について
このスキルは、視覚的ウェブアプリケーションの実行時検証を提供します。設定済みのヘッドレスChromiumブラウザでアプリケーションを実行し、実際のピクセル出力とコンソールエラーをチェックします。特に、ユニットテストでは検出できないWebGL、Canvas、GPGPU、WebAudioの変更を検証し、アプリケーションが可視の非黒色ピクセルをレンダリングすることを保証します。シェーダー、パーティクルシステム、またはキャンバスレンダリングに関連するPRのCIで使用することで、アプリケーションが機能するかどうかの自動化された合格/不合格チェックを得ることができます。
クイックインストール
Claude Code
推奨npx skills add pjt222/agent-almanac -a claude-code/plugin add https://github.com/pjt222/agent-almanacgit clone https://github.com/pjt222/agent-almanac.git ~/.claude/skills/verify-web-app-runtimeこのコマンドをClaude Codeにコピー&ペーストしてスキルをインストールします
ドキュメント
Verify Web App Runtime
Prove that a web app's visual runtime actually works — not that its elements exist, not that its unit tests pass, but that real pixels land on a real canvas with the expected GPU capabilities, in a headless browser with a CI-friendly exit code. This skill encodes the gotchas that bite every cold start of headless WebGL verification, each as an explicit procedure step.
When to Use
- A PR changes WebGL/WebGL2 shaders, a GPGPU simulation, canvas drawing, or a WebAudio graph, and "does it still render?" must be answered before merge
- A static code review or unit test suite cannot observe the failure mode (black canvas, silent GPGPU fallback, visibility-gated render loop)
- CI needs a one-command runtime gate: exits 0 when the app draws, 1 when not
- A headless machine (no GPU) must verify a WebGL2 app end-to-end
- NOT for extracting data from third-party pages — that is headless-web-scraping
Inputs
- Required: URL of the running app, including any base path
(e.g.
http://localhost:5173/myapp/— a missing base path 404s silently) - Required: Interaction steps to reach the surface under test — a JSON
array of
click/wait/assert_attrstep objects (clicks double as the WebAudio user gesture) - Optional: Output directory for screenshots and reports
(default:
./verify-runtime-out) - Optional: Luminance thresholds (default: a grayscale pixel > 25 counts as lit; >= 1% lit pixels passes)
- Optional: Console error keywords
(default:
gpgpu,webgl error,nan,fallback) - Optional: Whether
EXT_color_buffer_half_floatis required (default: yes; pass--skip-half-floatfor apps without a GPGPU path)
The packaged verifier at scripts/verify_runtime.py implements every step below; the procedure explains what it asserts and why, so each check can also be reproduced or adapted standalone.
Procedure
Step 1: Start the app and insist on a fresh-load test surface (never HMR)
HMR does not reset GPGPU textures: a hot-reloaded tab keeps the previous simulation state (e.g. particle position textures survive the module swap), so anything observed through HMR proves nothing about a cold start. Only a fresh page load in a fresh browser is a valid test surface.
npm run dev & # or: npm run build && npm run preview
curl -sf http://localhost:5173/myapp/ >/dev/null && echo "server up"
Expected: The app answers on a stable URL. All verification below happens
via fresh browser.launch() + page.goto() — never by observing a tab that
hot-reloaded.
On failure: If the URL 404s, check the base path — Vite apps often serve
under /<repo-name>/, and the bare origin silently serves nothing. If only an
HMR-updated tab shows the change, restart the dev server or build a preview
before verifying.
Step 2: Launch headless Chromium with the SwiftShader/ANGLE flags
Headless Chromium has no GPU; without software rendering flags,
getContext('webgl2') returns null and every downstream check fails for
the wrong reason.
CHROMIUM_ARGS = [
"--use-gl=angle",
"--use-angle=swiftshader",
"--enable-unsafe-swiftshader",
"--ignore-gpu-blocklist",
]
browser = playwright.chromium.launch(headless=True, args=CHROMIUM_ARGS)
Expected: A WebGL2 context is obtainable in the headless page (verified by the probe in Step 3).
On failure: Copy the four flags exactly — a misspelled flag is silently
ignored by Chromium and the symptom is identical (webgl2: false). If the
flags are correct and WebGL2 is still unavailable, update the browser:
python -m playwright install chromium.
Step 3: Probe EXT_color_buffer_half_float before trusting a GPGPU path
GPGPU pipelines render to half-float framebuffers. If the extension is missing, well-built apps fall back silently to a non-GPGPU mode — the page looks alive, but the screenshots verify the wrong code path.
webgl = page.evaluate("""() => {
const gl = document.createElement('canvas').getContext('webgl2');
return gl ? {webgl2: true, halfFloat: !!gl.getExtension('EXT_color_buffer_half_float')}
: {webgl2: false};
}""")
Expected: {"webgl2": true, "halfFloat": true}. SwiftShader supports the
extension, so a headless pass here is representative.
On failure: webgl2: false means Step 2's flags did not take effect.
halfFloat: false means any GPGPU verification is invalid — the app is
exercising its fallback. Fail the run (the packaged script does) rather than
screenshot the fallback; relax with --skip-half-float only for apps that
have no GPGPU path at all.
Step 4: Assert the page is visible — rAF loops are visibility-gated
Well-behaved apps pause their requestAnimationFrame loop when
document.hidden is true. A hidden page produces a black canvas that looks
exactly like a rendering bug.
info = page.evaluate("() => ({visibility: document.visibilityState})")
assert info["visibility"] == "visible"
Expected: visibilityState === 'visible' in every context, checked before
trusting any pixel assertion.
On failure: A black canvas with a hidden page is a test-harness artifact,
not an app bug. Call page.bring_to_front() and re-check; do not run other
foreground automation against the same headless browser while verifying.
Step 5: Drive interactions with real clicks (the WebAudio user gesture)
Reach the surface under test through the UI, exactly as a user would.
Browsers refuse to start an AudioContext without a user gesture — real
Playwright clicks count, programmatic dispatchEvent calls do not.
[
{"action": "click", "role": "button", "name": "Switch to 3D view", "settle": 2.5},
{"action": "click", "role": "button", "name": "Sand", "exact": true, "settle": 4},
{"action": "assert_attr", "role": "button", "name": "Sand", "exact": true,
"attr": "aria-pressed", "equals": "true"}
]
The "Switch to 3D view" / "Sand" names above are examples from the origin
project — replace them with your app's controls. settle (seconds) lets a
simulation reach a representative state before pixels are judged.
python3 scripts/verify_runtime.py --url http://localhost:5173/myapp/ \
--steps steps.json --out /tmp/verify-out
Expected: Every click resolves its target (ARIA role + accessible name
preferred, CSS selector as fallback), and every assert_attr step passes —
e.g. the mode button reports aria-pressed="true", proving the app accepted
the mode switch rather than ignoring the click.
On failure: A click timeout usually means a wrong accessible name — dump
candidates with page.get_by_role("button").all_inner_texts(). An
assert_attr mismatch means the UI ignored the interaction: check the
console log for errors thrown by the click handler.
Step 6: Screenshot the canvas and assert non-black luminance
A present canvas can still be empty. Element presence, canvas dimensions, and even a running rAF loop all pass while the app draws nothing. Only sampled pixels prove rendering.
from PIL import Image
grayscale = Image.open("norm_canvas.png").convert("L")
histogram = grayscale.histogram()
total_pixels = grayscale.size[0] * grayscale.size[1]
lit = sum(histogram[26:]) / total_pixels * 100 # pixels brighter than 25
assert lit >= 1.0, f"canvas effectively black (lit={lit:.1f}%)"
Expected: More than 1% of canvas pixels brighter than grayscale 25. The origin run measured ~9% lit for a healthy particle scene — comfortably above threshold without being tuned to the content.
On failure: Screenshot the full page too and compare: if the page shows
content but the canvas crop is black, the draw loop is not producing pixels
(check Steps 3 and 4 first). If a legitimately sparse scene sits under 1%,
raise settle so more of the scene accumulates, or lower --lit-percent-min
deliberately and note why.
Step 7: Run a second reduced_motion: reduce context and confirm the settled pose
Apps that honor prefers-reduced-motion must show a calm, settled state.
Verify it in a separate fresh context — never by toggling emulation on the
already-running page.
context = browser.new_context(viewport={"width": 1280, "height": 900},
reduced_motion="reduce")
The packaged script re-runs all previous assertions in this context, then takes two canvas screenshots one second apart and requires them near-identical (<= 2% of pixels changed) — a still pose, not a paused-by-accident one.
Expected: Reduced-motion context passes the same luminance/visibility checks and the two screenshots differ by <= 2% of pixels.
On failure: A large diff means animation continues despite
prefers-reduced-motion — check that the app queries the media feature and
that the settled pose is reachable without animation. If the app deliberately
keeps subtle motion, raise --still-max-changed-percent and document the
decision.
Step 8: Scan console output for GPU error signals
Pixels can look right while the console reports a degraded path. Collect
console and pageerror events for the whole run and match them against
error keywords.
gpu_errors = [line for line in logs
if any(keyword in line.lower()
for keyword in ["gpgpu", "webgl error", "nan", "fallback"])]
Expected: Zero matching lines. The packaged script appends any hits to the
failure list and writes the full log to <out>/console.log.
On failure: Read the matched lines in console.log. nan in a shader or
simulation log means numeric blow-up even if this frame looked fine;
fallback means Step 3's guarantee was violated at app level. Substring
matches on innocent words (e.g. "nan" inside a longer token) are tuned away
with explicit --console-error-keyword flags.
Step 9: Run the packaged verifier end-to-end and read the verdict
pip install playwright pillow
python -m playwright install chromium
python3 skills/verify-web-app-runtime/scripts/verify_runtime.py \
--url http://localhost:5173/myapp/ \
--steps steps.json \
--out /tmp/verify-out
Expected: Per-context summary lines, FAILURES: none, exit code 0.
/tmp/verify-out/ contains norm_page.png, norm_canvas.png, rm_page.png,
rm_canvas.png, rm_canvas_still.png, console.log, and report.json. The
origin run against a healthy build reported: half-float present, mode active,
visibility visible, ~9% pixels lit, zero GPGPU errors.
On failure: The exit code is 1 and every failed assertion is listed —
work through them in procedure order (flags before probe, probe before
pixels), since early failures cause misleading later ones. report.json
holds the raw evidence for each context.
Validation
- Verifier exits 0 against a known-good build of the app
- Verifier exits 1 when pointed at a deliberately broken surface (e.g. a blank page), proving the assertions can fail
-
report.jsonshows"webgl2": trueand"halfFloat": true(or the run explicitly used--skip-half-float) -
visibilityStateisvisiblein both contexts - Lit-pixel percentage >= 1% on the canvas crop in both contexts
- Reduced-motion stillness diff <= 2% changed pixels
-
console.logcontains no lines matching the error keywords - Screenshots for both contexts exist in the output directory
Common Pitfalls
- Verifying through HMR: Hot module replacement preserves GPGPU state (position textures survive the swap), so an HMR-updated tab can render correctly while a cold start is broken — or vice versa. Always verify a fresh page load in a fresh browser.
- Trusting element presence:
expect(canvas).toBeVisible()passes on a pitch-black canvas. Only the luminance assertion (Step 6) proves rendering. Know its converse limit too: an undrawn canvas composites as transparent, so over a bright page background it screenshots as fully lit — which is why Validation requires demonstrating the verifier can also fail. - Missing SwiftShader flags: Without the four Step 2 flags, headless
WebGL2 is simply
null— and the resulting black canvas is indistinguishable from an app bug. Rule the harness out first. - Ignoring the half-float probe: Skipping Step 3 lets a silent GPGPU fallback masquerade as a pass — the pixels came from the wrong code path.
- Asserting mid-animation: A simulation needs settle time before its
pixels are representative; screenshots taken during a transition flake.
Use per-step
settlevalues, not fixed global sleeps. - Expecting audio without a gesture:
AudioContextstays suspended until a user gesture; drive the UI with real Playwright clicks (Step 5) before asserting anything audio-dependent. - Overbroad console keywords: The default
nankeyword substring-matches innocent tokens. Tune with--console-error-keywordinstead of ignoring console failures wholesale.
Related Skills
- run-copilot-review-loop — companion skill from the same review workflow: after runtime verification passes, drive the bot review of the PR to a clean pass
- headless-web-scraping — the distinction: scraping is data extraction from (usually third-party) pages; this skill is runtime verification of your own app's pixels, GPU capabilities, and console — same headless browser, opposite purpose
- Headless WebGL Verification guide — background on why headless WebGL needs SwiftShader/ANGLE and how the verification pattern generalizes
This skill is a core skill of the frontend-runtime-verifier agent and is
exercised by the visual-pr-review team. See
references/EXAMPLES.md for the full origin recipe,
the steps-DSL reference, a no-GPGPU variant, and a CI integration example.
GitHub リポジトリ
Frequently asked questions
What is the verify-web-app-runtime skill?
verify-web-app-runtime is a Claude Skill by pjt222. Skills package instructions and resources that Claude loads on demand, so Claude can perform verify-web-app-runtime-related tasks without extra prompting.
How do I install verify-web-app-runtime?
Use the install commands on this page: add verify-web-app-runtime to Claude Code as a plugin, or clone its repository into your skills directory, then restart Claude so it picks up the skill.
What category does verify-web-app-runtime belong to?
verify-web-app-runtime is in the Testing category, tagged testing and design.
Is verify-web-app-runtime free to use?
Yes. verify-web-app-runtime is listed on AIMCP and free to install. It runs inside Claude, so no separate service account is required to use the skill itself.
関連スキル
このClaudeスキルは、lm-evaluation-harnessを実行し、MMLUやGSM8Kなど60以上の標準化学術タスクでLLMをベンチマークします。開発者がモデルの品質を比較し、トレーニングの進捗を追跡し、学術的な結果を報告するために設計されています。このツールはHuggingFaceやvLLMモデルを含む様々なバックエンドをサポートしています。
このスキルは、cron式を使用してWorkersをスケジュールするためのCloudflare Cron Triggersの実装に関する包括的な知識を提供します。定期的なタスクの設定、メンテナンスジョブ、自動化されたワークフローの構築を網羅し、無効なcron式やタイムゾーン問題といった一般的な課題への対処法も含みます。開発者はこれを使用して、スケジュールされたハンドラーの設定、cronトリガーのテスト、WorkflowsやGreen Computeとの連携を構成できます。
このClaude Skillは、Playwrightベースのツールキットを提供し、Pythonスクリプトを通じてローカルWebアプリケーションのテストを可能にします。フロントエンドの検証、UIデバッグ、スクリーンショット撮影、ログ表示を実現し、サーバーライフサイクルを管理します。ブラウザ自動化タスクにご利用いただけますが、コンテキストの汚染を避けるため、スクリプトのソースコードを読むのではなく直接実行してください。
このスキルは、開発者がテストの合格を確認し、構造化された統合オプションを提示することで、完成した作業を仕上げることを支援します。実装が完了した後のマージ、PR作成、ブランチの整理といったワークフローを案内します。コードが準備できてテスト済みの際に使用し、開発プロセスを体系的に完了させましょう。
