Debug interruptions that leave an agent speaking over callers

Describe the failure the caller hears

An interruption defect can mean several different things: audio never stops, it stops and then restarts, the next answer ignores the correction, or ordinary background sound cuts off every sentence. Capture which behavior occurred before changing thresholds. The Lupin gateway tracks interruptions and interrupted response contexts, but those counters describe events rather than subjective success. Compare the event timeline with a synthetic dialogue that has a known correction. A transcript alone can hide overlapping playback because it often records text without the precise time each sound reached a listener.

Check whether speech onset reaches the gateway

Barge-in begins with detecting new caller speech while output is active. Examine the inbound audio stream and voice activity decisions before blaming response generation. Missing frames, wrong audio encoding or an overaggressive noise filter can prevent reliable onset detection. The current cascade uses gateway voice activity and a turn policy; native audio has its own provider-managed events. Do not apply a threshold adjustment intended for one mode to another without checking the captured configuration. First establish that the relevant onset or interruption event actually arrived.

Cancel queued sound and fence late responses

Stopping a synthesis request does not necessarily remove audio already queued for playback. Lupin’s interruption path clears playback, cancels synthesis and fences the pending response so late chunks cannot restart the previous answer. Test this deliberately by delaying a synthetic audio chunk until after the interruption. The old chunk should be discarded rather than treated as new output. In native mode, event identity also matters because a repeated provider interruption must not repeatedly reset the fresh response. Log response identifiers and ordering instead of raw customer speech.

Preserve the meaning of the interrupted exchange

A caller saying Tuesday, not Thursday changes the task even if the assistant already produced a complete sentence in its internal context. The subsequent response should use Tuesday and should not claim that the abandoned Thursday action happened. Lupin marks interrupted response contexts separately from audible interruption counts. Check that distinction when output stops correctly but the conversation still behaves as though the full answer was delivered. Any externally consequential tool action remains behind exact human authorization; interrupting spoken output does not automatically undo a provider write that was already approved and completed.

Retest false positives and recovery together

Use three fixtures: a deliberate correction, a brief acknowledgement during speech and background noise without a new instruction. Confirm that corrections stop stale output, acknowledgements do not destroy the task, and noise does not trigger an interruption loop. After each case, ask a follow-up requiring the corrected fact. This tests conversational recovery rather than only cancellation. Record media mode, interruption count, end reason and the test’s expected state. Escalate transport or synthesis failures with those narrow facts, avoiding customer recordings and assumptions about the provider’s private internal model.

Set a concrete interruption outcome

For the correction Wednesday instead of Tuesday, expect no stale Tuesday audio after cancellation, a Wednesday proposal in the next turn and zero executed calendar changes. Those three assertions cover audible behavior, retained meaning and authority separately. A lower interruption count alone cannot demonstrate that the correction was handled successfully.

Sources & verification

Reviewed against current checked-in Lupin routes, shared contracts and official provider references. No live calls, OAuth authorizations, app actions or purchases were performed.

• Lupin phone, SMS and calls ↗Checked 2026-11-05
• ElevenLabs model reference ↗Checked 2026-11-05

Prepared with AI assistance by Lupin editorial. How we research and maintain these guides.