Decision models answer named, typed questions — they are not generative completions (no streaming, no tools, no free text) and do not ride the LLM provider path. They have their own connection (see decision-providers) and their own client.

The feature is off by default — it reaches a network endpoint — so it stays inert until decisions { enabled #true } opts in. Decision checks are used by permission rules.

decisions

Each decision is a named, reusable question:

decisions {
    enabled #true
    decision "is-this-safe" {
        type "noul"                        // probability the answer is yes
        instructions "Decide whether this shell command is safe to run."
        yes "The command is safe."
        no "The command is risky."
    }
    decision "pick-verb" {
        type "choice"
        instructions "Pick how the tool call should be treated."
        option "read-only" { description "Reads only; allow it." }
        option "destructive" { description "Changes state; deny it." }
    }
    decision "severity" {
        type "score"
        instructions "Rate the severity of the finding."
        level "info"
        level "warn" { description "Worth a look." }
        level "error"
    }
}

• type is required: noul (a yes/no judgment), choice (pick one labeled option) or score (rank ordered levels).
• instructions is required and must be non-empty.
• noul takes optional yes / no descriptions (both or neither).
• choice takes one or more option "label" { description "…" }.
• score takes one or more level children, lowest first; each level is either the wire text itself (level "…") or a local label with a description (level "warn" { description "…" }).

noul answers are probabilities in 0..=1 (not booleans), and confidence is the distribution’s concentration — not the chance the answer is correct. Threshold policy belongs to the caller (the permission check’s threshold).

ranking

ranking names which decision ranks the options the question tool asks about. It is cosmetic and off unless a decision is named:

ranking {
    decision "severity"
}

A scene can override the top-level ranking in its own block — see scenes.

The provider dialog

/decision-provider opens the only model-selection surface for decision models (name, endpoint, optional key, free-text model), because no registry enumerates decision models. It writes the decision-providers and decision blocks to connections.kdl.