XMarkdown status and remaining work

Status checked on 8 November 2026, MVP version 0.2.1. The extended renderer, comparison fixtures, and local demo exist. APIs and syntax remain experimental before 1.0.

Implemented

Ordinary Markdown through configured markdown-it, including tables and strikethrough. Main uses 15.0.3; the v0.2.0 release and checked-in benchmark baseline use 14.3.3. The core syntax has not been replaced.
:::note, :::tip, and :::warning callouts containing Markdown.
Closed x-chart fences with validated JSON for categorical bar and line charts.
Closed x-diagram fences with validated JSON for basic directed graphs, automatic placement, and SVG output.
Inline/display TeX and x-math, compiled by lazily loaded KaTeX 0.19.1 to native MathML.
x-widget controls with a bounded expression interpreter; a drawing widget with undo/clear and SVG polylines.
x-svg scenes and existing SVG imports with paths, groups, gradients, clipping, masks, markers, patterns, and reusable symbols.
Bounded caching of validated specifications and generated geometry; optional widget modules/styles and gzip/ETag delivery in the demo.
Diagnostics and escaped-source fallback for invalid graphics; source line maps and explicit resource limits.
Provisional streaming section patches and one complete parse at completion. The playground currently displays a complete document and inspection counters.
An HTML/SVG/MathML renderer, shared document styling, and a sandbox without same-origin permission. Interactive documents load only the trusted widget module.
A reproducible comparison with twelve examples, two named tokenizers, complete source artifacts, and a gallery. The HTML JSON adapter reuses the chart/diagram renderer; controls use the shared widget module. SVG scenes remain literal SVG in this HTML adapter comparison.
Version 0.2.1 package named xmarkdown-renderer, with Node/browser/React/CSS exports and TypeScript declarations. Both parser entries share the native core; browser optional processors use dynamic imports. The older v0.2.0 GitHub release retains its original name and imports.
Dark showcase with eighteen examples, five larger fixtures, native local edits, a live React component page, CI, maintenance policies, and installation guides. Full framework applications beyond the React smoke page remain untested.

Raw author HTML is escaped; ordinary Markdown images show alt text without fetching. Unsafe link schemes are blocked. Task-list checkboxes and automatic linkification remain absent. Mermaid stays source. Arbitrary author programs, external resource loads, SVG filters/animation, and unrestricted application logic remain outside the profile. SVG import accepts a validated subset and rejects unsupported content. See extended capabilities for exact syntax and limits.

Measured and unmeasured

The comparison still counts the original demo as 839 UTF-8 bytes / 208 o200k_base tokens of XMarkdown, 3,492 bytes / 1,219 tokens of compact HTML/SVG, or 1,044 bytes / 261 tokens of shared-component HTML. Source-token reduction stays 82.9% versus compact HTML/SVG and 20.3% versus shared-component HTML. Extended cases now include math, controls, drawing, scenes, and imports. Raw SVG imports add fence overhead. Counts exclude reusable runtimes and instructions; the extended guide adds 383 tokens if sent. No current Claude tokenizer, model API, generation latency, billed usage, or repair rate was measured.

The parser benchmark and extended capability benchmark describe the v0.2.0 baseline using markdown-it 14.3.3. They have not been rerun for the 15.0.3 dependency in v0.2.1. Feature measurements separate first use, empty-cache, disabled-cache, cached rendering, and streaming. Cached source is cheaper to process, but new capabilities have real costs and short documents may gain nothing from streaming. The requested 50× single-parse improvement remains unachieved. Browser layout/painting and model generation remain unmeasured.

Remaining work, in recommended order

Priority	Work	Completion evidence
1	Validate the product hypothesis against styled Markdown/plugins, Markdoc/MyST, shared-component HTML, and standalone HTML.	Capability-matched model trials with total prompt/output tokens, latency, valid output, repairs, and reader task results.
2	Stabilize the small grammar and versioned API. Evaluate JSON versus a more compact graphics notation before changing syntax.	Written compatibility policy, versioned normalized nodes independent of markdown-it tokens, and measured generation/error tradeoffs.
3	Integrate actual model streaming in a browser with a cancellable worker.	Patch-based UI updates, pending-content display, final equivalence, cancellation, and measured first useful display/interaction latency.
4	Complete remaining Markdown and asset features.	Explicit choices for task lists, Mermaid rendering, citations, and images with a controlled host policy. Math, controls, drawing, and supported SVG imports are implemented.
5	Improve graphics quality and accessibility.	Better edge routing and labels, additional chart types if justified, accessible data tables, keyboard/screen-reader review, responsive checks, and reliable static export.
6	Harden the parser and renderer.	Chosen CommonMark/GFM conformance corpus, fuzzing, truncation and malformed-input tests, network-boundary checks, memory/time measurements, and dependency/security review. Current allocation limits are not hard isolation.
7	Broaden integration verification and release maintenance.	Package exports, React, TypeScript and CI are implemented. Test complete framework applications and expand conformance/migration coverage before a stable release.

The implementation should not grow into a second HTML/CSS/JavaScript language. Known document widgets can be added through trusted schemas; unrestricted application design can use HTML and JavaScript directly. A new native or WASM parser is justified only if profiling identifies a remaining bottleneck and conformance can be maintained.