Jev decision tools

Calibrated yes/no, choice and score judgements from TypeSafe Jev — direct with a TypeSafe key, or through OpenRouter — probabilities instead of prose, about a second and $0.00002 a call.

TypeSafe Jev decisions for Elyra -- calibrated yes/no, choice, and score judgements through OpenRouter.

Jev is a decision model, not a chat model. It never generates text. You give it a piece of content (state) and one or more typed questions, and it returns probabilities. A call takes about one second and costs roughly $0.00002.

Install

elyra install npm:@elyracode/jev-tools

Requirements

One of two API keys:

• TypeSafe direct (preferred): a key from typesafe.ai. Set TYPESAFE_API_KEY. Requests go to https://api.typesafe.ai/v1/systemone with model jev-latest.
• OpenRouter: log in with /login and pick OpenRouter, or set OPENROUTER_API_KEY. Requests go to OpenRouter's alpha decisions endpoint with model ~typesafe/jev-latest.

When both are present, TypeSafe direct is used. Force one with JEV_BACKEND=typesafe|openrouter. The request and response shape is identical on both; only endpoint, model slug, and billing differ.

Tools

Commands

• /jev -- show active backend, key status, and bash gate state
• /jev gate on|off -- toggle the bash gate for the current session

Bash gate (opt-in)

When enabled, every bash tool call is first sent to Jev with the question "would running this command destroy data or make a change that is hard to reverse?". If the probability is at or above the threshold, Elyra asks you to confirm before running it. In non-interactive mode (print, RPC) the command is blocked instead.

Enable with elyra --jev-gate, JEV_BASH_GATE=1, or /jev gate on. Off by default because it adds about one second of latency to each bash call.

If the key is missing or the API fails, the gate logs a warning and lets the command through rather than blocking all shell access. It is a second opinion, not a security boundary.

Usage

> Is this diff a breaking change? Ask Jev before you answer.
> Classify these 20 log lines by severity using decide, then only show me the ones rated high or critical.
> Before running that migration, use decide to check if the command is destructive.

Question types

noul -- yes/no with a calibrated probability:

{ "name": "refund", "type": "noul", "instructions": "Is the user asking for a refund?",
  "criteria": { "true": "explicitly wants money back", "false": "anything else" } }

choice -- pick one option, get the full distribution:

{ "name": "route", "type": "choice", "instructions": "Who should handle this?",
  "criteria": { "code": "write or change code", "research": "needs web lookup", "other": "none of the above" } }

score -- ordinal scale, 2-10 labels from low to high:

{ "name": "severity", "type": "score", "instructions": "How severe is this log line?",
  "criteria": ["debug", "info", "warning", "error", "critical"] }

Writing good questions

Jev reads criteria literally and does not infer intent beyond what is stated.

• Describe observable conditions, not goals. Spell out both sides for noul via criteria.
• Make choice options exhaustive, or add an other option.
• Keep state to what the questions need. The context window is about 32K tokens.
• Ask several questions in one call when they concern the same state; the cost is the same.

Configuration

Notes

• The request contract follows the TypeSafe API reference. OpenRouter's decisions endpoint is an alpha API mirroring the same shape and may change.
• Retries with backoff on 429/529/5xx, up to three attempts. Requests time out after 60 seconds.
• Question design guidance: see the primitives docs. A Noul near 0.5 means yes and no are equally likely, not "medium"; use a Score for degree.
• Unofficial community integration. Not affiliated with TypeSafe or OpenRouter.