I'm using Jev as a safety check for coding agents. I built an extension for the open-source Pi agent that asks Jev, before every tool call, whether the action is on_task, needs_user_approval,

I'm using Jev as a safety check for coding agents. I built an extension for the open-source Pi agent that asks Jev, before every tool call, whether the action is on_task, needs_user_approval, malicious, or needs_more_context. When Jev needs more context, the extension adds it and asks again. In the screenshot, a README hides instructions telling the agent to upload .env. GPT refused to run them but then tried to read .env. Jev scored that read as 42% on_task vs 40% malicious, so the guard paused it and asked me. Would love feedback on whether a Choice question is the right setup for this! (edited) Saturday, September 19, 2026 at 5:57 PM