review-retrolisted

# Wingman: Review Retro Perform a retrospective analysis of accumulated review data. Identify what's working, what's not, and refine the feedback loop. Run this weekly or when review findings feel repetitive. ## Step 1: Load review data Read all `.json` files in `.reviews/` with `"status": "categorized"`. Also read the current `.claude/rules/review-patterns.md`. If fewer than 3 categorized reviews exist, tell the user there isn't enough data yet and suggest running more `/review:loop` cycles first. ### Schema-aware reading Files written by wingman v2 (`wingman_schema_version: "2"`) carry a top-level `reviewer` block with `tool_version`, `model`, `provider`, `reasoning_effort`, `session_id`, and `wall_seconds`. Use these when analysing trends: - Group findings by `reviewer.tool` (codex/gemini/claude) and by `reviewer.model` to compare findings produced by different reviewers and models (e.g. did gemini surface different patterns than codex?). - Plot `reviewer.wall_seconds` over time to spot reviewer performance regressions or model-tier upgrades. - Group by `reviewer.tool_version` when interpreting findings — a tool upgrade can change which categories surface. Older v1 files (no `wingman_schema_version`) lack these fields. Either run `python3 scripts/migrate-reviews.py` from the wingman repo to backfill (best-effort extraction from `raw_review` prose), or treat v1 files as `model: "unknown"` for trend analysis. ## Step 2: Frequency analysis Identify the top