dsh-generative-ui
 
Generative UI for DeepSeek Harness: the agent answers with a live React interface instead of prose. It streams — the component renders while the model is still typing it.
Two places it shows up:
- Inline — a fenced ```
ui4a/tsx``` block renders in place, between the paragraphs of the reply. Right for a chart, a form, a set of options to click, a calculation the reader will want to change a number in. - Canvas — a file at
ui4a/canvases/<id>.ui4a.tsxopens in a panel beside the conversation and stays there across turns. Right for a tool the user will come back to.
Generated code imports anything on npm (resolved from esm.sh at render time), shares the host's single React instance, and takes its colours from the app's own design tokens, so it follows the light/dark theme.
The ui4a in that fence is the harness this implements — UI for Agent, from Mind Lab: rather than coaxing an agent into a fixed UI schema, let it write ordinary frontend code and have the runtime enforce the boundaries. The reasoning, and the benchmarks behind it, are in UI4A: A Component-Native Harness for Generative UI. This package is that harness wired into dsh's web client.
Install
dsh plugin --profile web add dsh-generative-uidsh plugin forwards to the profile's package manager, so that installs the package. Mounting it also takes one line in ~/.dsh/profiles/web/package.json — the profile's bundle list is what dsh actually boots:
{
"dsh": {
"profile": {
"bundles": [
"@deepseek-ai/dsh-base",
"@deepseek-ai/dsh-web-app",
"dsh-generative-ui"
]
}
}
}Then restart dsh web — plugins are mounted at boot, and there is no hot-reload for adding one.
How it works
The package is one plugin with two halves, which is how dsh plugins reach the browser:
lib/index.js (node) | injects the system-prompt section, registers the generative-ui skill, serves the compiler wasm and canvas file reads over its own webServer routes |
lib/client.js (web) | claims ui4a/tsx code blocks in the transcript, compiles TSX in-browser, mounts the canvas panel |
The model is taught in two layers, split by what each costs:
src/prompt.tsrides in every request, so it carries only the trigger, the fence syntax, the canvas path, and the colour tokens.src/skill.tsis registered throughctx.skills.register()and loads only when the model reaches for it. It carries the judgement: whether the answer wants an interface at all, inline or canvas, how to frame and lay one out.
That split is measured, not assumed — see [CLAUDE.md](./CLAUDE.md) §4.5 for the 40-prompt evaluation behind it.
Development
bun install
bun run check # lint + typecheck + build + smokebun run build bundles both halves with Bun.build. [CLAUDE.md](./CLAUDE.md) is the design document — it records the host constraints this plugin was built against, all of them found the hard way.
License
MIT