dsh-thinking-level

A thinking-level (reasoning effort) selector for the DeepSeek Harness Web UI.

The plugin adds an independent full-width strip above the composer card (the conversation.input.dock slot), so it never replaces the official model selector. It shows a persistent status hint and offers exactly the thinking levels the current model declares — nothing is fabricated, and models that declare no levels are reported honestly.

• Default-max policy — when the current model declares max and no explicit effort is in force (the model default is active), the plugin submits reasoningEffort: "max" once and persists it through session.selectModel. When the model does not declare max, the model's declared default stays untouched.
• Shared model directory — the strip rides the same per-session model directory as the official composer model seat (ctx.modelDirectories, from @deepseek-ai/dsh-client-ui-model-selection), so switching models in either surface re-renders the other, and the Host stays the single fact source.
• Failure safety — a rejected selection keeps the last valid effort and shows a human-readable inline error.
• Bilingual — Simplified Chinese and English copy, following the active DSH locale.

Status hints

State	Chinese	English
Explicit effort in force	思考强度：Max	Thinking level: Max
On the model's declared default	模型默认思考强度：High	Model default thinking level: High
No default declared (provider default)	模型默认思考强度：Default	Model default thinking level: Default
Model declares no levels	当前模型不支持思考等级	This model does not support thinking levels
Directory loading	正在加载模型目录…	Loading model directory…

Requirements

• DeepSeek Harness with the default web profile (@deepseek-ai/dsh-base + @deepseek-ai/dsh-web-app).
• @deepseek-ai/dsh-client-ui-model-selection present (it is part of the default web composition; the strip depends on its shared model directory).
• A model that declares reasoningEfforts (pi-ai adapters) or adapter reasoning metadata — the strip only offers what the model declares.

Install

From the npm registry (pre-built; no build authorization needed):

dsh plugin --profile <name> add dsh-thinking-level

From GitHub (source install; pnpm runs the prepare build script — pnpm ≥10 will ask you to allowlist the package first, see docs/user/develop/basic/publish.zh.md):

dsh plugin --profile <name> add github:<you>/dsh-thinking-level

From a local checkout:

pnpm install && pnpm build
dsh plugin --profile <name> add ./dsh-thinking-level

Restart (or refresh) the Web UI — the strip appears above the composer once a workspace/session is open.

Declaring thinking levels for a model

The plugin never invents levels: they come from the model's declared reasoning metadata. For pi-ai adapters, declare them per model in settings.yaml:

llm-pi-ai:
  providers:
    my-gateway:
      api: openai-responses
      baseURL: https://gateway.example/v1
      models:
        - id: my-think-model
          name: My Think Model
          reasoningEfforts:
            off: null      # "supported, send nothing"
            low: low
            high: high
            max: max       # the level the default-max policy prefers
        - id: my-plain-model
          reasoningEfforts: false   # explicitly non-reasoning

With max declared and no explicit effort set, the strip auto-selects and persists max for the model.

How it works

• Slot: conversation.input.dock (list, session scope, replaceRisk: none) — registered with id community-thinking-level, order 100.
• State: the session's shared ModelDirectory store (same instance the official model seat renders), subscribed via useSyncExternalStore.
• Selection: session.selectModel({ provider, model, reasoningEffort }).
• The default-max decision is a pure function in src/client/policy.ts (unit-tested without a browser).

Development

pnpm install        # tooling only; the bundle itself has zero runtime deps
pnpm build          # tsdown: lib/index.js (host) + lib/client.js (browser)
pnpm test           # vitest: policy + component + plugin-registration suites
pnpm typecheck      # needs the DSH checkout (see scripts/link-dsh.mjs)

• The client bundle's only runtime externals are react and react/jsx-runtime; every DSH import is type-only and erased at build time, so the published artifact needs no DSH packages installed.
• scripts/link-dsh.mjs junctions the local DSH checkout's built lib/types output into node_modules/@dsh-types for tsc (set DSH_CHECKOUT to point at your checkout).

Limitations

• Requires the default web composition (in particular @deepseek-ai/dsh-client-ui-model-selection); removing it leaves the strip absent (the plugin stays pending on the service).
• Addressed subagent sessions hide the strip (same rule as the official model seat).
• The plugin reports what the model metadata declares; it cannot add levels a model does not advertise (see “Declaring thinking levels”).
• Explicitly chosen efforts are never overridden by the default-max policy.

License

MIT