English | 简体中文
The UDI framework treats the agent as one way for ordinary people to shape an app's presentation, supported interactions, and information handling. Direct controls are another path to the same persistent experience. The framework applies across app categories.
The current research demo works without a model. Its hosted assistant recognizes a bounded command grammar; this is not general AI understanding. The optional development bridge below lets a locally installed Ollama model translate natural-language edits into the same bounded operation vocabulary.
The Agent tab is available on desktop and phone. Chat keeps the last 40 conversation messages locally and prepares supported English/Chinese commands for review. Controls exposes layout, experiment filters, visible views, connected data sources, and requested capability grants in a settings layout. Both use the same controller and exact-state guard; changes apply only after Apply change.
Examples: Use Atlas lab, Revoke experiment access, Grant experiment access, 只显示失败实验, and 撤销实验访问权限. Binding commands can use an imported source name or ID. Grants are per instance, and notes.write still requires a separate exact-note confirmation before a write. Undo restores a definition while preserving current bindings, grants, and notes. Unsupported requests do not change state. Clear the local conversation from the Agent panel. The hosted mode is labeled Offline command assistant, with no LLM.
Proposal summaries and human-readable access labels appear first; operation JSON is under Technical details. Use Personalize on desktop to open Agent. On either device, Share this interface sends the resulting definition through a link, and a recipient previews before choosing Use and their own data access. Sharing does not include the conversation. Create my workspace in Data provides a simple first-entry form; file import remains in Advanced data tools. The agent cannot connect music accounts, build actual playlists, run background workflows, or publish posts autonomously. See sharing and data.
Install and start Ollama, then install a model that supports structured JSON outputs and record its exact local model name. Model selection depends on your hardware; this project does not bundle or endorse a specific model.
In one terminal, set UDI_OLLAMA_MODEL to that installed name and run:
npm run agentPowerShell example, replacing the quoted name with your installed model:
$env:UDI_OLLAMA_MODEL = 'your-installed-model-name'
npm run agentIn a second terminal:
npm run dev -- --port 4178Select Local model · Ollama in the intent composer's planner selector. The bridge listens only on 127.0.0.1:4179; Vite proxies /api/plan to it. Ollama is expected on 127.0.0.1:11434. The static production build deliberately offers only the offline planner because it has no local proxy or hosted model service.
Browser → bridge: { intent, spec, revision }. No dataset, note, grant, credential, or instance-binding object is accepted. The bridge validates the definition before model invocation.
Bridge → Ollama: /api/chat, stream: false, the planner output schema in format, bounded output length, and a low-temperature setting. This follows Ollama's structured-output interface.
Model → bridge: { summary, operations }. The host rejects unknown fields and operations, computes capability additions, simulates the candidate, and validates the resulting definition. An empty operation list represents an unsupported request and is surfaced as an error with the model's explanation. The browser validates the returned operation document again against its original base definition.
No model output can execute a handler, grant access, read fixture records, supply a URL, or bypass the user's change review through this interface. A valid proposed change can still be wrong for the user's intent.
Unit tests exercise malformed model output, semantic failures, capability additions, request minimization, and failure propagation using an injected mock transport. Browser tests also exercise model mode with a fixture-backed HTTP response.
On October 17, 2026, two direct smoke cases and one browser → proxy → bridge → model → review → render case passed using the locally installed qwen3.5:9b-q8_0 model and Ollama 0.40.4. See raw smoke output, browser integration record, and development failures and observations. The generation request explicitly disables thinking and caps output. Recorded timings describe those calls only; the inference device was not measured. No broad quality benchmark or multi-turn study is claimed.
To repeat the two direct smoke cases after setting the environment variable, run npm run smoke:model. To repeat the browser integration, start both servers and run npx tsx scripts/capture-model-demo.ts (set PLAYWRIGHT_CHANNEL=msedge to use installed Edge on Windows).
To evaluate a real model, record the exact model identifier and digest, Ollama version, hardware, prompt/schema versions, every request/output, schema success, semantic task success, refusal behavior, latency, and human correction rate. The offline conformance cases are insufficient evidence for language-model performance.