dsh-client-ui-model-reasoning

Graphical reasoning-effort editor for third-party models in the DeepSeek Harness Web UI.

The Web UI's built-in model form deliberately has no reasoning-effort control (effort is a per-model capability), so third-party models added through the Models page cannot declare thinking levels without hand-editing $DSH_HOME/settings.yaml. This plugin adds a Model reasoning settings section that edits llm-pi-ai model reasoningEfforts graphically — one card per provider, one row per model, with an offer checkbox and a wire-value input for each of the seven thinking levels (off / minimal / low / medium / high / xhigh / max).

How it works

• The section is registered into the existing settings.section slot (no changes to the built-in Models page or any host package).
• It reads and writes the llm-pi-ai settings namespace through the standard settings.describe / settings.mutate wire calls.
• Validation happens in the llm-pi-ai adapter's assertServiceable at write time: an unserviceable value is refused with settings-rejected naming the route and model, so the plugin needs no duplicate validator.
• Declared levels surface in the composer's model picker (Default / Off / High / Max … exactly what you declare), because the adapter's resolveModel() reports them back through the LLM capability metadata.

Install on a user machine (no source checkout)

Requires an installed dsh (any install form). The plugin's lib/client.js ships prebuilt in the package; the browser half is discovered from the package dsh.client declaration.

# From npm
dsh plugin --profile web add dsh-client-ui-model-reasoning

# Or from a tarball
dsh plugin --profile web add ./dsh-client-ui-model-reasoning-0.1.0.tgz

Then restart the web server:

dsh web

Open http://127.0.0.1:3080, go to Settings → Model reasoning, pick the levels per model, and save. The model picker now offers those levels for the configured third-party models.

To remove:

dsh plugin --profile web remove dsh-client-ui-model-reasoning

Build and publish

The package depends on @deepseek-ai/dsh-client-* packages published at 0.1.0-rc.6 — use the matching ^0.1.0-rc.6 range in peerDependencies/devDependencies so the browser loader resolves platform modules from the installed dsh's module table instead of bundling duplicates.

npm install          # or pnpm install
npm run build        # tsc types + tsdown → lib/index.js, lib/invariant.js, lib/client.js
npm pack             # produce the publishable tarball
npm publish          # publish (npm login first)

For a git-hosted install, the package needs a prepare script that builds lib/ from source (a git install fetches sources, not built artifacts); a prebuilt tarball or npm publish needs none.

Requirements

• The target models must be on an llm-pi-ai provider (the generic OpenAI-compatible / pi-ai adapter). Models on the official deepseek-official route already expose thinking through llm-deepseek and need no declaration.
• The section is read-only when the settings provider is read-only; it always loads even when llm-pi-ai is not mounted (it then shows the empty state and no providers).

Development notes

• Node half (src/index.ts) is deliberately empty: this plugin contributes only a settings section.
• The client bundle is built with the same closure-factory format as the in-repo packages/client/tsdown.client.ts preset (window.__ModuleLoader__.load({id, factory}), CSS Modules inlined with lightningcss).
• Platform modules (react, @deepseek-ai/cordis, @deepseek-ai/dsh-client-web-react, slot/locale/runtime contracts, …) stay external and resolve from the host's module table at runtime.

Known limitations

• Writes replace the whole declared dict per model (declared levels are the complete offer; undeclared levels become unsupported). Existing hand-written reasoningEfforts in settings.yaml are shown and editable, but a save rewrites the dict from the form.
• Catalog-route models are listed under modelOverrides; a model id that collides with a hand-declared models entry on the same route appears twice (two distinct rows).