@deepseek-ai/dsh-client-ui-desktop-customization

Desktop customization

English | 中文

Desktop-only browser plugin for learner-facing visual enhancement, background selection, the visible update center, and the Beyondata attribution entry. The package is mounted only when the Desktop Host exports DSH_DESKTOP=1; persistence and update operations cross the fixed Electron preload bridge. The attribution entry occupies the sidebar footer action seat above Settings, keeps its text in the wide column, and contracts to a tooltip-backed logo in the rail so it never overlays conversation controls. The update card labels the Studio shell version and embedded Harness core version separately.

The package ships five named background themes plus the image-free original UI: Whale Maid is the first-run default, while Cloud Cat, Jiutian Deep-Space Compute Observatory, Jiutian Quantum Glass Laboratory, and Jiutian Dawn Compute Horizon remain selectable. Their stable identifiers are persisted without duplicating bundled images in userData. The custom-background path still accepts PNG, JPEG, or WebP up to 16 MB, renders a 1920×1080 WebP locally, persists it under Electron userData, and applies ThemeRuntime token overrides. No selected image is uploaded.

The visual-enhancement Settings row and composer shortcut consume one Host-backed status source. When disabled, the shortcut first selects the built-in DeepSeek vision model for the session and future blank sessions and then activates its native route; when enabled, it disables the capability through the same Settings namespace. Exact model metadata selects one request path after later model changes: an image-capable primary model receives native image content, while a text-only model may use an already verified compatible sidecar. The built-in DeepSeek vision route normally sends deterministic request images through reusable Files API references and falls back to bounded all-inline images when file resolution fails. The setup dialog offers Bailian, OpenRouter, Ollama, vLLM, SGLang, and a generic OpenAI-compatible route. Bailian remains fixed to qwen3.8-max; OpenRouter defaults to openai/gpt-4.1-mini; every self-hosted route requires an explicit vision-model id and exposes an editable API base with presets for Ollama (127.0.0.1:11434/v1), vLLM (127.0.0.1:8000/v1), and SGLang (127.0.0.1:30000/v1). Local API keys are optional, while supplied keys are stored under application-owned credential references; ambient DASHSCOPE_API_KEY and OPENROUTER_API_KEY values remain read-only fallbacks. Compatible routes append /chat/completions, send OpenAI-compatible image parts without redirects, and persist the endpoint only after real-image verification succeeds. Host-pushed settings and credential updates refresh both entries together.

Model Experience

None, as this browser-side package only controls the Host-owned visual capability and registers no model-facing context itself.

KV Cache effect

The package itself adds no tokens or KV-cache entries; after this UI enables visual enhancement, the Host-owned capability governs all Skill, Tool, and visual-observation effects.

Known Limitations and Deferred Work

• The current learner-facing update card keeps online actions disabled; it reports both installed versions and directs learners to the release page.
• Signed installers, platform release metadata, and release publishing are deferred to the three-platform packaging phase.