rekol-bootstraplisted

# rekol-bootstrap The cold-start memory bootstrap. Indexing makes your Claude Code history *searchable*; this routine *distils* it — turning recurring instructions, preferences, and decisions buried in past transcripts into ambient, curated REKOL memory. **You** (the user's own Claude) are the precision pass: the engine recalls and batches candidates with no LLM; you cluster, classify, fix frontmatter, and present an approve/edit/skip gate before anything is captured. This is the LLM-precision half of the onboarding flow. The recall half (T2) already ran inside `rekol bootstrap` — it reused transcript search to surface candidate instruction-bearing messages, deduped them against existing memory, and narrowed to a bounded, reviewable shortlist. ## Non-negotiables - **Review-gated.** You NEVER auto-write memory. Every captured item passes a human approve/edit/skip gate. `rekol bootstrap` itself never captures — it only recalls, scopes, batches, and checkpoints. The *only* command that writes memory is `rekol capture`, and you run it only on an approved item. - **Conservative.** Capture only high-confidence, recurring, durable facts — "would a future session genuinely benefit from knowing this?" Skip trivia, one-offs, and anything you can't verify against its cited transcript. - **Provenance.** Each candidate carries its source (session id, transcript file, line). Open the source if a candidate is borderline; don't trust recall blindly. - **Resumable.** The run i