Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.
JavaScript 3d-generationai-agentsanimationavatardigital-humangenerative-aiglbgltfimage-to-3dllmmcpmcp-servermodel-context-protocolsolanatext-to-3dthreejsweb-componentswebglwebxrx402
docs(performance): say what the mobile sweep could and could not measure
nirholas4ed70f0eafa7
diff --git a/docs/site-performance.md b/docs/site-performance.mdindex 09cb140db..794ddcd0a 100644--- a/docs/site-performance.md+++ b/docs/site-performance.md@@ -82,7 +82,7 @@ Against commit `c2148462e` (stamped `2026-09-03T19:16:29Z`). Desktop, best of twBlocking time is the whole story again, and it came back through the assets rather than through the scheduling the rules already cover. Two thirds of `/`'s 29,980ms was one 4.6 MB hero GLB (rule 7); `/discover` was building a WebGL renderer for a directory whose visible cards are all still images (rule 4); `/forge`'s largest single bootup entry was the walk companion booting on top of the page (rule 8).-The mobile form factor was swept too and is recorded in the same directory, but **every mobile run landed below the 800 `benchmarkIndex` floor** (176 to 760), so those scores are a floor rather than a reading: `/` 26, `/create` 42, `/forge` 20, `/marketplace` 18, `/play` `PAGE_HUNG`, `/docs` 61, `/discover` 44, `/chat` 75.+**The mobile form factor could not be measured cleanly on this machine, and that is a finding about the machine.** A first pass put every run between `benchmarkIndex` 176 and 760, all under the floor: `/` 26, `/create` 42, `/forge` 20, `/marketplace` 18, `/play` `PAGE_HUNG`, `/docs` 61, `/discover` 44, `/chat` 75. A best-of-two repeat rescued exactly two of them before the machine degraded further and the renderer itself began dying: **`/` 31 at 1,384 and `/create` 40 at 1,318**, both quotable, and then `/marketplace` and `/discover` returned `TARGET_CRASHED` and `/chat` returned nothing at all. Mobile applies a 4x CPU slowdown on top of whatever the host is already doing, so it is the form factor that fails first when the build machine is oversubscribed. **Sweep mobile on an idle machine or not at all**; a crashed target is not a score of zero.**`/play` is not comparable to its 83 and is not a regression on its own evidence.** Profiled against production with a `longtask` observer, its render loop emits a **2.0 to 2.1 second long task every ~2.1 seconds, indefinitely**: one frame per two seconds, which is what rasterizing a coin world on the CPU costs. Chrome's own unresponsiveness detector fires before Lighthouse can finish, which is `PAGE_HUNG`. That figure is a property of `--disable-gpu`, not of the page, and it is why the "How to measure" section refuses to trade that flag for software WebGL. `/play` needs a machine with a GPU to be measured at all.