dsh-todo-freshness-guard

The toughest moment in any long agent session is when it is clearly still working, yet the todo list on your screen is a frozen snapshot from an hour ago: finished items still show as pending, new blockers are absent, and the next step is pure guesswork. dsh-todo-freshness-guard exists to close exactly that gap.

The mechanism is deliberately simple. As soon as a todo_write call leaves unfinished items behind, the Guard starts counting non-bookkeeping tool attempts for the session. At the reminder threshold it injects a model-visible nudge asking the agent to reconcile the full list; at the block threshold it refuses ordinary tools entirely until a fresh, complete todo_write replaces it. todo_write and the outer run_code recovery path stay reachable at all times, so the agent is never dead-ended. Native calls and Code Mode SDK sub-dispatches share the same counter, and enforcement lifts automatically once the list is completed or emptied.

This is for developers who run multi-step, long-running tasks on DeepSeek Harness and need a lightweight safety net that does not change their workflow. Drop the Guard into a profile, set two positive integers, and every subsequent tool call happens on top of a plan the agent actually keeps up to date.