Verso/TODO.md

Verso — Next Alpha Roadmap

Ideas and features deferred from the current alpha.

---

Next alpha (post-current)

Typst editing experience (inspired by Collabst)

• typst.ts WASM preview — Run the Typst compiler in the browser via WebAssembly (typst.ts). This would give instant, sub-second preview without a server round-trip, and would eliminate the entire class of race conditions in the CLSI watcher (files written → typst compiles → resolver missed). Could coexist with the CLSI watcher for PDF export while using the WASM path for live preview.

• Tinymist LSP integration — Wire up Tinymist (the Typst language server) behind a WebSocket proxy. Would give Typst files first-class autocomplete, hover docs, go-to-definition, and inline error diagnostics — the main editing comfort gap vs. a native editor.

Editor UX for non-LaTeX formats (.typ, .qmd, .md)

• Visual/rich-text editing mode — A toggle between raw source and a rendered-in-place view for .typ, .qmd, and .md files (similar to Overleaf's rich-text mode for LaTeX). Users who don't know Typst or Markdown syntax should be able to edit content without seeing markup. CodeMirror 6 already supports this pattern via a custom NodeView layer or a separate Prosemirror bridge.

• Toolbar / insertion shortcuts — A formatting toolbar and keyboard shortcuts for common operations, adapted per file type:

  • All formats: bold, italic, underline, headings, bullet/numbered lists, inline code, links.
  • Quarto / Markdown: insert image, insert table, insert code block with language tag.
  • Quarto RevealJS: insert slide divider (---), insert speaker notes (::: notes), insert columns layout, insert video embed (using Quarto's {{< video >}} shortcode).

AI writing assistant

• In-editor AI assistant — Inline writing help similar to what Overleaf and CoCalc offer: suggest completions, rephrase selections, explain LaTeX errors, and generate boilerplate (figures, tables, equations). Should work across all three formats (.tex, .typ, .qmd). Backend would proxy requests to a configurable model API (Claude, OpenAI-compatible) so self-hosters can bring their own key.