DSH Effort Tweak

English | 中文

DSH Effort Tweak adds a small settings page for the per-model capabilities of each model in a DeepSeek Harness custom provider: input types, context window, maximum output, and the reasoning levels the endpoint accepts. It is designed for the case where the model capability must be edited in llm-pi-ai configuration but the official Web UI does not yet expose that control.

What it does

• Adds Settings → Model capabilities to the DSH Web UI.
• Lists models explicitly configured in llm-pi-ai.providers.<provider>.models.
• Edits four model fields together: input (text and/or image), contextWindow, maxTokens, and reasoningEfforts.
• Lets you choose off, minimal, low, medium, high, xhigh, or max for the reasoning levels.
• Accepts typed capacities such as 131072, 256K, or 1M, and leaves a blank field inherited instead of copying the fallback into the model.
• Shows what a blank field would inherit from the provider's defaultContextWindow, defaultMaxTokens, and defaultInput.
• Preserves the model's other fields and existing provider-specific wire values.
• Saves through DSH's settings service, so no manual configuration-file editing is needed.

The page intentionally does not show catalog-only providers. Add a model list to a custom provider first, then open this page to tune that model.

Install from a release

Windows PowerShell

Run PowerShell as your normal user and execute:

irm https://raw.githubusercontent.com/Toukaiteio/dsh-effort-tweak/main/scripts/Install-DshEffortTweak.ps1 | iex

To install into another Profile without starting Web UI:

$script = "$env:TEMP\Install-DshEffortTweak.ps1"
Invoke-WebRequest https://raw.githubusercontent.com/Toukaiteio/dsh-effort-tweak/main/scripts/Install-DshEffortTweak.ps1 -OutFile $script
& $script -Profile work -NoStart

macOS and Linux

curl -fsSL https://raw.githubusercontent.com/Toukaiteio/dsh-effort-tweak/main/scripts/install-dsh-effort-tweak.sh | bash

For another Profile:

curl -fsSL https://raw.githubusercontent.com/Toukaiteio/dsh-effort-tweak/main/scripts/install-dsh-effort-tweak.sh -o install-dsh-effort-tweak.sh
bash install-dsh-effort-tweak.sh --profile work --no-start

The scripts download the latest GitHub Release archive, verify the GitHub-provided SHA-256 digest when available, keep it under DSH_HOME/plugin-archives/dsh-effort-tweak/ for future dependency resolution, install it with dsh plugin, and start the default web Profile unless --no-start or -NoStart is supplied.

Install a local package

pnpm install
pnpm build
dsh plugin --profile web add ./dsh-effort-tweak-0.2.0.tgz

Restart DSH Web after installing or updating the plugin so the new client bundle is loaded.

Development

Requirements: Node.js 22.19 or newer and pnpm 10.

pnpm install
pnpm check
pnpm pack --pack-destination artifacts

The CI workflow checks types, runs unit and bundle integration tests, builds the host and browser bundles, and uploads the package archive. Pushing a matching tag such as v0.1.0 runs the release workflow and publishes the archive to GitHub Releases.

Configuration shape

The plugin edits this part of the llm-pi-ai namespace:

providers:
  my-gateway:
    defaultContextWindow: 128000   # fallback for a model that sizes nothing
    defaultMaxTokens: 8192
    defaultInput: [text]
    models:
      - id: my-reasoning-model
        contextWindow: 200000
        maxTokens: 16384
        input: [text, image]
        reasoningEfforts:
          off: null
          low: low
          high: high

• input declares the request modalities the endpoint accepts. Blank leaves the field out, so the installed catalog entry, then the provider's defaultInput, decide; llm-pi-ai treats an empty list the same way, so this page never writes one.
• contextWindow and maxTokens must be positive integers. Blank removes the field, so the provider's defaultContextWindow and defaultMaxTokens apply, and the page shows those values as the inherited hint.
• reasoningEfforts decides which effort levels the DSH model picker can offer. The plugin uses the existing value for a level when one is already present; a newly enabled level is initialized with its canonical name.

License

MIT. See LICENSE.