meituan-longcat/wbench/happy

Happy Oyster Benchmark Automation

Automates one benchmark case on https://www.happyoyster.cn/create. Input: case_<id>.json from the dataset's cases/ directory. Output: happy_video/case_<id>.mp4 under the repo root.

Paths below assume the repo is checked out and cd-ed into. Replace <repo> with the absolute checkout path when calling scripts from a non-repo cwd.

Key insight

happyoyster.cn does NOT enforce Private Network Access (unlike Project Genie). fetch('http://127.0.0.1:18888/...') works from the page, so images upload via DataTransfer — no native file dialog, no pyautogui coordinates, no window-handle juggling. Every interaction is a single browser-automation call.

Prerequisite

The local HTTP server must be running, serving the directory that contains the dataset's images/ tree. From the repo root:

python serve.py 18888

Verify:

curl -o /dev/null -s -w "%{http_code}\n" http://127.0.0.1:18888/data/images/case_2.jpg

Should return 200.

Before each new case

Kill any leftover `auto_interact.py` process before navigating to /create. Stale key-press processes will leak into the next world.

Get-Process python -ErrorAction SilentlyContinue | ForEach-Object {
    $cmd = (Get-WmiObject Win32_Process -Filter "ProcessId=$($_.Id)").CommandLine
    if ($cmd -like "*auto_interact*") { Stop-Process -Id $_.Id -Force }
}

Workflow

1. Navigate & inject prompt

const env = `<environment_prompt>`;
const char = `<character_prompt>`;
const prompt = `Environment: ${env}\n\nCharacter: ${char}\n\nMaintain visual consistency with the reference image. No HUD. No UI elements.`;
const ta = document.querySelector('textarea');
const setter = Object.getOwnPropertyDescriptor(window.HTMLTextAreaElement.prototype, 'value').set;
setter.call(ta, prompt);
ta.dispatchEvent(new Event('input', { bubbles: true }));

For open-world navigation cases, append no obstacles, move freely to the environment clause — Happy Oyster otherwise tends to spawn blocking geometry.

2. Upload image (fetch + DataTransfer)

(async () => {
  const url = 'http://127.0.0.1:18888/data/images/case_<id>.jpg';
  const blob = await (await fetch(url)).blob();
  const file = new File([blob], 'case_<id>.jpg', { type: 'image/jpeg' });
  const input = document.querySelector('input[type=file]');
  const dt = new DataTransfer(); dt.items.add(file); input.files = dt.files;
  input.dispatchEvent(new Event('change', { bubbles: true }));
  window.__uploadDone = `uploaded ${file.size}`;
})();

Use the path in settings.initial_image (the public dataset has no perspective subdirs). Wait ~1s, then screenshot to verify a thumbnail appears in the Image slot.

3. Flip perspective if first_person

Default is Third person. If the case is first_person, click the toggle once:

const persp = [...document.querySelectorAll('button,[role=button]')]
  .find(b => b.textContent.trim() === 'Third person');
if (persp) persp.click();

4. Click send

Small (~36×36) unlabeled icon button in the bottom-right of the prompt bar. Find by geometry, not pixel coordinates:

const send = [...document.querySelectorAll('button,[role=button]')].filter(b => {
  const r = b.getBoundingClientRect();
  return r.width >= 30 && r.width <= 45 && r.height >= 30 && r.height <= 45
         && r.top > 600 && !b.textContent.trim();
});
if (send.length === 1) send[0].click();

URL changes to /explore/wander/.... Recording auto-starts with a ~55s timer.

5. Start auto_interact in background

Start immediately once the URL leaves /create — the script internally waits for t2.png to appear before sending keys.

python auto_interact.py data/cases/case_<id>.json happy_oyster/t2.png 3

• happy_oyster/t2.png — single "A" key template (not Genie3's WASD cluster t.png).
• 3 — seconds per turn. Recording is time-limited so don't bump this above 4.

6. Wait for recording to end

Poll the active tab every ~10s until the URL transitions from /explore/wander/... to /end/travel/.... Typical duration: ~90-100s total from send to end (30-60s world generation + ~55s recording).

If the page shows 99% loading text for more than ~2 minutes without showing REC, navigate back to /create and retry from step 1.

7. Download

[...document.querySelectorAll('button,[role=button]')]
  .find(b => b.getAttribute('aria-label') === 'Download').click();

A "Download options" modal appears. Select "Without BGM", then click Download in the modal:

const noBgm = [...document.querySelectorAll('*')]
  .find(d => d.textContent.trim() === 'Without BGM' && d.getBoundingClientRect().width > 50);
let el = noBgm; while (el && !['DIV','BUTTON'].includes(el.tagName)) el = el.parentElement;
el.click();
setTimeout(() => {
  [...document.querySelectorAll('button,[role=button]')]
    .find(b => b.textContent.trim() === 'Download').click();
}, 500);

A "Preparing download..." banner shows at the top. Server rendering takes 20-60s.

8. Move file

File downloads to ~/Downloads/ as <world_title>.mp4 (e.g. CanineUrbanPulse.mp4, The Corinthian Archive.mp4). Find the newest non-Genie* mp4 and move:

mv "<Downloads>/<newest_title>.mp4" "happy_video/case_<id>.mp4"

9. Update progress

Optional — if you maintain a happy_progress.json:

import json
with open('happy_progress.json') as f: p = json.load(f)
p['case_<id>']['status'] = 'done'
with open('happy_progress.json','w') as f: json.dump(p, f, indent=2)

Things that do NOT work (and why)

• pyautogui coordinate clicks: Unreliable. Multiple Chrome windows for the same tab can coexist; their rects change when SW_RESTORE fires; the Image button's screen position varies with toolbar height, DevTools state, and window maximization. Use the JS-only flow instead.
• `mcp__claude-in-chrome__upload_image`: Only accepts imageId from prior screenshots, not disk paths.
• Inline base64 injection: A 500 KB image is ~680 KB of base64 — too long to paste into a single JS-evaluation call.

Gotchas

• Recording is a fixed ~55s timer. With KEY_DURATION=3, cases with >15 navigation turns won't fit.
• The "Without BGM" click must land on the parent <div>, not the inner text node — walk up to the first DIV/BUTTON ancestor.
• After download, the URL stays at /end/travel/.... Navigate fresh to /create for the next case.
• Different cases produce different world_title names — find by "newest non-Genie mp4" rather than an expected filename.

Files referenced

• auto_interact.py — shared keyboard driver (accepts template path as 2nd arg, turn duration as 3rd)
• happy_oyster/t2.png — Happy Oyster single-"A" key template (~636 bytes)
• serve.py — local HTTP server with CORS + Private Network Access headers
• happy_video/ — output directory (create if missing)