Scene capabilities: the action catalog
The complete set of actions an agent can run against a published scene — visibility, selection, camera, animation, colour — and the rules that govern them.
A published scene is not a viewer with an assistant bolted onto it. The assistant operates the scene through a fixed, enumerable set of actions — the same set every time, against targets that were resolved before the action ran. This page is that set.
The mental model worth having: the agent never touches your geometry directly. It emits operations. Each one names an action, a set of already-resolved mesh targets, and an optional value. The viewer applies it and acknowledges it.
The shape of an operation
Every operation carries the same fields, whatever the action:
Field | What it holds |
| The canonical verb — one of the tracks below |
| Concrete mesh keys, resolved before the action runs |
| Where the verb takes one: a colour, an animation name |
| Action-specific extras: viewpoint, framing, magnitudes |
Targets resolve once, up front. An action never re-derives what you meant from your sentence — by the time it runs, the question of which parts is already closed.
Visibility
Action | Meaning |
| Make targets invisible |
| Make them visible again; ghosted targets return to solid |
| Translucent, still in place |
| Show only the targets, hide everything else, and frame what is left |
| Restore the visibility that preceded an isolate |
There is no “hide everything except” action, and the omission is deliberate. Isolate takes the set you want kept and hides the rest, so the thing you name is always the thing you get rather than the thing you have to describe by exclusion.
Selection
Action | Meaning |
| Replace the selection |
| Add targets to it |
| Empty it |
Selecting a hidden part shows it. Hiding a selected part deselects it. The coupling runs in both directions on purpose, so selection and visibility can never disagree about what you are actually looking at.
Camera
Action | Meaning |
| Frame named targets — the default “look at this” verb |
| Frame whatever is currently selected |
| Frame the whole assembly |
| Orbit for a read on all sides |
| Tighten onto targets |
| Reposition, including named viewpoints |
| Relative motion at canonical magnitudes |
| Return to the default view |
Named viewpoints are top, bottom, front, back, left, right and three_quarter. “Underside” resolves to bottom.
“Zoom in” tightens the framing. It does not touch the lens. A request for a closer look moves the camera rather than narrowing the field of view, because narrowing the field of view distorts the product you are trying to show.
Animation and state
Action | Meaning |
| Play a named animation |
| Playback control |
Animations are retrieved by the name you gave them when authoring, which is the practical reason naming matters: an animation called “Animation 3” can be played on request but is very hard for anyone to ask for.
Colour
Action | Meaning |
| Set target colour — the value is free-form |
| Restore original colours |
Free-form means the value does not have to be a hex code or a named CSS colour. “Brass-ish” resolves. Resetting colours is scene-wide by default, and it resets only colour — it is not a general scene reset in disguise.
Callouts
A callout.set places a bounded spatial annotation on a target: pay attention to this part, for this reason. It is not the answer — the answer is the reply — it is the marker that goes with it. Warning callouts ship today. A callout whose target cannot be resolved does nothing rather than leaving a card floating in space.
Scene-level
Action | Meaning |
| Restore the whole scene to its initial state |
| Exploded view |
How things get referred to
Four referent forms cover essentially every request:
Form | Example |
A named target | “the boiler”, “the rear hub” |
The current selection | “them”, “those”, “the selected parts” |
The whole scene | “everything”, “the assembly” |
A carve-out | “all of it except the sensors” |
A carve-out that cannot be grounded is refused rather than approximated. If the agent cannot work out what “the sensors” means in your model, it says so — it does not fall back to matching the word against part names and hoping.
Answers, not just operations
Alongside the actions above, the agent reads a searchable corpus built from the scene: the part and group names you confirmed, the descriptions you wrote, and any manuals, spec sheets or service bulletins attached during authoring. That corpus is why two scenes built from identical geometry can differ enormously in what they can answer.
What it will not do
Invent a target. If the request does not resolve to real meshes, nothing runs.
Act on half a carve-out. Partial grounding is treated as no grounding.
Produce a confident answer from nothing. Where the scene has no basis for an answer, saying so is the correct output.