Foreword¶
When configuring a custom llm-pi-ai provider in DSH, beyond the provider and model, it’s often necessary to handle reasoning intensity: which levels each model supports, what wire value is actually sent for each level, and what the provider’s default is.
dsh-reasoning-settings is a DSH plugin designed for this scenario. It provides a dedicated reasoning settings page for custom providers and supports selecting the provider, model, and reasoning intensity in subagent calls.
What This Is¶
dsh-reasoning-settings is maintained by JuneLearn under the MIT license.
It focuses on model reasoning configuration and custom provider settings, primarily addressing two types of issues:
- Declaring selectable reasoning intensity levels for custom
llm-pi-aiproviders and configuring the actual wire value sent. - In subagent calls, making model routing and reasoning intensity inheritance better align with the parent’s current selection.
Core Capabilities¶
Settings Page Capabilities¶
The plugin’s settings page capabilities include:
- Automatically reading custom providers and models added in the official Models page.
- Declaring
off / minimal / low / medium / high / xhigh / maxreasoning intensity levels for each model. - Customizing the wire value sent to the API for each level.
- Setting a provider-wide default reasoning intensity and restricting it to levels supported by all models under that provider.
- Configuring reasoning wire format for
openai-completionsproviders. - Writing back to
llm-pi-ai.providers.*via the officialsettings.mutateAPI without touching API keys.
Subagent Call Capabilities¶
The plugin provides the following capabilities for subagent calls:
- Fixing the issue where in-process subagents inherit the official model from Agent creation instead of the parent’s current third-party provider/model selection.
- Allowing subagents to inherit an explicit parent reasoning intensity or fall back to the default reasoning intensity configured for the target provider.
- Resolving model-only sub-overrides (e.g.,
grok-4.5) to the user-configured third-party provider instead of an unrelated official provider. - Extending the standard
subagentandsubagent_forktools to includeprovider,model, andreasoning_effortfields per call without replacing their official lifecycle, background, or result behavior. - Server-side entry point fixes for in-process
subagent,subagent_fork, and Workflow children targeting Harness0.1.0-rc.5.
Installation and Enablement¶
Check the runtime environment before installation.
The runtime requires Node.js 22.19.x or 24 and above; Node.js 24 LTS is recommended. It also requires Git, pnpm, and access to registry.npmjs.org and github.com.
npx Method¶
The npx method does not require checking out the DeepSeek Harness source code.
First, install the plugin:
npx --yes -p @deepseek-ai/dsh dsh plugin --profile web add github:JuneLearn/dsh-reasoning-settings
After installation, start the Web:
npx --yes -p @deepseek-ai/dsh dsh web
The Web listens on http://127.0.0.1:3080 by default; other ports are used only when the default port is occupied or explicitly selected.
pnpm Source Method¶
The pnpm source method requires the DeepSeek Harness source tree.
Navigate to D:\deepseek-harness and install the plugin:
cd D:\deepseek-harness
pnpm dsh plugin --profile web add github:JuneLearn/dsh-reasoning-settings
Then start the Web:
cd D:\deepseek-harness
pnpm dsh web
Upgrade¶
To upgrade, run the corresponding installation command again; no need to uninstall or maintain profile-patch.
For the npx method, run again:
npx --yes -p @deepseek-ai/dsh dsh plugin --profile web add github:JuneLearn/dsh-reasoning-settings
For the pnpm source method, run again:
cd D:\deepseek-harness
pnpm dsh plugin --profile web add github:JuneLearn/dsh-reasoning-settings
Uninstall¶
To uninstall via the npx method:
npx --yes -p @deepseek-ai/dsh dsh plugin --profile web remove dsh-reasoning-settings
Typical Usage¶
Here is a complete usage path:
- First, add custom providers and models in the official Models page.
- Open
Settings > Reasoning effort. - Select supported reasoning intensity levels for each model.
- Set the provider default intensity.
- Click Save under that provider.
- After starting a new session, select the model and reasoning intensity in the model selector.
It’s important to distinguish between “can send” and “accepted”: the plugin only declares reasoning intensity levels that Harness can select and send; whether the relay API actually supports them depends on that relay.
If the request returns HTTP 400, cancel the unsupported levels for that model or adjust its wire value.
Subagent Tool Fields¶
The plugin extends the standard subagent and subagent_fork tools to include per-call fields:
provider: The configured provider ID.model: The model ID owned by that provider.reasoning_effort: An optionaloff / minimal / low / medium / high / xhigh / maxlevel.
Usage rules are as follows:
providerandmodelmust be provided together.- Omitting both maintains normal inheritance.
- When
reasoning_effortis omitted, it is treated as the default reasoning intensity configured for the target provider.
Applicable Scenarios and Notes¶
Suitable for the following scenarios:
- DSH usage scenarios where custom
llm-pi-aiproviders and models are configured. - Scenarios requiring declaration of reasoning intensity levels for different models and configuration of the actual wire value sent.
- Scenarios where provider, model, and reasoning intensity need to be specified per call in
subagentorsubagent_forkinvocations.
Before using, note:
- The plugin runs with the current dsh process permissions; check the source code, dependencies, and license before installation.
- The plugin does not touch API keys and only writes back to
llm-pi-ai.providers.*via the officialsettings.mutateAPI. - The plugin is independent of any specific relay; the materials include third-party API relay promotion links and redemption codes, but the plugin does not require any specific relay.
- Whether to use a third-party relay service should be judged based on its price, reliability, and privacy terms.
References¶
- GitHub:
https://github.com/JuneLearn/dsh-reasoning-settings - Community directory page:
https://www.skillhub.cn/plugins/JuneLearn/dsh-reasoning-settings
The community directory page is an independent site, not an official app store; do not interpret the plugin directory as having an official affiliation with DeepSeek / High-Flyer.