跪拜 Guibai
← Back to the summary

DeepSeek Harness: An Agent Runtime Where Even the Main Loop Is a Plugin

0. Conclusion First (TL;DR)

  1. What is DeepSeek Harness (dsh): An open-source Agent runtime framework (MIT license, GitHub: deepseek-ai/deepseek-harness) released by DeepSeek on 2026-08-13, with the core philosophy of "Everything is a Plugin." Its positioning is not another "programming assistant product," but a runtime base for freely assembling Agents. The official formula: Agent = Model + Harness. [VERIFIED - Official Repository/README]

  2. What is the "plugin" it refers to: Plugins are the capability units of dsh, loaded through the underlying Cordis framework. Model adapters, tool registries, session storage, sandboxes, and even the Agent main loop itself are all plugins. Extending dsh = mounting a plugin; there is no "privileged core" that requires patching. [VERIFIED - Official Repository/README + The New Stack]

  3. Difference from Claude Code's Skills/MCP/Subagent/Agent Team (in one sentence): Claude Code is a finished product + external extension points (Skills/MCP/Subagent/Agent Team/Hooks are all interfaces attached outside a fixed core); dsh breaks the product itself into plugins—the granularity of extension goes deep into the Agent Loop. You cannot replace Claude Code's main loop, but you can with dsh. [VERIFIED - Cross-referenced from multiple sources]

  4. How to create a plugin: Write a TypeScript file, export an apply(ctx) function, register a tool inside the function using ctx.tools.register(defineTool({...})), and then mount it with dsh web --patch ./cordis.yml. A tool plugin can be up and running in about 20 lines of code. [VERIFIED - CSDN DeepSeek Tech Community + Datawhale Tutorial + Juejin Testing]

  5. Do you need to create a plugin for project code development with dsh: Not for basic use. Install dsh → configure the model Key → select a workspace → start chatting directly. The Standard preset comes with a full suite of capabilities: files, Shell, search, planning, sub-agents, and workflows. You only need to write a plugin when you need custom tools, custom UI, or to swap the model adapter; for lightweight "instructions/workflows," Skills are sufficient and don't require writing a plugin. [VERIFIED - Official Web UI Guide + Multiple Tests]

  6. What does "the 0.8 version mentioning the use of Claude Code, Codex" mean: RC.8 made Claude Code and Codex into Profile Bundles that can be installed on demand, allowing dsh to call them as sub-agents. That is: dsh acts as the scheduling layer, responsible for breaking down tasks; Claude Code / Codex act as the execution layer, each doing their specialized work. The prerequisite is that you have these two CLIs installed locally and logged in (dsh finds the binaries from PATH). Note this is "dsh calling Claude Code," which is different from the other path (putting the DeepSeek model into Claude Code for use). Don't mix them up. [VERIFIED - GitHub Release v0.1.0-rc.8 Original Text + Jiqizhixin/Sina Tech + The New Stack]

  7. Important Risk Warning: dsh is a Developer Preview, and the official documentation warns in capital letters "THERE WILL BE COMPATIBILITY-BREAKING CHANGES." Use with caution in production environments. [VERIFIED - Official README]

1. What is DeepSeek Harness: Positioning and Working Mode

1.1 Basic Facts (Verified)

1.2 Core Positioning: Agent = Model + Harness

The official definition is straightforward: the model is the "soul" of the Agent, and the Harness is the "body" that enables it to work—context management, tool invocation, task planning, file reading/writing, code execution, permission control, memory, retry mechanisms, etc., are all here. [VERIFIED - Official README + Official API Documentation]

A plain-language translation: A large model is a thousand-mile horse, fast but unable to find its way or carry things; the Harness is the saddle, reins, and rider. The industry consensus in 2026 is that model capabilities are converging; what truly creates a gap is this "engineering shell."

1.3 Working Modes: Four Presets

dsh comes with four "factory combinations," embodying the "Everything is a Plugin" assembly philosophy: [VERIFIED - Official Documentation + The New Stack + Prompt Genius]

1.4 Three Entry Points

1.5 Model-Agnostic

The model adapter itself is also a plugin. The official provider directory covers DeepSeek, OpenAI, Anthropic, AWS Bedrock, Azure, Google Gemini (documentation still refers to it as Vertex), Kimi, and any OpenAI-compatible endpoint (including local Ollama, OpenRouter). Changing models = changing configuration, no recompilation needed. [VERIFIED - The New Stack + Prompt Genius + Official Documentation]

1.6 Session Logs: Append-Only Event Stream (Trajectory)

Everything the model sees—system prompts, reasoning processes, tool calls, sub-agent dispatches, every context injection—is written into an append-only session log. Recovery, forking, retrieval, replay, and auditing are all based on this event stream. The official principle is called "model-visible means logged" (everything that enters the model must be reconstructable from the log). [VERIFIED - Official Architecture Documentation + The New Stack]

1.7 Security and Sandbox

2. What Exactly is a "Plugin": Plugin Mechanism Detailed

2.1 Underlying: Cordis Framework

dsh runs on Cordis—a plugin meta-framework authored by Cui Tianyi (creator of the Koishi chatbot framework, former Jane Street engineer), accompanied by the paper "A Programming Paradigm for Spatiotemporal Composability" (88 pages, co-authored by researchers from Peking University and DeepSeek). The Cordis source code is copied into the dsh repository's vendor/ directory and renamed to the @deepseek-ai scope. [VERIFIED - ArceApps + Official Repository + CSDN Tech Community]

Cordis does only three things: loading, unloading, and dependency management of plugins. The core is "reversible side effects": when a plugin registers a tool, listens to an event, or mounts a web route, it simultaneously registers a cleanup action; when the plugin is unloaded/reloaded, all these registrations are revoked—hot swapping leaves no residue like "old tools still hanging" or "listeners piling up." [VERIFIED - ArceApps + WeChat Technical Article (Orange's AI Cafe)]

2.2 You Only Need to Understand 5 Concepts to Write a Plugin

[VERIFIED - CSDN DeepSeek Tech Community "DeepSeek Harness First Experience: Getting Started, Installation, and Custom Plugins"]

2.3 Three Roles: Capability Seam

A replaceable capability is split into three decoupled roles: [VERIFIED - Official ADR-0009 + ArceApps + dev.to]

After decoupling these three parts, changing the Provider does not affect the tool Schema seen by the Consumer. For example, switching Shell from local to a remote sandbox is imperceptible to the model.

2.4 Two Types of Plugins: Host Plugin vs. Client Plugin

[VERIFIED - TreeRouter Blog "Build DeepSeek Harness Plugins with Cordis Tutorial" + Official Package Structure]

Hard constraint: Host plugins cannot manipulate the DOM; client plugins cannot directly access the local file system; cross-environment communication goes through the dsh internal event bus.

2.5 Plugin Distribution Format: Bundle

Plugins are packaged as npm packages with two extra things: [VERIFIED - TreeRouter + ArceApps + CSDN]

Installation command: dsh plugin --profile web add <plugin package name>. After installation, the Harness service must be restarted (both plugin host code and browser code are loaded at startup; just refreshing the page is not enough). [VERIFIED - Consistent across multiple tests]

2.6 How Far "Everything is a Plugin" Goes

The official list of replaceable items (official original text): Model, Tools, Skills, Sessions, Sandbox, File System, Storage, Agent Loop, Scheduling, UI. ArceApps' comment hits the nail on the head: Agent Loop being a plugin means dsh is not "another Claude Code," but "a manifesto on how an Agent runtime should be structured to support self-modifying Agents." [VERIFIED - Official README + ArceApps]

3. How to Create a Plugin (Complete Hands-On)

The following code and commands are all from publicly tested tutorials (CSDN DeepSeek Tech Community, Datawhale Tutorial, Juejin, WeChat Official Account "DeepSeek Harness Plugin Development Complete Guide"), based on the official package API. [VERIFIED - Cross-referenced from multiple sources]

3.1 Environment Setup (Source Repository Needed for TS Plugin Development)

Daily use only requires npx @deepseek-ai/dsh web; developing original TS plugins requires cloning the source code (the official getting-started documentation method):

 git clone https://github.com/deepseek-ai/deepseek-harness.git
 cd deepseek-harness
 corepack enable
 pnpm install
 pnpm run build        # Cannot be omitted, otherwise the Web side lacks build artifacts and the plugin won't take effect

Node.js version requirement: ^22.19.0 || >=24.0.0, actual testing suggests using Node 24 directly (community testing v24.19.0). [VERIFIED - Official + Community Testing]

3.2 Minimal Plugin: Registering Nothing

A minimal plugin has only three things:

 import type { Context } from '@deepseek-ai/cordis'
 
 export const name = 'hello'        // Plugin name, for diagnostics only, can be omitted
 export function apply(ctx: Context) {
   console.log('hello from my first plugin')
 }

3.3 Tool Plugin: greet Example (Recommended Reference)

 import type { Context } from '@deepseek-ai/cordis'
 import { defineTool } from '@deepseek-ai/dsh-tools'
 
 export const name = 'greet-tool'
 export const inject = ['tools']    // Declares dependency on the tool registration service
 
 export function apply(ctx: Context) {
   ctx.tools.register(defineTool({
     name: 'greet',
     description: 'Greet someone by name.',
     parameters: {
       name: { type: 'string', required: true, description: 'The name to greet' }
     },
     async execute(args) {
       return `Hello, ${args.name}!`
     }
   }))
 }

The model knows the tool exists through the description, and knows how to call it through the JSON Schema of parameters. [VERIFIED - Datawhale Tutorial + Juejin]

3.4 Mounting to Web Service

Create a new cordis.yml (or cordis.patch.yml):

 - insert:
     - id: greet-tool
       name: "/your/absolute/path/deepseek-harness/scratch-plugin/src/greet-tool.ts"

Startup:

 pnpm dsh web --patch ./scratch-plugin/cordis.yml
 # If port conflicts: pnpm dsh web --patch ... --port 3082

Verification: Settings → Plugin list confirms "Enabled," then in a session, let the Agent call the greet tool. You can see the input and output of the tool call fully expanded, confirming the plugin loop runs through. [VERIFIED - Juejin + CSDN Testing]

3.5 Hard Rules for Production-Grade Plugins

[VERIFIED - WeChat Official Account "DeepSeek Harness Plugin Development Complete Guide" (based on 0.1.0-rc.6 source code testing) + CSDN]

3.6 Using Ready-Made Plugins (More Common Daily)

 dsh plugin --profile web add @dsh-external/dsh-git-workflow   # Example: Git workflow plugin
 dsh plugin --profile web update                                # Update all
 dsh plugin --profile web remove <plugin name>                  # Uninstall
 dsh restart web                                                # Restart to take effect

Community plugin discovery channels: GitHub topic dsh-plugin; third-party directory deepseek-harness-plugin.com. Note: You must review the source code before installing third-party plugins—plugins run within the host process, are trusted code, and can call tools, run programs, and read the workspace. [VERIFIED - Community Consensus + Efficiency Jun Review]

4. Comparison with Claude Code / Codex Extension Mechanisms

4.1 Clarify a Prerequisite First

Claude Code and OpenAI Codex are finished programming Agents for end-users; DeepSeek Harness is a runtime base for Agent builders. This determines the difference in extension philosophy. [VERIFIED - VentureBeat + The New Stack + Consistent across multiple comparison articles]

4.2 Claude Code's Extension Mechanisms (Per Official Documentation)

Claude Code's official documentation lists these extension points: [VERIFIED - Claude Code Official Documentation features-overview]

4.3 Core Difference: Extension Granularity is Not on the Same Level

Claude Code's extension = adding interfaces outside a finished product; dsh's plugins = disassembling the product itself.

4.4 dsh Plugins vs. Claude Code Skills/MCP/Subagent/Agent Team

Item-by-item comparison (in plain language):

4.5 One-Sentence Summary of Differences

Claude Code is a precision Swiss watch: the gears cannot be swapped, but the dial, strap, and crown have standard interfaces (Skills/MCP/Hooks/Subagents/Agent Teams). Codex is a rugged engineering vehicle: Rust core + dual-mode (local CLI/Cloud), extension via MCP + AGENTS.md + configuration. DeepSeek Harness is a box of Lego: no finished product, all building blocks, and even the "assembly method" itself can be swapped. [VERIFIED - Industry Comparison Article Consensus]

4.6 Known Shortcomings Comparison (Honest List)

Dimension DeepSeek Harness Claude Code / Codex
Maturity Developer Preview, breaking change warnings Mature commercial products
Managed Background Agent Not provided (not officially documented) Provided
Native GitHub PR Workflow Integration not complete Claude Code has GitHub Actions; Codex has Cloud tasks/auto PR
Extension Onboarding Cost High (many concepts: Profile/Bundle/Patch) Low (one-time install, ready to use out of the box)
Configuration Complexity cordis.yml assembly, patch whole-line replacement easy to misstep Configuration files relatively simple

[VERIFIED - VentureBeat Comparison Table + Prompt Genius + Community Criticism ("for Agent builders" not average users)]

5. Claude Code / Codex in RC.8 ("Version 0.8"): Meaning and Operation

5.1 Version Clarification

5.2 RC.8 Official Original Text (Key Item)

GitHub Release v0.1.0-rc.8 original text (New Features):

Make Claude Code and Codex subagents installable on demand as Profile Bundles, with non-interactive permission modes and named instances for Codex

Chinese official release notes:

Claude Code and Codex sub-agents can both be installed on demand as Profile Bundles, with Codex also supporting non-interactive permission modes and multiple named instances

[VERIFIED - GitHub Release Original Text]

5.3 What This Means: dsh Turns Claude Code/Codex into "Dispatchable Teammates"

Core meaning (consistent across multiple source interpretations) [VERIFIED - Jiqizhixin/Sina Tech + KAD + NetEase/AppSo + CSDN]:

  1. dsh becomes a unified scheduling layer: The upper layer (dsh + main model, e.g., DeepSeek V4 Pro) is responsible for breaking down tasks and orchestrating workflows; the lower layer pulls in Claude Code or Codex as sub-agents to work based on task needs. Architecture evolution path: RC.7 first made both sub-agent tasks appear in the Job Panel, RC.8 upgraded them to independent Profile Bundles for on-demand installation.

  2. Why do this: Different coding Agents have different strengths—Claude Code excels at complex reasoning/large codebase exploration (SWE-Bench Verified 80.8%, but token consumption is about 4x that of Codex); Codex executes directly, has high token efficiency, and is suitable for implementation tasks with clear acceptance criteria. In one task, you can "hand frontend styling to Claude Code, backend API refactoring to Codex, and orchestrate the whole thing with DeepSeek V4 Pro."

  3. Two new capabilities for Codex:

    • Non-interactive permission mode: Runs the entire process without manual confirmation, suitable for CI/CD and Harness batch processing.

    • Multiple named instances: Multiple Codex sub-agents for different purposes (different permissions, different working directories) can be configured in parallel on the same machine.

  4. Supporting mechanisms: reportDelivery allows sub-agents to actively report back after completing a task and wake up the parent task; web_search supports concurrent queries.

5.4 Prerequisites (Must Be Met)

[VERIFIED - The New Stack + Prompt Genius + Community Testing]

5.5 How to Operate (Community-Verified Path + Official Release Notes)

Since the official Profile Bundle installation command details for RC.8 are not yet uniformly documented in public materials, the following two paths are explained:

Path A: Official Built-in Sub-agent Providers (Already in RC.7/RC.8, Need Explicit Enabling)

  1. Install and log in to Claude Code / Codex CLI (resolvable in PATH).

  2. In dsh's settings.yaml, find the codex and claude-code provider lines within the standard preset (default disabled: true), and remove the disabled flag to enable them. [VERIFIED - CSDN User Test (fating__): "In the standard preset, there prominently lie the codex and claude-code provider lines, disabled by default"]

  3. In a session, let the main Agent break down the task and delegate to sub-agents; the sub-agent execution process can be viewed in the Job Panel (supported since RC.7).

Path B: Community Plugin dsh-plugin-product-subagents (Mature Solution with Complete Documentation)

This community plugin connects Codex, Claude Code, and any ACP CLI to dsh's sub-agent channel, supporting "continuable chat" sub-agents (retaining remote session IDs, appending messages over multiple rounds): [VERIFIED - Yeyupiaoling Blog + GitHub shaokeyibb/dsh-plugin-product-subagents]

 # Requirement: At least one logged-in product CLI (claude, codex, or ACP CLI) on PATH
 dsh plugin --profile web add dsh-plugin-product-subagents
 dsh restart web   # Plugin loads only after restart

Typical usage: In a session, let the model call product_roles to list the role library, product_agents to see which Providers are available on PATH, then assign tasks based on review/exploration/troubleshooting/implementation roles. The persistent registry defaults to ~/.dsh/product-subagents-registry.json (runtime state, do not commit to git). Note that manual package installation must use pnpm (npm will overwrite @deepseek-ai/dsh-tools symlinks, causing tool call errors).

5.6 Don't Confuse This with the Other Path: Connecting DeepSeek Model to Claude Code

The official DeepSeek API documentation (api-docs.deepseek.com/guides/coding_agents) records the opposite direction: running DeepSeek models (deepseek-v4-pro[1m], etc.) inside the Claude Code client via the Anthropic-compatible endpoint provided by DeepSeek (https://api.deepseek.com/anthropic). The configuration method involves setting environment variables like ANTHROPIC_BASE_URL / ANTHROPIC_AUTH_TOKEN. [VERIFIED - DeepSeek Official API Documentation]

The essential difference between the two paths:

6. Using dsh for Project Code Development: Practical Path

6.1 Conclusion: Basic Development Does Not Require Creating Plugins

The official Web UI guide and numerous tests confirm: daily project development follows four steps—"install → configure → select → chat." The Standard preset comes with complete programming Agent capabilities, no plugin writing needed. [VERIFIED - Official Web UI Guide + Alibaba Cloud Developer Community + Efficiency Jun Review]

6.2 Complete Process (Official Recommended Path + Tested Supplements)

Step 1: Installation

 # Requires Node.js ≥ 22 (24 LTS recommended)
 npx @deepseek-ai/dsh web
 # Or global installation
 npm install -g @deepseek-ai/dsh
 dsh web

The first run pulls the entire runtime from npm (not a lightweight script). It takes several minutes even on a good network, during which the process occupies one CPU core and prints no output—this is normal; don't kill it at the 60-second mark. After startup, the browser automatically opens http://127.0.0.1:3080 (auto-open since rc.8; use dsh web --no-open to disable). [VERIFIED - Official + Community Testing + Tencent Cloud Interpretation]

Step 2: Configure Model

Settings → Models → Enter DeepSeek API Key (starts with sk-) → Save takes effect immediately, no restart needed. The Key is only written to the local credentials file $DSH_HOME/.credentials.yaml (default ~/.dsh/), and the interface only displays a masked descriptor. You can also add Anthropic, OpenAI, Bedrock, Vertex, Azure, Codex, or a custom OpenAI-compatible gateway. [VERIFIED - Official Web UI Guide]

Step 3: Select Workspace

Click Choose workspace to add a project directory and select it. The workspace is the security boundary—the Agent can only operate within directories you explicitly add. The input box is disabled before a workspace is selected. It is strongly recommended to use Git version control for the project first; the Agent will genuinely modify and delete files, and Git is the final insurance. [VERIFIED - Official Guide + Practical Tutorial]

Step 4: Initiate Task

Create a new session and start chatting directly, e.g., "Summarize the structure of this repository and identify the main modules." The Agent can read files, modify code, run commands, and break down sub-tasks; sensitive operations (deleting files, dangerous commands) will trigger a manual approval popup in the Web UI. The interface language can be switched to Chinese in settings. [VERIFIED - Official Guide]

6.3 When Do You Actually Need a Plugin / Skill

 # api-doc-agent.cordis.patch.yml
 - id: model-override
   config:
     model: z-ai/glm-5.2
     provider: qiniu
 - id: tool-whitelist
   config:
     allowed: [read_file, bash]
     bash_whitelist: ["node scripts/parse-openapi.js"]
 - id: system-prompt-override
   config:
     sections:
       - id: role
         content: "You are an API documentation specialist. Only read files and run the OpenAPI parser script."

Startup: dsh --profile web --patch api-doc-agent.cordis.patch.yml

6.4 Efficiency Test Reference (Third-Party, Not Official)

7. Risks, Limitations, and Selection Advice

7.1 Must-Know Limitations (All Sourced)

  1. Developer Preview: The official README warns in capital letters "THERE WILL BE COMPATIBILITY-BREAKING CHANGES"; VentureBeat explicitly advises "enterprise developers should not treat it as a stable production platform." [VERIFIED]

  2. RC.8 Storage Incompatibility: SQLite backend refactored (Schema upgrade), old session data cannot be read, no migration tool provided officially; must back up ~/.dsh before upgrading. [VERIFIED - GitHub Release + Tencent Cloud Interpretation]

  3. npm Version Lag: As of the time multiple interpretations were published, npm latest was still rc.7, rc.8 is on the @next tag (npm install -g @deepseek-ai/dsh@next) or requires source build (git checkout dsh-v0.1.0-rc.8). [VERIFIED - Tencent Cloud Interpretation; another CSDN test succeeded with npm install @0.1.0-rc.8, both coexist, recommend following the official Release]

  4. No Managed Background Agent: No officially documented managed service managed by DeepSeek (Claude Code/Codex have this). [VERIFIED - VentureBeat/Prompt Genius]

  5. GitHub PR Workflow Not Complete. [VERIFIED - Prompt Genius]

  6. Plugin Security: Third-party plugins run within the host process, are trusted code, and can read/write the workspace and execute commands. Must review source code, license, and update frequency before installation. Official stance: does not consider packages from the official repository more authoritative than community packages. [VERIFIED - Official + Community Consensus]

  7. Configuration Complexity: Accurate community criticism—"Everything is a Plugin" is a selling point for developers, but means configuration complexity for average users; patch replaces entire lines by id rather than deep merging, covering one line requires restating all keys, easy for newbies to misstep. [VERIFIED - Community Comparison Articles + CSDN Technical Analysis]

  8. Price Change: DeepSeek V4-Pro switched from flat-rate to peak/off-peak pricing on 2026-08-16. Independent analysis (aitoolsreview) points out: even at the "off-peak" price, the output price is about 2.28x the old flat-rate system, peak about 4.55x—this is an actual price increase, not a decrease. Choosing dsh + DeepSeek API requires modeling based on actual usage. [VERIFIED - aitoolsreview independent analysis, marked as third-party calculation]

7.2 Selection Advice (Judgment Based on Above Facts)

8. Community-Recognized Recommended Plugin List

8.0 First, Clarify the Recommendation Criteria (Must Read)

8.1 Batch 1: Solve the "Usability" Problem First (Essential Level)

1. dsh-market / dshmarket —— Plugin Marketplace (Ecosystem Entry)

2. dsh-at-file —— @ File Reference in Input Box

3. dsh-context —— Context Usage Visualization

4. dsh-plugin-check —— Plugin Health Check

8.2 Batch 2: Feature Enhancements (Install by Scenario)

5. DeepSeek Harness Desktop —— Desktop Client (Unofficial)

6. ModLens (@liustack/modlens) —— Giving "Eyes" to Text-Only Models

7. dsh-genui —— Generative UI (Letting Answers "Grow Hands and Feet")

8. dsh-TUI —— Claude Code Style Terminal Interface

9. dsh-files / dsh-office-tools —— File Upload and Office Generation

10. dsh-browser —— Browser Operations (Codex-Level Capability)

11. dsh-explorer —— File Tree + Git Status Panel

8.3 Batch 3: Power Users / Advanced (Install as Needed)

12. dsh-agent-teams —— Multi-Agent Team Orchestration

13. dsh-anchored-standard —— Model Startup State Calibration (Community Called "Enlightenment" Plugin)

14. dsh-usage-stats —— Token Cost Statistics

15. dsh-turn-rewind / dsh-undo —— Context and File Rollback

16. dsh-im-bridge (WeChat Bridge) —— Remote Control dsh Anytime, Anywhere

17. dsh-plugin-cc —— Bridge Claude Code

18. Model Routing / Money-Saving Group: dsh-tier-router, dsh-llm-fallbacks, dsh-shift-router, dsh-model-failover

8.4 Strictly Not "Plugins," but Often Recommended as Plugins Major Projects

These projects themselves have a large user base and high stars, but are not necessarily native bundles installable via dsh plugin add, connected in their own ways. [VERIFIED - JB51/Qiniu Cloud Classification Criteria]

8.5 Scenario-Based Selection Advice (Community Recommended Combinations)

8.6 Installation Commands and Security Reminders (Unified Criteria)

 # Native bundle (npm package)
 dsh plugin --profile web add <plugin package name>
 # Native bundle (GitHub source)
 dsh plugin --profile web add git+https://github.com/<owner>/<repo>.git
 # Must restart Harness after installation to take effect
 dsh restart web

9. Reference Source List

Official Primary Sources (Highest Confidence)

  1. DeepSeek Harness GitHub Repository: https://github.com/deepseek-ai/deepseek-harness (README, docs/, packages/, vendor/cordis)

  2. GitHub Release v0.1.0-rc.8 Official Release Notes (including Chinese and English): https://github.com/deepseek-ai/deepseek-harness/releases/tag/dsh-v0.1.0-rc.8

  3. DeepSeek Official API Documentation "Connecting Agent Tools" (Claude Code/OpenCode/OpenClaw integration): Connecting Agent Tools | DeepSeek API Docs

  4. Claude Code Official Documentation Extend Claude Code (Skills/MCP/Subagents/Agent Teams/Hooks/Plugins): https://code.claude.com/docs/en/features-overview

International Authoritative Media / In-Depth Analysis (Higher Confidence)

  1. The New Stack, Frederic Lardinois "DeepSeek open sources an agent harness where everything is a plugin" (2026-08-13): DeepSeek open sources an agent harness where everything is a plugin - The New Stack

  2. VentureBeat, Carl Franzen "DeepSeek Harness launches as open source rival to Claude Code..." (2026-08-13): DeepSeek Harness launches as open source rival to Claude Code, alongside V4-Pro on API with higher prices | VentureBeat

  3. ArceApps Blog "DeepSeek Harness: the runtime where everything is a plugin" (including Cordis paper interpretation, star count snapshots, source code details): DeepSeek Harness: the runtime where everything is a plugin - ArceApps Blog

  4. Prompt Genius "DeepSeek Harness: Everything Is a Plugin": DeepSeek Harness: Everything Is a Plugin | Prompt Genius

  5. ScriptByAI "DeepSeek Harness: Open-Source Plugin-Based AI Agent Harness" (four modes, command quick reference): DeepSeek Harness: Open-Source Plugin-Based AI Agent Harness

Domestic Mainstream Media / Community In-Depth Content (Medium Confidence, Cross-Verified)

  1. Sina Tech/Jiqizhixin Pro "Just Now, DeepSeek Harness Updated! Enhanced Multimodality" (RC.8 report): Just Now, DeepSeek Harness Updated! Enhanced Multimodality|Session|Workflow|Tool|Long Chain|Image-Text_Mobile Sina

  2. Tencent Cloud Developer Community "DeepSeek Harness rc.8 Update Interpretation" (including npm version lag, SQLite incompatibility tips): DeepSeek Harness rc.8 Update Interpretation: Image Input, Auto-Open Webpage, and Two Practical Community Tools-Tencent Cloud Developer Community-Tencent Cloud

  3. CSDN DeepSeek Tech Community "DeepSeek Harness First Experience: Getting Started, Installation, and Custom Plugins" (5 plugin concepts + minimal plugin source code): DeepSeek Harness First Experience: Getting Started, Installation, and Custom Plugins_Natural Language Processing_qq_42859625-DeepSeek Tech Community

  4. Juejin "DeepSeek Harness Open Sourced, How to Get Started" (greet tool plugin complete process, Datawhale tutorial reference): https://juejin.cn/post/7674783303249887283

  5. CSDN "[DeepSeek Harness Technical Analysis] How Everything is a Plugin Reconstructs Agent Runtime" (four-dimensional comparison table with Claude Code/Codex/OpenCode): [DeepSeek Harness Technical Analysis] How Everything is a Plugin Reconstructs Agent Runtime-CSDN Blog

  6. Yeyupiaoling Blog "Using dsh-plugin-product-subagents to Connect Codex, Claude Code to Continuable Chat DSH Sub-Agents" (community solution complete documentation): Using dsh-plugin-product-subagents to Connect Codex, Claude Code to Continuable Chat DSH Sub-Agents - Yeyupiaoling

  7. CSDN "It Can Call Claude Code as a Subordinate—DeepSeek Harness's 10 Most Underestimated Usages" (sub-agent orchestration scenarios + settings.yaml enabling method): It Can Call Claude Code as a Subordinate—DeepSeek Harness's 10 Most Underestimated Usages_Artificial Intelligence_Qiniu Cloud Industry Application-MCP Tech Community

  8. CSDN "Thoughts on Using DeepSeek Harness" (tested settings.yaml codex/claude-code provider default disabled): Thoughts on Using DeepSeek Harness-CSDN Blog

  9. TreeRouter Blog "Build DeepSeek Harness Plugins with Cordis Tutorial" (host/client plugin two-type division, bundle structure): Build DeepSeek Harness Plugins with Cordis Tutorial - TreeRouter Blog

  10. NetEase/AppSo "Just Now, DeepSeek Harness Major Update, Multimodal Capabilities Enhanced": Just Now, DeepSeek Harness Major Update, Multimodal Capabilities Enhanced|Call|Workflow|Context|deepseek_NetEase Subscribe

  11. MoClaw Blog "DeepSeek Harness Plugins vs Skills" (plugin vs Skill security review differences): DeepSeek Harness Plugins vs Skills | MoClaw Blog

Chapter 8 (Plugin Recommendation List) New Sources Added (Supplemented 2026-08-21)

  1. Tencent News "I Installed 16 Plugins for DeepSeek Harness in One Go, The Most Ridiculous One Was Playing Web Game Ads" (2026-08-17, 16 plugins tested in batches + security reminders): I Installed 16 Plugins for DeepSeek Harness in One Go, The Most Ridiculous One Was Playing Web Game Ads_Tencent News

  2. JB51/Qiniu Cloud Industry Application "DeepSeek Harness 10 Truly Practical Plugins Detailed Explanation (High-Star Projects and Installation Guide)" (2026-08-20, 10 plugins categorized, three connection methods analysis, Star snapshots): DeepSeek Harness 10 Truly Practical Plugins Detailed Explanation (High-Star Projects and Installation Guide)_Other_AI_JB51

  3. JB51 "DeepSeek Harness Recommended 6 Plugins (Tested Without Pitfalls)" (dshmarket/dsh-vision-router/dsh-TUI/dsh-context/dsh-chat-import/dsh-at-file tested version numbers and installation order): DeepSeek Harness Recommended 6 Plugins (Tested Without Pitfalls)_Other_AI_JB51

  4. Tencent "Netizens Turned DeepSeek Harness into Cyber Lego, We Picked the 11 Most Interesting Plugins" (dsh-browser/dsh-agent-teams/dsh-genui, etc.): Search Information Page

  5. 53AI "DeepSeek Harness Gained 166k Stars in One Week, What Have Developers Turned It Into?" (dsh-browser/dsh-genui/dsh-openpencil/dsh-context/dsh-agent-teams): DeepSeek Harness Gained 166k Stars in One Week, What Have Developers Turned It Into? - 53AI-AI Knowledge Base|Enterprise AI Knowledge Base|Large Model Knowledge Base|Frontline Deployment Engineer|FDE|AIHub

  6. CSDN DeepSeek Tech Community "ModLens / dsh-genui Mechanism Analysis" (visual evidence pipeline, dsh-ui code fence + event loop source-level interpretation): DSH Best 10 Plugins_Artificial Intelligence_Everything is Lovely^-DeepSeek Tech Community

  7. Yeyupiaoling "Using dsh-agent-teams to Orchestrate DeepSeek Harness Sessions into Agent Teams" (source-level technical documentation: event-driven scheduling, file-level persistence, lock serialization, permission scope): Using dsh-agent-teams to Orchestrate DeepSeek Harness Sessions into Agent Teams - Yeyupiaoling

  8. CSDN "DeepSeek Harness Desktop: Nearly 6K Stars in 2 Days Open Source, a Desktop Client That Lets Ordinary People Use Local AI Agents" (Electron architecture, security mechanisms, pros/cons analysis): DeepSeek Harness Desktop: Nearly 6K Stars in 2 Days Open Source, a Desktop Client That Lets Ordinary People Use Local AI Agents.-CSDN Blog

  9. ima Knowledge Base/AI Trainer "Why 'Everything is a Plugin'..." (deepseek-harness-desktop zero-config solution, log as single source of truth): ima.copilot-AI Trainer

  10. TuohuangzheIT "DeepSeek Harness Desktop Hands-On Record" (command-line vs desktop version comparison table, tray/notifications/auto-update): DSH Desktop (Official Website)

  11. DeepSeek Harness Desktop Official Repository (anywhere-labs/deepseek-harness-desktop, MIT, README declares unofficial product): https://github.com/anywhere-labs/deepseek-harness-desktop

  12. Qiniu Cloud News "DeepSeek Harness Complete Getting Started Guide: Deployment Installation, Model Configuration to Must-Install Community Plugins" (open-design/OpenViking/archify/EverOS Star data): DeepSeek Harness Complete Getting Started Guide: Deployment Installation, Model Configuration to Must-Install Community Plugins | Qiniu Cloud