A friendly, LLM-ready sim layer. Objects say what they can do, characters carry things, look where you point, and can even fail — and a swappable "mind" decides, scripted today, a language model tomorrow. Meet CATAPHRACT.
Needs decay, the character tends the most urgent, the need refills, and on to the next — all discovered from what objects advertise. The exact same objects answer typed commands.
The Sims "smart object" pattern: a fridge offers eat, a bed sleep, a cup grab. Characters discover what's possible — no per-object code.
Verbs like pull, clean, serve and watch tv are just more rows on the object; each row (an InteractionOffer) carries its animation, duration, reach, the need it refills, and any flags.
Everything a director wants runs through one small verb API — exactly what a language model would call. Gaze and locomotion are separate channels, so she can look at you while walking to the couch.
Sleep, hunger, entertainment decay over time. In the sim_house demo one character autonomously cycles fridge → couch → bed, tending whichever is lowest — the same objects the command layer uses, driven instead by a utility brain.
An offer can carry an attribute check. Threshold gates on a stat — drag a bed only if you're strong enough. Chance rolls the dice, tilted by luck — carry food across the room without dropping it.
Fail and nothing happens: she says so, and the world hears an ActionFailed signal it can react to.
A successful grab lifts the object into her hand and it rides along as she walks. A full cup is kept upright so it won't spill; an empty one she carries any old way. Too big to grip and the grab fails.
Her arms hang relaxed at her sides until she reaches — the reach IK is layered on top of the locomotion pose, so it starts natural and returns natural.
"Look at the couch" doesn't just crank her neck. Her head keeps a comfort cone; past it the body eases around to meet the target — and while seated or laying she re-poses more gently.
Look neutral settles her gaze back forward and lets the body relax to its resting facing.
She can put things down precisely — and the grid appears as a helper so you can see exactly where.
Ask, and a snap grid lights up on the floor: cell lines, a marker at every object's center point, and the target cell highlighted. Say "grid finer" / "coarser" to change density for fine placement.
"Move the couch to the bed", "push the fridge to the middle", or explicit coordinates "push the couch to 2 3" (x z) — even "2 1 3" (x y z). She walks behind the object and slides it onto the snapped cell.
Pushing checks footprints: she won't shove one object through another, and stops with "something's in the way." A step-up assist lets her climb stairs to a second floor when the nav mesh routes her there.
Objects can carry their own little grids — the levels of a shelf, the zones of a desk. Address a slot by area and cell: "place at shelf, area 3, 2 1" puts the carried item on the third shelf, column 2, row 1.
Pick who you're addressing from a dropdown, type, and only the right character listens — the sim character you command, the companion CATAPHRACT, or the floating POD that shadows her.
The console tags every line with a speaker and fires Submitted(speaker, text). Each brain picks up only its own lines, so telling CATAPHRACT to go "defensive" never makes the sim character try to interpret it — the old cross-talk that broke stances and the help action is gone.
The same box drives all three. Change the dropdown, keep typing.
Coordination is a tiny three-step handshake anyone can join. CATAPHRACT broadcasts a task onto a shared bus; the POD (or anyone) answers with help and attends to the requester; when it's handled the request clears and the initiator finishes the action.
broadcast eat fridge — CATAPHRACT posts a help request (15s) to the log.
help from the POD — it shadows the requester and logs "listening to Cataphract".
The request clears from the bus and the initiator completes the action.
A line flows CommandConsole → (CommandBrain | CompanionBrain) → verbs. The console only tags who is speaking; each brain turns its own lines into verb calls or stance changes, reading the scene from that character's own vantage — bearings, distances, only the verbs actually available, never god-mode. Swap a brain for a language model and nothing else moves: the bodies, objects, grid and broadcast bus stay exactly as they are.
Asking for a hand is a real, two-part decision the helper owns — and a crew of them can crowd an object without stepping on each other.
Ask "help pull the fridge" and CATAPHRACT first decides when to walk — her follow pauses (her movement doesn't), she rounds the furniture, and braces in position. Then she decides when to try: after a beat solo, once enough hands are up for a co-op lift, or on your go-signal — "try" / "now" ("wait" holds her). Ask "who's helping?" for the live roster.
Objects carry role stations — the corners of a car, the ends of a couch — so helpers fan out to their own spot and play that spot's animation. Foot IK plants their feet on stairs and uneven ground, a footprint-aware standoff keeps them beside what they push (not clipping in), and personal-space steering means two of them never shove into each other. Danger zones are avoided outright.
She picks up the cup with a real two-bone IK grab — the hand lands on the object — and holds it with a pose while the IK keeps it in her grip. Her eyes track you with look blend-shapes layered over the head-turn, and her face blends Happy / Sad / Angry from mood. Author those expressions visually in the Expressions dock.