Typed answers with Jev
Not every question deserves an essay. “Which of these three?”, “how risky is this?”, “is that true?” have shapes — and a decision model answers them in that shape, with a probability attached, for a fraction of what a language model costs.
Setting it up
Get a key from typesafe.ai, then paste it into Settings → TypeSafe (Jev). It is stored in the Keychain alongside your other credentials.
Without a key nothing changes anywhere in the app. This is an optimisation, never a dependency — if Jev is unreachable, slow, or unsure, iCoderAgent behaves exactly as it did before.
Asking directly
jev is python is dynamically typed
→ Yes (93% true)
jev pick swift|kotlin|dart :: which is native on iOS?
→ swift · confidence 81%
swift 88%
kotlin 11%
dart 1%
jev rate low|medium|high :: how risky is a force push?
→ 1.74 high · confidence 92%Options and rubric levels are separated by |, and the question follows ::. The same words work after agent — agent noul …, agent choice a|b :: … — and go straight to Jev rather than your model. A plain yes/no question is routed automatically.
It cannot read your files. Jev only sees the sentence you give it. Anything about this project, what you have installed, or the current state of your code still goes to the agent, which can actually look. “Is numpy installed?” is a question about your machine, not a fact about the world.
What the agent uses it for
| Decision | Why it helps |
|---|---|
| Which specialist role fits a subtask | Installing a catalogue of sub-agents used to put every role’s description into every prompt. Jev matches the subtask instead, so the list stops being re-sent on each turn. |
| Whether a command is about to destroy something | Shell tools are judged on the command, not just the tool. A confident “this is destructive” raises an approval prompt even when you have auto-approve on. |
| Whether finished work actually works | The verification pass is graded on a rubric rather than searched for the word “FAIL”, which only catches a verifier that phrased itself the expected way. |
What it costs
$0.042 per million input tokens; output tokens are free. A typical question costs a small fraction of a cent — far less than the same question answered by a language model.
It is metered separately and never folded into your model’s totals: the header shows a jev line once it has been used, and the Usage report lists it by name. That separation is the point — it is the only way to see whether the decision model pays for itself.