dsh-loop-continue

In a DeepSeek Harness agent workflow there is a quiet failure mode: the model writes "now I will re-apply that change" and then stops. No tool call is ever emitted. The loop sees a step with text but no tool call, marks the turn as completed, and the half-finished work is silently closed. You are left with what reads like a confident narrative, but the actual job is only half done.

dsh-loop-continue is built for exactly this case. It hooks into the agent/turn-stopping event and applies two deterministic gates first — the turn must end on a text-only step, and at least one tool must have been called earlier in the turn — so ordinary one-shot answers are never touched. When both gates pass, the plugin hands a compact summary of the turn's session to a judge model and asks a single strict yes-or-no question: did the model narrate an action it still needs to execute? A yes steers the same turn to take one more step; a no, or an unparseable reply, lets the turn close normally. Each turn is capped at ten continuations, and if the model narrates again after a steer without issuing a real tool call, the guard stops trying rather than burning the remaining budget on the same refusal. The judge prompt and the steer message are plain config fields, editable from the settings card without any code change.

It is aimed at users who run multi-step tool-calling tasks inside Harness and regularly hit the "it said it would, but didn't" truncation. The plugin does not rewrite model behaviour; it simply glances at the turn one last moment before it shuts and pulls the unfinished work back into the loop.