Modernizes the project's tooling and brings every dependency up to the latest version across both the agent and the web frontend. Also includes a livekit.toml for deployment to livekit cloud.

Agent

Tooling → uv

• Replace requirements.txt and ruff.toml with a uv-managed pyproject.toml + committed uv.lock.
• Fold the old ruff config into [tool.ruff].

Model → LiveKit Inference

• Switch the vision (guessing) and judging LLM calls from the OpenAI SDK directly to LiveKit Inference (openai/gpt-4o-mini).
• Inference authenticates with the existing LIVEKIT_API_KEY / LIVEKIT_API_SECRET, so OPENAI_API_KEY is no longer required.
• Guesses/judgements now build a ChatContext with ImageContent and stream through a small _llm_complete() helper; JSON output uses response_format via extra_kwargs.

Deployment

• Deploy agent to LiveKit Cloud instead of fly.io
• Rewrite the Dockerfile as a multi-stage uv build (build toolchain stays out of the runtime image; runs as non-root appuser; ~456MB). Stop git-ignoring it so it's tracked.
• Update the README to use uv and drop the OpenAI key requirement.

Dependencies

• Latest LiveKit Agents SDK (livekit-agents 1.5.17), openai 2.41, and refreshed transitive deps.

Web frontend

Every dependency updated to latest, including major versions:

Migration work required by the major bumps:

• Next 16 / React 19 — move viewport/themeColor out of metadata into a dedicated viewport export; pin the Turbopack workspace root in next.config.mjs (avoids inferring an unrelated parent-dir lockfile). tsconfig.json updated by Next (jsx → react-jsx). No forwardRef/legacy-render changes needed.
• Tailwind v4 — delete tailwind.config.ts (v4 auto-detects sources), switch PostCSS to @tailwindcss/postcss, and use v4 layer imports in globals.css. Deliberately preserved the "no preflight" setup so Tailwind's reset doesn't clobber the 98.css retro styling (verified the built CSS contains the utilities but none of Tailwind's universal reset).
• ESLint 9 flat config — next lint was removed in Next 16, so .eslintrc was converted to eslint.config.mjs and the lint scripts now call the ESLint CLI. Pinned ESLint to 9 because eslint-config-next 16's bundled plugins require eslint <= 9 (10 would break linting). Added a .prettierignore so build output isn't format-checked.

Note on a lint rule

The React 19-era react-hooks plugin adds a set-state-in-effect rule that flags 5 intentional, SSR-safe patterns in the providers (initializing state from localStorage / window.location.hash in mount effects). Rewriting those is a behavior-risky refactor that's out of scope for a dependency bump, so the rule is set to warn (with an explanatory comment) rather than failing lint.

Testing

• Agent: uv sync, uv lock --locked, ruff check pass; modules import against the new SDK; Docker image builds and the runtime image smoke-tests clean. Live-tested LiveKit Inference end-to-end (vision guess + JSON winners).
• Web: pnpm build (TypeScript 6), pnpm lint (0 errors), and pnpm format:check all pass.
• End-to-end: booted the agent + frontend together; agent registers as a worker, the frontend SSRs the full app shell with Tailwind utilities and 98.css applied, and the token API returns 200. A full game round was played to confirm guessing and winner detection work.