Ask Jev

TypeSafe Jev is a System One decision model, not a chat model. It returns typed Choice / Score / Noul answers with probabilities. Use it as a tiebreaker when two or more paths are still plausible after tools and user intent are accounted for. Code (this skill's script) owns the workflow; Jev only supplies the judgment.

Do not use Jev to generate text, plans, or explanations. It will not.

When to Use

Load and call this skill when:

• Two or more next actions, implementations, classifications, or candidates are still reasonable.
• You are about to pick arbitrarily (coin-flip, "feels slightly better," first option in a list).
• You need a calibrated yes/no (sufficient evidence? on-topic? blocked page vs empty?) before branching.
• You want a second axis: what (choice/score/noul) plus whether to act (confidence / grey-zone).

Don't use for:

• Facts you can look up, compute, or read from a file in one tool call.
• Anything the user already decided.
• Safety, policy, legal, or permission questions (Hermes rules win; Jev cannot override them).
• Open-ended reasoning, writing, or multi-hop plans — decompose or just think.
• Secrets, credentials, or raw private payloads as state.

If a tool answer in ~30s would settle it, skip Jev.

Prerequisites

• TYPESAFE_API_KEY in $HERMES_HOME/.env (reload: the script reads that file if the var is unset).
• Disable (any one is enough): set ASKJEV_DISABLED=1 in the same .env, or delete this skill directory.
• No extra packages. Script is stdlib-only against POST https://api.typesafe.ai/v1/systemone.
• Default model: jev-latest (override with TYPESAFE_DEFAULT_MODEL or --model).

How to Run

Resolve the skill dir from $HERMES_HOME (fallback ~/.hermes):

ASKJEV="$HERMES_HOME/skills/autonomous-ai-agents/askjev/scripts/askjev.py"

Canonical: write a payload JSON, then terminal:

terminal(command="python3 \"$HERMES_HOME/skills/autonomous-ai-agents/askjev/scripts/askjev.py\" --payload /tmp/jev.json")

Payload shape (same as TypeSafe HTTP):

{
  "state": {"goal": "...", "options_context": "...", "constraints": "..."},
  "questions": {
    "path": {
      "type": "choice",
      "instructions": "Which next action should Hermes take?",
      "criteria": {
        "a": "...",
        "b": "...",
        "other": "None of the named options fit"
      }
    }
  }
}

Helpers:

python3 "$ASKJEV" choice --state '...' --question 'Which path?' --option curl="static page" --option browser="JS app"
python3 "$ASKJEV" noul --state '...' --question 'Is evidence sufficient to answer without more tools?'
python3 "$ASKJEV" score --state '...' --question 'How on-topic is this source?' --level 'off' --level 'partial' --level 'direct'

Stdout is one JSON object: answers, verdicts (act/escalate), usage, model. Exit 0 all gates act; 3 at least one escalate (answers still present); 2 disabled/missing key; 1 request error.

Live eval (does not change config):

terminal(command="python3 \"$HERMES_HOME/skills/autonomous-ai-agents/askjev/scripts/eval.py\"", timeout=300)

Procedure

1. Confirm it is a judgment. Lookup / user intent / policy → do that instead. Done when you can name the fork in one sentence.
2. Pack state as named JSON fields, not a vibes paragraph: goal, constraints, evidence, the options under consideration. No secrets. Include an other / none choice when the list may not cover it.
3. Ask atomic questions in one call. Independent dimensions = separate questions (they run in parallel). Do not cram a plan into one Choice.
4. Run scripts/askjev.py. Completion: JSON on stdout with verdicts.
5. Gate on verdicts, not vibes.
   • Choice/Score gate=act (default min confidence 0.6): take decision.
   • gate=escalate: gather more state, ask the user, or pick a reversible default. Do not silently take the argmax.
   • Noul yes (noul>=0.7) / no (<=0.3) / uncertain (grey). Uncertain = escalate.
   • Destructive or hard-to-undo work: require confidence >=0.85 or user confirmation even if act.
6. Log one line in your reply when Jev moved the decision: question, decision, confidence/noul, whether you acted. Skip the essay.

Pitfalls

• Jev does not explain itself. If you need a rationale, you already have it in state; do not prompt Jev for prose.
• Docs examples can disagree with the current jev-* version. Trust live confidence more than a remembered label.
• Low confidence often means the options overlap or the state is thin, not that the model is broken. Tighten criteria (what / not_for) or split the question.
• A multi-issue ticket can still come back as one Choice with high confidence (Jev picks a primary). Detect several labels with separate Noul questions, not by waiting for Choice confidence to drop.
• Score can land on the right level with mediocre confidence when probability leaks to neighbors. Gate on both the score and confidence for high-stakes branches.
• Noul has no confidence field. A value near 0.5 is a coin-flip, not "medium yes."
• Question ids are not sent to the model. Put the full meaning in instructions + criteria.
• Do not install typesafe-sdk just for this skill. HTTP via the script keeps the whole thing one directory.
• Rate limits: 429 / 529 — the script retries with backoff. Persistent failure → proceed without Jev and say so.

Verification

• python3 scripts/askjev.py --payload ... returns model starting with jev- and one answer per question.
• ASKJEV_DISABLED=1 python3 scripts/askjev.py --payload ... exits 2 and does not call the network.
• python3 scripts/eval.py reports per-case pass/fail against live Jev. Treat a failing expected label with high confidence as a real reliability issue; a failing label with low confidence is an escalate-path success if the eval marks it that way.