What it is
Bounded Jev decisions around a coding agent, with shadow mode first.
What it does: Hooks Claude Code so Jev scores risk on shell commands no rule decides, classifies failures and checks diffs at stop time, logging every decision for /jev-report.
Why it belongs in the showcase: It publishes which decisions helped and which did not, and turns on only command-risk gating by default.
Caveats: Author-run benchmark; thresholds must be fitted on your own labelled data.
Evidence
Public MIT repo. Projects opt in with /jev-init and start in shadow mode, where Jev decisions are only logged to receipts. The author's study (244 labelled decisions, 162 end-to-end Claude Code runs on two models, write-up linked from the README) supports one active feature: command-risk gating, where policy rules plus Jev on unrecognised commands caught every deny-worthy command in the test set with 3% false alarms (balanced accuracy 0.92). Skill selection, context ranking and patch verification stay in shadow mode because they did not improve end-to-end results.
Caveats
- Benchmark numbers come from the author's own study, not an independent evaluation
- Jev can only add an ask; deny and allow stay with regex policy
- Needs Node.js 20+, Claude Code and a TypeSafe API key
Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.