arguseye skill
Preview React (incl. React Router and Remix), browser-renderable Next.js, Preact, Solid, Vue 3 (incl. Nuxt) Svelte 5 (incl. SvelteKit), HTML/Web Components, and native Django/Jinja/Handlebars/PHP/Ruby templates from project source without starting the app. Use to inspect UI edits, literal-prop usage variants, component states and visual differences through Arguseye MCP tools or its CLI. Reports missing context; use application-level browser tests for complete journeys.
- GitHub stars
- 7
- Files
- 1
- Last commit
- Nov 10, 2026
- License
- MIT
- Allowed tools
- —
What does the arguseye skill do?
arguseye is an agent skill. Preview React (incl. React Router and Remix), browser-renderable Next.js, Preact, Solid, Vue 3 (incl. Nuxt) Svelte 5 (incl. SvelteKit), HTML/Web Components, and native Django/Jinja/Handlebars/PHP/Ruby templates from project source without starting the app. Use to inspect UI edits, literal-prop usage variants, component states and visual differences through Arguseye MCP tools or its CLI. Reports missing context; use application-level browser tests for complete journeys.
How to install the arguseye skill
Each command copies the skill folder into the client's skills directory. Restart the client or start a new session to load it.Install the arguseye skill in Plugin
/plugin marketplace add DarkWorldCoder/arguseye /plugin install arguseye@arguseye
Install the arguseye skill in Claude Code
~/.claude/skills/arguseye/SKILL.md
$ npx degit DarkWorldCoder/arguseye/skills/arguseye ~/.claude/skills/arguseyeInstall the arguseye skill in Claude Code (project)
.claude/skills/arguseye/SKILL.md
$ npx degit DarkWorldCoder/arguseye/skills/arguseye .claude/skills/arguseyeInstall the arguseye skill in Codex
~/.agents/skills/arguseye/SKILL.md
$ npx degit DarkWorldCoder/arguseye/skills/arguseye ~/.agents/skills/arguseyeInstall the arguseye skill in Gemini CLI
~/.gemini/skills/arguseye/SKILL.md
$ npx degit DarkWorldCoder/arguseye/skills/arguseye ~/.gemini/skills/arguseyeInstall the arguseye skill in Claude.ai
npx degit DarkWorldCoder/arguseye/skills/arguseye arguseye && zip -r arguseye.zip arguseye # claude.ai → Customize → Skills → + → Create skill → Upload a skill → arguseye.zip
SKILL.md
View on GitHubArguseye
Render component source in Chromium with the project's packages, styles, assets and statically reconstructable layouts. Inspect a cropped PNG, accessibility snapshot, resolved inputs, warnings and errors. The application server and authored stories are unnecessary; project dependencies and a compatible Chromium browser are still required. Startup and warm render times vary with the project; read the returned timings.
When to use it
- After editing a component: render it again; the result says what changed (
changes) since the last render of the same target. - For a shared component: inspect relevant call-site variants with
render_usages, following pagination when broader coverage is needed. Useusageplusin_context: trueto reconstruct its parent elements. - When a component needs a state: pass
actions(click, hover, fill, press, scroll, wait) andstorage(localStorage/sessionStorage flags). Storage seeds do not reconstruct a backend or an authenticated application journey. - When a component has no usage yet:
describe_propsgives the schema and sample props;framegives it a parent box of a chosen size and background. - After a refactor, a dependency upgrade or a change to shared styles:
audit_componentsrenders every component once and lists which throw and which come out empty.
Tools (MCP server arguseye)
| Tool | Use |
|---|---|
inspect_project {} |
Check the selected application root, installed framework versions and missing/computed configuration before browser startup. Read each diagnostic's recovery advice. This does not test a browser mount or execute Vite plugins. |
render_component {file, engine?, profile?, export?, usage?, props?, viewport?, theme?, in_context?, frame?, actions?, storage?, fixtures?, inspection?, timeoutMs?, output?} |
Render with a discovered usage or explicit props. viewport: mobile, tablet, desktop or WxH. theme: light or dark forces the colour scheme (the dark class on <html>, color-scheme, prefers-color-scheme); check both when the app has a dark mode. Without supplied inputs, required props may be generated from TypeScript samples. |
list_usages {file, export?} |
Every call site with literal props, unresolved runtime props, and the JSX ancestors (context). |
render_usages {file, export?, viewport?, theme?, cursor?, limit?, output?} |
A page of isolated literal-prop variants, not full-page integration tests. limit is 1–12; pass the returned nextCursor as cursor until it is null. |
describe_props {file, export?} |
Props from the TypeScript type, with sample values. |
audit_components {dir?, limit?, cursor?} |
Render a bounded page under dir with sample props. Follow nextCursor using the same root and directory; changed source invalidates it. |
render_diff {baselineId, output?} |
Compare current code with a successful render's captureId, reusing resolved props and viewport. Returns before/after/highlighted images and change evidence. |
Function props are passed as { "$fn": "(x) => ..." }, dates as { "$date": "2026-01-15T10:30:00Z" }. Paths are relative to the project root the server was started with.
For render tools, output is "image", "summary" or "both" (default). Use summary for a compact text response, image for visual inspection, or both when needed. Structured evidence remains available separately from image bytes.
Render evidence options (0.0.19+)
These additions require 0.0.19 or later. Inspect the connected tool's advertised schema before using them. Read diagnostics[].recovery as advice while retaining the original errors and warnings; do not execute suggested source changes automatically. An audit's ok mount count is separate from its complete/incomplete evidence counts.
For real below-fold observers, use capturePolicy: {scroll: {}} on rendering tools. An optional selector names one nested scroll container; maxSteps, timeoutMs and maxDistance bound traversal. Read scrollCapture and preserve incomplete coverage. Traversal restores the original view and does not force hidden content to become visible; visited range is not proof every animation completed. Baseline comparisons replay the policy.
Use inspection: {layout: true, fonts: true} on render_component for bounded geometry, clipping/overflow observations and actual used-font families/glyph counts. Requested CSS families alone do not establish font fidelity. Inspect styleInspection coverage and cost: only the first 100 DOM descendants are covered, excluding shadow and iframe documents. Treat clipping/ellipsis as observations requiring review. CLI source builds expose --scroll, --scroll-container, --capture-policy, --layout and --fonts.
Focused evidence (0.0.20+)
Check the connected schema before using inspection.selector. Use {selector:"h2",fonts:true,sources:true} to inspect at most 100 matching nodes within the capture, including late page targets. Selector-only requests enable layout evidence. Missing/invalid targets make requested inspection incomplete while preserving the image. Shadow/iframe documents remain excluded and axe keeps its full captured scope. CLI: shot app/page.tsx --inspect-selector 'h2' --fonts --sources --json.
Read environment.capture and diff environmentChanges. Browser/tool version, platform/architecture, OS release, pixel ratio or browser-policy differences make comparisons incomplete; legacy baselines without capture identity need recapturing. This is not a complete font/GPU/external-state fingerprint. Byte-identical PNGs skip pixel decoding but still compare semantic/context evidence; their diff image is an unhighlighted copy of the after frame.
Smarty setup (0.0.21+)
Explicit PHP setup may return configure(smarty) to register native modifiers or compiler filters on the project’s real Smarty instance. Compilation uses a stable source timestamp within each render, including slow filters; unsaved helper/partial edits still need verification. Preserve lookup-tracing and other context warnings. Earlier packages do not provide this hook or clock-boundary fix.
Native templates (0.0.16+)
Template adapters require Arguseye 0.0.16 or later and native compilation mode. Supply engine when needed and JSON props as context; an optional profile selects project-owned preview data. A Python/PHP/Ruby project does not need React or package.json. Install its real engine/runtime dependencies; do not substitute another engine with similar delimiters. inspect_project reports declared template configuration before runtime execution.
For example, render_component {file:"templates/card.html",engine:"django",props:{name:"Ada"},output:"both"} renders native Django. Capture its ID, edit the template or partial, and use render_diff to reuse the resolved data and viewport. Review template engine/runtime versions, actual dependencies and unresolved lookups as well as errors and warnings. Failed or still-loading fonts remain incomplete evidence. When lookup tracing is unavailable, preserve the warning even if the HTML looks correct.
VS Code template edits remount the document and reset interactive state. Explicit setup registers project filters/helpers; views, request handlers, auth and database state are not reconstructed. Built-in loaders honor unsaved buffers and canonical roots; executable template/setup code is trusted local code, not an OS sandbox. Unknown engines need a native adapter and conformance evidence. See the native template guide for tested engines, configuration and limits.
What it handles without setup
Providers from the root layout or from src/main.tsx (React Query, Redux, i18n set up by a side-effect import, theme providers), Vue plugins installed in src/main.ts (Pinia, vue-router, vue-i18n, Element Plus, directives, app.provide), <Suspense>, top-level await and <Teleport> in Vue, SvelteKit's $app/*, $env/*, $lib and kit.alias, auto-imports and auto-registered components (unplugin-auto-import, unplugin-vue-components, Nuxt's declaration files), Nuxt composables and built-in components (data calls return their default in a preview), Next's next/router, next/head, next/dynamic, next/headers and async server components, React.lazy/Suspense and data loaded on mount (the capture waits for them, up to 3 s), animations, charts and WebGL canvases, portals (dialogs, menus, toasts), virtualized lists, JSX in .js, tsconfig paths and vite.config aliases, import.meta.env/define, CSS modules, Sass, Tailwind v3/v4, JSON and image imports, custom elements. A component that renders nothing for the given props comes back with a warning instead of a blank image.
Reading a result
- For context static analysis cannot reconstruct, use an existing
arguseye.setup.tsxdefault wrapper acceptingchildren(React/Preact/Solid), orarguseye.setup.vue.tsdefault function receiving the Vue app. These replace inferred providers/plugins while retaining inferred global CSS. Create the file before server startup; existing files accept unsaved edits. ReadsetupAppliedin the result. For Svelte,arguseye.setup.svelte.tsreturns a fresh context Map. Setup runs trusted project code and does not restore a backend. Saved config/setup changes invalidate the next verification; refresh the live panel after adding a setup file. - MCP results include
schemaVersion: 1anderrors[], even for requests that fail before producing evidence. InspectisErroranderrorsbefore using tool-specific fields. Text-only clients can pass the printedcaptureIdback torender_diff. errors[]are real: a resolve error names the import and the importer; a render error carries the component stack. Fix them before trusting the pixels.- Read
warnings[],unresolved[],propsSourceandcompletebefore trusting context. Generated samples are not real application data; provider gaps, runtime values or unsettled assets can make a successful render incomplete. changesand explicit diffs reportunchanged,changedorincomplete, with pixel, accessibility and error evidence. None of these establishes that the intended behavior is correct.newWarningsandresolvedWarningsdistinguish current context from repaired failures. A comparison stays incomplete if its baseline was incomplete even whenafter.completebecomes true; capture a new baseline. Replayed generated/sample inputs retain their original provenance. Virtual plugin stand-ins remain warnings across warm renders; they do not become real plugin output after caching.timings(prepare, show, settle, capture) tell you whether a slow render is loading, rendering or waiting for images.layoutsAppliedandcontextAppliedlist the layouts/templates and the call-site ancestors the component was wrapped in.sourceRevisionfingerprints files read by this preview session, not the whole application's external state or a Git commit.
Baselines and comparison limits
Successful renders return an immutable captureId. The renderer retains at most 64 captures, four warm pages, 16 previous results and 64 MiB of retained PNGs; byte-budget eviction can occur before 64 captures. Thus restarting the server or eviction makes an old ID unavailable. Capture again when the error names an unknown or expired baseline.
An explicit render_diff also replays fixtures and inspection requests when supplied. It replays the baseline's file, export, resolved props, viewport, theme, frame, in_context parent box, actions and storage (since 0.0.8), so a stateful or contextual baseline compares like for like. It does not restore the live editor's interaction state. Next.js server execution, Nuxt context and arbitrary runtime providers require the application environment.
Reproducible and deeper checks
fixturescan fixnow,randomSeed,locale,timezoneId, and exact network responses. It uses a fresh browser context, including module initialization. Example:{"now":"2026-01-01T00:00:00Z","randomSeed":42,"network":{"policy":"block-unmatched","responses":[{"url":"/api/user","contentType":"application/json","body":"{\"name\":\"Example\"}"}]}}. ReadfixtureProvenanceand actualenvironment. Timers, crypto randomness and server state remain real. Baselines store fixture bodies and storage; avoid sharing credentials.- Pending animation frames remain stability warnings.
scheduledTimerscounts application/dependency timers without treating housekeeping as failure. A comparison that varies with unchanged source/inputs while timers remain scheduled is incomplete; engine deadlines, Vite transport and cancelled timers do not count. - MCP cancellation propagates to in-flight rendering.
timeoutMsbounds execution (30 s default, 120 s maximum). Cancelled pages are discarded; retry to continue. Node config code is trusted and is not an OS sandbox. inspection: {accessibility:true,sources:true}adds axe violations, checks needing review, and bounded element references. ReaderrorLocations.originalwhere available. An owning component file is not an exact source line, and unavailable maps stay explicit. No violations do not prove accessibility conformance. Inspecttimings.inspectionseparately.- Native mode remains default. Plugin-dependent virtual SVGs/macros need the project's own installed Vite/plugins and
--build-mode viteon the MCP command. This executes trusted config in a separate process, not the app entry/backend; compatibility edits reload and reset state. A native stand-in stays incomplete. Readbuildprovenance; restore baselines using the same compiler mode. - Existing
.htmlfiles use an independent native adapter. Real module scripts and Lit/Web Components use project packages; no React is needed. With one custom element, props become properties; with several, pass its tag asexport. Undefined elements and pending updates remain warnings. Classic<script>s run in document order before modules, on a fresh page per render. Closed/nested async shadow work and server templates require integration checks. - A plain-script widget (a
.js/.cjsfile with noimport/export: an embed or IIFE script such as a chat bubble) renders by passing its file: it runs as a classic script on a fresh page per render, so globals and "already loaded" guards never leak between renders. The capture and snapshot include what it mounts on<body>and in open shadow roots; closed shadow roots are a warning (image only).propsbecomewindow.__ARGUSEYE_WIDGET_CONFIG__, anddata-*props become attributes on its<script>(document.currentScript.dataset). For widgets that read a named global (window.intercomSettings) or need a host element, render a small.htmlpage with that config<script>or host<div data-…>before<script src="widget.js">. Errors carry the script's file and line; failed requests stay warnings. Native build mode only. - Pixel changes use a 0.1 YIQ noise threshold, so subtle color differences may not count. Inspect before/after images; unchanged never proves correctness.
Without MCP (CLI)
npx @ayushniroula/arguseye shot src/components/Card.tsx --export Card --usage app/page.tsx:12 --in-context --out card.png
npx @ayushniroula/arguseye usages src/components/Card.tsx --export Card
npx @ayushniroula/arguseye serve src/components/Card.tsx --export Card --port 4321
npx @ayushniroula/arguseye audit src/components --out audit.json
npx @ayushniroula/arguseye doctor --root . --json
Use shot --fixtures inputs.json, --accessibility, --sources and --timeout-ms N for explicit deeper checks. Use shot --json for structured metadata and --output summary to omit the ordinary PNG output. CLI baselines survive separate commands when explicitly saved in a chosen directory:
npx @ayushniroula/arguseye shot src/components/Card.tsx --export Card --capture-dir /tmp/arguseye-before --json --output summary
# Edit the component in the same project, then compare:
npx @ayushniroula/arguseye diff --baseline /tmp/arguseye-before --out /tmp/arguseye-diff --json
The baseline directory contains capture.json and capture.png; comparison writes before.png, after.png, diff.png and diff.json. The same replay limits apply to saved CLI baselines.
Workflow
list_usagesfor the component you changed.render_componentwith the first usage andin_context: true; readerrors, then the snapshot, then the PNG.- Edit; render the same target again and read
changes. - Inspect relevant
render_usagespages when the component is shared; report incomplete context and the variants actually checked.
Files in this skill
1arguseye skill questions
How do I install the arguseye skill in Claude Code?
Run npx degit DarkWorldCoder/arguseye/skills/arguseye ~/.claude/skills/arguseye. Claude Code loads skills from ~/.claude/skills (all projects) or .claude/skills (one project) and uses a skill when its description matches the task.
Which clients can use the arguseye skill?
Any client that reads the Agent Skills format: Claude Code, Codex, Gemini CLI and claude.ai (upload as a ZIP). Scripts in the skill run only where the client allows code execution.