Scope Doors preview browser checks to changed pages

LU
6h ago
+787-211

Doors preview resolves the production data and announcements artifacts from successful main workflow runs. The candidate, PR base and main reference build in parallel from those pinned inputs. The Contributors page comes from the pinned data artifact. For scoped PRs, the base job reuses a successful push build at the exact base SHA when its manifest records the same artifacts and Contributors provenance. The comparison waits for both sites and checks URLs, structured data, SEO tags and internal links across the whole site.

The browser checks compare rendered HTML. CSS and JS changes add the core sample and retain all changed articles, including generated articles outside that sample. Changes to either the home page or Resources run the topic/facet assertion across both pages. Pushes, check infrastructure changes, static files and non-bundle assets, the full-checks label and diffs over 60 pages run the full suite.

Parallel reference benchmark (UTC)

Draft #1286's Doors preview run changed one Markdown page and selected only /replication-hub/submit/. The temporary base enabled the workflow trigger and had no reusable push build. The draft was closed and both temporary branches were deleted.

JobStartEndResult
Resolve build inputs13:37:2513:37:28Pass
Build13:37:3113:40:35Pass
Build PR base13:37:3113:40:35Pass
Build main reference13:37:3113:39:11Pass
Compare with main13:40:3813:41:55Pass
Designed screens13:40:3813:41:44Pass
Accessibility rules13:40:3813:41:44Pass
Deploy preview13:40:3613:40:36Skipped
Check results13:41:5813:42:00Pass

The gate passed 4m 35s after the first job started. Compare with main took 1m 17s, down from 2m 27s in the earlier content-only run. Candidate build times also differed between the runs.

Local validation passed for actionlint, 209 unit tests, the Hugo build, SEO HTML, sitemap lastmod against source history, and the comparison with main using production artifacts.

Earlier job timings on 2026-11-21 (UTC)

The draft PRs targeted a temporary base branched from this head, with deployment disabled and only the branch triggers changed. All three reused the same successful push build. In the previous cold runs the base job took 4m 32s for content and 5m 36s for CSS. With reuse it took 22–24s, and the browser jobs started as soon as the candidate was ready.

JobContent onlyCSS onlyCSS + content
Resolve build inputs13:10:27–13:10:30 (3s)13:17:12–13:17:15 (3s)13:17:11–13:17:15 (4s)
Build13:10:38–13:15:26 (4m 48s)13:17:17–13:21:54 (4m 37s)13:17:18–13:21:41 (4m 23s)
Build PR base13:10:33–13:10:55 (22s)13:17:17–13:17:41 (24s)13:17:18–13:17:40 (22s)
Designed screens13:15:29–13:16:17 (48s)13:21:57–13:26:42 (4m 45s)13:21:44–13:26:08 (4m 24s)
Accessibility rules13:15:29–13:16:19 (50s)13:21:57–13:25:09 (3m 12s)13:21:44–13:25:32 (3m 48s)
Compare with main13:15:29–13:17:56 (2m 27s)13:21:57–13:24:51 (2m 54s)13:21:45–13:24:57 (3m 12s)
Check results13:17:59–13:18:03 (4s)13:26:44–13:26:46 (2s)13:26:11–13:26:20 (9s)
First job start to gate end7m 36s9m 34s9m 09s

All three Doors checks passed. The content trial selected only the edited educators-corner page. The CSS trial selected core mode with zero changed HTML pages and 46 article pages. The mixed trial retained the core sections, added the edited article to the article set for 47 pages, and ran its call-out and phone-reading sections. The initial CSS fixture did not change the emitted bundle, so the table uses the corrected stylesheet-token trials. All three drafts were closed and their branches and the temporary base removed.

Earlier workflow totals, including queueing
WorkflowContent-only PRCSS-only PRFull
Before #1278~11 min~11 min~11 min
First #1278 version7m 51s10m 28s—
Updated, cold cache6m 43s11m 00s11m 13s
Updated, warm-cache rerun7m 00s5m 45s, Contributors only—

Implemented with GPT-6.2 Sol in Codex.