A typed HTML AST: parse any page or fragment into readonly nodes, render them back to canonical HTML or plain text, sanitize them against a floor no option can lower, and distill a page down to the prose a reader — or a language model — actually wants.
- Total parsing. Every input produces a document. No parse options, no issue list, no error path: malformed markup recovers per a documented table instead of throwing.
- One AST, many projections. Nodes are plain readonly data keyed by
category; querying, rewriting, folding, streaming, sanitizing, and rendering are all operations over it. - A real security floor.
sanitizeremoves unsafe subtrees whole, strips every handler and styling attribute, and decodes a URL before judging its scheme — whatever the options say. - Bounded by design. Every traversal and renderer is iterative and depth-capped, so hostile input degrades instead of exhausting the stack.
npm install @orkestrel/html- Node.js >= 22.12
- Ships ES and CommonJS builds with its own
.d.tstypes - One runtime dependency,
@orkestrel/contract
import { createHTML, renderHTML, renderText } from '@orkestrel/html'
const page = createHTML(
'<nav>Menu</nav><main><h1>Title</h1><p>Read the <a href="/b">guide</a>.</p></main>',
)
const safe = page.sanitize() // the whole page, made safe
const article = page.distill({ base: 'https://x.dev/docs/page' }) // its content, sanitized and extracted
// `distill` returns a handle, so the projection stays your choice.
renderHTML(article.document) // '<h1>Title</h1><p>Read the <a href="https://x.dev/b">guide</a>.</p>'
renderText(article.document) // 'Title\nRead the guide.'For a fail-closed source boundary instead of document recovery, use parseStartTag. It retains the
package's narrow ASCII tag-name grammar and returns an exact UTF-16 end offset:
import { parseStartTag } from '@orkestrel/html'
parseStartTag('<html lang="en" data-note="a>b">', 0)
// { name: 'html', attributes: [{ name: 'lang', value: 'en' }, { name: 'data-note', value: 'a>b' }], slashed: false, next: 32 }
parseStartTag('<html data-note="unterminated>', 0) // undefinedslashed reports a trailing solidus only when it is outside an attribute value; it does not make
an HTML element semantically self-closing.
renderText is flat but structural: tabs preserve table cells, newlines preserve table rows and
block boundaries, and whitespace beneath pre remains verbatim. Heading levels, link destinations,
list markers and ordinals, nesting depth, and image alt attributes do not survive. Read the
distilled AST, or serialize it with renderHTML, whenever those semantics are the point.
- AST fixpoint —
parseDocument(renderHTML(document))deep-equals a parser-produceddocument. - Canonical idempotence —
renderHTML(parseDocument(renderHTML(document)))equalsrenderHTML(document). - Sanitize fixpoint — sanitizing sanitized output changes nothing, directly or through a reparse.
What roundtrips is the AST, not the input bytes: canonical output lowercases names, quotes values,
canonicalizes character references, writes <br/> as <br>, and keeps dropped constructs dropped.
A hand-built AST that breaks an invariant — a void element with children, a comment body that could
close itself — is rendered for safety instead of fidelity.
For the full surface, the recovery table, the sanitize floor, and the distill pass, see
guides/src/html.md.
Published as a single typed entry point per the exports field in package.json.