From agent-tooling
Verify a change in a real HEADLESS browser: figure out what to look for, drive headless Chromium with Playwright, and either confirm a bug still reproduces or verify it's fixed AND that nothing on the risky adjacent paths broke. Use whenever the user wants to "check this fix in the browser", "confirm/repro this bug headlessly", "test the UI locally", "did the fix work and break anything", "run a headless browser test", "verify the arguments tab / voting / a page works", or hands you specific things to test in the running app. Prefer this over a screenshot or a code-only review when the question is "does it actually behave correctly in a browser?". Lighter than a full PR gate; pairs with pr-check's local-verification step.
How this skill is triggered — by the user, by Claude, or both
Slash command
/agent-tooling:browser-verifyThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Claude Code adapter over the agent-neutral **browser-verify playbook**. The method lives in the playbook; this file is the Claude-specific wiring.
Claude Code adapter over the agent-neutral browser-verify playbook. The method lives in the playbook; this file is the Claude-specific wiring.
Read the playbook: ../../playbooks/browser-verify.md. Follow its steps 1–4 and verdict labels.
Accept, in priority order:
Combine the user's items + the auto-derived claim + 1–3 risky neighbors into an explicit CHECK list. Use the project config (.claude/pr-check.json, same file pr-check uses) for the scope→risky-path mapping. Print the plan and let the user adjust before launching the browser.
Read the config's browser block: { base_url, serve, playwright }. Bring the stack up via serve (for wiki-polis that's the local-e2e skill — reuse it for the stack lifecycle). Note the real base URL it reports.
Ensure Playwright is present (once):
python3 -c "import playwright" 2>/dev/null || pip install playwright
python3 -c "import playwright" && playwright install chromium
Copy references/playwright-template.py to a scratch probe.py, set BASE_URL, and fill one check(...) per plan item (login/navigate in SETUP). Keep the console/pageerror/requestfailed listeners — console errors are failures. Wait on locators, never networkidle. Run headless:
python3 probe.py # exit 0 = all checks pass + no console/page errors
Allowlist Bash(python3 *probe.py*) and Bash(playwright install*) to keep it prompt-free.
Report each check as CONFIRMED / FIXED-AND-CLEAN / REGRESSION / COULD-NOT-RUN, with evidence (assertion text, console dump, probe.png). State explicitly whether the risky-path / user-supplied checks passed — that's the "no side damage" finding. One-line overall verdict. Trust the browser over the code.
Memory note: headless keeps this light, but it still launches Chromium — if OS memory pressure is critical, free memory or /compact first (the memory_guard hook only gates fan-outs, not this).
npx claudepluginhub lgelauff/wikimedia-coding-agent-lessons --plugin agent-toolingProvides behavioral guidelines to reduce common LLM coding mistakes, focusing on simplicity, surgical changes, assumption surfacing, and verifiable success criteria.
Searches, retrieves, and installs Agent Skills from prompts.chat registry using MCP tools like search_skills and get_skill. Activates for finding skills, browsing catalogs, or extending Claude.