跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
其他技能

capture-gpu-tier

Use when capturing a GPU capability tier for the stealth crate's WebGL spoofing tables — running the probe on a machine whose GPU backend is not yet represented, then regenerating and committing the tables. Triggers on a request to "capture a tier", "add a GPU tier", "probe this machine's GPU", or a `platform_skew` warning about a persona whose OS has no matching tier. Keywords: probe_gpu, gpu-tier-gen, tiers.rs, D3D11, Metal, SwiftShader, data/gpu-tiers.

TurtIeSocks · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add TurtIeSocks/zendriver-rs --skill capture-gpu-tier

技能说明


name: capture-gpu-tier description: >- Use when capturing a GPU capability tier for the stealth crate's WebGL spoofing tables — running the probe on a machine whose GPU backend is not yet represented, then regenerating and committing the tables. Triggers on a request to "capture a tier", "add a GPU tier", "probe this machine's GPU", or a platform_skew warning about a persona whose OS has no matching tier. Keywords: probe_gpu, gpu-tier-gen, tiers.rs, D3D11, Metal, SwiftShader, data/gpu-tiers.

Capture a GPU capability tier

Overview

crates/zendriver-stealth/src/gpu/tiers.rs holds the WebGL capability values the stealth patch reports. It is generated from committed probe captures in crates/zendriver-stealth/data/gpu-tiers/*.json, one per backend tier.

Whether a capture generalizes beyond the machine it was taken on depends entirely on the backend, because ANGLE decides these values differently per backend. Know which case you are in before you name the tier:

  • D3D11 — generalizes by feature level, with two measured exceptions. renderer11_utils.cpp branches on the D3D_FEATURE_LEVEL for most of its caps, so FL11+ cards agree on nearly all of them. Three tiers ship — d3d11-fl11, d3d11-fl11-nvidia, d3d11-fl11-intel-gen9 — because two things escape that branch:

    1. A vendor-conditional workaround, in the same file: ANGLE_FEATURE_CONDITION(features, skipVSConstantRegisterZero, isNvidia), which then applies caps->maxVertexUniformVectors -= 1. NVIDIA parts report MAX_VERTEX_UNIFORM_VECTORS 4095 where everyone else reports 4096, and the two values derived from it shift with it. The condition is the vendor and nothing else, so one capture per side covers every card.
    2. Two genuinely device-derived values. MAX_SAMPLES is not a D3D11_REQ_* constant — ANGLE fills the per-format caps by calling CheckMultisampleQualityLevels on the device — and the WebGPU feature list is Dawn's answer for the physical adapter. An Intel HD Graphics 520 reports 16 samples and 16 features where an AMD Radeon reports 8 and 19.

    Everything else really is shared. An RTX 4090, an AMD Radeon and an Intel Gen9 iGPU, probed under one Chrome build, matched byte-for-byte on every other WebGL parameter, both extension lists (content and order), all shader precisions and all 36 WebGPU limits.

    Both exceptions were found by probing, not predicted, which is the whole argument for taking a capture on hardware no tier has seen even when the theory says it will match. Note the asymmetry that follows: a vendor condition can be read out of ANGLE's source and trusted to cover every card on that side, while a device-derived value covers exactly the silicon that was probed. That is why only Gen9 routes to the Intel tier and Gen11/Gen12 stay on the catch-all one — they are unprobed, and guessing a device-derived value is the failure this whole skill exists to prevent.

  • Metal on macOS — generalizes across Macs, confirmed by measurement. DisplayMtl.mm's TARGET_OS_OSX arm sets its caps from plain compile-time constants. (The runtime supportsAppleGPUFamily test that would vary them is in the iOS arm — see step 2.) An M2 Pro and an M4 Pro, on different Chrome patches and different macOS builds, agree on every value: 82 WebGL1 parameters, 132 WebGL2, both extension lists, every precision, all 36 WebGPU limits and all 22 features. Only the renderer string differs. A second Apple silicon capture is therefore not worth taking — it will reproduce metal-macos exactly, and the useful Mac capture now is an Intel one, which takes a different Dawn path entirely.

  • Vulkan — does NOT generalize. vk_caps_utils.cpp reads the caps straight off the physical device: max2DTextureSize is min(limitsVk.maxFramebufferWidth, limitsVk.maxImageDimension2D), the viewport bounds come from limitsVk.maxViewportDimensions, and that one file reads limitsVk. around 99 times. A Linux/Vulkan capture describes that GPU under that driver, and nothing else.

So on D3D11 and Metal, capturing on the right backend matters more than on the right card, which is why a handful of captures is enough. On Vulkan there is no such shortcut: name and document the tier for the specific device and driver it came from (step 2), record the Mesa/driver version in its provenance (step 1), and never present it as covering Vulkan generally.

Do not read "generalizes" as "one capture is sufficient". The D3D11 split above was not predicted from the source; it was found by probing a second vendor and diffing, and the ANGLE condition explaining it was only looked up afterwards. A capture on a vendor no shipped tier has seen is worth taking even when the theory says it will match — a confirmed match costs one probe, and an unexamined assumption costs a wrong value served under a real name.

A capture that is not a tier may still be evidence. Some probes settle a claim without producing a table anybody should be served: a second generation that reproduces an existing tier exactly, or a backend no persona can claim. Those go in crates/zendriver-stealth/data/gpu-confirmations/ — never in CAPTURES, never compiled into tiers.rs — with a row in that directory's README saying which claim they support. The bar is that a shipped rustdoc cites them; a capture that merely duplicates a tier is clutter, not corroboration, and belongs nowhere. Pin whatever they establish with a test in gpu-tier-gen so the evidence cannot rot silently when the capture format moves.

What they have established so far, and the reason the Gen9 tier exists:

ArchitectureD3D11ANGLE-GLANGLE-Vulkan
AMD RDNA2888
Intel Gen916—16

MAX_SAMPLES is constant per architecture across every backend measured, and differs between architectures on the backen

阅读完整来源 ↗