Curly Nguyễn

dream

the night learner

the guru bank is a living system: real agent transcripts show what actually worked, and every guru's learned layer should carry that, not just curated upstream docs. dream reads what the agents did, mines recurring lessons, gates every edit against real evidence, and stages adoption -- an agent self-report is never fact, so staged proposals need human review and a separately authorized adoption.

harvests lessons while idlekeeps what was provenforgets the noise

Use when running the sleep cycle, harvesting agent transcripts for lessons, self-improving the guru bank, or reviewing and adopting staged trained-layer edits -- gsleep sleep, gsleep adopt <cycle>, gsleep status. This is the living-skill sleep cycle.

methodology

  1. run the cycle through the orchestrator, never by hand-editing a guru's trained files: node references/scripts/gsleep.mjs sleep (add --harvest-only or --mine-only during the calibration window).
  2. harvesting is read-only over transcripts; the harvest guide owns the extraction, attribution, and redaction rules. every candidate edit must cite real task ids from the cycle's task database.
  3. nothing reaches a guru without the gate -- replay tier for gurus with checkable outcomes, rubric tier otherwise. the gate guide owns the budgets and the rubric text.
  4. gate-passing edits are staged under the cycle's report page. gsleep adopt <cycle> applies staged replacements, appends changelog lines, and attempts to rebuild skillsearch. it does not create a Git commit or enforce review; check the proposal and current authority before invoking it.
  5. inspect state any time with gsleep status -- tasks, held-out split, and edit verdicts per guru.
  6. inspect scheduling separately; the gsleep CLI does not install or prove a nightly job.

contents

search pages

go to any page