A Lossless Group package · MIT
MDX power without MDX's opinions. A polyglot extended-markdown pipeline — authors keep authoring, LFM normalizes the variations.
Let content creators be content creators. Identify the syntaxes worth supporting — and the variations on intent inside each one — then parse and render beautifully. LFM flexibly handles the rest.
pnpm add jsr:@lossless-group/lfm STC — Syntax · Trigger · Component
Add the syntax you want rendered as a component, plus the params or props it should carry.
Wire the parser to recognize it — a peer remark plugin file that emits one canonical AST node.
Map the AST node to a component, or a chain of them. Authors keep authoring; the renderer takes over.
Live example
bare URL or directive
youtu.be/share={id} :::youtube-share[https://youtu.be/jCe2wg1ulus?si=oplqTdsbv8sv2JfH] Features
Triggers, normalizers, and build-time enrichment. All wired together by
remarkLfm; cherry-pick what you need.
Many authoring conventions, one canonical AST. Authors pick the syntax their tool prefers; renderers see one shape.
:::callout{type="warning"}, > [!warning], and (eventually) Markdoc {% callout %}
all produce the same MDAST node. Adding a new authoring syntax means a new
normalizer plugin — no consumer changes, no renderer rewrites.
This is the headline pitch: LFM owns the trigger → tokenize → normalize path. Your framework owns rendering.
Tables, task lists, strikethrough, autolinks — GitHub-flavored markdown enabled by default.
GFM is the floor, not the ceiling. Tables and task lists come for free; they’re also the substrate for the bare-link auto-unfurl detector — a paragraph with a single autolink child is the GFM signal that LFM uses to dispatch embeds.
Directive syntax for tagging blocks, leaves, and inline runs as components.
Block, leaf, and text directives — :::box, ::badge, :abbr — give authors a
typographically lightweight way to name a region. LFM’s normalizers emit
directive nodes from many input syntaxes, so directives are also the lingua
franca your renderer dispatches on.
Obsidian `> [!type] Title` blocks normalized into directive nodes — same downstream shape, two upstream syntaxes.
Authors writing in Obsidian get callout previews in their editor. Authors
writing in plain markdown can use :::callout{type="warning"}. Both end up as
the same :::callout directive node — your <Callout> component renders one
shape regardless of which tool wrote the file.
Hex-code footnotes get renumbered to display indices and lifted into a structured citation dataset.
Author with stable hex IDs ([^a1b2c3]); LFM renumbers them to readable indices
at parse time and assembles the full dataset on tree.data.citations.ordered —
parsed dates, source domains, raw text — so a Sources component can render the
canonical bibliography at the bottom of the article.
Build-time OpenGraph enrichment. Cards and rollups carry full metadata before the page ships.
remarkOgFetcher walks the tree, finds external links and :::link-preview
directives, fetches OG metadata, and annotates the AST with LinkPreviewData.
Direct + OpenGraph.io backends, configurable cache, separate TTLs for hits and
failures so a transient upstream error doesn’t poison results for a week.
The LinkPreviewData shape aligns with the canonical Sources schema in
cite-wide, so promoting an OG card to a canonical citation is additive — not a
rename pass.
A URL on its own line becomes the embedded player or rich card the author meant. The catalog ships in the package.
src/plugins/Bare-Link-Provider-Catalog.md is the canonical record of supported
providers. v0.2.2 ships YouTube video / shorts / playlist and Vimeo as stable,
with Loom, Spotify, and SoundCloud planned.
The matching rule is strict: the paragraph must have exactly one child that’s
a link whose visible text equals its URL. So Check this out https://...
mid-sentence stays a clickable link; only an isolated bare URL becomes an embed.
Assign tasks with a plain checkbox line, then roll a whole repo's tasks into one file for GitHub.
- [ ] @mpstaton Task title #Docs 2026-11-15 reads correctly in Obsidian and on
GitHub with no plugin. Turn on taskMetadata and LFM also exposes the assignees,
tags, estimate and due date as data, leaving the text as written. The lfm-tasks
script then rolls a repo’s tasks into one file, each linking to its source line.
A fence registry that ships empty — you name the formats you want, and pay for nothing else.
Diagram fences normally mean a routing table inside somebody’s renderer, edited once per format and reimplemented once per site. This inverts it: the plugin knows no formats at all, and a handler is plain data plus an optional pure function.
Seven ship, none adding a dependency — yang and jsonSchema render as trees,
plantuml encodes to a server URL with no renderer at all, vegaLite parses
for a client chart, mermaid and graphviz just claim their language.
Code-fence formats
remarkCodeFences knows no formats until you name them.
Seven handlers are included and none add a dependency,
because "support a diagram language" means three different things
depending on who does the drawing.
yang yang RFC 8340 tree diagram nobody — it renders as text json-schema json-schemajsonschema schema tree, $refs expanded nobody — it renders as text plantuml plantumlpumluml server URL a PlantUML server, via <img> vega-lite vega-litevegalitevl spec + mark/channels summary vega-embed, client-side mermaid mermaid claims the language mermaid.js, client-side graphviz graphvizdot claims the language @viz-js/viz (WASM) jsoncanvas jsoncanvasjson-canvascanvas normalized nodes + edges your canvas renderer Recent
Code blocks get real syntax colors that follow each site's modes, and any site can swap in its own.
The README drops from 898 lines to a short front door with one feature table, and each feature's guide now lives in docs-v.
A new highlight import colors code with Shiki, using a default palette that follows every mode and CSS variables a site can override.