Pigeon Problems
A 1–4 player co-op rooftop platformer in the browser. Pigeons haul oversized cargo, throw and catch it, on a server-authoritative 60 Hz simulation.
- 60 Hzserver-authoritative simulation
- 300 msinjected lag, still playable
- 127engine/session tests + browser checks
The game
A very professional delivery company, staffed entirely by pigeons. 1–4 players create a private room by link, carry opposite ends of oversized cargo, throw and catch it, and gust each other across rooftops. 5 deliveries and 6 platforming courses (50 platforms), with procedural 3D models, generated textures, movement-driven animation and synthesized sound. No imported assets.
Engineering
- A Three.js/WebGL client with a Cloudflare Workers backend that runs each room as its own Durable Object and persists room state. A 1–4 player game playable from a link, no account or install.
- Server-authoritative netcode: a fixed 60 Hz simulation, 20 Hz state snapshots and 30 Hz client input, with client-side prediction and sequence-based reconciliation. Movement and checkpoint recovery verified under 300 ms of injected round-trip latency.
- One JavaScript physics and level engine shared across the browser, a local Node server and Cloudflare, including a custom constrained-load solver for cargo carried by one or two players, and a throw preview that predicts impact up to ~3 s ahead using the same rules as gameplay.
- Hardened sessions: secret reconnect tickets kept out of broadcasts, input-sequence and duplicate-action checks, rate and payload limits, and a 2-minute reserved seat on disconnect. Optional WebRTC voice with room-scoped signaling.
- Geometry batching kept local 1440×900 Chrome samples at ~60 fps in 100–141 draw calls.
- Keyboard and mouse with remappable keys, touch controls and the standard Gamepad API, plus reduced-motion settings and HUD layout checks down to a 320 px screen.
Verification
- 62 automated engine and session tests, 65 scripted Playwright browser checks, and complete runs of all 6 courses with 2 and 4 real WebSocket clients on the public deployment.
- Client and real-time backend deploy independently, with saved release versions and a documented rollback path.
How it was made
I directed the concept, the cooperative mechanics, visual and movement revisions, level difficulty and review; OpenAI Codex assisted with code, procedural art, tests and docs.