SkillOpt Sleep for Cursor reads local Cursor agent transcripts, mines the tasks you repeat and proposes bounded edits to the project skill at .cursor/skills/skillopt-sleep-learned/SKILL.md.
A proposal is accepted only when it improves a held-out score with the default gate on. A normal run stages it for review, and nothing live changes until you adopt it. The plugin has no session-end hook and no MCP server, so it runs only when asked.
What You Get
- A fixed Cursor target path for learned skill edits, passed to harvest, dry-run and run
- Three command forms: a source checkout runner for macOS and Linux, one for Windows, and an installed skillopt-sleep command
- A mock backend for smoke checks and a cursor backend that uses your authenticated Cursor Agent CLI
- A reviewed-task workflow: export tasks, redact them, mark them reviewed, then replay with a real backend
Common Use Cases
- Teach a Cursor project skill from the work you repeat in Cursor agent sessions
- Export and review mined tasks before sending anything to a provider
- Schedule a nightly Cursor cycle after setting source, target and backend in the config file
- Check Cursor model names with the agent CLI and keep live runs small with session and task limits
What Makes It Unique
It is built around Cursor’s constraints. Backend calls run in an empty temporary workspace in read-only Ask mode, tool-aware replay is disabled and a task with a tool_called check fails before Agent mode starts, and the skill tells the agent not to claim that repository-dependent behaviour was validated. Real backends refuse a task file that is not marked reviewed.
Under The Hood
The skill is prose that tells the agent to run skillopt-sleep (installed) or the bundled runner scripts, which start the Python engine. It reads Cursor transcript files under the Cursor home folder, read-only, keeping message text and tool names and dropping tool arguments and outputs; secret-shaped strings are redacted but the skill says that is not a guarantee. The cursor backend calls the Cursor Agent CLI and other backends call their own providers, so task excerpts leave the machine. Runs write staging files into the project, adopt overwrites the target skill after a backup, and schedule installs a nightly cron job on Unix or a Task Scheduler entry on Windows.