dsh-restart-resume

Restarting a DeepSeek Harness backend for plugin updates, configuration changes, or bundle upgrades usually means killing the process outright, which severs active sessions, published sub-agents, and background jobs along with the user context. dsh-restart-resume introduces a safety protocol for this high-risk path: it waits for all in-flight work to drain, requests a supervised relaunch via exit code 75, persists the resume marker, and automatically reconnects the Web page so the user never has to refresh or re-issue a task.

The plugin registers three fixed tools, dsh_restart, dsh_restart_status, and dsh_restart_cancel, always exposed in the same order and schema, so the model-request tool prefix never shifts with state changes. Before the final countdown the plugin enters a drain-wait phase; if new work appears it reverts gracefully, and a drain timeout cancels the restart rather than force-killing sessions. Rate limiting, atomic markers, TTLs, and idempotent messages together prevent restart loops. The resumeMode setting distinguishes continue (resuming unfinished configuration, validation, and delivery in the original task) from report-only (isolated test runs), keeping test cycles from being mistaken for production resumes.

It is aimed at administrators or teams that already run Harness under a supervisor process. The launcher must mint a fresh boot ID before every spawn and only relaunch on exit code 75, with frequency limits enforced on both the plugin and the supervisor side. In that architecture dsh-restart-resume turns a dangerous hard-kill into an observable, cancellable, resumable standard flow, letting the Agent pick up unfinished work seamlessly instead of leaving the user staring at a blank page.