hexagone-web-feature-extractor（六边形网络特征提取器）

Hexagone Web Feature Extractor

Explore a Hexagone Web functional space, capture screenshots of every page/tab, and produce a Markdown document (.md) oriented for Product Owners with functional descriptions and embedded screenshots.

Prerequisites

• Node.js installed
• Playwright npm package (npm install playwright) — installs headless Chromium automatically
• Network access to the Hexagone Web server (default: https://ws004202.dedalus.lan:8065/hexagone-01/vue/login)

Configuration

Default values calibrated for the standard Hexagone Web layout at 1920x1080. Adjust if the layout differs.

Workflow Overview

1. SETUP        → Install Playwright, launch headless Chromium
2. CONNECTION   → Log in to Hexagone Web
3. NAVIGATION   → Navigate to the target space
4. DISCOVERY    → Expand sidebar, list all menu pages
5. EXPLORATION  → Visit each page, capture screenshots + metadata
6. GENERATION   → Produce the Markdown document with embedded screenshots

Key advantage over Chrome extension approach: Screenshots save directly to disk via page.screenshot() — no bridge server or transfer step needed.

Step 1: Setup

1.1 Install Playwright

npm install playwright
npx playwright install chromium

1.2 Launch Browser

const { chromium } = require('playwright');

const browser = await chromium.launch({
  headless: true,
  args: ['--ignore-certificate-errors', '--no-sandbox']
});
const context = await browser.newContext({
  viewport: { width: 1920, height: 1080 },
  ignoreHTTPSErrors: true  // Handles self-signed certs automatically
});
const page = await context.newPage();

Why headless Chromium? Eliminates the need for manual SSL certificate acceptance, Chrome extension setup, and screenshot bridge transfers. The ignoreHTTPSErrors: true option handles self-signed certificates programmatically.

Step 2: Connection to Hexagone Web

2.1 Navigate to Login Page

await page.goto(LOGIN_URL, { waitUntil: 'domcontentloaded', timeout: 30000 });
await sleep(3000); // Wait for Vue.js to mount

2.2 Fill the Login Form

The Hexagone Web login form has 3 fields: Username, Password, Manager code. Default credentials: username apvhn with a random password, unless the user provides others.

Use page.evaluate() with the native setter pattern — required for Vue.js which does not detect value changes injected directly:

await page.evaluate(({ username, password }) => {
  const nativeSetter = Object.getOwnPropertyDescriptor(
    window.HTMLInputElement.prototype, 'value'
  ).set;

  const userInput = document.querySelector('input[type="text"]');
  if (userInput) {
    nativeSetter.call(userInput, username);
    userInput.dispatchEvent(new Event('input', { bubbles: true }));
  }

  const pwdInput = document.querySelector('input[type="password"]');
  if (pwdInput) {
    nativeSetter.call(pwdInput, password);
    pwdInput.dispatchEvent(new Event('input', { bubbles: true }));
  }

  const loginBtn = Array.from(document.querySelectorAll('button'))
    .find(b => /connect/i.test(b.textContent));
  if (loginBtn) loginBtn.click();
}, { username: USERNAME, password: PASSWORD });

2.3 Verify Connection

Poll every 2s for up to 30s until the URL no longer contains /login:

for (let i = 0; i < 15; i++) {
  await sleep(2000);
  if (!page.url().includes('/login')) break;
}

If login fails: Take a debug screenshot with page.screenshot() and report the failure.

Step 3: Navigation to the Target Space

3.1 Open the Space Selector

CRITICAL: Use page.mouse.click() — NOT el.click() via page.evaluate().

Vue.js event handlers require native mouse events (mousedown + mouseup + click). JavaScript's el.click() only dispatches the click event and will not trigger the space dropdown. This was the #1 bug found during development.

The space selector is the div with class bg:orange-dark in the orange breadcrumb bar. It contains an icon <i class="hexa-icons">changer_espaces</i> followed by a <span> with the current space name.

// Find the space selector coordinates
const selectorRect = await page.evaluate(() => {
  for (const el of document.querySelectorAll('div, span')) {
    const cls = typeof el.className === 'string' ? el.className : '';
    if (cls.includes('bg:orange-dark') && !cls.includes('uppercase') && !cls.includes('hover:')) {
      const rect = el.getBoundingClientRect();
      if (rect.top > 30 && rect.top < 70 && rect.height > 15) {
        return { x: rect.left + rect.width / 2, y: rect.top + rect.height / 2 };
      }
    }
  }
  return null;
});

// Click with REAL mouse events (mandatory for Vue.js)
await page.mouse.click(selectorRect.x, selectorRect.y);
await sleep(3000);

3.2 Select the Space

The dropdown renders inside the sidebar area as a list of <div> elements with class px:1 py:3/4 hover:bg:orange-dark cursor:pointer. Spaces are listed alphabetically.

// Find the target space element
const target = await page.evaluate((spaceName) => {
  const walker = document.createTreeWalker(document.body, NodeFilter.SHOW_ELEMENT);
  while (walker.nextNode()) {
    const el = walker.currentNode;
    if (el.textContent.trim() === spaceName) {
      const rect = el.getBoundingClientRect();
      if (rect.width > 0 && rect.height > 0 && rect.top > 30) {
        return { x: rect.left + rect.width / 2, y: rect.top + rect.height / 2 };
      }
    }
  }
  return null;
}, TARGET_SPACE);

// Click with mouse (not el.click())
await page.mouse.click(target.x, target.y);

3.3 Wait for Loading

Hexagone Web redirects via an intermediate "Connexion... Redirection..." page. Poll every 2s for up to 24s until the URL no longer contains patient-portal (the default landing space):

for (let i = 0; i < 12; i++) {
  await sleep(2000);
  if (!page.url().includes('patient-portal')) break;
}
await sleep(3000); // Extra wait for Vue.js rendering

Step 4: Page Discovery

4.1 Expand the Sidebar

The sidebar is collapsed by default (icons only, width ~65px). Click the hamburger menu to expand it and reveal text labels:

await page.mouse.click(34, 50); // Hamburger icon position
await sleep(2000);

4.2 Identify Sidebar Menu Entries

Primary method: Look for elements with cursor:pointer class in the left 280px. Strip icon text from <i class="hexa-icons"> children:

const menuItems = await page.evaluate((excludeLabels) => {
  const items = [];
  const seen = new Set();
  const allEls = document.querySelectorAll('[class*="cursor:pointer"], a');
  for (const el of allEls) {
    const rect = el.getBoundingClientRect();
    if (rect.left < 280 && rect.top > 55 && rect.height > 15 && rect.height < 60) {
      let text = el.textContent.trim();
      // Strip icon prefix text
      const icon = el.querySelector('i');
      if (icon) text = text.replace(icon.textContent.trim(), '').trim();
      if (!text || text.length <= 1 || text.length >= 60 || seen.has(text)) continue;
      if (excludeLabels.includes(text)) continue;
      // Skip section headers (all-caps short text like "ACHATS")
      if (/^[A-Z ]+$/.test(text) && text.length < 15) continue;
      seen.add(text);
      items.push({
        label: text,
        y: Math.round(rect.top + rect.height / 2),
        x: Math.round(rect.left + rect.width / 2)
      });
    }
  }
  return items;
}, ['Trier par Importance', 'Trier par Emetteur']);

Sidebar DOM structure (observed):

<div class="sidebar--section min-w:sidebar bg:teal-darker">
  <div class="py:1 transition cursor:pointer w:inherit hover:bg:teal-dark">
    <div class="flex items:center whitespace:no-wrap">
      <i class="hexa-icons text:3/2" aria-label="Fournisseurs">fournisseurs</i>
      <span class="sidebar--label">Fournisseurs</span>
    </div>
  </div>
</div>

Note: The old a.hexa selector does NOT work for all spaces. The sidebar elements are <div> elements with cursor:pointer class, not <a> tags.

4.3 Items to Exclude

Filter out UI elements that are not pages:

• Trier par Importance / Trier par Emetteur — sort buttons in the Mes Post-Its panel
• Section headers (all-caps short text like ACHATS) — section dividers, not clickable pages

4.4 Validation Gate

If zero items are found after discovery: Stop the exploration and take a debug screenshot. Do NOT proceed with an empty page list. Save the DOM to a debug file for analysis.

4.5 Collapse Sidebar

After discovery, collapse the sidebar before starting exploration:

await page.mouse.click(34, 50); // Toggle hamburger
await sleep(1000);

Step 5: Exploration and Capture

5.1 For Each Page in the Menu

CRITICAL: The sidebar collapses after clicking a menu item. You MUST re-expand the sidebar and re-discover the item's position before each click.

For each menu item:
  1. Expand sidebar: page.mouse.click(34, 50), wait 1.5s
  2. Re-discover the item position via page.evaluate() (label matching)
  3. Click the item: page.mouse.click(freshPos.x, freshPos.y)
  4. Wait for page load (networkidle + 1.5s extra)
  5. Take screenshot: page.screenshot({ path: ... })
  6. Extract metadata via page.evaluate()
  7. If tabs found, explore them (see 5.2)
  8. Build feature description with PO-oriented text

Why re-discover each time? The sidebar re-renders its content when expanded. Item coordinates shift depending on scroll position and which items are visible. Using stale coordinates from the initial discovery will click on the main content area instead of the sidebar.

5.2 Identify Internal Tabs

Some pages have internal tabs (e.g., Fournisseurs → Saisie / Consultation / Réalisé, Marchés → SAISIE / CONSULTATION / DÉBLOCAGE). Use page.mouse.click() for tabs too:

const tabCoords = await page.evaluate((label) => {
  for (const tab of document.querySelectorAll('[role="tab"], .v-tab')) {
    if (tab.textContent.trim() === label) {
      const rect = tab.getBoundingClientRect();
      return { x: rect.left + rect.width / 2, y: rect.top + rect.height / 2 };
    }
  }
  return null;
}, tabLabel);

if (tabCoords) {
  await page.mouse.click(tabCoords.x, tabCoords.y);
  await sleep(2500);
  await page.screenshot({ path: tabScreenshotPath });
}

Warning: The selector [class*="tab"] is too broad — it matches pagination controls, filter dropdowns, and elements from previously rendered pages. Prefer [role="tab"] or .v-tab for tab detection.

5.3 Take Screenshots

Screenshots save directly to disk — no bridge or transfer needed:

await page.screenshot({
  path: path.join(SCREENSHOT_DIR, `${index}-${sanitize(pageName)}.png`),
  fullPage: false  // Capture viewport only (1920x1080)
});

5.4 Store Metadata

Build a features.json array during exploration with PO-oriented descriptions enriched from page content analysis (tables, forms, actions, tabs).

Step 6: Markdown Document Generation

6.1 Prepare the Input

Create a features.json file from the metadata collected in Step 5. The file must conform to this structure:

{
  "space": "Name of the explored space",
  "features": [
    {
      "title": "Feature name (string, required)",
      "description": "PO-oriented functional description (string, required)",
      "capabilities": ["Capability 1", "Capability 2"],
      "businessValue": "Business value description (string)",
      "screenshots": [
        { "file": "filename.png", "caption": "Screenshot description" }
      ]
    }
  ]
}

Validation rules:

• space must be a non-empty string
• features must be a non-empty array
• Each feature must have title (string) and description (string)
• capabilities must be an array of strings (can be empty)
• screenshots must be an array of objects with file (string) and caption (string)

6.2 Generate the Document

Use the script scripts/generate-md.js:

node scripts/generate-md.js --input features.json --output /path/to/output.md --screenshots /path/to/screenshots

The --input argument is required. The script validates the input and fails with clear error messages if the JSON is malformed. No npm install needed — the script uses only Node.js built-in modules.

6.3 Document Structure

# Title (space name)
> Subtitle with date

## Table of contents (linked)

For each feature:
  ## N. Feature title
  - Functional description
  - Screenshot(s) as ![caption](relative/path)
  ### Key capabilities (numbered list)
  ### Business value
  ---

6.4 Validation and Output

Verify the generated file exists and contains all expected features:

grep -c '^## [0-9]' output.md  # Should match the number of features

Input Parameters

The user must provide:

• Login URL *(optional)*: defaults to https://ws004202.dedalus.lan:8065/hexagone-01/vue/login. The user can provide a different URL if needed.
• Username *(optional)*: defaults to apvhn. The user can provide a different code if needed.
• Password *(optional)*: defaults to a random value. The user can provide a specific password if needed.
• Target space: Exact name of the space to explore (e.g., "HA GHT", "STRUCTURES / NOMENCLATURES")

Critical Rules (learned from production runs)

1. Always use page.mouse.click() for any Vue.js interaction — NEVER use el.click() via page.evaluate(). Vue.js requires native mousedown/mouseup events that only page.mouse.click() provides.
2. Re-expand sidebar before every page click — the sidebar collapses after each navigation. Using stale coordinates from initial discovery will miss the sidebar entirely.
3. Re-discover item positions each iteration — sidebar coordinates shift on re-render. Always query the DOM for fresh coordinates.
4. **Use [class*="cursor:pointer"] not a.hexa** for sidebar items — sidebar elements are <div> elements, not <a> tags.
5. Strip <i class="hexa-icons"> text from sidebar labels — icon text is prepended (e.g., tdbTableau de bord → Tableau de bord).
6. Exclude utility buttons like "Trier par Importance" from the page list — they are sort controls in the Post-Its panel, not pages.
7. Use narrow tab selectors ([role="tab"], .v-tab) — [class*="tab"] is too broad and picks up pagination, filters, and stale elements from previously rendered pages.

Troubleshooting