DEMO / RECORDING
Overview & Result
Roshan Chandna's open-source Auto-Guard sends proposed coding-agent tool calls and the user request to Jev; code maps typed risk, scope and sensitivity answers into allow, human approval or block.
How Jev fits in the loop
- A coding agent proposes a tool call before it runs.
- Jev evaluates destructive potential, task scope, sensitivity, action class and risk.
- Local policy lets it run, asks a human, or blocks it; native agent permissions still apply.
How to reproduce
- Read the project's public README and choose a supported coding-agent hook.
- Install agent-autoguard and configure a supported API key as documented.
- Run the project's sample check and inspect allow, escalate and block outputs; keep normal agent permissions enabled.
Why this build matters
A concrete example of bounded Jev decisions placed at a consequential tool boundary, with code, demo and an author-run evaluation.
Limitations
- Author-run holdout tests are small and not independently reproduced here.
- The repository states Auto-Guard is a filter, not a sandbox, and can miss a risky action.
- The guard cannot control a command after its pre-execution check.