ostat

0.5.0

Sat, 01 Aug 2026 08:45:00 +0000

Every page on this site sat under a breadcrumb that said blog. The section is called Release Notes. It has been called Release Notes since it existed, in the title of its _index.md, which is what the section heading has always shown.

The trail printed p.section, which is the directory name. That is a URL segment. It is only the section’s name when the two happen to be the same word, and the coincidence held for long enough to look deliberate: every directory anyone had made was already lowercase, so the trail looked like a considered lowercase style rather than a path being printed where a name belonged.

The crumb is the section’s title now, lowercased the way a crumb at the content root already lowercased its page title. The link is unchanged: the href is still the directory, because that is what a URL is made of.

<a href="/blog/">release notes</a>

Upgrading

A site whose section directories match their titles sees no change. A site where they differ sees the title, which is the point. If a section should be called what its directory is called, that is the title in its _index.md.

0.4.0

Sat, 01 Aug 2026 08:30:00 +0000

Three of the release notes below this one were published on the same day, and came out in the order 0.2.0, 0.3.0, 0.2.1. That is not their order, or their file order, or any order at all.

Two things were wrong, and the second is the one that matters.

A date is a day, so three posts on 2026-08-01 were equal, and no comparison of their dates could separate them. And slice.sort_by is not stable, so equal posts did not merely come out in an arbitrary order: they came out in an order that could change when an unrelated post was added.

A date may carry a time

"date": "2026-01-30"
"date": "2026-01-30T14:30"
"date": "2026-01-30T14:30:05"

Optional, and midnight when absent, so every stamp written before this means what it meant. Written large unit first a stamp sorts chronologically as a string, which is what the sort already relied on; a bare date sorts at the start of its day because "" is less than "T".

The slug now breaks a tie between identical stamps, so the comparison is total and an unstable sort cannot reorder anything.

What else it fixed

<pubDate> said 00:00:00 for every post ever published, because the day was all the front matter could say. It now carries the real time.

A listing shows the day and puts the whole stamp in the <time datetime> attribute, which is what that attribute is for. <lastmod> in the sitemap stays a date: a W3C datetime needs a timezone once it carries a time, and ostat does not know one.

Upgrading

Nothing to do. A site whose posts are dated to the day builds identically, and gains an order only if it starts writing times.

0.3.0

Sat, 01 Aug 2026 01:08:23 +0000

A section listing was a link and, for the blog, a date. That is enough for posts, where a title and a date are the conventional index. It is thin for a section whose pages carry no date, where the link is the entire entry.

Section listings now carry the page’s description under its title, in a <p class="entry-summary">.

The home page’s recent list does not. It is five entries sitting under two other sections, and its job is to say what is new rather than what each thing is. A section index has the room; the front page does not.

The summary is the front matter’s description and nothing else. page_description would fall back to the opening paragraph, which is right for a meta tag nobody reads in place and wrong here: an author who wrote no description gets no line rather than an arbitrary one.

Upgrading

Nothing breaks. A site with no rule for .entry-summary renders it as an ordinary paragraph, which is legible if not styled. site/static/css/style.css sets it muted and slightly smaller, the same treatment <time> already gets in a listing.

0.2.1

Sat, 01 Aug 2026 00:58:08 +0000

One fix, found by using the generator rather than reading it.

write_post_entries wrote a <time> for every entry in a listing, so a page without a date got <time datetime=""></time>. An empty datetime is not a valid date string, and a <time> carrying no date is not a time.

Only the blog is dated¹. The blog was also the only section anyone had put pages in, and a post is required to carry a date, so the broken shape had never been rendered. It appeared the first time a second section got a page.

write_single has guarded this since it was written; this is the same guard in the other place that writes a time. The fixture site gains a notes/ section holding one undated page, because the golden tree only knows the shapes the fixture contains and this was not one of them.

section_pages says so in a comment two procedures below the bug. ↩

0.2.0

Sat, 01 Aug 2026 00:15:41 +0000

Ten issues, closed together. Most of them were the same bug wearing different clothes: a build that exited 0 and wrote a site that was wrong in a way nothing on the site revealed.

Breaking

The home page is no longer described in site.json. portrait, contact, elsewhere and home are gone, along with the three structs behind them and the procedure that turned them back into the markup they started as. A site now writes html/home.html and gets the whole page, with  marking where the recent posts go¹.

html/ also holds header-01.html and header-02.html, the two halves of every page’s <head>, and not-found-404.html. Each is optional and falls back to a built-in default, so a bare content directory still builds. This is the arrangement gingerBill’s generator uses, read at runtime rather than compiled in, because compiling it in is what would undo multi-site support.

Three content trees that used to build now fail, on purpose:

A directory inside a section directory. Section membership was an exact string comparison, so a page under blog/2024/ was built and crawlable and reachable from nowhere on the site.
A section directory with no _index.md. Every page inside it linked a breadcrumb to a page that was never written.
A root content/_index.md. It was loaded, written, and then overwritten by the synthesised home page, and the sitemap advertised the front page twice.

An _index.md marked draft removes its section, so that now fails too.

Fixed

Two pages resolving to one output path both stayed in every listing, the feed and the sitemap, while the later write took the path
A file in static/ silently replaced the generated page at the same path, including the home page, the feed and the sitemap
-o, -base-url, -today and ostat new -s took a following flag as their value, so ostat build . -o -drafts created a directory named -drafts and reported success
truncate_words measured its limit in bytes and cut on a byte offset, which landed mid-codepoint and put invalid UTF-8 in the meta description of any page whose summary had no ASCII space in it
Feed items kept root-relative URLs in single-quoted attributes, srcset, poster, cite and CSS url(), where they resolve against the reader’s host
A bare & made strip_tags search the rest of the document for a ; and swallow everything between it as one entity
base_url and locale reached six URLs and attributes unescaped, and -base-url replaced the value after it was checked

Also

ostat writes a 404.html, which it did not have at all.

The test suite went from 82 to 98. The golden tree is unchanged except for the new 404 page, which is what says the home page migration was faithful rather than merely finished.

The data model had to grow a field for every idea a home page might have, and the generator a branch to draw it. A fourth section meant editing write_home and recompiling; a paragraph or a table had no representation at all. A document is stored as a document now. ↩

0.1.1

Wed, 29 Jul 2026 00:00:00 +0000

A whole-project audit, then the work it turned up. The headline was that the preprocessing pass could not see inline code, so a regex in backticks aborted the build¹ with an error naming a note nobody had written.

Also fixed:

Tildes inside a code span no longer become <del>
A fence closes only on a matching run, so a longer fence holds a shorter one
A fenced block indented under a list item stays in its item
Front-matter slugs and note labels are validated before they reach the output
Repeating one note reference no longer repeats its id

And the shape of the thing: building a document is separate from writing it, so the layouts are testable, and there is now a fixture site with a committed tree of expected output covering the third of the source that no test reached.

[^a-z] was read as a reference to a note called a-z. An unresolved marker is prose now, and a mistyped label is caught from the other side instead. ↩

0.1.0

Tue, 28 Jul 2026 00:00:00 +0000

The first version that generates a whole site.

Content discovery with drafts, future-dating and slug overrides
Margin notes in the Tufte shape, collapsing to numbered endnotes in the feed
Tables, strikethrough and syntax highlighting, none of which CommonMark has
An RSS feed at two addresses, and a sitemap
Layouts as Odin procedures, with no template language

Set in ET Book, served from this domain with font-display: optional, so the page never trades one typeface for another while you are reading it.