Environment Windows 11 x64, pt-BR; Desktop/CLI 1.1.10; Host 1.1.11; bundled Claude Code 2.1.220; standalone Desktop. Impact The Host crashes and the supervisor restarts it. Around recovery, collaboration rooms flap: agents/artifacts disappear from the GUI, provider visibility becomes unstable, and work is interrupted. Local chat/artifact files remain intact. Evidence Collected on 2026-08-23 at approximately 11:16 UTC-4.

traycer host status

retains 8 crash markers from 2026-08-21T02:19Z to 2026-08-23T02:35Z: 7 exited with 3221226505 / 0xC0000409; 1 with code 1. One crash tail reached survivorCount=17. The latest crash (2026-08-23T02:35:01.338Z) auto-recovered. Its stderr ends with 3 Claude probe children surviving teardown: 2 were force-killed; 1 did not exit after SIGKILL. [ephemeral-probe] ... survived query teardown; force-killed with SIGKILL [ephemeral-probe] ... did not exit after SIGKILL Important: this is temporal correlation only. One 0xC0000409 marker had no probe text, so I am not claiming that probes caused the crash or that this proves OOM. Aug 23, Host log local time (UTC-4): - 1,045 log entries containing getTaskCollabTokens -> 401 UNAUTHORIZED (01:36–09:13) - 29 WebSocket 503 occurrences (00:20–10:57) - 519 "stayed disconnected; rebuilding provider" warnings (00:02–11:08), across artifact, conversation, and notification rooms - 26 Claude probe-survivor entries; 12 still did not exit after SIGKILL (10:09–11:10) Expected Probe timeouts and transient cloud/auth errors should be bounded, self-heal, and never crash the Host. Questions 1. Is this known on Host 1.1.11 / Claude 2.1.220? 2. Can you identify whether the fail-fast is V8/Bun/native teardown, or request a dump/bundle? 3. Is there a fixed build and a safe recovery path that does not interrupt active sessions? Four in-app report attempts were rejected with HTTP 429. I can send their IDs and a sanitized bundle privately.

Thanks for the report, cc: @rapid-van-6755 @proud-printer-9482