new-api-custom/relay/responses_handler.go

package relay

import (
	"encoding/json"
	"fmt"
	"io"
	"net/http"
	"strings"

	"github.com/QuantumNous/new-api/common"
	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/relaykit/dto"
	"github.com/QuantumNous/new-api/relaykit/types"
	"github.com/QuantumNous/new-api/service"

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

// Independent cursors retain the payload metadata used by net/http redirects
// and transport retries without closing the shared request storage.
type responsesAttemptBody struct {
	common.ReplayableBody
	reader io.Reader
}

func (b responsesAttemptBody) Read(p []byte) (int, error) { return b.reader.Read(p) }

func shouldEnableAstraResponseGuard(path string, info *relaycommon.RelayInfo) bool {
	return path == "/v1/responses" &&
		info.OriginModelName == "gpt-6-astra" &&
		info.AstraAutoSwitch && !info.AstraDetectionDisabled &&
		info.ResponseGuard == nil
}

func ResponsesHelper(c *gin.Context, info *relaycommon.RelayInfo) (newAPIError *types.NewAPIError) {
	var responsesReq *dto.OpenAIResponsesRequest
	switch req := info.Request.(type) {
	case *dto.OpenAIResponsesRequest:
		responsesReq = req
	case *dto.OpenAIResponsesCompactionRequest:
		// Only fields documented for POST /v1/responses/compact are forwarded:
		// model, input, instructions, previous_response_id, prompt_cache_key,
		// prompt_cache_options, prompt_cache_retention, service_tier.
		// Undocumented Codex-parity fields (tools, reasoning, text) are parsed
		// for client compatibility but intentionally not sent upstream.
		responsesReq = &dto.OpenAIResponsesRequest{
			Model:                req.Model,
			Input:                req.Input,
			Instructions:         req.Instructions,
			PreviousResponseID:   req.PreviousResponseID,
			ParallelToolCalls:    req.ParallelToolCalls,
			ServiceTier:          req.ServiceTier,
			PromptCacheKey:       req.PromptCacheKey,
			PromptCacheOptions:   req.PromptCacheOptions,
			PromptCacheRetention: req.PromptCacheRetention,
		}
	default:
		return types.NewErrorWithStatusCode(
			fmt.Errorf("invalid request type, expected dto.OpenAIResponsesRequest or dto.OpenAIResponsesCompactionRequest, got %T", info.Request),
			types.ErrorCodeInvalidRequest,
			http.StatusBadRequest,
			types.ErrOptionWithSkipRetry(),
		)
	}

	adaptor, requestBody, closer, apiErr := PrepareResponsesRequest(c, info, responsesReq)
	if apiErr != nil {
		return apiErr
	}
	defer closer.Close()
	// Missing-usage estimation follows the actual prepared payload, including
	// parameter overrides, rather than the client's pre-transformation request.
	if info.DurableAttemptBilling {
		if cursor, err := requestBody.NewReader(); err == nil {
			var sent dto.OpenAIResponsesRequest
			decodeErr := json.NewDecoder(cursor).Decode(&sent)
			_ = cursor.Close()
			if decodeErr == nil {
				if tokens, err := service.CountRequestToken(c, sent.GetTokenCountMeta(), info); err == nil {
					info.SetEstimatePromptTokens(tokens)
				}
			}
		}
	}
	if shouldEnableAstraResponseGuard(c.Request.URL.Path, info) {
		info.ResponseGuard = &relaycommon.ResponseModelGuard{Deadline: info.GuardDeadline}
	}

	guardEnabled := info.ResponseGuard != nil && info.AstraAutoSwitch

	body, err := requestBody.NewReader()
	if err != nil {
		return types.NewOpenAIError(err, types.ErrorCodeReadRequestBodyFailed, http.StatusInternalServerError)
	}
	resp, requestErr := adaptor.DoRequest(c, info, responsesAttemptBody{ReplayableBody: requestBody, reader: body})
	_ = body.Close()
	if requestErr != nil {
		apiErr := types.NewError(requestErr, types.ErrorCodeDoRequestFailed)
		guard := info.ResponseGuard
		if guard == nil || apiErr.GetErrorCode() != relaycommon.ResponseGuardErrorCode {
			return apiErr
		}
		attempt := guard.Last()
		if attempt != nil {
			service.RequestPolicy(c).AddEvent(service.PolicyEvent{ChannelID: info.ChannelId, ErrorCode: string(relaycommon.ResponseGuardErrorCode), ErrorSource: "model_guard", Health: "unchanged", Decision: service.PolicyDecision{Action: "reject", Reason: attempt.Reason, Source: "model_guard"}})
		}
		return guard.APIError()
	}
	var httpResp *http.Response
	if resp != nil {
		httpResp = resp.(*http.Response)
	}
	if guardEnabled {
		defer func() {
			if attempt := info.ResponseGuard.Last(); attempt != nil {
				attempt.Finish(c.Writer.Written())
			}
		}()
	}

	statusCodeMappingStr := c.GetString("status_code_mapping")

	if httpResp != nil {
		if httpResp.StatusCode != http.StatusOK {
			newAPIError = service.RelayErrorHandler(c.Request.Context(), httpResp, false)
			// reset status code 重置状态码
			service.ResetStatusCode(newAPIError, statusCodeMappingStr)
			return newAPIError
		}
	}

	usage, newAPIError := adaptor.DoResponse(c, httpResp, info)
	if guardEnabled {
		if attempt := info.ResponseGuard.Last(); attempt != nil {
			attempt.Finish(c.Writer.Written())
			if guardErr := attempt.APIError(); guardErr != nil {
				if info.StreamStatus != nil {
					info.StreamStatus.MarkFailed(string(relaycommon.ResponseGuardErrorCode), "api_error", http.StatusConflict)
				}
				if c.Writer.Written() && c.Request.Context().Err() == nil {
					message := common.MessageWithRequestId(attempt.PublicErrorMessage(), c.GetString(common.RequestIdKey))
					c.SSEvent("response.failed", gin.H{"type": "response.failed", "sequence_number": attempt.Sequence + 1, "response": gin.H{"id": attempt.ResponseID, "object": "response", "status": "failed", "error": gin.H{"code": relaycommon.ResponseGuardErrorCode, "message": message}}})
					c.Writer.Flush()
					c.Set("response_guard_error_sent", true)
				}
				return guardErr
			}
		}
	}
	if newAPIError != nil {
		// reset status code 重置状态码
		service.ResetStatusCode(newAPIError, statusCodeMappingStr)
		return newAPIError
	}

	usageDto := usage.(*dto.Usage)
	if info.RelayMode == relayconstant.RelayModeResponsesCompact {
		originModelName := info.OriginModelName
		originPriceData := info.PriceData

		_, err := helper.ModelPriceHelper(c, info, info.GetEstimatePromptTokens(), &types.TokenCountMeta{})
		if err != nil {
			info.OriginModelName = originModelName
			info.PriceData = originPriceData
			return types.NewError(err, types.ErrorCodeModelPriceError, types.ErrOptionWithSkipRetry(), types.ErrOptionWithStatusCode(http.StatusBadRequest))
		}
		service.PostTextConsumeQuota(c, info, usageDto, nil)

		info.OriginModelName = originModelName
		info.PriceData = originPriceData
		return nil
	}

	ConsumeResponsesQuota(c, info, usageDto)
	return nil
}

// ConsumeResponsesQuota applies the same settlement dispatch to HTTP and
// WebSocket Responses usage. Compact requests keep their separate repricing.
func ConsumeResponsesQuota(c *gin.Context, info *relaycommon.RelayInfo, usage *dto.Usage) {
	if strings.HasPrefix(info.OriginModelName, "gpt-4o-audio") {
		service.PostAudioConsumeQuota(c, info, usage, nil)
		return
	}
	service.PostTextConsumeQuota(c, info, usage, nil)
}