SkillOpt Sleep for the dsh agent explains how to run the sleep cycle through the skillopt tools the dsh-skillopt plugin exposes: status, dry run, run, adopt, harvest, schedule and unschedule.
It harvests past sessions, mines recurring tasks, replays them through a backend and stages skill proposals that passed a held-out validation gate. Live files change only when the adopt tool runs.
What You Get
- A table of the seven skillopt tools and what each does
- A step-by-step example flow with a mock default and a real backend run
- A parameter table covering project, backend, source, model, task cap and preferences
- A configuration snippet for the plugin and the list of advanced engine keys
Common Use Cases
- Have the agent learn recurring preferences across sessions and store them in skill documents
- Preview a cycle with the mock backend before using real model budget
- Pass house rules such as preferred test tooling through the preferences parameter
- Schedule a nightly run for a project and review the staged report later
What Makes It Unique
Auto-adopt is an operator-only setting in the plugin configuration that the model cannot turn on, so adoption through the skillopt_adopt tool is the one live-change boundary. It also supports a long list of backends and transcript sources, including pi and opencode.
Under The Hood
The skill is prose that tells the agent to call the skillopt tools, which run the Python skillopt_sleep engine. Harvesting scans supported local session records read-only. A real backend sends truncated transcript excerpts and derived tasks to its provider, so the skill describes exporting tasks, redacting them and marking them reviewed first. Runs write staging files in the project folder, the adopt tool copies staged files over live skill documents after a backup, and the schedule tool installs a nightly cron entry. The validation example installs the skillopt package with pip and runs a Python demo script.