dream trained

Harvest signals are leads, not proof. Mine candidates, gate them, stage a proposal, review the evidence, then adopt only under current authority.

Run these from the guru folder; dry mode uses fixture transcripts and model commands by default and writes under a temporary folder.

• Inspect the staged diff and task evidence before adoption.
• Keep training limited to the gurus' learned-layer files.
• Check skillsearch check after adoption; index rebuild failure is logged but currently non-fatal.

Fixture-backed phases and a staged proposal allow bounded inspection before any learned-layer mutation.

Transcript outcome markers are heuristic. The rubric is a model judgment, and even a passing gate is not proof of product behavior.

• gsleep adopt does not commit or enforce human review.
• gsleep status initializes its database if missing.
• Dry mode uses fixture transcript sources unless the environment's transcript-directory overrides are set.

This is a local-only procedure. The sleep orchestrator, its attribution module, and its adoption module own the behavior; the observed tests are cited below. No internet source establishes its runtime contract. Current user authority governs adoption even though the CLI can apply staged edits without a review check.

No remaining version-specific defect is verified. Attribution includes flattened guide files alongside directory-based guides.

• On Node v24.14.1 with experimental SQLite, node --test over the sleep and library tests passes 45/45. A fixture-only dry sleep scans 5 files, extracts 7 tasks, and writes a proposal under a temporary folder.
• A real transcript cycle, model gate, and adoption remain unexercised; the measured run covers the fixture path only.