Learn

How thumbs, comments, and the miner become suggested rules.

Learn is not part of the review path. A review can write an architecture draft. It does not invent rules from findings.

gegenlesen harvest is a separate first-run pass over a packed tree. It scans lint configs as suppressions, the miner drafts cited conventions, the suggestion judge drops most of them, and Ledger gets disabled drafts tagged source: harvest. If the miner or suggestion judge fails or times out, harvest does not write the ordinary inbox; drafts go to needs_rejudge (job harvest_judge_failed). Send those to the inbox from Ledger, or dismiss. Re-running harvest retries the judge. A later judged harvest dismisses leftover needs_rejudge harvest drafts that the retry did not keep. Context notes are capped at 2000 characters. Whole README.md / docs/*.md dumps are dropped. Nothing auto-enables. Harvest finishes after persist; it does not walk the whole tree for embeddings.

When it runs

• Button on the job page (POST /api/jobs/:id/learn)
• gegenlesen has no separate learn CLI yet. Use the API or the button.
• LearnSweepJob on limits.learn_interval_minutes (default 0 / off). Set a positive interval (or GEGENLESEN_LEARN_INTERVAL_MINUTES) to run a tick that often. Thumbs and merge-intent are eligibility only; they do not start a miner by themselves.

What becomes a candidate

From the job, findings you 👍 or marked should_be_rule, including judge-dropped ones. Endorsement is learn eligibility only — dropped findings stay out of the kept inbox. Suggestion judge still default-drops.

A rule lands in the inbox only after ≥2 distinct jobs endorsed the same normalized title. Context notes can land after one job (still after the suggestion judge). Nothing auto-enables.

Dismissing a learning stores an optional reason (duplicate, already_covered, too_specific, not_a_rule, other) plus a comment. That title-hash stays out of the inbox until you restore the dismiss.

A job-level merge-intent label (would you have merged unread?) is also learn eligibility, same as thumbs. It does not start a miner until a schedule tick or the Learn button. Would-merge is a positive exemplar for that class of diff. Would-not treats kept errors as mine-worthy even without thumbs. Auto-approve then “no” is the strongest would-not. The label never auto-drops a finding and never enables a rule.

Pipeline

1. Host filter (this-job endorsement, or would-not-merge kept errors)
2. Miner uses the mine or learn engine from Setup (same ACP engines as reviewers/judge, or OpenCode). Harvest and corpus use the mine slot. Job Learn and the architecture card use the learn slot.
3. Suggestion judge uses that same slot and defaults to drop. A keep may include rewrite.title and rewrite.body.
4. Kept rules upsert as disabled only after two distinct-job endorsements. Context notes may inbox after one job.

Nothing auto-enables.

The original miner title and body stay in the learning payload as original_title / original_body when the suggestion judge rewrites.