Performance Test

When to use

Invoke this skill when:

• A release candidate URL needs to be gated on Core Web Vitals thresholds.
• A reported regression ("home page feels slow") needs quantified data.
• A new feature ships a heavyweight component and you need to check its cost.
• You need a mobile vs. desktop comparison on an identical route.

Do NOT use when: the target needs auth you cannot provide to Chrome DevTools; you need multi-hour RUM-style measurement (this skill is lab-based, not field-based); you need a load test against the origin (outside scope).

Inputs

• url (required): Fully qualified URL including path.
• form_factor (optional): mobile or desktop. Default mobile (matches CrUX reporting).
• throttling (optional): simulated-3g, applied-4g, none. Default simulated-3g for mobile, none for desktop.
• runs (optional): Number of Lighthouse runs; report median. Default 3.
• budget (optional): JSON of metric thresholds; defaults to CWV "good" per references/web-vitals-thresholds.md.
• auth_cookie (optional): Session cookie set before audit for gated routes.

Outputs

• A Markdown audit report including:
• Metric table (LCP / INP / CLS / TTFB / FCP) with value, threshold, verdict.
• Top 5 opportunities from Lighthouse with estimated savings.
• Long-task list (name, duration, source) from the trace.
• Render-blocking resources with suggested remediation.
• A prioritized fix list (P0/P1/P2) keyed to the worst-offender metric.
• A raw trace file path (saved via DevTools MCP) for engineers to open in Chrome.

Tool dependencies

• mcp__chrome-devtools__new_page, mcp__chrome-devtools__navigate_page, mcp__chrome-devtools__emulate (for form factor).
• mcp__chrome-devtools__lighthouse_audit (core Lighthouse run).
• mcp__chrome-devtools__performance_start_trace, mcp__chrome-devtools__performance_stop_trace, mcp__chrome-devtools__performance_analyze_insight.
• mcp__chrome-devtools__list_network_requests for render-blocking analysis.
• Write for the report.
• references/web-vitals-thresholds.md for thresholds and fix playbook.

Procedure

1. Open a clean page. mcp__chrome-devtools__new_page with an incognito-equivalent context. Set cookie if auth_cookie is supplied.
2. Emulate device. Call mcp__chrome-devtools__emulate with mobile (375x667, 2x DPR, iOS UA) or desktop (1366x768).
3. Warm a baseline navigation. mcp__chrome-devtools__navigate_page to about:blank, then to url. Discard the first load.
4. Run Lighthouse. Invoke mcp__chrome-devtools__lighthouse_audit with categories ["performance"], desired throttling, and onlyAudits empty for a full run. Repeat runs times; record median.
5. Capture a performance trace. mcp__chrome-devtools__performance_start_trace → navigate → performance_stop_trace.
6. Analyze insights. For each insight id (LCPBreakdown, INPBreakdown, LayoutShift, LongTask, RenderBlocking, DocumentLatency), call mcp__chrome-devtools__performance_analyze_insight.
7. Collect network data. mcp__chrome-devtools__list_network_requests filtered by resourceType=document,script,stylesheet to identify render-blocking assets.
8. Compare to budget. Use thresholds from references/web-vitals-thresholds.md (good/needs-improvement/poor per metric).
9. Rank fixes. Map each failing metric to its playbook entry; order P0 (core-path blocker) > P1 (single-metric fix) > P2 (polish).
10. Write report. Markdown to reports/perf/--.md.

Examples

Example 1 — Mobile audit of a product detail page

Input:

{ "url": "https://staging.shop.example.com/products/blue-hoodie", "form_factor": "mobile", "runs": 3 }

Report (excerpt):

# Performance audit — /products/blue-hoodie (mobile, 3G)
Run date: 2026-04-19  |  Median of 3 runs

| Metric | Value  | Good  | Needs improvement | Verdict |
| ------ | ------ | ----- | ----------------- | ------- |
| LCP    | 3.8s   | 50ms on main thread)
- 420ms — hydrateProductGallery (bundle.js:4221)
- 180ms — analytics.init (analytics.js:12)
- 90ms  — layoutShiftObserver (bundle.js:1980)

## Render-blocking resources
- /css/legacy.css (58KB, 740ms blocking)
- /vendor/analytics.js (112KB, 740ms blocking)

## Prioritized fixes
- P0 (LCP): ``. Move hero  above first script.
- P0 (TTFB): origin TTFB 1.1s — enable edge caching for product HTML.
- P1 (INP): move analytics.init behind `requestIdleCallback`, chunk hydrateProductGallery with `scheduler.yield()`.
- P1 (FCP): code-split `/bundle.js`; remove `/css/legacy.css` from critical path via `media="print" onload="..."`.
- P2: convert hero to AVIF with WebP fallback.

Example 2 — Desktop audit of a dashboard behind auth

Input:

{
  "url": "https://app.example.com/dashboard",
  "form_factor": "desktop",
  "throttling": "none",
  "auth_cookie": "session=abc; Domain=.example.com; Path=/"
}

Procedure notes:

• Set the auth cookie before navigation via mcp__chrome-devtools__evaluate_script with document.cookie = ... on a same-origin page, or inject through DevTools MCP cookie API.
• Desktop runs without throttling to reflect internal-tools reality.

Report focuses on INP (dashboards are interaction-heavy) and total blocking time from widget hydration. Typical fixes: virtualize long lists, memoize expensive selectors, offload CSV export to a worker.

Constraints

• Always run at least 3 times and report the median; single-run Lighthouse is too noisy to gate on.
• Never gate merges on "needs improvement"; only fail CI on "poor" (per thresholds in reference).
• Do not conflate lab metrics (this skill) with field metrics (CrUX / RUM). State it in the report.
• Simulated throttling is the default because it is deterministic; note this in the report so readers do not treat it as field-accurate.
• Do not audit pages that are rate-limited or behind a CAPTCHA without an explicit auth_cookie and a staging host.

Quality checks

• [ ] Median of >=3 runs reported, not single-shot.
• [ ] Every failing metric has at least one mapped fix.
• [ ] Top-5 opportunities are sourced from Lighthouse output, not invented.
• [ ] Long-task list cites script URL + line/function.
• [ ] Trace file path recorded in the report for engineer deep-dive.
• [ ] Device/throttling/run-count stated in the report header.
• [ ] Verdicts use thresholds from references/web-vitals-thresholds.md verbatim.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: tahirraufkeeyu
• Source: tahirraufkeeyu/software-development-agent-stack--sdas
• License: MIT

Install and usage instructions live in the source repository linked above.