RTL extension injects Unicode control characters (U+200F) into Claude Chat messages, corrupting technical text

[Ruth-Nadel-ns]

Description

When using the RTL extension in Claude Chat, invisible Unicode RTL control characters (specifically U+200F / Right-to-Left Mark) are unexpectedly inserted into the message content.

This breaks technical text by injecting control characters into identifiers, arrows, and code-like notation. For example, the following text:
על פני שני BDDs — header \u200Ft, גוף

is transformed into content containing hidden U+200F characters,
The inserted control characters are not visible to the user but are copied with the text and can corrupt prompts, code snippets, identifiers, and technical documentation.

Expected behavior

• RTL rendering should affect only visual presentation.
• The extension should not modify the underlying text buffer or inject Unicode directionality marks into the copied/generated content.

Actual behavior

• U+200F characters are inserted into the text content itself.
• Copying the message preserves these hidden characters, causing formatting and parsing issues in technical content.

Impact
This is particularly problematic for Claude Code and other technical workflows because invisible control characters can break prompts, source code, parsing, searches, and debugging of generated content.

[yechielby]

Thanks for the detailed report! A couple of clarifications and an update:

1. On the source of the U+200F characters — the RTL extension itself does not inject any characters into message content. It only injects CSS (and a small MutationObserver in Auto mode) into Claude Code's webview files; nothing in it modifies the text buffer or writes into message bubbles. The U+200F marks you're seeing come from Claude's model output itself — Claude sometimes emits BiDi control characters (or their visible ‏ escape-code form) to hint direction, and those end up in the text pipeline regardless of whether the RTL extension is installed.

2. v0.5.0 adds a BiDi stripper (Auto mode) — just released. In Auto mode the extension now scans message text nodes as they render (including live-streamed replies) and silently removes the visible ‎, ‏, ‪–‮, and ⁦–⁩ escape sequences that Claude sometimes leaks into responses. Text inside code blocks is left untouched so real technical content isn't affected.

Please update to v0.5.0 (from the VS Code Marketplace / Open VSX) and try again. If you still see BiDi control chars getting copied out with your messages after that — please reopen with a fresh reproduction and I'll take another look.

Also: v0.5.0 bundles a related fix for the reversed BOM detection issue, and a new Force-LTR mode if the auto-flipping direction is what's actually bothering you.