dsh-lark

@mention the bot in Feishu or Lark, drop a direct message, and the local DeepSeek Harness agent gets to work with replies landing right back in the same conversation. dsh-lark compresses that entire IM-to-agent loop into roughly 600 lines across four source files. No card pipelines, no pairing codes, no multi-project routing. A WebSocket long connection means it runs from a laptop with zero public callback URLs; each topic maps to its own agent session, and sessions resume cleanly after restarts.

Interaction centers on a single status card per task: it updates in place from running through completed, stopped, or error, showing one line per tool call, with a dsh Web button at the bottom. Sensitive actions pop an approval card with one-tap allow or reject buttons. Images sent in chat reach vision-capable models natively, and the agent can deliver files from its working directory back into the conversation, path-confined by design. Security defaults are strict, with an open_id allowlist on out of the box and unrecognized users receiving only a one-line ID hint.

If you want a Feishu gateway you can read end-to-end in ten minutes, dsh-lark is the leanest option. For progress cards, multi-project routing, or cross-platform support such as WeCom and Telegram, look at dsh-lark-bridge or dsh-im-hub instead.