hermes-workspace/playground-ws-worker

Hermes Playground multiplayer hub (Cloudflare Worker)

Drop-in port of scripts/playground-ws.mjs to Cloudflare Workers + Durable Objects. Free tier covers a hackathon and well beyond. Zero cold starts. Edge-deployed.

Why CF + DO over Fly.io / Render / Railway

• Free tier: 100k req/day on Workers, 1M+ DO req/mo. WebSocket connections count per-message, not per-connection — a presence broadcast every 200 ms across 20 players is well under the limit.
• Zero cold starts (vs Fly.io free tier which idles VMs after inactivity, and Render free which has a 30-50s cold start that kills demos).
• One Durable Object instance is the canonical "lobby/room" pattern — strong consistency, no Redis needed.
• Globally edge-deployed: a player in Tokyo and a player in NYC both connect to the closest edge, then route to the single DO holding game state.

Files

• src/worker.ts — entry + PlaygroundHub Durable Object class
• wrangler.toml — DO binding + migration
• package.json / tsconfig.json — build deps

Endpoints

• GET /playground — WebSocket upgrade (presence + chat fan-out, mirrors the Node sidecar protocol)
• GET /stats — JSON { online, byWorld, peakToday, peakDay, ts } for the HUD badge
• GET /health — JSON { ok: true, online, ts }

Deploy (~10 min, requires Eric's CF account)

cd playground-ws-worker
pnpm install                 # installs wrangler
pnpm wrangler login          # one-time browser auth
pnpm deploy                  # publishes to <name>.<your-subdomain>.workers.dev

Then in workspace .env.production:

VITE_PLAYGROUND_WS_URL=wss://hermes-playground-ws.<your-subdomain>.workers.dev/playground
VITE_PLAYGROUND_STATS_URL=https://hermes-playground-ws.<your-subdomain>.workers.dev/stats

(Custom domain optional via Workers Routes — wss://hub.hermes-playground.app.)

Local dev

pnpm wrangler dev            # hot-reload on http://localhost:8787
# In hermes-workspace:
VITE_PLAYGROUND_WS_URL=ws://localhost:8787/playground pnpm dev

Protocol parity

Wire format is identical to scripts/playground-ws.mjs. The client (src/screens/playground/hooks/use-playground-multiplayer.ts) connects unchanged.

Messages:

• { kind: 'presence', id, x, y, z, yaw, name, color, worldId, avatar?, ... }
• { kind: 'chat', id, text, ts }
• { kind: 'leave', id }
• Server-emitted: { kind: 'hello', server, ts }

State model

• presence: Map<id, lastWire> — fan-out + bootstrap on connect
• chatRing: array<chat> — last 50 messages, replayed to newcomers
• peakToday — persisted in DO storage for stats endpoint

Stale presence is pruned every 1 s via DO alarms (cheaper than setInterval, and survives instance hibernation).

Cost ceiling

For 100 concurrent players sending presence at 5 Hz:

• 100 × 5 × 60 × 60 × 24 = 43.2M msgs/day → still inside free tier for outbound (Workers count requests, not WS messages). If usage explodes:
  • Workers Paid: $5/mo for 10M req/day baseline.
  • DO storage: trivial (presence in memory, peak in storage).

Hardening (post-hackathon)

• Add rate limiting per playerId (token bucket in DO state).
• Multi-room: route by ?room= to idFromName(roomId).
• Anti-spoof: require signed JWT for presence.id.
• Replace presence ring with state.storage.transaction for crash recovery.