Jensen-Yao/dsh-model-palette
Global provider-aware model command palette for the DeepSeek Harness Web UI: Alt+M fuzzy search, provider filter and favourites.
Replaces the native model selector with a global, provider-aware command palette opened with Alt+M from anywhere in the Web UI or via the composer trigger. It offers a unified fuzzy search over model names/IDs and provider names/IDs, a provider filter rail with live model counts, favorites and recents persisted in localStorage, and a universal reasoning-effort selector (off, minimal, low, medium, high, xhigh, max) that can add a live reasoningEfforts declaration for models that do not advertise levels. A configuration panel adds/edits/removes provider profiles — visual provider card catalog, template gallery, base URL/protocol/credential config, connection testing via llm.discoverModels, OpenRouter free-:free catalog discovery and auto-sync, protocol probing and collision-safe -completions splitting, and per-provider/per-model retry rules. It bundles 469 model presets (v4) with online refresh and auto-fill, and optional OpenRouter media tools (five agent tools for image/video generation, status polling and download) that reuse the stored OPENROUTER_API_KEY. It also includes a loopback-only B.AI relay for direct-connection recovery without a VPN.
Install
dsh plugin --profile web add github:Jensen-Yao/dsh-model-palette#v0.12.0Copied verbatim from the README's '🚀 Quick Start > Install' section (pinned to the v0.12.0 tag); the README adds 'Restart dsh web, then press Alt+M or click the model trigger.' Caveat: the npm package name dsh-model-palette is not published — registry.npmjs.org returns 404 for dsh-model-palette (checked 2026-10-06) — which is why the README installs from GitHub; prebuilt lib/ files are committed so a GitHub install needs no build scripts.
Compatibility
Node.js >=22.19 and DeepSeek Harness @deepseek-ai/dsh >=0.1.2-alpha.3 (README Requirements); the README notes v0.11.0 targets @deepseek-ai/dsh@0.1.6-alpha.2. DeepSeek Harness Web: a client React palette plus a server-side plugin (src/index.js) registered through cordis.patch.yml (row id: dsh-model-palette); pnpm >=11.21 is required for development only.
Details
- Repo: Jensen-Yao/dsh-model-palette
- Category: Models & Providers
- Stars: 1
- Version: GitHub Jensen-Yao/dsh-model-palette — current release v0.12.0 (install pins #v0.12.0); the npm name dsh-model-palette is not published (registry.npmjs.org 404, 2026-10-06); MIT (LICENSE) © Jensen Yao.
FAQ
- What are the requirements?
- Node.js >=22.19 and DeepSeek Harness @deepseek-ai/dsh >=0.1.2-alpha.3 (pnpm >=11.21 only for development). v0.11.0 targeted @deepseek-ai/dsh@0.1.6-alpha.2.
- Are the OpenRouter media tools enabled by default?
- Yes — they are enabled by default with zero configuration, reusing the OPENROUTER_API_KEY already stored in DSH credentials and writing to ~/.dsh/media-output. Free endpoints run automatically; paid generation asks for per-run confirmation unless allowed in config. They can be turned off via the openrouterMedia.enabled: false patch.
- Can I switch reasoning effort per model?
- Yes. The footer offers off through max; choosing a level for a model that does not advertise levels creates a live reasoningEfforts declaration before selection, and every level/provider wire value can be configured per model.
Alternatives
ringoage/dsh-subagent-model-picker · BruceLanLan/dsh-tier-router · jiay98528-dev/dsh-model-sync