<p align="center"> <img src="assets/logo.jpg" width="128" alt="dsh-feishu"> </p>

<h1 align="center">dsh-feishu</h1> <p align="center">Harness 长在飞书里，不是飞书遥控网页</p>

<p align="center"> <img src="https://img.shields.io/badge/dsh--plugin-Feishu-3370ff" alt="dsh-plugin"> <img src="https://img.shields.io/badge/UI-Feishu_only-00d6b9" alt="Feishu only"> <img src="https://img.shields.io/badge/deploy-headless-222222" alt="headless"> <img src="https://img.shields.io/badge/runtime-Node_22-339933" alt="Node 22"> <img src="https://img.shields.io/badge/license-MIT-green" alt="MIT"> </p>

<p align="center"> <a href="#中文">中文</a> · <a href="#english">English</a> · <a href="#features">功能</a> · <a href="#who">适合谁</a> · <a href="#install">五分钟</a> </p>

<p align="center"> <img src="assets/hero.jpg" width="100%" alt="dsh-feishu: Harness 长在飞书里，不是飞书遥控网页"> </p>

---

中文

插件市场里 feishu / lark 已经三十多个。多数是通知器、多 IM 网关，或 dsh plugin --profile web add 之后挂在官方网页旁边的桥。飞书只是遥控器：网页还在，关掉浏览器，产品就少了一半。不少还默认你坐在个人 PC 前，终端里弹二维码。

这套是跑在飞书里的 harness，部署面是 无头常驻：macOS、Linux、WSL 都能跑，主打一台能出网的机器当守护进程，不是笔记本终端扫码。一个 dsh --profile feishu 常驻，日志进文件，没有要人盯着的 TTY。飞书长连接进、飞书卡片出。模型、斜杠命令、进度、Goal、会话，全部发生在聊天框。没有 3080，没有给本机用的 TUI。高危操作出飞书审批卡，不是网页弹窗。

首个 bot 只能在开放平台建好，把 App ID / Secret 写进环境变量再启动。服务器上没处可贴二维码。已经通了的 bot 再开第二个，码可以打回飞书；那是后话，不是第一次上手。

不跟货架比按钮数量。没有打字机答案卡。要比的四件事：

1. 长在飞书里 — harness 的面就是飞书，不是遥控网页 2. Goal 卡片 — 设定、改目标、暂停、恢复、清除，点卡片就能做；重启不偷着续 3. 一进程多 bot — 2～3 个飞书身份共用一套 Agent，同一条 session 不能两头同时握 4. 会变的进度头 — [dsf-23% G1/3] Using bash...(2)，不是「正在思考…」四个字一直挂着

<p align="center"> <img src="assets/cards.png" width="960" alt="Goal card, approval card, ask card"> </p>

GitHub 上另有一个 PGZXB/dsh-feishu，飞书控制台 + 卡内审批，官方 Web 还在。owner 不同，产品也不同。

<a id="features"></a>

卖点

完全不依赖 Web

不装 dsh-web-app，不开 3080，不跑 headless 一次性进程。Cordis 插件直接 create / followup / steer / cancel / resume。飞书是面，不是壳外面再套的壳。

Goal：一张卡片管到底

发 /goal，或直接在卡片里填目标点「设定并开始」。句末可写「最多跑 3 轮」。

Agent 自己多轮推进。进程重启后 不会 偷偷接着烧，要在卡片上点「恢复自动续跑」，或发 /goal resume。

卡片上能做的事：

• 设定 / 修改目标（卡内表单）
• 暂停 / 恢复自动续跑
• 清除当前目标
• 完成变绿、阻塞变红，原地改，不另刷一堆纯文字

斜杠命令仍然全部可用。先到的算数。

会变的进度头

进度卡片原地改：

[dsf-23% G1/3] Using bash...(2)

模型别名、上下文占比、Goal 轮次（快到上限标 near）。跑完变 Done，失败变 Failed，正文另发一条，不跟进度糊在一起。

一进程多 Bot

同一个进程挂 2～3 个独立飞书应用。会话按 (appId, chat_id) 分开。历史可以互相 /resume，同一条 session 不能被两个窗口同时握住。

斜杠命令就在聊天框

/stop /clear /model /status /compact /resume /rename /goal /bye

不支持的 / 会直接告诉你本期有哪些。完整表见 [docs/handbook.md](docs/handbook.md)。

高危审核卡 / 提问卡

未点名的 rm、find -delete、破坏性 git 会挂起，飞书出一张橙卡。点「允许这一次」才跑；打字同意不算批准。

需要选择时模型调 ask_user_question，出蓝卡。选项两列一排带序号，说明是灰字菜单；可以点选项，也可以在卡内填空提交（输入框默认一行），也可以直接打字。先到的算数。开放平台须订 card.action.trigger。

重启自己接上

进程重启后不用你先说话：插件主动 resume 所有已知会话，发一张绿卡（会话名 + 模型名）。重启包装脚本可以把一句「续跑指令」写进 $DSH_HOME/restart-continue.json——绿卡里直接显示这条指令，并作为新消息投进会话，模型自己接着干。指令 10 分钟过期、读后即焚。

上下文会自己收

累计过 100K 之后，名单内（read / glob / grep / bash / 搜索读页）且超过 8K 的工具结果，完整进模型 3 次，第 4 次发送前走官方头 4K 尾 1K。skill 不过。冷启动和开门前的历史不回砍。glob 自动滤掉 node_modules / .git / .pnpm。可选：生图 / 多步浏览类 MCP 藏到 enable_mcp 之后。规格见 [PRUNE.md](PRUNE.md)。

多模态按飞书的方式走

图 / 语音 / 文件落到本地 inbox，把路径写给模型。不把图片塞进 session image block。回图、回文件、回语音走 send_file。

<a id="who"></a>

适合谁 / 不适合谁

适合：有一台能出网的机器，要把 dsh 做成 7×24 云端同事；已经在飞书办公、不想再为 Agent 开网页或守着笔记本终端；同一套 Agent 多个飞书身份；Goal 这种「交代完让它自己盯」，但中断和续跑必须自己说了算。

不适合：个人 PC 上扫码 30 秒上手、还想留官方 Web 控制台、需要打字机流式答案卡、或只把飞书当通知渠道。那些货架上已经很多，去搜 feishu 即可。

<a id="install"></a>

五分钟跑起来

面向一台已经能出网的机器（Linux / macOS / WSL；Windows 原生机也能起进程，Agent 调 bash 会别扭）。需要 Node 22 和已安装的 dsh（npm i -g @deepseek-ai/dsh）。

先在飞书开放平台建一个企业自建应用，事件订阅用长连接（WebSocket），不要 webhook。把 App ID / Secret 备好再部署。步骤见 [docs/feishu-setup.md](docs/feishu-setup.md)。

从 GitHub 装到独立 profile

dsh plugin --profile feishu add github:shoxiy-danny/dsh-feishu

建议钉扫描过的 commit，和市场页证据对齐：

dsh plugin --profile feishu add github:shoxiy-danny/dsh-feishu#<sha>

然后把 App ID / Secret 和 DEEPSEEK_API_KEY 放进环境，再：

dsh --profile feishu

从本仓脚本装（开发 / 自托管）

cp .env.example .env
# 填 FEISHU_APP_ID / FEISHU_APP_SECRET / DEEPSEEK_API_KEY

chmod +x scripts/*.sh scripts/dsh-feishu-cli
./scripts/setup.sh
./scripts/start.sh

启动后到这个应用的单聊打一句。应先收到表情，再看到进度头，然后是正文。发 /goal 会出一张可点的目标卡。

工作目录默认家目录；附件落 workspace/inbox/。第二个 bot 再填 FEISHU_APP_ID_2 / FEISHU_APP_SECRET_2。

文档

默认模型只有 DeepSeek 官方 dsf / dsp。加自己的供应商：手册「模型」一节按字段说明，不按厂商抄作业。样例骨架在 examples/。

需要 DeepSeek Harness。发现页：dsh-plugin.market。

---

English

The harness lives in Feishu, not a remote for the webpage.

The plugin market already has thirty-plus feishu / lark listings. Most are notifiers, multi-IM gateways, or bridges you add with dsh plugin --profile web add. Feishu is only a remote: the webpage is still there, and closing the browser takes half the product with it.

This is the harness, running in Feishu, meant as a headless daemon on a box that stays on: Linux, macOS, or WSL. Not a laptop you scan a terminal QR on. One dsh --profile feishu. Logs go to a file. Nobody has to watch a TTY. Feishu WebSocket in, Feishu card out. Models, slash commands, progress, Goal, sessions: all in chat. No port 3080. No desktop TUI. High-risk commands pause on a Feishu approval card, not a web popup.

The first bot has to be created on the open platform and dropped into env vars. A headless server has nowhere to put a QR code. Extra bots can later be opened from an already-working chat; that is not how you bootstrap.

We do not compete on button count. No typewriter answer card. Four things instead:

1. Lives in Feishu — the harness surface is chat, not a remote for the webpage 2. Goal card — set, edit, pause, resume, clear on the card; a restart will not keep burning tokens until you resume 3. Several bots, one process — 2–3 Feishu identities, one agent; one session cannot be held by two windows 4. A progress header that changes — [dsf-23% G1/3] Using bash...(2), not a frozen "Thinking…"

The three cards below are drawn from the real card JSON. They are not Feishu screenshots.

There is another GitHub repo named PGZXB/dsh-feishu: a Feishu console with in-card approvals, official web still on. Different owner, different product.

Features

No web dependency

No dsh-web-app, no port 3080, no one-shot headless runner. A Cordis plugin calls create / followup / steer / cancel / resume in-process. Feishu is the surface, not a wrapper around a wrapper.

Goal: one card, the whole loop

Send /goal, or type the objective on the card and tap Set and start. You can append "max 3 rounds" at the end.

The agent continues across turns. After a process restart it will not keep burning tokens until you tap Resume or send /goal resume.

On the card: create / edit (form), pause, resume, clear. Complete turns green, blocked turns red. Slash commands still work.

A progress header that actually changes

The card edits in place:

[dsf-23% G1/3] Using bash...(2)

Model alias, context %, Goal round (near when the cap is close). Done / Failed on the card; the reply is a separate message.

Several bots, one process

The same process holds 2–3 Feishu apps. Sessions are keyed by (appId, chat_id). History can /resume across bots; one session cannot be held by two windows at once.

Slash commands in the composer

/stop /clear /model /status /compact /resume /rename /goal /bye

Unknown / replies with the supported list. Full table: [docs/handbook.md](docs/handbook.md). /resume with no number sends a session card. The model can rename via SlashCommand({ command: "/rename …" }).

Context prunes itself

After 100K tokens, large read / glob / grep / bash / search results stay intact for 3 model views, then official head 4K + tail 1K. Skill results are never cut. Spec: [PRUNE.md](PRUNE.md).

Approval cards / ask-user cards

Unnamed rm, find -delete, and destructive git pause on an orange card. Tap Allow once to run. Typing yes does not approve. Choices go through ask_user_question (blue card): options sit in a numbered two-per-row grid with a grey description menu, tap one, submit the one-row form, or type a reply. First answer wins. Subscribe to card.action.trigger on the open platform.

Restart reattaches itself

After a restart you don't have to speak first: the plugin resumes every known session and posts a green card (session title + model). A restart wrapper can drop one LLM-written continue instruction into $DSH_HOME/restart-continue.json; the card shows it inline and delivers it into the session as a new message, so the agent picks up where it left off. Expires in 10 minutes, consumed on read.

Five minutes

Aimed at a machine that can reach the internet (Linux / macOS / WSL; a raw Windows box will start, but the agent’s bash calls will fight you). Need Node 22 and dsh (npm i -g @deepseek-ai/dsh).

Create a Feishu enterprise self-built app first. Subscribe to events over WebSocket, not webhook. Have App ID / Secret ready before you deploy. Details: [docs/feishu-setup.md](docs/feishu-setup.md).

Install from GitHub

dsh plugin --profile feishu add github:shoxiy-danny/dsh-feishu

Pin a scanned commit if you want the same evidence the market page shows:

dsh plugin --profile feishu add github:shoxiy-danny/dsh-feishu#<sha>

Put FEISHU_APP_ID / FEISHU_APP_SECRET / DEEPSEEK_API_KEY in the environment, then:

dsh --profile feishu

Install from this repo (dev / self-host)

cp .env.example .env
# fill FEISHU_APP_ID / FEISHU_APP_SECRET / DEEPSEEK_API_KEY

chmod +x scripts/*.sh scripts/dsh-feishu-cli
./scripts/setup.sh
./scripts/start.sh

Then send one message in a p2p chat with that app. You should get a reaction, a progress header, then the reply. /goal opens a tappable Goal card.

Default cwd is the home directory; attachments land in workspace/inbox/. A second bot is FEISHU_APP_ID_2 / FEISHU_APP_SECRET_2.

Who this is for / not for

For: you have a machine that stays on and want dsh as a 24/7 coworker; you already live in Feishu and do not want another Agent webpage or a laptop terminal to babysit; one agent, several Feishu identities; Goal ("leave it running") with pause / resume that you say out loud.

Not for: 30-second QR setup on a personal PC; people who still want the official web console or typewriter streaming cards; or Feishu-as-notifications-only. The market already has those. Search feishu.

Docs

Ships DeepSeek official aliases only: dsf / dsp. Extra providers: handbook §模型 (per-field, not per-vendor). Skeleton in examples/.

Requires DeepSeek Harness. Discovery: dsh-plugin.market.

MIT. Issues and PRs welcome.