Passer au contenu principal
S

dsh-latex

sbvan/dsh-latex

Renders `latex` / `tex` code blocks in the chat as KaTeX formulae, and injects a system-prompt section that tells the model to write `$...$` and `$$...$$` instead of fenced code blocks.

Installer

dsh plugin --profile web add github:sbvan/dsh-latex

README

dsh-latex

English | 中文

Make the DeepSeek Harness chat render formulae as formulae, and make the model write LaTeX by default instead of wrapping math in latex fences.

What it fixes

DSH's frontend already renders TeX through KaTeX (MarkdownText in @deepseek-ai/dsh-client-ui-primitives), but:

SymptomCauseWhat this plugin does
Formulae show up as a grey code blockThe model writes them in ```latex / ```tex fences, which the frontend never renders (only $...$, $$...$$, \(...\), \[...\] and ```math are supported)The host half states the output contract in the system prompt; the browser half renders legacy latex/tex fences in place
During streaming the answer shows raw $ and \fracThe frontend deliberately keeps TeX literal while streaming and swaps in KaTeX once the message settlesLeft alone (upstream design); renders correctly after settling
A few formulae turn into harsh red textKaTeX strict mode: align, \si{}, \ce{} are outside the supported subsetForbidden by the prompt; the browser half turns the red error into a readable badge
Wide display math overflows the bubbleKaTeX display blocks do not scroll by defaultThe browser half adds horizontal scrolling and padding

Install

dsh plugin --profile desktop add github:sbVAN/dsh-latex

Restart DeepSeek Harness afterwards, then reload the page once (the browser half needs to load the new bundle).

# 1. register a link dependency in the profile package.json:
#    "dsh-latex": "link:Z:/文件留存/deepseek project/dsh-latex"
# 2. create the junction (the equivalent of pnpm's link:)
New-Item -ItemType Junction `
  -Path "$env:USERPROFILE\.dsh\profiles\desktop\node_modules\dsh-latex" `
  -Target "Z:\文件留存\deepseek project\dsh-latex"
# 3. append to the profile's cordis.patch.yml:
#    - insert:
#        - id: latex
#          name: dsh-latex
# 4. restart DeepSeek Harness (the host half needs the new composition)

After that the browser half can hot-reload through HMR when hmr.root covers the plugin sources.

Configuration

- insert:
    - id: latex
      name: dsh-latex
      config:
        strict: true
        compact: false
        locale: zh
        extraRules:
          - Write water as $\mathrm{H_2O}$, never \ce{}.
KeyDefaultMeaning
enabledtrueDisables the prompt section entirely (the browser half still runs)
localeautozh / en / auto (auto follows DSH_LOCALE, LANG, defaulting to Chinese)
stricttruetrue forbids fenced formulae; false only recommends
compactfalseShorter prompt, roughly 1/3 of the tokens
order145Section order inside the system prompt (tool guidance owns 100–199)
extraRules[]Extra bullet lines appended to the contract

Behaviour

Host half (lib/index.js)

Registers ctx.systemPrompt.section({ name: 'dsh-latex:math-output', order, text }) with a contract covering inline $...$, display $$...$$, a ban on fenced formulae, the KaTeX-supported subset (aligned over align, pmatrix/bmatrix/cases, \mathrm{} over \si{}/\pu{}/\ce{}, no \label/\tag), CJK only inside \text{}, and splitting long derivations.

The section is disposed with the plugin fiber, so hot reload cannot leave duplicates. The rendered contract is written to $DSH_HOME/cache/dsh-latex/prompt.txt, mount diagnostics to mounted.json.

Browser half (lib/client.js)

  • display-math scrolling and an error badge;
  • pre > code.language-latex|tex gets a preview below it, rendered with the shared MarkdownText (no bundled KaTeX), source folded by default;
  • toolbar with show/hide source and copy LaTeX;
  • 300 ms debounce over a full scan, orphan cleanup, and root unmount when React replaces nodes.

Limitations

  • No math while streaming — that is the frontend's deliberate behaviour and this plugin does not fight it.
  • Only latex / tex fences are handled — math fences are already rendered upstream.
  • No mhchem — \ce{} / \pu{} still fail; the prompt routes them to \mathrm{}.
  • No conversation.chat.node takeover — replacing the assistant renderer means duplicating the whole message structure and breaks on upstream upgrades.
  • No history rewriting — the prompt constrains future output only; old sessions are covered by the browser half.
  • A bare | in a table cell breaks the formula — GFM reads it as a column separator, so $\oint_{|z|=1}$ gets split across cells and renders as raw text. The prompt now requires \vert / \lvert ... \rvert; messages already sent cannot be repaired because the table was split at parse time.

Note: the browser half anchors on .md-code-block and reads the language from the header banner, because the rendered code block carries no language-* class on pre/code. See lib/client.js for the exact DOM contract.

Rollback

Remove the junction, drop the insert line from cordis.patch.yml, restore the profile package.json, restart. No DSH file is modified.

License

MIT

Plugins associés