image: the browser is invisible-playwright-mcp 0.72.2 (invisible-playwright 0.35.7, invisible-core 45.39.0, firefox-45)

[feder-cr]

The image's browser moves to invisible-playwright-mcp 0.72.2, with invisible-playwright 0.35.7 and invisible-core 45.39.0 on the firefox-45 engine.

What changes for a Dot

• Reads after a self-replacing page. browser_snapshot and browser_read_html read again on the page that replaced the one they went to (page.evaluate_read).
• The model's code runs once. browser_evaluate runs the model's code once when its page navigates under it, and answers with Playwright's navigation error instead of running it again (B285).
• firefox-42 to firefox-45 fixes:
  • A page that replaces itself keeps its main world. Before, a read in it could wait 30 s and fail (B286).
  • The real mouse no longer turns a click into a drag.
  • A session keeps focus when another starts beside it.
  • Iframes stay in the frame tree after a few navigations.
  • go_back/go_forward to a same-document entry answer.
  • A browser that dies ends every wait at once.
• Behaviour a page can measure:
  • Every input event is born a few ms before the page sees it, as with real hardware.
  • The first enumerateDevices of a session takes a time the seed declares, 240 to 640 ms.

The tools' schemas and descriptions are unchanged. The recaptured invisible_playwright_mcp.json differs only in its version.

Verification

Check	Result
Image tests (golden, runtime)	28/28
Browser smoke (real MCP and firefox-45)	74 passed, 0 failed or skipped
Engine smoke	257 passed, 0 failed
MCP probe in a real Dot, 7 pages that replace themselves, normal and with 16 cores saturated	0 read errors over 126 reads
reload-body probe in a real Dot	0 errors over 60 rounds
Real-VM e2e	not run locally (see below)

• In the MCP probe, read_text, snapshot and read_html took turns as the first read after browser_navigate.
• Every "interrupted" answer was genuine: after it, href and the document's age showed the page that replaced ours.
• The real-VM e2e did not run locally because snapshot.ubuntu.com answered 503 to the image build's apt downloads. CI's golden-image job builds it from this lock on merge.

Known

• One browser_navigate to a page that rewrites itself with document.write waited 45 s instead of answering. That was 1 round in 12 of the full probe, and 0 in 60 when the page was run alone.
• It is a rare race in the library, not specific to this lock, and is being looked at there.