Translation Assistant

A DeepSeek Harness plugin that turns translation into an AI workflow: the model drafts, you direct. Text, files, screenshots, and whole folders are translated by AI; your glossary and style rules ride along on every request; every score and comment you leave becomes the instruction for an AI rewrite; and each pass keeps a rollback point. The job shifts from typing translations to judging them.

中文

What the AI workflow changes

Before	With Translation Assistant
Type every segment; look terms up as you go	AI drafts the full text in one pass — a .translation-spec/ glossary and style rules are attached to every request automatically
Mark problems, then retype the fixes	Score a segment and write one comment; Apply suggestion turns both into the rewrite instruction for that segment
Copy-paste files into a translator tab, one at a time	Point the workspace at a folder: batch translation walks every pending file and writes the results back to disk while you review what has landed
One model has to do everything	One model translates, another reviews: the draft comes from a fast model, and a stronger one scores it and rewrites from its own comments
Switch windows to translate a snippet from chat	The same engine is an Agent tool — the assistant translates and reviews inside the conversation

Highlights

• AI at every step — drafting, retranslating with your prompt, scoring, rewriting from your comments, and reading text out of screenshots are all model calls; the plugin orchestrates them into one loop.
• Review instead of retyping — the side-by-side view shows the draft; you touch only the segments that need it, and AI applies your notes to the rest.
• A two-model loop — one model produces the draft, another scores it, and every score becomes a rewrite instruction; each pass leaves a revision you can roll back to.
• Guidelines travel with every request — the workspace's .translation-spec/ folder holds glossary and style documents; translation, retranslation, scoring, and batch jobs all include them.
• Not just a page, an Agent tool — the engine answers translation_assistant calls (translate | retranslate | score | score_document) inside chat.
• Local first — history and edits never leave the browser; bring your own models.

Install

One line from a terminal:

dsh plugin --profile web add dsh-translation-assistant

Or from the GUI: Plugins → Add plugin, enter dsh-translation-assistant, press Install, then Enable now.

Local checkout, Git address, manual registry, and upgrading or uninstalling: docs/install.md.

What you get

Translate

• By paragraph or By sentence, both returned in one pass; speed-first mode runs chunks concurrently and shows a Translating n/m counter.

• Three AI translation modes, picked on the new-translation form and in the full-text card. When a response does not match the chosen protocol, the pass falls back to batching without raising an error.

  Mode	Trait	Best for
  Coherence first	One request for the full text, the fullest context	Essays, reports — the default
  Alignment first	JSON-array protocol, the strongest paragraph alignment	Glossaries, subtitles, strict side-by-side layouts
  Speed first	Chunked three-way concurrency, fastest on long texts	Time-boxed bulk jobs

• Files — drop or pick .txt, .md, .log, .html, .htm, .docx (Word files are unpacked in the browser; the original never leaves the machine).

• Images — paste or pick a screenshot; AI transcribes it, then translates.

Workspace

• Workspaces are the projects you already added in Harness: a picker at the top of the drawer lists every project folder, selection switches instantly, and the last one is remembered. Reads and writes stay confined to the selected project; the drawer never covers what you are proofreading.
• When the host does not expose the project list (e.g. dsh web), a collapsible manual path entry serves as the fallback; an outdated Host side (404) points to a Harness restart.
• The tree folds by folder; clicking a file name opens it — untranslated files start translating immediately, translated ones reopen for proofreading.
• Output goes to a translated/ subfolder (default), a suffixed sibling copy, over the original, or stays on the page only without touching disk; .docx translations always save as .md. Proofread edits can auto-save or wait for the save button.
• Translate all walks the pending files in order and saves each result; a stop ends the run after the file in flight. While workspace content is open, the workspace button shows the project's name.
• A .translation-spec/ folder in the workspace holds guideline documents (glossary.md, style.md, …) that ride along on every AI pass; the drawer can view, edit, create, and delete them, and the banner shows which guidelines are active.

Review and refine

• Source and translation sit side by side; every segment is editable, and edits re-join sentences and paragraphs correctly.
• AI retranslate redoes one segment; Retranslate with prompt adds an instruction such as "more colloquial". Both follow the translation mode and leave one revision per pass.
• Scores are on demand only: score a segment, the current view, or the whole document, then reveal them from the toolbar. Apply suggestion sends a score and its comment back as the instruction for a fresh AI pass; Apply all rewrites every scored segment three at a time.
• A Stop button sits beside every running pass; in-flight requests are cancelled, and rewrites that already landed are kept.

Keep track

• On-device history — up to 200 documents with search, pinning, and two-click deletion.
• Revisions — up to 100 snapshots per document; the drawer shows how far each one differs from the current text, and restoring is undoable.

Integrate

• The header picker groups models by provider; Follow default tracks the profile, any other entry pins the route for translation, retranslation, scoring, and OCR — so a cheap model can draft while a stronger one reviews. Outside providers register through Harness settings — a worked example in docs/bilibili-index-translate.md.
• The same engine is exposed as a translation_assistant Agent tool (translate | retranslate | score | score_document), so the chat assistant can translate or review text as part of its own workflow.

Use cases

One model drafts, another reviews. Translate with a fast model, then switch to your strongest model and run Score all, then Apply all. Each rewrite carries the score and its comment as the instruction, so the strong model fixes exactly what it criticised — and every pass leaves a revision, so comparing and rolling back is one click.

A whole folder under your rules. Drop a glossary.md and a style.md into .translation-spec/, point the workspace at the project, and press Translate all: every file is translated with the same guidelines attached and written back to disk, and you review the results side by side afterwards.

A whole document, end to end. Drop a .docx, .md, or .html file: it is translated in one request, and Download translation writes the result back in the same shape with the paragraph structure intact.

How it works

host/index.js registers a same-origin /dsh-translation/api route and the Agent tool, host/engine.js does the model work (no Harness imports), and client.js is the page. POSTs are origin-checked and capped at 24 MiB, and every request can pin a route: { provider, model }.

Endpoints, configuration fields, on-device storage, the full limits, and maintainer notes: docs/reference.md.

Privacy

History, edits, and preferences never leave the browser; only the text being translated, retranslated, scored, or read from an image is sent to the model route. Clearing site data for the Harness origin erases the history — there is no server-side copy.

Development and verification

node --test 'test/*.test.mjs'

The repository doubles as the installed bundle: client.js hot-swaps into the open page, Host changes need a Harness restart. The screenshots in docs/screenshots/ are real renders against mocked data — retake them after layout changes.

Troubleshooting

See docs/troubleshooting.md.