Preface

In DeepSeek Harness (DSH), models are primarily used for writing code, modifying configurations, and running tasks. To intuitively assess a model’s reasoning and strategic judgment, there often lacks a repeatable and comparable “small battlefield.” Gomoku (Five in a Row) has simple rules, a readable board state, and each move depends on weighing threats and opportunities, making it suitable as an observational side channel.

omdsh-dev/dsh-gomoku is a community-maintained DSH plugin: it places a 15×15 board on the Harness sidebar, allowing you to play against DeepSeek or any configured model, or even pit two models against each other to compare their chess strength.

What Is This

The plugin’s npm package is @yejiming/dsh-gomoku, with its source repository maintained by omdsh-dev and hosted at omdsh-dev/dsh-gomoku. The SkillHub directory classifies it under “Fun Customizations.”

The plugin does not provide search algorithms or heuristic pruning: every move is derived from the LLM’s reasoning and judgment. The game state, turns, and win/draw determination are handled on the browser side; the server only arbitrates AI moves, rejecting invalid responses without disrupting the frontend board.

Core Features

Game Modes

Below the board, there are options to switch modes and start a new game. Three play styles are supported:

  1. Human vs. AI: You play as black or white, and the AI takes the other side.
  2. Dual AI Battle: Select a model for black and another for white, and observe the entire game.
  3. Pause and Manual Override: Clicking [Pause] interrupts AI thinking, allowing you to manually alternate moves for both sides; clicking [Resume] afterward restores AI control. The button highlights in amber during the pause.

The rules follow freestyle Gomoku (no forbidden moves): double threes, double fours, and overlines are all legal. Five or more consecutive stones in any direction win the game; if the board fills with no winner, it’s a draw.

Models and Thinking Levels

Black and white sides can independently choose their model and thinking level (Off / High / Max) without affecting each other. If a selected model does not support a given level, it automatically falls back to the model’s default level without causing request failures.

Thinking Process and Prompts

The thinking records for black and white are displayed separately on the left and right sides of the board, collapsed by default; clicking a move expands the full reasoning text for that step. The reasoning block is returned via the reasoning field upon a move, draw, or error.

The system prompts for black and white are persistently displayed in their respective side panels, editable at any time, with an option to restore defaults with one click. The default prompts include rules, terminology (e.g., live two, live three, sleeping three, double live three, four-three), a mandatory thinking process, and few-shot examples; any modified text is sent as-is as the system prompt.

Background Games and Error Tolerance

Closing the board popup does not reset the game or interrupt ongoing AI thinking; the game state is saved in the browser’s store, allowing you to use Harness’s main features while playing.

The end-to-end timeout for a single move is fixed at 3,000,000 milliseconds (3,000 seconds), and the output token limit is fixed at 32,000. These values are sent by the browser with each request and cannot be adjusted in the interface. Streaming responses that are interrupted due to connection drops, rate limiting, or other transient faults will automatically retry within the attempt budget.

Installation and Activation

The plugin repository includes pre-built artifacts in lib/, so no local build is required during installation. Requires DSH ≥ 0.1.0-rc.6: the plugin depends on the webServer service from @deepseek-ai/dsh-host-webserver; earlier versions may cause the plugin entry to remain in a pending state.

dsh plugin --profile demo add @yejiming/dsh-gomoku

Method 2: GitHub Source Installation

dsh plugin --profile demo add github:omdsh-dev/dsh-gomoku

Method 3: Local or Tarball Installation

# Inside the plugin directory
dsh plugin --profile demo add .

# Or pack first then install
pnpm pack
dsh plugin --profile demo add ./yejiming-dsh-gomoku-0.0.1.tgz

The first use of dsh plugin initializes the profile. After installation, confirm the layer is written with --dump-config, then start:

dsh --profile demo --dump-config   # The output should include the gomoku layer
dsh --profile demo

To remove the plugin:

dsh plugin --profile demo remove @yejiming/dsh-gomoku

If pnpm prompts a peer dependency warning, it can be ignored: the required services are provided at runtime by the host DSH.

Typical Usage

  1. Install and start DSH with the demo profile, then open the Gomoku board from the sidebar.
  2. Select “Play as Black,” “Play as White,” or “Dual AI” mode, and click New Game.
  3. In the left and right panels, choose models and thinking levels for black and white; edit system prompts as needed to observe changes in chess strength.
  4. In human vs. AI mode, click an empty intersection to place a stone and wait for the AI’s response; in dual AI mode, observe both sides’ reasoning and moves.
  5. To intervene, click [Pause], manually place stones for both sides, then [Resume]; you can close the popup to handle other tasks, as the game continues in the background.

Each AI move is an independent assistant request: the user message includes a 15×15 board text with row and column annotations, the side to move, and (on retry) the previous invalid response and rejection reason. The model must return strict JSON, such as {"move": [row, col]}, {"draw": true}, or {"error": "reason"}. This request is not logged in the agent’s session log.

Configuration Options

All fields have loader defaults; there are no library-level defaults:

Key Description
moveTimeoutMs End-to-end timeout for a single AI move, default 3,000,000 ms
maxMoveOutputTokens Output token limit per move response, default 32,000
maxMoveAttempts Total AI attempts per move request, default 3; the last attempt includes correction feedback

The browser sends fixed values for timeout and output limits, with no interface for adjustment.

Applicable Scenarios and Notes

This plugin is suitable for developers who wish to observe model reasoning in a side channel within the DSH workflow, or compare the tactical performance of different models or prompts. The dual AI mode facilitates comparing reasoning thoroughness within a single game, rather than relying on subjective conversation impressions.

Before installation, it is recommended to review the source code and license (MIT). The plugin runs with the privileges of the current DSH process, and model calls route through your configured providers, generating token consumption independent of the main conversation. SkillHub is a community directory site and has no official affiliation with DeepSeek or High-Flyer.

Links

  • Directory Page: https://www.skillhub.cn/plugins/omdsh-dev/dsh-gomoku
  • GitHub: https://github.com/omdsh-dev/dsh-gomoku