feat(agent): support dedicated coding model id and detail pipeline hardening

2026-02-22 18:19:31 +08:00 · 2026-02-22 18:19:31 +08:00 · ac76f31878
commit ac76f31878
parent 9c4e51d3fd
17 changed files with 977 additions and 92 deletions
--- a/app/(admin)/settings/admin-settings-client.tsx
+++ b/app/(admin)/settings/admin-settings-client.tsx
@ -158,6 +158,7 @@ type AiModel = {
  isDefault: boolean;
  sortOrder: number;
  isAgentModel?: boolean;
+  agentCodingModelId?: string;
  isMultimodal?: boolean;
  imageGenConfig?: ImageGenConfig; // 🆕 图片生成特性配置
 };
@ -1072,6 +1073,7 @@ export function AdminSettingsClient() {
    isHidden: false,
    isDefault: false,
    isAgentModel: false,
+    agentCodingModelId: "",
    isMultimodal: false,
    sortOrder: 0,
    fallbackModelId1: "", // 备用模型 ID 1
@ -2007,6 +2009,7 @@ export function AdminSettingsClient() {
      isHidden: false,
      isDefault: false,
      isAgentModel: false,
+      agentCodingModelId: "",
      isMultimodal: false,
      sortOrder: 0,
      fallbackModelId1: "",
@ -6138,6 +6141,24 @@ export function AdminSettingsClient() {
                          />
                          <Label>Agent 自动模型（前端隐藏，仅后台 Agent 使用）</Label>
                        </div>
+                        {modelForm.isAgentModel && (
+                          <div className="space-y-2">
+                            <Label className="text-sm">编程模型 ID（可选）</Label>
+                            <Input
+                              value={modelForm.agentCodingModelId}
+                              onChange={(e) =>
+                                setModelForm({
+                                  ...modelForm,
+                                  agentCodingModelId: e.target.value,
+                                })
+                              }
+                              placeholder="例如: doubao-seed-2-0-code-preview-260215"
+                            />
+                            <p className="text-xs text-muted-foreground">
+                              Agent 需要结构化规划/编程输出时优先使用该模型 ID，留空则使用当前 Agent 模型。
+                            </p>
+                          </div>
+                        )}
                        <div className="flex items-center gap-2">
                          <Switch
                            checked={modelForm.isMultimodal || false}
@ -6355,6 +6376,7 @@ export function AdminSettingsClient() {
                            isHidden: model.isHidden || false,
                            isDefault: model.isDefault,
                            isAgentModel: model.isAgentModel || false,
+                            agentCodingModelId: model.agentCodingModelId || "",
                            isMultimodal: model.isMultimodal || false,
                            sortOrder: model.sortOrder || 0,
                            fallbackModelId1: (model as any).fallbackModelId1 || "",
--- a/app/api/admin/ai-models/route.ts
+++ b/app/api/admin/ai-models/route.ts
@ -10,7 +10,11 @@ import {
  getAiProviderById,
 } from "@/lib/db/queries";
 import { ChatSDKError } from "@/lib/errors";
-import { isAiModelAgentAuto, isAiModelMultimodal } from "@/lib/db/ai-model-flags";
+import {
+  getAiModelAgentCodingModelId,
+  isAiModelAgentAuto,
+  isAiModelMultimodal,
+} from "@/lib/db/ai-model-flags";

 // 🆕 图片生成模型特性配置 Schema
 const imageGenConfigSchema = z.object({
@ -41,6 +45,7 @@ const createModelSchema = z.object({
  fallbackModelId1: z.string().max(128).optional(),
  fallbackModelId2: z.string().max(128).optional(),
  isAgentModel: z.boolean().optional(),
+  agentCodingModelId: z.string().max(128).optional(),
  isMultimodal: z.boolean().optional(),
  imageGenConfig: imageGenConfigSchema, // 🆕 图片生成特性配置
 });
@ -61,6 +66,7 @@ const updateModelSchema = z.object({
  fallbackModelId1: z.string().max(128).optional(),
  fallbackModelId2: z.string().max(128).optional(),
  isAgentModel: z.boolean().optional(),
+  agentCodingModelId: z.string().max(128).optional(),
  isMultimodal: z.boolean().optional(),
  imageGenConfig: imageGenConfigSchema, // 🆕 图片生成特性配置
 });
@ -69,6 +75,7 @@ function normalizeModel(model: any) {
  return {
    ...model,
    isAgentModel: isAiModelAgentAuto(model),
+    agentCodingModelId: getAiModelAgentCodingModelId(model) || "",
    isMultimodal: isAiModelMultimodal(model),
  };
 }
--- a/app/api/canvas/detail-plan/route.ts
+++ b/app/api/canvas/detail-plan/route.ts
@ -5,8 +5,12 @@ import {
  getAgentAiModelWithProvider,
  getEnabledAiModelsWithProvider,
 } from "@/lib/db/queries";
-import { isAiModelMultimodal } from "@/lib/db/ai-model-flags";
+import {
+  getAiModelAgentCodingModelId,
+  isAiModelMultimodal,
+} from "@/lib/db/ai-model-flags";
 import { getLanguageModelById } from "@/lib/ai/dynamic-provider";
+import { inlineVisionImages } from "@/lib/ai/vision-inline";

 export const runtime = "nodejs";
 export const maxDuration = 180;
@ -1598,6 +1602,7 @@ export async function POST(request: Request) {
      id: plannerConfig.model.id,
      displayName: plannerConfig.model.displayName,
      modelId: plannerConfig.model.modelId,
+      codingModelId: getAiModelAgentCodingModelId(plannerConfig.model),
      providerType: plannerConfig.provider.type,
      supportsVision: plannerConfig.model.supportsVision,
      multimodalFlag: isAiModelMultimodal({
@ -1607,7 +1612,11 @@ export async function POST(request: Request) {
      hasReferenceImages: referenceImages.length > 0,
    });

-    const model = await getLanguageModelById(plannerConfig.model.id);
+    const codingModelId = getAiModelAgentCodingModelId(plannerConfig.model);
+    const model = await getLanguageModelById(
+      plannerConfig.model.id,
+      codingModelId
+    );
    if (!model) {
      return NextResponse.json({ success: true, data: fallback });
    }
@ -1638,7 +1647,15 @@ export async function POST(request: Request) {

    const imageInputs =
      supportsVisionPlanning(plannerConfig.model) && referenceImages.length > 0
-        ? referenceImages.map((url) => ({ type: "image" as const, image: url }))
+        ? (
+            await inlineVisionImages(referenceImages, {
+              maxImages: 6,
+              maxSide: 900,
+              timeoutMs: 12000,
+              retries: 2,
+              loggerPrefix: "[Canvas Detail Plan]",
+            })
+          ).inlineDataUrls.map((dataUrl) => ({ type: "image" as const, image: dataUrl }))
        : [];

    let text = "";
--- a/app/api/canvas/detail-review/route.ts
+++ b/app/api/canvas/detail-review/route.ts
@ -5,8 +5,12 @@ import {
  getAgentAiModelWithProvider,
  getEnabledAiModelsWithProvider,
 } from "@/lib/db/queries";
-import { isAiModelMultimodal } from "@/lib/db/ai-model-flags";
+import {
+  getAiModelAgentCodingModelId,
+  isAiModelMultimodal,
+} from "@/lib/db/ai-model-flags";
 import { getLanguageModelById } from "@/lib/ai/dynamic-provider";
+import { inlineVisionImages } from "@/lib/ai/vision-inline";

 export const runtime = "nodejs";
 export const maxDuration = 60;
@ -186,7 +190,11 @@ export async function POST(request: Request) {
      });
    }

-    const model = await getLanguageModelById(reviewerConfig.model.id);
+    const codingModelId = getAiModelAgentCodingModelId(reviewerConfig.model);
+    const model = await getLanguageModelById(
+      reviewerConfig.model.id,
+      codingModelId
+    );
    if (!model) {
      return NextResponse.json({
        success: true,
@ -238,10 +246,18 @@ export async function POST(request: Request) {
      .filter(Boolean)
      .join("\n");

+    const { inlineDataUrls } = await inlineVisionImages([imageUrl, ...referenceImages], {
+      maxImages: 6,
+      maxSide: 900,
+      timeoutMs: 12000,
+      retries: 2,
+      loggerPrefix: "[Canvas Detail Review]",
+    });
+    const effectiveReviewImages = inlineDataUrls.length > 0 ? inlineDataUrls : [imageUrl, ...referenceImages];
+
    const content: Array<{ type: "text"; text: string } | { type: "image"; image: string }> = [
      { type: "text", text: promptParts },
-      { type: "image", image: imageUrl },
-      ...referenceImages.map((url) => ({ type: "image" as const, image: url })),
+      ...effectiveReviewImages.map((image) => ({ type: "image" as const, image })),
    ];

    const { text } = await generateText({
--- a/components/canvas/CanvasChatPanel.tsx
+++ b/components/canvas/CanvasChatPanel.tsx
@ -5397,10 +5397,10 @@ export function CanvasChatPanel({ chatId, initialMessages = [], onClose, onOpenF
  }, [activeSessionTitle, projectTitle]);

  return (
-    <div className="relative flex h-full min-h-0 flex-col bg-background">
+    <div className="relative flex h-full min-h-0 flex-col bg-background/95 backdrop-blur-sm">
      {/* 全局遮罩提示 - 当处于画布选图模式时显示 */}
      {detailSlotPickerIndex !== null && (
-        <div className="absolute inset-0 z-50 flex flex-col items-center justify-center bg-background/80 backdrop-blur-sm animate-in fade-in duration-200">
+        <div className="absolute inset-0 z-50 flex flex-col items-center justify-center bg-background/90 backdrop-blur-sm animate-in fade-in duration-200">
          <div className="flex flex-col items-center gap-4 p-6 text-center">
            <div className="flex h-12 w-12 items-center justify-center rounded-full bg-primary/10 text-primary ring-4 ring-primary/5 animate-pulse">
              <ImagePlus className="h-6 w-6" />
@ -5463,7 +5463,7 @@ export function CanvasChatPanel({ chatId, initialMessages = [], onClose, onOpenF
              <History className="h-4 w-4" />
            </Button>
            {isSessionMenuOpen && (
-              <div className="absolute right-0 top-10 w-64 rounded-xl border bg-background/90 backdrop-blur-md shadow-xl p-2">
+              <div className="absolute right-0 top-10 w-64 rounded-xl border bg-background/95 backdrop-blur-sm shadow-xl p-2">
                <div className="px-2 py-1 text-xs text-muted-foreground">聊天窗口记录</div>
                <div className="max-h-56 overflow-y-auto space-y-1">
                  {chatSessions.length === 0 ? (
@ -6052,7 +6052,7 @@ export function CanvasChatPanel({ chatId, initialMessages = [], onClose, onOpenF
      )}
      
      {/* 底部输入区域 - 混合输入（图片标签+文字可自由混合） */}
-      <div className="p-4 bg-background border-t">
+      <div className="p-4 bg-background/95 backdrop-blur-sm border-t">
        <div
          className="relative flex flex-col bg-card rounded-3xl border transition-all duration-200 focus-within:border-primary/50 focus-within:bg-background"
          data-canvas-chat-input-anchor="true"
--- a/components/canvas/CanvasFileList.tsx
+++ b/components/canvas/CanvasFileList.tsx
@ -99,7 +99,7 @@ export function CanvasFileList({ open, onOpenChange, currentId }: CanvasFileList

  return (
    <Dialog open={open} onOpenChange={onOpenChange}>
-      <DialogContent className="max-w-md max-h-[80vh] flex flex-col p-0 gap-0 overflow-hidden bg-background">
+      <DialogContent className="max-w-md max-h-[80vh] flex flex-col p-0 gap-0 overflow-hidden bg-background/95 backdrop-blur-sm">
        <DialogHeader className="p-4 border-b">
          <div className="flex items-center justify-between">
            <DialogTitle>已生成文件列表</DialogTitle>
--- a/components/canvas/RealtimeCanvasDialog.tsx
+++ b/components/canvas/RealtimeCanvasDialog.tsx
@ -956,7 +956,7 @@ export function RealtimeCanvasDialog({
        </DialogHeader>
        {/* Top Header - Minimal */}
        <div className="absolute left-0 right-0 top-0 z-50 flex h-14 items-center justify-start px-6 bg-gradient-to-b from-black/20 to-transparent pointer-events-none">
-          <div className="pointer-events-auto flex items-center gap-3 rounded-full bg-background px-4 py-1.5 border border-border/50 shadow-sm">
+          <div className="pointer-events-auto flex items-center gap-3 rounded-full bg-background/95 backdrop-blur-sm px-4 py-1.5 border border-border/50 shadow-sm">
            <span className="font-semibold text-sm">Realtime Canvas</span>
            <div className="h-4 w-[1px] bg-border" />
            <span className={cn("flex items-center gap-1.5 text-xs font-medium", statusBadge.className, "bg-transparent p-0")}>
--- a/docs/AI详情生成-AGENT可编程渲染开发文档.md
+++ b/docs/AI详情生成-AGENT可编程渲染开发文档.md
@ -0,0 +1,523 @@
+# AI详情生成（AGENT + 生图模型 + 可编程渲染）开发文档
+
+## 1. 文档目的
+
+本文档用于落地一套稳定的 AI 详情生成系统：
+
+- 保留现有传统生图模式（快速产图路线）
+- 保留生图模型能力（画面创作、质感、场景）
+- 用 AGENT 负责策划、文案、结构和审稿
+- 用编程渲染保证版式统一和可编辑
+- 支持用户切换两种模式（传统生图 / AGENT排版）
+- 最终结果直接进入无限画布，支持手动改字、拖拽图标、单屏重生
+
+目标是解决当前主要问题：生图模型理解偏差、全套风格不统一、失败不可控、返工成本高。
+
+---
+
+## 2. 核心结论与原则
+
+### 2.1 核心结论
+
+不能只靠生图模型“整页一次生成”。推荐架构是：
+
+`AGENT 策划 -> 分屏生成 -> 程序化合成 -> 画布可编辑 -> 审核回路`
+
+同时保留现有传统生图链路，采用双路线并行：
+
+- 路线 A：`传统生图模式`（保持现状，快速出图）
+- 路线 B：`AGENT排版模式`（生图只负责画面，文字与排版交给程序）
+
+### 2.2 设计原则
+
+1. **模型出画面，代码定版式**  
+2. **背景锁定，前景可生**（避免模型重画结构）  
+3. **每屏可独立重试**（不整套重跑）  
+4. **所有输入输出可追溯**（项目、版本、日志、成本）  
+5. **画布内可编辑优先**（文字和图标非烤图）  
+6. **双模式可切换且互不影响**（老链路稳定，新链路增量上线）  
+
+---
+
+## 3. 业务范围
+
+## 3.1 本期范围
+
+- AI详情模式固定 14 屏流程（可配置）
+- 每屏固定尺寸 `790x1300`
+- 用户上传：
+  - 5 张产品图（必填）
+  - 1 张模特图（可用于卖家视角场景）
+  - 3 张搭配图（可选，触发关联销售屏）
+- 输入业务信息：人群、场景、核心需求、SPU 等
+- 自动生成详情文案与视觉版式
+- 结果进入画布后可修改文字、拖拽图标、替换素材
+- 自动存档、历史恢复、继续生成
+
+## 3.2 非本期范围
+
+- 移动端编辑体验优化
+- 多人协同编辑冲突合并
+- 自动电商平台一键上架
+
+## 3.3 模式策略（新增）
+
+- `传统生图模式（classic）`
+  - 使用现有生图流程
+  - 保留当前用户习惯与响应节奏
+  - 适合快速尝试和低编辑需求场景
+- `AGENT排版模式（agent_layout）`
+  - 生图模型只产出画面素材层，不产出文字排版
+  - AGENT + 程序完成文案结构、版式和 UI 统一
+  - 适合稳定量产、可编辑和可复用场景
+- 前端必须提供模式切换，并把模式写入任务上下文，后端按模式分流 pipeline
+
+---
+
+## 4. 技术架构
+
+```mermaid
+flowchart LR
+    A["用户输入(图片+业务信息)"] --> B["AGENT 规划器"]
+    B --> C["ScreenSpec 14屏结构化脚本"]
+    C --> D["生图编排器(逐屏调用模型)"]
+    D --> E["程序化合成器(背景/文案/图标)"]
+    E --> F["双总监审稿(设计+营销)"]
+    F -->|通过| G["写入S3/MinIO + DB版本"]
+    F -->|不通过| D
+    G --> H["画布落地(垂直排版)"]
+    H --> I["手动编辑(改字/拖拽/重生单屏)"]
+```
+
+双路线分流（逻辑）：
+
+```mermaid
+flowchart LR
+    U["用户选择模式"] --> M{"generation_mode"}
+    M -->|classic| C1["传统生图Pipeline"]
+    M -->|agent_layout| C2["AGENT排版Pipeline"]
+    C1 --> R["统一写入画布与历史记录"]
+    C2 --> R
+```
+
+---
+
+## 5. 推荐开源技术选型
+
+## 5.1 画布编辑层
+
+- `react-konva + konva`：用于画布图层编辑、拖拽、缩放、导出
+- `Moveable` 或 `interact.js`：吸附对齐、参考线、间距控制
+
+## 5.2 渲染层
+
+- 主方案：HTML/CSS 模板 + Playwright 截图（最接近设计稿控制）
+- 备选：`Satori + resvg-js`（高一致性服务端矢量渲染）
+
+## 5.3 模型编排层
+
+- 继续使用当前生图供应商体系（支持路由和故障迁移）
+- AGENT 负责：
+  - 详情策略规划
+  - Prompt 生成与优化
+  - 结果审稿和重试建议
+
+## 5.4 可选参考（不直接引入）
+
+- Excalidraw：交互/白板体验参考
+- Penpot：组件化设计体系参考
+- tldraw：能力强，但许可证需谨慎评估
+
+---
+
+## 6. 数据模型设计（数据库）
+
+建议新增或扩展以下表（命名可按现有规范调整）：
+
+## 6.1 `detail_projects`
+
+- `id`
+- `user_id`
+- `canvas_id`
+- `spu`
+- `demand_type`
+- `project_name`（默认 `SPU+需求类型`）
+- `status`（draft/planning/generating/reviewing/completed/failed）
+- `generation_mode`（classic/agent_layout，默认 classic）
+- `created_at` / `updated_at`
+
+## 6.2 `detail_project_inputs`
+
+- `id`
+- `project_id`
+- `core_audience`
+- `usage_scenarios`
+- `core_needs`
+- `selling_points`
+- `tone_style`
+- `extra_requirements`
+- `raw_form_json`
+
+## 6.3 `detail_assets`
+
+- `id`
+- `project_id`
+- `asset_type`（product/model/combo/reference）
+- `source_url`
+- `storage_url`
+- `width` / `height`
+- `meta_json`
+
+## 6.4 `detail_screen_specs`
+
+- `id`
+- `project_id`
+- `screen_index`（1..14）
+- `screen_type`
+- `spec_json`（可编辑图层结构）
+- `prompt_json`
+- `status`
+
+## 6.5 `detail_screen_renders`
+
+- `id`
+- `project_id`
+- `screen_index`
+- `version_no`
+- `image_url`
+- `thumb_url`
+- `model_provider`
+- `model_name`
+- `cost_credits`
+- `latency_ms`
+- `review_score_design`
+- `review_score_marketing`
+- `review_notes`
+
+## 6.6 `detail_generation_jobs`
+
+- `id`
+- `project_id`
+- `job_type`（plan/generate/review/retry）
+- `generation_mode`（classic/agent_layout）
+- `progress`（0~100）
+- `stage`（规划中/生成中/审稿中/完成）
+- `error_code`
+- `error_message`
+- `started_at` / `finished_at`
+
+## 6.7 现有聊天/消息表扩展（建议）
+
+- 在会话或任务表增加 `generation_mode`
+- 在消息元数据增加 `mode_snapshot`（防止切换模式导致历史重放错乱）
+- 在生成记录增加 `resolved_pipeline`（最终实际执行链路）
+
+---
+
+## 7. ScreenSpec（核心协议）
+
+AGENT 和渲染器之间只通过结构化协议通信，不直接拼字符串 UI。
+
+```json
+{
+  "projectId": "dp_xxx",
+  "canvasId": "cv_xxx",
+  "size": { "width": 790, "height": 1300 },
+  "globalStyle": {
+    "fontFamily": "PingFang SC",
+    "titleSize": 68,
+    "subtitleSize": 36,
+    "bodySize": 28,
+    "lineHeight": 1.35,
+    "palette": { "primary": "#5B4DFF", "text": "#111111" }
+  },
+  "screen": {
+    "index": 1,
+    "type": "hero_poster",
+    "copy": {
+      "name": "安敏裤",
+      "title": "安敏裤 | 敏宝整夜安睡不红屁屁",
+      "subtitle": "过敏体质宝宝专研，夜用安心不反渗"
+    },
+    "layers": [
+      { "id": "bg", "kind": "backgroundImage", "editable": false, "src": "https://..." },
+      { "id": "title", "kind": "text", "editable": true, "x": 56, "y": 120, "w": 678, "h": 180 },
+      { "id": "badge", "kind": "icon", "editable": true, "x": 56, "y": 340, "w": 160, "h": 52 }
+    ],
+    "gen": {
+      "needModelImage": true,
+      "prompt": "只生成产品前景和光效，不生成文字，不改背景结构",
+      "negativePrompt": "杂乱文字, 水印, logo"
+    }
+  }
+}
+```
+
+---
+
+## 8. 14屏生产流程（标准）
+
+1. 首屏海报（命名+核心需求一句话）
+2. 用户痛点
+3. 核心卖点机制
+4. 卖点证据/参数
+5. 场景 1（日常）
+6. 场景 2（高频痛点）
+7. 细节特写
+8. 对比页
+9. 使用方法
+10. 搭配销售（用户上传搭配图时启用）
+11. 买家评价 1（自动生成评价+3图）
+12. 买家评价 2（可再次生成新2条）
+13. 品牌背书/服务承诺
+14. 收口 CTA
+
+约束：必须完整到第 14 屏才能标记 `completed`。
+
+---
+
+## 9. AGENT 职责拆分
+
+## 9.1 Planner Agent（规划师）
+
+- 解析输入信息和图片素材
+- 生成 14 屏文案骨架与屏幕目标
+- 生成命名（大白话、可记忆、带卖点）
+
+## 9.2 Prompt Agent（生图导演）
+
+- 将每屏目标转换为生图 prompt
+- 强约束“仅出前景元素，不写文案”
+- 设定风格一致性关键词
+
+## 9.3 Review Agent（双总监审稿）
+
+- 设计总监评分：版式层级、视觉统一、可读性
+- 营销总监评分：卖点清晰、转化逻辑、购买驱动
+- 低于阈值自动重生单屏
+
+---
+
+## 10. 画布集成要求
+
+## 10.1 渲染落位
+
+- 14 屏按顺序垂直插入画布
+- 固定屏间距（建议 80px）
+- 自动居中对齐
+- 生成时先放占位卡（含进度）
+
+## 10.2 编辑能力
+
+- 文本层可直接双击编辑
+- 图标层可拖拽/缩放
+- 支持吸附对齐线（含最小间距，避免粘连）
+- 单屏“重生”按钮（保留版式，仅替换画面）
+
+## 10.3 存档与恢复
+
+- 自动保存当前编辑状态（去抖）
+- 恢复项目时还原全部输入、素材、屏幕版本和图层状态
+- 支持“另存为新版本”
+
+---
+
+## 11. API 设计（建议）
+
+## 11.1 项目接口
+
+- `POST /api/canvas/[id]/detail-projects`：创建项目
+- `GET /api/canvas/[id]/detail-projects`：项目列表
+- `GET /api/canvas/[id]/detail-projects/:projectId`：项目详情
+- `DELETE /api/canvas/[id]/detail-projects/:projectId`：删除存档
+
+## 11.2 生成接口
+
+- `POST /api/canvas/[id]/detail-projects/:projectId/plan`
+- `POST /api/canvas/[id]/detail-projects/:projectId/generate`
+- `POST /api/canvas/[id]/detail-projects/:projectId/retry-screen`
+- `POST /api/canvas/[id]/detail-projects/:projectId/review`
+- 请求体需带 `generation_mode`，或由项目默认模式继承
+
+## 11.3 进度接口
+
+- `GET /api/canvas/[id]/detail-projects/:projectId/progress`（SSE 推荐）
+- 前端展示阶段：规划中 -> 生成中 -> 审稿中 -> 完成/失败
+
+## 11.4 模式切换接口（新增）
+
+- `POST /api/canvas/[id]/generation-mode`
+  - 入参：`conversationId`、`generation_mode`
+  - 作用：更新当前会话默认模式
+- 模式切换只影响新任务，不回写旧任务执行链路
+
+---
+
+## 12. 失败与兜底策略
+
+1. 单屏失败优先重试，不整套失败  
+2. 主模型失败自动故障迁移  
+3. 审稿失败回到该屏 prompt 优化重生  
+4. 图片 URL 拉取失败时切换缩略图/中转 URL  
+5. 超时任务可恢复继续跑，不丢进度  
+
+---
+
+## 13. 性能与成本控制
+
+- 生图并发建议：2~3（避免过载和限流）
+- 文字层和图标层走本地渲染，减少模型调用
+- 单屏缓存：输入未变则复用结果
+- 记录每屏耗时与积分，输出成本报表
+
+---
+
+## 13.1 排版遮挡与可读性自动质检（新增）
+
+为避免“文字压住产品主体、文案不可读、版式失衡”，在每屏渲染后新增 `Layout QA` 阶段。
+
+### A. 质检输入
+
+- 渲染结果图（790x1300）
+- `ScreenSpec` 图层信息（文本框、图标框、背景层）
+- 产品主体掩码 `product_mask`（由分割模型产出，建议 SAM2 或 YOLO Seg）
+- 图像显著区域 `saliency_map`（可选）
+
+### B. 核心判定指标
+
+- `occlusion_ratio = area(text_box ∩ product_mask) / area(product_mask)`
+- `text_intrusion = area(text_box ∩ product_mask) / area(text_box)`
+- 文本区域对比度（背景亮度与文字颜色对比）
+- 安全区越界（顶部标题区、边距、栅格）
+
+### C. 建议阈值（首版）
+
+- `occlusion_ratio <= 0.03`
+- `text_intrusion <= 0.10`
+- 主标题与产品主体最小距离 >= 24px
+- 正文对比度 >= 4.5:1；大标题对比度 >= 3.0:1
+- 文本不得覆盖显著区域 Top 20%
+
+### D. 自动修复策略
+
+1. 先移动文本框（上 -> 下 -> 侧）  
+2. 再缩短文案（调用 AGENT copy shortener）  
+3. 再调整字号和行高（保持全局层级）  
+4. 最后加底板/描边增强可读性  
+5. 超过重排次数仍失败 -> 标记人工介入
+
+### E. 结果落库
+
+- 每屏记录 `qa_passed`、`qa_score`、`qa_issues`、`qa_fix_actions`
+- 支持在画布中打开“质检报告”，可一键触发该屏重排/重生
+
+---
+
+## 14. 安全与合规
+
+- 所有 API Key 仅服务端读取（数据库加密存储）
+- 前端不暴露供应商密钥
+- 生成内容日志脱敏（URL、用户隐私）
+- 存储地址统一走可控域名，避免跨域拉图波动
+
+---
+
+## 15. 里程碑计划
+
+## 阶段 A（1 周）：可跑通 MVP
+
+- 前端模式切换（传统生图 / AGENT排版）
+- 后端按 `generation_mode` 分流
+- ScreenSpec 协议
+- 14 屏生成状态机
+- 画布垂直落位 + 占位卡
+- 单屏重试
+
+## 阶段 B（1~2 周）：可编辑与存档
+
+- 文本层编辑、图标拖拽、吸附线
+- 项目存档、恢复、版本化
+- 失败恢复和续跑
+
+## 阶段 C（1 周）：质量增强
+
+- 双总监审稿自动回路
+- 屏幕质量评分
+- 成本和速度优化
+
+---
+
+## 16. 验收标准（必须满足）
+
+1. 14 屏完整生成成功率 >= 95%（可重试后）
+2. 每屏版式和字体风格一致，无明显错位
+3. 画布中可改字、可拖拽图标、可单屏重生
+4. 刷新页面后项目状态和结果不丢失
+5. 失败有明确原因并可一键重试
+
+---
+
+## 17. 风险与应对
+
+- **风险：模型仍偏题** -> 强化 prompt 约束 + 审稿回路  
+- **风险：图片链路不稳定** -> 同域代理/中转 + 缩略图兜底  
+- **风险：编辑体验复杂** -> 分层上线，先文本编辑再高级控件  
+- **风险：成本高** -> 单屏缓存、只重跑失败屏、限制并发  
+
+---
+
+## 18. 开发清单（执行版）
+
+- [ ] 新增详情项目数据表与迁移脚本
+- [ ] 实现 ScreenSpec 生成器（AGENT Planner）
+- [ ] 实现每屏生图编排器
+- [ ] 实现程序化合成器（固定 790x1300）
+- [ ] 实现 Layout QA（遮挡检测、可读性检测、自动重排）
+- [ ] 实现双总监审稿器和重试策略
+- [ ] 接入画布垂直落位与占位进度
+- [ ] 接入文本编辑/图标拖拽/吸附线
+- [ ] 接入存档、恢复、删除、另存版本
+- [ ] 接入监控日志与成本统计
+
+---
+
+## 19. 编程生成PPT技术路线参考（新增）
+
+本项目主渲染仍以“画布图层渲染”为核心，PPT 相关能力建议定位为“导出能力”。
+
+### 20.1 推荐路线（与当前栈匹配）
+
+- `PptxGenJS`（Node/TS 生态）
+  - 优点：与现有前后端栈一致、可编程布局能力强、便于把 `ScreenSpec` 映射为幻灯片对象
+  - 适用：把 14 屏详情导出为可编辑 PPTX 交付给运营团队
+
+### 20.2 可选路线
+
+- `python-pptx`
+  - 优点：批处理成熟，适合离线报告任务
+  - 缺点：需要额外 Python 服务，增加运维与调用链复杂度
+- `Open XML SDK`（.NET）/ `Apache POI`（Java）
+  - 适合企业系统深度集成，不作为当前首选
+
+### 20.3 参考价值判断
+
+- 有价值点：
+  - `ScreenSpec -> PPTX` 的结构化映射
+  - 模板化母版（Master）和统一字号体系
+  - 批量导出与版本归档
+- 无法替代点：
+  - PPT 方案本身不解决“文字遮挡主体”问题，仍需 Layout QA
+  - 不适合替代当前画布内精细编辑与逐屏重生流程
+
+### 20.4 落地建议
+
+1. 先完成画布内稳定生产（ScreenSpec + Layout QA + 自动重排）  
+2. 再增加 `Export to PPTX`，作为交付层功能  
+3. 导出时保留标题/副标题/正文可编辑，背景图锁定，避免误改布局  
+
+---
+
+## 20. 结语
+
+这套方案不是“取消生图模型”，而是把生图模型放到最擅长的位置（视觉生成），把结构一致性、文案逻辑和生产稳定性交给 AGENT + 编程系统。  
+对你当前业务目标（稳定出高转化详情、可持续迭代）是更可控、可扩展、可运营的路径。
--- a/lib/ai/auto-agent-graph.ts
+++ b/lib/ai/auto-agent-graph.ts
@ -3,7 +3,10 @@ import { generateText } from "ai";
 import sharp from "sharp";
 import type { AnalysisResult } from "@/lib/ai/smart-analyzer";
 import type { RouteResult } from "@/lib/ai/model-router";
-import { isAiModelMultimodal } from "@/lib/db/ai-model-flags";
+import {
+  getAiModelAgentCodingModelId,
+  isAiModelMultimodal,
+} from "@/lib/db/ai-model-flags";

 interface HistoryMessage {
  role: "user" | "assistant";
@ -688,6 +691,26 @@ function parseJsonBlock<T = Record<string, unknown>>(text: string): T | null {
  }
 }

+function resolveAgentCodingModelOverride(
+  selectedModelId: string,
+  configuredAgentModel:
+    | {
+        model?: {
+          id?: string | null;
+          defaultParams?: unknown;
+          capabilities?: unknown;
+        } | null;
+      }
+    | null
+    | undefined,
+): string | null {
+  const configuredId = configuredAgentModel?.model?.id || null;
+  if (!configuredId || configuredId !== selectedModelId) {
+    return null;
+  }
+  return getAiModelAgentCodingModelId(configuredAgentModel?.model || {}) || null;
+}
+
 async function generateAgentConversation(
  analysis: AnalysisResult,
  state: AutoAgentStateType
@ -714,7 +737,10 @@ async function generateAgentConversation(
    if (!reasoningModelConfig) {
      return analysis.conversationalReply || null;
    }
-    const reasoningModel = await getLanguageModelById(reasoningModelConfig.model.id);
+    const reasoningModel = await getLanguageModelById(
+      reasoningModelConfig.model.id,
+      resolveAgentCodingModelOverride(reasoningModelConfig.model.id, agentModel),
+    );
    if (!reasoningModel) {
      return analysis.conversationalReply || null;
    }
@ -858,7 +884,10 @@ async function recommendImageParamsWithAgent(
    if (!plannerModelConfig) {
      return heuristicImageParams(state.text, analysis);
    }
-    const plannerModel = await getLanguageModelById(plannerModelConfig.model.id);
+    const plannerModel = await getLanguageModelById(
+      plannerModelConfig.model.id,
+      resolveAgentCodingModelOverride(plannerModelConfig.model.id, agentModel),
+    );
    if (!plannerModel) {
      return heuristicImageParams(state.text, analysis);
    }
@ -933,7 +962,10 @@ async function recommendVideoPlanWithAgent(
    if (!plannerModelConfig) {
      return heuristicVideoPlan(state);
    }
-    const plannerModel = await getLanguageModelById(plannerModelConfig.model.id);
+    const plannerModel = await getLanguageModelById(
+      plannerModelConfig.model.id,
+      resolveAgentCodingModelOverride(plannerModelConfig.model.id, agentModel),
+    );
    if (!plannerModel) {
      return heuristicVideoPlan(state);
    }
@ -1041,7 +1073,10 @@ async function refineRouteWithAgent(
    });
    const plannerModelConfig = pickAgentReasoningModel(primaryAgentModel, allModelsRaw);
    if (!plannerModelConfig) return baseRoute;
-    const plannerModel = await getLanguageModelById(plannerModelConfig.model.id);
+    const plannerModel = await getLanguageModelById(
+      plannerModelConfig.model.id,
+      resolveAgentCodingModelOverride(plannerModelConfig.model.id, agentModel),
+    );
    if (!plannerModel) return baseRoute;

    const candidatesText = imageCandidates
--- a/lib/ai/dynamic-provider.ts
+++ b/lib/ai/dynamic-provider.ts
@ -228,7 +228,8 @@ function getOrCreateProvider(provider: Pick<AiProvider, "id" | "type" | "baseUrl

 // 获取动态配置的语言模型
 export async function getDynamicModel(
-  modelId: string
+  modelId: string,
+  overrideModelId?: string | null
 ): Promise<LanguageModel | null> {
  try {
    const model = await getAiModelById({ id: modelId });
@ -256,8 +257,10 @@ export async function getDynamicModel(
      return null;
    }

+    const resolvedModelId = String(overrideModelId || "").trim() || model.modelId;
+
    // 创建基础模型
-    console.log(`[getDynamicModel] Creating model: ${model.modelId} (${model.displayName}) using provider: ${provider.name} (${provider.type})`);
+    console.log(`[getDynamicModel] Creating model: ${resolvedModelId} (${model.displayName}) using provider: ${provider.name} (${provider.type})`);
    
    // 根据 provider 类型选择正确的方法和模型ID
    // customProvider (gemini, local) 使用 .languageModel()
@ -266,56 +269,56 @@ export async function getDynamicModel(
    if (provider.type === "gemini") {
      // Gemini 图片生成使用 chat() 方法（与 local-qwen 相同的调用方式）
      console.log("[getDynamicModel] Getting Gemini chat model..."); // 调试日志
-      baseModel = providerInstance.chat(model.modelId);
+      baseModel = providerInstance.chat(resolvedModelId);
      console.log("[getDynamicModel] Gemini model retrieved:", baseModel ? "success" : "failed"); // 调试日志
      console.log("[getDynamicModel] Gemini model type:", typeof baseModel); // 调试
      console.log("[getDynamicModel] Gemini model keys:", baseModel ? Object.keys(baseModel).join(', ') : 'N/A'); // 调试
    } else if (provider.type === "local") {
      // 本地模型使用 chat() 方法
      console.log("[getDynamicModel] Getting local chat model..."); // 调试日志
-      baseModel = providerInstance.chat(model.modelId);
+      baseModel = providerInstance.chat(resolvedModelId);
      console.log("[getDynamicModel] Local model retrieved:", baseModel ? "success" : "failed"); // 调试日志
    } else if (provider.type === "qianwen-video") {
      // 千问视频生成使用 chat() 方法
      console.log("[getDynamicModel] Getting Qianwen Video chat model..."); // 调试日志
-      baseModel = providerInstance.chat(model.modelId);
+      baseModel = providerInstance.chat(resolvedModelId);
      console.log("[getDynamicModel] Qianwen Video model retrieved:", baseModel ? "success" : "failed"); // 调试日志
    } else if (provider.type === "fal-video") {
      // FAL 视频生成使用 chat() 方法
      console.log("[getDynamicModel] Getting Fal Video chat model..."); // 调试日志
-      baseModel = providerInstance.chat(model.modelId);
+      baseModel = providerInstance.chat(resolvedModelId);
      console.log("[getDynamicModel] Fal Video model retrieved:", baseModel ? "success" : "failed"); // 调试日志
    } else if (provider.type === "volcano-vision") {
      // 火山引擎视觉生成使用 chat() 方法
      console.log("[getDynamicModel] Getting Volcano Vision chat model..."); // 调试日志
-      baseModel = providerInstance.chat(model.modelId);
+      baseModel = providerInstance.chat(resolvedModelId);
      console.log("[getDynamicModel] Volcano Vision model retrieved:", baseModel ? "success" : "failed"); // 调试日志
    } else if (provider.type === "alibaba-image") {
      // 阿里云千问图片生成使用 chat() 方法
      console.log("[getDynamicModel] Getting Alibaba Image chat model..."); // 调试日志
-      baseModel = providerInstance.chat(model.modelId);
+      baseModel = providerInstance.chat(resolvedModelId);
      console.log("[getDynamicModel] Alibaba Image model retrieved:", baseModel ? "success" : "failed"); // 调试日志
    } else if (provider.type === "ruxa-image") {
      // RUXA 图片编辑生成使用 chat() 方法
      console.log("[getDynamicModel] Getting Ruxa Image chat model..."); // 调试日志
-      baseModel = providerInstance.chat(model.modelId);
+      baseModel = providerInstance.chat(resolvedModelId);
      console.log("[getDynamicModel] Ruxa Image model retrieved:", baseModel ? "success" : "failed"); // 调试日志
    } else if (provider.type === "fal-image") {
      // FAL 图片编辑生成使用 chat() 方法
      console.log("[getDynamicModel] Getting Fal Image chat model..."); // 调试日志
-      baseModel = providerInstance.chat(model.modelId);
+      baseModel = providerInstance.chat(resolvedModelId);
      console.log("[getDynamicModel] Fal Image model retrieved:", baseModel ? "success" : "failed"); // 调试日志
    } else if (provider.type === "huanke-gemini") {
      // 环科 Gemini 原生格式使用 chat() 方法
      console.log("[getDynamicModel] Getting Huanke Gemini chat model..."); // 调试日志
-      baseModel = providerInstance.chat(model.modelId);
+      baseModel = providerInstance.chat(resolvedModelId);
      console.log("[getDynamicModel] Huanke Gemini model retrieved:", baseModel ? "success" : "failed"); // 调试日志
      console.log("[getDynamicModel] Huanke Gemini model type:", typeof baseModel); // 调试
      console.log("[getDynamicModel] Huanke Gemini model keys:", baseModel ? Object.keys(baseModel).join(', ') : 'N/A'); // 调试
    } else {
      // OpenAI-compatible providers
      console.log("[getDynamicModel] Getting OpenAI-compatible model..."); // 调试日志
-      baseModel = (providerInstance as any).chat(model.modelId);
+      baseModel = (providerInstance as any).chat(resolvedModelId);
      console.log("[getDynamicModel] OpenAI model retrieved:", baseModel ? "success" : "failed"); // 调试日志
    }

@ -340,9 +343,10 @@ export async function getDynamicModel(

 // 通过模型的数据库 ID 获取语言模型实例
 export async function getLanguageModelById(
-  modelId: string
+  modelId: string,
+  overrideModelId?: string | null
 ): Promise<LanguageModel | null> {
-  return getDynamicModel(modelId);
+  return getDynamicModel(modelId, overrideModelId);
 }

 // 获取所有可用的模型列表（用于前端展示）
--- a/lib/ai/prompt-optimizer.ts
+++ b/lib/ai/prompt-optimizer.ts
@ -544,6 +544,7 @@ export async function optimizePromptWithLLM(params: {
      getEnabledAiModelsWithProvider,
      getAgentAiModelWithProvider,
    } = await import("@/lib/db/queries");
+    const { getAiModelAgentCodingModelId } = await import("@/lib/db/ai-model-flags");
    const { getLanguageModelById } = await import("./dynamic-provider");
    
    // 提示词优化优先使用 Agent 模型（建议智普），失败时回退
@ -562,13 +563,22 @@ export async function optimizePromptWithLLM(params: {
      (m) => m.provider.type === "alibaba" && m.model.supportsVision
    );
    const selectedOptimizerModel = usableAgentModel || zhipuModel || qwenModel;
+    const agentCodingModelId =
+      selectedOptimizerModel &&
+      usableAgentModel &&
+      selectedOptimizerModel.model.id === usableAgentModel.model.id
+        ? getAiModelAgentCodingModelId(usableAgentModel.model)
+        : null;

    if (!selectedOptimizerModel) {
      console.log("[Prompt Optimizer LLM] No optimizer model available, falling back to template");
      return optimizePrompt({ originalPrompt, intent, subIntent });
    }

-    const model = await getLanguageModelById(selectedOptimizerModel.model.id);
+    const model = await getLanguageModelById(
+      selectedOptimizerModel.model.id,
+      agentCodingModelId
+    );
    if (!model) {
      console.log("[Prompt Optimizer LLM] Failed to get model, falling back to template");
      return optimizePrompt({ originalPrompt, intent, subIntent });
@ -598,7 +608,8 @@ export async function optimizePromptWithLLM(params: {
    console.log(
      "[Prompt Optimizer LLM] Calling optimizer model:",
      selectedOptimizerModel.model.displayName,
-      `(${selectedOptimizerModel.provider.type})`
+      `(${selectedOptimizerModel.provider.type})`,
+      agentCodingModelId ? `override=${agentCodingModelId}` : ""
    );

    const { text: responseText } = await generateText({
--- a/lib/ai/vision-inline.ts
+++ b/lib/ai/vision-inline.ts
@ -0,0 +1,117 @@
+import sharp from "sharp";
+
+type InlineVisionOptions = {
+  maxImages?: number;
+  maxSide?: number;
+  timeoutMs?: number;
+  retries?: number;
+  loggerPrefix?: string;
+  jpegQuality?: number;
+};
+
+type InlineVisionResult = {
+  inlineDataUrls: string[];
+  skipped: Array<{ url: string; reason: string }>;
+};
+
+const DEFAULT_MAX_IMAGES = 6;
+const DEFAULT_MAX_SIDE = 900;
+const DEFAULT_TIMEOUT_MS = 12000;
+const DEFAULT_RETRIES = 2;
+const DEFAULT_JPEG_QUALITY = 82;
+
+async function fetchImageWithRetry(url: string, timeoutMs: number, retries: number) {
+  let lastError: unknown;
+
+  for (let attempt = 1; attempt <= retries; attempt++) {
+    const controller = new AbortController();
+    const timeout = setTimeout(() => controller.abort(), timeoutMs);
+    try {
+      const response = await fetch(url, {
+        method: "GET",
+        signal: controller.signal,
+      });
+
+      if (!response.ok) {
+        throw new Error(`HTTP ${response.status}`);
+      }
+
+      const arrayBuffer = await response.arrayBuffer();
+      return Buffer.from(arrayBuffer);
+    } catch (error) {
+      lastError = error;
+      if (attempt < retries) {
+        await new Promise((resolve) => setTimeout(resolve, 250 * attempt));
+      }
+    } finally {
+      clearTimeout(timeout);
+    }
+  }
+
+  throw lastError instanceof Error ? lastError : new Error("图片拉取失败");
+}
+
+async function downscaleForVision(
+  inputBuffer: Buffer,
+  maxSide: number,
+  jpegQuality: number,
+) {
+  const image = sharp(inputBuffer, { failOn: "none" });
+  const metadata = await image.metadata();
+  const hasAlpha = Boolean(metadata.hasAlpha);
+
+  const resized = image
+    .rotate()
+    .resize({
+      width: maxSide,
+      height: maxSide,
+      fit: "inside",
+      withoutEnlargement: true,
+    });
+
+  if (hasAlpha) {
+    const output = await resized.png({ compressionLevel: 9 }).toBuffer();
+    return { mime: "image/png", buffer: output };
+  }
+
+  const output = await resized.jpeg({ quality: jpegQuality, mozjpeg: true }).toBuffer();
+  return { mime: "image/jpeg", buffer: output };
+}
+
+export async function inlineVisionImages(
+  urls: string[],
+  options: InlineVisionOptions = {},
+): Promise<InlineVisionResult> {
+  const maxImages = Math.max(1, options.maxImages ?? DEFAULT_MAX_IMAGES);
+  const maxSide = Math.max(320, options.maxSide ?? DEFAULT_MAX_SIDE);
+  const timeoutMs = Math.max(3000, options.timeoutMs ?? DEFAULT_TIMEOUT_MS);
+  const retries = Math.max(1, options.retries ?? DEFAULT_RETRIES);
+  const jpegQuality = Math.max(55, Math.min(95, options.jpegQuality ?? DEFAULT_JPEG_QUALITY));
+  const loggerPrefix = options.loggerPrefix ?? "[Vision Inline]";
+
+  const inlineDataUrls: string[] = [];
+  const skipped: Array<{ url: string; reason: string }> = [];
+
+  const candidates = urls
+    .map((item) => String(item || "").trim())
+    .filter((url) => /^https?:\/\//i.test(url))
+    .slice(0, maxImages);
+
+  for (const url of candidates) {
+    try {
+      const buffer = await fetchImageWithRetry(url, timeoutMs, retries);
+      const processed = await downscaleForVision(buffer, maxSide, jpegQuality);
+      const dataUrl = `data:${processed.mime};base64,${processed.buffer.toString("base64")}`;
+      inlineDataUrls.push(dataUrl);
+    } catch (error) {
+      const reason = error instanceof Error ? error.message : "unknown_error";
+      skipped.push({ url, reason });
+      console.warn(`${loggerPrefix} skip reference image:`, {
+        url,
+        reason,
+      });
+    }
+  }
+
+  return { inlineDataUrls, skipped };
+}
--- a/lib/db/ai-model-flags.ts
+++ b/lib/db/ai-model-flags.ts
@ -1,5 +1,6 @@
 const AGENT_AUTO_MODEL_KEY = "agentAutoModel";
 const MULTIMODAL_MODEL_KEY = "multimodalModel";
+const AGENT_CODING_MODEL_ID_KEY = "agentCodingModelId";

 function toRecord(value: unknown): Record<string, unknown> | null {
  if (!value || typeof value !== "object" || Array.isArray(value)) {
@ -62,6 +63,43 @@ export function withAgentAutoModelFlag(
  return Object.keys(next).length > 0 ? next : null;
 }

+export function getAiModelAgentCodingModelId(input: {
+  defaultParams?: unknown;
+  capabilities?: unknown;
+}): string | null {
+  const defaultParams = toRecord(input.defaultParams);
+  const defaultValue = defaultParams?.[AGENT_CODING_MODEL_ID_KEY];
+  if (typeof defaultValue === "string") {
+    const normalized = defaultValue.trim();
+    if (normalized) return normalized;
+  }
+
+  const capabilities = toRecord(input.capabilities);
+  const capabilityValue = capabilities?.[AGENT_CODING_MODEL_ID_KEY];
+  if (typeof capabilityValue === "string") {
+    const normalized = capabilityValue.trim();
+    if (normalized) return normalized;
+  }
+
+  return null;
+}
+
+export function withAgentCodingModelId(
+  defaultParams: unknown,
+  codingModelId?: string | null
+): Record<string, unknown> | null {
+  const next = { ...(toRecord(defaultParams) || {}) };
+  const normalized = String(codingModelId || "").trim();
+
+  if (normalized) {
+    next[AGENT_CODING_MODEL_ID_KEY] = normalized;
+  } else {
+    delete next[AGENT_CODING_MODEL_ID_KEY];
+  }
+
+  return Object.keys(next).length > 0 ? next : null;
+}
+
 export function isAiModelMultimodal(input: {
  capabilities?: unknown;
  defaultParams?: unknown;
--- a/lib/db/queries.ts
+++ b/lib/db/queries.ts
@ -57,6 +57,7 @@ import { generateHashedPassword } from "./utils";
 import {
  isAiModelAgentAuto,
  withAgentAutoModelFlag,
+  withAgentCodingModelId,
  withMultimodalModelFlag,
 } from "./ai-model-flags";
 import {
@ -1440,6 +1441,7 @@ export async function createAiModel({
  imageGenConfig,
  isAgentModel,
  isMultimodal,
+  agentCodingModelId,
 }: {
  providerId: string;
  modelId: string;
@ -1457,6 +1459,7 @@ export async function createAiModel({
  imageGenConfig?: ImageGenConfig;
  isAgentModel?: boolean;
  isMultimodal?: boolean;
+  agentCodingModelId?: string;
 }): Promise<AiModel[]> {
  try {
    // 如果设置为默认，先取消其他默认
@ -1467,6 +1470,16 @@ export async function createAiModel({
        .where(eq(aiModel.isDefault, true));
    }

+    const defaultParams =
+      agentCodingModelId !== undefined
+        ? withAgentCodingModelId(
+            isAgentModel === undefined ? null : withAgentAutoModelFlag(null, isAgentModel),
+            agentCodingModelId
+          )
+        : isAgentModel === undefined
+          ? null
+          : withAgentAutoModelFlag(null, isAgentModel);
+
    return await db
      .insert(aiModel)
      .values({
@ -1484,10 +1497,7 @@ export async function createAiModel({
        fallbackModelId1: fallbackModelId1 || null,
        fallbackModelId2: fallbackModelId2 || null,
        imageGenConfig: imageGenConfig || null,
-        defaultParams:
-          isAgentModel === undefined
-            ? null
-            : withAgentAutoModelFlag(null, isAgentModel),
+        defaultParams,
        capabilities:
          isMultimodal === undefined
            ? null
@ -1520,6 +1530,7 @@ export async function updateAiModel({
  imageGenConfig,
  isAgentModel,
  isMultimodal,
+  agentCodingModelId,
 }: {
  id: string;
  providerId?: string;
@ -1538,6 +1549,7 @@ export async function updateAiModel({
  imageGenConfig?: ImageGenConfig;
  isAgentModel?: boolean;
  isMultimodal?: boolean;
+  agentCodingModelId?: string;
 }): Promise<AiModel[]> {
  try {
    // 如果设置为默认，先取消其他默认
@ -1566,7 +1578,11 @@ export async function updateAiModel({
    if (fallbackModelId2 !== undefined) updateData.fallbackModelId2 = fallbackModelId2 || null;
    // 🆕 图片生成配置
    if (imageGenConfig !== undefined) updateData.imageGenConfig = imageGenConfig || null;
-    if (isAgentModel !== undefined || isMultimodal !== undefined) {
+    if (
+      isAgentModel !== undefined ||
+      isMultimodal !== undefined ||
+      agentCodingModelId !== undefined
+    ) {
      const [existingModel] = await db
        .select({
          defaultParams: aiModel.defaultParams,
@ -1580,12 +1596,22 @@ export async function updateAiModel({
        return [];
      }

+      let nextDefaultParams = existingModel.defaultParams;
      if (isAgentModel !== undefined) {
-        updateData.defaultParams = withAgentAutoModelFlag(
-          existingModel.defaultParams,
+        nextDefaultParams = withAgentAutoModelFlag(
+          nextDefaultParams,
          isAgentModel
        );
      }
+      if (agentCodingModelId !== undefined) {
+        nextDefaultParams = withAgentCodingModelId(
+          nextDefaultParams,
+          agentCodingModelId
+        );
+      }
+      if (isAgentModel !== undefined || agentCodingModelId !== undefined) {
+        updateData.defaultParams = nextDefaultParams;
+      }

      if (isMultimodal !== undefined) {
        updateData.capabilities = withMultimodalModelFlag(
--- a/lib/security/rate-limiter.ts
+++ b/lib/security/rate-limiter.ts
@ -77,6 +77,9 @@ export const rateLimiters = {
  // 通用API限制：每分钟100次请求
  api: new RateLimiter(60000, 100),

+  // 高频只读接口（首页/画布轮询）：每分钟1200次请求
+  uiPolling: new RateLimiter(60000, 1200),
+
  // 画布保存接口：每分钟300次请求（含自动保存）
  canvas: new RateLimiter(60000, 300),

--- a/lib/storage/service.ts
+++ b/lib/storage/service.ts
@ -4,6 +4,17 @@ import { S3Client, PutObjectCommand } from "@aws-sdk/client-s3";
 import { put } from "@vercel/blob";
 import { getDefaultStorageConfig } from "@/lib/db/queries";

+type S3UploadConfig = {
+  endpoint: string;
+  region: string;
+  bucket: string;
+  accessKeyId: string;
+  secretAccessKey: string;
+  publicUrl?: string;
+};
+
+let lastKnownGoodDbS3Config: S3UploadConfig | null = null;
+
 // 获取文件的 Content-Type
 function getContentType(filename: string): string {
  const ext = filename.toLowerCase().split(".").pop();
@ -28,14 +39,7 @@ function getContentType(filename: string): string {
 async function uploadToS3(
  filename: string,
  buffer: ArrayBuffer,
-  config: {
-    endpoint: string;
-    region: string;
-    bucket: string;
-    accessKeyId: string;
-    secretAccessKey: string;
-    publicUrl?: string;
-  }
+  config: S3UploadConfig
 ) {
  const { endpoint, region, bucket, accessKeyId, secretAccessKey, publicUrl } = config;

@ -101,6 +105,49 @@ async function uploadToS3(
  }
 }

+function toDbS3UploadConfig(raw: unknown): S3UploadConfig | null {
+  const value = raw as Record<string, unknown> | null | undefined;
+  if (
+    !value ||
+    value.type !== "s3" ||
+    !value.isEnabled ||
+    !value.endpoint ||
+    !value.bucket ||
+    !value.accessKeyId ||
+    !value.secretAccessKey
+  ) {
+    return null;
+  }
+
+  return {
+    endpoint: String(value.endpoint),
+    region: String(value.region || "auto"),
+    bucket: String(value.bucket),
+    accessKeyId: String(value.accessKeyId),
+    secretAccessKey: String(value.secretAccessKey),
+    publicUrl: value.publicUrl ? String(value.publicUrl) : undefined,
+  };
+}
+
+function buildEnvS3Config(): S3UploadConfig | null {
+  if (
+    process.env.S3_ENDPOINT &&
+    process.env.S3_BUCKET &&
+    process.env.S3_ACCESS_KEY_ID &&
+    process.env.S3_SECRET_ACCESS_KEY
+  ) {
+    return {
+      endpoint: process.env.S3_ENDPOINT,
+      region: process.env.S3_REGION ?? "auto",
+      bucket: process.env.S3_BUCKET,
+      accessKeyId: process.env.S3_ACCESS_KEY_ID,
+      secretAccessKey: process.env.S3_SECRET_ACCESS_KEY,
+      publicUrl: process.env.S3_PUBLIC_URL,
+    };
+  }
+  return null;
+}
+
 // 上传文件到存储服务
 export async function uploadFile(
  filename: string,
@ -108,55 +155,41 @@ export async function uploadFile(
 ): Promise<{ url: string; pathname: string; contentType: string }> {
  console.log(`[Upload] Starting upload for file: ${filename}`);

+  let dbS3Config: S3UploadConfig | null = null;
  try {
-    // 尝试从数据库获取存储配置
    const storageConfig = await getDefaultStorageConfig();
-
-    if (
-      storageConfig &&
-      storageConfig.type === "s3" &&
-      storageConfig.endpoint &&
-      storageConfig.bucket &&
-      storageConfig.accessKeyId &&
-      storageConfig.secretAccessKey &&
-      storageConfig.isEnabled
-    ) {
-      console.log(`[Upload] Using database S3 config: ${storageConfig.name}`);
-      console.log(`[Upload] Endpoint: ${storageConfig.endpoint}`);
-      console.log(`[Upload] Bucket: ${storageConfig.bucket}`);
-
-      // 使用数据库配置的 S3/MinIO 存储
-      return await uploadToS3(filename, buffer, {
-        endpoint: storageConfig.endpoint,
-        region: storageConfig.region ?? "auto",
-        bucket: storageConfig.bucket,
-        accessKeyId: storageConfig.accessKeyId,
-        secretAccessKey: storageConfig.secretAccessKey,
-        publicUrl: storageConfig.publicUrl ?? undefined,
-      });
-    } else {
-      console.warn("[Upload] No valid S3 config in database");
+    dbS3Config = toDbS3UploadConfig(storageConfig);
+    if (dbS3Config) {
+      console.log(`[Upload] Using database S3 config: ${storageConfig?.name}`);
+      console.log(`[Upload] Endpoint: ${dbS3Config.endpoint}`);
+      console.log(`[Upload] Bucket: ${dbS3Config.bucket}`);
+      const uploaded = await uploadToS3(filename, buffer, dbS3Config);
+      lastKnownGoodDbS3Config = dbS3Config;
+      return uploaded;
    }
+    console.warn("[Upload] No valid S3 config in database");
  } catch (error) {
-    console.error("[Upload] Failed to get storage config from database:", error);
+    console.error("[Upload] Database storage path failed:", error);
+    if (lastKnownGoodDbS3Config) {
+      try {
+        console.log("[Upload] Retrying with last known good database S3 config");
+        return await uploadToS3(filename, buffer, lastKnownGoodDbS3Config);
+      } catch (retryError) {
+        console.error("[Upload] Last known S3 config upload failed:", retryError);
+      }
+    }
  }

-  // 回退到环境变量配置的 S3
-  if (
-    process.env.S3_ENDPOINT &&
-    process.env.S3_BUCKET &&
-    process.env.S3_ACCESS_KEY_ID &&
-    process.env.S3_SECRET_ACCESS_KEY
-  ) {
+  const envS3Config = buildEnvS3Config();
+  if (envS3Config) {
    console.log("[Upload] Using environment variable S3 config");
-    return await uploadToS3(filename, buffer, {
-      endpoint: process.env.S3_ENDPOINT,
-      region: process.env.S3_REGION ?? "auto",
-      bucket: process.env.S3_BUCKET,
-      accessKeyId: process.env.S3_ACCESS_KEY_ID,
-      secretAccessKey: process.env.S3_SECRET_ACCESS_KEY,
-      publicUrl: process.env.S3_PUBLIC_URL,
-    });
+    return await uploadToS3(filename, buffer, envS3Config);
+  }
+
+  if (!process.env.BLOB_READ_WRITE_TOKEN) {
+    throw new Error(
+      "No available storage backend: database S3 failed, env S3 missing, and BLOB_READ_WRITE_TOKEN is not configured."
+    );
  }

  // 最终回退到 Vercel Blob
--- a/proxy.ts
+++ b/proxy.ts
@ -4,6 +4,24 @@ import { isBlockedIP, getClientIP } from "@/lib/security/ip-blocklist";
 import { checkRateLimit, rateLimiters } from "@/lib/security/rate-limiter";
 import { isSuspiciousRequest, logSuspiciousRequest } from "@/lib/security/request-monitor";

+function isHighFrequencyReadApi(pathname: string, method: string) {
+  if (method !== "GET") return false;
+  if (/^\/api\/canvas\/[^/]+\/detail-archives$/.test(pathname)) return true;
+
+  return [
+    "/api/models",
+    "/api/canvas",
+    "/api/nav-links",
+    "/api/site-settings",
+    "/api/messages",
+    "/api/history",
+    "/api/video-tasks",
+    "/api/tutorial-videos",
+    "/api/user/membership-summary",
+    "/api/user/account-center",
+  ].includes(pathname);
+}
+
 export async function proxy(request: NextRequest) {
  const { pathname } = request.nextUrl;

@ -67,6 +85,21 @@ export async function proxy(request: NextRequest) {
        { status: 429 }
      );
    }
+  } else if (isHighFrequencyReadApi(pathname, request.method)) {
+    const pollingRateLimit = checkRateLimit(request, rateLimiters.uiPolling);
+    if (!pollingRateLimit.allowed) {
+      console.warn(`[Security] UI polling rate limit exceeded for IP: ${clientIP}`, {
+        pathname,
+        resetTime: new Date(pollingRateLimit.resetTime).toISOString(),
+      });
+      return NextResponse.json(
+        {
+          error: "Too many polling requests",
+          retryAfter: Math.ceil((pollingRateLimit.resetTime - Date.now()) / 1000),
+        },
+        { status: 429 }
+      );
+    }
  } else if (pathname.startsWith("/api/")) {
    const rateLimit = checkRateLimit(request, rateLimiters.api);
    if (!rateLimit.allowed) {