DeepSeek Harness Plugin Hub

Publish and manage complete Harness Profiles. Discover Plugins for your next setup.

Explore

PluginsPresetsDocsNews

Community

Publish a pluginContactReport an issue

Resources

Plugin Hub on GitHubDeepSeek HarnessSystem statusPrivacy notice
© 2026 DeepSeek Harness Plugin HubPowered byPaxTech

Independent and unofficial. Not affiliated with, authorized by, or endorsed by DeepSeek.

Model Thinking Levels — DSH Plugin for DeepSeek Harness
DeepSeek Harness Plugin Hub
ProfilesPluginsCategoriesNewsDocsSign inManage Profiles
ProfilesPluginsCategoriesNewsDocsSign in
← Plugins
M

@local/model-thinking-levels

Model Thinking Levels

Expose per-model thinking levels (reasoningEfforts) inside the Models settings card for a custom model API route.

The plugin will be installed here. Keep web if you are unsure.

npx -y @deepseek-ai/dsh plugin --profile web add github:Zessi-C/dsh-model-thinking-levels#22709a2b04b82773250705a546787ea4ac221f96
READMECompatibilityVersions

Compatibility and provenance

Model Thinking Levels is published as @local/model-thinking-levels and currently resolves to version 1.0.0. The Hub verifies its manifest and preserves the exact installation source for reproducible installs.

DSH compatibility
*
Runtime surfaces
web
Release source
github
Registry updated
9/27/2026

Versions

1.0.0stable
9/27/2026

Related plugins

Loading related plugins…

Latest
1.0.0
DSH
*
HMR
Process restart
Tree shaking
Safe tree shaking not declared
Unpacked size
Unavailable
Files
Unavailable
Surface
web
License
Not declared
Source
github
GitHub
★ 0
Weekly downloads
0
Last push
9/29/2026
View source ↗
README badge

Click the badge to copy Markdown for your README.

Do you maintain this Plugin?Claim benefit · Priority security scan

Verify the GitHub repository declared in package.json to manage this listing. After you claim it, Hub will prioritize a security scan of the current version and publish the result when it passes.

Claim this Plugin →
Report an issue

Related plugins

More verified plugins in models-usage.

Usage@linxin666/dsh-usageUsage statistics plugin for the dsh web GUI: per-provider balance and coding-plan quota detection plus a live token usage ledger, with a compact price glance card above the sidebar's Settings rowClient Ui Model Capabilities@linxin666/dsh-client-ui-model-capabilitiesPer-model capability declarations (image input and reasoning efforts) for custom pi-ai providers, edited in place on the Models settings cards; writes the official llm-pi-ai settings namespace, no DSH source changes.Codex Subscriptiondsh-codex-subscriptionUse ChatGPT and Codex subscriptions in DeepSeek Harness with OAuth, quota, safe resets, web search, images, and Fast modeCommandcode Provider@mars-sea/dsh-commandcode-providerUnofficial DeepSeek Harness LLM provider plugin for Command Code, ported from pi-commandcode-provider (MIT). Registers the 'commandcode' provider route with a Models-page card and live model catalog.

README

dsh-model-thinking-levels

给 DSH 的自定义模型 API 路由补上「思考强度」设置。

Unofficial DSH plugin. Per-model thinking levels for custom llm-pi-ai routes, with gateway-side capability probing. MIT.

功能

  • 在「设置 → 模型」的自定义 API 卡片里勾选每个模型支持的思考档位,写入该路由的 reasoningEfforts;
  • 打开面板时自动探测网关接受哪些档位(每个模型一次极小请求),并据此预填。网关没声明的档位仍然可选,但画成虚线,悬停提示「会被归并到相邻档位、勾了不生效」;
  • 未声明的模型预填成网关的档位,通常点一次「保存」就完事;把档位全部取消勾选则写入 reasoningEfforts: false(显式非推理模型)。

安装

dsh plugin --profile <profile> add github:Zessi-C/dsh-model-thinking-levels

也可以从本地目录装:dsh plugin --profile <profile> add /absolute/path/to/dsh-model-thinking-levels。

装完重启 DSH(Host 半是进程内模块,只刷新页面不生效)。

用法

设置 → 模型 → 你的自定义 API 卡片,里面会出现一行 ▸ 思考强度:

  1. 点开这一行。每个模型一行,右侧写着网关接受的档位,例如 网关接受:低 / 中 / 高;
  2. 勾选档位 → 保存;
  3. 回对话页,输入框上出现 Effort 行。

想让某个模型不显示 Effort 行:把它的档位全部取消勾选再保存。

命令行探测

面板的自动探测用的是同一个办法:给 reasoning_effort 传一个非法值,表驱动的网关会把 自己的档位表写进 400 报错里(valid levels: low, medium, high)。也可以单独跑:

node tools/probe-levels.mjs <baseURL> <model> [<model>...]

一个坑:CLIProxyAPI 对配置里手写的自定义 provider 模型,若模型条目没写 thinking, 会默认按 low, medium, high 校验——这是网关的默认值,不是模型的能力。 探测到恰好这三个档位时先怀疑是默认值,去上游 /v1/models 或网关配置里确认。 在网关侧显式声明即可修正:

    models:
      - name: upstream-model-id
        alias: exposed-model-id
        thinking:
          levels: [high]

不装插件时的等效写法

- id: llm-pi-ai
  config:
    providers:
      my-gateway:
        baseURL: https://gateway.example.com/v1
        apiKeyEnv: MY_GATEWAY_KEY
        models:
          - id: model-a
            reasoningEfforts:      # 只声明网关真正接受的档位
              off:                 # 空值 = 支持关闭,且不发送思考字段
              low: low
              medium: medium
              high: high

卸载

dsh plugin --profile <profile> remove @local/model-thinking-levels

License

MIT