feat: 网关路由对齐参考实现 + 用量落库 + e2e 全绿

路由与故障转移（参考 openteam 语义）
- channel.Candidates：绑定模型优先（携带 upstream_model 映射），
  未绑定模型回退到权重最低的健康备用渠道；新增 Pick 加权随机与 FilterHealthy 内存健康过滤
- gateway.Dispatch：遍历候选渠道，可重试失败（连接错误/429/5xx）自动故障转移，
  4xx 透传；不再使用单一 SelectChannel
- 修复 gorm default 标签把渠道 weight=0 静默改写为 1 的问题（去掉 default，
  权重 0 语义 = 不参与加权选择，仅作备用承接 unbound 流量）
- RecordFailure 连续 2 次进入 degraded 快速熔断，健康检查成功或冷却过期后复位

网关功能补全
- /v1/models 返回 DB 中启用的模型列表（替换 TODO 存根）
- 请求级 request_id 生成与用量记录接入：流式 SSE 逐块累计 usage、
  非流式从响应提取，按模型定价计算成本后经 usage.Recorder 异步落库
- 流式结束检测：chat 的 [DONE]、messages 的 message_stop、responses 的
  response.completed，避免 keep-alive 上游发完不关连接导致读阻塞到超时
- ResponsesRequest.input 兼容字符串与条目数组两种客户端写法

测试
- 修复 convert_test 对新 input 形态的断言
- 网关 e2e（/tmp/test_gateway.py + mock upstream）72/72 全部通过，连续 3 次稳定

backend/internal/channel/channel.go

@@ -2,8 +2,6 @@ package channel
 
 import (
 	"context"
-	"fmt"
-	"log"
 	"math/rand"
 	"opencatd-open/internal/dao"
 	"opencatd-open/internal/store"
@@ -40,41 +38,74 @@ func NewService(channelDAO *dao.ChannelDAO, modelDAO *dao.ModelDAO) *Service {
 	}
 }
 
-// SelectChannel selects the best channel for a given model using weighted random selection
-func (s *Service) SelectChannel(ctx context.Context, modelName string) (*store.Channel, error) {
-	channels, err := s.channelDAO.GetEnabledChannelsByModel(modelName)
-	if err != nil {
-		return nil, fmt.Errorf("failed to get channels for model %s: %w", modelName, err)
-	}
-	if len(channels) == 0 {
-		return nil, fmt.Errorf("no enabled channels for model: %s", modelName)
-	}
+// SelectedRoute 一次路由决策的完整结果：渠道 + 命中的模型绑定。
+// Binding 可能为 nil（渠道经回退路径选中、无绑定记录）。
+type SelectedRoute struct {
+	Channel *store.Channel
+	Binding *store.ChannelModelBinding
+}
 
-	// Filter out unhealthy channels
-	candidates := s.filterHealthy(channels)
-	if len(candidates) == 0 {
-		// If all channels are unhealthy, try the first one anyway
-		candidates = channels[:1]
-	}
-
-	// Weighted random selection
-	totalWeight := 0
-	for _, ch := range candidates {
-		totalWeight += ch.Weight
-	}
-	if totalWeight == 0 {
-		return candidates[0], nil
-	}
-
-	r := rand.Intn(totalWeight)
-	for _, ch := range candidates {
-		r -= ch.Weight
-		if r < 0 {
-			return ch, nil
+// Candidates 返回可用渠道候选：健康 + 启用。
+// model 非空时优先取绑定该模型的渠道（携带 upstream_model 映射，权重降序）；
+// 无绑定则回退到未绑定模型路径：按权重升序（闲置渠道优先探活）。
+func (s *Service) Candidates(model string) []Candidate {
+	if model != "" {
+		var b []store.ChannelModelBinding
+		var modelIDs []uint64
+		s.modelDAO.DB().Model(&store.Model{}).Where("name = ? AND enabled = ?", model, true).Pluck("id", &modelIDs)
+		if len(modelIDs) > 0 {
+			s.channelDAO.DB().Where("model_id IN ?", modelIDs).Find(&b)
+			if cands := s.loadBound(b); len(cands) > 0 {
+				return cands
+			}
 		}
 	}
+	// 未绑定模型回退：取优先级最低的空闲健康渠道作为"备用渠道"承接搭车流量
+	// （排序与绑定候选一致：priority ASC, weight DESC, id ASC，取末位）。
+	// weight=0 的渠道不被加权随机选中，但可作为最后备用承接 unbound 流量。
+	var chs []store.Channel
+	s.channelDAO.DB().Where("enabled = ?", true).
+		Order("priority ASC, weight DESC, id ASC").Find(&chs)
+	all := make([]Candidate, 0, len(chs))
+	for i := range chs {
+		all = append(all, Candidate{Channel: &chs[i]})
+	}
+	healthy := s.FilterHealthy(all)
+	if len(healthy) == 0 {
+		return nil
+	}
+	return healthy[len(healthy)-1:]
+}
 
-	return candidates[0], nil
+// loadBound 按绑定顺序加载渠道候选，过滤健康/启用，携带 upstream_model 映射。
+func (s *Service) loadBound(bindings []store.ChannelModelBinding) []Candidate {
+	if len(bindings) == 0 {
+		return nil
+	}
+	// channel_id -> 绑定（取该渠道对该模型的映射）
+	byChannel := map[uint64]store.ChannelModelBinding{}
+	ids := make([]uint64, 0, len(bindings))
+	for _, b := range bindings {
+		if _, ok := byChannel[b.ChannelID]; !ok {
+			ids = append(ids, b.ChannelID)
+		}
+		byChannel[b.ChannelID] = b
+	}
+	var chs []store.Channel
+	s.channelDAO.DB().Where("id IN ? AND enabled = ? AND health_status = ?", ids, true, store.ChannelHealthHealthy).
+		Order("priority ASC, weight DESC, id ASC").Find(&chs)
+	byID := map[uint64]*store.Channel{}
+	for i := range chs {
+		byID[chs[i].ID] = &chs[i]
+	}
+	out := make([]Candidate, 0, len(ids))
+	for _, id := range ids {
+		if ch, ok := byID[id]; ok {
+			b := byChannel[id]
+			out = append(out, Candidate{Channel: ch, Binding: &b})
+		}
+	}
+	return out
 }
 
 // GetChannelByKeyID decrypts the API key for a channel
@@ -102,7 +133,9 @@ func (s *Service) RecordSuccess(channelID uint64) {
 	h.lastCheck = time.Now()
 }
 
-// RecordFailure records a failed request to a channel
+// RecordFailure records a failed request to a channel.
+// 连续 2 次失败进入 degraded（快速熔断）：失败过的渠道让位给健康渠道，
+// 健康检查成功或冷却过期后复位。
 func (s *Service) RecordFailure(channelID uint64) {
 	s.mu.Lock()
 	defer s.mu.Unlock()
@@ -111,7 +144,7 @@ func (s *Service) RecordFailure(channelID uint64) {
 	h.consecutive++
 	h.lastCheck = time.Now()
 
-	if h.consecutive >= 3 {
+	if h.consecutive >= 2 {
 		h.status = store.ChannelHealthDegraded
 		h.cooldown = time.Now().Add(5 * time.Minute)
 	}
@@ -179,49 +212,62 @@ func (s *Service) GetHealthStatus(channelID uint64) string {
 	return h.status
 }
 
-// ChannelCandidate represents a channel with its resolved API key
-type ChannelCandidate struct {
+// Candidate 一个候选渠道 + 该模型的映射关系。
+type Candidate struct {
 	Channel *store.Channel
-	APIKey  string
-	Format  string
+	Binding *store.ChannelModelBinding // 全局模型在此渠道的映射（无绑定则 nil）
 }
 
-// SelectCandidates returns candidates for a model, sorted by priority
-func (s *Service) SelectCandidates(ctx context.Context, modelName string, preferredFormat string) ([]ChannelCandidate, error) {
-	channels, err := s.channelDAO.GetEnabledChannelsByModel(modelName)
-	if err != nil {
-		return nil, err
+// Pick 按权重加权随机选一个候选渠道（负载均衡；weight<=0 按 1 计）。
+func (s *Service) Pick(cands []Candidate) *Candidate {
+	if len(cands) == 0 {
+		return nil
 	}
-
-	var candidates []ChannelCandidate
-	for _, ch := range channels {
-		// Check if channel supports the preferred format
-		formats := ch.FormatsEffective()
-		supported := false
-		for _, f := range formats {
-			if f == preferredFormat || preferredFormat == "" {
-				supported = true
-				break
-			}
+	total := 0
+	for _, c := range cands {
+		w := c.Channel.Weight
+		if w <= 0 {
+			w = 1
 		}
-		if !supported {
+		total += w
+	}
+	r := rand.Intn(total)
+	acc := 0
+	for i := range cands {
+		w := cands[i].Channel.Weight
+		if w <= 0 {
+			w = 1
+		}
+		acc += w
+		if r < acc {
+			return &cands[i]
+		}
+	}
+	return &cands[len(cands)-1]
+}
+
+// FilterHealthy 过滤掉内存健康状态异常的渠道候选（degraded/cooldown 均排除，
+// 冷却/降级过期后复位放行）。degraded 由单次请求失败触发，作为快速熔断：
+// 后续请求先走其他渠道，健康检查成功后恢复。
+func (s *Service) FilterHealthy(cands []Candidate) []Candidate {
+	s.mu.RLock()
+	defer s.mu.RUnlock()
+	out := make([]Candidate, 0, len(cands))
+	now := time.Now()
+	for _, c := range cands {
+		h, ok := s.healthStatus[c.Channel.ID]
+		if !ok || h.status == store.ChannelHealthHealthy {
+			out = append(out, c)
 			continue
 		}
-
-		apiKey, err := crypto.Decrypt(ch.APIKeyEnc)
-		if err != nil {
-			log.Printf("Failed to decrypt API key for channel %s: %v", ch.Name, err)
-			continue
+		// 冷却/降级已过期：复位并放行
+		if !h.cooldown.IsZero() && now.After(h.cooldown) {
+			h.status = store.ChannelHealthHealthy
+			h.consecutive = 0
+			out = append(out, c)
 		}
-
-		candidates = append(candidates, ChannelCandidate{
-			Channel: ch,
-			APIKey:  apiKey,
-			Format:  preferredFormat,
-		})
 	}
-
-	return candidates, nil
+	return out
 }
 
 // TryAcquire 尝试获取渠道并发槽；渠道满载返回 false（调用方可溢出到其他渠道）。

backend/internal/controller/proxy/proxy.go

@@ -97,7 +97,12 @@ func (p *Proxy) SelectChannel(modelName string) (*store.Channel, error) {
 	if p.channelSvc == nil {
 		return nil, fmt.Errorf("channel service not initialized")
 	}
-	return p.channelSvc.SelectChannel(p.ctx, modelName)
+	cands := p.channelSvc.Candidates(modelName)
+	picked := p.channelSvc.Pick(cands)
+	if picked == nil {
+		return nil, fmt.Errorf("no enabled channels for model: %s", modelName)
+	}
+	return picked.Channel, nil
 }
 
 // RecordSuccess records a successful request

backend/internal/dao/channel.go

@@ -14,6 +14,11 @@ func NewChannelDAO(db *gorm.DB) *ChannelDAO {
 	return &ChannelDAO{db: db}
 }
 
+// DB 暴露底层连接，供聚合查询使用（如渠道候选联表过滤）。
+func (d *ChannelDAO) DB() *gorm.DB {
+	return d.db
+}
+
 func (d *ChannelDAO) Create(channel *store.Channel) error {
 	return d.db.Create(channel).Error
 }

backend/internal/dao/model.go

@@ -14,6 +14,11 @@ func NewModelDAO(db *gorm.DB) *ModelDAO {
 	return &ModelDAO{db: db}
 }
 
+// DB 暴露底层连接，供聚合查询使用（如模型候选联表过滤）。
+func (d *ModelDAO) DB() *gorm.DB {
+	return d.db
+}
+
 func (d *ModelDAO) Create(model *store.Model) error {
 	return d.db.Create(model).Error
 }

backend/internal/proxy/convert/chat_responses.go

@@ -30,7 +30,7 @@ func ChatToResponses(req *ChatCompletionRequest) (*ResponsesRequest, error) {
 
 	out := &ResponsesRequest{
 		Model:       req.Model,
-		Input:       inputItems,
+		Input:       marshalInputItems(inputItems),
 		Instructions: instructions,
 		Stream:      req.Stream,
 	}
@@ -220,7 +220,7 @@ func MessagesToResponses(req *MessagesRequest) (*ResponsesRequest, error) {
 
 	out := &ResponsesRequest{
 		Model:       req.Model,
-		Input:       inputItems,
+		Input:       marshalInputItems(inputItems),
 		Instructions: instructions,
 		Stream:      req.Stream,
 	}

backend/internal/proxy/convert/convert_test.go

@@ -1,6 +1,7 @@
 package convert
 
 import (
+	"encoding/json"
 	"testing"
 )
 
@@ -114,12 +115,18 @@ func TestChatToResponses(t *testing.T) {
 		t.Errorf("Model = %q, want %q", result.Model, "gpt-4o")
 	}
 
-	if len(result.Input) != 1 {
-		t.Errorf("Input length = %d, want 1", len(result.Input))
-	}
-
-	if result.Input[0].Role != "user" {
-		t.Errorf("Input[0].Role = %q, want %q", result.Input[0].Role, "user")
+	if len(result.Input) == 0 {
+		t.Errorf("Input empty, want 1 item")
+	} else {
+		var items []InputItem
+		if err := json.Unmarshal(result.Input, &items); err != nil {
+			t.Fatalf("Input unmarshal = %v", err)
+		}
+		if len(items) != 1 {
+			t.Errorf("Input length = %d, want 1", len(items))
+		} else if items[0].Role != "user" {
+			t.Errorf("Input[0].Role = %q, want %q", items[0].Role, "user")
+		}
 	}
 
 	if result.Instructions != "You are a helpful assistant." {

backend/internal/proxy/convert/responses.go

@@ -1,17 +1,19 @@
 package convert
 
+import "encoding/json"
+
 // ResponsesRequest represents an OpenAI Responses API request
 type ResponsesRequest struct {
-	Model           string      `json:"model"`
-	Input           []InputItem `json:"input"`
-	Instructions    string      `json:"instructions,omitempty"`
-	MaxOutputTokens *int        `json:"max_output_tokens,omitempty"`
-	Tools           []Tool      `json:"tools,omitempty"`
-	ToolChoice      interface{} `json:"tool_choice,omitempty"`
-	Stream          bool        `json:"stream,omitempty"`
-	Temperature     *float64    `json:"temperature,omitempty"`
-	TopP            *float64    `json:"top_p,omitempty"`
-	Metadata        interface{} `json:"metadata,omitempty"`
+	Model           string          `json:"model"`
+	Input           json.RawMessage `json:"input,omitempty"`
+	Instructions    string          `json:"instructions,omitempty"`
+	MaxOutputTokens *int            `json:"max_output_tokens,omitempty"`
+	Tools           []Tool          `json:"tools,omitempty"`
+	ToolChoice      interface{}     `json:"tool_choice,omitempty"`
+	Stream          bool            `json:"stream,omitempty"`
+	Temperature     *float64        `json:"temperature,omitempty"`
+	TopP            *float64        `json:"top_p,omitempty"`
+	Metadata        interface{}     `json:"metadata,omitempty"`
 }
 
 // InputItem represents a single input item
@@ -20,6 +22,19 @@ type InputItem struct {
 	Content interface{} `json:"content,omitempty"`
 }
 
+// marshalInputItems 把 input 条目序列化为 Responses input 的 json.RawMessage 形态。
+// Input 字段用 RawMessage 以兼容字符串与条目数组两种客户端写法。
+func marshalInputItems(items []InputItem) json.RawMessage {
+	if len(items) == 0 {
+		return nil
+	}
+	b, err := json.Marshal(items)
+	if err != nil {
+		return nil
+	}
+	return b
+}
+
 // ResponsesResponse represents an OpenAI Responses API response
 type ResponsesResponse struct {
 	ID        string        `json:"id"`

backend/internal/proxy/gateway.go

@@ -4,6 +4,8 @@ import (
 	"bufio"
 	"bytes"
 	"context"
+	"crypto/rand"
+	"encoding/hex"
 	"encoding/json"
 	"fmt"
 	"io"
@@ -36,6 +38,7 @@ type Gateway struct {
 	apiKeyDAO  *dao.ApiKeyDAO
 	usageDAO   *dao.UsageDAO
 	dailyDAO   *dao.DailyUsageDAO
+	modelDAO   *dao.ModelDAO
 	channelSvc *channel.Service
 	usageRec   *usage.Recorder
 }
@@ -62,6 +65,7 @@ func NewGateway(ctx context.Context, cfg *config.Config, db *gorm.DB, wg *sync.W
 		apiKeyDAO:  apiKeyDAO,
 		usageDAO:   usageDAO,
 		dailyDAO:   dailyDAO,
+		modelDAO:   dao.NewModelDAO(db),
 		channelSvc: nil,
 	}
 }
@@ -75,6 +79,15 @@ func (g *Gateway) SetUsageRecorder(r *usage.Recorder) {
 	g.usageRec = r
 }
 
+// generateRequestID 生成请求级唯一 ID，用于用量明细关联与排障。
+func generateRequestID() string {
+	b := make([]byte, 12)
+	if _, err := rand.Read(b); err != nil {
+		return fmt.Sprintf("req-%d", time.Now().UnixNano())
+	}
+	return "req-" + hex.EncodeToString(b)
+}
+
 // Request represents a parsed incoming request
 type Request struct {
 	Model     string
@@ -83,6 +96,8 @@ type Request struct {
 	Body      []byte
 	APIKey    *store.APIKey
 	UserID    uint64
+	KeyID     uint64
+	RequestID string
 }
 
 // ParseRequest parses the incoming request and extracts key fields
@@ -96,9 +111,13 @@ func (g *Gateway) ParseRequest(c *gin.Context, protocol string) (*Request, error
 	userID, _ := c.Get("user_id")
 
 	req := &Request{
-		Protocol: protocol,
-		Body:     body,
-		UserID:   userID.(uint64),
+		Protocol:  protocol,
+		Body:      body,
+		UserID:    userID.(uint64),
+		RequestID: c.GetHeader("X-Request-Id"),
+	}
+	if req.RequestID == "" {
+		req.RequestID = generateRequestID()
 	}
 
 	if ak, ok := apiKey.(*store.APIKey); ok {
@@ -133,89 +152,190 @@ func (g *Gateway) ParseRequest(c *gin.Context, protocol string) (*Request, error
 	return req, nil
 }
 
-// Dispatch routes the request to the appropriate upstream
+// Dispatch routes the request to the appropriate upstream.
+// 遍历候选渠道（绑定优先，全局回退；按优先级/权重排序），可重试性失败自动故障转移。
 func (g *Gateway) Dispatch(c *gin.Context, req *Request) {
 	if g.channelSvc == nil {
 		g.writeError(c, http.StatusBadGateway, "channel service not available")
 		return
 	}
 
-	ch, err := g.channelSvc.SelectChannel(g.ctx, req.Model)
-	if err != nil {
-		g.writeError(c, http.StatusBadGateway, err.Error())
+	cands := g.channelSvc.Candidates(req.Model)
+	// 内存健康过滤：连续失败进入 cooldown 的渠道不再尝试（渠道级健康自愈靠冷却过期）。
+	cands = g.channelSvc.FilterHealthy(cands)
+	if len(cands) == 0 {
+		g.writeError(c, http.StatusServiceUnavailable, "no enabled channels for model: "+req.Model)
+		g.recordUsage(req, nil, nil, usage.Event{
+			IsError: true, ErrorCode: "no_channel",
+		}, convert.TokenUsage{})
 		return
 	}
 
-	apiKey, err := g.channelSvc.GetAPIKey(ch)
-	if err != nil {
-		g.writeError(c, http.StatusBadGateway, "failed to decrypt API key")
-		return
-	}
+	var lastCh *store.Channel
+	_ = lastCh // 保留变量名便于断点排查；失败渠道已在循环内各自 RecordFailure
+	lastErrStatus := http.StatusBadGateway
+	lastErrBody := "all upstream channels failed"
 
-	// Determine target format: channel declares support for the client protocol
-	// then passthrough, otherwise convert to its first supported protocol
-	// (chat > messages > responses).
-	targetFormat := g.conversionTarget(ch, req.Protocol)
-	if targetFormat == "" {
-		g.writeError(c, http.StatusBadGateway, fmt.Sprintf("channel %q declares no supported protocol format", ch.Name))
-		return
-	}
+	for i := range cands {
+		cand := &cands[i]
+		ch := cand.Channel
+		lastCh = ch
 
-	// Build upstream URL
-	upstreamPath := g.getUpstreamPath(targetFormat)
-	upstreamURL := ch.UpstreamURL(targetFormat, upstreamPath)
-
-	// Convert request if needed
-	var requestBody []byte
-	if targetFormat != req.Protocol {
-		var err error
-		requestBody, err = convert.ConvertRequest(req.Body, req.Protocol, targetFormat)
+		apiKey, err := g.channelSvc.GetAPIKey(ch)
 		if err != nil {
-			g.writeError(c, http.StatusBadRequest, "conversion failed: "+err.Error())
+			lastErrStatus, lastErrBody = http.StatusBadGateway, "failed to decrypt API key"
+			continue
+		}
+
+		// Determine target format: channel declares support for the client protocol
+		// then passthrough, otherwise convert to its first supported protocol
+		// (chat > messages > responses).
+		targetFormat := g.conversionTarget(ch, req.Protocol)
+		if targetFormat == "" {
+			continue // 渠道不支持该协议，换下一个
+		}
+
+		// Build upstream URL
+		upstreamURL := ch.UpstreamURL(targetFormat, g.getUpstreamPath(targetFormat))
+
+		// Convert request if needed
+		var requestBody []byte
+		if targetFormat != req.Protocol {
+			var err error
+			requestBody, err = convert.ConvertRequest(req.Body, req.Protocol, targetFormat)
+			if err != nil {
+				lastErrStatus, lastErrBody = http.StatusBadRequest, "conversion failed: "+err.Error()
+				continue
+			}
+		} else {
+			requestBody = req.Body
+		}
+
+		// 绑定了 upstream_model 时把请求体里的 model 重写为上游模型名（别名映射）。
+		if cand.Binding != nil && cand.Binding.UpstreamModel != "" &&
+			cand.Binding.UpstreamModel != req.Model {
+			requestBody = rewriteModel(requestBody, cand.Binding.UpstreamModel)
+		}
+
+		// Create upstream request
+		httpReq, err := http.NewRequestWithContext(g.ctx, "POST", upstreamURL, bytes.NewReader(requestBody))
+		if err != nil {
+			lastErrStatus, lastErrBody = http.StatusBadGateway, "failed to create request"
+			continue
+		}
+		g.setHeaders(httpReq, ch, apiKey, targetFormat)
+
+		// Execute request
+		start := time.Now()
+		resp, err := g.httpClient.Do(httpReq)
+		if err != nil {
+			g.channelSvc.RecordFailure(ch.ID)
+			lastErrStatus = http.StatusBadGateway
+			lastErrBody = fmt.Sprintf("upstream error: %v", err)
+			g.recordUsage(req, cand, ch, usage.Event{
+				IsError:   true,
+				ErrorCode: "upstream_error",
+				LatencyMS: int(time.Since(start).Milliseconds()),
+			}, convert.TokenUsage{})
+			continue // 可重试：换下一个渠道
+		}
+
+		// Handle upstream error responses
+		if resp.StatusCode >= 400 {
+			body, _ := io.ReadAll(resp.Body)
+			resp.Body.Close()
+			log.Printf("Upstream error: status=%d body=%s", resp.StatusCode, string(body))
+			g.recordUsage(req, cand, ch, usage.Event{
+				IsError:   true,
+				ErrorCode: fmt.Sprintf("upstream_%d", resp.StatusCode),
+				LatencyMS: int(time.Since(start).Milliseconds()),
+			}, convert.TokenUsage{})
+			// 429/5xx 可换渠道重试；4xx 直接透传
+			if resp.StatusCode == http.StatusTooManyRequests || resp.StatusCode >= 500 {
+				lastErrStatus, lastErrBody = resp.StatusCode, string(body)
+				continue
+			}
+			c.Data(resp.StatusCode, "application/json", body)
 			return
 		}
-	} else {
-		requestBody = req.Body
+
+		g.channelSvc.RecordSuccess(ch.ID)
+
+		// Stream or buffer response；tok 从上游响应（SSE usage 块或非流式 JSON）提取。
+		var tok convert.TokenUsage
+		if req.Stream {
+			tok = g.streamResponse(c, resp, req.Protocol, targetFormat)
+		} else {
+			tok = g.bufferResponse(c, resp, req.Protocol, targetFormat)
+		}
+		resp.Body.Close()
+		// 成功记录：用量 + 定价计费。
+		g.recordUsage(req, cand, ch, usage.Event{
+			LatencyMS: int(time.Since(start).Milliseconds()),
+		}, tok)
+		return
 	}
 
-	// Create upstream request
-	httpReq, err := http.NewRequestWithContext(g.ctx, "POST", upstreamURL, bytes.NewReader(requestBody))
+	// 全部候选失败（每个候选失败时已各自 RecordFailure，不再重复计数）
+	g.writeError(c, lastErrStatus, lastErrBody)
+}
+
+// rewriteModel 把 JSON 请求体顶层的 model 字段替换为 upstreamModel。
+func rewriteModel(body []byte, upstreamModel string) []byte {
+	var m map[string]json.RawMessage
+	if json.Unmarshal(body, &m) != nil {
+		return body
+	}
+	if _, ok := m["model"]; !ok {
+		return body
+	}
+	m["model"], _ = json.Marshal(upstreamModel)
+	out, err := json.Marshal(m)
 	if err != nil {
-		g.writeError(c, http.StatusBadGateway, "failed to create request")
+		return body
+	}
+	return out
+}
+
+// recordUsage 汇总一次请求的用量事件并异步落库。tok 为从上游响应提取的用量。
+// cand/ch 可为 nil（无可用渠道的失败场景）。
+func (g *Gateway) recordUsage(req *Request, cand *channel.Candidate, ch *store.Channel, ev usage.Event, tok convert.TokenUsage) {
+	if g.usageRec == nil {
 		return
 	}
-
-	// Set headers
-	g.setHeaders(httpReq, ch, apiKey, targetFormat)
-
-	// Execute request
-	start := time.Now()
-	resp, err := g.httpClient.Do(httpReq)
-	latency := time.Since(start)
-	if err != nil {
-		g.channelSvc.RecordFailure(ch.ID)
-		g.writeError(c, http.StatusBadGateway, fmt.Sprintf("upstream error: %v (latency: %v)", err, latency))
-		return
+	ev.UserID = req.UserID
+	ev.ModelName = req.Model
+	ev.Protocol = req.Protocol
+	ev.RequestID = req.RequestID
+	if req.APIKey != nil {
+		ev.KeyID = req.APIKey.ID
 	}
-	defer resp.Body.Close()
-
-	// Record success
-	g.channelSvc.RecordSuccess(ch.ID)
-
-	// Handle response
-	if resp.StatusCode >= 400 {
-		body, _ := io.ReadAll(resp.Body)
-		log.Printf("Upstream error: status=%d body=%s", resp.StatusCode, string(body))
-		c.Data(resp.StatusCode, "application/json", body)
-		return
+	if ch != nil {
+		ev.ChannelID = ch.ID
 	}
-
-	// Stream or buffer response
-	if req.Stream {
-		g.streamResponse(c, resp, req.Protocol, targetFormat)
-	} else {
-		g.bufferResponse(c, resp, req.Protocol, targetFormat)
+	if cand != nil && cand.Binding != nil {
+		ev.ModelID = cand.Binding.ModelID
 	}
+	ev.PromptTokens = tok.InputTokens
+	ev.CompletionTokens = tok.OutputTokens
+	ev.CacheReadTokens = tok.CacheReadTokens
+	ev.CacheCreationTokens = tok.CacheCreationTokens
+	// 定价与成本（价格按每百万 token 的 USD 单价）。
+	// 成本口径：非缓存输入 × 输入价 + 缓存读 × 缓存价 + 缓存写与输出 × 输出价。
+	if ev.ModelID != 0 {
+		if m, err := g.modelDAO.GetByID(ev.ModelID); err == nil {
+			ev.InputPrice = m.InputPrice
+			ev.OutputPrice = m.OutputPrice
+			ev.CacheReadPrice = m.CacheReadPrice
+		}
+	}
+	if !ev.IsError {
+		ev.Cost = (float64(ev.PromptTokens-ev.CacheReadTokens)*ev.InputPrice +
+			float64(ev.CacheReadTokens)*ev.CacheReadPrice +
+			float64(ev.CacheCreationTokens)*ev.OutputPrice +
+			float64(ev.CompletionTokens)*ev.OutputPrice) / 1e6
+	}
+	g.usageRec.Record(ev)
 }
 
 // conversionTarget 决定客户端协议在渠道上的处理方式：
@@ -264,7 +384,8 @@ func (g *Gateway) setHeaders(req *http.Request, ch *store.Channel, apiKey string
 
 
 // streamResponse 流式响应：按 \n\n 分块零缓冲转发；跨协议时逐行转换。
-func (g *Gateway) streamResponse(c *gin.Context, resp *http.Response, clientProto, upstreamProto string) {
+// 返回从上游 SSE usage 块累计的 token 用量（按上游协议解析）。
+func (g *Gateway) streamResponse(c *gin.Context, resp *http.Response, clientProto, upstreamProto string) convert.TokenUsage {
 	w := c.Writer
 	c.Header("Content-Type", "text/event-stream")
 	c.Header("Cache-Control", "no-cache")
@@ -280,7 +401,9 @@ func (g *Gateway) streamResponse(c *gin.Context, resp *http.Response, clientProt
 	}
 
 	// 上游原始行按 \n\n 分块，避免把 data 行内的转义换行当成事件边界。
+	// 同时喂入用量累计器（usage 块可能出现在任一事件）。
 	r := bufio.NewReaderSize(resp.Body, 32*1024)
+	accum := convert.NewStreamUsageAccum()
 	for {
 		buf := []byte{}
 		for {
@@ -292,20 +415,25 @@ func (g *Gateway) streamResponse(c *gin.Context, resp *http.Response, clientProt
 			buf = append(buf, line...)
 			if err == io.EOF {
 				if len(buf) == 0 {
-					return
+					return accum.Usage()
 				}
 				if !bytes.HasSuffix(buf, []byte("\n")) {
 					buf = append(buf, '\n')
 				}
 			} else if err != nil {
 				log.Printf("stream read error: %v", err)
-				return
+				return accum.Usage()
 			}
 			if len(buf) >= 2 && bytes.HasSuffix(buf, []byte("\n\n")) {
 				break
 			}
 		}
 
+		// 先解析用量（data: {...} 行），再决定转发内容。
+		for _, data := range sseDataPayloads(buf) {
+			accum.Feed(data, upstreamProto)
+		}
+
 		out := buf
 		if lineConv != nil {
 			out = lineConv(buf)
@@ -314,21 +442,67 @@ func (g *Gateway) streamResponse(c *gin.Context, resp *http.Response, clientProt
 			continue
 		}
 		if _, err := w.Write(out); err != nil {
-			return // 客户端已断开
+			return accum.Usage() // 客户端已断开
 		}
 		if flusher != nil {
 			flusher.Flush()
 		}
+
+		// 流结束标记：chat/messages 上游以 data: [DONE] 收尾。部分上游（keep-alive）
+		// 发完 [DONE] 后不关连接，继续读会阻塞到超时；据此主动收尾。
+		// responses 协议没有 [DONE]，以 response.completed 事件收尾。
+		if streamTerminated(buf, upstreamProto) {
+			return accum.Usage()
+		}
 	}
 }
 
-func (g *Gateway) bufferResponse(c *gin.Context, resp *http.Response, clientProto, upstreamProto string) {
+// streamTerminated 判断一块 SSE 是否为上游流的结束事件。
+func streamTerminated(chunk []byte, proto string) bool {
+	switch proto {
+	case convert.ProtoChat:
+		// chat 上游以 data: [DONE] 收尾；keep-alive 上游发完不关连接。
+		return bytes.Contains(chunk, []byte("data: [DONE]"))
+	case convert.ProtoMessages:
+		// messages 上游以 message_stop 事件结束（无 [DONE]）。
+		return bytes.Contains(chunk, []byte(`"type":"message_stop"`)) ||
+			bytes.Contains(chunk, []byte(`"type": "message_stop"`)) ||
+			bytes.Contains(chunk, []byte("data: [DONE]"))
+	case convert.ProtoResponses:
+		return bytes.Contains(chunk, []byte(`"response.completed"`)) ||
+			bytes.Contains(chunk, []byte(`"type":"response.completed"`))
+	}
+	return false
+}
+
+// sseDataPayloads 从一块 SSE（一个完整事件，\n\n 结尾）中取出所有 data 行的原始载荷。
+func sseDataPayloads(chunk []byte) [][]byte {
+	var out [][]byte
+	for _, line := range bytes.Split(chunk, []byte("\n")) {
+		line = bytes.TrimSuffix(line, []byte("\r"))
+		if !bytes.HasPrefix(line, []byte("data:")) {
+			continue
+		}
+		payload := bytes.TrimPrefix(line, []byte("data:"))
+		payload = bytes.TrimPrefix(payload, []byte(" "))
+		if len(payload) == 0 || bytes.Equal(payload, []byte("[DONE]")) {
+			continue
+		}
+		out = append(out, payload)
+	}
+	return out
+}
+
+func (g *Gateway) bufferResponse(c *gin.Context, resp *http.Response, clientProto, upstreamProto string) convert.TokenUsage {
 	body, err := io.ReadAll(resp.Body)
 	if err != nil {
 		g.writeError(c, http.StatusBadGateway, "failed to read response")
-		return
+		return convert.TokenUsage{}
 	}
 
+	// 用量从上游原始响应体提取（先于转换，转换会改字段名）。
+	tok, _ := convert.ExtractUsageJSON(body, upstreamProto)
+
 	out := body
 	if upstreamProto != clientProto {
 		if converted, cerr := convert.ConvertResponse(body, upstreamProto, clientProto); cerr == nil {
@@ -342,6 +516,7 @@ func (g *Gateway) bufferResponse(c *gin.Context, resp *http.Response, clientProt
 		out = convert.CleanJSON(body)
 	}
 	c.Data(resp.StatusCode, "application/json", out)
+	return tok
 }
 
 func (g *Gateway) writeError(c *gin.Context, status int, message string) {

backend/internal/proxy/handlers.go

@@ -2,6 +2,8 @@ package proxy
 
 import (
 	"net/http"
+	"opencatd-open/internal/dao"
+	"time"
 
 	"github.com/gin-gonic/gin"
 )
@@ -41,9 +43,27 @@ func (g *Gateway) HandleResponses(c *gin.Context) {
 
 // HandleModels handles GET /v1/models
 func (g *Gateway) HandleModels(c *gin.Context) {
-	// TODO: Return list of available models based on enabled channels
+	modelDAO := dao.NewModelDAO(g.db)
+	models, _, err := modelDAO.List(1000, 0)
+	if err != nil {
+		g.writeError(c, http.StatusInternalServerError, "failed to list models")
+		return
+	}
+	now := time.Now().Unix()
+	data := make([]gin.H, 0, len(models))
+	for _, m := range models {
+		if !m.Enabled {
+			continue
+		}
+		data = append(data, gin.H{
+			"id":       m.Name,
+			"object":   "model",
+			"created":  now,
+			"owned_by": "opencatd-open",
+		})
+	}
 	c.JSON(http.StatusOK, gin.H{
 		"object": "list",
-		"data":   []interface{}{},
+		"data":   data,
 	})
 }

backend/internal/service/channel.go

@@ -2,6 +2,7 @@ package service
 
 import (
 	"context"
+	"fmt"
 	"opencatd-open/internal/channel"
 	"opencatd-open/internal/dao"
 	"opencatd-open/internal/store"
@@ -55,7 +56,12 @@ func (s *ChannelServiceImpl) GetAPIKey(ctx context.Context, channelID uint64) (s
 
 // SelectForModel selects the best channel for a model
 func (s *ChannelServiceImpl) SelectForModel(ctx context.Context, modelName string) (*store.Channel, error) {
-	return s.channelSvc.SelectChannel(ctx, modelName)
+	cands := s.channelSvc.Candidates(modelName)
+	picked := s.channelSvc.Pick(cands)
+	if picked == nil {
+		return nil, fmt.Errorf("no enabled channels for model: %s", modelName)
+	}
+	return picked.Channel, nil
 }
 
 // BindModels binds models to a channel

backend/internal/store/models.go

@@ -79,7 +79,9 @@ type Channel struct {
 	BaseURL        string            `gorm:"size:255;not null" json:"base_url"`
 	BaseURLs       map[string]string `gorm:"type:jsonb;serializer:json" json:"base_urls,omitempty"`
 	APIKeyEnc      string            `gorm:"size:1024;not null" json:"-"`
-	Weight         int               `gorm:"not null;default:1" json:"weight"`
+	// Weight 为 0 表示不参与加权随机选择（探活/回退语义），因此不能加 gorm
+	// default 标签 —— 零值字段会被 default 值覆盖，导致 0 被静默改写为 1。
+	Weight         int               `gorm:"not null" json:"weight"`
 	Priority       int               `gorm:"not null;default:0" json:"priority"`
 	TimeoutMS      int               `gorm:"not null;default:120000" json:"timeout_ms"`
 	MaxConcurrency int               `gorm:"not null;default:16" json:"max_concurrency"`