You can interrupt it while it is talking
High-priority input now wakes the agent out of a wait on an unrelated running action — a "stop talking" said mid-narration used to sit for up to 58 seconds. Narration itself also queues instead of cutting in. Subagent round budgets rise from a hardcoded 10 to a configurable 50, and running out now writes up what it found instead of discarding it.
The robot is reading out a long passage and you say "stop talking" — it used to take a long time to react. Up to 58 seconds, measured.
Why the wait
While an action is running, the robot waits for it to actually finish before doing the next thing. That part is right — without it, it would start the next sentence before finishing the current one, or issue a new command mid-motion.
The problem was that the wait could not be cut short. Your words landed in the queue, but the loop was stuck waiting on an action that had nothing to do with them, so it never even got to reconsider.
High-priority input now wakes that wait: the robot returns to deciding immediately and reads the new instruction. The action being waited on keeps running normally — it just no longer blocks you from speaking.
Narration queues up too
Progress narration used to travel the emergency-interrupt path — the one built for e-stop, which by design means "whatever you are doing, do this now". That is wrong for narration: it talked over whatever the robot was already saying, and because it registered nothing of its own, the very next line could talk over it in turn.
Narration now queues like anything else. One sentence finishes before the next begins.
Also: long tasks no longer end in nothing
Research handed to a subagent had a hardcoded 10-round ceiling. Hitting it almost always happened mid-step, and four minutes of searching came back with nothing at all.
| Before | Now | |
|---|---|---|
| Round ceiling | 10, hardcoded | 50, configurable |
| On running out | "max rounds reached" — findings discarded | Writes up what it has, then stops |
| Idle timeout | 10 minutes, hardcoded | Configurable |
Along the way it turned out the subagent was re-searching the same question repeatedly: context compaction was aggressive enough to lose the record of what had already been searched, so it searched again. With more rounds retained, the same task produced a complete report on its own by round 9.