The agent index

This page holds the contract of docs/llms.txt and docs/llms-full.txt, the agent index the roster of generated projections lists with its regeneration command and its freshness gate.

Inputs

The docs root is SITE_KIT_DOCS_DIR. The site title is the first title: line of its _config.yml and the host the first line of its CNAME, so a page’s URL is https://<host>/ plus its path as the Pages build serves it: index.md as its directory, any other page .html for .md. The summary is the product statement’s plain rendering, read through that statement’s own parser. The pages are the nav model check-docs-nav-reachable reads, one crate module both call (§Site chrome and the nav contract), and a title, an order or a note is a front-matter value with one pair of enclosing quotes dropped.

llms.txt

The file opens # <site title>, a blank line and > <summary>. A section per top-level page follows in nav_order, index.md excepted since the header stands for it. The section’s heading is the page’s title. Its items are the page, then its nav children in nav_child_order, each a list item linking the page’s title to its URL, followed by : <note> where the page sets description:.

## Optional comes last. It lists each item’s generated directory siblings by title and its nav_suffix pairs as written, item by item in the order above, labelled <item title> <sibling title> or <item title> <pair label>. Then come the top-level pages naming nav_children_key, which take no section and leave their derived children unlisted. A page off the menu carries no nav key, so it reaches neither part, and no exclusion list is read.

llms-full.txt

The file opens with a door-contributor declaration, because every door in it is copied from a menu page the door gate already sweeps at its source, and a generated file’s doors take their declaration from the emitter (guard-kit/SPEC.md §check-door-binding). The same header follows, then index.md and every item outside ## Optional in the order llms.txt gives them. Each body is preceded by Source: <url> and a blank line. A body drops its front matter and every HTML comment outside a fence, and a comment filling its lines takes them and one blank line after them.

Every inline link target outside a fence or a code span becomes absolute. A scheme-led target stays. An #anchor takes the page’s own URL before it, and any other target resolves against the page’s directory, or against the docs root where it opens with /: a .md path takes the URL rule above, any other path becomes https://<host>/<path>, and its anchor is kept.

The arm and its gate

Without --write the arm prints llms.txt, and --write writes neither file until both render. Both refuse at exit 2 on an unreadable input, a _config.yml with no title: line, an empty CNAME, a listed page with no title:, or a nav_suffix pair naming no docs page. They also refuse on a link target resolving outside the docs root, and on a body carrying Liquid template syntax, which its raw source would publish. check-llms-fresh compares both files byte for byte with the arm’s rendering, in process. It also exits 2 on an absent projection file and names each stale one. Its positional form is check-llms-fresh [docs-root].