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
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
One command → Audit → You approve scope + output → Convert → Validate
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.
Astra plans. Jev chooses the actions. The Ender Dragon goes down.
▶ 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.
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.
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 →
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.
27 commits
Python
60.1%
TypeScript
31.9%
JavaScript
8.0%
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
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
One command → Audit → You approve scope + output → Convert → Validate
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.
Astra plans. Jev chooses the actions. The Ender Dragon goes down.
▶ 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.
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.
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 →
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.
27 commits
Python
60.1%
TypeScript
31.9%
JavaScript
8.0%