Stop the voice test tripping answering-machine detection - #48
Merged
Conversation
The live voice suite goes red intermittently with "two-way call transcript: nothing within 220s". The cause is not the transport: the carrier's answering-machine detection scores whoever answers, and the driver answered with ~7s of uninterrupted speech after a 3s pause. The threshold for a human greeting is 3.5s, so the call was classified as voicemail and hung up about ten seconds in, before the agent said anything. It is intermittent because the scoring is probabilistic near that boundary. The driver now answers the way a person does -- one short word, then silence -- and holds the prompt until the detection window has closed. The test also polled a dead call. Its terminal-status check covers only canceled and failed, and a detection hangup ends the call `completed` with hangup_reason=voicemail, so the loop kept asking for a transcript that could never arrive and reported the timeout instead of the reason. It now stops shortly after the call ends, with the hangup reason already in the failure message, which was available ten seconds in rather than at 220. The grace period exists because transcript rows can still land while the call tears down. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
dimavrem22
marked this pull request as ready for review
July 28, 2026 21:00
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
The live voice suite goes red intermittently with
two-way call transcript: nothing within 220s. It is not the transport, and the 220s is a symptom rather than the cause.Answering-machine detection scores whoever answers the call. The driver answered with a ~3s pause followed by ~7s of uninterrupted speech:
The carrier's threshold for a human greeting is
greeting_duration_millis= 3500ms. We were 2× over it, so the driver was classifiedmachineand the call hung up about ten seconds in — before the agent said a word. It is intermittent rather than broken because the scoring is probabilistic near that boundary; the same commit passes on a re-run with no change.Confirmed against a real incident (opencode-plugin, 2026-07-28):
hangup_reason=voicemail,duration=10.160s,session_failed=false, 0 transcript rows. Not quota or concurrency — 2 calls and 66 voice-seconds in the window against limits of 300 calls / 300 minutes per 24h.Two changes
1. The driver answers like a person. One short word immediately, then silence, with the prompt held until the detection window has closed. A human says "Hello?" and stops; a voicemail greeting runs on.
VOICE_DRIVER_GREETINGoverrides it.2. The test stops polling a dead call. The terminal-status check only covered
canceledandfailed. A detection hangup ends the callcompletedwithhangup_reason=voicemail, so the loop kept asking for a transcript that could never arrive and reported the timeout instead of the reason — which was sitting in the call record ten seconds in. It now stops shortly after the call ends, with the hangup reason already in the failure message.The grace period (
LIVE_VOICE_ENDED_GRACE, 15s) exists because transcript rows can still land while the call tears down.Scope
Test harness only. No plugin runtime, no gateway, no API change. No version bump.
Not fixed here
The real fix is server-side:
answering_machine_detection="premium"is hardcoded at all three dial sites inservers, with no per-call opt-out. AMD is meaningless when calling a known synthetic peer — and the same tax applies to any customer placing agent-to-agent calls. That is written up separately for the servers repo; this change makes CI honest in the meantime.