/harvest — session learning capture

Adapted from dwarves-kit harvest hook + vault lizard pipeline. Prevents tacit knowledge dying in the transcript.

Triggers

• /harvest — manual end-of-session
• /harvest promote — curate staging → lizard drafts (uses harvest-curator)
• Staging file written by stop hook (.cursor/hooks/harvest-stager.sh)

Phase 1 — Collect (automatic or manual)

Scan the current session for:

1. Corrections you made ("no, actually X", rejected deliverables)
2. New stable facts (IDs, URLs, decisions) not yet in 05-knowledge
3. Process lessons (what worked, what failed, friction)
4. Skill gaps (something that should be a Verify step or CLAUDE.md rule)

Append to 04-projects/harness/harvest/staging-<YYYY-MM-DD>.md:

## <HH:MM> — <trigger>
- type: correction|fact|process|skill-gap
- source: <file or "session">
- text: <one paragraph max>
- proposed_home: <path or "new lizard note">

Dedup: skip if same fact already in staging today or in 05-knowledge.

Phase 2 — Curate (/harvest promote)

Spawn harvest-curator (Sonnet). It writes /tmp/harvest-curate-<date>.md.

Lead (Opus) presents to the user:

• Promote list → approve which become 05-knowledge/lizard/YYYY-MM-DD-<slug>.md
• Fold list → approve inline edits to existing notes
• Skill patches → approve CLAUDE.md / SKILL.md diffs

Only after approval: write durable files. Update 05-knowledge/lizard/index.md.

Phase 3 — Retro line

One line to 01-daily/journal/<today>.md:

Harvest: <n> staged, <m> promoted, <k> folded

Rules

• Harvest is advisory. It never auto-writes 05-knowledge.
• Contradictions with existing knowledge → flag both, ask the user.
• Pair with /memory-hygiene monthly for environment-dependent facts in ~/.claude/.../memory/.

Nightly enhance (optional)

If you want harvest to run unattended, schedule a job that invokes the skills in this order. There is nothing to install first, since COG ships no hooks.

/memory-hygiene
/harvest promote     # only if the staging file is non-empty
/content-factory     # only if scheduled