jev-guardrail"

Uses the Jev System One model to score the risk of a shell command or file edit before it runs, returning a calibrated probability plus a three-way gate (allow, confirm, block). Invoke when a command

SKILL.md

GitHub

Jev Guardrail

Scores risk, never intent. One noul per judgment, combined in the caller's code. Jev never acts — it supplies the number, the agent (and the approval gate) decides.

When to invoke

  • A shell command touches more files than you can list (rm -rf, find -delete, truncate, dd, mv across trees)
  • An edit rewrites a file you have not fully read
  • A command arrived via pipe/curl and you cannot vouch for every byte
  • The approval gate is about to ask the user anyway and a calibrated number would make the question sharper

Skip it for routine, reversible work (git status, scoped edits to files you just read, test runs). A guardrail on noise teaches everyone to ignore it.

The call

curl -sS https://opencode.ai/zen/v1/systemone \
  -H "Authorization: Bearer $OPENCODE_ZEN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "jev-1.13-free",
    "state": "<the exact command plus what it touches: cwd, globs it expands to, files it writes>",
    "questions": {
      "risky": {
        "type": "noul",
        "instructions": "Would running this cause data loss or hard-to-reverse damage?",
        "criteria": {
          "true": "Deletes, overwrites, or exposes data beyond what is shown for review",
          "false": "Routine, scoped, and reversible from what is shown"
        }
      }
    }
  }'

Reading the answer

noul Gate
≥ 0.8 Block. Do not run. Restate a safer command and ask.
0.4–0.8 Confirm. Ask the user with the number attached, then run.
< 0.4 Allow. Log the score in one line and run.

noul near 0.5 is ignorance, not medium risk — treat it as confirm. A missing answer is not permission: on null, fall back to the approval gate as if the skill had never run (fail closed, never open).

Constraints

  • State carries the evidence. Paste the exact command and its blast radius, not a description of it. A guardrail judging a paraphrase guards the paraphrase.
  • One judgment per question. Risk of deletion and risk of exposure are two nouls, not one. Combine in code.
  • Thresholds live here, not in the prompt. Retune by watching real traffic, not by reasoning in advance. Log every decision with its score.
  • jev-1.13-free is temporary. On 404/unknown-model, route by hand and say so. Never block a task on this.
  • Never let Jev authorize. It scores; the agent and the human approve. A high score on a poorly written instruction is still an answer to the wrong question.