qwert702/dsh-auto-translate

English | 中文

Full-pipeline Chinese translation for the DeepSeek Harness Web GUI.

Compatibility — this release (0.3.0) targets harness 0.2.0-rc.2 (client contract 0.2.0). For harness 0.1.7-rc.2 and earlier use version 0.2.1.

When the model is thinking (the Think disclosure), writing its final reply, or producing tool output, English content gets a Chinese rendering in place — translated live while the model is still streaming. Every translation is a standalone provider request that never enters the session transcript: no context window cost, no conversation-history pollution.

Install (one command):

dsh plugin add qwert702/dsh-auto-translate

Restart the harness (dsh web) and refresh the page.

Features

• Live reasoning translation — inside each Think disclosure row, the expanded body shows the original thinking text followed by a Chinese translation; while the model is still reasoning the translation updates on a debounced cadence, and once the step settles the full text is translated once and cached.
• Inline reply translation — every English text block gets a small Chinese line (with a 译 badge) directly beneath it. Chinese replies and code-dominated output are skipped heuristically.
• Tool-output summaries — for every settled tool call, a line tool — 中文含义 plus a 译 summary renders at the turn tail, summarizing the tool's text output in Chinese (a summary, not a line-by-line translation of raw output). The summary lives at the turn tail instead of inside the card because the chat view has no additive seat inside a tool card, and the tool-call renderer cannot be shadowed without losing the dedicated tool views (bash / web / read …) — see below.
• Zero context cost — translation and summary requests are standalone provider /chat/completions calls made by the node half, fully outside the agent loop; the plugin adds no prompts, tools, or messages to any conversation and never rewrites a model request.
• Configurable model — defaults to the current conversation's model; an override can be set in the profile entry config. When the conversation model id is unknown to the provider, the host retries once with deepseek-chat.

How it works (shadow rendering, no slot replacement)

The chat view has no additive seat inside a message, so the plugin uses the ecosystem's shadow registration technique: it registers the assistant-step renderer on conversation.chat.node at priority: -100, which shadows the harness's priority: 0 default and owns the in-place reasoning and reply translations. In 0.2.0 that seat belongs to @deepseek-ai/dsh-client-ui-chat.

The shadow renderer mirrors the 0.2.0 AssistantNodeView / AssistantMarkdown contract:

• It rebuilds the structure from public primitives only — MarkdownText / DisclosureRow / JsonBlock from @deepseek-ai/dsh-client-ui-primitives — and takes the framework-synthesized t seat from the entry's declared locale: 'chat', falling back to Chinese when a key is missing.
• It honours the groupPart split (Turn-process renders the same assistant node twice, once for reasoning and once for response) and the Turn-process fold: while folded, the inline reasoning block is marked hidden="until-found" exactly like the official renderer, so in-page find expands it.
• It uses the real 0.2.0 data shapes: the three-state status: running | settled | interrupted, blocks (text / reasoning / image / tool-call / other) and finalNode.providerMetadata.model.
• Image blocks go through the official message-images seat by calling the renderMessageImages callback the Chat view hands down (that is conversation.message.images, filled by @deepseek-ai/dsh-client-ui-attachment). 0.2.0's ui-attachment no longer exports ImageGallery, so the plugin no longer renders images itself.

The tool-call node is deliberately NOT shadowed: its default renderer owns the sole declaration of the tool.call.toolview child slot (slot declarations are globally unique), so shadowing it either loses the dedicated tool views (bash / web / read …) or throws a duplicate-declaration error at startup. Tool summaries therefore ride the additive conversation.chat.turnTail seat — a list seat in 0.2.0, whose owner props are { turn: TurnLocation, seq, openFile } — and read the Chat snapshot through the useChat standard hook (locations.getTurn + nodes.get), rendering before the turn's action row.

If a shadow renderer crashes — or meets a node shape it does not know (it throws instead of rendering blank) — the renderer's error boundary abdicates the entry and the official priority: 0 renderer takes over transparently; the message is never lost.

Configuration (optional)

Configuration uses the harness 0.2.0 profile-entry config channel (the plugin's Config export; the settings.register() namespace was removed in 0.1.7+). Override the ui-auto-translate entry's config in your profile patch layer ~/.dsh/profiles/<profile>/cordis.patch.yml:

- id: ui-auto-translate
  config:
    enabled: true                # master switch
    model: ''                    # '' = follow the current conversation's model
    apiKeyRef: DEEPSEEK_API_KEY  # credential reference resolved by the harness credentials service
    baseURL: https://api.deepseek.com
    temperature: 0.3             # sampling temperature for translation requests
    maxInputChars: 4000          # input truncation guard

All fields default as above; restart the harness after changing them. The API key stays server-side — it is resolved through the harness credentials service and never leaves the host.

Repo layout

• lib/index.js — node half: the Config export (loader config channel) plus the POST /api/auto-translate/translate route (standalone translate/summarize requests, mode: translate | summarize, with a model fallback); inject = ['webServer', 'credentials'].
• lib/client.js — browser half: the assistant-step shadow registration (priority: -100 with the locale: 'chat' seat, in-place reasoning/reply translations) plus the additive conversation.chat.turnTail entry (tool-output summaries).
• test/smoke.cjs — node test/smoke.cjs: syntax checks; host (Config export, every route path: bad-request / empty / no-key / ok / summarize / model fallback / config overrides / disabled); client (both seat registrations, label fallback, collectTurnTools, and ReactDOMServer renders: reasoning / text / unknown block / groupPart split / Turn-process fold / collapsed reasoning / images / streaming / interrupted / malformed-node abdication / tool summary / missing-useChat degradation).

Known limitations

• Conflicts with dsh-better-markdown — both shadow assistant-step at priority: -100; installing both throws a duplicate-key error at startup. Install one; to coexist, change one side's priority.
• Images depend on the official seat — message images render through the official conversation.message.images seat (ui-attachment). When that plugin is not part of the composition, renderMessageImages is absent and image blocks are skipped (the official view omits images under the same condition); nothing crashes.
• Local-path images in Markdown are not resolved — the official 0.2.0 AssistantMarkdown hands down a pathImages resolver that turns local file paths into displayable URLs; this plugin does not take that resolver over, so images authored as local paths stay inert alt text (absolute http(s) images render normally).
• The "work details" presentation policy is not read — the official reasoning row consults that policy to decide whether a settled summary preview shows; this plugin's shadow always shows the summary line. The difference is limited to that one cosmetic detail.
• Tool summaries live at the turn tail, not inside cards — the chat view has no additive seat inside a tool card and the tool-call renderer cannot be shadowed without losing the dedicated tool views, so summaries render at the turn tail.
• Owns the default rendering — shadowing means the thinking/text block presentation is maintained by this plugin; a dsh upgrade that changes the internal rendering may need an adaptation (a crash abdicates to the harness default, and an unrecognized node shape throws on purpose so abdication always happens).
• Streaming translations are provisional — while output is still streaming, the translation reflects the text seen so far and updates on a debounced cadence; when output stops it is re-translated from the full text.
• Tool summaries cover text blocks only — images and binary output are not translated; long outputs are truncated (maxInputChars) first.
• Heuristics — replies shorter than ~10 letters, CJK/latin mixed text, and code-dominated output are not translated (isEnglishProse / isMostlyCode in lib/client.js).
• Cache — translations are cached in browser memory per node/call identity; a page refresh re-requests the same content.
• No settings UI — configured through the profile entry config; no settings panel yet.

License

MIT