Draft: Desktop message publishes intermittently dropped while typing indicators deliver
Component: Buzz Desktop (v0.5.3), possibly relay interaction
Severity: message loss (silent)
Summary
Messages composed and "sent" in Desktop sometimes never arrive at the relay: subscribers (agent harnesses with live subscriptions, plus buzz messages get polling the same relay) receive the sender's kind-20002 typing indicators but the kind-9 message itself never appears. Observed 7 times in one evening session on a self-hosted deployment; the sender's client showed the message as sent.
Evidence pattern
- Sender types: bursts of kind-20002 events arrive at subscribers every ~3s (so the client↔relay socket is demonstrably up in that window).
- The final kind-9 publish never arrives; polling the relay directly for the channel shows no such event.
- No error surfaced to the sender.
Expected
Either the kind-9 lands on the relay, or the client surfaces a failed-publish state (retry/pending indicator). Silent loss with a healthy-looking socket is the worst case.
Ask
If the client has publish-ack handling (relay OK/NOTICE), a visible failed/pending state and automatic retry would eliminate the silent variant. Happy to provide timestamps/event-id gaps from our relay logs.
Draft: Desktop message publishes intermittently dropped while typing indicators deliver
Component: Buzz Desktop (v0.5.3), possibly relay interaction
Severity: message loss (silent)
Summary
Messages composed and "sent" in Desktop sometimes never arrive at the relay: subscribers (agent harnesses with live subscriptions, plus
buzz messages getpolling the same relay) receive the sender's kind-20002 typing indicators but the kind-9 message itself never appears. Observed 7 times in one evening session on a self-hosted deployment; the sender's client showed the message as sent.Evidence pattern
Expected
Either the kind-9 lands on the relay, or the client surfaces a failed-publish state (retry/pending indicator). Silent loss with a healthy-looking socket is the worst case.
Ask
If the client has publish-ack handling (relay OK/NOTICE), a visible failed/pending state and automatic retry would eliminate the silent variant. Happy to provide timestamps/event-id gaps from our relay logs.