Jev Git

That is a p50 the project reports, and the comparison it draws is against 4 to 6 seconds for a full LLM per commit. The architecture note explains the speed: a single statically linked Rust binary under 2MB using blocking HTTP through ureq for zero-runtime startup overhead, evaluating noul decisions on the SystemOne endpoint.

The README does not document that. It names no retries, no offline behaviour, and no bypass flag for a failing or unreachable endpoint. What it does describe is git jev install for a pre-commit hook, git jev check on demand, and piping any diff into it.

That is the stated aim: regex-based git hooks miss subtle semantic context, so the gate evaluates the staged diff with calibrated probabilities for unredacted secrets, prompt-injection attacks and destructive commands before they enter git history. You get a probability rather than a rule match, and it is explicitly not a heavy linter or syntax parser.

Not documented. The README describes the checks, the invocation and the non-goals, but names no confidence cut, no scoring scale and no configuration file, so you would be accepting the project's own default without being able to see or change it. Per-call cost is not given either.

It screens staged git diffs for unredacted secrets, prompt-injection attacks, and destructive commands, returning calibrated Jev probabilities. Run git jev install to add a pre-commit hook, git jev check on demand, or pipe a diff such as git diff HEAD~1 | git jev check.

A single statically linked Rust binary under 2 MB with no daemons and no background memory. It uses blocking HTTP through ureq for zero-runtime startup overhead and runs in about 80 ms p50.

MIT. It is explicitly not a multi-agent framework, not an interactive code-writing LLM, and not a heavy linter or syntax parser.