jev-router"

Uses the Jev System One model to choose which OpenCode subagent should handle an ambiguously-routed request, returning a typed choice with a calibrated confidence. Invoke when the correct subagent is

SKILL.md

GitHub

Jev Router

Picks which subagent handles a request. It never decides what to build — that stays with the calling agent and the approval gate.

Jev is a System One model: it evaluates a state against typed questions and returns typed values with probabilities. It does not generate text, so there is nothing to parse and no way for it to invent a subagent that is not in the criteria you listed.

When to invoke

Only when routing is genuinely ambiguous:

  • Two or more subagents plausibly apply and you cannot justify picking one
  • A previous delegation landed in the wrong place and you need a second opinion
  • The request is cross-cutting (e.g. "fix the bug and document it")

Skip it when the task maps cleanly to one subagent. A confident answer to a question you already knew the answer to costs ~0.8s and nothing else. On a clear-cut task this skill is pure overhead.

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 request, verbatim, plus what you already know about it>",
    "questions": {
      "subagent": {
        "type": "choice",
        "instructions": "Which subagent should handle this request?",
        "criteria": {
          "subagents/code/coder-agent": "Writing or changing source code files",
          "subagents/code/test-engineer": "Writing or fixing tests",
          "subagents/code/reviewer": "Reviewing code for correctness or security, without changing it",
          "explore": "Finding code, tracing flows, answering questions about the codebase — no changes",
          "subagents/meta/external-scout": "Upstream docs, library APIs, dependency source — outside this repo",
          "direct": "Small change you can do yourself without delegating",
          "none_of_these": "None of the listed options clearly fits"
        }
      },
      "needs_approval": {
        "type": "noul",
        "instructions": "Should a human approve the plan before any edit happens?",
        "criteria": {
          "true": "The change is risky, destructive, or hard to review",
          "false": "The change is routine and reversible"
        }
      }
    }
  }'

Reading the answer

{
  "answers": {
    "subagent": {
      "type": "choice",
      "choice": "subagents/code/coder-agent",
      "confidence": 0.94,
      "probabilities": {
        "subagents/code/coder-agent": 0.94,
        "subagents/core/documentation": 0.04,
        "direct": 0.02
      }
    },
    "needs_approval": { "type": "noul", "noul": 0.81 }
  },
  "usage": { "input_tokens": 340, "output_tokens": 60 },
  "cost": "0"
}

confidence is not probabilities[choice]. It is a separate number derived from how spread out the whole distribution is. A single option at 0.4 while the rest are at 0.3 is a confident answer in the wrong direction; 0.94 with the rest near zero is a confident answer. none_of_these winning — or landing second close behind — means the taxonomy missed; treat it as < 0.5 below regardless of the scalar.

Apply this policy:

confidence What to do
≥ 0.8 Delegate to the chosen subagent
0.5 – 0.8 Do not guess. Present the top two options to the user and ask.
< 0.5 Ignore the answer. Your own reading of the task was probably better.

needs_approval at or above 0.8 means the approval gate applies even if the confidence in routing is high — routing confidence and risk are separate questions, and that is why they are asked as two.

Constraints

  • One atomic question per judgment. Measured on this machine: a compound question about a command sequence returned noul=0.22, while the same judgment asked atomically returned noul=0.99. If a question needs the word "and", it is two questions.
  • jev-1.13-free is temporary. The OpenCode catalog marks it as available for a limited time. If the call returns 404 or an unknown-model error, fall back to routing by your own judgment and say so. Never block a task on this.
  • No network means no answer, and that is fine. On a null response, route by hand. This skill accelerates a decision; it is not allowed to prevent one.
  • Never let Jev pick what to build or whether to build. Scope comes from the user. Routing is the only thing on the table here.

Why a separate call instead of just deciding

Because the interesting case is the one where your own first instinct and the evidence disagree. When they agree, this skill is noise. When they do not, a calibrated choice with a full probability distribution is a better basis for asking the user a useful question than a coin flip between two plausible options — and confidence tells you when the question itself is not worth asking.