Move the consumer side of the OpenBiliClaw local recommendation Agent into the DSH Web interface. Left sidebar buttons trigger drawer panels, and register 22 openbiliclaw_* tools with the Agent to complete the recommendation, feedback, and portrait loop.
- Language
- JavaScript
- License
- BSD-3-Clause
- Branch
- main
Install
$ dsh plugin --profile web add github:whiteguo233/dsh-openbiliclawRun the command above in your terminal to install this plugin via the dsh CLI. You can switch Profile in the top-right corner. New to dsh? Read the beginner tutorial
Install via your agent
Install the DeepSeek Harness plugin whiteguo233/dsh-openbiliclaw for me: review the repository at https://github.com/whiteguo233/dsh-openbiliclaw first, then run the install command and verify the plugin loads successfully.
Paste this instruction to the DSH Web GUI assistant — it will install and verify for you.
One-Line Positioning
Integrate the locally running OpenBiliClaw personalized recommendation Agent into DSH's Web GUI—click the left sidebar to open a right drawer showing recommendations/favorites/dialogue/profile/settings, while registering 22 tools with the DSH Agent, enabling the Agent to read recommendations, respond to probes, save content, forming a closed loop of "recommendation → feedback → profile → more accurate recommendations".
Core Capabilities
- Append OpenBiliClaw button at the bottom of DSH's left sidebar; clicking slides out a drawer panel from the right (recommendations/content library/dialogue/profile/settings/notifications)
- Register 22
openbiliclaw_*tools with the DSH Agent, covering recommendation fetching, appending, refresh, feedback submission, delight cards, dialogue, profile, save/remove/list, runtime status, etc. - Auto-load
openbiliclaw-adapterskill (parses SKILL.md frontmatter in the repository), enabling the Agent to invoke the bridge as a long-term instruction when needed - Communicate with running OpenBiliClaw backend
serve-apivia HTTP calls to/api/agent-bridgeendpoint (based on agent-bridge/v2 JSON protocol), avoiding forking Python processes each time - Panel experience covers dark mode following, cross-platform cover image proxy (with fallback to text/media placeholder on failure), WebSocket real-time push, message drawer and "sync to platform" progress
Technical Implementation
- Language: TypeScript (Node side ESM) + React 18 (browser side panel)
- Key Dependencies:
@deepseek-ai/cordis ^4.0.1(plugin mounting),@deepseek-ai/dsh-tools ^0.1.0-rc.6(tool registration),@deepseek-ai/dsh-client-runtime / dsh-client-ui-slots ^0.1.0-rc.6(slot/theme injection), nativefetch(HTTP communication with backend) - Architecture Pattern: Simultaneously declare
dsh.bundle.patch(cordis insert Node half) anddsh.client.inject(browser half inject runtime + ui-slots); Node half registers 22 tools and 1 skill viactx.tools.register/ctx.skills.register; Browser half usesPanelLayoutControllerto append a column in the frame grid and rewritegrid-template-columns(push center column rather than overlay), only sidebar button usessidebar.footer.actionslot - Entry Files:
src/index.ts(node half) /src/client/index.ts(browser half) /cordis.patch.yml(cordis injection declaration)
Use Cases
Users who have locally deployed OpenBiliClaw and want to complete consumer-side operations "without leaving DSH": while chatting about requirements in a DSH session, view cross-platform recommendations via the right drawer (Xiaohongshu/Bilibili/Douyin/YouTube/X/Zhihu/Reddit/Linux.do/V2EX/Weibo), handle favorites/watch later, respond to interest/avoidance probes, view profile. The Agent can also use the 22 registered tools to incorporate recommendation context into responses, making "what I watch → what you think → what I recommend next" a complete closed loop within DSH.
Prerequisites & Compatibility
| Dependency | Min Version | Description |
|---|---|---|
| DeepSeek Harness (DSH) | 0.1.0-rc.6+ | Plugin peer ABI aligned with dsh-* packages @deepseek-ai/cordis ^4.0.1; README warns not to mix tool packages from 0.0.1 era in same profile |
| OpenBiliClaw Backend | Enable Agent Bridge v2 | Default listening 127.0.0.1:8420; this plugin is consumer-side only, crawling/source management still handled by main project |
| Node | Not declared | package.json doesn't declare engines field |
| Platform | Cross-platform | Only declares dsh.client.platform: "web", no os/cpu restrictions |
| Native Modules | None | All dependencies are JS/TS packages, no native deps like node-pty / node:sqlite |
Installation
dsh plugin --profile web add github:whiteguo233/dsh-openbiliclaw
Configuration Options
| Config | Type | Description | Default |
|---|---|---|---|
apiUrl | string | Root address of OpenBiliClaw backend serve-api; all bridge tools on Node side go to <apiUrl>/api/agent-bridge. Don't append /api at the end of address. | http://127.0.0.1:8420 |
workdir | string (required) | OpenBiliClaw main project root directory; used to parse config.toml / data/ and read SKILL.md. | /Users/white/workspace/OpenBiliClaw |
skillPath | string | Absolute path to adapter skill file; when not found only logs warn without error, tools still work but skill metadata absent in DSH. | <workdir>/skills/openbiliclaw-adapter/SKILL.md |
timeoutMs | number | Timeout for single bridge HTTP call (ms), applied to all 22 tools. | 300000 (5 minutes) |
"Settings → General → Connection" in the panel can change backend address (saved to localStorage, takes effect immediately); that address is used for browser-side direct
/api/*calls, separate fromapiUrl(Node side) in the table above.
FAQ
Q: Does this plugin conflict with DSH's official settings panel?
A: No conflict. The plugin only adds content: a button at bottom of sidebar, appends the rightmost column in the frame grid (pushing center column, not covering), doesn't modify DSH's native column layout. Panel mount failure only console.error, doesn't affect DSH main flow.
Q: Will the 22 tools be used automatically by Agent, or do I need to mention them in conversation?
A: DSH Agent calls them as needed. openbiliclaw_recommend / openbiliclaw_chat / openbiliclaw_get_profile etc. can be selected as tool calls by LLM when needed; to strictly control scope can also explicitly mention them in conversation.
Q: Is requestId required?
A: Mandatory only for write operation tools (openbiliclaw_submit_feedback / openbiliclaw_respond_delight). Retry of the same action must reuse the same requestId (max 400 characters), don't reuse across different actions—this is the backend's idempotency key, wrong reuse will lose feedback or write wrong entries.
Q: How to see "delight recommendations" (proactive delight)?
A: Two ways: panel receives delight.candidate event via WebSocket push and renders directly at top of drawer; Agent side can call openbiliclaw_get_delight to fetch current pending cards, or call openbiliclaw_respond_delight to respond with view/like/dislike/dismiss/chat (dismiss permanently removes that item from recommendations).
Q: Will interest/avoidance probes pop up proactively for me?
A: Backend pushes interest.probe / avoidance.probe events via WebSocket, panel uses optimistic state machine for immediate feedback; Agent side can also call openbiliclaw_next_probe / openbiliclaw_next_avoidance_probe to fetch next pending question, and respond with openbiliclaw_respond_interest_probe / openbiliclaw_respond_avoidance_probe (confirm/reject/defer/chat, rejecting interest has 30-day cooldown).
Q: Will saved content sync to Bilibili account?
A: openbiliclaw_save_local / openbiliclaw_remove_saved / openbiliclaw_list_saved all write to local SQLite only, won't sync to external accounts. "Sync to platform" button in panel calls main project's sync interface, requires Bilibili account login to succeed; non-Bilibili platform items marked as "local save only".
Getting Started Difficulty
Beginner — fill in workdir correctly, ensure OpenBiliClaw backend has Agent Bridge v2 enabled, other configs have reasonable defaults, out-of-the-box usable with panel and Agent tools.
Known Issues & Limitations
- Configuration items
pythonBin/stdoutMaxByteslisted in README do not exist in source code: current implementation uses HTTP (src/bridge.ts:74-99), doesn't invoke Python subprocess, no output byte limit; writing these configs per README won't take effect nor error. - Plugin consumer-side intentionally excludes crawling, platform source management, account sync,
sync-savedand other write side-effect capabilities (explicitly noted insrc/index.ts:7-10comments), these still need to be operated in main project / browser extension / mobile app. - When file pointed by
skillPathcan't be read, DSH skill registration silently skips (only logs warn), tools work but LLM can't see skill metadata; wrongworkdirtriggers this branch. - Neither README nor package.json declare Node version requirements; if DSH host Node version is too low, may affect other
dsh-*peer dependencies; recommended to follow DSH docs for Node version. - Browser panel uses DOM layer extension of frame grid (
PanelLayoutControllerrewritesgrid-template-columns), doesn't depend on officialasideslot; if DSH frame structure changes in future, may need to adjustpanel.module.css/layout.ts(source code commentsrc/client/index.ts:4-7already notes this risk).
OpenBiliClaw 是本地运行、跨平台、可调教的个性化内容推荐 Agent;本仓库是它的 DeepSeek Harness 客户端插件——DSH 左侧栏一个 OpenBiliClaw 按钮,点开右侧滑出抽屉(推荐/内容库/对话/画像/设置),并注册 22 个 Agent Bridge 工具,让 Agent 读推荐、答探测、闭环学习。
English | 中文
这是什么
OpenBiliClaw 是一个本地运行、跨平台、可调教的个性化内容推荐 Agent:从你在 B 站 / 小红书 / 抖音 / YouTube / X / 知乎 / Reddit / Linux.do / Bangumi / V2EX / 微博的使用、反馈与对话里持续深化你的心理画像,再主动去这些平台为你找内容。
本仓库是它的 DeepSeek Harness(DSH)客户端插件,把「消费侧」搬进 DSH 的 Web GUI:
- 人用侧:在 DSH 左侧栏底部加一个 OpenBiliClaw 按钮,点开从右侧滑出一个与浏览器插件 / 手机版视觉一致的抽屉面板——推荐流、惊喜推荐、内容库、苏格拉底式对话、用户画像、后端设置,全部在 DSH 里点开即用;
- Agent 用侧:注册 22 个
openbiliclaw_*工具(Agent Bridge v2 CLI)和openbiliclaw-adapterskill,让 DSH 里的 Agent 能读取推荐、回答探测、保存内容、和用户对话,形成「推荐 → 反馈 → 画像 → 更准的推荐」的闭环。
典型场景:左边和 Agent 聊需求,点左下角的 OpenBiliClaw 按钮,右侧抽屉滑出来顺手刷两条推荐,点个「收藏」,过一会儿弹出兴趣探测,和它聊两句,画像就更懂你了——全程不离开 DSH。
边界说明:本插件只做消费侧。爬取 / 平台源管理 / 账号同步等能力仍属于主项目(后端 + 浏览器插件 / 手机版),有意不带入 DSH。
截图(真实使用场景)
| 推荐 · 亮色 | 推荐 · 深色(跟随 DSH 主题) |
|---|---|
![]() | ![]() |
| 内容库 | 对话 | 画像 |
|---|---|---|
![]() | ![]() | ![]() |
| 设置 · 模型 | 设置 · 调度 | 设置 · 通用 |
|---|---|---|
![]() | ![]() | ![]() |
功能
推荐(For You)
- 惊喜推荐 hero 大图:整卡宽的 16:9 封面 + ✨ 角标 + 「💗 91% 匹配」分数胶囊,展开可见推荐理由、看看/喜欢/稍后看/收藏/少来点/聊一聊
- 跨平台封面兼容:与浏览器插件 / PC Web 统一经后端
/api/image-proxy加载,归一化小红书、YouTube 等 CDN URL;代理失败时自动回退为文本或媒体占位,不留空白卡片 - 无限滚动:滚到底自动加载下一页;距离底部约 800px 时提前预加载;加载中显示底部转圈指示
- 换一批 / 追加一批 / 刷新;卡片内嵌喜欢/不感兴趣/评论反馈;底部动态流(探测、惊喜、保存事件)
- 未初始化时提供「开始初始化」入口,自动轮询初始化状态
内容库
- 稍后再看 / 收藏:打开、移除,与主项目共享同一份本地列表
- 顶部「同步到平台」:全量同步到 B 站收藏夹 / 稍后再看,轮询展示逐条结果(成功 / 失败 / 需登录)
- 每条可单独「同步 / 重试同步」;卡片内嵌 喜欢 / 不感兴趣 / 聊一聊,以及与另一列表的互切(收藏 ⇄ 稍后再看)
- 已同步(
synced/already_synced)的条目自动从侧栏隐藏(数据保留);空态区分「已全部同步」与「还没有收藏」 - 非 B 站平台条目标注同步状态(如「仅本地保存」),不误报为已同步
- 历史记录:近 30 天「点开过 / 看过 / 移除的」,光标分页 + 触底自动加载,带收藏/稍后/不再推荐等上下文徽章;「移除的」分类可一键恢复收藏 / 稍后再看
对话
- 苏格拉底式对话引擎:一问一答之间持续深化画像
- 兴趣 / 回避探测卡片:确认、拒绝、聊聊、稍后,乐观状态机即时反馈
- 待聊确认卡片:准 / 不准 / 聊聊,回复会带回被确认的原话上下文
画像
- 与 popup 同构的画像卡:MBTI、核心特质、深层需求、兴趣树、风格、洞察与觉察
- 新增「阿B 最近新记住了什么」认知卡,可展开影响/推理/依据,并支持加载更早的认知更新
设置(与浏览器插件后端设置页对齐)
- 模型:v2 实例模型(新建/编辑/删除实例、provider 条件字段、获取模型)、默认调用链(排序 + 测试整链)、模块路由(画像理解/内容发现/推荐表达/内容评估)、Embedding(含备选 provider + 测试)、LLM 并发/超时
- 调度:版本与更新(检查/应用)、全部调度参数、猜测兴趣参数
- 高级功能:P1 视觉画像 / P2 弹幕 / P3 关键帧、多模态处理、搜索词生成模式
- 通用:连接地址、语言、保存同步、数据目录、海外网络代理(带测试)、局域网访问密码、开机自启动、重新初始化
- 日志;全局「保存配置」栏 + 脏状态提示(未保存修改提醒)
面板体验
- 深色模式跟随 DSH 主题实时切换(保留 OpenBiliClaw 自己的粉/蓝配色体系)
- 后端在线状态点带迟滞:连续两次探测失败才置离线,单次抖动不再闪
- 图标使用插件(浏览器扩展)原版图标:品牌 logo + 四个 tab 图标逐字节对齐
与主项目的关系
┌───────────────────────────── DSH Web GUI ─────────────────────────────┐
│ 左边栏 │ 会话区 │ 详情栏 │ │
│ (DSH 自带, │ (DSH 自带) │ (DSH 自带) │ ← 右侧滑出 OpenBiliClaw │
│ 底部含 │ │ │ 抽屉(推荐/内容库/ │
│ OpenBiliClaw │ │ │ 对话/画像/设置) │
│ 按钮) │ │ │ │
└─────────────────────────────┬─────────────────────────────────────────┘
│ HTTP + WebSocket (默认 http://127.0.0.1:8420)
┌─────────────────────────────▼─────────────────────────────────────────┐
│ OpenBiliClaw 后端(主项目,同一份 config.toml / SQLite / 画像)│
└────────────────────────────────────────────────────────────────────────┘
▲
DSH Agent(本插件注册的工具 + skill)── Agent Bridge v2 CLI ──┘
- 面板与 Agent 工具走同一套后端状态:面板里收藏的,Agent 工具立刻看得到;Agent 代答的探测,面板同步生效
- Agent Bridge 要求后端开启 Agent Bridge v2(
python -m openbiliclaw.integrations.openclaw.cli <command>),插件默认通过<workdir>/.venv/bin/python调用 - 插件注册的 skill 读取
<workdir>/skills/openbiliclaw-adapter/SKILL.md(本仓库skills/openbiliclaw-adapter/SKILL.md附了同一份副本,方便对照)
插件组成
| 部分 | 说明 |
|---|---|
src/index.ts + src/tools.ts | node 半:22 个 openbiliclaw_* 工具(bridge CLI,经 ctx.shell 执行器——新版 DSH 中 bash 更名为 shell)+ 注册 adapter skill |
src/skill.ts | 解析 SKILL.md frontmatter 并注册为 DSH skill |
src/bridge.ts | Agent Bridge v2 CLI 调用封装(超时 / 输出上限) |
src/client/* | browser 半:左侧栏按钮(sidebar.footer.action)+ 右侧抽屉(shell.overlay)面板(React),含推荐/内容库/对话/画像/设置/消息抽屉、WebSocket 实时流、深色模式 |
lib/ | 构建产物(lib/index.js node 半 + lib/client.js 浏览器半),可直接安装 |
安装
前置
- 一个可用的 DeepSeek Harness 部署(含 Web GUI)
- 一个运行中的 OpenBiliClaw 后端(主项目,开启 Agent Bridge v2,默认监听 127.0.0.1:8420)
插件要求 DSH 0.1.0-rc.6 或更高版本;插件的 peer ABI 已与该版本的 dsh-* 包及 @deepseek-ai/cordis ^4.0.1 对齐。请不要在同一个 profile 中混用 0.0.1 时代的 DSH 工具包。
通过 DSH 插件 bundle 安装
本仓库声明了 dsh.bundle,因此可以作为完整插件包交给 dsh plugin add 或插件市场安装。bundle 会自动提供 openbiliclaw 配置行;安装后仍需按下面的配置说明,把 workdir 指向本地 OpenBiliClaw 主项目目录。
通过 npm/pnpm 安装(已发布版本)
npm install @openbiliclaw/dsh-plugin
# 或 pnpm add @openbiliclaw/dsh-plugin
安装后同样在 cordis.patch.yml 中增加 openbiliclaw 配置行(见下文)。
1. 界面槽位(当前 DSH 免配置)
面板渲染在两个加法槽位上,无需改 DSH 源码:
- 左侧按钮:
sidebar.footer.action(侧边栏底部、设置旁的附加动作) - 右侧抽屉:
shell.overlay(帧级浮层,滑出在详情/文件/变更面板之上,不占用任何列)
两个槽位都是官方 DSH 自带的 list 槽(当前官方 DSH 已移除旧的 aside 列),此步直接跳过。
2. 安装插件包
把本仓库放进 web profile 的依赖目录并声明插件行:
cp -r <本仓库> ~/.dsh/profiles/<profile>/node_modules/@openbiliclaw/dsh-plugin
在 ~/.dsh/profiles/<profile>/cordis.patch.yml 增加:
# DSH 0.1.0-rc.6+: 新增配置行必须放在 insert 下
- insert:
- id: openbiliclaw
name: '@openbiliclaw/dsh-plugin'
config:
workdir: '/你的/OpenBiliClaw/项目目录' # 后端项目根目录(含 .venv 与 skills/)
cordis.patch.yml 的顶层是 patch 列表;不带 insert 的 id 条目会被当作已有配置行覆盖,找不到 openbiliclaw 时会被跳过。
3. 重启
dsh web # 重启 DSH Web 进程,刷新页面
面板内的「设置 → 通用 → 连接」可改后端地址(默认 http://127.0.0.1:8420,面板本地保存、立即生效;地址本身不要再附加 /api)。
配置项(cordis 行 config)
| 键 | 默认 | 说明 |
|---|---|---|
workdir | (必填) | OpenBiliClaw 后端项目根目录;bridge CLI 与 SKILL.md 都从这里解析 |
pythonBin | <workdir>/.venv/bin/python | 用于调用 bridge CLI 的 Python |
skillPath | <workdir>/skills/openbiliclaw-adapter/SKILL.md | adapter skill 文件 |
timeoutMs | 300000 | 单次 CLI 调用超时 |
stdoutMaxBytes | 2000000 | CLI 输出上限 |
构建
构建需要 DSH 源码 checkout(类型与打包 preset 都从那里解析;本仓库不是 pnpm workspace 成员):
tsc -p tsconfig.json # 先产出 lib/types
tsdown --env.DSH_BUILD_FACE client # node 半 + 浏览器半(window.__ModuleLoader__ 闭包)
tsdown.config.ts 引用了 DSH checkout 的 packages/client/tsdown.client.ts(clientBundle preset),首次构建前把该路径改成你的 checkout。
Agent 侧工具一览
openbiliclaw_recommend / openbiliclaw_append_recommendations / openbiliclaw_reshuffle / openbiliclaw_get_delight / openbiliclaw_respond_delight / openbiliclaw_submit_feedback / openbiliclaw_get_activity_feed / openbiliclaw_chat / openbiliclaw_get_chat_history / openbiliclaw_next_probe / openbiliclaw_respond_interest_probe / openbiliclaw_next_avoidance_probe / openbiliclaw_respond_avoidance_probe / openbiliclaw_get_profile / openbiliclaw_get_profile_edit_state / openbiliclaw_edit_profile / openbiliclaw_list_saved / openbiliclaw_save_local / openbiliclaw_remove_saved / openbiliclaw_get_runtime_status / openbiliclaw_get_platform_availability / openbiliclaw_get_capabilities
边界(有意不做)
- 平台源 / 爬取配置、源状态、池配比(这些属于主项目的「平台源」设置页)
- 同步到外部平台账号(本地收藏/稍后看始终本地优先,手动同步在主项目)
- 浏览器插件专属的设备配对、断开暂停等
相关链接
- 主项目:OpenBiliClaw · 项目主页
- DSH:DeepSeek Harness
License
友情链接
English
A DeepSeek Harness (DSH) client plugin for OpenBiliClaw, the local-first cross-platform content-discovery agent. It adds a left-sidebar OpenBiliClaw button that opens a right-side drawer over the DSH web GUI (the sidebar.footer.action + shell.overlay seats) with the consumer side of OpenBiliClaw — recommendations with a hero delight banner, infinite scroll with prefetch, saved/history library, Socratic dialogue with interest/avoidance probes, the user profile card, and a settings surface aligned with the browser extension — and registers 22 openbiliclaw_* tools plus the openbiliclaw-adapter skill so DSH agents can drive the same backend in a closed loop. Cover images use the same backend image proxy as the browser and PC Web clients, with CDN URL normalization and a local fallback for failed loads. Crawling and source management intentionally stay in the main project. Requires DSH 0.1.0-rc.6 or newer (with the matching dsh-* ABI and @deepseek-ai/cordis ^4.0.1) plus a running OpenBiliClaw backend (Agent Bridge v2, default http://127.0.0.1:8420). When adding the plugin to cordis.patch.yml, wrap the row in a top-level insert entry; a bare id row is treated as an override and skipped when it does not already exist. See the Chinese section above for install, build and configuration details.
Read the usage guide →
Install steps, key points, FAQ and compatibility for this plugin — auto-derived from indexed fields.
Listing badge
[](https://deepseek-plugin.org/plugins/whiteguo233/dsh-openbiliclaw)Paste this markdown into your GitHub README to link back to this listing. The badge only states the listing — not a security endorsement.







