mirror of
https://github.com/vitorpamplona/amethyst.git
synced 2026-10-29 23:27:44 +00:00
Vitor PamplonaandClaude Opus 5.5
f335b8722d
KtorRelayFrameLimitTest.aFrameOverTheLimitClosesTheConnection failed on CI (close code null instead of 1009). Most of it was the relay, not the test. When Ktor refuses a frame it queues the close (1009, or 1002 for a protocol violation) and ends the incoming stream with FrameTooBigException. WebSocketSessionPump let that escape the handler, so Ktor logged the handler as failed and dropped the socket without waiting for the queued close: the client saw a bare EOF. The pump now returns normally on those two exceptions, and Ktor closes the session and waits for it. The rest was the test's client. OkHttp writes the whole 128 KB message, so the relay closes with the payload unread and the reset can reach the client before the 1009, which OkHttp reports as a broken pipe. The test now sends the frame header alone (claiming 2 GB, as the class comment describes) over a raw socket and reads the close frame. Linux (Docker, --cpus=0.5), 600 runs each: - without the pump fix: raw test 15 failures (EOF, no 1009); OkHttp 15 EOF and 1 broken pipe - with it: raw test 600 passes; OkHttp 600 closes with 1009 macOS never reproduced it (0 of 300). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>