Tools & Plugins / Tools

TypeSafe Jev

TypeSafe Jev is an evaluation model. It takes state plus typed questions and returns calibrated probabilities. It cannot chat, write text, or call tools, so it is not a chat model in Genesis. The bundled typesafe plugin uses it in three opt-in ways:

• Tool router. Inside a browser or computer-use loop (the latest tool result came from browser or a computer-use style MCP server such as computer-use, desktop, or claude-in-chrome), Jev checks whether the next step can run without an LLM call. One request decides the whole step.
  • Browser: Jev reads the numbered elements ([ref=eN]) from the latest snapshot and picks a fresh snapshot or the element to click.
  • Computer use: MCP use tools whose arguments are all enum, const, or boolean (for example a screenshot) run with Jev-picked arguments. Coordinate clicks, typing, and navigation stay with the main model.
  • Otherwise, and on low confidence, errors, or timeouts, the main model is called exactly as it would be without the router. The router never changes tool_choice, thinking, or the tool list, so provider prompt caches keep hitting. Outside use loops Jev is not called at all.
• Exec safeguard. Before every exec call, Jev rates the command's system impact as high, medium, or low. A high verdict at or above blockConfidence blocks the command; medium (or a less certain high) asks for approval; low leaves it alone. It runs alongside the built-in tools.exec.safeguard heuristics and approvals and only ever tightens them. On Jev errors or timeouts the command runs under the normal exec rules.
• jev_evaluate tool. Lets the agent ask Jev boolean, choice, and score questions directly.

Backends

backend: "auto" (default) uses the first backend with a key, in the order above. Vercel and OpenRouter reuse the keys you already configured for those providers.

Configure

In the Control UI, open Config → Plugins → TypeSafe Jev, pick the backend, and switch Enable Jev tool router on or off. Turning the router on needs a gateway restart (the Control UI offers one when you apply). Turning it off, switching backends, and changing thresholds apply to the next model call. Enable Jev exec safeguard works the same way: on needs a restart, off and threshold edits apply to the next exec call.

{
  plugins: {
    entries: {
      typesafe: {
        config: {
          backend: "auto", // or "typesafe" | "vercel-ai-gateway" | "openrouter"
          jevRouter: {
            enabled: true,
            minConfidence: 0.6, // below this, the model decides
            maxConsecutiveDirectCalls: 8, // model-free tool calls in a row per run
            timeoutMs: 3000,
          },
          jevSafeguard: {
            enabled: true,
            blockConfidence: 0.8, // below this, a high verdict asks for approval instead
            timeoutMs: 5000,
          },
        },
      },
    },
  },
  // jev_evaluate is optional, so allowlist it explicitly.
  tools: { alsoAllow: ["jev_evaluate"] },
}

The router works with any main model and provider.

Related

• Plugin hooks: before_model_call
• Vercel AI Gateway
• OpenRouter