docs: use Fumadocs TypeTable on data model pages

[marcklingen]

Simplified rebase of #3767 onto current main.

Uses Fumadocs TypeTable on two data-model pages:

• /docs/evaluation/scores/data-model — Score and ScoreConfig fields
• /docs/evaluation/experiments/data-model — Dataset, DatasetItem, DatasetItemMediaReference, DatasetRun, DatasetRunItem

Path/use-case tables stay markdown. Field names, types, required flags, and documented defaults match the previous tables.

What changed vs #3767

#3767 still includes the already-merged Fumadocs 16.15.4 upgrade, so the GitHub diff is 18 files and mixes upgrade fallout with TypeTable work.

This branch:

• Replays only the TypeTable commits onto current main (7 files)
• Replaces the ~150-line hand-rolled TypeTable field lexer with new Function evaluation of the authored object literal
• Keeps the flat table chrome, MDX registration, and markdown/PDF export

TypeTable is registered in mdx-components.tsx. lib/markdown-component-renderers.js still exports each TypeTable as a markdown field table so .md / PDF / md-src keep the same attributes.

Styling is the flat table (no radius, 1px structure border, no card shadow). AGENTS.md recommends TypeTable for object/field data models.

  

Confidence Score: 4/5

The PR appears safe to merge, with a non-blocking recommendation to add regression coverage for TypeTable Markdown exports.

Summary

• Adds and globally registers a Langfuse-styled TypeTable wrapper.
• Converts the score and experiment object schemas to TypeTable definitions.
• Adds Markdown serialization for TypeTable props.
• Adds flat table styling and contributor guidance for future data-model pages.
• The export renderer would benefit from focused regression coverage because it is responsible for preserving component-only documentation outside the rendered site.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A[Data-model MDX] --> B[Registered TypeTable wrapper]
  B --> C[Fumadocs web rendering]
  A --> D[TypeTable Markdown renderer]
  D --> E[Plain Markdown / md-src]
  E --> F[PDF output]

Loading

_{Reviews (1) · Last reviewed commit: "refactor: evaluate TypeTable literals in..."}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
langfuse-docs	Ready	Preview	Sep 12, 2026 7:14pm UTC

[github-actions]

@claude review

[claude]

Claude Code Review

No review was started: this request came from a bot account. Manual reviews can only be requested by someone with write access to this repository. Ask a maintainer to comment @claude review, or have your automation post the comment from a user account with write access.

_{Tip: disable this comment in your organization's Code Review settings.}

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-12T19:18:04.544373Z	5f9ad6c	Draft marked ready

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" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[claude]

I reviewed this PR and didn't find any bugs. Because it introduces a new build-time JSX→Markdown export path (renderTypeTable in lib/markdown-component-renderers.js, using new Function to evaluate the TypeTable type literal) that two live docs pages now depend on, a quick human look would still be worthwhile.

What was reviewed:

• renderTypeTable/extractJsxExpression in lib/markdown-component-renderers.js — scoped to plain JS object literals authored in-repo, not external input; throws clearly on non-literal expressions.
• Pipe-escaping in code-span cells: the media field's object | null type (content/docs/evaluation/experiments/data-model.mdx:167) is escaped to \| and wrapped in backticks — CommonMark doesn't process backslash escapes inside code spans, so the exported .md/PDF table may show a literal backslash there; cosmetic only, table structure isn't broken.
• Heading structure in both migrated .mdx pages — single H1, proper ##/###/#### nesting, explicit [#anchor] markers preserved.
• src/overrides.css additions — scoped to the new .lf-type-table class, using existing CSS variables, no changes to other selectors.

Extended reasoning...

Overview

The PR adds a Fumadocs TypeTable convention (components/docs/type-table.tsx), registers it in mdx-components.tsx, migrates the Score/ScoreConfig and Dataset/DatasetItem/DatasetRun/DatasetRunItem data-model pages to it, adds matching CSS in src/overrides.css, and — critically for parity — adds a Markdown-export renderer (replaceTypeTablesWithMarkdown/renderTypeTable/extractJsxExpression) in lib/markdown-component-renderers.js so .md/PDF/md-src consumers still see the same attribute tables. A small shim file (.agents/AGENTS.md) is also synced to match root guidance already documenting this exact convention.

Security risks

Low. new Function is used to evaluate the type prop's object-literal source, but that source is always repository-authored MDX content compiled at build time, not user- or request-supplied input, so this isn't an injection vector in the traditional sense — it's equivalent in risk to existing hand-rolled JSX parsers already in this file (e.g. for CardGroup/Academy components). No auth, secrets, or data-exposure paths are touched.

Level of scrutiny

Medium. The page content and CSS changes are simple and mechanical, but the new renderTypeTable markdown-export logic is genuinely new parsing/rendering code with edge cases (non-literal expressions, JSX in descriptions, table-cell escaping) that a multi-round automated review already exercised without finding a bug it flagged. My own reading turned up one concrete, low-severity cosmetic case: the media field's "object | null" type value gets escaped to \| and wrapped in backticks, and per CommonMark, backslash escapes don't apply inside code spans — so the rendered .md/PDF output may literally show a backslash rather than a clean pipe. It doesn't corrupt table structure and is far from severe, but it's a concrete, currently-present instance in this exact diff rather than a purely hypothetical future risk, so I think it's worth a human glancing at the exported Markdown for that one page before merging.

Other factors

The overall design directly implements a convention the project's own AGENTS.md/CLAUDE.md already calls for (TypeTable for data models, with Score/Dataset pages named as the intended examples), and the CSS/heading/anchor conventions are all followed correctly. Given the change is well-scoped and no functional bug was found, but it does introduce new export-path code with at least one minor, real rendering nuance, deferring with a specific note feels more honest than an unqualified approval.