UncleK/dsh-think-translate

🐋 dsh-think-translate

Languages: English · 中文 · 日本語 · 한국어 · Español · Français · Deutsch · Русский

      

Translate the reasoning / thinking chain (chain-of-thought), task cards and answers of the DeepSeek Harness Web UI into one of 8 target languages — in real time, on the display layer only. The originals stay untouched in the transcript, and the translated text never enters the model context.

✨ Why dsh-think-translate

DeepSeek-class models often reason in Chinese — or in whatever language they happen to think in. dsh-think-translate renders the Think row, task cards and answer in *your* language while you watch, like subtitles for the model's thinking.

• 🕵️ Read any thinking chain — reasoning, chain-of-thought, task cards and answers translated in real time, streamed batch by batch
• 🌍 8 languages, one consistent UI — 中文 / English / 日本語 / 한국어 / Español / Français / Deutsch / Русский; the settings panel, thinking rows and task cards all follow your choice, and it persists across reloads
• 🔗 Dynamic provider chain — order providers by drag-and-drop and enable/disable each one. Built-ins (google gtx, bing, local Ollama) plus any number of custom providers; the chain already ends with the free providers, so a second fallback pass is a config-file option, not a UI one
• 🔌 Custom providers (OpenAI & Anthropic) — add arbitrary OpenAI-compatible endpoints (any /v1/chat/completions gateway) or native Anthropic Messages API endpoints (Claude) from the settings panel: name, type, base URL, API key, model
• 🪄 DSH provider discovery — providers already configured in DSH's profiles//cordis.patch.yml / legacy settings.yaml (llm-pi-ai.providers, e.g. linuxdo-hub, coding-hub) are auto-discovered and appear in the chain as read-only "DSH" entries, and one button in settings rescans and adds them all at once; keys are resolved from .credentials.yaml at runtime and never written to the plugin's config. The harness's own default route counts too: when default-model / agent-default-model points at deepseek-official, DeepSeek's API is offered as one more DSH entry
• 🔒 Local Ollama — reuse an existing installation and model. When enabled Ollama is first in priority, or you click Test, the plugin starts the service and loads the model as needed. Missing software/models require the setup button before installation/download, with live progress
• 🧠 Zero context cost — pure display layer: the model still sees the original text, and translated text never consumes the context window
• ☁️ Google / Bing fallback — automatic switch when other providers are unavailable (google goes through a Node CONNECT tunnel using the system proxy, bypassing anti-bot blocks)
• 🛡️ Code-safe — file paths, commands, URLs, regexes and pure-code lines are never translated
• 🚨 Auto-review reasons — when Auto review denies a tool call, the approval card appends the reviewer's reason in your target language, so the decision prompt is readable before choosing Reject / Allow once
• 🧩 Paragraph & sentence-aware chunking — long thinking chains are split on blank lines (paragraph structure preserved) and further batched by sentence, so even a small local model keeps quality
• ⏱️ Resilient — failed translations are cached in memory for 30 seconds to prevent retry storms, then retried automatically; each provider has its own Test button
• 🎚️ Adjustable translation timing — pre-translate everything, lazy-load historical chains (default), or translate only the expanded chain

📦 Installation

# Option 1: npm (recommended)
dsh plugin --profile web add dsh-think-translate
# then restart web

# Option 2: GitHub
dsh plugin --profile web add github:UncleK/dsh-think-translate

# Option 3: manual (junction + patch)
#  1. link the package into the profile's node_modules
New-Item -ItemType Junction -Path "$HOME\.dsh\profiles\node_modules\dsh-think-translate" `
  -Target "<repo path>"
#  2. add to "$HOME\.dsh\profiles\web\cordis.patch.yml":
# - insert:
#     - id: dsh-think-translate
#       name: dsh-think-translate
#  3. restart web

🧯 After a DSH upgrade

Third-party client plugins load through DSH's client module graph, and that graph is composed once per process — a composition that failed is remembered in memory until the process restarts. So these three things commonly bite right after an upgrade:

• Starting from a source checkout fails with client bundles not found; run pnpm run build before launch — the new client packages ship unbuilt: run pnpm run build in the harness checkout, then start dsh web again.
• The plugin's UI is gone (no translated Think row, no *Think Translation* section in Settings) — restart `dsh web`; refreshing the page alone is sometimes not enough.
• The local model list is empty — the ollama service is not serving that model directory: check ollama list (or GET /api/tags) and the OLLAMA_MODELS the running service actually uses. When the model files live on another drive, a directory junction can point the service's default directory at them.

Nothing to configure on the plugin side: it declares no ordering dependency on DSH internals (it binds only to the slots service, plus an optional @deepseek-ai/dsh-client-ui-primitives), so it runs on both older DSH (≤ 0.1.1-rc) and the current line (≥ 0.1.2-alpha.1, 0.1.5-rc.1 included).

For DSH 0.2.0-rc.2, use dsh-think-translate 1.2.5 or later. Version 1.2.5 fixes the todo_write slot collision and declares the 0.2 peer range. Run dsh plugin --profile web add dsh-think-translate@latest, then restart DSH; no version exemption is needed.

Version 1.2.6 fixes duplicated thinking and answer text in DSH 0.2: the assistant renderer now respects the separate reasoning/response groups, including its original-text error fallback.

1.2.7 local setup: enabled Ollama in first priority, or its Test button, finds an existing installation and starts/loads it automatically. Missing software/models require the labelled setup action, with staged download/install/load progress. Automatic software installation supports Windows; other systems link to the official guide. “In use” follows the first enabled, verified provider in priority order; testing a lower row does not override a verified higher row.

1.2.8 transport and recovery: ordinary HTTP/HTTPS translation requests use native Node networking instead of spawning curl per text batch; SOCKS proxies and the Ollama installer retain curl compatibility. Failures stay in memory for 30 seconds, then reasoning and answer translation resume automatically. Request deadlines allow time for local-model loading and provider fallback.

1.2.14: clearer target-language prompts and conservative output checks reject obvious language mismatches and try the next enabled provider. Invalid translations are not kept as successful cache entries; old translation caches are cleared while preferences remain. Code, quoted terms and ambiguous short text are preserved. Fully restart DSH Web or Desktop after upgrading. See #14.

1.2.13: Auto review rejection cards retain the original reason and asynchronously append its translation. An independent, default-on switch controls this feature. Disabling it, changing the target language or removing the card discards stale results; non-English reasons translate correctly to English. Thanks to llg1634 for the report, initial patch and tests in #13. Fully restart DSH Web or Desktop after upgrading.

1.2.12: provider switches are authoritative: the browser no longer sends direct Google requests after Host failure. Fenced, indented and inline code is excluded from translation payloads and restored exactly, including unfinished streaming spans. Old translation caches are discarded; preferences are retained. After upgrading, fully restart DSH Web or Desktop; a page refresh alone may still reference old client resources.

1.2.11 Bing fixes: the free Bing provider is retained. Translation requests use the session page's final origin, preventing the mainland China www.bing.com → cn.bing.com redirect from dropping the POST body. SOCKS/curl now shares the native browser headers to fix HTTP 401; empty and non-JSON responses report clear errors.

1.2.10 DSH profile compatibility: discovery follows the running Desktop/Web profile and reads profiles//cordis.patch.yml (id + config), with home patches and launcher overlays applied afterward. Legacy settings.yaml is still supported. Default model rows use default-model; credentials remain in the DSH home .credentials.yaml and stay in memory. Rescan diagnostics show the file actually read.

1.2.9 language fixes: Japanese vocabulary hints now apply only to Japanese targets; OpenAI/Anthropic requests include a full language name. Bing uses auto-detect and maps Chinese targets to zh-Hans/zh-Hant. Chinese detection examines prose after removing paths, code and JSON: already-Simplified prose is skipped, common Traditional script cues request conversion, and Japanese/Korean prose is translated. Old persisted translation results are cleared once; settings remain intact.

🚀 Usage

1. Open Settings → Think Translation
2. Pick the target language (e.g. 日本語) — the settings panel, thinking rows and task cards all switch to it
3. Manage the provider chain — the list below it *is* the delivery order; drag rows to reorder, and the checkbox is the on/off switch:

• Built-ins: google gtx / bing (free, via the system proxy) and local Ollama — existing software/models are reused; missing components require the setup button (qwen2.5:7b / 14b or custom), with progress displayed. Built-ins are permanent: uncheck one to stop using it, there is nothing to delete
• DSH providers: endpoints already configured in DSH (llm-pi-ai.providers) appear automatically as read-only entries (badge "DSH"); their checkbox adds/removes them from the chain. The Import from DSH config button right below the list rescans on click and adds all of them in one go — there is no base URL and no key to retype, because the key is resolved from DSH's own credentials at request time
• Custom providers: "Add custom provider" registers any OpenAI-compatible or Anthropic Messages endpoint; the form offers presets for common low-cost models (DeepSeek, OpenAI, Qwen, GLM, Kimi, SiliconFlow, OpenRouter, Anthropic) that fill base URL + model + env name, all still editable; the model lists were checked against the vendors' documentation in September 2026 and are suggestions only — any model id can be typed in. Each row has a test button, an edit form and a × to delete it (a DSH row cannot be deleted — it is managed by the harness)
• A provider can name an environment variable (apiKeyEnv) instead of carrying the key: presets fill that in, and a DSH row brings its own. The key is resolved per request and never written to config.json, and such a row shows an env:NAME badge. The edit form has no env field (the value survives an edit untouched), but emptying a field really removes it — the field is sent as an explicit delete
• The former fallback chain option is gone: it was redundant once the chain itself ends with the free providers, and unchecking a provider is how you opt out of it. fallback still exists in config.json for anyone who wants it (off by default)

1. Send a message that makes the model think, then expand the Think row to read the translation and compare with the original

⚙️ How it works

browser → POST /_xlate/translate (same-origin, no CORS)
  → host provider chain (fail-open, user-ordered):
      chain: [provider1, provider2, ...]   ← drag-reordered in settings
        each provider is one of:
          google   (gtx via native Node HTTP/HTTPS and the system proxy)
          bing     (ttranslatev3 via native Node HTTP/HTTPS)
          openai   (OpenAI-compatible /chat/completions — Ollama local or any gateway)
          anthropic(Anthropic Messages API /v1/messages)
      fallback chain (config-only, off by default) tried when the primary chain fails entirely
  → original text if all enabled providers fail

• Provider config lives in config.json (runtime, gitignored): chain (ordered ids), fallback (enabled + chain, config-only), providers (per-provider type/enabled/baseURL/apiKey/apiKeyEnv/model). Old priority-based configs auto-migrate. A provider that declares apiKeyEnv resolves its key from that environment variable at request time (the literal apiKey stays as the fallback), and no env-resolved key is ever written back to config.json. A null field in a config patch deletes that field, which is how the UI clears one.
• DSH discovery reads the harness profiles//cordis.patch.yml / legacy settings.yaml (llm-pi-ai.providers) and .credentials.yaml (refs) on load; discovered providers are marked source: "dsh", resolved keys stay in memory (never written to config.json), and a /_xlate/dsh-scan route re-reads them on demand.
• Host half (lib/index.js): provider adapters, ordered chain + fallback execution, LRU cache (600), per-provider override for tests, /_xlate/models listing, /_xlate/model/pull + pull-status model download management (auto-configures on completion)
• Client half (lib/client.js): 8-language UI, drag-reorderable provider list, add/edit/delete custom providers, per-provider test buttons, sentence/paragraph-batched translation, streaming Think rows, localStorage persistence (settings + translation cache)
• Pure display layer: originals remain in the transcript and model context

🛠 Development

• No build step: lib/client.js is the browser bundle (source = artifact), lib/index.js is the host ESM
• Client changes apply on page refresh; host changes need a web restart
• The 8-language strings live in the UI_TEXT dictionary in lib/client.js

📄 License

MIT