scene-analysis

name	scene-analysis
description	Analyze and optimize Roblox scenes using SceneAnalysisService — rendering, memory, instance composition, unparented instances, and animation/audio assets. Use when investigating performance, memory, or leaks in a place.

Execution in this repo (rodeo / Studio MCP)

SceneAnalysisService only works at runtime, so run every query against a play DM: rodeo run --target test:client for rendering/triangle queries (view-dependent), --target test:server for memory/unparented/leak queries (server-side leaks are common). Start play via MCP start_stop_play or a rodeo play target. commands/ holds the four analysis routines; QuerySpecs.md documents the return shapes.

Scene Analysis

Uses SceneAnalysisService to analyze and optimize Roblox game scenes. All commands follow the same cycle:

1. Start play mode (required — SceneAnalysisService only works at runtime)
2. Run queries via SceneAnalysisService
3. Stop play mode
4. Analyze results and present findings or make edits

When to Use

• User asks to optimize a scene, check performance, trace memory, or find leaks
• Keywords: performance, memory, draw calls, triangles, scene health, unparented, leak, optimize, SceneAnalysisService

Commands

Command	What it does
/scene-health	Full overview — runs all 6 queries, summarizes rendering, memory, instances, and unparented objects
/optimize-rendering	Deep rendering analysis — auto-finds geometry hotspots, sweeps camera at player height, measures triangles and draw calls per view, places hotspot markers in the scene
/optimize-memory	Memory analysis — script VM memory, animation clip memory, audio asset memory with reference counts
/fix-leaks	Traces unparented instances to host scripts, reads the source, identifies cleanup patterns, offers to apply fixes

SceneAnalysisService queries

Query	Returns
GetInstanceCompositionAsync()	Instance counts by category and class
GetTriangleCompositionAsync()	Triangle and draw call counts by render pass (view-dependent)
GetScriptMemoryAsync()	Per-script Luau VM heap memory
GetUnparentedInstancesAsync()	Instances held in memory but not in the DataModel, traced to host scripts
GetAnimationMemoryAsync()	Loaded animation clip memory with owner Animators
GetAudioMemoryAsync()	Loaded audio asset memory with owner Sound/AudioPlayer instances

See QuerySpecs.md in this skill for full data shape documentation.

Key facts

• Play mode is required. SceneAnalysisService queries runtime data (GPU render passes, Luau VM heap, loaded assets). It doesn't work in edit mode.
• Triangle composition is view-dependent. A single measurement only represents one viewpoint — /optimize-rendering handles this with an automatic hotspot sweep.
• Set CameraType = Scriptable before programmatically moving the camera, or the PlayerModule will override it each frame.
• Shadows should be excluded from triangle/draw call budgets — engine-driven, not creator-controllable.
• Animation and audio assets are reference-counted. Memory can only be freed by removing ALL references.
• Server mode — running memory and unparented instance queries from a server session can surface server-side issues (server-side animation leaks are especially common).

Tone

Frame findings as opportunities, not problems. Be helpful and collaborative. Avoid language like "CRITICAL", "WARNING", "bloat", "problem", "offender", "over budget", "failing". Don't assign grades or severity labels. Present data clearly and let the creator decide what to act on.

Reference points (rendering)

~800k triangles and ~600 draw calls is roughly the range where most devices can maintain 60fps. Present these as context, not pass/fail criteria — some creators target higher fidelity and accept lower framerates. Ask about framerate/device target before making strong recommendations.

Internally, each draw call costs roughly the equivalent of 20,000 triangles in GPU overhead — use this to prioritize draw call reduction in recommendations, but don't show the ratio to the creator.