Port generation to OpenUI Cloud

[devin-ai-integration]

Summary

AppLess no longer talks to Cerebras directly. Screens are generated through OpenUI Cloud (https://api.thesys.dev/v1/embed/chat/completions) with the app's own native component library handed over as a managed chatLibrary config block, so the Cloud generates the system prompt (generateSystemPrompt from @openuidev/thesys-server) and validates the streamed openui-lang against that same library, while routing the model call to Cerebras only. No behaviour beyond the provider swap changes: same contract, same renderer, same local tool loop, no new src modules and no test changes.

Request shape:

POST https://api.thesys.dev/v1/embed/chat/completions   // Authorization: Bearer <thesys key>
{
  model: "google/gemma-4-31b-it",
  messages: [{ role: "system", content: SYSTEM_PROMPT /* ]]>openui:config {…} + tools/date prose */ }, …],
  stream: true,
  provider: { only: ["cerebras"], allow_fallbacks: false },
  metadata: { thesys: '{"c1_openrouter_provider":{"only":["cerebras"],"allow_fallbacks":false}}' },
}

Routing travels twice on purpose: with a managed config block the Cloud builds its metadata from the config and honours the request-level provider field, while metadata.thesys is the documented form for plain embed requests — both agree, so either path routes Cerebras-only.

@openuidev/thesys-server is Node-only, so prompt generation stays a build step: scripts/generate-prompt.mjs replaces scripts/embed-prompt.mjs (which required a sibling checkout of the web repo to copy a prompt from). It bundles the native contract with esbuild, serialises it to a chatLibrary (library.toJSONSchema() + per-component descriptions + componentGroups), passes the AppLess-specific prose from the old prompt as systemPromptOptions, and embeds the result:

]]>openui:config
{"chatLibrary":{"schema":…,"root":"Card","componentGroups":…,"id":"appless-native-0.1.0"},"systemPromptOptions":{"preamble":…,"additionalRules":…,"examples":…}}

src/genos/ui/contract.tsx is now the single source for both the renderer and the wire library, so npm run generate:prompt is the only step when components change.

Responses arrive wrapped in the Cloud's inline section markers, so cleanLang() strips them before parsing (stripInlineMarkers() in src/genos/store.ts) — including a marker split across chunk boundaries, since a trailing ]]>openui… fragment is never valid openui-lang:

]]>openui:content?thesys=true
```openui-lang
root = Card([…])

]]>openui:end

The local tool loop is untouched: the Cloud passes `tools`/`tool_calls` turns through unwrapped, so `web_search` still runs on-device.

BYOK is unchanged in shape, only the provider: `EXPO_PUBLIC_THESYS_API_KEY` / device-stored Thesys key (`genos.thesys-key`) instead of the Cerebras key, and the first-launch gate points at the Thesys console.

## Verification

- `npm run typecheck`, `npm test` (existing 22 tests, unchanged).
- Replayed the generated config block through muse's gateway parser and validator: config parses with zero `validateChatLibrary` issues, the trailing prose survives as instructions, and muse's `buildChatLibrarySystemPrompt` renders a 27k-char prompt containing the AppLess preamble, rules and examples.
- A live request to the endpoint returned `ERR_BILLING_THRESHOLD_EXCEEDED` (org billing suspended), so end-to-end streaming against production is still unverified.


Link to Devin session: https://app.devin.ai/sessions/35baf7b209b041ab83ae2917408cbb33
Open in Devin Desktop: https://app.devin.ai/desktop/session/35baf7b209b041ab83ae2917408cbb33?variant=devin

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring