Webapp Testing

Web Application Testing

Test local web applications by writing native Python Playwright scripts, or by using the browser MCP tools available in VS Code Copilot Chat (open_browser_page, click_element, read_page, screenshot_page, type_in_page).

This skill teaches the reconnaissance-then-action pattern and the static vs dynamic decision tree. It does not bundle any helper scripts — the agent writes Playwright code directly or uses the browser MCP tools.

Decision Tree: Choosing Your Approach

flowchart TD
    A[User task] --> B{Is it static HTML?}
    B -- Yes --> C[Read HTML file directly<br/>to identify selectors]
    C --> D{Success?}
    D -- Yes --> E[Write Playwright script<br/>using selectors]
    D -- Fails/Incomplete --> F[Treat as dynamic]
    B -- No, dynamic webapp --> G{Is server already running?}
    G -- No --> H[Tell user to start server<br/>or use MCP browser tools]
    H --> I[Reconnaissance-then-action]
    G -- Yes --> I
    I --> J[1. Navigate and wait for networkidle]
    J --> K[2. Take screenshot or inspect DOM]
    K --> L[3. Identify selectors from rendered state]
    L --> M[4. Execute actions with discovered selectors]

Two execution paths

Path A: Native Playwright (Python)

Write a self-contained Playwright script. The agent manages server lifecycle itself — it does NOT shell out to a bundled helper. If a dev server needs to start, the user starts it in a separate terminal, or the agent uses the run_in_terminal tool.

Single server example (user starts npm run dev in a terminal first):

from playwright.sync_api import sync_playwright

with sync_playwright() as p:
    browser = p.chromium.launch(headless=True)
    page = browser.new_page()
    page.goto('http://localhost:5173')
    page.wait_for_load_state('networkidle')  # CRITICAL: wait for JS to execute
    # ... your automation logic
    browser.close()

Static HTML example (no server needed):

from playwright.sync_api import sync_playwright
from pathlib import Path

with sync_playwright() as p:
    browser = p.chromium.launch(headless=True)
    page = browser.new_page()
    page.goto(f'file://{Path("index.html").resolve()}')
    page.wait_for_load_state('networkidle')
    # ... your automation logic
    browser.close()

Path B: Browser MCP tools (VS Code Copilot)

If the browser MCP tools are available (open_browser_page, click_element, read_page, screenshot_page, type_in_page), prefer them for interactive inspection — they require no script file and give live feedback.

1. open_browser_page → http://localhost:5173
2. read_page → get accessibility snapshot
3. screenshot_page → capture current state
4. click_element / type_in_page → interact
5. read_page → verify result

Reconnaissance-Then-Action Pattern

1. Inspect rendered DOM (after waiting for networkidle):

   page.screenshot(path='/tmp/inspect.png', full_page=True)
   content = page.content()
   page.locator('button').all()
   

2. Identify selectors from inspection results

3. Execute actions using discovered selectors

Common Pitfall

Don't inspect the DOM before waiting for networkidle on dynamic apps. Do wait for page.wait_for_load_state('networkidle') before inspection.

Best Practices

• Use sync_playwright() for synchronous scripts
• Always close the browser when done
• Use descriptive selectors: text=, role=, CSS selectors, or IDs
• Add appropriate waits: page.wait_for_selector() or page.wait_for_timeout()
• If a dev server is needed, start it in a separate terminal or via run_in_terminal — do not bundle server management into the test script

What was removed from the upstream skill

The original webapp-testing skill (JZKK720/oz-skills) bundled scripts/with_server.py to manage server lifecycle. SkillSpector blocked it at scripts/with_server.py:69 for HIGH TM1 — tool parameter abuse (shell=True, which allows arbitrary command injection through user-provided --server arguments).

This port removes all bundled scripts. The agent writes native Playwright code directly or uses the browser MCP tools. Server lifecycle is the user's responsibility (start npm run dev in a terminal) or the agent's (via run_in_terminal), never a bundled helper that shells out with user input.

The methodology — decision tree, reconnaissance-then-action, static vs dynamic handling, common pitfalls — is preserved verbatim. Only the unsafe script execution path was removed.