DeepSeek User Message Markdown Renderer

A userscript that renders your own messages on DeepSeek web with the same native-style Markdown, LaTeX math, and code blocks as the assistant's replies — without breaking editing, re-rendering, or the history-item highlight.

Tested against the current DeepSeek web build. The script relies on some hashed CSS class names (e.g. _9663006) that DeepSeek occasionally changes; a small header/class update may be needed after their big UI releases.

Features

• Native-style Markdown: headings, paragraphs, lists, inline code, links, and blockquotes rendered like DeepSeek's own markdown; a single newline is a soft line break (GFM style), so multi-line input keeps its line breaks.
• LaTeX math via KaTeX: $...$, $$...$$, \(...\), \[...\].
• Code blocks rebuilt into DeepSeek's official md-code-block structure: banner with the language label, native light/dark theme, corner decorations, and Prism-style token colors from the page's own stylesheet.
• Hard line breaks preserved in code blocks; unknown languages (e.g. mermaid) stay as clean code blocks without console warnings.
• Safe editing: clicking "edit" restores the original message before DeepSeek reads it, so the editor never crashes; cancel re-renders the message; empty edit placeholders are cleaned up.
• In-place rendering: the original message text element (hashed class fbb737a4; _8271fc3 only marks messages with an attachment) is transformed into DeepSeek's native ds-markdown structure (paragraphs carry ds-markdown-paragraph), so no extra bubble is created and nothing is hidden — attachment cards and the native bubble layout stay intact.
• Native history highlight: because the original bubble is never replaced, DeepSeek's history-item highlight works as-is without any mirroring.
• Never removes or hides DeepSeek's original nodes — the text element is rendered in place, so the references held by the host app (React) stay valid and re-rendering never throws NotFoundError.

Install

1. Install Tampermonkey.

2. Open the raw script below — Tampermonkey will offer to install it:

   https://raw.githubusercontent.com/NIyueeE/deepseek-user-message-renderer/master/src/deepseek-user-message-renderer.user.js

   Or copy the contents of src/deepseek-user-message-renderer.user.js into a new Tampermonkey script manually.

3. Open https://chat.deepseek.com. The script loads marked / highlight.js / KaTeX from CDNs via @require.

The script auto-updates from GitHub when @updateURL / @downloadURL are enabled in Tampermonkey.

Development

bun install       # install dependencies
bun test          # run tests in a simulated Tampermonkey + browser environment
bun run lint      # Biome static checks
bun run lint:fix  # auto-fix formatting and lint issues

Test structure

• test/env.ts: simulates the browser DOM / MutationObserver / events with happy-dom, stubs the Tampermonkey APIs GM_addStyle and GM_getResourceText, and exposes the same marked version as the production script. Each test file runs in an isolated process.
• test/render.test.ts: Markdown, native md-code-block structure, hard line breaks, style classes, resource injection, dark mode, and keeping original nodes intact.
• test/security.test.ts: dangerous HTML is escaped — blocked tags (iframe / base / meta / form / style / ...), event handlers, dangerous URL schemes (including character-reference smuggling), and unsafe Markdown links/images; safe tags and harmless attribute values are kept.
• test/edit-restore.test.ts: restoring the message box on edit click, re-rendering after submit, skipping rendering in edit state.
• test/marked-quirk.test.ts: regression guard for marked 18 (a code fence directly after a paragraph whose content is --- parses as a code block; marked <=12 misparsed it as a setext heading).

CI / Release

• .github/workflows/ci.yml runs lint + tests on every push and pull request.
• .github/workflows/release.yml builds the script and creates a GitHub Release for every v* tag. It also publishes to GreasyFork when the GFU / GFP / GREASYFORK_TOTP_SECRET secrets are configured (GreasyFork has no official API, so the workflow signs in with these credentials and imports the script from the raw GitHub URL).
• OpenUserJS has no publishing API for regular users (its /api endpoints are admin-only), so publishing there is manual: upload the script at https://openuserjs.org/user/add/scripts, or log in with GitHub and import it from this repository. The script metadata includes @license MIT, which OpenUserJS requires.

License

Released under the MIT License.