dsh-feishu-bridge
Chat with DeepSeek Harness (DSH) through a Feishu (Lark) bot. Messages you send the bot enter DSH sessions; replies come back as live streaming interactive cards — reasoning, body text, and tool chain in separate sections with typewriter updates. No public webhook needed (uses the Feishu Open Platform long connection).
通过飞书机器人与 DeepSeek Harness(DSH)对话的桥接插件。回复以流式交互卡片实时发回(思维链/正文/工具链分区、打字机效果),无需公网 webhook。
English
Features
| Capability | Description |
|---|---|
| Direct chat | messages to the bot reach DSH with multi-turn context |
| Group @ | respond when @-mentioned in groups (needs botOpenId) |
| Streaming cards | CardKit 2.0 cards update in real time: reasoning + body + tool chain sections, token-level typewriter; native collapsible panels (collapsed when done) |
| Multi-turn context | same Feishu chat reuses the same DSH session; auto-resume after DSH restart (agents.resume) |
| Slash commands | /help, /reset, /status |
| Long-reply chunking | >8000 chars auto-splits into follow-up text messages |
| Reconnect | consume subprocess auto-backoff reconnects (2s → 30s cap) |
| Allowlist | only specified open_ids may chat |
| Idempotent dedup | per message_id dedup against event replays |
| Channel direct-send | registers a feishu channel in DSH's global registry (globalThis.__dshChannelNotify), so de_channel_send / de_notify can push text/images/files to Feishu (defaults to the most recent chat; target: 'p2p:oc_xxx' supported) |
Architecture
Feishu App ──long connection──> lark-cli event consume im.message.receive_v1 --as bot (child process)
│ NDJSON (chat_id / sender_id / content / mentions …)
▼
feishu-bridge plugin (DSH host process, cordis)
│ agents.create / agents.resume (chat_id → sessionId)
▼
DSH agent session (same models/tools as GUI)
│ llm/stream waterfall (per sessionId, token deltas)
▼
streaming accumulator (text / reasoning / tool-call deltas)
│ throttled PATCH (default ≥1.2s)
▼
lark-cli api patch im/v1/messages/:id ──▶ interactive card updates
└── turn end ──▶ final card (✅ done) + sessions.flush
- Event listening:
lark-cli event consume im.message.receive_v1 --as bot(long-connection WebSocket, no public network required) - Streaming hook: DSH's
llm/streamwaterfall events, matched persessionId - Card updates:
PATCH /open-apis/im/v1/messages/:message_idwith the card JSON string - Session mapping:
<stateFile>(default~/.dsh/plugins/dsh-feishu-bridge/state.json),chat_id → sessionId
Prerequisites
- DeepSeek Harness running
dsh web(plugin runs in the host process) lark-cliinstalled + Feishu app configured (~/.lark-cli/config.json), bot identity available- A Feishu custom app with bot capability, scope covering target users/groups
npm install -g @larksuite/cli
lark-cli auth login # choose bot identity (tenant_access_token)
lark-cli auth status # bot: ready
The app needs the
im:messagepermission; events come via long-connection subscription — no callback URL needed.
Install
git clone https://github.com/fengs2021/dsh-feishu-bridge.git ~/.dsh/plugins/dsh-feishu-bridge
cd ~/.dsh/plugins/dsh-feishu-bridge && npm install
# register into the web profile:
# dependencies += "dsh-feishu-bridge": "link:/root/.dsh/plugins/dsh-feishu-bridge"
# dsh.profile.bundles += "dsh-feishu-bridge"
cd ~/.dsh/profiles/web && pnpm install
systemctl restart dsh-web
Send the bot a message in Feishu — a card reply means success.
Config (all optional, cordis.patch.yml)
- id: feishu-bridge
config:
botOpenId: 'ou_xxxxxx' # bot open_id (group @ detection)
allowlist: ['ou_xxxxxx'] # open_id allowlist; empty = everyone
cwd: '~' # working dir for new sessions
enableGroup: true # respond to group @mentions
maxReplyChars: 3500 # text-mode reply cap (chunked beyond)
typingHint: true # "thinking" hint in text mode
replyMarkdown: true # markdown formatting in text mode
streamCard: true # streaming card mode (default on)
cardPollMs: 600 # stream poll interval (turn-end fallback)
cardMinIntervalMs: 1200 # min card update interval (rate-limit guard)
maxTurnMs: 600000 # max wait per turn
larkBin: 'lark-cli' # lark-cli executable
stateFile: '~/.dsh/plugins/dsh-feishu-bridge/state.json'
Getting botOpenId: auto-probed at startup (/open-apis/bot/v3/info); or @ the bot in a group and read the log (learned botOpenId=ou_xxx from group mention); or check the Feishu console → app → bot.
Usage
- Open the bot chat in Feishu (or add the bot to a group)
- Message it directly; in groups, @ the bot
- Commands:
/help,/reset,/status
Channel direct-send (DSH → Feishu): de_channel_send channels=feishu content=...; attachments via attachments=[{kind:'image'|'file', path|url|base64, fileName?}]; target defaults to the most recent Feishu chat, or explicit target: 'p2p:oc_xxx'. Implemented in lib/channel-registry.js (standalone, zero deps on the notification module).
FAQ highlights
- Card stuck at "thinking" → check
bridge.log:patchCard failed= app permission;target=miss= session mapping stream is not async iterable→ known bug of old versions; upgrade to 1.0.0 (⚠️ the listener affects ALL LLM calls — do not revert to async)- Group @ not responding → check
botOpenId+ app scope - Card update 230001/230099 → must use
PATCH /open-apis/im/v1/messages/:idwith card JSON body (neverPUT+msg_type) - Context lost after restart → plugin auto-
agents.resume; failed resumes rebuild on next message
Development
npm test # pure-function unit tests
npm run check # static checks
Design notes for contributors: all pure functions are exported (collectReply / buildCard / createStreamAccumulator / toolSummary / splitChunks / cleanContent / shouldHandle); llm/stream is a cordis waterfall event — the listener must be a sync function returning an AsyncIterable (yield chunks through, never swallow exceptions); all lark-cli calls go through runLark() (timeout + output capture, degraded paths); concurrent resume on the same session is coordinated by a per-session promise lock.
中文
通过飞书机器人与 DeepSeek Harness(DSH) 对话的桥接插件。
在飞书里给机器人发消息 → 消息进入 DSH 会话由 AI 处理 → 回复以流式交互卡片实时发回:思维链、正文、工具链分区展示,打字机效果逐字更新。无需公网 webhook(使用飞书开放平台长连接)。
效果示例(交互卡片,随生成实时更新):
┌──────────────────────────────────┐ │ 🤖 DSH 助手 │ ├──────────────────────────────────┤ │ ▶ 🧠 思维链 · 212 字(点击展开) │ │ 用户想查 /root/dsh 目录… │ │ ──────────────────────────────── │ │ 好的,我来查看一下目录内容… │ │ ──────────────────────────────── │ │ 🔧 工具链(2) │ │ - ✅ bash ls -la /root/dsh │ │ - ✅ memory 写入今日日志 │ │ ──────────────────────────────── │ │ ✅ 已完成 │ └──────────────────────────────────┘
功能特性
| 能力 | 说明 |
|---|---|
| 私聊对话 | 与机器人单聊,消息直达 DSH,多轮上下文连续 |
| 群聊 @ | 群聊中 @ 机器人即响应(需配置 botOpenId) |
| 流式卡片 | CardKit 2.0 交互卡片实时更新:思维链 + 正文 + 工具链分区,token 级打字机效果;思维链/工具链为原生折叠面板(完结后默认收起,点击展开) |
| 多轮上下文 | 同一飞书会话固定复用同一 DSH 会话;DSH 重启后自动恢复(agents.resume) |
| 斜杠命令 | /help 帮助、/reset 清空上下文、/status 会话状态 |
| 长回复分段 | 超过卡片预览上限(8000 字)自动补发文本消息 |
| 断线重连 | consume 子进程异常退出自动退避重连(2s 起、30s 封顶) |
| 白名单 | 只允许指定 open_id 对话 |
| 幂等去重 | 按 message_id 去重,避免事件重放导致重复处理 |
| 渠道直发 | 注册 feishu 渠道到 DSH 全局注册表(globalThis.__dshChannelNotify),de_channel_send / de_notify 可直接把文本/图片/文件发到飞书(默认发往最近交互的会话,也支持 target: 'p2p:oc_xxx' 显式指定) |
架构原理
飞书 App ──长连接──> lark-cli event consume im.message.receive_v1 --as bot(子进程)
│ NDJSON(chat_id / sender_id / content / mentions …)
▼
feishu-bridge 插件(DSH host 进程内,cordis 插件)
│ agents.create / agents.resume(chat_id → sessionId 映射)
▼
DSH agent 会话(模型、工具与 GUI 同款)
│ llm/stream waterfall(按 sessionId 匹配,token 级 delta)
▼
流式累积器(text-delta / reasoning-delta / tool-call-delta)
│ 节流 PATCH(默认 ≥1.2s 一次)
▼
lark-cli api patch im/v1/messages/:id ──> 交互卡片实时更新(打字机效果)
│
└── turn 结束 ──> 最终态卡片(✅ 已完成)+ sessions.flush 持久化
- 事件监听:
lark-cli event consume im.message.receive_v1 --as bot(飞书长连接 WebSocket,无公网要求) - 流式接入:DSH 的
llm/streamwaterfall 事件(每次模型调用都会经过),按请求携带的sessionId匹配到飞书 turn,拦截增量 chunk - 卡片更新:
PATCH /open-apis/im/v1/messages/:message_id,content为交互卡片 JSON 字符串 - 会话映射:
<stateFile>(默认~/.dsh/plugins/dsh-feishu-bridge/state.json),记录chat_id → sessionId
前置条件
| 依赖 | 说明 |
|---|---|
| DeepSeek Harness | 已安装并运行 dsh web(插件运行在 host 进程内) |
| lark-cli | 已安装并完成飞书应用配置(~/.lark-cli/config.json),bot 身份可用 |
| 飞书自建应用 | 具备机器人能力,可用范围包含目标用户/群 |
lark-cli 安装与授权:
npm install -g @larksuite/cli
lark-cli auth login # 选择 bot 身份(tenant_access_token)
lark-cli auth status # bot: ready
bot 身份要求应用具备
im:message(收发消息)权限;消息事件走长连接订阅,无需在开放平台配置回调地址。
安装
# 1. 获取插件源码(二选一)
git clone https://github.com/fengs2021/dsh-feishu-bridge.git ~/.dsh/plugins/dsh-feishu-bridge
# 或手动放置到 ~/.dsh/plugins/dsh-feishu-bridge/
# 2. 安装插件依赖(@deepseek-ai/dsh-agent 等)
cd ~/.dsh/plugins/dsh-feishu-bridge
npm install
# 3. 注册进 web profile
cd ~/.dsh/profiles/web
# 编辑 package.json:
# - dependencies 增加 "dsh-feishu-bridge": "link:/root/.dsh/plugins/dsh-feishu-bridge"
# - dsh.profile.bundles 数组增加 "dsh-feishu-bridge"
pnpm install
# 4. 重启 dsh web(插件在 host 进程内启动 consume 子进程)
systemctl restart dsh-web # systemd 托管
# 或手动重启你的 dsh web 进程
安装成功后,在飞书中给机器人发一条消息测试;回复以卡片形式出现即成功。 若
~/.dsh/plugins/下已有其他插件(如 dsh-novel-studio),参照其安装方式即可。
配置
在 ~/.dsh/profiles/web/cordis.patch.yml 中覆盖配置(全部可选,均有默认值):
- id: feishu-bridge
config:
botOpenId: 'ou_xxxxxx' # 机器人 open_id(群聊 @ 判断用;见下方获取方式)
allowlist: ['ou_xxxxxx'] # open_id 白名单;空数组 = 允许所有人
cwd: '~' # 新会话工作目录(默认用户主目录)
enableGroup: true # 是否响应群聊中 @ 机器人的消息
maxReplyChars: 3500 # 文本模式单条回复上限(超出分段)
typingHint: true # 文本模式下收到先回「思考中」提示
replyMarkdown: true # 文本模式回复使用 markdown 排版
streamCard: true # 流式卡片模式(默认开;关闭则退回文本分段回复)
cardPollMs: 600 # 流式轮询间隔(毫秒,兜底检测 turn 结束)
cardMinIntervalMs: 1200 # 卡片更新最小间隔(毫秒,飞书接口限频保护)
maxTurnMs: 600000 # 单轮最长等待(毫秒),超时停止更新并提示
larkBin: 'lark-cli' # lark-cli 可执行文件路径
stateFile: '~/.dsh/plugins/dsh-feishu-bridge/state.json'
botOpenId 获取方式
- 插件启动时自动探测(
/open-apis/bot/v3/info;部分应用权限下返回为空); - 在任意群里 @ 机器人发一条消息,插件日志会打印
learned botOpenId=ou_xxx from group mention,填入配置即可; - 飞书开放平台后台 → 应用 → 机器人,查看机器人 open_id。
使用
- 在飞书中搜索并打开机器人会话(或让管理员把机器人拉进群聊);
- 直接发消息即可对话;群聊中需 @ 机器人;
渠道直发(DSH → 飞书)
插件在 apply 时把主动发送能力登记到 DSH 渠道注册表(与 dsh-memory-evolve 通知模块的 globalThis.__dshChannelNotify 约定一致),因此 DSH 的 de_channel_send / de_notify 工具可直接发到飞书:
- 文本:
de_channel_send channels=feishu content=... - 附件:
attachments=[{kind:'image'|'file', path|url|base64, fileName?}](本地路径/base64 经临时目录 + 相对路径发送;图片走--image,其余走--file) - 目标:缺省 = 最近交互的飞书会话(插件 state 记录);显式传
target: 'p2p:oc_xxx' - 实现:
lib/channel-registry.js(独立模块,零依赖通知模块)
- 命令:
/help— 帮助/reset— 清空当前对话上下文,重新开始/status— 查看会话状态(sessionId / 模型 / 已处理消息数)
运维
| 事项 | 说明 |
|---|---|
| 插件日志 | ~/.dsh/plugins/dsh-feishu-bridge/bridge.log(事件接收 / 会话创建 / 流式匹配 / 错误) |
| 会话映射 | state.json;删除某条映射并重启即与该飞书会话「断连」(也可在飞书里发 /reset) |
| consume 异常 | 子进程自动退避重连;kill -9 可能泄漏服务端订阅,勿用 |
| 卸载 | 从 profile package.json 的 dependencies/bundles 移除,pnpm install 后重启 |
| 升级 | git -C ~/.dsh/plugins/dsh-feishu-bridge pull 后重启 dsh web |
常见问题(FAQ)
Q:飞书里一直显示「思考中」,卡片不更新?
检查 bridge.log:若出现 patchCard failed,确认飞书应用权限与卡片消息是否可更新;
若出现 target=miss,确认消息对应会话的 sessionId 与 state.json 一致(重启后会自动恢复)。
Q:收到「stream is not async iterable」?
历史版本的已知 bug(llm/stream listener 误用 async 函数),升级到 1.0.0 即可。
⚠️ 该 listener 影响所有 LLM 调用,请勿改回 async 函数。
Q:群聊里 @ 机器人没反应?
确认 botOpenId 已配置正确(见上文获取方式),且应用已在群内、可用范围包含该群。
Q:卡片更新失败(230001 / 230099)?
卡片更新接口固定为 PATCH /open-apis/im/v1/messages/:id(body 为
{"content": "<卡片JSON字符串>"});不要改用 PUT + msg_type(会返回 230001)。
Q:重启 DSH 后上下文丢失?
重启后插件会 agents.resume 恢复持久化会话(日志可见 preloaded session ...)。
若 resume 失败(如会话文件损坏),映射会被保留,下次消息到来时自动重建。
开发
# 单元测试(纯函数:流式累积器 / 事件聚合 / 卡片构建 / 辅助函数)
npm test
# 静态检查
npm run check
目录结构
dsh-feishu-bridge/
├── lib/index.js # 插件主体(cordis 插件)
├── cordis.patch.yml # bundle patch(插入 web profile roster)
├── test/functions.test.mjs # 纯函数单元测试
├── package.json
├── README.md
├── CHANGELOG.md
└── LICENSE
设计要点(贡献者必读)
lib/index.js导出全部纯函数(collectReply/buildCard/createStreamAccumulator/toolSummary/splitChunks/cleanContent/shouldHandle),便于测试与复用;llm/stream是 cordis waterfall 事件:listener 必须为同步函数且返回AsyncIterable(包装流时逐 chunkyield透传,不得吞异常);- 所有
lark-cli子进程调用走runLark()(超时 + 输出捕获),失败均有降级路径; - 同一持久化会话的并发
resume通过 per-session promise 协调锁去重。
许可证
相关项目
- DeepSeek Harness — 本插件运行的宿主
- @larksuite/cli — 飞书开放平台 CLI(事件订阅 / 消息收发 / 卡片更新)
No comments yet. Be the first to write one.