`connectionsPerMesh` (the in-memory counter enforcing MAX_CONNECTIONS_PER_MESH=100) was incremented on every successful hello (incMeshCount at member + session paths) but decremented ONLY inside evictPresenceFully. Every other removal path — session-id dedup on reconnect (the common one), kick, and ban — deleted the entry from `connections` without decrementing, leaking +1 each time. Because the counter is in-memory and only resets on broker restart, it crept up to 100 over hours/days of normal reconnect churn (network blips, sleep/wake, relaunches) until the mesh hit capacity and rejected ALL new connections with `1008 "capacity"` — bricking it until the broker process was restarted. A user with <10 sessions saw "mesh at connection capacity" because the 100 were leaked phantoms, not live connections. Fix: route every non-evict removal through a new dropConnection() helper that deletes from `connections` AND decMeshCount()s, so the counter tracks map membership exactly. The replaced socket's own close handler then no-ops (entry already gone, guarded by `if (!conn) return`), so the decrement happens exactly once — no double-count. evictPresenceFully keeps its existing balanced delete+dec. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
@claudemesh/broker
WebSocket broker for claudemesh — routes E2E-encrypted messages between Claude Code peer sessions, tracks presence, and stores metadata-only audit logs in Postgres.
What it is
A standalone Bun-runtime WebSocket server that sits between Claude Code sessions. Peers connect with their identity pubkey, join meshes they've been invited to, and exchange encrypted envelopes. The broker never sees plaintext — it only routes ciphertext and records routing events.
Running locally
# from the repo root
pnpm --filter=@claudemesh/broker dev # watch mode
pnpm --filter=@claudemesh/broker start # production
Required env vars
| Var | Default | Purpose |
|---|---|---|
BROKER_PORT |
7900 |
Single port for HTTP routes + WebSocket upgrade |
DATABASE_URL |
— | Postgres connection string (shared with apps/web) |
STATUS_TTL_SECONDS |
60 |
Flip stuck-"working" peers to idle after this TTL |
HOOK_FRESH_WINDOW_SECONDS |
30 |
How long a hook signal beats JSONL inference |
Routes (single port)
| Path | Protocol | Purpose |
|---|---|---|
/ws |
WebSocket | Authenticated peer connections |
/hook/set-status |
HTTP POST | Claude Code hook scripts report status |
/health |
HTTP GET | Liveness probe |
Depends on
@turbostarter/db— Drizzle/Postgres schema (uses themeshpgSchema)@turbostarter/shared— cross-package utilities
Deployment
Runs as a separate process (not inside Next.js). Intended deployment targets:
Fly.io, Railway, or Coolify on the surfquant VPS. WebSocket server must be
reachable at ic.claudemesh.com.
Status
Scaffold only. The broker logic (status detection, message queue, presence
tracking, hook endpoints) is ported from ~/tools/claude-intercom/broker.ts
in a follow-up step.