dsh-models-import

dsh-models-import

English | 中文

A DeepSeek Harness (DSH) plugin that adds a "Models Pro" settings section with capability-aware model import and editing — the stock Models section stays untouched.

• "Fetch available models" recognizes the capabilities extensions OpenAI-compatible gateways return in /v1/models (vision, reasoning/thinking, tools, search, thinking dialect, …) and adopts them into the imported model entries;
• a model row's disclosure edits more than capacities — can-think (with thinking levels) and accepts images are editable beside context window and max output.

The problem it solves

Stock dsh model discovery reads only the plain OpenAI fields (id, context_window, max_output_tokens) and drops the gateway capabilities object; the stock row editor edits capacities only. This plugin:

1. Host half registers POST /plugins/models-import/models on the web server, fetches <baseURL>/models for you (resolving the provider's stored credential), parses capabilities, and maps them to llm-pi-ai model fields;
2. Browser half adds a "Models Pro" settings section (built on an extended copy of the stock page): the same provider cards and editors plus capability-aware fetch and per-model capability editing. The stock Models section keeps running the official implementation — both pages edit the same settings.yaml, so you can switch back at any time.

Capability mapping

gateway capabilities	llm-pi-ai model field
vision: true	input: ['text', 'image']
reasoning: true / thinking: true	reasoningEfforts (default off+high; off offered only when thinkingCanDisable is true)
thinkingFormat (openai / deepseek / openrouter / together / zai / qwen / string-thinking / ant-ling)	compat.thinkingFormat (other formats are left unset for pi-ai's own inference)
contextWindow / context_length	contextWindow
maxOutput / max_completion_tokens	maxTokens

tools / search / pdf / audioInput … have no llm-pi-ai configuration target yet and ride along as badges in the fetch dialog only.

Install

The dsh CLI does not need to be installed globally — every command below goes through npx.

Option 1: npm install (recommended — one command, no approvals)

The npm package ships prebuilt, so installing runs no build scripts at all:

npx -y @deepseek-ai/dsh plugin --profile web add dsh-models-import

Restart dsh web to activate. Upgrades are the same one command: ... add dsh-models-import@latest.

Mirror-registry users (e.g. npmmirror.com): a freshly published package takes a while to sync; if the install reports 404 or No matching version, pin the official registry for the install:

npx -y @deepseek-ai/dsh plugin --profile web add dsh-models-import --registry=https://registry.npmjs.org

Option 2: install from GitHub (sources; needs build approval)

Prerequisites: pnpm (npm i -g pnpm if missing).

Step 1 — run the install command (no global dsh needed; npx fetches it):

npx -y @deepseek-ai/dsh plugin --profile web add github:Dudo-m/dsh-models-import

Step 2 — the first run fails while pnpm waits for build approval (a git install pulls sources and must run the prepare build script). That is expected. Copy the exact key from your own error message (it looks like dsh-models-import@https://codeload.github.com/Dudo-m/dsh-models-import/tar.gz/<commit-sha> — the key embeds the commit, so copy it verbatim) into the web profile's pnpm-workspace.yaml:

• Windows: C:\Users\<you>\.dsh\profiles\web\pnpm-workspace.yaml
• macOS / Linux: ~/.dsh/profiles/web/pnpm-workspace.yaml

allowBuilds:
  'dsh-models-import@https://codeload.github.com/Dudo-m/dsh-models-import/tar.gz/<commit-sha>': true

Step 3 — re-run the same command from step 1. A + dsh-models-import github:Dudo-m/dsh-models-import line means it installed (prepare builds automatically during install).

If the retry fails with ERR_PNPM_ENOENT ... dsh-models-import_tmp_... (leftover state from the first, blocked attempt), remove and re-add once:

npx -y @deepseek-ai/dsh plugin --profile web remove dsh-models-import
npx -y @deepseek-ai/dsh plugin --profile web add github:Dudo-m/dsh-models-import

Take the approval seriously: it allows the package's code to run a build on your machine at install time. Only approve repositories you trust, and pin a commit (github:Dudo-m/dsh-models-import#<sha>).

Option 3: local clone (development)

git clone https://github.com/Dudo-m/dsh-models-import
cd dsh-models-import
pnpm install && pnpm run build

npx -y @deepseek-ai/dsh plugin --profile web add .

dsh plugin records it as a link: dependency in the web profile (~/.dsh/profiles/web) and appends it to dsh.profile.bundles; update later with git pull && pnpm run build inside the clone.

Option 4: tarball (no build approval needed)

Each release carries a prebuilt tarball on GitHub; install straight from the download URL:

npx -y @deepseek-ai/dsh plugin --profile web add https://github.com/Dudo-m/dsh-models-import/releases/download/v0.1.0/dsh-models-import-0.1.0.tgz

You can also pnpm pack inside a clone and install from the local path.

Start / verify

npx -y @deepseek-ai/dsh web

Open http://127.0.0.1:3080 → Settings: seeing both Models and Models Pro means the plugin loaded (restart a running server once so the new plugin loads).

Uninstall

npx -y @deepseek-ai/dsh plugin --profile web remove dsh-models-import

The Models Pro section goes with it; the stock Models section was never touched. Model entries and capability fields already written to settings.yaml are ordinary llm-pi-ai configuration — they keep working (the stock editor just shows them as unknown fields).

Use

1. Settings → Models Pro → edit a custom provider (or create one);
2. click "Fetch available models" — the request first takes this plugin's capability-aware route, falling back to stock discovery (with a note) when it cannot answer;
3. pick models in the dialog; badges show what the endpoint reported (thinking / vision / tools / search);
4. expand any row to edit: the thinking toggle, thinking levels (off…max), image input, display name, context window, max output;
5. saving writes llm-pi-ai.providers.<route>.models in settings.yaml; the next request picks it up (no restart).

Releasing (maintainer)

Publishing to npm is what gives users the zero-approval install. Automated flow:

1. add NPM_TOKEN under the GitHub repo's Settings → Secrets → Actions (from npmjs.com → Access Tokens → Automation);
2. cut a release:

npm version patch      # or minor / major
git push --follow-tags

The tag triggers .github/workflows/publish.yml: build, publish to npm, and attach the tarball to the GitHub release.

A one-off manual publish works too (prepare builds automatically):

npm login
npm publish

If the npm name dsh-models-import is taken, rename the package to @dudo-m/dsh-models-import (scope = your npm username) and adjust the publish/install commands accordingly.

Development

pnpm install
pnpm run build      # emits lib/index.js (host half) and lib/client.js (browser half)
pnpm run watch
pnpm run typecheck

Layout:

src/index.ts                      # host half: the /plugins/models-import HTTP route
src/mapping.ts                    # capabilities → llm-pi-ai field mapping (pure)
src/client/index.ts               # browser half entry (re-exports the settings page)
src/client/settings-models/       # vendored, extended copy of the stock Models page
src/client/helpers.ts             # capability-field read/write helpers

After a DSH upgrade that changes the stock Models page, refresh the vendored copy (src/client/settings-models/ mirrors @deepseek-ai/dsh-client-ui-settings-models's src/client); the integration points are ModelListEditor.tsx and locales.ts.

License

MIT — see LICENSE.

src/client/settings-models/ is a vendored, modified copy of the browser half of @deepseek-ai/dsh-client-ui-settings-models from DeepSeek's deepseek-harness (MIT); file headers and the third-party notice in LICENSE retain the upstream attribution.