dsh-tool-cassette: Recorder for DeepSeek Harness tool calling
Writing tests for Agent workflows typically involves handling two sources of instability simultaneously: model outputs and external tools. The official `dsh-llm-replay` can already replay the model stream; however, tools such as HTTP, MCP, and databases may still access the network, modify data, or rely on a service that is currently down during testing. Additionally, third-party APIs sometimes return `429` (Too Many Requests) in CI environments. `dsh-tool-cassette` is a community plugin for DeepSeek Harness (hereinafter referred to as DSH). Its approach is straightforward: the first time, the selected tools execute for real, and the normalized results are recorded into a cassette file. Subsequently, the system switches to replay mode, sets real tool invocations to zero, and returns the recorded results to the tool chain based on exact matching.
Read Moredsh-plugin-wiki-skills: Porting the claude-obsidian knowledge base skills suite to DeepSeek Harness
Maintained by Lion-1209 (v0.2.2, MIT), dsh-plugin-wiki-skills ported the prompt-centric vault workflow from claude-obsidian to the DeepSeek Harness plugin, allowing the dsh agent to autonomously maintain a knowledge wiki. It consists of five core skills: `wiki` is responsible for vault setup and scaffolding; `wiki-ingest` ingests materials, extracts entities, and establishes cross-references; `wiki-query` retrieves and answers with citations following the hot cache → index → pages order; `wiki-lint` checks for health issues such as orphans and dead links; and `save` writes down conversation insights.
Read MoreDHS-multi-agent-plugin: Integrating multi-agent orchestration into DeepSeek Harness
# Preface When writing agents using DSH (DeepSeek Harness), a single agent loop can cover most linear tasks. However, once tasks become more complex—such as researching first, then reviewing, and finally summarizing—you must handle task dependencies, parallel scheduling, and failure recovery yourself. Scattered logic is difficult to test and reuse. Below, I introduce Li3NGa's DHS-multi-agent-plugin, which transforms this orchestration infrastructure into a standalone plugin. # What is it? DHS-multi-agent-plugin is a DeepSeek Harness plugin. Its README positioning is succinctly defined as: "Turning DeepSeek Harness into a powerful multi-agent orchestration engine." The repository is maintained by Li3NGa.
Read More@lanbaolu/dsh-wechat-bridge: Integrate WeChat into local DSH Agent
# Introduction The philosophy of DSH (DeepSeek Harness) is "everything is a plugin," with the Agent running locally to handle code, files, and various automated tasks. The challenge is that these tasks often take more than a few seconds: running a test suite or processing a batch of files requires you to stay at the terminal to monitor progress; if the Agent requests permission midway, a human must be present to confirm. While using SSH or remote desktop to connect back to the machine solves the "operability," the experience is disjointed—typing commands and scrolling through sessions on a small screen is not as natural as daily chatting. The @lanbaolu/dsh-wechat-bridge introduced below takes a different approach: it integrates WeChat into DSH. Once you scan the code to bind your personal WeChat account, you can chat with the local Agent directly within WeChat.
Read Moredsh-plugin-doctor: Do a health check before installing the DSH plugin
dsh-plugin-doctor is an open-source DSH plugin health check tool maintained by lin-cheng-lab (MIT license), designed to address installation risks caused by release candidate (rc) version mismatches within the DSH plugin ecosystem. It registers as a DSH tool, enabling a quick health check with a single command before plugin installation, without requiring additional scripts or permissions. The health check comprises five checks: local environment version baseline, repository accessibility, bundle declaration, peer dependency compatibility with local versions (the core check, correctly handling prerelease rules), and whether build artifacts are committed. The results are categorized into three levels—ok, warn, and danger—and include a structured JSON report along with handling suggestions. Installation is supported via GitHub (not yet released) or local source code, with profi needing to be restarted after installation.
Read Moredsh-workspace-upload: Moving workspace file management into the dsh chat interface
# Introduction When running agents using dsh's Web profile, the session workspace is where agents read and write files. However, the GUI itself lacks a file entry point: to hand a local CSV to the agent, one must first log in to the server and copy it into the workspace; to verify the artifacts just generated by the agent, one can only switch back to the terminal and run `ls`. dsh-workspace-upload fills this gap—allowing direct browsing, uploading, downloading, renaming, creating, and deleting files in the session workspace directly from the chat interface. The following sections introduce this plugin in the order of features, architecture, installation, and API. # What is it? dsh-workspace-upload (version 0.1.0, MIT License) is maintained by LI-Huaa and is
Read Moredsh-3d-model-viewer: View 3D Models in DSH, enabling LLMs to understand parts.
# Introduction When using Large Language Models (LLMs) to address mechanical and structural problems, a common obstacle is the model files themselves. Formats such as OBJ, STL, and STEP are either pure geometric data or BREP entities that require a CAD kernel for discretization, meaning pure LLMs often receive nothing but a stream of bytes they cannot understand. Previous solutions either required running a local CAD toolchain to extract results or simply gave up. `dsh-3d-model-viewer` is a plugin in the DeepSeek Harness (DSH) ecosystem. The approach involves splitting the process into two steps: first, rendering the model directly in a web interface for human visualization, and second, translating the same model into a standardized, LLM-readable JSON to be passed to the language model. The following sections introduce its positioning and installation.
Read Moredsh-task-worktree: Isolated git worktree checkouts for each task
# Introduction A common issue when letting agents work in a repository is that all changes end up in the same workspace: changes from the previous task haven't been committed yet while the next task has started, mixing uncommitted content together. If you try to use branch switching for isolation, you must first deal with the current state. Git's native worktree allows maintaining multiple checkouts of the same repository simultaneously, each with its own attached branch. `dsh-task-worktree` integrates this capability into the DeepSeek Harness (hereinafter referred to as DSH) session model: it provides one isolated checkout per task, keeps the main workspace static, and leaves handover and cleanup actions to be performed by humans. Below, we introduce its positioning, features, and usage. # What is this `dsh-task-worktree` is Letter202
Read Moredsh-deepseek-balance: Real-time view of DeepSeek API balance and usage fees at the bottom right of DSH
# Introduction When running agent tasks with DSH (DeepSeek Harness), checking API balances and costs typically requires opening the official console separately: switching tabs, logging in, and scrolling through invoices disrupts your workflow. To conveniently keep track of your "remaining balance" and "daily spending," you need a monitor that persists within the client. The dsh-deepseek-balance plugin introduced below solves this problem: it places a floating badge in the bottom-right corner of the window, displaying the account balance in real-time along with usage and cost charts for the past 7 and 30 days. # What is it dsh-deepseek-balance is a DSH plugin maintained by lin-cheng-lab, currently at version 0.1.
Read Moredsh-eyes: Enable text-only models in DeepSeek Harness to process images on the fly
# Introduction When using pure text models like DeepSeek as the main model in DeepSeek Harness (DSH), there is an unavoidable limitation: Harness does not allow sending messages with images to sessions where the current model does not support images. If you want the model to view an error screenshot, an interface layout, or a photo of a table, you must either convert the content in the image to text and send it first, or switch to a multimodal main model. dsh-eyes provides an alternative approach: images can be pasted and sent as usual, and are persisted in the background; the main model invokes the `view_image` tool when it needs to view the image, and the recognition part is handled by any OpenAI-compatible vision interface (defaulting to Bailian Qwen). In terms of usage experience, the effect is close to the model's native capabilities...
Read Moredsh-experts: Represent the expert team as a data catalog and automatically register as a routable dsh skill
# Introduction A common challenge in agent orchestration is that expert personas, division of labor, and reporting rules are scattered across prompts and code; altering the review panel necessitates modifying the engine. The philosophy behind dsh (DeepSeek Harness) is "everything is a plugin." Following this logic, libinyam/dsh-experts transforms the "expert team" into a user-editable data directory: each team is automatically registered as a model-routable skill (`experts-<team-name>`), decoupling the engine from the lineup. The following sections introduce its design, installation, and usage. # What is it dsh-experts is a multi-expert plugin for dsh maintained by libinyam, released under the MIT license. One-sentence positioning: Translating the expert persona list into native dsh skills—
Read Moredsh-knowledge: DeepSeek Harness Knowledge Base Plugin with Local and Remote Dual Backends
# Introduction Using DSH (DeepSeek Harness) for daily development, an unavoidable issue arises: project standards, architectural decisions, and lessons learned occur within specific sessions and vanish once the session ends. The typical solution is to manually copy important content into system prompts, but this leads to increasingly bloated contexts, making it impossible to control retrieval scopes and write permissions. The dsh-knowledge introduced below turns this into a standalone component: without modifying the DSH Agent Loop, the same plugin can utilize local SQLite or connect to a remote centralized knowledge base, providing on-demand retrieval, controlled backfilling, and a Web management console. This article proceeds in the order of features, installation, and typical usage.
Read Morezotero-mcp-dsh: Connect local Zotero library to DSH
When writing academic papers, asking the model to help organize the main text is not difficult, but citing is challenging: the model does not know what literature exists in your library, and the DOIs and years provided casually are often fabricated. Existing approaches either involve manually organizing `.bib` files or pasting literature metadata piece by piece into the context, both of which interrupt the workflow. The `zotero-mcp-dsh` introduced below is a DSH plugin maintained by LiJunfeng000000000000, currently at version 0.1.0 and released under the MIT license. It connects a local Zotero library to DSH, enabling the model to retrieve literature, read basic information and full PDFs, and export BibTeX directly within the session, facilitating the construction of LaTeX reference libraries and the insertion of LaTeX citation commands.
Read Moredsh-deepseek-chat: Add an independent "Ask DeepSeek" chat window to the DeepSeek Harness Web UI
dsh-deepseek-chat is a DSH community plugin maintained by lhenlihai-hub (MIT License, v0.2.2). It adds an independent "Ask DeepSeek" chat window to the DeepSeek Harness Web UI, solving the need to ask DeepSeek separately without polluting the main conversation context. The plugin reuses the configured DeepSeek model in dsh (via `ctx.llm`), eliminating the need to enter an API key separately. Core features include opening an independent chat window by clicking on the page's right side, switching between configured models at the top, directly quoting selected text from the main session for questions, supporting streaming output and the thinking process, stopping generation, creating new conversations, and automatically adapting to Chinese/English interfaces and light/dark themes. The plugin conversation and main...
Read Moredsh-timer: DSH scheduler plugin to automatically resume tasks in the original session at the scheduled time.
dsh-timer is DSH's "scheduled task" plugin, designed to address the issue where messages are executed immediately upon sending and cannot be scheduled for later execution. It allows messages to be queued for a specific time, at which point they will automatically resume execution in the original session with full access permissions. The toggle is located on the toolbar below the message input box; upon enabling it, clicking "Send" with the mouse will trigger a schedule confirmation dialog (pressing Enter will not trigger it). Users can set execution times and modes, and a "Miss Peak Hours" feature is provided to skip the two peak periods of Beijing Time (9:00–12:00 and 14:00–18:00) with one click. The task list displays real-time statuses for pending, running, completed, and failed tasks, supports canceling and clearing, and supports unlimited concurrency. Installation requires a locally running DSH and is performed via `dsh plugin --profile web add github:life1996cou/d`.
Read Moredsh-ptc-cordis-preset: Add Creative Mode to DSH PTC
# Introduction When developing with DSH (DeepSeek Harness), there are two types of tasks that are very common. One type is tool orchestration: using the Code Mode SDK to write a `run_code` program to combine multiple tool calls into a single execution. The other type is creative work centered around the runtime: inspecting the live runtime, defining and experimenting with dynamic plugin packages, and even creating new agent presets. These two capabilities currently belong to two built-in presets in DSH: the built-in PTC preset (ID is `code`, renamed to `ptc` starting from dsh 0.1.2) provides Code Mode orchestration; the built-in creative preset (`cordis`) provides self-referential Cordis tools...
Read Morereview-workflow: Add a multi-reviewer blind review workflow to DeepSeek Harness
# Introduction Common problems with using a single agent for review include the preparation of materials and scoring occurring in the same context, which makes the scoring easily swayed by previous discussions. Even if multiple sessions are opened, there is no isolation between roles, nor are there forced checkpoints to guarantee that every step is actually completed. For developers building agent applications, achieving reproducibility and traceability for the "review" process requires a fixed workflow combined with role isolation. review-workflow implements this as a DSH (DeepSeek Harness) plugin, aligning with DSH's philosophy of "everything is a plugin." Below, I introduce its workflow, installation method, and the configuration required before integration. # What is it? review-workflow is a workflow-type DSH plugin maintained by LeslieWylie, p
Read Moredsh-web-speech-input: Add a microphone button to the DeepSeek Harness Web input field
dsh-web-speech-input is a speech-to-text plugin for the DeepSeek Harness Web UI (maintained by leozou320-ai, v1.0.0, MIT license). Its core functionality is adding a microphone button to the left side of the prompt composer. Upon clicking, it uses the browser's Web Speech API to write real-time speech recognition results into the current draft. It supports interim and final recognition text, automatic language selection (Page Language -> Browser Language -> zh-CN fallback), clear status indicators, and accessibility support. The plugin never sends a prompt automatically; the submission action is always performed by the user. In terms of privacy, microphone permission is only requested after the user clicks. There are no server-side components, no audio storage, and no network requests are sent; audio processing depends on browser implementation. Installation Requirements
Read Moredsh-whale-avatar: Add avatars and names to DSH Web UI assistant messages
dsh-whale-avatar is a pure client-side plugin for the DSH Web UI with a single function: adding avatars (the App icon "Blue Fat Fish") and names to assistant messages, solving the issue of distinguishing message sources in long chat histories. It is maintained by lhjlol and hosted on GitHub. The implementation involves injecting a style tag into the page and using a CSS rule via the ::before pseudo-element to render the avatar and name. It does not alter the DOM structure, so the effect persists after React re-rendering, and the styles adapt to the day/night theme. The avatar is a 56px PNG embedded in a 28x28 SVG data URI, ensuring clarity on Retina displays. Installation is available via the Oh-DSH Desktop plugin marketplace.
Read Moredsh-qwen-voice: Real-time voice control for DeepSeek Harness Web
# Introduction When developing with DSH Web, creating sessions, switching sessions, and dispatching tasks require step-by-step manual interaction on the page. To manage several sessions simultaneously, you have to switch back and forth in the sidebar. dsh-qwen-voice integrates real-time voice into the Harness page: you can create, continue, and schedule multiple named DSH sessions simply by speaking, and the voice connection remains uninterrupted when switching sessions. Below is an introduction to its features, installation steps, and usage considerations. # What is it dsh-qwen-voice (`leaveimagination/dsh-qwen-voice`) is an experimental voice control plugin for DeepSeek Harness Web, built upon Qwen Audio Agent. It uses...
Read Moredsh-taffy-theme (Taffy Live Studio): Apply a Pink and Gold Acrylic Theme for DSH Web
# Introduction When using DSH Web for daily development, if you want the interface to be more distinctive—for example, for live streams, screen recordings, or simply to change the mood—modifying the frontend styles directly is quite costly. DSH's philosophy is "everything is a plugin," and styling can be handled by plugins as well. The `lengzhanbao/dsh-taffy-theme` introduced below is a theme plugin designed solely for changing the look: install it to change the skin, and uninstall it to revert to the original. # What is this? `dsh-taffy-theme` (full name: Taffy Live Atelier / 塔菲直播工房), created by lengzhanbao, is categorized under "Fun Dress-up" in the community directory. It provides a virtual livestream workshop-style interface for DSH Web: a light "Flower House" and a dark...
Read Moredsh-messages-sanitizer: Automatically fix orphan tool_calls in DeepSeek Harness sessions
dsh-messages-sanitizer is a DeepSeek Harness plugin developed by Leeminjing (MIT licensed, v0.1.1) that addresses the issue where tool scheduling crashes leave orphaned `tool_calls` in a session, causing subsequent requests to be continuously rejected with `400 INVALID_REQUEST` and resulting in a deadlock. The plugin implements four layers of fixes: first, injecting a synthetic error tool-result to automatically continue when the agent returns to idle; second, proactively inserting response messages before the next round of requests; third, stripping dangling `tool_calls`, neutralizing orphaned tool messages, collapsing duplicate user messages, and forcing a retry when the API still returns a 400 error; and fourth, performing pure array correction as a fallback for non-frozen requests. The fix is idempotent, and...
Read Moredsh-lark-bridge: Get notified on Feishu when DSH stops working
**Introduction:** Running a long DSH session is most troublesome not because tasks are slow, but because it tends to silently halt and wait for you: the model may be stuck on `ask_user_question` awaiting a response, tools may be pending sandbox upgrade approvals, or rounds may end abruptly due to 4xx errors. While you are occupied with other matters, you might not notice until returning ten minutes later that it has already stopped. `dsh-lark-bridge` addresses this issue by invoking `lark-cli` in real-time to send Feishu/Lark notifications whenever a DSH session ceases working. The goal is to ensure that "DSH stops working = Must receive notification." Unlike existing methods (screen polling or scripts covering only single events), it covers scenarios such as waiting for interaction, task completion, being blocked, request backoff, process stagnation, and process termination.
Read Moredsh-claudecode-tips: Replace "Deep diving..." in DSH web profile with Claude Code style loading words
dsh-claudecode-tips (or @claude-code-tips/dsh-plugin) is a DSH plugin developed by ldgeng, designed for the web profile. It replaces the static status word "Deep diving..." during the turn runtime with the original 56 randomly switching loading words from Claude Code (such as "Thinking" and "Pondering"), while retaining the elapsed time timer. It is a pure browser-side client plugin that monitors the chat DOM.
Read Moredsh-mcp-lazy: DSH's Lazy-Loading MCP Plugin, Reducing Token Waste by Disclosing Tool Schemas on Demand
# Introduction The philosophy of DSH (DeepSeek Harness) is "everything is a plugin." Developers often install multiple MCPs on it, such as files, browsers, and databases. This incurs a specific cost: even if the current problem only requires file tools, all MCP tool names, descriptions, and parameters may enter the model context simultaneously. The more MCPs installed, the more tokens are consumed per round. Below is an introduction to leaforbook/dsh-mcp-lazy, a DSH plugin designed to hide descriptions of temporarily unused tools and load them when the task requires them. # What is it dsh-mcp-lazy (npm package name `@yilinxiao/dsh-mcp-lazy`) is an MCP for DeepSeek Harness (DSH).
Read More