Contracts
AgentAssert
Design-by-contract for agents: preconditions, invariants, governance and recovery checked at runtime, so a violation is caught at the step that caused it.
Loading content…
Decisions · Jev Decision Layer
Jev Decision Layer for AI Agents
Give your assistant a bounded task, tool, skill, or review choice. Receive a typed answer with confidence and a local receipt. The host keeps execution authority and verifies the work.
Offline first
Try the local gate before connecting a model
On macOS with Python 3.11 or newer and Git, replay the shipped synthetic answers through the real checking logic. No key or provider call is needed.
git clone https://github.com/qualixar/jev-decision-layer.gitcd jev-decision-layerplugins/qualixar-jev-decision-layer/scripts/jev selftest
The self-test checks gate contracts. It does not measure provider accuracy or savings. With Node.js 20 or newer, node tools/demo.mjs replays twenty recipe examples across nominal, uncertain and adversarial cases.
Install
Install for Codex
codex plugin marketplace add qualixar/jev-decision-layercodex plugin add qualixar-jev-decision-layer@qualixar-jev-layer
Restart Codex, then ask to set up Jev for a folder. Enter any provider key only in the private setup page and review the scope yourself. Other hosts and upgrades.
Release 1.0.15; product page reviewed 2026-10-31. Source and host verification. macOS (supported); Linux (experimental); Windows (disabled). Jev is an AI Reliability Engineering decision primitive. Hosted decisions use the provider and scope you approve.
Evidence and adoption boundaries
- Confidence and probabilities are advisory; thresholds are uncalibrated defaults.
- Synthetic self-tests check local gates, not decision-model accuracy or end-to-end savings.
- Host adapters have different verification evidence. Review the host guide before adopting.
- Hosted TypeSafe or OpenRouter requests use provider billing and the reviewed workspace scope.
- Local Laya requires Apple Silicon; Jev-only never silently falls back to Laya.
- Secret screening is best-effort and does not replace approval to share confidential material.
Built by Varun Pratap Bhardwaj under Qualixar. This is an independent integration, not an official TypeSafe or host-provider product.
Recipes
Find a workflow for your task
55 typed decision recipes for agent routing, research, content and delivery. Choose a specification, supply its inputs and review the advisory result.
What it does
How it works.
Typed questions
Choose from known options, score against a rubric, or check a statement with a yes/no likelihood.
Local policy gates
Check reviewed workspace scope, recognizable secrets and request limits before a hosted call. Screening is best-effort.
Inspectable receipts
Review the recommendation and local policy result. An advisory answer never grants tool permission.
Hosted or local
TypeSafe Jev is the primary hosted route. Optional local Laya requires a separately verified Apple-Silicon installation.
Reusable recipes
Use bounded contracts for routing, extraction checks, retrieval relevance and review attention. Confidence thresholds remain uncalibrated defaults.
Host-specific adapters
Codex, Claude Code, VS Code, Hermes and Antigravity have different verification evidence. Read the host guide for the exact boundary.
Use it for
Where it fits.
- Choose a task, tool or skill from your own shortlist
- Check an extraction against its source
- Assess whether retrieved passages answer a question
- Prioritize attention during a code review
Changelog
What shipped.
The latest releases on GitHub.
Decisions sent through OpenRouter are no longer refused. OpenRouter names the dated version of the model that answered, and Jev treated that as a different model. A dated version of the requested model is now accepted, nothing else is, and the local receipt records which…
On Linux, the setup and workbench commands exit as soon as they finish.
Local Laya setup, seventeen new recipes, and hardening throughout.
Qualixar Jev Decision Layer 1.0.12 — Name the Policy That Blocks It
When an organization manages Claude Code, its policy can stop the Jev plugin's hooks, keep the plugin from loading, or leave its MCP server off an allowlist. Until now the plugin just went quiet and nothing said why. A tool that stops working without a word is a reliability…
Qualixar Jev Decision Layer 1.0.11 — Approve Once, Every Harness
Qualixar Jev Decision Layer lets a coding agent hand small, consequential choices (which tool, which task, which files matter, whether an extraction matches its source) to TypeSafe Jev, then continue through the host's own permissions. It is part of Qualixar's AI Reliability…
Qualixar Jev Decision Layer 1.0.8 — Fast, Typed AI Agent Decisions
When an agent reaches a small but consequential choice—tool, task, test, or review path—it can ask TypeSafe Jev, a model built for fast typed decisions, and continue through the host's normal controls. Qualixar Jev Decision Layer makes that flow reusable across harnesses.…