v0.3.0

Since 0.2.0.

Review-history analysis (the mine feature is now analyze)

• The feature that turns your PR-review history into reusable pitfalls is renamed end to end: CLI plex analyze, MCP analyze_scan / analyze_history, backed by a dedicated @plex/distill pipeline. Clearer intent, no data-extraction connotation.

Knowledge is learned-only

• Removed plex.md, markdown seeding, and the entire knowledge→rule promotion / external-runner surface (ADR-37). Knowledge now comes only from analyzing your PR history and your review verdicts — one provenance story, no hand-authored files, no wasted per-review tokens. (A committed plex.json policy file is the planned successor.)

Engine & ranking

• Key-less installs still get their accumulated pitfalls: retrieval falls back to lexical (IDF) ranking when no embedding provider is set.
• Outcome-weighted consolidation (a reverted change counts as the strongest confirmation) and accept→pitfall inference with learning idempotency (no double-counted evidence).
• Deletion blast radius, cross-window co-change accumulation, generated-artifact exclusion (lockfiles/bundles are never read line-by-line), measured deterministic prevalence, and an auto-accept audit trail.
• Quality-floor benches (ranking nDCG + retrieval recall@k) guard against ranking/retrieval regressions.

Breaking changes

• MCP tools renamed: mine_scan → analyze_scan, mine_history → analyze_history. Removed: seed_knowledge, propose_promotions. Update any saved agent prompt that references the old names.
• CLI: plex mine → plex analyze; plex seed removed; plex promote → plex consolidate.
• Requires Node ≥ 22.22.

Install: npm i -g @sshanzel/plex, or run ad-hoc with npx @sshanzel/plex <cmd>. The plugin tracks latest.