xinghuoapi/backend/internal/service/antigravity_gateway_service.go

package service

import (
	"bufio"
	"bytes"
	"context"
	"encoding/json"
	"errors"
	"fmt"
	"io"
	"log"
	"log/slog"
	mathrand "math/rand"
	"net"
	"net/http"
	"os"
	"strconv"
	"strings"
	"sync/atomic"
	"time"

	"github.com/Wei-Shaw/sub2api/internal/pkg/antigravity"
	"github.com/gin-gonic/gin"
	"github.com/google/uuid"
)

const (
	antigravityStickySessionTTL = time.Hour
	antigravityMaxRetries       = 3
	antigravityRetryBaseDelay   = 1 * time.Second
	antigravityRetryMaxDelay    = 16 * time.Second

	// 限流相关常量
	// antigravityRateLimitThreshold 限流等待/切换阈值
	// - 智能重试：retryDelay < 此阈值时等待后重试，>= 此阈值时直接限流模型
	// - 预检查：剩余限流时间 < 此阈值时等待，>= 此阈值时切换账号
	antigravityRateLimitThreshold       = 7 * time.Second
	antigravitySmartRetryMinWait        = 1 * time.Second  // 智能重试最小等待时间
	antigravitySmartRetryMaxAttempts    = 1                // 智能重试最大次数（仅重试 1 次，防止重复限流/长期等待）
	antigravityDefaultRateLimitDuration = 30 * time.Second // 默认限流时间（无 retryDelay 时使用）

	// Google RPC 状态和类型常量
	googleRPCStatusResourceExhausted      = "RESOURCE_EXHAUSTED"
	googleRPCStatusUnavailable            = "UNAVAILABLE"
	googleRPCTypeRetryInfo                = "type.googleapis.com/google.rpc.RetryInfo"
	googleRPCTypeErrorInfo                = "type.googleapis.com/google.rpc.ErrorInfo"
	googleRPCReasonModelCapacityExhausted = "MODEL_CAPACITY_EXHAUSTED"
	googleRPCReasonRateLimitExceeded      = "RATE_LIMIT_EXCEEDED"
)

// antigravityPassthroughErrorMessages 透传给客户端的错误消息白名单（小写）
// 匹配时使用 strings.Contains，无需完全匹配
var antigravityPassthroughErrorMessages = []string{
	"prompt is too long",
}

const (
	antigravityBillingModelEnv    = "GATEWAY_ANTIGRAVITY_BILL_WITH_MAPPED_MODEL"
	antigravityFallbackSecondsEnv = "GATEWAY_ANTIGRAVITY_FALLBACK_COOLDOWN_SECONDS"
)

// AntigravityAccountSwitchError 账号切换信号
// 当账号限流时间超过阈值时，通知上层切换账号
type AntigravityAccountSwitchError struct {
	OriginalAccountID int64
	RateLimitedModel  string
	IsStickySession   bool // 是否为粘性会话切换（决定是否缓存计费）
}

func (e *AntigravityAccountSwitchError) Error() string {
	return fmt.Sprintf("account %d model %s rate limited, need switch",
		e.OriginalAccountID, e.RateLimitedModel)
}

// IsAntigravityAccountSwitchError 检查错误是否为账号切换信号
func IsAntigravityAccountSwitchError(err error) (*AntigravityAccountSwitchError, bool) {
	var switchErr *AntigravityAccountSwitchError
	if errors.As(err, &switchErr) {
		return switchErr, true
	}
	return nil, false
}

// PromptTooLongError 表示上游明确返回 prompt too long
type PromptTooLongError struct {
	StatusCode int
	RequestID  string
	Body       []byte
}

func (e *PromptTooLongError) Error() string {
	return fmt.Sprintf("prompt too long: status=%d", e.StatusCode)
}

// antigravityRetryLoopParams 重试循环的参数
type antigravityRetryLoopParams struct {
	ctx             context.Context
	prefix          string
	account         *Account
	proxyURL        string
	accessToken     string
	action          string
	body            []byte
	c               *gin.Context
	httpUpstream    HTTPUpstream
	settingService  *SettingService
	accountRepo     AccountRepository // 用于智能重试的模型级别限流
	handleError     func(ctx context.Context, prefix string, account *Account, statusCode int, headers http.Header, body []byte, requestedModel string, groupID int64, sessionHash string, isStickySession bool) *handleModelRateLimitResult
	requestedModel  string // 用于限流检查的原始请求模型
	isStickySession bool   // 是否为粘性会话（用于账号切换时的缓存计费判断）
	groupID         int64  // 用于模型级限流时清除粘性会话
	sessionHash     string // 用于模型级限流时清除粘性会话
}

// antigravityRetryLoopResult 重试循环的结果
type antigravityRetryLoopResult struct {
	resp *http.Response
}

// smartRetryAction 智能重试的处理结果
type smartRetryAction int

const (
	smartRetryActionContinue      smartRetryAction = iota // 继续默认重试逻辑
	smartRetryActionBreakWithResp                         // 结束循环并返回 resp
	smartRetryActionContinueURL                           // 继续 URL fallback 循环
)

// smartRetryResult 智能重试的结果
type smartRetryResult struct {
	action      smartRetryAction
	resp        *http.Response
	err         error
	switchError *AntigravityAccountSwitchError // 模型限流时返回账号切换信号
}

// handleSmartRetry 处理 OAuth 账号的智能重试逻辑
// 将 429/503 限流处理逻辑抽取为独立函数，减少 antigravityRetryLoop 的复杂度
func (s *AntigravityGatewayService) handleSmartRetry(p antigravityRetryLoopParams, resp *http.Response, respBody []byte, baseURL string, urlIdx int, availableURLs []string) *smartRetryResult {
	// "Resource has been exhausted" 是 URL 级别限流，切换 URL（仅 429）
	if resp.StatusCode == http.StatusTooManyRequests && isURLLevelRateLimit(respBody) && urlIdx < len(availableURLs)-1 {
		log.Printf("%s URL fallback (429): %s -> %s", p.prefix, baseURL, availableURLs[urlIdx+1])
		return &smartRetryResult{action: smartRetryActionContinueURL}
	}

	// 判断是否触发智能重试
	shouldSmartRetry, shouldRateLimitModel, waitDuration, modelName := shouldTriggerAntigravitySmartRetry(p.account, respBody)

	// 情况1: retryDelay >= 阈值，限流模型并切换账号
	if shouldRateLimitModel {
		rateLimitDuration := waitDuration
		if rateLimitDuration <= 0 {
			rateLimitDuration = antigravityDefaultRateLimitDuration
		}
		log.Printf("%s status=%d oauth_long_delay model=%s account=%d upstream_retry_delay=%v body=%s (model rate limit, switch account)",
			p.prefix, resp.StatusCode, modelName, p.account.ID, rateLimitDuration, truncateForLog(respBody, 200))

		resetAt := time.Now().Add(rateLimitDuration)
		if !setModelRateLimitByModelName(p.ctx, p.accountRepo, p.account.ID, modelName, p.prefix, resp.StatusCode, resetAt, false) {
			p.handleError(p.ctx, p.prefix, p.account, resp.StatusCode, resp.Header, respBody, p.requestedModel, p.groupID, p.sessionHash, p.isStickySession)
			log.Printf("%s status=%d rate_limited account=%d (no model mapping)", p.prefix, resp.StatusCode, p.account.ID)
		} else {
			s.updateAccountModelRateLimitInCache(p.ctx, p.account, modelName, resetAt)
		}

		// 返回账号切换信号，让上层切换账号重试
		return &smartRetryResult{
			action: smartRetryActionBreakWithResp,
			switchError: &AntigravityAccountSwitchError{
				OriginalAccountID: p.account.ID,
				RateLimitedModel:  modelName,
				IsStickySession:   p.isStickySession,
			},
		}
	}

	// 情况2: retryDelay < 阈值，智能重试（最多 antigravitySmartRetryMaxAttempts 次）
	if shouldSmartRetry {
		var lastRetryResp *http.Response
		var lastRetryBody []byte

		for attempt := 1; attempt <= antigravitySmartRetryMaxAttempts; attempt++ {
			log.Printf("%s status=%d oauth_smart_retry attempt=%d/%d delay=%v model=%s account=%d",
				p.prefix, resp.StatusCode, attempt, antigravitySmartRetryMaxAttempts, waitDuration, modelName, p.account.ID)

			select {
			case <-p.ctx.Done():
				log.Printf("%s status=context_canceled_during_smart_retry", p.prefix)
				return &smartRetryResult{action: smartRetryActionBreakWithResp, err: p.ctx.Err()}
			case <-time.After(waitDuration):
			}

			// 智能重试：创建新请求
			retryReq, err := antigravity.NewAPIRequestWithURL(p.ctx, baseURL, p.action, p.accessToken, p.body)
			if err != nil {
				log.Printf("%s status=smart_retry_request_build_failed error=%v", p.prefix, err)
				p.handleError(p.ctx, p.prefix, p.account, resp.StatusCode, resp.Header, respBody, p.requestedModel, p.groupID, p.sessionHash, p.isStickySession)
				return &smartRetryResult{
					action: smartRetryActionBreakWithResp,
					resp: &http.Response{
						StatusCode: resp.StatusCode,
						Header:     resp.Header.Clone(),
						Body:       io.NopCloser(bytes.NewReader(respBody)),
					},
				}
			}

			retryResp, retryErr := p.httpUpstream.Do(retryReq, p.proxyURL, p.account.ID, p.account.Concurrency)
			if retryErr == nil && retryResp != nil && retryResp.StatusCode != http.StatusTooManyRequests && retryResp.StatusCode != http.StatusServiceUnavailable {
				log.Printf("%s status=%d smart_retry_success attempt=%d/%d", p.prefix, retryResp.StatusCode, attempt, antigravitySmartRetryMaxAttempts)
				return &smartRetryResult{action: smartRetryActionBreakWithResp, resp: retryResp}
			}

			// 网络错误时，继续重试
			if retryErr != nil || retryResp == nil {
				log.Printf("%s status=smart_retry_network_error attempt=%d/%d error=%v", p.prefix, attempt, antigravitySmartRetryMaxAttempts, retryErr)
				continue
			}

			// 重试失败，关闭之前的响应
			if lastRetryResp != nil {
				_ = lastRetryResp.Body.Close()
			}
			lastRetryResp = retryResp
			if retryResp != nil {
				lastRetryBody, _ = io.ReadAll(io.LimitReader(retryResp.Body, 2<<20))
				_ = retryResp.Body.Close()
			}

			// 解析新的重试信息，用于下次重试的等待时间
			if attempt < antigravitySmartRetryMaxAttempts && lastRetryBody != nil {
				newShouldRetry, _, newWaitDuration, _ := shouldTriggerAntigravitySmartRetry(p.account, lastRetryBody)
				if newShouldRetry && newWaitDuration > 0 {
					waitDuration = newWaitDuration
				}
			}
		}

		// 所有重试都失败，限流当前模型并切换账号
		rateLimitDuration := waitDuration
		if rateLimitDuration <= 0 {
			rateLimitDuration = antigravityDefaultRateLimitDuration
		}
		retryBody := lastRetryBody
		if retryBody == nil {
			retryBody = respBody
		}
		log.Printf("%s status=%d smart_retry_exhausted attempts=%d model=%s account=%d upstream_retry_delay=%v body=%s (switch account)",
			p.prefix, resp.StatusCode, antigravitySmartRetryMaxAttempts, modelName, p.account.ID, rateLimitDuration, truncateForLog(retryBody, 200))

		resetAt := time.Now().Add(rateLimitDuration)
		if p.accountRepo != nil && modelName != "" {
			if err := p.accountRepo.SetModelRateLimit(p.ctx, p.account.ID, modelName, resetAt); err != nil {
				log.Printf("%s status=%d model_rate_limit_failed model=%s error=%v", p.prefix, resp.StatusCode, modelName, err)
			} else {
				log.Printf("%s status=%d model_rate_limited_after_smart_retry model=%s account=%d reset_in=%v",
					p.prefix, resp.StatusCode, modelName, p.account.ID, rateLimitDuration)
				s.updateAccountModelRateLimitInCache(p.ctx, p.account, modelName, resetAt)
			}
		}

		// 清除粘性会话绑定，避免下次请求仍命中限流账号
		if s.cache != nil && p.sessionHash != "" {
			_ = s.cache.DeleteSessionAccountID(p.ctx, p.groupID, p.sessionHash)
		}

		// 返回账号切换信号，让上层切换账号重试
		return &smartRetryResult{
			action: smartRetryActionBreakWithResp,
			switchError: &AntigravityAccountSwitchError{
				OriginalAccountID: p.account.ID,
				RateLimitedModel:  modelName,
				IsStickySession:   p.isStickySession,
			},
		}
	}

	// 未触发智能重试，继续默认重试逻辑
	return &smartRetryResult{action: smartRetryActionContinue}
}

// antigravityRetryLoop 执行带 URL fallback 的重试循环
func (s *AntigravityGatewayService) antigravityRetryLoop(p antigravityRetryLoopParams) (*antigravityRetryLoopResult, error) {
	// 预检查：如果账号已限流，直接返回切换信号
	if p.requestedModel != "" {
		if remaining := p.account.GetRateLimitRemainingTimeWithContext(p.ctx, p.requestedModel); remaining > 0 {
			log.Printf("%s pre_check: rate_limit_switch remaining=%v model=%s account=%d",
				p.prefix, remaining.Truncate(time.Millisecond), p.requestedModel, p.account.ID)
			return nil, &AntigravityAccountSwitchError{
				OriginalAccountID: p.account.ID,
				RateLimitedModel:  p.requestedModel,
				IsStickySession:   p.isStickySession,
			}
		}
	}

	availableURLs := antigravity.DefaultURLAvailability.GetAvailableURLs()
	if len(availableURLs) == 0 {
		availableURLs = antigravity.BaseURLs
	}

	var resp *http.Response
	var usedBaseURL string
	logBody := p.settingService != nil && p.settingService.cfg != nil && p.settingService.cfg.Gateway.LogUpstreamErrorBody
	maxBytes := 2048
	if p.settingService != nil && p.settingService.cfg != nil && p.settingService.cfg.Gateway.LogUpstreamErrorBodyMaxBytes > 0 {
		maxBytes = p.settingService.cfg.Gateway.LogUpstreamErrorBodyMaxBytes
	}
	getUpstreamDetail := func(body []byte) string {
		if !logBody {
			return ""
		}
		return truncateString(string(body), maxBytes)
	}

urlFallbackLoop:
	for urlIdx, baseURL := range availableURLs {
		usedBaseURL = baseURL
		for attempt := 1; attempt <= antigravityMaxRetries; attempt++ {
			select {
			case <-p.ctx.Done():
				log.Printf("%s status=context_canceled error=%v", p.prefix, p.ctx.Err())
				return nil, p.ctx.Err()
			default:
			}

			upstreamReq, err := antigravity.NewAPIRequestWithURL(p.ctx, baseURL, p.action, p.accessToken, p.body)
			if err != nil {
				return nil, err
			}

			// Capture upstream request body for ops retry of this attempt.
			if p.c != nil && len(p.body) > 0 {
				p.c.Set(OpsUpstreamRequestBodyKey, string(p.body))
			}

			resp, err = p.httpUpstream.Do(upstreamReq, p.proxyURL, p.account.ID, p.account.Concurrency)
			if err == nil && resp == nil {
				err = errors.New("upstream returned nil response")
			}
			if err != nil {
				safeErr := sanitizeUpstreamErrorMessage(err.Error())
				appendOpsUpstreamError(p.c, OpsUpstreamErrorEvent{
					Platform:           p.account.Platform,
					AccountID:          p.account.ID,
					AccountName:        p.account.Name,
					UpstreamStatusCode: 0,
					Kind:               "request_error",
					Message:            safeErr,
				})
				if shouldAntigravityFallbackToNextURL(err, 0) && urlIdx < len(availableURLs)-1 {
					log.Printf("%s URL fallback (connection error): %s -> %s", p.prefix, baseURL, availableURLs[urlIdx+1])
					continue urlFallbackLoop
				}
				if attempt < antigravityMaxRetries {
					log.Printf("%s status=request_failed retry=%d/%d error=%v", p.prefix, attempt, antigravityMaxRetries, err)
					if !sleepAntigravityBackoffWithContext(p.ctx, attempt) {
						log.Printf("%s status=context_canceled_during_backoff", p.prefix)
						return nil, p.ctx.Err()
					}
					continue
				}
				log.Printf("%s status=request_failed retries_exhausted error=%v", p.prefix, err)
				setOpsUpstreamError(p.c, 0, safeErr, "")
				return nil, fmt.Errorf("upstream request failed after retries: %w", err)
			}

			// 统一处理错误响应
			if resp.StatusCode >= 400 {
				respBody, _ := io.ReadAll(io.LimitReader(resp.Body, 2<<20))
				_ = resp.Body.Close()

				// ★ 统一入口：自定义错误码 + 临时不可调度
				if handled, outStatus, policyErr := s.applyErrorPolicy(p, resp.StatusCode, resp.Header, respBody); handled {
					if policyErr != nil {
						return nil, policyErr
					}
					resp = &http.Response{
						StatusCode: outStatus,
						Header:     resp.Header.Clone(),
						Body:       io.NopCloser(bytes.NewReader(respBody)),
					}
					break urlFallbackLoop
				}

				// 429/503 限流处理：区分 URL 级别限流、智能重试和账户配额限流
				if resp.StatusCode == http.StatusTooManyRequests || resp.StatusCode == http.StatusServiceUnavailable {
					// 尝试智能重试处理（OAuth 账号专用）
					smartResult := s.handleSmartRetry(p, resp, respBody, baseURL, urlIdx, availableURLs)
					switch smartResult.action {
					case smartRetryActionContinueURL:
						continue urlFallbackLoop
					case smartRetryActionBreakWithResp:
						if smartResult.err != nil {
							return nil, smartResult.err
						}
						// 模型限流时返回切换账号信号
						if smartResult.switchError != nil {
							return nil, smartResult.switchError
						}
						resp = smartResult.resp
						break urlFallbackLoop
					}
					// smartRetryActionContinue: 继续默认重试逻辑

					// 账户/模型配额限流，重试 3 次（指数退避）- 默认逻辑（非 OAuth 账号或解析失败）
					if attempt < antigravityMaxRetries {
						upstreamMsg := strings.TrimSpace(extractAntigravityErrorMessage(respBody))
						upstreamMsg = sanitizeUpstreamErrorMessage(upstreamMsg)
						appendOpsUpstreamError(p.c, OpsUpstreamErrorEvent{
							Platform:           p.account.Platform,
							AccountID:          p.account.ID,
							AccountName:        p.account.Name,
							UpstreamStatusCode: resp.StatusCode,
							UpstreamRequestID:  resp.Header.Get("x-request-id"),
							Kind:               "retry",
							Message:            upstreamMsg,
							Detail:             getUpstreamDetail(respBody),
						})
						log.Printf("%s status=%d retry=%d/%d body=%s", p.prefix, resp.StatusCode, attempt, antigravityMaxRetries, truncateForLog(respBody, 200))
						if !sleepAntigravityBackoffWithContext(p.ctx, attempt) {
							log.Printf("%s status=context_canceled_during_backoff", p.prefix)
							return nil, p.ctx.Err()
						}
						continue
					}

					// 重试用尽，标记账户限流
					p.handleError(p.ctx, p.prefix, p.account, resp.StatusCode, resp.Header, respBody, p.requestedModel, p.groupID, p.sessionHash, p.isStickySession)
					log.Printf("%s status=%d rate_limited base_url=%s body=%s", p.prefix, resp.StatusCode, baseURL, truncateForLog(respBody, 200))
					resp = &http.Response{
						StatusCode: resp.StatusCode,
						Header:     resp.Header.Clone(),
						Body:       io.NopCloser(bytes.NewReader(respBody)),
					}
					break urlFallbackLoop
				}

				// 其他可重试错误（500/502/504/529，不包括 429 和 503）
				if shouldRetryAntigravityError(resp.StatusCode) {
					if attempt < antigravityMaxRetries {
						upstreamMsg := strings.TrimSpace(extractAntigravityErrorMessage(respBody))
						upstreamMsg = sanitizeUpstreamErrorMessage(upstreamMsg)
						appendOpsUpstreamError(p.c, OpsUpstreamErrorEvent{
							Platform:           p.account.Platform,
							AccountID:          p.account.ID,
							AccountName:        p.account.Name,
							UpstreamStatusCode: resp.StatusCode,
							UpstreamRequestID:  resp.Header.Get("x-request-id"),
							Kind:               "retry",
							Message:            upstreamMsg,
							Detail:             getUpstreamDetail(respBody),
						})
						log.Printf("%s status=%d retry=%d/%d body=%s", p.prefix, resp.StatusCode, attempt, antigravityMaxRetries, truncateForLog(respBody, 500))
						if !sleepAntigravityBackoffWithContext(p.ctx, attempt) {
							log.Printf("%s status=context_canceled_during_backoff", p.prefix)
							return nil, p.ctx.Err()
						}
						continue
					}
				}

				// 其他 4xx 错误或重试用尽，直接返回
				resp = &http.Response{
					StatusCode: resp.StatusCode,
					Header:     resp.Header.Clone(),
					Body:       io.NopCloser(bytes.NewReader(respBody)),
				}
				break urlFallbackLoop
			}

			// 成功响应（< 400）
			break urlFallbackLoop
		}
	}

	if resp != nil && resp.StatusCode < 400 && usedBaseURL != "" {
		antigravity.DefaultURLAvailability.MarkSuccess(usedBaseURL)
	}

	return &antigravityRetryLoopResult{resp: resp}, nil
}

// shouldRetryAntigravityError 判断是否应该重试
func shouldRetryAntigravityError(statusCode int) bool {
	switch statusCode {
	case 429, 500, 502, 503, 504, 529:
		return true
	default:
		return false
	}
}

// isURLLevelRateLimit 判断是否为 URL 级别的限流（应切换 URL 重试）
// "Resource has been exhausted" 是 URL/节点级别限流，切换 URL 可能成功
// "exhausted your capacity on this model" 是账户/模型配额限流，切换 URL 无效
func isURLLevelRateLimit(body []byte) bool {
	// 快速检查：包含 "Resource has been exhausted" 且不包含 "capacity on this model"
	bodyStr := string(body)
	return strings.Contains(bodyStr, "Resource has been exhausted") &&
		!strings.Contains(bodyStr, "capacity on this model")
}

// isAntigravityConnectionError 判断是否为连接错误（网络超时、DNS 失败、连接拒绝）
func isAntigravityConnectionError(err error) bool {
	if err == nil {
		return false
	}

	// 检查超时错误
	var netErr net.Error
	if errors.As(err, &netErr) && netErr.Timeout() {
		return true
	}

	// 检查连接错误（DNS 失败、连接拒绝）
	var opErr *net.OpError
	return errors.As(err, &opErr)
}

// shouldAntigravityFallbackToNextURL 判断是否应切换到下一个 URL
// 仅连接错误和 HTTP 429 触发 URL 降级
func shouldAntigravityFallbackToNextURL(err error, statusCode int) bool {
	if isAntigravityConnectionError(err) {
		return true
	}
	return statusCode == http.StatusTooManyRequests
}

// getSessionID 从 gin.Context 获取 session_id（用于日志追踪）
func getSessionID(c *gin.Context) string {
	if c == nil {
		return ""
	}
	return c.GetHeader("session_id")
}

// logPrefix 生成统一的日志前缀
func logPrefix(sessionID, accountName string) string {
	if sessionID != "" {
		return fmt.Sprintf("[antigravity-Forward] session=%s account=%s", sessionID, accountName)
	}
	return fmt.Sprintf("[antigravity-Forward] account=%s", accountName)
}

// AntigravityGatewayService 处理 Antigravity 平台的 API 转发
type AntigravityGatewayService struct {
	accountRepo       AccountRepository
	tokenProvider     *AntigravityTokenProvider
	rateLimitService  *RateLimitService
	httpUpstream      HTTPUpstream
	settingService    *SettingService
	cache             GatewayCache // 用于模型级限流时清除粘性会话绑定
	schedulerSnapshot *SchedulerSnapshotService
}

func NewAntigravityGatewayService(
	accountRepo AccountRepository,
	cache GatewayCache,
	schedulerSnapshot *SchedulerSnapshotService,
	tokenProvider *AntigravityTokenProvider,
	rateLimitService *RateLimitService,
	httpUpstream HTTPUpstream,
	settingService *SettingService,
) *AntigravityGatewayService {
	return &AntigravityGatewayService{
		accountRepo:       accountRepo,
		tokenProvider:     tokenProvider,
		rateLimitService:  rateLimitService,
		httpUpstream:      httpUpstream,
		settingService:    settingService,
		cache:             cache,
		schedulerSnapshot: schedulerSnapshot,
	}
}

// GetTokenProvider 返回 token provider
func (s *AntigravityGatewayService) GetTokenProvider() *AntigravityTokenProvider {
	return s.tokenProvider
}

// getLogConfig 获取上游错误日志配置
// 返回是否记录日志体和最大字节数
func (s *AntigravityGatewayService) getLogConfig() (logBody bool, maxBytes int) {
	maxBytes = 2048 // 默认值
	if s.settingService == nil || s.settingService.cfg == nil {
		return false, maxBytes
	}
	cfg := s.settingService.cfg.Gateway
	if cfg.LogUpstreamErrorBodyMaxBytes > 0 {
		maxBytes = cfg.LogUpstreamErrorBodyMaxBytes
	}
	return cfg.LogUpstreamErrorBody, maxBytes
}

// getUpstreamErrorDetail 获取上游错误详情（用于日志记录）
func (s *AntigravityGatewayService) getUpstreamErrorDetail(body []byte) string {
	logBody, maxBytes := s.getLogConfig()
	if !logBody {
		return ""
	}
	return truncateString(string(body), maxBytes)
}

// checkErrorPolicy nil 安全的包装
func (s *AntigravityGatewayService) checkErrorPolicy(ctx context.Context, account *Account, statusCode int, body []byte) ErrorPolicyResult {
	if s.rateLimitService == nil {
		return ErrorPolicyNone
	}
	return s.rateLimitService.CheckErrorPolicy(ctx, account, statusCode, body)
}

// applyErrorPolicy 应用错误策略结果，返回是否应终止当前循环及应返回的状态码。
// ErrorPolicySkipped 时 outStatus 为 500（前端约定：未命中的错误返回 500）。
func (s *AntigravityGatewayService) applyErrorPolicy(p antigravityRetryLoopParams, statusCode int, headers http.Header, respBody []byte) (handled bool, outStatus int, retErr error) {
	switch s.checkErrorPolicy(p.ctx, p.account, statusCode, respBody) {
	case ErrorPolicySkipped:
		return true, http.StatusInternalServerError, nil
	case ErrorPolicyMatched:
		_ = p.handleError(p.ctx, p.prefix, p.account, statusCode, headers, respBody,
			p.requestedModel, p.groupID, p.sessionHash, p.isStickySession)
		return true, statusCode, nil
	case ErrorPolicyTempUnscheduled:
		slog.Info("temp_unschedulable_matched",
			"prefix", p.prefix, "status_code", statusCode, "account_id", p.account.ID)
		return true, statusCode, &AntigravityAccountSwitchError{OriginalAccountID: p.account.ID, IsStickySession: p.isStickySession}
	}
	return false, statusCode, nil
}

// mapAntigravityModel 获取映射后的模型名
// 完全依赖映射配置：账户映射（通配符）→ 默认映射兜底（DefaultAntigravityModelMapping）
// 注意：返回空字符串表示模型不被支持，调度时会过滤掉该账号
func mapAntigravityModel(account *Account, requestedModel string) string {
	if account == nil {
		return ""
	}

	// 获取映射表（未配置时自动使用 DefaultAntigravityModelMapping）
	mapping := account.GetModelMapping()
	if len(mapping) == 0 {
		return "" // 无映射配置（非 Antigravity 平台）
	}

	// 通过映射表查询（支持精确匹配 + 通配符）
	mapped := account.GetMappedModel(requestedModel)

	// 判断是否映射成功（mapped != requestedModel 说明找到了映射规则）
	if mapped != requestedModel {
		return mapped
	}

	// 如果 mapped == requestedModel，检查是否在映射表中配置（精确或通配符）
	// 这区分两种情况：
	// 1. 映射表中有 "model-a": "model-a"（显式透传）→ 返回 model-a
	// 2. 通配符匹配 "claude-*": "claude-sonnet-4-5" 恰好目标等于请求名 → 返回 model-a
	// 3. 映射表中没有 model-a 的配置 → 返回空（不支持）
	if account.IsModelSupported(requestedModel) {
		return requestedModel
	}

	// 未在映射表中配置的模型，返回空字符串（不支持）
	return ""
}

// getMappedModel 获取映射后的模型名
// 完全依赖映射配置：账户映射（通配符）→ 默认映射兜底
func (s *AntigravityGatewayService) getMappedModel(account *Account, requestedModel string) string {
	return mapAntigravityModel(account, requestedModel)
}

// applyThinkingModelSuffix 根据 thinking 配置调整模型名
// 当映射结果是 claude-sonnet-4-5 且请求开启了 thinking 时，改为 claude-sonnet-4-5-thinking
func applyThinkingModelSuffix(mappedModel string, thinkingEnabled bool) string {
	if !thinkingEnabled {
		return mappedModel
	}
	if mappedModel == "claude-sonnet-4-5" {
		return "claude-sonnet-4-5-thinking"
	}
	return mappedModel
}

// IsModelSupported 检查模型是否被支持
// 所有 claude- 和 gemini- 前缀的模型都能通过映射或透传支持
func (s *AntigravityGatewayService) IsModelSupported(requestedModel string) bool {
	return strings.HasPrefix(requestedModel, "claude-") ||
		strings.HasPrefix(requestedModel, "gemini-")
}

// TestConnectionResult 测试连接结果
type TestConnectionResult struct {
	Text        string // 响应文本
	MappedModel string // 实际使用的模型
}

// TestConnection 测试 Antigravity 账号连接（非流式，无重试、无计费）
// 支持 Claude 和 Gemini 两种协议，根据 modelID 前缀自动选择
func (s *AntigravityGatewayService) TestConnection(ctx context.Context, account *Account, modelID string) (*TestConnectionResult, error) {

	// 获取 token
	if s.tokenProvider == nil {
		return nil, errors.New("antigravity token provider not configured")
	}
	accessToken, err := s.tokenProvider.GetAccessToken(ctx, account)
	if err != nil {
		return nil, fmt.Errorf("获取 access_token 失败: %w", err)
	}

	// 获取 project_id（部分账户类型可能没有）
	projectID := strings.TrimSpace(account.GetCredential("project_id"))

	// 模型映射
	mappedModel := s.getMappedModel(account, modelID)
	if mappedModel == "" {
		return nil, fmt.Errorf("model %s not in whitelist", modelID)
	}

	// 构建请求体
	var requestBody []byte
	if strings.HasPrefix(modelID, "gemini-") {
		// Gemini 模型：直接使用 Gemini 格式
		requestBody, err = s.buildGeminiTestRequest(projectID, mappedModel)
	} else {
		// Claude 模型：使用协议转换
		requestBody, err = s.buildClaudeTestRequest(projectID, mappedModel)
	}
	if err != nil {
		return nil, fmt.Errorf("构建请求失败: %w", err)
	}

	// 代理 URL
	proxyURL := ""
	if account.ProxyID != nil && account.Proxy != nil {
		proxyURL = account.Proxy.URL()
	}

	// URL fallback 循环
	availableURLs := antigravity.DefaultURLAvailability.GetAvailableURLs()
	if len(availableURLs) == 0 {
		availableURLs = antigravity.BaseURLs // 所有 URL 都不可用时，重试所有
	}

	var lastErr error
	for urlIdx, baseURL := range availableURLs {
		// 构建 HTTP 请求（总是使用流式 endpoint，与官方客户端一致）
		req, err := antigravity.NewAPIRequestWithURL(ctx, baseURL, "streamGenerateContent", accessToken, requestBody)
		if err != nil {
			lastErr = err
			continue
		}

		// 调试日志：Test 请求信息
		log.Printf("[antigravity-Test] account=%s request_size=%d url=%s", account.Name, len(requestBody), req.URL.String())

		// 发送请求
		resp, err := s.httpUpstream.Do(req, proxyURL, account.ID, account.Concurrency)
		if err != nil {
			lastErr = fmt.Errorf("请求失败: %w", err)
			if shouldAntigravityFallbackToNextURL(err, 0) && urlIdx < len(availableURLs)-1 {
				log.Printf("[antigravity-Test] URL fallback: %s -> %s", baseURL, availableURLs[urlIdx+1])
				continue
			}
			return nil, lastErr
		}

		// 读取响应
		respBody, err := io.ReadAll(io.LimitReader(resp.Body, 2<<20))
		_ = resp.Body.Close() // 立即关闭，避免循环内 defer 导致的资源泄漏
		if err != nil {
			return nil, fmt.Errorf("读取响应失败: %w", err)
		}

		// 检查是否需要 URL 降级
		if shouldAntigravityFallbackToNextURL(nil, resp.StatusCode) && urlIdx < len(availableURLs)-1 {
			log.Printf("[antigravity-Test] URL fallback (HTTP %d): %s -> %s", resp.StatusCode, baseURL, availableURLs[urlIdx+1])
			continue
		}

		if resp.StatusCode >= 400 {
			return nil, fmt.Errorf("API 返回 %d: %s", resp.StatusCode, string(respBody))
		}

		// 解析流式响应，提取文本
		text := extractTextFromSSEResponse(respBody)

		// 标记成功的 URL，下次优先使用
		antigravity.DefaultURLAvailability.MarkSuccess(baseURL)
		return &TestConnectionResult{
			Text:        text,
			MappedModel: mappedModel,
		}, nil
	}

	return nil, lastErr
}

// buildGeminiTestRequest 构建 Gemini 格式测试请求
// 使用最小 token 消耗：输入 "." + maxOutputTokens: 1
func (s *AntigravityGatewayService) buildGeminiTestRequest(projectID, model string) ([]byte, error) {
	payload := map[string]any{
		"contents": []map[string]any{
			{
				"role": "user",
				"parts": []map[string]any{
					{"text": "."},
				},
			},
		},
		// Antigravity 上游要求必须包含身份提示词
		"systemInstruction": map[string]any{
			"parts": []map[string]any{
				{"text": antigravity.GetDefaultIdentityPatch()},
			},
		},
		"generationConfig": map[string]any{
			"maxOutputTokens": 1,
		},
	}
	payloadBytes, _ := json.Marshal(payload)
	return s.wrapV1InternalRequest(projectID, model, payloadBytes)
}

// buildClaudeTestRequest 构建 Claude 格式测试请求并转换为 Gemini 格式
// 使用最小 token 消耗：输入 "." + MaxTokens: 1
func (s *AntigravityGatewayService) buildClaudeTestRequest(projectID, mappedModel string) ([]byte, error) {
	claudeReq := &antigravity.ClaudeRequest{
		Model: mappedModel,
		Messages: []antigravity.ClaudeMessage{
			{
				Role:    "user",
				Content: json.RawMessage(`"."`),
			},
		},
		MaxTokens: 1,
		Stream:    false,
	}
	return antigravity.TransformClaudeToGemini(claudeReq, projectID, mappedModel)
}

func (s *AntigravityGatewayService) getClaudeTransformOptions(ctx context.Context) antigravity.TransformOptions {
	opts := antigravity.DefaultTransformOptions()
	if s.settingService == nil {
		return opts
	}
	opts.EnableIdentityPatch = s.settingService.IsIdentityPatchEnabled(ctx)
	opts.IdentityPatch = s.settingService.GetIdentityPatchPrompt(ctx)
	return opts
}

// extractTextFromSSEResponse 从 SSE 流式响应中提取文本
func extractTextFromSSEResponse(respBody []byte) string {
	var texts []string
	lines := bytes.Split(respBody, []byte("\n"))

	for _, line := range lines {
		line = bytes.TrimSpace(line)
		if len(line) == 0 {
			continue
		}

		// 跳过 SSE 前缀
		if bytes.HasPrefix(line, []byte("data:")) {
			line = bytes.TrimPrefix(line, []byte("data:"))
			line = bytes.TrimSpace(line)
		}

		// 跳过非 JSON 行
		if len(line) == 0 || line[0] != '{' {
			continue
		}

		// 解析 JSON
		var data map[string]any
		if err := json.Unmarshal(line, &data); err != nil {
			continue
		}

		// 尝试从 response.candidates[0].content.parts[].text 提取
		response, ok := data["response"].(map[string]any)
		if !ok {
			// 尝试直接从 candidates 提取（某些响应格式）
			response = data
		}

		candidates, ok := response["candidates"].([]any)
		if !ok || len(candidates) == 0 {
			continue
		}

		candidate, ok := candidates[0].(map[string]any)
		if !ok {
			continue
		}

		content, ok := candidate["content"].(map[string]any)
		if !ok {
			continue
		}

		parts, ok := content["parts"].([]any)
		if !ok {
			continue
		}

		for _, part := range parts {
			if partMap, ok := part.(map[string]any); ok {
				if text, ok := partMap["text"].(string); ok && text != "" {
					texts = append(texts, text)
				}
			}
		}
	}

	return strings.Join(texts, "")
}

// injectIdentityPatchToGeminiRequest 为 Gemini 格式请求注入身份提示词
// 如果请求中已包含 "You are Antigravity" 则不重复注入
func injectIdentityPatchToGeminiRequest(body []byte) ([]byte, error) {
	var request map[string]any
	if err := json.Unmarshal(body, &request); err != nil {
		return nil, fmt.Errorf("解析 Gemini 请求失败: %w", err)
	}

	// 检查现有 systemInstruction 是否已包含身份提示词
	if sysInst, ok := request["systemInstruction"].(map[string]any); ok {
		if parts, ok := sysInst["parts"].([]any); ok {
			for _, part := range parts {
				if partMap, ok := part.(map[string]any); ok {
					if text, ok := partMap["text"].(string); ok {
						if strings.Contains(text, "You are Antigravity") {
							// 已包含身份提示词，直接返回原始请求
							return body, nil
						}
					}
				}
			}
		}
	}

	// 获取默认身份提示词
	identityPatch := antigravity.GetDefaultIdentityPatch()

	// 构建新的 systemInstruction
	newPart := map[string]any{"text": identityPatch}

	if existing, ok := request["systemInstruction"].(map[string]any); ok {
		// 已有 systemInstruction，在开头插入身份提示词
		if parts, ok := existing["parts"].([]any); ok {
			existing["parts"] = append([]any{newPart}, parts...)
		} else {
			existing["parts"] = []any{newPart}
		}
	} else {
		// 没有 systemInstruction，创建新的
		request["systemInstruction"] = map[string]any{
			"parts": []any{newPart},
		}
	}

	return json.Marshal(request)
}

// wrapV1InternalRequest 包装请求为 v1internal 格式
func (s *AntigravityGatewayService) wrapV1InternalRequest(projectID, model string, originalBody []byte) ([]byte, error) {
	var request any
	if err := json.Unmarshal(originalBody, &request); err != nil {
		return nil, fmt.Errorf("解析请求体失败: %w", err)
	}

	wrapped := map[string]any{
		"project":     projectID,
		"requestId":   "agent-" + uuid.New().String(),
		"userAgent":   "antigravity", // 固定值，与官方客户端一致
		"requestType": "agent",
		"model":       model,
		"request":     request,
	}

	return json.Marshal(wrapped)
}

// unwrapV1InternalResponse 解包 v1internal 响应
func (s *AntigravityGatewayService) unwrapV1InternalResponse(body []byte) ([]byte, error) {
	var outer map[string]any
	if err := json.Unmarshal(body, &outer); err != nil {
		return nil, err
	}

	if resp, ok := outer["response"]; ok {
		return json.Marshal(resp)
	}

	return body, nil
}

// isModelNotFoundError 检测是否为模型不存在的 404 错误
func isModelNotFoundError(statusCode int, body []byte) bool {
	if statusCode != 404 {
		return false
	}

	bodyStr := strings.ToLower(string(body))
	keywords := []string{"model not found", "unknown model", "not found"}
	for _, keyword := range keywords {
		if strings.Contains(bodyStr, keyword) {
			return true
		}
	}
	return true // 404 without specific message also treated as model not found
}

// Forward 转发 Claude 协议请求（Claude → Gemini 转换）
//
// 限流处理流程:
//
//	请求 → antigravityRetryLoop → 预检查(remaining>0? → 切换账号) → 发送上游
//	  ├─ 成功 → 正常返回
//	  └─ 429/503 → handleSmartRetry
//	      ├─ retryDelay >= 7s → 设置模型限流 + 清除粘性绑定 → 切换账号
//	      └─ retryDelay <  7s → 等待后重试 1 次
//	          ├─ 成功 → 正常返回
//	          └─ 失败 → 设置模型限流 + 清除粘性绑定 → 切换账号
func (s *AntigravityGatewayService) Forward(ctx context.Context, c *gin.Context, account *Account, body []byte, isStickySession bool) (*ForwardResult, error) {
	// 上游透传账号直接转发，不走 OAuth token 刷新
	if account.Type == AccountTypeUpstream {
		return s.ForwardUpstream(ctx, c, account, body)
	}

	startTime := time.Now()

	sessionID := getSessionID(c)
	prefix := logPrefix(sessionID, account.Name)

	// 解析 Claude 请求
	var claudeReq antigravity.ClaudeRequest
	if err := json.Unmarshal(body, &claudeReq); err != nil {
		return nil, s.writeClaudeError(c, http.StatusBadRequest, "invalid_request_error", "Invalid request body")
	}
	if strings.TrimSpace(claudeReq.Model) == "" {
		return nil, s.writeClaudeError(c, http.StatusBadRequest, "invalid_request_error", "Missing model")
	}

	originalModel := claudeReq.Model
	mappedModel := s.getMappedModel(account, claudeReq.Model)
	if mappedModel == "" {
		return nil, s.writeClaudeError(c, http.StatusForbidden, "permission_error", fmt.Sprintf("model %s not in whitelist", claudeReq.Model))
	}
	// 应用 thinking 模式自动后缀：如果 thinking 开启且目标是 claude-sonnet-4-5，自动改为 thinking 版本
	thinkingEnabled := claudeReq.Thinking != nil && claudeReq.Thinking.Type == "enabled"
	mappedModel = applyThinkingModelSuffix(mappedModel, thinkingEnabled)

	// 获取 access_token
	if s.tokenProvider == nil {
		return nil, s.writeClaudeError(c, http.StatusBadGateway, "api_error", "Antigravity token provider not configured")
	}
	accessToken, err := s.tokenProvider.GetAccessToken(ctx, account)
	if err != nil {
		return nil, s.writeClaudeError(c, http.StatusBadGateway, "authentication_error", "Failed to get upstream access token")
	}

	// 获取 project_id（部分账户类型可能没有）
	projectID := strings.TrimSpace(account.GetCredential("project_id"))

	// 代理 URL
	proxyURL := ""
	if account.ProxyID != nil && account.Proxy != nil {
		proxyURL = account.Proxy.URL()
	}

	// 获取转换选项
	// Antigravity 上游要求必须包含身份提示词，否则会返回 429
	transformOpts := s.getClaudeTransformOptions(ctx)
	transformOpts.EnableIdentityPatch = true // 强制启用，Antigravity 上游必需

	// 转换 Claude 请求为 Gemini 格式
	geminiBody, err := antigravity.TransformClaudeToGeminiWithOptions(&claudeReq, projectID, mappedModel, transformOpts)
	if err != nil {
		return nil, s.writeClaudeError(c, http.StatusBadRequest, "invalid_request_error", "Invalid request")
	}

	// Antigravity 上游只支持流式请求，统一使用 streamGenerateContent
	// 如果客户端请求非流式，在响应处理阶段会收集完整流式响应后转换返回
	action := "streamGenerateContent"

	// 执行带重试的请求
	result, err := s.antigravityRetryLoop(antigravityRetryLoopParams{
		ctx:             ctx,
		prefix:          prefix,
		account:         account,
		proxyURL:        proxyURL,
		accessToken:     accessToken,
		action:          action,
		body:            geminiBody,
		c:               c,
		httpUpstream:    s.httpUpstream,
		settingService:  s.settingService,
		accountRepo:     s.accountRepo,
		handleError:     s.handleUpstreamError,
		requestedModel:  originalModel,
		isStickySession: isStickySession, // Forward 由上层判断粘性会话
		groupID:         0,               // Forward 方法没有 groupID，由上层处理粘性会话清除
		sessionHash:     "",              // Forward 方法没有 sessionHash，由上层处理粘性会话清除
	})
	if err != nil {
		// 检查是否是账号切换信号，转换为 UpstreamFailoverError 让 Handler 切换账号
		if switchErr, ok := IsAntigravityAccountSwitchError(err); ok {
			return nil, &UpstreamFailoverError{
				StatusCode:        http.StatusServiceUnavailable,
				ForceCacheBilling: switchErr.IsStickySession,
			}
		}
		return nil, s.writeClaudeError(c, http.StatusBadGateway, "upstream_error", "Upstream request failed after retries")
	}
	resp := result.resp
	defer func() { _ = resp.Body.Close() }()

	if resp.StatusCode >= 400 {
		respBody, _ := io.ReadAll(io.LimitReader(resp.Body, 2<<20))

		// 优先检测 thinking block 的 signature 相关错误（400）并重试一次：
		// Antigravity /v1internal 链路在部分场景会对 thought/thinking signature 做严格校验，
		// 当历史消息携带的 signature 不合法时会直接 400；去除 thinking 后可继续完成请求。
		if resp.StatusCode == http.StatusBadRequest && isSignatureRelatedError(respBody) {
			upstreamMsg := strings.TrimSpace(extractAntigravityErrorMessage(respBody))
			upstreamMsg = sanitizeUpstreamErrorMessage(upstreamMsg)
			logBody, maxBytes := s.getLogConfig()
			upstreamDetail := s.getUpstreamErrorDetail(respBody)
			appendOpsUpstreamError(c, OpsUpstreamErrorEvent{
				Platform:           account.Platform,
				AccountID:          account.ID,
				AccountName:        account.Name,
				UpstreamStatusCode: resp.StatusCode,
				UpstreamRequestID:  resp.Header.Get("x-request-id"),
				Kind:               "signature_error",
				Message:            upstreamMsg,
				Detail:             upstreamDetail,
			})

			// Conservative two-stage fallback:
			// 1) Disable top-level thinking + thinking->text
			// 2) Only if still signature-related 400: also downgrade tool_use/tool_result to text.

			retryStages := []struct {
				name  string
				strip func(*antigravity.ClaudeRequest) (bool, error)
			}{
				{name: "thinking-only", strip: stripThinkingFromClaudeRequest},
				{name: "thinking+tools", strip: stripSignatureSensitiveBlocksFromClaudeRequest},
			}

			for _, stage := range retryStages {
				retryClaudeReq := claudeReq
				retryClaudeReq.Messages = append([]antigravity.ClaudeMessage(nil), claudeReq.Messages...)

				stripped, stripErr := stage.strip(&retryClaudeReq)
				if stripErr != nil || !stripped {
					continue
				}

				log.Printf("Antigravity account %d: detected signature-related 400, retrying once (%s)", account.ID, stage.name)

				retryGeminiBody, txErr := antigravity.TransformClaudeToGeminiWithOptions(&retryClaudeReq, projectID, mappedModel, s.getClaudeTransformOptions(ctx))
				if txErr != nil {
					continue
				}
				retryResult, retryErr := s.antigravityRetryLoop(antigravityRetryLoopParams{
					ctx:             ctx,
					prefix:          prefix,
					account:         account,
					proxyURL:        proxyURL,
					accessToken:     accessToken,
					action:          action,
					body:            retryGeminiBody,
					c:               c,
					httpUpstream:    s.httpUpstream,
					settingService:  s.settingService,
					accountRepo:     s.accountRepo,
					handleError:     s.handleUpstreamError,
					requestedModel:  originalModel,
					isStickySession: isStickySession,
					groupID:         0,  // Forward 方法没有 groupID，由上层处理粘性会话清除
					sessionHash:     "", // Forward 方法没有 sessionHash，由上层处理粘性会话清除
				})
				if retryErr != nil {
					appendOpsUpstreamError(c, OpsUpstreamErrorEvent{
						Platform:           account.Platform,
						AccountID:          account.ID,
						AccountName:        account.Name,
						UpstreamStatusCode: 0,
						Kind:               "signature_retry_request_error",
						Message:            sanitizeUpstreamErrorMessage(retryErr.Error()),
					})
					log.Printf("Antigravity account %d: signature retry request failed (%s): %v", account.ID, stage.name, retryErr)
					continue
				}

				retryResp := retryResult.resp
				if retryResp.StatusCode < 400 {
					_ = resp.Body.Close()
					resp = retryResp
					respBody = nil
					break
				}

				retryBody, _ := io.ReadAll(io.LimitReader(retryResp.Body, 2<<20))
				_ = retryResp.Body.Close()
				if retryResp.StatusCode == http.StatusTooManyRequests {
					retryBaseURL := ""
					if retryResp.Request != nil && retryResp.Request.URL != nil {
						retryBaseURL = retryResp.Request.URL.Scheme + "://" + retryResp.Request.URL.Host
					}
					log.Printf("%s status=429 rate_limited base_url=%s retry_stage=%s body=%s", prefix, retryBaseURL, stage.name, truncateForLog(retryBody, 200))
				}
				kind := "signature_retry"
				if strings.TrimSpace(stage.name) != "" {
					kind = "signature_retry_" + strings.ReplaceAll(stage.name, "+", "_")
				}
				retryUpstreamMsg := strings.TrimSpace(extractAntigravityErrorMessage(retryBody))
				retryUpstreamMsg = sanitizeUpstreamErrorMessage(retryUpstreamMsg)
				retryUpstreamDetail := ""
				if logBody {
					retryUpstreamDetail = truncateString(string(retryBody), maxBytes)
				}
				appendOpsUpstreamError(c, OpsUpstreamErrorEvent{
					Platform:           account.Platform,
					AccountID:          account.ID,
					AccountName:        account.Name,
					UpstreamStatusCode: retryResp.StatusCode,
					UpstreamRequestID:  retryResp.Header.Get("x-request-id"),
					Kind:               kind,
					Message:            retryUpstreamMsg,
					Detail:             retryUpstreamDetail,
				})

				// If this stage fixed the signature issue, we stop; otherwise we may try the next stage.
				if retryResp.StatusCode != http.StatusBadRequest || !isSignatureRelatedError(retryBody) {
					respBody = retryBody
					resp = &http.Response{
						StatusCode: retryResp.StatusCode,
						Header:     retryResp.Header.Clone(),
						Body:       io.NopCloser(bytes.NewReader(retryBody)),
					}
					break
				}

				// Still signature-related; capture context and allow next stage.
				respBody = retryBody
				resp = &http.Response{
					StatusCode: retryResp.StatusCode,
					Header:     retryResp.Header.Clone(),
					Body:       io.NopCloser(bytes.NewReader(retryBody)),
				}
			}
		}

		// 处理错误响应（重试后仍失败或不触发重试）
		if resp.StatusCode >= 400 {
			// 检测 prompt too long 错误，返回特殊错误类型供上层 fallback
			if resp.StatusCode == http.StatusBadRequest && isPromptTooLongError(respBody) {
				upstreamMsg := strings.TrimSpace(extractAntigravityErrorMessage(respBody))
				upstreamMsg = sanitizeUpstreamErrorMessage(upstreamMsg)
				upstreamDetail := s.getUpstreamErrorDetail(respBody)
				logBody, maxBytes := s.getLogConfig()
				if logBody {
					log.Printf("%s status=400 prompt_too_long=true upstream_message=%q request_id=%s body=%s", prefix, upstreamMsg, resp.Header.Get("x-request-id"), truncateForLog(respBody, maxBytes))
				}
				appendOpsUpstreamError(c, OpsUpstreamErrorEvent{
					Platform:           account.Platform,
					AccountID:          account.ID,
					AccountName:        account.Name,
					UpstreamStatusCode: resp.StatusCode,
					UpstreamRequestID:  resp.Header.Get("x-request-id"),
					Kind:               "prompt_too_long",
					Message:            upstreamMsg,
					Detail:             upstreamDetail,
				})
				return nil, &PromptTooLongError{
					StatusCode: resp.StatusCode,
					RequestID:  resp.Header.Get("x-request-id"),
					Body:       respBody,
				}
			}

			s.handleUpstreamError(ctx, prefix, account, resp.StatusCode, resp.Header, respBody, originalModel, 0, "", isStickySession)

			if s.shouldFailoverUpstreamError(resp.StatusCode) {
				upstreamMsg := strings.TrimSpace(extractAntigravityErrorMessage(respBody))
				upstreamMsg = sanitizeUpstreamErrorMessage(upstreamMsg)
				upstreamDetail := s.getUpstreamErrorDetail(respBody)
				appendOpsUpstreamError(c, OpsUpstreamErrorEvent{
					Platform:           account.Platform,
					AccountID:          account.ID,
					AccountName:        account.Name,
					UpstreamStatusCode: resp.StatusCode,
					UpstreamRequestID:  resp.Header.Get("x-request-id"),
					Kind:               "failover",
					Message:            upstreamMsg,
					Detail:             upstreamDetail,
				})
				return nil, &UpstreamFailoverError{StatusCode: resp.StatusCode, ResponseBody: respBody}
			}

			return nil, s.writeMappedClaudeError(c, account, resp.StatusCode, resp.Header.Get("x-request-id"), respBody)
		}
	}

	requestID := resp.Header.Get("x-request-id")
	if requestID != "" {
		c.Header("x-request-id", requestID)
	}

	var usage *ClaudeUsage
	var firstTokenMs *int
	var clientDisconnect bool
	if claudeReq.Stream {
		// 客户端要求流式，直接透传转换
		streamRes, err := s.handleClaudeStreamingResponse(c, resp, startTime, originalModel)
		if err != nil {
			log.Printf("%s status=stream_error error=%v", prefix, err)
			return nil, err
		}
		usage = streamRes.usage
		firstTokenMs = streamRes.firstTokenMs
		clientDisconnect = streamRes.clientDisconnect
	} else {
		// 客户端要求非流式，收集流式响应后转换返回
		streamRes, err := s.handleClaudeStreamToNonStreaming(c, resp, startTime, originalModel)
		if err != nil {
			log.Printf("%s status=stream_collect_error error=%v", prefix, err)
			return nil, err
		}
		usage = streamRes.usage
		firstTokenMs = streamRes.firstTokenMs
	}

	return &ForwardResult{
		RequestID:        requestID,
		Usage:            *usage,
		Model:            originalModel, // 使用原始模型用于计费和日志
		Stream:           claudeReq.Stream,
		Duration:         time.Since(startTime),
		FirstTokenMs:     firstTokenMs,
		ClientDisconnect: clientDisconnect,
	}, nil
}

func isSignatureRelatedError(respBody []byte) bool {
	msg := strings.ToLower(strings.TrimSpace(extractAntigravityErrorMessage(respBody)))
	if msg == "" {
		// Fallback: best-effort scan of the raw payload.
		msg = strings.ToLower(string(respBody))
	}

	// Keep this intentionally broad: different upstreams may use "signature" or "thought_signature".
	if strings.Contains(msg, "thought_signature") || strings.Contains(msg, "signature") {
		return true
	}

	// Also detect thinking block structural errors:
	// "Expected `thinking` or `redacted_thinking`, but found `text`"
	if strings.Contains(msg, "expected") && (strings.Contains(msg, "thinking") || strings.Contains(msg, "redacted_thinking")) {
		return true
	}

	return false
}

// isPromptTooLongError 检测是否为 prompt too long 错误
func isPromptTooLongError(respBody []byte) bool {
	msg := strings.ToLower(strings.TrimSpace(extractAntigravityErrorMessage(respBody)))
	if msg == "" {
		msg = strings.ToLower(string(respBody))
	}
	return strings.Contains(msg, "prompt is too long") ||
		strings.Contains(msg, "request is too long") ||
		strings.Contains(msg, "context length exceeded") ||
		strings.Contains(msg, "max_tokens")
}

// isPassthroughErrorMessage 检查错误消息是否在透传白名单中
func isPassthroughErrorMessage(msg string) bool {
	lower := strings.ToLower(msg)
	for _, pattern := range antigravityPassthroughErrorMessages {
		if strings.Contains(lower, pattern) {
			return true
		}
	}
	return false
}

// getPassthroughOrDefault 若消息在白名单内则返回原始消息，否则返回默认消息
func getPassthroughOrDefault(upstreamMsg, defaultMsg string) string {
	if isPassthroughErrorMessage(upstreamMsg) {
		return upstreamMsg
	}
	return defaultMsg
}

func extractAntigravityErrorMessage(body []byte) string {
	var payload map[string]any
	if err := json.Unmarshal(body, &payload); err != nil {
		return ""
	}

	// Google-style: {"error": {"message": "..."}}
	if errObj, ok := payload["error"].(map[string]any); ok {
		if msg, ok := errObj["message"].(string); ok && strings.TrimSpace(msg) != "" {
			return msg
		}
	}

	// Fallback: top-level message
	if msg, ok := payload["message"].(string); ok && strings.TrimSpace(msg) != "" {
		return msg
	}

	return ""
}

// stripThinkingFromClaudeRequest converts thinking blocks to text blocks in a Claude Messages request.
// This preserves the thinking content while avoiding signature validation errors.
// Note: redacted_thinking blocks are removed because they cannot be converted to text.
// It also disables top-level `thinking` to avoid upstream structural constraints for thinking mode.
func stripThinkingFromClaudeRequest(req *antigravity.ClaudeRequest) (bool, error) {
	if req == nil {
		return false, nil
	}

	changed := false
	if req.Thinking != nil {
		req.Thinking = nil
		changed = true
	}

	for i := range req.Messages {
		raw := req.Messages[i].Content
		if len(raw) == 0 {
			continue
		}

		// If content is a string, nothing to strip.
		var str string
		if json.Unmarshal(raw, &str) == nil {
			continue
		}

		// Otherwise treat as an array of blocks and convert thinking blocks to text.
		var blocks []map[string]any
		if err := json.Unmarshal(raw, &blocks); err != nil {
			continue
		}

		filtered := make([]map[string]any, 0, len(blocks))
		modifiedAny := false
		for _, block := range blocks {
			t, _ := block["type"].(string)
			switch t {
			case "thinking":
				thinkingText, _ := block["thinking"].(string)
				if thinkingText != "" {
					filtered = append(filtered, map[string]any{
						"type": "text",
						"text": thinkingText,
					})
				}
				modifiedAny = true
			case "redacted_thinking":
				modifiedAny = true
			case "":
				if thinkingText, hasThinking := block["thinking"].(string); hasThinking {
					if thinkingText != "" {
						filtered = append(filtered, map[string]any{
							"type": "text",
							"text": thinkingText,
						})
					}
					modifiedAny = true
				} else {
					filtered = append(filtered, block)
				}
			default:
				filtered = append(filtered, block)
			}
		}

		if !modifiedAny {
			continue
		}

		if len(filtered) == 0 {
			filtered = append(filtered, map[string]any{
				"type": "text",
				"text": "(content removed)",
			})
		}

		newRaw, err := json.Marshal(filtered)
		if err != nil {
			return changed, err
		}
		req.Messages[i].Content = newRaw
		changed = true
	}

	return changed, nil
}

// stripSignatureSensitiveBlocksFromClaudeRequest is a stronger retry degradation that additionally converts
// tool blocks to plain text. Use this only after a thinking-only retry still fails with signature errors.
func stripSignatureSensitiveBlocksFromClaudeRequest(req *antigravity.ClaudeRequest) (bool, error) {
	if req == nil {
		return false, nil
	}

	changed := false
	if req.Thinking != nil {
		req.Thinking = nil
		changed = true
	}

	for i := range req.Messages {
		raw := req.Messages[i].Content
		if len(raw) == 0 {
			continue
		}

		// If content is a string, nothing to strip.
		var str string
		if json.Unmarshal(raw, &str) == nil {
			continue
		}

		// Otherwise treat as an array of blocks and convert signature-sensitive blocks to text.
		var blocks []map[string]any
		if err := json.Unmarshal(raw, &blocks); err != nil {
			continue
		}

		filtered := make([]map[string]any, 0, len(blocks))
		modifiedAny := false
		for _, block := range blocks {
			t, _ := block["type"].(string)
			switch t {
			case "thinking":
				// Convert thinking to text, skip if empty
				thinkingText, _ := block["thinking"].(string)
				if thinkingText != "" {
					filtered = append(filtered, map[string]any{
						"type": "text",
						"text": thinkingText,
					})
				}
				modifiedAny = true
			case "redacted_thinking":
				// Remove redacted_thinking (cannot convert encrypted content)
				modifiedAny = true
			case "tool_use":
				// Convert tool_use to text to avoid upstream signature/thought_signature validation errors.
				// This is a retry-only degradation path, so we prioritise request validity over tool semantics.
				name, _ := block["name"].(string)
				id, _ := block["id"].(string)
				input := block["input"]
				inputJSON, _ := json.Marshal(input)
				text := "(tool_use)"
				if name != "" {
					text += " name=" + name
				}
				if id != "" {
					text += " id=" + id
				}
				if len(inputJSON) > 0 && string(inputJSON) != "null" {
					text += " input=" + string(inputJSON)
				}
				filtered = append(filtered, map[string]any{
					"type": "text",
					"text": text,
				})
				modifiedAny = true
			case "tool_result":
				// Convert tool_result to text so it stays consistent when tool_use is downgraded.
				toolUseID, _ := block["tool_use_id"].(string)
				isError, _ := block["is_error"].(bool)
				content := block["content"]
				contentJSON, _ := json.Marshal(content)
				text := "(tool_result)"
				if toolUseID != "" {
					text += " tool_use_id=" + toolUseID
				}
				if isError {
					text += " is_error=true"
				}
				if len(contentJSON) > 0 && string(contentJSON) != "null" {
					text += "\n" + string(contentJSON)
				}
				filtered = append(filtered, map[string]any{
					"type": "text",
					"text": text,
				})
				modifiedAny = true
			case "":
				// Handle untyped block with "thinking" field
				if thinkingText, hasThinking := block["thinking"].(string); hasThinking {
					if thinkingText != "" {
						filtered = append(filtered, map[string]any{
							"type": "text",
							"text": thinkingText,
						})
					}
					modifiedAny = true
				} else {
					filtered = append(filtered, block)
				}
			default:
				filtered = append(filtered, block)
			}
		}

		if !modifiedAny {
			continue
		}

		if len(filtered) == 0 {
			// Keep request valid: upstream rejects empty content arrays.
			filtered = append(filtered, map[string]any{
				"type": "text",
				"text": "(content removed)",
			})
		}

		newRaw, err := json.Marshal(filtered)
		if err != nil {
			return changed, err
		}
		req.Messages[i].Content = newRaw
		changed = true
	}

	return changed, nil
}

// ForwardGemini 转发 Gemini 协议请求
//
// 限流处理流程:
//
//	请求 → antigravityRetryLoop → 预检查(remaining>0? → 切换账号) → 发送上游
//	  ├─ 成功 → 正常返回
//	  └─ 429/503 → handleSmartRetry
//	      ├─ retryDelay >= 7s → 设置模型限流 + 清除粘性绑定 → 切换账号
//	      └─ retryDelay <  7s → 等待后重试 1 次
//	          ├─ 成功 → 正常返回
//	          └─ 失败 → 设置模型限流 + 清除粘性绑定 → 切换账号
func (s *AntigravityGatewayService) ForwardGemini(ctx context.Context, c *gin.Context, account *Account, originalModel string, action string, stream bool, body []byte, isStickySession bool) (*ForwardResult, error) {
	startTime := time.Now()

	sessionID := getSessionID(c)
	prefix := logPrefix(sessionID, account.Name)

	if strings.TrimSpace(originalModel) == "" {
		return nil, s.writeGoogleError(c, http.StatusBadRequest, "Missing model in URL")
	}
	if strings.TrimSpace(action) == "" {
		return nil, s.writeGoogleError(c, http.StatusBadRequest, "Missing action in URL")
	}
	if len(body) == 0 {
		return nil, s.writeGoogleError(c, http.StatusBadRequest, "Request body is empty")
	}

	// 解析请求以获取 image_size（用于图片计费）
	imageSize := s.extractImageSize(body)

	switch action {
	case "generateContent", "streamGenerateContent":
		// ok
	case "countTokens":
		// 直接返回空值，不透传上游
		c.JSON(http.StatusOK, map[string]any{"totalTokens": 0})
		return &ForwardResult{
			RequestID:    "",
			Usage:        ClaudeUsage{},
			Model:        originalModel,
			Stream:       false,
			Duration:     time.Since(time.Now()),
			FirstTokenMs: nil,
		}, nil
	default:
		return nil, s.writeGoogleError(c, http.StatusNotFound, "Unsupported action: "+action)
	}

	mappedModel := s.getMappedModel(account, originalModel)
	if mappedModel == "" {
		return nil, s.writeGoogleError(c, http.StatusForbidden, fmt.Sprintf("model %s not in whitelist", originalModel))
	}

	// 获取 access_token
	if s.tokenProvider == nil {
		return nil, s.writeGoogleError(c, http.StatusBadGateway, "Antigravity token provider not configured")
	}
	accessToken, err := s.tokenProvider.GetAccessToken(ctx, account)
	if err != nil {
		return nil, s.writeGoogleError(c, http.StatusBadGateway, "Failed to get upstream access token")
	}

	// 获取 project_id（部分账户类型可能没有）
	projectID := strings.TrimSpace(account.GetCredential("project_id"))

	// 代理 URL
	proxyURL := ""
	if account.ProxyID != nil && account.Proxy != nil {
		proxyURL = account.Proxy.URL()
	}

	// Antigravity 上游要求必须包含身份提示词，注入到请求中
	injectedBody, err := injectIdentityPatchToGeminiRequest(body)
	if err != nil {
		return nil, s.writeGoogleError(c, http.StatusBadRequest, "Invalid request body")
	}

	// 清理 Schema
	if cleanedBody, err := cleanGeminiRequest(injectedBody); err == nil {
		injectedBody = cleanedBody
		log.Printf("[Antigravity] Cleaned request schema in forwarded request for account %s", account.Name)
	} else {
		log.Printf("[Antigravity] Failed to clean schema: %v", err)
	}

	// 包装请求
	wrappedBody, err := s.wrapV1InternalRequest(projectID, mappedModel, injectedBody)
	if err != nil {
		return nil, s.writeGoogleError(c, http.StatusInternalServerError, "Failed to build upstream request")
	}

	// Antigravity 上游只支持流式请求，统一使用 streamGenerateContent
	// 如果客户端请求非流式，在响应处理阶段会收集完整流式响应后返回
	upstreamAction := "streamGenerateContent"

	// 执行带重试的请求
	result, err := s.antigravityRetryLoop(antigravityRetryLoopParams{
		ctx:             ctx,
		prefix:          prefix,
		account:         account,
		proxyURL:        proxyURL,
		accessToken:     accessToken,
		action:          upstreamAction,
		body:            wrappedBody,
		c:               c,
		httpUpstream:    s.httpUpstream,
		settingService:  s.settingService,
		accountRepo:     s.accountRepo,
		handleError:     s.handleUpstreamError,
		requestedModel:  originalModel,
		isStickySession: isStickySession, // ForwardGemini 由上层判断粘性会话
		groupID:         0,               // ForwardGemini 方法没有 groupID，由上层处理粘性会话清除
		sessionHash:     "",              // ForwardGemini 方法没有 sessionHash，由上层处理粘性会话清除
	})
	if err != nil {
		// 检查是否是账号切换信号，转换为 UpstreamFailoverError 让 Handler 切换账号
		if switchErr, ok := IsAntigravityAccountSwitchError(err); ok {
			return nil, &UpstreamFailoverError{
				StatusCode:        http.StatusServiceUnavailable,
				ForceCacheBilling: switchErr.IsStickySession,
			}
		}
		return nil, s.writeGoogleError(c, http.StatusBadGateway, "Upstream request failed after retries")
	}
	resp := result.resp
	defer func() {
		if resp != nil && resp.Body != nil {
			_ = resp.Body.Close()
		}
	}()

	// 处理错误响应
	if resp.StatusCode >= 400 {
		respBody, _ := io.ReadAll(io.LimitReader(resp.Body, 2<<20))
		contentType := resp.Header.Get("Content-Type")
		// 尽早关闭原始响应体，释放连接；后续逻辑仍可能需要读取 body，因此用内存副本重新包装。
		_ = resp.Body.Close()
		resp.Body = io.NopCloser(bytes.NewReader(respBody))

		// 模型兜底：模型不存在且开启 fallback 时，自动用 fallback 模型重试一次
		if s.settingService != nil && s.settingService.IsModelFallbackEnabled(ctx) &&
			isModelNotFoundError(resp.StatusCode, respBody) {
			fallbackModel := s.settingService.GetFallbackModel(ctx, PlatformAntigravity)
			if fallbackModel != "" && fallbackModel != mappedModel {
				log.Printf("[Antigravity] Model not found (%s), retrying with fallback model %s (account: %s)", mappedModel, fallbackModel, account.Name)

				fallbackWrapped, err := s.wrapV1InternalRequest(projectID, fallbackModel, injectedBody)
				if err == nil {
					fallbackReq, err := antigravity.NewAPIRequest(ctx, upstreamAction, accessToken, fallbackWrapped)
					if err == nil {
						fallbackResp, err := s.httpUpstream.Do(fallbackReq, proxyURL, account.ID, account.Concurrency)
						if err == nil && fallbackResp.StatusCode < 400 {
							_ = resp.Body.Close()
							resp = fallbackResp
						} else if fallbackResp != nil {
							_ = fallbackResp.Body.Close()
						}
					}
				}
			}
		}

		// fallback 成功：继续按正常响应处理
		if resp.StatusCode < 400 {
			goto handleSuccess
		}

		requestID := resp.Header.Get("x-request-id")
		if requestID != "" {
			c.Header("x-request-id", requestID)
		}

		unwrapped, unwrapErr := s.unwrapV1InternalResponse(respBody)
		unwrappedForOps := unwrapped
		if unwrapErr != nil || len(unwrappedForOps) == 0 {
			unwrappedForOps = respBody
		}
		s.handleUpstreamError(ctx, prefix, account, resp.StatusCode, resp.Header, respBody, originalModel, 0, "", isStickySession)
		upstreamMsg := strings.TrimSpace(extractAntigravityErrorMessage(unwrappedForOps))
		upstreamMsg = sanitizeUpstreamErrorMessage(upstreamMsg)
		upstreamDetail := s.getUpstreamErrorDetail(unwrappedForOps)

		// Always record upstream context for Ops error logs, even when we will failover.
		setOpsUpstreamError(c, resp.StatusCode, upstreamMsg, upstreamDetail)

		if s.shouldFailoverUpstreamError(resp.StatusCode) {
			appendOpsUpstreamError(c, OpsUpstreamErrorEvent{
				Platform:           account.Platform,
				AccountID:          account.ID,
				AccountName:        account.Name,
				UpstreamStatusCode: resp.StatusCode,
				UpstreamRequestID:  requestID,
				Kind:               "failover",
				Message:            upstreamMsg,
				Detail:             upstreamDetail,
			})
			return nil, &UpstreamFailoverError{StatusCode: resp.StatusCode, ResponseBody: unwrappedForOps}
		}
		if contentType == "" {
			contentType = "application/json"
		}
		appendOpsUpstreamError(c, OpsUpstreamErrorEvent{
			Platform:           account.Platform,
			AccountID:          account.ID,
			AccountName:        account.Name,
			UpstreamStatusCode: resp.StatusCode,
			UpstreamRequestID:  requestID,
			Kind:               "http_error",
			Message:            upstreamMsg,
			Detail:             upstreamDetail,
		})
		log.Printf("[antigravity-Forward] upstream error status=%d body=%s", resp.StatusCode, truncateForLog(unwrappedForOps, 500))
		c.Data(resp.StatusCode, contentType, unwrappedForOps)
		return nil, fmt.Errorf("antigravity upstream error: %d", resp.StatusCode)
	}

handleSuccess:
	requestID := resp.Header.Get("x-request-id")
	if requestID != "" {
		c.Header("x-request-id", requestID)
	}

	var usage *ClaudeUsage
	var firstTokenMs *int
	var clientDisconnect bool

	if stream {
		// 客户端要求流式，直接透传
		streamRes, err := s.handleGeminiStreamingResponse(c, resp, startTime)
		if err != nil {
			log.Printf("%s status=stream_error error=%v", prefix, err)
			return nil, err
		}
		usage = streamRes.usage
		firstTokenMs = streamRes.firstTokenMs
		clientDisconnect = streamRes.clientDisconnect
	} else {
		// 客户端要求非流式，收集流式响应后返回
		streamRes, err := s.handleGeminiStreamToNonStreaming(c, resp, startTime)
		if err != nil {
			log.Printf("%s status=stream_collect_error error=%v", prefix, err)
			return nil, err
		}
		usage = streamRes.usage
		firstTokenMs = streamRes.firstTokenMs
	}

	if usage == nil {
		usage = &ClaudeUsage{}
	}

	// 判断是否为图片生成模型
	imageCount := 0
	if isImageGenerationModel(mappedModel) {
		// Gemini 图片生成 API 每次请求只生成一张图片（API 限制）
		imageCount = 1
	}

	return &ForwardResult{
		RequestID:        requestID,
		Usage:            *usage,
		Model:            originalModel,
		Stream:           stream,
		Duration:         time.Since(startTime),
		FirstTokenMs:     firstTokenMs,
		ClientDisconnect: clientDisconnect,
		ImageCount:       imageCount,
		ImageSize:        imageSize,
	}, nil
}

func (s *AntigravityGatewayService) shouldFailoverUpstreamError(statusCode int) bool {
	switch statusCode {
	case 401, 403, 429, 529:
		return true
	default:
		return statusCode >= 500
	}
}

// sleepAntigravityBackoffWithContext 带 context 取消检查的退避等待
// 返回 true 表示正常完成等待，false 表示 context 已取消
func sleepAntigravityBackoffWithContext(ctx context.Context, attempt int) bool {
	delay := antigravityRetryBaseDelay * time.Duration(1<<uint(attempt-1))
	if delay > antigravityRetryMaxDelay {
		delay = antigravityRetryMaxDelay
	}

	// +/- 20% jitter
	r := mathrand.New(mathrand.NewSource(time.Now().UnixNano()))
	jitter := time.Duration(float64(delay) * 0.2 * (r.Float64()*2 - 1))
	sleepFor := delay + jitter
	if sleepFor < 0 {
		sleepFor = 0
	}

	select {
	case <-ctx.Done():
		return false
	case <-time.After(sleepFor):
		return true
	}
}

// setModelRateLimitByModelName 使用官方模型 ID 设置模型级限流
// 直接使用上游返回的模型 ID（如 claude-sonnet-4-5）作为限流 key
// 返回是否已成功设置（若模型名为空或 repo 为 nil 将返回 false）
func setModelRateLimitByModelName(ctx context.Context, repo AccountRepository, accountID int64, modelName, prefix string, statusCode int, resetAt time.Time, afterSmartRetry bool) bool {
	if repo == nil || modelName == "" {
		return false
	}
	// 直接使用官方模型 ID 作为 key，不再转换为 scope
	if err := repo.SetModelRateLimit(ctx, accountID, modelName, resetAt); err != nil {
		log.Printf("%s status=%d model_rate_limit_failed model=%s error=%v", prefix, statusCode, modelName, err)
		return false
	}
	if afterSmartRetry {
		log.Printf("%s status=%d model_rate_limited_after_smart_retry model=%s account=%d reset_in=%v", prefix, statusCode, modelName, accountID, time.Until(resetAt).Truncate(time.Second))
	} else {
		log.Printf("%s status=%d model_rate_limited model=%s account=%d reset_in=%v", prefix, statusCode, modelName, accountID, time.Until(resetAt).Truncate(time.Second))
	}
	return true
}

func antigravityFallbackCooldownSeconds() (time.Duration, bool) {
	raw := strings.TrimSpace(os.Getenv(antigravityFallbackSecondsEnv))
	if raw == "" {
		return 0, false
	}
	seconds, err := strconv.Atoi(raw)
	if err != nil || seconds <= 0 {
		return 0, false
	}
	return time.Duration(seconds) * time.Second, true
}

// antigravitySmartRetryInfo 智能重试所需的信息
type antigravitySmartRetryInfo struct {
	RetryDelay time.Duration // 重试延迟时间
	ModelName  string        // 限流的模型名称（如 "claude-sonnet-4-5"）
}

// parseAntigravitySmartRetryInfo 解析 Google RPC RetryInfo 和 ErrorInfo 信息
// 返回解析结果，如果解析失败或不满足条件返回 nil
//
// 支持两种情况：
// 1. 429 RESOURCE_EXHAUSTED + RATE_LIMIT_EXCEEDED：
//   - error.status == "RESOURCE_EXHAUSTED"
//   - error.details[].reason == "RATE_LIMIT_EXCEEDED"
//
// 2. 503 UNAVAILABLE + MODEL_CAPACITY_EXHAUSTED：
//   - error.status == "UNAVAILABLE"
//   - error.details[].reason == "MODEL_CAPACITY_EXHAUSTED"
//
// 必须满足以下条件才会返回有效值：
// - error.details[] 中存在 @type == "type.googleapis.com/google.rpc.RetryInfo" 的元素
// - 该元素包含 retryDelay 字段，格式为 "数字s"（如 "0.201506475s"）
func parseAntigravitySmartRetryInfo(body []byte) *antigravitySmartRetryInfo {
	var parsed map[string]any
	if err := json.Unmarshal(body, &parsed); err != nil {
		return nil
	}

	errObj, ok := parsed["error"].(map[string]any)
	if !ok {
		return nil
	}

	// 检查 status 是否符合条件
	// 情况1: 429 RESOURCE_EXHAUSTED (需要进一步检查 reason == RATE_LIMIT_EXCEEDED)
	// 情况2: 503 UNAVAILABLE (需要进一步检查 reason == MODEL_CAPACITY_EXHAUSTED)
	status, _ := errObj["status"].(string)
	isResourceExhausted := status == googleRPCStatusResourceExhausted
	isUnavailable := status == googleRPCStatusUnavailable

	if !isResourceExhausted && !isUnavailable {
		return nil
	}

	details, ok := errObj["details"].([]any)
	if !ok {
		return nil
	}

	var retryDelay time.Duration
	var modelName string
	var hasRateLimitExceeded bool      // 429 需要此 reason
	var hasModelCapacityExhausted bool // 503 需要此 reason

	for _, d := range details {
		dm, ok := d.(map[string]any)
		if !ok {
			continue
		}

		atType, _ := dm["@type"].(string)

		// 从 ErrorInfo 提取模型名称和 reason
		if atType == googleRPCTypeErrorInfo {
			if meta, ok := dm["metadata"].(map[string]any); ok {
				if model, ok := meta["model"].(string); ok {
					modelName = model
				}
			}
			// 检查 reason
			if reason, ok := dm["reason"].(string); ok {
				if reason == googleRPCReasonModelCapacityExhausted {
					hasModelCapacityExhausted = true
				}
				if reason == googleRPCReasonRateLimitExceeded {
					hasRateLimitExceeded = true
				}
			}
			continue
		}

		// 从 RetryInfo 提取重试延迟
		if atType == googleRPCTypeRetryInfo {
			delay, ok := dm["retryDelay"].(string)
			if !ok || delay == "" {
				continue
			}
			// 使用 time.ParseDuration 解析，支持所有 Go duration 格式
			// 例如: "0.5s", "10s", "4m50s", "1h30m", "200ms" 等
			dur, err := time.ParseDuration(delay)
			if err != nil {
				log.Printf("[Antigravity] failed to parse retryDelay: %s error=%v", delay, err)
				continue
			}
			retryDelay = dur
		}
	}

	// 验证条件
	// 情况1: RESOURCE_EXHAUSTED 需要有 RATE_LIMIT_EXCEEDED reason
	// 情况2: UNAVAILABLE 需要有 MODEL_CAPACITY_EXHAUSTED reason
	if isResourceExhausted && !hasRateLimitExceeded {
		return nil
	}
	if isUnavailable && !hasModelCapacityExhausted {
		return nil
	}

	// 必须有模型名才返回有效结果
	if modelName == "" {
		return nil
	}

	// 如果上游未提供 retryDelay，使用默认限流时间
	if retryDelay <= 0 {
		retryDelay = antigravityDefaultRateLimitDuration
	}

	return &antigravitySmartRetryInfo{
		RetryDelay: retryDelay,
		ModelName:  modelName,
	}
}

// shouldTriggerAntigravitySmartRetry 判断是否应该触发智能重试
// 返回：
//   - shouldRetry: 是否应该智能重试（retryDelay < antigravityRateLimitThreshold）
//   - shouldRateLimitModel: 是否应该限流模型（retryDelay >= antigravityRateLimitThreshold）
//   - waitDuration: 等待时间（智能重试时使用，shouldRateLimitModel=true 时为 0）
//   - modelName: 限流的模型名称
func shouldTriggerAntigravitySmartRetry(account *Account, respBody []byte) (shouldRetry bool, shouldRateLimitModel bool, waitDuration time.Duration, modelName string) {
	if account.Platform != PlatformAntigravity {
		return false, false, 0, ""
	}

	info := parseAntigravitySmartRetryInfo(respBody)
	if info == nil {
		return false, false, 0, ""
	}

	// retryDelay >= 阈值：直接限流模型，不重试
	// 注意：如果上游未提供 retryDelay，parseAntigravitySmartRetryInfo 已设置为默认 30s
	if info.RetryDelay >= antigravityRateLimitThreshold {
		return false, true, info.RetryDelay, info.ModelName
	}

	// retryDelay < 阈值：智能重试
	waitDuration = info.RetryDelay
	if waitDuration < antigravitySmartRetryMinWait {
		waitDuration = antigravitySmartRetryMinWait
	}

	return true, false, waitDuration, info.ModelName
}

// handleModelRateLimitParams 模型级限流处理参数
type handleModelRateLimitParams struct {
	ctx             context.Context
	prefix          string
	account         *Account
	statusCode      int
	body            []byte
	cache           GatewayCache
	groupID         int64
	sessionHash     string
	isStickySession bool
}

// handleModelRateLimitResult 模型级限流处理结果
type handleModelRateLimitResult struct {
	Handled      bool                           // 是否已处理
	ShouldRetry  bool                           // 是否等待后重试
	WaitDuration time.Duration                  // 等待时间
	SwitchError  *AntigravityAccountSwitchError // 账号切换错误
}

// handleModelRateLimit 处理模型级限流（在原有逻辑之前调用）
// 仅处理 429/503，解析模型名和 retryDelay
// - retryDelay < antigravityRateLimitThreshold: 返回 ShouldRetry=true，由调用方等待后重试
// - retryDelay >= antigravityRateLimitThreshold: 设置模型限流 + 清除粘性会话 + 返回 SwitchError
func (s *AntigravityGatewayService) handleModelRateLimit(p *handleModelRateLimitParams) *handleModelRateLimitResult {
	if p.statusCode != 429 && p.statusCode != 503 {
		return &handleModelRateLimitResult{Handled: false}
	}

	info := parseAntigravitySmartRetryInfo(p.body)
	if info == nil || info.ModelName == "" {
		return &handleModelRateLimitResult{Handled: false}
	}

	// < antigravityRateLimitThreshold: 等待后重试
	if info.RetryDelay < antigravityRateLimitThreshold {
		log.Printf("%s status=%d model_rate_limit_wait model=%s wait=%v",
			p.prefix, p.statusCode, info.ModelName, info.RetryDelay)
		return &handleModelRateLimitResult{
			Handled:      true,
			ShouldRetry:  true,
			WaitDuration: info.RetryDelay,
		}
	}

	// >= antigravityRateLimitThreshold: 设置限流 + 清除粘性会话 + 切换账号
	s.setModelRateLimitAndClearSession(p, info)

	return &handleModelRateLimitResult{
		Handled: true,
		SwitchError: &AntigravityAccountSwitchError{
			OriginalAccountID: p.account.ID,
			RateLimitedModel:  info.ModelName,
			IsStickySession:   p.isStickySession,
		},
	}
}

// setModelRateLimitAndClearSession 设置模型限流并清除粘性会话
func (s *AntigravityGatewayService) setModelRateLimitAndClearSession(p *handleModelRateLimitParams, info *antigravitySmartRetryInfo) {
	resetAt := time.Now().Add(info.RetryDelay)
	log.Printf("%s status=%d model_rate_limited model=%s account=%d reset_in=%v",
		p.prefix, p.statusCode, info.ModelName, p.account.ID, info.RetryDelay)

	// 设置模型限流状态（数据库）
	if err := s.accountRepo.SetModelRateLimit(p.ctx, p.account.ID, info.ModelName, resetAt); err != nil {
		log.Printf("%s model_rate_limit_failed model=%s error=%v", p.prefix, info.ModelName, err)
	}

	// 立即更新 Redis 快照中账号的限流状态，避免并发请求重复选中
	s.updateAccountModelRateLimitInCache(p.ctx, p.account, info.ModelName, resetAt)

	// 清除粘性会话绑定
	if p.cache != nil && p.sessionHash != "" {
		_ = p.cache.DeleteSessionAccountID(p.ctx, p.groupID, p.sessionHash)
	}
}

// updateAccountModelRateLimitInCache 立即更新 Redis 中账号的模型限流状态
func (s *AntigravityGatewayService) updateAccountModelRateLimitInCache(ctx context.Context, account *Account, modelKey string, resetAt time.Time) {
	if s.schedulerSnapshot == nil || account == nil || modelKey == "" {
		return
	}

	// 更新账号对象的 Extra 字段
	if account.Extra == nil {
		account.Extra = make(map[string]any)
	}

	limits, _ := account.Extra["model_rate_limits"].(map[string]any)
	if limits == nil {
		limits = make(map[string]any)
		account.Extra["model_rate_limits"] = limits
	}

	limits[modelKey] = map[string]any{
		"rate_limited_at":     time.Now().UTC().Format(time.RFC3339),
		"rate_limit_reset_at": resetAt.UTC().Format(time.RFC3339),
	}

	// 更新 Redis 快照
	if err := s.schedulerSnapshot.UpdateAccountInCache(ctx, account); err != nil {
		log.Printf("[antigravity-Forward] cache_update_failed account=%d model=%s err=%v", account.ID, modelKey, err)
	}
}

func (s *AntigravityGatewayService) handleUpstreamError(
	ctx context.Context, prefix string, account *Account,
	statusCode int, headers http.Header, body []byte,
	requestedModel string,
	groupID int64, sessionHash string, isStickySession bool,
) *handleModelRateLimitResult {
	// 模型级限流处理（优先）
	result := s.handleModelRateLimit(&handleModelRateLimitParams{
		ctx:             ctx,
		prefix:          prefix,
		account:         account,
		statusCode:      statusCode,
		body:            body,
		cache:           s.cache,
		groupID:         groupID,
		sessionHash:     sessionHash,
		isStickySession: isStickySession,
	})
	if result.Handled {
		return result
	}

	// 503 仅处理模型限流（MODEL_CAPACITY_EXHAUSTED），非模型限流不做额外处理
	// 避免将普通的 503 错误误判为账号问题
	if statusCode == 503 {
		return nil
	}

	// 429：尝试解析模型级限流，解析失败时兜底为账号级限流
	if statusCode == 429 {
		if logBody, maxBytes := s.getLogConfig(); logBody {
			log.Printf("[Antigravity-Debug] 429 response body: %s", truncateString(string(body), maxBytes))
		}

		resetAt := ParseGeminiRateLimitResetTime(body)
		defaultDur := s.getDefaultRateLimitDuration()

		// 尝试解析模型 key 并设置模型级限流
		modelKey := resolveAntigravityModelKey(requestedModel)
		if modelKey != "" {
			ra := s.resolveResetTime(resetAt, defaultDur)
			if err := s.accountRepo.SetModelRateLimit(ctx, account.ID, modelKey, ra); err != nil {
				log.Printf("%s status=429 model_rate_limit_set_failed model=%s error=%v", prefix, modelKey, err)
			} else {
				log.Printf("%s status=429 model_rate_limited model=%s account=%d reset_at=%v reset_in=%v",
					prefix, modelKey, account.ID, ra.Format("15:04:05"), time.Until(ra).Truncate(time.Second))
				s.updateAccountModelRateLimitInCache(ctx, account, modelKey, ra)
			}
			return nil
		}

		// 无法解析模型 key，兜底为账号级限流
		ra := s.resolveResetTime(resetAt, defaultDur)
		log.Printf("%s status=429 rate_limited account=%d reset_at=%v reset_in=%v (fallback)",
			prefix, account.ID, ra.Format("15:04:05"), time.Until(ra).Truncate(time.Second))
		if err := s.accountRepo.SetRateLimited(ctx, account.ID, ra); err != nil {
			log.Printf("%s status=429 rate_limit_set_failed account=%d error=%v", prefix, account.ID, err)
		}
		return nil
	}
	// 其他错误码继续使用 rateLimitService
	if s.rateLimitService == nil {
		return nil
	}
	shouldDisable := s.rateLimitService.HandleUpstreamError(ctx, account, statusCode, headers, body)
	if shouldDisable {
		log.Printf("%s status=%d marked_error", prefix, statusCode)
	}
	return nil
}

// getDefaultRateLimitDuration 获取默认限流时间
func (s *AntigravityGatewayService) getDefaultRateLimitDuration() time.Duration {
	defaultDur := antigravityDefaultRateLimitDuration
	if s.settingService != nil && s.settingService.cfg != nil && s.settingService.cfg.Gateway.AntigravityFallbackCooldownMinutes > 0 {
		defaultDur = time.Duration(s.settingService.cfg.Gateway.AntigravityFallbackCooldownMinutes) * time.Minute
	}
	if override, ok := antigravityFallbackCooldownSeconds(); ok {
		defaultDur = override
	}
	return defaultDur
}

// resolveResetTime 根据解析的重置时间或默认时长计算重置时间点
func (s *AntigravityGatewayService) resolveResetTime(resetAt *int64, defaultDur time.Duration) time.Time {
	if resetAt != nil {
		return time.Unix(*resetAt, 0)
	}
	return time.Now().Add(defaultDur)
}

type antigravityStreamResult struct {
	usage            *ClaudeUsage
	firstTokenMs     *int
	clientDisconnect bool // 客户端是否在流式传输过程中断开
}

// antigravityClientWriter 封装流式响应的客户端写入，自动检测断开并标记。
// 断开后所有写入操作变为 no-op，调用方通过 Disconnected() 判断是否继续 drain 上游。
type antigravityClientWriter struct {
	w            gin.ResponseWriter
	flusher      http.Flusher
	disconnected bool
	prefix       string // 日志前缀，标识来源方法
}

func newAntigravityClientWriter(w gin.ResponseWriter, flusher http.Flusher, prefix string) *antigravityClientWriter {
	return &antigravityClientWriter{w: w, flusher: flusher, prefix: prefix}
}

// Write 写入数据到客户端，写入失败时标记断开并返回 false
func (cw *antigravityClientWriter) Write(p []byte) bool {
	if cw.disconnected {
		return false
	}
	if _, err := cw.w.Write(p); err != nil {
		cw.markDisconnected()
		return false
	}
	cw.flusher.Flush()
	return true
}

// Fprintf 格式化写入数据到客户端，写入失败时标记断开并返回 false
func (cw *antigravityClientWriter) Fprintf(format string, args ...any) bool {
	if cw.disconnected {
		return false
	}
	if _, err := fmt.Fprintf(cw.w, format, args...); err != nil {
		cw.markDisconnected()
		return false
	}
	cw.flusher.Flush()
	return true
}

func (cw *antigravityClientWriter) Disconnected() bool { return cw.disconnected }

func (cw *antigravityClientWriter) markDisconnected() {
	cw.disconnected = true
	log.Printf("Client disconnected during streaming (%s), continuing to drain upstream for billing", cw.prefix)
}

// handleStreamReadError 处理上游读取错误的通用逻辑。
// 返回 (clientDisconnect, handled)：handled=true 表示错误已处理，调用方应返回已收集的 usage。
func handleStreamReadError(err error, clientDisconnected bool, prefix string) (disconnect bool, handled bool) {
	if errors.Is(err, context.Canceled) || errors.Is(err, context.DeadlineExceeded) {
		log.Printf("Context canceled during streaming (%s), returning collected usage", prefix)
		return true, true
	}
	if clientDisconnected {
		log.Printf("Upstream read error after client disconnect (%s): %v, returning collected usage", prefix, err)
		return true, true
	}
	return false, false
}

func (s *AntigravityGatewayService) handleGeminiStreamingResponse(c *gin.Context, resp *http.Response, startTime time.Time) (*antigravityStreamResult, error) {
	c.Status(resp.StatusCode)
	c.Header("Cache-Control", "no-cache")
	c.Header("Connection", "keep-alive")
	c.Header("X-Accel-Buffering", "no")

	contentType := resp.Header.Get("Content-Type")
	if contentType == "" {
		contentType = "text/event-stream; charset=utf-8"
	}
	c.Header("Content-Type", contentType)

	flusher, ok := c.Writer.(http.Flusher)
	if !ok {
		return nil, errors.New("streaming not supported")
	}

	// 使用 Scanner 并限制单行大小，避免 ReadString 无上限导致 OOM
	scanner := bufio.NewScanner(resp.Body)
	maxLineSize := defaultMaxLineSize
	if s.settingService.cfg != nil && s.settingService.cfg.Gateway.MaxLineSize > 0 {
		maxLineSize = s.settingService.cfg.Gateway.MaxLineSize
	}
	scanner.Buffer(make([]byte, 64*1024), maxLineSize)
	usage := &ClaudeUsage{}
	var firstTokenMs *int

	type scanEvent struct {
		line string
		err  error
	}
	// 独立 goroutine 读取上游，避免读取阻塞影响超时处理
	events := make(chan scanEvent, 16)
	done := make(chan struct{})
	sendEvent := func(ev scanEvent) bool {
		select {
		case events <- ev:
			return true
		case <-done:
			return false
		}
	}
	var lastReadAt int64
	atomic.StoreInt64(&lastReadAt, time.Now().UnixNano())
	go func() {
		defer close(events)
		for scanner.Scan() {
			atomic.StoreInt64(&lastReadAt, time.Now().UnixNano())
			if !sendEvent(scanEvent{line: scanner.Text()}) {
				return
			}
		}
		if err := scanner.Err(); err != nil {
			_ = sendEvent(scanEvent{err: err})
		}
	}()
	defer close(done)

	// 上游数据间隔超时保护（防止上游挂起长期占用连接）
	streamInterval := time.Duration(0)
	if s.settingService.cfg != nil && s.settingService.cfg.Gateway.StreamDataIntervalTimeout > 0 {
		streamInterval = time.Duration(s.settingService.cfg.Gateway.StreamDataIntervalTimeout) * time.Second
	}
	var intervalTicker *time.Ticker
	if streamInterval > 0 {
		intervalTicker = time.NewTicker(streamInterval)
		defer intervalTicker.Stop()
	}
	var intervalCh <-chan time.Time
	if intervalTicker != nil {
		intervalCh = intervalTicker.C
	}

	cw := newAntigravityClientWriter(c.Writer, flusher, "antigravity gemini")

	// 仅发送一次错误事件，避免多次写入导致协议混乱
	errorEventSent := false
	sendErrorEvent := func(reason string) {
		if errorEventSent || cw.Disconnected() {
			return
		}
		errorEventSent = true
		_, _ = fmt.Fprintf(c.Writer, "event: error\ndata: {\"error\":\"%s\"}\n\n", reason)
		flusher.Flush()
	}

	for {
		select {
		case ev, ok := <-events:
			if !ok {
				return &antigravityStreamResult{usage: usage, firstTokenMs: firstTokenMs, clientDisconnect: cw.Disconnected()}, nil
			}
			if ev.err != nil {
				if disconnect, handled := handleStreamReadError(ev.err, cw.Disconnected(), "antigravity gemini"); handled {
					return &antigravityStreamResult{usage: usage, firstTokenMs: firstTokenMs, clientDisconnect: disconnect}, nil
				}
				if errors.Is(ev.err, bufio.ErrTooLong) {
					log.Printf("SSE line too long (antigravity): max_size=%d error=%v", maxLineSize, ev.err)
					sendErrorEvent("response_too_large")
					return &antigravityStreamResult{usage: usage, firstTokenMs: firstTokenMs}, ev.err
				}
				sendErrorEvent("stream_read_error")
				return nil, ev.err
			}

			line := ev.line
			trimmed := strings.TrimRight(line, "\r\n")
			if strings.HasPrefix(trimmed, "data:") {
				payload := strings.TrimSpace(strings.TrimPrefix(trimmed, "data:"))
				if payload == "" || payload == "[DONE]" {
					cw.Fprintf("%s\n", line)
					continue
				}

				// 解包 v1internal 响应
				inner, parseErr := s.unwrapV1InternalResponse([]byte(payload))
				if parseErr == nil && inner != nil {
					payload = string(inner)
				}

				// 解析 usage
				var parsed map[string]any
				if json.Unmarshal(inner, &parsed) == nil {
					if u := extractGeminiUsage(parsed); u != nil {
						usage = u
					}
					// Check for MALFORMED_FUNCTION_CALL
					if candidates, ok := parsed["candidates"].([]any); ok && len(candidates) > 0 {
						if cand, ok := candidates[0].(map[string]any); ok {
							if fr, ok := cand["finishReason"].(string); ok && fr == "MALFORMED_FUNCTION_CALL" {
								log.Printf("[Antigravity] MALFORMED_FUNCTION_CALL detected in forward stream")
								if content, ok := cand["content"]; ok {
									if b, err := json.Marshal(content); err == nil {
										log.Printf("[Antigravity] Malformed content: %s", string(b))
									}
								}
							}
						}
					}
				}

				if firstTokenMs == nil {
					ms := int(time.Since(startTime).Milliseconds())
					firstTokenMs = &ms
				}

				cw.Fprintf("data: %s\n\n", payload)
				continue
			}

			cw.Fprintf("%s\n", line)

		case <-intervalCh:
			lastRead := time.Unix(0, atomic.LoadInt64(&lastReadAt))
			if time.Since(lastRead) < streamInterval {
				continue
			}
			if cw.Disconnected() {
				log.Printf("Upstream timeout after client disconnect (antigravity gemini), returning collected usage")
				return &antigravityStreamResult{usage: usage, firstTokenMs: firstTokenMs, clientDisconnect: true}, nil
			}
			log.Printf("Stream data interval timeout (antigravity)")
			sendErrorEvent("stream_timeout")
			return &antigravityStreamResult{usage: usage, firstTokenMs: firstTokenMs}, fmt.Errorf("stream data interval timeout")
		}
	}
}

// handleGeminiStreamToNonStreaming 读取上游流式响应，合并为非流式响应返回给客户端
// Gemini 流式响应是增量的，需要累积所有 chunk 的内容
func (s *AntigravityGatewayService) handleGeminiStreamToNonStreaming(c *gin.Context, resp *http.Response, startTime time.Time) (*antigravityStreamResult, error) {
	scanner := bufio.NewScanner(resp.Body)
	maxLineSize := defaultMaxLineSize
	if s.settingService.cfg != nil && s.settingService.cfg.Gateway.MaxLineSize > 0 {
		maxLineSize = s.settingService.cfg.Gateway.MaxLineSize
	}
	scanner.Buffer(make([]byte, 64*1024), maxLineSize)

	usage := &ClaudeUsage{}
	var firstTokenMs *int
	var last map[string]any
	var lastWithParts map[string]any
	var collectedImageParts []map[string]any // 收集所有包含图片的 parts
	var collectedTextParts []string          // 收集所有文本片段

	type scanEvent struct {
		line string
		err  error
	}

	// 独立 goroutine 读取上游，避免读取阻塞影响超时处理
	events := make(chan scanEvent, 16)
	done := make(chan struct{})
	sendEvent := func(ev scanEvent) bool {
		select {
		case events <- ev:
			return true
		case <-done:
			return false
		}
	}

	var lastReadAt int64
	atomic.StoreInt64(&lastReadAt, time.Now().UnixNano())
	go func() {
		defer close(events)
		for scanner.Scan() {
			atomic.StoreInt64(&lastReadAt, time.Now().UnixNano())
			if !sendEvent(scanEvent{line: scanner.Text()}) {
				return
			}
		}
		if err := scanner.Err(); err != nil {
			_ = sendEvent(scanEvent{err: err})
		}
	}()
	defer close(done)

	// 上游数据间隔超时保护（防止上游挂起长期占用连接）
	streamInterval := time.Duration(0)
	if s.settingService.cfg != nil && s.settingService.cfg.Gateway.StreamDataIntervalTimeout > 0 {
		streamInterval = time.Duration(s.settingService.cfg.Gateway.StreamDataIntervalTimeout) * time.Second
	}
	var intervalTicker *time.Ticker
	if streamInterval > 0 {
		intervalTicker = time.NewTicker(streamInterval)
		defer intervalTicker.Stop()
	}
	var intervalCh <-chan time.Time
	if intervalTicker != nil {
		intervalCh = intervalTicker.C
	}

	for {
		select {
		case ev, ok := <-events:
			if !ok {
				// 流结束，返回收集的响应
				goto returnResponse
			}
			if ev.err != nil {
				if errors.Is(ev.err, bufio.ErrTooLong) {
					log.Printf("SSE line too long (antigravity non-stream): max_size=%d error=%v", maxLineSize, ev.err)
				}
				return nil, ev.err
			}

			line := ev.line
			trimmed := strings.TrimRight(line, "\r\n")

			if !strings.HasPrefix(trimmed, "data:") {
				continue
			}

			payload := strings.TrimSpace(strings.TrimPrefix(trimmed, "data:"))
			if payload == "" || payload == "[DONE]" {
				continue
			}

			// 解包 v1internal 响应
			inner, parseErr := s.unwrapV1InternalResponse([]byte(payload))
			if parseErr != nil {
				continue
			}

			var parsed map[string]any
			if err := json.Unmarshal(inner, &parsed); err != nil {
				continue
			}

			// 记录首 token 时间
			if firstTokenMs == nil {
				ms := int(time.Since(startTime).Milliseconds())
				firstTokenMs = &ms
			}

			last = parsed

			// 提取 usage
			if u := extractGeminiUsage(parsed); u != nil {
				usage = u
			}

			// Check for MALFORMED_FUNCTION_CALL
			if candidates, ok := parsed["candidates"].([]any); ok && len(candidates) > 0 {
				if cand, ok := candidates[0].(map[string]any); ok {
					if fr, ok := cand["finishReason"].(string); ok && fr == "MALFORMED_FUNCTION_CALL" {
						log.Printf("[Antigravity] MALFORMED_FUNCTION_CALL detected in forward non-stream collect")
						if content, ok := cand["content"]; ok {
							if b, err := json.Marshal(content); err == nil {
								log.Printf("[Antigravity] Malformed content: %s", string(b))
							}
						}
					}
				}
			}

			// 保留最后一个有 parts 的响应
			if parts := extractGeminiParts(parsed); len(parts) > 0 {
				lastWithParts = parsed
				// 收集包含图片和文本的 parts
				for _, part := range parts {
					if inlineData, ok := part["inlineData"].(map[string]any); ok {
						collectedImageParts = append(collectedImageParts, part)
						_ = inlineData // 避免 unused 警告
					}
					if text, ok := part["text"].(string); ok && text != "" {
						collectedTextParts = append(collectedTextParts, text)
					}
				}
			}

		case <-intervalCh:
			lastRead := time.Unix(0, atomic.LoadInt64(&lastReadAt))
			if time.Since(lastRead) < streamInterval {
				continue
			}
			log.Printf("Stream data interval timeout (antigravity non-stream)")
			return nil, fmt.Errorf("stream data interval timeout")
		}
	}

returnResponse:
	// 选择最后一个有效响应
	finalResponse := pickGeminiCollectResult(last, lastWithParts)

	// 处理空响应情况
	if last == nil && lastWithParts == nil {
		log.Printf("[antigravity-Forward] warning: empty stream response, no valid chunks received")
	}

	// 如果收集到了图片 parts，需要合并到最终响应中
	if len(collectedImageParts) > 0 {
		finalResponse = mergeImagePartsToResponse(finalResponse, collectedImageParts)
	}

	// 如果收集到了文本，需要合并到最终响应中
	if len(collectedTextParts) > 0 {
		finalResponse = mergeTextPartsToResponse(finalResponse, collectedTextParts)
	}

	respBody, err := json.Marshal(finalResponse)
	if err != nil {
		return nil, fmt.Errorf("failed to marshal response: %w", err)
	}
	c.Data(http.StatusOK, "application/json", respBody)

	return &antigravityStreamResult{usage: usage, firstTokenMs: firstTokenMs}, nil
}

// getOrCreateGeminiParts 获取 Gemini 响应的 parts 结构，返回深拷贝和更新回调
func getOrCreateGeminiParts(response map[string]any) (result map[string]any, existingParts []any, setParts func([]any)) {
	// 深拷贝 response
	result = make(map[string]any)
	for k, v := range response {
		result[k] = v
	}

	// 获取或创建 candidates
	candidates, ok := result["candidates"].([]any)
	if !ok || len(candidates) == 0 {
		candidates = []any{map[string]any{}}
	}

	// 获取第一个 candidate
	candidate, ok := candidates[0].(map[string]any)
	if !ok {
		candidate = make(map[string]any)
		candidates[0] = candidate
	}

	// 获取或创建 content
	content, ok := candidate["content"].(map[string]any)
	if !ok {
		content = map[string]any{"role": "model"}
		candidate["content"] = content
	}

	// 获取现有 parts
	existingParts, ok = content["parts"].([]any)
	if !ok {
		existingParts = []any{}
	}

	// 返回更新回调
	setParts = func(newParts []any) {
		content["parts"] = newParts
		result["candidates"] = candidates
	}

	return result, existingParts, setParts
}

// mergeCollectedPartsToResponse 将收集的所有 parts 合并到 Gemini 响应中
// 这个函数会合并所有类型的 parts：text、thinking、functionCall、inlineData 等
// 保持原始顺序，只合并连续的普通 text parts
func mergeCollectedPartsToResponse(response map[string]any, collectedParts []map[string]any) map[string]any {
	if len(collectedParts) == 0 {
		return response
	}

	result, _, setParts := getOrCreateGeminiParts(response)

	// 合并策略：
	// 1. 保持原始顺序
	// 2. 连续的普通 text parts 合并为一个
	// 3. thinking、functionCall、inlineData 等保持原样
	var mergedParts []any
	var textBuffer strings.Builder

	flushTextBuffer := func() {
		if textBuffer.Len() > 0 {
			mergedParts = append(mergedParts, map[string]any{
				"text": textBuffer.String(),
			})
			textBuffer.Reset()
		}
	}

	for _, part := range collectedParts {
		// 检查是否是普通 text part
		if text, ok := part["text"].(string); ok {
			// 检查是否有 thought 标记
			if thought, _ := part["thought"].(bool); thought {
				// thinking part，先刷新 text buffer，然后保留原样
				flushTextBuffer()
				mergedParts = append(mergedParts, part)
			} else {
				// 普通 text，累积到 buffer
				_, _ = textBuffer.WriteString(text)
			}
		} else {
			// 非 text part（functionCall、inlineData 等），先刷新 text buffer，然后保留原样
			flushTextBuffer()
			mergedParts = append(mergedParts, part)
		}
	}

	// 刷新剩余的 text
	flushTextBuffer()

	setParts(mergedParts)
	return result
}

// mergeImagePartsToResponse 将收集到的图片 parts 合并到 Gemini 响应中
func mergeImagePartsToResponse(response map[string]any, imageParts []map[string]any) map[string]any {
	if len(imageParts) == 0 {
		return response
	}

	result, existingParts, setParts := getOrCreateGeminiParts(response)

	// 检查现有 parts 中是否已经有图片
	for _, p := range existingParts {
		if pm, ok := p.(map[string]any); ok {
			if _, hasInline := pm["inlineData"]; hasInline {
				return result // 已有图片，不重复添加
			}
		}
	}

	// 添加收集到的图片 parts
	for _, imgPart := range imageParts {
		existingParts = append(existingParts, imgPart)
	}
	setParts(existingParts)
	return result
}

// mergeTextPartsToResponse 将收集到的文本合并到 Gemini 响应中
func mergeTextPartsToResponse(response map[string]any, textParts []string) map[string]any {
	if len(textParts) == 0 {
		return response
	}

	mergedText := strings.Join(textParts, "")
	result, existingParts, setParts := getOrCreateGeminiParts(response)

	// 查找并更新第一个 text part，或创建新的
	newParts := make([]any, 0, len(existingParts)+1)
	textUpdated := false

	for _, p := range existingParts {
		pm, ok := p.(map[string]any)
		if !ok {
			newParts = append(newParts, p)
			continue
		}
		if _, hasText := pm["text"]; hasText && !textUpdated {
			// 用累积的文本替换
			newPart := make(map[string]any)
			for k, v := range pm {
				newPart[k] = v
			}
			newPart["text"] = mergedText
			newParts = append(newParts, newPart)
			textUpdated = true
		} else {
			newParts = append(newParts, pm)
		}
	}

	if !textUpdated {
		newParts = append([]any{map[string]any{"text": mergedText}}, newParts...)
	}

	setParts(newParts)
	return result
}

func (s *AntigravityGatewayService) writeClaudeError(c *gin.Context, status int, errType, message string) error {
	c.JSON(status, gin.H{
		"type":  "error",
		"error": gin.H{"type": errType, "message": message},
	})
	return fmt.Errorf("%s", message)
}

// WriteMappedClaudeError 导出版本，供 handler 层使用（如 fallback 错误处理）
func (s *AntigravityGatewayService) WriteMappedClaudeError(c *gin.Context, account *Account, upstreamStatus int, upstreamRequestID string, body []byte) error {
	return s.writeMappedClaudeError(c, account, upstreamStatus, upstreamRequestID, body)
}

func (s *AntigravityGatewayService) writeMappedClaudeError(c *gin.Context, account *Account, upstreamStatus int, upstreamRequestID string, body []byte) error {
	upstreamMsg := strings.TrimSpace(extractUpstreamErrorMessage(body))
	upstreamMsg = sanitizeUpstreamErrorMessage(upstreamMsg)
	logBody, maxBytes := s.getLogConfig()
	upstreamDetail := s.getUpstreamErrorDetail(body)
	setOpsUpstreamError(c, upstreamStatus, upstreamMsg, upstreamDetail)
	appendOpsUpstreamError(c, OpsUpstreamErrorEvent{
		Platform:           account.Platform,
		AccountID:          account.ID,
		AccountName:        account.Name,
		UpstreamStatusCode: upstreamStatus,
		UpstreamRequestID:  upstreamRequestID,
		Kind:               "http_error",
		Message:            upstreamMsg,
		Detail:             upstreamDetail,
	})

	// 记录上游错误详情便于排障（可选：由配置控制；不回显到客户端）
	if logBody {
		log.Printf("[antigravity-Forward] upstream_error status=%d body=%s", upstreamStatus, truncateForLog(body, maxBytes))
	}

	var statusCode int
	var errType, errMsg string

	switch upstreamStatus {
	case 400:
		statusCode = http.StatusBadRequest
		errType = "invalid_request_error"
		errMsg = getPassthroughOrDefault(upstreamMsg, "Invalid request")
	case 401:
		statusCode = http.StatusBadGateway
		errType = "authentication_error"
		errMsg = "Upstream authentication failed"
	case 403:
		statusCode = http.StatusBadGateway
		errType = "permission_error"
		errMsg = "Upstream access forbidden"
	case 429:
		statusCode = http.StatusTooManyRequests
		errType = "rate_limit_error"
		errMsg = "Upstream rate limit exceeded"
	case 529:
		statusCode = http.StatusServiceUnavailable
		errType = "overloaded_error"
		errMsg = "Upstream service overloaded"
	default:
		statusCode = http.StatusBadGateway
		errType = "upstream_error"
		errMsg = "Upstream request failed"
	}

	c.JSON(statusCode, gin.H{
		"type":  "error",
		"error": gin.H{"type": errType, "message": errMsg},
	})
	if upstreamMsg == "" {
		return fmt.Errorf("upstream error: %d", upstreamStatus)
	}
	return fmt.Errorf("upstream error: %d message=%s", upstreamStatus, upstreamMsg)
}

func (s *AntigravityGatewayService) writeGoogleError(c *gin.Context, status int, message string) error {
	statusStr := "UNKNOWN"
	switch status {
	case 400:
		statusStr = "INVALID_ARGUMENT"
	case 404:
		statusStr = "NOT_FOUND"
	case 429:
		statusStr = "RESOURCE_EXHAUSTED"
	case 500:
		statusStr = "INTERNAL"
	case 502, 503:
		statusStr = "UNAVAILABLE"
	}

	c.JSON(status, gin.H{
		"error": gin.H{
			"code":    status,
			"message": message,
			"status":  statusStr,
		},
	})
	return fmt.Errorf("%s", message)
}

// handleClaudeStreamToNonStreaming 收集上游流式响应，转换为 Claude 非流式格式返回
// 用于处理客户端非流式请求但上游只支持流式的情况
func (s *AntigravityGatewayService) handleClaudeStreamToNonStreaming(c *gin.Context, resp *http.Response, startTime time.Time, originalModel string) (*antigravityStreamResult, error) {
	scanner := bufio.NewScanner(resp.Body)
	maxLineSize := defaultMaxLineSize
	if s.settingService.cfg != nil && s.settingService.cfg.Gateway.MaxLineSize > 0 {
		maxLineSize = s.settingService.cfg.Gateway.MaxLineSize
	}
	scanner.Buffer(make([]byte, 64*1024), maxLineSize)

	var firstTokenMs *int
	var last map[string]any
	var lastWithParts map[string]any
	var collectedParts []map[string]any // 收集所有 parts（包括 text、thinking、functionCall、inlineData 等）

	type scanEvent struct {
		line string
		err  error
	}

	// 独立 goroutine 读取上游，避免读取阻塞影响超时处理
	events := make(chan scanEvent, 16)
	done := make(chan struct{})
	sendEvent := func(ev scanEvent) bool {
		select {
		case events <- ev:
			return true
		case <-done:
			return false
		}
	}

	var lastReadAt int64
	atomic.StoreInt64(&lastReadAt, time.Now().UnixNano())
	go func() {
		defer close(events)
		for scanner.Scan() {
			atomic.StoreInt64(&lastReadAt, time.Now().UnixNano())
			if !sendEvent(scanEvent{line: scanner.Text()}) {
				return
			}
		}
		if err := scanner.Err(); err != nil {
			_ = sendEvent(scanEvent{err: err})
		}
	}()
	defer close(done)

	// 上游数据间隔超时保护（防止上游挂起长期占用连接）
	streamInterval := time.Duration(0)
	if s.settingService.cfg != nil && s.settingService.cfg.Gateway.StreamDataIntervalTimeout > 0 {
		streamInterval = time.Duration(s.settingService.cfg.Gateway.StreamDataIntervalTimeout) * time.Second
	}
	var intervalTicker *time.Ticker
	if streamInterval > 0 {
		intervalTicker = time.NewTicker(streamInterval)
		defer intervalTicker.Stop()
	}
	var intervalCh <-chan time.Time
	if intervalTicker != nil {
		intervalCh = intervalTicker.C
	}

	for {
		select {
		case ev, ok := <-events:
			if !ok {
				// 流结束，转换并返回响应
				goto returnResponse
			}
			if ev.err != nil {
				if errors.Is(ev.err, bufio.ErrTooLong) {
					log.Printf("SSE line too long (antigravity claude non-stream): max_size=%d error=%v", maxLineSize, ev.err)
				}
				return nil, ev.err
			}

			line := ev.line
			trimmed := strings.TrimRight(line, "\r\n")

			if !strings.HasPrefix(trimmed, "data:") {
				continue
			}

			payload := strings.TrimSpace(strings.TrimPrefix(trimmed, "data:"))
			if payload == "" || payload == "[DONE]" {
				continue
			}

			// 解包 v1internal 响应
			inner, parseErr := s.unwrapV1InternalResponse([]byte(payload))
			if parseErr != nil {
				continue
			}

			var parsed map[string]any
			if err := json.Unmarshal(inner, &parsed); err != nil {
				continue
			}

			// 记录首 token 时间
			if firstTokenMs == nil {
				ms := int(time.Since(startTime).Milliseconds())
				firstTokenMs = &ms
			}

			last = parsed

			// 保留最后一个有 parts 的响应，并收集所有 parts
			if parts := extractGeminiParts(parsed); len(parts) > 0 {
				lastWithParts = parsed

				// 收集所有 parts（text、thinking、functionCall、inlineData 等）
				collectedParts = append(collectedParts, parts...)
			}

		case <-intervalCh:
			lastRead := time.Unix(0, atomic.LoadInt64(&lastReadAt))
			if time.Since(lastRead) < streamInterval {
				continue
			}
			log.Printf("Stream data interval timeout (antigravity claude non-stream)")
			return nil, fmt.Errorf("stream data interval timeout")
		}
	}

returnResponse:
	// 选择最后一个有效响应
	finalResponse := pickGeminiCollectResult(last, lastWithParts)

	// 处理空响应情况
	if last == nil && lastWithParts == nil {
		log.Printf("[antigravity-Forward] warning: empty stream response, no valid chunks received")
		return nil, s.writeClaudeError(c, http.StatusBadGateway, "upstream_error", "Empty response from upstream")
	}

	// 将收集的所有 parts 合并到最终响应中
	if len(collectedParts) > 0 {
		finalResponse = mergeCollectedPartsToResponse(finalResponse, collectedParts)
	}

	// 序列化为 JSON（Gemini 格式）
	geminiBody, err := json.Marshal(finalResponse)
	if err != nil {
		return nil, fmt.Errorf("failed to marshal gemini response: %w", err)
	}

	// 转换 Gemini 响应为 Claude 格式
	claudeResp, agUsage, err := antigravity.TransformGeminiToClaude(geminiBody, originalModel)
	if err != nil {
		log.Printf("[antigravity-Forward] transform_error error=%v body=%s", err, string(geminiBody))
		return nil, s.writeClaudeError(c, http.StatusBadGateway, "upstream_error", "Failed to parse upstream response")
	}

	c.Data(http.StatusOK, "application/json", claudeResp)

	// 转换为 service.ClaudeUsage
	usage := &ClaudeUsage{
		InputTokens:              agUsage.InputTokens,
		OutputTokens:             agUsage.OutputTokens,
		CacheCreationInputTokens: agUsage.CacheCreationInputTokens,
		CacheReadInputTokens:     agUsage.CacheReadInputTokens,
	}

	return &antigravityStreamResult{usage: usage, firstTokenMs: firstTokenMs}, nil
}

// handleClaudeStreamingResponse 处理 Claude 流式响应（Gemini SSE → Claude SSE 转换）
func (s *AntigravityGatewayService) handleClaudeStreamingResponse(c *gin.Context, resp *http.Response, startTime time.Time, originalModel string) (*antigravityStreamResult, error) {
	c.Header("Content-Type", "text/event-stream")
	c.Header("Cache-Control", "no-cache")
	c.Header("Connection", "keep-alive")
	c.Header("X-Accel-Buffering", "no")
	c.Status(http.StatusOK)

	flusher, ok := c.Writer.(http.Flusher)
	if !ok {
		return nil, errors.New("streaming not supported")
	}

	processor := antigravity.NewStreamingProcessor(originalModel)
	var firstTokenMs *int
	// 使用 Scanner 并限制单行大小，避免 ReadString 无上限导致 OOM
	scanner := bufio.NewScanner(resp.Body)
	maxLineSize := defaultMaxLineSize
	if s.settingService.cfg != nil && s.settingService.cfg.Gateway.MaxLineSize > 0 {
		maxLineSize = s.settingService.cfg.Gateway.MaxLineSize
	}
	scanner.Buffer(make([]byte, 64*1024), maxLineSize)

	// 辅助函数：转换 antigravity.ClaudeUsage 到 service.ClaudeUsage
	convertUsage := func(agUsage *antigravity.ClaudeUsage) *ClaudeUsage {
		if agUsage == nil {
			return &ClaudeUsage{}
		}
		return &ClaudeUsage{
			InputTokens:              agUsage.InputTokens,
			OutputTokens:             agUsage.OutputTokens,
			CacheCreationInputTokens: agUsage.CacheCreationInputTokens,
			CacheReadInputTokens:     agUsage.CacheReadInputTokens,
		}
	}

	type scanEvent struct {
		line string
		err  error
	}
	// 独立 goroutine 读取上游，避免读取阻塞影响超时处理
	events := make(chan scanEvent, 16)
	done := make(chan struct{})
	sendEvent := func(ev scanEvent) bool {
		select {
		case events <- ev:
			return true
		case <-done:
			return false
		}
	}
	var lastReadAt int64
	atomic.StoreInt64(&lastReadAt, time.Now().UnixNano())
	go func() {
		defer close(events)
		for scanner.Scan() {
			atomic.StoreInt64(&lastReadAt, time.Now().UnixNano())
			if !sendEvent(scanEvent{line: scanner.Text()}) {
				return
			}
		}
		if err := scanner.Err(); err != nil {
			_ = sendEvent(scanEvent{err: err})
		}
	}()
	defer close(done)

	streamInterval := time.Duration(0)
	if s.settingService.cfg != nil && s.settingService.cfg.Gateway.StreamDataIntervalTimeout > 0 {
		streamInterval = time.Duration(s.settingService.cfg.Gateway.StreamDataIntervalTimeout) * time.Second
	}
	var intervalTicker *time.Ticker
	if streamInterval > 0 {
		intervalTicker = time.NewTicker(streamInterval)
		defer intervalTicker.Stop()
	}
	var intervalCh <-chan time.Time
	if intervalTicker != nil {
		intervalCh = intervalTicker.C
	}

	cw := newAntigravityClientWriter(c.Writer, flusher, "antigravity claude")

	// 仅发送一次错误事件，避免多次写入导致协议混乱
	errorEventSent := false
	sendErrorEvent := func(reason string) {
		if errorEventSent || cw.Disconnected() {
			return
		}
		errorEventSent = true
		_, _ = fmt.Fprintf(c.Writer, "event: error\ndata: {\"error\":\"%s\"}\n\n", reason)
		flusher.Flush()
	}

	// finishUsage 是获取 processor 最终 usage 的辅助函数
	finishUsage := func() *ClaudeUsage {
		_, agUsage := processor.Finish()
		return convertUsage(agUsage)
	}

	for {
		select {
		case ev, ok := <-events:
			if !ok {
				// 上游完成，发送结束事件
				finalEvents, agUsage := processor.Finish()
				if len(finalEvents) > 0 {
					cw.Write(finalEvents)
				}
				return &antigravityStreamResult{usage: convertUsage(agUsage), firstTokenMs: firstTokenMs, clientDisconnect: cw.Disconnected()}, nil
			}
			if ev.err != nil {
				if disconnect, handled := handleStreamReadError(ev.err, cw.Disconnected(), "antigravity claude"); handled {
					return &antigravityStreamResult{usage: finishUsage(), firstTokenMs: firstTokenMs, clientDisconnect: disconnect}, nil
				}
				if errors.Is(ev.err, bufio.ErrTooLong) {
					log.Printf("SSE line too long (antigravity): max_size=%d error=%v", maxLineSize, ev.err)
					sendErrorEvent("response_too_large")
					return &antigravityStreamResult{usage: convertUsage(nil), firstTokenMs: firstTokenMs}, ev.err
				}
				sendErrorEvent("stream_read_error")
				return nil, fmt.Errorf("stream read error: %w", ev.err)
			}

			// 处理 SSE 行，转换为 Claude 格式
			claudeEvents := processor.ProcessLine(strings.TrimRight(ev.line, "\r\n"))
			if len(claudeEvents) > 0 {
				if firstTokenMs == nil {
					ms := int(time.Since(startTime).Milliseconds())
					firstTokenMs = &ms
				}
				cw.Write(claudeEvents)
			}

		case <-intervalCh:
			lastRead := time.Unix(0, atomic.LoadInt64(&lastReadAt))
			if time.Since(lastRead) < streamInterval {
				continue
			}
			if cw.Disconnected() {
				log.Printf("Upstream timeout after client disconnect (antigravity claude), returning collected usage")
				return &antigravityStreamResult{usage: finishUsage(), firstTokenMs: firstTokenMs, clientDisconnect: true}, nil
			}
			log.Printf("Stream data interval timeout (antigravity)")
			sendErrorEvent("stream_timeout")
			return &antigravityStreamResult{usage: convertUsage(nil), firstTokenMs: firstTokenMs}, fmt.Errorf("stream data interval timeout")
		}
	}
}

// extractImageSize 从 Gemini 请求中提取 image_size 参数
func (s *AntigravityGatewayService) extractImageSize(body []byte) string {
	var req antigravity.GeminiRequest
	if err := json.Unmarshal(body, &req); err != nil {
		return "2K" // 默认 2K
	}

	if req.GenerationConfig != nil && req.GenerationConfig.ImageConfig != nil {
		size := strings.ToUpper(strings.TrimSpace(req.GenerationConfig.ImageConfig.ImageSize))
		if size == "1K" || size == "2K" || size == "4K" {
			return size
		}
	}

	return "2K" // 默认 2K
}

// isImageGenerationModel 判断模型是否为图片生成模型
// 支持的模型：gemini-3-pro-image, gemini-3-pro-image-preview, gemini-2.5-flash-image 等
func isImageGenerationModel(model string) bool {
	modelLower := strings.ToLower(model)
	// 移除 models/ 前缀
	modelLower = strings.TrimPrefix(modelLower, "models/")

	// 精确匹配或前缀匹配
	return modelLower == "gemini-3-pro-image" ||
		modelLower == "gemini-3-pro-image-preview" ||
		strings.HasPrefix(modelLower, "gemini-3-pro-image-") ||
		modelLower == "gemini-2.5-flash-image" ||
		modelLower == "gemini-2.5-flash-image-preview" ||
		strings.HasPrefix(modelLower, "gemini-2.5-flash-image-")
}

// cleanGeminiRequest 清理 Gemini 请求体中的 Schema
func cleanGeminiRequest(body []byte) ([]byte, error) {
	var payload map[string]any
	if err := json.Unmarshal(body, &payload); err != nil {
		return nil, err
	}

	modified := false

	// 1. 清理 Tools
	if tools, ok := payload["tools"].([]any); ok && len(tools) > 0 {
		for _, t := range tools {
			toolMap, ok := t.(map[string]any)
			if !ok {
				continue
			}

			// function_declarations (snake_case) or functionDeclarations (camelCase)
			var funcs []any
			if f, ok := toolMap["functionDeclarations"].([]any); ok {
				funcs = f
			} else if f, ok := toolMap["function_declarations"].([]any); ok {
				funcs = f
			}

			if len(funcs) == 0 {
				continue
			}

			for _, f := range funcs {
				funcMap, ok := f.(map[string]any)
				if !ok {
					continue
				}

				if params, ok := funcMap["parameters"].(map[string]any); ok {
					antigravity.DeepCleanUndefined(params)
					cleaned := antigravity.CleanJSONSchema(params)
					funcMap["parameters"] = cleaned
					modified = true
				}
			}
		}
	}

	if !modified {
		return body, nil
	}

	return json.Marshal(payload)
}

// filterEmptyPartsFromGeminiRequest 过滤掉 parts 为空的消息
// Gemini API 不接受空 parts，需要在请求前过滤
func filterEmptyPartsFromGeminiRequest(body []byte) ([]byte, error) {
	var payload map[string]any
	if err := json.Unmarshal(body, &payload); err != nil {
		return nil, err
	}

	contents, ok := payload["contents"].([]any)
	if !ok || len(contents) == 0 {
		return body, nil
	}

	filtered := make([]any, 0, len(contents))
	modified := false

	for _, c := range contents {
		contentMap, ok := c.(map[string]any)
		if !ok {
			filtered = append(filtered, c)
			continue
		}

		parts, hasParts := contentMap["parts"]
		if !hasParts {
			filtered = append(filtered, c)
			continue
		}

		partsSlice, ok := parts.([]any)
		if !ok {
			filtered = append(filtered, c)
			continue
		}

		// 跳过 parts 为空数组的消息
		if len(partsSlice) == 0 {
			modified = true
			continue
		}

		filtered = append(filtered, c)
	}

	if !modified {
		return body, nil
	}

	payload["contents"] = filtered
	return json.Marshal(payload)
}

// ForwardUpstream 使用 base_url + /v1/messages + 双 header 认证透传上游 Claude 请求
func (s *AntigravityGatewayService) ForwardUpstream(ctx context.Context, c *gin.Context, account *Account, body []byte) (*ForwardResult, error) {
	startTime := time.Now()
	sessionID := getSessionID(c)
	prefix := logPrefix(sessionID, account.Name)

	// 获取上游配置
	baseURL := strings.TrimSpace(account.GetCredential("base_url"))
	apiKey := strings.TrimSpace(account.GetCredential("api_key"))
	if baseURL == "" || apiKey == "" {
		return nil, fmt.Errorf("upstream account missing base_url or api_key")
	}
	baseURL = strings.TrimSuffix(baseURL, "/")

	// 解析请求获取模型信息
	var claudeReq antigravity.ClaudeRequest
	if err := json.Unmarshal(body, &claudeReq); err != nil {
		return nil, fmt.Errorf("parse claude request: %w", err)
	}
	if strings.TrimSpace(claudeReq.Model) == "" {
		return nil, fmt.Errorf("missing model")
	}
	originalModel := claudeReq.Model
	billingModel := originalModel

	// 构建上游请求 URL
	upstreamURL := baseURL + "/v1/messages"

	// 创建请求
	req, err := http.NewRequestWithContext(ctx, http.MethodPost, upstreamURL, bytes.NewReader(body))
	if err != nil {
		return nil, fmt.Errorf("create upstream request: %w", err)
	}

	// 设置请求头
	req.Header.Set("Content-Type", "application/json")
	req.Header.Set("Authorization", "Bearer "+apiKey)
	req.Header.Set("x-api-key", apiKey) // Claude API 兼容

	// 透传 Claude 相关 headers
	if v := c.GetHeader("anthropic-version"); v != "" {
		req.Header.Set("anthropic-version", v)
	}
	if v := c.GetHeader("anthropic-beta"); v != "" {
		req.Header.Set("anthropic-beta", v)
	}

	// 代理 URL
	proxyURL := ""
	if account.ProxyID != nil && account.Proxy != nil {
		proxyURL = account.Proxy.URL()
	}

	// 发送请求
	resp, err := s.httpUpstream.Do(req, proxyURL, account.ID, account.Concurrency)
	if err != nil {
		log.Printf("%s upstream request failed: %v", prefix, err)
		return nil, fmt.Errorf("upstream request failed: %w", err)
	}
	defer func() { _ = resp.Body.Close() }()

	// 处理错误响应
	if resp.StatusCode >= 400 {
		respBody, _ := io.ReadAll(io.LimitReader(resp.Body, 2<<20))

		// 429 错误时标记账号限流
		if resp.StatusCode == http.StatusTooManyRequests {
			s.handleUpstreamError(ctx, prefix, account, resp.StatusCode, resp.Header, respBody, originalModel, 0, "", false)
		}

		// 透传上游错误
		c.Header("Content-Type", resp.Header.Get("Content-Type"))
		c.Status(resp.StatusCode)
		_, _ = c.Writer.Write(respBody)

		return &ForwardResult{
			Model: billingModel,
		}, nil
	}

	// 处理成功响应（流式/非流式）
	var usage *ClaudeUsage
	var firstTokenMs *int
	var clientDisconnect bool

	if claudeReq.Stream {
		// 流式响应：透传
		c.Header("Content-Type", "text/event-stream")
		c.Header("Cache-Control", "no-cache")
		c.Header("Connection", "keep-alive")
		c.Header("X-Accel-Buffering", "no")
		c.Status(http.StatusOK)

		streamRes := s.streamUpstreamResponse(c, resp, startTime)
		usage = streamRes.usage
		firstTokenMs = streamRes.firstTokenMs
		clientDisconnect = streamRes.clientDisconnect
	} else {
		// 非流式响应：直接透传
		respBody, err := io.ReadAll(resp.Body)
		if err != nil {
			return nil, fmt.Errorf("read upstream response: %w", err)
		}

		// 提取 usage
		usage = s.extractClaudeUsage(respBody)

		c.Header("Content-Type", resp.Header.Get("Content-Type"))
		c.Status(http.StatusOK)
		_, _ = c.Writer.Write(respBody)
	}

	// 构建计费结果
	duration := time.Since(startTime)
	log.Printf("%s status=success duration_ms=%d", prefix, duration.Milliseconds())

	return &ForwardResult{
		Model:            billingModel,
		Stream:           claudeReq.Stream,
		Duration:         duration,
		FirstTokenMs:     firstTokenMs,
		ClientDisconnect: clientDisconnect,
		Usage: ClaudeUsage{
			InputTokens:              usage.InputTokens,
			OutputTokens:             usage.OutputTokens,
			CacheReadInputTokens:     usage.CacheReadInputTokens,
			CacheCreationInputTokens: usage.CacheCreationInputTokens,
		},
	}, nil
}

// streamUpstreamResponse 透传上游 SSE 流并提取 Claude usage
func (s *AntigravityGatewayService) streamUpstreamResponse(c *gin.Context, resp *http.Response, startTime time.Time) *antigravityStreamResult {
	usage := &ClaudeUsage{}
	var firstTokenMs *int

	scanner := bufio.NewScanner(resp.Body)
	maxLineSize := defaultMaxLineSize
	if s.settingService.cfg != nil && s.settingService.cfg.Gateway.MaxLineSize > 0 {
		maxLineSize = s.settingService.cfg.Gateway.MaxLineSize
	}
	scanner.Buffer(make([]byte, 64*1024), maxLineSize)

	type scanEvent struct {
		line string
		err  error
	}
	events := make(chan scanEvent, 16)
	done := make(chan struct{})
	sendEvent := func(ev scanEvent) bool {
		select {
		case events <- ev:
			return true
		case <-done:
			return false
		}
	}
	var lastReadAt int64
	atomic.StoreInt64(&lastReadAt, time.Now().UnixNano())
	go func() {
		defer close(events)
		for scanner.Scan() {
			atomic.StoreInt64(&lastReadAt, time.Now().UnixNano())
			if !sendEvent(scanEvent{line: scanner.Text()}) {
				return
			}
		}
		if err := scanner.Err(); err != nil {
			_ = sendEvent(scanEvent{err: err})
		}
	}()
	defer close(done)

	streamInterval := time.Duration(0)
	if s.settingService.cfg != nil && s.settingService.cfg.Gateway.StreamDataIntervalTimeout > 0 {
		streamInterval = time.Duration(s.settingService.cfg.Gateway.StreamDataIntervalTimeout) * time.Second
	}
	var intervalTicker *time.Ticker
	if streamInterval > 0 {
		intervalTicker = time.NewTicker(streamInterval)
		defer intervalTicker.Stop()
	}
	var intervalCh <-chan time.Time
	if intervalTicker != nil {
		intervalCh = intervalTicker.C
	}

	flusher, _ := c.Writer.(http.Flusher)
	cw := newAntigravityClientWriter(c.Writer, flusher, "antigravity upstream")

	for {
		select {
		case ev, ok := <-events:
			if !ok {
				return &antigravityStreamResult{usage: usage, firstTokenMs: firstTokenMs, clientDisconnect: cw.Disconnected()}
			}
			if ev.err != nil {
				if disconnect, handled := handleStreamReadError(ev.err, cw.Disconnected(), "antigravity upstream"); handled {
					return &antigravityStreamResult{usage: usage, firstTokenMs: firstTokenMs, clientDisconnect: disconnect}
				}
				log.Printf("Stream read error (antigravity upstream): %v", ev.err)
				return &antigravityStreamResult{usage: usage, firstTokenMs: firstTokenMs}
			}

			line := ev.line

			// 记录首 token 时间
			if firstTokenMs == nil && len(line) > 0 {
				ms := int(time.Since(startTime).Milliseconds())
				firstTokenMs = &ms
			}

			// 尝试从 message_delta 或 message_stop 事件提取 usage
			s.extractSSEUsage(line, usage)

			// 透传行
			cw.Fprintf("%s\n", line)

		case <-intervalCh:
			lastRead := time.Unix(0, atomic.LoadInt64(&lastReadAt))
			if time.Since(lastRead) < streamInterval {
				continue
			}
			if cw.Disconnected() {
				log.Printf("Upstream timeout after client disconnect (antigravity upstream), returning collected usage")
				return &antigravityStreamResult{usage: usage, firstTokenMs: firstTokenMs, clientDisconnect: true}
			}
			log.Printf("Stream data interval timeout (antigravity upstream)")
			return &antigravityStreamResult{usage: usage, firstTokenMs: firstTokenMs}
		}
	}
}

// extractSSEUsage 从 SSE data 行中提取 Claude usage（用于流式透传场景）
func (s *AntigravityGatewayService) extractSSEUsage(line string, usage *ClaudeUsage) {
	if !strings.HasPrefix(line, "data: ") {
		return
	}
	dataStr := strings.TrimPrefix(line, "data: ")
	var event map[string]any
	if json.Unmarshal([]byte(dataStr), &event) != nil {
		return
	}
	u, ok := event["usage"].(map[string]any)
	if !ok {
		return
	}
	if v, ok := u["input_tokens"].(float64); ok && int(v) > 0 {
		usage.InputTokens = int(v)
	}
	if v, ok := u["output_tokens"].(float64); ok && int(v) > 0 {
		usage.OutputTokens = int(v)
	}
	if v, ok := u["cache_read_input_tokens"].(float64); ok && int(v) > 0 {
		usage.CacheReadInputTokens = int(v)
	}
	if v, ok := u["cache_creation_input_tokens"].(float64); ok && int(v) > 0 {
		usage.CacheCreationInputTokens = int(v)
	}
}

// extractClaudeUsage 从非流式 Claude 响应提取 usage
func (s *AntigravityGatewayService) extractClaudeUsage(body []byte) *ClaudeUsage {
	usage := &ClaudeUsage{}
	var resp map[string]any
	if json.Unmarshal(body, &resp) != nil {
		return usage
	}
	if u, ok := resp["usage"].(map[string]any); ok {
		if v, ok := u["input_tokens"].(float64); ok {
			usage.InputTokens = int(v)
		}
		if v, ok := u["output_tokens"].(float64); ok {
			usage.OutputTokens = int(v)
		}
		if v, ok := u["cache_read_input_tokens"].(float64); ok {
			usage.CacheReadInputTokens = int(v)
		}
		if v, ok := u["cache_creation_input_tokens"].(float64); ok {
			usage.CacheCreationInputTokens = int(v)
		}
	}
	return usage
}