lesson-harvest

Run scripts/lesson-harvest.py with --queue to list unprocessed sessions, read each with vendor-insights.py cat-session, extract candidate lessons, route them to existing owners, then mark the session with --mark and the lesson list.

Use the --vendor claude flag on the queue and cat-session commands to enumerate and read Claude transcripts. Sessions are processed newest first in tranches of ten, and each is stamped processed in censor/lesson-harvest.jsonl.

Lessons route only to existing termini: repeatable corrections become precedents in censor/precedents.jsonl, human-only actions become levers in his-hand-levers.json, behavioral knowledge becomes memory files indexed in MEMORY.md, and buildable work becomes board tasks via the conduct broker.

No. The skill forbids inventing new ledgers or registries; a lesson binds only when it names an existing registry path, predicate script, lever id, or precedent id. If nothing fits, the lesson is not recorded as a new destination.

Mark it with --none-found, since processed is a state rather than a success claim. Marking is idempotent, so re-marking an already processed session is a no-op.

Only vendors with a working enumerator appear in the queue; others are printed as excluded. Their backlog is unknown rather than zero, and exclusions must be carried verbatim into the final report.