writeTimeout is applied — Server.Builder.writeTimeout(Duration) was accepted and never read (chappe#13). It now bounds each blocking write of an HTTP/1.1 response: when the client stops reading and the socket refuses the bytes for longer than the timeout (30 s by default), Chappe closes the connection and the client gets a truncated body, where the connection’s virtual thread used to stay blocked with no bound. It is not an inactivity bound: a Server-Sent Events stream that stays silent for minutes between events is not affected. If a client of yours legitimately stalls reading for longer than 30 s, raise the value. Reference — timeouts, Migration — upgrading from 0.3.0.

The last response survives the close — when an HTTP/1.1 connection ends after a response (a 431, a 400, or Connection: close) while request bytes are still unread, Chappe used to close at once; the TCP stack then answered with a reset, and the reset destroyed the response before the client read it ("Connection reset"). Chappe now closes in stages, as RFC 9112 §9.6 describes: it half-closes, drains what the client still sends for at most 2 s or 1 MiB, then closes, so the response ends with a FIN (chappe#19). Over TLS it sends close_notify first and drains the ciphertext without decrypting it (chappe#21). A lost connection, a client that closed first, the idle timeout and a WebSocket takeover still close at once. Internals — threading model, HTTP capabilities.