon.LastMessageTypeText
+			}
+		}
+
+		// 如果首块就带 finish_reason，需要立即发送停止块
+		if len(openAIResponse.Choices) > 0 && openAIResponse.Choices[0].FinishReason != nil && *openAIResponse.Choices[0].FinishReason != "" {
+			info.FinishReason = *openAIResponse.Choices[0].FinishReason
+			stopOpenBlocks()
+			oaiUsage := openAIResponse.Usage
+			if oaiUsage == nil {
+				oaiUsage = info.ClaudeConvertInfo.Usage
+			}
+			if oaiUsage != nil {
+				claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+					Type:  "message_delta",
+					Usage: buildClaudeUsageFromOpenAIUsage(oaiUsage),
+					Delta: &dto.ClaudeMediaMessage{
+						StopReason: common.GetPointer[string](stopReasonOpenAI2Claude(info.FinishReason)),
+					},
+				})
+			}
+			claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+				Type: "message_stop",
+			})
+			info.ClaudeConvertInfo.Done = true
+		}
+		return claudeResponses
+	}
+
+	if len(openAIResponse.Choices) == 0 {
+		// no choices
+		// 可能为非标准的 OpenAI 响应，判断是否已经完成
+		if info.ClaudeConvertInfo.Done {
+			stopOpenBlocks()
+			oaiUsage := info.ClaudeConvertInfo.Usage
+			if oaiUsage != nil {
+				claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+					Type:  "message_delta",
+					Usage: buildClaudeUsageFromOpenAIUsage(oaiUsage),
+					Delta: &dto.ClaudeMediaMessage{
+						StopReason: common.GetPointer[string](stopReasonOpenAI2Claude(info.FinishReason)),
+					},
+				})
+			}
+			claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+				Type: "message_stop",
+			})
+		}
+		return claudeResponses
+	} else {
+		chosenChoice := openAIResponse.Choices[0]
+		doneChunk := chosenChoice.FinishReason != nil && *chosenChoice.FinishReason != ""
+		if doneChunk {
+			info.FinishReason = *chosenChoice.FinishReason
+		}
+
+		var claudeResponse dto.ClaudeResponse
+		var isEmpty bool
+		claudeResponse.Type = "content_block_delta"
+		if len(chosenChoice.Delta.ToolCalls) > 0 {
+			toolCalls := chosenChoice.Delta.ToolCalls
+			if info.ClaudeConvertInfo.LastMessagesType != relaycommon.LastMessageTypeTools {
+				stopOpenBlocksAndAdvance()
+				info.ClaudeConvertInfo.ToolCallBaseIndex = info.ClaudeConvertInfo.Index
+				info.ClaudeConvertInfo.ToolCallMaxIndexOffset = 0
+			}
+			info.ClaudeConvertInfo.LastMessagesType = relaycommon.LastMessageTypeTools
+			base := info.ClaudeConvertInfo.ToolCallBaseIndex
+			maxOffset := info.ClaudeConvertInfo.ToolCallMaxIndexOffset
+
+			for i, toolCall := range toolCalls {
+				offset := 0
+				if toolCall.Index != nil {
+					offset = *toolCall.Index
+				} else {
+					offset = i
+				}
+				if offset > maxOffset {
+					maxOffset = offset
+				}
+				blockIndex := base + offset
+
+				idx := blockIndex
+				if toolCall.Function.Name != "" {
+					claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+						Index: &idx,
+						Type:  "content_block_start",
+						ContentBlock: &dto.ClaudeMediaMessage{
+							Id:    toolCall.ID,
+							Type:  "tool_use",
+							Name:  toolCall.Function.Name,
+							Input: map[string]interface{}{},
+						},
+					})
+				}
+
+				if len(toolCall.Function.Arguments) > 0 {
+					claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+						Index: &idx,
+						Type:  "content_block_delta",
+						Delta: &dto.ClaudeMediaMessage{
+							Type:        "input_json_delta",
+							PartialJson: &toolCall.Function.Arguments,
+						},
+					})
+				}
+			}
+			info.ClaudeConvertInfo.ToolCallMaxIndexOffset = maxOffset
+			info.ClaudeConvertInfo.Index = base + maxOffset
+		} else {
+			reasoning := chosenChoice.Delta.GetReasoningContent()
+			textContent := chosenChoice.Delta.GetContentString()
+			if reasoning != "" || textContent != "" {
+				if reasoning != "" {
+					if info.ClaudeConvertInfo.LastMessagesType != relaycommon.LastMessageTypeThinking {
+						stopOpenBlocksAndAdvance()
+						idx := info.ClaudeConvertInfo.Index
+						claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+							Index: &idx,
+							Type:  "content_block_start",
+							ContentBlock: &dto.ClaudeMediaMessage{
+								Type:     "thinking",
+								Thinking: common.GetPointer[string](""),
+							},
+						})
+					}
+					info.ClaudeConvertInfo.LastMessagesType = relaycommon.LastMessageTypeThinking
+					claudeResponse.Delta = &dto.ClaudeMediaMessage{
+						Type:     "thinking_delta",
+						Thinking: &reasoning,
+					}
+				} else {
+					if info.ClaudeConvertInfo.LastMessagesType != relaycommon.LastMessageTypeText {
+						stopOpenBlocksAndAdvance()
+						idx := info.ClaudeConvertInfo.Index
+						claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+							Index: &idx,
+							Type:  "content_block_start",
+							ContentBlock: &dto.ClaudeMediaMessage{
+								Type: "text",
+								Text: common.GetPointer[string](""),
+							},
+						})
+					}
+					info.ClaudeConvertInfo.LastMessagesType = relaycommon.LastMessageTypeText
+					claudeResponse.Delta = &dto.ClaudeMediaMessage{
+						Type: "text_delta",
+						Text: common.GetPointer[string](textContent),
+					}
+				}
+			} else {
+				isEmpty = true
+			}
+		}
+
+		claudeResponse.Index = common.GetPointer[int](info.ClaudeConvertInfo.Index)
+		if !isEmpty && claudeResponse.Delta != nil {
+			claudeResponses = append(claudeResponses, &claudeResponse)
+		}
+
+		if doneChunk || info.ClaudeConvertInfo.Done {
+			stopOpenBlocks()
+			oaiUsage := openAIResponse.Usage
+			if oaiUsage == nil {
+				oaiUsage = info.ClaudeConvertInfo.Usage
+			}
+			if oaiUsage != nil {
+				claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+					Type:  "message_delta",
+					Usage: buildClaudeUsageFromOpenAIUsage(oaiUsage),
+					Delta: &dto.ClaudeMediaMessage{
+						StopReason: common.GetPointer[string](stopReasonOpenAI2Claude(info.FinishReason)),
+					},
+				})
+			}
+			claudeResponses = append(claudeResponses, &dto.ClaudeResponse{
+				Type: "message_stop",
+			})
+			info.ClaudeConvertInfo.Done = true
+			return claudeResponses
+		}
+	}
+
+	return claudeResponses
+}
+
+func ResponseOpenAI2Claude(openAIResponse *dto.OpenAITextResponse, info *relaycommon.RelayInfo) *dto.ClaudeResponse {
+	var stopReason string
+	contents := make([]dto.ClaudeMediaMessage, 0)
+	claudeResponse := &dto.ClaudeResponse{
+		Id:    openAIResponse.Id,
+		Type:  "message",
+		Role:  "assistant",
+		Model: openAIResponse.Model,
+	}
+	for _, choice := range openAIResponse.Choices {
+		stopReason = stopReasonOpenAI2Claude(choice.FinishReason)
+		if choice.FinishReason == "tool_calls" {
+			for _, toolUse := range choice.Message.ParseToolCalls() {
+				claudeContent := dto.ClaudeMediaMessage{}
+				claudeContent.Type = "tool_use"
+				claudeContent.Id = toolUse.ID
+				claudeContent.Name = toolUse.Function.Name
+				var mapParams map[string]interface{}
+				if err := common.Unmarshal([]byte(toolUse.Function.Arguments), &mapParams); err == nil {
+					claudeContent.Input = mapParams
+				} else {
+					claudeContent.Input = toolUse.Function.Arguments
+				}
+				contents = append(contents, claudeContent)
+			}
+		} else {
+			claudeContent := dto.ClaudeMediaMessage{}
+			claudeContent.Type = "text"
+			claudeContent.SetText(choice.Message.StringContent())
+			contents = append(contents, claudeContent)
+		}
+	}
+	claudeResponse.Content = contents
+	claudeResponse.StopReason = stopReason
+	claudeResponse.Usage = &dto.ClaudeUsage{
+		InputTokens:  openAIResponse.PromptTokens,
+		OutputTokens: openAIResponse.CompletionTokens,
+	}
+
+	return claudeResponse
+}
+
+func stopReasonOpenAI2Claude(reason string) string {
+	return reasonmap.OpenAIFinishReasonToClaudeStopReason(reason)
+}
+
+func toJSONString(v interface{}) string {
+	b, err := json.Marshal(v)
+	if err != nil {
+		return "{}"
+	}
+	return string(b)
+}
+
+func GeminiToOpenAIRequest(geminiRequest *dto.GeminiChatRequest, info *relaycommon.RelayInfo) (*dto.GeneralOpenAIRequest, error) {
+	openaiRequest := &dto.GeneralOpenAIRequest{
+		Model:  info.UpstreamModelName,
+		Stream: lo.ToPtr(info.IsStream),
+	}
+
+	// 转换 messages
+	var messages []dto.Message
+	for _, content := range geminiRequest.Contents {
+		message := dto.Message{
+			Role: convertGeminiRoleToOpenAI(content.Role),
+		}
+
+		// 处理 parts
+		var mediaContents []dto.MediaContent
+		var toolCalls []dto.ToolCallRequest
+		for _, part := range content.Parts {
+			if part.Text != "" {
+				mediaContent := dto.MediaContent{
+					Type: "text",
+					Text: part.Text,
+				}
+				mediaContents = append(mediaContents, mediaContent)
+			} else if part.InlineData != nil {
+				mediaContent := dto.MediaContent{
+					Type: "image_url",
+					ImageUrl: &dto.MessageImageUrl{
+						Url:      fmt.Sprintf("data:%s;base64,%s", part.InlineData.MimeType, part.InlineData.Data),
+						Detail:   "auto",
+						MimeType: part.InlineData.MimeType,
+					},
+				}
+				mediaContents = append(mediaContents, mediaContent)
+			} else if part.FileData != nil {
+				mediaContent := dto.MediaContent{
+					Type: "image_url",
+					ImageUrl: &dto.MessageImageUrl{
+						Url:      part.FileData.FileUri,
+						Detail:   "auto",
+						MimeType: part.FileData.MimeType,
+					},
+				}
+				mediaContents = append(mediaContents, mediaContent)
+			} else if part.FunctionCall != nil {
+				// 处理 Gemini 的工具调用
+				toolCall := dto.ToolCallRequest{
+					ID:   fmt.Sprintf("call_%d", len(toolCalls)+1), // 生成唯一ID
+					Type: "function",
+					Function: dto.FunctionRequest{
+						Name:      part.FunctionCall.FunctionName,
+						Arguments: toJSONString(part.FunctionCall.Arguments),
+					},
+				}
+				toolCalls = append(toolCalls, toolCall)
+			} else if part.FunctionResponse != nil {
+				// 处理 Gemini 的工具响应，创建单独的 tool 消息
+				toolMessage := dto.Message{
+					Role:       "tool",
+					ToolCallId: fmt.Sprintf("call_%d", len(toolCalls)), // 使用对应的调用ID
+				}
+				toolMessage.SetStringContent(toJSONString(part.FunctionResponse.Response))
+				messages = append(messages, toolMessage)
+			}
+		}
+
+		// 设置消息内容
+		if len(toolCalls) > 0 {
+			// 如果有工具调用，设置工具调用
+			message.SetToolCalls(toolCalls)
+		} else if len(mediaContents) == 1 && mediaContents[0].Type == "text" {
+			// 如果只有一个文本内容，直接设置字符串
+			message.Content = mediaContents[0].Text
+		} else if len(mediaContents) > 0 {
+			// 如果有多个内容或包含媒体，设置为数组
+			message.SetMediaContent(mediaContents)
+		}
+
+		// 只有当消息有内容或工具调用时才添加
+		if len(message.ParseContent()) > 0 || len(message.ToolCalls) > 0 {
+			messages = append(messages, message)
+		}
+	}
+
+	openaiRequest.Messages = messages
+
+	if geminiRequest.GenerationConfig.Temperature != nil {
+		openaiRequest.Temperature = geminiRequest.GenerationConfig.Temperature
+	}
+	if geminiRequest.GenerationConfig.TopP != nil && *geminiRequest.GenerationConfig.TopP > 0 {
+		openaiRequest.TopP = lo.ToPtr(*geminiRequest.GenerationConfig.TopP)
+	}
+	if geminiRequest.GenerationConfig.TopK != nil && *geminiRequest.GenerationConfig.TopK > 0 {
+		openaiRequest.TopK = lo.ToPtr(int(*geminiRequest.GenerationConfig.TopK))
+	}
+	if geminiRequest.GenerationConfig.MaxOutputTokens != nil && *geminiRequest.GenerationConfig.MaxOutputTokens > 0 {
+		openaiRequest.MaxTokens = lo.ToPtr(*geminiRequest.GenerationConfig.MaxOutputTokens)
+	}
+	// gemini stop sequences 最多 5 个，openai stop 最多 4 个
+	if len(geminiRequest.GenerationConfig.StopSequences) > 0 {
+		openaiRequest.Stop = geminiRequest.GenerationConfig.StopSequences[:4]
+	}
+	if geminiRequest.GenerationConfig.CandidateCount != nil && *geminiRequest.GenerationConfig.CandidateCount > 0 {
+		openaiRequest.N = lo.ToPtr(*geminiRequest.GenerationConfig.CandidateCount)
+	}
+
+	// 转换工具调用
+	if len(geminiRequest.GetTools()) > 0 {
+		var tools []dto.ToolCallRequest
+		for _, tool := range geminiRequest.GetTools() {
+			if tool.FunctionDeclarations != nil {
+				functionDeclarations, err := common.Any2Type[[]dto.FunctionRequest](tool.FunctionDeclarations)
+				if err != nil {
+					common.SysError(fmt.Sprintf("failed to parse gemini function declarations: %v (type=%T)", err, tool.FunctionDeclarations))
+					continue
+				}
+				for _, function := range functionDeclarations {
+					openAITool := dto.ToolCallRequest{
+						Type: "function",
+						Function: dto.FunctionRequest{
+							Name:        function.Name,
+							Description: function.Description,
+							Parameters:  function.Parameters,
+						},
+					}
+					tools = append(tools, openAITool)
+				}
+			}
+		}
+		if len(tools) > 0 {
+			openaiRequest.Tools = tools
+		}
+	}
+
+	// gemini system instructions
+	if geminiRequest.SystemInstructions != nil {
+		// 将系统指令作为第一条消息插入
+		systemMessage := dto.Message{
+			Role:    "system",
+			Content: extractTextFromGeminiParts(geminiRequest.SystemInstructions.Parts),
+		}
+		openaiRequest.Messages = append([]dto.Message{systemMessage}, openaiRequest.Messages...)
+	}
+
+	return openaiRequest, nil
+}
+
+func convertGeminiRoleToOpenAI(geminiRole string) string {
+	switch geminiRole {
+	case "user":
+		return "user"
+	case "model":
+		return "assistant"
+	case "function":
+		return "function"
+	default:
+		return "user"
+	}
+}
+
+func extractTextFromGeminiParts(parts []dto.GeminiPart) string {
+	var texts []string
+	for _, part := range parts {
+		if part.Text != "" {
+			texts = append(texts, part.Text)
+		}
+	}
+	return strings.Join(texts, "\n")
+}
+
+// ResponseOpenAI2Gemini 将 OpenAI 响应转换为 Gemini 格式
+func ResponseOpenAI2Gemini(openAIResponse *dto.OpenAITextResponse, info *relaycommon.RelayInfo) *dto.GeminiChatResponse {
+	geminiResponse := &dto.GeminiChatResponse{
+		Candidates: make([]dto.GeminiChatCandidate, 0, len(openAIResponse.Choices)),
+		UsageMetadata: dto.GeminiUsageMetadata{
+			PromptTokenCount:     openAIResponse.PromptTokens,
+			CandidatesTokenCount: openAIResponse.CompletionTokens,
+			TotalTokenCount:      openAIResponse.PromptTokens + openAIResponse.CompletionTokens,
+		},
+	}
+
+	for _, choice := range openAIResponse.Choices {
+		candidate := dto.GeminiChatCandidate{
+			Index:         int64(choice.Index),
+			SafetyRatings: []dto.GeminiChatSafetyRating{},
+		}
+
+		// 设置结束原因
+		var finishReason string
+		switch choice.FinishReason {
+		case "stop":
+			finishReason = "STOP"
+		case "length":
+			finishReason = "MAX_TOKENS"
+		case "content_filter":
+			finishReason = "SAFETY"
+		case "tool_calls":
+			finishReason = "STOP"
+		default:
+			finishReason = "STOP"
+		}
+		candidate.FinishReason = &finishReason
+
+		// 转换消息内容
+		content := dto.GeminiChatContent{
+			Role:  "model",
+			Parts: make([]dto.GeminiPart, 0),
+		}
+
+		// 处理工具调用
+		toolCalls := choice.Message.ParseToolCalls()
+		if len(toolCalls) > 0 {
+			for _, toolCall := range toolCalls {
+				// 解析参数
+				var args map[string]interface{}
+				if toolCall.Function.Arguments != "" {
+					if err := json.Unmarshal([]byte(toolCall.Function.Arguments), &args); err != nil {
+						args = map[string]interface{}{"arguments": toolCall.Function.Arguments}
+					}
+				} else {
+					args = make(map[string]interface{})
+				}
+
+				part := dto.GeminiPart{
+					FunctionCall: &dto.FunctionCall{
+						FunctionName: toolCall.Function.Name,
+						Arguments:    args,
+					},
+				}
+				content.Parts = append(content.Parts, part)
+			}
+		} else {
+			// 处理文本内容
+			textContent := choice.Message.StringContent()
+			if textContent != "" {
+				part := dto.GeminiPart{
+					Text: textContent,
+				}
+				content.Parts = append(content.Parts, part)
+			}
+		}
+
+		candidate.Content = content
+		geminiResponse.Candidates = append(geminiResponse.Candidates, candidate)
+	}
+
+	return geminiResponse
+}
+
+// StreamResponseOpenAI2Gemini 将 OpenAI 流式响应转换为 Gemini 格式
+func StreamResponseOpenAI2Gemini(openAIResponse *dto.ChatCompletionsStreamResponse, info *relaycommon.RelayInfo) *dto.GeminiChatResponse {
+	// 检查是否有实际内容或结束标志
+	hasContent := false
+	hasFinishReason := false
+	for _, choice := range openAIResponse.Choices {
+		if len(choice.Delta.GetContentString()) > 0 || (choice.Delta.ToolCalls != nil && len(choice.Delta.ToolCalls) > 0) {
+			hasContent = true
+		}
+		if choice.FinishReason != nil {
+			hasFinishReason = true
+		}
+	}
+
+	// 如果没有实际内容且没有结束标志，跳过。主要针对 openai 流响应开头的空数据
+	if !hasContent && !hasFinishReason {
+		return nil
+	}
+
+	geminiResponse := &dto.GeminiChatResponse{
+		Candidates: make([]dto.GeminiChatCandidate, 0, len(openAIResponse.Choices)),
+		UsageMetadata: dto.GeminiUsageMetadata{
+			PromptTokenCount:     info.GetEstimatePromptTokens(),
+			CandidatesTokenCount: 0, // 流式响应中可能没有完整的 usage 信息
+			TotalTokenCount:      info.GetEstimatePromptTokens(),
+		},
+	}
+
+	if openAIResponse.Usage != nil {
+		geminiResponse.UsageMetadata.PromptTokenCount = openAIResponse.Usage.PromptTokens
+		geminiResponse.UsageMetadata.CandidatesTokenCount = openAIResponse.Usage.CompletionTokens
+		geminiResponse.UsageMetadata.TotalTokenCount = openAIResponse.Usage.TotalTokens
+	}
+
+	for _, choice := range openAIResponse.Choices {
+		candidate := dto.GeminiChatCandidate{
+			Index:         int64(choice.Index),
+			SafetyRatings: []dto.GeminiChatSafetyRating{},
+		}
+
+		// 设置结束原因
+		if choice.FinishReason != nil {
+			var finishReason string
+			switch *choice.FinishReason {
+			case "stop":
+				finishReason = "STOP"
+			case "length":
+				finishReason = "MAX_TOKENS"
+			case "content_filter":
+				finishReason = "SAFETY"
+			case "tool_calls":
+				finishReason = "STOP"
+			default:
+				finishReason = "STOP"
+			}
+			candidate.FinishReason = &finishReason
+		}
+
+		// 转换消息内容
+		content := dto.GeminiChatContent{
+			Role:  "model",
+			Parts: make([]dto.GeminiPart, 0),
+		}
+
+		// 处理工具调用
+		if choice.Delta.ToolCalls != nil {
+			for _, toolCall := range choice.Delta.ToolCalls {
+				// 解析参数
+				var args map[string]interface{}
+				if toolCall.Function.Arguments != "" {
+					if err := json.Unmarshal([]byte(toolCall.Function.Arguments), &args); err != nil {
+						args = map[string]interface{}{"arguments": toolCall.Function.Arguments}
+					}
+				} else {
+					args = make(map[string]interface{})
+				}
+
+				part := dto.GeminiPart{
+					FunctionCall: &dto.FunctionCall{
+						FunctionName: toolCall.Function.Name,
+						Arguments:    args,
+					},
+				}
+				content.Parts = append(content.Parts, part)
+			}
+		} else {
+			// 处理文本内容
+			textContent := choice.Delta.GetContentString()
+			if textContent != "" {
+				part := dto.GeminiPart{
+					Text: textContent,
+				}
+				content.Parts = append(content.Parts, part)
+			}
+		}
+
+		candidate.Content = content
+		geminiResponse.Candidates = append(geminiResponse.Candidates, candidate)
+	}
+
+	return geminiResponse
+}

service/download.go

@@ -0,0 +1,70 @@
+package service
+
+import (
+	"bytes"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"strings"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/setting/system_setting"
+)
+
+// WorkerRequest Worker请求的数据结构
+type WorkerRequest struct {
+	URL     string            `json:"url"`
+	Key     string            `json:"key"`
+	Method  string            `json:"method,omitempty"`
+	Headers map[string]string `json:"headers,omitempty"`
+	Body    json.RawMessage   `json:"body,omitempty"`
+}
+
+// DoWorkerRequest 通过Worker发送请求
+func DoWorkerRequest(req *WorkerRequest) (*http.Response, error) {
+	if !system_setting.EnableWorker() {
+		return nil, fmt.Errorf("worker not enabled")
+	}
+	if !system_setting.WorkerAllowHttpImageRequestEnabled && !strings.HasPrefix(req.URL, "https") {
+		return nil, fmt.Errorf("only support https url")
+	}
+
+	// SSRF防护：验证请求URL
+	fetchSetting := system_setting.GetFetchSetting()
+	if err := common.ValidateURLWithFetchSetting(req.URL, fetchSetting.EnableSSRFProtection, fetchSetting.AllowPrivateIp, fetchSetting.DomainFilterMode, fetchSetting.IpFilterMode, fetchSetting.DomainList, fetchSetting.IpList, fetchSetting.AllowedPorts, fetchSetting.ApplyIPFilterForDomain); err != nil {
+		return nil, fmt.Errorf("request reject: %v", err)
+	}
+
+	workerUrl := system_setting.WorkerUrl
+	if !strings.HasSuffix(workerUrl, "/") {
+		workerUrl += "/"
+	}
+
+	// 序列化worker请求数据
+	workerPayload, err := json.Marshal(req)
+	if err != nil {
+		return nil, fmt.Errorf("failed to marshal worker payload: %v", err)
+	}
+
+	return GetHttpClient().Post(workerUrl, "application/json", bytes.NewBuffer(workerPayload))
+}
+
+func DoDownloadRequest(originUrl string, reason ...string) (resp *http.Response, err error) {
+	if system_setting.EnableWorker() {
+		common.SysLog(fmt.Sprintf("downloading file from worker: %s, reason: %s", originUrl, strings.Join(reason, ", ")))
+		req := &WorkerRequest{
+			URL: originUrl,
+			Key: system_setting.WorkerValidKey,
+		}
+		return DoWorkerRequest(req)
+	} else {
+		// SSRF防护：验证请求URL（非Worker模式）
+		fetchSetting := system_setting.GetFetchSetting()
+		if err := common.ValidateURLWithFetchSetting(originUrl, fetchSetting.EnableSSRFProtection, fetchSetting.AllowPrivateIp, fetchSetting.DomainFilterMode, fetchSetting.IpFilterMode, fetchSetting.DomainList, fetchSetting.IpList, fetchSetting.AllowedPorts, fetchSetting.ApplyIPFilterForDomain); err != nil {
+			return nil, fmt.Errorf("request reject: %v", err)
+		}
+
+		common.SysLog(fmt.Sprintf("downloading from origin: %s, reason: %s", common.MaskSensitiveInfo(originUrl), strings.Join(reason, ", ")))
+		return GetHttpClient().Get(originUrl)
+	}
+}

service/epay.go

@@ -0,0 +1,13 @@
+package service
+
+import (
+	"github.com/QuantumNous/new-api/setting/operation_setting"
+	"github.com/QuantumNous/new-api/setting/system_setting"
+)
+
+func GetCallbackAddress() string {
+	if operation_setting.CustomCallbackAddress == "" {
+		return system_setting.ServerAddress
+	}
+	return operation_setting.CustomCallbackAddress
+}

service/error.go

@@ -0,0 +1,221 @@
+package service
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"math"
+	"net/http"
+	"strconv"
+	"strings"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/dto"
+	"github.com/QuantumNous/new-api/logger"
+	"github.com/QuantumNous/new-api/types"
+)
+
+func MidjourneyErrorWrapper(code int, desc string) *dto.MidjourneyResponse {
+	return &dto.MidjourneyResponse{
+		Code:        code,
+		Description: desc,
+	}
+}
+
+func MidjourneyErrorWithStatusCodeWrapper(code int, desc string, statusCode int) *dto.MidjourneyResponseWithStatusCode {
+	return &dto.MidjourneyResponseWithStatusCode{
+		StatusCode: statusCode,
+		Response:   *MidjourneyErrorWrapper(code, desc),
+	}
+}
+
+//// OpenAIErrorWrapper wraps an error into an OpenAIErrorWithStatusCode
+//func OpenAIErrorWrapper(err error, code string, statusCode int) *dto.OpenAIErrorWithStatusCode {
+//	text := err.Error()
+//	lowerText := strings.ToLower(text)
+//	if !strings.HasPrefix(lowerText, "get file base64 from url") && !strings.HasPrefix(lowerText, "mime type is not supported") {
+//		if strings.Contains(lowerText, "post") || strings.Contains(lowerText, "dial") || strings.Contains(lowerText, "http") {
+//			common.SysLog(fmt.Sprintf("error: %s", text))
+//			text = "请求上游地址失败"
+//		}
+//	}
+//	openAIError := dto.OpenAIError{
+//		Message: text,
+//		Type:    "new_api_error",
+//		Code:    code,
+//	}
+//	return &dto.OpenAIErrorWithStatusCode{
+//		Error:      openAIError,
+//		StatusCode: statusCode,
+//	}
+//}
+//
+//func OpenAIErrorWrapperLocal(err error, code string, statusCode int) *dto.OpenAIErrorWithStatusCode {
+//	openaiErr := OpenAIErrorWrapper(err, code, statusCode)
+//	openaiErr.LocalError = true
+//	return openaiErr
+//}
+
+func ClaudeErrorWrapper(err error, code string, statusCode int) *dto.ClaudeErrorWithStatusCode {
+	text := err.Error()
+	lowerText := strings.ToLower(text)
+	if !strings.HasPrefix(lowerText, "get file base64 from url") {
+		if strings.Contains(lowerText, "post") || strings.Contains(lowerText, "dial") || strings.Contains(lowerText, "http") {
+			common.SysLog(fmt.Sprintf("error: %s", text))
+			text = "请求上游地址失败"
+		}
+	}
+	claudeError := types.ClaudeError{
+		Message: text,
+		Type:    "new_api_error",
+	}
+	return &dto.ClaudeErrorWithStatusCode{
+		Error:      claudeError,
+		StatusCode: statusCode,
+	}
+}
+
+func ClaudeErrorWrapperLocal(err error, code string, statusCode int) *dto.ClaudeErrorWithStatusCode {
+	claudeErr := ClaudeErrorWrapper(err, code, statusCode)
+	claudeErr.LocalError = true
+	return claudeErr
+}
+
+func RelayErrorHandler(ctx context.Context, resp *http.Response, showBodyWhenFail bool) (newApiErr *types.NewAPIError) {
+	newApiErr = types.InitOpenAIError(types.ErrorCodeBadResponseStatusCode, resp.StatusCode)
+
+	responseBody, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return
+	}
+	CloseResponseBodyGracefully(resp)
+	var errResponse dto.GeneralErrorResponse
+	buildErrWithBody := func(message string) error {
+		if message == "" {
+			return fmt.Errorf("bad response status code %d, body: %s", resp.StatusCode, string(responseBody))
+		}
+		return fmt.Errorf("bad response status code %d, message: %s, body: %s", resp.StatusCode, message, string(responseBody))
+	}
+
+	err = common.Unmarshal(responseBody, &errResponse)
+	if err != nil {
+		if showBodyWhenFail {
+			newApiErr.Err = buildErrWithBody("")
+		} else {
+			logger.LogError(ctx, fmt.Sprintf("bad response status code %d, body: %s", resp.StatusCode, string(responseBody)))
+			newApiErr.Err = fmt.Errorf("bad response status code %d", resp.StatusCode)
+		}
+		return
+	}
+
+	if common.GetJsonType(errResponse.Error) == "object" {
+		// General format error (OpenAI, Anthropic, Gemini, etc.)
+		oaiError := errResponse.TryToOpenAIError()
+		if oaiError != nil {
+			newApiErr = types.WithOpenAIError(*oaiError, resp.StatusCode)
+			if showBodyWhenFail {
+				newApiErr.Err = buildErrWithBody(newApiErr.Error())
+			}
+			return
+		}
+	}
+	newApiErr = types.NewOpenAIError(errors.New(errResponse.ToMessage()), types.ErrorCodeBadResponseStatusCode, resp.StatusCode)
+	if showBodyWhenFail {
+		newApiErr.Err = buildErrWithBody(newApiErr.Error())
+	}
+	return
+}
+
+func ResetStatusCode(newApiErr *types.NewAPIError, statusCodeMappingStr string) {
+	if newApiErr == nil {
+		return
+	}
+	if statusCodeMappingStr == "" || statusCodeMappingStr == "{}" {
+		return
+	}
+	statusCodeMapping := make(map[string]any)
+	err := common.Unmarshal([]byte(statusCodeMappingStr), &statusCodeMapping)
+	if err != nil {
+		return
+	}
+	if newApiErr.StatusCode == http.StatusOK {
+		return
+	}
+	codeStr := strconv.Itoa(newApiErr.StatusCode)
+	if value, ok := statusCodeMapping[codeStr]; ok {
+		intCode, ok := parseStatusCodeMappingValue(value)
+		if !ok {
+			return
+		}
+		newApiErr.StatusCode = intCode
+	}
+}
+
+func parseStatusCodeMappingValue(value any) (int, bool) {
+	switch v := value.(type) {
+	case string:
+		if v == "" {
+			return 0, false
+		}
+		statusCode, err := strconv.Atoi(v)
+		if err != nil {
+			return 0, false
+		}
+		return statusCode, true
+	case float64:
+		if v != math.Trunc(v) {
+			return 0, false
+		}
+		return int(v), true
+	case int:
+		return v, true
+	case json.Number:
+		statusCode, err := strconv.Atoi(v.String())
+		if err != nil {
+			return 0, false
+		}
+		return statusCode, true
+	default:
+		return 0, false
+	}
+}
+
+func TaskErrorWrapperLocal(err error, code string, statusCode int) *dto.TaskError {
+	openaiErr := TaskErrorWrapper(err, code, statusCode)
+	openaiErr.LocalError = true
+	return openaiErr
+}
+
+func TaskErrorWrapper(err error, code string, statusCode int) *dto.TaskError {
+	text := err.Error()
+	lowerText := strings.ToLower(text)
+	if strings.Contains(lowerText, "post") || strings.Contains(lowerText, "dial") || strings.Contains(lowerText, "http") {
+		common.SysLog(fmt.Sprintf("error: %s", text))
+		//text = "请求上游地址失败"
+		text = common.MaskSensitiveInfo(text)
+	}
+	//避免暴露内部错误
+	taskError := &dto.TaskError{
+		Code:       code,
+		Message:    text,
+		StatusCode: statusCode,
+		Error:      err,
+	}
+
+	return taskError
+}
+
+// TaskErrorFromAPIError 将 PreConsumeBilling 返回的 NewAPIError 转换为 TaskError。
+func TaskErrorFromAPIError(apiErr *types.NewAPIError) *dto.TaskError {
+	if apiErr == nil {
+		return nil
+	}
+	return &dto.TaskError{
+		Code:       string(apiErr.GetErrorCode()),
+		Message:    apiErr.Err.Error(),
+		StatusCode: apiErr.StatusCode,
+		Error:      apiErr.Err,
+	}
+}

service/error_test.go

@@ -0,0 +1,57 @@
+package service
+
+import (
+	"testing"
+
+	"github.com/QuantumNous/new-api/types"
+	"github.com/stretchr/testify/require"
+)
+
+func TestResetStatusCode(t *testing.T) {
+	t.Parallel()
+
+	testCases := []struct {
+		name             string
+		statusCode       int
+		statusCodeConfig string
+		expectedCode     int
+	}{
+		{
+			name:             "map string value",
+			statusCode:       429,
+			statusCodeConfig: `{"429":"503"}`,
+			expectedCode:     503,
+		},
+		{
+			name:             "map int value",
+			statusCode:       429,
+			statusCodeConfig: `{"429":503}`,
+			expectedCode:     503,
+		},
+		{
+			name:             "skip invalid string value",
+			statusCode:       429,
+			statusCodeConfig: `{"429":"bad-code"}`,
+			expectedCode:     429,
+		},
+		{
+			name:             "skip status code 200",
+			statusCode:       200,
+			statusCodeConfig: `{"200":503}`,
+			expectedCode:     200,
+		},
+	}
+
+	for _, tc := range testCases {
+		tc := tc
+		t.Run(tc.name, func(t *testing.T) {
+			t.Parallel()
+
+			newAPIError := &types.NewAPIError{
+				StatusCode: tc.statusCode,
+			}
+			ResetStatusCode(newAPIError, tc.statusCodeConfig)
+			require.Equal(t, tc.expectedCode, newAPIError.StatusCode)
+		})
+	}
+}

service/file_decoder.go

@@ -0,0 +1,203 @@
+package service
+
+import (
+	"bytes"
+	"fmt"
+	"image"
+	_ "image/gif"
+	_ "image/jpeg"
+	_ "image/png"
+	"io"
+	"net/http"
+	"strings"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/logger"
+	"github.com/QuantumNous/new-api/types"
+
+	"github.com/gin-gonic/gin"
+)
+
+// GetFileTypeFromUrl 获取文件类型，返回 mime type， 例如 image/jpeg, image/png, image/gif, image/bmp, image/tiff, application/pdf
+// 如果获取失败，返回 application/octet-stream
+func GetFileTypeFromUrl(c *gin.Context, url string, reason ...string) (string, error) {
+	response, err := DoDownloadRequest(url, []string{"get_mime_type", strings.Join(reason, ", ")}...)
+	if err != nil {
+		common.SysLog(fmt.Sprintf("fail to get file type from url: %s, error: %s", url, err.Error()))
+		return "", err
+	}
+	defer response.Body.Close()
+
+	if response.StatusCode != 200 {
+		logger.LogError(c, fmt.Sprintf("failed to download file from %s, status code: %d", url, response.StatusCode))
+		return "", fmt.Errorf("failed to download file, status code: %d", response.StatusCode)
+	}
+
+	if headerType := strings.TrimSpace(response.Header.Get("Content-Type")); headerType != "" {
+		if i := strings.Index(headerType, ";"); i != -1 {
+			headerType = headerType[:i]
+		}
+		if headerType != "application/octet-stream" {
+			return headerType, nil
+		}
+	}
+
+	if cd := response.Header.Get("Content-Disposition"); cd != "" {
+		parts := strings.Split(cd, ";")
+		for _, part := range parts {
+			part = strings.TrimSpace(part)
+			if strings.HasPrefix(strings.ToLower(part), "filename=") {
+				name := strings.TrimSpace(strings.TrimPrefix(part, "filename="))
+				if len(name) > 2 && name[0] == '"' && name[len(name)-1] == '"' {
+					name = name[1 : len(name)-1]
+				}
+				if dot := strings.LastIndex(name, "."); dot != -1 && dot+1 < len(name) {
+					ext := strings.ToLower(name[dot+1:])
+					if ext != "" {
+						mt := GetMimeTypeByExtension(ext)
+						if mt != "application/octet-stream" {
+							return mt, nil
+						}
+					}
+				}
+				break
+			}
+		}
+	}
+
+	cleanedURL := url
+	if q := strings.Index(cleanedURL, "?"); q != -1 {
+		cleanedURL = cleanedURL[:q]
+	}
+	if slash := strings.LastIndex(cleanedURL, "/"); slash != -1 && slash+1 < len(cleanedURL) {
+		last := cleanedURL[slash+1:]
+		if dot := strings.LastIndex(last, "."); dot != -1 && dot+1 < len(last) {
+			ext := strings.ToLower(last[dot+1:])
+			if ext != "" {
+				mt := GetMimeTypeByExtension(ext)
+				if mt != "application/octet-stream" {
+					return mt, nil
+				}
+			}
+		}
+	}
+
+	var readData []byte
+	limits := []int{512, 8 * 1024, 24 * 1024, 64 * 1024}
+	for _, limit := range limits {
+		logger.LogDebug(c, fmt.Sprintf("Trying to read %d bytes to determine file type", limit))
+		if len(readData) < limit {
+			need := limit - len(readData)
+			tmp := make([]byte, need)
+			n, _ := io.ReadFull(response.Body, tmp)
+			if n > 0 {
+				readData = append(readData, tmp[:n]...)
+			}
+		}
+
+		if len(readData) == 0 {
+			continue
+		}
+
+		sniffed := http.DetectContentType(readData)
+		if sniffed != "" && sniffed != "application/octet-stream" {
+			return sniffed, nil
+		}
+
+		if _, format, err := image.DecodeConfig(bytes.NewReader(readData)); err == nil {
+			switch strings.ToLower(format) {
+			case "jpeg", "jpg":
+				return "image/jpeg", nil
+			case "png":
+				return "image/png", nil
+			case "gif":
+				return "image/gif", nil
+			case "bmp":
+				return "image/bmp", nil
+			case "tiff":
+				return "image/tiff", nil
+			default:
+				if format != "" {
+					return "image/" + strings.ToLower(format), nil
+				}
+			}
+		}
+	}
+
+	// Fallback
+	return "application/octet-stream", nil
+}
+
+// GetFileBase64FromUrl 从 URL 获取文件的 base64 编码数据
+// Deprecated: 请使用 GetBase64Data 配合 types.NewURLFileSource 替代
+// 此函数保留用于向后兼容，内部已重构为调用统一的文件服务
+func GetFileBase64FromUrl(c *gin.Context, url string, reason ...string) (*types.LocalFileData, error) {
+	source := types.NewURLFileSource(url)
+	cachedData, err := LoadFileSource(c, source, reason...)
+	if err != nil {
+		return nil, err
+	}
+
+	// 转换为旧的 LocalFileData 格式以保持兼容
+	base64Data, err := cachedData.GetBase64Data()
+	if err != nil {
+		return nil, err
+	}
+	return &types.LocalFileData{
+		Base64Data: base64Data,
+		MimeType:   cachedData.MimeType,
+		Size:       cachedData.Size,
+		Url:        url,
+	}, nil
+}
+
+func GetMimeTypeByExtension(ext string) string {
+	// Convert to lowercase for case-insensitive comparison
+	ext = strings.ToLower(ext)
+	switch ext {
+	// Text files
+	case "txt", "md", "markdown", "csv", "json", "xml", "html", "htm":
+		return "text/plain"
+
+	// Image files
+	case "jpg", "jpeg":
+		return "image/jpeg"
+	case "png":
+		return "image/png"
+	case "gif":
+		return "image/gif"
+	case "jfif":
+		return "image/jpeg"
+
+	// Audio files
+	case "mp3":
+		return "audio/mp3"
+	case "wav":
+		return "audio/wav"
+	case "mpeg":
+		return "audio/mpeg"
+
+	// Video files
+	case "mp4":
+		return "video/mp4"
+	case "wmv":
+		return "video/wmv"
+	case "flv":
+		return "video/flv"
+	case "mov":
+		return "video/mov"
+	case "mpg":
+		return "video/mpg"
+	case "avi":
+		return "video/avi"
+	case "mpegps":
+		return "video/mpegps"
+
+	// Document files
+	case "pdf":
+		return "application/pdf"
+
+	default:
+		return "application/octet-stream" // Default for unknown types
+	}
+}

service/file_service.go

@@ -0,0 +1,471 @@
+package service
+
+import (
+	"bytes"
+	"encoding/base64"
+	"fmt"
+	"image"
+	_ "image/gif"
+	_ "image/jpeg"
+	_ "image/png"
+	"io"
+	"net/http"
+	"strings"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/constant"
+	"github.com/QuantumNous/new-api/logger"
+	"github.com/QuantumNous/new-api/types"
+
+	"github.com/gin-gonic/gin"
+	"golang.org/x/image/webp"
+)
+
+// FileService 统一的文件处理服务
+// 提供文件下载、解码、缓存等功能的统一入口
+
+// getContextCacheKey 生成 context 缓存的 key
+func getContextCacheKey(url string) string {
+	return fmt.Sprintf("file_cache_%s", common.GenerateHMAC(url))
+}
+
+// LoadFileSource 加载文件源数据
+// 这是统一的入口，会自动处理缓存和不同的来源类型
+func LoadFileSource(c *gin.Context, source *types.FileSource, reason ...string) (*types.CachedFileData, error) {
+	if source == nil {
+		return nil, fmt.Errorf("file source is nil")
+	}
+
+	if common.DebugEnabled {
+		logger.LogDebug(c, fmt.Sprintf("LoadFileSource starting for: %s", source.GetIdentifier()))
+	}
+
+	// 1. 快速检查内部缓存
+	if source.HasCache() {
+		// 即使命中内部缓存，也要确保注册到清理列表（如果尚未注册）
+		if c != nil {
+			registerSourceForCleanup(c, source)
+		}
+		return source.GetCache(), nil
+	}
+
+	// 2. 加锁保护加载过程
+	source.Mu().Lock()
+	defer source.Mu().Unlock()
+
+	// 3. 双重检查
+	if source.HasCache() {
+		if c != nil {
+			registerSourceForCleanup(c, source)
+		}
+		return source.GetCache(), nil
+	}
+
+	// 4. 如果是 URL，检查 Context 缓存
+	var contextKey string
+	if source.IsURL() && c != nil {
+		contextKey = getContextCacheKey(source.URL)
+		if cachedData, exists := c.Get(contextKey); exists {
+			data := cachedData.(*types.CachedFileData)
+			source.SetCache(data)
+			registerSourceForCleanup(c, source)
+			return data, nil
+		}
+	}
+
+	// 5. 执行加载逻辑
+	var cachedData *types.CachedFileData
+	var err error
+
+	if source.IsURL() {
+		cachedData, err = loadFromURL(c, source.URL, reason...)
+	} else {
+		cachedData, err = loadFromBase64(source.Base64Data, source.MimeType)
+	}
+
+	if err != nil {
+		return nil, err
+	}
+
+	// 6. 设置缓存
+	source.SetCache(cachedData)
+	if contextKey != "" && c != nil {
+		c.Set(contextKey, cachedData)
+	}
+
+	// 7. 注册到 context 以便请求结束时自动清理
+	if c != nil {
+		registerSourceForCleanup(c, source)
+	}
+
+	return cachedData, nil
+}
+
+// registerSourceForCleanup 注册 FileSource 到 context 以便请求结束时清理
+func registerSourceForCleanup(c *gin.Context, source *types.FileSource) {
+	if source.IsRegistered() {
+		return
+	}
+
+	key := string(constant.ContextKeyFileSourcesToCleanup)
+	var sources []*types.FileSource
+	if existing, exists := c.Get(key); exists {
+		sources = existing.([]*types.FileSource)
+	}
+	sources = append(sources, source)
+	c.Set(key, sources)
+	source.SetRegistered(true)
+}
+
+// CleanupFileSources 清理请求中所有注册的 FileSource
+// 应在请求结束时调用（通常由中间件自动调用）
+func CleanupFileSources(c *gin.Context) {
+	key := string(constant.ContextKeyFileSourcesToCleanup)
+	if sources, exists := c.Get(key); exists {
+		for _, source := range sources.([]*types.FileSource) {
+			if cache := source.GetCache(); cache != nil {
+				cache.Close()
+			}
+		}
+		c.Set(key, nil) // 清除引用
+	}
+}
+
+// loadFromURL 从 URL 加载文件
+func loadFromURL(c *gin.Context, url string, reason ...string) (*types.CachedFileData, error) {
+	// 下载文件
+	var maxFileSize = constant.MaxFileDownloadMB * 1024 * 1024
+
+	if common.DebugEnabled {
+		logger.LogDebug(c, "loadFromURL: initiating download")
+	}
+	resp, err := DoDownloadRequest(url, reason...)
+	if err != nil {
+		return nil, fmt.Errorf("failed to download file from %s: %w", url, err)
+	}
+	defer resp.Body.Close()
+
+	if resp.StatusCode != 200 {
+		return nil, fmt.Errorf("failed to download file, status code: %d", resp.StatusCode)
+	}
+
+	// 读取文件内容（限制大小）
+	if common.DebugEnabled {
+		logger.LogDebug(c, "loadFromURL: reading response body")
+	}
+	fileBytes, err := io.ReadAll(io.LimitReader(resp.Body, int64(maxFileSize+1)))
+	if err != nil {
+		return nil, fmt.Errorf("failed to read file content: %w", err)
+	}
+	if len(fileBytes) > maxFileSize {
+		return nil, fmt.Errorf("file size exceeds maximum allowed size: %dMB", constant.MaxFileDownloadMB)
+	}
+
+	// 转换为 base64
+	base64Data := base64.StdEncoding.EncodeToString(fileBytes)
+
+	// 智能获取 MIME 类型
+	mimeType := smartDetectMimeType(resp, url, fileBytes)
+
+	// 判断是否使用磁盘缓存
+	base64Size := int64(len(base64Data))
+	var cachedData *types.CachedFileData
+
+	if shouldUseDiskCache(base64Size) {
+		// 使用磁盘缓存
+		diskPath, err := writeToDiskCache(base64Data)
+		if err != nil {
+			// 磁盘缓存失败，回退到内存
+			logger.LogWarn(c, fmt.Sprintf("Failed to write to disk cache, falling back to memory: %v", err))
+			cachedData = types.NewMemoryCachedData(base64Data, mimeType, int64(len(fileBytes)))
+		} else {
+			cachedData = types.NewDiskCachedData(diskPath, mimeType, int64(len(fileBytes)))
+			cachedData.DiskSize = base64Size
+			cachedData.OnClose = func(size int64) {
+				common.DecrementDiskFiles(size)
+			}
+			common.IncrementDiskFiles(base64Size)
+			if common.DebugEnabled {
+				logger.LogDebug(c, fmt.Sprintf("File cached to disk: %s, size: %d bytes", diskPath, base64Size))
+			}
+		}
+	} else {
+		// 使用内存缓存
+		cachedData = types.NewMemoryCachedData(base64Data, mimeType, int64(len(fileBytes)))
+	}
+
+	// 如果是图片，尝试获取图片配置
+	if strings.HasPrefix(mimeType, "image/") {
+		if common.DebugEnabled {
+			logger.LogDebug(c, "loadFromURL: decoding image config")
+		}
+		config, format, err := decodeImageConfig(fileBytes)
+		if err == nil {
+			cachedData.ImageConfig = &config
+			cachedData.ImageFormat = format
+			// 如果通过图片解码获取了更准确的格式，更新 MIME 类型
+			if mimeType == "application/octet-stream" || mimeType == "" {
+				cachedData.MimeType = "image/" + format
+			}
+		}
+	}
+
+	return cachedData, nil
+}
+
+// shouldUseDiskCache 判断是否应该使用磁盘缓存
+func shouldUseDiskCache(dataSize int64) bool {
+	return common.ShouldUseDiskCache(dataSize)
+}
+
+// writeToDiskCache 将数据写入磁盘缓存
+func writeToDiskCache(base64Data string) (string, error) {
+	return common.WriteDiskCacheFileString(common.DiskCacheTypeFile, base64Data)
+}
+
+// smartDetectMimeType 智能检测 MIME 类型
+func smartDetectMimeType(resp *http.Response, url string, fileBytes []byte) string {
+	// 1. 尝试从 Content-Type header 获取
+	mimeType := resp.Header.Get("Content-Type")
+	if idx := strings.Index(mimeType, ";"); idx != -1 {
+		mimeType = strings.TrimSpace(mimeType[:idx])
+	}
+	if mimeType != "" && mimeType != "application/octet-stream" {
+		return mimeType
+	}
+
+	// 2. 尝试从 Content-Disposition header 的 filename 获取
+	if cd := resp.Header.Get("Content-Disposition"); cd != "" {
+		parts := strings.Split(cd, ";")
+		for _, part := range parts {
+			part = strings.TrimSpace(part)
+			if strings.HasPrefix(strings.ToLower(part), "filename=") {
+				name := strings.TrimSpace(strings.TrimPrefix(part, "filename="))
+				// 移除引号
+				if len(name) > 2 && name[0] == '"' && name[len(name)-1] == '"' {
+					name = name[1 : len(name)-1]
+				}
+				if dot := strings.LastIndex(name, "."); dot != -1 && dot+1 < len(name) {
+					ext := strings.ToLower(name[dot+1:])
+					if ext != "" {
+						mt := GetMimeTypeByExtension(ext)
+						if mt != "application/octet-stream" {
+							return mt
+						}
+					}
+				}
+				break
+			}
+		}
+	}
+
+	// 3. 尝试从 URL 路径获取扩展名
+	mt := guessMimeTypeFromURL(url)
+	if mt != "application/octet-stream" {
+		return mt
+	}
+
+	// 4. 使用 http.DetectContentType 内容嗅探
+	if len(fileBytes) > 0 {
+		sniffed := http.DetectContentType(fileBytes)
+		if sniffed != "" && sniffed != "application/octet-stream" {
+			// 去除可能的 charset 参数
+			if idx := strings.Index(sniffed, ";"); idx != -1 {
+				sniffed = strings.TrimSpace(sniffed[:idx])
+			}
+			return sniffed
+		}
+	}
+
+	// 5. 尝试作为图片解码获取格式
+	if len(fileBytes) > 0 {
+		if _, format, err := decodeImageConfig(fileBytes); err == nil && format != "" {
+			return "image/" + strings.ToLower(format)
+		}
+	}
+
+	// 最终回退
+	return "application/octet-stream"
+}
+
+// loadFromBase64 从 base64 字符串加载文件
+func loadFromBase64(base64String string, providedMimeType string) (*types.CachedFileData, error) {
+	var mimeType string
+	var cleanBase64 string
+
+	// 处理 data: 前缀
+	if strings.HasPrefix(base64String, "data:") {
+		idx := strings.Index(base64String, ",")
+		if idx != -1 {
+			header := base64String[:idx]
+			cleanBase64 = base64String[idx+1:]
+
+			if strings.Contains(header, ":") && strings.Contains(header, ";") {
+				mimeStart := strings.Index(header, ":") + 1
+				mimeEnd := strings.Index(header, ";")
+				if mimeStart < mimeEnd {
+					mimeType = header[mimeStart:mimeEnd]
+				}
+			}
+		} else {
+			cleanBase64 = base64String
+		}
+	} else {
+		cleanBase64 = base64String
+	}
+
+	if providedMimeType != "" {
+		mimeType = providedMimeType
+	}
+
+	decodedData, err := base64.StdEncoding.DecodeString(cleanBase64)
+	if err != nil {
+		return nil, fmt.Errorf("failed to decode base64 data: %w", err)
+	}
+
+	base64Size := int64(len(cleanBase64))
+	var cachedData *types.CachedFileData
+
+	if shouldUseDiskCache(base64Size) {
+		diskPath, err := writeToDiskCache(cleanBase64)
+		if err != nil {
+			cachedData = types.NewMemoryCachedData(cleanBase64, mimeType, int64(len(decodedData)))
+		} else {
+			cachedData = types.NewDiskCachedData(diskPath, mimeType, int64(len(decodedData)))
+			cachedData.DiskSize = base64Size
+			cachedData.OnClose = func(size int64) {
+				common.DecrementDiskFiles(size)
+			}
+			common.IncrementDiskFiles(base64Size)
+		}
+	} else {
+		cachedData = types.NewMemoryCachedData(cleanBase64, mimeType, int64(len(decodedData)))
+	}
+
+	if mimeType == "" || strings.HasPrefix(mimeType, "image/") {
+		config, format, err := decodeImageConfig(decodedData)
+		if err == nil {
+			cachedData.ImageConfig = &config
+			cachedData.ImageFormat = format
+			if mimeType == "" {
+				cachedData.MimeType = "image/" + format
+			}
+		}
+	}
+
+	return cachedData, nil
+}
+
+// GetImageConfig 获取图片配置
+func GetImageConfig(c *gin.Context, source *types.FileSource) (image.Config, string, error) {
+	cachedData, err := LoadFileSource(c, source, "get_image_config")
+	if err != nil {
+		return image.Config{}, "", err
+	}
+
+	if cachedData.ImageConfig != nil {
+		return *cachedData.ImageConfig, cachedData.ImageFormat, nil
+	}
+
+	base64Str, err := cachedData.GetBase64Data()
+	if err != nil {
+		return image.Config{}, "", fmt.Errorf("failed to get base64 data: %w", err)
+	}
+	decodedData, err := base64.StdEncoding.DecodeString(base64Str)
+	if err != nil {
+		return image.Config{}, "", fmt.Errorf("failed to decode base64 for image config: %w", err)
+	}
+
+	config, format, err := decodeImageConfig(decodedData)
+	if err != nil {
+		return image.Config{}, "", err
+	}
+
+	cachedData.ImageConfig = &config
+	cachedData.ImageFormat = format
+
+	return config, format, nil
+}
+
+// GetBase64Data 获取 base64 编码的数据
+func GetBase64Data(c *gin.Context, source *types.FileSource, reason ...string) (string, string, error) {
+	cachedData, err := LoadFileSource(c, source, reason...)
+	if err != nil {
+		return "", "", err
+	}
+	base64Str, err := cachedData.GetBase64Data()
+	if err != nil {
+		return "", "", fmt.Errorf("failed to get base64 data: %w", err)
+	}
+	return base64Str, cachedData.MimeType, nil
+}
+
+// GetMimeType 获取文件的 MIME 类型
+func GetMimeType(c *gin.Context, source *types.FileSource) (string, error) {
+	if source.HasCache() {
+		return source.GetCache().MimeType, nil
+	}
+
+	if source.IsURL() {
+		mimeType, err := GetFileTypeFromUrl(c, source.URL, "get_mime_type")
+		if err == nil && mimeType != "" && mimeType != "application/octet-stream" {
+			return mimeType, nil
+		}
+	}
+
+	cachedData, err := LoadFileSource(c, source, "get_mime_type")
+	if err != nil {
+		return "", err
+	}
+	return cachedData.MimeType, nil
+}
+
+// DetectFileType 检测文件类型
+func DetectFileType(mimeType string) types.FileType {
+	if strings.HasPrefix(mimeType, "image/") {
+		return types.FileTypeImage
+	}
+	if strings.HasPrefix(mimeType, "audio/") {
+		return types.FileTypeAudio
+	}
+	if strings.HasPrefix(mimeType, "video/") {
+		return types.FileTypeVideo
+	}
+	return types.FileTypeFile
+}
+
+// decodeImageConfig 从字节数据解码图片配置
+func decodeImageConfig(data []byte) (image.Config, string, error) {
+	reader := bytes.NewReader(data)
+
+	config, format, err := image.DecodeConfig(reader)
+	if err == nil {
+		return config, format, nil
+	}
+
+	reader.Seek(0, io.SeekStart)
+	config, err = webp.DecodeConfig(reader)
+	if err == nil {
+		return config, "webp", nil
+	}
+
+	return image.Config{}, "", fmt.Errorf("failed to decode image config: unsupported format")
+}
+
+// guessMimeTypeFromURL 从 URL 猜测 MIME 类型
+func guessMimeTypeFromURL(url string) string {
+	cleanedURL := url
+	if q := strings.Index(cleanedURL, "?"); q != -1 {
+		cleanedURL = cleanedURL[:q]
+	}
+
+	if slash := strings.LastIndex(cleanedURL, "/"); slash != -1 && slash+1 < len(cleanedURL) {
+		last := cleanedURL[slash+1:]
+		if dot := strings.LastIndex(last, "."); dot != -1 && dot+1 < len(last) {
+			ext := strings.ToLower(last[dot+1:])
+			return GetMimeTypeByExtension(ext)
+		}
+	}
+
+	return "application/octet-stream"
+}

service/funding_source.go

@@ -0,0 +1,140 @@
+package service
+
+import (
+	"time"
+
+	"github.com/QuantumNous/new-api/model"
+)
+
+// ---------------------------------------------------------------------------
+// FundingSource — 资金来源接口（钱包 or 订阅）
+// ---------------------------------------------------------------------------
+
+// FundingSource 抽象了预扣费的资金来源。
+type FundingSource interface {
+	// Source 返回资金来源标识："wallet" 或 "subscription"
+	Source() string
+	// PreConsume 从该资金来源预扣 amount 额度
+	PreConsume(amount int) error
+	// Settle 根据差额调整资金来源（正数补扣，负数退还）
+	Settle(delta int) error
+	// Refund 退还所有预扣费
+	Refund() error
+}
+
+// ---------------------------------------------------------------------------
+// WalletFunding — 钱包资金来源实现
+// ---------------------------------------------------------------------------
+
+type WalletFunding struct {
+	userId   int
+	consumed int // 实际预扣的用户额度
+}
+
+func (w *WalletFunding) Source() string { return BillingSourceWallet }
+
+func (w *WalletFunding) PreConsume(amount int) error {
+	if amount <= 0 {
+		return nil
+	}
+	if err := model.DecreaseUserQuota(w.userId, amount); err != nil {
+		return err
+	}
+	w.consumed = amount
+	return nil
+}
+
+func (w *WalletFunding) Settle(delta int) error {
+	if delta == 0 {
+		return nil
+	}
+	if delta > 0 {
+		return model.DecreaseUserQuota(w.userId, delta)
+	}
+	return model.IncreaseUserQuota(w.userId, -delta, false)
+}
+
+func (w *WalletFunding) Refund() error {
+	if w.consumed <= 0 {
+		return nil
+	}
+	// IncreaseUserQuota 是 quota += N 的非幂等操作，不能重试，否则会多退额度。
+	// 订阅的 RefundSubscriptionPreConsume 有 requestId 幂等保护所以可以重试。
+	return model.IncreaseUserQuota(w.userId, w.consumed, false)
+}
+
+// ---------------------------------------------------------------------------
+// SubscriptionFunding — 订阅资金来源实现
+// ---------------------------------------------------------------------------
+
+type SubscriptionFunding struct {
+	requestId      string
+	userId         int
+	modelName      string
+	amount         int64 // 预扣的订阅额度（subConsume）
+	group          string
+	subscriptionId int
+	preConsumed    int64
+	// 以下字段在 PreConsume 成功后填充，供 RelayInfo 同步使用
+	AmountTotal     int64
+	AmountUsedAfter int64
+	PlanId          int
+	PlanTitle       string
+}
+
+func (s *SubscriptionFunding) Source() string { return BillingSourceSubscription }
+
+func (s *SubscriptionFunding) PreConsume(_ int) error {
+	// amount 参数被忽略，使用内部 s.amount（已在构造时根据 preConsumedQuota 计算）
+	res, err := model.PreConsumeUserSubscription(s.requestId, s.userId, s.modelName, 0, s.amount, s.group)
+	if err != nil {
+		return err
+	}
+	s.subscriptionId = res.UserSubscriptionId
+	s.preConsumed = res.PreConsumed
+	s.AmountTotal = res.AmountTotal
+	s.AmountUsedAfter = res.AmountUsedAfter
+	// 获取订阅计划信息
+	if planInfo, err := model.GetSubscriptionPlanInfoByUserSubscriptionId(res.UserSubscriptionId); err == nil && planInfo != nil {
+		s.PlanId = planInfo.PlanId
+		s.PlanTitle = planInfo.PlanTitle
+	}
+	return nil
+}
+
+func (s *SubscriptionFunding) Settle(delta int) error {
+	if delta == 0 {
+		return nil
+	}
+	return model.PostConsumeUserSubscriptionDelta(s.subscriptionId, int64(delta))
+}
+
+func (s *SubscriptionFunding) Refund() error {
+	if s.preConsumed <= 0 {
+		return nil
+	}
+	return refundWithRetry(func() error {
+		return model.RefundSubscriptionPreConsume(s.requestId)
+	})
+}
+
+// refundWithRetry 尝试多次执行退款操作以提高成功率，只能用于基于事务的退款函数！！！！！！
+// try to refund with retries, only for refund functions based on transactions!!!
+func refundWithRetry(fn func() error) error {
+	if fn == nil {
+		return nil
+	}
+	const maxAttempts = 3
+	var lastErr error
+	for i := 0; i < maxAttempts; i++ {
+		if err := fn(); err == nil {
+			return nil
+		} else {
+			lastErr = err
+		}
+		if i < maxAttempts-1 {
+			time.Sleep(time.Duration(200*(i+1)) * time.Millisecond)
+		}
+	}
+	return lastErr
+}

service/group.go

@@ -0,0 +1,65 @@
+package service
+
+import (
+	"strings"
+
+	"github.com/QuantumNous/new-api/setting"
+	"github.com/QuantumNous/new-api/setting/ratio_setting"
+)
+
+func GetUserUsableGroups(userGroup string) map[string]string {
+	groupsCopy := setting.GetUserUsableGroupsCopy()
+	if userGroup != "" {
+		specialSettings, b := ratio_setting.GetGroupRatioSetting().GroupSpecialUsableGroup.Get(userGroup)
+		if b {
+			// 处理特殊可用分组
+			for specialGroup, desc := range specialSettings {
+				if strings.HasPrefix(specialGroup, "-:") {
+					// 移除分组
+					groupToRemove := strings.TrimPrefix(specialGroup, "-:")
+					delete(groupsCopy, groupToRemove)
+				} else if strings.HasPrefix(specialGroup, "+:") {
+					// 添加分组
+					groupToAdd := strings.TrimPrefix(specialGroup, "+:")
+					groupsCopy[groupToAdd] = desc
+				} else {
+					// 直接添加分组
+					groupsCopy[specialGroup] = desc
+				}
+			}
+		}
+		// 如果userGroup不在UserUsableGroups中，返回UserUsableGroups + userGroup
+		if _, ok := groupsCopy[userGroup]; !ok {
+			groupsCopy[userGroup] = "用户分组"
+		}
+	}
+	return groupsCopy
+}
+
+func GroupInUserUsableGroups(userGroup, groupName string) bool {
+	_, ok := GetUserUsableGroups(userGroup)[groupName]
+	return ok
+}
+
+// GetUserAutoGroup 根据用户分组获取自动分组设置
+func GetUserAutoGroup(userGroup string) []string {
+	groups := GetUserUsableGroups(userGroup)
+	autoGroups := make([]string, 0)
+	for _, group := range setting.GetAutoGroups() {
+		if _, ok := groups[group]; ok {
+			autoGroups = append(autoGroups, group)
+		}
+	}
+	return autoGroups
+}
+
+// GetUserGroupRatio 获取用户使用某个分组的倍率
+// userGroup 用户分组
+// group 需要获取倍率的分组
+func GetUserGroupRatio(userGroup, group string) float64 {
+	ratio, ok := ratio_setting.GetGroupGroupRatio(userGroup, group)
+	if ok {
+		return ratio
+	}
+	return ratio_setting.GetGroupRatio(group)
+}

service/http.go

@@ -0,0 +1,61 @@
+package service
+
+import (
+	"bytes"
+	"fmt"
+	"io"
+	"net/http"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/logger"
+
+	"github.com/gin-gonic/gin"
+)
+
+func CloseResponseBodyGracefully(httpResponse *http.Response) {
+	if httpResponse == nil || httpResponse.Body == nil {
+		return
+	}
+	err := httpResponse.Body.Close()
+	if err != nil {
+		common.SysError("failed to close response body: " + err.Error())
+	}
+}
+
+func IOCopyBytesGracefully(c *gin.Context, src *http.Response, data []byte) {
+	if c.Writer == nil {
+		return
+	}
+
+	body := io.NopCloser(bytes.NewBuffer(data))
+
+	// We shouldn't set the header before we parse the response body, because the parse part may fail.
+	// And then we will have to send an error response, but in this case, the header has already been set.
+	// So the httpClient will be confused by the response.
+	// For example, Postman will report error, and we cannot check the response at all.
+	if src != nil {
+		for k, v := range src.Header {
+			// avoid setting Content-Length
+			if k == "Content-Length" {
+				continue
+			}
+			c.Writer.Header().Set(k, v[0])
+		}
+	}
+
+	// set Content-Length header manually BEFORE calling WriteHeader
+	c.Writer.Header().Set("Content-Length", fmt.Sprintf("%d", len(data)))
+
+	// Write header with status code (this sends the headers)
+	if src != nil {
+		c.Writer.WriteHeader(src.StatusCode)
+	} else {
+		c.Writer.WriteHeader(http.StatusOK)
+	}
+
+	_, err := io.Copy(c.Writer, body)
+	if err != nil {
+		logger.LogError(c, fmt.Sprintf("failed to copy response body: %s", err.Error()))
+	}
+	c.Writer.Flush()
+}

service/http_client.go

@@ -0,0 +1,169 @@
+package service
+
+import (
+	"context"
+	"fmt"
+	"net"
+	"net/http"
+	"net/url"
+	"sync"
+	"time"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/setting/system_setting"
+
+	"golang.org/x/net/proxy"
+)
+
+var (
+	httpClient      *http.Client
+	proxyClientLock sync.Mutex
+	proxyClients    = make(map[string]*http.Client)
+)
+
+func checkRedirect(req *http.Request, via []*http.Request) error {
+	fetchSetting := system_setting.GetFetchSetting()
+	urlStr := req.URL.String()
+	if err := common.ValidateURLWithFetchSetting(urlStr, fetchSetting.EnableSSRFProtection, fetchSetting.AllowPrivateIp, fetchSetting.DomainFilterMode, fetchSetting.IpFilterMode, fetchSetting.DomainList, fetchSetting.IpList, fetchSetting.AllowedPorts, fetchSetting.ApplyIPFilterForDomain); err != nil {
+		return fmt.Errorf("redirect to %s blocked: %v", urlStr, err)
+	}
+	if len(via) >= 10 {
+		return fmt.Errorf("stopped after 10 redirects")
+	}
+	return nil
+}
+
+func InitHttpClient() {
+	transport := &http.Transport{
+		MaxIdleConns:        common.RelayMaxIdleConns,
+		MaxIdleConnsPerHost: common.RelayMaxIdleConnsPerHost,
+		ForceAttemptHTTP2:   true,
+		Proxy:               http.ProxyFromEnvironment, // Support HTTP_PROXY, HTTPS_PROXY, NO_PROXY env vars
+	}
+	if common.TLSInsecureSkipVerify {
+		transport.TLSClientConfig = common.InsecureTLSConfig
+	}
+
+	if common.RelayTimeout == 0 {
+		httpClient = &http.Client{
+			Transport:     transport,
+			CheckRedirect: checkRedirect,
+		}
+	} else {
+		httpClient = &http.Client{
+			Transport:     transport,
+			Timeout:       time.Duration(common.RelayTimeout) * time.Second,
+			CheckRedirect: checkRedirect,
+		}
+	}
+}
+
+func GetHttpClient() *http.Client {
+	return httpClient
+}
+
+// GetHttpClientWithProxy returns the default client or a proxy-enabled one when proxyURL is provided.
+func GetHttpClientWithProxy(proxyURL string) (*http.Client, error) {
+	if proxyURL == "" {
+		return GetHttpClient(), nil
+	}
+	return NewProxyHttpClient(proxyURL)
+}
+
+// ResetProxyClientCache 清空代理客户端缓存，确保下次使用时重新初始化
+func ResetProxyClientCache() {
+	proxyClientLock.Lock()
+	defer proxyClientLock.Unlock()
+	for _, client := range proxyClients {
+		if transport, ok := client.Transport.(*http.Transport); ok && transport != nil {
+			transport.CloseIdleConnections()
+		}
+	}
+	proxyClients = make(map[string]*http.Client)
+}
+
+// NewProxyHttpClient 创建支持代理的 HTTP 客户端
+func NewProxyHttpClient(proxyURL string) (*http.Client, error) {
+	if proxyURL == "" {
+		if client := GetHttpClient(); client != nil {
+			return client, nil
+		}
+		return http.DefaultClient, nil
+	}
+
+	proxyClientLock.Lock()
+	if client, ok := proxyClients[proxyURL]; ok {
+		proxyClientLock.Unlock()
+		return client, nil
+	}
+	proxyClientLock.Unlock()
+
+	parsedURL, err := url.Parse(proxyURL)
+	if err != nil {
+		return nil, err
+	}
+
+	switch parsedURL.Scheme {
+	case "http", "https":
+		transport := &http.Transport{
+			MaxIdleConns:        common.RelayMaxIdleConns,
+			MaxIdleConnsPerHost: common.RelayMaxIdleConnsPerHost,
+			ForceAttemptHTTP2:   true,
+			Proxy:               http.ProxyURL(parsedURL),
+		}
+		if common.TLSInsecureSkipVerify {
+			transport.TLSClientConfig = common.InsecureTLSConfig
+		}
+		client := &http.Client{
+			Transport:     transport,
+			CheckRedirect: checkRedirect,
+		}
+		client.Timeout = time.Duration(common.RelayTimeout) * time.Second
+		proxyClientLock.Lock()
+		proxyClients[proxyURL] = client
+		proxyClientLock.Unlock()
+		return client, nil
+
+	case "socks5", "socks5h":
+		// 获取认证信息
+		var auth *proxy.Auth
+		if parsedURL.User != nil {
+			auth = &proxy.Auth{
+				User:     parsedURL.User.Username(),
+				Password: "",
+			}
+			if password, ok := parsedURL.User.Password(); ok {
+				auth.Password = password
+			}
+		}
+
+		// 创建 SOCKS5 代理拨号器
+		// proxy.SOCKS5 使用 tcp 参数，所有 TCP 连接包括 DNS 查询都将通过代理进行。行为与 socks5h 相同
+		dialer, err := proxy.SOCKS5("tcp", parsedURL.Host, auth, proxy.Direct)
+		if err != nil {
+			return nil, err
+		}
+
+		transport := &http.Transport{
+			MaxIdleConns:        common.RelayMaxIdleConns,
+			MaxIdleConnsPerHost: common.RelayMaxIdleConnsPerHost,
+			ForceAttemptHTTP2:   true,
+			DialContext: func(ctx context.Context, network, addr string) (net.Conn, error) {
+				return dialer.Dial(network, addr)
+			},
+		}
+		if common.TLSInsecureSkipVerify {
+			transport.TLSClientConfig = common.InsecureTLSConfig
+		}
+
+		client := &http.Client{Transport: transport, CheckRedirect: checkRedirect}
+		client.Timeout = time.Duration(common.RelayTimeout) * time.Second
+		proxyClientLock.Lock()
+		proxyClients[proxyURL] = client
+		proxyClientLock.Unlock()
+		return client, nil
+
+	default:
+		return nil, fmt.Errorf("unsupported proxy scheme: %s, must be http, https, socks5 or socks5h", parsedURL.Scheme)
+	}
+}

service/image.go

@@ -0,0 +1,178 @@
+package service
+
+import (
+	"bytes"
+	"encoding/base64"
+	"errors"
+	"fmt"
+	"image"
+	"io"
+	"net/http"
+	"strings"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/constant"
+
+	"golang.org/x/image/webp"
+)
+
+// return image.Config, format, clean base64 string, error
+func DecodeBase64ImageData(base64String string) (image.Config, string, string, error) {
+	// 去除base64数据的URL前缀（如果有）
+	if idx := strings.Index(base64String, ","); idx != -1 {
+		base64String = base64String[idx+1:]
+	}
+
+	if len(base64String) == 0 {
+		return image.Config{}, "", "", errors.New("base64 string is empty")
+	}
+
+	// 将base64字符串解码为字节切片
+	decodedData, err := base64.StdEncoding.DecodeString(base64String)
+	if err != nil {
+		fmt.Println("Error: Failed to decode base64 string")
+		return image.Config{}, "", "", fmt.Errorf("failed to decode base64 string: %s", err.Error())
+	}
+
+	// 创建一个bytes.Buffer用于存储解码后的数据
+	reader := bytes.NewReader(decodedData)
+	config, format, err := getImageConfig(reader)
+	return config, format, base64String, err
+}
+
+func DecodeBase64FileData(base64String string) (string, string, error) {
+	var mimeType string
+	var idx int
+	idx = strings.Index(base64String, ",")
+	if idx == -1 {
+		_, file_type, base64, err := DecodeBase64ImageData(base64String)
+		return "image/" + file_type, base64, err
+	}
+	mimeType = base64String[:idx]
+	base64String = base64String[idx+1:]
+	idx = strings.Index(mimeType, ";")
+	if idx == -1 {
+		_, file_type, base64, err := DecodeBase64ImageData(base64String)
+		return "image/" + file_type, base64, err
+	}
+	mimeType = mimeType[:idx]
+	idx = strings.Index(mimeType, ":")
+	if idx == -1 {
+		_, file_type, base64, err := DecodeBase64ImageData(base64String)
+		return "image/" + file_type, base64, err
+	}
+	mimeType = mimeType[idx+1:]
+	return mimeType, base64String, nil
+}
+
+// GetImageFromUrl 获取图片的类型和base64编码的数据
+func GetImageFromUrl(url string) (mimeType string, data string, err error) {
+	resp, err := DoDownloadRequest(url)
+	if err != nil {
+		return "", "", fmt.Errorf("failed to download image: %w", err)
+	}
+	defer resp.Body.Close()
+
+	// Check HTTP status code
+	if resp.StatusCode != http.StatusOK {
+		return "", "", fmt.Errorf("failed to download image: HTTP %d", resp.StatusCode)
+	}
+
+	contentType := resp.Header.Get("Content-Type")
+	if contentType != "application/octet-stream" && !strings.HasPrefix(contentType, "image/") {
+		return "", "", fmt.Errorf("invalid content type: %s, required image/*", contentType)
+	}
+	maxImageSize := int64(constant.MaxFileDownloadMB * 1024 * 1024)
+
+	// Check Content-Length if available
+	if resp.ContentLength > maxImageSize {
+		return "", "", fmt.Errorf("image size %d exceeds maximum allowed size of %d bytes", resp.ContentLength, maxImageSize)
+	}
+
+	// Use LimitReader to prevent reading oversized images
+	limitReader := io.LimitReader(resp.Body, maxImageSize)
+	buffer := &bytes.Buffer{}
+
+	written, err := io.Copy(buffer, limitReader)
+	if err != nil {
+		return "", "", fmt.Errorf("failed to read image data: %w", err)
+	}
+	if written >= maxImageSize {
+		return "", "", fmt.Errorf("image size exceeds maximum allowed size of %d bytes", maxImageSize)
+	}
+
+	data = base64.StdEncoding.EncodeToString(buffer.Bytes())
+	mimeType = contentType
+
+	// Handle application/octet-stream type
+	if mimeType == "application/octet-stream" {
+		_, format, _, err := DecodeBase64ImageData(data)
+		if err != nil {
+			return "", "", err
+		}
+		mimeType = "image/" + format
+	}
+
+	return mimeType, data, nil
+}
+
+func DecodeUrlImageData(imageUrl string) (image.Config, string, error) {
+	response, err := DoDownloadRequest(imageUrl)
+	if err != nil {
+		common.SysLog(fmt.Sprintf("fail to get image from url: %s", err.Error()))
+		return image.Config{}, "", err
+	}
+	defer response.Body.Close()
+
+	if response.StatusCode != 200 {
+		err = errors.New(fmt.Sprintf("fail to get image from url: %s", response.Status))
+		return image.Config{}, "", err
+	}
+
+	mimeType := response.Header.Get("Content-Type")
+
+	if mimeType != "application/octet-stream" && !strings.HasPrefix(mimeType, "image/") {
+		return image.Config{}, "", fmt.Errorf("invalid content type: %s, required image/*", mimeType)
+	}
+
+	var readData []byte
+	for _, limit := range []int64{1024 * 8, 1024 * 24, 1024 * 64} {
+		common.SysLog(fmt.Sprintf("try to decode image config with limit: %d", limit))
+
+		// 从response.Body读取更多的数据直到达到当前的限制
+		additionalData := make([]byte, limit-int64(len(readData)))
+		n, _ := io.ReadFull(response.Body, additionalData)
+		readData = append(readData, additionalData[:n]...)
+
+		// 使用io.MultiReader组合已经读取的数据和response.Body
+		limitReader := io.MultiReader(bytes.NewReader(readData), response.Body)
+
+		var config image.Config
+		var format string
+		config, format, err = getImageConfig(limitReader)
+		if err == nil {
+			return config, format, nil
+		}
+	}
+
+	return image.Config{}, "", err // 返回最后一个错误
+}
+
+func getImageConfig(reader io.Reader) (image.Config, string, error) {
+	// 读取图片的头部信息来获取图片尺寸
+	config, format, err := image.DecodeConfig(reader)
+	if err != nil {
+		err = errors.New(fmt.Sprintf("fail to decode image config(gif, jpg, png): %s", err.Error()))
+		common.SysLog(err.Error())
+		config, err = webp.DecodeConfig(reader)
+		if err != nil {
+			err = errors.New(fmt.Sprintf("fail to decode image config(webp): %s", err.Error()))
+			common.SysLog(err.Error())
+		}
+		format = "webp"
+	}
+	if err != nil {
+		return image.Config{}, "", err
+	}
+	return config, format, nil
+}

service/log_info_generate.go

@@ -0,0 +1,236 @@
+package service
+
+import (
+	"strings"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/constant"
+	"github.com/QuantumNous/new-api/dto"
+	relaycommon "github.com/QuantumNous/new-api/relay/common"
+	"github.com/QuantumNous/new-api/types"
+
+	"github.com/gin-gonic/gin"
+)
+
+func appendRequestPath(ctx *gin.Context, relayInfo *relaycommon.RelayInfo, other map[string]interface{}) {
+	if other == nil {
+		return
+	}
+	if ctx != nil && ctx.Request != nil && ctx.Request.URL != nil {
+		if path := ctx.Request.URL.Path; path != "" {
+			other["request_path"] = path
+			return
+		}
+	}
+	if relayInfo != nil && relayInfo.RequestURLPath != "" {
+		path := relayInfo.RequestURLPath
+		if idx := strings.Index(path, "?"); idx != -1 {
+			path = path[:idx]
+		}
+		other["request_path"] = path
+	}
+}
+
+func GenerateTextOtherInfo(ctx *gin.Context, relayInfo *relaycommon.RelayInfo, modelRatio, groupRatio, completionRatio float64,
+	cacheTokens int, cacheRatio float64, modelPrice float64, userGroupRatio float64) map[string]interface{} {
+	other := make(map[string]interface{})
+	other["model_ratio"] = modelRatio
+	other["group_ratio"] = groupRatio
+	other["completion_ratio"] = completionRatio
+	other["cache_tokens"] = cacheTokens
+	other["cache_ratio"] = cacheRatio
+	other["model_price"] = modelPrice
+	other["user_group_ratio"] = userGroupRatio
+	other["frt"] = float64(relayInfo.FirstResponseTime.UnixMilli() - relayInfo.StartTime.UnixMilli())
+	if relayInfo.ReasoningEffort != "" {
+		other["reasoning_effort"] = relayInfo.ReasoningEffort
+	}
+	if relayInfo.IsModelMapped {
+		other["is_model_mapped"] = true
+		other["upstream_model_name"] = relayInfo.UpstreamModelName
+	}
+
+	isSystemPromptOverwritten := common.GetContextKeyBool(ctx, constant.ContextKeySystemPromptOverride)
+	if isSystemPromptOverwritten {
+		other["is_system_prompt_overwritten"] = true
+	}
+
+	adminInfo := make(map[string]interface{})
+	adminInfo["use_channel"] = ctx.GetStringSlice("use_channel")
+	isMultiKey := common.GetContextKeyBool(ctx, constant.ContextKeyChannelIsMultiKey)
+	if isMultiKey {
+		adminInfo["is_multi_key"] = true
+		adminInfo["multi_key_index"] = common.GetContextKeyInt(ctx, constant.ContextKeyChannelMultiKeyIndex)
+	}
+
+	isLocalCountTokens := common.GetContextKeyBool(ctx, constant.ContextKeyLocalCountTokens)
+	if isLocalCountTokens {
+		adminInfo["local_count_tokens"] = isLocalCountTokens
+	}
+
+	AppendChannelAffinityAdminInfo(ctx, adminInfo)
+
+	other["admin_info"] = adminInfo
+	appendRequestPath(ctx, relayInfo, other)
+	appendRequestConversionChain(relayInfo, other)
+	appendFinalRequestFormat(relayInfo, other)
+	appendBillingInfo(relayInfo, other)
+	appendParamOverrideInfo(relayInfo, other)
+	return other
+}
+
+func appendParamOverrideInfo(relayInfo *relaycommon.RelayInfo, other map[string]interface{}) {
+	if relayInfo == nil || other == nil || len(relayInfo.ParamOverrideAudit) == 0 {
+		return
+	}
+	other["po"] = relayInfo.ParamOverrideAudit
+}
+
+func appendBillingInfo(relayInfo *relaycommon.RelayInfo, other map[string]interface{}) {
+	if relayInfo == nil || other == nil {
+		return
+	}
+	// billing_source: "wallet" or "subscription"
+	if relayInfo.BillingSource != "" {
+		other["billing_source"] = relayInfo.BillingSource
+	}
+	if relayInfo.UserSetting.BillingPreference != "" {
+		other["billing_preference"] = relayInfo.UserSetting.BillingPreference
+	}
+	if relayInfo.BillingSource == "subscription" {
+		if relayInfo.SubscriptionId != 0 {
+			other["subscription_id"] = relayInfo.SubscriptionId
+		}
+		if relayInfo.SubscriptionPreConsumed > 0 {
+			other["subscription_pre_consumed"] = relayInfo.SubscriptionPreConsumed
+		}
+		// post_delta: settlement delta applied after actual usage is known (can be negative for refund)
+		if relayInfo.SubscriptionPostDelta != 0 {
+			other["subscription_post_delta"] = relayInfo.SubscriptionPostDelta
+		}
+		if relayInfo.SubscriptionPlanId != 0 {
+			other["subscription_plan_id"] = relayInfo.SubscriptionPlanId
+		}
+		if relayInfo.SubscriptionPlanTitle != "" {
+			other["subscription_plan_title"] = relayInfo.SubscriptionPlanTitle
+		}
+		// Compute "this request" subscription consumed + remaining
+		consumed := relayInfo.SubscriptionPreConsumed + relayInfo.SubscriptionPostDelta
+		usedFinal := relayInfo.SubscriptionAmountUsedAfterPreConsume + relayInfo.SubscriptionPostDelta
+		if consumed < 0 {
+			consumed = 0
+		}
+		if usedFinal < 0 {
+			usedFinal = 0
+		}
+		if relayInfo.SubscriptionAmountTotal > 0 {
+			remain := relayInfo.SubscriptionAmountTotal - usedFinal
+			if remain < 0 {
+				remain = 0
+			}
+			other["subscription_total"] = relayInfo.SubscriptionAmountTotal
+			other["subscription_used"] = usedFinal
+			other["subscription_remain"] = remain
+		}
+		if consumed > 0 {
+			other["subscription_consumed"] = consumed
+		}
+		// Wallet quota is not deducted when billed from subscription.
+		other["wallet_quota_deducted"] = 0
+	}
+}
+
+func appendRequestConversionChain(relayInfo *relaycommon.RelayInfo, other map[string]interface{}) {
+	if relayInfo == nil || other == nil {
+		return
+	}
+	if len(relayInfo.RequestConversionChain) == 0 {
+		return
+	}
+	chain := make([]string, 0, len(relayInfo.RequestConversionChain))
+	for _, f := range relayInfo.RequestConversionChain {
+		switch f {
+		case types.RelayFormatOpenAI:
+			chain = append(chain, "OpenAI Compatible")
+		case types.RelayFormatClaude:
+			chain = append(chain, "Claude Messages")
+		case types.RelayFormatGemini:
+			chain = append(chain, "Google Gemini")
+		case types.RelayFormatOpenAIResponses:
+			chain = append(chain, "OpenAI Responses")
+		default:
+			chain = append(chain, string(f))
+		}
+	}
+	if len(chain) == 0 {
+		return
+	}
+	other["request_conversion"] = chain
+}
+
+func appendFinalRequestFormat(relayInfo *relaycommon.RelayInfo, other map[string]interface{}) {
+	if relayInfo == nil || other == nil {
+		return
+	}
+	if relayInfo.GetFinalRequestRelayFormat() == types.RelayFormatClaude {
+		// claude indicates the final upstream request format is Claude Messages.
+		// Frontend log rendering uses this to keep the original Claude input display.
+		other["claude"] = true
+	}
+}
+
+func GenerateWssOtherInfo(ctx *gin.Context, relayInfo *relaycommon.RelayInfo, usage *dto.RealtimeUsage, modelRatio, groupRatio, completionRatio, audioRatio, audioCompletionRatio, modelPrice, userGroupRatio float64) map[string]interface{} {
+	info := GenerateTextOtherInfo(ctx, relayInfo, modelRatio, groupRatio, completionRatio, 0, 0.0, modelPrice, userGroupRatio)
+	info["ws"] = true
+	info["audio_input"] = usage.InputTokenDetails.AudioTokens
+	info["audio_output"] = usage.OutputTokenDetails.AudioTokens
+	info["text_input"] = usage.InputTokenDetails.TextTokens
+	info["text_output"] = usage.OutputTokenDetails.TextTokens
+	info["audio_ratio"] = audioRatio
+	info["audio_completion_ratio"] = audioCompletionRatio
+	return info
+}
+
+func GenerateAudioOtherInfo(ctx *gin.Context, relayInfo *relaycommon.RelayInfo, usage *dto.Usage, modelRatio, groupRatio, completionRatio, audioRatio, audioCompletionRatio, modelPrice, userGroupRatio float64) map[string]interface{} {
+	info := GenerateTextOtherInfo(ctx, relayInfo, modelRatio, groupRatio, completionRatio, 0, 0.0, modelPrice, userGroupRatio)
+	info["audio"] = true
+	info["audio_input"] = usage.PromptTokensDetails.AudioTokens
+	info["audio_output"] = usage.CompletionTokenDetails.AudioTokens
+	info["text_input"] = usage.PromptTokensDetails.TextTokens
+	info["text_output"] = usage.CompletionTokenDetails.TextTokens
+	info["audio_ratio"] = audioRatio
+	info["audio_completion_ratio"] = audioCompletionRatio
+	return info
+}
+
+func GenerateClaudeOtherInfo(ctx *gin.Context, relayInfo *relaycommon.RelayInfo, modelRatio, groupRatio, completionRatio float64,
+	cacheTokens int, cacheRatio float64,
+	cacheCreationTokens int, cacheCreationRatio float64,
+	cacheCreationTokens5m int, cacheCreationRatio5m float64,
+	cacheCreationTokens1h int, cacheCreationRatio1h float64,
+	modelPrice float64, userGroupRatio float64) map[string]interface{} {
+	info := GenerateTextOtherInfo(ctx, relayInfo, modelRatio, groupRatio, completionRatio, cacheTokens, cacheRatio, modelPrice, userGroupRatio)
+	info["claude"] = true
+	info["cache_creation_tokens"] = cacheCreationTokens
+	info["cache_creation_ratio"] = cacheCreationRatio
+	if cacheCreationTokens5m != 0 {
+		info["cache_creation_tokens_5m"] = cacheCreationTokens5m
+		info["cache_creation_ratio_5m"] = cacheCreationRatio5m
+	}
+	if cacheCreationTokens1h != 0 {
+		info["cache_creation_tokens_1h"] = cacheCreationTokens1h
+		info["cache_creation_ratio_1h"] = cacheCreationRatio1h
+	}
+	return info
+}
+
+func GenerateMjOtherInfo(relayInfo *relaycommon.RelayInfo, priceData types.PriceData) map[string]interface{} {
+	other := make(map[string]interface{})
+	other["model_price"] = priceData.ModelPrice
+	other["group_ratio"] = priceData.GroupRatioInfo.GroupRatio
+	if priceData.GroupRatioInfo.HasSpecialRatio {
+		other["user_group_ratio"] = priceData.GroupRatioInfo.GroupSpecialRatio
+	}
+	appendRequestPath(nil, relayInfo, other)
+	return other
+}

service/midjourney.go

@@ -0,0 +1,259 @@
+package service
+
+import (
+	"context"
+	"encoding/json"
+	"io"
+	"log"
+	"net/http"
+	"strconv"
+	"strings"
+	"time"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/constant"
+	"github.com/QuantumNous/new-api/dto"
+	relayconstant "github.com/QuantumNous/new-api/relay/constant"
+	"github.com/QuantumNous/new-api/setting"
+
+	"github.com/gin-gonic/gin"
+)
+
+func CovertMjpActionToModelName(mjAction string) string {
+	modelName := "mj_" + strings.ToLower(mjAction)
+	if mjAction == constant.MjActionSwapFace {
+		modelName = "swap_face"
+	}
+	return modelName
+}
+
+func GetMjRequestModel(relayMode int, midjRequest *dto.MidjourneyRequest) (string, *dto.MidjourneyResponse, bool) {
+	action := ""
+	if relayMode == relayconstant.RelayModeMidjourneyAction {
+		// plus request
+		err := CoverPlusActionToNormalAction(midjRequest)
+		if err != nil {
+			return "", err, false
+		}
+		action = midjRequest.Action
+	} else {
+		switch relayMode {
+		case relayconstant.RelayModeMidjourneyImagine:
+			action = constant.MjActionImagine
+		case relayconstant.RelayModeMidjourneyVideo:
+			action = constant.MjActionVideo
+		case relayconstant.RelayModeMidjourneyEdits:
+			action = constant.MjActionEdits
+		case relayconstant.RelayModeMidjourneyDescribe:
+			action = constant.MjActionDescribe
+		case relayconstant.RelayModeMidjourneyBlend:
+			action = constant.MjActionBlend
+		case relayconstant.RelayModeMidjourneyShorten:
+			action = constant.MjActionShorten
+		case relayconstant.RelayModeMidjourneyChange:
+			action = midjRequest.Action
+		case relayconstant.RelayModeMidjourneyModal:
+			action = constant.MjActionModal
+		case relayconstant.RelayModeSwapFace:
+			action = constant.MjActionSwapFace
+		case relayconstant.RelayModeMidjourneyUpload:
+			action = constant.MjActionUpload
+		case relayconstant.RelayModeMidjourneySimpleChange:
+			params := ConvertSimpleChangeParams(midjRequest.Content)
+			if params == nil {
+				return "", MidjourneyErrorWrapper(constant.MjRequestError, "invalid_request"), false
+			}
+			action = params.Action
+		case relayconstant.RelayModeMidjourneyTaskFetch, relayconstant.RelayModeMidjourneyTaskFetchByCondition, relayconstant.RelayModeMidjourneyNotify:
+			return "", nil, true
+		default:
+			return "", MidjourneyErrorWrapper(constant.MjRequestError, "unknown_relay_action"), false
+		}
+	}
+	modelName := CovertMjpActionToModelName(action)
+	return modelName, nil, true
+}
+
+func CoverPlusActionToNormalAction(midjRequest *dto.MidjourneyRequest) *dto.MidjourneyResponse {
+	// "customId": "MJ::JOB::upsample::2::3dbbd469-36af-4a0f-8f02-df6c579e7011"
+	customId := midjRequest.CustomId
+	if customId == "" {
+		return MidjourneyErrorWrapper(constant.MjRequestError, "custom_id_is_required")
+	}
+	splits := strings.Split(customId, "::")
+	var action string
+	if splits[1] == "JOB" {
+		action = splits[2]
+	} else {
+		action = splits[1]
+	}
+
+	if action == "" {
+		return MidjourneyErrorWrapper(constant.MjRequestError, "unknown_action")
+	}
+	if strings.Contains(action, "upsample") {
+		index, err := strconv.Atoi(splits[3])
+		if err != nil {
+			return MidjourneyErrorWrapper(constant.MjRequestError, "index_parse_failed")
+		}
+		midjRequest.Index = index
+		midjRequest.Action = constant.MjActionUpscale
+	} else if strings.Contains(action, "variation") {
+		midjRequest.Index = 1
+		if action == "variation" {
+			index, err := strconv.Atoi(splits[3])
+			if err != nil {
+				return MidjourneyErrorWrapper(constant.MjRequestError, "index_parse_failed")
+			}
+			midjRequest.Index = index
+			midjRequest.Action = constant.MjActionVariation
+		} else if action == "low_variation" {
+			midjRequest.Action = constant.MjActionLowVariation
+		} else if action == "high_variation" {
+			midjRequest.Action = constant.MjActionHighVariation
+		}
+	} else if strings.Contains(action, "pan") {
+		midjRequest.Action = constant.MjActionPan
+		midjRequest.Index = 1
+	} else if strings.Contains(action, "reroll") {
+		midjRequest.Action = constant.MjActionReRoll
+		midjRequest.Index = 1
+	} else if action == "Outpaint" {
+		midjRequest.Action = constant.MjActionZoom
+		midjRequest.Index = 1
+	} else if action == "CustomZoom" {
+		midjRequest.Action = constant.MjActionCustomZoom
+		midjRequest.Index = 1
+	} else if action == "Inpaint" {
+		midjRequest.Action = constant.MjActionInPaint
+		midjRequest.Index = 1
+	} else {
+		return MidjourneyErrorWrapper(constant.MjRequestError, "unknown_action:"+customId)
+	}
+	return nil
+}
+
+func ConvertSimpleChangeParams(content string) *dto.MidjourneyRequest {
+	split := strings.Split(content, " ")
+	if len(split) != 2 {
+		return nil
+	}
+
+	action := strings.ToLower(split[1])
+	changeParams := &dto.MidjourneyRequest{}
+	changeParams.TaskId = split[0]
+
+	if action[0] == 'u' {
+		changeParams.Action = "UPSCALE"
+	} else if action[0] == 'v' {
+		changeParams.Action = "VARIATION"
+	} else if action == "r" {
+		changeParams.Action = "REROLL"
+		return changeParams
+	} else {
+		return nil
+	}
+
+	index, err := strconv.Atoi(action[1:2])
+	if err != nil || index < 1 || index > 4 {
+		return nil
+	}
+	changeParams.Index = index
+	return changeParams
+}
+
+func DoMidjourneyHttpRequest(c *gin.Context, timeout time.Duration, fullRequestURL string) (*dto.MidjourneyResponseWithStatusCode, []byte, error) {
+	var nullBytes []byte
+	//var requestBody io.Reader
+	//requestBody = c.Request.Body
+	// read request body to json, delete accountFilter and notifyHook
+	var mapResult map[string]interface{}
+	// if get request, no need to read request body
+	if c.Request.Method != "GET" {
+		err := json.NewDecoder(c.Request.Body).Decode(&mapResult)
+		if err != nil {
+			return MidjourneyErrorWithStatusCodeWrapper(constant.MjErrorUnknown, "read_request_body_failed", http.StatusInternalServerError), nullBytes, err
+		}
+		if !setting.MjAccountFilterEnabled {
+			delete(mapResult, "accountFilter")
+		}
+		if !setting.MjNotifyEnabled {
+			delete(mapResult, "notifyHook")
+		}
+		//req, err := http.NewRequest(c.Request.Method, fullRequestURL, requestBody)
+		// make new request with mapResult
+	}
+	if setting.MjModeClearEnabled {
+		if prompt, ok := mapResult["prompt"].(string); ok {
+			prompt = strings.Replace(prompt, "--fast", "", -1)
+			prompt = strings.Replace(prompt, "--relax", "", -1)
+			prompt = strings.Replace(prompt, "--turbo", "", -1)
+
+			mapResult["prompt"] = prompt
+		}
+	}
+	reqBody, err := json.Marshal(mapResult)
+	if err != nil {
+		return MidjourneyErrorWithStatusCodeWrapper(constant.MjErrorUnknown, "marshal_request_body_failed", http.StatusInternalServerError), nullBytes, err
+	}
+	req, err := http.NewRequest(c.Request.Method, fullRequestURL, strings.NewReader(string(reqBody)))
+	if err != nil {
+		return MidjourneyErrorWithStatusCodeWrapper(constant.MjErrorUnknown, "create_request_failed", http.StatusInternalServerError), nullBytes, err
+	}
+	ctx, cancel := context.WithTimeout(context.Background(), timeout)
+	// 使用带有超时的 context 创建新的请求
+	req = req.WithContext(ctx)
+	req.Header.Set("Content-Type", c.Request.Header.Get("Content-Type"))
+	req.Header.Set("Accept", c.Request.Header.Get("Accept"))
+	auth := common.GetContextKeyString(c, constant.ContextKeyChannelKey)
+	if auth != "" {
+		auth = strings.TrimPrefix(auth, "Bearer ")
+		req.Header.Set("mj-api-secret", auth)
+	}
+	defer cancel()
+	resp, err := GetHttpClient().Do(req)
+	if err != nil {
+		common.SysLog("do request failed: " + err.Error())
+		return MidjourneyErrorWithStatusCodeWrapper(constant.MjErrorUnknown, "do_request_failed", http.StatusInternalServerError), nullBytes, err
+	}
+	statusCode := resp.StatusCode
+	//if statusCode != 200  {
+	//	return MidjourneyErrorWithStatusCodeWrapper(constant.MjErrorUnknown, "bad_response_status_code", statusCode), nullBytes, nil
+	//}
+	err = req.Body.Close()
+	if err != nil {
+		return MidjourneyErrorWithStatusCodeWrapper(constant.MjErrorUnknown, "close_request_body_failed", statusCode), nullBytes, err
+	}
+	err = c.Request.Body.Close()
+	if err != nil {
+		return MidjourneyErrorWithStatusCodeWrapper(constant.MjErrorUnknown, "close_request_body_failed", statusCode), nullBytes, err
+	}
+	var midjResponse dto.MidjourneyResponse
+	var midjourneyUploadsResponse dto.MidjourneyUploadResponse
+	responseBody, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return MidjourneyErrorWithStatusCodeWrapper(constant.MjErrorUnknown, "read_response_body_failed", statusCode), nullBytes, err
+	}
+	CloseResponseBodyGracefully(resp)
+	respStr := string(responseBody)
+	log.Printf("respStr: %s", respStr)
+	if respStr == "" {
+		return MidjourneyErrorWithStatusCodeWrapper(constant.MjErrorUnknown, "empty_response_body", statusCode), responseBody, nil
+	} else {
+		err = json.Unmarshal(responseBody, &midjResponse)
+		if err != nil {
+			err2 := json.Unmarshal(responseBody, &midjourneyUploadsResponse)
+			if err2 != nil {
+				return MidjourneyErrorWithStatusCodeWrapper(constant.MjErrorUnknown, "unmarshal_response_body_failed", statusCode), responseBody, err
+			}
+		}
+	}
+	//log.Printf("midjResponse: %v", midjResponse)
+	//for k, v := range resp.Header {
+	//	c.Writer.Header().Set(k, v[0])
+	//}
+	return &dto.MidjourneyResponseWithStatusCode{
+		StatusCode: statusCode,
+		Response:   midjResponse,
+	}, responseBody, nil
+}

service/notify-limit.go

@@ -0,0 +1,118 @@
+package service
+
+import (
+	"fmt"
+	"strconv"
+	"sync"
+	"time"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/constant"
+	"github.com/bytedance/gopkg/util/gopool"
+)
+
+// notifyLimitStore is used for in-memory rate limiting when Redis is disabled
+var (
+	notifyLimitStore sync.Map
+	cleanupOnce      sync.Once
+)
+
+type limitCount struct {
+	Count     int
+	Timestamp time.Time
+}
+
+func getDuration() time.Duration {
+	minute := constant.NotificationLimitDurationMinute
+	return time.Duration(minute) * time.Minute
+}
+
+// startCleanupTask starts a background task to clean up expired entries
+func startCleanupTask() {
+	gopool.Go(func() {
+		for {
+			time.Sleep(time.Hour)
+			now := time.Now()
+			notifyLimitStore.Range(func(key, value interface{}) bool {
+				if limit, ok := value.(limitCount); ok {
+					if now.Sub(limit.Timestamp) >= getDuration() {
+						notifyLimitStore.Delete(key)
+					}
+				}
+				return true
+			})
+		}
+	})
+}
+
+// CheckNotificationLimit checks if the user has exceeded their notification limit
+// Returns true if the user can send notification, false if limit exceeded
+func CheckNotificationLimit(userId int, notifyType string) (bool, error) {
+	if common.RedisEnabled {
+		return checkRedisLimit(userId, notifyType)
+	}
+	return checkMemoryLimit(userId, notifyType)
+}
+
+func checkRedisLimit(userId int, notifyType string) (bool, error) {
+	key := fmt.Sprintf("notify_limit:%d:%s:%s", userId, notifyType, time.Now().Format("2006010215"))
+
+	// Get current count
+	count, err := common.RedisGet(key)
+	if err != nil && err.Error() != "redis: nil" {
+		return false, fmt.Errorf("failed to get notification count: %w", err)
+	}
+
+	// If key doesn't exist, initialize it
+	if count == "" {
+		err = common.RedisSet(key, "1", getDuration())
+		return true, err
+	}
+
+	currentCount, _ := strconv.Atoi(count)
+	limit := constant.NotifyLimitCount
+
+	// Check if limit is already reached
+	if currentCount >= limit {
+		return false, nil
+	}
+
+	// Only increment if under limit
+	err = common.RedisIncr(key, 1)
+	if err != nil {
+		return false, fmt.Errorf("failed to increment notification count: %w", err)
+	}
+
+	return true, nil
+}
+
+func checkMemoryLimit(userId int, notifyType string) (bool, error) {
+	// Ensure cleanup task is started
+	cleanupOnce.Do(startCleanupTask)
+
+	key := fmt.Sprintf("%d:%s:%s", userId, notifyType, time.Now().Format("2006010215"))
+	now := time.Now()
+
+	// Get current limit count or initialize new one
+	var currentLimit limitCount
+	if value, ok := notifyLimitStore.Load(key); ok {
+		currentLimit = value.(limitCount)
+		// Check if the entry has expired
+		if now.Sub(currentLimit.Timestamp) >= getDuration() {
+			currentLimit = limitCount{Count: 0, Timestamp: now}
+		}
+	} else {
+		currentLimit = limitCount{Count: 0, Timestamp: now}
+	}
+
+	// Increment count
+	currentLimit.Count++
+
+	// Check against limits
+	limit := constant.NotifyLimitCount
+
+	// Store updated count
+	notifyLimitStore.Store(key, currentLimit)
+
+	return currentLimit.Count <= limit, nil
+}

service/openai_chat_responses_compat.go

@@ -0,0 +1,18 @@
+package service
+
+import (
+	"github.com/QuantumNous/new-api/dto"
+	"github.com/QuantumNous/new-api/service/openaicompat"
+)
+
+func ChatCompletionsRequestToResponsesRequest(req *dto.GeneralOpenAIRequest) (*dto.OpenAIResponsesRequest, error) {
+	return openaicompat.ChatCompletionsRequestToResponsesRequest(req)
+}
+
+func ResponsesResponseToChatCompletionsResponse(resp *dto.OpenAIResponsesResponse, id string) (*dto.OpenAITextResponse, *dto.Usage, error) {
+	return openaicompat.ResponsesResponseToChatCompletionsResponse(resp, id)
+}
+
+func ExtractOutputTextFromResponses(resp *dto.OpenAIResponsesResponse) string {
+	return openaicompat.ExtractOutputTextFromResponses(resp)
+}

service/openai_chat_responses_mode.go

@@ -0,0 +1,14 @@
+package service
+
+import (
+	"github.com/QuantumNous/new-api/service/openaicompat"
+	"github.com/QuantumNous/new-api/setting/model_setting"
+)
+
+func ShouldChatCompletionsUseResponsesPolicy(policy model_setting.ChatCompletionsToResponsesPolicy, channelID int, channelType int, model string) bool {
+	return openaicompat.ShouldChatCompletionsUseResponsesPolicy(policy, channelID, channelType, model)
+}
+
+func ShouldChatCompletionsUseResponsesGlobal(channelID int, channelType int, model string) bool {
+	return openaicompat.ShouldChatCompletionsUseResponsesGlobal(channelID, channelType, model)
+}

service/openaicompat/chat_to_responses.go

@@ -0,0 +1,402 @@
+package openaicompat
+
+import (
+	"encoding/json"
+	"errors"
+	"fmt"
+	"strings"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/dto"
+	"github.com/samber/lo"
+)
+
+func normalizeChatImageURLToString(v any) any {
+	switch vv := v.(type) {
+	case string:
+		return vv
+	case map[string]any:
+		if url := common.Interface2String(vv["url"]); url != "" {
+			return url
+		}
+		return v
+	case dto.MessageImageUrl:
+		if vv.Url != "" {
+			return vv.Url
+		}
+		return v
+	case *dto.MessageImageUrl:
+		if vv != nil && vv.Url != "" {
+			return vv.Url
+		}
+		return v
+	default:
+		return v
+	}
+}
+
+func convertChatResponseFormatToResponsesText(reqFormat *dto.ResponseFormat) json.RawMessage {
+	if reqFormat == nil || strings.TrimSpace(reqFormat.Type) == "" {
+		return nil
+	}
+
+	format := map[string]any{
+		"type": reqFormat.Type,
+	}
+
+	if reqFormat.Type == "json_schema" && len(reqFormat.JsonSchema) > 0 {
+		var chatSchema map[string]any
+		if err := common.Unmarshal(reqFormat.JsonSchema, &chatSchema); err == nil {
+			for key, value := range chatSchema {
+				if key == "type" {
+					continue
+				}
+				format[key] = value
+			}
+
+			if nested, ok := format["json_schema"].(map[string]any); ok {
+				for key, value := range nested {
+					if _, exists := format[key]; !exists {
+						format[key] = value
+					}
+				}
+				delete(format, "json_schema")
+			}
+		} else {
+			format["json_schema"] = reqFormat.JsonSchema
+		}
+	}
+
+	textRaw, _ := common.Marshal(map[string]any{
+		"format": format,
+	})
+	return textRaw
+}
+
+func ChatCompletionsRequestToResponsesRequest(req *dto.GeneralOpenAIRequest) (*dto.OpenAIResponsesRequest, error) {
+	if req == nil {
+		return nil, errors.New("request is nil")
+	}
+	if req.Model == "" {
+		return nil, errors.New("model is required")
+	}
+	if lo.FromPtrOr(req.N, 1) > 1 {
+		return nil, fmt.Errorf("n>1 is not supported in responses compatibility mode")
+	}
+
+	var instructionsParts []string
+	inputItems := make([]map[string]any, 0, len(req.Messages))
+
+	for _, msg := range req.Messages {
+		role := strings.TrimSpace(msg.Role)
+		if role == "" {
+			continue
+		}
+
+		if role == "tool" || role == "function" {
+			callID := strings.TrimSpace(msg.ToolCallId)
+
+			var output any
+			if msg.Content == nil {
+				output = ""
+			} else if msg.IsStringContent() {
+				output = msg.StringContent()
+			} else {
+				if b, err := common.Marshal(msg.Content); err == nil {
+					output = string(b)
+				} else {
+					output = fmt.Sprintf("%v", msg.Content)
+				}
+			}
+
+			if callID == "" {
+				inputItems = append(inputItems, map[string]any{
+					"role":    "user",
+					"content": fmt.Sprintf("[tool_output_missing_call_id] %v", output),
+				})
+				continue
+			}
+
+			inputItems = append(inputItems, map[string]any{
+				"type":    "function_call_output",
+				"call_id": callID,
+				"output":  output,
+			})
+			continue
+		}
+
+		// Prefer mapping system/developer messages into `instructions`.
+		if role == "system" || role == "developer" {
+			if msg.Content == nil {
+				continue
+			}
+			if msg.IsStringContent() {
+				if s := strings.TrimSpace(msg.StringContent()); s != "" {
+					instructionsParts = append(instructionsParts, s)
+				}
+				continue
+			}
+			parts := msg.ParseContent()
+			var sb strings.Builder
+			for _, part := range parts {
+				if part.Type == dto.ContentTypeText && strings.TrimSpace(part.Text) != "" {
+					if sb.Len() > 0 {
+						sb.WriteString("\n")
+					}
+					sb.WriteString(part.Text)
+				}
+			}
+			if s := strings.TrimSpace(sb.String()); s != "" {
+				instructionsParts = append(instructionsParts, s)
+			}
+			continue
+		}
+
+		item := map[string]any{
+			"role": role,
+		}
+
+		if msg.Content == nil {
+			item["content"] = ""
+			inputItems = append(inputItems, item)
+
+			if role == "assistant" {
+				for _, tc := range msg.ParseToolCalls() {
+					if strings.TrimSpace(tc.ID) == "" {
+						continue
+					}
+					if tc.Type != "" && tc.Type != "function" {
+						continue
+					}
+					name := strings.TrimSpace(tc.Function.Name)
+					if name == "" {
+						continue
+					}
+					inputItems = append(inputItems, map[string]any{
+						"type":      "function_call",
+						"call_id":   tc.ID,
+						"name":      name,
+						"arguments": tc.Function.Arguments,
+					})
+				}
+			}
+			continue
+		}
+
+		if msg.IsStringContent() {
+			item["content"] = msg.StringContent()
+			inputItems = append(inputItems, item)
+
+			if role == "assistant" {
+				for _, tc := range msg.ParseToolCalls() {
+					if strings.TrimSpace(tc.ID) == "" {
+						continue
+					}
+					if tc.Type != "" && tc.Type != "function" {
+						continue
+					}
+					name := strings.TrimSpace(tc.Function.Name)
+					if name == "" {
+						continue
+					}
+					inputItems = append(inputItems, map[string]any{
+						"type":      "function_call",
+						"call_id":   tc.ID,
+						"name":      name,
+						"arguments": tc.Function.Arguments,
+					})
+				}
+			}
+			continue
+		}
+
+		parts := msg.ParseContent()
+		contentParts := make([]map[string]any, 0, len(parts))
+		for _, part := range parts {
+			switch part.Type {
+			case dto.ContentTypeText:
+				textType := "input_text"
+				if role == "assistant" {
+					textType = "output_text"
+				}
+				contentParts = append(contentParts, map[string]any{
+					"type": textType,
+					"text": part.Text,
+				})
+			case dto.ContentTypeImageURL:
+				contentParts = append(contentParts, map[string]any{
+					"type":      "input_image",
+					"image_url": normalizeChatImageURLToString(part.ImageUrl),
+				})
+			case dto.ContentTypeInputAudio:
+				contentParts = append(contentParts, map[string]any{
+					"type":        "input_audio",
+					"input_audio": part.InputAudio,
+				})
+			case dto.ContentTypeFile:
+				contentParts = append(contentParts, map[string]any{
+					"type": "input_file",
+					"file": part.File,
+				})
+			case dto.ContentTypeVideoUrl:
+				contentParts = append(contentParts, map[string]any{
+					"type":      "input_video",
+					"video_url": part.VideoUrl,
+				})
+			default:
+				contentParts = append(contentParts, map[string]any{
+					"type": part.Type,
+				})
+			}
+		}
+		item["content"] = contentParts
+		inputItems = append(inputItems, item)
+
+		if role == "assistant" {
+			for _, tc := range msg.ParseToolCalls() {
+				if strings.TrimSpace(tc.ID) == "" {
+					continue
+				}
+				if tc.Type != "" && tc.Type != "function" {
+					continue
+				}
+				name := strings.TrimSpace(tc.Function.Name)
+				if name == "" {
+					continue
+				}
+				inputItems = append(inputItems, map[string]any{
+					"type":      "function_call",
+					"call_id":   tc.ID,
+					"name":      name,
+					"arguments": tc.Function.Arguments,
+				})
+			}
+		}
+	}
+
+	inputRaw, err := common.Marshal(inputItems)
+	if err != nil {
+		return nil, err
+	}
+
+	var instructionsRaw json.RawMessage
+	if len(instructionsParts) > 0 {
+		instructions := strings.Join(instructionsParts, "\n\n")
+		instructionsRaw, _ = common.Marshal(instructions)
+	}
+
+	var toolsRaw json.RawMessage
+	if req.Tools != nil {
+		tools := make([]map[string]any, 0, len(req.Tools))
+		for _, tool := range req.Tools {
+			switch tool.Type {
+			case "function":
+				tools = append(tools, map[string]any{
+					"type":        "function",
+					"name":        tool.Function.Name,
+					"description": tool.Function.Description,
+					"parameters":  tool.Function.Parameters,
+				})
+			default:
+				// Best-effort: keep original tool shape for unknown types.
+				var m map[string]any
+				if b, err := common.Marshal(tool); err == nil {
+					_ = common.Unmarshal(b, &m)
+				}
+				if len(m) == 0 {
+					m = map[string]any{"type": tool.Type}
+				}
+				tools = append(tools, m)
+			}
+		}
+		toolsRaw, _ = common.Marshal(tools)
+	}
+
+	var toolChoiceRaw json.RawMessage
+	if req.ToolChoice != nil {
+		switch v := req.ToolChoice.(type) {
+		case string:
+			toolChoiceRaw, _ = common.Marshal(v)
+		default:
+			var m map[string]any
+			if b, err := common.Marshal(v); err == nil {
+				_ = common.Unmarshal(b, &m)
+			}
+			if m == nil {
+				toolChoiceRaw, _ = common.Marshal(v)
+			} else if t, _ := m["type"].(string); t == "function" {
+				// Chat: {"type":"function","function":{"name":"..."}}
+				// Responses: {"type":"function","name":"..."}
+				if name, ok := m["name"].(string); ok && name != "" {
+					toolChoiceRaw, _ = common.Marshal(map[string]any{
+						"type": "function",
+						"name": name,
+					})
+				} else if fn, ok := m["function"].(map[string]any); ok {
+					if name, ok := fn["name"].(string); ok && name != "" {
+						toolChoiceRaw, _ = common.Marshal(map[string]any{
+							"type": "function",
+							"name": name,
+						})
+					} else {
+						toolChoiceRaw, _ = common.Marshal(v)
+					}
+				} else {
+					toolChoiceRaw, _ = common.Marshal(v)
+				}
+			} else {
+				toolChoiceRaw, _ = common.Marshal(v)
+			}
+		}
+	}
+
+	var parallelToolCallsRaw json.RawMessage
+	if req.ParallelTooCalls != nil {
+		parallelToolCallsRaw, _ = common.Marshal(*req.ParallelTooCalls)
+	}
+
+	textRaw := convertChatResponseFormatToResponsesText(req.ResponseFormat)
+
+	maxOutputTokens := lo.FromPtrOr(req.MaxTokens, uint(0))
+	maxCompletionTokens := lo.FromPtrOr(req.MaxCompletionTokens, uint(0))
+	if maxCompletionTokens > maxOutputTokens {
+		maxOutputTokens = maxCompletionTokens
+	}
+	// OpenAI Responses API rejects max_output_tokens < 16 when explicitly provided.
+	//if maxOutputTokens > 0 && maxOutputTokens < 16 {
+	//	maxOutputTokens = 16
+	//}
+
+	var topP *float64
+	if req.TopP != nil {
+		topP = common.GetPointer(lo.FromPtr(req.TopP))
+	}
+
+	out := &dto.OpenAIResponsesRequest{
+		Model:             req.Model,
+		Input:             inputRaw,
+		Instructions:      instructionsRaw,
+		Stream:            req.Stream,
+		Temperature:       req.Temperature,
+		Text:              textRaw,
+		ToolChoice:        toolChoiceRaw,
+		Tools:             toolsRaw,
+		TopP:              topP,
+		User:              req.User,
+		ParallelToolCalls: parallelToolCallsRaw,
+		Store:             req.Store,
+		Metadata:          req.Metadata,
+	}
+	if req.MaxTokens != nil || req.MaxCompletionTokens != nil {
+		out.MaxOutputTokens = lo.ToPtr(maxOutputTokens)
+	}
+
+	if req.ReasoningEffort != "" {
+		out.Reasoning = &dto.Reasoning{
+			Effort:  req.ReasoningEffort,
+			Summary: "detailed",
+		}
+	}
+
+	return out, nil
+}

service/openaicompat/policy.go

@@ -0,0 +1,19 @@
+package openaicompat
+
+import "github.com/QuantumNous/new-api/setting/model_setting"
+
+func ShouldChatCompletionsUseResponsesPolicy(policy model_setting.ChatCompletionsToResponsesPolicy, channelID int, channelType int, model string) bool {
+	if !policy.IsChannelEnabled(channelID, channelType) {
+		return false
+	}
+	return matchAnyRegex(policy.ModelPatterns, model)
+}
+
+func ShouldChatCompletionsUseResponsesGlobal(channelID int, channelType int, model string) bool {
+	return ShouldChatCompletionsUseResponsesPolicy(
+		model_setting.GetGlobalSettings().ChatCompletionsToResponsesPolicy,
+		channelID,
+		channelType,
+		model,
+	)
+}

service/openaicompat/regex.go

@@ -0,0 +1,33 @@
+package openaicompat
+
+import (
+	"regexp"
+	"sync"
+)
+
+var compiledRegexCache sync.Map // map[string]*regexp.Regexp
+
+func matchAnyRegex(patterns []string, s string) bool {
+	if len(patterns) == 0 || s == "" {
+		return false
+	}
+	for _, pattern := range patterns {
+		if pattern == "" {
+			continue
+		}
+		re, ok := compiledRegexCache.Load(pattern)
+		if !ok {
+			compiled, err := regexp.Compile(pattern)
+			if err != nil {
+				// Treat invalid patterns as non-matching to avoid breaking runtime traffic.
+				continue
+			}
+			re = compiled
+			compiledRegexCache.Store(pattern, re)
+		}
+		if re.(*regexp.Regexp).MatchString(s) {
+			return true
+		}
+	}
+	return false
+}

service/openaicompat/responses_to_chat.go

@@ -0,0 +1,133 @@
+package openaicompat
+
+import (
+	"errors"
+	"strings"
+
+	"github.com/QuantumNous/new-api/dto"
+)
+
+func ResponsesResponseToChatCompletionsResponse(resp *dto.OpenAIResponsesResponse, id string) (*dto.OpenAITextResponse, *dto.Usage, error) {
+	if resp == nil {
+		return nil, nil, errors.New("response is nil")
+	}
+
+	text := ExtractOutputTextFromResponses(resp)
+
+	usage := &dto.Usage{}
+	if resp.Usage != nil {
+		if resp.Usage.InputTokens != 0 {
+			usage.PromptTokens = resp.Usage.InputTokens
+			usage.InputTokens = resp.Usage.InputTokens
+		}
+		if resp.Usage.OutputTokens != 0 {
+			usage.CompletionTokens = resp.Usage.OutputTokens
+			usage.OutputTokens = resp.Usage.OutputTokens
+		}
+		if resp.Usage.TotalTokens != 0 {
+			usage.TotalTokens = resp.Usage.TotalTokens
+		} else {
+			usage.TotalTokens = usage.PromptTokens + usage.CompletionTokens
+		}
+		if resp.Usage.InputTokensDetails != nil {
+			usage.PromptTokensDetails.CachedTokens = resp.Usage.InputTokensDetails.CachedTokens
+			usage.PromptTokensDetails.ImageTokens = resp.Usage.InputTokensDetails.ImageTokens
+			usage.PromptTokensDetails.AudioTokens = resp.Usage.InputTokensDetails.AudioTokens
+		}
+		if resp.Usage.CompletionTokenDetails.ReasoningTokens != 0 {
+			usage.CompletionTokenDetails.ReasoningTokens = resp.Usage.CompletionTokenDetails.ReasoningTokens
+		}
+	}
+
+	created := resp.CreatedAt
+
+	var toolCalls []dto.ToolCallResponse
+	if text == "" && len(resp.Output) > 0 {
+		for _, out := range resp.Output {
+			if out.Type != "function_call" {
+				continue
+			}
+			name := strings.TrimSpace(out.Name)
+			if name == "" {
+				continue
+			}
+			callId := strings.TrimSpace(out.CallId)
+			if callId == "" {
+				callId = strings.TrimSpace(out.ID)
+			}
+			toolCalls = append(toolCalls, dto.ToolCallResponse{
+				ID:   callId,
+				Type: "function",
+				Function: dto.FunctionResponse{
+					Name:      name,
+					Arguments: out.Arguments,
+				},
+			})
+		}
+	}
+
+	finishReason := "stop"
+	if len(toolCalls) > 0 {
+		finishReason = "tool_calls"
+	}
+
+	msg := dto.Message{
+		Role:    "assistant",
+		Content: text,
+	}
+	if len(toolCalls) > 0 {
+		msg.SetToolCalls(toolCalls)
+		msg.Content = ""
+	}
+
+	out := &dto.OpenAITextResponse{
+		Id:      id,
+		Object:  "chat.completion",
+		Created: created,
+		Model:   resp.Model,
+		Choices: []dto.OpenAITextResponseChoice{
+			{
+				Index:        0,
+				Message:      msg,
+				FinishReason: finishReason,
+			},
+		},
+		Usage: *usage,
+	}
+
+	return out, usage, nil
+}
+
+func ExtractOutputTextFromResponses(resp *dto.OpenAIResponsesResponse) string {
+	if resp == nil || len(resp.Output) == 0 {
+		return ""
+	}
+
+	var sb strings.Builder
+
+	// Prefer assistant message outputs.
+	for _, out := range resp.Output {
+		if out.Type != "message" {
+			continue
+		}
+		if out.Role != "" && out.Role != "assistant" {
+			continue
+		}
+		for _, c := range out.Content {
+			if c.Type == "output_text" && c.Text != "" {
+				sb.WriteString(c.Text)
+			}
+		}
+	}
+	if sb.Len() > 0 {
+		return sb.String()
+	}
+	for _, out := range resp.Output {
+		for _, c := range out.Content {
+			if c.Text != "" {
+				sb.WriteString(c.Text)
+			}
+		}
+	}
+	return sb.String()
+}

service/passkey/service.go

@@ -0,0 +1,177 @@
+package passkey
+
+import (
+	"errors"
+	"fmt"
+	"net"
+	"net/http"
+	"net/url"
+	"strings"
+	"time"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/setting/system_setting"
+
+	"github.com/go-webauthn/webauthn/protocol"
+	webauthn "github.com/go-webauthn/webauthn/webauthn"
+)
+
+const (
+	RegistrationSessionKey = "passkey_registration_session"
+	LoginSessionKey        = "passkey_login_session"
+	VerifySessionKey       = "passkey_verify_session"
+)
+
+// BuildWebAuthn constructs a WebAuthn instance using the current passkey settings and request context.
+func BuildWebAuthn(r *http.Request) (*webauthn.WebAuthn, error) {
+	settings := system_setting.GetPasskeySettings()
+	if settings == nil {
+		return nil, errors.New("未找到 Passkey 设置")
+	}
+
+	displayName := strings.TrimSpace(settings.RPDisplayName)
+	if displayName == "" {
+		displayName = common.SystemName
+	}
+
+	origins, err := resolveOrigins(r, settings)
+	if err != nil {
+		return nil, err
+	}
+
+	rpID, err := resolveRPID(r, settings, origins)
+	if err != nil {
+		return nil, err
+	}
+
+	selection := protocol.AuthenticatorSelection{
+		ResidentKey:        protocol.ResidentKeyRequirementRequired,
+		RequireResidentKey: protocol.ResidentKeyRequired(),
+		UserVerification:   protocol.UserVerificationRequirement(settings.UserVerification),
+	}
+	if selection.UserVerification == "" {
+		selection.UserVerification = protocol.VerificationPreferred
+	}
+	if attachment := strings.TrimSpace(settings.AttachmentPreference); attachment != "" {
+		selection.AuthenticatorAttachment = protocol.AuthenticatorAttachment(attachment)
+	}
+
+	config := &webauthn.Config{
+		RPID:                   rpID,
+		RPDisplayName:          displayName,
+		RPOrigins:              origins,
+		AuthenticatorSelection: selection,
+		Debug:                  common.DebugEnabled,
+		Timeouts: webauthn.TimeoutsConfig{
+			Login: webauthn.TimeoutConfig{
+				Enforce:    true,
+				Timeout:    2 * time.Minute,
+				TimeoutUVD: 2 * time.Minute,
+			},
+			Registration: webauthn.TimeoutConfig{
+				Enforce:    true,
+				Timeout:    2 * time.Minute,
+				TimeoutUVD: 2 * time.Minute,
+			},
+		},
+	}
+
+	return webauthn.New(config)
+}
+
+func resolveOrigins(r *http.Request, settings *system_setting.PasskeySettings) ([]string, error) {
+	originsStr := strings.TrimSpace(settings.Origins)
+	if originsStr != "" {
+		originList := strings.Split(originsStr, ",")
+		origins := make([]string, 0, len(originList))
+		for _, origin := range originList {
+			trimmed := strings.TrimSpace(origin)
+			if trimmed == "" {
+				continue
+			}
+			if !settings.AllowInsecureOrigin && strings.HasPrefix(strings.ToLower(trimmed), "http://") {
+				return nil, fmt.Errorf("Passkey 不允许使用不安全的 Origin: %s", trimmed)
+			}
+			origins = append(origins, trimmed)
+		}
+		if len(origins) == 0 {
+			// 如果配置了Origins但过滤后为空，使用自动推导
+			goto autoDetect
+		}
+		return origins, nil
+	}
+
+autoDetect:
+	scheme := detectScheme(r)
+	if scheme == "http" && !settings.AllowInsecureOrigin && r.Host != "localhost" && r.Host != "127.0.0.1" && !strings.HasPrefix(r.Host, "127.0.0.1:") && !strings.HasPrefix(r.Host, "localhost:") {
+		return nil, fmt.Errorf("Passkey 仅支持 HTTPS，当前访问: %s://%s，请在 Passkey 设置中允许不安全 Origin 或配置 HTTPS", scheme, r.Host)
+	}
+	// 优先使用请求的完整Host（包含端口）
+	host := r.Host
+
+	// 如果无法从请求获取Host，尝试从ServerAddress获取
+	if host == "" && system_setting.ServerAddress != "" {
+		if parsed, err := url.Parse(system_setting.ServerAddress); err == nil && parsed.Host != "" {
+			host = parsed.Host
+			if scheme == "" && parsed.Scheme != "" {
+				scheme = parsed.Scheme
+			}
+		}
+	}
+	if host == "" {
+		return nil, fmt.Errorf("无法确定 Passkey 的 Origin，请在系统设置或 Passkey 设置中指定。当前 Host: '%s', ServerAddress: '%s'", r.Host, system_setting.ServerAddress)
+	}
+	if scheme == "" {
+		scheme = "https"
+	}
+	origin := fmt.Sprintf("%s://%s", scheme, host)
+	return []string{origin}, nil
+}
+
+func resolveRPID(r *http.Request, settings *system_setting.PasskeySettings, origins []string) (string, error) {
+	rpID := strings.TrimSpace(settings.RPID)
+	if rpID != "" {
+		return hostWithoutPort(rpID), nil
+	}
+	if len(origins) == 0 {
+		return "", errors.New("Passkey 未配置 Origin，无法推导 RPID")
+	}
+	parsed, err := url.Parse(origins[0])
+	if err != nil {
+		return "", fmt.Errorf("无法解析 Passkey Origin: %w", err)
+	}
+	return hostWithoutPort(parsed.Host), nil
+}
+
+func hostWithoutPort(host string) string {
+	host = strings.TrimSpace(host)
+	if host == "" {
+		return ""
+	}
+	if strings.Contains(host, ":") {
+		if host, _, err := net.SplitHostPort(host); err == nil {
+			return host
+		}
+	}
+	return host
+}
+
+func detectScheme(r *http.Request) string {
+	if r == nil {
+		return ""
+	}
+	if proto := r.Header.Get("X-Forwarded-Proto"); proto != "" {
+		parts := strings.Split(proto, ",")
+		return strings.ToLower(strings.TrimSpace(parts[0]))
+	}
+	if r.TLS != nil {
+		return "https"
+	}
+	if r.URL != nil && r.URL.Scheme != "" {
+		return strings.ToLower(r.URL.Scheme)
+	}
+	if r.Header.Get("X-Forwarded-Protocol") != "" {
+		return strings.ToLower(strings.TrimSpace(r.Header.Get("X-Forwarded-Protocol")))
+	}
+	return "http"
+}

service/passkey/session.go

@@ -0,0 +1,50 @@
+package passkey
+
+import (
+	"encoding/json"
+	"errors"
+
+	"github.com/gin-contrib/sessions"
+	"github.com/gin-gonic/gin"
+	webauthn "github.com/go-webauthn/webauthn/webauthn"
+)
+
+var errSessionNotFound = errors.New("Passkey 会话不存在或已过期")
+
+func SaveSessionData(c *gin.Context, key string, data *webauthn.SessionData) error {
+	session := sessions.Default(c)
+	if data == nil {
+		session.Delete(key)
+		return session.Save()
+	}
+	payload, err := json.Marshal(data)
+	if err != nil {
+		return err
+	}
+	session.Set(key, string(payload))
+	return session.Save()
+}
+
+func PopSessionData(c *gin.Context, key string) (*webauthn.SessionData, error) {
+	session := sessions.Default(c)
+	raw := session.Get(key)
+	if raw == nil {
+		return nil, errSessionNotFound
+	}
+	session.Delete(key)
+	_ = session.Save()
+	var data webauthn.SessionData
+	switch value := raw.(type) {
+	case string:
+		if err := json.Unmarshal([]byte(value), &data); err != nil {
+			return nil, err
+		}
+	case []byte:
+		if err := json.Unmarshal(value, &data); err != nil {
+			return nil, err
+		}
+	default:
+		return nil, errors.New("Passkey 会话格式无效")
+	}
+	return &data, nil
+}

service/passkey/user.go

@@ -0,0 +1,71 @@
+package passkey
+
+import (
+	"fmt"
+	"strconv"
+	"strings"
+
+	"github.com/QuantumNous/new-api/model"
+
+	webauthn "github.com/go-webauthn/webauthn/webauthn"
+)
+
+type WebAuthnUser struct {
+	user       *model.User
+	credential *model.PasskeyCredential
+}
+
+func NewWebAuthnUser(user *model.User, credential *model.PasskeyCredential) *WebAuthnUser {
+	return &WebAuthnUser{user: user, credential: credential}
+}
+
+func (u *WebAuthnUser) WebAuthnID() []byte {
+	if u == nil || u.user == nil {
+		return nil
+	}
+	return []byte(strconv.Itoa(u.user.Id))
+}
+
+func (u *WebAuthnUser) WebAuthnName() string {
+	if u == nil || u.user == nil {
+		return ""
+	}
+	name := strings.TrimSpace(u.user.Username)
+	if name == "" {
+		return fmt.Sprintf("user-%d", u.user.Id)
+	}
+	return name
+}
+
+func (u *WebAuthnUser) WebAuthnDisplayName() string {
+	if u == nil || u.user == nil {
+		return ""
+	}
+	display := strings.TrimSpace(u.user.DisplayName)
+	if display != "" {
+		return display
+	}
+	return u.WebAuthnName()
+}
+
+func (u *WebAuthnUser) WebAuthnCredentials() []webauthn.Credential {
+	if u == nil || u.credential == nil {
+		return nil
+	}
+	cred := u.credential.ToWebAuthnCredential()
+	return []webauthn.Credential{cred}
+}
+
+func (u *WebAuthnUser) ModelUser() *model.User {
+	if u == nil {
+		return nil
+	}
+	return u.user
+}
+
+func (u *WebAuthnUser) PasskeyCredential() *model.PasskeyCredential {
+	if u == nil {
+		return nil
+	}
+	return u.credential
+}

service/quota.go

@@ -0,0 +1,507 @@
+package service
+
+import (
+	"errors"
+	"fmt"
+	"log"
+	"math"
+	"strings"
+	"time"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/constant"
+	"github.com/QuantumNous/new-api/dto"
+	"github.com/QuantumNous/new-api/logger"
+	"github.com/QuantumNous/new-api/model"
+	relaycommon "github.com/QuantumNous/new-api/relay/common"
+	"github.com/QuantumNous/new-api/setting/ratio_setting"
+	"github.com/QuantumNous/new-api/setting/system_setting"
+	"github.com/QuantumNous/new-api/types"
+
+	"github.com/bytedance/gopkg/util/gopool"
+
+	"github.com/gin-gonic/gin"
+	"github.com/shopspring/decimal"
+)
+
+type TokenDetails struct {
+	TextTokens  int
+	AudioTokens int
+}
+
+type QuotaInfo struct {
+	InputDetails  TokenDetails
+	OutputDetails TokenDetails
+	ModelName     string
+	UsePrice      bool
+	ModelPrice    float64
+	ModelRatio    float64
+	GroupRatio    float64
+}
+
+func hasCustomModelRatio(modelName string, currentRatio float64) bool {
+	defaultRatio, exists := ratio_setting.GetDefaultModelRatioMap()[modelName]
+	if !exists {
+		return true
+	}
+	return currentRatio != defaultRatio
+}
+
+func calculateAudioQuota(info QuotaInfo) int {
+	if info.UsePrice {
+		modelPrice := decimal.NewFromFloat(info.ModelPrice)
+		quotaPerUnit := decimal.NewFromFloat(common.QuotaPerUnit)
+		groupRatio := decimal.NewFromFloat(info.GroupRatio)
+
+		quota := modelPrice.Mul(quotaPerUnit).Mul(groupRatio)
+		return int(quota.IntPart())
+	}
+
+	completionRatio := decimal.NewFromFloat(ratio_setting.GetCompletionRatio(info.ModelName))
+	audioRatio := decimal.NewFromFloat(ratio_setting.GetAudioRatio(info.ModelName))
+	audioCompletionRatio := decimal.NewFromFloat(ratio_setting.GetAudioCompletionRatio(info.ModelName))
+
+	groupRatio := decimal.NewFromFloat(info.GroupRatio)
+	modelRatio := decimal.NewFromFloat(info.ModelRatio)
+	ratio := groupRatio.Mul(modelRatio)
+
+	inputTextTokens := decimal.NewFromInt(int64(info.InputDetails.TextTokens))
+	outputTextTokens := decimal.NewFromInt(int64(info.OutputDetails.TextTokens))
+	inputAudioTokens := decimal.NewFromInt(int64(info.InputDetails.AudioTokens))
+	outputAudioTokens := decimal.NewFromInt(int64(info.OutputDetails.AudioTokens))
+
+	quota := decimal.Zero
+	quota = quota.Add(inputTextTokens)
+	quota = quota.Add(outputTextTokens.Mul(completionRatio))
+	quota = quota.Add(inputAudioTokens.Mul(audioRatio))
+	quota = quota.Add(outputAudioTokens.Mul(audioRatio).Mul(audioCompletionRatio))
+
+	quota = quota.Mul(ratio)
+
+	// If ratio is not zero and quota is less than or equal to zero, set quota to 1
+	if !ratio.IsZero() && quota.LessThanOrEqual(decimal.Zero) {
+		quota = decimal.NewFromInt(1)
+	}
+
+	return int(quota.Round(0).IntPart())
+}
+
+func PreWssConsumeQuota(ctx *gin.Context, relayInfo *relaycommon.RelayInfo, usage *dto.RealtimeUsage) error {
+	if relayInfo.UsePrice {
+		return nil
+	}
+	userQuota, err := model.GetUserQuota(relayInfo.UserId, false)
+	if err != nil {
+		return err
+	}
+
+	token, err := model.GetTokenByKey(strings.TrimPrefix(relayInfo.TokenKey, "sk-"), false)
+	if err != nil {
+		return err
+	}
+
+	modelName := relayInfo.OriginModelName
+	textInputTokens := usage.InputTokenDetails.TextTokens
+	textOutTokens := usage.OutputTokenDetails.TextTokens
+	audioInputTokens := usage.InputTokenDetails.AudioTokens
+	audioOutTokens := usage.OutputTokenDetails.AudioTokens
+	groupRatio := ratio_setting.GetGroupRatio(relayInfo.UsingGroup)
+	modelRatio, _, _ := ratio_setting.GetModelRatio(modelName)
+
+	autoGroup, exists := common.GetContextKey(ctx, constant.ContextKeyAutoGroup)
+	if exists {
+		groupRatio = ratio_setting.GetGroupRatio(autoGroup.(string))
+		log.Printf("final group ratio: %f", groupRatio)
+		relayInfo.UsingGroup = autoGroup.(string)
+	}
+
+	actualGroupRatio := groupRatio
+	userGroupRatio, ok := ratio_setting.GetGroupGroupRatio(relayInfo.UserGroup, relayInfo.UsingGroup)
+	if ok {
+		actualGroupRatio = userGroupRatio
+	}
+
+	quotaInfo := QuotaInfo{
+		InputDetails: TokenDetails{
+			TextTokens:  textInputTokens,
+			AudioTokens: audioInputTokens,
+		},
+		OutputDetails: TokenDetails{
+			TextTokens:  textOutTokens,
+			AudioTokens: audioOutTokens,
+		},
+		ModelName:  modelName,
+		UsePrice:   relayInfo.UsePrice,
+		ModelRatio: modelRatio,
+		GroupRatio: actualGroupRatio,
+	}
+
+	quota := calculateAudioQuota(quotaInfo)
+
+	if userQuota < quota {
+		return fmt.Errorf("user quota is not enough, user quota: %s, need quota: %s", logger.FormatQuota(userQuota), logger.FormatQuota(quota))
+	}
+
+	if !token.UnlimitedQuota && token.RemainQuota < quota {
+		return fmt.Errorf("token quota is not enough, token remain quota: %s, need quota: %s", logger.FormatQuota(token.RemainQuota), logger.FormatQuota(quota))
+	}
+
+	err = PostConsumeQuota(relayInfo, quota, 0, false)
+	if err != nil {
+		return err
+	}
+	logger.LogInfo(ctx, "realtime streaming consume quota success, quota: "+fmt.Sprintf("%d", quota))
+	return nil
+}
+
+func PostWssConsumeQuota(ctx *gin.Context, relayInfo *relaycommon.RelayInfo, modelName string,
+	usage *dto.RealtimeUsage, extraContent string) {
+
+	useTimeSeconds := time.Now().Unix() - relayInfo.StartTime.Unix()
+	textInputTokens := usage.InputTokenDetails.TextTokens
+	textOutTokens := usage.OutputTokenDetails.TextTokens
+
+	audioInputTokens := usage.InputTokenDetails.AudioTokens
+	audioOutTokens := usage.OutputTokenDetails.AudioTokens
+
+	tokenName := ctx.GetString("token_name")
+	completionRatio := decimal.NewFromFloat(ratio_setting.GetCompletionRatio(modelName))
+	audioRatio := decimal.NewFromFloat(ratio_setting.GetAudioRatio(relayInfo.OriginModelName))
+	audioCompletionRatio := decimal.NewFromFloat(ratio_setting.GetAudioCompletionRatio(modelName))
+
+	modelRatio := relayInfo.PriceData.ModelRatio
+	groupRatio := relayInfo.PriceData.GroupRatioInfo.GroupRatio
+	modelPrice := relayInfo.PriceData.ModelPrice
+	usePrice := relayInfo.PriceData.UsePrice
+
+	quotaInfo := QuotaInfo{
+		InputDetails: TokenDetails{
+			TextTokens:  textInputTokens,
+			AudioTokens: audioInputTokens,
+		},
+		OutputDetails: TokenDetails{
+			TextTokens:  textOutTokens,
+			AudioTokens: audioOutTokens,
+		},
+		ModelName:  modelName,
+		UsePrice:   usePrice,
+		ModelRatio: modelRatio,
+		GroupRatio: groupRatio,
+	}
+
+	quota := calculateAudioQuota(quotaInfo)
+
+	totalTokens := usage.TotalTokens
+	var logContent string
+	if !usePrice {
+		logContent = fmt.Sprintf("模型倍率 %.2f，补全倍率 %.2f，音频倍率 %.2f，音频补全倍率 %.2f，分组倍率 %.2f",
+			modelRatio, completionRatio.InexactFloat64(), audioRatio.InexactFloat64(), audioCompletionRatio.InexactFloat64(), groupRatio)
+	} else {
+		logContent = fmt.Sprintf("模型价格 %.2f，分组倍率 %.2f", modelPrice, groupRatio)
+	}
+
+	// record all the consume log even if quota is 0
+	if totalTokens == 0 {
+		// in this case, must be some error happened
+		// we cannot just return, because we may have to return the pre-consumed quota
+		quota = 0
+		logContent += fmt.Sprintf("（可能是上游超时）")
+		logger.LogError(ctx, fmt.Sprintf("total tokens is 0, cannot consume quota, userId %d, channelId %d, "+
+			"tokenId %d, model %s， pre-consumed quota %d", relayInfo.UserId, relayInfo.ChannelId, relayInfo.TokenId, modelName, relayInfo.FinalPreConsumedQuota))
+	} else {
+		model.UpdateUserUsedQuotaAndRequestCount(relayInfo.UserId, quota)
+		model.UpdateChannelUsedQuota(relayInfo.ChannelId, quota)
+	}
+
+	logModel := modelName
+	if extraContent != "" {
+		logContent += ", " + extraContent
+	}
+	other := GenerateWssOtherInfo(ctx, relayInfo, usage, modelRatio, groupRatio,
+		completionRatio.InexactFloat64(), audioRatio.InexactFloat64(), audioCompletionRatio.InexactFloat64(), modelPrice, relayInfo.PriceData.GroupRatioInfo.GroupSpecialRatio)
+	model.RecordConsumeLog(ctx, relayInfo.UserId, model.RecordConsumeLogParams{
+		ChannelId:        relayInfo.ChannelId,
+		PromptTokens:     usage.InputTokens,
+		CompletionTokens: usage.OutputTokens,
+		ModelName:        logModel,
+		TokenName:        tokenName,
+		Quota:            quota,
+		Content:          logContent,
+		TokenId:          relayInfo.TokenId,
+		UseTimeSeconds:   int(useTimeSeconds),
+		IsStream:         relayInfo.IsStream,
+		Group:            relayInfo.UsingGroup,
+		Other:            other,
+	})
+}
+
+func CalcOpenRouterCacheCreateTokens(usage dto.Usage, priceData types.PriceData) int {
+	if priceData.CacheCreationRatio == 1 {
+		return 0
+	}
+	quotaPrice := priceData.ModelRatio / common.QuotaPerUnit
+	promptCacheCreatePrice := quotaPrice * priceData.CacheCreationRatio
+	promptCacheReadPrice := quotaPrice * priceData.CacheRatio
+	completionPrice := quotaPrice * priceData.CompletionRatio
+
+	cost, _ := usage.Cost.(float64)
+	totalPromptTokens := float64(usage.PromptTokens)
+	completionTokens := float64(usage.CompletionTokens)
+	promptCacheReadTokens := float64(usage.PromptTokensDetails.CachedTokens)
+
+	return int(math.Round((cost -
+		totalPromptTokens*quotaPrice +
+		promptCacheReadTokens*(quotaPrice-promptCacheReadPrice) -
+		completionTokens*completionPrice) /
+		(promptCacheCreatePrice - quotaPrice)))
+}
+
+func PostAudioConsumeQuota(ctx *gin.Context, relayInfo *relaycommon.RelayInfo, usage *dto.Usage, extraContent string) {
+
+	useTimeSeconds := time.Now().Unix() - relayInfo.StartTime.Unix()
+	textInputTokens := usage.PromptTokensDetails.TextTokens
+	textOutTokens := usage.CompletionTokenDetails.TextTokens
+
+	audioInputTokens := usage.PromptTokensDetails.AudioTokens
+	audioOutTokens := usage.CompletionTokenDetails.AudioTokens
+
+	tokenName := ctx.GetString("token_name")
+	completionRatio := decimal.NewFromFloat(ratio_setting.GetCompletionRatio(relayInfo.OriginModelName))
+	audioRatio := decimal.NewFromFloat(ratio_setting.GetAudioRatio(relayInfo.OriginModelName))
+	audioCompletionRatio := decimal.NewFromFloat(ratio_setting.GetAudioCompletionRatio(relayInfo.OriginModelName))
+
+	modelRatio := relayInfo.PriceData.ModelRatio
+	groupRatio := relayInfo.PriceData.GroupRatioInfo.GroupRatio
+	modelPrice := relayInfo.PriceData.ModelPrice
+	usePrice := relayInfo.PriceData.UsePrice
+
+	quotaInfo := QuotaInfo{
+		InputDetails: TokenDetails{
+			TextTokens:  textInputTokens,
+			AudioTokens: audioInputTokens,
+		},
+		OutputDetails: TokenDetails{
+			TextTokens:  textOutTokens,
+			AudioTokens: audioOutTokens,
+		},
+		ModelName:  relayInfo.OriginModelName,
+		UsePrice:   usePrice,
+		ModelRatio: modelRatio,
+		GroupRatio: groupRatio,
+	}
+
+	quota := calculateAudioQuota(quotaInfo)
+
+	totalTokens := usage.TotalTokens
+	var logContent string
+	if !usePrice {
+		logContent = fmt.Sprintf("模型倍率 %.2f，补全倍率 %.2f，音频倍率 %.2f，音频补全倍率 %.2f，分组倍率 %.2f",
+			modelRatio, completionRatio.InexactFloat64(), audioRatio.InexactFloat64(), audioCompletionRatio.InexactFloat64(), groupRatio)
+	} else {
+		logContent = fmt.Sprintf("模型价格 %.2f，分组倍率 %.2f", modelPrice, groupRatio)
+	}
+
+	// record all the consume log even if quota is 0
+	if totalTokens == 0 {
+		// in this case, must be some error happened
+		// we cannot just return, because we may have to return the pre-consumed quota
+		quota = 0
+		logContent += fmt.Sprintf("（可能是上游超时）")
+		logger.LogError(ctx, fmt.Sprintf("total tokens is 0, cannot consume quota, userId %d, channelId %d, "+
+			"tokenId %d, model %s， pre-consumed quota %d", relayInfo.UserId, relayInfo.ChannelId, relayInfo.TokenId, relayInfo.OriginModelName, relayInfo.FinalPreConsumedQuota))
+	} else {
+		model.UpdateUserUsedQuotaAndRequestCount(relayInfo.UserId, quota)
+		model.UpdateChannelUsedQuota(relayInfo.ChannelId, quota)
+	}
+
+	if err := SettleBilling(ctx, relayInfo, quota); err != nil {
+		logger.LogError(ctx, "error settling billing: "+err.Error())
+	}
+
+	logModel := relayInfo.OriginModelName
+	if extraContent != "" {
+		logContent += ", " + extraContent
+	}
+	other := GenerateAudioOtherInfo(ctx, relayInfo, usage, modelRatio, groupRatio,
+		completionRatio.InexactFloat64(), audioRatio.InexactFloat64(), audioCompletionRatio.InexactFloat64(), modelPrice, relayInfo.PriceData.GroupRatioInfo.GroupSpecialRatio)
+	model.RecordConsumeLog(ctx, relayInfo.UserId, model.RecordConsumeLogParams{
+		ChannelId:        relayInfo.ChannelId,
+		PromptTokens:     usage.PromptTokens,
+		CompletionTokens: usage.CompletionTokens,
+		ModelName:        logModel,
+		TokenName:        tokenName,
+		Quota:            quota,
+		Content:          logContent,
+		TokenId:          relayInfo.TokenId,
+		UseTimeSeconds:   int(useTimeSeconds),
+		IsStream:         relayInfo.IsStream,
+		Group:            relayInfo.UsingGroup,
+		Other:            other,
+	})
+}
+
+func PreConsumeTokenQuota(relayInfo *relaycommon.RelayInfo, quota int) error {
+	if quota < 0 {
+		return errors.New("quota 不能为负数！")
+	}
+	if relayInfo.IsPlayground {
+		return nil
+	}
+	//if relayInfo.TokenUnlimited {
+	//	return nil
+	//}
+	token, err := model.GetTokenByKey(relayInfo.TokenKey, false)
+	if err != nil {
+		return err
+	}
+	if !relayInfo.TokenUnlimited && token.RemainQuota < quota {
+		return fmt.Errorf("token quota is not enough, token remain quota: %s, need quota: %s", logger.FormatQuota(token.RemainQuota), logger.FormatQuota(quota))
+	}
+	err = model.DecreaseTokenQuota(relayInfo.TokenId, relayInfo.TokenKey, quota)
+	if err != nil {
+		return err
+	}
+	return nil
+}
+
+func PostConsumeQuota(relayInfo *relaycommon.RelayInfo, quota int, preConsumedQuota int, sendEmail bool) (err error) {
+
+	// 1) Consume from wallet quota OR subscription item
+	if relayInfo != nil && relayInfo.BillingSource == BillingSourceSubscription {
+		if relayInfo.SubscriptionId == 0 {
+			return errors.New("subscription id is missing")
+		}
+		delta := int64(quota)
+		if delta != 0 {
+			if err := model.PostConsumeUserSubscriptionDelta(relayInfo.SubscriptionId, delta); err != nil {
+				return err
+			}
+			relayInfo.SubscriptionPostDelta += delta
+		}
+	} else {
+		// Wallet
+		if quota > 0 {
+			err = model.DecreaseUserQuota(relayInfo.UserId, quota)
+		} else {
+			err = model.IncreaseUserQuota(relayInfo.UserId, -quota, false)
+		}
+		if err != nil {
+			return err
+		}
+	}
+
+	if !relayInfo.IsPlayground {
+		if quota > 0 {
+			err = model.DecreaseTokenQuota(relayInfo.TokenId, relayInfo.TokenKey, quota)
+		} else {
+			err = model.IncreaseTokenQuota(relayInfo.TokenId, relayInfo.TokenKey, -quota)
+		}
+		if err != nil {
+			return err
+		}
+	}
+
+	if sendEmail {
+		if (quota + preConsumedQuota) != 0 {
+			checkAndSendQuotaNotify(relayInfo, quota, preConsumedQuota)
+		}
+	}
+
+	return nil
+}
+
+func checkAndSendQuotaNotify(relayInfo *relaycommon.RelayInfo, quota int, preConsumedQuota int) {
+	gopool.Go(func() {
+		userSetting := relayInfo.UserSetting
+		threshold := common.QuotaRemindThreshold
+		if userSetting.QuotaWarningThreshold != 0 {
+			threshold = int(userSetting.QuotaWarningThreshold)
+		}
+
+		//noMoreQuota := userCache.Quota-(quota+preConsumedQuota) <= 0
+		quotaTooLow := false
+		consumeQuota := quota + preConsumedQuota
+		if relayInfo.UserQuota-consumeQuota < threshold {
+			quotaTooLow = true
+		}
+		if quotaTooLow {
+			prompt := "您的额度即将用尽"
+			topUpLink := fmt.Sprintf("%s/console/topup", system_setting.ServerAddress)
+
+			// 根据通知方式生成不同的内容格式
+			var content string
+			var values []interface{}
+
+			notifyType := userSetting.NotifyType
+			if notifyType == "" {
+				notifyType = dto.NotifyTypeEmail
+			}
+
+			if notifyType == dto.NotifyTypeBark {
+				// Bark推送使用简短文本，不支持HTML
+				content = "{{value}}，剩余额度：{{value}}，请及时充值"
+				values = []interface{}{prompt, logger.FormatQuota(relayInfo.UserQuota)}
+			} else if notifyType == dto.NotifyTypeGotify {
+				content = "{{value}}，当前剩余额度为 {{value}}，请及时充值。"
+				values = []interface{}{prompt, logger.FormatQuota(relayInfo.UserQuota)}
+			} else {
+				// 默认内容格式，适用于Email和Webhook（支持HTML）
+				content = "{{value}}，当前剩余额度为 {{value}}，为了不影响您的使用，请及时充值。<br/>充值链接：<a href='{{value}}'>{{value}}</a>"
+				values = []interface{}{prompt, logger.FormatQuota(relayInfo.UserQuota), topUpLink, topUpLink}
+			}
+
+			err := NotifyUser(relayInfo.UserId, relayInfo.UserEmail, relayInfo.UserSetting, dto.NewNotify(dto.NotifyTypeQuotaExceed, prompt, content, values))
+			if err != nil {
+				common.SysError(fmt.Sprintf("failed to send quota notify to user %d: %s", relayInfo.UserId, err.Error()))
+			}
+		}
+	})
+}
+
+func checkAndSendSubscriptionQuotaNotify(relayInfo *relaycommon.RelayInfo) {
+	gopool.Go(func() {
+		if relayInfo == nil {
+			return
+		}
+		if relayInfo.SubscriptionId == 0 || relayInfo.SubscriptionAmountTotal <= 0 {
+			return
+		}
+
+		userSetting := relayInfo.UserSetting
+		threshold := common.QuotaRemindThreshold
+		if userSetting.QuotaWarningThreshold != 0 {
+			threshold = int(userSetting.QuotaWarningThreshold)
+		}
+
+		usedAfter := relayInfo.SubscriptionAmountUsedAfterPreConsume + relayInfo.SubscriptionPostDelta
+		remaining := relayInfo.SubscriptionAmountTotal - usedAfter
+		if remaining >= int64(threshold) {
+			return
+		}
+
+		prompt := "您的订阅额度即将用尽"
+		topUpLink := fmt.Sprintf("%s/console/topup", system_setting.ServerAddress)
+
+		var content string
+		var values []interface{}
+		notifyType := userSetting.NotifyType
+		if notifyType == "" {
+			notifyType = dto.NotifyTypeEmail
+		}
+
+		if notifyType == dto.NotifyTypeBark {
+			content = "{{value}}，剩余额度：{{value}}，请及时充值"
+			values = []interface{}{prompt, logger.FormatQuota(int(remaining))}
+		} else if notifyType == dto.NotifyTypeGotify {
+			content = "{{value}}，当前剩余额度为 {{value}}，请及时充值。"
+			values = []interface{}{prompt, logger.FormatQuota(int(remaining))}
+		} else {
+			content = "{{value}}，当前剩余额度为 {{value}}，为了不影响您的使用，请及时充值。<br/>充值链接：<a href='{{value}}'>{{value}}</a>"
+			values = []interface{}{prompt, logger.FormatQuota(int(remaining)), topUpLink, topUpLink}
+		}
+
+		if err := NotifyUser(relayInfo.UserId, relayInfo.UserEmail, relayInfo.UserSetting, dto.NewNotify(dto.NotifyTypeQuotaExceed, prompt, content, values)); err != nil {
+			common.SysError(fmt.Sprintf("failed to send subscription quota notify to user %d: %s", relayInfo.UserId, err.Error()))
+		}
+	})
+}

service/sensitive.go

@@ -0,0 +1,77 @@
+package service
+
+import (
+	"errors"
+	"strings"
+
+	"github.com/QuantumNous/new-api/dto"
+	"github.com/QuantumNous/new-api/setting"
+)
+
+func CheckSensitiveMessages(messages []dto.Message) ([]string, error) {
+	if len(messages) == 0 {
+		return nil, nil
+	}
+
+	for _, message := range messages {
+		arrayContent := message.ParseContent()
+		for _, m := range arrayContent {
+			if m.Type == "image_url" {
+				// TODO: check image url
+				continue
+			}
+			// 检查 text 是否为空
+			if m.Text == "" {
+				continue
+			}
+			if ok, words := SensitiveWordContains(m.Text); ok {
+				return words, errors.New("sensitive words detected")
+			}
+		}
+	}
+	return nil, nil
+}
+
+func CheckSensitiveText(text string) (bool, []string) {
+	return SensitiveWordContains(text)
+}
+
+// SensitiveWordContains 是否包含敏感词，返回是否包含敏感词和敏感词列表
+func SensitiveWordContains(text string) (bool, []string) {
+	if len(setting.SensitiveWords) == 0 {
+		return false, nil
+	}
+	if len(text) == 0 {
+		return false, nil
+	}
+	checkText := strings.ToLower(text)
+	return AcSearch(checkText, setting.SensitiveWords, true)
+}
+
+// SensitiveWordReplace 敏感词替换，返回是否包含敏感词和替换后的文本
+func SensitiveWordReplace(text string, returnImmediately bool) (bool, []string, string) {
+	if len(setting.SensitiveWords) == 0 {
+		return false, nil, text
+	}
+	checkText := strings.ToLower(text)
+	m := getOrBuildAC(setting.SensitiveWords)
+	hits := m.MultiPatternSearch([]rune(checkText), returnImmediately)
+	if len(hits) > 0 {
+		words := make([]string, 0, len(hits))
+		var builder strings.Builder
+		builder.Grow(len(text))
+		lastPos := 0
+
+		for _, hit := range hits {
+			pos := hit.Pos
+			word := string(hit.Word)
+			builder.WriteString(text[lastPos:pos])
+			builder.WriteString("**###**")
+			lastPos = pos + len(word)
+			words = append(words, word)
+		}
+		builder.WriteString(text[lastPos:])
+		return true, words, builder.String()
+	}
+	return false, nil, text
+}

service/str.go

@@ -0,0 +1,152 @@
+package service
+
+import (
+	"bytes"
+	"fmt"
+	"hash/fnv"
+	"sort"
+	"strings"
+	"sync"
+
+	goahocorasick "github.com/anknown/ahocorasick"
+)
+
+func SundaySearch(text string, pattern string) bool {
+	// 计算偏移表
+	offset := make(map[rune]int)
+	for i, c := range pattern {
+		offset[c] = len(pattern) - i
+	}
+
+	// 文本串长度和模式串长度
+	n, m := len(text), len(pattern)
+
+	// 主循环，i表示当前对齐的文本串位置
+	for i := 0; i <= n-m; {
+		// 检查子串
+		j := 0
+		for j < m && text[i+j] == pattern[j] {
+			j++
+		}
+		// 如果完全匹配，返回匹配位置
+		if j == m {
+			return true
+		}
+
+		// 如果还有剩余字符，则检查下一位字符在偏移表中的值
+		if i+m < n {
+			next := rune(text[i+m])
+			if val, ok := offset[next]; ok {
+				i += val // 存在于偏移表中，进行跳跃
+			} else {
+				i += len(pattern) + 1 // 不存在于偏移表中，跳过整个模式串长度
+			}
+		} else {
+			break
+		}
+	}
+	return false // 如果没有找到匹配，返回-1
+}
+
+func RemoveDuplicate(s []string) []string {
+	result := make([]string, 0, len(s))
+	temp := map[string]struct{}{}
+	for _, item := range s {
+		if _, ok := temp[item]; !ok {
+			temp[item] = struct{}{}
+			result = append(result, item)
+		}
+	}
+	return result
+}
+
+func InitAc(dict []string) *goahocorasick.Machine {
+	m := new(goahocorasick.Machine)
+	runes := readRunes(dict)
+	if err := m.Build(runes); err != nil {
+		fmt.Println(err)
+		return nil
+	}
+	return m
+}
+
+var acCache sync.Map
+
+func acKey(dict []string) string {
+	if len(dict) == 0 {
+		return ""
+	}
+	normalized := make([]string, 0, len(dict))
+	for _, w := range dict {
+		w = strings.ToLower(strings.TrimSpace(w))
+		if w != "" {
+			normalized = append(normalized, w)
+		}
+	}
+	if len(normalized) == 0 {
+		return ""
+	}
+	sort.Strings(normalized)
+	hasher := fnv.New64a()
+	for _, w := range normalized {
+		hasher.Write([]byte{0})
+		hasher.Write([]byte(w))
+	}
+	return fmt.Sprintf("%x", hasher.Sum64())
+}
+
+func getOrBuildAC(dict []string) *goahocorasick.Machine {
+	key := acKey(dict)
+	if key == "" {
+		return nil
+	}
+	if v, ok := acCache.Load(key); ok {
+		if m, ok2 := v.(*goahocorasick.Machine); ok2 {
+			return m
+		}
+	}
+	m := InitAc(dict)
+	if m == nil {
+		return nil
+	}
+	if actual, loaded := acCache.LoadOrStore(key, m); loaded {
+		if cached, ok := actual.(*goahocorasick.Machine); ok {
+			return cached
+		}
+	}
+	return m
+}
+
+func readRunes(dict []string) [][]rune {
+	var runes [][]rune
+
+	for _, word := range dict {
+		word = strings.ToLower(word)
+		l := bytes.TrimSpace([]byte(word))
+		runes = append(runes, bytes.Runes(l))
+	}
+
+	return runes
+}
+
+func AcSearch(findText string, dict []string, stopImmediately bool) (bool, []string) {
+	if len(dict) == 0 {
+		return false, nil
+	}
+	if len(findText) == 0 {
+		return false, nil
+	}
+	m := getOrBuildAC(dict)
+	if m == nil {
+		return false, nil
+	}
+	hits := m.MultiPatternSearch([]rune(findText), stopImmediately)
+	if len(hits) > 0 {
+		words := make([]string, 0)
+		for _, hit := range hits {
+			words = append(words, string(hit.Word))
+		}
+		return true, words
+	}
+	return false, nil
+}

service/subscription_reset_task.go

@@ -0,0 +1,93 @@
+package service
+
+import (
+	"context"
+	"fmt"
+	"sync"
+	"sync/atomic"
+	"time"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/logger"
+	"github.com/QuantumNous/new-api/model"
+
+	"github.com/bytedance/gopkg/util/gopool"
+)
+
+const (
+	subscriptionResetTickInterval = 1 * time.Minute
+	subscriptionResetBatchSize    = 300
+	subscriptionCleanupInterval   = 30 * time.Minute
+)
+
+var (
+	subscriptionResetOnce    sync.Once
+	subscriptionResetRunning atomic.Bool
+	subscriptionCleanupLast  atomic.Int64
+)
+
+func StartSubscriptionQuotaResetTask() {
+	subscriptionResetOnce.Do(func() {
+		if !common.IsMasterNode {
+			return
+		}
+		gopool.Go(func() {
+			logger.LogInfo(context.Background(), fmt.Sprintf("subscription quota reset task started: tick=%s", subscriptionResetTickInterval))
+			ticker := time.NewTicker(subscriptionResetTickInterval)
+			defer ticker.Stop()
+
+			runSubscriptionQuotaResetOnce()
+			for range ticker.C {
+				runSubscriptionQuotaResetOnce()
+			}
+		})
+	})
+}
+
+func runSubscriptionQuotaResetOnce() {
+	if !subscriptionResetRunning.CompareAndSwap(false, true) {
+		return
+	}
+	defer subscriptionResetRunning.Store(false)
+
+	ctx := context.Background()
+	totalReset := 0
+	totalExpired := 0
+	for {
+		n, err := model.ExpireDueSubscriptions(subscriptionResetBatchSize)
+		if err != nil {
+			logger.LogWarn(ctx, fmt.Sprintf("subscription expire task failed: %v", err))
+			return
+		}
+		if n == 0 {
+			break
+		}
+		totalExpired += n
+		if n < subscriptionResetBatchSize {
+			break
+		}
+	}
+	for {
+		n, err := model.ResetDueSubscriptions(subscriptionResetBatchSize)
+		if err != nil {
+			logger.LogWarn(ctx, fmt.Sprintf("subscription quota reset task failed: %v", err))
+			return
+		}
+		if n == 0 {
+			break
+		}
+		totalReset += n
+		if n < subscriptionResetBatchSize {
+			break
+		}
+	}
+	lastCleanup := time.Unix(subscriptionCleanupLast.Load(), 0)
+	if time.Since(lastCleanup) >= subscriptionCleanupInterval {
+		if _, err := model.CleanupSubscriptionPreConsumeRecords(7 * 24 * 3600); err == nil {
+			subscriptionCleanupLast.Store(time.Now().Unix())
+		}
+	}
+	if common.DebugEnabled && (totalReset > 0 || totalExpired > 0) {
+		logger.LogDebug(ctx, "subscription maintenance: reset_count=%d, expired_count=%d", totalReset, totalExpired)
+	}
+}

service/task.go

@@ -0,0 +1,11 @@
+package service
+
+import (
+	"strings"
+
+	"github.com/QuantumNous/new-api/constant"
+)
+
+func CoverTaskActionToModelName(platform constant.TaskPlatform, action string) string {
+	return strings.ToLower(string(platform)) + "_" + strings.ToLower(action)
+}

service/task_billing.go

@@ -0,0 +1,285 @@
+package service
+
+import (
+	"context"
+	"fmt"
+	"strings"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/constant"
+	"github.com/QuantumNous/new-api/logger"
+	"github.com/QuantumNous/new-api/model"
+	relaycommon "github.com/QuantumNous/new-api/relay/common"
+	"github.com/QuantumNous/new-api/setting/ratio_setting"
+	"github.com/gin-gonic/gin"
+)
+
+// LogTaskConsumption 记录任务消费日志和统计信息（仅记录，不涉及实际扣费）。
+// 实际扣费已由 BillingSession（PreConsumeBilling + SettleBilling）完成。
+func LogTaskConsumption(c *gin.Context, info *relaycommon.RelayInfo) {
+	tokenName := c.GetString("token_name")
+	logContent := fmt.Sprintf("操作 %s", info.Action)
+	// 支持任务仅按次计费
+	if common.StringsContains(constant.TaskPricePatches, info.OriginModelName) {
+		logContent = fmt.Sprintf("%s，按次计费", logContent)
+	} else {
+		if len(info.PriceData.OtherRatios) > 0 {
+			var contents []string
+			for key, ra := range info.PriceData.OtherRatios {
+				if 1.0 != ra {
+					contents = append(contents, fmt.Sprintf("%s: %.2f", key, ra))
+				}
+			}
+			if len(contents) > 0 {
+				logContent = fmt.Sprintf("%s, 计算参数：%s", logContent, strings.Join(contents, ", "))
+			}
+		}
+	}
+	other := make(map[string]interface{})
+	other["request_path"] = c.Request.URL.Path
+	other["model_price"] = info.PriceData.ModelPrice
+	other["group_ratio"] = info.PriceData.GroupRatioInfo.GroupRatio
+	if info.PriceData.GroupRatioInfo.HasSpecialRatio {
+		other["user_group_ratio"] = info.PriceData.GroupRatioInfo.GroupSpecialRatio
+	}
+	if info.IsModelMapped {
+		other["is_model_mapped"] = true
+		other["upstream_model_name"] = info.UpstreamModelName
+	}
+	model.RecordConsumeLog(c, info.UserId, model.RecordConsumeLogParams{
+		ChannelId: info.ChannelId,
+		ModelName: info.OriginModelName,
+		TokenName: tokenName,
+		Quota:     info.PriceData.Quota,
+		Content:   logContent,
+		TokenId:   info.TokenId,
+		Group:     info.UsingGroup,
+		Other:     other,
+	})
+	model.UpdateUserUsedQuotaAndRequestCount(info.UserId, info.PriceData.Quota)
+	model.UpdateChannelUsedQuota(info.ChannelId, info.PriceData.Quota)
+}
+
+// ---------------------------------------------------------------------------
+// 异步任务计费辅助函数
+// ---------------------------------------------------------------------------
+
+// resolveTokenKey 通过 TokenId 运行时获取令牌 Key（用于 Redis 缓存操作）。
+// 如果令牌已被删除或查询失败，返回空字符串。
+func resolveTokenKey(ctx context.Context, tokenId int, taskID string) string {
+	token, err := model.GetTokenById(tokenId)
+	if err != nil {
+		logger.LogWarn(ctx, fmt.Sprintf("获取令牌 key 失败 (tokenId=%d, task=%s): %s", tokenId, taskID, err.Error()))
+		return ""
+	}
+	return token.Key
+}
+
+// taskIsSubscription 判断任务是否通过订阅计费。
+func taskIsSubscription(task *model.Task) bool {
+	return task.PrivateData.BillingSource == BillingSourceSubscription && task.PrivateData.SubscriptionId > 0
+}
+
+// taskAdjustFunding 调整任务的资金来源（钱包或订阅），delta > 0 表示扣费，delta < 0 表示退还。
+func taskAdjustFunding(task *model.Task, delta int) error {
+	if taskIsSubscription(task) {
+		return model.PostConsumeUserSubscriptionDelta(task.PrivateData.SubscriptionId, int64(delta))
+	}
+	if delta > 0 {
+		return model.DecreaseUserQuota(task.UserId, delta)
+	}
+	return model.IncreaseUserQuota(task.UserId, -delta, false)
+}
+
+// taskAdjustTokenQuota 调整任务的令牌额度，delta > 0 表示扣费，delta < 0 表示退还。
+// 需要通过 resolveTokenKey 运行时获取 key（不从 PrivateData 中读取）。
+func taskAdjustTokenQuota(ctx context.Context, task *model.Task, delta int) {
+	if task.PrivateData.TokenId <= 0 || delta == 0 {
+		return
+	}
+	tokenKey := resolveTokenKey(ctx, task.PrivateData.TokenId, task.TaskID)
+	if tokenKey == "" {
+		return
+	}
+	var err error
+	if delta > 0 {
+		err = model.DecreaseTokenQuota(task.PrivateData.TokenId, tokenKey, delta)
+	} else {
+		err = model.IncreaseTokenQuota(task.PrivateData.TokenId, tokenKey, -delta)
+	}
+	if err != nil {
+		logger.LogWarn(ctx, fmt.Sprintf("调整令牌额度失败 (delta=%d, task=%s): %s", delta, task.TaskID, err.Error()))
+	}
+}
+
+// taskBillingOther 从 task 的 BillingContext 构建日志 Other 字段。
+func taskBillingOther(task *model.Task) map[string]interface{} {
+	other := make(map[string]interface{})
+	if bc := task.PrivateData.BillingContext; bc != nil {
+		other["model_price"] = bc.ModelPrice
+		other["group_ratio"] = bc.GroupRatio
+		if len(bc.OtherRatios) > 0 {
+			for k, v := range bc.OtherRatios {
+				other[k] = v
+			}
+		}
+	}
+	props := task.Properties
+	if props.UpstreamModelName != "" && props.UpstreamModelName != props.OriginModelName {
+		other["is_model_mapped"] = true
+		other["upstream_model_name"] = props.UpstreamModelName
+	}
+	return other
+}
+
+// taskModelName 从 BillingContext 或 Properties 中获取模型名称。
+func taskModelName(task *model.Task) string {
+	if bc := task.PrivateData.BillingContext; bc != nil && bc.OriginModelName != "" {
+		return bc.OriginModelName
+	}
+	return task.Properties.OriginModelName
+}
+
+// RefundTaskQuota 统一的任务失败退款逻辑。
+// 当异步任务失败时，将预扣的 quota 退还给用户（支持钱包和订阅），并退还令牌额度。
+func RefundTaskQuota(ctx context.Context, task *model.Task, reason string) {
+	quota := task.Quota
+	if quota == 0 {
+		return
+	}
+
+	// 1. 退还资金来源（钱包或订阅）
+	if err := taskAdjustFunding(task, -quota); err != nil {
+		logger.LogWarn(ctx, fmt.Sprintf("退还资金来源失败 task %s: %s", task.TaskID, err.Error()))
+		return
+	}
+
+	// 2. 退还令牌额度
+	taskAdjustTokenQuota(ctx, task, -quota)
+
+	// 3. 记录日志
+	other := taskBillingOther(task)
+	other["task_id"] = task.TaskID
+	other["reason"] = reason
+	model.RecordTaskBillingLog(model.RecordTaskBillingLogParams{
+		UserId:    task.UserId,
+		LogType:   model.LogTypeRefund,
+		Content:   "",
+		ChannelId: task.ChannelId,
+		ModelName: taskModelName(task),
+		Quota:     quota,
+		TokenId:   task.PrivateData.TokenId,
+		Group:     task.Group,
+		Other:     other,
+	})
+}
+
+// RecalculateTaskQuota 通用的异步差额结算。
+// actualQuota 是任务完成后的实际应扣额度，与预扣额度 (task.Quota) 做差额结算。
+// reason 用于日志记录（例如 "token重算" 或 "adaptor调整"）。
+func RecalculateTaskQuota(ctx context.Context, task *model.Task, actualQuota int, reason string) {
+	if actualQuota <= 0 {
+		return
+	}
+	preConsumedQuota := task.Quota
+	quotaDelta := actualQuota - preConsumedQuota
+
+	if quotaDelta == 0 {
+		logger.LogInfo(ctx, fmt.Sprintf("任务 %s 预扣费准确（%s，%s）",
+			task.TaskID, logger.LogQuota(actualQuota), reason))
+		return
+	}
+
+	logger.LogInfo(ctx, fmt.Sprintf("任务 %s 差额结算：delta=%s（实际：%s，预扣：%s，%s）",
+		task.TaskID,
+		logger.LogQuota(quotaDelta),
+		logger.LogQuota(actualQuota),
+		logger.LogQuota(preConsumedQuota),
+		reason,
+	))
+
+	// 调整资金来源
+	if err := taskAdjustFunding(task, quotaDelta); err != nil {
+		logger.LogError(ctx, fmt.Sprintf("差额结算资金调整失败 task %s: %s", task.TaskID, err.Error()))
+		return
+	}
+
+	// 调整令牌额度
+	taskAdjustTokenQuota(ctx, task, quotaDelta)
+
+	task.Quota = actualQuota
+
+	var logType int
+	var logQuota int
+	if quotaDelta > 0 {
+		logType = model.LogTypeConsume
+		logQuota = quotaDelta
+		model.UpdateUserUsedQuotaAndRequestCount(task.UserId, quotaDelta)
+		model.UpdateChannelUsedQuota(task.ChannelId, quotaDelta)
+	} else {
+		logType = model.LogTypeRefund
+		logQuota = -quotaDelta
+	}
+	other := taskBillingOther(task)
+	other["task_id"] = task.TaskID
+	//other["reason"] = reason
+	other["pre_consumed_quota"] = preConsumedQuota
+	other["actual_quota"] = actualQuota
+	model.RecordTaskBillingLog(model.RecordTaskBillingLogParams{
+		UserId:    task.UserId,
+		LogType:   logType,
+		Content:   reason,
+		ChannelId: task.ChannelId,
+		ModelName: taskModelName(task),
+		Quota:     logQuota,
+		TokenId:   task.PrivateData.TokenId,
+		Group:     task.Group,
+		Other:     other,
+	})
+}
+
+// RecalculateTaskQuotaByTokens 根据实际 token 消耗重新计费（异步差额结算）。
+// 当任务成功且返回了 totalTokens 时，根据模型倍率和分组倍率重新计算实际扣费额度，
+// 与预扣费的差额进行补扣或退还。支持钱包和订阅计费来源。
+func RecalculateTaskQuotaByTokens(ctx context.Context, task *model.Task, totalTokens int) {
+	if totalTokens <= 0 {
+		return
+	}
+
+	modelName := taskModelName(task)
+
+	// 获取模型价格和倍率
+	modelRatio, hasRatioSetting, _ := ratio_setting.GetModelRatio(modelName)
+	// 只有配置了倍率(非固定价格)时才按 token 重新计费
+	if !hasRatioSetting || modelRatio <= 0 {
+		return
+	}
+
+	// 获取用户和组的倍率信息
+	group := task.Group
+	if group == "" {
+		user, err := model.GetUserById(task.UserId, false)
+		if err == nil {
+			group = user.Group
+		}
+	}
+	if group == "" {
+		return
+	}
+
+	groupRatio := ratio_setting.GetGroupRatio(group)
+	userGroupRatio, hasUserGroupRatio := ratio_setting.GetGroupGroupRatio(group, group)
+
+	var finalGroupRatio float64
+	if hasUserGroupRatio {
+		finalGroupRatio = userGroupRatio
+	} else {
+		finalGroupRatio = groupRatio
+	}
+
+	// 计算实际应扣费额度: totalTokens * modelRatio * groupRatio
+	actualQuota := int(float64(totalTokens) * modelRatio * finalGroupRatio)
+
+	reason := fmt.Sprintf("token重算：tokens=%d, modelRatio=%.2f, groupRatio=%.2f", totalTokens, modelRatio, finalGroupRatio)
+	RecalculateTaskQuota(ctx, task, actualQuota, reason)
+}

service/task_billing_test.go

@@ -0,0 +1,714 @@
+package service
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"os"
+	"testing"
+	"time"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/model"
+	relaycommon "github.com/QuantumNous/new-api/relay/common"
+	"github.com/glebarez/sqlite"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"gorm.io/gorm"
+)
+
+func TestMain(m *testing.M) {
+	db, err := gorm.Open(sqlite.Open(":memory:"), &gorm.Config{})
+	if err != nil {
+		panic("failed to open test db: " + err.Error())
+	}
+	sqlDB, err := db.DB()
+	if err != nil {
+		panic("failed to get sql.DB: " + err.Error())
+	}
+	sqlDB.SetMaxOpenConns(1)
+
+	model.DB = db
+	model.LOG_DB = db
+
+	common.UsingSQLite = true
+	common.RedisEnabled = false
+	common.BatchUpdateEnabled = false
+	common.LogConsumeEnabled = true
+
+	if err := db.AutoMigrate(
+		&model.Task{},
+		&model.User{},
+		&model.Token{},
+		&model.Log{},
+		&model.Channel{},
+		&model.UserSubscription{},
+	); err != nil {
+		panic("failed to migrate: " + err.Error())
+	}
+
+	os.Exit(m.Run())
+}
+
+// ---------------------------------------------------------------------------
+// Seed helpers
+// ---------------------------------------------------------------------------
+
+func truncate(t *testing.T) {
+	t.Helper()
+	t.Cleanup(func() {
+		model.DB.Exec("DELETE FROM tasks")
+		model.DB.Exec("DELETE FROM users")
+		model.DB.Exec("DELETE FROM tokens")
+		model.DB.Exec("DELETE FROM logs")
+		model.DB.Exec("DELETE FROM channels")
+		model.DB.Exec("DELETE FROM user_subscriptions")
+	})
+}
+
+func seedUser(t *testing.T, id int, quota int) {
+	t.Helper()
+	user := &model.User{Id: id, Username: "test_user", Quota: quota, Status: common.UserStatusEnabled}
+	require.NoError(t, model.DB.Create(user).Error)
+}
+
+func seedToken(t *testing.T, id int, userId int, key string, remainQuota int) {
+	t.Helper()
+	token := &model.Token{
+		Id:          id,
+		UserId:      userId,
+		Key:         key,
+		Name:        "test_token",
+		Status:      common.TokenStatusEnabled,
+		RemainQuota: remainQuota,
+		UsedQuota:   0,
+	}
+	require.NoError(t, model.DB.Create(token).Error)
+}
+
+func seedSubscription(t *testing.T, id int, userId int, amountTotal int64, amountUsed int64) {
+	t.Helper()
+	sub := &model.UserSubscription{
+		Id:          id,
+		UserId:      userId,
+		AmountTotal: amountTotal,
+		AmountUsed:  amountUsed,
+		Status:      "active",
+		StartTime:   time.Now().Unix(),
+		EndTime:     time.Now().Add(30 * 24 * time.Hour).Unix(),
+	}
+	require.NoError(t, model.DB.Create(sub).Error)
+}
+
+func seedChannel(t *testing.T, id int) {
+	t.Helper()
+	ch := &model.Channel{Id: id, Name: "test_channel", Key: "sk-test", Status: common.ChannelStatusEnabled}
+	require.NoError(t, model.DB.Create(ch).Error)
+}
+
+func makeTask(userId, channelId, quota, tokenId int, billingSource string, subscriptionId int) *model.Task {
+	return &model.Task{
+		TaskID:    "task_" + time.Now().Format("150405.000"),
+		UserId:    userId,
+		ChannelId: channelId,
+		Quota:     quota,
+		Status:    model.TaskStatus(model.TaskStatusInProgress),
+		Group:     "default",
+		Data:      json.RawMessage(`{}`),
+		CreatedAt: time.Now().Unix(),
+		UpdatedAt: time.Now().Unix(),
+		Properties: model.Properties{
+			OriginModelName: "test-model",
+		},
+		PrivateData: model.TaskPrivateData{
+			BillingSource:  billingSource,
+			SubscriptionId: subscriptionId,
+			TokenId:        tokenId,
+			BillingContext: &model.TaskBillingContext{
+				ModelPrice:      0.02,
+				GroupRatio:      1.0,
+				OriginModelName: "test-model",
+			},
+		},
+	}
+}
+
+// ---------------------------------------------------------------------------
+// Read-back helpers
+// ---------------------------------------------------------------------------
+
+func getUserQuota(t *testing.T, id int) int {
+	t.Helper()
+	var user model.User
+	require.NoError(t, model.DB.Select("quota").Where("id = ?", id).First(&user).Error)
+	return user.Quota
+}
+
+func getTokenRemainQuota(t *testing.T, id int) int {
+	t.Helper()
+	var token model.Token
+	require.NoError(t, model.DB.Select("remain_quota").Where("id = ?", id).First(&token).Error)
+	return token.RemainQuota
+}
+
+func getTokenUsedQuota(t *testing.T, id int) int {
+	t.Helper()
+	var token model.Token
+	require.NoError(t, model.DB.Select("used_quota").Where("id = ?", id).First(&token).Error)
+	return token.UsedQuota
+}
+
+func getSubscriptionUsed(t *testing.T, id int) int64 {
+	t.Helper()
+	var sub model.UserSubscription
+	require.NoError(t, model.DB.Select("amount_used").Where("id = ?", id).First(&sub).Error)
+	return sub.AmountUsed
+}
+
+func getLastLog(t *testing.T) *model.Log {
+	t.Helper()
+	var log model.Log
+	err := model.LOG_DB.Order("id desc").First(&log).Error
+	if err != nil {
+		return nil
+	}
+	return &log
+}
+
+func countLogs(t *testing.T) int64 {
+	t.Helper()
+	var count int64
+	model.LOG_DB.Model(&model.Log{}).Count(&count)
+	return count
+}
+
+// ===========================================================================
+// RefundTaskQuota tests
+// ===========================================================================
+
+func TestRefundTaskQuota_Wallet(t *testing.T) {
+	truncate(t)
+	ctx := context.Background()
+
+	const userID, tokenID, channelID = 1, 1, 1
+	const initQuota, preConsumed = 10000, 3000
+	const tokenRemain = 5000
+
+	seedUser(t, userID, initQuota)
+	seedToken(t, tokenID, userID, "sk-test-key", tokenRemain)
+	seedChannel(t, channelID)
+
+	task := makeTask(userID, channelID, preConsumed, tokenID, BillingSourceWallet, 0)
+
+	RefundTaskQuota(ctx, task, "task failed: upstream error")
+
+	// User quota should increase by preConsumed
+	assert.Equal(t, initQuota+preConsumed, getUserQuota(t, userID))
+
+	// Token remain_quota should increase, used_quota should decrease
+	assert.Equal(t, tokenRemain+preConsumed, getTokenRemainQuota(t, tokenID))
+	assert.Equal(t, -preConsumed, getTokenUsedQuota(t, tokenID))
+
+	// A refund log should be created
+	log := getLastLog(t)
+	require.NotNil(t, log)
+	assert.Equal(t, model.LogTypeRefund, log.Type)
+	assert.Equal(t, preConsumed, log.Quota)
+	assert.Equal(t, "test-model", log.ModelName)
+}
+
+func TestRefundTaskQuota_Subscription(t *testing.T) {
+	truncate(t)
+	ctx := context.Background()
+
+	const userID, tokenID, channelID, subID = 2, 2, 2, 1
+	const preConsumed = 2000
+	const subTotal, subUsed int64 = 100000, 50000
+	const tokenRemain = 8000
+
+	seedUser(t, userID, 0)
+	seedToken(t, tokenID, userID, "sk-sub-key", tokenRemain)
+	seedChannel(t, channelID)
+	seedSubscription(t, subID, userID, subTotal, subUsed)
+
+	task := makeTask(userID, channelID, preConsumed, tokenID, BillingSourceSubscription, subID)
+
+	RefundTaskQuota(ctx, task, "subscription task failed")
+
+	// Subscription used should decrease by preConsumed
+	assert.Equal(t, subUsed-int64(preConsumed), getSubscriptionUsed(t, subID))
+
+	// Token should also be refunded
+	assert.Equal(t, tokenRemain+preConsumed, getTokenRemainQuota(t, tokenID))
+
+	log := getLastLog(t)
+	require.NotNil(t, log)
+	assert.Equal(t, model.LogTypeRefund, log.Type)
+}
+
+func TestRefundTaskQuota_ZeroQuota(t *testing.T) {
+	truncate(t)
+	ctx := context.Background()
+
+	const userID = 3
+	seedUser(t, userID, 5000)
+
+	task := makeTask(userID, 0, 0, 0, BillingSourceWallet, 0)
+
+	RefundTaskQuota(ctx, task, "zero quota task")
+
+	// No change to user quota
+	assert.Equal(t, 5000, getUserQuota(t, userID))
+
+	// No log created
+	assert.Equal(t, int64(0), countLogs(t))
+}
+
+func TestRefundTaskQuota_NoToken(t *testing.T) {
+	truncate(t)
+	ctx := context.Background()
+
+	const userID, channelID = 4, 4
+	const initQuota, preConsumed = 10000, 1500
+
+	seedUser(t, userID, initQuota)
+	seedChannel(t, channelID)
+
+	task := makeTask(userID, channelID, preConsumed, 0, BillingSourceWallet, 0) // TokenId=0
+
+	RefundTaskQuota(ctx, task, "no token task failed")
+
+	// User quota refunded
+	assert.Equal(t, initQuota+preConsumed, getUserQuota(t, userID))
+
+	// Log created
+	log := getLastLog(t)
+	require.NotNil(t, log)
+	assert.Equal(t, model.LogTypeRefund, log.Type)
+}
+
+// ===========================================================================
+// RecalculateTaskQuota tests
+// ===========================================================================
+
+func TestRecalculate_PositiveDelta(t *testing.T) {
+	truncate(t)
+	ctx := context.Background()
+
+	const userID, tokenID, channelID = 10, 10, 10
+	const initQuota, preConsumed = 10000, 2000
+	const actualQuota = 3000 // under-charged by 1000
+	const tokenRemain = 5000
+
+	seedUser(t, userID, initQuota)
+	seedToken(t, tokenID, userID, "sk-recalc-pos", tokenRemain)
+	seedChannel(t, channelID)
+
+	task := makeTask(userID, channelID, preConsumed, tokenID, BillingSourceWallet, 0)
+
+	RecalculateTaskQuota(ctx, task, actualQuota, "adaptor adjustment")
+
+	// User quota should decrease by the delta (1000 additional charge)
+	assert.Equal(t, initQuota-(actualQuota-preConsumed), getUserQuota(t, userID))
+
+	// Token should also be charged the delta
+	assert.Equal(t, tokenRemain-(actualQuota-preConsumed), getTokenRemainQuota(t, tokenID))
+
+	// task.Quota should be updated to actualQuota
+	assert.Equal(t, actualQuota, task.Quota)
+
+	// Log type should be Consume (additional charge)
+	log := getLastLog(t)
+	require.NotNil(t, log)
+	assert.Equal(t, model.LogTypeConsume, log.Type)
+	assert.Equal(t, actualQuota-preConsumed, log.Quota)
+}
+
+func TestRecalculate_NegativeDelta(t *testing.T) {
+	truncate(t)
+	ctx := context.Background()
+
+	const userID, tokenID, channelID = 11, 11, 11
+	const initQuota, preConsumed = 10000, 5000
+	const actualQuota = 3000 // over-charged by 2000
+	const tokenRemain = 5000
+
+	seedUser(t, userID, initQuota)
+	seedToken(t, tokenID, userID, "sk-recalc-neg", tokenRemain)
+	seedChannel(t, channelID)
+
+	task := makeTask(userID, channelID, preConsumed, tokenID, BillingSourceWallet, 0)
+
+	RecalculateTaskQuota(ctx, task, actualQuota, "adaptor adjustment")
+
+	// User quota should increase by abs(delta) = 2000 (refund overpayment)
+	assert.Equal(t, initQuota+(preConsumed-actualQuota), getUserQuota(t, userID))
+
+	// Token should be refunded the difference
+	assert.Equal(t, tokenRemain+(preConsumed-actualQuota), getTokenRemainQuota(t, tokenID))
+
+	// task.Quota updated
+	assert.Equal(t, actualQuota, task.Quota)
+
+	// Log type should be Refund
+	log := getLastLog(t)
+	require.NotNil(t, log)
+	assert.Equal(t, model.LogTypeRefund, log.Type)
+	assert.Equal(t, preConsumed-actualQuota, log.Quota)
+}
+
+func TestRecalculate_ZeroDelta(t *testing.T) {
+	truncate(t)
+	ctx := context.Background()
+
+	const userID = 12
+	const initQuota, preConsumed = 10000, 3000
+
+	seedUser(t, userID, initQuota)
+
+	task := makeTask(userID, 0, preConsumed, 0, BillingSourceWallet, 0)
+
+	RecalculateTaskQuota(ctx, task, preConsumed, "exact match")
+
+	// No change to user quota
+	assert.Equal(t, initQuota, getUserQuota(t, userID))
+
+	// No log created (delta is zero)
+	assert.Equal(t, int64(0), countLogs(t))
+}
+
+func TestRecalculate_ActualQuotaZero(t *testing.T) {
+	truncate(t)
+	ctx := context.Background()
+
+	const userID = 13
+	const initQuota = 10000
+
+	seedUser(t, userID, initQuota)
+
+	task := makeTask(userID, 0, 5000, 0, BillingSourceWallet, 0)
+
+	RecalculateTaskQuota(ctx, task, 0, "zero actual")
+
+	// No change (early return)
+	assert.Equal(t, initQuota, getUserQuota(t, userID))
+	assert.Equal(t, int64(0), countLogs(t))
+}
+
+func TestRecalculate_Subscription_NegativeDelta(t *testing.T) {
+	truncate(t)
+	ctx := context.Background()
+
+	const userID, tokenID, channelID, subID = 14, 14, 14, 2
+	const preConsumed = 5000
+	const actualQuota = 2000 // over-charged by 3000
+	const subTotal, subUsed int64 = 100000, 50000
+	const tokenRemain = 8000
+
+	seedUser(t, userID, 0)
+	seedToken(t, tokenID, userID, "sk-sub-recalc", tokenRemain)
+	seedChannel(t, channelID)
+	seedSubscription(t, subID, userID, subTotal, subUsed)
+
+	task := makeTask(userID, channelID, preConsumed, tokenID, BillingSourceSubscription, subID)
+
+	RecalculateTaskQuota(ctx, task, actualQuota, "subscription over-charge")
+
+	// Subscription used should decrease by delta (refund 3000)
+	assert.Equal(t, subUsed-int64(preConsumed-actualQuota), getSubscriptionUsed(t, subID))
+
+	// Token refunded
+	assert.Equal(t, tokenRemain+(preConsumed-actualQuota), getTokenRemainQuota(t, tokenID))
+
+	assert.Equal(t, actualQuota, task.Quota)
+
+	log := getLastLog(t)
+	require.NotNil(t, log)
+	assert.Equal(t, model.LogTypeRefund, log.Type)
+}
+
+// ===========================================================================
+// CAS + Billing integration tests
+// Simulates the flow in updateVideoSingleTask (service/task_polling.go)
+// ===========================================================================
+
+// simulatePollBilling reproduces the CAS + billing logic from updateVideoSingleTask.
+// It takes a persisted task (already in DB), applies the new status, and performs
+// the conditional update + billing exactly as the polling loop does.
+func simulatePollBilling(ctx context.Context, task *model.Task, newStatus model.TaskStatus, actualQuota int) {
+	snap := task.Snapshot()
+
+	shouldRefund := false
+	shouldSettle := false
+	quota := task.Quota
+
+	task.Status = newStatus
+	switch string(newStatus) {
+	case model.TaskStatusSuccess:
+		task.Progress = "100%"
+		task.FinishTime = 9999
+		shouldSettle = true
+	case model.TaskStatusFailure:
+		task.Progress = "100%"
+		task.FinishTime = 9999
+		task.FailReason = "upstream error"
+		if quota != 0 {
+			shouldRefund = true
+		}
+	default:
+		task.Progress = "50%"
+	}
+
+	isDone := task.Status == model.TaskStatus(model.TaskStatusSuccess) || task.Status == model.TaskStatus(model.TaskStatusFailure)
+	if isDone && snap.Status != task.Status {
+		won, err := task.UpdateWithStatus(snap.Status)
+		if err != nil {
+			shouldRefund = false
+			shouldSettle = false
+		} else if !won {
+			shouldRefund = false
+			shouldSettle = false
+		}
+	} else if !snap.Equal(task.Snapshot()) {
+		_, _ = task.UpdateWithStatus(snap.Status)
+	}
+
+	if shouldSettle && actualQuota > 0 {
+		RecalculateTaskQuota(ctx, task, actualQuota, "test settle")
+	}
+	if shouldRefund {
+		RefundTaskQuota(ctx, task, task.FailReason)
+	}
+}
+
+func TestCASGuardedRefund_Win(t *testing.T) {
+	truncate(t)
+	ctx := context.Background()
+
+	const userID, tokenID, channelID = 20, 20, 20
+	const initQuota, preConsumed = 10000, 4000
+	const tokenRemain = 6000
+
+	seedUser(t, userID, initQuota)
+	seedToken(t, tokenID, userID, "sk-cas-refund-win", tokenRemain)
+	seedChannel(t, channelID)
+
+	task := makeTask(userID, channelID, preConsumed, tokenID, BillingSourceWallet, 0)
+	task.Status = model.TaskStatus(model.TaskStatusInProgress)
+	require.NoError(t, model.DB.Create(task).Error)
+
+	simulatePollBilling(ctx, task, model.TaskStatus(model.TaskStatusFailure), 0)
+
+	// CAS wins: task in DB should now be FAILURE
+	var reloaded model.Task
+	require.NoError(t, model.DB.First(&reloaded, task.ID).Error)
+	assert.EqualValues(t, model.TaskStatusFailure, reloaded.Status)
+
+	// Refund should have happened
+	assert.Equal(t, initQuota+preConsumed, getUserQuota(t, userID))
+	assert.Equal(t, tokenRemain+preConsumed, getTokenRemainQuota(t, tokenID))
+
+	log := getLastLog(t)
+	require.NotNil(t, log)
+	assert.Equal(t, model.LogTypeRefund, log.Type)
+}
+
+func TestCASGuardedRefund_Lose(t *testing.T) {
+	truncate(t)
+	ctx := context.Background()
+
+	const userID, tokenID, channelID = 21, 21, 21
+	const initQuota, preConsumed = 10000, 4000
+	const tokenRemain = 6000
+
+	seedUser(t, userID, initQuota)
+	seedToken(t, tokenID, userID, "sk-cas-refund-lose", tokenRemain)
+	seedChannel(t, channelID)
+
+	// Create task with IN_PROGRESS in DB
+	task := makeTask(userID, channelID, preConsumed, tokenID, BillingSourceWallet, 0)
+	task.Status = model.TaskStatus(model.TaskStatusInProgress)
+	require.NoError(t, model.DB.Create(task).Error)
+
+	// Simulate another process already transitioning to FAILURE
+	model.DB.Model(&model.Task{}).Where("id = ?", task.ID).Update("status", model.TaskStatusFailure)
+
+	// Our process still has the old in-memory state (IN_PROGRESS) and tries to transition
+	// task.Status is still IN_PROGRESS in the snapshot
+	simulatePollBilling(ctx, task, model.TaskStatus(model.TaskStatusFailure), 0)
+
+	// CAS lost: user quota should NOT change (no double refund)
+	assert.Equal(t, initQuota, getUserQuota(t, userID))
+	assert.Equal(t, tokenRemain, getTokenRemainQuota(t, tokenID))
+
+	// No billing log should be created
+	assert.Equal(t, int64(0), countLogs(t))
+}
+
+func TestCASGuardedSettle_Win(t *testing.T) {
+	truncate(t)
+	ctx := context.Background()
+
+	const userID, tokenID, channelID = 22, 22, 22
+	const initQuota, preConsumed = 10000, 5000
+	const actualQuota = 3000 // over-charged, should get partial refund
+	const tokenRemain = 8000
+
+	seedUser(t, userID, initQuota)
+	seedToken(t, tokenID, userID, "sk-cas-settle-win", tokenRemain)
+	seedChannel(t, channelID)
+
+	task := makeTask(userID, channelID, preConsumed, tokenID, BillingSourceWallet, 0)
+	task.Status = model.TaskStatus(model.TaskStatusInProgress)
+	require.NoError(t, model.DB.Create(task).Error)
+
+	simulatePollBilling(ctx, task, model.TaskStatus(model.TaskStatusSuccess), actualQuota)
+
+	// CAS wins: task should be SUCCESS
+	var reloaded model.Task
+	require.NoError(t, model.DB.First(&reloaded, task.ID).Error)
+	assert.EqualValues(t, model.TaskStatusSuccess, reloaded.Status)
+
+	// Settlement should refund the over-charge (5000 - 3000 = 2000 back to user)
+	assert.Equal(t, initQuota+(preConsumed-actualQuota), getUserQuota(t, userID))
+	assert.Equal(t, tokenRemain+(preConsumed-actualQuota), getTokenRemainQuota(t, tokenID))
+
+	// task.Quota should be updated to actualQuota
+	assert.Equal(t, actualQuota, task.Quota)
+}
+
+func TestNonTerminalUpdate_NoBilling(t *testing.T) {
+	truncate(t)
+	ctx := context.Background()
+
+	const userID, channelID = 23, 23
+	const initQuota, preConsumed = 10000, 3000
+
+	seedUser(t, userID, initQuota)
+	seedChannel(t, channelID)
+
+	task := makeTask(userID, channelID, preConsumed, 0, BillingSourceWallet, 0)
+	task.Status = model.TaskStatus(model.TaskStatusInProgress)
+	task.Progress = "20%"
+	require.NoError(t, model.DB.Create(task).Error)
+
+	// Simulate a non-terminal poll update (still IN_PROGRESS, progress changed)
+	simulatePollBilling(ctx, task, model.TaskStatus(model.TaskStatusInProgress), 0)
+
+	// User quota should NOT change
+	assert.Equal(t, initQuota, getUserQuota(t, userID))
+
+	// No billing log
+	assert.Equal(t, int64(0), countLogs(t))
+
+	// Task progress should be updated in DB
+	var reloaded model.Task
+	require.NoError(t, model.DB.First(&reloaded, task.ID).Error)
+	assert.Equal(t, "50%", reloaded.Progress)
+}
+
+// ===========================================================================
+// Mock adaptor for settleTaskBillingOnComplete tests
+// ===========================================================================
+
+type mockAdaptor struct {
+	adjustReturn int
+}
+
+func (m *mockAdaptor) Init(_ *relaycommon.RelayInfo) {}
+func (m *mockAdaptor) FetchTask(string, string, map[string]any, string) (*http.Response, error) {
+	return nil, nil
+}
+func (m *mockAdaptor) ParseTaskResult([]byte) (*relaycommon.TaskInfo, error) { return nil, nil }
+func (m *mockAdaptor) AdjustBillingOnComplete(_ *model.Task, _ *relaycommon.TaskInfo) int {
+	return m.adjustReturn
+}
+
+// ===========================================================================
+// PerCallBilling tests — settleTaskBillingOnComplete
+// ===========================================================================
+
+func TestSettle_PerCallBilling_SkipsAdaptorAdjust(t *testing.T) {
+	truncate(t)
+	ctx := context.Background()
+
+	const userID, tokenID, channelID = 30, 30, 30
+	const initQuota, preConsumed = 10000, 5000
+	const tokenRemain = 8000
+
+	seedUser(t, userID, initQuota)
+	seedToken(t, tokenID, userID, "sk-percall-adaptor", tokenRemain)
+	seedChannel(t, channelID)
+
+	task := makeTask(userID, channelID, preConsumed, tokenID, BillingSourceWallet, 0)
+	task.PrivateData.BillingContext.PerCallBilling = true
+
+	adaptor := &mockAdaptor{adjustReturn: 2000}
+	taskResult := &relaycommon.TaskInfo{Status: model.TaskStatusSuccess}
+
+	settleTaskBillingOnComplete(ctx, adaptor, task, taskResult)
+
+	// Per-call: no adjustment despite adaptor returning 2000
+	assert.Equal(t, initQuota, getUserQuota(t, userID))
+	assert.Equal(t, tokenRemain, getTokenRemainQuota(t, tokenID))
+	assert.Equal(t, preConsumed, task.Quota)
+	assert.Equal(t, int64(0), countLogs(t))
+}
+
+func TestSettle_PerCallBilling_SkipsTotalTokens(t *testing.T) {
+	truncate(t)
+	ctx := context.Background()
+
+	const userID, tokenID, channelID = 31, 31, 31
+	const initQuota, preConsumed = 10000, 4000
+	const tokenRemain = 7000
+
+	seedUser(t, userID, initQuota)
+	seedToken(t, tokenID, userID, "sk-percall-tokens", tokenRemain)
+	seedChannel(t, channelID)
+
+	task := makeTask(userID, channelID, preConsumed, tokenID, BillingSourceWallet, 0)
+	task.PrivateData.BillingContext.PerCallBilling = true
+
+	adaptor := &mockAdaptor{adjustReturn: 0}
+	taskResult := &relaycommon.TaskInfo{Status: model.TaskStatusSuccess, TotalTokens: 9999}
+
+	settleTaskBillingOnComplete(ctx, adaptor, task, taskResult)
+
+	// Per-call: no recalculation by tokens
+	assert.Equal(t, initQuota, getUserQuota(t, userID))
+	assert.Equal(t, tokenRemain, getTokenRemainQuota(t, tokenID))
+	assert.Equal(t, preConsumed, task.Quota)
+	assert.Equal(t, int64(0), countLogs(t))
+}
+
+func TestSettle_NonPerCall_AdaptorAdjustWorks(t *testing.T) {
+	truncate(t)
+	ctx := context.Background()
+
+	const userID, tokenID, channelID = 32, 32, 32
+	const initQuota, preConsumed = 10000, 5000
+	const adaptorQuota = 3000
+	const tokenRemain = 8000
+
+	seedUser(t, userID, initQuota)
+	seedToken(t, tokenID, userID, "sk-nonpercall-adj", tokenRemain)
+	seedChannel(t, channelID)
+
+	task := makeTask(userID, channelID, preConsumed, tokenID, BillingSourceWallet, 0)
+	// PerCallBilling defaults to false
+
+	adaptor := &mockAdaptor{adjustReturn: adaptorQuota}
+	taskResult := &relaycommon.TaskInfo{Status: model.TaskStatusSuccess}
+
+	settleTaskBillingOnComplete(ctx, adaptor, task, taskResult)
+
+	// Non-per-call: adaptor adjustment applies (refund 2000)
+	assert.Equal(t, initQuota+(preConsumed-adaptorQuota), getUserQuota(t, userID))
+	assert.Equal(t, tokenRemain+(preConsumed-adaptorQuota), getTokenRemainQuota(t, tokenID))
+	assert.Equal(t, adaptorQuota, task.Quota)
+
+	log := getLastLog(t)
+	require.NotNil(t, log)
+	assert.Equal(t, model.LogTypeRefund, log.Type)
+}

service/task_polling.go

@@ -0,0 +1,560 @@
+package service
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"io"
+	"net/http"
+	"sort"
+	"strings"
+	"time"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/constant"
+	"github.com/QuantumNous/new-api/dto"
+	"github.com/QuantumNous/new-api/logger"
+	"github.com/QuantumNous/new-api/model"
+	"github.com/QuantumNous/new-api/relay/channel/task/taskcommon"
+	relaycommon "github.com/QuantumNous/new-api/relay/common"
+
+	"github.com/samber/lo"
+)
+
+// TaskPollingAdaptor 定义轮询所需的最小适配器接口，避免 service -> relay 的循环依赖
+type TaskPollingAdaptor interface {
+	Init(info *relaycommon.RelayInfo)
+	FetchTask(baseURL string, key string, body map[string]any, proxy string) (*http.Response, error)
+	ParseTaskResult(body []byte) (*relaycommon.TaskInfo, error)
+	// AdjustBillingOnComplete 在任务到达终态（成功/失败）时由轮询循环调用。
+	// 返回正数触发差额结算（补扣/退还），返回 0 保持预扣费金额不变。
+	AdjustBillingOnComplete(task *model.Task, taskResult *relaycommon.TaskInfo) int
+}
+
+// GetTaskAdaptorFunc 由 main 包注入，用于获取指定平台的任务适配器。
+// 打破 service -> relay -> relay/channel -> service 的循环依赖。
+var GetTaskAdaptorFunc func(platform constant.TaskPlatform) TaskPollingAdaptor
+
+// sweepTimedOutTasks 在主轮询之前独立清理超时任务。
+// 每次最多处理 100 条，剩余的下个周期继续处理。
+// 使用 per-task CAS (UpdateWithStatus) 防止覆盖被正常轮询已推进的任务。
+func sweepTimedOutTasks(ctx context.Context) {
+	if constant.TaskTimeoutMinutes <= 0 {
+		return
+	}
+	cutoff := time.Now().Unix() - int64(constant.TaskTimeoutMinutes)*60
+	tasks := model.GetTimedOutUnfinishedTasks(cutoff, 100)
+	if len(tasks) == 0 {
+		return
+	}
+
+	const legacyTaskCutoff int64 = 1740182400 // 2026-02-22 00:00:00 UTC
+	reason := fmt.Sprintf("任务超时（%d分钟）", constant.TaskTimeoutMinutes)
+	legacyReason := "任务超时（旧系统遗留任务，不进行退款，请联系管理员）"
+	now := time.Now().Unix()
+	timedOutCount := 0
+
+	for _, task := range tasks {
+		isLegacy := task.SubmitTime > 0 && task.SubmitTime < legacyTaskCutoff
+
+		oldStatus := task.Status
+		task.Status = model.TaskStatusFailure
+		task.Progress = "100%"
+		task.FinishTime = now
+		if isLegacy {
+			task.FailReason = legacyReason
+		} else {
+			task.FailReason = reason
+		}
+
+		won, err := task.UpdateWithStatus(oldStatus)
+		if err != nil {
+			logger.LogError(ctx, fmt.Sprintf("sweepTimedOutTasks CAS update error for task %s: %v", task.TaskID, err))
+			continue
+		}
+		if !won {
+			logger.LogInfo(ctx, fmt.Sprintf("sweepTimedOutTasks: task %s already transitioned, skip", task.TaskID))
+			continue
+		}
+		timedOutCount++
+		if !isLegacy && task.Quota != 0 {
+			RefundTaskQuota(ctx, task, reason)
+		}
+	}
+
+	if timedOutCount > 0 {
+		logger.LogInfo(ctx, fmt.Sprintf("sweepTimedOutTasks: timed out %d tasks", timedOutCount))
+	}
+}
+
+// TaskPollingLoop 主轮询循环，每 15 秒检查一次未完成的任务
+func TaskPollingLoop() {
+	for {
+		time.Sleep(time.Duration(15) * time.Second)
+		common.SysLog("任务进度轮询开始")
+		ctx := context.TODO()
+		sweepTimedOutTasks(ctx)
+		allTasks := model.GetAllUnFinishSyncTasks(constant.TaskQueryLimit)
+		platformTask := make(map[constant.TaskPlatform][]*model.Task)
+		for _, t := range allTasks {
+			platformTask[t.Platform] = append(platformTask[t.Platform], t)
+		}
+		for platform, tasks := range platformTask {
+			if len(tasks) == 0 {
+				continue
+			}
+			taskChannelM := make(map[int][]string)
+			taskM := make(map[string]*model.Task)
+			nullTaskIds := make([]int64, 0)
+			for _, task := range tasks {
+				upstreamID := task.GetUpstreamTaskID()
+				if upstreamID == "" {
+					// 统计失败的未完成任务
+					nullTaskIds = append(nullTaskIds, task.ID)
+					continue
+				}
+				taskM[upstreamID] = task
+				taskChannelM[task.ChannelId] = append(taskChannelM[task.ChannelId], upstreamID)
+			}
+			if len(nullTaskIds) > 0 {
+				err := model.TaskBulkUpdateByID(nullTaskIds, map[string]any{
+					"status":   "FAILURE",
+					"progress": "100%",
+				})
+				if err != nil {
+					logger.LogError(ctx, fmt.Sprintf("Fix null task_id task error: %v", err))
+				} else {
+					logger.LogInfo(ctx, fmt.Sprintf("Fix null task_id task success: %v", nullTaskIds))
+				}
+			}
+			if len(taskChannelM) == 0 {
+				continue
+			}
+
+			DispatchPlatformUpdate(platform, taskChannelM, taskM)
+		}
+		common.SysLog("任务进度轮询完成")
+	}
+}
+
+// DispatchPlatformUpdate 按平台分发轮询更新
+func DispatchPlatformUpdate(platform constant.TaskPlatform, taskChannelM map[int][]string, taskM map[string]*model.Task) {
+	switch platform {
+	case constant.TaskPlatformMidjourney:
+		// MJ 轮询由其自身处理，这里预留入口
+	case constant.TaskPlatformSuno:
+		_ = UpdateSunoTasks(context.Background(), taskChannelM, taskM)
+	default:
+		if err := UpdateVideoTasks(context.Background(), platform, taskChannelM, taskM); err != nil {
+			common.SysLog(fmt.Sprintf("UpdateVideoTasks fail: %s", err))
+		}
+	}
+}
+
+// UpdateSunoTasks 按渠道更新所有 Suno 任务
+func UpdateSunoTasks(ctx context.Context, taskChannelM map[int][]string, taskM map[string]*model.Task) error {
+	for channelId, taskIds := range taskChannelM {
+		err := updateSunoTasks(ctx, channelId, taskIds, taskM)
+		if err != nil {
+			logger.LogError(ctx, fmt.Sprintf("渠道 #%d 更新异步任务失败: %s", channelId, err.Error()))
+		}
+	}
+	return nil
+}
+
+func updateSunoTasks(ctx context.Context, channelId int, taskIds []string, taskM map[string]*model.Task) error {
+	logger.LogInfo(ctx, fmt.Sprintf("渠道 #%d 未完成的任务有: %d", channelId, len(taskIds)))
+	if len(taskIds) == 0 {
+		return nil
+	}
+	ch, err := model.CacheGetChannel(channelId)
+	if err != nil {
+		common.SysLog(fmt.Sprintf("CacheGetChannel: %v", err))
+		// Collect DB primary key IDs for bulk update (taskIds are upstream IDs, not task_id column values)
+		var failedIDs []int64
+		for _, upstreamID := range taskIds {
+			if t, ok := taskM[upstreamID]; ok {
+				failedIDs = append(failedIDs, t.ID)
+			}
+		}
+		err = model.TaskBulkUpdateByID(failedIDs, map[string]any{
+			"fail_reason": fmt.Sprintf("获取渠道信息失败，请联系管理员，渠道ID：%d", channelId),
+			"status":      "FAILURE",
+			"progress":    "100%",
+		})
+		if err != nil {
+			common.SysLog(fmt.Sprintf("UpdateSunoTask error: %v", err))
+		}
+		return err
+	}
+	adaptor := GetTaskAdaptorFunc(constant.TaskPlatformSuno)
+	if adaptor == nil {
+		return errors.New("adaptor not found")
+	}
+	proxy := ch.GetSetting().Proxy
+	resp, err := adaptor.FetchTask(*ch.BaseURL, ch.Key, map[string]any{
+		"ids": taskIds,
+	}, proxy)
+	if err != nil {
+		common.SysLog(fmt.Sprintf("Get Task Do req error: %v", err))
+		return err
+	}
+	if resp.StatusCode != http.StatusOK {
+		logger.LogError(ctx, fmt.Sprintf("Get Task status code: %d", resp.StatusCode))
+		return fmt.Errorf("Get Task status code: %d", resp.StatusCode)
+	}
+	defer resp.Body.Close()
+	responseBody, err := io.ReadAll(resp.Body)
+	if err != nil {
+		common.SysLog(fmt.Sprintf("Get Suno Task parse body error: %v", err))
+		return err
+	}
+	var responseItems dto.TaskResponse[[]dto.SunoDataResponse]
+	err = common.Unmarshal(responseBody, &responseItems)
+	if err != nil {
+		logger.LogError(ctx, fmt.Sprintf("Get Suno Task parse body error2: %v, body: %s", err, string(responseBody)))
+		return err
+	}
+	if !responseItems.IsSuccess() {
+		common.SysLog(fmt.Sprintf("渠道 #%d 未完成的任务有: %d, 成功获取到任务数: %s", channelId, len(taskIds), string(responseBody)))
+		return err
+	}
+
+	for _, responseItem := range responseItems.Data {
+		task := taskM[responseItem.TaskID]
+		if !taskNeedsUpdate(task, responseItem) {
+			continue
+		}
+
+		task.Status = lo.If(model.TaskStatus(responseItem.Status) != "", model.TaskStatus(responseItem.Status)).Else(task.Status)
+		task.FailReason = lo.If(responseItem.FailReason != "", responseItem.FailReason).Else(task.FailReason)
+		task.SubmitTime = lo.If(responseItem.SubmitTime != 0, responseItem.SubmitTime).Else(task.SubmitTime)
+		task.StartTime = lo.If(responseItem.StartTime != 0, responseItem.StartTime).Else(task.StartTime)
+		task.FinishTime = lo.If(responseItem.FinishTime != 0, responseItem.FinishTime).Else(task.FinishTime)
+		if responseItem.FailReason != "" || task.Status == model.TaskStatusFailure {
+			logger.LogInfo(ctx, task.TaskID+" 构建失败，"+task.FailReason)
+			task.Progress = "100%"
+			RefundTaskQuota(ctx, task, task.FailReason)
+		}
+		if responseItem.Status == model.TaskStatusSuccess {
+			task.Progress = "100%"
+		}
+		task.Data = responseItem.Data
+
+		err = task.Update()
+		if err != nil {
+			common.SysLog("UpdateSunoTask task error: " + err.Error())
+		}
+	}
+	return nil
+}
+
+// taskNeedsUpdate 检查 Suno 任务是否需要更新
+func taskNeedsUpdate(oldTask *model.Task, newTask dto.SunoDataResponse) bool {
+	if oldTask.SubmitTime != newTask.SubmitTime {
+		return true
+	}
+	if oldTask.StartTime != newTask.StartTime {
+		return true
+	}
+	if oldTask.FinishTime != newTask.FinishTime {
+		return true
+	}
+	if string(oldTask.Status) != newTask.Status {
+		return true
+	}
+	if oldTask.FailReason != newTask.FailReason {
+		return true
+	}
+
+	if (oldTask.Status == model.TaskStatusFailure || oldTask.Status == model.TaskStatusSuccess) && oldTask.Progress != "100%" {
+		return true
+	}
+
+	oldData, _ := common.Marshal(oldTask.Data)
+	newData, _ := common.Marshal(newTask.Data)
+
+	sort.Slice(oldData, func(i, j int) bool {
+		return oldData[i] < oldData[j]
+	})
+	sort.Slice(newData, func(i, j int) bool {
+		return newData[i] < newData[j]
+	})
+
+	if string(oldData) != string(newData) {
+		return true
+	}
+	return false
+}
+
+// UpdateVideoTasks 按渠道更新所有视频任务
+func UpdateVideoTasks(ctx context.Context, platform constant.TaskPlatform, taskChannelM map[int][]string, taskM map[string]*model.Task) error {
+	for channelId, taskIds := range taskChannelM {
+		if err := updateVideoTasks(ctx, platform, channelId, taskIds, taskM); err != nil {
+			logger.LogError(ctx, fmt.Sprintf("Channel #%d failed to update video async tasks: %s", channelId, err.Error()))
+		}
+	}
+	return nil
+}
+
+func updateVideoTasks(ctx context.Context, platform constant.TaskPlatform, channelId int, taskIds []string, taskM map[string]*model.Task) error {
+	logger.LogInfo(ctx, fmt.Sprintf("Channel #%d pending video tasks: %d", channelId, len(taskIds)))
+	if len(taskIds) == 0 {
+		return nil
+	}
+	cacheGetChannel, err := model.CacheGetChannel(channelId)
+	if err != nil {
+		// Collect DB primary key IDs for bulk update (taskIds are upstream IDs, not task_id column values)
+		var failedIDs []int64
+		for _, upstreamID := range taskIds {
+			if t, ok := taskM[upstreamID]; ok {
+				failedIDs = append(failedIDs, t.ID)
+			}
+		}
+		errUpdate := model.TaskBulkUpdateByID(failedIDs, map[string]any{
+			"fail_reason": fmt.Sprintf("Failed to get channel info, channel ID: %d", channelId),
+			"status":      "FAILURE",
+			"progress":    "100%",
+		})
+		if errUpdate != nil {
+			common.SysLog(fmt.Sprintf("UpdateVideoTask error: %v", errUpdate))
+		}
+		return fmt.Errorf("CacheGetChannel failed: %w", err)
+	}
+	adaptor := GetTaskAdaptorFunc(platform)
+	if adaptor == nil {
+		return fmt.Errorf("video adaptor not found")
+	}
+	info := &relaycommon.RelayInfo{}
+	info.ChannelMeta = &relaycommon.ChannelMeta{
+		ChannelBaseUrl: cacheGetChannel.GetBaseURL(),
+	}
+	info.ApiKey = cacheGetChannel.Key
+	adaptor.Init(info)
+	for _, taskId := range taskIds {
+		if err := updateVideoSingleTask(ctx, adaptor, cacheGetChannel, taskId, taskM); err != nil {
+			logger.LogError(ctx, fmt.Sprintf("Failed to update video task %s: %s", taskId, err.Error()))
+		}
+		// sleep 1 second between each task to avoid hitting rate limits of upstream platforms
+		time.Sleep(1 * time.Second)
+	}
+	return nil
+}
+
+func updateVideoSingleTask(ctx context.Context, adaptor TaskPollingAdaptor, ch *model.Channel, taskId string, taskM map[string]*model.Task) error {
+	baseURL := constant.ChannelBaseURLs[ch.Type]
+	if ch.GetBaseURL() != "" {
+		baseURL = ch.GetBaseURL()
+	}
+	proxy := ch.GetSetting().Proxy
+
+	task := taskM[taskId]
+	if task == nil {
+		logger.LogError(ctx, fmt.Sprintf("Task %s not found in taskM", taskId))
+		return fmt.Errorf("task %s not found", taskId)
+	}
+	key := ch.Key
+
+	privateData := task.PrivateData
+	if privateData.Key != "" {
+		key = privateData.Key
+	}
+	resp, err := adaptor.FetchTask(baseURL, key, map[string]any{
+		"task_id": task.GetUpstreamTaskID(),
+		"action":  task.Action,
+	}, proxy)
+	if err != nil {
+		return fmt.Errorf("fetchTask failed for task %s: %w", taskId, err)
+	}
+	defer resp.Body.Close()
+	responseBody, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return fmt.Errorf("readAll failed for task %s: %w", taskId, err)
+	}
+
+	logger.LogDebug(ctx, fmt.Sprintf("updateVideoSingleTask response: %s", string(responseBody)))
+
+	snap := task.Snapshot()
+
+	taskResult := &relaycommon.TaskInfo{}
+	// try parse as New API response format
+	var responseItems dto.TaskResponse[model.Task]
+	if err = common.Unmarshal(responseBody, &responseItems); err == nil && responseItems.IsSuccess() {
+		logger.LogDebug(ctx, fmt.Sprintf("updateVideoSingleTask parsed as new api response format: %+v", responseItems))
+		t := responseItems.Data
+		taskResult.TaskID = t.TaskID
+		taskResult.Status = string(t.Status)
+		taskResult.Url = t.GetResultURL()
+		taskResult.Progress = t.Progress
+		taskResult.Reason = t.FailReason
+		task.Data = t.Data
+	} else if taskResult, err = adaptor.ParseTaskResult(responseBody); err != nil {
+		return fmt.Errorf("parseTaskResult failed for task %s: %w", taskId, err)
+	}
+
+	task.Data = redactVideoResponseBody(responseBody)
+
+	logger.LogDebug(ctx, fmt.Sprintf("updateVideoSingleTask taskResult: %+v", taskResult))
+
+	now := time.Now().Unix()
+	if taskResult.Status == "" {
+		//taskResult = relaycommon.FailTaskInfo("upstream returned empty status")
+		errorResult := &dto.GeneralErrorResponse{}
+		if err = common.Unmarshal(responseBody, &errorResult); err == nil {
+			openaiError := errorResult.TryToOpenAIError()
+			if openaiError != nil {
+				// 返回规范的 OpenAI 错误格式，提取错误信息，判断错误是否为任务失败
+				if openaiError.Code == "429" {
+					// 429 错误通常表示请求过多或速率限制，暂时不认为是任务失败，保持原状态等待下一轮轮询
+					return nil
+				}
+
+				// 其他错误认为是任务失败，记录错误信息并更新任务状态
+				taskResult = relaycommon.FailTaskInfo("upstream returned error")
+			} else {
+				// unknown error format, log original response
+				logger.LogError(ctx, fmt.Sprintf("Task %s returned empty status with unrecognized error format, response: %s", taskId, string(responseBody)))
+				taskResult = relaycommon.FailTaskInfo("upstream returned unrecognized message")
+			}
+		}
+	}
+
+	shouldRefund := false
+	shouldSettle := false
+	quota := task.Quota
+
+	task.Status = model.TaskStatus(taskResult.Status)
+	switch taskResult.Status {
+	case model.TaskStatusSubmitted:
+		task.Progress = taskcommon.ProgressSubmitted
+	case model.TaskStatusQueued:
+		task.Progress = taskcommon.ProgressQueued
+	case model.TaskStatusInProgress:
+		task.Progress = taskcommon.ProgressInProgress
+		if task.StartTime == 0 {
+			task.StartTime = now
+		}
+	case model.TaskStatusSuccess:
+		task.Progress = taskcommon.ProgressComplete
+		if task.FinishTime == 0 {
+			task.FinishTime = now
+		}
+		if strings.HasPrefix(taskResult.Url, "data:") {
+			// data: URI (e.g. Vertex base64 encoded video) — keep in Data, not in ResultURL
+			task.PrivateData.ResultURL = taskcommon.BuildProxyURL(task.TaskID)
+		} else if taskResult.Url != "" {
+			// Direct upstream URL (e.g. Kling, Ali, Doubao, etc.)
+			task.PrivateData.ResultURL = taskResult.Url
+		} else {
+			// No URL from adaptor — construct proxy URL using public task ID
+			task.PrivateData.ResultURL = taskcommon.BuildProxyURL(task.TaskID)
+		}
+		shouldSettle = true
+	case model.TaskStatusFailure:
+		logger.LogJson(ctx, fmt.Sprintf("Task %s failed", taskId), task)
+		task.Status = model.TaskStatusFailure
+		task.Progress = taskcommon.ProgressComplete
+		if task.FinishTime == 0 {
+			task.FinishTime = now
+		}
+		task.FailReason = taskResult.Reason
+		logger.LogInfo(ctx, fmt.Sprintf("Task %s failed: %s", task.TaskID, task.FailReason))
+		taskResult.Progress = taskcommon.ProgressComplete
+		if quota != 0 {
+			shouldRefund = true
+		}
+	default:
+		return fmt.Errorf("unknown task status %s for task %s", taskResult.Status, task.TaskID)
+	}
+	if taskResult.Progress != "" {
+		task.Progress = taskResult.Progress
+	}
+
+	isDone := task.Status == model.TaskStatusSuccess || task.Status == model.TaskStatusFailure
+	if isDone && snap.Status != task.Status {
+		won, err := task.UpdateWithStatus(snap.Status)
+		if err != nil {
+			logger.LogError(ctx, fmt.Sprintf("UpdateWithStatus failed for task %s: %s", task.TaskID, err.Error()))
+			shouldRefund = false
+			shouldSettle = false
+		} else if !won {
+			logger.LogWarn(ctx, fmt.Sprintf("Task %s already transitioned by another process, skip billing", task.TaskID))
+			shouldRefund = false
+			shouldSettle = false
+		}
+	} else if !snap.Equal(task.Snapshot()) {
+		if _, err := task.UpdateWithStatus(snap.Status); err != nil {
+			logger.LogError(ctx, fmt.Sprintf("Failed to update task %s: %s", task.TaskID, err.Error()))
+		}
+	} else {
+		// No changes, skip update
+		logger.LogDebug(ctx, fmt.Sprintf("No update needed for task %s", task.TaskID))
+	}
+
+	if shouldSettle {
+		settleTaskBillingOnComplete(ctx, adaptor, task, taskResult)
+	}
+	if shouldRefund {
+		RefundTaskQuota(ctx, task, task.FailReason)
+	}
+
+	return nil
+}
+
+func redactVideoResponseBody(body []byte) []byte {
+	var m map[string]any
+	if err := common.Unmarshal(body, &m); err != nil {
+		return body
+	}
+	resp, _ := m["response"].(map[string]any)
+	if resp != nil {
+		delete(resp, "bytesBase64Encoded")
+		if v, ok := resp["video"].(string); ok {
+			resp["video"] = truncateBase64(v)
+		}
+		if vs, ok := resp["videos"].([]any); ok {
+			for i := range vs {
+				if vm, ok := vs[i].(map[string]any); ok {
+					delete(vm, "bytesBase64Encoded")
+				}
+			}
+		}
+	}
+	b, err := common.Marshal(m)
+	if err != nil {
+		return body
+	}
+	return b
+}
+
+func truncateBase64(s string) string {
+	const maxKeep = 256
+	if len(s) <= maxKeep {
+		return s
+	}
+	return s[:maxKeep] + "..."
+}
+
+// settleTaskBillingOnComplete 任务完成时的统一计费调整。
+// 优先级：1. adaptor.AdjustBillingOnComplete 返回正数 → 使用 adaptor 计算的额度
+//
+//  2. taskResult.TotalTokens > 0 → 按 token 重算
+//  3. 都不满足 → 保持预扣额度不变
+func settleTaskBillingOnComplete(ctx context.Context, adaptor TaskPollingAdaptor, task *model.Task, taskResult *relaycommon.TaskInfo) {
+	// 0. 按次计费的任务不做差额结算
+	if bc := task.PrivateData.BillingContext; bc != nil && bc.PerCallBilling {
+		logger.LogInfo(ctx, fmt.Sprintf("任务 %s 按次计费，跳过差额结算", task.TaskID))
+		return
+	}
+	// 1. 优先让 adaptor 决定最终额度
+	if actualQuota := adaptor.AdjustBillingOnComplete(task, taskResult); actualQuota > 0 {
+		RecalculateTaskQuota(ctx, task, actualQuota, "adaptor计费调整")
+		return
+	}
+	// 2. 回退到 token 重算
+	if taskResult.TotalTokens > 0 {
+		RecalculateTaskQuotaByTokens(ctx, task, taskResult.TotalTokens)
+		return
+	}
+	// 3. 无调整，保持预扣额度
+}

service/text_quota.go

@@ -0,0 +1,427 @@
+package service
+
+import (
+	"fmt"
+	"strings"
+	"time"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/constant"
+	"github.com/QuantumNous/new-api/dto"
+	"github.com/QuantumNous/new-api/logger"
+	"github.com/QuantumNous/new-api/model"
+	relaycommon "github.com/QuantumNous/new-api/relay/common"
+	"github.com/QuantumNous/new-api/setting/operation_setting"
+	"github.com/QuantumNous/new-api/types"
+
+	"github.com/gin-gonic/gin"
+	"github.com/shopspring/decimal"
+)
+
+type textQuotaSummary struct {
+	PromptTokens             int
+	CompletionTokens         int
+	TotalTokens              int
+	CacheTokens              int
+	CacheCreationTokens      int
+	CacheCreationTokens5m    int
+	CacheCreationTokens1h    int
+	ImageTokens              int
+	AudioTokens              int
+	ModelName                string
+	TokenName                string
+	UseTimeSeconds           int64
+	CompletionRatio          float64
+	CacheRatio               float64
+	ImageRatio               float64
+	ModelRatio               float64
+	GroupRatio               float64
+	ModelPrice               float64
+	CacheCreationRatio       float64
+	CacheCreationRatio5m     float64
+	CacheCreationRatio1h     float64
+	Quota                    int
+	IsClaudeUsageSemantic    bool
+	UsageSemantic            string
+	WebSearchPrice           float64
+	WebSearchCallCount       int
+	ClaudeWebSearchPrice     float64
+	ClaudeWebSearchCallCount int
+	FileSearchPrice          float64
+	FileSearchCallCount      int
+	AudioInputPrice          float64
+	ImageGenerationCallPrice float64
+}
+
+func cacheWriteTokensTotal(summary textQuotaSummary) int {
+	if summary.CacheCreationTokens5m > 0 || summary.CacheCreationTokens1h > 0 {
+		splitCacheWriteTokens := summary.CacheCreationTokens5m + summary.CacheCreationTokens1h
+		if summary.CacheCreationTokens > splitCacheWriteTokens {
+			return summary.CacheCreationTokens
+		}
+		return splitCacheWriteTokens
+	}
+	return summary.CacheCreationTokens
+}
+
+func isLegacyClaudeDerivedOpenAIUsage(relayInfo *relaycommon.RelayInfo, usage *dto.Usage) bool {
+	if relayInfo == nil || usage == nil {
+		return false
+	}
+	if relayInfo.GetFinalRequestRelayFormat() == types.RelayFormatClaude {
+		return false
+	}
+	if usage.UsageSource != "" || usage.UsageSemantic != "" {
+		return false
+	}
+	return usage.ClaudeCacheCreation5mTokens > 0 || usage.ClaudeCacheCreation1hTokens > 0
+}
+
+func calculateTextQuotaSummary(ctx *gin.Context, relayInfo *relaycommon.RelayInfo, usage *dto.Usage) textQuotaSummary {
+	summary := textQuotaSummary{
+		ModelName:            relayInfo.OriginModelName,
+		TokenName:            ctx.GetString("token_name"),
+		UseTimeSeconds:       time.Now().Unix() - relayInfo.StartTime.Unix(),
+		CompletionRatio:      relayInfo.PriceData.CompletionRatio,
+		CacheRatio:           relayInfo.PriceData.CacheRatio,
+		ImageRatio:           relayInfo.PriceData.ImageRatio,
+		ModelRatio:           relayInfo.PriceData.ModelRatio,
+		GroupRatio:           relayInfo.PriceData.GroupRatioInfo.GroupRatio,
+		ModelPrice:           relayInfo.PriceData.ModelPrice,
+		CacheCreationRatio:   relayInfo.PriceData.CacheCreationRatio,
+		CacheCreationRatio5m: relayInfo.PriceData.CacheCreation5mRatio,
+		CacheCreationRatio1h: relayInfo.PriceData.CacheCreation1hRatio,
+		UsageSemantic:        usageSemanticFromUsage(relayInfo, usage),
+	}
+	summary.IsClaudeUsageSemantic = summary.UsageSemantic == "anthropic"
+
+	if usage == nil {
+		usage = &dto.Usage{
+			PromptTokens:     relayInfo.GetEstimatePromptTokens(),
+			CompletionTokens: 0,
+			TotalTokens:      relayInfo.GetEstimatePromptTokens(),
+		}
+	}
+
+	summary.PromptTokens = usage.PromptTokens
+	summary.CompletionTokens = usage.CompletionTokens
+	summary.TotalTokens = usage.PromptTokens + usage.CompletionTokens
+	summary.CacheTokens = usage.PromptTokensDetails.CachedTokens
+	summary.CacheCreationTokens = usage.PromptTokensDetails.CachedCreationTokens
+	summary.CacheCreationTokens5m = usage.ClaudeCacheCreation5mTokens
+	summary.CacheCreationTokens1h = usage.ClaudeCacheCreation1hTokens
+	summary.ImageTokens = usage.PromptTokensDetails.ImageTokens
+	summary.AudioTokens = usage.PromptTokensDetails.AudioTokens
+	legacyClaudeDerived := isLegacyClaudeDerivedOpenAIUsage(relayInfo, usage)
+
+	if relayInfo.ChannelMeta != nil && relayInfo.ChannelType == constant.ChannelTypeOpenRouter {
+		summary.PromptTokens -= summary.CacheTokens
+		isUsingCustomSettings := relayInfo.PriceData.UsePrice || hasCustomModelRatio(summary.ModelName, relayInfo.PriceData.ModelRatio)
+		if summary.CacheCreationTokens == 0 && relayInfo.PriceData.CacheCreationRatio != 1 && usage.Cost != 0 && !isUsingCustomSettings {
+			maybeCacheCreationTokens := CalcOpenRouterCacheCreateTokens(*usage, relayInfo.PriceData)
+			if maybeCacheCreationTokens >= 0 && summary.PromptTokens >= maybeCacheCreationTokens {
+				summary.CacheCreationTokens = maybeCacheCreationTokens
+			}
+		}
+		summary.PromptTokens -= summary.CacheCreationTokens
+	}
+
+	dPromptTokens := decimal.NewFromInt(int64(summary.PromptTokens))
+	dCacheTokens := decimal.NewFromInt(int64(summary.CacheTokens))
+	dImageTokens := decimal.NewFromInt(int64(summary.ImageTokens))
+	dAudioTokens := decimal.NewFromInt(int64(summary.AudioTokens))
+	dCompletionTokens := decimal.NewFromInt(int64(summary.CompletionTokens))
+	dCachedCreationTokens := decimal.NewFromInt(int64(summary.CacheCreationTokens))
+	dCompletionRatio := decimal.NewFromFloat(summary.CompletionRatio)
+	dCacheRatio := decimal.NewFromFloat(summary.CacheRatio)
+	dImageRatio := decimal.NewFromFloat(summary.ImageRatio)
+	dModelRatio := decimal.NewFromFloat(summary.ModelRatio)
+	dGroupRatio := decimal.NewFromFloat(summary.GroupRatio)
+	dModelPrice := decimal.NewFromFloat(summary.ModelPrice)
+	dCacheCreationRatio := decimal.NewFromFloat(summary.CacheCreationRatio)
+	dCacheCreationRatio5m := decimal.NewFromFloat(summary.CacheCreationRatio5m)
+	dCacheCreationRatio1h := decimal.NewFromFloat(summary.CacheCreationRatio1h)
+	dQuotaPerUnit := decimal.NewFromFloat(common.QuotaPerUnit)
+
+	ratio := dModelRatio.Mul(dGroupRatio)
+
+	var dWebSearchQuota decimal.Decimal
+	if relayInfo.ResponsesUsageInfo != nil {
+		if webSearchTool, exists := relayInfo.ResponsesUsageInfo.BuiltInTools[dto.BuildInToolWebSearchPreview]; exists && webSearchTool.CallCount > 0 {
+			summary.WebSearchCallCount = webSearchTool.CallCount
+			summary.WebSearchPrice = operation_setting.GetWebSearchPricePerThousand(summary.ModelName, webSearchTool.SearchContextSize)
+			dWebSearchQuota = decimal.NewFromFloat(summary.WebSearchPrice).
+				Mul(decimal.NewFromInt(int64(webSearchTool.CallCount))).
+				Div(decimal.NewFromInt(1000)).Mul(dGroupRatio).Mul(dQuotaPerUnit)
+		}
+	} else if strings.HasSuffix(summary.ModelName, "search-preview") {
+		searchContextSize := ctx.GetString("chat_completion_web_search_context_size")
+		if searchContextSize == "" {
+			searchContextSize = "medium"
+		}
+		summary.WebSearchCallCount = 1
+		summary.WebSearchPrice = operation_setting.GetWebSearchPricePerThousand(summary.ModelName, searchContextSize)
+		dWebSearchQuota = decimal.NewFromFloat(summary.WebSearchPrice).
+			Div(decimal.NewFromInt(1000)).Mul(dGroupRatio).Mul(dQuotaPerUnit)
+	}
+
+	var dClaudeWebSearchQuota decimal.Decimal
+	summary.ClaudeWebSearchCallCount = ctx.GetInt("claude_web_search_requests")
+	if summary.ClaudeWebSearchCallCount > 0 {
+		summary.ClaudeWebSearchPrice = operation_setting.GetClaudeWebSearchPricePerThousand()
+		dClaudeWebSearchQuota = decimal.NewFromFloat(summary.ClaudeWebSearchPrice).
+			Div(decimal.NewFromInt(1000)).Mul(dGroupRatio).Mul(dQuotaPerUnit).
+			Mul(decimal.NewFromInt(int64(summary.ClaudeWebSearchCallCount)))
+	}
+
+	var dFileSearchQuota decimal.Decimal
+	if relayInfo.ResponsesUsageInfo != nil {
+		if fileSearchTool, exists := relayInfo.ResponsesUsageInfo.BuiltInTools[dto.BuildInToolFileSearch]; exists && fileSearchTool.CallCount > 0 {
+			summary.FileSearchCallCount = fileSearchTool.CallCount
+			summary.FileSearchPrice = operation_setting.GetFileSearchPricePerThousand()
+			dFileSearchQuota = decimal.NewFromFloat(summary.FileSearchPrice).
+				Mul(decimal.NewFromInt(int64(fileSearchTool.CallCount))).
+				Div(decimal.NewFromInt(1000)).Mul(dGroupRatio).Mul(dQuotaPerUnit)
+		}
+	}
+
+	var dImageGenerationCallQuota decimal.Decimal
+	if ctx.GetBool("image_generation_call") {
+		summary.ImageGenerationCallPrice = operation_setting.GetGPTImage1PriceOnceCall(ctx.GetString("image_generation_call_quality"), ctx.GetString("image_generation_call_size"))
+		dImageGenerationCallQuota = decimal.NewFromFloat(summary.ImageGenerationCallPrice).Mul(dGroupRatio).Mul(dQuotaPerUnit)
+	}
+
+	var audioInputQuota decimal.Decimal
+	if !relayInfo.PriceData.UsePrice {
+		baseTokens := dPromptTokens
+
+		var cachedTokensWithRatio decimal.Decimal
+		if !dCacheTokens.IsZero() {
+			if !summary.IsClaudeUsageSemantic && !legacyClaudeDerived {
+				baseTokens = baseTokens.Sub(dCacheTokens)
+			}
+			cachedTokensWithRatio = dCacheTokens.Mul(dCacheRatio)
+		}
+
+		var cachedCreationTokensWithRatio decimal.Decimal
+		hasSplitCacheCreationTokens := summary.CacheCreationTokens5m > 0 || summary.CacheCreationTokens1h > 0
+		if !dCachedCreationTokens.IsZero() || hasSplitCacheCreationTokens {
+			if !summary.IsClaudeUsageSemantic && !legacyClaudeDerived {
+				baseTokens = baseTokens.Sub(dCachedCreationTokens)
+				cachedCreationTokensWithRatio = dCachedCreationTokens.Mul(dCacheCreationRatio)
+			} else {
+				remaining := summary.CacheCreationTokens - summary.CacheCreationTokens5m - summary.CacheCreationTokens1h
+				if remaining < 0 {
+					remaining = 0
+				}
+				cachedCreationTokensWithRatio = decimal.NewFromInt(int64(remaining)).Mul(dCacheCreationRatio)
+				cachedCreationTokensWithRatio = cachedCreationTokensWithRatio.Add(decimal.NewFromInt(int64(summary.CacheCreationTokens5m)).Mul(dCacheCreationRatio5m))
+				cachedCreationTokensWithRatio = cachedCreationTokensWithRatio.Add(decimal.NewFromInt(int64(summary.CacheCreationTokens1h)).Mul(dCacheCreationRatio1h))
+			}
+		}
+
+		var imageTokensWithRatio decimal.Decimal
+		if !dImageTokens.IsZero() {
+			baseTokens = baseTokens.Sub(dImageTokens)
+			imageTokensWithRatio = dImageTokens.Mul(dImageRatio)
+		}
+
+		if !dAudioTokens.IsZero() {
+			summary.AudioInputPrice = operation_setting.GetGeminiInputAudioPricePerMillionTokens(summary.ModelName)
+			if summary.AudioInputPrice > 0 {
+				baseTokens = baseTokens.Sub(dAudioTokens)
+				audioInputQuota = decimal.NewFromFloat(summary.AudioInputPrice).
+					Div(decimal.NewFromInt(1000000)).Mul(dAudioTokens).Mul(dGroupRatio).Mul(dQuotaPerUnit)
+			}
+		}
+
+		promptQuota := baseTokens.Add(cachedTokensWithRatio).Add(imageTokensWithRatio).Add(cachedCreationTokensWithRatio)
+		completionQuota := dCompletionTokens.Mul(dCompletionRatio)
+		quotaCalculateDecimal := promptQuota.Add(completionQuota).Mul(ratio)
+		quotaCalculateDecimal = quotaCalculateDecimal.Add(dWebSearchQuota)
+		quotaCalculateDecimal = quotaCalculateDecimal.Add(dClaudeWebSearchQuota)
+		quotaCalculateDecimal = quotaCalculateDecimal.Add(dFileSearchQuota)
+		quotaCalculateDecimal = quotaCalculateDecimal.Add(audioInputQuota)
+		quotaCalculateDecimal = quotaCalculateDecimal.Add(dImageGenerationCallQuota)
+
+		if len(relayInfo.PriceData.OtherRatios) > 0 {
+			for _, otherRatio := range relayInfo.PriceData.OtherRatios {
+				quotaCalculateDecimal = quotaCalculateDecimal.Mul(decimal.NewFromFloat(otherRatio))
+			}
+		}
+
+		if !ratio.IsZero() && quotaCalculateDecimal.LessThanOrEqual(decimal.Zero) {
+			quotaCalculateDecimal = decimal.NewFromInt(1)
+		}
+		summary.Quota = int(quotaCalculateDecimal.Round(0).IntPart())
+	} else {
+		quotaCalculateDecimal := dModelPrice.Mul(dQuotaPerUnit).Mul(dGroupRatio)
+		quotaCalculateDecimal = quotaCalculateDecimal.Add(dWebSearchQuota)
+		quotaCalculateDecimal = quotaCalculateDecimal.Add(dClaudeWebSearchQuota)
+		quotaCalculateDecimal = quotaCalculateDecimal.Add(dFileSearchQuota)
+		quotaCalculateDecimal = quotaCalculateDecimal.Add(audioInputQuota)
+		quotaCalculateDecimal = quotaCalculateDecimal.Add(dImageGenerationCallQuota)
+		if len(relayInfo.PriceData.OtherRatios) > 0 {
+			for _, otherRatio := range relayInfo.PriceData.OtherRatios {
+				quotaCalculateDecimal = quotaCalculateDecimal.Mul(decimal.NewFromFloat(otherRatio))
+			}
+		}
+		summary.Quota = int(quotaCalculateDecimal.Round(0).IntPart())
+	}
+
+	if summary.TotalTokens == 0 {
+		summary.Quota = 0
+	} else if !ratio.IsZero() && summary.Quota == 0 {
+		summary.Quota = 1
+	}
+
+	return summary
+}
+
+func usageSemanticFromUsage(relayInfo *relaycommon.RelayInfo, usage *dto.Usage) string {
+	if usage != nil && usage.UsageSemantic != "" {
+		return usage.UsageSemantic
+	}
+	if relayInfo != nil && relayInfo.GetFinalRequestRelayFormat() == types.RelayFormatClaude {
+		return "anthropic"
+	}
+	return "openai"
+}
+
+func PostTextConsumeQuota(ctx *gin.Context, relayInfo *relaycommon.RelayInfo, usage *dto.Usage, extraContent []string) {
+	originUsage := usage
+	if usage == nil {
+		extraContent = append(extraContent, "上游无计费信息")
+	}
+	if originUsage != nil {
+		ObserveChannelAffinityUsageCacheByRelayFormat(ctx, usage, relayInfo.GetFinalRequestRelayFormat())
+	}
+
+	adminRejectReason := common.GetContextKeyString(ctx, constant.ContextKeyAdminRejectReason)
+	summary := calculateTextQuotaSummary(ctx, relayInfo, usage)
+
+	if summary.WebSearchCallCount > 0 {
+		extraContent = append(extraContent, fmt.Sprintf("Web Search 调用 %d 次，调用花费 %s", summary.WebSearchCallCount, decimal.NewFromFloat(summary.WebSearchPrice).Mul(decimal.NewFromInt(int64(summary.WebSearchCallCount))).Div(decimal.NewFromInt(1000)).Mul(decimal.NewFromFloat(summary.GroupRatio)).Mul(decimal.NewFromFloat(common.QuotaPerUnit)).String()))
+	}
+	if summary.ClaudeWebSearchCallCount > 0 {
+		extraContent = append(extraContent, fmt.Sprintf("Claude Web Search 调用 %d 次，调用花费 %s", summary.ClaudeWebSearchCallCount, decimal.NewFromFloat(summary.ClaudeWebSearchPrice).Div(decimal.NewFromInt(1000)).Mul(decimal.NewFromFloat(summary.GroupRatio)).Mul(decimal.NewFromFloat(common.QuotaPerUnit)).Mul(decimal.NewFromInt(int64(summary.ClaudeWebSearchCallCount))).String()))
+	}
+	if summary.FileSearchCallCount > 0 {
+		extraContent = append(extraContent, fmt.Sprintf("File Search 调用 %d 次，调用花费 %s", summary.FileSearchCallCount, decimal.NewFromFloat(summary.FileSearchPrice).Mul(decimal.NewFromInt(int64(summary.FileSearchCallCount))).Div(decimal.NewFromInt(1000)).Mul(decimal.NewFromFloat(summary.GroupRatio)).Mul(decimal.NewFromFloat(common.QuotaPerUnit)).String()))
+	}
+	if summary.AudioInputPrice > 0 && summary.AudioTokens > 0 {
+		extraContent = append(extraContent, fmt.Sprintf("Audio Input 花费 %s", decimal.NewFromFloat(summary.AudioInputPrice).Div(decimal.NewFromInt(1000000)).Mul(decimal.NewFromInt(int64(summary.AudioTokens))).Mul(decimal.NewFromFloat(summary.GroupRatio)).Mul(decimal.NewFromFloat(common.QuotaPerUnit)).String()))
+	}
+	if summary.ImageGenerationCallPrice > 0 {
+		extraContent = append(extraContent, fmt.Sprintf("Image Generation Call 花费 %s", decimal.NewFromFloat(summary.ImageGenerationCallPrice).Mul(decimal.NewFromFloat(summary.GroupRatio)).Mul(decimal.NewFromFloat(common.QuotaPerUnit)).String()))
+	}
+
+	if summary.TotalTokens == 0 {
+		extraContent = append(extraContent, "上游没有返回计费信息，无法扣费（可能是上游超时）")
+		logger.LogError(ctx, fmt.Sprintf("total tokens is 0, cannot consume quota, userId %d, channelId %d, tokenId %d, model %s， pre-consumed quota %d", relayInfo.UserId, relayInfo.ChannelId, relayInfo.TokenId, summary.ModelName, relayInfo.FinalPreConsumedQuota))
+	} else {
+		model.UpdateUserUsedQuotaAndRequestCount(relayInfo.UserId, summary.Quota)
+		model.UpdateChannelUsedQuota(relayInfo.ChannelId, summary.Quota)
+	}
+
+	if err := SettleBilling(ctx, relayInfo, summary.Quota); err != nil {
+		logger.LogError(ctx, "error settling billing: "+err.Error())
+	}
+
+	logModel := summary.ModelName
+	if strings.HasPrefix(logModel, "gpt-4-gizmo") {
+		logModel = "gpt-4-gizmo-*"
+		extraContent = append(extraContent, fmt.Sprintf("模型 %s", summary.ModelName))
+	}
+	if strings.HasPrefix(logModel, "gpt-4o-gizmo") {
+		logModel = "gpt-4o-gizmo-*"
+		extraContent = append(extraContent, fmt.Sprintf("模型 %s", summary.ModelName))
+	}
+
+	logContent := strings.Join(extraContent, ", ")
+	var other map[string]interface{}
+	if summary.IsClaudeUsageSemantic {
+		other = GenerateClaudeOtherInfo(ctx, relayInfo,
+			summary.ModelRatio, summary.GroupRatio, summary.CompletionRatio,
+			summary.CacheTokens, summary.CacheRatio,
+			summary.CacheCreationTokens, summary.CacheCreationRatio,
+			summary.CacheCreationTokens5m, summary.CacheCreationRatio5m,
+			summary.CacheCreationTokens1h, summary.CacheCreationRatio1h,
+			summary.ModelPrice, relayInfo.PriceData.GroupRatioInfo.GroupSpecialRatio)
+		other["usage_semantic"] = "anthropic"
+	} else {
+		other = GenerateTextOtherInfo(ctx, relayInfo, summary.ModelRatio, summary.GroupRatio, summary.CompletionRatio, summary.CacheTokens, summary.CacheRatio, summary.ModelPrice, relayInfo.PriceData.GroupRatioInfo.GroupSpecialRatio)
+	}
+	if adminRejectReason != "" {
+		other["reject_reason"] = adminRejectReason
+	}
+	if summary.ImageTokens != 0 {
+		other["image"] = true
+		other["image_ratio"] = summary.ImageRatio
+		other["image_output"] = summary.ImageTokens
+	}
+	if summary.WebSearchCallCount > 0 {
+		other["web_search"] = true
+		other["web_search_call_count"] = summary.WebSearchCallCount
+		other["web_search_price"] = summary.WebSearchPrice
+	} else if summary.ClaudeWebSearchCallCount > 0 {
+		other["web_search"] = true
+		other["web_search_call_count"] = summary.ClaudeWebSearchCallCount
+		other["web_search_price"] = summary.ClaudeWebSearchPrice
+	}
+	if summary.FileSearchCallCount > 0 {
+		other["file_search"] = true
+		other["file_search_call_count"] = summary.FileSearchCallCount
+		other["file_search_price"] = summary.FileSearchPrice
+	}
+	if summary.AudioInputPrice > 0 && summary.AudioTokens > 0 {
+		other["audio_input_seperate_price"] = true
+		other["audio_input_token_count"] = summary.AudioTokens
+		other["audio_input_price"] = summary.AudioInputPrice
+	}
+	if summary.ImageGenerationCallPrice > 0 {
+		other["image_generation_call"] = true
+		other["image_generation_call_price"] = summary.ImageGenerationCallPrice
+	}
+	if summary.CacheCreationTokens > 0 {
+		other["cache_creation_tokens"] = summary.CacheCreationTokens
+		other["cache_creation_ratio"] = summary.CacheCreationRatio
+	}
+	if summary.CacheCreationTokens5m > 0 {
+		other["cache_creation_tokens_5m"] = summary.CacheCreationTokens5m
+		other["cache_creation_ratio_5m"] = summary.CacheCreationRatio5m
+	}
+	if summary.CacheCreationTokens1h > 0 {
+		other["cache_creation_tokens_1h"] = summary.CacheCreationTokens1h
+		other["cache_creation_ratio_1h"] = summary.CacheCreationRatio1h
+	}
+	cacheWriteTokens := cacheWriteTokensTotal(summary)
+	if cacheWriteTokens > 0 {
+		// cache_write_tokens: normalized cache creation total for UI display.
+		// If split 5m/1h values are present, this is their sum; otherwise it falls back
+		// to cache_creation_tokens.
+		other["cache_write_tokens"] = cacheWriteTokens
+	}
+	if relayInfo.GetFinalRequestRelayFormat() != types.RelayFormatClaude && usage != nil && usage.UsageSource != "" && usage.InputTokens > 0 {
+		// input_tokens_total: explicit normalized total input used by the usage log UI.
+		// Only write this field when upstream/current conversion has already provided a
+		// reliable total input value and tagged the usage source. Do not infer it from
+		// prompt/cache fields here, otherwise old upstream payloads may be double-counted.
+		other["input_tokens_total"] = usage.InputTokens
+	}
+
+	model.RecordConsumeLog(ctx, relayInfo.UserId, model.RecordConsumeLogParams{
+		ChannelId:        relayInfo.ChannelId,
+		PromptTokens:     summary.PromptTokens,
+		CompletionTokens: summary.CompletionTokens,
+		ModelName:        logModel,
+		TokenName:        summary.TokenName,
+		Quota:            summary.Quota,
+		Content:          logContent,
+		TokenId:          relayInfo.TokenId,
+		UseTimeSeconds:   int(summary.UseTimeSeconds),
+		IsStream:         relayInfo.IsStream,
+		Group:            relayInfo.UsingGroup,
+		Other:            other,
+	})
+}

service/text_quota_test.go

@@ -0,0 +1,206 @@
+package service
+
+import (
+	"net/http/httptest"
+	"testing"
+	"time"
+
+	"github.com/QuantumNous/new-api/dto"
+	relaycommon "github.com/QuantumNous/new-api/relay/common"
+	"github.com/QuantumNous/new-api/types"
+
+	"github.com/gin-gonic/gin"
+	"github.com/stretchr/testify/require"
+)
+
+func TestCalculateTextQuotaSummaryUnifiedForClaudeSemantic(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	w := httptest.NewRecorder()
+	ctx, _ := gin.CreateTestContext(w)
+
+	usage := &dto.Usage{
+		PromptTokens:     1000,
+		CompletionTokens: 200,
+		PromptTokensDetails: dto.InputTokenDetails{
+			CachedTokens:         100,
+			CachedCreationTokens: 50,
+		},
+		ClaudeCacheCreation5mTokens: 10,
+		ClaudeCacheCreation1hTokens: 20,
+	}
+
+	priceData := types.PriceData{
+		ModelRatio:           1,
+		CompletionRatio:      2,
+		CacheRatio:           0.1,
+		CacheCreationRatio:   1.25,
+		CacheCreation5mRatio: 1.25,
+		CacheCreation1hRatio: 2,
+		GroupRatioInfo: types.GroupRatioInfo{
+			GroupRatio: 1,
+		},
+	}
+
+	chatRelayInfo := &relaycommon.RelayInfo{
+		RelayFormat:             types.RelayFormatOpenAI,
+		FinalRequestRelayFormat: types.RelayFormatClaude,
+		OriginModelName:         "claude-3-7-sonnet",
+		PriceData:               priceData,
+		StartTime:               time.Now(),
+	}
+	messageRelayInfo := &relaycommon.RelayInfo{
+		RelayFormat:             types.RelayFormatClaude,
+		FinalRequestRelayFormat: types.RelayFormatClaude,
+		OriginModelName:         "claude-3-7-sonnet",
+		PriceData:               priceData,
+		StartTime:               time.Now(),
+	}
+
+	chatSummary := calculateTextQuotaSummary(ctx, chatRelayInfo, usage)
+	messageSummary := calculateTextQuotaSummary(ctx, messageRelayInfo, usage)
+
+	require.Equal(t, messageSummary.Quota, chatSummary.Quota)
+	require.Equal(t, messageSummary.CacheCreationTokens5m, chatSummary.CacheCreationTokens5m)
+	require.Equal(t, messageSummary.CacheCreationTokens1h, chatSummary.CacheCreationTokens1h)
+	require.True(t, chatSummary.IsClaudeUsageSemantic)
+	require.Equal(t, 1488, chatSummary.Quota)
+}
+
+func TestCalculateTextQuotaSummaryUsesSplitClaudeCacheCreationRatios(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	w := httptest.NewRecorder()
+	ctx, _ := gin.CreateTestContext(w)
+
+	relayInfo := &relaycommon.RelayInfo{
+		RelayFormat:             types.RelayFormatOpenAI,
+		FinalRequestRelayFormat: types.RelayFormatClaude,
+		OriginModelName:         "claude-3-7-sonnet",
+		PriceData: types.PriceData{
+			ModelRatio:           1,
+			CompletionRatio:      1,
+			CacheRatio:           0,
+			CacheCreationRatio:   1,
+			CacheCreation5mRatio: 2,
+			CacheCreation1hRatio: 3,
+			GroupRatioInfo: types.GroupRatioInfo{
+				GroupRatio: 1,
+			},
+		},
+		StartTime: time.Now(),
+	}
+
+	usage := &dto.Usage{
+		PromptTokens:     100,
+		CompletionTokens: 0,
+		PromptTokensDetails: dto.InputTokenDetails{
+			CachedCreationTokens: 10,
+		},
+		ClaudeCacheCreation5mTokens: 2,
+		ClaudeCacheCreation1hTokens: 3,
+	}
+
+	summary := calculateTextQuotaSummary(ctx, relayInfo, usage)
+
+	// 100 + remaining(5)*1 + 2*2 + 3*3 = 118
+	require.Equal(t, 118, summary.Quota)
+}
+
+func TestCalculateTextQuotaSummaryUsesAnthropicUsageSemanticFromUpstreamUsage(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	w := httptest.NewRecorder()
+	ctx, _ := gin.CreateTestContext(w)
+
+	relayInfo := &relaycommon.RelayInfo{
+		RelayFormat:     types.RelayFormatOpenAI,
+		OriginModelName: "claude-3-7-sonnet",
+		PriceData: types.PriceData{
+			ModelRatio:           1,
+			CompletionRatio:      2,
+			CacheRatio:           0.1,
+			CacheCreationRatio:   1.25,
+			CacheCreation5mRatio: 1.25,
+			CacheCreation1hRatio: 2,
+			GroupRatioInfo: types.GroupRatioInfo{
+				GroupRatio: 1,
+			},
+		},
+		StartTime: time.Now(),
+	}
+
+	usage := &dto.Usage{
+		PromptTokens:     1000,
+		CompletionTokens: 200,
+		UsageSemantic:    "anthropic",
+		PromptTokensDetails: dto.InputTokenDetails{
+			CachedTokens:         100,
+			CachedCreationTokens: 50,
+		},
+		ClaudeCacheCreation5mTokens: 10,
+		ClaudeCacheCreation1hTokens: 20,
+	}
+
+	summary := calculateTextQuotaSummary(ctx, relayInfo, usage)
+
+	require.True(t, summary.IsClaudeUsageSemantic)
+	require.Equal(t, "anthropic", summary.UsageSemantic)
+	require.Equal(t, 1488, summary.Quota)
+}
+
+func TestCacheWriteTokensTotal(t *testing.T) {
+	t.Run("split cache creation", func(t *testing.T) {
+		summary := textQuotaSummary{
+			CacheCreationTokens:   50,
+			CacheCreationTokens5m: 10,
+			CacheCreationTokens1h: 20,
+		}
+		require.Equal(t, 50, cacheWriteTokensTotal(summary))
+	})
+
+	t.Run("legacy cache creation", func(t *testing.T) {
+		summary := textQuotaSummary{CacheCreationTokens: 50}
+		require.Equal(t, 50, cacheWriteTokensTotal(summary))
+	})
+
+	t.Run("split cache creation without aggregate remainder", func(t *testing.T) {
+		summary := textQuotaSummary{
+			CacheCreationTokens5m: 10,
+			CacheCreationTokens1h: 20,
+		}
+		require.Equal(t, 30, cacheWriteTokensTotal(summary))
+	})
+}
+
+func TestCalculateTextQuotaSummaryHandlesLegacyClaudeDerivedOpenAIUsage(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	w := httptest.NewRecorder()
+	ctx, _ := gin.CreateTestContext(w)
+
+	relayInfo := &relaycommon.RelayInfo{
+		RelayFormat:     types.RelayFormatOpenAI,
+		OriginModelName: "claude-3-7-sonnet",
+		PriceData: types.PriceData{
+			ModelRatio:           1,
+			CompletionRatio:      5,
+			CacheRatio:           0.1,
+			CacheCreationRatio:   1.25,
+			CacheCreation5mRatio: 1.25,
+			CacheCreation1hRatio: 2,
+			GroupRatioInfo:       types.GroupRatioInfo{GroupRatio: 1},
+		},
+		StartTime: time.Now(),
+	}
+
+	usage := &dto.Usage{
+		PromptTokens:     62,
+		CompletionTokens: 95,
+		PromptTokensDetails: dto.InputTokenDetails{
+			CachedTokens: 3544,
+		},
+		ClaudeCacheCreation5mTokens: 586,
+	}
+
+	summary := calculateTextQuotaSummary(ctx, relayInfo, usage)
+
+	// 62 + 3544*0.1 + 586*1.25 + 95*5 = 1624.9 => 1624
+	require.Equal(t, 1624, summary.Quota)
+}

service/token_counter.go

@@ -0,0 +1,411 @@
+package service
+
+import (
+	"errors"
+	"fmt"
+	"log"
+	"math"
+	"path/filepath"
+	"strings"
+	"unicode/utf8"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/constant"
+	"github.com/QuantumNous/new-api/dto"
+	relaycommon "github.com/QuantumNous/new-api/relay/common"
+	constant2 "github.com/QuantumNous/new-api/relay/constant"
+	"github.com/QuantumNous/new-api/types"
+
+	"github.com/gin-gonic/gin"
+)
+
+func getImageToken(c *gin.Context, fileMeta *types.FileMeta, model string, stream bool) (int, error) {
+	if fileMeta == nil || fileMeta.Source == nil {
+		return 0, fmt.Errorf("image_url_is_nil")
+	}
+
+	// Defaults for 4o/4.1/4.5 family unless overridden below
+	baseTokens := 85
+	tileTokens := 170
+
+	// Model classification
+	lowerModel := strings.ToLower(model)
+
+	// Special cases from existing behavior
+	if strings.HasPrefix(lowerModel, "glm-4") {
+		return 1047, nil
+	}
+
+	// Patch-based models (32x32 patches, capped at 1536, with multiplier)
+	isPatchBased := false
+	multiplier := 1.0
+	switch {
+	case strings.Contains(lowerModel, "gpt-4.1-mini"):
+		isPatchBased = true
+		multiplier = 1.62
+	case strings.Contains(lowerModel, "gpt-4.1-nano"):
+		isPatchBased = true
+		multiplier = 2.46
+	case strings.HasPrefix(lowerModel, "o4-mini"):
+		isPatchBased = true
+		multiplier = 1.72
+	case strings.HasPrefix(lowerModel, "gpt-5-mini"):
+		isPatchBased = true
+		multiplier = 1.62
+	case strings.HasPrefix(lowerModel, "gpt-5-nano"):
+		isPatchBased = true
+		multiplier = 2.46
+	}
+
+	// Tile-based model tokens and bases per doc
+	if !isPatchBased {
+		if strings.HasPrefix(lowerModel, "gpt-4o-mini") {
+			baseTokens = 2833
+			tileTokens = 5667
+		} else if strings.HasPrefix(lowerModel, "gpt-5-chat-latest") || (strings.HasPrefix(lowerModel, "gpt-5") && !strings.Contains(lowerModel, "mini") && !strings.Contains(lowerModel, "nano")) {
+			baseTokens = 70
+			tileTokens = 140
+		} else if strings.HasPrefix(lowerModel, "o1") || strings.HasPrefix(lowerModel, "o3") || strings.HasPrefix(lowerModel, "o1-pro") {
+			baseTokens = 75
+			tileTokens = 150
+		} else if strings.Contains(lowerModel, "computer-use-preview") {
+			baseTokens = 65
+			tileTokens = 129
+		} else if strings.Contains(lowerModel, "4.1") || strings.Contains(lowerModel, "4o") || strings.Contains(lowerModel, "4.5") {
+			baseTokens = 85
+			tileTokens = 170
+		}
+	}
+
+	// Respect existing feature flags/short-circuits
+	if fileMeta.Detail == "low" && !isPatchBased {
+		return baseTokens, nil
+	}
+
+	// Whether to count image tokens at all
+	if !constant.GetMediaToken {
+		return 3 * baseTokens, nil
+	}
+
+	if !constant.GetMediaTokenNotStream && !stream {
+		return 3 * baseTokens, nil
+	}
+	// Normalize detail
+	if fileMeta.Detail == "auto" || fileMeta.Detail == "" {
+		fileMeta.Detail = "high"
+	}
+
+	// 使用统一的文件服务获取图片配置
+	config, format, err := GetImageConfig(c, fileMeta.Source)
+	if err != nil {
+		return 0, err
+	}
+	fileMeta.MimeType = format
+
+	if config.Width == 0 || config.Height == 0 {
+		// not an image, but might be a valid file
+		if format != "" {
+			// file type
+			return 3 * baseTokens, nil
+		}
+		return 0, errors.New(fmt.Sprintf("fail to decode image config: %s", fileMeta.GetIdentifier()))
+	}
+
+	width := config.Width
+	height := config.Height
+	log.Printf("format: %s, width: %d, height: %d", format, width, height)
+
+	if isPatchBased {
+		// 32x32 patch-based calculation with 1536 cap and model multiplier
+		ceilDiv := func(a, b int) int { return (a + b - 1) / b }
+		rawPatchesW := ceilDiv(width, 32)
+		rawPatchesH := ceilDiv(height, 32)
+		rawPatches := rawPatchesW * rawPatchesH
+		if rawPatches > 1536 {
+			// scale down
+			area := float64(width * height)
+			r := math.Sqrt(float64(32*32*1536) / area)
+			wScaled := float64(width) * r
+			hScaled := float64(height) * r
+			// adjust to fit whole number of patches after scaling
+			adjW := math.Floor(wScaled/32.0) / (wScaled / 32.0)
+			adjH := math.Floor(hScaled/32.0) / (hScaled / 32.0)
+			adj := math.Min(adjW, adjH)
+			if !math.IsNaN(adj) && adj > 0 {
+				r = r * adj
+			}
+			wScaled = float64(width) * r
+			hScaled = float64(height) * r
+			patchesW := math.Ceil(wScaled / 32.0)
+			patchesH := math.Ceil(hScaled / 32.0)
+			imageTokens := int(patchesW * patchesH)
+			if imageTokens > 1536 {
+				imageTokens = 1536
+			}
+			return int(math.Round(float64(imageTokens) * multiplier)), nil
+		}
+		// below cap
+		imageTokens := rawPatches
+		return int(math.Round(float64(imageTokens) * multiplier)), nil
+	}
+
+	// Tile-based calculation for 4o/4.1/4.5/o1/o3/etc.
+	// Step 1: fit within 2048x2048 square
+	maxSide := math.Max(float64(width), float64(height))
+	fitScale := 1.0
+	if maxSide > 2048 {
+		fitScale = maxSide / 2048.0
+	}
+	fitW := int(math.Round(float64(width) / fitScale))
+	fitH := int(math.Round(float64(height) / fitScale))
+
+	// Step 2: scale so that shortest side is exactly 768
+	minSide := math.Min(float64(fitW), float64(fitH))
+	if minSide == 0 {
+		return baseTokens, nil
+	}
+	shortScale := 768.0 / minSide
+	finalW := int(math.Round(float64(fitW) * shortScale))
+	finalH := int(math.Round(float64(fitH) * shortScale))
+
+	// Count 512px tiles
+	tilesW := (finalW + 512 - 1) / 512
+	tilesH := (finalH + 512 - 1) / 512
+	tiles := tilesW * tilesH
+
+	if common.DebugEnabled {
+		log.Printf("scaled to: %dx%d, tiles: %d", finalW, finalH, tiles)
+	}
+
+	return tiles*tileTokens + baseTokens, nil
+}
+
+func EstimateRequestToken(c *gin.Context, meta *types.TokenCountMeta, info *relaycommon.RelayInfo) (int, error) {
+	// 是否统计token
+	if !constant.CountToken {
+		return 0, nil
+	}
+
+	if meta == nil {
+		return 0, errors.New("token count meta is nil")
+	}
+
+	if info.RelayFormat == types.RelayFormatOpenAIRealtime {
+		return 0, nil
+	}
+	if info.RelayMode == constant2.RelayModeAudioTranscription || info.RelayMode == constant2.RelayModeAudioTranslation {
+		multiForm, err := common.ParseMultipartFormReusable(c)
+		if err != nil {
+			return 0, fmt.Errorf("error parsing multipart form: %v", err)
+		}
+		fileHeaders := multiForm.File["file"]
+		totalAudioToken := 0
+		for _, fileHeader := range fileHeaders {
+			file, err := fileHeader.Open()
+			if err != nil {
+				return 0, fmt.Errorf("error opening audio file: %v", err)
+			}
+			defer file.Close()
+			// get ext and io.seeker
+			ext := filepath.Ext(fileHeader.Filename)
+			duration, err := common.GetAudioDuration(c.Request.Context(), file, ext)
+			if err != nil {
+				return 0, fmt.Errorf("error getting audio duration: %v", err)
+			}
+			// 一分钟 1000 token，与 $price / minute 对齐
+			totalAudioToken += int(math.Round(math.Ceil(duration) / 60.0 * 1000))
+		}
+		return totalAudioToken, nil
+	}
+
+	model := common.GetContextKeyString(c, constant.ContextKeyOriginalModel)
+	tkm := 0
+
+	if meta.TokenType == types.TokenTypeTextNumber {
+		tkm += utf8.RuneCountInString(meta.CombineText)
+	} else {
+		tkm += CountTextToken(meta.CombineText, model)
+	}
+
+	if info.RelayFormat == types.RelayFormatOpenAI {
+		tkm += meta.ToolsCount * 8
+		tkm += meta.MessagesCount * 3 // 每条消息的格式化token数量
+		tkm += meta.NameCount * 3
+		tkm += 3
+	}
+
+	shouldFetchFiles := true
+
+	if info.RelayFormat == types.RelayFormatGemini {
+		shouldFetchFiles = false
+	}
+
+	// 是否本地计算媒体token数量
+	if !constant.GetMediaToken {
+		shouldFetchFiles = false
+	}
+
+	// 是否在非流模式下本地计算媒体token数量
+	if !constant.GetMediaTokenNotStream && !info.IsStream {
+		shouldFetchFiles = false
+	}
+
+	// 使用统一的文件服务获取文件类型
+	for _, file := range meta.Files {
+		if file.Source == nil {
+			continue
+		}
+
+		// 如果文件类型未知且需要获取，通过 MIME 类型检测
+		if file.FileType == "" || (file.Source.IsURL() && shouldFetchFiles) {
+			// 注意：这里我们直接调用 LoadFileSource 而不是 GetMimeType
+			// 因为 GetMimeType 内部可能会调用 GetFileTypeFromUrl (HEAD 请求)
+			// 而我们这里既然要计算 token，通常需要完整数据
+			cachedData, err := LoadFileSource(c, file.Source, "token_counter")
+			if err != nil {
+				if shouldFetchFiles {
+					return 0, fmt.Errorf("error getting file type: %v", err)
+				}
+				continue
+			}
+			file.MimeType = cachedData.MimeType
+			file.FileType = DetectFileType(cachedData.MimeType)
+		}
+	}
+
+	for i, file := range meta.Files {
+		switch file.FileType {
+		case types.FileTypeImage:
+			if common.IsOpenAITextModel(model) {
+				token, err := getImageToken(c, file, model, info.IsStream)
+				if err != nil {
+					return 0, fmt.Errorf("error counting image token, media index[%d], identifier[%s], err: %v", i, file.GetIdentifier(), err)
+				}
+				tkm += token
+			} else {
+				tkm += 520
+			}
+		case types.FileTypeAudio:
+			tkm += 256
+		case types.FileTypeVideo:
+			tkm += 4096 * 2
+		case types.FileTypeFile:
+			tkm += 4096
+		default:
+			tkm += 4096 // Default case for unknown file types
+		}
+	}
+
+	common.SetContextKey(c, constant.ContextKeyPromptTokens, tkm)
+	return tkm, nil
+}
+
+func CountTokenRealtime(info *relaycommon.RelayInfo, request dto.RealtimeEvent, model string) (int, int, error) {
+	audioToken := 0
+	textToken := 0
+	switch request.Type {
+	case dto.RealtimeEventTypeSessionUpdate:
+		if request.Session != nil {
+			msgTokens := CountTextToken(request.Session.Instructions, model)
+			textToken += msgTokens
+		}
+	case dto.RealtimeEventResponseAudioDelta:
+		// count audio token
+		atk, err := CountAudioTokenOutput(request.Delta, info.OutputAudioFormat)
+		if err != nil {
+			return 0, 0, fmt.Errorf("error counting audio token: %v", err)
+		}
+		audioToken += atk
+	case dto.RealtimeEventResponseAudioTranscriptionDelta, dto.RealtimeEventResponseFunctionCallArgumentsDelta:
+		// count text token
+		tkm := CountTextToken(request.Delta, model)
+		textToken += tkm
+	case dto.RealtimeEventInputAudioBufferAppend:
+		// count audio token
+		atk, err := CountAudioTokenInput(request.Audio, info.InputAudioFormat)
+		if err != nil {
+			return 0, 0, fmt.Errorf("error counting audio token: %v", err)
+		}
+		audioToken += atk
+	case dto.RealtimeEventConversationItemCreated:
+		if request.Item != nil {
+			switch request.Item.Type {
+			case "message":
+				for _, content := range request.Item.Content {
+					if content.Type == "input_text" {
+						tokens := CountTextToken(content.Text, model)
+						textToken += tokens
+					}
+				}
+			}
+		}
+	case dto.RealtimeEventTypeResponseDone:
+		// count tools token
+		if !info.IsFirstRequest {
+			if info.RealtimeTools != nil && len(info.RealtimeTools) > 0 {
+				for _, tool := range info.RealtimeTools {
+					toolTokens := CountTokenInput(tool, model)
+					textToken += 8
+					textToken += toolTokens
+				}
+			}
+		}
+	}
+	return textToken, audioToken, nil
+}
+
+func CountTokenInput(input any, model string) int {
+	switch v := input.(type) {
+	case string:
+		return CountTextToken(v, model)
+	case []string:
+		text := ""
+		for _, s := range v {
+			text += s
+		}
+		return CountTextToken(text, model)
+	case []interface{}:
+		text := ""
+		for _, item := range v {
+			text += fmt.Sprintf("%v", item)
+		}
+		return CountTextToken(text, model)
+	}
+	return CountTokenInput(fmt.Sprintf("%v", input), model)
+}
+
+func CountAudioTokenInput(audioBase64 string, audioFormat string) (int, error) {
+	if audioBase64 == "" {
+		return 0, nil
+	}
+	duration, err := parseAudio(audioBase64, audioFormat)
+	if err != nil {
+		return 0, err
+	}
+	return int(duration / 60 * 100 / 0.06), nil
+}
+
+func CountAudioTokenOutput(audioBase64 string, audioFormat string) (int, error) {
+	if audioBase64 == "" {
+		return 0, nil
+	}
+	duration, err := parseAudio(audioBase64, audioFormat)
+	if err != nil {
+		return 0, err
+	}
+	return int(duration / 60 * 200 / 0.24), nil
+}
+
+// CountTextToken 统计文本的token数量，仅OpenAI模型使用tokenizer，其余模型使用估算
+func CountTextToken(text string, model string) int {
+	if text == "" {
+		return 0
+	}
+	if common.IsOpenAITextModel(model) {
+		tokenEncoder := getTokenEncoder(model)
+		return getTokenNum(tokenEncoder, text)
+	} else {
+		// 非openai模型，使用tiktoken-go计算没有意义，使用估算节省资源
+		return EstimateTokenByModel(model, text)
+	}
+}

service/token_estimator.go

@@ -0,0 +1,230 @@
+package service
+
+import (
+	"math"
+	"strings"
+	"sync"
+	"unicode"
+)
+
+// Provider 定义模型厂商大类
+type Provider string
+
+const (
+	OpenAI  Provider = "openai"  // 代表 GPT-3.5, GPT-4, GPT-4o
+	Gemini  Provider = "gemini"  // 代表 Gemini 1.0, 1.5 Pro/Flash
+	Claude  Provider = "claude"  // 代表 Claude 3, 3.5 Sonnet
+	Unknown Provider = "unknown" // 兜底默认
+)
+
+// multipliers 定义不同厂商的计费权重
+type multipliers struct {
+	Word       float64 // 英文单词 (每词)
+	Number     float64 // 数字 (每连续数字串)
+	CJK        float64 // 中日韩字符 (每字)
+	Symbol     float64 // 普通标点符号 (每个)
+	MathSymbol float64 // 数学符号 (∑,∫,∂,√等，每个)
+	URLDelim   float64 // URL分隔符 (/,:,?,&,=,#,%) - tokenizer优化好
+	AtSign     float64 // @符号 - 导致单词切分，消耗较高
+	Emoji      float64 // Emoji表情 (每个)
+	Newline    float64 // 换行符/制表符 (每个)
+	Space      float64 // 空格 (每个)
+	BasePad    int     // 基础起步消耗 (Start/End tokens)
+}
+
+var (
+	multipliersMap = map[Provider]multipliers{
+		Gemini: {
+			Word: 1.15, Number: 2.8, CJK: 0.68, Symbol: 0.38, MathSymbol: 1.05, URLDelim: 1.2, AtSign: 2.5, Emoji: 1.08, Newline: 1.15, Space: 0.2, BasePad: 0,
+		},
+		Claude: {
+			Word: 1.13, Number: 1.63, CJK: 1.21, Symbol: 0.4, MathSymbol: 4.52, URLDelim: 1.26, AtSign: 2.82, Emoji: 2.6, Newline: 0.89, Space: 0.39, BasePad: 0,
+		},
+		OpenAI: {
+			Word: 1.02, Number: 1.55, CJK: 0.85, Symbol: 0.4, MathSymbol: 2.68, URLDelim: 1.0, AtSign: 2.0, Emoji: 2.12, Newline: 0.5, Space: 0.42, BasePad: 0,
+		},
+	}
+	multipliersLock sync.RWMutex
+)
+
+// getMultipliers 根据厂商获取权重配置
+func getMultipliers(p Provider) multipliers {
+	multipliersLock.RLock()
+	defer multipliersLock.RUnlock()
+
+	switch p {
+	case Gemini:
+		return multipliersMap[Gemini]
+	case Claude:
+		return multipliersMap[Claude]
+	case OpenAI:
+		return multipliersMap[OpenAI]
+	default:
+		// 默认兜底 (按 OpenAI 的算)
+		return multipliersMap[OpenAI]
+	}
+}
+
+// EstimateToken 计算 Token 数量
+func EstimateToken(provider Provider, text string) int {
+	m := getMultipliers(provider)
+	var count float64
+
+	// 状态机变量
+	type WordType int
+	const (
+		None WordType = iota
+		Latin
+		Number
+	)
+	currentWordType := None
+
+	for _, r := range text {
+		// 1. 处理空格和换行符
+		if unicode.IsSpace(r) {
+			currentWordType = None
+			// 换行符和制表符使用Newline权重
+			if r == '\n' || r == '\t' {
+				count += m.Newline
+			} else {
+				// 普通空格使用Space权重
+				count += m.Space
+			}
+			continue
+		}
+
+		// 2. 处理 CJK (中日韩) - 按字符计费
+		if isCJK(r) {
+			currentWordType = None
+			count += m.CJK
+			continue
+		}
+
+		// 3. 处理Emoji - 使用专门的Emoji权重
+		if isEmoji(r) {
+			currentWordType = None
+			count += m.Emoji
+			continue
+		}
+
+		// 4. 处理拉丁字母/数字 (英文单词)
+		if isLatinOrNumber(r) {
+			isNum := unicode.IsNumber(r)
+			newType := Latin
+			if isNum {
+				newType = Number
+			}
+
+			// 如果之前不在单词中，或者类型发生变化（字母<->数字），则视为新token
+			// 注意：对于OpenAI，通常"version 3.5"会切分，"abc123xyz"有时也会切分
+			// 这里简单起见，字母和数字切换时增加权重
+			if currentWordType == None || currentWordType != newType {
+				if newType == Number {
+					count += m.Number
+				} else {
+					count += m.Word
+				}
+				currentWordType = newType
+			}
+			// 单词中间的字符不额外计费
+			continue
+		}
+
+		// 5. 处理标点符号/特殊字符 - 按类型使用不同权重
+		currentWordType = None
+		if isMathSymbol(r) {
+			count += m.MathSymbol
+		} else if r == '@' {
+			count += m.AtSign
+		} else if isURLDelim(r) {
+			count += m.URLDelim
+		} else {
+			count += m.Symbol
+		}
+	}
+
+	// 向上取整并加上基础 padding
+	return int(math.Ceil(count)) + m.BasePad
+}
+
+// 辅助：判断是否为 CJK 字符
+func isCJK(r rune) bool {
+	return unicode.Is(unicode.Han, r) ||
+		(r >= 0x3040 && r <= 0x30FF) || // 日文
+		(r >= 0xAC00 && r <= 0xD7A3) // 韩文
+}
+
+// 辅助：判断是否为单词主体 (字母或数字)
+func isLatinOrNumber(r rune) bool {
+	return unicode.IsLetter(r) || unicode.IsNumber(r)
+}
+
+// 辅助：判断是否为Emoji字符
+func isEmoji(r rune) bool {
+	// Emoji的Unicode范围
+	// 基本范围：0x1F300-0x1F9FF (Emoticons, Symbols, Pictographs)
+	// 补充范围：0x2600-0x26FF (Misc Symbols), 0x2700-0x27BF (Dingbats)
+	// 表情符号：0x1F600-0x1F64F (Emoticons)
+	// 其他：0x1F900-0x1F9FF (Supplemental Symbols and Pictographs)
+	return (r >= 0x1F300 && r <= 0x1F9FF) ||
+		(r >= 0x2600 && r <= 0x26FF) ||
+		(r >= 0x2700 && r <= 0x27BF) ||
+		(r >= 0x1F600 && r <= 0x1F64F) ||
+		(r >= 0x1F900 && r <= 0x1F9FF) ||
+		(r >= 0x1FA00 && r <= 0x1FAFF) // Symbols and Pictographs Extended-A
+}
+
+// 辅助：判断是否为数学符号
+func isMathSymbol(r rune) bool {
+	// 数学运算符和符号
+	// 基本数学符号：∑ ∫ ∂ √ ∞ ≤ ≥ ≠ ≈ ± × ÷
+	// 上下标数字：² ³ ¹ ⁴ ⁵ ⁶ ⁷ ⁸ ⁹ ⁰
+	// 希腊字母等也常用于数学
+	mathSymbols := "∑∫∂√∞≤≥≠≈±×÷∈∉∋∌⊂⊃⊆⊇∪∩∧∨¬∀∃∄∅∆∇∝∟∠∡∢°′″‴⁺⁻⁼⁽⁾ⁿ₀₁₂₃₄₅₆₇₈₉₊₋₌₍₎²³¹⁴⁵⁶⁷⁸⁹⁰"
+	for _, m := range mathSymbols {
+		if r == m {
+			return true
+		}
+	}
+	// Mathematical Operators (U+2200–U+22FF)
+	if r >= 0x2200 && r <= 0x22FF {
+		return true
+	}
+	// Supplemental Mathematical Operators (U+2A00–U+2AFF)
+	if r >= 0x2A00 && r <= 0x2AFF {
+		return true
+	}
+	// Mathematical Alphanumeric Symbols (U+1D400–U+1D7FF)
+	if r >= 0x1D400 && r <= 0x1D7FF {
+		return true
+	}
+	return false
+}
+
+// 辅助：判断是否为URL分隔符（tokenizer对这些优化较好）
+func isURLDelim(r rune) bool {
+	// URL中常见的分隔符，tokenizer通常优化处理
+	urlDelims := "/:?&=;#%"
+	for _, d := range urlDelims {
+		if r == d {
+			return true
+		}
+	}
+	return false
+}
+
+func EstimateTokenByModel(model, text string) int {
+	// strings.Contains(model, "gpt-4o")
+	if text == "" {
+		return 0
+	}
+
+	model = strings.ToLower(model)
+	if strings.Contains(model, "gemini") {
+		return EstimateToken(Gemini, text)
+	} else if strings.Contains(model, "claude") {
+		return EstimateToken(Claude, text)
+	} else {
+		return EstimateToken(OpenAI, text)
+	}
+}

service/tokenizer.go

@@ -0,0 +1,63 @@
+package service
+
+import (
+	"sync"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/tiktoken-go/tokenizer"
+	"github.com/tiktoken-go/tokenizer/codec"
+)
+
+// tokenEncoderMap won't grow after initialization
+var defaultTokenEncoder tokenizer.Codec
+
+// tokenEncoderMap is used to store token encoders for different models
+var tokenEncoderMap = make(map[string]tokenizer.Codec)
+
+// tokenEncoderMutex protects tokenEncoderMap for concurrent access
+var tokenEncoderMutex sync.RWMutex
+
+func InitTokenEncoders() {
+	common.SysLog("initializing token encoders")
+	defaultTokenEncoder = codec.NewCl100kBase()
+	common.SysLog("token encoders initialized")
+}
+
+func getTokenEncoder(model string) tokenizer.Codec {
+	// First, try to get the encoder from cache with read lock
+	tokenEncoderMutex.RLock()
+	if encoder, exists := tokenEncoderMap[model]; exists {
+		tokenEncoderMutex.RUnlock()
+		return encoder
+	}
+	tokenEncoderMutex.RUnlock()
+
+	// If not in cache, create new encoder with write lock
+	tokenEncoderMutex.Lock()
+	defer tokenEncoderMutex.Unlock()
+
+	// Double-check if another goroutine already created the encoder
+	if encoder, exists := tokenEncoderMap[model]; exists {
+		return encoder
+	}
+
+	// Create new encoder
+	modelCodec, err := tokenizer.ForModel(tokenizer.Model(model))
+	if err != nil {
+		// Cache the default encoder for this model to avoid repeated failures
+		tokenEncoderMap[model] = defaultTokenEncoder
+		return defaultTokenEncoder
+	}
+
+	// Cache the new encoder
+	tokenEncoderMap[model] = modelCodec
+	return modelCodec
+}
+
+func getTokenNum(tokenEncoder tokenizer.Codec, text string) int {
+	if text == "" {
+		return 0
+	}
+	tkm, _ := tokenEncoder.Count(text)
+	return tkm
+}

service/usage_helpr.go

@@ -0,0 +1,33 @@
+package service
+
+import (
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/constant"
+	"github.com/QuantumNous/new-api/dto"
+	"github.com/gin-gonic/gin"
+)
+
+//func GetPromptTokens(textRequest dto.GeneralOpenAIRequest, relayMode int) (int, error) {
+//	switch relayMode {
+//	case constant.RelayModeChatCompletions:
+//		return CountTokenMessages(textRequest.Messages, textRequest.Model)
+//	case constant.RelayModeCompletions:
+//		return CountTokenInput(textRequest.Prompt, textRequest.Model), nil
+//	case constant.RelayModeModerations:
+//		return CountTokenInput(textRequest.Input, textRequest.Model), nil
+//	}
+//	return 0, errors.New("unknown relay mode")
+//}
+
+func ResponseText2Usage(c *gin.Context, responseText string, modeName string, promptTokens int) *dto.Usage {
+	common.SetContextKey(c, constant.ContextKeyLocalCountTokens, true)
+	usage := &dto.Usage{}
+	usage.PromptTokens = promptTokens
+	usage.CompletionTokens = EstimateTokenByModel(modeName, responseText)
+	usage.TotalTokens = usage.PromptTokens + usage.CompletionTokens
+	return usage
+}
+
+func ValidUsage(usage *dto.Usage) bool {
+	return usage != nil && (usage.PromptTokens != 0 || usage.CompletionTokens != 0)
+}

service/user_notify.go

@@ -0,0 +1,281 @@
+package service
+
+import (
+	"bytes"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"net/url"
+	"strings"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/dto"
+	"github.com/QuantumNous/new-api/model"
+	"github.com/QuantumNous/new-api/setting/system_setting"
+)
+
+func NotifyRootUser(t string, subject string, content string) {
+	user := model.GetRootUser().ToBaseUser()
+	err := NotifyUser(user.Id, user.Email, user.GetSetting(), dto.NewNotify(t, subject, content, nil))
+	if err != nil {
+		common.SysLog(fmt.Sprintf("failed to notify root user: %s", err.Error()))
+	}
+}
+
+func NotifyUpstreamModelUpdateWatchers(subject string, content string) {
+	var users []model.User
+	if err := model.DB.
+		Select("id", "email", "role", "status", "setting").
+		Where("status = ? AND role >= ?", common.UserStatusEnabled, common.RoleAdminUser).
+		Find(&users).Error; err != nil {
+		common.SysLog(fmt.Sprintf("failed to query upstream update notification users: %s", err.Error()))
+		return
+	}
+
+	notification := dto.NewNotify(dto.NotifyTypeChannelUpdate, subject, content, nil)
+	sentCount := 0
+	for _, user := range users {
+		userSetting := user.GetSetting()
+		if !userSetting.UpstreamModelUpdateNotifyEnabled {
+			continue
+		}
+		if err := NotifyUser(user.Id, user.Email, userSetting, notification); err != nil {
+			common.SysLog(fmt.Sprintf("failed to notify user %d for upstream model update: %s", user.Id, err.Error()))
+			continue
+		}
+		sentCount++
+	}
+	common.SysLog(fmt.Sprintf("upstream model update notifications sent: %d", sentCount))
+}
+
+func NotifyUser(userId int, userEmail string, userSetting dto.UserSetting, data dto.Notify) error {
+	notifyType := userSetting.NotifyType
+	if notifyType == "" {
+		notifyType = dto.NotifyTypeEmail
+	}
+
+	// Check notification limit
+	canSend, err := CheckNotificationLimit(userId, data.Type)
+	if err != nil {
+		common.SysLog(fmt.Sprintf("failed to check notification limit: %s", err.Error()))
+		return err
+	}
+	if !canSend {
+		return fmt.Errorf("notification limit exceeded for user %d with type %s", userId, notifyType)
+	}
+
+	switch notifyType {
+	case dto.NotifyTypeEmail:
+		// 优先使用设置中的通知邮箱，如果为空则使用用户的默认邮箱
+		emailToUse := userSetting.NotificationEmail
+		if emailToUse == "" {
+			emailToUse = userEmail
+		}
+		if emailToUse == "" {
+			common.SysLog(fmt.Sprintf("user %d has no email, skip sending email", userId))
+			return nil
+		}
+		return sendEmailNotify(emailToUse, data)
+	case dto.NotifyTypeWebhook:
+		webhookURLStr := userSetting.WebhookUrl
+		if webhookURLStr == "" {
+			common.SysLog(fmt.Sprintf("user %d has no webhook url, skip sending webhook", userId))
+			return nil
+		}
+
+		// 获取 webhook secret
+		webhookSecret := userSetting.WebhookSecret
+		return SendWebhookNotify(webhookURLStr, webhookSecret, data)
+	case dto.NotifyTypeBark:
+		barkURL := userSetting.BarkUrl
+		if barkURL == "" {
+			common.SysLog(fmt.Sprintf("user %d has no bark url, skip sending bark", userId))
+			return nil
+		}
+		return sendBarkNotify(barkURL, data)
+	case dto.NotifyTypeGotify:
+		gotifyUrl := userSetting.GotifyUrl
+		gotifyToken := userSetting.GotifyToken
+		if gotifyUrl == "" || gotifyToken == "" {
+			common.SysLog(fmt.Sprintf("user %d has no gotify url or token, skip sending gotify", userId))
+			return nil
+		}
+		return sendGotifyNotify(gotifyUrl, gotifyToken, userSetting.GotifyPriority, data)
+	}
+	return nil
+}
+
+func sendEmailNotify(userEmail string, data dto.Notify) error {
+	// make email content
+	content := data.Content
+	// 处理占位符
+	for _, value := range data.Values {
+		content = strings.Replace(content, dto.ContentValueParam, fmt.Sprintf("%v", value), 1)
+	}
+	return common.SendEmail(data.Title, userEmail, content)
+}
+
+func sendBarkNotify(barkURL string, data dto.Notify) error {
+	// 处理占位符
+	content := data.Content
+	for _, value := range data.Values {
+		content = strings.Replace(content, dto.ContentValueParam, fmt.Sprintf("%v", value), 1)
+	}
+
+	// 替换模板变量
+	finalURL := strings.ReplaceAll(barkURL, "{{title}}", url.QueryEscape(data.Title))
+	finalURL = strings.ReplaceAll(finalURL, "{{content}}", url.QueryEscape(content))
+
+	// 发送GET请求到Bark
+	var req *http.Request
+	var resp *http.Response
+	var err error
+
+	if system_setting.EnableWorker() {
+		// 使用worker发送请求
+		workerReq := &WorkerRequest{
+			URL:    finalURL,
+			Key:    system_setting.WorkerValidKey,
+			Method: http.MethodGet,
+			Headers: map[string]string{
+				"User-Agent": "OneAPI-Bark-Notify/1.0",
+			},
+		}
+
+		resp, err = DoWorkerRequest(workerReq)
+		if err != nil {
+			return fmt.Errorf("failed to send bark request through worker: %v", err)
+		}
+		defer resp.Body.Close()
+
+		// 检查响应状态
+		if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+			return fmt.Errorf("bark request failed with status code: %d", resp.StatusCode)
+		}
+	} else {
+		// SSRF防护：验证Bark URL（非Worker模式）
+		fetchSetting := system_setting.GetFetchSetting()
+		if err := common.ValidateURLWithFetchSetting(finalURL, fetchSetting.EnableSSRFProtection, fetchSetting.AllowPrivateIp, fetchSetting.DomainFilterMode, fetchSetting.IpFilterMode, fetchSetting.DomainList, fetchSetting.IpList, fetchSetting.AllowedPorts, fetchSetting.ApplyIPFilterForDomain); err != nil {
+			return fmt.Errorf("request reject: %v", err)
+		}
+
+		// 直接发送请求
+		req, err = http.NewRequest(http.MethodGet, finalURL, nil)
+		if err != nil {
+			return fmt.Errorf("failed to create bark request: %v", err)
+		}
+
+		// 设置User-Agent
+		req.Header.Set("User-Agent", "OneAPI-Bark-Notify/1.0")
+
+		// 发送请求
+		client := GetHttpClient()
+		resp, err = client.Do(req)
+		if err != nil {
+			return fmt.Errorf("failed to send bark request: %v", err)
+		}
+		defer resp.Body.Close()
+
+		// 检查响应状态
+		if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+			return fmt.Errorf("bark request failed with status code: %d", resp.StatusCode)
+		}
+	}
+
+	return nil
+}
+
+func sendGotifyNotify(gotifyUrl string, gotifyToken string, priority int, data dto.Notify) error {
+	// 处理占位符
+	content := data.Content
+	for _, value := range data.Values {
+		content = strings.Replace(content, dto.ContentValueParam, fmt.Sprintf("%v", value), 1)
+	}
+
+	// 构建完整的 Gotify API URL
+	// 确保 URL 以 /message 结尾
+	finalURL := strings.TrimSuffix(gotifyUrl, "/") + "/message?token=" + url.QueryEscape(gotifyToken)
+
+	// Gotify优先级范围0-10，如果超出范围则使用默认值5
+	if priority < 0 || priority > 10 {
+		priority = 5
+	}
+
+	// 构建 JSON payload
+	type GotifyMessage struct {
+		Title    string `json:"title"`
+		Message  string `json:"message"`
+		Priority int    `json:"priority"`
+	}
+
+	payload := GotifyMessage{
+		Title:    data.Title,
+		Message:  content,
+		Priority: priority,
+	}
+
+	// 序列化为 JSON
+	payloadBytes, err := json.Marshal(payload)
+	if err != nil {
+		return fmt.Errorf("failed to marshal gotify payload: %v", err)
+	}
+
+	var req *http.Request
+	var resp *http.Response
+
+	if system_setting.EnableWorker() {
+		// 使用worker发送请求
+		workerReq := &WorkerRequest{
+			URL:    finalURL,
+			Key:    system_setting.WorkerValidKey,
+			Method: http.MethodPost,
+			Headers: map[string]string{
+				"Content-Type": "application/json; charset=utf-8",
+				"User-Agent":   "OneAPI-Gotify-Notify/1.0",
+			},
+			Body: payloadBytes,
+		}
+
+		resp, err = DoWorkerRequest(workerReq)
+		if err != nil {
+			return fmt.Errorf("failed to send gotify request through worker: %v", err)
+		}
+		defer resp.Body.Close()
+
+		// 检查响应状态
+		if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+			return fmt.Errorf("gotify request failed with status code: %d", resp.StatusCode)
+		}
+	} else {
+		// SSRF防护：验证Gotify URL（非Worker模式）
+		fetchSetting := system_setting.GetFetchSetting()
+		if err := common.ValidateURLWithFetchSetting(finalURL, fetchSetting.EnableSSRFProtection, fetchSetting.AllowPrivateIp, fetchSetting.DomainFilterMode, fetchSetting.IpFilterMode, fetchSetting.DomainList, fetchSetting.IpList, fetchSetting.AllowedPorts, fetchSetting.ApplyIPFilterForDomain); err != nil {
+			return fmt.Errorf("request reject: %v", err)
+		}
+
+		// 直接发送请求
+		req, err = http.NewRequest(http.MethodPost, finalURL, bytes.NewBuffer(payloadBytes))
+		if err != nil {
+			return fmt.Errorf("failed to create gotify request: %v", err)
+		}
+
+		// 设置请求头
+		req.Header.Set("Content-Type", "application/json; charset=utf-8")
+		req.Header.Set("User-Agent", "NewAPI-Gotify-Notify/1.0")
+
+		// 发送请求
+		client := GetHttpClient()
+		resp, err = client.Do(req)
+		if err != nil {
+			return fmt.Errorf("failed to send gotify request: %v", err)
+		}
+		defer resp.Body.Close()
+
+		// 检查响应状态
+		if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+			return fmt.Errorf("gotify request failed with status code: %d", resp.StatusCode)
+		}
+	}
+
+	return nil
+}

service/violation_fee.go

@@ -0,0 +1,164 @@
+package service
+
+import (
+	"fmt"
+	"strings"
+	"time"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/logger"
+	"github.com/QuantumNous/new-api/model"
+	relaycommon "github.com/QuantumNous/new-api/relay/common"
+	"github.com/QuantumNous/new-api/setting/model_setting"
+	"github.com/QuantumNous/new-api/types"
+
+	"github.com/shopspring/decimal"
+
+	"github.com/gin-gonic/gin"
+)
+
+const (
+	ViolationFeeCodePrefix     = "violation_fee."
+	CSAMViolationMarker        = "Failed check: SAFETY_CHECK_TYPE"
+	ContentViolatesUsageMarker = "Content violates usage guidelines"
+)
+
+func IsViolationFeeCode(code types.ErrorCode) bool {
+	return strings.HasPrefix(string(code), ViolationFeeCodePrefix)
+}
+
+func HasCSAMViolationMarker(err *types.NewAPIError) bool {
+	if err == nil {
+		return false
+	}
+	if strings.Contains(err.Error(), CSAMViolationMarker) || strings.Contains(err.Error(), ContentViolatesUsageMarker) {
+		return true
+	}
+	msg := err.ToOpenAIError().Message
+	return strings.Contains(msg, CSAMViolationMarker) || strings.Contains(err.Error(), ContentViolatesUsageMarker)
+}
+
+func WrapAsViolationFeeGrokCSAM(err *types.NewAPIError) *types.NewAPIError {
+	if err == nil {
+		return nil
+	}
+	oai := err.ToOpenAIError()
+	oai.Type = string(types.ErrorCodeViolationFeeGrokCSAM)
+	oai.Code = string(types.ErrorCodeViolationFeeGrokCSAM)
+	return types.WithOpenAIError(oai, err.StatusCode, types.ErrOptionWithSkipRetry())
+}
+
+// NormalizeViolationFeeError ensures:
+// - if the CSAM marker is present, error.code is set to a stable violation-fee code and skip-retry is enabled.
+// - if error.code already has the violation-fee prefix, skip-retry is enabled.
+//
+// It must be called before retry decision logic.
+func NormalizeViolationFeeError(err *types.NewAPIError) *types.NewAPIError {
+	if err == nil {
+		return nil
+	}
+
+	if HasCSAMViolationMarker(err) {
+		return WrapAsViolationFeeGrokCSAM(err)
+	}
+
+	if IsViolationFeeCode(err.GetErrorCode()) {
+		oai := err.ToOpenAIError()
+		return types.WithOpenAIError(oai, err.StatusCode, types.ErrOptionWithSkipRetry())
+	}
+
+	return err
+}
+
+func shouldChargeViolationFee(err *types.NewAPIError) bool {
+	if err == nil {
+		return false
+	}
+	if err.GetErrorCode() == types.ErrorCodeViolationFeeGrokCSAM {
+		return true
+	}
+	// In case some callers didn't normalize, keep a safety net.
+	return HasCSAMViolationMarker(err)
+}
+
+func calcViolationFeeQuota(amount, groupRatio float64) int {
+	if amount <= 0 {
+		return 0
+	}
+	if groupRatio <= 0 {
+		return 0
+	}
+	quota := decimal.NewFromFloat(amount).
+		Mul(decimal.NewFromFloat(common.QuotaPerUnit)).
+		Mul(decimal.NewFromFloat(groupRatio)).
+		Round(0).
+		IntPart()
+	if quota <= 0 {
+		return 0
+	}
+	return int(quota)
+}
+
+// ChargeViolationFeeIfNeeded charges an additional fee after the normal flow finishes (including refund).
+// It uses Grok fee settings as the fee policy.
+func ChargeViolationFeeIfNeeded(ctx *gin.Context, relayInfo *relaycommon.RelayInfo, apiErr *types.NewAPIError) bool {
+	if ctx == nil || relayInfo == nil || apiErr == nil {
+		return false
+	}
+	//if relayInfo.IsPlayground {
+	//	return false
+	//}
+	if !shouldChargeViolationFee(apiErr) {
+		return false
+	}
+
+	settings := model_setting.GetGrokSettings()
+	if settings == nil || !settings.ViolationDeductionEnabled {
+		return false
+	}
+
+	groupRatio := relayInfo.PriceData.GroupRatioInfo.GroupRatio
+	feeQuota := calcViolationFeeQuota(settings.ViolationDeductionAmount, groupRatio)
+	if feeQuota <= 0 {
+		return false
+	}
+
+	if err := PostConsumeQuota(relayInfo, feeQuota, 0, true); err != nil {
+		logger.LogError(ctx, fmt.Sprintf("failed to charge violation fee: %s", err.Error()))
+		return false
+	}
+
+	model.UpdateUserUsedQuotaAndRequestCount(relayInfo.UserId, feeQuota)
+	model.UpdateChannelUsedQuota(relayInfo.ChannelId, feeQuota)
+
+	useTimeSeconds := time.Now().Unix() - relayInfo.StartTime.Unix()
+	tokenName := ctx.GetString("token_name")
+	oai := apiErr.ToOpenAIError()
+
+	other := map[string]any{
+		"violation_fee":        true,
+		"violation_fee_code":   string(types.ErrorCodeViolationFeeGrokCSAM),
+		"fee_quota":            feeQuota,
+		"base_amount":          settings.ViolationDeductionAmount,
+		"group_ratio":          groupRatio,
+		"status_code":          apiErr.StatusCode,
+		"upstream_error_type":  oai.Type,
+		"upstream_error_code":  fmt.Sprintf("%v", oai.Code),
+		"violation_fee_marker": CSAMViolationMarker,
+	}
+
+	model.RecordConsumeLog(ctx, relayInfo.UserId, model.RecordConsumeLogParams{
+		ChannelId:      relayInfo.ChannelId,
+		ModelName:      relayInfo.OriginModelName,
+		TokenName:      tokenName,
+		Quota:          feeQuota,
+		Content:        "Violation fee charged",
+		TokenId:        relayInfo.TokenId,
+		UseTimeSeconds: int(useTimeSeconds),
+		IsStream:       relayInfo.IsStream,
+		Group:          relayInfo.UsingGroup,
+		Other:          other,
+	})
+
+	return true
+}

service/webhook.go

@@ -0,0 +1,126 @@
+package service
+
+import (
+	"bytes"
+	"crypto/hmac"
+	"crypto/sha256"
+	"encoding/hex"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"time"
+
+	"github.com/QuantumNous/new-api/common"
+	"github.com/QuantumNous/new-api/dto"
+	"github.com/QuantumNous/new-api/setting/system_setting"
+)
+
+// WebhookPayload webhook 通知的负载数据
+type WebhookPayload struct {
+	Type      string        `json:"type"`
+	Title     string        `json:"title"`
+	Content   string        `json:"content"`
+	Values    []interface{} `json:"values,omitempty"`
+	Timestamp int64         `json:"timestamp"`
+}
+
+// generateSignature 生成 webhook 签名
+func generateSignature(secret string, payload []byte) string {
+	h := hmac.New(sha256.New, []byte(secret))
+	h.Write(payload)
+	return hex.EncodeToString(h.Sum(nil))
+}
+
+// SendWebhookNotify 发送 webhook 通知
+func SendWebhookNotify(webhookURL string, secret string, data dto.Notify) error {
+	// 处理占位符
+	content := data.Content
+	for _, value := range data.Values {
+		content = fmt.Sprintf(content, value)
+	}
+
+	// 构建 webhook 负载
+	payload := WebhookPayload{
+		Type:      data.Type,
+		Title:     data.Title,
+		Content:   content,
+		Values:    data.Values,
+		Timestamp: time.Now().Unix(),
+	}
+
+	// 序列化负载
+	payloadBytes, err := json.Marshal(payload)
+	if err != nil {
+		return fmt.Errorf("failed to marshal webhook payload: %v", err)
+	}
+
+	// 创建 HTTP 请求
+	var req *http.Request
+	var resp *http.Response
+
+	if system_setting.EnableWorker() {
+		// 构建worker请求数据
+		workerReq := &WorkerRequest{
+			URL:    webhookURL,
+			Key:    system_setting.WorkerValidKey,
+			Method: http.MethodPost,
+			Headers: map[string]string{
+				"Content-Type": "application/json",
+			},
+			Body: payloadBytes,
+		}
+
+		// 如果有secret，添加签名到headers
+		if secret != "" {
+			signature := generateSignature(secret, payloadBytes)
+			workerReq.Headers["X-Webhook-Signature"] = signature
+			workerReq.Headers["Authorization"] = "Bearer " + secret
+		}
+
+		resp, err = DoWorkerRequest(workerReq)
+		if err != nil {
+			return fmt.Errorf("failed to send webhook request through worker: %v", err)
+		}
+		defer resp.Body.Close()
+
+		// 检查响应状态
+		if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+			return fmt.Errorf("webhook request failed with status code: %d", resp.StatusCode)
+		}
+	} else {
+		// SSRF防护：验证Webhook URL（非Worker模式）
+		fetchSetting := system_setting.GetFetchSetting()
+		if err := common.ValidateURLWithFetchSetting(webhookURL, fetchSetting.EnableSSRFProtection, fetchSetting.AllowPrivateIp, fetchSetting.DomainFilterMode, fetchSetting.IpFilterMode, fetchSetting.DomainList, fetchSetting.IpList, fetchSetting.AllowedPorts, fetchSetting.ApplyIPFilterForDomain); err != nil {
+			return fmt.Errorf("request reject: %v", err)
+		}
+
+		req, err = http.NewRequest(http.MethodPost, webhookURL, bytes.NewBuffer(payloadBytes))
+		if err != nil {
+			return fmt.Errorf("failed to create webhook request: %v", err)
+		}
+
+		// 设置请求头
+		req.Header.Set("Content-Type", "application/json")
+
+		// 如果有 secret，生成签名
+		if secret != "" {
+			signature := generateSignature(secret, payloadBytes)
+			req.Header.Set("X-Webhook-Signature", signature)
+		}
+
+		// 发送请求
+		client := GetHttpClient()
+		resp, err = client.Do(req)
+		if err != nil {
+			return fmt.Errorf("failed to send webhook request: %v", err)
+		}
+		defer resp.Body.Close()
+
+		// 检查响应状态
+		if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+			return fmt.Errorf("webhook request failed with status code: %d", resp.StatusCode)
+		}
+	}
+
+	return nil
+}