skill-sentinellisted

# skill-sentinel The safety layer for growing your Claude Code setup. Every directory online ranks **popularity**; none tells you whether an extension will read your `.env`, or rug-pull after you trust it. skill-sentinel adds the missing signal: it finds the good ones **and vets them**, combining a deterministic static scan, public-reputation signals, and an LLM semantic judge that reads the extension's text surface the way a careful human reviewer would. `SKILL_DIR` below = the folder this file lives in. Run the CLI as `python3 SKILL_DIR/scripts/scout.py ...` (no `cd` needed). It is Python 3.9+ standard library only. Requires Python 3.9+ and git; the `gh` CLI is optional (it lifts GitHub rate limits) and the Task tool is optional (used for the semantic judge; there is an inline fallback). ## Golden rules (never break) These exist because the whole value of this skill is honest safety judgement. If you cut a corner here, the tool becomes worse than useless: it gives false confidence. 1. **Never promise "safe".** This is static analysis + reputation + a semantic text review, not a proof. Always give the mandatory caveat from `references/verdict-template.md`, and always separate "**what I scanned in the code**" (facts) from "**what is a reputation signal**" (inference about trust, not the code). 2. **Never install without explicit approval**, and **never offer install for band 4-5.** For those bands, state plainly that it is not recommended, and why, in one cle