← Syamjith NK

8 September 2026 · Syamjith NK

3,168 files use the broken Arabic recipe. Most of them are fine.

Searching GitHub for get_display arabic_reshaper.reshape language:python returns 3,168 files. On matplotlib 3.11 and on Pillow built with Raqm, that recipe reverses your text. It would be easy to conclude there are three thousand broken projects out there and start filing.

That conclusion is wrong, and the reason it is wrong is the only interesting thing here.

The recipe is not a bug. Doing it twice is.

arabic_reshaper.reshape() converts Arabic to presentation forms. get_display() puts the string into visual order. Together they do the work a text engine would do, which is exactly right if the thing that draws your text does not do it for you, and exactly wrong if it does.

renderer	shapes and reorders?	the recipe is
matplotlib ≥ 3.11	yes	a bug
matplotlib < 3.11	no	required
Pillow built with Raqm	yes	a bug
Pillow without Raqm	no	required
ReportLab, fpdf	no	required
print() to a terminal	terminal-dependent	not mine to judge

So the search result is not a list of bugs. It is a list of places where somebody has to look at what happens next.

I looked at six

Ranked by stars and recency, then read rather than pattern-matched. Three were genuinely broken:

• A Discord bot, actively maintained, whose helper carries the docstring “Shape Arabic for matplotlib (which doesn’t run bidi itself)”. That was true for a decade. Its pin is matplotlib>=3.10.5 with no upper bound, so a fresh install resolves to 3.11 and every Arabic player name in its charts comes out backwards.
• A second bot, a sibling of the first, carrying the identical helper on the identical line.
• A ComfyUI node built specifically to render Persian and Arabic, drawing through Pillow with Pillow unpinned.
• An Arabic poetry dataset with no requirements.txt at all, so every chart label and title it draws is reversed on a current install.

Three were not:

• A transcription tool that renders through ReportLab, which does no shaping. The recipe is correct there and removing it would break the output.
• An Arabic benchmark from a well-known lab where the helper is defined and never called. The import and the function are left over; the data is written untouched.
• An OCR demo that pre-shapes only for print(). Terminals vary, and this is a forty-line script.

Two of the six matched the search and were fine. One matched, and the matching code does not execute. If I had filed on all six I would have been wrong half the time, and being wrong in public on someone else’s repository is expensive in a way that being quiet is not.

What a useful checker has to do

I put this into arabic-lint 0.2.0, and writing it taught me more than the search did. Every one of these came from testing against the six real files rather than snippets I made up:

• Follow import aliases. Real code writes from bidi.algorithm import get_display as _bidi_get_display. A checker that only knows the canonical spelling passes the file in silence. This missed the first bot entirely.
• Follow the recipe across statements. reshape() on one line and get_display() on the next is the common shape. Looking only inside the get_display() call missed the ComfyUI node.
• Name the renderer that actually draws. A file that imports both Pillow and matplotlib, sorted alphabetically, blames Pillow for a matplotlib bug.
• Require that something draws. Importing Pillow to load an image and then printing the Arabic is not a rendering bug.
• Do not call a helper dead just because nothing in the file calls it. That is true of a script; it is false of a library module, where the caller is in another file. Getting this wrong means going quiet on a real bug, which is the more expensive mistake.

The tool now stays silent on all three of the projects that are fine, and flags all three that are not.

The part worth keeping

A grep gives you 3,168 results and no judgement. The judgement is the work, and most of it is deciding what not to report. A checker that flagged all 3,168 would be switched off within a day, and rightly.

The same is true of the search itself. Three thousand hits is not three thousand bugs; it is a population, and the interesting question is what fraction of it is actually broken. On a sample of six, it was half.