fix(cli): expose benign globals in run-code vm context

[Pavel Feldman (pavelfeldman)]

Summary

• Expose timers, fetch, URL, Buffer, crypto, AbortController, TextEncoder/TextDecoder and similar globals in the run-code vm context; require and process stay unavailable.
• Document the available globals in the running-code.md skill reference.

Fixes microsoft/playwright-cli#471

[github-actions]

Test results for "MCP"

2 failed
❌ [chromium] › mcp/webmcp-dynamic.spec.ts:189 › a tool registered in an iframe can be called after the page is reloaded @mcp-macos-latest-chromium
❌ [firefox] › mcp/annotate.spec.ts:446 › should switch screencast to -s session on show --annotate @mcp-windows-latest-firefox

8827 passed, 1480 skipped

---

Merge workflow run.

[github-actions]

Hi, I'm the Playwright bot and I took a first look at the MCP failures here.

🟢 Both failures look like pre-existing flakes

Both tests have failed with the same assertion on main and on other PRs. Neither test goes through the run-code VM context this PR changes.

Details

This PR only changes browser_run_code_unsafe / run-code: it adds timers, URL, fetch, Buffer and other globals to the vm context in tools/backend/runCode.ts. It also updates the docs and the run-code specs. Neither failing test uses that tool. I checked against the test-results DB snapshot. The incremental update step hit a GitHub API 500, so this PR's own runs aren't in it.

Pre-existing flake / infra

• [firefox] › mcp/annotate.spec.ts:446 › should switch screencast to -s session on show --annotate (mcp-windows-latest-firefox). This is a known flake. Its final verdict was a failure in 40 of 723 runs, and a retry passed it in 37 more. Those runs were mostly on Windows Firefox, on unrelated PRs (fix(mcp): do not crash when download fails after browser close #42922, fix(tracing): do not hang on unresponsive page when capturing snapshots #42919, chore(proxy): drop legacy 'per-context' launch proxy handling #42892, feat(test-runner): add --shuffle to run tests in random order #42877, docs(triage): also check previous similar issues and related PRs #42844, …) and on main pushes. The signature is always the same: the toHaveAccessibleName check on the Session … region fails on the first attempt.
• [chromium] › mcp/webmcp-dynamic.spec.ts:189 › a tool registered in an iframe can be called after the page is reloaded (mcp-macos-latest-chromium). This test is newer and flips verdict in the DB. It failed with the same toEqual / StringContaining "\"text\": \"3…" mismatch on a main push (run 36076490274, the same macOS Chromium bot) and on feat(chromium): roll to r1247 #42915. It also failed on chrome on main (run 36037888265). The WebMCP iframe/reload flow doesn't go through the run-code VM context.

_{Triaged by the Playwright bot - agent run}