lesson-harvest

The operator's law (memory feedback-codify-every-session-lesson): every discussion carries lessons, and a discussion not processed into a durable owner was wasted. scripts/ lesson-harvest.py owns everything deterministic (queue, cursor, backlog); this skill owns only judgment — reading a session and deciding what it taught, where that lesson binds, and whether it is already codified.

Iron rules:

• Route ONLY to existing termini. Never invent a new ledger, doc, or registry for a lesson (PREC-2026-07-30-plan-decisions-dont-bind: a lesson binds only when it names a registry path, predicate script, lever id, or precedent id).
• PII firewall: owner refs and one-line summaries only. No verbatim personal content, no third-party names, no secrets, no /Users/... paths in anything tracked. Person-scoped material (people reviews, relationship content) is NOT a lesson-harvest output — it belongs to the _people-private review lane (ARCA-sealed), and in a background session its writes are classifier-gated; cite existing and move on.
• Bounded reads only: vendor-insights.py cat-session excerpts, never whole-file loads.
• Tier the fan-out: session reading is scan-class work — use cheap-tier subagents for the first pass; reserve the session model for judgment on candidate lessons.

Steps

1. Queue. python3 scripts/lesson-harvest.py --queue [--vendor <vendor>] [--limit N] → TSV: vendor, session id, project label, day, bytes. Newest first. Note the excluded vendors and the coverage: lines it prints — carry both into your report verbatim (backlog unknown ≠ zero; every non-Claude lane is listed from a rolling-window index, so a session older than that window is not in the queue).
2. Read. For each session, with the vendor from its queue row: python3 scripts/vendor-insights.py cat-session --vendor <vendor> --session <sid> (raise --max-chars only when the excerpt truncates mid-incident). Look for: a correction from the operator; a workaround that was narrated but not codified; a wrong confident claim and what refuted it; a gate/tool/method fought twice; a decision that binds future sessions. Skip pleasantries and routine execution — a lesson is something the NEXT session would otherwise get wrong.
3. Dedupe against the owners first. Before writing anything, check whether the lesson already lives somewhere: grep censor/precedents.jsonl, his-hand-levers.json, the memory index (.agent-runtime/claude/projects/<project>/memory/MEMORY.md), and CLAUDE.md. Already owned → record {"owner_kind":"existing","owner_ref":"<the id/path>"} and write nothing new.
4. Route each genuinely new lesson by the cascade:
   • Repeatable correction with a binding site → precedent: append one JSON line to censor/precedents.jsonl matching the live schema (id: PREC-YYYY-MM-DD-<slug>, ts, type, subject, outcome, reversible, action naming WHERE it binds, authorised_by (the file's uniform spelling — all 36 records), review). Any script it cites must exist (check-runner-coverage.py E).
   • Irreducible human atom → lever in his-hand-levers.json (id L-…, label with the concrete ~minutes action, owner, cost, unlocks, source_task, issue).
   • Behavioral/project knowledge for future sessions → memory file (frontmatter name/description/metadata.type) + one pointer line in MEMORY.md (evocator.py-verifiable).
   • Buildable work → board task through the conduct broker (limen conduct … / TABVLARIVS) — never by editing tasks.yaml.
5. Mark. python3 scripts/lesson-harvest.py --mark <sid> --vendor <vendor> --lessons '[{"owner_kind":"precedent","owner_ref":"PREC-…"}, …]' — or --none-found when a session genuinely taught nothing (processed is a state, not a success claim). Marking is idempotent; re-marking is a no-op.
6. Report. End with python3 scripts/lesson-harvest.py --check output: per-vendor remaining/processed, coverage windows, exclusions, TOTAL. The backlog number is the deliverable — never summarize a tranche as if it drained the corpus.

Tranche discipline

Default tranche: 10 sessions, newest first (recency = highest un-codified density; the oldest sessions' lessons mostly got codified since). The daily lesson-harvest-backlog sensor keeps the remaining count on the beat, so stopping mid-backlog is safe — the state is in the cursor, not in anyone's memory.