dsh-skillopt

Agent skill quality rarely improves on its own. You write a prompt, use it for a while, watch it degrade, then hand-tune again. dsh-skillopt closes that loop with a nightly "sleep cycle": the agent reviews your past sessions offline, mines the recurring tasks you actually performed, replays them on your own API budget, and feeds the resulting changes through a held-out gate. Only edits that pass the regression check are promoted into validated, reusable skills; the rest stay staged and never touch live files.

Seven native tools give you full control over the cycle—at a glance (status), in dry-run preview (dry-run), end-to-end execution that stages a proposal (run), the single live-change boundary with automatic backup (adopt), read-only task export (harvest), and cron install/remove (schedule / unschedule). The data boundary is deliberately tight: harvest is read-only, mock and handoff backends make zero network calls, and real backends accept only truncated transcript excerpts and derived tasks that have been explicitly marked reviewed.

If you run DeepSeek Harness for daily development or research and are tired of manually tweaking prompts every time a skill starts to drift, this plugin turns "gets better the more you use it" into an auditable, reversible overnight routine.