BRAVO Search Skill

Search for and download building records from the Hong Kong Buildings

Department BRAVO (Building Records Access and Viewing On-line) system.

Works for any project — not tied to a specific building or lot.

Prerequisites

• Use playwright-cli for browser automation.
• agent-browser is NOT suitable for BRAVO (element interaction issues).
• Ask the user for BRAVO credentials if not already provided.

Login Flow

BRAVO has a multi-step login. These steps are identical for all users:

1. Navigate to https://bravo.bd.gov.hk/login?request_locale=zh_HK
2. Fill textbox "登入名稱" with username
3. Fill textbox "密碼" with password
4. Click button "登入"
5. On the notice page (/notice):
• Check checkbox "本人確認已閱讀..."
• Click button "接受"
6. On the declaration page (/noticeProcess.action):
• Check checkbox "選項 A" (building construction matters)
• Click button "遞交"
7. On the multi-login page (/checkMulitpleSignOn):
• Click button "是"
8. Now at the home page (/home)

Always take a fresh snapshot after each page transition to get current

element refs. Refs change between sessions.

Location Search with Autocomplete

BRAVO uses jQuery UI autocomplete for address search.

Correct method:

1. Ask the user for the address or lot to search
2. Find textbox "位置" on the home page, click to focus
3. Use playwright-cli type "
   " — type character-by-character
4. Wait ~2-3 seconds for autocomplete dropdown
5. Click the .ui-menu-item-wrapper element in the dropdown
6. Click button "搜尋" or press Enter
7. If results panel is collapsed, click button "只顯示搜尋結果"

Wrong methods (will NOT work):

• fill command — no autocomplete trigger
• Direct API calls — require Daeja-specific auth, not session cookies
• goto to internal pages — kills the session

Navigating Search Results

Results appear in a table with columns:

• 圖則/文件/小型工程 | 檔案編號 | 地址 | 大廈名稱 | 地段編號 | 檔案類型 | OP編號 | 備註

To open a record's plan list:

• Find the row matching the target file number (e.g., 3/4003/72)
• Click the button "圖則" in that row
• A new tab opens with the plan list at viewPlanList?recId=
• Switch to the new tab with tab-select

Each plan in the list has:

• button "Preview" — thumbnail view (low res, full plan)
• button "Inspect Full Image" — Daeja ViewONE viewer

Downloading High-Resolution Plans via Tile API

The Daeja ViewONE viewer often shows "Lost connection to the server"

while the actual tile API remains functional. This is the most reliable

method for full-resolution plan extraction.

Step 1: Navigate and Get Base Resolution + Auth Tokens

Navigate to the inspection page for a plan and extract both the base

dimensions and auth tokens from the tile request:

var entries = performance.getEntriesByType('resource');
for (var e of entries) {
  if (e.name.includes('op=tile') && e.name.includes('tw=')) {
    var url = e.name;
    var getP = function(n) {
      var m = url.match(new RegExp('[&?]' + n + '=([^&]*)'));
      return m ? m[1] : '';
    };
    var baseW = parseInt(getP('tw'));  // e.g., 794
    var baseH = parseInt(getP('th'));  // e.g., 554
    var tokens = {
      fch: getP('fch'), xCCH: getP('X-CCH'), xCDH: getP('X-CDH'),
      xCId: getP('X-CId'), xLCId: getP('X-LCId'),
      version: getP('v') || '45571'
    };
    // Use these for the HD request
  }
}

The tw and th values tell you the base resolution at 96 DPI. Multiply

by the desired scale factor to get target HD dimensions.

Step 2: Request HD Tile

Construct a tile URL with increased DPI. The scale factor determines output

quality — 6x gives ~4764px for a typical 794px base, suitable for 300 DPI

printing at A3 size.

https://bravo.bd.gov.hk/v1files/?v=<version>&op=tile
  &tw=<baseW * scale>&th=<baseH * scale>&x=0&y=0&page=1
  &DPI=<96 * scale>&rxr=<96 * scale>&ryr=<96 * scale>
  &w=<baseW * scale>&h=<baseH * scale>&enh=1&iv=0
  &fch=<fch>&err=0
  &X-CCH=<xCCH>&X-CDH=<xCDH>&X-CId=<xCId>&X-LCId=<xLCId>
  &X-Client-Version=<version>

This returns the FULL plan as a single PNG tile. Unlike zoom+canvas

extraction, this captures the entire drawing at once.

Step 3: Base64 Chunking and Download

Large HD images require chunked base64 encoding inside page.evaluate()

to avoid "Maximum call stack size exceeded":

const arr = new Uint8Array(await blob.arrayBuffer());
let b64 = '';
for (let i = 0; i < arr.length; i += 8192) {
  b64 += String.fromCharCode.apply(null, arr.subarray(i, i + 8192));
}
return { b64: btoa(b64), size: arr.length };

Then decode in Python and save via Pillow:

data = base64.b64decode(result["b64"])
with open(output_path, "wb") as f:
    f.write(data)

Step 4: Combine to PDF

Use Pillow to combine all downloaded PNGs:

from PIL import Image
images = [Image.open(f).convert("RGB") for f in png_files]
images[0].save("output.pdf", "PDF", save_all=True,
               append_images=images[1:], resolution=300)

Key Constraints

• Auth tokens are unique per inspection page load — must extract fresh

for each plan

• Different plans may have different base resolutions (tw×th)
• Canvas extraction at any zoom level captures only visible tiles
• Fit-to-window gives full plan but at base 96 DPI resolution
• Direct getInspectionPlanImage URL requires Daeja auth headers
• Scale factor choice: 6x is a safe default; higher may hit server limits

Common Issues

1. Modal backdrop: Close with eval on #systemMsgModal button
2. Autocomplete 503: Transient server issue; retry in a few minutes
3. Session lost on navigation: Use history.back() or tab switching,

never goto to internal URLs

1. "Lost connection to the server": Daeja viewer UI issue only; tile

API still works — proceed with HD tile download method