@starxer/chatterbox4dsh
适用于 DeepSeek Harness (dsh) 的 Feishu/Lark 频道插件：在 Feishu 聊天中使用同一个代理，并将每个步骤（推理、工具调用、结果和耗时）显示为卡片。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:Starxer/chatterbox4dsh#42f15d2044c3ead6adc8006c47297179948dbeffChatterbox4dsh 以 @starxer/chatterbox4dsh 发布，当前版本为 0.2.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
中文 ｜ English
把 DeepSeek Harness 接进飞书/Lark 的唠叨型插件。 A chatty Feishu/Lark channel for DeepSeek Harness.
中文
把 DeepSeek Harness 的 agent 能力接入飞书/Lark 聊天，并把 agent 的每一步都唠叨给你看（推理、工具调用、结果、时长/token）。安装后，用户直接从飞书与 Harness Agent 对话，共享 DSH 的模型、工具、工作区和会话存储。
定位：
chatterbox4dsh（chatterbox = 唠叨话痨）只做"DSH 原生特性 → 飞书聊天"这一层接入，DSH 本身才是 agent 本体。我们不做一个独立的 agent 平台或 24h 常驻助手——那是另一个产品。飞书和 DSH Web UI 共享同一套服务，飞书只是多出一个聊天端口。设计理念：功能对齐优先于功能创新。新增能力前先对照 DSH Web UI 是否有同名/等价能力，有则对齐，无再讨论。
差异化：step 级过程透明——每个 step 一张卡片（💬 Reasoning / 📝 Message / 🛠 Tool call + 结果），底部两行 footer（时长/token · 速度/上下文），标题标注「第几轮 · 第几步」，reasoning 标题标注思考耗时与思考 token；快步骤自动合并为一张卡。弱模型下也能观察 Agent 行为并即时干预（
/steer/stop），其余 IM 桥接多收敛到"结果交付/审批"。
English
Bridges DeepSeek Harness agents into Feishu/Lark chat and narrates every agent step (reasoning, tool calls, results, timing/tokens). Once installed, you talk to the same Harness agent from Feishu, sharing DSH's models, tools, workspaces and session storage.
Positioning:
chatterbox4dsh("chatterbox" = talkative) only bridges DSH's native capabilities into Feishu chat — DSH itself is the agent. It is not a standalone agent platform or a 24/7 always-on assistant; that would be a different product. Feishu and the DSH Web UI share the same services — Feishu is just another chat port.Design principle: feature parity first. Before adding anything, check whether the DSH Web UI already has an equivalent; if it does, match it; only discuss new ideas when it doesn't.
What's different: step-level transparency — one card per step (💬 Reasoning / 📝 Message / 🛠 Tool call + result), a two-line footer (duration/tokens · speed/context), the title tagged with "turn · step", and the reasoning header showing thinking time and thinking tokens; fast steps collapse into a single card. You can watch a weak model work and intervene immediately (
/steer,/stop), while most other IM bridges stop at "deliver the result / approve".
中文
| 能力 | 说明 |
|---|---|
| 单聊 / 群聊 / 话题群 | 单聊和群聊按聊天复用 Session；话题群按线程独立 Session |
| 统一 per-step 卡片 | 每个 agent step 一张卡片，包含推理、文本、工具调用、结果预览；标题带「第 N 轮 · 第 M 步」，reasoning 标题带思考耗时与思考 token，footer 两行（⏱ 时长 · 📥 in → 📤 out / 🚀 tok/s · 📊 上下文）。快步骤（思考→工具调用→结果落在 150ms 内）只发一张卡 |
| 工具调用展示 | 工具名内联代码 + args（独立 fenced 代码块防溢出）+ 结果预览（terminal/web/search/read/diff），原地更新 wathet→green/red |
| Turn Complete 卡片 | turn 结束后展示总时长/LLM 时间/工具时间、TTFT/吞吐量、token/缓存命中率，footer 另显示 Enter while busy；吞吐量口径与 DSH Web UI deriveTurnMetrics 对齐（首 token 判定含 tool-call delta，token 与 decode 时间同批配对） |
| 会话管理面板 | /session：交互式卡片下拉选会话 + 切换/detach/归档/fork/改名/列表/刷新；/session list 表格卡；/session N 快速切换 |
| 斜杠命令 | /model /new /session /status /stop /steer /queue /busy /permission /reasoning /approve /deny /help 等 |
| 审批 | 与 DSH Web UI 共享同一份 pending 审批状态；审批卡片 Approve 在上 / Reject 在下，并显示 Reason: 原因 |
ask_user_question 卡片 | 问题卡片（选项/自定义输入/跳过），一次多问时顺序迭代、整批返回答案 |
| 图片 / 文件接收 | 图片按真实字节判型（PNG/JPEG/WebP/GIF）经 attachment store 落盘；文件下载到 DSH 原生附件库（~/.dsh/attachments/v1/files/…）并附 fileHostPath 给 agent 读取 |
| agent 主动发文件 | feishu_send_file 模型工具：agent 可把工作区文件推送到当前飞书聊天（≤30MB） |
| agent 接收文件 | feishu_receive_file 模型工具：agent 按需/兜底直接下载入站飞书文件到 DSH 原生附件库，返回 fileHostPath |
| 运行中注入 steer | /steer <内容> 一次性注入当前 turn；/busy steer 可把"运行中发消息"默认设为注入（持久化）；agent 空闲时 /steer//queue 自动回退为发新消息。运行中发普通消息会先回一条纯文本提示（已插入 / 已排队），且被注入的消息不再单独回一张回复卡（由当前轮的回复卡回答） |
| 运行中排队（/queue） | /queue <内容> 强制走 queue 路径（即使处于 steer 模式）作为新一轮运行；空闲时回退为发新消息 |
English
| Capability | Description |
|---|---|
| DM / group / topic group | DMs and groups reuse one Session per chat; a topic group gives each thread its own Session |
| Unified per-step card | One card per agent step with reasoning, text, tool calls and result previews; the title carries "turn N · step M", the reasoning header shows thinking time and thinking tokens, and the footer is two lines (⏱ duration · 📥 in → 📤 out / 🚀 tok/s · 📊 context). A fast step (thinking → tool call → result inside 150 ms) sends only one card |
| Tool call display | Tool name in inline code + args (its own fenced block to avoid overflow) + result preview (terminal/web/search/read/diff), updated in place wathet→green/red |
| Turn Complete card | After a turn: total / LLM / tool time, TTFT, throughput, tokens, cache hit rate, plus Enter while busy in the footer. Throughput matches the DSH Web UI's deriveTurnMetrics (first-token detection includes tool-call deltas; tokens and decode time are paired) |
| Session panel | /session: an interactive card with a session dropdown plus switch / detach / archive / fork / rename / list / refresh; /session list renders a table card; /session N switches by index |
| Slash commands | /model, /new, /session, /status, /stop, /steer, /queue, /busy, /permission, /reasoning, /approve, /deny, /help and more |
| Approvals | Shares the same pending approvals as the DSH Web UI; the card puts Approve on top / Reject below and shows the Reason: |
ask_user_question card | Question cards with options, free-form input and a skip button; multiple questions are asked one at a time and returned as one batch |
| Image / file intake | Images are typed from their real bytes (PNG/JPEG/WebP/GIF) and stored via the attachment store; files are downloaded into the native DSH attachment store (~/.dsh/attachments/v1/files/…) and exposed to the agent as fileHostPath |
| Agent sends files |
npx @deepseek-ai/dsh plugin --profile web add @starxer/chatterbox4dsh
中文：然后在 DSH Settings → 飞书与 Lark 中配置 App ID 和 App Secret。支持扫码一键配置（推荐）或手动创建应用。详见 docs/feishu-setup.md。
English: Then configure the App ID and App Secret in DSH Settings → Feishu & Lark. You can use the QR-code one-click setup (recommended) or create the app manually. See docs/feishu-setup.md.
中文
npx @deepseek-ai/dsh webEnglish
npx @deepseek-ai/dsh web中文
| 命令 | 说明 |
|---|---|
/model [list|<provider>/<model>] | 查看/列出/切换模型 |
/new [--workspace <path>] [--preset <id>] | 新建会话（可指定工作区和 preset） |
/session [N|list] | 无参：交互式会话管理面板；list：表格；N：按下标快速切换 |
/status | 展示会话状态（token/TTFT/吞吐量/缓存命中率/权限模式/Enter while busy 等） |
/reasoning [off|low|high|max] | 设置推理强度 |
/stop | 中止当前轮次并丢弃排队消息（不再进入下一 turn） |
/steer <内容> | agent 运行中，把一条消息注入当前 turn；空闲时自动回退为发新消息 |
/queue <内容> | /steer 的共轭：强制把消息排队为新轮（即使处于 steer 模式）；空闲时回退为发新消息 |
/busy [queue|steer] | 设置运行中（busy）的 Enter 行为：排队发送（默认）或插话发送，持久化 |
/permission [模式] | 查看/切换会话权限（沙箱）模式；无参发交互式选择卡片 |
/approve /deny /approvals | 处理工具审批 |
/help | 卡片列出所有可用命令（分组：chatterbox4dsh 插件 / DSH 内置） |
English
| Command | Description |
|---|---|
/model [list|<provider>/<model>] | View / list / switch models |
/new [--workspace <path>] [--preset <id>] | Create a session (optionally with a workspace and preset) |
/session [N|list] | No args: interactive session panel; list: table; N: switch by index |
/status | Session status (tokens / TTFT / throughput / cache hit rate / permission mode / Enter while busy) |
/reasoning [off|low|high|max] | Set the reasoning effort |
/stop | Abort the current turn and drop queued messages (they no longer run as the next turn) |
/steer <text> | While the agent runs, inject a message into the current turn; falls back to a new message when idle |
/queue <text> | The counterpart of /steer: force the message to queue as a new turn (even in steer mode); falls back to a new message when idle |
/busy [queue|steer] | Set what Enter does while busy: queue (default) or steer, persisted |
/permission [mode] | View / switch the session permission (sandbox) mode; no args sends an interactive card |
/approve /deny /approvals | Handle tool approvals |
/help | A card listing every command (grouped: chatterbox4dsh plugin / DSH built-in) |
关于会话压缩 / On compaction：
compact命令由 DSH 内置提供（@deepseek-ai/dsh-command-compact，注册/compact并调用官方压缩），本插件不注册compact命令（避免与 DSH 内置同名冲突）。当某个会话历史过大时，用 DSH 内置的/compact压缩其上下文，或将会话归档/另开新会话。DSH 压缩只改写「进入模型上下文的部分」，会话的完整事件日志不会被删除。The
compactcommand is provided by DSH itself (@deepseek-ai/dsh-command-compact, registered as/compact). This plugin deliberately does not registercompact(a same-name conflict with the built-in would fail at boot). When a session's history grows too large, use the built-in/compactto compact its context, or archive the session / start a new one. DSH compaction only rewrites the part that enters the model context; the full event log is never deleted.
运行中发消息的行为（
/busy）：queue（排队发送，等待当前轮结束后作为新轮运行）或steer（插话发送，注入当前轮立即响应，persist）。/status的 Enter while busy 行显示当前值。一次性插话用/steer <内容>；一次性排队用/queue <内容>。agent 空闲时，/steer、/queue都会自动回退为「作为新消息发送」而不是报错。运行中发普通消息时会立即回一条纯文本提示（steer：已插入当前轮；queue：已排队，本轮结束后执行）——提示是文本消息而非卡片；steer 注入的消息由当前轮的回复卡统一回答，不会再单独回一张卡。/stop会中止当前轮并丢弃排队/等待中的消息（不再自动进入下一 turn）。What happens when you send while running (
/busy):queue(wait for the current turn, then run as a new turn) orsteer(inject into the current turn and respond immediately, persisted). The Enter while busy row in/statusshows the current value. Use/steer <text>for a one-off injection and/queue <text>for a one-off queue. When the agent is idle, both fall back to "send as a new message" instead of erroring. A plain message sent while running gets an immediate plain-text notice (steer: injected into the current turn; queue: queued for after this turn) — a text message, not a card; the injected message is answered by the running turn's reply card and never gets a second card./stopaborts the current turn and drops queued/waiting messages (they no longer run as the next turn).
- id: lark-channel
config:
appId: cli_xxxxxxxxxxxxxxxx
appSecretRef: DSH_LARK_APP_SECRET
domain: feishu # feishu（中国版）/ lark（国际版） · feishu (China) / lark (international)
requireMention: true # 群聊是否必须 @机器人 · require an @mention in groups
dmMode: open # open / allowlist / disabled
workspace: /path/to/project # 默认工作区 · default workspace
agentPreset: coding # 默认 agent preset · default agent preset
中文
| 配置项 | 默认值 | 说明 |
|---|---|---|
appId | — | 飞书应用 App ID |
appSecretRef | DSH_LARK_APP_SECRET | Harness Credentials 中的 Secret 引用名 |
domain | feishu | feishu 或 lark |
requireMention | true | 群聊是否必须 @机器人 |
dmMode | open | 单聊策略：open / allowlist / disabled |
groupAllowlist | [] | 允许的群 chat_id 列表 |
dmAllowlist | [] | allowlist 模式下的用户 open_id 列表 |
provider / model | Harness 默认 | 为飞书渠道指定模型 |
workspace | 第一个注册的 Workspace | Agent 工作目录 |
agentPreset | Harness 默认 Preset | Agent preset |
reactEmoji | THUMBSUP | 收到消息时的表情回应（空字符串关闭） |
English
| Option | Default | Description |
|---|---|---|
appId | — | Feishu app ID |
appSecretRef | DSH_LARK_APP_SECRET | Name of the Secret reference in Harness Credentials |
domain | feishu | feishu or lark |
requireMention | true | Whether groups must @mention the bot |
dmMode | open | DM policy: open / allowlist / disabled |
groupAllowlist | [] | Allowed group chat_ids |
dmAllowlist | [] | Allowed user open_ids in allowlist mode |
provider / model | Harness default | Pin a model for the Feishu channel |
workspace | First registered workspace | Agent working directory |
agentPreset | Harness default preset | Agent preset |
reactEmoji | THUMBSUP | Reaction added on incoming messages (empty string disables it) |
飞书用户 / Feishu user
→ Lark SDK (WebSocket)
→ chatterbox4dsh
→ Harness Agent
→ 回复卡片 / reply cards
中文：插件运行在 DSH Host 内部，不启动额外进程，不暴露 HTTP 端点。每个飞书聊天映射一个 DSH Session，Agent 在 turn 完成后复用。详见 docs/architecture.md。
English: The plugin runs inside the DSH host, starts no extra process and exposes no HTTP endpoint. Each Feishu chat maps to one DSH Session, reused across turns. See docs/architecture.md.
npm install
npm run test
npm run build
中文：修改源码后需要 npm run build 并重启 DSH 进程（如 systemctl --user restart dsh，具体取决于部署方式）。注意重启会中断正在运行的会话/turn，建议在空闲时执行。
English: After changing the source, run npm run build and restart the DSH process (e.g. systemctl --user restart dsh, depending on your deployment). Note that a restart interrupts running sessions/turns — do it while idle.
中文
header.template 始终存在且正确（服务端卡片实体为 green/wathet/red），但某些飞书客户端在卡片被更新后不重绘标题背景，表现为白底。属飞书客户端渲染问题（同一条消息在不同设备上表现不同），不是插件 bug；切换会话或重启客户端通常可恢复。im.v1.message.patch，标题颜色变化在部分客户端同样可能不重绘（同上）。English
header.template (the server-side card entity is green/wathet/red), but some Feishu clients fail to repaint the header background after a card update, showing it as white. This is a Feishu client rendering issue (the same message looks different across devices), not a plugin bug; switching chats or restarting the client usually restores it.im.v1.message.patch, and the title color change may likewise not repaint on some clients (same cause as above).中文：基于 sugarforever/dsh-lark（ee639df）fork，已独立维护，不再跟踪上游同步。本仓库的改动记录见 CHANGELOG.md。
English: Forked from sugarforever/dsh-lark (ee639df) and maintained independently — upstream is no longer tracked. See CHANGELOG.md for this repository's changes.
MIT — Copyright (c) 2026 sugarforever (upstream), modified work by Starxer.
| 权限模式选择 | /permission：查看/切换会话权限（沙箱）模式，名称与显示名对齐 WebUI（Read Only / Workspace Write / Full access），交互式卡片点选切换 |
| 报错带具体原因 | 出错回报带具体原因与错误码（errorText），不再只回统一道歉文案 |
| WebSocket 长连接 | 无需公网服务器，支持飞书中国版和国际版 Lark |
| 访问控制 | 群聊白名单、单聊白名单、@机器人 要求 |
The feishu_send_file model tool pushes a workspace file into the current Feishu chat (≤30 MB) |
| Agent receives files | The feishu_receive_file model tool downloads an inbound Feishu file into the native DSH attachment store on demand and returns its fileHostPath |
| Steer while running | /steer <text> injects one message into the current turn; /busy steer makes "send while running" mean inject by default (persisted); when the agent is idle, /steer and /queue fall back to a normal new message. A plain message sent while running gets a plain-text notice first (injected / queued), and the injected message does not get its own reply card — the running turn's reply answers it |
Queue while running (/queue) | /queue <text> forces the queue path (even in steer mode) and runs as a new turn; falls back to a normal message when idle |
| Permission mode | /permission views/switches the session permission (sandbox) mode with an interactive card; names match the Web UI (Read Only / Workspace Write / Full access) |
| Errors explain themselves | Failures come back with the concrete cause and error code (errorText) instead of a generic apology |
| WebSocket connection | No public server required; works with both Feishu (China) and Lark (international) |
| Access control | Group allowlist, DM allowlist, mention requirement |