dsh-timer-task

When running unattended long jobs in DeepSeek Harness, an agent often needs a follow-up action that fires at a specific time-for example, shutting down a cloud build instance after a nightly compile finishes. A bare cron job is unsafe: the agent may still be working, and forced execution can interrupt the workflow or lose results. dsh-timer-task solves this by letting the agent schedule its own recurring tasks, with a veto checkpoint before every firing so the command only runs when nobody objects.

Three model-facing tools power the plugin: timer_schedule creates a task, timer_cancel removes it, and timer_veto blocks a single firing. Tasks support three lifecycles-one-shot, untilDone, and repeat-covering retry and polling patterns. The veto window is dynamic: as long as any agent is still running, the window extends automatically; once all go idle, a fixed grace period applies. Tasks persist to disk across dsh restarts, and a result notification is injected into the owning session after each execution. Multiple veto channels are available: the primary agent can call the tool, a subagent can write a workspace marker file, and an external process can drop a file in the global veto directory.

It is built for developers and automation engineers who frequently orchestrate long-running background jobs in dsh and want the agent to decide for itself whether the moment is right, rather than being interrupted by a rigid timer. The plugin ships no built-in external calls-the command is entirely specified by the caller, keeping the security boundary explicit.