Full walkthrough

Step 1: Grow from two users to many: the poll() event loop

The intermediate chat used one thread for one peer. Ten clients cannot each get a blocking thread cheaply here, so the server uses poll() from <poll.h>: one array of pollfd (one per socket), and poll() sleeps until SOME socket is readable. One thread, many clients, zero busy-waiting — the event-driven pattern behind real servers. The starter's struct Client { int fd; std::string name; std::string room; } is the per-connection state poll cannot give you.

Step 2: Accept and register clients

The listening socket sits in the poll array too; when IT is readable, accept() a new client, add their fd to the array and a Client to the vector, and prompt for a nickname. A nickname collision gets a polite refusal or a suffix — decide, because broadcast() will identify people by this name.

Step 3: broadcast(): fan-out with an exclusion

broadcast(msg, room, excludeFd) loops the clients in that room and send()s to everyone except the sender (clients echo their own messages locally). Centralising fan-out in one function means join/leave announcements, chat lines, and system notices all travel the same tested path.

void broadcast(const std::string& msg, const std::string& room, int excludeFd) {
    for (auto& c : clients)
        if (c.room == room && c.fd != excludeFd)
            send(c.fd, msg.c_str(), msg.size(), 0);
}

Step 4: Rooms and commands: /join, /msg, /list, /quit

Lines starting with '/' are commands, everything else is chat. /join #room moves the Client's room field and triggers leave/join announcements in both rooms. /msg name text finds ONE client by nickname and sends privately. /list prints the room's nicknames. Commands are parsed on the first token — the mini shell's splitArgs instinct, reused on a socket.

Step 5: Handle disconnects without poisoning the loop

recv() returning 0 (or a send failing with EPIPE) means that client is gone: close the fd, REMOVE the Client from the vector and its pollfd from the array, and announce the departure. A dead fd left in the poll set makes poll() return instantly forever — the server spins at 100% CPU mourning a ghost. Removal mid-iteration is the fiddly part; mark-and-sweep after the event pass is the safe shape.

Step 6: UDP mode: the stretch goal with honest trade-offs

The UDP variant replaces connections with datagrams: no accept, no per-client socket — one socket, recvfrom gives you the sender's address each time, and 'clients' are addresses you have seen. UDP can drop, duplicate, or reorder messages, so the starter adds sequence numbers per sender and acks for messages that matter. Run both modes side by side and FEEL the difference: TCP hides the network's chaos; UDP hands it to you in a bucket.

Step 7: Test like a crowd

Open four terminals: three clients (two joining #general, one in #random) and the server. Verify room isolation (the #random user must NOT see #general traffic), a /msg that reaches exactly one person, /list accuracy, and a mid-conversation quit that announces itself and leaves the server responsive. Every one of those is a feature in the checklist, tested by acting, not by reading code.