dsh-think-any-lang

A DeepSeek Harness plugin that adds a Thinking Language dropdown under Settings → General, letting you choose which language the model uses for its reasoning / chain of thought (中文, English, 日本語, 한국어, Deutsch, …). The final reply stays in the user's language.

Screenshots

Features

• Settings → General → Thinking Language: a dropdown to pick the language the model thinks in.
• Zero latency, zero extra calls: implemented as a system-prompt instruction (systemPrompt.section) — no extra model calls or response delay like "translate the reasoning" approaches.
• Instructions written in the target language: each instruction is written in its own language (e.g. the English instruction is written in English), which improves model adherence.
• Persisted: the choice is stored in the user settings document (settings service) and survives restarts.
• Plain JS, no build step: both the host and browser halves are hand-written JavaScript; install straight from GitHub with zero friction.

Supported languages: off (follow default, no instruction), Simplified Chinese, English, 日本語, 한국어, Deutsch, Français, Español, Português, Русский, Italiano, العربية, हिन्दी.

How it works

Why a private RPC channel instead of the settings RPC surface? dsh-host-apiproxy only exposes an explicit allowlist of settings namespaces to configuration clients (WEB_SETTINGS_NAMESPACES / PRODUCT_SETTINGS_NAMESPACES); a third-party namespace answers settings-not-exposed, so the selection would never persist. The browser half therefore talks to a private channel registered in-process (ctx.connection.rpc.handle), and in-process ctx.settings access is not filtered by that allowlist — persistence and the systemPrompt linkage stay unchanged.

The language table is maintained in two places — index.js (instruction texts) and client.js (display labels): the browser factory's require only resolves platform modules and cannot read local modules, so each half keeps its own language view. Keep both files in sync when adding a language.

Installation

Requirements: dsh CLI installed with an initialized profile (dsh plugin --profile <name> add auto-initializes).

From GitHub (recommended)

dsh plugin --profile web add github:lco117/dsh-think-any-lang

The package is pure JS with no prepare build script, so no pnpm ≥10 allowBuilds approval is needed — install straight from GitHub. Pinning to a commit is recommended as in the official docs: github:lco117/dsh-think-any-lang#<commit-sha>.

From a local directory (development)

dsh plugin --profile web add ./dsh-think-any-lang

Verify

dsh --profile web --dump-config   # should show a "# == dsh-think-any-lang" layer

Usage

1. Start dsh web.
2. Open Settings → General.
3. Pick a language in the Thinking Language dropdown (default: Chinese).
4. Subsequent model calls receive the thinking-language instruction in the system prompt; DeepSeek reasoning models typically output reasoning_content in the chosen language.

Development

dsh-think-any-lang/
├── package.json       # declares dsh.bundle and dsh.client
├── cordis.patch.yml   # plugin row: mounts both halves of this package
├── index.js           # host half: settings namespace + systemPrompt.section
├── client.js          # browser half: settings row UI (hand-written factory bundle)
├── README.md          # 中文说明
├── README.en-US.md    # English docs
└── screenshots/       # README screenshots (webp)

Tech Stack

License

MIT © 2026 lco117

⭐ Star History

If this project helps you, give it a ⭐ to help more people discover dsh-think-any-lang.