You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Follow-ups from review of #121 (fixes for #120). The Vycari Review shadow round raised the first three; the responder confirmed them against c5644ad. None of them blocked the merge.
resync caps its topic list at 40 but claims to be complete._RESYNC_TOPICS_SHOWN lists 40 topics, then "and N more", while the notice text and the bundled SKILL.md say every topic is listed. Either reword it so it doesn't claim completeness (and tell the agent how to get the rest, e.g. topics list), or list every topic.
first_run can stay set for the life of the process. It is cleared only in _flush_backlog, which returns early when nothing was skipped. Clear it when the first catch-up completes, whether or not anything was skipped.
--self is never checked against the scope frame's agent_person_id. A wrong --self now reaches the agent as self_id on every tag. If they differ, end with a clear error, or prefer the scope's id and warn.
Tighten the first-run cutoff against clock skew. The cutoff for unrecognised events is a timestamp test: created_at < started_at - 60s. A local clock more than 60 s ahead of the server can drop a live unrecognised event with no trace, whereas a skipped message at least shows in the backlog count. Requiring "no heartbeat seen yet" as well limits that exposure to the replay plus the first ping.
Activity
bniladridas commentedon Nov 13, 2026
Tracking the split:
fix(agents): end the catch-up on a ping, not on any other comment