Introduction

The core idea of DeepSeek Harness (DSH) is “everything is a plugin.” dsh-go-sensei is a workflow plugin maintained by Zhuang-A. Its purpose is to translate the mathematical judgments of a Go AI (win rate, score margin, candidate moves, and variations) into teacher-level explanations for self-studying players to review games.

The explanations and multi-turn follow-up questions are handled natively in DSH sessions, without requiring additional Go software. The plugin includes the KataGo engine and weights, and on Windows it is ready to use out of the box after installation.

Prerequisites

Before use, the following environment requirements must be met:

Item Requirement
DSH ≥ 0.1.5-rc.1
Node.js ≥ 22.19
Runtime environment Windows / macOS / Linux all supported
Bundled engine Windows ships with KataGo v1.16.4 + 18b weights

macOS / Linux users who have not configured an external engine will have the plugin operate in “pure Go principles” mode, without calculating win rate or score margin.

Installation

Run the following command in the terminal to install the plugin into the web profile:

dsh plugin --profile web add github:Zhuang-A/dsh-go-sensei

After installation is complete, restart dsh web for it to take effect.

Core Features

The plugin provides the following core capabilities:

  1. Move-by-move explanation: Give the game record to Sensei, and it will explain move by move, like a practice coach, covering intent, issues, and improvement suggestions.
  2. Multi-turn follow-up: Follow-up questions are supported, with diagrams included directly in the answers. For example, ask “What would happen if move 42 were changed to R16?” and the answer will draw the variation, mark key stones with triangles, and number the moves sequentially as 1-9, A-Z.
  3. Win rate and score margin visualization: Provides a win rate curve and a score margin curve (both from Black’s perspective). Problem moves are marked on the curves, and clicking a curve position jumps to the corresponding move.
  4. SGF annotation write-back: Explanations can be written move by move into the game record comments (C[] attribute), written into a -sensei copy in the same directory, while the source game record remains read-only.
  5. Markdown report generation: Can export a Markdown review report containing game information, a problem move table, and explanations.
  6. Rank-adaptive explanation: The density of terminology automatically adjusts according to both players’ ranks (18K~10K use everyday analogies, 9K~1D use standard terminology, 2D and above can directly discuss overall strategy).

Configuration

Configuration is done by editing the profile’s patch-layer file. The Windows path is %USERPROFILE%\.dsh\profiles\web\cordis.patch.yml, and macOS/Linux is ~/.dsh/profiles/web/cordis.patch.yml.

All default configuration items are optional; leaving them empty will use the built-in defaults:

# ── DeepGo Sensei ─────────────────────────────────────────
- id: go-sensei
  config:
    level: auto                 # 讲解难度：auto（按双方段位自适应）、18K..1K/1D..9D
    winrateThreshold: 0.03      # 问题手胜率落差阈值（0~1）
    scoreThreshold: 3           # 问题手目差阈值（目）
    maxCandidates: 10           # 单次复盘最多返回多少个问题手
    pvDepth: 6                  # 每条变化图保留多少手
    tokenBudget: 50000          # 单局讲解的 token 预算
    engineDir: ''               # 引擎目录；留空＝用插件自带的 engine/
    kataGoPath: ''              # 可选：可执行文件路径
    kataGoConfig: ''            # 可选：analysis 配置路径
    kataGoModel: ''             # 可选：权重文件；留空＝自动挑最大的 *.bin.gz
    maxVisits: 100              # 补算每手搜索量
    autoInstallSkill: true      # 加载时把随件讲棋技能装进技能根

Engine and Game Records

Engine

The plugin ships with an engine/ directory, including the Windows version of KataGo v1.16.4 and the 18b weights. macOS / Linux users need to download the engine for their platform themselves. When the engine is unavailable, the plugin automatically switches to pure Go principles mode.

Game Records

The plugin supports SGF game records. If the game record already contains AI analysis data (such as WV[] or LZ[] attributes), the plugin reads that data directly for explanations; if no data is present, the plugin automatically backfills the analysis.

Known Limitations

  1. Platform differences: Desktop and browser image forms differ. Desktop images are saved to disk and served through an authenticated path, while the browser uses absolute direct links.
  2. Dependency requirements: Depends on DSH >= 0.1.5-rc.1 and Node.js >= 22.19.
  3. Engine dependency: On non-Windows systems without an engine configured, win rate and score margin data cannot be provided.