@creait/dsh-to-english

Rewrites a market-installed plugin's Chinese copy into English, using a model you already have configured — once when it lands, then live-reloaded.

Most of the DSH plugin market is written in Chinese. Installing from it gives you a working plugin whose settings page, tool descriptions, log lines and README you cannot read. Translating by hand is a per-plugin chore that has to be redone on every update, and a blunt "replace every CJK run" pass produces something worse: a plugin that reads English and no longer works.

The untouched original is always kept beside the file as .zh.bak.

How it translates

One pass over the whole file is the default. The model is handed the entire file and returns the entire file. What stands behind it is not a diff gate but a parser: node --check for .js/.cjs/.mjs (on a sibling temp file, so the package's own type field decides script-vs-module exactly as it will at load time), JSON.parse for .json, a YAML parse for .yml/.yaml. A file that fails is handed back to the model with the parser's complaint, up to three repair round-trips. A reply shorter than 60% of the original is treated as truncated rather than as a translation — prose has no parser to catch a half-written README.

Files up to 48 KB go through this path. Above that the model would have to echo back more than the completion budget allows.

The segment path is the fallback. For a file too large to echo back, or one the model could never get past its parser, the older pipeline still runs: only line ranges containing Chinese are sent, the reply must have the same number of lines, and each line is checked against the original's code skeleton before it is accepted. A conservative partial translation beats none at all. Nothing above 512 KB is touched at all.

Why the default moved: the skeleton comparison is safe, and it is also why Chinese that has to change shape to become English never translated. /^[好嗯啊]*\s*继续/ has to become /^(ok|um|ah)*\s*continue/ — a character class turning into an alternation, because Chinese marks optionality per character and English words are longer than one character. Every such line came back rejected.

What that trades away is real and worth stating: node --check catches a file that stopped parsing, not one that still parses and now means something else. A renamed identifier, a reworded id: in cordis.patch.yml, a translated "main" in package.json all survive their parser. The contract asks the model to leave those alone; nothing enforces it. Read the diff against .zh.bak.

What it will not translate

translateEverything (default true) decides whether Chinese the program matches on is rewritten or preserved. On an English-only harness, preserving it just leaves a feature switched off — the phrases that would fire it can never be typed — so the default is to translate it, and to change its shape where English demands a different one. Turn it off and the segment path locks three shapes:

Match data. dsh-recall's auto-recall gate is ~324 characters of Chinese trigger phrases and Chinese regexes. A line holding a regex literal with CJK, or a line that is only quoted literals and commas, is locked.

The non-English half of an i18n table. A var zh = {…} next to a var en = {…} is already bilingual; rewriting zh serves English to the people who deliberately picked Chinese. A brace-depth scan from a head line bound to a non-English language tag locks the block.

Two things are locked regardless of that setting.

Names the code looks up. Variable and function names, object keys, import and route paths, settings namespaces, and the id, name and main fields of package.json and cordis.patch.yml. In the segment path keepsStructure enforces this by comparing the code skeleton, and YAML gets its own gate — a quoted scalar is writable only if the original held CJK, so prompt: "你是助手" translates and name: 'dsh-enhance' does not. In the whole-file path this is a contract clause, not a gate.

Protocol strings. lib/semantic.js's QUERY_PREFIX in dsh-recall is the BGE embedding model's required instruction prefix. It belongs to the checkpoint, not to the user: a model trained on Chinese instructions expects its Chinese instruction whatever language the query is in. No deterministic rule catches this — it is a prompt clause in both paths. Review the diff.

README.zh.md and its siblings are skipped outright — the English README is the file next to it, and translating it spends the budget producing a second English README under a name that says it is Chinese. A README.md that is itself the localized half (a Chinese README.md beside an English README_EN.md) is skipped the same way.

The one structural edit the segment path allows

"请求 " + <b>{n}</b> + " 次" puts the measure word after the number and English puts it in front, so the trailing literal has nothing left to hold and must go. dropsOnlyAffixes permits that deletion under narrow terms — only a short CJK literal, at most two per line, exactly one separator removed each, and only when another Chinese literal on the same line survives, so t("中文标题", fallback) cannot quietly lose an argument. Every such acceptance is logged and listed under relaxed in the file report.

Install

Not on npm yet, and pnpm cannot install a subdirectory of a git repo — so clone the repo and point the profile at the folder:

git clone https://github.com/CREAIT-nl/dsh-plugins.git
dsh plugin --profile web add ./dsh-plugins/to-english

That records a link: rather than copying the package in, which is what you want here: pnpm does not install a link target's dependencies into the linking project, so the profile's node_modules stays free of a second copy of any @deepseek-ai/* runtime. Run pnpm install inside to-english/ once, then restart dsh web — the boot manifest is built at startup.

The package ships its own cordis.patch.yml, so it inserts its roster row on its own — no manual profile edit. Add it to dsh.profile.bundles to activate the browser half.

Configure

Settings live in the dsh-to-english namespace, or in the GUI at Settings → To English.

| Key | Default | What it does | |---|---|---| | enabled | true | Run automatically when a plugin is installed. | | provider / model | '' | Which configured model translates. Empty means the first available. | | prompt | style guidance | Editable. Style only — the mechanical contract lives in WHOLE_FILE_RULES and PROTOCOL_SYSTEM, where an edited prompt cannot weaken it. | | rewriteRadius | 1 | Segment path only. Lines without Chinese on each side of a Chinese line that are also open to rewriting. 0 leaves a mixed passage reading like a graft; 1 lets the model repair the English clause the Chinese was joined to. Capped at 5. | | translateEverything | true | Translate the match data and locale tables described above rather than preserving them. The gates that stop the model renaming code are unaffected. |

The automatic run fires on a new top-level package directory appearing in the profile's node_modules. Anything installed before this plugin existed never crossed that trigger — use the settings section's Translate now box, which takes a package name and runs the same pipeline.

Routes

Loopback-only, mirroring the gen-limit settings-route pattern — the harness settings wire only exposes namespaces on its own allowlist, which a plugin cannot widen:

| Route | Purpose | |---|---| | GET/POST /api/dsh-to-english/config | read/write the settings above | | GET /api/dsh-to-english/catalog | live provider/model list | | POST /api/dsh-to-english/translate | translate + reload one installed plugin now | | GET /api/dsh-to-english/status | enabled, provider, model, last run |

What breaks this

The watcher, the settings provider and the live-reload seam are pre-1.0 internal dsh surfaces with no compatibility guarantee. peerDependencies pins the versions this was built against; a harness upgrade can move them.

Translation is a model call, so it is not deterministic. In the segment path a line that fails its check is left in Chinese and counted in the report. In the whole-file path a file that fails its parser three times is abandoned and falls back to segments. Neither can make a mediocre translation good, and neither catches a change that parses. Read the report, and for anything carrying a protocol string, read the diff.

The settings nav glyph is a deliberate reach past the API. settings.section has no icon option — the shell picks the glyph from a hardcoded section-id map and falls back to the gear for ids it does not know, ours included. So the client half repaints its own row: it finds the nav cell by label and replaces the gear's markup with the official IconListPenOutline16. It fails safe — if the shell's markup moves, nothing matches and the row keeps the gear.