Browser Environment fails to initialize on macOS with Chromium headless mode

🐛 Bug Description

AutoAgent fails to start on macOS because Playwright's Chromium browser crashes immediately in headless mode. The error occurs during browser environment initialization.

💻 Environment

• OS: macOS 26.0 (Build 25A354)
• Python: 3.12.11 (via pyenv)
• AutoAgent: 0.1.0
• Playwright: 1.39.0
• BrowserGym: 0.13.0
• browsergym-core: 0.13.0
• Architecture: ARM64 (Apple Silicon)

🔴 Error Output

____________________Browser Env____________________
[2025-10-07 00:49:22]
Starting browser env...
Process Process-1:
Traceback (most recent call last):
  File "/Users/vladislaviskov/.pyenv/versions/3.12.11/lib/python3.12/multiprocessing/process.py", line 314, in _bootstrap
    self.run()
  File "/Users/vladislaviskov/.pyenv/versions/3.12.11/lib/python3.12/multiprocessing/process.py", line 108, in run
    self._target(*self._args, **self._kwargs)
  File "/Users/vladislaviskov/Desktop/agent/AutoAgent/autoagent/environment/browser_env.py", line 475, in browser_process
    obs, info = env.reset()
                ^^^^^^^^^^^
  File "/Users/vladislaviskov/.pyenv/versions/3.12.11/lib/python3.12/site-packages/browsergym/core/env.py", line 221, in reset
    self.browser = pw.chromium.launch(
                   ^^^^^^^^^^^^^^^^^^^
playwright._impl._api_types.Error: Browser closed.
==================== Browser output: ====================
<launching> /Users/vladislaviskov/Library/Caches/ms-playwright/chromium-1084/chrome-mac/Chromium.app/Contents/MacOS/Chromium --disable-field-trial-config --disable-background-networking --enable-features=NetworkService,NetworkServiceInProcess --disable-background-timer-throttling --disable-backgrounding-occluded-windows --disable-back-forward-cache --disable-breakpad --disable-client-side-phishing-detection --disable-component-extensions-with-background-pages --disable-component-update --no-default-browser-check --disable-default-apps --disable-dev-shm-usage --disable-extensions --disable-features=ImprovedCookieControls,LazyFrameLoading,GlobalMediaControls,DestroyProfileOnBrowserClose,MediaRouter,DialMediaRouteProvider,AcceptCHFrame,AutoExpandDetailsElement,CertificateTransparencyComponentUpdater,AvoidUnnecessaryBeforeUnloadCheckSync,Translate --allow-pre-commit-input --disable-hang-monitor --disable-ipc-flooding-protection --disable-popup-blocking --disable-prompt-on-repost --disable-renderer-backgrounding --force-color-profile=srgb --metrics-recording-only --no-first-run --enable-automation --password-store=basic --use-mock-keychain --no-service-autorun --export-tagged-pdf --disable-search-engine-choice-screen --enable-use-zoom-for-dsf=false --use-angle --headless --hide-scrollbars --mute-audio --blink-settings=primaryHoverType=2,availableHoverTypes=2,primaryPointerType=4,availablePointerTypes=4 --no-sandbox --user-data-dir=/var/folders/rn/gym3bnf52pg94rdfjjxy6ryw0000gn/T/playwright_chromiumdev_profile-K7Uemr --remote-debugging-pipe --no-startup-window
<launched> pid=79873

🔄 Steps to Reproduce

1. Install AutoAgent on macOS (Apple Silicon):

   git clone https://github.com/HKUDS/AutoAgent.git
   cd AutoAgent
   pip install -e .
   playwright install

2. Create .env file with OpenAI API key:

   cp .env.template .env
   # Add OPENAI_API_KEY=your_key

3. Run AutoAgent:

   COMPLETION_MODEL=gpt-4o auto main

4. Browser environment fails to initialize

🧪 Additional Testing

✅ What Works:

• Chromium launches successfully in non-headless mode
• Running Chromium binary directly: /Users/.../chromium-1084/.../Chromium --version ✅
• Firefox launches successfully in headless mode
• Docker containers created successfully
• All other components work fine

❌ What Doesn't Work:

• Chromium in headless mode (--headless flag)
• Trying Firefox as alternative fails because BrowserGym requires CDP (Chrome DevTools Protocol), which is Chromium-only

Test Results:

# Firefox works but CDP not available
python3 -c "from playwright.sync_api import sync_playwright; p = sync_playwright().start(); browser = p.firefox.launch(headless=True); print('OK'); browser.close()"
# ✅ Output: OK

# But CDP fails with Firefox
playwright._impl._api_types.Error: CDP session is only available in Chromium

💡 Potential Causes

This appears to be a macOS security issue with Playwright's Chromium in headless mode on newer macOS versions (especially Apple Silicon). Similar issues reported:

• [BUG] No tests found using the test explorer - pw/[email protected] microsoft/playwright#17500
• [BUG] Using libsoup2 and libsoup3 in the same process is not supported. microsoft/playwright#27313

🔧 Attempted Workarounds

1. Removed quarantine attribute:

xattr -d com.apple.quarantine ~/Library/Caches/ms-playwright/chromium-1084/chrome-mac/Chromium.app
# Already no quarantine attribute

2. Tried Firefox:

• Patched browsergym/core/env.py and browsergym/core/chat.py to use pw.firefox.launch()
• Firefox launches successfully but fails with: CDP session is only available in Chromium

3. Updated Playwright:

• Attempted upgrade to 1.55.0 but incompatible with libwebarena/libvisualwebarena requirements

🙏 Proposed Solutions

Option 1: Add Firefox Support to BrowserGym

Make BrowserGym work without CDP for Firefox compatibility

Option 2: Optional Browser Environment

Add a flag to disable browser environment for users who don't need web browsing:

web_env = BrowserEnv(..., disable=True)  # Skip browser init

Option 3: Use headed mode on macOS

Add OS detection and use headed mode (with virtual display) on problematic macOS versions

Option 4: Documentation Update

Add clear documentation about:

• Known issue with macOS 25.0+
• Recommend Linux/Windows WSL2 for browser-based tasks
• Workarounds or limitations

📋 Workaround for Users

Currently, users on affected macOS versions cannot run AutoAgent with browser environment. Temporary solutions:

1. Use Linux VM or Docker
2. Use Windows with WSL2
3. Wait for fix

🤔 Questions

1. Is there a way to run AutoAgent without browser environment for non-web tasks?
2. Could we add an environment variable like AUTOAGENT_DISABLE_BROWSER=true?
3. Any plans to support newer macOS versions or alternative browsers?

Would love to help test any fixes! 🙌