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。

Gold Preset — 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
G

@local/dsh-gold-preset

Gold Preset

GOLD Auto 与工程专家,遵循共享指南并开展自适应协作

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:bigunmd/gold#981c5c7ba3617edc44e9b12890252ab7f97be5ba
README兼容性版本

兼容性与来源证明

Gold Preset 以 @local/dsh-gold-preset 发布,当前版本为 1.3.0。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
2026/10/4

版本

1.3.0stable
2026/10/4
1.2.0stable
2026/10/3
1.1.0stable
2026/10/2

相关插件

正在加载相关插件…

最新版
1.3.0
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
★ 1
周下载
0
最近提交
2026/10/4
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 developer-tools 分类下经过校验的插件。

Web App@deepseek-ai/dsh-web-appdsh 浏览器界面捆绑包:位于 dsh-base 之上的 Web 补丁层,加上运行时粘合插件(提供前端 dist、Web 界面提示符、bash 运行时变量和 URL 行)Sdk App@deepseek-ai/dsh-sdk-appdsh SDK 配置包:基于 dsh-base 提供 stdio JSON-RPC 服务和进程生命周期管理Experimental Inspector@deepseek-ai/dsh-experimental-inspector用于 Host 调试和 Client Runtime 检查的实验性跨 realm CDP hubExperimental Inspector Profile@deepseek-ai/dsh-experimental-inspector-profile用于原始 Session 日志和 Chat 节点检查的可选 Web 捆绑包

README

GOLD engineering presets

Installable Cordis bundle @local/dsh-gold-preset: one shared engineering core, selectable specialties, and adaptive collaboration. Version 1.3.0 adds reliable execution contracts and empirical evaluation tooling; all five existing preset identities remain compatible.

Documentation

Start with the user guide, quick start, or six task recipes. See validation and limitations, the changelog, and the release checklist.

These links target version-pinned documentation. In a source checkout, the same guides are under docs/. Guides are hosted in the repository rather than bundled into the runtime package.

Version 1.3.0: reliable execution

  • On-demand assignment/result contracts with parent integration checks, bounded execution budgets and explicit milestone policies.
  • Resume and closure reconcile outstanding children/jobs and uncertain effects before restarting work or claiming completion.
  • Verification ties acceptance criteria to content identities, including repeated uncommitted edits; maintainer fingerprint tooling excludes ignored inputs explicitly.
  • Real-model behavioral evaluation cases capture tool events, workspace differences, duration and provider-reported token usage. The initial adapter uses Codex CLI, not mounted DSH sessions; heuristics and sandbox effects require transcript review.
  • Git fixture isolation, packaged Markdown-link validation and an opt-in exact-package/runtime Loader gate. No new runtime plugins, permissions, or automatic agents.

See reliable execution, behavioral evaluations, and runtime gate.

Version 1.2.0: Git discipline and domain specialties

This source revision adds precise Conventional Commit/branch guidance, no-model-authorship rules and eight on-demand specialties: Git & Release, Security, API & Contracts, Database & Migrations, Performance, Documentation, Accessibility, and Incident Investigation. They do not add presets, permissions or automatic agents. See the repository's specialty and Git-conventions guides under docs/. Candidate Git metadata checks are available with npm run check:git -- --base BASE_SHA --head HEAD_SHA; branch naming is advisory in CI, and unknown identities/consent cannot be mechanically proven.

Selected specialty material is adapted from a pinned VoltAgent revision. Third-party notices ship with the runtime package. Required legal attribution and explicitly approved human coauthors are distinct from prohibited model authorship. These additions are included in v1.2.0, not in the earlier v1.1.0 tag.

Choose a starting point

PresetIDUse for
GOLD Auto (recommended)gold-autoRoute by your requested outcome and propose useful phases
GOLD DevelopmentgoldDevelopment, fixes and refactoring; preserves the existing identity
GOLD Architectgold-architectRequirements, architecture, trade-offs and planning
GOLD Reviewer / QAgold-reviewerEvidence-backed review, test strategy and validation
GOLD DevOpsgold-devopsCI/CD, infrastructure, releases and operations

A specialty is an initial focus, not a hard role wall. You can ask an Architect to move to implementation after approving its scope. Changing focus inside a session does not remount a preset or change tools/permissions. This bundle does not change your Harness default preset.

Choose how to collaborate

StyleHow GOLD works
GuidedExplains options and asks at consequential decisions; small meaningful stages
Balanced (default)Recommends an approach, handles routine approved details, asks on material changes
DelegatedResolves routine details within approved boundaries; reports milestones and escalates exceptions

Examples:

  • “Design this service with GOLD Architect, guided. No implementation.”
  • “Implement this in Balanced style; local code and tests only.”
  • “Switch to Delegated for the approved plan; no deployment.”
  • “Review only. Report findings; don't fix anything.”

For substantial work, GOLD proposes a style with the scope question if you have not already chosen one. It asks at meaningful decisions, not on timers or every few messages. Phases already covered by an approved complete brief do not trigger repeated approval. Pure questions stay lightweight. Delegated does not require subagents, remove acceptance gates, or authorize production actions.

Task preferences override session/project style preferences, but safety constraints remain constraints. Task overrides expire; preferences never prove action approval. This version uses conversation context and authorized local records, not global settings; no cross-session preference persistence is promised.

Scenarios and engineering guidelines

The packaged gold-standard-development skill routes Plan/Design, Develop, Debug/Investigate, Resolve issue, Review/Audit, Test/Validate and Release/Operate. Architecture/planning, software development, testing/QA, DevOps/reliability and security/privacy references load only when relevant.

Planning, diagnosis, review and test-execution reports do not imply repair permission or accepted delivery. Tests can change state: GOLD inspects their side effects and target before running them. Release preparation, publication, deployment and recovery are distinct scopes.

Optional process skills supplement the self-contained quality core but cannot impose automatic Git actions or committed planning artifacts. Project skills may shadow bundled resources; critical persona constraints remain in force and genuine conflicts must be surfaced.

Artifacts and boundaries

Local briefs, progress, investigations, evidence and bounded memory live in repository-root .gold/, ignored with /.gold/. Shared ADRs, C4 and runbooks stay tracked. Every closure assesses C4; update actual changes/proven inaccuracies, not diagrams or ADRs for ceremony.

Local records are a cache, not proof of approval. They do not travel through Git/worktrees. Ignore rules do not protect already tracked files or secrets. Read-only requests do not authorize notes files or .gitignore edits; non-Git workspaces have no ignore protection. Handoffs carry scope/provenance, revision/dirty state, decisions/evidence, risks and next authorized action.

Instruction-driven, not enforced isolation: Architect and Reviewer retain tools such as shell and are not read-only sandboxes. Runtime permission policy and explicit action scope remain authoritative. No preset or style grants publication, deployment, destructive recovery or remote issue closure permission.

Install and update

Use the Harness plugin_manager tool: action: install_bundle, target set to the absolute directory containing the reviewed bundle. Do not hand-edit profile manifests or install dependencies in the profile. Select the desired GOLD preset in a new session after activation; existing sessions retain their preset revision. Restart DSH when needed after package replacement.

For GitHub distribution, install github:bigunmd/gold#v1.3.0 from https://github.com/bigunmd/gold. The older github:bigunmd/gold#v1.0.0 does not contain this adaptive update. There is no automatic Git synchronization; private: true prevents accidental npm publication, not Git installation.

Local installations can link directly to the source directory. Keep that directory in place, and develop changes in an isolated workspace rather than accidentally modifying the installed source. Integration and activation are separate authorized actions. Skill paths resolve from the installed package identity, not an author's workstation path.

Development and verification

Maintainers edit preset-src/core-persona.txt, preset-src/roles.mjs and preset-src/shared-tools.yml; generate the declarative artifact with:

npm run build:presets
npm test
npm run check:presets
npm run check:docs
npm run check:package
npm pack --dry-run --json

cordis.patch.yml is generated and checked for freshness. Runtime consumers need only the generated artifact and packaged skills; generator inputs/scripts/tests and .gold/ do not ship. There are no install scripts, new dependencies, custom host plugins or native preset-inheritance assumptions. Shared Standard tool composition and planning/compaction/delegation isolation are preserved; compare against current Standard when upgrading Harness.

Tests cover generation behavior, stale checks, package contracts, bounded resources and relative links. tests/scenarios.md defines instruction evaluation cases. Distinguish structural tests, static consumer simulations and fresh-session runtime evidence: none alone proves model compliance or enforcement.

The 1.3.0 evidence record separates structural/runtime parsing checks from 28 repeated Codex-adapter model/tool runs. Strict regrading passed 11/14 runs for both baseline and initial candidate; semantic review retains command-scope and causal-RED gaps rather than claiming perfect compliance. Mounted DSH fresh-session behavior remains unverified. Historical evidence and current limitations are distinct.

Fresh-session smoke checklist after authorized activation

  1. Verify five selectable GOLD entries and no change to the global default.
  2. In a new session for each entry, load gold-standard-development and verify packaged resource discovery.
  3. Ask Auto to plan substantial work without selecting a style; expect a Balanced proposal combined with scope clarification.
  4. Choose Guided, then Delegated within approved local-only scope; expect meaningful checkpoints without an automatic team or production permission.
  5. Ask Architect for design only and Reviewer for findings only; expect no edits. Ask QA about side-effecting shared-DB tests and DevOps about release preparation; expect target/scope boundaries.
  6. Confirm covered phase transitions reuse approval, and completion reports current evidence without claiming acceptance.

Community and license

GOLD is available under the MIT license. See the contributing guidelines, Code of Conduct, security policy, and accessibility statement. Use GitHub's bug/feature forms for public reports; report vulnerabilities and conduct concerns through the private contacts in the policies.

Rollback

List bundles for the exact identifier, then disable @local/dsh-gold-preset with plugin_manager set_bundle, or remove it with remove_bundle, only when authorized. To restore a previous release, reinstall the explicitly chosen prior bundle/version. Existing sessions may retain loaded revisions. Never delete legacy files, linked source directories or local work as an automatic rollback step.