rh-quitlisted
Install: claude install-skill toolbeltross/rh-claude-framework
# rh-quit
User-invoked scribe curation pass. The Stop hook's inline regex extraction (`rh-scribe-prefilter.js`) captures low-fidelity rows to `recommendations.md` and `cleanup.md` on every turn. This skill is the **high-quality LLM path**: it dispatches `rh-scribe-multiscope` directly to curate, dedup, and write richer entries — especially learnings, which inline extraction skips entirely.
**Architecture (rev 2026-05-08, P1-4 follow-up):** the skill dispatches `rh-scribe-multiscope` DIRECTLY via the Agent tool. The supervisor agent is no longer in the path for `/rh-quit`. Prior architecture (skill → supervisor → multiscope) added 400+s of supervisor latency with no value-add — the multiscope agent already handles privacy, sentinel, and triage internally. The supervisor's failure-pattern / advisory roles are unrelated to scribe drain and don't belong on this critical path.
## What this skill does
1. **Read the full session text** — per-turn staging is on by default (P1-3). The multiscope agent should consume the staging file via `node ~/.claude/scripts/rh-scribe-staging-read.js <session-id>` for full-session coverage. Fall back to the 10K-char transcript tail only if staging is explicitly disabled (`oversight.json: scribeStaging:false` or `RH_SCRIBE_STAGING=0`) or no staging file exists for this session.
2. **Detect scribe-worthy content** — apply the same marker regexes from `rh-scribe-prefilter.js` (recommendations, cleanup, learnings) to decide whether to dispatch at a