chore(release): v2.2.1

补 v2.2.0 ghcr.io 镜像构建失败（pnpm 版本不兼容 Node 20）。

.github/workflows/commitlint.yml

@@ -26,7 +26,5 @@ jobs:
         uses: wagoid/commitlint-github-action@v6
         with:
           configFile: .commitlintrc.json
-          # PR 上检查 base..head 之间所有 commit；push 上只校验最新 commit
-          firstParent: false
           failOnWarnings: false
           helpURL: https://github.com/JefferyHcool/BiliNote/blob/develop/CONTRIBUTING.md#5-提交规范

.github/workflows/main.yml

@@ -13,8 +13,6 @@ jobs:
         include:
           - platform: macos-latest
             target: universal-apple-darwin
-          - platform: ubuntu-22.04
-            target: x86_64-unknown-linux-gnu
           - platform: windows-latest
             target: x86_64-pc-windows-msvc
 
@@ -24,13 +22,6 @@ jobs:
       - name: Checkout Code
         uses: actions/checkout@v4
 
-      # Linux 系统依赖（Tauri 需要）
-      - name: Install Linux Dependencies
-        if: matrix.platform == 'ubuntu-22.04'
-        run: |
-          sudo apt-get update
-          sudo apt-get install -y libwebkit2gtk-4.1-dev libappindicator3-dev librsvg2-dev patchelf
-
       # 设置 Python 环境（带 pip 缓存）
       - name: Set up Python
         uses: actions/setup-python@v5
@@ -103,9 +94,6 @@ jobs:
           # Windows: .msi, .exe (NSIS)
           find "$BUNDLE_DIR" -name "*.msi" -exec cp {} release-artifacts/ \; 2>/dev/null || true
           find "$BUNDLE_DIR/nsis" -name "*.exe" -exec cp {} release-artifacts/ \; 2>/dev/null || true
-          # Linux: .deb, .AppImage
-          find "$BUNDLE_DIR" -name "*.deb" -exec cp {} release-artifacts/ \; 2>/dev/null || true
-          find "$BUNDLE_DIR" -name "*.AppImage" -exec cp {} release-artifacts/ \; 2>/dev/null || true
 
           echo "=== Collected artifacts ==="
           ls -lh release-artifacts/

BillNote_extension/src/background/main.ts

@@ -70,6 +70,7 @@ async function startTask(url: string): Promise<{ ok: boolean, taskId?: string, e
   // B 站：先在浏览器里抓字幕（带本地登录态 cookie），随提交带过去
   const prefetched = platform === 'bilibili' ? await fetchBilibiliSubtitle(url) : null
 
+  const formats = settings.formats || []
   try {
     const res = await fetch(`${backend}/api/generate_note`, {
       method: 'POST',
@@ -80,13 +81,15 @@ async function startTask(url: string): Promise<{ ok: boolean, taskId?: string, e
         quality: settings.quality,
         provider_id: settings.providerId,
         model_name: settings.modelName,
-        screenshot: settings.screenshot,
-        link: settings.link,
+        // backend 同时接受 format 数组与 screenshot/link 单独布尔；从 formats 派生保持单一真相源
+        format: [...formats],
+        screenshot: formats.includes('screenshot'),
+        link: formats.includes('link'),
         style: settings.style || undefined,
-        format: [
-          ...(settings.screenshot ? ['screenshot'] : []),
-          ...(settings.link ? ['link'] : []),
-        ],
+        extras: settings.extras || undefined,
+        video_understanding: settings.video_understanding || undefined,
+        video_interval: settings.video_understanding ? settings.video_interval : undefined,
+        grid_size: settings.video_understanding ? settings.grid_size : undefined,
         prefetched_transcript: prefetched ?? undefined,
       }),
     })

BillNote_extension/src/logic/constants.ts

@@ -7,9 +7,14 @@ export const DEFAULT_SETTINGS: Settings = {
   providerId: '',
   modelName: '',
   quality: 'medium',
+  formats: ['toc', 'summary'],
   screenshot: false,
   link: false,
-  style: '',
+  style: 'minimal',
+  extras: '',
+  video_understanding: false,
+  video_interval: 6,
+  grid_size: [2, 2],
 }
 
 export const MAX_TASKS = 30

BillNote_extension/src/logic/types.ts

@@ -40,6 +40,9 @@ export interface GenerateRequest {
   format?: string[]
   style?: string
   extras?: string
+  video_understanding?: boolean
+  video_interval?: number
+  grid_size?: [number, number]
   // 客户端在浏览器里直接抓到的字幕，跳过后端的 download_subtitles + 音频转写
   prefetched_transcript?: {
     language: string
@@ -78,14 +81,52 @@ export interface TaskRecord {
   result?: NoteResult
 }
 
+// 与 backend/app/gpt/prompt_builder.py note_styles 一一对齐
+export type NoteStyle =
+  | 'minimal' | 'detailed' | 'academic' | 'tutorial'
+  | 'xiaohongshu' | 'life_journal' | 'task_oriented'
+  | 'business' | 'meeting_minutes'
+
+// 与 backend/app/gpt/prompt_builder.py note_formats 一一对齐
+export type NoteFormat = 'toc' | 'link' | 'screenshot' | 'summary'
+
+export const NOTE_STYLES: Array<{ value: NoteStyle, label: string }> = [
+  { value: 'minimal', label: '精简' },
+  { value: 'detailed', label: '详细' },
+  { value: 'tutorial', label: '教程' },
+  { value: 'academic', label: '学术' },
+  { value: 'xiaohongshu', label: '小红书' },
+  { value: 'life_journal', label: '生活向' },
+  { value: 'task_oriented', label: '任务导向' },
+  { value: 'business', label: '商业风格' },
+  { value: 'meeting_minutes', label: '会议纪要' },
+]
+
+export const NOTE_FORMATS: Array<{ value: NoteFormat, label: string }> = [
+  { value: 'toc', label: '目录' },
+  { value: 'summary', label: 'AI 总结' },
+  { value: 'screenshot', label: '原片截图' },
+  { value: 'link', label: '原片跳转' },
+]
+
 export interface Settings {
   backendUrl: string
   providerId: string
   modelName: string
   quality: Quality
+  // 输出 format 的 toggle 集合（screenshot / link 与下方两个布尔保持联动）
+  formats: NoteFormat[]
   screenshot: boolean
   link: boolean
-  style: string
+  style: NoteStyle
+  extras: string
+  // 多模态视频理解：抽帧拼图喂给视觉模型，提升画面相关问题的回答质量
+  // 要求所选 model 是视觉模型（如 gpt-4o / gemini / claude-opus 系列），文字模型会忽略图片
+  video_understanding: boolean
+  // 抽帧间隔（秒），范围 1-30，默认 6
+  video_interval: number
+  // 拼图网格 [rows, cols]，每张拼图最多 rows*cols 帧。默认 [2,2]
+  grid_size: [number, number]
 }
 
 export interface ProviderUpdatePayload {

BillNote_extension/src/options/pages/General.vue

@@ -3,9 +3,16 @@ import { onMounted, ref } from 'vue'
 import { getProviders, ping } from '~/logic/api'
 import { settings, settingsReady } from '~/logic/storage'
 import { getModelsByProvider } from '~/logic/api'
-import type { Model, Provider } from '~/logic/types'
+import { NOTE_FORMATS, NOTE_STYLES, type Model, type NoteFormat, type Provider } from '~/logic/types'
 import { watch } from 'vue'
 
+function toggleFormat(value: NoteFormat, checked: boolean) {
+  const cur = settings.value.formats || []
+  settings.value.formats = checked
+    ? Array.from(new Set([...cur, value]))
+    : cur.filter(v => v !== value)
+}
+
 const providers = ref<Provider[]>([])
 const models = ref<Model[]>([])
 const status = ref<{ kind: 'idle' | 'ok' | 'err', text: string }>({ kind: 'idle', text: '' })
@@ -128,13 +135,67 @@ onMounted(async () => {
         </label>
         <label class="flex flex-col gap-1">
           <span class="text-gray-600">笔记风格</span>
-          <input v-model="settings.style" class="input" placeholder="留空使用默认">
+          <select v-model="settings.style" class="input">
+            <option v-for="s in NOTE_STYLES" :key="s.value" :value="s.value">{{ s.label }}</option>
+          </select>
         </label>
-        <label class="flex items-center gap-2">
-          <input v-model="settings.screenshot" type="checkbox"> 自动插入截图
+      </div>
+
+      <div class="flex flex-col gap-1 text-sm">
+        <span class="text-gray-600">输出形式（与 web 端 NoteForm 对齐）</span>
+        <div class="flex flex-wrap gap-x-4 gap-y-2">
+          <label v-for="f in NOTE_FORMATS" :key="f.value" class="flex items-center gap-2">
+            <input
+              type="checkbox"
+              :checked="(settings.formats || []).includes(f.value)"
+              @change="toggleFormat(f.value, ($event.target as HTMLInputElement).checked)"
+            >
+            {{ f.label }}
+          </label>
+        </div>
+      </div>
+
+      <label class="flex flex-col gap-1 text-sm">
+        <span class="text-gray-600">额外提示词（追加到 prompt 末尾）</span>
+        <textarea
+          v-model="settings.extras"
+          class="input resize-y"
+          rows="3"
+          placeholder="例如：重点关注游戏开发部分；保留所有专业术语原文"
+        />
+      </label>
+    </section>
+
+    <section class="section-card">
+      <h2 class="font-semibold">视频理解（多模态）</h2>
+      <p class="text-xs text-gray-500">
+        启用后会按抽帧间隔截取视频帧拼成网格图，连同字幕一起喂给视觉模型，提升画面相关问题的回答质量。
+        <strong class="text-amber-700">需要选择视觉模型</strong>（GPT-4o / Gemini / Claude 等），文字模型会忽略图片。
+      </p>
+      <label class="flex items-center gap-2 text-sm">
+        <input v-model="settings.video_understanding" type="checkbox">
+        启用视频理解
+      </label>
+      <div v-if="settings.video_understanding" class="grid grid-cols-3 gap-3 text-sm">
+        <label class="flex flex-col gap-1">
+          <span class="text-gray-600">抽帧间隔(秒, 1-30)</span>
+          <input v-model.number="settings.video_interval" type="number" min="1" max="30" class="input">
         </label>
-        <label class="flex items-center gap-2">
-          <input v-model="settings.link" type="checkbox"> 插入原片跳转链接
+        <label class="flex flex-col gap-1">
+          <span class="text-gray-600">拼图行 (1-10)</span>
+          <input
+            :value="settings.grid_size?.[0] ?? 2"
+            type="number" min="1" max="10" class="input"
+            @input="settings.grid_size = [Number(($event.target as HTMLInputElement).value) || 2, settings.grid_size?.[1] ?? 2]"
+          >
+        </label>
+        <label class="flex flex-col gap-1">
+          <span class="text-gray-600">拼图列 (1-10)</span>
+          <input
+            :value="settings.grid_size?.[1] ?? 2"
+            type="number" min="1" max="10" class="input"
+            @input="settings.grid_size = [settings.grid_size?.[0] ?? 2, Number(($event.target as HTMLInputElement).value) || 2]"
+          >
         </label>
       </div>
     </section>

BillNote_extension/src/popup/Popup.vue

@@ -4,7 +4,7 @@ import { detectPlatform } from '~/logic/platform'
 import { settings, settingsReady, tasks, tasksReady, upsertTask } from '~/logic/storage'
 import { generateNote, getTaskStatus, resolveImageUrl } from '~/logic/api'
 import { fetchBilibiliSubtitle } from '~/logic/bilibili-subtitle'
-import type { TaskRecord } from '~/logic/types'
+import { NOTE_FORMATS, NOTE_STYLES, type NoteFormat, type TaskRecord } from '~/logic/types'
 
 const tabUrl = ref<string>('')
 const tabTitle = ref<string>('')
@@ -67,19 +67,22 @@ async function start() {
   try {
     // B 站：在用户浏览器里直接抓字幕（带本地登录态 cookie），跳过后端的 download_subtitles 与音频转写
     const prefetched = platform.value === 'bilibili' ? await fetchBilibiliSubtitle(tabUrl.value) : null
+    const formats = settings.value.formats || []
     const { task_id } = await generateNote({
       video_url: tabUrl.value,
       platform: platform.value!,
       quality: settings.value.quality,
       provider_id: settings.value.providerId,
       model_name: settings.value.modelName,
-      screenshot: settings.value.screenshot,
-      link: settings.value.link,
+      // backend VideoRequest 同时接受 format 数组与 screenshot/link 单独布尔，从 formats 派生保持单一真相源
+      format: [...formats],
+      screenshot: formats.includes('screenshot'),
+      link: formats.includes('link'),
       style: settings.value.style || undefined,
-      format: [
-        ...(settings.value.screenshot ? ['screenshot'] : []),
-        ...(settings.value.link ? ['link'] : []),
-      ],
+      extras: settings.value.extras || undefined,
+      video_understanding: settings.value.video_understanding || undefined,
+      video_interval: settings.value.video_understanding ? settings.value.video_interval : undefined,
+      grid_size: settings.value.video_understanding ? settings.value.grid_size : undefined,
       prefetched_transcript: prefetched ?? undefined,
     })
     activeTaskId.value = task_id
@@ -108,6 +111,13 @@ function openOptions() {
   browser.runtime.openOptionsPage()
 }
 
+function toggleFormat(value: NoteFormat, checked: boolean) {
+  const cur = settings.value.formats || []
+  settings.value.formats = checked
+    ? Array.from(new Set([...cur, value]))
+    : cur.filter(v => v !== value)
+}
+
 async function openSidePanel() {
   // 只能在用户操作触发的同步上下文里调，且需要明确的 tabId
   try {
@@ -176,7 +186,7 @@ onUnmounted(() => {
     </div>
 
     <fieldset class="border rounded p-2 flex flex-col gap-2" :disabled="!supported || submitting">
-      <div class="grid grid-cols-3 gap-2 text-xs">
+      <div class="grid grid-cols-2 gap-2 text-xs">
         <label class="flex flex-col gap-1">
           <span class="text-gray-600">画质</span>
           <select v-model="settings.quality" class="border rounded px-1 py-0.5">
@@ -185,14 +195,76 @@ onUnmounted(() => {
             <option value="slow">高质</option>
           </select>
         </label>
-        <label class="flex items-center gap-1 mt-4">
-          <input v-model="settings.screenshot" type="checkbox"> 截图
-        </label>
-        <label class="flex items-center gap-1 mt-4">
-          <input v-model="settings.link" type="checkbox"> 跳转
+        <label class="flex flex-col gap-1">
+          <span class="text-gray-600">笔记风格</span>
+          <select v-model="settings.style" class="border rounded px-1 py-0.5">
+            <option v-for="s in NOTE_STYLES" :key="s.value" :value="s.value">{{ s.label }}</option>
+          </select>
         </label>
       </div>
 
+      <div class="flex flex-col gap-1 text-xs">
+        <span class="text-gray-600">输出形式</span>
+        <div class="flex flex-wrap gap-x-3 gap-y-1">
+          <label v-for="f in NOTE_FORMATS" :key="f.value" class="flex items-center gap-1">
+            <input
+              type="checkbox"
+              :checked="(settings.formats || []).includes(f.value)"
+              @change="toggleFormat(f.value, ($event.target as HTMLInputElement).checked)"
+            >
+            {{ f.label }}
+          </label>
+        </div>
+      </div>
+
+      <details class="text-xs">
+        <summary class="cursor-pointer text-gray-500">高级</summary>
+        <label class="flex flex-col gap-1 mt-2">
+          <span class="text-gray-600">额外提示词（追加到 prompt 末尾）</span>
+          <textarea
+            v-model="settings.extras"
+            class="border rounded px-1 py-1 resize-y"
+            rows="2"
+            placeholder="例如：重点关注游戏开发部分；保留所有专业术语原文"
+          />
+        </label>
+        <label class="flex items-center gap-2 mt-2">
+          <input v-model="settings.video_understanding" type="checkbox">
+          <span class="text-gray-600">启用视频理解（抽帧拼图喂视觉模型）</span>
+        </label>
+        <div v-if="settings.video_understanding" class="grid grid-cols-3 gap-2 mt-2">
+          <label class="flex flex-col gap-1">
+            <span class="text-gray-600">抽帧间隔(秒)</span>
+            <input
+              v-model.number="settings.video_interval"
+              type="number" min="1" max="30"
+              class="border rounded px-1 py-0.5"
+            >
+          </label>
+          <label class="flex flex-col gap-1">
+            <span class="text-gray-600">拼图行</span>
+            <input
+              :value="settings.grid_size?.[0] ?? 2"
+              type="number" min="1" max="10"
+              class="border rounded px-1 py-0.5"
+              @input="settings.grid_size = [Number(($event.target as HTMLInputElement).value) || 2, settings.grid_size?.[1] ?? 2]"
+            >
+          </label>
+          <label class="flex flex-col gap-1">
+            <span class="text-gray-600">拼图列</span>
+            <input
+              :value="settings.grid_size?.[1] ?? 2"
+              type="number" min="1" max="10"
+              class="border rounded px-1 py-0.5"
+              @input="settings.grid_size = [settings.grid_size?.[0] ?? 2, Number(($event.target as HTMLInputElement).value) || 2]"
+            >
+          </label>
+        </div>
+        <p v-if="settings.video_understanding" class="text-amber-700 mt-1">
+          ⚠ 需要选择视觉模型（GPT-4o / Gemini / Claude 等），文字模型会忽略图片
+        </p>
+      </details>
+
       <div class="text-xs text-gray-600">
         <span v-if="settings.providerId && settings.modelName">
           模型：{{ settings.modelName }}

BillNote_frontend/Dockerfile

@@ -1,7 +1,9 @@
 # === 前端构建阶段 ===
-FROM node:18-alpine AS builder
+# Tailwind v4 / Vite 6 需要 Node 20+，alpine + pnpm 会按 lockfile 拉 musl native binary。
+FROM node:20-alpine AS builder
 
-RUN corepack enable && corepack prepare pnpm@latest --activate
+# pnpm pin 到 9.x：lockfile 是 v9 生成；pnpm 11 要求 Node 22+ 与 node:20 不兼容
+RUN corepack enable && corepack prepare pnpm@9.15.0 --activate
 
 WORKDIR /app
 

BillNote_frontend/src-tauri/src/lib.rs

@@ -1,8 +1,14 @@
-use tauri::{Manager, Emitter};
+use tauri::{Manager, Emitter, State};
 use tauri_plugin_shell::ShellExt;
-use tauri_plugin_shell::process::CommandEvent;
+use tauri_plugin_shell::process::{CommandChild, CommandEvent};
 use std::env;
 use std::collections::HashMap;
+use std::path::Path;
+use std::sync::Mutex;
+use serde::Serialize;
+
+// Sidecar 子进程句柄，用 Mutex 包裹方便 restart 时杀旧进程
+struct SidecarHandle(Mutex<Option<CommandChild>>);
 
 #[cfg_attr(mobile, tauri::mobile_entry_point)]
 pub fn run() {
@@ -18,77 +24,31 @@ pub fn run() {
             }
 
             let exe_path = env::current_exe().expect("无法获取当前可执行文件路径");
-            let sidecar_dir = exe_path.parent().expect("无法获取可执行文件的父目录");
 
-            // 收集所有系统环境变量
-            let mut all_env_vars = HashMap::new();
-            for (key, value) in env::vars() {
-                all_env_vars.insert(key, value);
+            // 安装路径诊断：PyInstaller sidecar 在含非 ASCII / 空格的路径下经常炸（README 已警告但缺主动防御）
+            // 命中时把诊断信息 emit 给前端，由顶端横幅展示，不阻断启动
+            let diag = analyze_install_path(&exe_path);
+            if diag.path_has_non_ascii || diag.path_has_space || !diag.parent_writable {
+                let app_handle = app.handle().clone();
+                // 等前端首屏挂载好 listener；setup 阶段 window 已存在但 React 还没 render
+                // 用独立线程 + 标准 sleep，不引入 tokio 依赖
+                std::thread::spawn(move || {
+                    std::thread::sleep(std::time::Duration::from_millis(1500));
+                    if let Some(window) = app_handle.get_webview_window("main") {
+                        let _ = window.emit("backend-warning", &diag);
+                    }
+                });
             }
 
-            // 增强 PATH 环境变量，添加常见的二进制路径
-            let current_path = all_env_vars.get("PATH").cloned().unwrap_or_default();
-            let additional_paths = get_additional_binary_paths();
-            let enhanced_path = enhance_path_variable(&current_path, &additional_paths);
-            all_env_vars.insert("PATH".to_string(), enhanced_path);
-
-            // 打印一些关键环境变量用于调试
-            println!("Enhanced PATH: {}", all_env_vars.get("PATH").unwrap_or(&"Not found".to_string()));
-            println!("Total environment variables: {}", all_env_vars.len());
-
             // 检查 ffmpeg 是否在 PATH 中可用
             check_ffmpeg_availability();
 
-            // 启动 Python 后端侧车
-            let mut sidecar_command = app.shell().sidecar("BiliNoteBackend").unwrap();
-
-            // 设置所有环境变量到 sidecar
-            for (key, value) in &all_env_vars {
-                sidecar_command = sidecar_command.env(key, value);
-            }
-
-            let (mut rx, _child) = sidecar_command
-                .current_dir(sidecar_dir)
-                .spawn()
-                .expect("Failed to spawn sidecar");
-
-            // 获取主窗口句柄用于发送事件
-            let window = app.get_webview_window("main").unwrap();
-
-            tauri::async_runtime::spawn(async move {
-                // 读取诸如 stdout 之类的事件
-                while let Some(event) = rx.recv().await {
-                    match event {
-                        CommandEvent::Stdout(line) => {
-                            let output = String::from_utf8_lossy(&line);
-                            println!("Backend stdout: {}", output);
-
-                            // 发送到前端
-                            window
-                                .emit("backend-message", Some(format!("'{}'", output)))
-                                .expect("failed to emit event");
-                        }
-                        CommandEvent::Stderr(line) => {
-                            let error = String::from_utf8_lossy(&line);
-                            eprintln!("Backend stderr: {}", error);
-
-                            window
-                                .emit("backend-error", Some(format!("'{}'", error)))
-                                .expect("failed to emit event");
-                        }
-                        CommandEvent::Terminated(payload) => {
-                            println!("Backend terminated with code: {:?}", payload.code);
-                            window
-                                .emit("backend-terminated", Some(payload.code))
-                                .expect("failed to emit event");
-                            break;
-                        }
-                        _ => {
-                            println!("Backend event: {:?}", event);
-                        }
-                    }
-                }
-            });
+            // 启动 Sidecar 并把 child handle 存到 state，方便后续 restart_backend_sidecar 使用
+            let child = spawn_backend_sidecar(app.handle()).map_err(|e| {
+                eprintln!("Sidecar 启动失败: {}", e);
+                e
+            })?;
+            app.manage(SidecarHandle(Mutex::new(Some(child))));
 
             Ok(())
         })
@@ -96,7 +56,9 @@ pub fn run() {
             get_system_env_vars,
             find_executable_path,
             run_command_with_env,
-            test_ffmpeg_access
+            test_ffmpeg_access,
+            get_install_path_diagnostics,
+            restart_backend_sidecar
         ])
         .run(tauri::generate_context!())
         .expect("error while running tauri application");
@@ -268,6 +230,150 @@ async fn test_ffmpeg_access() -> Result<String, String> {
     run_command_with_env("ffmpeg".to_string(), vec!["-version".to_string()]).await
 }
 
+// 启动后端 Sidecar：负责装环境变量、spawn、挂 stdout/stderr/terminated 监听并 emit 给前端。
+// 第一次启动 + restart_backend_sidecar 都走这里，保持单一启动路径。
+fn spawn_backend_sidecar(app_handle: &tauri::AppHandle) -> Result<CommandChild, String> {
+    let exe_path = env::current_exe().map_err(|e| format!("无法获取可执行文件路径: {}", e))?;
+    let sidecar_dir = exe_path
+        .parent()
+        .ok_or("无法获取可执行文件父目录")?
+        .to_path_buf();
+
+    // 收集所有系统环境变量并增强 PATH（含 ffmpeg 常见安装位置）
+    let mut all_env_vars = HashMap::new();
+    for (key, value) in env::vars() {
+        all_env_vars.insert(key, value);
+    }
+    let current_path = all_env_vars.get("PATH").cloned().unwrap_or_default();
+    let additional_paths = get_additional_binary_paths();
+    let enhanced_path = enhance_path_variable(&current_path, &additional_paths);
+    all_env_vars.insert("PATH".to_string(), enhanced_path);
+
+    let mut sidecar_command = app_handle
+        .shell()
+        .sidecar("BiliNoteBackend")
+        .map_err(|e| format!("找不到 BiliNoteBackend sidecar: {}", e))?;
+    for (key, value) in &all_env_vars {
+        sidecar_command = sidecar_command.env(key, value);
+    }
+
+    let (mut rx, child) = sidecar_command
+        .current_dir(sidecar_dir)
+        .spawn()
+        .map_err(|e| format!("spawn sidecar 失败: {}", e))?;
+
+    // 异步监听 stdout / stderr / terminated 事件，转发到前端 webview
+    let app_handle_for_listener = app_handle.clone();
+    tauri::async_runtime::spawn(async move {
+        while let Some(event) = rx.recv().await {
+            // window 句柄每次重新取，允许窗口关闭重开
+            let window = app_handle_for_listener.get_webview_window("main");
+            match event {
+                CommandEvent::Stdout(line) => {
+                    let output = String::from_utf8_lossy(&line).to_string();
+                    println!("Backend stdout: {}", output);
+                    if let Some(w) = window {
+                        let _ = w.emit("backend-message", Some(output));
+                    }
+                }
+                CommandEvent::Stderr(line) => {
+                    let error = String::from_utf8_lossy(&line).to_string();
+                    eprintln!("Backend stderr: {}", error);
+                    if let Some(w) = window {
+                        let _ = w.emit("backend-error", Some(error));
+                    }
+                }
+                CommandEvent::Terminated(payload) => {
+                    println!("Backend terminated with code: {:?}", payload.code);
+                    if let Some(w) = window {
+                        let _ = w.emit("backend-terminated", Some(payload.code));
+                    }
+                    break;
+                }
+                _ => {
+                    println!("Backend event: {:?}", event);
+                }
+            }
+        }
+    });
+
+    Ok(child)
+}
+
+// 重启 sidecar：杀旧 child，spawn 新 child，回写到 state。
+#[tauri::command]
+fn restart_backend_sidecar(
+    state: State<'_, SidecarHandle>,
+    app: tauri::AppHandle,
+) -> Result<(), String> {
+    // 1. 拿出旧 child 并 kill（kill 失败也继续，可能进程已经退了）
+    {
+        let mut guard = state.0.lock().map_err(|e| format!("锁 sidecar state 失败: {}", e))?;
+        if let Some(child) = guard.take() {
+            let _ = child.kill();
+        }
+    }
+    // 2. 重新 spawn
+    let new_child = spawn_backend_sidecar(&app)?;
+    {
+        let mut guard = state.0.lock().map_err(|e| format!("锁 sidecar state 失败: {}", e))?;
+        *guard = Some(new_child);
+    }
+    // 3. emit 一个事件让前端知道已重启
+    if let Some(window) = app.get_webview_window("main") {
+        let _ = window.emit("backend-restarted", ());
+    }
+    Ok(())
+}
+
+// 安装路径诊断：PyInstaller 在含非 ASCII / 空格的路径下加载 _internal/* 经常炸；
+// 父目录不可写时模型 / 配置 / 日志也无法落盘
+#[derive(Serialize, Clone)]
+struct InstallPathDiagnostics {
+    exe_path: String,
+    path_has_non_ascii: bool,
+    path_has_space: bool,
+    parent_writable: bool,
+    platform: String,
+}
+
+fn analyze_install_path(exe_path: &Path) -> InstallPathDiagnostics {
+    let path_str = exe_path.to_string_lossy().to_string();
+    // 不在 ASCII 范围内的字符（中文 / 日文 / 西里尔等都会命中 PyInstaller 路径解析坑）
+    let has_non_ascii = path_str.chars().any(|c| !c.is_ascii());
+    // 空格本身在 Windows shell 引号场景偶尔出问题，且 macOS path 里也偶尔触发 sidecar 启动失败
+    let has_space = path_str.contains(' ');
+    // 父目录可写：PyInstaller 解压 _internal/、写日志、写配置都需要这个
+    let parent = exe_path.parent();
+    let parent_writable = parent
+        .and_then(|p| {
+            let probe = p.join(".bilinote_write_probe");
+            match std::fs::write(&probe, b"x") {
+                Ok(_) => {
+                    let _ = std::fs::remove_file(&probe);
+                    Some(true)
+                }
+                Err(_) => Some(false),
+            }
+        })
+        .unwrap_or(false);
+
+    InstallPathDiagnostics {
+        exe_path: path_str,
+        path_has_non_ascii: has_non_ascii,
+        path_has_space: has_space,
+        parent_writable,
+        platform: std::env::consts::OS.to_string(),
+    }
+}
+
+// Tauri 命令：让前端按需重新查询诊断结果（比如用户卸载到新目录后重启）
+#[tauri::command]
+fn get_install_path_diagnostics() -> InstallPathDiagnostics {
+    let exe_path = env::current_exe().unwrap_or_default();
+    analyze_install_path(&exe_path)
+}
+
 // 可选：添加一个函数来动态更新 sidecar 的环境变量
 #[tauri::command]
 async fn update_sidecar_environment(

BillNote_frontend/src/App.tsx

@@ -5,11 +5,23 @@ import { useTaskPolling } from '@/hooks/useTaskPolling.ts'
 import { useCheckBackend } from '@/hooks/useCheckBackend.ts'
 import { systemCheck } from '@/services/system.ts'
 import BackendInitDialog from '@/components/BackendInitDialog'
+import StartupBanner from '@/components/SystemDiagnostic/StartupBanner'
+import BackendHealthIndicator from '@/components/BackendHealth/BackendHealthIndicator'
 import Index from '@/pages/Index.tsx'
 import { HomePage } from './pages/HomePage/Home.tsx'
 
 // 非首屏页面使用 React.lazy 按需加载
+const Onboarding = lazy(() => import('@/pages/Onboarding'))
 const SettingPage = lazy(() => import('./pages/SettingPage/index.tsx'))
+
+// 桌面端首启引导守卫：未完成 onboarding 时强制跳到 /onboarding
+function OnboardingGuard({ children }: { children: React.ReactNode }) {
+  const isTauri = typeof window !== 'undefined' && '__TAURI_INTERNALS__' in window
+  // 仅在 Tauri 桌面端拦截；纯 web 端不打扰用户
+  if (!isTauri) return <>{children}</>
+  if (localStorage.getItem('bilinote-onboarded') !== '1') return <Navigate to="/onboarding" replace />
+  return <>{children}</>
+}
 const Model = lazy(() => import('@/pages/SettingPage/Model.tsx'))
 const ProviderForm = lazy(() => import('@/components/Form/modelForm/Form.tsx'))
 const AboutPage = lazy(() => import('@/pages/SettingPage/about.tsx'))
@@ -34,6 +46,7 @@ function App() {
   if (!initialized) {
     return (
       <>
+        <StartupBanner />
         <BackendInitDialog open={loading} />
       </>
     )
@@ -42,10 +55,13 @@ function App() {
   // 后端已初始化，渲染主应用
   return (
     <>
+      <StartupBanner />
+      <BackendHealthIndicator />
       <BrowserRouter>
         <Suspense fallback={<div className="flex h-screen items-center justify-center">加载中…</div>}>
           <Routes>
-            <Route path="/" element={<Index />}>
+            <Route path="/onboarding" element={<Onboarding />} />
+            <Route path="/" element={<OnboardingGuard><Index /></OnboardingGuard>}>
               <Route index element={<HomePage />} />
               <Route path="settings" element={<SettingPage />}>
                 <Route index element={<Navigate to="model" replace />} />

BillNote_frontend/src/components/BackendHealth/BackendHealthIndicator.tsx

@@ -0,0 +1,111 @@
+import { useEffect, useState } from 'react'
+import { useBackendEvents } from './useBackendEvents'
+import BackendLogPanel from './BackendLogPanel'
+
+// 健康度判定：
+// - 绿：sidecar running 且 /sys_health 通
+// - 黄：sidecar running 但 /sys_health 失败 (ffmpeg 缺等)
+// - 红：sidecar terminated 或 /sys_health 连续 3 次失败
+
+type Health = 'green' | 'yellow' | 'red' | 'unknown'
+
+const HEALTH_POLL_MS = 5000
+const SYS_HEALTH_PATH = '/api/sys_health'
+
+function backendBase(): string {
+  // 与 services/request.ts 用的一致
+  const fromEnv = (import.meta as any).env?.VITE_API_BASE_URL as string | undefined
+  return (fromEnv ?? '').replace(/\/$/, '')
+}
+
+const BackendHealthIndicator = () => {
+  const { status, isTauri, exitCode, logs, restart, copyLogs } = useBackendEvents()
+  const [open, setOpen] = useState(false)
+  const [healthCheckFailures, setHealthCheckFailures] = useState(0)
+  const [lastHealthOk, setLastHealthOk] = useState<boolean | null>(null)
+
+  // 仅在 Tauri 环境挂指示器；纯 web 用户由 useCheckBackend 接管
+  useEffect(() => {
+    if (!isTauri) return
+    let mounted = true
+
+    async function ping() {
+      try {
+        const res = await fetch(`${backendBase()}${SYS_HEALTH_PATH}`)
+        const ok = res.ok
+        if (!mounted) return
+        if (ok) {
+          setHealthCheckFailures(0)
+          setLastHealthOk(true)
+        }
+        else {
+          setHealthCheckFailures(c => c + 1)
+          setLastHealthOk(false)
+        }
+      }
+      catch {
+        if (!mounted) return
+        setHealthCheckFailures(c => c + 1)
+        setLastHealthOk(false)
+      }
+    }
+
+    ping()
+    const t = setInterval(ping, HEALTH_POLL_MS)
+    return () => {
+      mounted = false
+      clearInterval(t)
+    }
+  }, [isTauri])
+
+  if (!isTauri) return null
+
+  const health: Health = (() => {
+    if (status === 'terminated') return 'red'
+    if (healthCheckFailures >= 3) return 'red'
+    if (lastHealthOk === false) return 'yellow'
+    if (lastHealthOk === true) return 'green'
+    return 'unknown'
+  })()
+
+  const colorMap: Record<Health, string> = {
+    green: 'bg-green-500',
+    yellow: 'bg-amber-500',
+    red: 'bg-red-500',
+    unknown: 'bg-gray-400',
+  }
+
+  const labelMap: Record<Health, string> = {
+    green: '后端运行正常',
+    yellow: '后端运行中（部分检查未通过）',
+    red: status === 'terminated' ? `后端已退出 (code=${exitCode ?? 'unknown'})` : '后端无响应',
+    unknown: '后端状态未知',
+  }
+
+  return (
+    <>
+      <button
+        className="fixed right-3 bottom-3 z-[9998] flex items-center gap-2 rounded-full border bg-white px-3 py-1.5 text-xs shadow hover:shadow-md"
+        title={labelMap[health]}
+        onClick={() => setOpen(true)}
+      >
+        <span className={`inline-block h-2 w-2 rounded-full ${colorMap[health]}${health === 'red' || health === 'yellow' ? ' animate-pulse' : ''}`} />
+        <span className="text-gray-700">后端</span>
+      </button>
+
+      {open && (
+        <BackendLogPanel
+          status={status}
+          exitCode={exitCode}
+          logs={logs}
+          health={health}
+          onRestart={restart}
+          onCopyLogs={copyLogs}
+          onClose={() => setOpen(false)}
+        />
+      )}
+    </>
+  )
+}
+
+export default BackendHealthIndicator

BillNote_frontend/src/components/BackendHealth/BackendLogPanel.tsx

@@ -0,0 +1,108 @@
+import { useEffect, useRef, useState } from 'react'
+import type { LogEntry, BackendStatus } from './useBackendEvents'
+
+interface Props {
+  status: BackendStatus
+  exitCode: number | null
+  logs: LogEntry[]
+  health: 'green' | 'yellow' | 'red' | 'unknown'
+  onRestart: () => Promise<void>
+  onCopyLogs: () => Promise<boolean>
+  onClose: () => void
+}
+
+const BackendLogPanel = ({ status, exitCode, logs, health, onRestart, onCopyLogs, onClose }: Props) => {
+  const [restarting, setRestarting] = useState(false)
+  const [copied, setCopied] = useState(false)
+  const scrollRef = useRef<HTMLDivElement>(null)
+
+  // 新日志进来自动滚到底
+  useEffect(() => {
+    if (scrollRef.current)
+      scrollRef.current.scrollTop = scrollRef.current.scrollHeight
+  }, [logs])
+
+  async function handleRestart() {
+    setRestarting(true)
+    try { await onRestart() }
+    catch { /* errors already in log via useBackendEvents */ }
+    finally { setRestarting(false) }
+  }
+
+  async function handleCopy() {
+    const ok = await onCopyLogs()
+    setCopied(ok)
+    setTimeout(() => setCopied(false), 1500)
+  }
+
+  return (
+    <>
+      {/* 半透明遮罩 */}
+      <div className="fixed inset-0 z-[9998] bg-black/20" onClick={onClose} />
+
+      <aside className="fixed right-0 bottom-0 top-0 z-[9999] flex w-[480px] max-w-[90vw] flex-col border-l bg-white shadow-2xl">
+        <header className="flex items-center justify-between border-b px-4 py-3">
+          <div>
+            <h2 className="text-base font-semibold">后端运行状态</h2>
+            <div className="mt-0.5 text-xs text-gray-500">
+              {status === 'terminated'
+                ? `已退出（退出码 ${exitCode ?? 'unknown'}）`
+                : health === 'red'
+                  ? '运行中但无响应'
+                  : health === 'yellow'
+                    ? '运行中，部分系统检查未通过'
+                    : '运行正常'}
+            </div>
+          </div>
+          <button className="rounded p-1 text-gray-500 hover:bg-gray-100" onClick={onClose}>✕</button>
+        </header>
+
+        <div className="flex items-center gap-2 border-b px-4 py-2">
+          <button
+            className="rounded bg-blue-600 px-3 py-1 text-sm text-white hover:bg-blue-700 disabled:opacity-50"
+            disabled={restarting}
+            onClick={handleRestart}
+          >
+            {restarting ? '重启中…' : '重启后端'}
+          </button>
+          <button
+            className="rounded bg-gray-100 px-3 py-1 text-sm text-gray-700 hover:bg-gray-200"
+            onClick={handleCopy}
+          >
+            {copied ? '已复制 ✓' : '复制日志'}
+          </button>
+          <span className="ml-auto text-xs text-gray-400">
+            最近 {logs.length} 行
+          </span>
+        </div>
+
+        <div
+          ref={scrollRef}
+          className="flex-1 overflow-auto bg-gray-900 p-3 font-mono text-xs text-gray-100"
+        >
+          {logs.length === 0 ? (
+            <div className="text-gray-500 italic">暂无日志输出</div>
+          ) : (
+            logs.map((l, i) => (
+              <div
+                key={`${l.ts}-${i}`}
+                className={`whitespace-pre-wrap break-all leading-snug ${l.level === 'error' ? 'text-red-300' : 'text-gray-100'}`}
+              >
+                <span className="mr-2 text-gray-500">
+                  {new Date(l.ts).toISOString().slice(11, 19)}
+                </span>
+                {l.text}
+              </div>
+            ))
+          )}
+        </div>
+
+        <footer className="border-t px-4 py-2 text-xs text-gray-500">
+          后端进程退出 / 无响应时，先点「重启后端」；仍不行复制日志去 issue 反馈。
+        </footer>
+      </aside>
+    </>
+  )
+}
+
+export default BackendLogPanel

BillNote_frontend/src/components/BackendHealth/useBackendEvents.ts

@@ -0,0 +1,119 @@
+import { useEffect, useRef, useState } from 'react'
+
+// 桌面端 Sidecar 健康度。监听 Tauri 侧的 backend-message / backend-error /
+// backend-terminated / backend-restarted 事件，把 stdout/stderr 缓冲成 ring buffer，
+// 同时维护进程运行状态。
+
+const isTauri = typeof window !== 'undefined' && '__TAURI_INTERNALS__' in window
+
+export type LogLevel = 'info' | 'error'
+
+export interface LogEntry {
+  level: LogLevel
+  text: string
+  ts: number
+}
+
+export type BackendStatus = 'running' | 'terminated'
+
+const MAX_LOG_LINES = 200
+
+interface BackendEvents {
+  status: BackendStatus
+  exitCode: number | null
+  logs: LogEntry[]
+  /** 调 Tauri 命令重启 sidecar */
+  restart: () => Promise<void>
+  /** 复制全部日志到剪贴板 */
+  copyLogs: () => Promise<boolean>
+  isTauri: boolean
+}
+
+export function useBackendEvents(): BackendEvents {
+  const [status, setStatus] = useState<BackendStatus>('running')
+  const [exitCode, setExitCode] = useState<number | null>(null)
+  const [logs, setLogs] = useState<LogEntry[]>([])
+  // 用 ref 持有最新 logs 数组，append 时不被闭包陷阱卡到旧值
+  const logsRef = useRef<LogEntry[]>([])
+
+  function append(entry: LogEntry) {
+    const next = logsRef.current.concat(entry)
+    if (next.length > MAX_LOG_LINES)
+      next.splice(0, next.length - MAX_LOG_LINES)
+    logsRef.current = next
+    setLogs(next)
+  }
+
+  useEffect(() => {
+    if (!isTauri) return
+
+    let unlisteners: Array<() => void> = []
+
+    ;(async () => {
+      const { listen } = await import('@tauri-apps/api/event')
+
+      const offMsg = await listen<string>('backend-message', event => {
+        append({ level: 'info', text: stripQuotes(event.payload), ts: Date.now() })
+      })
+      const offErr = await listen<string>('backend-error', event => {
+        append({ level: 'error', text: stripQuotes(event.payload), ts: Date.now() })
+      })
+      const offTerm = await listen<number | null>('backend-terminated', event => {
+        setStatus('terminated')
+        setExitCode(event.payload ?? null)
+        append({
+          level: 'error',
+          text: `[Backend terminated] code=${event.payload ?? 'unknown'}`,
+          ts: Date.now(),
+        })
+      })
+      const offRestart = await listen('backend-restarted', () => {
+        setStatus('running')
+        setExitCode(null)
+        append({ level: 'info', text: '[Backend restarted]', ts: Date.now() })
+      })
+
+      unlisteners = [offMsg, offErr, offTerm, offRestart]
+    })()
+
+    return () => {
+      unlisteners.forEach(fn => fn())
+    }
+  }, [])
+
+  async function restart() {
+    if (!isTauri) return
+    const { invoke } = await import('@tauri-apps/api/core')
+    try {
+      await invoke('restart_backend_sidecar')
+    }
+    catch (e) {
+      append({ level: 'error', text: `[Restart failed] ${(e as Error).message ?? e}`, ts: Date.now() })
+      throw e
+    }
+  }
+
+  async function copyLogs() {
+    const text = logsRef.current
+      .map(l => `${new Date(l.ts).toISOString().slice(11, 19)} ${l.level === 'error' ? 'E' : 'I'} ${l.text}`)
+      .join('\n')
+    try {
+      await navigator.clipboard.writeText(text)
+      return true
+    }
+    catch {
+      return false
+    }
+  }
+
+  return { status, exitCode, logs, restart, copyLogs, isTauri }
+}
+
+// Rust 早期版本 emit 时把 stdout 包了一层 '...'，新版本已经直接 emit 原文。
+// 这里做兼容：去掉外层单引号（如果有的话）。
+function stripQuotes(s: string): string {
+  if (typeof s !== 'string') return String(s)
+  if (s.length >= 2 && s.startsWith("'") && s.endsWith("'"))
+    return s.slice(1, -1)
+  return s
+}

BillNote_frontend/src/components/SystemDiagnostic/StartupBanner.tsx

@@ -0,0 +1,123 @@
+import { useEffect, useState } from 'react'
+
+// 桌面端启动诊断横幅。监听 Tauri 侧 emit 的 backend-warning / backend-error / backend-terminated。
+// 只在 Tauri 环境生效；纯 web 环境（无 window.__TAURI_INTERNALS__）下静默不挂载。
+
+type Severity = 'info' | 'warning' | 'error'
+
+interface DiagnosticPayload {
+  exe_path?: string
+  path_has_non_ascii?: boolean
+  path_has_space?: boolean
+  parent_writable?: boolean
+  platform?: string
+}
+
+interface BannerState {
+  severity: Severity
+  title: string
+  detail: string
+  payload?: DiagnosticPayload
+  dismissible: boolean
+}
+
+const isTauri = typeof window !== 'undefined' && '__TAURI_INTERNALS__' in window
+
+function describeWarning(payload: DiagnosticPayload): { title: string; detail: string } {
+  const parts: string[] = []
+  if (payload.path_has_non_ascii) {
+    parts.push('安装路径包含非 ASCII 字符（中文 / 日文等）')
+  }
+  if (payload.path_has_space) {
+    parts.push('安装路径包含空格')
+  }
+  if (payload.parent_writable === false) {
+    parts.push('安装目录不可写（缺少权限或只读）')
+  }
+  return {
+    title: '检测到可能导致后端启动失败的安装路径',
+    detail:
+      `${parts.join('；')}。\n` +
+      '建议把 BiliNote 重新安装到一个纯英文、无空格、可写的路径下（如 C:\\BiliNote\\ 或 /Applications/）。\n' +
+      `当前路径：${payload.exe_path || '未知'}`,
+  }
+}
+
+const StartupBanner = () => {
+  const [banner, setBanner] = useState<BannerState | null>(null)
+
+  useEffect(() => {
+    if (!isTauri) return
+
+    let unlisteners: Array<() => void> = []
+
+    ;(async () => {
+      const { listen } = await import('@tauri-apps/api/event')
+
+      const offWarning = await listen<DiagnosticPayload>('backend-warning', event => {
+        const { title, detail } = describeWarning(event.payload || {})
+        setBanner({
+          severity: 'warning',
+          title,
+          detail,
+          payload: event.payload,
+          dismissible: true,
+        })
+      })
+
+      const offTerminated = await listen<number | null>('backend-terminated', event => {
+        setBanner({
+          severity: 'error',
+          title: '后端进程已退出',
+          detail: `退出码：${event.payload ?? '未知'}。打开「部署监控」或重启应用以恢复。`,
+          dismissible: false,
+        })
+      })
+
+      // backend-error 是 sidecar stderr，量大噪音多，这里不直接展示，留给 P2 的日志面板。
+      unlisteners = [offWarning, offTerminated]
+    })()
+
+    return () => {
+      unlisteners.forEach(fn => fn())
+    }
+  }, [])
+
+  if (!banner) return null
+
+  const colorByLevel: Record<Severity, string> = {
+    info: 'bg-blue-50 border-blue-300 text-blue-900',
+    warning: 'bg-amber-50 border-amber-300 text-amber-900',
+    error: 'bg-red-50 border-red-300 text-red-900',
+  }
+
+  const iconByLevel: Record<Severity, string> = {
+    info: 'ℹ️',
+    warning: '⚠️',
+    error: '✕',
+  }
+
+  return (
+    <div
+      className={`fixed left-0 right-0 top-0 z-[9999] flex items-start gap-3 border-b px-4 py-2 text-sm shadow-sm ${colorByLevel[banner.severity]}`}
+    >
+      <span className="text-lg">{iconByLevel[banner.severity]}</span>
+      <div className="flex-1 min-w-0">
+        <div className="font-medium">{banner.title}</div>
+        <pre className="mt-0.5 whitespace-pre-wrap break-words font-sans text-xs opacity-90">
+          {banner.detail}
+        </pre>
+      </div>
+      {banner.dismissible && (
+        <button
+          className="shrink-0 rounded px-2 py-0.5 text-xs hover:bg-black/10"
+          onClick={() => setBanner(null)}
+        >
+          知道了
+        </button>
+      )}
+    </div>
+  )
+}
+
+export default StartupBanner

BillNote_frontend/src/pages/Onboarding/index.tsx

@@ -0,0 +1,265 @@
+import { useEffect, useState } from 'react'
+import { useNavigate } from 'react-router-dom'
+import { addProvider, addModel, getProviderList, testConnection } from '@/services/model'
+import { getTranscriberConfig, updateTranscriberConfig } from '@/services/transcriber'
+import logo from '@/assets/icon.svg'
+
+// 桌面端首启 4 步引导。完成后写 localStorage('bilinote-onboarded') = '1'，路由守卫不再拦。
+//
+// 1. 后端连通性自检
+// 2. LLM 供应商 + 模型（最简：只引导填一个 OpenAI-兼容供应商 + 一个 model 名）
+// 3. 转写引擎选择（推荐 Groq 在线，避开本地模型下载坑）
+// 4. （可选）Cookie 同步说明（仅当用户关注 B 站等需要登录态的平台时）
+
+const ONBOARD_KEY = 'bilinote-onboarded'
+
+export function isOnboarded(): boolean {
+  return localStorage.getItem(ONBOARD_KEY) === '1'
+}
+
+function markOnboarded() {
+  localStorage.setItem(ONBOARD_KEY, '1')
+}
+
+const Onboarding = () => {
+  const navigate = useNavigate()
+  const [step, setStep] = useState(1)
+  const [error, setError] = useState('')
+
+  // step 1
+  const [pinging, setPinging] = useState(false)
+  const [backendOk, setBackendOk] = useState<boolean | null>(null)
+
+  // step 2
+  const [providerName, setProviderName] = useState('OpenAI')
+  const [apiKey, setApiKey] = useState('')
+  const [baseUrl, setBaseUrl] = useState('https://api.openai.com/v1')
+  const [modelName, setModelName] = useState('gpt-4o-mini')
+  const [providerId, setProviderId] = useState<string | null>(null)
+  const [savingProvider, setSavingProvider] = useState(false)
+
+  // step 3
+  const [transcriberType, setTranscriberType] = useState<string>('groq')
+  const [savingTranscriber, setSavingTranscriber] = useState(false)
+
+  function next() {
+    setError('')
+    setStep(s => s + 1)
+  }
+  function prev() {
+    setError('')
+    setStep(s => Math.max(1, s - 1))
+  }
+
+  // step 1: ping 后端
+  useEffect(() => {
+    if (step !== 1) return
+    let cancelled = false
+    ;(async () => {
+      setPinging(true)
+      try {
+        await getProviderList()
+        if (!cancelled) setBackendOk(true)
+      }
+      catch {
+        if (!cancelled) setBackendOk(false)
+      }
+      finally {
+        if (!cancelled) setPinging(false)
+      }
+    })()
+    return () => { cancelled = true }
+  }, [step])
+
+  async function saveProvider() {
+    setError('')
+    if (!apiKey.trim()) { setError('请填 API Key'); return }
+    if (!baseUrl.trim()) { setError('请填 API 地址'); return }
+    if (!providerName.trim()) { setError('请填供应商名'); return }
+    if (!modelName.trim()) { setError('请填模型名'); return }
+    setSavingProvider(true)
+    try {
+      // 复用桌面 web 端的 add_provider；type 必须是 'custom'（backend 强制）
+      const res: any = await addProvider({
+        name: providerName.trim(),
+        api_key: apiKey.trim(),
+        base_url: baseUrl.trim(),
+        type: 'custom',
+        logo: 'custom',
+      })
+      const newId = (res?.data ?? res) as string | undefined
+      if (!newId) throw new Error('后端未返回 provider id')
+      setProviderId(newId)
+
+      // 加一个默认 model
+      await addModel({ provider_id: newId, model_name: modelName.trim() })
+
+      // 测试连通
+      try { await testConnection({ id: newId }) }
+      catch (e: any) {
+        // 测试失败不阻断流程，让用户自己决定继续
+        console.warn('测试连接失败：', e?.message ?? e)
+      }
+      next()
+    }
+    catch (e: any) {
+      setError(`保存失败：${e?.message ?? e}`)
+    }
+    finally {
+      setSavingProvider(false)
+    }
+  }
+
+  async function saveTranscriber() {
+    setError('')
+    setSavingTranscriber(true)
+    try {
+      // fast-whisper / mlx-whisper 需指定 model size；在线 (groq/bcut/kuaishou) 不用
+      const needsSize = transcriberType === 'fast-whisper' || transcriberType === 'mlx-whisper'
+      await updateTranscriberConfig({
+        transcriber_type: transcriberType,
+        ...(needsSize ? { whisper_model_size: 'tiny' } : {}),
+      } as any)
+      next()
+    }
+    catch (e: any) {
+      setError(`保存失败：${e?.message ?? e}`)
+    }
+    finally {
+      setSavingTranscriber(false)
+    }
+  }
+
+  function finish() {
+    markOnboarded()
+    navigate('/', { replace: true })
+  }
+
+  return (
+    <div className="min-h-screen flex items-center justify-center bg-gradient-to-br from-blue-50 to-pink-50 p-6">
+      <div className="w-full max-w-xl rounded-xl border bg-white p-6 shadow-lg">
+        <div className="flex items-center gap-3 mb-4">
+          <img src={logo} alt="logo" className="h-10 w-10" />
+          <div>
+            <h1 className="text-xl font-bold">欢迎使用 BiliNote</h1>
+            <p className="text-xs text-gray-500">几步配置后就可以开始把视频转笔记。</p>
+          </div>
+        </div>
+
+        {/* Stepper */}
+        <div className="mb-5 flex items-center gap-2 text-xs text-gray-500">
+          {[1, 2, 3, 4].map(s => (
+            <div key={s} className="flex items-center gap-2">
+              <div
+                className={`flex h-6 w-6 items-center justify-center rounded-full border ${step >= s ? 'border-blue-600 bg-blue-600 text-white' : 'border-gray-300 bg-white text-gray-400'}`}
+              >{s}</div>
+              {s < 4 && <div className={`h-px w-8 ${step > s ? 'bg-blue-600' : 'bg-gray-300'}`} />}
+            </div>
+          ))}
+        </div>
+
+        {step === 1 && (
+          <section className="flex flex-col gap-3">
+            <h2 className="font-semibold">第 1 步 · 后端连通性</h2>
+            <p className="text-sm text-gray-600">桌面端会自动启动 Python 后端进程。检查连通中…</p>
+            {pinging && <div className="text-sm text-gray-500">检测中…</div>}
+            {backendOk === true && <div className="rounded bg-green-50 p-2 text-sm text-green-700">✓ 后端已就绪</div>}
+            {backendOk === false && (
+              <div className="rounded bg-red-50 p-2 text-sm text-red-700">
+                ✗ 暂时连不上后端。可能正在初始化（首次启动会下载依赖），等 1-2 分钟再试。
+                右下角的「后端」状态点会持续监控。
+              </div>
+            )}
+            <div className="flex gap-2 justify-end">
+              <button className="px-4 py-1.5 text-sm rounded bg-blue-600 text-white hover:bg-blue-700 disabled:opacity-50" disabled={!backendOk} onClick={next}>
+                下一步
+              </button>
+            </div>
+          </section>
+        )}
+
+        {step === 2 && (
+          <section className="flex flex-col gap-3">
+            <h2 className="font-semibold">第 2 步 · 模型供应商</h2>
+            <p className="text-sm text-gray-600">填一个 OpenAI 兼容供应商：DeepSeek / Qwen / Claude / 自托管 / OpenAI 都行。</p>
+            <label className="flex flex-col gap-1 text-sm">
+              <span className="text-gray-600">供应商名（自取）</span>
+              <input className="input border rounded px-2 py-1" value={providerName} onChange={e => setProviderName(e.target.value)} />
+            </label>
+            <label className="flex flex-col gap-1 text-sm">
+              <span className="text-gray-600">API 地址</span>
+              <input className="input border rounded px-2 py-1" value={baseUrl} onChange={e => setBaseUrl(e.target.value)} />
+            </label>
+            <label className="flex flex-col gap-1 text-sm">
+              <span className="text-gray-600">API Key</span>
+              <input type="password" className="input border rounded px-2 py-1" value={apiKey} onChange={e => setApiKey(e.target.value)} />
+            </label>
+            <label className="flex flex-col gap-1 text-sm">
+              <span className="text-gray-600">模型名（如 gpt-4o-mini / deepseek-chat / qwen-turbo）</span>
+              <input className="input border rounded px-2 py-1" value={modelName} onChange={e => setModelName(e.target.value)} />
+            </label>
+            {error && <div className="text-xs text-red-600">{error}</div>}
+            <div className="flex gap-2 justify-between">
+              <button className="text-sm text-gray-500 hover:text-gray-800" onClick={prev}>上一步</button>
+              <button className="px-4 py-1.5 text-sm rounded bg-blue-600 text-white hover:bg-blue-700 disabled:opacity-50" disabled={savingProvider} onClick={saveProvider}>
+                {savingProvider ? '保存中…' : '保存并下一步'}
+              </button>
+            </div>
+          </section>
+        )}
+
+        {step === 3 && (
+          <section className="flex flex-col gap-3">
+            <h2 className="font-semibold">第 3 步 · 音频转写引擎</h2>
+            <p className="text-sm text-gray-600">把视频音频转成文字。<strong>推荐在线引擎</strong>，避免本地下载 ~600MB 的模型。</p>
+            <div className="grid gap-2">
+              {[
+                { value: 'groq', title: 'Groq（在线，推荐）', desc: '注册 https://groq.com/ 拿免费 key；速度快、英文语料佳。无需本地模型。' },
+                { value: 'bcut', title: '必剪（在线，免登）', desc: '免登，中文表现好；偶尔限流。' },
+                { value: 'kuaishou', title: '快手（在线，免登）', desc: '与必剪类似，备选。' },
+                { value: 'fast-whisper', title: 'Faster Whisper（本地）', desc: '完全离线但首次需下载 ~75MB（tiny）至 ~3GB（large-v3）的模型。CPU 慢。' },
+              ].map(opt => (
+                <label key={opt.value} className={`flex gap-3 p-3 rounded border cursor-pointer ${transcriberType === opt.value ? 'border-blue-600 bg-blue-50' : 'border-gray-200 hover:border-gray-300'}`}>
+                  <input type="radio" name="transcriber" value={opt.value} checked={transcriberType === opt.value} onChange={e => setTranscriberType(e.target.value)} />
+                  <div>
+                    <div className="text-sm font-medium">{opt.title}</div>
+                    <div className="text-xs text-gray-500 mt-0.5">{opt.desc}</div>
+                  </div>
+                </label>
+              ))}
+            </div>
+            {error && <div className="text-xs text-red-600">{error}</div>}
+            <div className="flex gap-2 justify-between">
+              <button className="text-sm text-gray-500 hover:text-gray-800" onClick={prev}>上一步</button>
+              <button className="px-4 py-1.5 text-sm rounded bg-blue-600 text-white hover:bg-blue-700 disabled:opacity-50" disabled={savingTranscriber} onClick={saveTranscriber}>
+                {savingTranscriber ? '保存中…' : '保存并下一步'}
+              </button>
+            </div>
+          </section>
+        )}
+
+        {step === 4 && (
+          <section className="flex flex-col gap-3">
+            <h2 className="font-semibold">第 4 步 · Cookie 同步（可选）</h2>
+            <p className="text-sm text-gray-600">
+              想总结 <strong>B 站 / 抖音 / 快手</strong> 等需要登录态的平台时，需要把浏览器 cookie 复制到「下载配置」页。
+              <br />
+              YouTube 一般不需要 cookie。先跳过也没问题，到时再去配。
+            </p>
+            <div className="rounded bg-gray-50 p-3 text-xs text-gray-600">
+              提示：插件版（<a className="text-blue-600 underline" href="https://github.com/JefferyHcool/BiliNote/tree/develop/BillNote_extension" target="_blank" rel="noreferrer">BillNote_extension</a>）支持一键 cookie 同步；桌面版需手动复制。
+            </div>
+            <div className="flex gap-2 justify-between">
+              <button className="text-sm text-gray-500 hover:text-gray-800" onClick={prev}>上一步</button>
+              <button className="px-4 py-1.5 text-sm rounded bg-blue-600 text-white hover:bg-blue-700" onClick={finish}>
+                完成，进入 BiliNote
+              </button>
+            </div>
+          </section>
+        )}
+      </div>
+    </div>
+  )
+}
+
+export default Onboarding

BillNote_frontend/src/pages/SettingPage/transcriber.tsx

@@ -73,6 +73,28 @@ export default function Transcriber() {
   }, [modelStatuses, mlxModelStatuses, fetchModelsStatus])
 
   const handleSave = async () => {
+    // 切到本地 whisper 引擎且选了未下载的模型时，提前 confirm，避免用户保存后到首次任务才发现要下 GB 级模型
+    if (isWhisperType(selectedType)) {
+      const pool = selectedType === 'mlx-whisper' ? mlxModelStatuses : modelStatuses
+      const target = pool.find(m => m.model_size === selectedModelSize)
+      if (target && !target.downloaded && !target.downloading) {
+        const sizeHint: Record<string, string> = {
+          'tiny': '~75MB',
+          'base': '~150MB',
+          'small': '~500MB',
+          'medium': '~1.5GB',
+          'large-v3': '~3GB',
+          'large-v3-turbo': '~1.6GB',
+        }
+        const ok = window.confirm(
+          `选择 ${selectedType} / ${selectedModelSize} 后，首次转写时会下载该模型（${sizeHint[selectedModelSize] || '体积未知'}）。\n` +
+          `网络较差时容易中断；推荐改用 Groq / 必剪 / 快手 等在线引擎。\n\n` +
+          '继续保存吗？',
+        )
+        if (!ok) return
+      }
+    }
+
     setSaving(true)
     try {
       const payload: { transcriber_type: string; whisper_model_size?: string } = {

CHANGELOG.md

@@ -2,6 +2,79 @@
 
 本项目所有重要变更记录于此。格式参考 [Keep a Changelog](https://keepachangelog.com/zh-CN/1.1.0/)，遵循 [语义化版本](https://semver.org/lang/zh-CN/)。
 
+## [2.2.1] - 2026-05-09
+
+补 v2.2.0 ghcr.io 镜像构建失败。
+
+### Fixed
+
+- Docker 镜像构建失败：`v2.2.0` tag 触发的 ghcr.io 推送在 frontend-builder 第 5/7 步 `pnpm install --frozen-lockfile` 报 `ERR_UNKNOWN_BUILTIN_MODULE`。根因：`corepack prepare pnpm@latest` 拉到了 pnpm 11.0.9，而 pnpm 11 要求 Node 22+，跟我们的 `node:20-alpine` 不兼容。
+  - `Dockerfile.complete` 与 `BillNote_frontend/Dockerfile` 的 pnpm 版本 pin 到 `9.15.0`（lockfile 由 pnpm 9 生成，匹配 Node 20）
+
+## [2.2.0] - 2026-05-09
+
+主线：浏览器插件功能与 web 端 NoteForm 完整对齐；桌面客户端 UX 与错误恢复一波重炼。
+
+### Added — 浏览器插件
+
+- 笔记选项与 web 端 NoteForm 完整对齐：
+  - `style` 由自由文本改成 9 个预设下拉（minimal / detailed / academic / tutorial / xiaohongshu / life_journal / task_oriented / business / meeting_minutes），与 backend `prompt_builder.note_styles` 严格匹配（之前自由文本不命中 enum 等于没传——隐性 bug）
+  - `format` 完整 4 个 checkbox（toc / link / screenshot / summary，原来只有 screenshot/link）
+  - `extras` 文本框：拼接到 prompt 末尾的 ad-hoc 提示
+- 多模态视频理解：`video_understanding` 开关 + `video_interval`（1-30 秒）+ `grid_size`（[r,c]，1-10），抽帧拼图喂视觉模型，提示需选视觉模型才生效
+
+### Added — 桌面客户端
+
+- **首启 4 步引导**（`/onboarding`）：后端连通性自检 → LLM 供应商 + 模型 → 转写引擎选择（默认推荐 Groq）→ Cookie 同步说明。完成后 `localStorage('bilinote-onboarded')` 标记，纯 web 端不打扰
+- **Sidecar 健康度面板**：右下角浮动状态点（绿/黄/红，5s 轮询 `/sys_health`），点开抽屉看最近 200 行后端日志、一键重启后端（新增 Tauri command `restart_backend_sidecar`）、复制日志
+- **启动期路径诊断**：Tauri `setup` 中检测安装路径含非 ASCII / 含空格 / 父目录不可写时，emit `backend-warning` 让前端顶端横幅显式告警，主动暴露 README 长期文字警告但无防御的"中文路径"等坑
+
+### Changed
+
+- Whisper 默认模型 size 从 `medium`（~1.5GB）改为 `tiny`（~75MB）：新装用户没主动设置时不再卡在首次大模型下载；高精度可在「音频转写配置」页主动切
+- 切到 `fast-whisper` / `mlx-whisper` 且当前 size 未下载时，「音频转写配置」页保存前 confirm 体积提示，并推荐改用在线引擎
+- Tauri sidecar 启动逻辑抽出 `spawn_backend_sidecar()`；child handle 存进 `SidecarHandle` state 以支持后续 restart
+- sidecar stdout/stderr emit 时不再用 `format!("'{}'", ...)` 包引号，原文直传（前端 hook 兼容旧格式兜底剥引号）
+
+### Fixed
+
+- WhisperTranscriber 在半成品模型目录上死循环报 `Unable to open file 'model.bin'`：判定从「目录存在」改为「`model.bin` 落盘」，半成品目录会被识别并重新下载（PR `fix/backend-deploy-resilience`）
+- `/api/deploy_status` 在没装 torch 的部署上 `ModuleNotFoundError: No module named 'torch'` 500：torch 改 try/except，未装时返回 `{available: false, torch_installed: false}`；transcriber 配置 + ffmpeg 也都裹 try，单项失败不再打死整个监控页（同上 PR）
+- `routers/config._check_whisper_model_exists` 同步改用 `model.bin` 判定，避免「已下载」状态在监控页误报
+
+## [2.1.4] - 2026-05-07
+
+CI 工程化修复，无运行时行为变化。
+
+### Internal
+
+- 桌面端 Tauri 构建矩阵去掉 Linux（`ubuntu-22.04 / x86_64-unknown-linux-gnu`）。Linux 桌面端构建持续 17m+，且无对应分发渠道；Linux 用户继续可以走 Docker 镜像 (`ghcr.io/jefferyhcool/bilinote`)
+- commitlint workflow 去掉无效的 `firstParent` input（wagoid/commitlint-github-action@v6 不支持，被忽略并打 warn）
+- 规范 release merge commit 标题：`chore(release): vX.Y.Z`（合 master）/ `chore(release): merge release/X.Y.Z back into develop`（回灌 develop），让 commitlint 能正确识别。`RELEASING.md` §3 与 `CONTRIBUTING.md` §6.3 同步更新
+
+## [2.1.3] - 2026-05-07
+
+### Fixed
+
+- DeepSeek 等非多模态供应商被 400 拒绝（issue #282）：`UniversalGPT.create_messages` 与 `_build_merge_messages` 此前**无条件**把 content 拼成 OpenAI 多模态数组 `[{"type":"text",...}]`，DeepSeek `deepseek-chat` 等模型不识别 `image_url` 变体直接报 `invalid_request_error`。`GPTFactory.from_config` 一律实例化 `UniversalGPT`，所以问题覆盖**所有**通过模型设置页接入的非多模态供应商，不止 DeepSeek。
+  - 现按 `video_img_urls` 是否非空切换 content 形态：有图保留多模态数组（视觉模型不退化），无图退回 string。合并阶段历来不带图，统一改 string。
+  - 与同包内 `deepseek_gpt.py` / `openai_gpt.py` / `qwen_gpt.py` 的 message builder 行为对齐。
+  - 新增 `backend/tests/test_universal_gpt_content_format.py` 6 个 case 回归覆盖（含 `image_url` 字面 not-in JSON 断言）。
+
+感谢 @voidborne-d 的修复（#345）。
+
+## [2.1.2] - 2026-05-07
+
+补 v2.1.1 上 ghcr.io 镜像构建失败的坑。
+
+### Fixed
+
+- Docker 镜像构建失败：v2.1.1 tag 触发的 ghcr.io 推送在 frontend-builder 第 7/7 步 `pnpm run build` 挂掉（vite `loadConfigFromBundledFile` 加载 `@tailwindcss/vite` plugin 时 1.5s 内异常退出）。
+  - `Dockerfile.complete` 与 `BillNote_frontend/Dockerfile` 升 `node:18-alpine` → `node:20-alpine`：Tailwind v4 已不再支持 Node 18，Vite 6 也推荐 Node 20+
+  - `Dockerfile.complete` 的 frontend 阶段同时复制 `pnpm-lock.yaml` 并改用 `--frozen-lockfile`，杜绝每次构建重解析 semver 拉到比本地新的 native dep
+  - `BillNote_frontend/pnpm-lock.yaml` 强制入库（之前一直未提交，导致 CI / 本地依赖图持续漂移）
+- README 联系社区段补上微信群二维码（之前只写"年会恢复更新以后放出最新社区地址"）
+
 ## [2.1.1] - 2026-05-07
 
 工程化与文档收尾，无运行时行为变化。

CONTRIBUTING.md

@@ -250,6 +250,7 @@ chore(ci): 优化 docker 构建缓存
 
 - `feature/*` / `fix/*` 合入 `develop`：推荐 **Squash and merge**，保持 develop 历史线性。
 - `release/*` 合入 `master` 与回灌 `develop`：使用 **Merge commit (--no-ff)**，保留发版结构。
+  · merge commit 标题用 `chore(release): vX.Y.Z`（合 master）或 `chore(release): merge release/X.Y.Z back into develop`（回灌 develop），保证 commitlint 通过。
 - `hotfix/*` 同上 release。
 
 ### 6.4 合并后

Dockerfile.complete

@@ -26,15 +26,22 @@ RUN pip install --no-cache-dir -i ${PIP_INDEX} -r requirements.txt
 COPY ./backend /tmp/backend
 
 # === 阶段2：构建 Frontend ===
-FROM node:18-alpine AS frontend-builder
+# Node 18-alpine 跑不动 Tailwind v4 / Vite 6（前者要求 Node 20+，后者推荐 Node 20+），
+# 升到 node:20-alpine。alpine 走 musl，pnpm 会按 lockfile 拉 *-linux-x64-musl native binary。
+FROM node:20-alpine AS frontend-builder
 
-RUN corepack enable && corepack prepare pnpm@latest --activate
+# pnpm 版本 pin 到 9 系列：
+# - lockfile (BillNote_frontend/pnpm-lock.yaml) 是 lockfileVersion '9.0'，由 pnpm 9 生成
+# - pnpm 11+ 要求 Node 22+，与 node:20 不兼容（ERR_UNKNOWN_BUILTIN_MODULE）
+# - 不用 @latest 避免上游 pnpm 升级悄悄破坏 CI
+RUN corepack enable && corepack prepare pnpm@9.15.0 --activate
 
 WORKDIR /tmp/frontend
 
-# 先复制 package.json 利用依赖层缓存
-COPY ./BillNote_frontend/package.json ./
-RUN pnpm install
+# 先复制 package.json + lockfile 利用依赖层缓存
+# --frozen-lockfile 保证 CI 与本地开发依赖版本一致，杜绝 semver 漂移引入的破坏性升级
+COPY ./BillNote_frontend/package.json ./BillNote_frontend/pnpm-lock.yaml ./
+RUN pnpm install --frozen-lockfile
 
 COPY ./BillNote_frontend /tmp/frontend
 

README.md

@@ -3,7 +3,7 @@
     <p align="center">
   <img src="./doc/icon.svg" alt="BiliNote Banner" width="50" height="50"  />
 </p>
-<h1 align="center" > BiliNote v2.1.1</h1>
+<h1 align="center" > BiliNote v2.2.1</h1>
 </div>
 
 <p align="center"><i>AI 视频笔记生成工具 让 AI 为你的视频做笔记</i></p>
@@ -53,6 +53,35 @@ BiliNote 是一个开源的 AI 视频笔记助手，支持通过哔哩哔哩、Y
 - 笔记顶部视频封面 Banner 展示
 - 工作区和生成历史面板支持折叠/展开
 
+### v2.2.1 修订
+
+- 修复 v2.2.0 ghcr.io 镜像构建失败（pnpm@latest 拉到 11，与 Node 20 不兼容；pin 到 pnpm 9.15.0）
+
+### v2.2.0 新增
+
+- **浏览器插件**笔记选项与 web 端完整对齐：style 9 个预设下拉、format 4 个 checkbox、extras 文本框、多模态视频理解开关
+- **桌面客户端**首启 4 步引导（连通自检 → 供应商/模型 → 转写引擎 → Cookie 提示）
+- **桌面客户端**右下角后端运行状态指示，点开看日志、一键重启
+- **桌面客户端**启动期主动检测中文 / 空格 / 不可写安装路径，弹横幅告警
+- Whisper 默认 size 从 medium（~1.5GB）改为 tiny（~75MB）；切大模型时显式 confirm
+- 修：whisper 半成品模型目录死循环；`/deploy_status` 在没装 torch 的部署 500
+- 详见 [CHANGELOG.md](./CHANGELOG.md)
+
+### v2.1.4 修订
+
+- CI：桌面端 Tauri 构建去掉 Linux（17m+ 慢线退役；Linux 用户继续走 Docker 镜像）
+- CI：commitlint workflow 修复 + 规范 release merge commit 标题约定
+
+### v2.1.3 修订
+
+- 修复 DeepSeek 等非多模态供应商被 400 拒绝的问题（issue #282）：`UniversalGPT` 的 message builder 按是否带图切换 string / 多模态数组形态
+- 感谢 @voidborne-d (#345)
+
+### v2.1.2 修订
+
+- 修复 v2.1.1 触发的 ghcr.io Docker 镜像构建失败（Node 18 + Tailwind v4 不兼容、缺 lockfile）
+- README 补上微信群二维码
+
 ### v2.1.1 修订
 
 - 工程化与文档收尾：CONTRIBUTING.md / RELEASING.md / issue + PR 模板 / commitlint CI / 插件发版工作流
@@ -205,7 +234,12 @@ docker-compose -f docker-compose.gpu.yml up -d
 - [ ] 笔记导出为 PDF / Word / Notion
 
 ### Contact and Join-联系和加入社区
-年会恢复更新以后放出最新社区地址
+
+扫码加入 BiliNote 交流微信群（如二维码失效，请到 [Issues](https://github.com/JefferyHcool/BiliNote/issues) 反馈）：
+
+<p align="center">
+  <img src="./doc/wechat.png" alt="BiliNote 交流微信群" width="240" />
+</p>
 
 
 

RELEASING.md

@@ -42,10 +42,13 @@ git push -u origin release/X.Y.Z
 
 在 GitHub 上发起两个 PR：
 
-| PR | base | 合并方式 |
-|---|---|---|
-| `release/X.Y.Z` → `master` | `master` | **Merge commit (--no-ff)** |
-| `release/X.Y.Z` → `develop` | `develop` | **Merge commit (--no-ff)** |
+| PR | base | 合并方式 | 合并后 commit 标题 |
+|---|---|---|---|
+| `release/X.Y.Z` → `master` | `master` | **Merge commit (--no-ff)** | `chore(release): vX.Y.Z` |
+| `release/X.Y.Z` → `develop` | `develop` | **Merge commit (--no-ff)** | `chore(release): merge release/X.Y.Z back into develop` |
+
+> ⚠️ Merge commit 的标题**必须**符合 `type(scope): subject` 格式（commitlint 在 push 到 master/develop 时会校验）。
+> 历史上用过 `Release vX.Y.Z` 这种形式，会被 commitlint 报 `type-empty` / `subject-empty`。
 
 `master` 分支保护要求 review 通过。回灌 `develop` 是为了把发版冻结期内的小修同步回来。
 

backend/app/gpt/universal_gpt.py

@@ -53,20 +53,26 @@ class UniversalGPT(GPT):
             extras=kwargs.get('extras'),
         )
 
-        # ⛳ 组装 content 数组，支持 text + image_url 混合
-        content: List[dict] = [{"type": "text", "text": content_text}]
         video_img_urls = kwargs.get('video_img_urls', [])
 
-        for url in video_img_urls:
-            content.append({
-                "type": "image_url",
-                "image_url": {
-                    "url": url,
-                    "detail": "auto"
-                }
-            })
+        content: list[dict] | str
+        if video_img_urls:
+            # 有截图时走 OpenAI 多模态 content 数组（text + image_url）
+            content = [{"type": "text", "text": content_text}]
+            for url in video_img_urls:
+                content.append({
+                    "type": "image_url",
+                    "image_url": {
+                        "url": url,
+                        "detail": "auto"
+                    }
+                })
+        else:
+            # 纯文本场景退回 string content：DeepSeek deepseek-chat 等非多模态模型
+            # 不识别 [{"type":"text",...}] 数组形态，会返回 invalid_request_error
+            # （issue #282）。OpenAI 规范本身也允许 content 为 string。
+            content = content_text
 
-        #  正确格式：整体包在一个 message 里，role + content array
         messages = [{
             "role": "user",
             "content": content
@@ -83,9 +89,10 @@ class UniversalGPT(GPT):
 
     def _build_merge_messages(self, partials: list) -> list:
         merge_text = MERGE_PROMPT + "\n\n" + "\n\n---\n\n".join(partials)
+        # 合并阶段没有图片，直接用 string content 兼容非多模态模型（issue #282）
         return [{
             "role": "user",
-            "content": [{"type": "text", "text": merge_text}]
+            "content": merge_text
         }]
 
     def _checkpoint_path(self, checkpoint_key: str) -> Path:

backend/app/services/transcriber_config_manager.py

@@ -25,7 +25,12 @@ class TranscriberConfigManager:
             json.dump(data, f, ensure_ascii=False, indent=2)
 
     def get_config(self) -> Dict[str, Any]:
-        """获取当前转写器配置，fallback 到环境变量默认值。"""
+        """获取当前转写器配置，fallback 到环境变量默认值。
+
+        whisper 默认 size 从 'medium' (~1.5GB) 改为 'tiny' (~75MB)：
+        新装用户没主动设置时不应该被首次下载卡住。想要更高精度可在「音频转写配置」
+        页主动切换。
+        """
         data = self._read()
         return {
             "transcriber_type": data.get(
@@ -34,7 +39,7 @@ class TranscriberConfigManager:
             ),
             "whisper_model_size": data.get(
                 "whisper_model_size",
-                os.getenv("WHISPER_MODEL_SIZE", "medium"),
+                os.getenv("WHISPER_MODEL_SIZE", "tiny"),
             ),
         }
 

backend/tests/test_universal_gpt_content_format.py

@@ -0,0 +1,189 @@
+"""issue #282 回归测试：UniversalGPT 拼装 content 时按是否有图片切换 string / array 形态。
+
+DeepSeek deepseek-chat 等非多模态模型只接受 ``content`` 为字符串，旧实现无条件
+emit ``[{"type":"text","text":...}]`` 导致 ``invalid_request_error``。
+"""
+import importlib.util
+import pathlib
+import sys
+import types
+import unittest
+
+
+def _install_stubs():
+    app_mod = types.ModuleType("app")
+    gpt_pkg = types.ModuleType("app.gpt")
+    models_pkg = types.ModuleType("app.models")
+
+    base_mod = types.ModuleType("app.gpt.base")
+
+    class _GPT:
+        pass
+
+    base_mod.GPT = _GPT
+
+    prompt_builder_mod = types.ModuleType("app.gpt.prompt_builder")
+
+    def _generate_base_prompt(**_kwargs):
+        return "PROMPT_BODY"
+
+    prompt_builder_mod.generate_base_prompt = _generate_base_prompt
+
+    prompt_mod = types.ModuleType("app.gpt.prompt")
+    prompt_mod.BASE_PROMPT = ""
+    prompt_mod.AI_SUM = ""
+    prompt_mod.SCREENSHOT = ""
+    prompt_mod.LINK = ""
+    prompt_mod.MERGE_PROMPT = "MERGE_HEAD"
+
+    utils_mod = types.ModuleType("app.gpt.utils")
+
+    def _fix_markdown(text):
+        return text
+
+    utils_mod.fix_markdown = _fix_markdown
+
+    request_chunker_mod = types.ModuleType("app.gpt.request_chunker")
+
+    class _RequestChunker:
+        def __init__(self, *_args, **_kwargs):
+            pass
+
+        def group_texts_by_budget(self, texts, _builder, **_kwargs):
+            return [texts]
+
+    request_chunker_mod.RequestChunker = _RequestChunker
+
+    gpt_model_mod = types.ModuleType("app.models.gpt_model")
+
+    class _GPTSource:
+        pass
+
+    gpt_model_mod.GPTSource = _GPTSource
+
+    transcriber_model_mod = types.ModuleType("app.models.transcriber_model")
+
+    class _TranscriptSegment:
+        def __init__(self, **kwargs):
+            self.start = kwargs.get("start", 0)
+            self.end = kwargs.get("end", 0)
+            self.text = kwargs.get("text", "")
+
+    transcriber_model_mod.TranscriptSegment = _TranscriptSegment
+
+    sys.modules.setdefault("app", app_mod)
+    sys.modules.setdefault("app.gpt", gpt_pkg)
+    sys.modules.setdefault("app.models", models_pkg)
+    sys.modules["app.gpt.base"] = base_mod
+    sys.modules["app.gpt.prompt_builder"] = prompt_builder_mod
+    sys.modules["app.gpt.prompt"] = prompt_mod
+    sys.modules["app.gpt.utils"] = utils_mod
+    sys.modules["app.gpt.request_chunker"] = request_chunker_mod
+    sys.modules["app.models.gpt_model"] = gpt_model_mod
+    sys.modules["app.models.transcriber_model"] = transcriber_model_mod
+
+
+def _load_universal_gpt_class():
+    _install_stubs()
+    root = pathlib.Path(__file__).resolve().parents[1]
+    module_path = root / "app" / "gpt" / "universal_gpt.py"
+    spec = importlib.util.spec_from_file_location(
+        "universal_gpt_content_format", module_path
+    )
+    if spec is None or spec.loader is None:
+        raise ImportError("universal_gpt module spec not found")
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module.UniversalGPT
+
+
+UniversalGPT = _load_universal_gpt_class()
+
+
+class _DummyClient:
+    """create_messages 不会真的调用 client，给个空壳即可。"""
+
+
+def _make_gpt():
+    return UniversalGPT(_DummyClient(), model="deepseek-chat")
+
+
+class TestCreateMessagesContentFormat(unittest.TestCase):
+    """覆盖 create_messages 在不同 video_img_urls 输入下的输出形态。"""
+
+    def test_no_images_emits_string_content(self):
+        """无图片时 content 为 str（DeepSeek / 非多模态模型可解析）。"""
+        gpt = _make_gpt()
+
+        messages = gpt.create_messages(segments=[])
+
+        self.assertEqual(len(messages), 1)
+        self.assertEqual(messages[0]["role"], "user")
+        self.assertIsInstance(messages[0]["content"], str)
+        self.assertEqual(messages[0]["content"], "PROMPT_BODY")
+
+    def test_empty_image_list_emits_string_content(self):
+        """显式传入空列表也要走纯文本分支，避免图片字段误触发。"""
+        gpt = _make_gpt()
+
+        messages = gpt.create_messages(segments=[], video_img_urls=[])
+
+        self.assertIsInstance(messages[0]["content"], str)
+
+    def test_with_images_emits_multimodal_array(self):
+        """有图片时保留多模态 array 形态，确保多模态模型功能不退化。"""
+        gpt = _make_gpt()
+
+        messages = gpt.create_messages(
+            segments=[],
+            video_img_urls=["https://example.com/a.jpg", "https://example.com/b.jpg"],
+        )
+
+        content = messages[0]["content"]
+        self.assertIsInstance(content, list)
+        self.assertEqual(len(content), 3)  # 1 text + 2 images
+        self.assertEqual(content[0], {"type": "text", "text": "PROMPT_BODY"})
+        self.assertEqual(content[1]["type"], "image_url")
+        self.assertEqual(content[1]["image_url"]["url"], "https://example.com/a.jpg")
+        self.assertEqual(content[1]["image_url"]["detail"], "auto")
+        self.assertEqual(content[2]["image_url"]["url"], "https://example.com/b.jpg")
+
+    def test_no_image_url_field_when_no_images(self):
+        """纯文本响应里不应该出现 image_url 关键字 —— 这是触发 DeepSeek 400 的根因。"""
+        gpt = _make_gpt()
+
+        messages = gpt.create_messages(segments=[])
+
+        import json
+        serialized = json.dumps(messages, ensure_ascii=False)
+        self.assertNotIn("image_url", serialized)
+
+
+class TestBuildMergeMessagesContentFormat(unittest.TestCase):
+    """合并阶段从不带图片，应该统一走 string content 路径。"""
+
+    def test_merge_messages_use_string_content(self):
+        """否则长视频 chunk 后的合并阶段还会复现 issue #282 错误。"""
+        gpt = _make_gpt()
+
+        messages = gpt._build_merge_messages(["partial-A", "partial-B"])
+
+        self.assertEqual(len(messages), 1)
+        self.assertEqual(messages[0]["role"], "user")
+        self.assertIsInstance(messages[0]["content"], str)
+        self.assertIn("MERGE_HEAD", messages[0]["content"])
+        self.assertIn("partial-A", messages[0]["content"])
+        self.assertIn("partial-B", messages[0]["content"])
+
+    def test_merge_messages_no_image_url_field(self):
+        gpt = _make_gpt()
+
+        messages = gpt._build_merge_messages(["x"])
+
+        import json
+        serialized = json.dumps(messages, ensure_ascii=False)
+        self.assertNotIn("image_url", serialized)
+
+
+if __name__ == "__main__":
+    unittest.main()