What problem does it solve? Production voice agents fail in ways that only surface in real calls, and manually turning those failures into reproducible regression tests is slow and error-prone. This Skill converts flagged call logs into evaluator scenarios on the Cekura platform so every production failure becomes a test the next regression run can catch. ## Core Features & Use Cases - Failure clustering: Groups flagged call logs by failure mode (hallucination, drop, tool_error, workflow_miss, refusal) into 3-8 scenario clusters, each grounded in verbatim evidence quotes from real transcripts. - Dual write paths: Creates turn-by-turn conditional_actions replays directly via the scenarios API, or generates free-form instruction scenarios server-side from call logs, with dedup against existing scenarios. - Single-call fast path: Reproduces one specific call log ID (or observe URL) as a single evaluator scenario, including test profile with camelCase dynamic variables, metric selection (reuse, predefined copy, or create), and evaluator phone attachment. - Use Case: After triaging a week of production calls with cekura-flag-call-log-failures, hand the flagged set to this Skill to draft one scenario per failure cluster, confirm the proposals, and create them with matching caller personalities and metrics. ## Quick Start Ask the assistant to create evaluator scenarios on Cekura agent 12345 from the flagged call logs produced by the failure-triage step, then review and approve the proposed scenario set before anything is written.
