loop-distill

Loop Distill

A project grows loops; each loop's memory is local to itself and knows nothing of its siblings. This is the reconciler: one pass over every self-learning-loop/*/ and meta-training/*/ (meta-trainer run) in a project, yielding one ledger, a health audit of the loops, and a small set of gated promotions outward. "Loop" below covers both unless a rule names meta-training explicitly. It is read-mostly, and it is never a gate — nothing found here can fail a loop.

Hierarchy

GLOBAL   global-learnings.md  +  mindgap
         ~/.claude/skills/**                        (tier-2, gated)
         mindgap-plugin/skills/**              (tier-2, gated)
   |
PROJECT  <project>/CLAUDE.md            <- one marked block, gated diff
         <project>/.claude/PROJECT-LEARNINGS.md     <- ledger, authoritative
   |
LOOP     self-learning-loop/<name>/{GOAL,STATE,STRATEGY}.md
         meta-training/<name>/{GOAL.md,STATE.json + STATE.md render}
         <project>/.claude/skills/<loop>-lessons/   (optional)

Each layer reads the one above it: a loop reads the ledger at RESUME, the ledger reads global learnings, and every session in the project reads the marked CLAUDE.md block. This skill owns the PROJECT layer — it rewrites the ledger outright, proposes the CLAUDE.md block, and touches the layers above and below it only under the phase-5 gates.

Phase 0 · discover & lock

Locate, in this order: the project root; every self-learning-loop/*/STATE.md AND every meta-training/*/STATE.json beneath it; the ledger at <project>/.claude/PROJECT-LEARNINGS.md; the project CLAUDE.md; <project>/.claude/skills/*-lessons/.

A meta-training run's STATE.md is a RENDER of STATE.json (meta_loop's state of record) — read both, but this skill NEVER writes either: no stuck reset, no ## Consult seeding, no memory-section edit. Every fix on a meta-training run is proposed in ledger §7, and the field mapping for the audit checks is in references/audit-checks.md § Meta-training runs.