Kevin-Zhouu/jevify

Python

1

27 commits

updated Sep 23, 2026

See the code

See what people are saying

README

jevify

Get your agents running on Jev today

Explore Jevify ↗ · Install · Docs

Illustrative Jevify flow: one command, audit plan, approval, conversion

Install

npx skills add Kevin-Zhouu/jevify

Then, in Claude Code (beta):

/jevify

In Codex: $jevify. Or say “jevify this repo.” Optional: /jevify src/agents/triage.py.

See the audit first. Then choose a new branch, a separate copy, report only, or individual calls. No key or setup questions needed to see the plan.

The animation illustrates the flow; its files, counts and test results are examples. Demo details · Real historical run

LLM vs Jev for decisions

Claude streams JSON while Jev returns the structured decisions

12 questions. 136 lines of JSON. Jev: 0.44s. Claude Haiku 4.5: 5.11s.

All 12 category choices match. One synthetic example via OpenRouter, replayed from real stream timestamps—not a general benchmark. Watch video · Raw data

How it works

One command → Audit → You approve scope + output → Convert → Validate

  • Your codebase: finds call sites and traces how their outputs are used.
  • Your choice: new branch, separate clone, or report only.
  • Your safety net: keeps generation with your LLM and adds fallback for uncertain decisions.
  • Your evidence: runs tests and, with a key, compares live outputs, latency, and cost.

Decisions such as routing, classification, and rubric scoring are candidates. Free-form generation stays with your LLM. You get a conversion plan and a results report.

Built with Jev · Minecraft

Astra plans. Jev chooses the actions. The Ender Dragon goes down.

Play: Jev and Astra take down the Minecraft Ender Dragon

▶ Watch full video · Original post · Source code

Community demo by Ronak Malde; not a Jevify conversion. Runs on a pre-surveyed seed in Peaceful mode.

Bring your key

Use TYPESAFE_API_KEY or OPENROUTER_API_KEY through your environment or the hidden-input launcher. Keep your original provider key for fallback.

No Jev key? Audit and approved code changes still work; live validation is skipped.

Status

Audit-first flow: new; full cross-harness re-evaluation is pending. Historical Codex DEV runs passed; Claude Code remains beta with known conversion validation failures. Evaluation results →

Installation, headless usage, and development
codex exec 'Use jevify with jevify.config.yaml'
claude -p 'Use jevify with jevify.config.yaml'

Development checks (Python 3.10+, Node 18+, and npm ci in eval/):

python3 -m unittest discover -s eval -p 'test_*.py'
python3 eval/freeze.py --verify

Skill helpers use Python's standard library. Vendored material retains its upstream licenses.

Contributors

Kevin-Zhouu

27 commits

Kevin-Zhouu/jevify

Python

1

27 commits

updated Sep 23, 2026

See the code

See what people are saying

README

jevify

Get your agents running on Jev today

Explore Jevify ↗ · Install · Docs

Illustrative Jevify flow: one command, audit plan, approval, conversion

Install

npx skills add Kevin-Zhouu/jevify

Then, in Claude Code (beta):

/jevify

In Codex: $jevify. Or say “jevify this repo.” Optional: /jevify src/agents/triage.py.

See the audit first. Then choose a new branch, a separate copy, report only, or individual calls. No key or setup questions needed to see the plan.

The animation illustrates the flow; its files, counts and test results are examples. Demo details · Real historical run

LLM vs Jev for decisions

Claude streams JSON while Jev returns the structured decisions

12 questions. 136 lines of JSON. Jev: 0.44s. Claude Haiku 4.5: 5.11s.

All 12 category choices match. One synthetic example via OpenRouter, replayed from real stream timestamps—not a general benchmark. Watch video · Raw data

How it works

One command → Audit → You approve scope + output → Convert → Validate

  • Your codebase: finds call sites and traces how their outputs are used.
  • Your choice: new branch, separate clone, or report only.
  • Your safety net: keeps generation with your LLM and adds fallback for uncertain decisions.
  • Your evidence: runs tests and, with a key, compares live outputs, latency, and cost.

Decisions such as routing, classification, and rubric scoring are candidates. Free-form generation stays with your LLM. You get a conversion plan and a results report.

Built with Jev · Minecraft

Astra plans. Jev chooses the actions. The Ender Dragon goes down.

Play: Jev and Astra take down the Minecraft Ender Dragon

▶ Watch full video · Original post · Source code

Community demo by Ronak Malde; not a Jevify conversion. Runs on a pre-surveyed seed in Peaceful mode.

Bring your key

Use TYPESAFE_API_KEY or OPENROUTER_API_KEY through your environment or the hidden-input launcher. Keep your original provider key for fallback.

No Jev key? Audit and approved code changes still work; live validation is skipped.

Status

Audit-first flow: new; full cross-harness re-evaluation is pending. Historical Codex DEV runs passed; Claude Code remains beta with known conversion validation failures. Evaluation results →

Installation, headless usage, and development
codex exec 'Use jevify with jevify.config.yaml'
claude -p 'Use jevify with jevify.config.yaml'

Development checks (Python 3.10+, Node 18+, and npm ci in eval/):

python3 -m unittest discover -s eval -p 'test_*.py'
python3 eval/freeze.py --verify

Skill helpers use Python's standard library. Vendored material retains its upstream licenses.

Contributors

Kevin-Zhouu

27 commits

Languages

Python

60.1%

TypeScript

31.9%

JavaScript

8.0%