Browser Testing#

⚠️ Ground rules (must be confirmed before every test — do not skip)#

1. Screenshots must use Playwright MCP — never the built-in browser's screenshot The built-in browser's browser screenshot only returns AI analysis text; MEDIA: can't render it, so the user won't see the image. The only correct way to take a screenshot: playwright__browser_take_screenshot → save to a file → MEDIA:<path>

2. Before using the built-in browser, check the ssrfPolicy allowlist first The target domain and the login domain (Auth0, etc.) must be in allowedHostnames in ~/.openclaw/openclaw.json, or navigation will fail with blocked by policy. After changing the config, you must run the full restart sequence: browser stop → gateway restart → wait 15 seconds → browser start

3. For font/environment issues, first check whether browser-setup has been run — don't troubleshoot manually For environment issues like mojibake in Chinese text or browser startup timeouts, first run bash testagent-browser-setup/scripts/setup.sh rather than manually installing fonts or debugging step by step.

---

Testing SOP#

Phase 1: Gather startup info#

After receiving a testing request, only ask for information that can't be obtained on your own before logging in:

1. Target URL (if not provided by the user)
2. Test account and password (for login)
3. Test goal description (e.g. "test the create-team feature")

Don't ask about product details (field formats, required fields, secondary accounts, etc.) before you've seen the product — those you can check yourself after logging in.

Phase 2: Explore the product and draft a test plan (requires user confirmation)#

1. Pick the primary tool per REFERENCE.md tool selection

2. If using the built-in browser, before navigating, confirm that browser.ssrfPolicy.allowedHostnames in ~/.openclaw/openclaw.json already includes the target domain and login domain; if not, add them and run the full restart sequence first (see REFERENCE.md pitfalls), then navigate

3. Navigate to the target URL and log in automatically

4. Viewport check (mandatory, use your own judgment): take a screenshot with Playwright MCP and judge for yourself whether the page is suitable for testing:

   • If content looks cramped, elements overlap, buttons are obscured, or the sidebar is crowding the main content area → run playwright__browser_resize → width=1440, height=900, then screenshot again to confirm
   • If the page looks fine → proceed directly
   • No need to ask the user — decide on your own whether to adjust

5. Browse the target feature area to understand the actual page structure and interactions

6. Based on the real page, produce 3-5 test points, and wait for the user to confirm or revise them before starting execution

text

(after logging in and viewing the product) I plan to test the following scenarios — let me know if this works:
1. Create team — fill in the name and required fields, verify creation succeeds
2. Form validation — check whether empty required fields show a prompt
3. Send a message within the team — verify the message displays correctly
4. Boundary input — behavior when the team name is excessively long

Only ask the user for additional info at this stage if exploration reveals it's genuinely needed (e.g. a test scenario requires a second account, or specific test data).

Phase 3: Execute the tests (fully autonomous, don't interrupt the user)#

• Execute each test point one by one
• Screenshot a bug the moment it's found — don't save it for the end (page state can change at any time)
• Screenshot naming: /root/.openclaw/workspace/bug_<number>_<short description>.png
• If the same action fails 2-3 times, switch tools or strategy immediately (see REFERENCE.md pitfalls)

Phase 4: Produce the test report#

Once all test points are complete, output the full report in chat (format in REFERENCE.md report format).

Phase 5: Confirm filing into Coding (requires user confirmation)#

After the report, ask:

text

Should these N bugs be filed into Coding?
You can adjust the suggested priority, assignee, or timeline, then reply "file it", or specify a change to a particular one.

Once the user confirms, call the coding-net skill's create_issue to bulk-file the bugs:

• issue_type="DEFECT"
• priority: per the priority rules (see REFERENCE.md)
• due_date: today's date + the suggested fix timeline (in days)
• assignee_id: look up the member ID per the assignment rules in AGENTS.md (using get_team_members_id_and_name or extract_members_from_issue_list)

Once filed, output the Coding issue number for each bug.

---

For detailed tool comparisons, screenshot workflow, and report templates, see REFERENCE.md