jev-code plugin
The Jev classifier as an agent tool: labels items, answers yes or no and ranks candidates with probabilities within milliseconds.
- 59 ★ on the repository
- Last commit
Extended access
Why this level
- Evidence text and the API key go to an external model host, and setup changes the configuration of several agents.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add plugins/jev-codeDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Evidence text and the API key go to an external model host, and setup changes the configuration of several agents.
This is third-party code. Review the repository files before installing.
What it does
jev-code wires the TypeSafe Jev decision model into Claude Code, Codex, OpenCode and Pi, together with a skill that tells the agent when to reach for it. The agent sends evidence and its own classes and gets back a label, a probability distribution and an `auto` or `review` decision to branch on. There are five tools: `jev_classify` labels many items, `jev_check` gives a yes or no for one piece of evidence, `jev_score` rates on an ordered scale, `jev_rank` sorts candidates by relevance and `jev_ask` covers other questions. The same calls work from the shell through the `jev-code` CLI. The `setup` command detects installed agents and wires the skill and tool into each.
Who it is for. Developers whose agents triage dozens of similar items: CI failures, review findings, files matching a question.
Good fit when
- After a CI run you need to separate flaky tests from real bugs.
- The agent must pick the files, out of forty, that answer a question.
- A pull request description is checked against the diff to see whether it does what it claims.
- Review findings need a severity score on a common scale.
Not a fit when
- You need a coherent prose answer: Jev returns only typed answers.
- The data must not leave for an external model host.
- A single regular expression is enough.
Example request
Triage the failing tests in the last CI run: which are flaky, which are real bugs?Limitations
The model runs on TypeSafe, OpenRouter, Vercel AI Gateway, OpenAI or your own System One gateway, so an API key and sending evidence over the network are required. Installation needs Node.js 20 or newer. The default auto-accept threshold is 0.85 with a minimum margin of 0.5, and the agent must check uncertain cases itself. Pi uses a native extension, other agents work through MCP over stdio.
How to disable. Remove the `jev` skill from the skills directory and the MCP server entry from the agent config. Delete the key from your shell profile by hand.
FAQ
How do I install jev-code?
Run `npx -y @french-castle/jev-code@latest setup`. It detects Claude Code, Codex, Pi and OpenCode, copies the skill and registers the tool. The `--dry-run` flag shows the plan without changes.
Which keys work?
Keys for TypeSafe, OpenRouter, Vercel AI Gateway and OpenAI, or the URL of your own System One gateway. `TYPESAFE_API_KEY`, `OPENROUTER_API_KEY` and `AI_GATEWAY_API_KEY` are read from the environment.
How do I check it works?
Run `npx -y @french-castle/jev-code@latest doctor --live`: it makes a live call and reports the connection state.
What do the auto and review decisions mean?
`auto` means probability and margin clear the thresholds and the answer can be accepted. `review` sends the case to a manual check by the agent.
Does it work without MCP?
Yes, Pi has a native extension, and any shell can use the `jev-code` CLI with the same five calls.
Discussion
No questions or reviews yet.
Related
PluginsA plugin and MCP server that hand no-text decisions to the Jev model and gate shell commands while the LLM writes content
An MCP server giving an agent three Jev tools: typed yes or no, choice and score questions with probabilities, offline by default.
Jev-Style runs small decision models on your machine: a probability per option, six skills, MCP and a Claude Code guard hook
A plugin for DeepSeek Harness and iPolloWork: structured choice of tools, skills and agents and scoring of outputs
A plugin with a catalogue of public Jev examples: a skill and an MCP server show how others solved the same decision task.
Five skills for the Jev model: sorting records, source checks, rubric scoring and choosing an agent's next action.
README in short
The project packages the TypeSafe classifier for four agents: five tools with identical names and JSON, a skill on writing classes and questions, and a CLI. Installation and the key are set up with one command.