Texuli for agents

Every emittable component is described by an authored manifest: id, props, use-when rules, accessibility notes and examples.

• Build-time validation — invalid manifests fail the catalog build
• Content-addressed catalog file: catalog-<sha256>.json
• Groups discover automatically (display, forms, overlays, layouts, …)

Developer tier may use app chrome. Untrusted model output is limited to Tier U content components and content layouts.

• Untrusted never gets dialogs, menus, shells or focus traps
• isTierU(catalog, id) answers the renderer’s emit check
• agentEmittable defaults to false so missing flags stay closed

Agent trees become safe render nodes: unknown ids become placeholders, schema failures drop nodes, class/style never pass through.

• No HTML sinks — no {@html}, innerHTML or string-built scripts
• Tree limits (depth, nodes, bytes) before render
• Hostile corpus asserts script URLs, bidi tricks and giant trees stay inert

A local stdio MCP server exposes catalog lookup so coding agents can search components without inventing props.

• Read-only tools over the closed catalog
• Runs in-process with the monorepo — no remote catalog fetch
• Pair with llms.txt for prose context and MCP for structured props

texuli add copies styled sources with provenance. llms generation writes agent docs. Updates stay on local paths only.

• Never copies primitives into consumer apps
• Path containment refuses .. escapes and symlink writes
• All file writers share packages/tools internal path rules

A curated set of components and flows every release must keep green: behavior checks, visual baselines, axe and the hostile corpus.

• Playwright behavior checks for every primitive, held to the WAI-ARIA APG patterns
• Visual snapshots on real themes in light and dark
• WCAG 2.x contrast is the contrast gate (APCA out of scope)