Static UI review

The remote MCP at https://ui.logic2b.com/mcp provides review_ui. The CLI review command is available in the source candidate; its npm publication is tracked separately. Check your installed CLI’s --help or a different server’s tools/list before using them.

Review reports source evidence and suggested fixes. It never executes source, resolves imports, reads a registry or modifies files. The current version has four rules in two scopes: tokens and a11y. Native HTML is valid; custom components and wrappers are unresolved without their runtime semantics.

Review a local composition

From a built checkout:

pnpm --filter logic2b build
node packages/cli/dist/index.js review src/CustomerPage.tsx --cwd /path/to/app --semantic-colors --json

Select files or source directories explicitly. The collector accepts TSX/JSX, rejects symbolic/hard-linked files and external/private/dependency paths, and does not run project scripts. Select at most 64 files and 256 KiB of total UTF-8 source. Overlapping selections are rejected.

--semantic-colors explicitly enables a project design policy. The rule identifies supported literal color utilities and inline style colors; stylesheets and intentional token overrides are outside this review. Without the flag, L2B-TOK-001 is disabled. Dynamic styling is unknown.

--complete-label-context asserts that external labels, caller ancestors and other composition cannot provide a missing accessible name. Use it only when you have verified that assertion. The default is partial context. Providing several files does not resolve labels across them. Missing local names remain unknown in partial context; even complete context cannot resolve spreads, dynamic names, custom wrappers or ambiguous references.

--scope tokens,a11y selects supported scopes. --fail-on error is the default; --fail-on warning also fails on warnings. Exit 1 means the selected finding threshold was reached. Exit 2 means invalid review input, parse failure or truncated processing. Command-line usage errors retain exit 1. Semantic unknowns alone do not fail the command: exit 0 is not an accessibility certificate.

Supply source through MCP

The host selects and supplies source; a remote MCP cannot read your files. Use project-relative paths and omit unrelated private files.

{
  "schemaVersion": 1,
  "files": [{
    "path": "src/CustomerPage.tsx",
    "content": "export const Page = () => <button aria-label=\"Remove customer\"><svg aria-hidden=\"true\" /></button>",
    "labelContext": "partial"
  }],
  "scope": ["tokens", "a11y"],
  "policy": { "semanticColors": true }
}

The request is the argument object for review_ui. Only version 1 and these fields are supported. Each path is at most 256 characters. The same file/count limits apply in CLI and MCP. Source is parsed as data, with a 50,000-node traversal budget, maximum AST depth 256 and a shared 250,000-step analysis budget. Exhaustion discards partial findings for the affected file, marks review as truncated and leaves explicit unknowns. No import, configuration or component is executed.

Read the evidence

• findings contain a stable rule id, severity, category, confidence, source location, evidence, message, suggested fix and documentation link.
• summary counts unsuppressed errors, warnings and informational findings.
• unknowns identify source or semantics the engine could not establish.
• evaluatedRules lists rules run on at least one parsed file; disabledRules explains exclusions. Neither means every file passed.
• suppressed retains each explicitly excepted finding and its reason.
• assumptions states the evidence boundary. truncated means review is incomplete; select a smaller source set. Each result array is capped at 512 entries, with at most 128 suppression directives per file.

Unknown is not pass. Zero findings can coexist with unresolved components, external labels or dynamic styles. Keep keyboard, focus, contrast, responsive behavior and screen-reader verification in the consuming application. This static result has no quality score or registry version.

Rules and examples

The following examples come from the engine’s shared rule catalog. A token example requires semanticColors: true; a missing-name example requires declared complete label context before it can become a defect. Each current finding has error severity and high confidence within those assumptions.

<div className="text-red-500" />

<div className="text-destructive" />

<dialog open />

<dialog aria-label="Edit customer" open />

<input />

<label>Customer name<input /></label>

<button><svg aria-hidden="true" /></button>

<button aria-label="Remove customer"><svg aria-hidden="true" /></button>

The naming rules cover supported native dialogs, inputs/selects/textareas, buttons and input actions. Static native labels, authored text, aria-label, same-context aria-labelledby and supported native fallbacks can provide the name. Dialog, Input or Button imports are not resolved by name alone.

Record an intentional exception

A known rule can be suppressed on the next source line with a 3–200 character reason. The finding remains available in suppressed.

// logic2b-review-disable-next-line L2B-TOK-001 -- Approved brand swatch example
const swatch = <div className="bg-red-500" />

Scope and independent verification

Icon, motion, state, validation, radius and touch-target rules are deferred. There is no automatic fix or custom component resolution. Apply a suggested change with your normal tools, review it and verify the application.

The agent composition benchmark includes shared review rule ids and evidence beside its existing score. Findings and unknowns do not change that score or replace independently observed builds and customer-journey outcomes. See agent benchmarks.