--- name: cartoonos-production description: Prepare and execute a bounded CartoonOS mascot shot or Higgsfield production handoff from approved references, evidence and budget. Use for CartoonOS shot planning, prompt repair and Golden Shot review. --- # CartoonOS production Read the repository's AGENTS.md and [current docs index](../../docs/INDEX.md). Read [the Higgsfield playbook](../../docs/integrations/HIGGSFIELD_PLAYBOOK.md) for provider execution or benchmark work. Load only the relevant episode script, character pack and evidence, not the whole catalog. Identify the episode/shot, exact mascot versions, language, approved factual mechanism, original reference binaries, intended workspace, quote and approved spending cap. Missing reference files, unsupported claims, insufficient account credit or unpriced generation leave the handoff blocked. Do not infer approval from a candidate manifest or fabricate Element IDs and QA scores. Plan one readable primary action, explicit camera movement and continuity facts. Use approved keyframes before motion when they can resolve composition cheaply. Dialogue is reviewed es-419. Use original visual traits; do not imitate protected franchise styles. Learning must change a character's action. Inspect current model/tool constraints and estimate the exact payload before any paid execution. Respect existing user authorization and the approved cap; a skill is not spending or publication authorization. For an ambiguous provider POST outcome, retain the intent/reservation and reconcile; do not retry blindly. Local deduplication does not make the current public API exactly-once. Export each completed output to owned storage, verify checksum and inspect actual media before marking it accepted. The ephemeral sandbox cannot hold canonical state. Record shot/request/provider IDs, prompt/version, capability snapshot, refs, estimate/actual cost, durable output, reviewer, defect taxonomy and next action. Production completion requires accepted evidence/safety/originality/continuity review, not merely provider completion. For repairs, record Observation → Hypothesis → changed variable → Result → Learning. Return a compact handoff with `state` (planned/blocked/submitted/unknown/awaiting_review/accepted/rejected), exact IDs, artifacts, costs, evidence, limitations and next owner. Do not install new agents or broaden to unrelated channels to solve a shot problem.