- Auto-promotion removed.
count:3+now stages to.learnings/promotion-queue.mdinstead of auto-writing to instruction files. Human approval required before any promotion.
- Promotion queue (
.learnings/promotion-queue.md) — staged candidates with source, evidence, proposed rule, and approval status - Source labels (
source:agent,source:user,source:external) — injection defense; external entries blocked from promotion - Severity field (
severity:low|medium|high|critical) — critical entries trigger immediate review at count:1 - Prevention tracking (
prevented:N) — loop closure; tracks whether recalled learnings changed behavior - Structured IDs (
id:ERR-YYYYMMDD-NNN,id:LRN-YYYYMMDD-NNN) — stable dedup by ID, fallback to keyword grep - Detail files (
.learnings/details/) — optional linked markdown for complex failures - Expiry field (
expires:YYYY-MM-DD) — optional staleness control with quarterly revalidation - install.sh — creates all v2 directories and files (replaces setup.sh, which now delegates)
- review.sh v2 — pending promotions, stale entries, expired entries, source distribution, prevention stats
- references/ — design tradeoffs, detail template, promotion queue format
- README.md, LICENSE (MIT), CHANGELOG.md, .gitignore
- Author attribution: Don Zurbrick (was Zye)
- Comparison table moved to
references/design-tradeoffs.mdand reframed as design tradeoffs - Version bumped to 2.0.0
All new fields are optional. Existing v1 one-line entries ([date] CATEGORY | what | action | count:N) continue to work without modification.
- Initial release
- One-line format for errors, learnings, and wishes
- grep-based dedup
- Auto-promotion at count:3+
- Pre-flight review checklist
- setup.sh and review.sh scripts