Behavior trees

evaluateBehaviorTree runs a small reusable control-flow model over any product-owned node schema. It recognizes selectors, binary conditions, and leaves through a typed adapter; it does not prescribe JSON field names, actions, targeting, or world state.

Node projection

const result = evaluateBehaviorTree(context, productTree, {
  inspect(node) {
    if ('select' in node) return { kind: 'selector', children: node.select };
    if ('when' in node) {
      return {
        kind: 'condition',
        condition: node.when,
        then: node.then,
        else: node.else,
      };
    }
    return { kind: 'leaf' };
  },
  test: (ctx, condition, node) => conditionHolds(ctx, condition, node),
  evaluateLeaf: (ctx, node) => chooseProductAction(ctx, node),
});

The adapter can project classes, JSON objects, compact tuples, or generated nodes into the common view without converting the authored tree first.

Evaluation semantics

• A selector evaluates children in array order and returns the first result that is not null.
• A condition tests once and evaluates only its selected branch.
• A false condition without an else branch returns null.
• A leaf delegates entirely to evaluateLeaf and may return a result or null to signal failure/fallthrough.

Only null is the failure sentinel. A valid result such as false, 0, or an empty string still stops selector traversal when represented by TResult.

Determinism and safety

Child ordering is behavior. Preserve authored arrays and ensure test and evaluateLeaf read the same authoritative snapshot. If a leaf mutates state, later evaluation and replay can depend on traversal; it is usually clearer for leaves to return an order that the reducer commits afterward.

Evaluation is recursive and has no built-in depth or cycle guard. Validate that authored trees are acyclic and reasonably bounded before running untrusted content.

Product responsibilities

The product retains condition vocabulary, target selection, action reuse, cooldowns, spawning, memory, blackboards, entity rules, and fallback policy. The SDK guarantees only the selector/condition/leaf control flow and the exact short-circuit behavior described above.

Zonoid example

Zonoid authors hostile and allied NPC behavior as product trees. A guard can select “attack if a target is visible,” otherwise “follow its standing order,” otherwise “patrol.” The adapter evaluates those conditions against the current world and returns a product order; the universal reducer then resolves all NPC orders together with the player’s action.

Focused Zonoid demo board: a guard patrols while the player is hidden, then attacks after the player enters its line of sight.