Mercury CLI
Say hello to Jev! But he can’t say hello back...?
I wanted to try Jev in Mercury because I was curious whether it could help the model make better decisions while working on code. Beta.20 adds it as a tool called JevEval.
Jev is a decision model from TypeSafe. It takes the facts you give it and returns a probability, a choice between supplied options, or a score. His replies are all numbers.
Why I wanted to try it
A model can collect the evidence for a bug and still have a few possible explanations. It can put those options to Jev before choosing what to investigate next.
Picking the wrong explanation can mean changing code and running tests, only to have to go back and do it again. I’d like to see whether getting a second opinion at that point helps catch some of those wrong turns before the edits start.
I’m also interested in using it to check a proposed change against decisions I’ve already made. Mercury would give Jev those decisions and the proposal, then ask how well they fit.
Using it on Mercury
Jev starts switched off and uses your own TypeSafe API key. You can set it up through /jev, set limits on spending and requests, and see its usage separately from the main model.
I’ll load up some credits and try it on Mercury’s own development. I want to see which questions are worth asking and whether the answers actually help with the next step. The tests will still be there to check the resulting code.
The other fixes in beta.20
This update also adds usage bars for Moonshot and Z.AI plans. Usage readings refresh expired sign-ins after a night idle, and a long-running shell command gets a hint for moving it to the background.
There are fixes for typed brackets being mistaken for mouse input, the config popup spilling into the neighbouring column, and a missing Jev figure crashing the screen.
I’m excited to see how Jev can help make Mercury better. Getting it into the toolset means I can start using it while I build, and hopefully add more support as I get a better feel for where it helps.