new-api/controller/relay.go

package controller

import (
	"errors"
	"fmt"
	"io"
	"log"
	"net/http"
	"strconv"
	"strings"
	"time"

	"github.com/QuantumNous/new-api/common"
	"github.com/QuantumNous/new-api/common/metrics"
	"github.com/QuantumNous/new-api/constant"
	"github.com/QuantumNous/new-api/dto"
	"github.com/QuantumNous/new-api/logger"
	"github.com/QuantumNous/new-api/middleware"
	"github.com/QuantumNous/new-api/model"
	"github.com/QuantumNous/new-api/relay"
	relaycommon "github.com/QuantumNous/new-api/relay/common"
	relayconstant "github.com/QuantumNous/new-api/relay/constant"
	"github.com/QuantumNous/new-api/relay/helper"
	"github.com/QuantumNous/new-api/service"
	"github.com/QuantumNous/new-api/setting"
	"github.com/QuantumNous/new-api/setting/operation_setting"
	"github.com/QuantumNous/new-api/setting/ratio_setting"
	"github.com/QuantumNous/new-api/types"

	"github.com/bytedance/gopkg/util/gopool"

	"github.com/gin-gonic/gin"
	"github.com/gorilla/websocket"
)

func relayHandler(c *gin.Context, info *relaycommon.RelayInfo) *types.NewAPIError {
	var err *types.NewAPIError
	switch info.RelayMode {
	case relayconstant.RelayModeImagesGenerations, relayconstant.RelayModeImagesEdits:
		err = relay.ImageHelper(c, info)
	case relayconstant.RelayModeAudioSpeech:
		fallthrough
	case relayconstant.RelayModeAudioTranslation:
		fallthrough
	case relayconstant.RelayModeAudioTranscription:
		err = relay.AudioHelper(c, info)
	case relayconstant.RelayModeRerank:
		err = relay.RerankHelper(c, info)
	case relayconstant.RelayModeEmbeddings:
		err = relay.EmbeddingHelper(c, info)
	case relayconstant.RelayModeResponses, relayconstant.RelayModeResponsesCompact:
		err = relay.ResponsesHelper(c, info)
	default:
		err = relay.TextHelper(c, info)
	}
	return err
}

func geminiRelayHandler(c *gin.Context, info *relaycommon.RelayInfo) *types.NewAPIError {
	var err *types.NewAPIError
	if strings.Contains(c.Request.URL.Path, "embed") {
		err = relay.GeminiEmbeddingHandler(c, info)
	} else {
		err = relay.GeminiHelper(c, info)
	}
	return err
}

func Relay(c *gin.Context, relayFormat types.RelayFormat) {

	requestId := c.GetString(common.RequestIdKey)
	//group := common.GetContextKeyString(c, constant.ContextKeyUsingGroup)
	//originalModel := common.GetContextKeyString(c, constant.ContextKeyOriginalModel)

	var (
		newAPIError *types.NewAPIError
		ws          *websocket.Conn
	)

	if relayFormat == types.RelayFormatOpenAIRealtime {
		var err error
		ws, err = upgrader.Upgrade(c.Writer, c.Request, nil)
		if err != nil {
			helper.WssError(c, ws, types.NewError(err, types.ErrorCodeGetChannelFailed, types.ErrOptionWithSkipRetry()).ToOpenAIError())
			return
		}
		defer ws.Close()
	}

	defer func() {
		if newAPIError != nil {
			logger.LogError(c, fmt.Sprintf("relay error: %s", newAPIError.Error()))

			// Prometheus: 错误计数
			if metrics.IsEnabled() {
				channelId := strconv.Itoa(c.GetInt("channel_id"))
				modelName := c.GetString("original_model")
				errorType, errorCode := metrics.ClassifyError(newAPIError.StatusCode, newAPIError.Error())
				metrics.GetMetrics().RequestErrorsTotal.WithLabelValues(channelId, modelName, errorType, errorCode).Inc()
			}

			newAPIError.SetMessage(common.MessageWithRequestId(newAPIError.Error(), requestId))
			switch relayFormat {
			case types.RelayFormatOpenAIRealtime:
				helper.WssError(c, ws, newAPIError.ToOpenAIError())
			case types.RelayFormatClaude:
				c.JSON(newAPIError.StatusCode, gin.H{
					"type":  "error",
					"error": newAPIError.ToClaudeError(),
				})
			default:
				c.JSON(newAPIError.StatusCode, gin.H{
					"error": newAPIError.ToOpenAIError(),
				})
			}
		}
	}()

	request, err := helper.GetAndValidateRequest(c, relayFormat)
	if err != nil {
		// Map "request body too large" to 413 so clients can handle it correctly
		if common.IsRequestBodyTooLargeError(err) || errors.Is(err, common.ErrRequestBodyTooLarge) {
			newAPIError = types.NewErrorWithStatusCode(err, types.ErrorCodeReadRequestBodyFailed, http.StatusRequestEntityTooLarge, types.ErrOptionWithSkipRetry())
		} else {
			newAPIError = types.NewError(err, types.ErrorCodeInvalidRequest)
		}
		return
	}

	relayInfo, err := relaycommon.GenRelayInfo(c, relayFormat, request, ws)
	if err != nil {
		newAPIError = types.NewError(err, types.ErrorCodeGenRelayInfoFailed)
		return
	}

	needSensitiveCheck := setting.ShouldCheckPromptSensitive()
	needCountToken := constant.CountToken
	// Avoid building huge CombineText (strings.Join) when token counting and sensitive check are both disabled.
	var meta *types.TokenCountMeta
	if needSensitiveCheck || needCountToken {
		meta = request.GetTokenCountMeta()
	} else {
		meta = fastTokenCountMetaForPricing(request)
	}

	if needSensitiveCheck && meta != nil {
		contains, words := service.CheckSensitiveText(meta.CombineText)
		if contains {
			logger.LogWarn(c, fmt.Sprintf("user sensitive words detected: %s", strings.Join(words, ", ")))
			newAPIError = types.NewError(err, types.ErrorCodeSensitiveWordsDetected)
			return
		}
	}

	tokens, err := service.EstimateRequestToken(c, meta, relayInfo)
	if err != nil {
		newAPIError = types.NewError(err, types.ErrorCodeCountTokenFailed)
		return
	}

	relayInfo.SetEstimatePromptTokens(tokens)

	priceData, err := helper.ModelPriceHelper(c, relayInfo, tokens, meta)
	if err != nil {
		newAPIError = types.NewError(err, types.ErrorCodeModelPriceError)
		return
	}

	// common.SetContextKey(c, constant.ContextKeyTokenCountMeta, meta)

	if priceData.FreeModel {
		logger.LogInfo(c, fmt.Sprintf("模型 %s 免费，跳过预扣费", relayInfo.OriginModelName))
	} else {
		common.SysLog(fmt.Sprintf("[RegionSync] relay PreConsumeBilling: userId=%d, QuotaToPreConsume=%d, FreeModel=%v, UsePrice=%v, ModelRatio=%.4f, GroupRatio=%.4f, ModelPrice=%.4f, model=%s",
			relayInfo.UserId, priceData.QuotaToPreConsume, priceData.FreeModel, priceData.UsePrice, priceData.ModelRatio, priceData.GroupRatioInfo.GroupRatio, priceData.ModelPrice, relayInfo.OriginModelName))
		newAPIError = service.PreConsumeBilling(c, priceData.QuotaToPreConsume, relayInfo)
		if newAPIError != nil {
			return
		}
	}

	defer func() {
		// Only return quota if downstream failed and quota was actually pre-consumed
		if newAPIError != nil {
			newAPIError = service.NormalizeViolationFeeError(newAPIError)
			if relayInfo.Billing != nil {
				relayInfo.Billing.Refund(c)
			}
			service.ChargeViolationFeeIfNeeded(c, relayInfo, newAPIError)
		}
	}()

	retryParam := &service.RetryParam{
		Ctx:                       c,
		TokenGroup:                relayInfo.TokenGroup,
		ModelName:                 relayInfo.OriginModelName,
		Retry:                     common.GetPointer(0),
		RequireMatrixUsageBilling: relayInfo.RequireMatrixUsageBilling,
	}

	for ; retryParam.GetRetry() <= common.RetryTimes; retryParam.IncreaseRetry() {
		channel, _, channelErr := getChannel(c, relayInfo, retryParam)
		if channelErr != nil {
			logger.LogError(c, channelErr.Error())
			newAPIError = channelErr
			break
		}

		addUsedChannel(c, channel.Id)
		bodyStorage, bodyErr := common.GetBodyStorage(c)
		if bodyErr != nil {
			// Ensure consistent 413 for oversized bodies even when error occurs later (e.g., retry path)
			if common.IsRequestBodyTooLargeError(bodyErr) || errors.Is(bodyErr, common.ErrRequestBodyTooLarge) {
				newAPIError = types.NewErrorWithStatusCode(bodyErr, types.ErrorCodeReadRequestBodyFailed, http.StatusRequestEntityTooLarge, types.ErrOptionWithSkipRetry())
			} else {
				newAPIError = types.NewErrorWithStatusCode(bodyErr, types.ErrorCodeReadRequestBodyFailed, http.StatusBadRequest, types.ErrOptionWithSkipRetry())
			}
			break
		}
		c.Request.Body = io.NopCloser(bodyStorage)

		switch relayFormat {
		case types.RelayFormatOpenAIRealtime:
			newAPIError = relay.WssHelper(c, relayInfo)
		case types.RelayFormatClaude:
			newAPIError = relay.ClaudeHelper(c, relayInfo)
		case types.RelayFormatGemini:
			newAPIError = geminiRelayHandler(c, relayInfo)
		default:
			newAPIError = relayHandler(c, relayInfo)
		}

		if newAPIError == nil {
			return
		}

		newAPIError = service.NormalizeViolationFeeError(newAPIError)

		processChannelError(c, *types.NewChannelError(channel.Id, channel.Type, channel.Name, channel.ChannelInfo.IsMultiKey, common.GetContextKeyString(c, constant.ContextKeyChannelKey), channel.GetAutoBan()), newAPIError)

		if !shouldRetry(c, newAPIError, common.RetryTimes-retryParam.GetRetry()) {
			break
		}

		// Prometheus: 重试计数
		if metrics.IsEnabled() {
			channelIdStr := strconv.Itoa(channel.Id)
			metrics.GetMetrics().RequestRetriesTotal.WithLabelValues(channelIdStr, relayInfo.OriginModelName).Inc()
		}
	}

	useChannel := c.GetStringSlice("use_channel")
	if len(useChannel) > 1 {
		retryLogStr := fmt.Sprintf("重试：%s", strings.Trim(strings.Join(strings.Fields(fmt.Sprint(useChannel)), "->"), "[]"))
		logger.LogInfo(c, retryLogStr)
	}
}

var upgrader = websocket.Upgrader{
	Subprotocols: []string{"realtime"}, // WS 握手支持的协议，如果有使用 Sec-WebSocket-Protocol，则必须在此声明对应的 Protocol TODO add other protocol
	CheckOrigin: func(r *http.Request) bool {
		return true // 允许跨域
	},
}

func addUsedChannel(c *gin.Context, channelId int) {
	useChannel := c.GetStringSlice("use_channel")
	useChannel = append(useChannel, fmt.Sprintf("%d", channelId))
	c.Set("use_channel", useChannel)
}

func fastTokenCountMetaForPricing(request dto.Request) *types.TokenCountMeta {
	if request == nil {
		return &types.TokenCountMeta{}
	}
	meta := &types.TokenCountMeta{
		TokenType: types.TokenTypeTokenizer,
	}
	switch r := request.(type) {
	case *dto.GeneralOpenAIRequest:
		if r.MaxCompletionTokens > r.MaxTokens {
			meta.MaxTokens = int(r.MaxCompletionTokens)
		} else {
			meta.MaxTokens = int(r.MaxTokens)
		}
	case *dto.OpenAIResponsesRequest:
		meta.MaxTokens = int(r.MaxOutputTokens)
	case *dto.ClaudeRequest:
		meta.MaxTokens = int(r.MaxTokens)
	case *dto.ImageRequest:
		// Pricing for image requests depends on ImagePriceRatio; safe to compute even when CountToken is disabled.
		return r.GetTokenCountMeta()
	default:
		// Best-effort: leave CombineText empty to avoid large allocations.
	}
	return meta
}

func concreteDoubaoVideoBindingGroup(c *gin.Context, group string) string {
	group = strings.TrimSpace(group)
	if group == "" {
		group = strings.TrimSpace(common.GetContextKeyString(c, constant.ContextKeyTokenGroup))
	}
	if group == "auto" {
		autoGroup := strings.TrimSpace(common.GetContextKeyString(c, constant.ContextKeyAutoGroup))
		if autoGroup == "" || autoGroup == "auto" {
			return ""
		}
		return autoGroup
	}
	return group
}

func persistDoubaoVideoTaskBindingIfNeeded(userId int, group string, channel *model.Channel) error {
	group = strings.TrimSpace(group)
	if group == "" || group == "auto" {
		return nil
	}
	if !service.IsUsableDoubaoAssetChannel(channel, group) {
		return nil
	}
	return model.BindUserAssetChannel(userId, service.DoubaoAssetChannelType, group, channel.Id)
}

func getChannel(c *gin.Context, info *relaycommon.RelayInfo, retryParam *service.RetryParam) (*model.Channel, string, *types.NewAPIError) {
	if info.ChannelMeta == nil {
		autoBan := c.GetBool("auto_ban")
		autoBanInt := 1
		if !autoBan {
			autoBanInt = 0
		}
		return &model.Channel{
			Id:      c.GetInt("channel_id"),
			Type:    c.GetInt("channel_type"),
			Name:    c.GetString("channel_name"),
			AutoBan: &autoBanInt,
		}, concreteDoubaoVideoBindingGroup(c, info.TokenGroup), nil
	}
	channel, selectGroup, err := service.CacheGetRandomSatisfiedChannel(retryParam)

	info.PriceData.GroupRatioInfo = helper.HandleGroupRatio(c, info)

	if err != nil {
		if apiErr, ok := err.(*types.NewAPIError); ok {
			return nil, selectGroup, apiErr
		}
		return nil, selectGroup, types.NewError(fmt.Errorf("获取分组 %s 下模型 %s 的可用渠道失败（retry）: %s", selectGroup, info.OriginModelName, err.Error()), types.ErrorCodeGetChannelFailed, types.ErrOptionWithSkipRetry())
	}
	if channel == nil {
		return nil, selectGroup, types.NewError(fmt.Errorf("分组 %s 下模型 %s 的可用渠道不存在（retry）", selectGroup, info.OriginModelName), types.ErrorCodeGetChannelFailed, types.ErrOptionWithSkipRetry())
	}

	newAPIError := middleware.SetupContextForSelectedChannel(c, channel, info.OriginModelName)
	if newAPIError != nil {
		return nil, selectGroup, newAPIError
	}
	return channel, selectGroup, nil
}

func requiredTaskChannelTypeForRequest(c *gin.Context) int {
	if strings.HasPrefix(c.Request.URL.Path, "/api/v3/contents/generations/tasks") {
		return constant.ChannelTypeDoubaoVideoCompatibleAiping
	}
	return 0
}

func shouldRetry(c *gin.Context, openaiErr *types.NewAPIError, retryTimes int) bool {
	if openaiErr == nil {
		return false
	}
	if service.ShouldSkipRetryAfterChannelAffinityFailure(c) {
		return false
	}
	if types.IsChannelError(openaiErr) {
		return true
	}
	if types.IsSkipRetryError(openaiErr) {
		return false
	}
	if retryTimes <= 0 {
		return false
	}
	code := openaiErr.StatusCode
	if code >= 200 && code < 300 {
		return false
	}
	if code < 100 || code > 599 {
		return true
	}
	return operation_setting.ShouldRetryByStatusCode(code)
}

func processChannelError(c *gin.Context, channelError types.ChannelError, err *types.NewAPIError) {
	logger.LogError(c, fmt.Sprintf("channel error (channel #%d, status code: %d): %s", channelError.ChannelId, err.StatusCode, err.Error()))
	// 不要使用context获取渠道信息，异步处理时可能会出现渠道信息不一致的情况
	// do not use context to get channel info, there may be inconsistent channel info when processing asynchronously
	if service.ShouldDisableChannel(channelError.ChannelType, err) && channelError.AutoBan {
		gopool.Go(func() {
			service.DisableChannel(channelError, err.ErrorWithStatusCode())
		})
	}

	if constant.ErrorLogEnabled && types.IsRecordErrorLog(err) {
		// 保存错误日志到mysql中
		userId := c.GetInt("id")
		tokenName := c.GetString("token_name")
		modelName := c.GetString("original_model")
		tokenId := c.GetInt("token_id")
		userGroup := c.GetString("group")
		channelId := c.GetInt("channel_id")
		other := make(map[string]interface{})
		if c.Request != nil && c.Request.URL != nil {
			other["request_path"] = c.Request.URL.Path
		}
		other["error_type"] = err.GetErrorType()
		other["error_code"] = err.GetErrorCode()
		other["status_code"] = err.StatusCode
		other["channel_id"] = channelId
		other["channel_name"] = c.GetString("channel_name")
		other["channel_type"] = c.GetInt("channel_type")
		if err.UpstreamRequestId != "" {
			other["upstream_request_id"] = err.UpstreamRequestId
		}
		if err.UpstreamBody != "" {
			other["upstream_body"] = err.UpstreamBody
		}
		adminInfo := make(map[string]interface{})
		adminInfo["use_channel"] = c.GetStringSlice("use_channel")
		isMultiKey := common.GetContextKeyBool(c, constant.ContextKeyChannelIsMultiKey)
		if isMultiKey {
			adminInfo["is_multi_key"] = true
			adminInfo["multi_key_index"] = common.GetContextKeyInt(c, constant.ContextKeyChannelMultiKeyIndex)
		}
		service.AppendChannelAffinityAdminInfo(c, adminInfo)
		other["admin_info"] = adminInfo
		startTime := common.GetContextKeyTime(c, constant.ContextKeyRequestStartTime)
		if startTime.IsZero() {
			startTime = time.Now()
		}
		useTimeSeconds := int(time.Since(startTime).Seconds())
		model.RecordErrorLog(c, userId, channelId, modelName, tokenName, err.MaskSensitiveErrorWithStatusCode(), tokenId, useTimeSeconds, false, userGroup, other)
	}

}

func RelayMidjourney(c *gin.Context) {
	relayInfo, err := relaycommon.GenRelayInfo(c, types.RelayFormatMjProxy, nil, nil)

	if err != nil {
		c.JSON(http.StatusInternalServerError, gin.H{
			"description": fmt.Sprintf("failed to generate relay info: %s", err.Error()),
			"type":        "upstream_error",
			"code":        4,
		})
		return
	}

	var mjErr *dto.MidjourneyResponse
	switch relayInfo.RelayMode {
	case relayconstant.RelayModeMidjourneyNotify:
		mjErr = relay.RelayMidjourneyNotify(c)
	case relayconstant.RelayModeMidjourneyTaskFetch, relayconstant.RelayModeMidjourneyTaskFetchByCondition:
		mjErr = relay.RelayMidjourneyTask(c, relayInfo.RelayMode)
	case relayconstant.RelayModeMidjourneyTaskImageSeed:
		mjErr = relay.RelayMidjourneyTaskImageSeed(c)
	case relayconstant.RelayModeSwapFace:
		mjErr = relay.RelaySwapFace(c, relayInfo)
	default:
		mjErr = relay.RelayMidjourneySubmit(c, relayInfo)
	}
	//err = relayMidjourneySubmit(c, relayMode)
	log.Println(mjErr)
	if mjErr != nil {
		statusCode := http.StatusBadRequest
		if mjErr.Code == 30 {
			mjErr.Result = "当前分组负载已饱和，请稍后再试，或升级账户以提升服务质量。"
			statusCode = http.StatusTooManyRequests
		}
		c.JSON(statusCode, gin.H{
			"description": fmt.Sprintf("%s %s", mjErr.Description, mjErr.Result),
			"type":        "upstream_error",
			"code":        mjErr.Code,
		})
		channelId := c.GetInt("channel_id")
		logger.LogError(c, fmt.Sprintf("relay error (channel #%d, status code %d): %s", channelId, statusCode, fmt.Sprintf("%s %s", mjErr.Description, mjErr.Result)))
	}
}

func RelayNotImplemented(c *gin.Context) {
	err := types.OpenAIError{
		Message: "API not implemented",
		Type:    "new_api_error",
		Param:   "",
		Code:    "api_not_implemented",
	}
	c.JSON(http.StatusNotImplemented, gin.H{
		"error": err,
	})
}

func RelayNotFound(c *gin.Context) {
	err := types.OpenAIError{
		Message: fmt.Sprintf("Invalid URL (%s %s)", c.Request.Method, c.Request.URL.Path),
		Type:    "invalid_request_error",
		Param:   "",
		Code:    "",
	}
	c.JSON(http.StatusNotFound, gin.H{
		"error": err,
	})
}

func RelayTaskFetch(c *gin.Context) {
	relayInfo, err := relaycommon.GenRelayInfo(c, types.RelayFormatTask, nil, nil)
	if err != nil {
		c.JSON(http.StatusInternalServerError, &dto.TaskError{
			Code:       "gen_relay_info_failed",
			Message:    err.Error(),
			StatusCode: http.StatusInternalServerError,
		})
		return
	}
	if taskErr := relay.RelayTaskFetch(c, relayInfo.RelayMode); taskErr != nil {
		respondTaskError(c, taskErr)
	}
}

func preloadTaskPricingConfig(c *gin.Context, info *relaycommon.RelayInfo) *dto.TaskError {
	modelName := strings.TrimSpace(info.OriginModelName)
	action := strings.TrimSpace(info.Action)
	contentType := c.Request.Header.Get("Content-Type")

	storage, err := common.GetBodyStorage(c)
	if err != nil {
		status := http.StatusBadRequest
		if common.IsRequestBodyTooLargeError(err) || errors.Is(err, common.ErrRequestBodyTooLarge) {
			status = http.StatusRequestEntityTooLarge
		}
		return service.TaskErrorWrapperLocal(err, "read_request_body_failed", status)
	}
	defer func() {
		_, _ = storage.Seek(0, io.SeekStart)
		c.Request.Body = io.NopCloser(storage)
	}()

	switch {
	case strings.HasPrefix(contentType, "application/json"):
		body, err := storage.Bytes()
		if err != nil {
			return service.TaskErrorWrapperLocal(err, "read_request_body_failed", http.StatusBadRequest)
		}
		if strings.TrimSpace(string(body)) != "" {
			var payload map[string]any
			if err := common.Unmarshal(body, &payload); err != nil {
				return service.TaskErrorWrapperLocal(err, "invalid_json", http.StatusBadRequest)
			}
			if raw, ok := payload["model"].(string); ok && strings.TrimSpace(raw) != "" {
				modelName = strings.TrimSpace(raw)
			}
			if raw, ok := payload["action"].(string); ok && strings.TrimSpace(raw) != "" {
				action = strings.TrimSpace(raw)
			}
		}
	case strings.Contains(contentType, gin.MIMEMultipartPOSTForm):
		form, err := common.ParseMultipartFormReusable(c)
		if err != nil {
			return service.TaskErrorWrapperLocal(err, "invalid_multipart_form", http.StatusBadRequest)
		}
		if vals := form.Value["model"]; len(vals) > 0 && strings.TrimSpace(vals[0]) != "" {
			modelName = strings.TrimSpace(vals[0])
		}
		if vals := form.Value["action"]; len(vals) > 0 && strings.TrimSpace(vals[0]) != "" {
			action = strings.TrimSpace(vals[0])
		}
	}

	if modelName == "" && action != "" {
		platform := constant.TaskPlatform(c.GetString("platform"))
		modelName = service.CoverTaskActionToModelName(platform, action)
	}
	if modelName != "" {
		info.OriginModelName = modelName
	}
	if action != "" {
		info.Action = action
	}

	info.PricingConfigSnapshotLoaded = true
	info.PricingConfigSnapshot = ratio_setting.GetPricingConfig(info.OriginModelName)
	info.RequireMatrixUsageBilling = info.PricingConfigSnapshot != nil &&
		info.PricingConfigSnapshot.BillingUnit == types.BillingUnitPer1MTokens
	return nil
}

func buildTaskBillingContext(info *relaycommon.RelayInfo) *model.TaskBillingContext {
	bc := &model.TaskBillingContext{
		ModelPrice:      info.PriceData.ModelPrice,
		GroupRatio:      info.PriceData.GroupRatioInfo.GroupRatio,
		ModelRatio:      info.PriceData.ModelRatio,
		OtherRatios:     info.PriceData.OtherRatios,
		OriginModelName: info.OriginModelName,
		PerCallBilling:  common.StringsContains(constant.TaskPricePatches, info.OriginModelName),
	}
	if decision := info.PricingDecisionFrozen; decision != nil && decision.BillingMode == types.BillingModeMatrix {
		bc.BillingMode = decision.BillingMode
		bc.PricingSnapshot = types.CloneMapAny(decision.Snapshot)
		bc.BillingUnit = decision.BillingUnit
		bc.TokenUnitPriceUSD = decision.TokenUnitPriceUSD
		bc.PerCallBilling = decision.PerCallBilling
		if decision.BillingUnit == types.BillingUnitPer1MTokens {
			bc.ModelPrice = decision.TokenUnitPriceUSD
		} else {
			bc.ModelPrice = decision.PriceUSD
		}
		bc.GroupRatio = decision.GroupRatioInfo.GroupRatio
		bc.OtherRatios = types.CloneRatios(decision.OtherRatios)
	}
	return bc
}

func RelayTask(c *gin.Context) {
	relayInfo, err := relaycommon.GenRelayInfo(c, types.RelayFormatTask, nil, nil)
	if err != nil {
		c.JSON(http.StatusInternalServerError, &dto.TaskError{
			Code:       "gen_relay_info_failed",
			Message:    err.Error(),
			StatusCode: http.StatusInternalServerError,
		})
		return
	}
	relayTaskWithInfo(c, relayInfo)
}

func relayTaskWithInfo(c *gin.Context, relayInfo *relaycommon.RelayInfo) {
	if taskErr := relay.ResolveOriginTask(c, relayInfo); taskErr != nil {
		respondTaskError(c, taskErr)
		return
	}
	if taskErr := preloadTaskPricingConfig(c, relayInfo); taskErr != nil {
		respondTaskError(c, taskErr)
		return
	}

	var result *relay.TaskSubmitResult
	var taskErr *dto.TaskError
	defer func() {
		if taskErr != nil && relayInfo.Billing != nil {
			relayInfo.Billing.Refund(c)
		}
	}()

	retryParam := &service.RetryParam{
		Ctx:                 c,
		TokenGroup:          relayInfo.TokenGroup,
		ModelName:           relayInfo.OriginModelName,
		Retry:               common.GetPointer(0),
		RequiredChannelType: requiredTaskChannelTypeForRequest(c),
	}

	for ; retryParam.GetRetry() <= common.RetryTimes; retryParam.IncreaseRetry() {
		var channel *model.Channel
		var selectedGroup string

		if lockedCh, ok := relayInfo.LockedChannel.(*model.Channel); ok && lockedCh != nil {
			channel = lockedCh
			selectedGroup = concreteDoubaoVideoBindingGroup(c, relayInfo.TokenGroup)
			if setupErr := middleware.SetupContextForSelectedChannel(c, channel, relayInfo.OriginModelName); setupErr != nil {
				taskErr = service.TaskErrorWrapperLocal(setupErr.Err, "setup_locked_channel_failed", http.StatusInternalServerError)
				break
			}
		} else {
			var channelErr *types.NewAPIError
			channel, selectedGroup, channelErr = getChannel(c, relayInfo, retryParam)
			if channelErr != nil {
				logger.LogError(c, channelErr.Error())
				taskErr = service.TaskErrorWrapperLocal(channelErr.Err, "get_channel_failed", http.StatusInternalServerError)
				break
			}
		}

		if bindErr := persistDoubaoVideoTaskBindingIfNeeded(c.GetInt("id"), selectedGroup, channel); bindErr != nil {
			taskErr = service.TaskErrorWrapperLocal(bindErr, "bind_doubao_video_channel_failed", http.StatusServiceUnavailable)
			break
		}

		addUsedChannel(c, channel.Id)
		bodyStorage, bodyErr := common.GetBodyStorage(c)
		if bodyErr != nil {
			if common.IsRequestBodyTooLargeError(bodyErr) || errors.Is(bodyErr, common.ErrRequestBodyTooLarge) {
				taskErr = service.TaskErrorWrapperLocal(bodyErr, "read_request_body_failed", http.StatusRequestEntityTooLarge)
			} else {
				taskErr = service.TaskErrorWrapperLocal(bodyErr, "read_request_body_failed", http.StatusBadRequest)
			}
			break
		}
		c.Request.Body = io.NopCloser(bodyStorage)

		result, taskErr = relay.RelayTaskSubmit(c, relayInfo)
		if taskErr == nil {
			break
		}

		if !taskErr.LocalError {
			processChannelError(c,
				*types.NewChannelError(channel.Id, channel.Type, channel.Name, channel.ChannelInfo.IsMultiKey,
					common.GetContextKeyString(c, constant.ContextKeyChannelKey), channel.GetAutoBan()),
				types.NewOpenAIError(taskErr.Error, types.ErrorCodeBadResponseStatusCode, taskErr.StatusCode))
		}

		if !shouldRetryTaskRelay(c, channel.Id, taskErr, common.RetryTimes-retryParam.GetRetry()) {
			break
		}
	}

	useChannel := c.GetStringSlice("use_channel")
	if len(useChannel) > 1 {
		retryLogStr := fmt.Sprintf("重试：%s", strings.Trim(strings.Join(strings.Fields(fmt.Sprint(useChannel)), "->"), "[]"))
		logger.LogInfo(c, retryLogStr)
	}

	// ── 成功：结算 + 日志 + 插入任务 ──
	if taskErr == nil {
		if settleErr := service.SettleBilling(c, relayInfo, result.Quota); settleErr != nil {
			common.SysError("settle task billing error: " + settleErr.Error())
		}
		service.LogTaskConsumption(c, relayInfo)

		task := model.InitTask(result.Platform, relayInfo)
		task.PrivateData.UpstreamTaskID = result.UpstreamTaskID
		task.PrivateData.BillingSource = relayInfo.BillingSource
		task.PrivateData.SubscriptionId = relayInfo.SubscriptionId
		task.PrivateData.TokenId = relayInfo.TokenId
		task.PrivateData.BillingContext = buildTaskBillingContext(relayInfo)
		task.PrivateData.UpstreamRequest = relay.BuildUpstreamRequestSnapshotForTask(result.UpstreamReqJSON)
		task.Quota = result.Quota
		task.Data = result.TaskData
		task.Action = relayInfo.Action
		if insertErr := task.Insert(); insertErr != nil {
			common.SysError("insert task error: " + insertErr.Error())
		}
	}

	if taskErr != nil {
		respondTaskError(c, taskErr)
	}
}

// respondTaskError 统一输出 Task 错误响应（含 429 限流提示改写）
func respondTaskError(c *gin.Context, taskErr *dto.TaskError) {
	if taskErr.StatusCode == http.StatusTooManyRequests {
		taskErr.Message = "当前分组上游负载已饱和，请稍后再试"
	}
	c.JSON(taskErr.StatusCode, taskErr)
}

func shouldRetryTaskRelay(c *gin.Context, channelId int, taskErr *dto.TaskError, retryTimes int) bool {
	if taskErr == nil {
		return false
	}
	if service.ShouldSkipRetryAfterChannelAffinityFailure(c) {
		return false
	}
	if retryTimes <= 0 {
		return false
	}
	if taskErr.StatusCode == http.StatusTooManyRequests {
		return true
	}
	if taskErr.StatusCode == 307 {
		return true
	}
	if taskErr.StatusCode/100 == 5 {
		// 超时不重试
		if operation_setting.IsAlwaysSkipRetryStatusCode(taskErr.StatusCode) {
			return false
		}
		return true
	}
	if taskErr.StatusCode == http.StatusBadRequest {
		return false
	}
	if taskErr.StatusCode == 408 {
		// azure处理超时不重试
		return false
	}
	if taskErr.LocalError {
		return false
	}
	if taskErr.StatusCode/100 == 2 {
		return false
	}
	return true
}