A suite of Node.js CLI tools for design system extraction, rendering audits, attention/hierarchy analysis, and site archival. Captures full design tokens, renders pages, measures visual hierarchy, and creates self-contained offline archives.
UIForge defines 7 tools with complete input schemas and descriptions. Naming follows verb_noun patterns (create-uiforge, uiforge-archive, uiforge-capture). Descriptions are detailed (100-300 chars) and explain purpose and use cases. However, output schemas are not documented, LLMs cannot predict return structure for chaining. Parameter descriptions are present but lack format constraints (e.g., viewport 'WxH' format not formalized as regex/pattern). Error handling is absent from tool definitions. No tool annotations (readOnlyHint/destructiveHint) despite clear risk levels (WRITE vs READ_ONLY). Composition is sound, each tool has one responsibility. Security: tools accept file paths and URLs without visible sanitization guidance.
Wire a project so slop can't land: drop in a token kit, the linter, an npm script, a pre-commit hook, and a CI workflow. Zero dependencies (Node standard library only). Does NOT scaffold the app; wires UIForge into an existing project.
Record a site's code + network, replay it so the REAL behavior runs offline. The COMPLETE behavior clone: not a re-render of one state but a record of the site's real code + all its network, replayed so the ORIGINAL JavaScript runs against cached responses. Creates a self-contained folder + zero-dependency static replay server.
Download a capture's external assets so the clone is self-contained. Stages 7 of the clone pipeline: make a clone SELF-CONTAINED by downloading every referenced asset (images, fonts, videos) into a local directory and providing rewrite helpers.
Predict where the eye LANDS, and check the hierarchy. Predicts a gaze order from a saliency proxy (size · contrast · position · accent · weight) on the rendered page, then checks: is there ONE clear focal point? does the eye reach the headline first? is the primary action buried?
Stage 1 of the clone pipeline. Render a reference and extract its FULL design: every reproduction-relevant computed style per element, geometry, text, and assets, plus a deduped token set (design system's palette, type scale, spacing, radii, shadows, and fonts).
Output schemas not documented. LLMs cannot predict return structure (fields, types, pagination) for downstream tool chaining or result extraction.
Parameter format constraints not formalized. 'viewport' accepts 'WxH' but no regex/pattern declared; 'command' enum values (stats|search|show|near) documented in description only, not as JSON Schema enum.
No tool annotations (readOnlyHint/destructiveHint/idempotentHint). Tools marked WRITE (create-uiforge, uiforge-archive, uiforge-assets) should declare destructiveHint=true; READ_ONLY tools should declare readOnlyHint=true.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 52 | 2026-07-28+ | v2 |
UIForge catalog query CLI over tools/catalog/catalog.db. Commands: stats, search, show, near (M2 preview signature distance ranking). Requires Node >= 22.5 for built-in node:sqlite.
Empirical validation. Run the render-audit (+ attention) over a LABELED corpus (designed vs template) and report whether the grades actually separate the classes. Honest by construction: external sites can 429 / cookie-wall / time out — those are recorded as failed, not silently dropped.
No error handling guidance in tool definitions. Tools that call external services (uiforge-archive, uiforge-capture, uiforge-corpus) should document timeout behavior, retry guidance, and recovery steps.
Path traversal and URL injection not addressed. Tools accept 'targetDir', 'capturePath', 'url', 'profile' as free-form strings without visible sanitization or validation constraints.