@bidilens/core

Framework-independent Unicode direction analysis, inline-isolation planning, streaming state, and bidi security auditing. Runtime analysis is offline and uses generated Unicode 17.0.0 bidi-class and natural-letter data rather than the host runtime's potentially older Unicode property tables.

npm install @bidilens/core

import {
  analyzeText,
  createBidiStream,
  planInlineIsolation,
  sanitizeBidiControls,
  scanBidiSecurity
} from '@bidilens/core';

const source = 'React یک کتابخانه جاوااسکریپت بسیار محبوب است.';
const analysis = analyzeText(source);
// analysis.direction === 'rtl'
// analysis.text === source

const isolations = planInlineIsolation(source, 'rtl');
// React is an LTR identifier isolation.

const security = scanBidiSecurity(source, { mode: 'strict' });
const sanitized = sanitizeBidiControls(source, {
  removeGroups: ['embedding-override', 'deprecated']
});
// Directional marks and complete isolate/PDI pairs are preserved.
// With removeGroups, risk filtering applies to the whole atomic family.

const stream = createBidiStream();
stream.push('old response');
stream.reset(source); // clear and analyze replacement text atomically

needsBidiIntervention(text, context) exposes the shared non-interference gate. LTR-only text in an LTR context returns false; RTL content, bidi formatting controls, an inherited RTL context, or { intervention: 'always' } returns true. Inline planning likewise returns no unnecessary ranges for ordinary LTR-only text.

The default newline paragraph separator is recognized incrementally. A custom paragraphSeparator regular expression is evaluated once by finish() so future-sensitive lookarounds, anchors, and extendable matches remain invariant across arbitrary source chunking. Until then, custom-separated input is exposed as one unresolved open paragraph. Set paragraphBoundary: 'markdown' to recognize blank lines incrementally while retaining a single soft line break inside the current paragraph. If both options are supplied, the explicit Markdown boundary policy takes precedence in core and framework rendering.

The default content-majority policy excludes technical tokens before counting natural-language evidence. Raw-text analysis recognizes closed multiline backtick/tilde fences embedded in surrounding prose; standalone and incomplete fences retain their literal evidence. Parser-aware Markdown integrations remain the preferred path for structured documents. Use first-strong or strict-uax9 only when compatibility with first-strong host behavior is required.

Built-in recognition covers conservative, unambiguous tool and product names. Add private or domain-specific single-token identifiers without changing global state:

analyzeText('internalplatform خوب است.', {
  technicalIdentifiers: ['InternalPlatform']
});

Matching is case-insensitive. Custom values must start with an ASCII letter and may then contain ASCII letters, digits, _, ., or -. firstStrong reports the first strong character after configured technical-token exclusion; rawFirstStrong reports the literal first Unicode bidi-strong character.

The default stream direction remains provisional and revisable until finish(). Choose sticky-majority only when UI stability is more important than correcting a misleading prefix before completion.

Run the packaged example after building with pnpm --filter @bidilens/core example.