omniverse-usd-performance-tuning

omniverse-usd-performance-tuning is a top-level NVIDIA Omniverse orchestrator skill for diagnosing and optimizing USD scene performance. It handles broad requests such as slow loading, high memory, low FPS, GPU crashes, conversion-quality triage, or a generic ask to optimize a scene. It solves the problem of coordinating a complex multi-phase optimization pipeline: it owns the user-facing performance request and delegates authentication and runtime setup to dedicated Phase 0 skills, keeping raw runtime artifacts on disk while working from compact summaries to stay within a token budget.

The skill enforces a mandatory session-start runtime context gate before producing tuning output, following the runtime-context-header reference that owns output_path, the canonical setup-preflight.json location, and Format A/Format B status displays. It classifies broad optimization requests as `ready_to_plan` and reserves `approval_required` for prompts that explicitly name a destructive operation such as flatten, decimate, merge prototypes, or delete unused prims. It plans the full canonical milestone chain through optimization-report, defaulting to 3 scoped iterations, with an exact subsequence: omniverse-usd-performance-tuning, profile-stage:baseline, usd-structure-assessment, usd-validation-runner, restructure-decision, apply-restructure, so-run-validators, so-interpret-validators, so-run-operations, profile-stage:after, compare-profiles, and optimization-report. Profile phases must use the exact labels profile-stage:baseline and profile-stage:after. Entry-skill rules route omniverse:// assets to omniverse-authentication first, and enter at setup-usd-performance-tuning only when no runtime path is verified; specific blocked_code values (for example blocked_missing_scene_optimizer) are returned instead of substituting another workflow when a requested tool is missing.

It targets Omniverse and USD engineers optimizing scenes for real-time performance. Downstream phases may require Kit, Scene Optimizer, Asset Validator, USD Python, writable output paths, and omniverse:// authentication as selected by setup-usd-performance-tuning. The distinction between authorizing a plan and authorizing a destructive action is central: a general optimization request authorizes planning (with per-operation approval requested alongside plan approval) but does not authorize execution of specific destructive operations. End-to-end execution should produce an optimized stage when mutation runs and a report conforming to the optimization-report schema.