Introduction

The DSH (DeepSeek Harness) plugin. It splits a project into one main document + several module documents, makes the AI proactively ask questions to turn uncertain items into decided items, and places a small button to the left of the model selector in the input box to display project health.

The problem it solves: in long conversations, key project decisions are scattered across the chat history—the AI may forget, and you have nowhere to check. Puzzle Mode writes these decisions into files you can open and inspect anytime: the main document serves as the lookup entry point, module documents store details, and every entry includes source references so it can be reviewed anytime.

Core Features

Document splitting and proactive asking

The plugin splits the project into one main document and several module documents. It injects rules into the conversation, making the AI ask questions proactively to convert uncertain items into decided items. The questioning rules include: at most 10 questions per round, at most 10 options per question; explore branching choices from five dimensions; if there is only one way to do something, do not force options; always end each round of questions with the fixed question, “Should we pause first?”

Standardized pipeline

Section 5 of the main document, ## 工作流, is the standardized pipeline. To complete a specific task, repeated steps, tools, and rules are chained in order into a reusable path. A workflow is written as a ### 名字 block, containing ordered steps. When trigger words are matched, it is injected automatically, and every step remains readable.

Multi-project binding

A single session can bind multiple projects simultaneously. Binding information is written in the main document, not in a plugin state file. It supports appending bindings, unbinding, switching the current project, and other operations. When bound projects exceed 8, the panel only shows a reminder and does not block.

Conversation interaction and guards

The plugin includes response correction (Chinese, no repetition), urging (sharp tone), and file observation guard features. Response correction restricts the model to making statements in Chinese and not repeating the same meaning that has already been stated. Urging is triggered when the model continuously only reads or reads files in segments, with a sharp tone. The file observation guard suggests, “read first, then retry,” when writes are rejected.

Repeated thinking breaker

The plugin has a built-in repeated-thinking breaker mechanism. If the same tool is called three consecutive times with identical parameters, it blocks subsequent duplicate calls.

Installation and Activation

Installation command

Install using the plugin manager:

python3 "$DSH_HOME/plugin-manager.py" github liancha22 dsh-puzzle-mode v1.1.3

Activation method

Puzzle Mode is not a standalone mode. After installing it into the host combination, standard mode (or any preset) sessions will have it included. After installation, you must restart that profile (patchReload: startup), then refresh the browser page.

Notes

Dependencies and compatibility

Besides the @deepseek-ai/dsh-tools provided by the runtime, the plugin does not consume any external packages. The document format is v7.

Themes and coexistence

Themes are not included in the plugin package and must be downloaded separately from the repository. When installed together with dsh-infinite-gen-5, they coexist according to the contract.