What problem does it solve? Machines accumulate long session histories that are never synthesized into a searchable, citable narrative record. This Skill bootstraps that derived narrative estate in one idempotent run and catches up after any daily-digest lapse, without re-processing already-digested sessions. ## Core Features & Use Cases - Idempotent journal-led digest: An append-only journal keys every skip and watermark, so re-runs only process the missing diff and crashed runs resume exactly where they stopped. - Rolling worker pool: A synchronous pool of ephemeral codex exec workers digests sessions oldest-first, with chunking and merge for sessions over 150K tokens, quota-aware backoff, and a single retry per target. - Owner-controlled spend cap: The run fails closed until the owner annotates the agenda item with an exact NS CAP: $<amount> line, and stops refilling workers when cumulative estimated cost crosses the cap. - Use Case: After adopting the narrative-synthesis cycle on a machine with months of session history, run this once to digest everything into cited Markdown digests with JSON sidecars, then let the daily digest take over. ## Quick Start Approve the narrative-backfill agenda item, annotate it with a spend cap such as NS CAP: $60, and ask the agent to run the narrative backfill to digest this machine's session history.
