Generated from fhir-harness/skills/fhir-ig-base/ig-render-jekyll.md — do not edit here.

✎ Edit this page’s source

ig-render-jekyll

Skill id: ig-render-jekyll · Package: fhir-ig-base · Instance: fhir-harness · Beans jut3, kn0t

Render an IG’s pages as folio pages — markdown with front matter, through Jekyll and just-the-docs — populated from the Publisher’s structured output, rather than copying the Publisher’s finished HTML into place.

Whether an IG repository gets this rendering at all is the owner’s call, asked when the repository is adopted, default yes. The question, and the manual-only workflow template it installs, are in ig-build-pipeline §”Who starts a build”.

Why, in one property

Mounted HTML is opaque to everything the harness does. It carries no front matter, so the navbar, the language bar, the QA badges and the translation surface all stop at its edge. Rendering through the pipeline makes an IG page an ordinary folio page and those surfaces reach it for free.

That is not a claim about looks. Bean hw9g exists only because a mounted page got no sidebar, and it was resolved by injecting a rail into HTML that had no business needing one.

The three contracts

1. JSON only

The render path takes JSON, wrapped in JSON-LD / JSON Schema. Not XML, not Turtle.

The Publisher emits all three per resource. Taking one is a decision, not an omission: three serialisations of one resource are three chances to disagree about what the resource says, and the JSON-LD wrapper is what carries the semantics the other two were being kept for. An instance that genuinely needs RDF gets it by projecting the JSON-LD, not by shipping the Publisher’s .ttl.

Record a refusal rather than silently dropping. A representation that arrives and is not taken is reported, so “this IG publishes no Turtle” and “we ignored its Turtle” stay distinguishable.

Not linked either. Owner, 2026-10-23: “excise XML and ttl everywhere for now”. The site does not link the Publisher’s .xml or .ttl copy of a resource. An artefact’s JSON-LD node carries only its JSON encodings, and the site.data artefact variables carry only url.json. The refusal record still counts what each IG published in those forms, so the excision stays visible.

2. Navigation is derived, never authored

sushi-config.yaml already carries both maps the site needs:

• pages: — an ordered map of source file to title. That is the page order.
• menu: — the grouping, including nested subsections.

Derive the navigation from them. Do not hand-author a nav file, and do not write entries back into sushi-config.yaml to make them visible — that is what an overlay’s pre-processing may do, and it is a round trip through the tool being removed.

When the config names a page that does not exist, that is a finding, not a reason to create a placeholder.

3. LHS rail, and the theme is captured rather than replaced

Navigation goes in the left-hand rail, not the FHIR IG’s top bar. The rest of the IG theme — its CSS, its typography, its resource-page furniture — is captured into the theme layer rather than discarded: the goal is an IG that reads like an IG, in a shell that behaves like the rest of the site.

Two mechanisms not to conflate: which side the nav is on is a layout decision made once, and what the nav contains is derived per build. A change to one is not licence to change the other.

Where the variables come from

generate_smart_liquid.py in the WHO build already computes IG metadata → Liquid variables from output/, naming them smart__<ResourceType>__<id>__<category>__<key> across url__canonical, url__page, url__json, text__display, link__html and elements__<key>.

Lift that, do not redesign it. It is aimed at the Publisher’s own Jekyll and needs re-pointing, which is a smaller change than it looks and is phase P0 of kn0t. Its one defect — the documentation page it writes is processed on the next build, so the variable surface takes two builds to converge — goes away when variables are computed and consumed in one pass.

Naming: the lifted variables are not smart__-prefixed at this layer. That prefix is WHO’s, and this layer does not know about WHO.

What parity means, and what it does not

Parity is a checklist of asset types and page kinds, stated before MVP is declared, not an impression formed by looking at a page. The ceiling is a data limit rather than an effort one: a page kind whose source data the Publisher does not emit cannot be rendered at any effort, and belongs on the AST ask rather than on the parity list.

Declaring MVP against the checklist is the owner’s call, not this skill’s.

Thin pages, the default (--full-pages to opt out)

The owner, 2026-10-11, option A: “tiny html w/ dynamic load from json(ld) KG”; and 2026-10-12: “make thin default”. Every page of an IG site is now thin unless a build asks for the old pages with --full-pages. The shell is cat-harness-tools’ thinPageHtml, imported and not copied: this layer needs core, which sits above cat-harness-tools.

Artefact pages (gen-ig-pages.ts)

• The address does not move. The page stays at artifact/<Type>-<id>.html and names that file as rel="canonical". The IG’s artifacts page, the narrative’s links and the footer’s reading order all still land.
• The asset is a JSON-LD node. The node, artifact/<Type>-<id>.jsonld with a .json twin, is written by ig-artifact-jsonld.ts from the index. It uses the FHIR RDF class, dcterms, schema.org and PROV, and mints nothing. It is kind ig-artifact-jsonld, and its QA is ig-artifact-jsonld.ts check over the built site. The page names the node as alternate, and the node names no rendering.
• One shared renderer. assets/ig-artifact.js fetches the node, then the resource the node names as a local copy. It draws the description, facts and identity table, the elements, concepts or includes, and then hands the narrative and JSON to ast-resource.js. The page itself carries only its title, its key and a <noscript> link to the node and the resource.
• The rail is linked. The page declares folio-navbar: linked, and the build’s rail pass (rail-standalone-pages.ts) links the rail’s style and scripts instead of inlining them. Measured on the iHRIS 5 IG (182 artefacts): 168 KB per full page, about 5.6 KB per railed thin page.
• The tab views are thin pages too. Owner, 2026-10-14, answering whether the IG Publisher’s per-artefact tab views are rendered: “3. Render them too”. This supersedes “IG API, JSON-view and tab pages stay full pages”. Wherever the site serves the artefact’s JSON, its views are written beside its page under artifact/. That JSON is either a compiled copy (--compiled-data, which ig-site:compose passes as ast-data: the AST’s copy) or the Publisher’s own <Type>-<id>.json carried as data (--publisher-data <dir>). Each view keeps the Publisher’s file name:
  • -definitions: the element table, from the snapshot, with the differential’s elements marked;
  • -mappings: the mapping declarations, each with the elements that map to it;
  • the JSON view, .profile.json for a StructureDefinition and .json for anything else, drawn by ast-resource.js (data-ast-view="json");
  • -examples: for every StructureDefinition, listing from the index node the instances that claim it in their meta.profile (see below).

  Each is the same shell (thinArtifactView) and the same renderer (assets/ig-artifact.js, the config’s view). The artefact’s page links them in a tab bar of plain links. The narrative’s links to those file names are relinked the same way as its links to the artefact pages. A full tab page at the same address is replaced. The change history, testing, XML and Turtle views are not rendered, and none of them is published. The IG API view pages stay full pages.

• Every profile gets its examples view (owner, 2026-10-14: “1a”). Each StructureDefinition gets -examples, as the Publisher writes it. Where no instance of the IG claims the profile, the page says there are no examples.
• Nothing under comparison-v*/ is published (owner, 2026-10-14: “2a”). The Publisher’s version comparisons are not rendered, and ig-site:publisher-pages copies none of their files: it serves the AST’s data and reads no Publisher output at all.
• In an IG site the docs root is the site root, so --compiled-data is ast-data. With ../ast-data, every resource link points one level above the site.

The artefact list and the table of contents (build-ig-site.ts)

• One index node. Beside the artefact nodes, artifact/index.jsonld (family folio-ig-artifact-index-jsonld/v1, same kind) lists every artefact once, in the Publisher’s order, each entry naming its node and carrying only what a list row shows. check holds the list to the artefact index and each entry to its node: key, class, canonical, profiles and code.
• artifacts.html is a thin shell over that node, drawn by the same renderer: a jump list of categories, a filter, then every category OPEN with its table. The owner could not find artefacts behind collapsed <details>, so nothing is collapsed. Category anchors are the full page’s (#artifacts-<slug>), so old links still land.
• What a list draws, and what it does not (owner, 2026-10-12):
  • The IG’s own ImplementationGuide resource is “This guide”, drawn first with its description and a link (#artifacts-this-guide), never as a “Not categorised” row at the end. It is the guide, not one of its artefacts; the toc lists it first the same way.
  • A category that mixes resource types is sub-headed by type (#artifacts-<category>-<type>; the Publisher’s “Other”: Basic, Questionnaire, DocumentReference…), and the jump list counts each type.
  • A table draws only the columns its rows fill. No row with a description: no Description column. An instance (an artefact with no canonical URL) has no description, so the index carries what the resource states instead, copied, never looked up: its meta.profile (dcterms:conformsTo, drawn as the profile’s last path segment, linked when it is an artefact of the same IG) and its code (dcterms:subject). A definition’s core profile (shareablevalueset) is not written: on a phone it cost a column for nothing.
• toc.html is a thin shell. The IG’s own pages (sushi-config.yaml pages:, the landing page first) stay in its HTML: they are the page’s structure. The artefacts under each category are drawn from the index node.
• Both are written by stageIgSite when StageOptions.artifacts.index is given, which stage-ig-sites.ts does unless --full-pages. A site whose artefact pages were generated with --full-pages has no index node, and staging it thin is refused rather than shipping two pages that draw nothing.

The narrative pages (composeIgSite)

• The IG’s own prose stays in the HTML, rendered by Jekyll as before. What changes is the layout: an IG page composed into a host site wears thinIgLayout (_layouts/ig-thin-<instance>.html) instead of the host’s layout, so none of the host’s chrome (the harness bar, the inline SVG sprite, the sidebar nav, the inline scripts) is written into it.
• Everything around the prose is LINKED: the shared assets/ig-artifact.css, the top bar’s assets/ig-topbar.css, the edit and source-link runtimes as assets/ig-source-links.js, and the rail through folio-navbar: linked. The IG’s top bar, edit links, footer and its TOC declaration for the rail stay. Measured on iHRIS 5: index.html 182 KB to about 11 KB.
• What the host’s chrome added and the thin page does not: the host’s search box and the QA badges. The thin page has its own search box instead, and the language picker’s block (below).

Search and the language picker (owner, 2026-10-15)

Layout “C”, chosen from wireframes: the folio rail stands in for the theme’s side menu (#site-nav is dropped by ruling, IG_DROPPED in stage-local-ig.ts), and the IG’s top bar carries a search box over the IG’s pages and artefacts (ig-search.ts):

• ig-repo-site.ts compose writes the index, assets/ig-search.json (each of the IG’s own pages, its title and its whole text — owner, 2026-10-15, “index whole pages”, where it had been cut at 6000 characters; each artefact, its title, id, name, key, canonical and description), and the client, assets/ig-search.js, under the docs root. The index is fetched on the first focus of the box, never on page load.
• The box (form.ig-search, input #search-input) is in the composed pages’ .ig-topbar and in the thin pages’ header.iga-top (artefact pages, their tab views, artifacts.html, toc.html). It is not .search, so docs-ui leaves it in the top bar rather than moving it into its band. A build that does not write the files (gen-ig-pages without --search, StageOptions.search unset) writes no box.
• Every other built page without a box gets one: publisher-pages puts a bar at the top of its body (header.ig-search-top: the IG’s name, the box, the picker’s slot) with its own styles and the client, paths relative to the page (stampSearchBox, withSearchBar). These are the pages a layer above or below writes outside the IG’s bar, such as an IG API hub page and openapi/index.html (owner, 2026-10-15). The pass is keyed on what a page lacks, never on its name, so this layer names none of them; it skips assets/ and does nothing when the site serves no index. The pages’ writers are left alone: cat-harness-tools, below, may not import the box, and the overlay above would have to repeat it per page kind.
• The language picker sits in the bar, beside the box (owner, 2026-10-15: “move the language picker into the IG top bar”). The box is followed by an empty <span data-fa-slot="locale">; docs-ui fills the first such slot on a page and, on a page with none, keeps the picker in its own band. So a folio site with no IG bar is unchanged, and this layer knows nothing of the picker beyond the place it offers.
• “what happened to search/language select at top of display window?”: docs-ui draws the language picker from a page’s fa-translation-meta block, which only a theme-laid-out page carried. publisher-pages writes the platform’s block for a page Jekyll never laid out (cat-harness-tools lib/translation-meta.ts) into every built page that lacks one, from the IG checkout’s translation declaration. The site build’s set-html-lang.ts reads it afterwards.

The root parity guard (cat-harness-tools site-parity.ts) requires both: a build without the search box or the language block is not published over a root that had them.

The opt-out

gen-ig-pages.ts --full-pages and stage-ig-sites.ts --full-pages build the pages as before 2026-10-12 (composeIgSite(…, { fullPages: true })). Pass both or neither. gen-ig-pages --thin is still accepted and changes nothing; --thin --full-pages is refused. Removing the full path is a separate decision.

What this does not replace

Which AST the pages are drawn from. An IG repository’s site is drawn from the AST on its cache branch, cat/fhir-harness/fhir-ast/<package id>. The publish refuses one that is not valid for the commit, and is never drawn from the Publisher’s HTML. The cache is refreshed by a Publisher run on a runner. ig-build-pipeline §”The AST cache lifecycle” has the rule and the workflow.

The Publisher still runs. This skill is about where pages are rendered, not about whether the IG is built: validation, dependency resolution and terminology expansion are the Publisher’s and are not reproducible from cached output. The phased reduction of the Publisher to AST + QA is bean kn0t; until it lands, assume a full build.

Processes that run this skill

process	step(s) that name it
An IG site, built from the AST on its cache branch	Build and publish the site from the AST