dsh-pause

Once DeepSeek Harness starts streaming at 100+ tok/s, there is simply no way to type a correction in time. Hard-interrupting the output stream, on the other hand, visibly deflates the model's willingness to keep executing — the next round arrives timid and reluctant. dsh-pause solves both problems by parking the agent in the narrow gap between two tool-call steps. The outbound request has not been sent yet, so nothing in-flight is disturbed. You finish a note in the composer and press Enter to interject, or hit bare Enter to resume silently. The model sees a perfectly continuous context with zero extra messages; it never knows a pause happened.

Core capabilities include a per-session toggle on the composer tool row, a composer that stays fully editable throughout the pause, and two release paths: bare Enter (empty draft = seamless resume, typed draft = interjection sent with the request) or the pause button click. After releasing, the draft is cleared and the loop continues as if nothing stopped. A global default can be set in config and overridden at runtime per session.

The plugin hooks into agent/pre-step, intercepting only requests that follow a tool result — the very first text-only request in a turn is never paused. The browser-side control renders through the conversation.input.right slot, and the host side self-registers a route that reuses the official trust and authentication pipeline. No DSH source files are modified.

It is built for anyone running agent workflows who needs to steer the model mid-run without racing its output speed, without breaking its flow, and without the awkwardness of a visible interruption.