Jewel Markdown

Jewel Markdown is a two-stage renderer: parse raw Markdown into Jewel's MarkdownBlock / InlineMarkdown model, then render that model with Jewel Compose renderers. Choose the smallest customization point that matches the job.

The Jewel Markdown sources live in JetBrains intellij-community under platform/jewel/markdown/. Confirm exact API/file/symbol names by reading those sources (grep/read) rather than trusting memory before proposing code. For deeper topic detail, read the matching reference only when that topic is in play:

• references/CODE-HIGHLIGHTING.md — fenced/indented code highlighting, CodeHighlighter, why code is unstyled by default.
• references/IMAGE-LOADING.md — ImageRendererExtension + ImageSourceResolver, Coil3, path resolution.
• references/SCROLL-SYNC.md — editor-preview scroll synchronization and its ScrollState-only limitation.
• references/HTML-PARSING.md — embedded HTML, built-in tag conversion, MarkdownHtmlConverterExtension, alignment.

First decision: what are you changing?

1. Visual styling only — customize MarkdownStyling and theme-specific styling helpers.
2. Existing block UI behavior — use a custom MarkdownBlockRenderer, usually by subclassing DefaultMarkdownBlockRenderer and overriding the specific Render* method.
3. Existing inline text behavior — use a custom InlineMarkdownRenderer, usually by subclassing DefaultInlineMarkdownRenderer.
4. New Markdown syntax — write a processor extension and a renderer extension.
5. Images — use an ImageRendererExtension, not a general block renderer.
6. Embedded HTML — enable parseEmbeddedHtml on the MarkdownProcessor, and for custom HTML→Markdown mapping add a MarkdownHtmlConverterExtension.
7. A live editor preview — use MarkdownMode.EditorPreview with a dedicated MarkdownProcessor instance and render with LazyMarkdown.

Do not start with a custom renderer when styling or an extension solves the problem.

High-level rendering flow

1. MarkdownProcessor parses raw Markdown into List<MarkdownBlock>.
2. Blocks may contain InlineMarkdown nodes.
3. Markdown renders short/simple content in a Column; LazyMarkdown renders larger or editor-preview content in a LazyColumn.
4. MarkdownBlockRenderer renders block nodes and delegates inline content to InlineMarkdownRenderer.
5. ProvideMarkdownStyling wires LocalMarkdownStyling, LocalMarkdownProcessor, LocalMarkdownBlockRenderer, and code/image support into JewelTheme.
6. Extensions can plug into processing, block rendering, and inline rendering (with some limitations).

Prefer processing Markdown outside composition for non-trivial or frequently changing text. The string-taking Markdown(...) overload is convenient, but it parses from composition and is best for small, mostly static snippets.

Built-in and existing extension choices

Use the existing extension modules before inventing syntax:

• Plain URLs are not becoming links: add AutolinkProcessorExtension to the MarkdownProcessor.
• GFM tables: add GitHubTableProcessorExtension and GitHubTableRendererExtension. The processor extension also bundles a MarkdownHtmlConverterExtension for <table> (active when parseEmbeddedHtml = true).
• GFM strikethrough: add GitHubStrikethroughProcessorExtension and GitHubStrikethroughRendererExtension. The processor extension also bundles a MarkdownHtmlConverterExtension for <s>, <strike>, and <del> (active when parseEmbeddedHtml = true).
• GFM alerts: add GitHubAlertProcessorExtension and GitHubAlertRendererExtension.
• Images: add an ImageRendererExtension, commonly Coil3ImageRendererExtension, to the renderer extensions and provide an image source resolver where needed. For details read references/IMAGE-LOADING.md.
• Code syntax highlighting: in a plugin, use the Project-aware bridge ProvideMarkdownStyling, which wires an IJPL-backed highlighter by default; in standalone (or a bridge overload without a Project) the default is a no-op and you must provide a CodeHighlighter. For details read references/CODE-HIGHLIGHTING.md.

For a worked example of combining several extensions (processor + renderer + theme-specific styling, parsing off composition, LazyMarkdown), read the standalone sample at platform/jewel/samples/standalone/src/main/kotlin/org/jetbrains/jewel/samples/standalone/markdown/MarkdownPreview.kt.

Keep parser and renderer extensions paired when a feature has both sides. If a processor creates a custom block or inline node and no renderer extension claims it, the content will not render usefully. Do not invent renderer extensions for parse-only features such as autolink, and do not invent processor extensions for renderer-only image loading.

Embedded HTML

HTML support is off by default: a MarkdownProcessor ignores embedded HTML unless you construct it with parseEmbeddedHtml = true. There are two distinct HTML paths, do not conflate them:

• Native HTML mapped to Markdown: when parseEmbeddedHtml is on, a fixed set of built-in tags is converted into normal MarkdownBlocks ( p, li, ol, ul, h1-h6, code, pre, img). These render through the standard block/inline renderers, not through a raw-HTML renderer.
• Custom tag conversion: to map an additional HTML tag onto a Markdown block your code understands (e.g., <table> to a tables block), implement a MarkdownHtmlConverterExtension and expose it from your MarkdownProcessorExtension.htmlConverterExtension. It declares supportedTags and returns an HtmlElementConverter per tag; the converter turns an HTML element into a MarkdownBlock (or inline list), delegating children/inlines via the provided lambdas. Existing GFM extensions ship these: gfm-tables converts <table> (block), and gfm-strikethrough converts <s>/<strike>/<del> (inline). Use them as references and add the extension whose tags you need rather than reimplementing it.

For full details, models, and wiring, read references/HTML-PARSING.md.

Key points and gotchas:

• htmlConverterExtensions are only collected when parseEmbeddedHtml = true; adding one without enabling embedded HTML does nothing.
• HTML that is not converted is preserved as an HtmlBlock (raw HTML string) or inline HTML; the default renderer shows it via the htmlBlock styling, it is not interpreted as native UI. Jewel does not do general inline-HTML rendering.
• HTML carrying layout attributes (e.g., an align attribute) becomes a MarkdownBlock.HtmlBlockWithAttributes wrapping the converted block; the default renderer reads the align attribute and propagates text alignment to children. A custom renderer that overrides this path must preserve that alignment behavior (see Gotchas for custom renderers).
• Custom HTML conversion is parse-side only. If a converter emits a custom block type, you still need a matching renderer extension to display it.

Live Markdown editor/preview

MarkdownMode.EditorPreview optimizes for small, frequent edits (a user typing): the MarkdownProcessor keeps state between calls and re-parses only the block(s) that changed instead of the whole document. That statefulness is why an editor-preview processor is tied to one evolving document and must not be shared across unrelated documents (doing so busts the cache and is slower than Standalone). Use MarkdownMode.Standalone (the default, stateless) for content that is parsed once and doesn't change.

For a side-by-side editor and preview:

1. Keep raw text/editing state in the presenter or state holder, not hidden inside the preview leaf.
2. Create one MarkdownProcessor(markdownMode = MarkdownMode.EditorPreview(scrollingSynchronizer = ...)) per editable document/preview stream. Do not share editor-preview processors across unrelated documents.
3. Process edits off the UI path when practical, debounce/drop stale work if the editor can emit faster than rendering.
4. Render with LazyMarkdown for documents and previews.
5. Add the expected dialect extensions: autolink is processor-only, tables/strikethrough/alerts need processor and renderer pieces, and images need an image renderer extension.
6. If scroll sync matters, use the scrolling synchronizer/rendering support rather than ad-hoc list scrolling. Read references/SCROLL-SYNC.md — note the synchronizer currently supports ScrollState, not LazyListState.

AI chat with Markdown

For chat timelines, do not make Markdown own the chat architecture.

Use a presenter/coordinator to project message state, roles, streaming status, and parsed Markdown blocks. The UI should render a LazyColumn of message rows and use Markdown or LazyMarkdown inside text-bearing rows as appropriate. Keep networking, agent protocol, message mutation, and Markdown parsing policy outside leaf composables. Use stable lazy-list keys and content types for heterogeneous chat rows.

Custom syntax: extension or renderer?

Write a custom extension when the Markdown input language changes, e.g., task list items, custom admonitions, ::file:/path/foo.kt::, or another syntax that must parse into a domain model.

Write a custom renderer when the existing parsed model is right, but the UI should change, e.g., headings need link buttons, code blocks need copy buttons, list markers need different layout, or inline images need product-specific treatment.

Many features need both: an extension to parse syntax into model nodes, and a renderer to display those nodes. Some may need more than two parts. See the examples below.

How to write a custom block extension

Use gfm-alerts and gfm-tables as templates.

1. Define a model type implementing MarkdownBlock.CustomBlock.
2. Add a MarkdownProcessorExtension.
3. If CommonMark needs help recognizing the syntax, expose a CommonMark ParserExtension.
4. Expose a MarkdownBlockProcessorExtension that checks canProcess(CustomBlock) and returns your MarkdownBlock.CustomBlock from processMarkdownBlock.
5. Set allowsMergingWithNextBlock = true only when incremental editor parsing must reparse a following block that can merge into this one, as tables do.
6. Add a MarkdownRendererExtension exposing a MarkdownBlockRendererExtension.
7. In RenderCustomBlock, render native Jewel Compose UI and delegate nested Markdown back to the provided MarkdownBlockRenderer / InlineMarkdownRenderer where possible.
8. Add standalone and bridge styling helpers when the feature has colors, spacing, borders, or icons.
9. If the block will be used in an editor preview and should be a scroll-sync target, opt in: the scroll-sync renderer does not wrap custom blocks automatically. Wrap your content in AutoScrollableBlock in RenderCustomBlock. See references/SCROLL-SYNC.md.
10. Add processor tests and renderer tests or validated sample coverage.

How to write a custom delimited inline extension

Use gfm-strikethrough as the template.

1. Define a model type implementing InlineMarkdown.CustomDelimitedNode.
2. Add a MarkdownProcessorExtension with a CommonMark parser extension if needed.
3. Expose a MarkdownDelimitedInlineProcessorExtension that turns supported CommonMark Delimited nodes into your custom inline node.
4. Add a MarkdownRendererExtension exposing MarkdownDelimitedInlineRendererExtension.
5. Render to an AnnotatedString, recursively delegating nested inline content to the supplied InlineMarkdownRenderer.
6. Add an HTML converter only if embedded HTML should map to the same inline model.

Example: inline icons and rich inline content

MarkdownDelimitedInlineRendererExtension returns only an AnnotatedString. If custom inline syntax must show an icon or other InlineTextContent — for example ::file:/my/path/foo.kt:: with the file icon and name — plan for an extension encompassing parsing plus renderer customization:

1. Parse the syntax into a custom inline node.
2. Provide an inline renderer or block renderer path that appends inline-content placeholders for those nodes.
3. Override the block rendering for text-bearing blocks (Paragraph, headings, table cells, or whichever contexts you support) so the Jewel Text receives the matching inlineContent map.
4. Use Jewel icon APIs (IconKey, AllIconsKeys, or platform file-type icon APIs in bridge/plugin context) inside the InlineTextContent.
5. Document supported contexts rather than pretending every Markdown block can host the custom inline content automatically.

Example: GFM task lists

Do not claim Jewel already supports GFM task lists. They are listed as missing/roadmap in Jewel docs. To support - [ ] / - [x] task list items, create and validate a new extension:

1. Decide whether the model is a custom block/list item wrapper or an extension of list rendering.
2. Parse checked/unchecked state into an explicit model.
3. Render with native Jewel checkbox-like visuals, preserving Markdown semantics and disabled/read-only behavior as appropriate.
4. Validate both processor output and rendering. Include editor-preview incremental parsing if task-list syntax can interact with neighboring list blocks.
5. If you're rendering interactive components such as checkboxes, make sure their status is managed, and that changes are reflected in the underlying Markdown (e.g., checked state for a checkbox should manifest as [ ] or [x] in text).

Gotchas for custom renderers

When overriding or replacing rendering, preserve the behaviors the default renderer relies on. Each item below is something a custom renderer can silently break.

• Text alignment: every text-bearing block in DefaultMarkdownBlockRenderer passes textAlign = LocalTextAlignment.current to its Jewel Text. That composition local is set when an aligned block (e.g., an HTML block with an align attribute) wraps its children, so alignment propagates into nested text. A custom Render* override or a renderer written from scratch that emits Text without forwarding the current text alignment will silently lose center/right alignment. Always apply the provided text alignment in custom text rendering. Note LocalTextAlignment is internal to DefaultMarkdownBlockRenderer; prefer subclassing it (and still forwarding textAlign in your override) rather than reimplementing block rendering and dropping this wiring.
• Content color: text-bearing rendering resolves color via the styling text style falling back to LocalContentColor. Preserve that fallback instead of hardcoding a color.
• Inline content / images: paragraphs that contain images render through an inlineContent map. If you override paragraph or other text-bearing rendering, keep passing the inlineContent map so images and other inline placeholders still appear.
• Enabled/disabled state: honor the enabled flag (links non-interactive, disabled styling) rather than always rendering the enabled appearance.
• Delegation: render nested blocks/inlines through the provided MarkdownBlockRenderer / InlineMarkdownRenderer so styling, extensions, and these behaviors stay consistent.
• Editor-preview scroll sync: scroll sync is implemented entirely by ScrollSyncMarkdownBlockRenderer (a DefaultMarkdownBlockRenderer subclass that wraps blocks in AutoScrollableBlock and reports layout). If you supply your own block renderer for an editor preview, you bypass that subclass and lose scroll sync. Subclass ScrollSyncMarkdownBlockRenderer instead of DefaultMarkdownBlockRenderer (calling super from your overrides), or re-implement the AutoScrollableBlock wiring yourself. See references/SCROLL-SYNC.md.

General rule: when you override one Render* method, mirror the default implementation's wiring (alignment, color, inline content, enabled state, and editor-preview scroll sync) and change only what you intend to change.

Validation checklist

Before finishing Jewel Markdown work:

• Confirm runtime context: standalone uses int-ui standalone styling; IntelliJ plugin/bridge uses IDE LaF bridge styling.
• Confirm required processor and renderer extensions are both wired.
• Confirm MarkdownProcessor editor-preview instances are not shared across unrelated documents.
• Confirm heavy parsing is not happening in a hot LazyColumn item body.
• Confirm code highlighting, image loading, and URL click handling are provided where the UX requires them.
• Confirm custom syntax has tests for parsing and rendering; for new extensions, use existing extension tests as guides.
• Confirm task lists or other missing GFM features were implemented/validated, not merely mentioned as supported.
• Confirm custom text rendering forwards LocalTextAlignment (text alignment), preserves the content-color fallback, keeps the inlineContent map, and honors enabled.
• Confirm HTML expectations: embedded HTML needs parseEmbeddedHtml = true; custom tag mapping needs a MarkdownHtmlConverterExtension; and unconverted/inline HTML is shown as raw text, not interpreted as native UI.