Merge pull request #353 from JefferyHcool/feature/extension-video-understanding

feat(extension): 多模态视频理解开关 + 抽帧/拼图参数（对齐 web NoteForm）

.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

@@ -1,5 +1,6 @@
 # === 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
 

CHANGELOG.md

@@ -2,6 +2,39 @@
 
 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/)。
 
+## [2.1.4] - 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 +26,18 @@ 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
 
 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 v2.1.4</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 +53,21 @@ 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
 
+### v2.1.4 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 +220,12 @@ 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/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()