Question 1

Is jev-axi a prettier wrapper around an LLM with a deny list?

Accepted Answer

No, and the demo transcripts show the split. Routine read-only and project commands are decided on your machine for free and report source: local, so `pnpm test` never costs a call. Only escalated Bash calls reach Jev, where a curl-pipe-bash is denied at p=0.99 remote_code for roughly 375 ms and $0.00004. The same judgment screens fetched text for injection, and guard exits 3 so a pipeline can gate on it.

Question 2

The README admits file ranking did not work, so why trust anything else?

Accepted Answer

Because that is the finding, not a footnote. On a 390k-line repo, file ranking moved cost and file reads by less than run-to-run noise: a 25% saving in one session, a 29% penalty in another, mean and median disagreeing in sign within a single session, six runs per condition. The conclusion is to use it for judgments, not as a replacement for reading code. Spend is tracked in a local ledger because the API has no spend endpoint.

Question 3

Will a cached answer go stale after TypeSafe ships a model update?

Accepted Answer

No. A cached answer is reused for up to 24 hours, configurable with cacheTtlHours, and only while jev-latest still resolves to the model version that produced it, so a model update invalidates old answers automatically. jev-axi cache shows what is stored and cache clear [--stale] empties it. Output also prints the concrete model version and estimated cost per call.

Question 4

Will it hang in CI waiting on a pipe that never speaks?

Accepted Answer

Commands read piped stdin only when there is content on it; an empty stdin, or one that stays open and silent, both count as no input, so jev-axi diff reviews the working tree instead. diff and progress wait one second for a pipe to start before falling back, and passing `-` waits for a slow producer. In CI keep exit codes apart, since 0 matched, 1 nothing matched and 2 could not run are different outcomes.

Question 5

How do I install it and what do the commands cover?

Accepted Answer

MIT-licensed and on npm: `npm install -g jev-axi` or `npx -y jev-axi`, with TYPESAFE_API_KEY from the environment, a .env.local or .env found by walking up from the current directory to the repo root, or `jev-axi config set apiKey`. Bare `jev-axi` prints live status. The primitives are pick, rate, check, ask, rank, filter, find, diff, files, triage, guard, progress, commit, and recipe.

Question 6

What do the numbers in the output actually mean?

Accepted Answer

p is a probability, and choice probabilities sum to 1 across options. confidence from 0 to 1 summarizes how peaked the distribution is, and for yes/no answers it is the distance from 0.5 rescaled. band is a policy you can branch on: act at 0.75 or above, confirm at 0.45 or above, escalate below, tunable with --act, --confirm, or config set. Every run also prints tokens, latency, model version and estimated cost.

Question 7

How does it handle hundreds of files in one request?

Accepted Answer

rank, filter, and find tag items or lines with short ids and send them as one request, chunking automatically at 255 options or the token budget. Each ranking is paired with a yes/no does-anything-here-match question, so an empty result is a real answer rather than the least bad option. pick, rate, check, and ask also take several questions in a single call.

Jev Axi

About this project

Highlights

Quickstart

Questions about this entry

Reactions & coverage

From the Jev guides

From the Jev AI model community