Playwright 1.64 waits for a Playwright MCP release built on it: the image keeps one Chromium for both #157

Closed
opened 2026-10-08 23:17:22 -04:00 by cmoriarty · 2 comments
Owner

#155 tried to move Playwright from 1.63.0 to 1.64.0 and the Playwright MCP from 0.0.80 to 0.0.83, and held both back. The MCP server must use the same Chromium revision as the image's Playwright, so it never downloads a browser mid-step (openspec-flow#17). Read from the packages' browsers.json on 2026-10-08:

Package playwright-core Chromium
Playwright 1.64.0 1.64.0 1248
MCP 0.0.81 1.64.0-alpha-2026-09-14 1244
MCP 0.0.82 1.64.0-alpha-1789764292000 1246
MCP 0.0.83 1.64.0-alpha-1790635538000 1247

What it needs

When an MCP release is built on a final Playwright 1.64.x with the same Chromium revision, move PLAYWRIGHT_VERSION and PLAYWRIGHT_MCP_VERSION in the Dockerfile, osf.config.PLAYWRIGHT_MCP_VERSION, and @playwright/test in ui/package.json and its lockfile, together.

Done when

The image's self-check reports the MCP server's Chromium at the image's revision and the browser subagent's server opens a page, and the fast lane's Playwright specs pass.

#155 tried to move Playwright from 1.63.0 to 1.64.0 and the Playwright MCP from 0.0.80 to 0.0.83, and held both back. The MCP server must use the same Chromium revision as the image's Playwright, so it never downloads a browser mid-step (openspec-flow#17). Read from the packages' `browsers.json` on 2026-10-08: | Package | playwright-core | Chromium | |---|---|---| | Playwright 1.64.0 | 1.64.0 | 1248 | | MCP 0.0.81 | 1.64.0-alpha-2026-09-14 | 1244 | | MCP 0.0.82 | 1.64.0-alpha-1789764292000 | 1246 | | MCP 0.0.83 | 1.64.0-alpha-1790635538000 | 1247 | ## What it needs When an MCP release is built on a final Playwright 1.64.x with the same Chromium revision, move `PLAYWRIGHT_VERSION` and `PLAYWRIGHT_MCP_VERSION` in the `Dockerfile`, `osf.config.PLAYWRIGHT_MCP_VERSION`, and `@playwright/test` in `ui/package.json` and its lockfile, together. ## Done when The image's self-check reports the MCP server's Chromium at the image's revision and the browser subagent's server opens a page, and the fast lane's Playwright specs pass.
Author
Owner

Still blocked upstream, checked against npm on 2026-10-16:

  • The newest @playwright/mcp is still 0.0.83 (2026-09-28). It is built on playwright-core 1.64.0-alpha with Chromium 1247.
  • Final Playwright 1.64.0 (2026-10-07) and 1.64.1 (2026-10-14) both use Chromium 1248.

No MCP release is built on a final 1.64.x yet, so the image keeps Playwright 1.63.0 and MCP 0.0.80, which share Chromium 1243.

Related, from #170: project-playwright-browsers now installs a project's own Playwright browsers into a writable cache on the state volume. So a project on 1.64 gets its own build, and no longer depends on this bump.

Still blocked upstream, checked against npm on 2026-10-16: - The newest `@playwright/mcp` is still 0.0.83 (2026-09-28). It is built on `playwright-core` 1.64.0-alpha with Chromium 1247. - Final Playwright 1.64.0 (2026-10-07) and 1.64.1 (2026-10-14) both use Chromium 1248. No MCP release is built on a final 1.64.x yet, so the image keeps Playwright 1.63.0 and MCP 0.0.80, which share Chromium 1243. Related, from #170: `project-playwright-browsers` now installs a project's own Playwright browsers into a writable cache on the state volume. So a project on 1.64 gets its own build, and no longer depends on this bump.
Author
Owner

Shipped in 4309dfe (merge 49af5fc), live since 15:53Z on 2026-10-17, and archived in OpenSpec as playwright-mcp-from-core in 874eb27.

No MCP release was ever built on a final 1.64.x, so the change took the other route: the MCP server is now playwright mcp, from Playwright itself.

  • The server. The image installs Playwright 1.64.1 once, with Chromium build 1248, and the browser subagent's server is <playwright> mcp (4c12d8d).
  • One pin. @playwright/mcp and its pin are gone, leaving one pin, PLAYWRIGHT_VERSION. OSF_PLAYWRIGHT_CLI replaces OSF_PLAYWRIGHT_MCP.
  • The console's tests. The browser tests run on 1.64.1 too (33ca45f), and one unit test keeps the Dockerfile, osf.config and the lockfile's versions in step (4c12d8d).
  • The self-check checks three things: the image's Playwright version, the options osfd passes, and which Chromium build the server actually launches (4c12d8d, 83f03ff).
  • Old links. osfd removes the browser links a previous image left in the state volume's cache, and never removes a folder (9da3eb1, bc4878c).

Checked on production after 4309dfe:

  • osfd logged "removed links to browsers that are gone ...: chromium-1243, chromium_headless_shell-1243, ffmpeg-1011".
  • It also logged "linked the image's browsers ...: chromium-1248, chromium_headless_shell-1248, ffmpeg-1013".
  • Run 74's judge's browser subagent used the new server: 2 navigations, 5 screenshots, 7 clicks, 7 typings and 9 evaluations, all completed.
  • Run 75 ran one judge step, agent.judge.r3. Its first attempt used the browser, through three browser subagents, before an osfd restart cut it off; the second attempt judged from the first one's screenshot.
  • In osfd's container, find /var/lib/openspecflow/cache/ms-playwright -maxdepth 1 -xtype l printed nothing: no link dangles. The 1248 builds are links into the image, and the 1243 builds are real folders, a project's own install, which osfd leaves alone.

The deploy log's own lines weren't recorded. The deploy only tags an image latest after its self-check passes, so the check did pass. Closing.

Shipped in 4309dfe (merge 49af5fc), live since 15:53Z on 2026-10-17, and archived in OpenSpec as `playwright-mcp-from-core` in 874eb27. No MCP release was ever built on a final 1.64.x, so the change took the other route: the MCP server is now `playwright mcp`, from Playwright itself. - **The server.** The image installs Playwright 1.64.1 once, with Chromium build 1248, and the `browser` subagent's server is `<playwright> mcp` (4c12d8d). - **One pin.** `@playwright/mcp` and its pin are gone, leaving one pin, `PLAYWRIGHT_VERSION`. `OSF_PLAYWRIGHT_CLI` replaces `OSF_PLAYWRIGHT_MCP`. - **The console's tests.** The browser tests run on 1.64.1 too (33ca45f), and one unit test keeps the Dockerfile, `osf.config` and the lockfile's versions in step (4c12d8d). - **The self-check** checks three things: the image's Playwright version, the options osfd passes, and which Chromium build the server actually launches (4c12d8d, 83f03ff). - **Old links.** osfd removes the browser links a previous image left in the state volume's cache, and never removes a folder (9da3eb1, bc4878c). **Checked on production** after 4309dfe: - osfd logged "removed links to browsers that are gone ...: chromium-1243, chromium_headless_shell-1243, ffmpeg-1011". - It also logged "linked the image's browsers ...: chromium-1248, chromium_headless_shell-1248, ffmpeg-1013". - Run 74's judge's browser subagent used the new server: 2 navigations, 5 screenshots, 7 clicks, 7 typings and 9 evaluations, all completed. - Run 75 ran one judge step, `agent.judge.r3`. Its first attempt used the browser, through three browser subagents, before an osfd restart cut it off; the second attempt judged from the first one's screenshot. - In osfd's container, `find /var/lib/openspecflow/cache/ms-playwright -maxdepth 1 -xtype l` printed nothing: no link dangles. The 1248 builds are links into the image, and the 1243 builds are real folders, a project's own install, which osfd leaves alone. The deploy log's own lines weren't recorded. The deploy only tags an image `latest` after its self-check passes, so the check did pass. Closing.
Sign in to join this conversation.
No labels
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
cmoriarty/braid#157
No description provided.