hatch-pet - SKILL.md Agent Skill

name: hatch-pet description: Create or update an Omnibot / Xiaowan custom pet from a concept, reference image, or bare company name. Use when the user wants a selectable pet package, animated 8x9 sprite atlas, brand-discovery-driven mascot, or static fallback pet.

Hatch Pet

Build a pet package that shows up in 外观设置 > 宠物 and is written to:

/workspace/.omnibot/pets/<pet-id>/pet.json
/workspace/.omnibot/pets/<pet-id>/spritesheet.webp

Prefer the animated atlas path. Static files such as current.svg or current.png are allowed only when the user explicitly asks for a static pet or when image_generate is unavailable after you have told the user.

Never write the selectable pet package under a desktop Codex directory such as ${HOME}/.codex/pets, ~/.codex/pets, or /workspace/.codex/pets.

If the user gives 宠物名称 (瀹犵墿鍚嶇О), use that exact name as displayName, and use the same name for the folder/id after replacing only path-unsafe characters such as /, \, :, *, ?, ", <, >, |, and whitespace.

Animated pet.json example:

{
  "id": "mimibear",
  "displayName": "mimibear",
  "description": "像素风小白熊，圆润可爱。",
  "spritesheetPath": "spritesheet.webp"
}

Generation Provider

Codex uses $imagegen for image creation, but Xiaowan does not have a desktop Codex $imagegen skill. In Xiaowan, use the image_generate tool for normal visual generation. It calls the configured OpenAI-compatible image provider, decodes the returned image, and writes a real PNG/WebP/JPEG file into the workspace.

Use file_write only for manifests, pet.json, static SVG fallbacks, or copying already-generated base64 image bytes. Do not use file_write to hand-write fake raster art when image_generate is available.

If no image-capable provider is available, fall back to a static SVG/PNG pet only after telling the user that animated image generation is unavailable. Do not fake an animated pet with placeholder files.

image_generate requires an OpenAI-compatible image credential. Xiaowan's default image endpoint is https://cloud.omnimind.com.cn, and the default image model is gpt-image-2. For first-party builds, the app can bundle the company image credential through the build-time OMNIBOT_IMAGE_API_KEY secret so users do not need to enter a key. Do not write API keys into skill files, pet manifests, prompts, or generated artifacts.

For animated pets, generate the same Codex-compatible atlas contract:

one base reference image
9 animation rows: idle, running-right, running-left, waving, jumping, failed, waiting, running, review
final atlas size 1536x1872, arranged as 8 columns x 9 rows with 192x208 cells
transparent background and fully transparent unused cells

Xiaowan can display this atlas dynamically: after the package lands under /workspace/.omnibot/pets/<pet-id>/, the settings page uses a generated preview for the pet list, while the floating pet uses the original spritesheet.webp or spritesheet.png for animation playback.

State row rules mirror the Codex hatch-pet contract:

idle: subtle breathing, blink, or tiny body bob only; the loop must visibly move.
running-right and running-left: directional drag movement, facing the correct direction; no speed lines, dust, shadows, or motion trails. Generate running-right first and mirror running-left only when identity and prop placement still make sense.
waving: show the wave through limb pose only; no floating wave marks or symbols.
jumping: show vertical body motion only; no floor marks, dust, or shadows.
failed: sad/error emotion through pose or attached opaque effects only; no detached symbols.
waiting: expectant help/approval pose, distinct from idle.
running: active task work or thinking motion, not literal foot-running.
review: focused review pose; avoid new props unless the base pet already has them.

References

Read these when needed:

references/codex-pet-contract.md
references/animation-rows.md
references/qa-rubric.md

Workflow

If the user gives only a brand, company, product, or prospect name, run a lightweight brand-discovery subagent first. Search the web narrowly, prefer official sources, and extract palette, tone, motifs, and mascot-safe cues. Save the brief before generation.
Prepare the run with scripts/prepare_pet_run.py. Pass the chosen name, short description, pet notes, references, style preset, and brand-discovery file when relevant. Let the script create the layout guides, prompt files, and imagegen-jobs.json.
Generate visuals with image_generate. Base first, then each row strip. Use the prompt file as the prompt, set outputPath to the job's decoded output path or a temporary workspace image path, and use model: gpt-image-2, format: png, background: transparent, and size: 1536x1024 for row strips when supported. If a job lists input images, summarize their identity/layout constraints into the prompt until Xiaowan has an image-edit/reference tool. Respect the transparency rules: flat removable background, no scenery, no labels, no detached effects, no shadows, no glow, no dust, no motion trails.
After selecting each output, copy it into the decoded path and mark the job complete. Derive running-left with scripts/derive_running_left_from_running_right.py only when mirroring preserves identity, cadence, and prop placement.
Run the deterministic post-pass scripts in order: extract_strip_frames.py, inspect_frames.py, compose_atlas.py, validate_atlas.py, make_contact_sheet.py, and render_animation_previews.py. Use stable-slots only if QA shows extraction-induced popping and the source strip itself was already stable.
Visually QA the contact sheet and preview GIFs. Repair the smallest failing row only. Keep the same silhouette, face, palette, materials, props, and style across all rows.
Package the final pet into /workspace/.omnibot/pets/<pet-id>/ with pet.json plus spritesheet.webp. If you must fall back to a static pet, write imagePath with current.svg or current.png instead.

Pet Rules

Keep the body compact and readable at 192x208 per cell.
Preserve identity across every row.
idle must stay calm and low-distraction.
running is task-work motion, not literal foot-running.
waving, jumping, failed, waiting, and review each need distinct semantics.
For brand-only requests, use the discovery brief as a mascot-safe cue set. Do not copy logos, readable marks, UI screenshots, slogans, or text.
Preserve Chinese pet names when the user provides them. Only remove path-unsafe characters and collapse whitespace as needed.

Output Rules

Do not create files outside /workspace/.omnibot/pets/<pet-id>/.
Do not stop after writing only the art file. pet.json must exist next to it.
Do not create a top-level /workspace/.omnibot/pets/current.svg or /workspace/.omnibot/pets/spritesheet.webp for a named pet.
If writing an image through file_write, supply real image bytes or a valid data URI/base64 payload.
If the pet is animated, pet.json should use spritesheetPath. If the pet is static, use imagePath.

Failure Handling

If any required hatch-pet write fails for /workspace/.omnibot/pets/<pet-id>/spritesheet.webp, /workspace/.omnibot/pets/<pet-id>/current.svg, /workspace/.omnibot/pets/<pet-id>/current.png, or /workspace/.omnibot/pets/<pet-id>/pet.json, report the exact failure and do not switch to an unrelated path.