Set up Cursor Cloud dev environment (AGENTS.md notes)

[vijit-lark]

What

Provisions and verifies a full development environment for openlily (the Pipecat voice assistant under server/) on the Cursor Cloud VM, and records durable, non-obvious startup/run guidance for future agents.

The only committed change is an AGENTS.md addition (## Cursor Cloud specific instructions). Environment provisioning is handled by the update script and the VM snapshot; .env/brains.yaml/.venv are git-ignored and not committed.

Environment

• Installed uv + Python 3.11 (pinned by server/.python-version) and PortAudio system libs (for PyAudio).
• uv sync for the server/ project with the local, web, and email extras. local-models is intentionally excluded (its mlx-whisper runtime is Apple-Silicon only).
• Update script: uv sync --project server --extra local --extra web --extra email (idempotent).

Verified

Task	Command	Result
Lint	uv run ruff check .	Runs; 3 pre-existing findings (UP035/UP017) in committed code
Type-check	uv run pyright	Runs; 5 pre-existing reportArgumentType in agent.py
Tests	uv run pytest	79 passed
Run app	uv run bot.py --mode webrtc	Server up on :7860; browser client connects to Client/Agent READY

Hello-world (running app, end to end)

Opened the Pipecat client at http://localhost:7860/client/, clicked Connect, and the WebRTC session reached Client: READY / Agent: READY. Server logs show the full cascade pipeline assembling (SmallWebRTCInputTransport → CartesiaSTT → LLMUserAggregator → OpenAIResponsesLLM → CartesiaTTS → … → LLMAssistantAggregator), Silero VAD + Smart Turn v3 loading, Client connected, and Pipeline ready; played readiness chime. With placeholder keys, the pipeline then hits HTTP 401 on Cartesia/OpenAI — which confirms the wiring reaches the external APIs; a real spoken/text reply needs valid provider keys.

openlily_webrtc_client_connects_ready.mp4

Notes for future agents (in AGENTS.md)

• No audio device on the VM → local/local-with-wake-word can't run headless; use --mode webrtc.
• cli.py uses load_dotenv(override=True), so a server/.env overrides real env-var Secrets — don't create a .env that sets keys you provide as Secrets.
• Provider keys gate the conversation (default brain needs OPENAI_API_KEY + CARTESIA_API_KEY), not server startup.

_{To show artifacts inline, enable in settings.}