BidiLens: an MIT toolkit and corpus for mixed RTL/LTR AI output

[CodeinScrubs]

I maintain BidiLens, a new MIT-licensed toolkit for a recurring AI-interface problem: text is logically correct in the model output, but mixed RTL/LTR prose is displayed in the wrong visual order.

A compact example is:

React یک کتابخانه جاوااسکریپت بسیار محبوب است.

React is logically word 1, the Persian sentence follows it, and the period should appear at the visual left edge. Common first-strong solutions (dir="auto" and unicode-bidi: plaintext) see React and choose an LTR paragraph base. A global RTL rule has the opposite regression on English-majority messages, code, paths, and controls.

BidiLens uses a host-adaptable policy:

• resolve each structural prose block independently;
• choose the base from Unicode strong-character majority, with first-strong for ties;
• isolate code, commands, paths, URLs, model names, and identifiers;
• preserve source in logical order;
• keep pure-LTR applications on an identity/no-op path;
• support explicit Auto/LTR/RTL overrides and stable streamed-output decisions.

The v0.1.1 release includes 12 focused packages covering core analysis, DOM, React, Vue, Svelte, Web Components, Markdown, HTML, terminal output, CLI, a shared spec, and Playwright verification. The repository includes 918 schema-validated direction fixtures and documents security, accessibility, performance, migration, and remaining limitations.

This is not a claim that one JavaScript package belongs everywhere. In a Rust terminal host such as the Codex CLI, the corpus and policy may be more useful than the runtime package; UAX #9, grapheme/style mapping, shaping, and terminal capabilities need a native design. In a browser/webview, the DOM/React adapters can be integrated at the render boundary without changing stored conversations or model output.

Current integration evidence:

• NousResearch/hermes-agent#72508 is a focused host-native implementation using Hermes's existing bidi dependency, with direction/tie/no-op regression tests.
• Codex app/extension issue #21563 contains the webview integration outline.
• Codex CLI issue #34871 separates terminal bidi ordering from Arabic shaping rather than recommending string reversal.

I would value review from Codex renderer/i18n maintainers and users of Persian, Arabic, Hebrew, Urdu, Kurdish, and other RTL scripts. I am happy to prepare a focused patch against any public renderer component the team identifies.

Known limits are stated rather than hidden: native-speaker review of the generated technical corpus is still pending; terminal shaping depends on host capabilities; and no renderer can reconstruct source that was already corrupted or stored in visual order.

Maintainer disclosure: I am the BidiLens maintainer. This is an invitation to review and pilot an independent open-source project, not an OpenAI endorsement.