hatch-pet

hatch-pet is an OpenAI/Codex skill (Apache-2.0) for creating, repairing, validating, visually QAing, and packaging Codex-compatible animated pets and pet spritesheets. It solves the problem of turning a concept, brand cue, company/prospect name, or reference image into a complete 8x9 animated pet atlas with transparent unused cells, QA contact sheets, and pet.json packaging. It combines an installed $imagegen system skill for visual generation with bundled deterministic Python scripts for spritesheet assembly, so creative generation and reproducible geometry stay separated.

The workflow accepts optional user inputs and infers a pet name, description, and base art when omitted (generating from text first, then reusing that base as the canonical reference for each animation row). All normal visual generation is delegated to $imagegen rather than direct image-API calls, and the skill's own scripts handle deterministic tasks: preparing layout guides and prompts, mirroring an approved running-left row, extracting frames, validating rows, composing the final atlas, and building contact-sheet and motion-preview QA media. It includes explicit storage controls to keep image generation isolated in lightweight workers and avoid bloating rollout payloads, plus an optional brand-discovery subagent that uses web search of official brand sources when only a company or product name is given. Reference files document the animation rows, the Codex pet contract, and a QA rubric.

It targets Codex users, brand and marketing teams, and creative developers who want a custom or mascot animated pet without hand-building sprite atlases. Operations are creative asset generation, local file assembly, and optional read-only web research; a CLI fallback path for image generation requires local API credentials and explicit user confirmation.