feat: 自定义 OpenAI 兼容上游聚合 + 权重/并发调度

- custom provider:把任意 v1 端点当账号(base_url+key),按 model id 路由,
  直连不走代理;支持 images/generations、images/edits、Sora 式异步视频
- 调度器全局权重优先 + 每账号并发感知(上游账号可配并发,其余系统固定)
- 后台:添加/编辑上游、自定义模型表单(每档普通/代理价,留空=不支持)、
  账号权重/并发列与导入权重
- grok-video / nano-banana-2(runway) 写入硬编码目录(开源自带);
  去掉 540p,Adobe 视频最低 720p
- UI:SelectMenu 与输入框等高、测试耗时改用秒

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

README.md

@@ -45,7 +45,7 @@
 
 ## ✨ 简介
 
-**image2api** 把 Adobe Firefly、OpenAI、Runway、Grok、Leonardo、Krea、Imagine 等平台的图像 / 视频能力,统一封装成**一套 OpenAI 兼容的 API**;背后用多账号池自动调度 —— 额度耗尽自动换号、认证失效自动刷新或判死、临时错误自动重试、token 到期前主动续期 —— 对外提供稳定服务。
+**image2api** 把 Adobe Firefly、OpenAI、Runway、Grok、Leonardo、Krea、Imagine 等平台,以及**任意 OpenAI 兼容上游**的图像 / 视频能力,统一封装成**一套 OpenAI 兼容的 API**;背后用多账号池自动调度 —— 权重优先 + 并发感知、额度耗尽自动换号、认证失效自动刷新或判死、临时错误自动重试、token 到期前主动续期 —— 对外提供稳定服务。
 
 它不只是 API 代理:自带**积分计费、CDK 充值、邀请奖励、用户体系、管理后台、现代化画图前端**,一条命令即可跑成一个对外运营的 AI 生成站点 —— 作者的线上实例 **[Vivid AI · vividai.run](https://vividai.run)**(品牌)即基于本项目搭建。
 
@@ -74,10 +74,17 @@
 - 图片结果 **base64 直返**,服务端不留存文件,隐私友好
 
 #### 🔁 多账号池 + 智能故障转移
-- 账号池轮询调度,单账号出错不影响整体
+- 账号池调度,单账号出错不影响整体
+- **权重优先 + 并发感知**:按账号权重从高到低调度,某账号并发满了才轮到下一个;同权重组内 round-robin 均摊。每账号并发数可配(上游账号),其余系统固定
 - **额度耗尽→换号** · **认证失效→刷新重试 / 判死** · **临时错误→同号重试 ×3** · **参数错→直接报错**
 - **预扣额度**:生成前原子扣减,失败自动退回,杜绝并发超额
 
+#### 🔗 自定义上游聚合(OpenAI 兼容)
+- 把任意 **OpenAI 兼容的 v1 端点**当成一个账号接入(填 `base_url` + `key`),无需写代码
+- **按 model id 自动路由**:上游声明支持哪些 id,生成该 id 时即走对应上游(可覆盖内置 provider);id 留空 = 全部
+- 模型管理里自由新建自定义模型(id / 类型 / 比例 / 分辨率·价 / 时长·价 / 参考图),按本地价计费
+- 调用**直连不走代理**;上游可配权重与并发,与内置池统一调度
+
 #### 🔐 Token 自动保活
 - 一次性轮换 token(Krea / Imagine)**到期前 10 分钟主动续期**,新 token 自动落库
 - Adobe cookie 定时换 token;纯 JWT 到期自动判死
@@ -110,8 +117,9 @@
 | **Leonardo.ai** | seedream-4.5 | 图像 |
 | **Krea.ai** | flux-klein-2 | 图像 |
 | **Imagine.art** | imagine-1.5 · imagine-1.5pro | 图像 |
+| **自定义上游** | 任意 OpenAI 兼容 v1 端点(按 id 路由) | 图像 / 视频 |
 
-> 模型由管理后台动态启用并定价,可随时增删。
+> 模型由管理后台动态启用并定价,可随时增删。自定义上游支持把任何 OpenAI 兼容服务接成账号,按 model id 路由调用。
 
 ## 🔌 OpenAI 兼容 API
 
@@ -238,7 +246,8 @@ backend/                       后端源码(Go)
 │   │   ├── grok/              Grok(grok.com,statsig 伪造,视频)
 │   │   ├── leonardo/          Leonardo
 │   │   ├── krea/              Krea
-│   │   └── imagine/           Imagine.art
+│   │   ├── imagine/           Imagine.art
+│   │   └── custom/            自定义上游(OpenAI 兼容 v1,按 id 路由,直连不走代理)
 │   ├── repo/                  数据访问层(用户 / 模型 / 账号 / 日志 / CDK…)
 │   ├── service/              业务逻辑(生成调度、计费、账号池、保活、维护)
 │   └── storage/               RustFS / S3 媒体存储

backend/internal/bootstrap/app.go

@@ -12,6 +12,7 @@ import (
 	"backend/internal/model"
 	"backend/internal/provider/adobe"
 	"backend/internal/provider/chatgpt"
+	"backend/internal/provider/custom"
 	"backend/internal/provider/grok"
 	"backend/internal/provider/imagine"
 	"backend/internal/provider/krea"
@@ -110,7 +111,8 @@ func NewApp(ctx context.Context) (*App, error) {
 	kreaClient := krea.NewClient("")
 	imagineClient := imagine.NewClient("")
 	grokClient := grok.NewClient("")
-	v1Svc := service.NewV1Service(cfg, modelRepo, userRepo, eventRepo, tokenRepo, siteRepo, adobeClient, chatGPTClient, runwayClient, leonardoClient, kreaClient, imagineClient, grokClient, rustfsClient)
+	customClient := custom.NewClient()
+	v1Svc := service.NewV1Service(cfg, modelRepo, userRepo, eventRepo, tokenRepo, siteRepo, adobeClient, chatGPTClient, runwayClient, leonardoClient, kreaClient, imagineClient, grokClient, customClient, rustfsClient)
 	siteSvc := service.NewSiteService(siteRepo, cfg.AppTitle)
 	showcaseSvc := service.NewShowcaseService(showcaseRepo)
 	adminReadSvc := service.NewAdminReadService(cfg, userRepo, modelRepo, eventRepo, siteRepo, tokenRepo, cdkRepo, rustfsClient)

backend/internal/http/handler/provider_admin.go

@@ -132,6 +132,38 @@ func (h *ProviderAdminHandler) ImportGrokToken(c *gin.Context) {
 	c.JSON(http.StatusOK, gin.H{"ok": true, "id": item.ID, "status": item.Status, "pending": item.Status == "pending"})
 }
 
+func (h *ProviderAdminHandler) ImportCustomAccount(c *gin.Context) {
+	var body struct {
+		BaseURL     string `json:"base_url"`
+		URL         string `json:"url"`
+		Key         string `json:"key"`
+		APIKey      string `json:"api_key"`
+		Models      string `json:"models"`
+		Name        string `json:"name"`
+		Weight      int    `json:"weight"`
+		Concurrency int    `json:"concurrency"`
+		ID          string `json:"id"`
+	}
+	if err := c.ShouldBindJSON(&body); err != nil {
+		c.JSON(http.StatusBadRequest, gin.H{"detail": "invalid request body"})
+		return
+	}
+	baseURL := body.BaseURL
+	if baseURL == "" {
+		baseURL = body.URL
+	}
+	key := body.Key
+	if key == "" {
+		key = body.APIKey
+	}
+	item, err := h.tokens.ImportCustomAccount(c.Request.Context(), baseURL, key, body.Models, body.Name, body.Weight, body.Concurrency, body.ID)
+	if err != nil {
+		c.JSON(http.StatusBadRequest, gin.H{"detail": err.Error()})
+		return
+	}
+	c.JSON(http.StatusOK, gin.H{"ok": true, "id": item.ID, "status": item.Status})
+}
+
 func (h *ProviderAdminHandler) ImportKreaCookie(c *gin.Context) {
 	var body struct {
 		Cookie string `json:"cookie"`

backend/internal/http/handler/user_generation.go

@@ -371,12 +371,14 @@ func (h *UserGenerationHandler) catalogEntries(c *gin.Context) ([]gin.H, error)
 		},
 		{
 			"id":                   "nano-banana-2",
-			"provider":       "adobe",
+			"provider":             "runway",
 			"type":                 "image",
-			"ratios":         []string{"1:1", "16:9", "9:16", "4:3", "3:4"},
+			"ratios":               []string{"16:9", "9:16", "1:1", "4:3", "3:4", "21:9", "3:2", "5:4", "4:5", "2:3"},
-			"resolutions":    []string{"1K"},
+			"resolutions":          []string{"1K", "2K", "4K"},
 			"image_to_image":       true,
-			"description":    "Adobe Gemini Flash Nano Banana",
+			"max_reference_images": 6,
+			"reference_mode":       "asset",
+			"description":          "Runway Nano Banana 2 (图/参考图)",
 		},
 		{
 			"id":                   "gemini-veo31",
@@ -405,7 +407,7 @@ func (h *UserGenerationHandler) catalogEntries(c *gin.Context) ([]gin.H, error)
 			"provider":             "adobe",
 			"type":                 "video",
 			"ratios":               []string{"16:9", "1:1", "9:16"},
-			"resolutions":          []string{"540p", "720p", "1080p"},
+			"resolutions":          []string{"720p", "1080p"},
 			"durations":            []string{"5s"},
 			"max_reference_images": 2,
 			"reference_mode":       "frame",
@@ -422,6 +424,17 @@ func (h *UserGenerationHandler) catalogEntries(c *gin.Context) ([]gin.H, error)
 			"reference_mode":       "frame",
 			"description":          "Runway Gen-4 Turbo video (图生视频)",
 		},
+		{
+			"id":                   "grok-video",
+			"provider":             "grok",
+			"type":                 "video",
+			"ratios":               []string{"2:3", "3:2", "1:1", "9:16", "16:9"},
+			"resolutions":          []string{"720p"},
+			"durations":            []string{"6s", "10s"},
+			"max_reference_images": 6,
+			"reference_mode":       "asset",
+			"description":          "Grok Imagine video (文/图生视频)",
+		},
 		{
 			"id":                   "seedream-4.5",
 			"provider":             "leonardo",
@@ -519,11 +532,11 @@ func (h *UserGenerationHandler) publicModels() ([]gin.H, error) {
 		},
 		{
 			"id":          "nano-banana-2",
-			"provider":    "adobe",
+			"provider":    "runway",
 			"kind":        "image",
-			"ratios":      []string{"1:1", "16:9", "9:16", "4:3", "3:4"},
+			"ratios":      []string{"16:9", "9:16", "1:1", "4:3", "3:4", "21:9", "3:2", "5:4", "4:5", "2:3"},
-			"resolutions": []string{"1K"},
+			"resolutions": []string{"1K", "2K", "4K"},
-			"description": "Adobe Gemini Flash Nano Banana",
+			"description": "Runway Nano Banana 2",
 			"stub":        false,
 		},
 		{
@@ -549,7 +562,7 @@ func (h *UserGenerationHandler) publicModels() ([]gin.H, error) {
 			"provider":    "adobe",
 			"kind":        "video",
 			"ratios":      []string{"16:9", "1:1", "9:16"},
-			"resolutions": []string{"540p", "720p", "1080p"},
+			"resolutions": []string{"720p", "1080p"},
 			"description": "Adobe Firefly Video",
 			"stub":        false,
 		},
@@ -562,6 +575,15 @@ func (h *UserGenerationHandler) publicModels() ([]gin.H, error) {
 			"description": "Runway Gen-4 Turbo video",
 			"stub":        false,
 		},
+		{
+			"id":          "grok-video",
+			"provider":    "grok",
+			"kind":        "video",
+			"ratios":      []string{"2:3", "3:2", "1:1", "9:16", "16:9"},
+			"resolutions": []string{"720p"},
+			"description": "Grok Imagine video",
+			"stub":        false,
+		},
 		{
 			"id":          "seedream-4.5",
 			"provider":    "leonardo",

backend/internal/http/router/router.go

@@ -109,6 +109,7 @@ func New(cfg *config.Config, auth *service.AuthService, handlers Handlers) *gin.
 		authed.POST("/tokens/import-krea-cookie", handlers.ProviderAdmin.ImportKreaCookie)
 		authed.POST("/tokens/import-imagine-token", handlers.ProviderAdmin.ImportImagineToken)
 		authed.POST("/tokens/import-grok-token", handlers.ProviderAdmin.ImportGrokToken)
+		authed.POST("/tokens/import-custom-account", handlers.ProviderAdmin.ImportCustomAccount)
 		authed.POST("/tokens/delete-bulk", handlers.ProviderAdmin.TokenDeleteBulk)
 		authed.PATCH("/tokens/:pool/:id", handlers.ProviderAdmin.TokenUpdate)
 		authed.DELETE("/tokens/:pool/:id", handlers.ProviderAdmin.TokenDelete)

backend/internal/model/models.go

@@ -103,6 +103,10 @@ type ModelConfig struct {
 	Durations          datatypes.JSON    `gorm:"type:jsonb"`
 	MaxReferenceImages int               `gorm:"not null;default:0"`
 	ReferenceMode      string            `gorm:"size:32;not null;default:'none'"`
+	// Custom-upstream models (provider="custom"): UpstreamModel is the model name
+	// sent to the upstream OpenAI-compatible API; the base_url + key live on the
+	// matching custom account (pool="custom", meta.base_url). Empty for built-ins.
+	UpstreamModel string `gorm:"size:255;not null;default:''"`
 	// Weight controls display order in the model dropdown / admin list: higher
 	// weight floats to the top (matches ShowcaseItem.Weight semantics). Ties fall
 	// back to created_at desc. Default 0.
@@ -146,6 +150,13 @@ type TokenAccount struct {
 	VideoLimited bool `gorm:"not null;default:false"`
 	AccountEmail          string `gorm:"size:255"`
 	AccountDisplayName    string `gorm:"size:255"`
+	// Weight biases scheduling order for ANY account — higher weight is picked
+	// first within its pool (ties fall back to round-robin). Default 0.
+	Weight int `gorm:"not null;default:0"`
+	// Concurrency is the max simultaneous jobs for THIS account. Only custom
+	// (upstream) accounts honor it; built-in pools use their system default
+	// (1 per account, grok 10). 0 = use the system default.
+	Concurrency           int `gorm:"not null;default:0"`
 	CreatedAt             time.Time
 	UpdatedAt             time.Time
 }

backend/internal/provider/custom/client.go

@@ -0,0 +1,298 @@
+// Package custom implements a generic OpenAI-compatible upstream client. A
+// "custom" model forwards generation to any OpenAI-compatible API: the upstream
+// base_url + api_key live on a custom account (pool="custom"), the upstream model
+// name on the model config (UpstreamModel). Calls go DIRECT (no tls-client, no
+// proxy) — the upstream is a normal API with no anti-bot.
+package custom
+
+import (
+	"bytes"
+	"context"
+	"encoding/base64"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"mime/multipart"
+	"net/http"
+	"strings"
+	"time"
+)
+
+var (
+	ErrAuth              = errors.New("custom upstream auth failed")
+	ErrQuotaExhausted    = errors.New("custom upstream quota exhausted")
+	ErrTemporaryUpstream = errors.New("custom upstream temporary error")
+)
+
+type Client struct{}
+
+func NewClient() *Client { return &Client{} }
+
+func httpClient() *http.Client { return &http.Client{Timeout: 10 * time.Minute} }
+
+// GenerateImage calls the upstream OpenAI image API. With reference images it
+// uses /v1/images/edits (multipart); otherwise /v1/images/generations. Returns
+// the raw image bytes (decoded from b64_json, or downloaded from url).
+func (c *Client) GenerateImage(ctx context.Context, baseURL, apiKey, model, prompt, size, quality string, refs [][]byte) ([]byte, error) {
+	baseURL = strings.TrimRight(strings.TrimSpace(baseURL), "/")
+	if baseURL == "" || apiKey == "" {
+		return nil, ErrAuth
+	}
+	var req *http.Request
+	var err error
+	if len(refs) > 0 {
+		body := &bytes.Buffer{}
+		w := multipart.NewWriter(body)
+		_ = w.WriteField("model", model)
+		_ = w.WriteField("prompt", prompt)
+		if size != "" {
+			_ = w.WriteField("size", size)
+		}
+		for i, r := range refs {
+			fw, e := w.CreateFormFile("image[]", fmt.Sprintf("ref_%d.png", i+1))
+			if e != nil {
+				return nil, e
+			}
+			_, _ = fw.Write(r)
+		}
+		_ = w.Close()
+		req, err = http.NewRequest(http.MethodPost, baseURL+"/v1/images/edits", body)
+		if err != nil {
+			return nil, err
+		}
+		req.Header.Set("Content-Type", w.FormDataContentType())
+	} else {
+		payload := map[string]any{"model": model, "prompt": prompt, "n": 1}
+		if size != "" {
+			payload["size"] = size
+		}
+		if quality != "" {
+			payload["quality"] = quality
+		}
+		raw, _ := json.Marshal(payload)
+		req, err = http.NewRequest(http.MethodPost, baseURL+"/v1/images/generations", bytes.NewReader(raw))
+		if err != nil {
+			return nil, err
+		}
+		req.Header.Set("Content-Type", "application/json")
+	}
+	req = req.WithContext(ctx)
+	req.Header.Set("Authorization", "Bearer "+apiKey)
+
+	resp, err := httpClient().Do(req)
+	if err != nil {
+		return nil, fmt.Errorf("%w: %v", ErrTemporaryUpstream, err)
+	}
+	defer resp.Body.Close()
+	body, _ := io.ReadAll(resp.Body)
+	if e := mapStatus(resp.StatusCode, body); e != nil {
+		return nil, e
+	}
+	return imageBytesFromResponse(ctx, body)
+}
+
+// GenerateVideo drives the upstream Sora-style async video API:
+// POST /v1/videos → poll GET /v1/videos/{id} → GET /v1/videos/{id}/content.
+// When downloadResult is false it returns the upstream content URL instead.
+func (c *Client) GenerateVideo(ctx context.Context, baseURL, apiKey, model, prompt, size string, seconds int, downloadResult bool) ([]byte, string, error) {
+	baseURL = strings.TrimRight(strings.TrimSpace(baseURL), "/")
+	if baseURL == "" || apiKey == "" {
+		return nil, "", ErrAuth
+	}
+	payload := map[string]any{"model": model, "prompt": prompt}
+	if size != "" {
+		payload["size"] = size
+	}
+	if seconds > 0 {
+		payload["seconds"] = fmt.Sprintf("%d", seconds)
+	}
+	raw, _ := json.Marshal(payload)
+	created, err := c.doJSON(ctx, http.MethodPost, baseURL+"/v1/videos", apiKey, raw)
+	if err != nil {
+		return nil, "", err
+	}
+	jobID := strings.TrimSpace(stringValue(created["id"]))
+	if jobID == "" {
+		return nil, "", fmt.Errorf("%w: video create missing id", ErrTemporaryUpstream)
+	}
+	// Poll until terminal.
+	for {
+		if err := ctx.Err(); err != nil {
+			return nil, "", err
+		}
+		job, err := c.doJSON(ctx, http.MethodGet, baseURL+"/v1/videos/"+jobID, apiKey, nil)
+		if err != nil {
+			if errors.Is(err, ErrTemporaryUpstream) {
+				if sleepCtx(ctx, 5*time.Second) != nil {
+					return nil, "", ctx.Err()
+				}
+				continue
+			}
+			return nil, "", err
+		}
+		switch strings.ToLower(strings.TrimSpace(stringValue(job["status"]))) {
+		case "completed", "succeeded", "success":
+			contentURL := baseURL + "/v1/videos/" + jobID + "/content"
+			if !downloadResult {
+				return nil, contentURL, nil
+			}
+			data, err := c.download(ctx, contentURL, apiKey)
+			if err != nil {
+				return nil, "", err
+			}
+			return data, contentURL, nil
+		case "failed", "error", "canceled", "cancelled":
+			reason := stringValue(job["error"])
+			if isCreditError(reason) {
+				return nil, "", fmt.Errorf("%w: %s", ErrQuotaExhausted, clip([]byte(reason), 160))
+			}
+			return nil, "", fmt.Errorf("custom: video %s", clip([]byte(reason), 160))
+		}
+		if sleepCtx(ctx, 5*time.Second) != nil {
+			return nil, "", ctx.Err()
+		}
+	}
+}
+
+func (c *Client) doJSON(ctx context.Context, method, url, apiKey string, body []byte) (map[string]any, error) {
+	var reader io.Reader
+	if body != nil {
+		reader = bytes.NewReader(body)
+	}
+	req, err := http.NewRequest(method, url, reader)
+	if err != nil {
+		return nil, err
+	}
+	req = req.WithContext(ctx)
+	req.Header.Set("Authorization", "Bearer "+apiKey)
+	if body != nil {
+		req.Header.Set("Content-Type", "application/json")
+	}
+	resp, err := httpClient().Do(req)
+	if err != nil {
+		return nil, fmt.Errorf("%w: %v", ErrTemporaryUpstream, err)
+	}
+	defer resp.Body.Close()
+	raw, _ := io.ReadAll(resp.Body)
+	if e := mapStatus(resp.StatusCode, raw); e != nil {
+		return nil, e
+	}
+	var out map[string]any
+	if len(raw) == 0 {
+		return map[string]any{}, nil
+	}
+	if err := json.Unmarshal(raw, &out); err != nil {
+		return nil, fmt.Errorf("%w: non-json: %s", ErrTemporaryUpstream, clip(raw, 120))
+	}
+	return out, nil
+}
+
+func (c *Client) download(ctx context.Context, url, apiKey string) ([]byte, error) {
+	req, _ := http.NewRequest(http.MethodGet, url, nil)
+	req = req.WithContext(ctx)
+	req.Header.Set("Authorization", "Bearer "+apiKey)
+	resp, err := httpClient().Do(req)
+	if err != nil {
+		return nil, fmt.Errorf("%w: %v", ErrTemporaryUpstream, err)
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+		return nil, fmt.Errorf("%w: download %d", ErrTemporaryUpstream, resp.StatusCode)
+	}
+	data, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return nil, err
+	}
+	if len(data) == 0 {
+		return nil, fmt.Errorf("%w: empty download", ErrTemporaryUpstream)
+	}
+	return data, nil
+}
+
+// imageBytesFromResponse extracts image bytes from an OpenAI images response:
+// data[0].b64_json (preferred) or data[0].url (downloaded).
+func imageBytesFromResponse(ctx context.Context, body []byte) ([]byte, error) {
+	var out struct {
+		Data []struct {
+			B64JSON string `json:"b64_json"`
+			URL     string `json:"url"`
+		} `json:"data"`
+	}
+	if err := json.Unmarshal(body, &out); err != nil || len(out.Data) == 0 {
+		return nil, fmt.Errorf("%w: bad image response: %s", ErrTemporaryUpstream, clip(body, 160))
+	}
+	d := out.Data[0]
+	if d.B64JSON != "" {
+		raw, err := base64.StdEncoding.DecodeString(strings.TrimSpace(d.B64JSON))
+		if err != nil {
+			return nil, fmt.Errorf("%w: bad b64: %v", ErrTemporaryUpstream, err)
+		}
+		return raw, nil
+	}
+	if d.URL != "" {
+		req, _ := http.NewRequestWithContext(ctx, http.MethodGet, d.URL, nil)
+		resp, err := httpClient().Do(req)
+		if err != nil {
+			return nil, fmt.Errorf("%w: %v", ErrTemporaryUpstream, err)
+		}
+		defer resp.Body.Close()
+		return io.ReadAll(resp.Body)
+	}
+	return nil, fmt.Errorf("%w: image response had no b64/url", ErrTemporaryUpstream)
+}
+
+func mapStatus(status int, body []byte) error {
+	switch {
+	case status >= 200 && status < 300:
+		return nil
+	case status == 401 || status == 403:
+		return fmt.Errorf("%w: %d %s", ErrAuth, status, clip(body, 160))
+	case status == 429:
+		return fmt.Errorf("%w: 429 %s", ErrQuotaExhausted, clip(body, 160))
+	case status >= 500:
+		return fmt.Errorf("%w: %d %s", ErrTemporaryUpstream, status, clip(body, 160))
+	default:
+		if isCreditError(string(body)) {
+			return fmt.Errorf("%w: %s", ErrQuotaExhausted, clip(body, 160))
+		}
+		return fmt.Errorf("custom: %d %s", status, clip(body, 160))
+	}
+}
+
+func isCreditError(s string) bool {
+	s = strings.ToLower(s)
+	return strings.Contains(s, "insufficient") || strings.Contains(s, "quota") ||
+		strings.Contains(s, "credit") || strings.Contains(s, "balance")
+}
+
+func stringValue(v any) string {
+	switch x := v.(type) {
+	case string:
+		return x
+	case nil:
+		return ""
+	default:
+		b, _ := json.Marshal(x)
+		return strings.TrimSpace(string(b))
+	}
+}
+
+func clip(b []byte, n int) string {
+	s := strings.TrimSpace(string(b))
+	if len(s) > n {
+		return s[:n]
+	}
+	return s
+}
+
+func sleepCtx(ctx context.Context, d time.Duration) error {
+	t := time.NewTimer(d)
+	defer t.Stop()
+	select {
+	case <-ctx.Done():
+		return ctx.Err()
+	case <-t.C:
+		return nil
+	}
+}

backend/internal/provider/grok/client.go

@@ -290,29 +290,29 @@ func parseCreditsConfig(buf []byte) (remaining int, resetUnix string, ok bool) {
 	return 0, "", false
 }
 
-func scanConfigMessage(msg []byte) (remaining int, resetUnix string, ok bool) {
+func scanConfigMessage(msg []byte) (used int, resetUnix string, ok bool) {
-	var remF float32
+	var usedF float32
-	haveRem := false
+	seen := false
 	for len(msg) > 0 {
 		fn, wt, val, rest, good := readField(msg)
 		if !good {
 			break
 		}
 		msg = rest
+		seen = true
 		switch {
-		case fn == 1 && wt == 5: // float32 remaining credits
+		case fn == 1 && wt == 5: // float32 credits USED this period
-			remF = float32FromLE(val)
+			usedF = float32FromLE(val)
-			haveRem = true
 		case fn == 5 && wt == 2: // reset timestamp message { #1 varint=seconds }
 			if sec, sok := firstVarint(val); sok {
 				resetUnix = strconv.FormatInt(sec, 10)
 			}
 		}
 	}
-	if haveRem {
+	// A valid config message may OMIT field #1 when used == 0 (proto3 drops zero
-		return int(remF), resetUnix, true
+	// scalars) — a full-quota account. So as long as the message had any field,
-	}
+	// treat it as parsed with used defaulting to 0 (= 100 remaining).
-	return 0, resetUnix, false
+	return int(usedF), resetUnix, seen
 }
 
 // readField reads one protobuf field: returns (fieldNum, wireType, value, rest, ok).

backend/internal/service/tokens.go

@@ -34,6 +34,7 @@ var validTokenPools = map[string]string{
 	"krea":     "krea",
 	"imagine":  "imagine",
 	"grok":     "grok",
+	"custom":   "custom",
 }
 
 type TokenService struct {
@@ -947,6 +948,68 @@ func (s *TokenService) checkPendingGrok(tokenID, ssoToken string) {
 	s.finishPending(ctx, "grok", tokenID, "active", false, quotaMeta)
 }
 
+// ImportCustomAccount adds an upstream as a custom account: base_url + key, the
+// csv list of model ids it serves (empty = all), plus optional weight and
+// per-account concurrency. No probe — the account goes active immediately and is
+// matched to custom models by id at generation time. Calls go direct (no proxy).
+func (s *TokenService) ImportCustomAccount(ctx context.Context, baseURL, apiKey, models, name string, weight, concurrency int, tokenID string) (*model.TokenAccount, error) {
+	baseURL = strings.TrimRight(strings.TrimSpace(baseURL), "/")
+	apiKey = strings.TrimSpace(apiKey)
+	// Edit mode: tokenID points at an existing custom account. base_url required;
+	// a blank key keeps the stored one.
+	if strings.TrimSpace(tokenID) != "" {
+		existing, gerr := s.tokens.Get(ctx, "custom", tokenID)
+		if gerr != nil {
+			return nil, gerr
+		}
+		if baseURL == "" {
+			return nil, errors.New("base_url required")
+		}
+		meta := datatypes.JSONMap{"base_url": baseURL}
+		if m := strings.TrimSpace(models); m != "" {
+			meta["models"] = m
+		}
+		patch := map[string]any{"meta": meta, "weight": weight, "concurrency": concurrency, "account_email": strings.TrimSpace(name)}
+		if apiKey != "" {
+			patch["value"] = apiKey
+		}
+		item, uerr := s.tokens.Update(ctx, "custom", tokenID, patch)
+		if uerr != nil {
+			return nil, uerr
+		}
+		_ = existing
+		return item, nil
+	}
+	if baseURL == "" || apiKey == "" {
+		return nil, errors.New("base_url and key required")
+	}
+	meta := datatypes.JSONMap{"base_url": baseURL}
+	if m := strings.TrimSpace(models); m != "" {
+		meta["models"] = m
+	}
+	tokenID = newTokenID("custom")
+	item, err := s.createToken(ctx, "custom", tokenID, apiKey, "active", meta)
+	if err != nil {
+		return nil, err
+	}
+	patch := map[string]any{}
+	if strings.TrimSpace(name) != "" {
+		patch["account_email"] = strings.TrimSpace(name)
+	}
+	if weight != 0 {
+		patch["weight"] = weight
+	}
+	if concurrency > 0 {
+		patch["concurrency"] = concurrency
+	}
+	if len(patch) > 0 {
+		if updated, uerr := s.tokens.Update(ctx, "custom", tokenID, patch); uerr == nil {
+			item = updated
+		}
+	}
+	return item, nil
+}
+
 // finishPending writes the terminal status/dead flag and clears the pending_check
 // marker (merging any cached quota) for a background import probe.
 func (s *TokenService) finishPending(ctx context.Context, pool, id, status string, dead bool, quotaMeta map[string]any) {
@@ -1546,6 +1609,10 @@ func accountRow(item model.TokenAccount, inFlight int64) map[string]any {
 		"pending":           pending,
 		"quota_supported":   hasQuota,
 		"needs_reset_fetch": typeLabel == "adobe" && item.Status == "active" && strings.TrimSpace(item.CachedQuotaResetAfter) == "",
+		"weight":            item.Weight,
+		"concurrency":       item.Concurrency,
+		"base_url":          emptyToNil(strings.TrimSpace(stringValue(item.Meta["base_url"]))),
+		"models":            strings.TrimSpace(stringValue(item.Meta["models"])),
 	}
 }
 

backend/internal/service/v1.go

@@ -15,10 +15,13 @@ import (
 	"sync/atomic"
 	"time"
 
+	"strconv"
+
 	"backend/internal/config"
 	"backend/internal/model"
 	"backend/internal/provider/adobe"
 	"backend/internal/provider/chatgpt"
+	"backend/internal/provider/custom"
 	"backend/internal/provider/grok"
 	"backend/internal/provider/imagine"
 	"backend/internal/provider/krea"
@@ -70,6 +73,7 @@ type V1Service struct {
 	krea     *krea.Client
 	imagine  *imagine.Client
 	grok     *grok.Client
+	custom   *custom.Client
 	store    *storage.Client
 	// refresh re-mints an Adobe access token from its cookie when a request hits a
 	// 401 mid-flight (set via SetRefresh — wired after construction to avoid an
@@ -195,7 +199,7 @@ type V1VideoRequest struct {
 	BaseURL string
 }
 
-func NewV1Service(cfg *config.Config, models *repo.ModelRepository, users *repo.UserRepository, events *repo.EventRepository, tokens *repo.TokenRepository, settings *repo.SiteSettingRepository, adobeClient *adobe.Client, chatGPTClient *chatgpt.Client, runwayClient *runway.Client, leonardoClient *leonardo.Client, kreaClient *krea.Client, imagineClient *imagine.Client, grokClient *grok.Client, store *storage.Client) *V1Service {
+func NewV1Service(cfg *config.Config, models *repo.ModelRepository, users *repo.UserRepository, events *repo.EventRepository, tokens *repo.TokenRepository, settings *repo.SiteSettingRepository, adobeClient *adobe.Client, chatGPTClient *chatgpt.Client, runwayClient *runway.Client, leonardoClient *leonardo.Client, kreaClient *krea.Client, imagineClient *imagine.Client, grokClient *grok.Client, customClient *custom.Client, store *storage.Client) *V1Service {
 	return &V1Service{
 		cfg:      cfg,
 		models:   models,
@@ -210,6 +214,7 @@ func NewV1Service(cfg *config.Config, models *repo.ModelRepository, users *repo.
 		krea:     kreaClient,
 		imagine:  imagineClient,
 		grok:     grokClient,
+		custom:   customClient,
 		store:    store,
 		inflight: &InflightRegistry{},
 	}
@@ -352,7 +357,7 @@ func (s *V1Service) prepareImageExecution(ctx context.Context, principal *APIPri
 	startedAt := time.Now()
 
 	var imageBytes []byte
-	switch modelItem.Provider {
+	switch s.effectiveProvider(genCtx, modelItem) {
 	case "adobe":
 		b, execErr := s.generateAdobeImage(genCtx, eventID, modelItem, in, aspectRatio, resolution)
 		if execErr != nil {
@@ -455,6 +460,23 @@ func (s *V1Service) prepareImageExecution(ctx context.Context, principal *APIPri
 			}
 		}
 		imageBytes = b
+	case "custom":
+		b, execErr := s.generateCustomImage(genCtx, eventID, modelItem, in, aspectRatio, resolution)
+		if execErr != nil {
+			_ = s.refundIfNeeded(ctx, principal, eventID, price)
+			_ = s.events.UpdateStatus(ctx, eventID, "failed", execErr.Error(), 0)
+			switch {
+			case errors.Is(execErr, custom.ErrAuth):
+				return nil, ErrProviderAuth
+			case errors.Is(execErr, custom.ErrQuotaExhausted):
+				return nil, ErrProviderQuota
+			case errors.Is(execErr, custom.ErrTemporaryUpstream):
+				return nil, ErrProviderTemporary
+			default:
+				return nil, fmt.Errorf("%w: %v", ErrProviderExecution, execErr)
+			}
+		}
+		imageBytes = b
 	default:
 		_ = s.refundIfNeeded(ctx, principal, eventID, price)
 		_ = s.events.UpdateStatus(ctx, eventID, "failed", "provider not implemented", 0)
@@ -552,13 +574,15 @@ func (s *V1Service) prepareVideoExecution(ctx context.Context, principal *APIPri
 
 	var videoBytes []byte
 	var execErr error
-	switch modelItem.Provider {
+	switch s.effectiveProvider(genCtx, modelItem) {
 	case "adobe":
 		videoBytes, _, execErr = s.generateAdobeVideo(genCtx, eventID, modelItem, in, aspectRatio, resolution, parseDurationSeconds(duration), true)
 	case "runway":
 		videoBytes, _, execErr = s.generateRunwayVideo(genCtx, eventID, modelItem, in, aspectRatio, parseDurationSeconds(duration), true)
 	case "grok":
 		videoBytes, _, execErr = s.generateGrokVideo(genCtx, eventID, modelItem, in, aspectRatio, resolution, parseDurationSeconds(duration), true)
+	case "custom":
+		videoBytes, _, execErr = s.generateCustomVideo(genCtx, eventID, modelItem, in, aspectRatio, resolution, parseDurationSeconds(duration), true)
 	default:
 		_ = s.refundIfNeeded(ctx, principal, eventID, price)
 		_ = s.events.UpdateStatus(ctx, eventID, "failed", "provider not implemented", 0)
@@ -570,11 +594,11 @@ func (s *V1Service) prepareVideoExecution(ctx context.Context, principal *APIPri
 		switch {
 		case errors.Is(execErr, ErrNoProviderAccount):
 			return nil, ErrNoProviderAccount
-		case errors.Is(execErr, adobe.ErrAuth), errors.Is(execErr, runway.ErrAuth), errors.Is(execErr, grok.ErrAuth):
+		case errors.Is(execErr, adobe.ErrAuth), errors.Is(execErr, runway.ErrAuth), errors.Is(execErr, grok.ErrAuth), errors.Is(execErr, custom.ErrAuth):
 			return nil, ErrProviderAuth
-		case errors.Is(execErr, adobe.ErrQuotaExhausted), errors.Is(execErr, runway.ErrQuotaExhausted), errors.Is(execErr, grok.ErrQuotaExhausted):
+		case errors.Is(execErr, adobe.ErrQuotaExhausted), errors.Is(execErr, runway.ErrQuotaExhausted), errors.Is(execErr, grok.ErrQuotaExhausted), errors.Is(execErr, custom.ErrQuotaExhausted):
 			return nil, ErrProviderQuota
-		case errors.Is(execErr, adobe.ErrTemporaryUpstream), errors.Is(execErr, runway.ErrTemporaryUpstream), errors.Is(execErr, grok.ErrTemporaryUpstream):
+		case errors.Is(execErr, adobe.ErrTemporaryUpstream), errors.Is(execErr, runway.ErrTemporaryUpstream), errors.Is(execErr, grok.ErrTemporaryUpstream), errors.Is(execErr, custom.ErrTemporaryUpstream):
 			return nil, ErrProviderTemporary
 		default:
 			return nil, fmt.Errorf("%w: %v", ErrProviderExecution, execErr)
@@ -658,13 +682,15 @@ func (s *V1Service) runVideoJob(ctx context.Context, principal *APIPrincipal, in
 
 	var videoURL string
 	var execErr error
-	switch modelItem.Provider {
+	switch s.effectiveProvider(genCtx, modelItem) {
 	case "adobe":
 		_, videoURL, execErr = s.generateAdobeVideo(genCtx, eventID, modelItem, in, aspectRatio, resolution, parseDurationSeconds(duration), false)
 	case "runway":
 		_, videoURL, execErr = s.generateRunwayVideo(genCtx, eventID, modelItem, in, aspectRatio, parseDurationSeconds(duration), false)
 	case "grok":
 		_, videoURL, execErr = s.generateGrokVideo(genCtx, eventID, modelItem, in, aspectRatio, resolution, parseDurationSeconds(duration), false)
+	case "custom":
+		_, videoURL, execErr = s.generateCustomVideo(genCtx, eventID, modelItem, in, aspectRatio, resolution, parseDurationSeconds(duration), false)
 	default:
 		_ = s.refundIfNeeded(ctx, principal, eventID, price)
 		_ = s.events.UpdateStatus(ctx, eventID, "failed", "provider not implemented", 0)
@@ -859,12 +885,17 @@ func (s *V1Service) prepareImage(ctx context.Context, principal *APIPrincipal, i
 	if !modelItem.Enabled || modelItem.Type != "image" {
 		return nil, "", "", 0, ErrUnknownModel
 	}
-	// Fail fast before charging if the provider has no usable account.
+	// Fail fast before charging if the provider has no usable account. Use the
-	if ok, err := s.hasActiveProviderToken(ctx, modelItem.Provider, "image"); err != nil {
+	// effective provider: a custom upstream serving this model id routes to
+	// "custom" (effectiveProvider only returns it when such an account exists, so
+	// the precheck is satisfied); otherwise check the native provider pool.
+	if eff := s.effectiveProvider(ctx, modelItem); eff != "custom" {
+		if ok, err := s.hasActiveProviderToken(ctx, eff, "image"); err != nil {
 			return nil, "", "", 0, err
 		} else if !ok {
 			return nil, "", "", 0, ErrNoProviderAccount
 		}
+	}
 	refLimit := 0
 	if modelItem.ImageToImage {
 		refLimit = modelItem.MaxReferenceImages
@@ -921,8 +952,10 @@ func (s *V1Service) prepareVideo(ctx context.Context, principal *APIPrincipal, i
 	if !modelItem.Enabled || modelItem.Type != "video" {
 		return nil, "", "", "", 0, ErrUnknownModel
 	}
-	// Fail fast before charging if the provider has no usable account.
+	// Fail fast before charging — effective provider (custom upstream by id, else native).
-	if ok, err := s.hasActiveProviderToken(ctx, modelItem.Provider, "video"); err != nil {
+	if eff := s.effectiveProvider(ctx, modelItem); eff == "custom" {
+		// custom serves this id (effectiveProvider guaranteed it) — precheck ok
+	} else if ok, err := s.hasActiveProviderToken(ctx, eff, "video"); err != nil {
 		return nil, "", "", "", 0, err
 	} else if !ok {
 		return nil, "", "", "", 0, ErrNoProviderAccount
@@ -1534,6 +1567,246 @@ func (s *V1Service) generateRunwayVideo(ctx context.Context, eventID string, mod
 	return nil, "", lastErr
 }
 
+// customAccountServes reports whether a custom (upstream) account is usable for a
+// given model id: active, not dead, has a base_url, and its meta.models list (csv
+// of model ids it serves) contains the id. An empty models list serves ALL ids.
+func customAccountServes(item model.TokenAccount, modelID string) bool {
+	if item.Status != "active" || item.Dead || strings.TrimSpace(item.Value) == "" {
+		return false
+	}
+	if item.Meta == nil || strings.TrimSpace(stringValue(item.Meta["base_url"])) == "" {
+		return false
+	}
+	list := strings.TrimSpace(stringValue(item.Meta["models"]))
+	if list == "" {
+		return true
+	}
+	for _, m := range strings.Split(list, ",") {
+		if strings.EqualFold(strings.TrimSpace(m), modelID) {
+			return true
+		}
+	}
+	return false
+}
+
+// customActive returns the custom accounts that serve modelID, ordered by weight
+// (higher first; ties by id) so heavier upstreams are preferred.
+func (s *V1Service) customActive(ctx context.Context, modelID string) ([]model.TokenAccount, error) {
+	items, err := s.tokens.ListByPool(ctx, "custom")
+	if err != nil {
+		return nil, err
+	}
+	var active []model.TokenAccount
+	for _, item := range items {
+		if customAccountServes(item, modelID) {
+			active = append(active, item)
+		}
+	}
+	s.rotateRoundRobin("custom", active) // weight priority + round-robin within ties
+	return active, nil
+}
+
+// accountConcurrency is the per-account simultaneous-job cap. Custom accounts use
+// their configured Concurrency (default 1); built-in pools use the system value.
+func accountConcurrency(item model.TokenAccount) int {
+	if item.Concurrency > 0 {
+		return item.Concurrency
+	}
+	return 1
+}
+
+// effectiveProvider routes a model to the "custom" upstream whenever a custom
+// account declares it serves that model id (id-based override of the model's
+// native provider) — so an upstream can take over any model by matching its id.
+// Otherwise the model's own provider is used.
+func (s *V1Service) effectiveProvider(ctx context.Context, modelItem *model.ModelConfig) string {
+	if s.custom != nil {
+		if active, err := s.customActive(ctx, modelItem.ID); err == nil && len(active) > 0 {
+			return "custom"
+		}
+	}
+	return modelItem.Provider
+}
+
+// generateCustomImage forwards an image generation to an OpenAI-compatible
+// upstream. The upstream (custom account) is matched by model id; calls go direct
+// (no proxy). Billing uses the local model price.
+func (s *V1Service) generateCustomImage(ctx context.Context, eventID string, modelItem *model.ModelConfig, in V1ImageRequest, aspectRatio, resolution string) ([]byte, error) {
+	if s.custom == nil {
+		return nil, errors.New("custom client not configured")
+	}
+	refs, err := decodeReferenceImages(in.ReferenceImages, max(1, modelItem.MaxReferenceImages))
+	if err != nil {
+		return nil, err
+	}
+	active, err := s.customActive(ctx, modelItem.ID)
+	if err != nil {
+		return nil, err
+	}
+	if len(active) == 0 {
+		return nil, ErrNoProviderAccount
+	}
+	size := upstreamSize(aspectRatio, resolution)
+	quality := upstreamQuality(resolution)
+	var lastErr error
+	busy := 0
+	for _, token := range active {
+		if !s.gate.tryAcquireN(token.ID, accountConcurrency(token)) {
+			busy++
+			continue
+		}
+		var data []byte
+		done, failover := func() (bool, bool) {
+			defer s.gate.release(token.ID)
+			_ = s.events.SetAccount(ctx, eventID, token.ID)
+			_ = s.tokens.TouchLastUsed(ctx, token.ID)
+			baseURL := stringValue(token.Meta["base_url"])
+			d, genErr := s.custom.GenerateImage(ctx, baseURL, token.Value, modelItem.ID, in.Prompt, size, quality, refs)
+			if genErr == nil {
+				_, _ = s.tokens.Update(ctx, "custom", token.ID, map[string]any{
+					"last_used_at": time.Now(), "success_total": gorm.Expr("success_total + 1"), "fails": 0,
+				})
+				data = d
+				return true, false
+			}
+			lastErr = genErr
+			switch {
+			case errors.Is(genErr, custom.ErrAuth):
+				s.markTokenFailure(ctx, "custom", token, "image", true, false)
+				return false, true
+			case errors.Is(genErr, custom.ErrQuotaExhausted):
+				s.markTokenFailure(ctx, "custom", token, "image", false, true)
+				return false, true
+			case errors.Is(genErr, custom.ErrTemporaryUpstream):
+				return false, true
+			default:
+				return false, false
+			}
+		}()
+		if done {
+			return data, nil
+		}
+		if failover {
+			continue
+		}
+		return nil, lastErr
+	}
+	if lastErr == nil {
+		if busy > 0 {
+			return nil, ErrConcurrencyFull
+		}
+		lastErr = ErrProviderExecution
+	}
+	return nil, lastErr
+}
+
+// generateCustomVideo forwards a video generation to an OpenAI-compatible
+// (Sora-style) upstream, matched by model id. No proxy; local-price billing.
+func (s *V1Service) generateCustomVideo(ctx context.Context, eventID string, modelItem *model.ModelConfig, in V1VideoRequest, aspectRatio, resolution string, durationSeconds int, downloadResult bool) ([]byte, string, error) {
+	if s.custom == nil {
+		return nil, "", errors.New("custom client not configured")
+	}
+	active, err := s.customActive(ctx, modelItem.ID)
+	if err != nil {
+		return nil, "", err
+	}
+	if len(active) == 0 {
+		return nil, "", ErrNoProviderAccount
+	}
+	size := upstreamSize(aspectRatio, resolution)
+	var lastErr error
+	var videoURL string
+	busy := 0
+	for _, token := range active {
+		if !s.gate.tryAcquireN(token.ID, accountConcurrency(token)) {
+			busy++
+			continue
+		}
+		var data []byte
+		done, failover := func() (bool, bool) {
+			defer s.gate.release(token.ID)
+			_ = s.events.SetAccount(ctx, eventID, token.ID)
+			_ = s.tokens.TouchLastUsed(ctx, token.ID)
+			baseURL := stringValue(token.Meta["base_url"])
+			d, url, genErr := s.custom.GenerateVideo(ctx, baseURL, token.Value, modelItem.ID, in.Prompt, size, durationSeconds, downloadResult)
+			if genErr == nil {
+				_, _ = s.tokens.Update(ctx, "custom", token.ID, map[string]any{
+					"last_used_at": time.Now(), "success_total": gorm.Expr("success_total + 1"), "fails": 0,
+				})
+				data = d
+				videoURL = url
+				return true, false
+			}
+			lastErr = genErr
+			switch {
+			case errors.Is(genErr, custom.ErrAuth):
+				s.markTokenFailure(ctx, "custom", token, "video", true, false)
+				return false, true
+			case errors.Is(genErr, custom.ErrQuotaExhausted):
+				s.markTokenFailure(ctx, "custom", token, "video", false, true)
+				return false, true
+			case errors.Is(genErr, custom.ErrTemporaryUpstream):
+				return false, true
+			default:
+				return false, false
+			}
+		}()
+		if done {
+			return data, videoURL, nil
+		}
+		if failover {
+			continue
+		}
+		return nil, "", lastErr
+	}
+	if lastErr == nil {
+		if busy > 0 {
+			return nil, "", ErrConcurrencyFull
+		}
+		lastErr = ErrProviderExecution
+	}
+	return nil, "", lastErr
+}
+
+// upstreamSize maps our (ratio, resolution) to an OpenAI-style "WxH" size string
+// for the upstream. The pixel base scales with the tier (1K/2K/4K); the ratio
+// sets the shape. Upstreams that key off ratio (our own /v1) read it fine.
+func upstreamSize(aspectRatio, resolution string) string {
+	base := 1024
+	switch strings.ToUpper(strings.TrimSpace(resolution)) {
+	case "2K":
+		base = 2048
+	case "4K":
+		base = 4096
+	}
+	w, h := 1, 1
+	parts := strings.Split(strings.ReplaceAll(strings.TrimSpace(aspectRatio), "x", ":"), ":")
+	if len(parts) == 2 {
+		if a, e1 := strconv.Atoi(strings.TrimSpace(parts[0])); e1 == nil && a > 0 {
+			if b, e2 := strconv.Atoi(strings.TrimSpace(parts[1])); e2 == nil && b > 0 {
+				w, h = a, b
+			}
+		}
+	}
+	if w >= h {
+		return fmt.Sprintf("%dx%d", base, base*h/w)
+	}
+	return fmt.Sprintf("%dx%d", base*w/h, base)
+}
+
+// upstreamQuality maps a resolution tier to the OpenAI quality enum.
+func upstreamQuality(resolution string) string {
+	switch strings.ToUpper(strings.TrimSpace(resolution)) {
+	case "2K":
+		return "medium"
+	case "4K":
+		return "high"
+	case "1K":
+		return "low"
+	}
+	return ""
+}
+
 // generateGrokVideo runs grok's imagine video pipeline across the grok pool.
 // Mirrors the runway policy: no pre-deduct, skip accounts known out of credits
 // (cached remaining <= 0), and treat an out-of-credits / auth failure as a dead
@@ -2567,19 +2840,35 @@ func (s *V1Service) nextCursor(pool string) uint64 {
 // retry chain is preserved — on failure the caller's loop simply continues to
 // the next account in rotation order.
 func (s *V1Service) rotateRoundRobin(pool string, items []model.TokenAccount) {
+	if len(items) <= 1 {
+		return
+	}
+	// Weight = priority: higher-weight accounts come first, so the scheduler tries
+	// them before lower-weight ones (and only falls through when they're at their
+	// concurrency cap). Within the SAME weight all accounts are equal, so they're
+	// rotated by the pool cursor for even distribution.
 	sort.SliceStable(items, func(i, j int) bool {
+		if items[i].Weight != items[j].Weight {
+			return items[i].Weight > items[j].Weight
+		}
 		return items[i].ID < items[j].ID
 	})
-	n := len(items)
+	start := int(s.nextCursor(pool))
-	if n <= 1 {
+	for i := 0; i < len(items); {
-		return
+		j := i + 1
+		for j < len(items) && items[j].Weight == items[i].Weight {
+			j++
 		}
-	start := int(s.nextCursor(pool) % uint64(n))
+		if g := j - i; g > 1 {
-	if start == 0 {
+			off := start % g
-		return
+			if off != 0 {
+				grp := items[i:j]
+				rot := make([]model.TokenAccount, 0, g)
+				rot = append(rot, grp[off:]...)
+				rot = append(rot, grp[:off]...)
+				copy(grp, rot)
+			}
+		}
+		i = j
 	}
-	rotated := make([]model.TokenAccount, 0, n)
-	rotated = append(rotated, items[start:]...)
-	rotated = append(rotated, items[:start]...)
-	copy(items, rotated)
 }

frontend/src/components/CustomModelModal.vue

@@ -0,0 +1,172 @@
+<script setup>
+import { ref, computed, watch } from 'vue'
+import { api, jsonBody } from '../api'
+import Icon from './Icon.vue'
+import SelectMenu from './SelectMenu.vue'
+
+const emit = defineEmits(['close', 'saved'])
+
+const RATIO_OPTS = ['1:1', '16:9', '9:16', '4:3', '3:4', '21:9', '3:2', '5:4', '4:5', '2:3', '2:1']
+const IMG_RES = ['1K', '2K', '4K']
+const VID_RES = ['720p', '1080p', '2K', '4K']
+const ALL_RES = ['1K', '2K', '4K', '720p', '1080p']
+const DUR_OPTS = ['5s', '6s', '8s', '10s']
+
+const id = ref('')
+const type = ref('image')
+const ratios = ref(['1:1', '16:9', '9:16'])
+const maxRefs = ref(0)
+const refMode = ref('none')
+const weight = ref(0)
+// tier -> { price, agent } — blank price means the tier is NOT supported.
+const res = ref(Object.fromEntries(ALL_RES.map((r) => [r, { price: '', agent: '' }])))
+const dur = ref(Object.fromEntries(DUR_OPTS.map((d) => [d, { price: '', agent: '' }])))
+const error = ref('')
+const saving = ref(false)
+
+const isVideo = computed(() => type.value === 'video')
+// Resolution tiers depend on type: image = 1K/2K/4K, video = 540p/720p/1080p/2K/4K.
+const resOpts = computed(() => (isVideo.value ? VID_RES : IMG_RES))
+
+// 首尾帧(frame) only has first+last slots → cap reference images at 2.
+const refsCap = computed(() => (refMode.value === 'frame' ? 2 : 99))
+watch([refMode, maxRefs], () => {
+  if (refMode.value === 'frame' && Number(maxRefs.value) > 2) maxRefs.value = 2
+})
+
+function toggleRatio(r) {
+  const i = ratios.value.indexOf(r)
+  if (i >= 0) ratios.value.splice(i, 1)
+  else ratios.value.push(r)
+}
+
+// collect checked tiers into { tier: price } and { tier: agentPrice }
+// blank price = tier not supported (skipped), matching the edit form.
+function collect(tiers, allowed) {
+  const prices = {}, agent = {}, keys = []
+  for (const [k, v] of Object.entries(tiers)) {
+    if (allowed && !allowed.includes(k)) continue
+    const raw = String(v.price ?? '').trim()
+    if (raw === '') continue
+    const n = Number(raw)
+    if (Number.isNaN(n) || n < 0) continue
+    prices[k] = n; keys.push(k)
+    const ar = String(v.agent ?? '').trim()
+    if (ar !== '') { const a = Number(ar); if (!Number.isNaN(a) && a >= 0) agent[k] = a }
+  }
+  return { prices, agent, keys }
+}
+
+async function save() {
+  const mid = id.value.trim()
+  if (!mid) { error.value = '请填写模型 id'; return }
+  if (refMode.value === 'frame' && Number(maxRefs.value) > 2) maxRefs.value = 2
+  const r = collect(res.value, resOpts.value)
+  if (!r.keys.length) { error.value = '请至少勾选一个分辨率并填价格'; return }
+  const body = {
+    id: mid,
+    name: mid,
+    type: type.value,
+    provider: 'custom',
+    prices: r.prices,
+    prices_agent: r.agent,
+    ratios: ratios.value.slice(),
+    max_reference_images: Number(maxRefs.value) || 0,
+    reference_mode: refMode.value,
+    weight: Number(weight.value) || 0,
+    image_to_image: (Number(maxRefs.value) || 0) > 0,
+  }
+  if (isVideo.value) {
+    body.resolutions = r.keys
+    const d = collect(dur.value)
+    if (!d.keys.length) { error.value = '视频请至少勾选一个时长并填价格'; return }
+    body.duration_prices = d.prices
+    body.duration_prices_agent = d.agent
+    body.durations = d.keys
+  }
+  saving.value = true; error.value = ''
+  try {
+    const resp = await api('/managed-models', jsonBody('POST', body))
+    if (resp.ok || resp.data?.ok || resp.status === 200) emit('saved')
+    else error.value = resp.data?.detail || '创建失败'
+  } catch (e) { error.value = String(e) }
+  finally { saving.value = false }
+}
+</script>
+
+<template>
+  <div class="fixed inset-0 z-50 bg-slate-900/40 backdrop-blur-sm flex items-start justify-center overflow-y-auto p-4"
+       @click.self="emit('close')">
+    <div class="card !shadow-xl mt-10 mb-10 w-full max-w-xl">
+      <div class="px-5 py-4 border-b border-slate-100 flex items-center justify-between">
+        <h2 class="text-sm font-semibold">添加自定义模型(上游 / provider=custom)</h2>
+        <button @click="emit('close')" class="text-slate-400 hover:text-slate-700"><Icon name="close" class="w-5 h-5" /></button>
+      </div>
+      <div class="p-5 space-y-4">
+        <p class="text-xs text-slate-500 leading-relaxed">
+          id 要与上游模型名<strong class="text-slate-700">一致</strong> —— 生成时按 id 自动路由到「支持该 id 的上游账号」。价格按本地价计费。
+        </p>
+
+        <div class="flex gap-3">
+          <div class="flex-1">
+            <label class="text-xs text-slate-500 block mb-1">模型 id <span class="text-rose-500">*</span></label>
+            <input v-model="id" class="field font-mono text-xs h-10" placeholder="gpt-image-2" />
+          </div>
+          <div class="w-28">
+            <label class="text-xs text-slate-500 block mb-1">类型</label>
+            <SelectMenu v-model="type" :options="[{value:'image',label:'图像'},{value:'video',label:'视频'}]" />
+          </div>
+          <div class="w-24">
+            <label class="text-xs text-slate-500 block mb-1">权重</label>
+            <input v-model.number="weight" type="number" class="field h-10" placeholder="0" />
+          </div>
+        </div>
+
+        <div>
+          <label class="text-xs text-slate-500 block mb-1.5">比例(多选)</label>
+          <div class="flex flex-wrap gap-1.5">
+            <button v-for="r in RATIO_OPTS" :key="r" type="button" @click="toggleRatio(r)"
+                    class="px-2 py-1 rounded text-xs ring-1 transition-colors"
+                    :class="ratios.includes(r) ? 'bg-indigo-500/15 text-indigo-600 ring-indigo-300' : 'bg-slate-50 text-slate-500 ring-slate-200 hover:ring-slate-300'">{{ r }}</button>
+          </div>
+        </div>
+
+        <div>
+          <label class="text-xs text-slate-500 block mb-1.5">分辨率 · 价格(填普通价 = 支持该档,<strong class="text-slate-600">留空 = 不支持</strong>)</label>
+          <div class="space-y-1.5">
+            <div v-for="r in resOpts" :key="r" class="flex items-center gap-2">
+              <span class="w-16 text-xs font-mono text-slate-500">{{ r }}</span>
+              <input v-model="res[r].price" type="number" class="field !py-1 flex-1" placeholder="普通价(留空=不支持)" />
+              <input v-model="res[r].agent" type="number" class="field !py-1 flex-1" placeholder="代理价(留空跟随)" />
+            </div>
+          </div>
+        </div>
+
+        <div v-if="isVideo">
+          <label class="text-xs text-slate-500 block mb-1.5">时长 · 价格(总价 = 分辨率价 + 时长价;<strong class="text-slate-600">留空 = 不支持</strong>)</label>
+          <div class="space-y-1.5">
+            <div v-for="d in DUR_OPTS" :key="d" class="flex items-center gap-2">
+              <span class="w-16 text-xs font-mono text-slate-500">{{ d }}</span>
+              <input v-model="dur[d].price" type="number" class="field !py-1 flex-1" placeholder="普通价(留空=不支持)" />
+              <input v-model="dur[d].agent" type="number" class="field !py-1 flex-1" placeholder="代理价(留空跟随)" />
+            </div>
+          </div>
+        </div>
+
+        <div class="flex gap-3">
+          <div class="flex-1">
+            <label class="text-xs text-slate-500 block mb-1">参考图张数<span v-if="refMode==='frame'" class="text-white/40">(首尾帧最多 2)</span></label>
+            <input v-model.number="maxRefs" type="number" min="0" :max="refsCap" class="field h-10" />
+          </div>
+          <div class="flex-1">
+            <label class="text-xs text-slate-500 block mb-1">参考模式</label>
+            <SelectMenu v-model="refMode" :options="[{value:'none',label:'无'},{value:'asset',label:'参考图'},{value:'frame',label:'首尾帧(视频)'}]" />
+          </div>
+        </div>
+
+        <button @click="save" :disabled="saving" class="btn-primary w-full">{{ saving ? '创建中…' : '创建模型' }}</button>
+        <p v-if="error" class="text-xs text-rose-600">{{ error }}</p>
+      </div>
+    </div>
+  </div>
+</template>

frontend/src/components/ImportModal.vue

@@ -7,10 +7,14 @@ import Icon from './Icon.vue'
 const emit = defineEmits(['close', 'imported'])
 
 const input = ref('')
+const weight = ref(0)
 const status = ref('')
 const isError = ref(false)
 const submitting = ref(false)
 
+// type → token pool (for the post-import weight PATCH).
+const TYPE_POOL = { openai: 'chatgpt', adobe: 'adobe', runway: 'runway', leonardo: 'leonardo', krea: 'krea', imagine: 'imagine', grok: 'grok' }
+
 // Live preview of what the parser would extract — updates as the user types
 // so they can see whether their paste was understood before clicking import.
 const detected = computed(() => {
@@ -56,8 +60,15 @@ async function doSmartImport() {
               : it.type === 'imagine'
                 ? await api('/tokens/import-imagine-token', jsonBody('POST', { value: it.value }))
                 : await api('/tokens/import-adobe-cookie', jsonBody('POST', { cookie: it.value }))
-      if (r.ok) ok++
+      if (r.ok) {
-      else { fail++; errs.push(`${it.type}: ${r.data?.detail || r.status}`) }
+        ok++
+        // Apply the chosen weight to the freshly-imported account (best-effort).
+        const w = Number(weight.value) || 0
+        const pool = TYPE_POOL[it.type]
+        if (w !== 0 && pool && r.data?.id) {
+          try { await api(`/tokens/${pool}/${r.data.id}`, jsonBody('PATCH', { weight: w })) } catch (_) {}
+        }
+      } else { fail++; errs.push(`${it.type}: ${r.data?.detail || r.status}`) }
     } catch (e) {
       fail++; errs.push(`${it.type}: ${e}`)
     }
@@ -131,6 +142,10 @@ async function doSmartImport() {
           </template>
           <span v-else class="text-rose-600">未识别到任何 Cookie 或 JWT</span>
         </div>
+        <div class="mt-3 flex items-center gap-2">
+          <label class="text-xs text-slate-500 whitespace-nowrap">权重(本批账号,高的优先)</label>
+          <input v-model.number="weight" type="number" class="field !w-24" placeholder="0" />
+        </div>
         <button @click="doSmartImport" :disabled="submitting || !detected.total" class="btn-primary w-full mt-3">
           {{ submitting ? '导入中…' : (detected.total ? `识别并导入 (${detected.total})` : '识别并导入') }}
         </button>

frontend/src/components/SelectMenu.vue

@@ -73,7 +73,7 @@ onUnmounted(() => document.removeEventListener('mousedown', onDocClick))
     <button type="button" @click="toggle" @keydown="onKeydown"
             :aria-expanded="open"
             :disabled="disabled"
-            class="field flex items-center justify-between gap-2 text-left disabled:opacity-50 disabled:cursor-not-allowed"
+            class="field flex items-center justify-between gap-2 text-left h-10 !py-0 disabled:opacity-50 disabled:cursor-not-allowed"
             :class="[mono ? 'font-mono' : '', selected ? '' : 'text-[color:var(--fg-faint)]']">
       <span class="truncate">{{ label }}</span>
       <Icon name="chevron"

frontend/src/components/TestModal.vue

@@ -142,7 +142,7 @@ async function run() {
     busy.value = false
     resultUrl.value = r.data.url
     resultKind.value = r.data.kind || (isVideo ? 'video' : 'image')
-    status.value = `完成 · ${r.data.provider} · ${r.data.elapsed_ms}ms`
+    status.value = `完成 · ${r.data.provider} · ${(r.data.elapsed_ms / 1000).toFixed(1)}s`
   } else if (GATEWAY_TIMEOUT.has(r.status)) {
     // CDN/代理回源超时(如 EdgeOne 524)—— 后端仍在生成。保持锁住,轮询恢复结果。
     status.value = isVideo ? '生成视频中 (约 1–3 分钟)…' : '生成中…'

frontend/src/components/UpstreamModal.vue

@@ -0,0 +1,123 @@
+<script setup>
+import { ref, onMounted } from 'vue'
+import { api, jsonBody } from '../api'
+import Icon from './Icon.vue'
+
+const props = defineProps({ account: { type: Object, default: null } }) // edit mode when set
+const emit = defineEmits(['close', 'imported'])
+
+const isEdit = !!props.account
+const name = ref(props.account?.email || '')
+const baseUrl = ref(props.account?.base_url || '')
+const key = ref('')            // edit: blank = keep existing key
+const allModels = ref([])      // existing models to pick from
+const selected = ref(props.account?.models ? String(props.account.models).split(',').map((x) => x.trim()).filter(Boolean) : [])
+const weight = ref(Number(props.account?.weight) || 0)
+const concurrency = ref(Number(props.account?.concurrency) || 1)
+const status = ref('')
+const isError = ref(false)
+const submitting = ref(false)
+
+onMounted(async () => {
+  try {
+    const r = await api('/managed-models')
+    allModels.value = (r.data?.data || []).map((m) => ({ id: m.id, type: m.type }))
+  } catch (_) {}
+})
+
+function toggle(id) {
+  const i = selected.value.indexOf(id)
+  if (i >= 0) selected.value.splice(i, 1)
+  else selected.value.push(id)
+}
+
+async function submit() {
+  if (!baseUrl.value.trim() || (!isEdit && !key.value.trim())) {
+    status.value = isEdit ? '请填写 URL' : '请填写 URL 和 Key'; isError.value = true; return
+  }
+  submitting.value = true; status.value = ''; isError.value = false
+  try {
+    const r = await api('/tokens/import-custom-account', jsonBody('POST', {
+      id: isEdit ? props.account.id : undefined,
+      name: name.value.trim(),
+      base_url: baseUrl.value.trim(),
+      key: key.value.trim(),       // blank in edit = keep existing
+      models: selected.value.join(','),
+      weight: Number(weight.value) || 0,
+      concurrency: Number(concurrency.value) || 1,
+    }))
+    if (r.ok) {
+      status.value = isEdit ? '✓ 已保存' : '✓ 已添加上游'; emit('imported')
+      setTimeout(() => emit('close'), 700)
+    } else {
+      status.value = r.data?.detail || '保存失败'; isError.value = true
+    }
+  } catch (e) {
+    status.value = String(e); isError.value = true
+  } finally {
+    submitting.value = false
+  }
+}
+</script>
+
+<template>
+  <div class="fixed inset-0 z-50 bg-slate-900/40 backdrop-blur-sm flex items-start justify-center overflow-y-auto p-4"
+       @click.self="emit('close')">
+    <div class="card !shadow-xl mt-14 mb-14 w-full max-w-lg">
+      <div class="px-5 py-4 border-b border-slate-100 flex items-center justify-between">
+        <h2 class="text-sm font-semibold">{{ isEdit ? '编辑上游' : '添加上游(自定义 OpenAI 兼容)' }}</h2>
+        <button @click="emit('close')" class="text-slate-400 hover:text-slate-700 transition-colors">
+          <Icon name="close" class="w-5 h-5" />
+        </button>
+      </div>
+      <div class="p-5 space-y-3">
+        <p class="text-xs text-slate-500 leading-relaxed">
+          上游就是一个账号:填 v1 URL + Key。模型按 <strong class="text-slate-700">id 相同</strong>自动路由 ——
+          在「模型管理」加一个 provider=custom、id 与上游一致的模型即可从这个上游调用。调用<strong class="text-slate-700">直连不走代理</strong>。
+        </p>
+        <div>
+          <label class="text-xs text-slate-500">备注名</label>
+          <input v-model="name" class="field" placeholder="例如:我的中转 / xx-api" />
+        </div>
+        <div>
+          <label class="text-xs text-slate-500">v1 URL <span class="text-rose-500">*</span></label>
+          <input v-model="baseUrl" class="field font-mono text-xs" placeholder="https://api.example.com(无需 /v1 结尾)" />
+        </div>
+        <div>
+          <label class="text-xs text-slate-500">Key <span v-if="!isEdit" class="text-rose-500">*</span><span v-else class="text-white/40">(留空=不改)</span></label>
+          <input v-model="key" class="field font-mono text-xs" :placeholder="isEdit ? '留空保持原 key' : 'sk-...'" />
+        </div>
+        <div>
+          <div class="flex items-center justify-between mb-1.5">
+            <label class="text-xs text-slate-500">支持的模型(多选,不选 = 全部)</label>
+            <span class="text-[11px] text-slate-400">{{ selected.length ? `已选 ${selected.length}` : '全部' }}</span>
+          </div>
+          <div v-if="!allModels.length" class="text-xs text-slate-400 rounded-lg ring-1 ring-slate-200 bg-slate-50/60 p-3">
+            暂无模型 —— 可先去模型管理加自定义模型
+          </div>
+          <div v-else class="flex flex-wrap gap-1.5 max-h-44 overflow-y-auto rounded-lg ring-1 ring-slate-200 bg-slate-50/60 p-2">
+            <button v-for="m in allModels" :key="m.id" type="button" @click="toggle(m.id)"
+                    class="inline-flex items-center gap-1.5 px-2.5 py-1 rounded-md text-xs ring-1 transition-colors"
+                    :class="selected.includes(m.id) ? 'bg-indigo-500/15 text-indigo-700 ring-indigo-300 font-medium' : 'bg-white text-slate-600 ring-slate-200 hover:ring-slate-300'">
+              <span class="w-1.5 h-1.5 rounded-full" :class="m.type === 'video' ? 'bg-violet-400' : 'bg-emerald-400'"></span>{{ m.id }}
+            </button>
+          </div>
+        </div>
+        <div class="flex gap-3">
+          <div class="flex-1">
+            <label class="text-xs text-slate-500">权重(高的优先)</label>
+            <input v-model.number="weight" type="number" class="field" placeholder="0" />
+          </div>
+          <div class="flex-1">
+            <label class="text-xs text-slate-500">并发数(单账号)</label>
+            <input v-model.number="concurrency" type="number" min="1" class="field" placeholder="1" />
+          </div>
+        </div>
+        <button @click="submit" :disabled="submitting" class="btn-primary w-full mt-1">
+          {{ submitting ? '保存中…' : (isEdit ? '保存' : '添加上游') }}
+        </button>
+        <p v-if="status" class="text-xs" :class="isError ? 'text-rose-600' : 'text-emerald-600'">{{ status }}</p>
+      </div>
+    </div>
+  </div>
+</template>

frontend/src/views/AccountsView.vue

@@ -3,12 +3,16 @@ import { ref, computed, onMounted, watch } from 'vue'
 import { api, jsonBody } from '../api'
 import { fmtTs, fmtIso, fmtDate, fmtClock } from '../utils/format'
 import ImportModal from '../components/ImportModal.vue'
+import UpstreamModal from '../components/UpstreamModal.vue'
 import Icon from '../components/Icon.vue'
 
 const rows = ref([])
 const loading = ref(false)
 const quotaStatus = ref('')
 const showImport = ref(false)
+const showUpstream = ref(false)
+const editingUpstream = ref(null)
+function editUpstream(a) { editingUpstream.value = a; showUpstream.value = true }
 
 const typeFilter = ref('')      // '' | 'openai' | 'adobe' | 'runway' | 'leonardo'
 const statusFilter = ref('')    // '' | 'active' | 'quota' | 'disabled'
@@ -97,6 +101,13 @@ const pageNumbers = computed(() => {
 
 let pendingTimer = null
 
+// Inline-edit an account field (weight / concurrency) → PATCH /tokens/{pool}/{id}.
+async function saveField(row, field, value) {
+  const v = Number(value)
+  if (Number.isNaN(v)) return
+  await api(`/tokens/${row.pool}/${row.id}`, jsonBody('PATCH', { [field]: v }))
+}
+
 async function loadAccounts() {
   loading.value = true
   quotaStatus.value = ''
@@ -361,10 +372,13 @@ onMounted(loadAccounts)
       <button @click="showImport = true" class="btn-primary">
         <Icon name="plus" class="w-3.5 h-3.5" /> 导入账号
       </button>
+      <button @click="editingUpstream = null; showUpstream = true" class="btn-soft">
+        <Icon name="plus" class="w-3.5 h-3.5" /> 添加上游
+      </button>
     </div>
 
     <!-- Table -->
-    <div class="card overflow-hidden">
+    <div class="card overflow-x-auto">
       <div v-if="loading && !rows.length" class="text-center text-sm text-white/40 py-20">加载中…</div>
       <div v-else-if="!filtered.length" class="flex flex-col items-center gap-3 text-white/40 py-20">
         <span class="w-14 h-14 rounded-2xl bg-white/[0.04] grid place-items-center">
@@ -374,18 +388,20 @@ onMounted(loadAccounts)
         <button v-if="!rows.length" @click="showImport = true" class="btn-soft mt-1">导入第一个</button>
       </div>
 
-      <table v-else class="w-full text-sm table-fixed">
+      <table v-else class="w-full text-sm table-fixed min-w-[1040px]">
         <colgroup>
           <col class="w-9" />      <!-- select -->
           <col />                  <!-- identity (flex) -->
           <col class="w-20" />     <!-- type -->
           <col class="w-24" />     <!-- remaining -->
+          <col class="w-16" />     <!-- weight -->
+          <col class="w-16" />     <!-- concurrency -->
           <col class="w-32" />     <!-- reset -->
           <col class="w-28" />     <!-- created -->
           <col class="w-28" />     <!-- last used -->
           <col class="w-40" />     <!-- inflight/success/fail -->
           <col class="w-16" />     <!-- status switch -->
-          <col class="w-16" />     <!-- actions -->
+          <col class="w-24" />     <!-- actions -->
         </colgroup>
         <thead>
           <tr class="text-[10px] uppercase tracking-[0.2em] text-white/40 border-b border-white/[0.06]">
@@ -396,6 +412,8 @@ onMounted(loadAccounts)
             <th class="text-left px-5 py-3 font-medium">账户</th>
             <th class="text-left px-3 py-3 font-medium">类型</th>
             <th class="text-right px-3 py-3 font-medium">额度</th>
+            <th class="text-center px-3 py-3 font-medium">权重</th>
+            <th class="text-center px-3 py-3 font-medium">并发</th>
             <th class="text-left px-3 py-3 font-medium">恢复时间</th>
             <th class="text-left px-3 py-3 font-medium">创建时间</th>
             <th class="text-left px-3 py-3 font-medium">最后使用</th>
@@ -445,6 +463,20 @@ onMounted(loadAccounts)
                     :class="a.remaining > 0 ? 'text-emerald-300' : 'text-rose-300'">{{ a.remaining }}{{ a.type === 'grok' ? '%' : '' }}</span>
               <span v-else class="text-white/25" :title="a._quotaError || ''">—</span>
             </td>
+            <!-- weight (editable, all accounts) -->
+            <td class="px-3 py-3.5 align-middle text-center whitespace-nowrap">
+              <input type="number" :value="a.weight" @change="saveField(a, 'weight', $event.target.value)"
+                     class="w-12 bg-white/5 border border-white/10 rounded px-1 py-0.5 text-xs text-center tabular-nums focus:border-indigo-400 outline-none"
+                     title="权重(高的优先)" />
+            </td>
+            <!-- concurrency (editable only for custom upstreams; others = system fixed) -->
+            <td class="px-3 py-3.5 align-middle text-center whitespace-nowrap">
+              <input v-if="a.type === 'custom'" type="number" min="1" :value="a.concurrency || 1"
+                     @change="saveField(a, 'concurrency', $event.target.value)"
+                     class="w-12 bg-white/5 border border-white/10 rounded px-1 py-0.5 text-xs text-center tabular-nums focus:border-indigo-400 outline-none"
+                     title="并发数(仅上游可设)" />
+              <span v-else class="text-white/25 text-xs" title="系统固定">{{ a.type === 'grok' ? 10 : 1 }}</span>
+            </td>
             <!-- reset_after -->
             <td class="px-3 py-3.5 align-middle text-xs whitespace-nowrap">
               <div v-if="a.reset_after" class="leading-tight" :title="fmtIso(a.reset_after)">
@@ -494,10 +526,15 @@ onMounted(loadAccounts)
               </button>
             </td>
             <!-- actions -->
-            <td class="px-3 py-3.5 align-middle text-right whitespace-nowrap">
+            <td class="px-3 py-3.5 align-middle whitespace-nowrap">
+              <div class="flex items-center justify-end gap-2">
+                <button v-if="a.type === 'custom'" @click="editUpstream(a)" class="act" title="编辑上游">
+                  <Icon name="config" class="w-3.5 h-3.5" />
+                </button>
                 <button @click="deleteAccount(a.pool, a.id)" class="act danger" title="删除">
                   <Icon name="trash" class="w-3.5 h-3.5" />
                 </button>
+              </div>
             </td>
           </tr>
         </tbody>
@@ -520,6 +557,7 @@ onMounted(loadAccounts)
     </div>
 
     <ImportModal v-if="showImport" @close="showImport = false" @imported="loadAccounts" />
+    <UpstreamModal v-if="showUpstream" :account="editingUpstream" @close="showUpstream = false; editingUpstream = null" @imported="loadAccounts" />
   </section>
 </template>
 

frontend/src/views/ModelsView.vue

@@ -3,12 +3,14 @@ import { ref, computed, onMounted } from 'vue'
 import { api } from '../api'
 import Icon from '../components/Icon.vue'
 import ModelFormModal from '../components/ModelFormModal.vue'
+import CustomModelModal from '../components/CustomModelModal.vue'
 import TestModal from '../components/TestModal.vue'
 import { points } from '../credits'
 
 const models = ref([])
 const loading = ref(false)
 const showForm = ref(false)
+const showCustom = ref(false)
 const editing = ref(null) // null = add, object = edit
 const testing = ref(null) // model being tested, or null
 
@@ -120,6 +122,9 @@ onMounted(loadModels)
       <button @click="openAdd" class="btn-primary">
         <Icon name="plus" class="w-3.5 h-3.5" /> 新增模型
       </button>
+      <button @click="showCustom = true" class="btn-soft">
+        <Icon name="plus" class="w-3.5 h-3.5" /> 自定义模型
+      </button>
     </div>
 
     <!-- Table -->
@@ -268,6 +273,7 @@ onMounted(loadModels)
     </div>
 
     <ModelFormModal v-if="showForm" :model="editing" @close="showForm = false" @saved="onSaved" />
+    <CustomModelModal v-if="showCustom" @close="showCustom = false" @saved="() => { showCustom = false; loadModels() }" />
     <TestModal v-if="testing" :model="testing" @close="testing = null" />
   </section>
 </template>