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。

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

dsh-custom-provider

Custom Provider

DeepSeek Harness 官方 Models 页增强插件:models.dev 元数据填充、配置导入导出和单模型 JSON 编辑

插件会安装到这里;不确定时保持 web。

npx -y @deepseek-ai/dsh plugin --profile web add dsh-custom-provider@0.2.1
README兼容性版本

兼容性与来源证明

Custom Provider 以 dsh-custom-provider 发布,当前版本为 0.2.1。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
web
发布来源
npm
Registry 更新时间
2026/11/4

版本

0.2.1stable
2026/10/19
0.2.0stable
2026/10/19
0.1.10stable
2026/10/15
查看其余 12 个版本收起版本
0.1.9stable
2026/9/29
0.1.8stable
2026/9/28
0.1.7stable
2026/9/28
0.1.6stable
2026/9/25
0.1.5stable
2026/9/25
0.1.4stable
2026/9/24
0.1.3stable
2026/9/23
0.1.2stable
2026/9/23
0.1.1stable
2026/9/23
0.1.0stable
2026/9/23
0.6.0stable
2026/9/13
0.4.1stable
2026/9/11
最新版
0.2.1
DSH
*
HMR
重启进程
Tree shaking
未声明可安全裁剪
解包体积
71 kB
文件数
8
Surface
web
许可证
MIT
发布源
npm
GitHub
★ 0
周下载
141
安全扫描
✓ v0.2.1 扫描通过
最近提交
2026/9/29
查看源码 ↗项目主页 ↗
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

README

dsh-custom-provider

Model tools embedded in the official DeepSeek Harness Models settings page.

简体中文

The official page manages providers, credentials, model lists, and reasoning capabilities. This plugin adds only:

  • models.dev metadata: fill a model's name, context window, output limit, and supported text/image inputs.
  • Single-model JSON editing: edit one model while preserving sibling models and advanced fields.
  • Configuration import/export: transfer selected llm-pi-ai provider profiles without credential values or request headers.

The plugin registers no separate settings page and stores no separate model configuration.

Requirements and installation

Tested with DeepSeek Harness 0.2.1-rc.2, with the Web Settings surface and the built-in llm-pi-ai adapter enabled. Older versions are not supported by this revision.

dsh plugin --profile web add dsh-custom-provider

For a local checkout, run from this repository:

dsh plugin --profile web add "$PWD"

Restart DSH and open Settings → Models. The existing official provider cards remain the primary configuration surface.

Model tools

Open a provider’s official Edit form, then expand a model’s Model options. Each llm-pi-ai model has Fetch from models.dev and an inline Edit JSON disclosure in its own details. There is no separate model picker or model-tools dialog. New custom providers use the same inline controls.

Fetching updates that row’s official form fields and JSON together, without saving. It imports only name, contextWindow, maxTokens, and supported input values, leaving reasoningEfforts, compat, protocol settings, and credentials unchanged. Matching prefers the original full model ID, then its recognizable official provider, then the first matching bare ID. Requests share a six-hour in-memory cache; a failed refresh never imports an expired snapshot.

Valid JSON edits immediately update the model’s native ID, name, capacity, input-type, and reasoning controls. Fields without a native control, such as compat, stay in the same model draft and are included when you use the official provider Save button. There is no separate apply action or automatic settings write. Advanced fields still depend on the adapter’s schema and runtime support.

Your JSON spacing and caret position are preserved while typing. Native form edits update JSON automatically. Incomplete JSON or invalid visible-field values leave the form at its last valid state, keep the original text, and block provider Save until corrected or reset. If native fields change while JSON is incomplete, independent changes merge when the JSON becomes valid; competing changes to the same field are reported without overwriting either draft. Reset JSON reloads the current form draft, not the last saved configuration. Collapsing a model retains its JSON text, and Save also waits for metadata requests to finish.

The official editor owns the provider draft, schema validation, inherited model behavior, persistence, and revision conflicts. Editing JSON or fetching metadata makes no separate settings write. Sibling model edits and unsaved provider fields stay in the same draft. Adapter refusals and revision conflicts are displayed by the official form without closing it.

This version uses a narrowly scoped adapter around the registered official Models components because DSH does not yet expose model-detail or page-header slots. It retains the original section, injection, child-slot ownership, controller, forms, and Save operations; it does not modify the installed application. Compatibility depends on the native component structure as well as slots. Recheck the real-page tests when upgrading DSH; do not assume a newer release is compatible solely from its version number.

Configuration transfer

Import configuration and Export configuration appear beside the official Models page title, in the same order and toolbar position as the previous plugin. Their preview opens directly below the title.

Export selects configured provider profiles from the effective llm-pi-ai.providers settings, including profiles configured in the official page or inherited from a base layer. It omits API key values and custom request headers. deepseek-account and deepseek-official are excluded. Endpoint URLs may contain sensitive parameters, so inspect and protect the resulting file.

Import accepts the existing version-1 dsh-custom-provider/providers JSON format. The preview shows source and destination model counts and whether each provider is new, user-configured, inherited, or layered. Existing providers default to Skip; new providers default to Import. Replace updates the transferred provider fields and model list, preserving destination headers. Unselected providers and the credential store are untouched.

The file retains apiKeyEnv names. An imported profile may therefore use an existing credential reference on the destination. Configure missing credentials through the official page after importing. Inherited layers and the installed catalog still determine the final effective settings.

Model writes use the official form’s save path. Imports use the official settings service, its schema, and the exact revision on which the import preview was based. An incompatible file or obsolete revision is rejected without silently adapting it.

Upgrading from the independent configuration page

Existing providers and models already live in the official llm-pi-ai namespace and require no migration. The old Model configuration navigation item is removed. The plugin no longer edits credentials or profile registrations, and no longer provides provider management, endpoint discovery, or its own reasoning form.

The package name stays dsh-custom-provider, so existing installations can upgrade in place. Keep one plugin registration; do not insert another llm-pi-ai adapter for this plugin.

Development

pnpm install
npm run build
npm test
npm pack --dry-run --json

lib/client.js is generated from lib/client.source.js, lib/config.js, lib/modelsdev.js, and lib/official-models.js and committed because DSH loads it directly. Tests load the pinned official Models client and cover native draft synchronization and Save, inline JSON and conflict handling, import/export boundaries, adapter cleanup, and persistence through real DSH settings services.

Use an isolated home for browser checks:

preview_home="$(mktemp -d "${TMPDIR:-/tmp}/dsh-model-tools.XXXXXX")"
DSH_HOME="$preview_home" dsh --profile web --patch "$PWD/test/local.patch.yml" --no-open --port 0

Attribution and license

This project is a fork of Luck9Star/dsh-gateway-provider. The MIT license and copyright attribution are preserved in LICENSE.