VFA-2026-0016
A delayed decision about whether an earlier overlap counted as an interruption gets wrongly applied to a later, unrelated moment, cutting the agent off for speech that was never actually judged an interruption.
What this is
Filed as Async interruption response misattribution, the name every report of this mode is collected under.
A failure mode is one named thing that goes wrong on a live call. This record carries every public report of it the atlas has found, in each reporter's own words, and a status that moves only when a re-check stores the artifact behind the move. Status open: reported and still open where it was reported.
Evidence
Every independent public report of this mode, newest first: who wrote it, where, when, what they said, and the link back to the post it came from.
...is accepted and attributed to that later overlap — which can emit a false isInterruption: true and cut off the agent for a user utterance the model never scored that...
https://github.com/livekit/agents-js/issues/2119
first seen 2026-07-26 · fetched 2026-07-26T03:05:24Z · body sha256 a5091d4d64d4c61e
Each quote is at most 30 words around the reported failure and links to the post it came from. If you wrote one of these and want it off this page, the removal path takes it down within 24 hours.
Hit this one yourself? Add your call to this entry: the stack, the versions, and what the caller heard. It lands as another independent report on this page, with your handle and a link back to what you wrote.
Report another occurrenceThe form is a GitHub issue on the hotato repo. Prefer email? [email protected] reaches a person either way.
What is measurable
Evidence tier: Tier 1 · dual-channel audio: caller and agent separated at capture, so overlap and timing are facts of the recording, exact to the sample.
What it establishes
- The seconds of overlap where the caller had taken the floor and the agent kept speaking, with the timestamp in the call where it started.
- The same recording produces the same overlap on every run, so the number is comparable across builds.
What your trace settles
- Your pipeline trace at that timestamp names what held the floor: the interrupt handler, the VAD threshold, or the realtime API's own turn logic.
- Whether the caller meant to interrupt is a judgement about the conversation; the measurement covers who held the floor and for how long.
# Printed as BARGE-IN: the caller took the floor and the agent kept speaking... hotato autopsy ./call.wav
A shipped hotato subcommand maps to this mode, as detector kind overlap_while_agent_talking. Printed as BARGE-IN: the caller took the floor and the agent kept speaking, reported with the overlap in seconds. Needs the caller and agent on separate channels. hotato is MIT and runs locally. getting started.
Status history
A daily re-check reads every source thread on this page. The status moves when that re-check stores the link, the quote and the date behind the move, so every state on this record is one you can open.
The record has carried one status since it opened: open, reported and still open where it was reported. A move lands here with the link, the quote and the date that show it.
Discussion
Reader discussion of this mode: what fixed it, what did not, and what the record here is missing. It is a conversation, not part of the evidence above. The reports on this page are attributed and linked to their sources, and a comment is a comment.
Open this discussion on GitHubThe record
References
Every link this entry stands on: the post behind each report, and the artifact behind each status move.
- reporthttps://github.com/livekit/agents-js/issues/2119 · toubatbrian · GitHub issues · 2026-07-26
Machine surfaces
Removal
If you wrote something quoted on this page and want it gone, one email does it. The removal path takes the quote, the handle and the link off this page, out of the feeds and out of the llms.txt within 24 hours, and tombstones the id so the entry cannot quietly return at a new address.