Scenario Chatgpt Pet Update

Scenario ChatGPT Pet: Update

Overview

An existing pet is its sprite sheet: a PNG or WebP the user downloaded from ChatGPT, a Codex pet folder (pet.json plus the sheet it names), or a Scenario asset id of a sheet made earlier. That file is the pet's identity here; there is no pet record to look up. The update never overwrites it: the result is a new package next to the run, with the same id and name unless the user changes them.

This skill runs the scripts and follows the sheet contract of scenario-chatgpt-pet-create, which must be installed; run each script by its path in that skill's scripts/ folder, from the user's working folder, and read the contract before generating a row or writing direction verdicts: pet_frames.py, pet_prepare.py, pet_build.py, pet_check.py, pet_preview.py, pet_package.py, and the contract. Generation follows that skill's loop and model order (model_openai-gpt-image-2-5-sunburst, then model_google-gemini-nano-banana-2-1, then recommend only when both are blocked); connection and scope: the scenario skill. If a sibling skill named here is missing from your available skills, ask the user to install it (npx skills add scenario-labs/skills --skill <name>); unattended, proceed from tool schemas and flag the gap.

Quick reference

Read the sheet first, then pick the smallest change that does the job. Every row below runs in a run made by pet_prepare.py init --from-split (step 3 of the example): it records the original sheet, which the build, the check and the package all need. Unattended, the approvals are your own judgment against the request, recorded in the run, and nothing is installed.

split refuses anything that is not 1536x1872 or 1536x2288: other art is a new pet, made with scenario-chatgpt-pet-create using that art as a reference.

Worked example: Pepper's stiff run, then a red scarf for next season

1. Get the sheet. The user points at ~/.codex/pets/pepper/: read pet.json (id, displayName, description, spritesheetPath) and copy the sheet into a work folder. From an asset id: asset_download with format: "png", then curl -L.
2. Read it. pet_frames.py split pepper.png --out work/split prints the version, a background key that contrasts with the pet, and pixel_grid (set when the sheet is pixel art with on/off alpha; the update then keeps that grid and palette). pet_check.py pepper.png --json-out work/before.json measures it: height per row against idle, jump lift, look-row drift. pet_preview.py contact and pet_preview.py gif show the current pet. Ask what should change if the request is vague, naming what the check found.
3. Redo the run. pet_prepare.py init --name Pepper --id pepper --description "<from pet.json>" --from-split work/split --only running-right,running-left --out work/run (every row of the table starts with an init --from-split like this; --only names the rows to generate). The match style keeps the sheet's look; each job's references are references/base.png (the idle frame, enlarged on the key) and references/rows/<row>.png (the current row). Each job maps onto one model_run: its references uploaded with upload_asset and passed as referenceImages (an array), its prompt verbatim, its sizes for the model, background: "opaque" on Sunburst, priced with dry_run first; download the result to the job's output. When the old row's motion or directions are what is wrong (a jump that never lifts, a run that does not alternate, a head that turns the wrong way), add --no-current-row to init so the old strip is not a reference, since the model copies what it is shown, and say what changes with pet_prepare.py note work/run --jobs <row> --text "..."; a redone look row still references the kept other half of the sweep. Then pet_frames.py extract work/run --job running-right, read the report and preview, and mirror it with --mirror-left or generate running-left the same way.
4. Rebuild. pet_build.py work/run --mirror-left. With the original sheet as the base (recorded by init), only the rebuilt rows change, scaled to the existing idle row's height, ground line and anchor; the edge cleanup touches only them. pet_check.py work/run/final/spritesheet.webp --run work/run must print ok: true; kept look rows only bring a warning that they were not reviewed, while a redone look row needs fresh verdicts in qa/direction-semantics.json (format in the contract) or the check fails.
5. Show and package. pet_preview.py gif work/run/final/spritesheet.webp --out-dir work/run/previews, show pet.gif, upload the sheet and both GIFs into a collection as the create skill does, then pet_package.py make work/run --id pepper. Ask before pet_package.py install work/run --force, which backs up the installed Pepper into the run before replacing it.
6. The scarf, later. A new look changes every frame, so it is a new generation grounded in the old pet: init --name Pepper --id pepper --description "<from pet.json>" --from-split work/split --change "add a red knitted scarf" --out work/scarf adds an edit job on the identity image. Approve the edited base, then run every row as scenario-chatgpt-pet-create does from its rows step (look rows included for a v2 sheet, with fresh direction verdicts). The old sheet still sets the pet's size, so the new pet does not grow or shrink.