A Claude Code mod that makes Persian, Arabic and Hebrew readable in the terminal transcript.
Most terminals have no UAX #9 bidi and no Arabic shaping, so Persian arrives reversed and with
its letters unjoined. This mod hooks the transcript's render events, runs each line through
fribidi, and draws the result: letters joined, order right-to-left, RTL paragraphs flush right.
Claude Code mods are early access and off by default. This one does nothing at all until you set
CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1on a recent enough Claude Code. See Requirements.
It shapes what Claude Code prints. It cannot fix what you type into the prompt box; see Limits.
1. fribidi
brew install fribidi # macOS
apt install fribidi # Debian/Ubuntu2. A Claude Code new enough to carry the function-hooks runtime.
claude --versionMods are an early-access feature. They are not in the public changelog and not in the official docs, so there is no published "available from" version to point at. What is known: 2.1.260 is the earliest build reported to carry the runtime, and this mod is tested on 2.1.271 through 2.1.273. If you are on something older and the mod does nothing, update before debugging anything else.
Because the feature is early access, the plugin API can change between releases without notice, so a Claude Code update may break this mod until it is rebuilt. That warning is Anthropic's own, from the generated type declarations.
3. Function hooks switched on. The feature is gated behind an environment variable even on a build that has it. Without it the plugin installs, loads and silently does nothing, which is the single most common reason this mod appears not to work.
The durable way is ~/.claude/settings.json, which applies to every session however you start it:
{
"env": {
"CLAUDE_CODE_ENABLE_FUNCTION_HOOKS": "1"
}
}Or export it in your shell (~/.zshrc, ~/.bashrc) if you only ever launch Claude Code from
a terminal:
export CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=14. A monospace font covering the Arabic Presentation Forms block (U+FB50-U+FEFF), which is what this mod emits. Monospace matters: a proportional Persian face forced into a terminal's cell grid pulls the letters of a word apart, where a monospace face has its joined forms drawn to meet at the cell edges.
Vazir Code works (the "Vazir Code Hack" variant pairs it with Hack for the Latin glyphs). Note the project is discontinued, though the released fonts are fine. It has one gap that matters: no U+FEF5-U+FEFC, the eight lam-alef ligature forms (the لا in سلام). Fill just those from Vazirmatn, which you will need installed as well. In Ghostty:
font-family = "JetBrains Mono"
font-family = "Vazir Code Hack"
font-codepoint-map = U+FEF5-U+FEFC=Vazirmatn
Do not add a proportional face like Vazirmatn as a plain font-family fallback. It wins the
Latin glyphs too and spoils your English text.
The mod itself is terminal-agnostic: it asks Claude Code for the terminal surface and nothing
in it knows which terminal you run. What decides whether it helps or hurts is whether your
terminal does its own bidi.
This mod emits text already reordered into visual order and already converted to presentation forms. Those characters still carry a strong RTL bidi class, so a terminal that runs its own UAX #9 pass will reorder them a second time and put you back where you started.
| Terminal | Use this mod? | |
|---|---|---|
| Ghostty | Yes — tested | No bidi shipped; #1442 open |
| kitty | Expected yes | #2109 open since 2019 |
| Alacritty | Expected yes | #663 open since 2017 |
| foot | Expected yes | #756, declined by the maintainer |
| Windows Terminal | Expected yes | #538 open since 2019 |
| VS Code terminal (xterm.js) | Expected yes | vscode#271615 |
| iTerm2 | Only with its own RTL off | 3.6+ has experimental RTL under Settings → General → Experimental; off by default |
| WezTerm | Only with bidi_enabled = false |
That is the default |
| macOS Terminal.app | No | Native bidi via CoreText; would double-reverse |
| GNOME Terminal / VTE | No | Bidi since VTE 0.58 |
| Konsole | No | Bug 403729 resolved fixed |
| mlterm | No | Bidi when built --enable-fribidi, as most packages are |
Only the Ghostty row is tested. Everything else is read off each project's own issue tracker, so treat it as a strong prior rather than a promise. If you try one, a PR correcting the row is welcome.
If your terminal is in the bottom group, you do not need this mod. Its rendering is already better than what this mod can offer, since it works on logical text and can handle the composer too.
claude plugin marketplace add aliir74/claude-code-rtl
claude plugin install rtl-text@claude-code-rtlOr run it straight from a clone, without installing:
CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1 claude --plugin-dir /path/to/claude-code-rtlInstalling writes the enabledPlugins entry itself, at user scope, so it is on in every project.
Later updates:
claude plugin update rtl-textRemember requirement 3: without CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1 the plugin installs
successfully and then does nothing.
| Component | Shaped |
|---|---|
AssistantMessage |
yes, as markdown |
UserMessage |
yes, as markdown, handed back to the engine so it keeps its background band |
CommandOutput |
yes, as plain text |
ToolResult (Bash only) |
yes, as plain text |
The composer is not covered. Typing Persian into the prompt box is still broken, and no mod
can fix it: the input is not a RenderComponent, and ui.input fires only for Input elements a
render hook itself drew, never for Claude Code's own composer. Only your terminal gaining real
bidi fixes that. For Ghostty that is PR #14142,
not merged as of September 2026.
Copying shaped text gives you presentation forms. What is on screen is what gets yanked, so text copied out of the transcript is in visual order and will not paste cleanly back into a logical-order editor.
Two things it deliberately leaves alone. A code block, an inline `code` span and a URL are
passed through untouched, so nothing reorders your commands. A Bash result whose output was too
large and got persisted to a file is left to the engine, because redrawing stdout would present
a truncated view as if it were the whole thing.
Set them in the plugin's config. Every one is optional and clamped in code, so a bad value degrades rather than failing the load.
| Option | Default | Meaning |
|---|---|---|
alignment |
auto |
auto right-aligns only paragraphs whose own base direction is RTL; left shapes without padding; right always right-aligns |
fribidiPath |
fribidi |
Path to the binary |
margin |
4 |
Cells held back from viewport.columns before wrapping. Keep it at 1 or more: it is the slack that stops wrap: 'truncate-end' clipping the start of a right-aligned line if the cell measure is ever off by one |
cacheSize |
256 |
Shaped messages kept in the LRU |
timeoutMs |
2000 |
How long a fribidi call may take before the row falls back to the engine's own drawing |
replyBullet |
⏺ |
Marker on the opening line of a reply. Drawing our own tree replaces the engine's whole row, marker included, so the mod redraws it, on the right edge for RTL where the sentence starts; an empty string leaves it off |
Nothing changes at all. The mod is failing silently by design: every hook falls back to the engine's own row rather than breaking your transcript. Work down this list.
fribidi --version # is the binary there?
grep FUNCTION_HOOKS ~/.claude/settings.json # is the gate set?
claude plugin list # is rtl-text installed and enabled?If you exported the variable in your shell instead of putting it in settings.json, check it with
echo $CLAUDE_CODE_ENABLE_FUNCTION_HOOKS. A settings.json entry will not show up there: it is
set inside the Claude Code process, not in your shell.
If fribidi is installed somewhere unusual, set fribidiPath to its absolute path.
A no runnable fribidi after 3 tries line in the transcript. The mod probes for the binary on
the first render and prints why each candidate failed. The message names the real reason, which is
usually a path problem.
Letters are joined but gappy. That is the font, not the mod. See requirement 4 above: you are almost certainly rendering with a proportional face.
Persian text is reversed. Your terminal probably has its own bidi, and it is undoing the mod's work. Check the terminal support table.
text -> hasRtl? -> split blocks -> split markdown prefix -> protect code/URLs
-> WRAP in logical order -> one fribidi call per direction -> restore -> pad -> Text rows
The ordering is the whole design. Wrapping happens in logical order against the viewport
width, and only the finished lines are shaped. Shaping first and wrapping the visual result puts
the paragraph's opening words on the last line, which is the bug harness/transform-real.check.ts
exists to catch.
Two other decisions worth knowing:
- fribidi never breaks lines for us. Its own
--widthbreaking is not word-aware and splits words mid-token, so every call passes--nobreakand the wrapping is ours. - Padding is ours too. fribidi's
--widthpadding is display-cell accurate, but it cannot express a per-lineautoalignment, socellWidthdoes it. That makescellWidthload-bearing, which is why the harness checks it against fribidi's own padding as an independent oracle.
One fribidi subprocess handles a whole message per base direction, never one per line, and the
result is cached by (markdown, width, text) because render hooks re-run on every redraw and
resize.
.claude-plugin/plugin.json manifest and userConfig
.claude-plugin/marketplace.json
hooks/ the mod; no npm dependencies, no Node imports
register.ts the four ui.render hooks and the engine adapter
transform.ts the pipeline
segment.ts blocks, markdown prefixes, inline protection, tables
wrap.ts cell-width.ts align.ts lru-cache.ts rtl-detect.ts fribidi-args.ts
render-tree.ts lines -> the surface's Box/Text/Code constructors
tests/ run by `claude plugin test`
harness/*.check.ts run by tsx; the only tests that touch the real binary
.claude/types/ generated by /plugin-types, do not hand-edit
docs/research/ measured findings this mod was built from
Four checks, because none covers another:
CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1 claude plugin test . # hooks + pure logic, fribidi mocked
npx --yes tsx --test harness/*.check.ts # the real fribidi binary
CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1 claude plugin validate . # structure: hooks, matchers, $ calls
npx --yes -p typescript@latest tsc -p tsconfig.json # typesThe hooks environment has no fs, network or process, so the real-binary tests cannot live under
claude plugin test. They are named *.check.ts rather than *.test.ts because that runner globs
the whole plugin directory and would refuse a module importing node:child_process.
- Elements are never object literals. They come from
$.ui.resolve(e). A{ type: 'Box' }literal typechecks fine and is then refused at runtime, and the engine quietly redraws its own row, which looks exactly like the mod not being installed. $may only be passed to a function declared at the top of the file.claude plugin validateenforces this so it can report which$calls a plugin makes. That is why the helpers inregister.tsare top-level functions taking aContextrather than closures.- No raw NUL or private-use characters in source. Use
String.fromCharCode. A literal escape can land as a real control byte, which turns the file binary and makesgrepsilently miss it. userConfigentries need atitle. The manifest schema rejects them otherwise.- Regenerate the type contract after a Claude Code update:
claude -p '/plugin-types'. - Bump the version in BOTH
plugin.jsonand themarketplace.jsonentry, in sync. The plugin cache is keyed on that string, so an unbumped release does not reach anyone who has already installed it, and a marketplace entry version overrides the manifest's if they differ.
Working, confirmed in Ghostty 1.3.2 on Claude Code 2.1.273. Persian renders joined, right-to-left and flush right, Latin runs inside a Persian sentence keep their own direction, and code blocks pass through.
MIT