DSH Models Input Modalities

English | 简体中文

A DeepSeek Harness (DSH) Web client plugin that adds a Model capabilities fold to every third-party (pi-ai) provider card on the Settings → Models page, declaring per model which inputs it accepts (whether images are allowed) and which reasoning levels it offers — the two per-model fields the page's own forms deliberately do not expose. Built against DSH 0.1.5-alpha.1 (peer range >=0.1.5-alpha.1 <0.2.0).

The providers themselves (provider ID, API base URL, protocol, API key, model list) are still created and edited entirely through the Models page forms; this plugin pre-writes no provider configuration.

What it does

Input modalities

• A three-way choice per model — Provider default, Text only, or Text and image.
• Exact adapter semantics — Text only writes input: [text], Text and image writes input: [text, image], and Provider default removes the field so the row inherits the installed catalog's modalities, then the route's defaultInput.

Reasoning levels

• A three-way choice per model — Catalog default, Not a reasoning model, or Declare levels, which unfolds the seven levels (off, minimal, low, medium, high, xhigh, max) as a tick box plus the wire spelling to send for it.
• Exact adapter semantics — a declared set writes the reasoningEfforts dict the adapter reads: every ticked level becomes a key whose value is the spelling dispatch sends, so max: ultra renames a level for a gateway with its own vocabulary, and an empty off stores a valueless off: — offered, and sent as no parameter at all. Not a reasoning model writes reasoningEfforts: false; Catalog default removes the field, keeping the installed catalog's capability, which for a hand-declared model is none.
• A seed that assumes nothing about "off" — switching a row to a declared set starts from low/medium/high, each spelled as its own name, with off left unticked: on the plain reasoning_effort wire an empty off is the very same request as naming no effort, so pre-offering it would promise a stop-thinking choice the endpoint may not honour. Tick off — and spell it — once the endpoint says how.
• This is what puts levels in the picker — the composer's model picker lists exactly the levels a model declares, so a gateway model added by hand gains its Off/Low/Medium/High choices here and nowhere else in the UI.
• Refused before the write — the adapter rejects a declared set with no level beyond off, and any level but off with no wire value; the card names the offending row and keeps Save closed rather than answering a rejected settings mutation.

Both claims

• One fold, one write — the two claims share the fold because they share the stored models array: two folds would each write it and fence each other into conflicts.
• Revision-fenced writes — the whole models array is written back under the revision read when the fold was opened, with the same array semantics and conflict handling as the Models page's own cards: a concurrent edit surfaces a conflict notice and reloads instead of silently clobbering.
• Field preservation — every other field of every row survives verbatim, including fields this card never shows.
• Follows the page live — add or remove a model in the catalog above and the rows follow it in place: the fold listens for the pushed settings/document-updated invalidation of its own namespace, so nothing has to be closed and reopened. A commit that outruns a running read is caught by revision comparison and re-read until the fold is level. An unsaved draft is never overwritten, not even by a reopen — a change that lands mid-draft parks until the draft settles: revert the edits and the fold re-reads silently, or save and the revision fence refuses the raced write and reloads behind a conflict notice.
• Localized UI — English and Chinese follow the Web UI's locale.

Usage

1. Install (see below) and restart dsh web.
2. Settings → Models: create your gateway provider with Add custom provider as usual, or open an existing one.
3. A Model capabilities fold appears under every third-party provider card; expand it and, per model, pick its input modalities and its reasoning state — for a model that reasons, choose Declare levels and tick the levels its endpoint serves.
4. Click Save. The claim lands in the user layer of $DSH_HOME/settings.yaml; the adapter picks it up on its next request — no restart required.

A model declared here writes exactly this (other fields of the row untouched):

models:
  - id: acme-think
    contextWindow: 262144
    maxTokens: 32768
    input:
      - text
      - image
    reasoningEfforts:
      off: null      # offered, and sent as no parameter at all
      low: low
      medium: medium
      high: high
      max: ultra     # a gateway with its own vocabulary

How it works

The plugin registers its component into the settings.models.provider-card extension seat exposed by the Models page (key llm-pi-ai, i.e. the cards of the whole pi-ai adapter family). On first expansion the fold reads the provider's stored models rows through the settings Remote, edits them locally, and writes the whole array back under the revision captured at read time — the same array semantics and conflict handling as the page's own cards (a concurrent edit prompts a conflict notice and a reload). All fields other than input and reasoningEfforts are preserved verbatim in every row.

The fold also subscribes to the Host's forwarded settings/document-updated event and filters it to llm-pi-ai, so the page's own model-list writes reach it without a remount. An open, clean fold re-reads silently, and keeps re-reading until its data is level with the newest announced revision; a closed one parks the notice and re-reads on the next expansion; one holding an unsaved draft parks the notice until the draft settles, so a reopen never costs edits. The card's own write is recognised by the revision it just committed and does not echo back as a refresh.

Installation

Prerequisites: DeepSeek Harness (dsh) >=0.1.5-alpha.1 <0.2.0 with the web profile.

From npm:

dsh plugin --profile web add @jcy2387/dsh-models-input-modalities
dsh web

From a local checkout (development):

git clone https://github.com/DamonBao/dsh-models-input-modalities.git
cd dsh-models-input-modalities
pnpm install && pnpm run build
dsh plugin --profile web add link:$PWD

• link: references this directory directly; after a change run pnpm run build and restart dsh. Without the link: prefix the install is a copy — update it with dsh plugin --profile web update.
• Uninstall: dsh plugin --profile web remove @jcy2387/dsh-models-input-modalities.

Development

Requirements: Node.js ^22.19.0 || >=24.0.0 and pnpm 11.7.

pnpm install
pnpm run check        # typecheck + test + build + publint, same as CI

Individual commands:

pnpm run typecheck    # host + client faces
pnpm test             # vitest suites over the pure row helpers
pnpm run build        # tsc d.ts + tsdown (lib/index.js & lib/client.cjs)

The client artifact is a self-contained bundle: React, Cordis, ui-slots, and ui-primitives are supplied by the Web shell, CSS Modules are inlined, and every other @deepseek-ai/* package is a type-only import (enforced by a build-time purity check).

CI (.github/workflows/ci.yml) runs the same gates on Node 22.22.0 and 24.x, verifies on tag pushes that the release tag matches the package version, audits the packed tarball's file list, and runs a consumer smoke test that installs the tarball into a scratch project (resolving the published peer ranges against the real registry) and imports every Node-side entry point.

Release

Publishing is automated by the Release workflow (.github/workflows/release.yml), which runs whenever a GitHub Release is published. It requires the release tag to equal the package version (an optional v prefix is stripped), re-runs the full quality gates, packs the tarball, and publishes to npm with provenance via OIDC trusted publishing — no long-lived NPM_TOKEN secret is involved.

One-time setup: configure trusted publishing on npmjs.com for @jcy2387/dsh-models-input-modalities, authorizing repository DamonBao/dsh-models-input-modalities with workflow release.yml (no environment).

The dist-tag follows the GitHub Release's pre-release flag: a full release (checkbox unchecked) publishes under latest — including rc versions — while a pre-release publishes under the channel tag derived from the version (0.1.1-alpha.2 → alpha, 0.1.1-rc.1 → rc). The workflow is idempotent — a version that already exists on npm is skipped, so a re-run after a partial failure republishes only what is missing.

A typical release:

# bump the version in package.json, then:
pnpm run check
VERSION="$(node -p "require('./package.json').version")"
git commit -am "release: $VERSION"
git tag "$VERSION"
git push origin main --tags

Then create and publish a GitHub Release for that tag.

Dependabot checks GitHub Actions dependencies weekly. npm version updates are intentionally not enabled for Dependabot: it regenerates pnpm-lock.yaml without the workspace overrides, so its PRs cannot pass pnpm install --frozen-lockfile — bump dependencies manually with pnpm update.

Repository layout

.
├─ src/
│  ├─ index.ts             # Host half: an intentionally empty apply (browser-only plugin)
│  ├─ model-row.ts         # the row vocabulary both claims share
│  ├─ image-input.ts       # pure row helpers for the per-model input claim
│  ├─ reasoning-efforts.ts # pure row helpers for the per-model reasoning claim
│  └─ client/              # Web half: the capability fold (controller, card, locales)
├─ tests/                  # vitest suites over the pure row helpers
├─ build/                  # tsdown preset for the self-contained client bundle
├─ .github/workflows/ci.yml       # validate + tarball audit + consumer smoke
├─ .github/workflows/release.yml  # npm publish on GitHub Release
├─ cordis.patch.yml
└─ README.md / README.zh.md

Known limitations

• Dormant (not yet configured) provider cards do not render the fold; a freshly created custom provider appears after it is saved.
• Route-level knobs — defaultInput, the default reasoning level, and the compat switches — plus modelOverrides for built-in provider-catalog models are out of this plugin's scope; set them directly in $DSH_HOME/settings.yaml.
• A declared level is a claim, not a check: nothing asks the gateway whether it serves that level or honors that spelling, and how the level travels on the wire (reasoning_effort, a thinking budget, chat-template kwargs) is compat's job. A level the endpoint refuses is refused by the provider, mid-turn.
• A level unticked and ticked again restarts from its default spelling — its own name, empty for off — rather than reviving what it carried before.
• In read-only settings deployments the fold is visible but cannot save.

License

MIT © jcy2387