dsh-jev-interceptor

Checked, not fully reviewed. We ran its own build, typecheck and tests, and read the guard and the client; we did not install it into a live DeepSeek Harness session.

The plugin sits in front of every tool call and asks one batch of typed questions about it — risk, irreversibility, task fit, injection suspicion — and it keeps the decision architecture honest in a way worth naming: “no objection” delegates via next() rather than returning allow, so this listener cannot swallow another listener’s veto, and every degraded path — no key, cooldown, timeout, parse mismatch — delegates as well. A guard that fails open is a choice; a guard that cannot accidentally overrule a stricter neighbour is a better-built one.

npm run build, npm run typecheck and npm test are all clean on a fresh clone: 64 tests across 7 files.

What it sends is deliberately small, and its own module comment says why: the state carries the tool name, a head-and-tail preview of the arguments, and a short digest of recent messages, because “full arguments can be megabytes and add cost without accuracy”. Answers are cached under a SHA-256 of model, state and questions, and a cache hit reuses the answers only — usage and latency stay attributed to the original call, so a reused answer cannot inflate the reported token count.

MIT, Node 20.3+, bilingual README.

For the same gate in other harnesses, see pi-jev and Jevbridge; for what a memory tool sends instead, jevmem.

See how it compares with other tools in Best Jev tools, tested hands-on.

Review updated Sep 26, 2026. Numbers quoted from the project are its author's own; we don't publish our own measurements of Jev.