| name | chrome-trace |
| description | Capture and analyze Chrome/Chromium performance traces with Playwright around a concrete browser interaction. Use when Codex needs to answer where frame time is spent during an update, drag, rotation, scroll, animation, camera movement, light movement, DOM/CSS render change, or other performance-sensitive UI action; especially when the right answer requires per-frame Chrome trace evidence instead of FPS-only guesses. |
Chrome Trace
Core Workflow
Use Playwright with Chromium and the Chrome DevTools Protocol Tracing domain to capture the exact interaction under investigation.
- Start from a reproducible page and action.
- Warm up the page until app-specific readiness is true.
- Start Chrome tracing and a
requestAnimationFrame sampler immediately before the action.
- Mark the action window with
performance.mark() and console.timeStamp().
- Perform the action with real Playwright input when the user is asking about input-driven behavior.
- Stop tracing after a short settle window.
- Align trace event timestamps to
performance.now() using the start/end marks.
- Report where time went:
Scripting, Style, Layout, PrePaint, Paint, compositor main-thread work, compositor impl-thread work, raster, GPU/viz events, and the slowest frame windows.
Do not draw conclusions from FPS alone. Use FPS/frame-time summaries only as the symptom; use trace groups and top events as the explanation.
PolyCSS Trace Runners
Use scripts/trace.mjs as the front door:
pnpm bench:build
node .agents/skills/chrome-trace/scripts/trace.mjs motion --page nonvoxel --mesh glb:Elephant.glb --variant baseline --dom-samples --label elephant-baseline
node .agents/skills/chrome-trace/scripts/trace.mjs motion --page nonvoxel --mesh teapot --variant baseline --dom-samples --frame-details --layer-details --gpu-details --trace-out bench/results/teapot.trace.json --label teapot-enriched --report
node .agents/skills/chrome-trace/scripts/trace.mjs motion --page nonvoxel --mesh teapot --variant baseline --gpu-details full --trace-out bench/results/teapot-full-gpu.trace.json --label teapot-full-gpu
node .agents/skills/chrome-trace/scripts/trace.mjs drag --mesh teapot --mode baked --frame-details --label teapot-drag
node .agents/skills/chrome-trace/scripts/trace.mjs motion --mesh garden --report --markdown-out bench/results/garden-trace.md
node .agents/skills/chrome-trace/scripts/trace.mjs compare bench/results/before.json bench/results/after.json --markdown-out bench/results/trace-compare.md
Use trace.mjs motion for steady bench motion across perf and nonvoxel pages, cadence buckets, DOM samples, render stats, and tag counts.
Add --frame-details to motion traces when you need slowest/fastest frame attribution instead of only bucket averages. On nonvoxel pages this also enables page-work samples for camera.update, scene.applyCamera, and input/control callbacks when available. Add --layer-details when compositor/layer shape is part of the question; it records LayerTree counts, layer aggregates by DOM tag/class (leaf:b, leaf:u, polycss-camera, etc.), largest layers, and compositing reasons. Add --trace-out when the raw Chrome trace should be preserved for DevTools.
Add --gpu-details when render pass timing is the current question and the trace still needs to stay reasonably sized. Light mode keeps the normal GPU/viz timeline categories and adds only disabled-by-default-viz.gpu_composite_time; render-pass attribution still comes from base events such as DirectRenderer::DrawFrame and DirectRenderer::DrawRenderPass. It intentionally avoids per-quad, Skia command, and GPU service spam.
Use --deep-gpu or --gpu-details full only for rare forensic runs that truly need per-quad/Skia/overdraw detail. Full mode also enables disabled-by-default-viz.quads, disabled-by-default-viz.triangles, disabled-by-default-viz.overdraw, disabled-by-default-gpu.debug, and disabled-by-default-skia.gpu; raw traces can become hundreds of MB and timing can be heavily perturbed.
Use trace.mjs drag for real PolyOrbitControls pointer-drag traces on nonvoxel-vanilla.html. This runner knows the non-voxel readiness hooks, camera state, interaction stats, and per-frame page-work samples.
Use trace.mjs generic for arbitrary pages and interactions that are not covered by a polycss bench page.
When interpreting PolyCSS traces, map the result back to the render model:
FunctionCall, EventDispatch, FireAnimationFrame: JS/input work. Unexpected sustained per-frame work is suspicious outside imported skeletal animation.
UpdateLayoutTree, RecalculateStyles: style recalculation, often CSS variable or selector invalidation cost.
Layout: layout; should stay low for transform/CSS-var-driven motion.
PrePaint, Paint, PaintArtifactCompositor::Update, Layerize: paint/compositing setup.
LayerTreeImpl::UpdateDrawProperties, draw_property_utils::ComputeDrawPropertiesOfVisibleLayers, LayerTreeHostImpl::PrepareToDraw, MainFrame.Draw, SubmitCompositorFrame: compositor-side cost.
Graphics.Pipeline, DisplayScheduler::DrawAndSwap, DirectRenderer::DrawFrame, DirectRenderer::DrawRenderPass: GPU/viz drawing pipeline. Treat these as browser output work, and compare them against layer details before changing app JS.
gpuVizRenderPass, gpuVizTiles, gpuVizGpuService: opt-in --gpu-details attribution buckets for render pass, coarse tile/raster playback, and GPU service events.
gpuVizQuads and gpuVizSkia: usually require --gpu-details full / --deep-gpu; treat them as high-overhead forensic signals.
RasterTask, image decode events: raster/bitmap work, usually atlas or tile work.
Trace event durations are inclusive and often nested, especially GPU/viz and scheduler events. Use group ms/frame as attribution evidence and for before/after deltas, not as exclusive slices that must add up to frame time.
Generic Capture
For arbitrary pages, use trace.mjs generic:
node .agents/skills/chrome-trace/scripts/trace.mjs generic \
--url http://127.0.0.1:3000 \
--ready-js "window.appReady === true" \
--action drag \
--selector "#viewport" \
--drag "480,0" \
--duration 1200 \
--steps 90 \
--summary-out trace-summary.json \
--trace-out trace.json
Useful alternatives:
node .agents/skills/chrome-trace/scripts/trace.mjs generic --url http://127.0.0.1:3000 --action wait --sample 3000
node .agents/skills/chrome-trace/scripts/trace.mjs generic --url http://127.0.0.1:3000 --action eval --eval "window.rotateScene?.(Math.PI / 2)"
node .agents/skills/chrome-trace/scripts/trace.mjs generic --url http://127.0.0.1:3000 --action scroll --scroll "0,900"
Comparing Runs
Use --report on a runner to generate a Markdown report after capture, or use trace.mjs report on an existing summary JSON:
node .agents/skills/chrome-trace/scripts/trace.mjs report bench/results/garden.json --markdown-out bench/results/garden.md
Use trace.mjs compare on summary JSON files from any runner:
node .agents/skills/chrome-trace/scripts/trace.mjs compare before.json after.json --markdown-out trace-compare.md
Read positive deltas in frame_time_*_ms and trace group ms/frame as more expensive after the change. Read positive FPS deltas as better.
Reports include a Quick Read section that calls out p95 frame time and the dominant trace group.
Reporting
Keep the report evidence-led and compact:
- State the exact command, page, viewport, action, warmup, and sample/settle windows.
- Include frame-time p50/p95/p99 and slowest-frame count.
- Identify the dominant trace groups in the action window.
- Name the top Chrome events, not only broad categories.
- For polycss, explicitly say whether the trace supports or violates the "no JS in the render loop" expectation.
- Mention artifacts written, especially the raw trace JSON that can be opened in Chrome DevTools Performance.
If trace markers are missing, say that alignment is weaker and rerun with marks before making a firm claim.