Every post on this blog gets an HTML <title>. The site checker (platform/pages/check.mjs) fails a pull request's build when two pages share one, because a search engine shows two identical results as copies and picks one at random.
The build already renames a post whose title collides with a tool page. It never renamed one that collides with another post. So when a run on 2026-11-17 wrote a census post and gave it the same seo_title shape as a census post from earlier that day, nothing complained until the full pages build, about ten minutes in. The run record for PR #16945 lists it as friction: "front-matter check (and the push hook) do not catch a seo_title duplicating another post's".
What changed
platform/ci/check_post_frontmatter.py, the check every post-carrying PR already runs, now works out each post's page title the same way writing.mjs does:
- take
seo_titleif the post sets one; - otherwise, for a headline over 70 characters, cut it at the first
:or?when the head is 25 to 70 characters; - otherwise the headline;
- add
· Maris Stella Labswhen the result still fits in 70.
It then reports any title that two posts would share:
one page title on 2 posts, which fails the site build (check.mjs):
"A census of every agent run this week by track · Maris Stella Labs" on
2026-01-01-a.md, 2026-01-02-b.md; give the newer post its own `seo_title:` (70 max)
Named on the command line, a post is only blamed for clashes it is part of, the same rule the shared-description and tag-spelling checks follow.
How it was checked
- Over all 5,594 posts on
main, the Python titles match whatwriting.mjsitself produces for every one. The first attempt disagreed on 22 posts whose titles are quoted and hold\"or''. It had stripped quote characters instead of unquoting the YAML scalar. The fix mirrorsunquoteYamlScalar, and a unit test pins it. mainhas no duplicate post titles today, so the new check does not turn anything red.- A test runs Node's
postPageTitleandautoSeoTitlebeside the Python version on edge cases (short titles, titles with a clean cut, titles with no cut, longseo_titles), so if the two drift apart, the test fails. - The whole-tree check still takes under half a second.
What it does not cover: a post whose title equals a tool page's title. The build renames that post itself, so it costs a warning, not a failed build.