Authoring a behaviour tree

A .vxbt is a behaviour tree as a file: a list of blackboard keys, a root node with its children in priority order, and the decorators and services attached to each one. The editor opens it on a canvas drawn top-down, with an execution-index badge on every node and the compiler's complaints listed beside it.

The document is read by the editor, checked by the importer and turned into a runnable BehaviorTreeTemplate by BehaviorTreeContentCompiler against the game's own action and sensor registries.

Everything a designer should be able to change without a programmer and without a recompile: which branch outranks which, what interrupts what, how long a cooldown lasts, which key a condition reads.

You do not want a file for a tree that only ever exists in one test or one sample. BehaviorTree's fluent builder makes one in code, and the compiler is the same one.

Double-click a .vxbt in the project browser. An empty file opens as a selector with one wait, so the tree compiles from the first frame rather than opening with two complaints about itself.

Nodes are drawn top-down, parents over the branch they own. Each box carries:

• A badge in the header — its execution index. That number is its priority: node 4 outranks node 9 because it is earlier in the pre-order walk, which is what "left to right, top to bottom" means. An author who cannot see it cannot reason about what a decorator interrupts.
• Decorator rows stacked above the body, in evaluation order.
• Service rows stacked below it, with their interval.

⚠ A decorator and a service are attached, not wired. There is no edge to draw, because an attachment is always exactly one edge to exactly one parent, can never be shared, and has no position of its own.

A composite's child order is the whole priority ordering of the tree, and it is stored in the file rather than derived from where the boxes sit. Drag a node onto a sibling to become its neighbour, or onto a composite to become its child; ↑/↓ on the selection moves it among its siblings.

⚠ Laying the tree out never changes what it does. Unreal derives child order from horizontal position, which makes three ordinary gestures dangerous: auto-layout silently reorders the tree, dragging a node six pixels to line it up changes which branch wins, and a merge that resolves two positions produces a tree neither author wrote with a diff showing only coordinates. Here Lay out writes positions and nothing else.

Keys are added, renamed, retyped and deleted in the panel. The type picker offers the six the runtime has and nothing else.

⚠ A rename rewrites every reference in the document, and the count is what the operation returns. That is why a file references a key by name and the compiled form by index: a rename that only changed the declaration would leave every decorator pointing at a key that no longer exists.

⚠ A delete leaves its references dangling, on purpose. Clearing them would throw away which key forty decorators used to read, which is exactly what somebody undoing a mistaken delete wants back. The compiler reports each one by name.

This is the payoff for the narrower abort rule. An observer reaches the siblings under its own parent composite and no further, which makes the region it can interrupt a subtree — and a subtree is a thing that can be drawn. Select a decorator with Aborts set to anything but None and the region is shaded on the canvas.

The selected node's settings are drawn from its declaration in BehaviorNodeSchema — label, type and tooltip — with no per-node editor code. A node type your project adds shows up in the search popup and gets an inspector for free.

Compile checks the tree and lists what it found. A tree whose sensors and actions are registered in your game's code will report those as unresolved, and that is not an error: laying out a tree before the code exists is the ordinary order of work, so those become remarks and the topology, the key references and the parallel's shape are all checked anyway.

The importer applies the same check at build time, and fails the build on everything except the names only a game can resolve.

A small tree as a file:

Loading one in a game — the tree is data, and the actions and sensors are yours:

⚠ The action registry is the resolver's, and the tree fills it. Two Waits with different durations are two actions — an action object carries its own settings and is shared by every agent — so the compiler registers one per distinct set of fields rather than one per type.

Adding a node type of your own, which the popup and the inspector then know about:

⚠ Declaring it is half the job; the compiler also has to be able to build it. The schema says what a node is called and what fields it takes, and Vixen.Ai cannot construct a type it does not reference — so a node whose implementation is elsewhere registers a factory too, and gets a BehaviorBuildContext that resolves a key name to the index the runtime uses:

AddDecorator and AddService are the same shape. The shipped nodes are matched first, so a factory that reuses a builtin name does nothing rather than silently changing what every existing file means; a type in the schema with no factory behind it is a diagnostic on the node, not a crash. Vixen.Ai.Perception is the worked example — its three nodes arrive exactly this way.

• Behaviour trees — what the tree does once it is running, and what an abort means.
• Perception — three nodes that live in another assembly and register themselves through the factories above.
• The blackboard — the six key types and what set-ness is.