# Content Workflow Scene Decomposition

> Decompose a large OpenUSD scene into finalized processable representatives, instance/prototype/payload family mappings, optional extracted assets, and a sealed Workflow 1 handoff. Use for the decomposition phase of a large-scene run, for repeated or heavily instanced scenes, or when material and physics tasks need explicit scene partitions.

- Skill: `nvidia-omniverse/content-workflow-scene-decomposition` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add nvidia-omniverse/content-workflow-scene-decomposition`
- Raw SKILL.md: https://api.skillmd.com/api/skills/nvidia-omniverse/content-workflow-scene-decomposition/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Productivity
- Author: NVIDIA-Omniverse (https://skillmd.com/u/nvidia-omniverse)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/nvidia-omniverse/content-workflow-scene-decomposition

---


# content-workflow-scene-decomposition

> **Workflow-owned decomposition:** Manifests, topology mappings, and the
> sealed Workflow 1 handoff are produced by
> `content_agent_workflows.scene_decomposition`. For ambiguity that structural
> USD evidence cannot resolve, inspect/render through usd-cli as frozen in the
> large-scene request.

Produce finalized manifest views and topology mappings. Do not make material or
physics assignments and do not mark the umbrella phase complete.

## Workflow

1. Read the source scene, requested tasks, phase input digest, and output
   directory from `large_scene_run.json`.
2. Inspect scene hierarchy, instances, prototypes, payloads, geometry counts,
   and prior diagnostics. Use usd-cli only for ambiguity
   that structural evidence cannot resolve; do not move decomposition policy
   into the scene tool.
3. Choose a decomposition intent and stable manifest ID. Create separate views
   only when requested domains genuinely need different processable boundaries.
4. Run deterministic decomposition:

   ```bash
   content-workflow-cli scene decompose SCENE.usd \
     --output-dir RUN/01-decomposition \
     --manifest-id MANIFEST_ID \
     --intent INTENT \
     --extract-assets \
     --input-digest PHASE_INPUT_DIGEST
   ```

   Add include/exclude paths, payload/prototype controls, structural duplicate
   detection, and extraction worker limits as scene evidence requires. Keep
   extraction concurrency conservative for large scenes.

   For a `material_processing` view, default to
   `--exclude-invisible-assets --min-mesh-count 1` unless the task explicitly
   requires hidden variants. A physics or other domain may need a separate
   view with a different visibility policy.

5. Do not enable external LLM refinement. The driving agent owns ambiguous
   boundary decisions and may revise explicit decomposition controls and rerun
   after inspecting the frontier.
6. Inspect `scene_manifest.json`, `manifest_catalog.json`, extracted assets, and
   diagnostics. Require one processable source representative per validated
   family; keep non-representative instance members non-processable.
7. Confirm `decomposition_result.json` reports `success: true`, no unresolved
   issues, and a non-empty output digest.
8. Return the result path to `content-workflow-cli scene phase` for handoff
   validation and phase completion.

## Required Outputs

Preserve:

- `scene_decomposition_request.json`
- `manifest_catalog.json`
- every finalized `scene_manifest.json`
- extracted representative USDs when requested
- `decomposition_result.json`

The catalog and manifests must identify the same original scene. Never expose
an instance proxy or temporary extraction path as a durable authoring target.

When changing decomposition behavior, preserve manifest schema compatibility,
source-scene identity, original-path mappings, and deterministic handoff
validation.