What it is
Claude Code mod that puts Jev in front of the auto-mode classifier: one request with eight yes/no questions per tool call, then allow (plausibly serves the request, every risk check low), deny (risky and unrequested), or defer to the built-in classifier.
What it does: Shell-aware prefilter parses commands the way a shell would and blocklists destructive / exfil shapes; Jev decides the rest with calibrated thresholds.
Why it belongs in the showcase: Clean allow / deny / defer split for tool-call guardrails, with author-published thresholds and an eval corpus you can inspect.
Caveats: Author says 7,860 labeled calls and ~2× faster than the built-in classifier on the calls it decides — author-reported; not a sandbox replacement.
Evidence
Public MIT README; allow / deny / defer thresholds plus a shell-aware blocklist in front of Jev. Author says evaluated on 7,860 labeled tool calls.
Caveats
- Author says 7,860 labeled tool calls and a sealed test split — author-reported, not independently audited
- Defers undecided calls to Claude Code's built-in classifier; a second lock, not a sandbox
- Hosted Jev sees recent messages and the tool call; README says secret-bearing calls are blocked locally
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.