reality-review-weeklylisted

Layered eval architecture, Layer 4 weekly review. Cron-fired weekly. Queries reality_feedback.sqlite for entries where outcome IS NULL AND timestamp + window has elapsed. Surfaces N due entries via a chat message for operator labeling. Outcomes feed a Layer 1 classifier's training set over months. Decision gate at month 1 - >=50% labeling rate then continue Layer 4; else DROP.
0xUrsanomics/utopia-os · ★ 2 · AI & Automation · score 76

Install: claude install-skill 0xUrsanomics/utopia-os

# Reality-feedback weekly review skill Layer 4's labeling cadence. Once per week, surfaces high-stakes outputs whose outcome window has elapsed but the operator hasn't yet labeled. The discipline is: a 5-minute weekly review session, mark outcomes via reply, accumulate labeled training data over months. ## Conversation context (prior) **Prior conversation: N/A**. cron-fired, no prior turn context. Operator interaction happens via `/reality-review` replies AFTER the skill surfaces due entries. ## Output format The skill produces a chat message with bulleted due-entries (one per line: timestamp + skill + output summary + days-elapsed), each with an inline label-reply suggestion. Format: ``` 🔍 reality-review (5 due, week of YYYY-MM-DD): 1. {skill} ({N}d ago) "{output summary 80 chars}" reply: /label 1 win | loss | neutral 2. ... ``` Updates `reality_feedback.sqlite` table `outcomes` only when the operator replies with `/label N <outcome>`. this skill itself does NOT auto-mark anything. ## When to fire **Auto-fired** by cron task `reality-review-weekly` (weekly). Cron config: - `cron_expr`: `0 9 * * 0` - `task_type`: `claude_code` - `task_config.prompt`: "Run skills/reality-review-weekly/SKILL.md. Surface due entries to the inbox channel." **Manual invocation**: operator can run `python3 scripts/pipeline/reality_review_weekly.py --digest` anytime to see the pending review queue. ## Weekly flow 1. **Cron fires** weekly 2. **Skill runs** `scripts/pipeline/reali