What changed in Buzz in the last 24 hours — a wall of relay-reachability bugs on 0.5.5, agents that are not community-scoped, and the token-burn question

Still on v0.5.5, no new release. Around 15 fresh open issues overnight and the pile now sits near 905 open.

The loud cluster today is relay reachability and all of it is on 0.5.5. Desktop cannot reach a relay over TLS on #5157. A fresh install grabs the local dev relay and hangs on connecting to relay with no error on #5141. A hosted community relay is unreachable and every agent fails to connect on #5144. Plain DNS failure on #5121. Signup confirmation email never arrives on Windows on #5125. None of these are touched by the 0.5.5 reconnect fix. That one was desktop session recovery, not first dial and not hosted DNS.

The one to read in full is #5128. The Agents tab is not scoped to the active community. A builtin agent gets a separate key per machine and community, so one persona shows up as several identities the UI cannot tell apart. The Auto-start toggle writes to whichever instance the profile preselected, and it keeps preselecting the one with zero channels in the community you are looking at. Two agents silently failed to come back the night before because the flag landed on the wrong identity. Removing an agent from a channel does not stick across a community switch either. The reporter's own fix is the cheap one. Give builtin agents the single-key-per-machine model the custom agents already use.

BYO side did not get better. Custom ACP harnesses still do not onboard cleanly. Empty model dropdown, the config file corrupts and duplicates entries when you change harness, the agent dies a few seconds after the bridge errors on #5155 and #5156. Workflow call_webhook fires into the void with no run record on #5122. buzz dms list comes back empty for a live DM on #5137.

And the question that keeps coming up. Does Buzz burn 2x the tokens of Claude Code direct? Short version from people running it, most of the early burn is agents running their own test loops plus the large base prompt, and tight team instructions cut it hard. Longer breakdown in the token-usage thread.