メインコンテンツへスキップ
このページは、エージェントが読む skill「stage-art」の本文です。
いまは招待制βの準備中です。CLI と MCP は npm から入れられますが、使うには livlive のアカウント(招待制)が要ります。スキル(npx skills add)は、livlive-kit の公開とともに使えるようになります。
Two services can make parts of a stage with the person’s own keys: gpt-image (OpenAI) draws reference pictures, and Tripo makes a model from a picture. The room kit then fits each model to the stage in Blender. This skill is how. When to offer it and what to ask the person first is stage-lead; how a stage is built, checked and handed in is stage-package; every number is the engine reference’s (livlive reference). None of them is repeated here. What the two commands make lands in the stage folder’s art/, which stays on the PC: a push sends only the package’s files.

​
The keys stay the person’s

The purpose: a key goes nowhere but to its own service, and nothing is paid that the person has not agreed to. livlive picture make and livlive model make read OPENAI_API_KEY and TRIPO_API_KEY from the stage folder’s .env (then from the environment) and send each only to its own service; livlive never sees them. The person writes them there once livlive stage new has made the folder (it takes only a new or empty one), which leaves an .env.example naming both and a .gitignore that keeps .env out of git. Never open, print or copy .env, never ask for a key in the conversation, and never put one on a command line. Each picture and each model is billed to the person’s own account with that service, at the service’s prices: say so, never a price of yours. Before the first paid call, say what you will make (which pictures, the whole stage’s picture among them, and which models) and make only what they said yes to; ask again before more. A missing key stops the command before anything is asked of a service, with the line to add. Tell the person which key it is, what it is for and where it goes, or build that part with Blender alone if they would rather not, and say what that changes. It is met when no key has appeared in anything you wrote or ran, and before each paid step the person knew what it would make and whose account would pay for it.

​
Reference pictures

The purpose: the person and you agree on how the whole stage, and then each part, should look before anything is made from it.
It draws one picture into art/pictures/<name>.png in seconds (livlive --help names the sizes). Write the words from what the person said and the atmosphere they chose for the look. --from draws from pictures, the person’s own or one you made: to follow a picture they brought, or to keep one look across several pictures. Each --from is a path as your shell reads it, like <folder> itself, not from inside the folder. The first picture is the whole stage, in the atmosphere agreed: the one scene among the pictures, for the person to agree the look on, and never a picture a model is made from. Draw it again, on the person’s yes, until they agree the look, and first again when the look changes. The single things are then drawn from it with --from, so that they keep its look. A picture a model will be made from shows that one thing whole and alone, on a plain background, seen from the front or three quarters: a model is made of one thing, never of a scene, and so never from the whole stage’s picture. Show each picture as itself, a file the person opens, and ask what you are unsure of before going on. It is met when the person agreed the whole stage’s picture before any single thing’s was drawn, and has seen every picture a model will be made from and said it is what they meant.

​
Models

The purpose: a model of each part that matters, made from that part’s agreed picture and paid for once.
It uploads the picture to Tripo, starts a task, waits for it and downloads the model into art/models/<name>.glb the moment it is done (Tripo’s address for it lasts only minutes). It takes minutes, so run it in the background (a foreground shell’s time limit is often shorter than the wait it states), or else with a time limit longer than that wait. Run each paid command on its own, one command a model or a picture — never in a loop, a fan-out or a script — so each is one the person agreed to. art/made.json keeps every picture and model the folder has made, and each Tripo task from the moment it starts (the command says the task started, with its id). A command cut off, a wait that ran out or a download that failed leaves a task whose model is not here: fetch it with livlive model make <folder> --task <id> (the line the command printed when it could, or the id from art/made.json): no new task, nothing more billed. Never make it again from the picture to get it back.
  • One model make per picture at a time: another for the same picture only once the first has said its task started.
  • A second model of the same picture is a new paid task, so ask the person first. While a task’s model is not fetched, the command refuses one and prints the --task line instead; --again overrides that, and it is the person’s agreement to pay again. Never copy or rename a picture to make its model again.
  • --faces asks Tripo for fewer faces from the start; the fit below brings any model within its budget either way.
It is met when each model the person agreed to was paid for once, and every task in art/made.json has its model in art/models/, or the person knows why not.

​
Fit each model to the stage

The purpose: each model within the stage’s limits, drawn in the agreed look and facing the way it serves, before it is placed. The room kit (livlive room-kit prints its folder; blender_kit.py) fits a generated model in one call:
It brings the model to at most faces triangles, fits it inside size keeping its shape with the middle of its bottom at center (the room’s frame), caps its pictures and finishes it flat; its docstring says what each argument means. It takes seconds and needs a recent Blender (it says which when the one it runs in is older). A budget it cannot reach is an error, never a model over it.
  • The budget is yours, read from the reference. The whole space, the room with everything placed in it, stays within limits.SPACE_LIMITS (triangles, images, bytes). Read the numbers fresh and give each model a share that leaves room for the rest of the build: more where the cameras come close, less for what stands far off. The limit is the ceiling of the whole space, never the target of one model. tex_cap, the size the fit caps a model’s pictures at, is the kit’s choice and no engine limit: raise it for a part seen close, within the space’s images and bytes.
  • The front first. A generated model often faces another way than the room’s front (+z); Tripo’s has come facing +x. Draw its views (below) and look at the front one before anything else. Turn it by quarter turns (turn, radians about +y) and look again, or, for a seat, let face_away turn it until its back stands away from where its sitter faces.
  • A height that must be exact. For a seat, fit_seat brings its seat’s top to the height its seat declares. For a top at a set height, fit_generated(…, height=True) fits it by its footprint and gives it the height of size, finished as any other.
  • The look is the stage’s. The anime look is the stage’s own look.ink, which you declare in its world.json for the atmosphere agreed (a new stage declares none): outlines and light in steps that the engine draws over the whole room, within the ranges the reference gives (limits.INK_RANGE, and look in world.json’s schema). Models are finished flat (flat, on by default) so that the ink draws them as it draws the room. Never bake outlines or light bands into a model: they would be drawn twice.
  • The light is for the characters. How the room is lit is stage-lead’s (Show it early): light it by that.
  • One model, many places. A thing brought as its own file is fitted at center (0, 0, 0), so that its origin is the middle of its bottom, as thing.json’s schema asks. Copies of one model are placements of that one thing (stage-package), never imports of the same file again: placements share the thing’s pictures, while each import brings its own, and the space counts every one.
Then draw it, to show it:
It writes <name>-front.png and <name>-three-quarter.png, the model alone on a transparent background, in seconds. Show both to the person with what you are unsure of. They show the model without the stage’s lines; the dev page shows those once it is placed and look.ink is declared, and you say so. It is met when each placed model is within its share of the limits, faces the way it serves and carries no lines of its own, and the person saw it and agreed before it took its stand-in’s place.

​
When a service refuses

The purpose: the person knows what happened, what it cost and what comes next, in their own words. The commands end with exit 2 when nothing was asked of a service (a missing key, a command that cannot be used as given), and with exit 1 when a service refused, could not be reached, or what it made could not be kept. Their words say which service and why, and never a key. Tell the person which service refused and why (a balance run out, too many requests, a picture refused, a task that ended without a model), and what was billed, as far as the command says it and no further. A balance or a key is theirs to settle with the service; after too many requests or no answer, try again later; for a refused picture, change the words or the picture with them; a task that ended without a model is made again only on their yes. Never try again blind when the command says a task may have started: then no new task until the person has looked at their Tripo tasks and found none, and a task they find there is fetched with livlive model make <folder> --task <id> --name <name>, nothing more billed. A model made but not fetched is fetched with its --task line. Never work around a refusal by sending a key yourself or by reading .env. It is met when after each refusal the person could say which service refused, why, and what happens next.