ViktorsBaikers /rite-learn: review durable lessons
Use reviewed Markdown/native memory, not miners or registries. This proposes maintenance of one authority; it cannot promote a rule alone.
Modes
/rite-learnreviews recurring evidence across shipped work./rite-learn "<lesson>"evaluates one lesson or rejected direction.
Workflow
- Read durable promotion;
bound the archive; inspect applicable
AGENTS.md/CLAUDE.md, accepted ADRs, and relevant.devrites/archive/*/{decisions,drift,review,seal}.md. - Broad mode dispatches exact fresh/read-only
devrites-retrospector; reconcile its claims against cited files. - Keep corrections repeated in two features, a judgement call made twice, or a defect class seen twice; drop one-off preferences, task-specific detail, generic advice.
- Verify claims against live authoritative sources; state currentness signal, applies/does-not-apply scope,
unknownwhere unverifiable. Research promotion requires: each external claim carries a dated citation pertooling.md§ Research provenance. Failing case: "best practice is X" with no source → reject promotion. - Apply durable promotion's recoverability test; search same/contrary rules, choose one existing owner (nearest instruction/standard, ADR, or feature
decisions.md), and name discovery. - Show the exact edit + duplicate/conflict/supersession disposition; update/narrow/replace/retire contradictions; apply only after user approval of exact edits.
Rules
- Live repository evidence outranks memory; unverifiable is unknown, not false.
- A lesson's premise is graded established / working / open with its source (retrospector output): only established premises — or working premises whose confirming check is named — may be proposed; open premises return as recorded assumptions, not lessons.
- Never create a learning ledger/index/queue, score, timeline, or parallel authority; rejected directions return only when evidence changes their rationale.
- A declined lesson persists as a declined decision entry (reason recorded) in the nearest owning decisions file in the same round — a refusal recorded only in conversation or on an unmerged branch is lost and will be re-litigated. Failing case: the same rejected proposal returns next round because no tracked entry exists.
- A lesson proposing a new recurring check names its controls: the real past instance it would have caught (positive) and a near-miss class it must not fire on (negative). Failing case: a trigger patterned on an imagined command that never matched real history.
- Contradiction outranks staleness: actively misleading guidance outranks merely old guidance.
- A proposal names the retrospective failing case: the concrete past feature/artifact the rule would have caught. None → generic advice — drop.
- ≤3 accepted lessons per round; proposals extend/narrow but never lower an existing bar (revisions show old text beside new); duplicates consolidate into one canonical edit — simplification (deletions/merges) counts toward the cap.
- An accepted lesson ships with a follow-through owner and deadline: the exact edit lands in the named canonical file in the same round; a lesson unapplied at round end returns to candidates with its blocker recorded. Failing case: an accepted lesson with no applied edit and no recorded blocker — promotion failed; re-raise it.
Output
Done: reviewed <scope>.
Changed: none (proposal only)
Candidate: <specific proposal or none>
Currentness: <live source + signal | unknown>
Scope: applies <trigger>; does not apply <boundary>
Authority: existing <path|none>; canonical <path>; consumers/discovery <route>
Disposition: <no conflict | update/narrow/replace/retire path + reason>
Evidence: <feature/file references>
Awaiting: <approval for exact edits | none>
Next: <single action>
Record: <nearest owning decisions file | chat-only>