I set a TypeSafe key and every answer still comes back instantly, what happened?

Almost certainly the silent heuristic fallback. TypeSafe's API is a single POST /v1/systemone endpoint with no /chat/completions at api.typesafe.ai, so without TYPESAFE_JEV_NATIVE=1 the chat-proxy transport 404s and the provider chain falls through to the offline heuristic. Run node dist/cli.js info to see the backend chain and which stage answered; JEV_BACKEND_PROVIDER=auto is the default and will mask a misconfiguration.

Is the offline simulator actually testing the model, or just the plumbing?

The plumbing. The deterministic offline simulator exists so every primitive stays valid and schema-safe without a key, and that is what the unit and datacenter tests exercise. It is not a model: a decision made under JEV_BACKEND_PROVIDER=heuristic is not a Jev answer, which is why the gatekeeper and the calibration check matter once you point it at a real backend.

Will this replace the LLM for planning and research writing?

No. The hosted API only answers Choice, Score and Noul, so JEV_LLM_BASE_URL still has to point at Ollama or vLLM for planner, red-team and research-synthesis text. Jev sits in front of the micro-decisions while System 2 generation stays wherever you run it, and the backend table in the README makes that split explicit.

The README quotes specific costs and latencies, are those measured?

They are the author's own figures on his own runs, not a benchmark on your workload. One real API cost for a single research invocation is quoted, and the latency and token-cost-per-session numbers are framing for the problem being solved. There is a calibration check tool that compares claimed confidence against real accuracy per threshold bucket and checks option-order bias, which is the honest next step.

How do I install jev-x-kit as a Claude Code plugin?

Git clone the repo, run npm install and npm run build, then point Claude Code at the folder. The shipped .claude-plugin/plugin.json registers the jev skill and the jev-super-agent MCP server.

Which backends can jev-x-kit run against for free?

OpenJev on vLLM, LayA, Ollama for the System-2 planner, the Vercel AI Gateway free tier, and TypeSafe Jev native. JEV_BACKEND_PROVIDER=auto picks one automatically and falls back to a fully offline heuristic.

What are the gatekeeper's three confidence bands?

Above 0.85 executes directly at zero LLM cost, 0.60 to 0.85 splits into speculative sub-decisions, and below 0.60 escalates to System 2 or a human.

CommunityJev Practices & Patterns

Jev X Kit

A TypeScript framework that puts a System 1 decision layer in front of agent micro-decisions, shipped as a Claude Code plugin, an MCP server and a CLI, with an offline simulator.

See highlights

Open github.com

Stars: 3
Language: TypeScript
License: MIT
Published by: Community
Author: Kadihx
Last verified: Oct 2026

Kadihx/jev-x-kit

User repository on GitHub

View on GitHub

Offline $0 decision layer for coding agents: Choice/Score/Noul primitives, BELKI confidence gatekeeper, ultra-planning, red-teaming, research and RLVR self-improvement -- as an MCP server + CLI + Claude Code skill.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

About this project

Universal TypeSafe Jev / OpenJev autonomous decision, deep research, ultra-planning & self-improving agent framework — packaged as a Claude Code plugin/skill, an MCP server, and a standalone CLI. Then point Claude Code at this folder as a plugin (.claude-plugin/plugin.json is already wired up: it registers the jev skill and the jev-super-agent MCP server with 21 tools). See skills/jev/SKILL.md for the command cheat-sheet Claude reads to decide when to reach for Jev instead of free-text reasoning.

Quoted from the project's README. Read the full README.

Tagscommunity awesomejev catalog-import

Works forJev for Classification Jev for Ranking Jev for Tool calling

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

A confidence gatekeeper executes above 0.85, splits into speculative sub-decisions between 0.60 and 0.85, and escalates below 0.60.
Winnow compaction deletes irrelevant log lines and keeps file paths, commands, error codes, URLs and diff headers byte-identical.
Backend resolution falls back from TypeSafe to OpenJev on vLLM, then LayA, then a heuristic that is always available.
Unit tests, datacenter tests and a 50-check MCP smoke test all run with no network, no GPU and no API key.

Watch out

TYPESAFE_JEV_NATIVE=1 is required for the hosted API, because api.typesafe.ai exposes a single POST /v1/systemone endpoint and no /chat/completions; without the flag the chat-proxy transport 404s and every call silently falls back to the offline heuristic, which is easy to mistake for a working local model. The hosted API answers only Choice, Score and Noul, so text generation for planner and red-team work needs Ollama or vLLM regardless. The tool count is given as 21 in the plugin description and 28 in the MCP table. The latency, cost and per-run figures are the author's own rather than a benchmark on your workload, and the source text reviewed here was cut off partway through, so later sections were not read. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

tenbin

Design-time toolkit for using Jev from a coding agent, combining docs, a stdio MCP server, and an agent skill for decomposing judgments into Choice, Score, and Noul.