FAQ

Is this an LLM?

It is a call to a System One style endpoint: you send questions plus an input and get back probabilities, a chosen key or a level, no text. The hosted Jev model is the reference; jevx does not care what is behind the URL as long as it speaks that request shape.

Why not just let Claude decide?

Because the small calls add up. Deciding whether each of 300 log lines is an error costs the agent context and tokens for every line; one batch call gives a number per line that the agent can filter with jq. A number is also something an if can use.

Does it explain its answer?

No. You get yes 0.92. If you need the why, ask a better question (ask about the reason directly: “Does this line show a timeout?”) or use a --choice over the candidate reasons.

How accurate is it?

There is no published number, because none would hold for your questions and your model. Run a few dozen of your own items through it, compare with what you would have said, and set --yes / --no per question. An eval command for exactly this is on the maintainer’s list, not shipped.

What does it cost?

jevx itself is free and open source. A hosted endpoint charges per call; jevx stats shows calls and tokens per day. A self-hosted or local model costs whatever it costs you to run.

What does the hosted endpoint see?

The question, the options, the input and the merged context. Never your config, your other files or your keys. See Privacy.

Can I run it fully offline?

Yes, if you have a model that serves the System One request shape locally; point a profile at it. jevx makes no other network calls: no telemetry, no update check.

Does it work with Codex, Cursor, Gemini CLI?

The skill is a Markdown file in ~/.agents/skills and ~/.codex/skills, so any agent that reads those gets it. The plugins use Claude Code’s hook protocol and only run there today.

Why are the plugins disabled after install?

Because a probability threshold that can refuse a shell command deserves a look at its log first. Enable one, watch plugin log in shadow mode, then plugin mode NAME act.

Can it edit my CLAUDE.md?

No, by design. It reads the ## Jev section, and jevx context shows what it read. There is no command that writes it.

What is "noul"?

The endpoint’s name for a yes/no question, kept as the flag name (--noul) so the CLI matches the API and the model’s docs.

Why exit 3 and 4 rather than 1 for everything that is not yes?

So if jevx is … cannot take the “no” branch because the network was down. 1 is a real no; 3 is “not sure, look yourself”; 4 is “the call failed”.

Where is the config?

~/.config/jevx/config.json (or JEVX_CONFIG), repo-level .jevx/questions.json and .jevx/plugins.json, logs under ~/.local/share/jevx/. See Config file.

I used jevcli before. What changed?

The command is now jevx. The binary still reads the old JEVCLI_CONFIG and JEVCLI_LEDGER variables as a fallback.