dsh-im-hub

English | 简体中文

A multi-platform IM gateway plugin for DeepSeek Harness (dsh): connect your dsh agent to Feishu (Lark), WeCom (WeChat Work), and Telegram, and chat with it from the messaging apps you already use.

One agent per chat. Multi-turn context. Whitelist access control. Idle reaping. No public endpoint required (Feishu long connection / Telegram long polling; WeCom uses an HTTP callback).

Highlights

• Feishu (Lark) — official WebSocket long connection (/callback/ws/endpoint + protobuf frames, client-driven ping keepalive, 3s event ACK, message-id dedup), or webhook mode. No public URL needed for the default mode. The same adapter also serves the international Lark edition (open.larksuite.com) via adapters.lark.
• WeCom (WeChat Work) — app message callback with full WXBizMsgCrypt AES-256-CBC decryption and SHA1 signature verification, plus active replies via the messaging API.
• Telegram — Bot API long polling (getUpdates), automatic message splitting at the 4096-char limit.
• Mock adapter — stdin + local HTTP endpoint for testing without any real platform credentials.
• Per-chat agent sessions: conversation context is retained, each chat is serialized (no interleaved turns), and idle agents are disposed after a configurable timeout.
• Whitelist enforcement (allowedUserIds) on every adapter; empty means everyone — set it in production.
• Slack-style commands: /help, /reset, /status, /model.

Install

# From npm (recommended):
dsh plugin --profile im add dsh-im-hub

# From a Git repo or a local checkout (development):
dsh plugin --profile im add link:D:/projects/dsh-im-hub

This creates a headless profile im whose bundles are @deepseek-ai/dsh-base + dsh-im-hub. Boot it with:

dsh --profile im

Configure

Option A — Web GUI settings card (v0.2.0+)

When the plugin runs inside the Web GUI (dsh web profile), a visual configuration card appears at Settings → Plugins → Configurable plugins → IM Gateway (dsh-im-hub). It edits the same configuration live — no cordis.patch.yml hand-editing, no restart needed (the bridge hot-reloads on save). Credential fields are stored server-side and shown as write-only "configured / not set" badges.

Guided configuration

Every control states its purpose and every empty input includes a safe example. Credential values remain write-only: the screenshots below use placeholders, not real tokens or secrets.

Telegram: Bot Token and access whitelist

Feishu: App ID, App Secret, and Open ID whitelist

WeCom: enterprise ID, application credentials, and callback Token

Option B — YAML (cordis.patch.yml)

The plugin row is disabled by default. Enable it from the profile's own cordis.patch.yml ($DSH_HOME/profiles/im/cordis.patch.yml):

- id: dsh-im-hub
  disabled: false
  config:
    adapters:
      telegram:
        enabled: true
        token: '123456:ABC-DEF...'
        allowedUserIds: [123456789]        # numeric Telegram user ids; empty = everyone

Full configuration reference:

Feishu / Lark prerequisites

• Create a custom app in the Feishu Open Platform (or the Lark Open Platform for the international edition), enable the im.message.receive_v1 event subscription, and grant the message permissions (im:message:send_as_bot, im:message:p2p_msg, im:message:group_msg / group_at_msg).
• Long-connection mode is available to self-built (enterprise) apps. In the admin console choose Event subscription → 使用长连接接收事件 (long connection) or configure the webhook URL for webhook mode.

WeCom prerequisites

• In the WeCom admin console create an app, configure 接收消息服务器 (callback server) with the URL https://your-public-host/wecom, a random Token and a 43-char EncodingAESKey, and copy them into the config.
• The callback server needs a public HTTPS URL (or a tunnel) — WeCom does not offer a long-connection mode.

Commands

How it works

IM platform ──(adapter)──► Bridge ──► ctx.agents.create({ sessionId })
   ▲                          │                │
   └──── reply text ◄─────────┴── session/event listener ◄── agent turn

• Each platform:chatId maps to one agent session (like @deepseek-ai/dsh-headless, but kept alive per chat).
• Inbound IM messages enter the session with source.kind = 'plugin' / form = 'relay' (community norm), and outbound text is read back from session/event (assistant/message, aggregated per turn, then chunked to maxMessageLength).
• Turns per chat are serialized through a busy-promise chain; a message queue prevents interleaving.
• Idle agents are disposed after agent.idleTimeoutMs and re-created on the next message.

Security notes

• Force a whitelist. Set allowedUserIds on every enabled adapter before exposing the bot publicly. An empty list means anyone can drive your agent — which can execute tools on the host.
• IM messages are injected into the agent session as plugin-originated user messages; they do not bypass the deployment's own approval/guardrail policy — treat them like any other user input.
• Platform secrets (token, appSecret, encodingAesKey) live in the profile's cordis.patch.yml; keep that file private.

Development

node --test test/                          # unit tests (protobuf frame codec)
dsh plugin --profile im add link:D:/projects/dsh-im-hub   # install from checkout
dsh --profile im --patch test/disable-skin.overlay.yml         # boot with mock adapter
# POST a message: curl -X POST http://127.0.0.1:9099/mock -H 'content-type: application/json' -d '{"text":"hi","chatId":"test"}'

Tip: if the dsh-skin manager ($DSH_HOME/cordis.patch.yml) inserted a UI-skin row that your headless profile cannot resolve, disable it via a --patch overlay (see test/disable-skin.overlay.yml) — the home layer outranks the profile layer, so an overlay is the reliable place to turn it off.

License

MIT