Open source · Community toolkit

Small decisions.
Explicit boundaries.

Bring typed Jev decisions into your own workflow. Start with a reusable skill and an offline test—not a production API key.

Install the pluginRead the guide
CLAUDE CODE / QUICK START
/plugin marketplace add cathy-kim/jev-workflow
/plugin install jev@jev-workflow

/jev:jev-decisions Design a support-routing workflow. Keep it offline.

No hooks, credentials or automatic external calls. The opt-in client calls OpenAI Decisions, TypeSafe or OpenRouter only with --live; the stdio MCP server starts offline. Your agent runtime's own permissions still apply.

Judgment ≠ permission.Jev supplies typed judgments. Your application owns authorization, policy, execution and review. Confidence is not a guarantee of accuracy.
01 / Set up

One skill.
Your workflow.

01

Install in Claude Code

Run the marketplace commands above in Claude Code. Start a fresh session if the command is not yet visible, then invoke /jev:jev-decisions.

02

Or use a portable skill

Download the skill bundle and add its folder to your agent's supported skill directory. It includes the references, templates and Python scripts—no private runtime required.

03

Run the offline tests

Requires Python 3.10+. No Python packages or API key are needed for the core starter.

git clone https://github.com/cathy-kim/jev-workflow.git
cd jev-workflow
python3 -B -m unittest discover \
  -s plugins/jev/skills/jev-decisions/scripts \
  -p 'test_decision_kit.py' -v
python3 examples/offline_demo.py

The demo uses synthetic probabilities. A shadow-mode REVIEW recommendation is expected; no downstream action is executed.

Full installation & troubleshooting →
02 / Apply

Make the next decision narrow.

Define the evidence, ask one question per dimension, validate the answer, and return a policy recommendation. Test failure paths before introducing live inference.

Support routing“Design typed questions for billing, technical, account and other. Treat missing evidence as review, never as automatic denial.”
RAG evidence gates“Separate passage relevance from answer sufficiency. Preserve citations and access boundaries. Write offline parser tests.”
Agent evaluation“Check tool order and schema in code; reserve semantic judgments for the model. Keep refusal and N/A distinct from a pass.”
Shadow adoption“Produce an evaluation plan with labeled holdout data, uncertainty, cost caps and a rollback path. Do not call a live API.”
03 / Read & reuse

The guide.
The working parts.

04 / Know the limits

Offline first.
Honest about scope.

  • The offline starter validates a limited, normalized text contract. The separate live client covers three documented provider routes and sends nothing without --live.
  • Synthetic fixtures prove local behavior, not model quality, production compatibility or calibrated thresholds.
  • Live use still needs authorized data, explicit budgets and independent evaluation; live authentication and answer quality were not tested by this project.
  • The guide is a dated technical reference. Check current official APIs, model versions, pricing and media support before adoption.
  • No internal project data, personal conversation logs, secrets or private Git history are included.