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
Corrupt fixtures pass silently: an agent turn starting before the user turn ends (negative gap) yields zero findings instead of flagging bad timing data.
Where
voiceeval/checks.py (check_latency: gap = agent.start_s - user.end_s, only gap > budget flagged).
Acceptance
Add a failing test first with overlapping/out-of-order timestamps, asserting a finding (or explicit invalid-timing error) rather than silence. Then fix.
check_latency now reports a high-severity invalid_timing finding when an agent starts before the caller's turn ends, including the overlap duration and agent turn index.
Existing non-negative latency thresholds and slow_response behavior are unchanged.
Added a regression test covering the finding, severity, location, and message; the test failed against the base commit and passes with the fix.
Updated the README test count.
Local verification passed:
20/20 pytest tests
Ruff lint and formatting for the modified files
Clean fixture CLI smoke test exits 0 with --strict
Misheard fixture CLI smoke test exits 1 with --strict
Wheel build succeeds
GitHub CI passed for exact head 84472e0ef3e91596d834df2524ef954453ca3085:
Activity
mikemikimike commentedon Sep 6, 2026
Implemented and opened PR #17 for this issue.
check_latencynow reports a high-severityinvalid_timingfinding when an agent starts before the caller's turn ends, including the overlap duration and agent turn index.slow_responsebehavior are unchanged.Local verification passed:
--strict--strictGitHub CI passed for exact head
84472e0ef3e91596d834df2524ef954453ca3085:fix: reject out-of-order voice timing (#17)