DeepSeek Harness Plugin Hub

发布与管理完整 Harness Profiles,发现适合你的插件。

探索

插件目录环境预设文档中心动态

社区

发布插件联系我们报告问题

相关链接

Plugin Hub GitHubDeepSeek Harness 官方项目系统状态隐私说明
© 2026 DeepSeek Harness Plugin HubPowered byPaxTech

独立、非官方社区项目,与 DeepSeek 官方无隶属、授权或背书关系。

Boost — DeepSeek Harness 插件(DSH Plugin)
DeepSeek Harness Plugin Hub
ProfilesPlugins分类动态文档登录管理 Profiles
ProfilesPlugins分类动态文档登录
← Plugins
B

@local/dsh-boost

Boost

DeepSeek Harness 的 Quorum 模式套件:通过角色隔离的子代理实现多代理深度推理、对抗性验证、持久返回通道和上下文预算管理。

插件会安装到这里;不确定时保持 web。

npx -y @deepseek-ai/dsh plugin --profile web add github:improveTheWorld/dsh-quorum#2dba4048f3ae583813b5ce34e50958278053152f
README兼容性版本

兼容性与来源证明

Boost 以 @local/dsh-boost 发布,当前版本为 0.1.0。Plugin Hub 会校验它的 manifest,并保存精确安装来源,便于复现安装结果。

DSH 兼容范围
*
运行环境
any
发布来源
github
Registry 更新时间
2026/10/7

版本

0.1.0stable
2026/10/7

相关插件

正在加载相关插件…

最新版
0.1.0
DSH
*
HMR
重启进程
Tree shaking
未声明可安全裁剪
解包体积
未提供
文件数
未提供
Surface
any
许可证
MIT
发布源
github
GitHub
★ 0
周下载
0
最近提交
2026/10/13
查看源码 ↗
README Badge

点击下方 Badge 复制 Markdown,粘贴到 README 即可。

这是你的 Plugin?认领权益 · 优先安全扫描

验证 package.json 声明的 GitHub 仓库,即可管理这个公开页面。认领后,Hub 会优先安排当前版本的安全扫描,并在通过后公开展示结果。

认领这个 Plugin →
报告问题

相关插件

继续浏览 agents-orchestration 分类下经过校验的插件。

Headless@deepseek-ai/dsh-headlessdsh one-shot bundle:基于 dsh-base 的直接核心 Agent/Session 运行器,不包含 Host、HTTP 或浏览器层Experimental Agent Team Profile@deepseek-ai/dsh-experimental-agent-team-profile在一个实验性捆绑包中提供 Agent Teams 协作、工具和 Web UIExperimental Ralph Bundle@deepseek-ai/dsh-experimental-ralph-bundle使用全新代理执行明确请求的重复工作。Hindsight Coding Agents@vectorize-io/hindsight-coding-agents面向编程代理的仅反思式 Hindsight 长期记忆(可插拔支持多种运行框架:opencode、……),自动在后台导入(无需设置 CLI)。

README

Quorum

Deep multi-agent reasoning with mandatory adversarial verification for DeepSeek Harness (inspired by Antigravity /boost).

English | Français


English

What is Quorum?

If you are coming from Google Antigravity or multi-agent orchestration frameworks, Quorum brings the deep-reasoning /boost protocol to DeepSeek Harness (dsh).

In standard harness sessions, LLMs often exhibit complacency: declaring tasks "fixed" without running test suites, missing regressions, or getting stuck in conversational loops.

Quorum solves this by construction:

  1. Strategy First — The orchestrator inspects the workspace, decomposes the objective, and writes verifiable acceptance criteria before delegating.
  2. Parallel Fan-out — Work is dispatched in a single turn to role-isolated subagents with zero context contamination.
  3. Mandatory Adversarial Verification — An independent verifier runs in the foreground to aggressively attempt to falsify the solution against real test suites and entry points. No conclusion is delivered without raw execution output proving the verdict.

Three Tailored Presets

Quorum installs three selectable agent presets under Settings → Agent Presets:

PresetBase Harness PresetTool Presentation & ExecutionBest For
Quorum (PTC)ptc.patch.ymlProgrammatic Tool Calling (run_code + generated TypeScript SDK) inherited recursively by all subagentsMaximum token efficiency, complex workflows, high-speed automated code generation
Quorum (Standard)standard.patch.ymlNative tool calls (direct JSON schema calls) + workflow engineStandard models and native tool inspection
Quorum (Shell)minimal.patch.ymlBare persistent shell (pwsh/bash with persistent state across turns)Interactive shell tasks, stateful build environments

Three Isolated Worker Roles

Subagents are created with strict mechanical boundaries (toolFilter):

                     ┌───────────────────────────────┐
                     │    Quorum Orchestrator Lead   │
                     │  (Strategy, Plan, Integrate)  │
                     └───────────────┬───────────────┘
                                     │
           ┌─────────────────────────┼─────────────────────────┐
           ▼                         ▼                         ▼
┌─────────────────────┐   ┌─────────────────────┐   ┌─────────────────────┐
│ subagent_investigate│   │ subagent_implement  │   │   subagent_verify   │
│  (DeepInvestigator) │   │     (DeepCoder)     │   │(Adversarial Verifier│
├─────────────────────┤   ├─────────────────────┤   ├─────────────────────┤
│ • Read-only         │   │ • Scoped diffs      │   │ • Foreground run    │
│ • Write denied      │   │ • Unit tests        │   │ • Write denied      │
│ • Delegation denied │   │ • Delegation denied │   │ • Falsification     │
└─────────────────────┘   └─────────────────────┘   └─────────────────────┘
  • subagent_investigate (DeepInvestigator): Strictly read-only (write, edit, git commit denied). Traces call graphs, inspects logs, and eliminates hypotheses with raw proof.
  • subagent_implement (DeepCoder): Owns bounded file diffs and writes tests. Cannot delegate.
  • subagent_verify (Adversarial Verifier): Runs independently in the foreground. Attacks the implementation with real suites and boundary inputs. Rejects paraphrases — only raw command output is admissible.

The Seven Companion Host Services

Quorum is not just a prompt; it mounts seven battle-tested infrastructure services:

  1. boost-job-relay — Re-routes background job settlement notices to the root session when child workers complete early.
  2. dsh-detached-jobs (run_detached) — Allows long-running builds/tests to survive disposable workers by attaching ownership to the root session.
  3. dsh-boost-channel (channel_post / channel_read) — Typed, throttled asynchronous back-channel between workers and the lead (decouverte, avancement, question, resultat, echec).
  4. dsh-boost-context-budget — Context occupancy meter (context_occupancy) and fork protection threshold. Supports requested compaction (context_compact).
  5. dsh-guard-surrogate — Intercepts and repairs unpaired UTF-16 surrogates in tool results before session logging, preventing fatal HTTP 400 INVALID_REQUEST errors.
  6. agent-teams-shield — Automatically shields Quorum sessions from tool shadowing when @deepseek-ai/dsh-experimental-agent-team is active in the host profile.
  7. boost-status-command (/boost-status) — Command-plane status monitor that responds even while a tool call is in flight.

Quick Start & Installation

Option 1: Install from GitHub clone

git clone https://github.com/improveTheWorld/dsh-quorum.git
dsh plugin add ./dsh-quorum

Option 2: Install into a specific profile

dsh plugin --profile web add C:\path\to\dsh-quorum

Select in the Web GUI:

Go to Settings → Agent Presets, then select Quorum (PTC), Quorum (Standard), or Quorum (Shell).


Présentation en Français

Le mode Quorum pour DeepSeek Harness

Ce dépôt consolide les sources et l'infrastructure du mode Quorum : une famille de trois presets d'agent DSH pour le raisonnement profond multi-agents avec vérification adversariale obligatoire, conçue pour éliminer les complaisances et les hallucinations d'ingénierie.

Pourquoi « Quorum » ?

En systèmes distribués, un quorum est le nombre minimal de membres devant s'accorder pour qu'une décision soit valide. Dans ce mode, l'orchestrateur ne peut rien livrer tant que le vérificateur indépendant n'a pas falsifié et validé le résultat sur pièces brutes.


Les 10 lignes montées par le bundle agrégateur

cordis.patch.yml monte exactement les 10 lignes suivantes :

preset-quorum-ptc · preset-quorum-standard · preset-quorum-shell ·
boost-job-relay · boost-status-command · dsh-detached-jobs ·
dsh-guard-surrogate · dsh-boost-channel · dsh-boost-context-budget · dsh-boost-lessons
PaquetRôleTests
packages/boost-mode/Les 3 presets Quorum (quorum-ptc, quorum-standard, quorum-shell)Anti-dérive racine
packages/boost-relay/Relais hôte des règlements de jobs orphelins vers la racine29/29
packages/boost-status/Commande /boost-status lisible même en cours d'appel d'outil10/10
packages/detached-jobs/Outil run_detached pour jobs persistants rattachés à la racine58/58
packages/guard-surrogate/Réparation des surrogates UTF-16 isolés + Bouclier d'immunité Agent Teams31/31
packages/boost-channel/Canal typé à double bourse (channel_post, channel_read)52/52
packages/boost-context-budget/Mesure d'occupation, garde du fork et compaction demandée36/36
packages/boost-lessons/Journalisation passive des compactions de racine14/14

Tests et Validation

La suite complète s'exécute avec le runner natif de Node.js :

node --test

Résultat : 291/291 tests passés, 0 échec.

  • test/aggregate.test.mjs : Test anti-dérive strict garantissant la cohérence absolue entre le patch agrégateur racine et les sous-paquets.
  • Éprouvé en conditions réelles sur un corpus mesuré de plus d'un milliard de tokens.

Licence

MIT © bilel GATRI (@improveTheWorld)