dsh-open-efforts

Choose reasoning effort for additional models in DeepSeek Harness.

中文 · English

Overview

dsh-open-efforts adds official per-model reasoningEfforts metadata to models explicitly configured through @deepseek-ai/dsh-llm-pi-ai. DeepSeek Harness displays the native Effort control in its existing model menu and submits each selection through the standard request path.

Screenshots

Figure 1: Models connected through Chat Completions, Responses, or Messages can use the same reasoning-effort selector.

Figure 2: The menu provides Off and five standard levels. Each model and provider defines the levels it actually accepts.

Effort levels

The plugin exposes these six choices:

UI label	reasoningEfforts key	Wire value
Off	off	none
Low	low	low
Mid	medium	medium
High	high	high
XHigh	xhigh	xhigh
Max	max	max

The protocol field uses the official id medium, while the UI displays Mid. The selectable range consists of the six rows above; minimal sits outside that range.

Unofficial community plugin: This independent project is developed and maintained by community contributors.

Scope

The plugin handles models explicitly declared in the llm-pi-ai settings section:

• providers.<route>.models[]
• providers.<route>.modelOverrides.<modelId>

Catalog-only routes retain their existing configuration. The official deepseek-official adapter also retains its existing configuration.

DSH supports three protocols for custom routes:

• openai-completions
• openai-responses
• anthropic-messages

For openai-completions, the plugin writes the official compatibility switch:

compat:
  supportsReasoningEffort: true

Responses and Messages use their protocol-native reasoning fields directly.

Model compatibility

The UI always displays all six choices. Each model and provider determines which values it accepts. A model may accept only High and Max, or it may reject every effort value. Provider compatibility errors surface through the standard DSH request path.

Off uses the supported off: none wire mapping. Generic Chat Completions routes explicitly send reasoning_effort: "none". Responses and Messages translate the same logical choice into their protocol-native disable formats.

Installation

Requirements:

• Node.js ^22.19.0 or >=24.0.0
• DeepSeek Harness 0.1.0-rc.7 or a newer release with per-model reasoningEfforts
• @deepseek-ai/dsh-llm-pi-ai enabled

Run these commands from the repository directory:

npm run verify:release
dsh plugin --profile web add .

dsh plugin resolves the local path from the current working directory and adds the package's dsh.bundle declaration to the selected profile.

Verify the profile configuration:

dsh --profile web --dump-config

The output should contain open-efforts and dsh-open-efforts. Restart the matching DSH Web process and refresh the page.

For another profile, replace web with its name. Install the plugin separately for each profile.

Usage

Once a custom model appears in llm-pi-ai under models or modelOverrides:

1. Open the model menu in the composer.
2. Select the target model.
3. Open Effort.
4. Select Off / Low / Mid / High / XHigh / Max.

The plugin watches model-topology and settings updates. Reopen the model menu after adding or editing a model to load the latest metadata.

Configuration written by the plugin

A Chat Completions model receives this capability configuration:

models:
  - id: your-model
    reasoningEfforts:
      off: none
      low: low
      medium: medium
      high: high
      xhigh: xhigh
      max: max
    compat:
      supportsReasoningEffort: true

Model names, context windows, output limits, existing compat.thinkingFormat values, and unrelated fields remain intact. The plugin replaces each models array as a whole to follow DSH settings path-mutation rules for arrays.

Data and permissions

• Runtime access is limited to llm-pi-ai model-capability settings and resolveModel metadata.
• API keys, conversations, system prompts, messages, and model output stay outside the plugin's access scope.
• Network requests generated by the plugin: zero.
• Native DSH adapters continue to handle all model traffic.

Uninstall and full rollback

Remove the package from the profile:

dsh plugin --profile web remove dsh-open-efforts

Then inspect the profile and restart it:

dsh --profile web --dump-config

Uninstalling restores the runtime resolveModel method. Official declarations already stored in the settings backend remain available. For a full rollback, remove these fields from affected model entries:

• reasoningEfforts
• the plugin-added compat.supportsReasoningEffort

Keep all other compat fields.

Compatibility baseline

• Last verified: 2026-08-18
• DeepSeek Harness: 0.1.0-rc.7
• Official repository commit checked: 99f6f02fecdb7dff40c3fbc9470f5907c29f74ca
• Node.js: ^22.19.0 or >=24.0.0

The supported baseline starts with DeepSeek Harness 0.1.0-rc.7 and relies on per-model reasoningEfforts, reasoningEffort call configuration, and the native Model/Effort UI.

Development

The package uses TypeScript source, keeps zero runtime npm dependencies, and compiles ESM plus declarations into lib/:

npm run lint
npm test
npm run build
npm run verify:package
npm pack --dry-run

License

MIT. Report security issues through the private channel described in SECURITY.md. Credentials and pending vulnerabilities belong in the same private channel.