dsh-client-ui-rtl

Right-to-left text direction for the DeepSeek Harness Web client.

The client renders left-to-right by default. Arabic, Hebrew, Persian, Urdu and other right-to-left content therefore arrives correct as text but wrong on screen: bullets sit on the left, table columns run the wrong way, and any sentence mixing Latin and RTL words is reordered against its meaning.

This package marks right-to-left content as it renders, and leaves everything else alone.

Before and after

An Arabic answer mixing English technical terms:

Before	After

A short prompt opening with a Latin product name — the case dir="auto" gets wrong:

Before	After

Install

dsh plugin --profile web add github:haythamat/dsh-client-ui-rtl

Restart dsh web afterwards. The package ships its own patch layer, so no configuration file needs editing.

Why not dir="auto"

dir="auto" resolves direction from the first strong character in the element. That works for most Hebrew and Arabic text, but it fails on a pattern that is very common in technical writing in both languages: a sentence opening with an English product name. The first strong character is Latin, the whole paragraph resolves left-to-right, and every word after it is misordered.

The estimator contract

Direction is estimated by script dominance. This is a product heuristic, not universal direction detection — it is stated explicitly here so it can be tested, versioned, and disagreed with.

• Unit — one whitespace-delimited token. Punctuation does not split a token, so identifiers, paths and package names count once rather than once per segment.
• Classification — a token holding any right-to-left character is an RTL word; otherwise a token holding any Latin letter is an LTR word. Mixed tokens resolve to RTL, since RTL prose embeds Latin terms far more often than the reverse.
• Neutral — tokens with no strong letters (numbers, punctuation, symbols) count as neither.
• Tie — an equal count resolves to RTL.
• Fallback — a block with no RTL words at all is left untouched, so left-to-right content is never marked.
• Override — an element carrying a dir attribute this package did not set is left as authored. That is the opt-out.

Counting words rather than characters matters because RTL words are short while Latin technical terms are long: اشرح لي ال Agentic AI is eight Arabic characters against nine Latin ones, but three Arabic words against two Latin ones.

Elements are judged on the text they hold directly, not on their descendants, so a wrapper containing many children is not flipped on account of its contents, and an inline code span inside a paragraph does not contribute Latin words. Tables and lists are the exception: they are judged on their whole subtree, because column order and list markers only reorder when the container itself flips.

Known failures

Word dominance has irreducible failure modes. Documented rather than hidden:

• Short RTL clause, long Latin command. شغّل npx @deepseek-ai/dsh web is one Arabic word against three Latin tokens and resolves left-to-right, which is wrong. No word-count rule fixes this; it needs authored direction or inline isolation of the command.
• Standalone parenthetical. A block that is mostly a Latin term in parentheses resolves left-to-right even inside RTL prose.
• Evenly balanced blocks resolve to RTL by the tie rule, which is a choice, not a derivation.
• Text split across child elements. Direction is decided per element on the text it holds directly, so a wrapper whose text lives entirely in children is never evaluated — no element sees the whole sentence. A bubble rendering as <div><div>كيف أستخدم</div><div>dsh-client-ui-brand-official</div></div> leaves the identifier child correctly LTR in isolation while the sentence it belongs to is never judged. This is a placement problem, not an estimator problem; inheriting direction at the root with container-level overrides avoids it by construction.

Set dir explicitly on any block the estimator gets wrong; the package will leave it alone.

What is left alone

code, pre, kbd, samp, var, form controls, and embedded media are never marked — their direction is meaningful as authored. No bidi control characters are inserted into content; the package only sets DOM attributes.

Streaming

A MutationObserver watches for added nodes and changed text, batched to one pass per animation frame, so assistant messages are corrected as they stream rather than only on load. Elements are reconciled rather than marked once: a block whose balance shifts while it streams has its direction withdrawn or applied to match.

Attribute mutations are deliberately not observed, since this package writes attributes and watching them would feed its own writes back in.

Model Experience

None, as the package contributes browser presentation only; nothing here reaches a model request.

KV Cache effect

None; this package neither assembles nor sends a provider request.

Known Limitations and Deferred Work

• Presentation only, not layout — the shell itself stays left-to-right. The sidebar remains on the left and controls keep their positions. Mirroring the application shell belongs in a separate package occupying the layout slots.
• The composer is untouched — the input textarea is deliberately excluded to avoid disturbing IME and selection state.
• The browser title is independent — DSH_CLIENT_TITLE selects title text at build time rather than through a UI slot.
• Validated in Arabic only. Hebrew, Persian, Syriac and Urdu are in the character ranges but have not been reviewed by fluent readers of those languages. A DOM attribute cannot prove readability.

License

MIT