JEV via OpenRouter

JEV is not a chatbot. It answers exactly one of three shapes: noul (boolean probability), choice (pick one option from a fixed list), score (position on an ordered scale). No output tokens, no free text. Model: typesafe/jev-1.13 on OpenRouter, $0.042/M input tokens, output free, ~32k token context.

Use it as a tool Claude Code calls, not as a replacement for reasoning. The pattern: Claude (system 2) decides what to ask and why; JEV (system 1) answers fast and cheap.

When to reach for it

• Routing: which subagent/skill/tool should handle this request
• Triage/classification: ticket severity, file type, invoice vs not-invoice, spam vs not
• Filtering a large batch: "which of these 500 files are candidates for X" (loop calls, or batch via multiple questions per state)
• Guardrail gate: does this input look like a prompt injection / policy violation, yes or no
• Scoring: urgency, sentiment, confidence, relevance, on a defined scale
• Pruning agent context: keep/drop a text block

If the task needs generated prose, code, or an answer outside a fixed set — don't use JEV, just answer directly.

Setup

Requires OPENROUTER_API_KEY in the environment. Check first:

echo $OPENROUTER_API_KEY   # bash
$env:OPENROUTER_API_KEY    # powershell

If unset, ask the user for it before proceeding — never hardcode a key in scripts or commit it.

How to call it

Use scripts/jev_client.py from this skill's folder (stdlib only, no deps, works from any cwd). Below, <skill-dir> means this skill's base directory — the one shown when the skill loads (plugin install: inside the plugin cache; manual install: ~/.claude/skills/jev). Always use the absolute path.

If the base directory isn't shown or the script isn't there, find it instead of guessing (newest install wins):

ls -d ~/.claude/plugins/cache/*/*/*/skills/jev/scripts/jev_client.py ~/.claude/skills/jev/scripts/jev_client.py 2>/dev/null | sort -V | tail -1

(Get-ChildItem "$HOME/.claude/plugins/cache","$HOME/.claude/skills" -Recurse -Filter jev_client.py -ErrorAction SilentlyContinue | Sort-Object LastWriteTime | Select-Object -Last 1).FullName

Three ways to call it:

Inline, one-off:

python "<skill-dir>/scripts/jev_client.py" --state "Customer: my Stripe connection keeps failing after 3 days" \
  --noul urgent="Is human intervention needed right now?" \
  --choice team="Which team handles this?:technical,billing,sales,other"

From a JSON file or stdin (preferred for anything with criteria descriptions or multiple questions — see references/api.md for the full schema):

python "<skill-dir>/scripts/jev_client.py" request.json
cat request.json | python "<skill-dir>/scripts/jev_client.py" -

Bulk (many items, different states) — write one request per line to a .jsonl file; runs in parallel (default 8 workers, -w N to change), prints one result per line in input order. A failed item prints {"error": ...} without stopping the rest. Transient 429/5xx/network errors retry automatically with backoff.

python "<skill-dir>/scripts/jev_client.py" items.jsonl -w 16

Measured: 12 tickets classified in ~1.6s wall total, $0.00018.

Write request files to the session scratchpad, not the user's project.

For Python-side batch loops, import it directly instead of shelling out per item:

import sys; sys.path.insert(0, "<skill-dir>/scripts")
from jev_client import decide
decide(state, questions)  # returns parsed JSON dict, raises JevError
decide_many([{"state": s, "questions": q}, ...], workers=8)  # parallel, per-item {"error"} on failure

Ready-made patterns

Read references/patterns.md for the exact request shape and how to act on each answer:

1. Rerank search/research results before fetching (fetch only the top few)
2. Pick the next element in browser automation (+ irreversibility check)
3. Review a diff before commit (plain rules first, then risk per hunk)
4. Triage failing tests / logs (real bug vs flaky vs environment…)
5. Verify citations after research (claim vs fetched source)
6. Screen external content for prompt injection / exfiltration
7. Triage PR review comments (must-fix / nit / question / outdated)
8. Find duplicates (pre-filtered pairs only)
9. Organize files into folders (propose, user confirms, never auto-move)

Question-writing rules (from TypeSafe's official skill + cookbooks)

• One narrow judgment per question; put several independent questions about the same state in ONE request (13 batched ≈ 12× cheaper, 10× faster than 13 calls).
• Put structured context in state as named JSON fields and reference them in instructions with backticked paths, e.g. ticket.messages[0].text.
• Always include a no-match option (none/unknown) when nothing may fit. JEV can't pick a value you didn't list.
• Never ask JEV to count or do math: ask one noul per item and sum probabilities in code.
• For ranking with several criteria: ask each criterion as its own score once, apply weights in code — reweight later without re-calling.
• Don't use JEV for things code does exactly: known rules, lookups, arithmetic.

Always batch every question you need about the same state into ONE request — JEV answers all questions in parallel in a single call, which is both faster and cheaper than N separate calls.

Reading the answer

• noul type → response has a probability 0–1 that the proposition is true. Don't hard-threshold at 0.5 for anything consequential; read references/api.md on calibration and adding an explicit "I don't know" option to choice/score criteria to avoid confident-but-wrong answers on out-of-distribution input.
• choice type → response has the chosen option plus a probability distribution over all options and a confidence score.
• score type → response has a probability-weighted position on the scale plus per-level probabilities.

Full field-level schema, request/response examples, and known limitations are in references/api.md — read it before hand-writing a raw request.

What NOT to build

Don't reimplement a local model (that's what Rizzo Flow / SemIf / other open clones are for — different tool, different tradeoff: this skill talks to the real hosted JEV over the network). Don't add retries/caching/queueing beyond what jev_client.py already does unless the task actually needs it.