Render a web page on the host and measure it - screenshots, computed styles, element boxes and post-script text - when a page only exists after its scripts run or when the question is about layout. Covers choosing between web_fetch and a render, reading a design's real measurements, comparing viewports, and the egress policy that decides what may be reached.
インストールする前に、エージェントに与えられる指示の中身を確認できます。
browser_render, browser_query and browser_text open a URL in a throwaway
headless Chrome/Edge on the host and bring back what a browser saw: a PNG, the
page's own numbers, and measured elements. The remote page never runs in the
user's browser, and every socket the engine opens goes through an egress gate that
allows only https on port 443 to a validated public destination.
| The question | Use |
|---|---|
| What does this text SAY? A page that needs no script, an API, a plain document | web_fetch - cheaper, and it is the harness's own text path |
| What does the page look like / what is where / how big is it | browser_render (a picture) or browser_query (numbers) |
| Numbers a script fetched, a dashboard, a page that showed a consent wall to a plain fetch | browser_text |
| The design system's real type scale, spacing, or why something overflows | browser_query with what: ["box", "style"] |
A render costs a browser start (about 1-2 seconds on this class of machine) plus
the page's own load. Prefer browser_query when a picture is not needed: it is
the same render with less to carry.
browser_query reports, per selector: count, the first match's box
(x/y/width/height in CSS pixels, plus offset), its computed style
(font-family, size, weight, line-height, colour, background, display, gap,
padding, margin, border-radius, and more), its text, attrs or html.
font-size: 22.5px and
line-height: 28.125px are the computed values, not the authored ones - a
rem-based system shows what it resolved to at that viewport.390 and
1440 and diff the boxes. That is how a responsive claim gets checked.count to see how many nodes matched, and narrow the selector.dpr: 2 measures the same layout but asks for a retina screenshot; it does not
change CSS pixels.image.width/image.height are read back out of the PNG's own header, so they
are the file's true size. clipped: true means the page was taller than the
capture ceiling: raise the viewport or ask for a section instead of the whole
page.scroll is the content size, which may be much taller than the viewport; a
page whose scroll.height is enormous is usually an infinite feed.pageHosts lists the hosts the page asked for. gate.allowed lists what
the gate let through, and it can name one extra host: the engine's own
connectivity probe. That is not a page request, and its presence is evidence
the gate is the only way out.blocked is what failed to load (a subresource refused by policy, or a site
error). A page whose text is a consent wall or a CAPTCHA is reported as it is -
never guessed at.The policy is deliberately narrow, and every refusal names its rule:
http:, file:, data:, blob: and javascript: addresses
are refused by scheme;localhost, intranet), and no private, loopback or link-local
destination, because the whole answer set is refused when any address in it
is not public (one private answer in a round-robin is enough to send a request
somewhere it must not go);So a page that requires a sign-in, or lives on an intranet, is refused or shown without its authenticated content. Say so plainly rather than working around it.
title strings.まだレビューはありません。使ってみた感想をお寄せください。
概要と使いどころ
Author an editorial HTML+SVG diagram with the Diagram Design kit - choose the visual type, read its reference, start from an example, and satisfy the safety and accessibility contract that diagram_write({kind:"html"}) enforces.
日本語の概要は準備中です。原文の説明を表示しています。
Drive ffmpeg through the media_run tool as an argv array - remux instead of re-encode, map streams explicitly, trim at keyframes, scale and crop, extract audio and frames, and build GIFs, contact sheets and concatenations. Covers the flags this pack injects, reading a failure from its exit code and last log lines, and the version-dependent spellings to hedge.
日本語の概要は準備中です。原文の説明を表示しています。
Read a media file's real structure with ffprobe through media_run - the JSON invocation and how to read it, what every family of field means, how to count frames, sample without decoding everything, detect HDR and interlacing, and diagnose a missing duration, a stream with no frames or a container that lies. Covers when media_probe's own report is already enough.
日本語の概要は準備中です。原文の説明を表示しています。
Author Mermaid diagrams that parse the first time: choose the right diagram type for the question, write syntax the parser accepts, and keep the picture readable in a narrow panel. A bundled reference covers complex, large diagrams.
日本語の概要は準備中です。原文の説明を表示しています。
Write openFrameworks prototypes into the lab — real C++ (ofApp.h, ofApp.cpp, main.cpp) that builds into an app with the projectGenerator and make. Use for native/desktop sketches, performance-sensitive work, OpenGL, audio input, sensors, anything that has to be an application rather than a page.
日本語の概要は準備中です。原文の説明を表示しています。
Write p5.js prototypes into the lab — sketches that actually run in a browser, seed-reproducible, with controls generated from the sketch's own parameters. Use for generative art, flow fields, particle systems, noise, animation, anything sketched visually in the browser.
日本語の概要は準備中です。原文の説明を表示しています。