claude-automation-recommenderlisted

# Claude Automation Recommender The repo's automation layer has three failure modes, and only the third is obvious: 1. **A guard that never fires.** The script is present, complete, and executable — but nothing invokes it. It reads as active protection while protecting nothing. 2. **A guard that fires wrongly.** It runs on everything, or on the wrong thing, so its signal stops carrying information and reviewers learn to ignore it. 3. **A failure with no guard.** Something keeps breaking and nothing watches for it. Audit in that order. A dead guard is worse than a missing one, because everyone believes it is working. **Output is a proposal, never a silent write.** Show findings, get a decision, then edit. ## Arguments - (none) — full audit of all four surfaces. - `--surface hooks|agents|skills|workflows` — audit one surface. - `--apply` — after review, apply the accepted recommendations. - `--dry-run` — report only. ## Process ### Step 0: Track the run Use TodoWrite: reachability → correctness → coverage gaps → rank → propose → apply. ### Step 1: Reachability — guards that never fire Every hook must be reachable from something. Check all invocation sites, not just `settings.json`: ```bash for f in .claude/hooks/*; do b=$(basename "$f"); found="" grep -qF "$b" .claude/settings.json 2>/dev/null && found="settings.json" [ -z "$found" ] && found=$(grep -rlF "$b" .github/workflows/ .claude/hooks/ 2>/dev/null | grep -v "/$b$" | head -1) [ -n "$found" ] && echo "R