dsh-thincoder-suite: Four Self-Discipline Plugins for DeepSeek Harness
`shenhuanageshei/dsh-thincoder-suite` is a DSH plugin ported from thincoder, designed around "everything is a plugin" to enforce four self-discipline rules for AI during development and turn trust-dependent workflows into auditable ones. It includes four core tools—advisor, eng/eng_coder, escalate, and consult_start—which address review loops, code direction, capability overflow, and blockers, respectively. Its core mechanisms include mandatory output reconciliation tables and citation verification to achieve review convergence; a "design review token" mechanism to control code-writing permissions; support for parallel multi-model consultation; and background task handling for long-running jobs. Installation requires DSH 0.2.0-rc.1 and job.
Read Moredsh-theme-tuner: Fine-Tuning the Interface Theme in DeepSeek Harness
`dsh-theme-tuner` is a Web extension for DeepSeek Harness (DSH) that allows users to fine-tune interface colors based on an existing theme. Mounted under "General Settings → Appearance," it supports independent adjustments for accent color, background, contrast, and gradient intensity, and saves configurations separately for light and dark modes. With real-time updates and a one-click reset feature, users can precisely customize interface details without extra extensions, addressing the difficulty of theme customization. A Web Profile restart is required after installation for changes to take effect.
Read MoreDeepSeek Harness Main Program Version Management Plugin dsh-versions: Safe Upgrades and Rollbacks
The dsh-versions plugin addresses the lack of a rollback mechanism when upgrading the DeepSeek Harness main program. It allows users to view versions and check for updates in the web interface, and supports online installation, reinstallation, and downgrading. With automatic snapshots and second-level rollback, it significantly reduces upgrade risk. The installation requires dsh ≥ 0.1.0 and is Windows-only (as it depends on Windows tools); users on other platforms need to run npm commands manually.
Read Moredsh-skills-mcp-group-manager: DeepSeek Harness Skill Grouping and MCP Management Plugin
The dsh-skills-mcp-group-manager plugin is designed to address the growing management complexity caused by the increasing number of skills and MCP servers in DeepSeek Harness (DSH). It provides a left-side panel that supports grouped skill management (including create, read, update, and delete operations, collapsing, and batch operations), independent toggles and dynamic mounting for MCP servers, and filtering of injected skills based on group states. Configuration and state are persistently saved, and the interface is a pure floating panel that does not modify the host layout. The plugin has minimal dependencies, is ready to use after installation, and is suitable for developers who need fine-grained control over skill injection scope and MCP service enablement.
Read Moredsh-anydoc-markdown: DeepSeek Harness Document-to-Markdown and Image Description Plugin
`dsh-anydoc-markdown` is a plugin for DeepSeek Harness designed to address the challenge of agents understanding images in non-text files (such as Word and PDF documents). The plugin converts documents into GitHub-style Markdown, uses VLM to describe images and replace them with placeholders, and supports drag-and-drop upload, batch requests, and offline mode. It provides a `convert_document` tool for agents to invoke on demand, with results either returned inline or saved to a local directory.
Read Moredsh-composer-dynamic-island: An Upward-Expanding Control Island for Web Chat Interfaces
The DSH Web plugin `dsh-composer-dynamic-island` provides a compact upward-expanding island mode for collapsing cluttered plugin controls in the DSH Web chat interface. It keeps the interface tidy without moving DOM nodes or changing the React tree structure. It supports automatic control discovery, manual opt-in, and locally persisted configuration, and allows users to expand or pin the island via mouse hover or touch. The plugin declares that it makes no network requests and does not read sensitive data. It is suitable for scenarios where users want to simplify the interface and reduce visual clutter.
Read More