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
Drain the mid-run queue on inference.done, not reactor.done
reactor.done only fires once, at agent shutdown, never between turns,
so a plain-text reply with no tool calls left queued messages stranded
forever (reproduced in the tool-less-turn regression test). inference.done
fires on every reactor cycle regardless of what the director does next,
so it also covers goal-governor and workflow continuations that never
produce a settled connector.reply.
Also replace the internal-state readout ("queue +1 -> pending N", the
"queue" meta label repeated on the same row) with the queued message's
own text and attachments, and add Ctrl+X to cancel the most recently
queued item before it dispatches.
0 commit comments