What problem does it solve? Voice-mode pipelines transform speech twice (cleanup of the user's raw transcript, and summarization of the agent's reply into spoken form), and each hop can silently change meaning or drop facts. This Skill audits those per-turn logs so fidelity regressions are caught, evidenced locally, and reported without leaking private voice content. ## Core Features & Use Cases - Two-hop fidelity judgment: Compares RawTranscript vs CleanedTranscript (inbound) and AgentReply vs WingmanSpoken (outbound), assigning each turn an ok, minor, or flag verdict. - Privacy-safe archiving and reporting: Copies flagged turns with full quoted analysis into a local archive under %LOCALAPPDATA%\cc-director\voice-review, while GitHub issues contain only abstract defect classes, turn ids, and local paths. - Incremental, deduplicated issue filing: Skips already-reviewed turns via reviewed.json markers, dedupes against open voice-review issues with gh, and labels issues as bug or enhancement. - Use Case: Run it nightly via a Windows Scheduled Task to review the day's voice turns, archive three flagged turns where numbers were dropped from spoken replies, and file one sanitized GitHub issue recommending a fix to the outbound summarizer prompt. ## Quick Start Ask the agent to run /voice-review to review the voice log and report any fidelity problems it finds.
