What it does
The review router scores diffs, not filenames, and caught thirteen CVE fixes that filename rules had routed to the cheap path. The command gate sees only the command, so nothing in the session context can talk it out of a block.
Agent tooling
jev-engineering
Working notes on putting guardrails around coding agents, each with a benchmark in the repo that reproduces the claim.
Bookmark: jev-engineering
Keep this in your collection.
Leave a noteWhat would you try with this?
: jev-engineering
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
How you can use it
A developer can install this tool as a plugin for a coding assistant like Claude. They provide an access key that connects the app to the TypeSafe AI service. The tool starts by simply watching the assistant. It records proposed commands without changing anything.
A developer can set the tool to block risky commands. It can also wait for a person when the answer is uncertain. This catches normal mistakes rather than stopping real attackers. The tool can still be tricked if an assistant falsely claims someone already approved the work.