Agent tooling
retsu-AI's open-source coding-agent harness can ask Jev to pick a model for each task, check tool results and final answers, and approve held tool calls.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
Each Jev feature is switched on separately, and a stored TypeSafe key turns nothing on by itself. When Jev gives no confident approval, the request falls through to a reviewer model and then a person. QQ's own design notes say users report that well over half of what reaches Jev comes back to them; the rate is unmeasured, and the maintainers' tests point mostly to the integration, not necessarily the model.
QQ suits someone who leaves an AI coding helper working alone and wants it to stop less often to ask permission for routine steps. Switch on only the Jev feature you need, and only for work you are happy to send to TypeSafe. A yes from Jev is support, not proof that the work is right.