dsh-feishu

dsh (DeepSeek Harness) excels at driving long, multi-step AI agent tasks on your desktop, but the moment you close the laptop that thinking, tool-calling, approval-waiting session loses both your eyes and your hands — you cannot go back, and it may stall. dsh-feishu bridges the entire interaction chain into Feishu or Lark over an outbound-only WebSocket: no open ports, no tunnels, just your phone becoming a second console for the running session.

The experience centers on live round cards: every LLM round-trip produces a state card showing whether the agent is thinking, calling tools, or running a subagent, with the in-flight message tail refreshed every five seconds. When the host approval waterfall requests a sandbox escalation, the phone surfaces an allow-once or deny card so an unattended run never stalls at the desk; when the agent invokes ask-user-question, a dropdown, multi-select, or free-text card captures your answer and pushes it straight back into the session. Beyond DMs, the bot joins Feishu groups for @-mention dispatch, accepts images in DM and injects them as image blocks, renders /model, /think, /permission, /select-skill, and /profile-switch as one-tap pickers, supports /btw to fire a side question alongside the running task without disturbing the main line, and offers an interactive /resume picker with one-tap corrupt-log repair.

It is built for developers who run extended dsh agent workflows on the desktop but want to monitor, steer, and approve from anywhere — especially teams that coordinate in Feishu groups and need mobile approvals or multi-surface prompting (pair with ask-router so the first response wins across desktop and phone). With a single published Feishu enterprise app and an open_id allowlist, it turns the session only advances when I am at the desk problem into a pocket-sized session remote.