Results & Impact

Where it stands

Huddl is live at huddl.gg, running two full games (Player Imposter and Wavelength) end to end: Google sign-in, room creation, guest joins, real-time lobby and gameplay over WebSockets, and reconnect handling, all served from Cloudflare Workers with Durable Objects backing each room. It shipped as a solo project in 85 commits over 17 days (Jul 1–17, 2026).

The backend and frontend test suites are both green: 203 tests across 14 files in game-service (including full Imposter and Wavelength flow tests through the lobby Durable Object, not just isolated unit tests) and 84 tests across 8 files in the frontend, 287 passing tests total. They run in CI on every PR before either app is allowed to deploy.

What I learned

  • A plugin contract only proves itself once something un-special uses it. Imposter alone couldn't tell me whether GameDefinition was a real abstraction or just "what Imposter happens to need." Wavelength stress-tested it with a real second game is what actually gave me confidence the lobby engine doesn't secretly know anything about any particular game.
  • StrictMode races are a preview of real double-tab races, not a development-only nuisance. Fixing the WebSocket store to survive StrictMode's mount/unmount/remount properly is the same fix as making it survive a user genuinely opening the same room in two tabs. Treating the dev-mode symptom as the whole problem would have shipped a real bug.
  • Hibernation changes what "keep this connection alive" has to mean. Reaching for an in-memory Map of live sockets is the natural first instinct and would have worked, right up until it silently pinned every Durable Object with an open connection in memory, permanently, which defeats the entire cost model Durable Objects are supposed to offer.

What I'd do next

In rough priority order:

  1. Ship Team Auction, the third game named in the README's launch lineup. Imposter and Wavelength shipped; this one hasn't started yet.
  2. Move sessions to JWT. Today the game Worker verifies a session cookie's HMAC and then looks it up in KV or D1. JWT-based sessions would remove the storage lookup entirely. The tradeoff is losing the ability to revoke a session, but for an app with non-critical data the reduced complexity seems worth it.
  3. Add real observability. Sentry catches unhandled exceptions today, but there's no metric for how many rooms are live, how long they last, or message latency. This is the same gap PulseFM had, and the same fix: wire it in from the start on the next project rather than as a follow-up.
  4. Formalize infra-as-code across both Workers. Each app's Cloudflare resources (D1, KV, the Worker itself) live in its own wrangler.jsonc. There's no single plan/apply step that provisions all of it together the way Terraform does for a multi-service system.
  5. Add lobby config. Right now rooms can only be configured for the game the host wants to play. I want to extend the configuration to allow for the changing of rules, lobby settings, etc
LinkedIn