Browser automation using nodriver for anti-bot/Cloudflare bypass. Maintains page state across script executions. Write small, focused scripts to accomplish tasks incrementally.

Key difference from dev-browser: Uses nodriver instead of Playwright, which bypasses most anti-bot detection including Cloudflare Turnstile.

• Local/source-available sites: Read the source code first to write selectors directly
• Unknown page layouts: Use get_ai_snapshot() to discover elements and select_snapshot_ref() to interact with them
• Visual feedback: Take screenshots to see what the user sees

Important: Each session needs its own server. Do not reuse servers started by other sessions.

Start your server:

./skills/stealth-browser/server.sh &

• Default port is 6222. If in use, the script will tell you which port to use instead.
• Add --headless flag if user requests headless mode.
• Wait for the Ready on port XXXX message before running scripts.
• Remember the port number - you'll need it for connect(port=XXXX).

If you see "Port 6222 is in use", that's another session's server. Start your own on the suggested port:

./skills/stealth-browser/server.sh --port 6223 &

Execute scripts inline using heredocs. Use the port from your server startup:

cd skills/stealth-browser && source venv/bin/activate && python <<'EOF'
from client import connect

client = connect(port=6222)  # Use YOUR server's port!
page = client.page("example")  # descriptive name like "cnn-homepage"

page.goto("https://example.com")

print(f"Title: {page.title()}")
print(f"URL: {page.url}")

client.disconnect()
EOF

Write to tmp/ files only when the script needs reuse, is complex, or user explicitly requests it.

1. Small scripts: Each script does ONE thing (navigate, click, fill, check)
2. Evaluate state: Print state at the end to decide next steps
3. Descriptive page names: Use "checkout", "login", not "main"
4. Disconnect to exit: client.disconnect() - pages persist on server
5. Plain JS in evaluate: page.evaluate() runs in browser - use JavaScript syntax

Follow this pattern for complex tasks:

1. Write a script to perform one action
2. Run it and observe the output
3. Evaluate - did it work? What's the current state?
4. Decide - is the task complete or do we need another script?
5. Repeat until task is done

from client import connect

client = connect(port=6222)                 # Connect to YOUR server's port!
page = client.page("name")                  # Get or create named page
pages = client.list()                       # List all page names
client.close("name")                        # Close a page
client.disconnect()                         # Disconnect (pages persist)

# ARIA Snapshot methods
snapshot = client.get_ai_snapshot("name")   # Get accessibility tree (YAML)
ref = client.select_snapshot_ref("name", "e5")  # Get element proxy by ref
ref.click()                                 # Click the ref
ref.fill("text")                            # Fill the ref with text

client = connect(port=6222)  # Your server's port
page = client.page("mypage")

# Navigation
page.goto("https://example.com")

# Get state
url = page.url                    # Current URL (property)
title = page.title()              # Page title

# Interaction
page.click("button.submit")       # Click by CSS selector
page.fill("input[name=q]", "search term")  # Fill input

# JavaScript
result = page.evaluate("document.title")   # Run JS, get result

# Screenshots
page.screenshot("screenshot.png")          # Saves to tmp/ directory

# Waiting
page.wait_for_selector(".results")         # Wait for element

page.screenshot("screenshot.png")      # Saves to tmp/screenshot.png
page.screenshot("tmp/full.png")        # Explicit path

Use get_ai_snapshot() to discover page elements. Returns YAML-formatted accessibility tree:

- banner:
  - link "Hacker News" [ref=e1]
  - navigation:
    - link "new" [ref=e2]
- main:
  - list:
    - listitem:
      - link "Article Title" [ref=e8]
      - link "328 comments" [ref=e9]
- contentinfo:
  - textbox [ref=e10]
    - /placeholder: "Search"

Interpreting refs:

• [ref=eN] - Element reference for interaction (visible, clickable elements only)
• [checked], [disabled], [expanded] - Element states
• [level=N] - Heading level
• /url:, /placeholder: - Element properties

Interacting with refs:

snapshot = client.get_ai_snapshot("hackernews")
print(snapshot)  # Find the ref you need

element = client.select_snapshot_ref("hackernews", "e2")
element.click()

Page state persists after failures. Debug with:

cd skills/stealth-browser && source venv/bin/activate && python <<'EOF'
from client import connect

client = connect(port=6222)  # Your server's port
page = client.page("debug")

page.screenshot("debug.png")
print(f"URL: {page.url}")
print(f"Title: {page.title()}")
print(f"Body: {page.evaluate('document.body.innerText.substring(0, 200)')}")

client.disconnect()
EOF

This skill uses nodriver which has built-in stealth patches. Sites protected by Cloudflare (like lowendtalk.com) work automatically:

page.goto("https://lowendtalk.com/")
# No Cloudflare challenge - loads directly!
print(page.title())  # "LowEndTalk" (not "Just a moment...")