fix(kimi-web): render inline KaTeX math with conservative $...$ delimiters

[nothankyouzzz]

Related Issue

Resolve #1804

Problem

See linked issue. In short: kimi web disables markstream's inline math rule entirely (disableInlineMath in Markdown.vue), so $...$ inline formulas — the form models emit most — render as raw source. Only standalone $$...$$ blocks render.

What changed

Markdown.vue no longer disables the inline math rule; it replaces it (md.inline.ruler.at('math', …)) with a conservative pandoc-style variant (apps/kimi-web/src/lib/mathInline.ts):

• the opening delimiter must be followed by a non-space character;
• a closing $ must be preceded by a non-space and must not be followed by a digit; a candidate closer that fails these rules abandons the opener (no scanning across it), so math content never contains an unescaped $ — spans never cross;
• the content must be non-empty, contain no backtick, and not be a bare number ($100$ stays a currency amount, while $3+4=7$ still renders);
• \(...\) and inline $$...$$ keep working; math_block is untouched;
• while streaming, an unclosed $ still renders progressively — but only when the partial content is unambiguously TeX (details below).

Why this rule set (trade-off analysis)

$ carries no type information in plain text, so any rule here is a heuristic — the question is which errors you accept. Options considered:

Option	Verdict
Keep inline math disabled (status quo)	Real formulas degrade to raw source — the reported bug
Re-enable markstream's built-in rule as-is	Accepts any closed $...$ with no delimiter-adjacency checks: $HOME/bin and $PATH (two env vars in one sentence) becomes "math". Its mid-stream loading math has the same problem: any lone $ with math-ish text after it flickers as a formula
Enable only \(...\) (the proposal in #1804)	Unambiguous, but models overwhelmingly emit $...$; fixes a fraction of real output
pandoc delimiter rules (this PR)	The GitHub / GitLab / VS Code markdown-math standard; covers real $...$ math and the common false-positive shapes
pandoc + content sniffing (require TeX signals like \cmd, ^, _, =)	Kills the residual $HOME/bin:$PATH edge, but also kills legitimate signal-free formulas ($x$, $t$, $D/R$); every exception carved out makes the rule set more special-cased
Try KaTeX render, fall back to text on error	Doesn't discriminate — HOME/bin: is valid TeX and renders "successfully"

The chosen rule set is the pandoc standard plus one extra guard (bare-number content ⇒ currency). It is a sweet spot because the realistic false-positive shapes line up with what the rules already reject:

• Prices ($5, $5 and $10, $1,000.50) — no valid closing delimiter, closing $ followed by a digit, or bare-number content;
• Env vars & paths in prose ($PATH, $HOME/bin and $PATH) — no closing $, or closing $ preceded by a space;
• Env vars & paths in code spans (`$HOME/bin:$PATH` — how models actually write shell content) — never reach the math rule at all: markdown-it's backticks rule consumes the whole span first, and the backtick guard stops a math span from closing across a code span.

Known residual edge: bare-text $HOME/bin:$PATH (no backticks, closing $ adjacent to a non-space) still parses as math — identical behavior to pandoc and GitHub; the escape hatch is \$, as on those platforms.

Streaming: guarded progressive rendering

The built-in rule rendered any unclosed $... as loading math mid-stream; the naive fix would be to wait for the closing $ and accept the raw-text flash. This PR keeps progressive rendering but guards it harder than the closed case, since partial content has no closing delimiter to vouch for it:

• only a backslash command (\Sigma, \frac, …) qualifies as the TeX signal — the one signal with zero overlap with prices and shell syntax (_, ^, =, {} all appear in $MY_VAR, $PATH=$HOME, ${PATH});
• the partial render stops at the first line break, so trailing prose is never swallowed;
• settled messages (__markstreamFinal) never take this path — an unclosed $ in history stays literal;
• when KaTeX can't parse a partial formula yet, MathInlineNode's fallback shows the raw source (its loading spinner is hidden via CSS — the raw text alone is the honest state).

Net effect: $\Sigma_t + \cdots$ renders live as it streams; $5, $PATH, $x never flash as math; simple signal-free formulas like $r_t$ appear the moment they close.

Testing

• apps/kimi-web/src/lib/mathInline.test.ts: 19 unit tests — real formulas, prices, env vars, escaped dollars, unclosed $ (settled vs. streaming), TeX-signal gating, line-break cutoff, silent mode, $$...$$, \(...\), and a regression test for the opener-abandon semantics found in review;
• property-based fuzzing in the same file: 80k seeded adversarial inputs (structured fragments + character noise, settled and streaming) asserting lossless round-trip, delimiter adjacency, no backtick/$ crossing, bare-number rejection, and TeX-signal gating of loading math;
• verified through the real parser (getMarkdown + parseMarkdownToStructure) that inputs like $\Sigma_t$, $\leq t-1$, $w^\top\Sigma w$ produce math_inline nodes while $5, $PATH, $HOME/bin and $PATH stay literal — including inside **bold** spans, which exercise markstream's strong/math token fixups;
• pnpm --filter kimi-web typecheck, pnpm --filter kimi-web test, pnpm --filter kimi-web build, and oxlint all green; manually verified in the running web app.

Checklist

• I have read the CONTRIBUTING document.
• I have linked a related issue, or explained the problem above.
• I have added tests that prove my feature works.
• Ran gen-changesets skill, or this PR needs no changeset. (Changeset included: @moonshot-ai/kimi-code patch — kimi-web is internal, but its build output ships inside the CLI bundle.)
• Ran gen-docs skill, or this PR needs no doc update. (No doc update: bug fix restoring expected rendering; no docs page documents math delimiters.)

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 637a984ee4

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[changeset-bot]

🦋 Changeset detected

Latest commit: d3f3aae

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 1 package

Name	Type
@moonshot-ai/kimi-code	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR