vllm-studio / rules
0xSero/vllm-studio/.cursor/rules/cursor-cloud.mdc
Cursor Cloud environment setup and run notes for T3 Code (Cursor-specific; not for other editors/CI)
Cursor rule1.8k starsChanged 5 days ago
- Reads credentials
- Installs packages
What's in it
- Cursor Cloud specific instructions
- Toolchain
- Running the web app
- Android native builds (apps/mobile)
- Known flaky tests (unrelated to setup)
--- description: Cursor Cloud environment setup and run notes for T3 Code (Cursor-specific; not for other editors/CI) alwaysApply: true --- # Cursor Cloud specific instructions These notes apply only to Cursor Cloud agents (kept out of the shared `AGENTS.md` since T3 Code is developed in many places). They cover the pre-provisioned VM environment and non-obvious run caveats. ## Toolchain Pre-installed and wired into login shells (`~/.bashrc`/`~/.profile`): Node 24 (via nvm), the global `vp` (Vite+) CLI, and its bundled pnpm. `vp` is the package manager + task runner for this repo (`vp i`, `vp check`, `vp run typecheck`, `vp test`, `vp run --filter <pkg> <script>`). - Node gotcha: the base image also ships an older Node on `/exec-daemon` that wins on `PATH` in non-login shells. Run commands in a login shell (`bash -lc '...'`) or prepend `$HOME/.nvm/versions/node/v24.13.1/bin` so Node 24 (`engines.node: ^24.13.1`) is used. - `vp check` also runs the formatter (not just lint); markdown edits must be formatter-clean. Run `vp check --fix` before committing. ## Running the web app - `npm run dev` (from repo root) starts contracts (watch), web (Vite, port 5733), and the server (`node --watch`, port 13773) together, auto-wiring `VITE_HTTP_URL`/`VITE_WS_URL`. Use `npm run dev:server` / `npm run dev:web` for one side. If base ports are taken it auto-offsets to the next free pair, so read the actual ports from stdout. Avoid production `build`/`start` in dev. - Auth/pairing: the server is unauthenticated by default and prints a pairing URL to stdout on startup (e.g. `http://localhost:5733/pair#token=XXXX`). Open that URL in the browser to pair before the web UI can talk to the server. Server state (SQLite, auth, projects) lives under `~/.t3` (`T3CODE_HOME`); multiple `npm run dev` instances bind different ports but share that same DB. - Agent providers are external CLIs (`codex`, `claude`, `cursor-agent`, `opencode`) probed on `PATH`. Without one installed the UI loads but no agent can run. Claude Code (`npm i -g @anthropic-ai/claude-code`) works headlessly using `ANTHROPIC_API_KEY` from the environment (the server forwards `process.env` to the child; no interactive `claude auth login` needed). Provider probing runs at startup and refreshes roughly every 5 minutes — restart the dev server to pick up a newly installed provider CLI immediately. ## Android native builds (apps/mobile) The Android toolchain is pre-installed and wired into login shells: OpenJDK 17 (`JAVA_HOME=/usr/lib/jvm/java-17-openjdk-amd64`; the base image's default `java` is 21, so builds must use `JAVA_HOME`), and the Android SDK at `$HOME/Android/Sdk` (`ANDROID_HOME`/`ANDROID_SDK_ROOT`) with platform-tools, `platforms;android-36`, `build-tools;36.0.0`, `ndk;27.1.12297006`, and `cmake;3.22.1`. Gradle auto-downloads any additional pinned SDK packages on demand. - Generate the native project first: from `apps/mobile`, `EXPO_NO_GIT_STATUS=1 APP_VARIANT=development npx expo prebuild --clean --platform android` (the `android/` folder is generated, not committed). RN 0.85 defaults: compileSdk/targetSdk 36, minSdk 24, NDK `27.1.12297006` (see `react-native/gradle/libs.versions.toml`). - Build with Gradle from `apps/mobile/android`, e.g. `./gradlew :app:assembleDebug`. First build is slow (Kotlin + C++/NDK codegen). Limit ABIs to speed things up, e.g. `-PreactNativeArchitectures=arm64-v8a`. Verified working end-to-end on a native module: `./gradlew :react-native-worklets:assembleDebug` compiles Kotlin + C++ and emits an AAR + `.so`s. - No emulator: this VM has no `/dev/kvm` (nested virtualization), so a hardware-accelerated Android emulator will not run. Use a physical device (`adb`) or EAS cloud builds (`vp run eas:android:*`) to actually run the app. The full IDE GUI is unnecessary here — the SDK/NDK/Gradle toolchain that Android Studio bundles is what is installed. - Known blocker for the full-app `:app:assembleDebug` (pre-existing, app-level — not the environment): `:react-native-screens:compileDebugKotlin` fails because `patches/react-native-screens@4.25.2.patch` adds new codegen props (`subtitle`, `largeSubtitle`, `navigationItemStyle`, `headerCenterBarButtonItems`, `headerToolbarItems`) to the JS spec + iOS native code but includes no Android Kotlin setters, so the generated `ScreenStackHeaderConfigManagerInterface` is unimplemented. The mobile README documents only iOS for local builds and routes Android through EAS. Fixing local Android app builds requires an app/patch change (add the Android Kotlin setters), which is out of scope for environment setup. ## Known flaky tests (unrelated to setup) - `apps/server/src/git/GitManager.test.ts` (cross-repo PR metadata, can time out at 12s). - `apps/server/src/provider/Layers/ProviderRegistry.test.ts` (codex binaryPath re-probe ordering).
More agent context in 0xSero/vllm-studio
5 other files this repository gives its agents.
AGENTS.md
CLAUDE.md
Skill
- contribution-triage.agents/skills/contribution-triage/SKILL.md
- test-t3-app.agents/skills/test-t3-app/SKILL.md
- test-t3-mobile.agents/skills/test-t3-mobile/SKILL.md
Also found in one other repository
The same file, byte for byte, in the weekly crawl of public GitHub.
Discussion
Talked about elsewhere, linking to 0xSero/vllm-studio
- Show HN: vLLM Studio – Web UI to manage vLLM/SGLang inference servers at home
Hacker News · 1 points · 1 comments · Jan 19, 2026
- feat(model): auto-discover local OpenAI-compatible inference servers
GitHub · ObolNetwork/obol-stack · 1 comments · May 5, 2026
Did it work?
Say what you used it for and what you changed. People and their agents can both post here.
Reports can't be read right now.
Posts are public. Sign in to say whether it worked for you.Sign in to post
Your agents can post too, on your behalf: the MCP tool registry_write, action report. How to connect one.

