Chrome DevTools Agent Skill

Browser automation via executable Puppeteer scripts. All scripts output JSON for easy parsing.

Quick Start

CRITICAL: Always check pwd before running scripts.

Installation

Step 1: Install System Dependencies (Linux/WSL only)

On Linux/WSL, Chrome requires system libraries. Install them first:

pwd  # Should show current working directory
cd .claude/skills/chrome-devtools/scripts
./install-deps.sh  # Auto-detects OS and installs required libs

Supports: Ubuntu, Debian, Fedora, RHEL, CentOS, Arch, Manjaro

macOS/Windows: Skip this step (dependencies bundled with Chrome)

Step 2: Install Node Dependencies

npm install  # Installs puppeteer, debug, yargs

Step 3: Install ImageMagick (Optional, Recommended)

ImageMagick enables automatic screenshot compression to keep files under 5MB:

macOS:

brew install imagemagick

Ubuntu/Debian/WSL:

sudo apt-get install imagemagick

Verify:

magick -version  # or: convert -version

Without ImageMagick, screenshots >5MB will not be compressed (may fail to load in Gemini/Claude).

Test

node navigate.js --url https://example.com
# Output: {"success": true, "url": "https://example.com", "title": "Example Domain"}

Available Scripts

All scripts are in .claude/skills/chrome-devtools/scripts/

CRITICAL: Always check pwd before running scripts.

Script Usage

• ./scripts/README.md

Core Automation

• navigate.js - Navigate to URLs
• screenshot.js - Capture screenshots (full page or element)
• click.js - Click elements
• fill.js - Fill form fields
• evaluate.js - Execute JavaScript in page context

Analysis & Monitoring

• snapshot.js - Extract interactive elements with metadata
• console.js - Monitor console messages/errors
• network.js - Track HTTP requests/responses
• performance.js - Measure Core Web Vitals + record traces

Usage Patterns

Single Command

pwd  # Should show current working directory
cd .claude/skills/chrome-devtools/scripts
node screenshot.js --url https://example.com --output ./docs/screenshots/page.png

Important: Always save screenshots to ./docs/screenshots directory.

Automatic Image Compression

Screenshots are automatically compressed if they exceed 5MB to ensure compatibility with Gemini API and Claude Code (which have 5MB limits). This uses ImageMagick internally:

# Default: auto-compress if >5MB
node screenshot.js --url https://example.com --output page.png

# Custom size threshold (e.g., 3MB)
node screenshot.js --url https://example.com --output page.png --max-size 3

# Disable compression
node screenshot.js --url https://example.com --output page.png --no-compress

Compression behavior:

• PNG: Resizes to 90% + quality 85 (or 75% + quality 70 if still too large)
• JPEG: Quality 80 + progressive encoding (or quality 60 if still too large)
• Other formats: Converted to JPEG with compression
• Requires ImageMagick installed (see imagemagick skill)

Output includes compression info:

{
  "success": true,
  "output": "/path/to/page.png",
  "compressed": true,
  "originalSize": 8388608,
  "size": 3145728,
  "compressionRatio": "62.50%",
  "url": "https://example.com"
}

Chain Commands (reuse browser)

# Keep browser open with --close false
node navigate.js --url https://example.com/login --close false
node fill.js --selector "#email" --value "user@example.com" --close false
node fill.js --selector "#password" --value "secret" --close false
node click.js --selector "button[type=submit]"

Parse JSON Output

# Extract specific fields with jq
node performance.js --url https://example.com | jq '.vitals.LCP'

# Save to file
node network.js --url https://example.com --output /tmp/requests.json

Execution Protocol

Working Directory Verification

BEFORE executing any script:

1. Check current working directory with pwd
2. Verify in .claude/skills/chrome-devtools/scripts/ directory
3. If wrong directory, cd to correct location
4. Use absolute paths for all output files

Example:

pwd  # Should show: .../chrome-devtools/scripts
# If wrong:
cd .claude/skills/chrome-devtools/scripts

Output Validation

AFTER screenshot/capture operations:

1. Verify file created with ls -lh <output-path>
2. Read screenshot using Read tool to confirm content
3. Check JSON output for success:true
4. Report file size and compression status

Example:

node screenshot.js --url https://example.com --output ./docs/screenshots/page.png
ls -lh ./docs/screenshots/page.png  # Verify file exists
# Then use Read tool to visually inspect

5. Restart working directory to the project root.

Error Recovery

If script fails:

1. Check error message for selector issues
2. Use snapshot.js to discover correct selectors
3. Try XPath selector if CSS selector fails
4. Verify element is visible and interactive

Example:

# CSS selector fails
node click.js --url https://example.com --selector ".btn-submit"
# Error: waiting for selector ".btn-submit" failed

# Discover correct selector
node snapshot.js --url https://example.com | jq '.elements[] | select(.tagName=="BUTTON")'

# Try XPath
node click.js --url https://example.com --selector "//button[contains(text(),'Submit')]"

Common Mistakes

❌ Wrong working directory → output files go to wrong location ❌ Skipping output validation → silent failures ❌ Using complex CSS selectors without testing → selector errors ❌ Not checking element visibility → timeout errors

✅ Always verify pwd before running scripts ✅ Always validate output after screenshots ✅ Use snapshot.js to discover selectors ✅ Test selectors with simple commands first

Common Workflows

Web Scraping

node evaluate.js --url https://example.com --script "
  Array.from(document.querySelectorAll('.item')).map(el => ({
    title: el.querySelector('h2')?.textContent,
    link: el.querySelector('a')?.href
  }))
" | jq '.result'

Performance Testing

PERF=$(node performance.js --url https://example.com)
LCP=$(echo $PERF | jq '.vitals.LCP')
if (( $(echo "$LCP < 2500" | bc -l) )); then
  echo "✓ LCP passed: ${LCP}ms"
else
  echo "✗ LCP failed: ${LCP}ms"
fi

Form Automation

node fill.js --url https://example.com --selector "#search" --value "query" --close false
node click.js --selector "button[type=submit]"

Error Monitoring

node console.js --url https://example.com --types error,warn --duration 5000 | jq '.messageCount'

Script Options

All scripts support:

• --headless false - Show browser window
• --close false - Keep browser open for chaining
• --timeout 30000 - Set timeout (milliseconds)
• --wait-until networkidle2 - Wait strategy

See ./scripts/README.md for complete options.

Output Format

All scripts output JSON to stdout:

{
  "success": true,
  "url": "https://example.com",
  ... // script-specific data
}

Errors go to stderr:

{
  "success": false,
  "error": "Error message"
}

Finding Elements

Use snapshot.js to discover selectors:

node snapshot.js --url https://example.com | jq '.elements[] | {tagName, text, selector}'

Troubleshooting

Common Errors

"Cannot find package 'puppeteer'"

• Run: npm install in the scripts directory

"error while loading shared libraries: libnss3.so" (Linux/WSL)

• Missing system dependencies
• Fix: Run ./install-deps.sh in scripts directory
• Manual install: sudo apt-get install -y libnss3 libnspr4 libasound2t64 libatk1.0-0 libatk-bridge2.0-0 libcups2 libdrm2 libxkbcommon0 libxcomposite1 libxdamage1 libxfixes3 libxrandr2 libgbm1

"Failed to launch the browser process"

• Check system dependencies installed (Linux/WSL)
• Verify Chrome downloaded: ls ~/.cache/puppeteer
• Try: npm rebuild then npm install

Chrome not found

• Puppeteer auto-downloads Chrome during npm install
• If failed, manually trigger: npx puppeteer browsers install chrome

Script Issues

Element not found

• Get snapshot first to find correct selector: node snapshot.js --url <url>

Script hangs

• Increase timeout: --timeout 60000
• Change wait strategy: --wait-until load or --wait-until domcontentloaded

Blank screenshot

• Wait for page load: --wait-until networkidle2
• Increase timeout: --timeout 30000

Permission denied on scripts

• Make executable: chmod +x *.sh

Screenshot too large (>5MB)

• Install ImageMagick for automatic compression
• Manually set lower threshold: --max-size 3
• Use JPEG format instead of PNG: --format jpeg --quality 80
• Capture specific element instead of full page: --selector .main-content

Compression not working

• Verify ImageMagick installed: magick -version or convert -version
• Check file was actually compressed in output JSON: "compressed": true
• For very large pages, use --selector to capture only needed area

Resilience and Subagent Stability

When using the browser subagent or Puppeteer scripts, you may encounter connectivity issues (e.g., "connection is reset") or tab loss. Follow these guardrails for professional-grade reliability:

1. State Verification First

Rule: Always call list_pages (or a similar verification script) at the start of a new subagent task or if an error occurs.

• Do not assume the tab you just used is still active.
• If a "Connection Reset" error occurs, it usually means the browser instance crashed or closed. Immediately attempt to re-launch or re-initiate the session.

2. Handling Authenticated Sessions

Pattern: Use Persistent Browser Mode for sites requiring login or complex state.

• Launch with node launch-persistent.js --headless false.
• This ensures the browser process stays alive even if the subagent session times out.
• Subsequent subagents can reconnect via the WebSocket endpoint in .browser-endpoint.

3. Progressive Handoff

Rule: Use --close false consistently if the workflow requires multiple steps.

• When finishing a subagent task, explicitly check if the browser should remain open for the next phase.
• Provide the user with the URL and a screenshot of the final state to confirm continuity.

4. Error Recovery Strategy

If a "Page not found" or "Target closed" error occurs:

1. List all open pages to see if the tab was accidentally closed or navigated away.
2. Check for session tokens (cookies/local storage) if the site allows session persistence.
3. Re-navigate to the last known URL if the tab is truly lost.
4. Notify the user clearly about the reset instead of retrying silently 10 times.

5. Interaction Heartbeats

Before a complex click or fill sequence, take a small snapshot or evaluate a simple script (e.g., () => document.readyState) to "ping" the page. If it fails, trigger the recovery strategy early.

---

Reference Documentation

Detailed guides available in ./references/:

• CDP Domains Reference - 47 Chrome DevTools Protocol domains
• Puppeteer Quick Reference - Complete Puppeteer API patterns
• Performance Analysis Guide - Core Web Vitals optimization

Advanced Usage

Custom Scripts

Create custom scripts using shared library:

import { getBrowser, getPage, closeBrowser, outputJSON } from './lib/browser.js';
// Your automation logic

Direct CDP Access

const client = await page.createCDPSession();
await client.send('Emulation.setCPUThrottlingRate', { rate: 4 });

See reference documentation for advanced patterns and complete API coverage.

External Resources

• Puppeteer Documentation
• Chrome DevTools Protocol
• Scripts README