dsh-model-reasoning

Smoothly Model Reasoning (Smoothly MR)

English | 简体中文

Brand
English brand	Smoothly
中文品牌	思磨力
English product name	Smoothly Model Reasoning — short: Smoothly MR
中文产品名	思磨力提供方参数

Smoothly Model Reasoning (Smoothly MR / 思磨力提供方参数) is an external DeepSeek Harness web client plugin: a Settings page that manages the provider-route parameters the built-in Models page deliberately does not expose — retry & backoff policy, timeouts, transport, caching, thinking budgets, capacities, request image budgets — plus per-model thinking levels (reasoning efforts) for third-party (pi-ai) providers. It writes the exact llm-pi-ai.providers.<route>.* fields the adapter reads, so changes take effect with no other configuration.

Why an external plugin: adding fields to the built-in ui-settings-models package would be overwritten by the next official release. This package ships as an installable bundle that never touches repository source, so official updates cannot clobber it.

What it manages

A new Settings section, 「思磨力提供方参数 / Smoothly MR」, placed after the built-in Models page. Pick any provider route — every entry in the llm-pi-ai providers dict is editable here, catalog routes included — and manage it in five parameter groups:

• Per model — for routes carrying an explicit models list, edit the SELECTED model's own declaration: input modalities (text/image, so a text-only model stays text-only beside its vision siblings), context and output caps (contextWindow, maxTokens), and thinking levels (reasoningEfforts: inherit / non-reasoning / a level set over off minimal low medium high xhigh max) with custom wire spelling per level. The picker fuses search into the dropdown itself — open it and type to narrow long model lists by name / id (display-only; stored order and write path untouched). Apply to all models copies only the CHECKED dimensions from the editor to every model on the route; this group also holds the route default thinking level (providers.<route>.reasoning) that unset models inherit.
• Retry & backoff: retryPolicy.mode (normal bounded transient retries vs always unbounded retries), maxRetries, retryableCodes (the five stable preset codes plus custom entries), and the shared exponential backoff (initialDelayMs, maxDelayMs, jitterRatio). Unset values fall back to the adapter defaults (5 retries, 500 ms initial, 10 s ceiling, 10% jitter).
• Timeouts & transport: timeoutMs, websocketConnectTimeoutMs, streamIdleTimeoutMs (default 300 000 ms), and transport (auto/sse/websocket/websocket-cached).
• Caching & thinking budgets: cacheRetention (none/short/long) and thinkingBudgets token budgets per level (minimal/low/medium/high).
• Capacities & budgets: defaultContextWindow, defaultMaxTokens, defaultInput modalities (text/image) — these three are FALLBACKS read only by models that declare nothing of their own (set the real values per model in the Per model tab) — plus the per-request image payload caps (maxRequestImageBytes, requestImagePixelBudget, requestImageMaxBytes).

Every group's panel opens with a scope badge: 「按模型 / Per model」 writes into the selected model's own declaration, and a dimension you leave unset inherits the route fallback or catalog value; the four other groups are 「整条 路由 / Whole route」 — those fields exist ONLY at route level in the llm-pi-ai schema, so one shared value is the whole truth (the badge's hover tooltip says so; no duplicate text sits beside it).

A field with an adapter default shows it as the placeholder while unset; clearing a field removes the override instead of writing an echo of the default. Local validation mirrors the host's own resolution rules, so most mistakes are caught before the write; anything the host still refuses surfaces verbatim from settings.mutate. The write path uses revision fencing, so a concurrent change is refused rather than silently overwritten. An Apply to all models action copies the checked dimensions of the current model's editor to every model on the route at once.

Custom wire spelling (adapt to any upstream vocabulary)

Each selected reasoning level has a wire-spelling field (defaults to the level name). Change it to remap what that level sends — e.g. max → ultra for a model that calls its top level Ultra, or high → turbo. off can either send nothing (default) or a custom value of its own.

⚠️ DSH does not support inventing new level names. The pi-ai schema pins reasoningEfforts keys to the seven canonical levels and resolution only reads those keys, so a bare ultra: key is rejected at write time. "Ultra" is expressed by remapping an existing level's wire spelling (max: ultra).

Empty state

With no third-party provider configured the page shows a friendly placeholder pointing at Settings → Models → Add a custom provider. Selecting a catalog route (no explicit models list) explains that its MODELS stay read-only here — catalog reasoning levels remain selectable in the composer — while the route's parameter groups above stay fully editable. Loading / unavailable states are hinted while the settings document loads.

Install

Prerequisites: a DeepSeek Harness install with the dsh CLI, plus pnpm (the dsh plugin command runs pnpm under the hood). This is an installable bundle — it is loaded by dsh, not imported as a library.

Compatibility: since v0.2.5 the page reads through the settings-forms service (ctx.configForms) and writes through the settings Remote namespace (ctx.remote.settings, the write path since v0.2.2); dsh builds older than 0.1.7 — which removed the predecessor settingsScope client seam — are not supported (v0.2.4 remains the release for dsh 0.1.2–0.1.6). The floor is also declared in package.json as an optional @deepseek-ai/dsh-client-ui-settings peer dependency (>=0.1.7-rc.1), making it machine-readable for the peer gate that ships from dsh 0.1.7-rc.1 on: an out-of-range bundle is refused with the exact dsh plugin allow-version remedy. Runtimes older than that gate evaluate no peers, so nothing refuses them: on 0.1.2–0.1.6 the page simply never loads, while the 0.1.7 alphas (seam present, gate absent) work as before.

From npm (recommended)

The package is published to npm as dsh-model-reasoning:

dsh plugin --profile web add dsh-model-reasoning

This installs the prebuilt bundle and appends it to the web profile. Then restart dsh web and open Settings → 思磨力提供方参数 / Smoothly MR.

From git

dsh plugin --profile web add github:karoc/dsh-model-reasoning#<sha>

A git install runs the package's prepare script to build the bundle. pnpm ≥ 10 requires allowlisting that build once — copy the exact package key pnpm prints into the profile's pnpm-workspace.yaml under allowBuilds, then re-run add (see docs/user/develop/basic/publish.md in the DSH repo).

Updating

Bump to the newest release with pnpm update (or re-add to pick up a newer git ref):

dsh plugin --profile web update dsh-model-reasoning
# or, if the dependency spec is pinned: dsh plugin --profile web add dsh-model-reasoning

Then restart dsh web so the new client bundle loads.

Removing

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

This removes both the dependency and its bundle layer from the web profile. Restart dsh web for the section to disappear.

Layout

cordis.patch.yml                # bundle layer: mounts the row that the client-modules
                                # service discovers (dsh.client manifest)
package.json                    # dsh.bundle (patch) + dsh.client (web) + exports["./client"]
tsdown.config.ts                # self-contained build: node half + module-table client bundle
src/index.ts                    # host apply (no-op)
src/client/index.ts             # client apply: configForms.get(llm-pi-ai) + register settings.section
src/client/ProviderParamsSection.tsx  # the settings page (route → group tabs → editors)
src/client/params.ts            # managed-field registry: domains, defaults, host-mirroring
                                # validators, draft model, minimal-op diff engine
src/client/styles.ts            # design-token styles (--dsw-alias-*) + injection
src/client/locales.ts           # en/zh copy
tests/params.test.ts            # unit tests for the pure registry logic

Build & test

npm install
npm run bundle       # emits lib/index.js + lib/client.js
npm test             # unit tests for params.ts (node:test runs TypeScript directly)
pnpm release:check   # release gate: docs/changelog/tag/tree/build/registry must all pass
npm publish          # runs the gate (prepack/prepublishOnly), then postpublish verifies the live release

The bundle leaves the platform packages (react, @deepseek-ai/cordis, @deepseek-ai/dsh-client-*) external — they resolve at runtime from the loader's module table; everything else is inlined.

Notes / limitations

• Only routes carrying an explicit models list expose per-MODEL editors here (the installed catalog is not reachable from the client); every route's route-level parameter groups are always editable. Catalog-only per-model customization via modelOverrides is future work.
• Credential management (apiKeyEnv), displayName, baseURL, protocol (api), and the models list structure stay on the built-in Models page; this plugin does not duplicate them.
• Wire spellings default to the level name and are edited per selected level in the page; the same per-model field can be set directly in settings.yaml (e.g. max: ultra).
• Legacy profile keys removed upstream (provider, maxRetries, maxRetryDelayMs at route level) are never written by this plugin.
• Section nav icon is shell-assigned, not plugin-assigned. The built-in ui-settings-general SettingsRoot.tsx navIcon(id) maps known ids and falls back to a gear for every other id — including this section's provider-params. When DSH exposes a per-section icon, use IconThinkOutlineRegular from dsh-client-ui-primitives for this section.

License

MIT

Contributing

See CONTRIBUTING.md (development + release checklist) and CHANGELOG.md for version history.