Why Tracking a Coding Agent's Completion Event Alone Can Mislead Developers

A terminal monitor detecting a task_complete signal does not guarantee the product state is correct, as the event may belong to an old session, background worker, or already-resumed turn. To address this, a local parser can return a completion flag, a turn key for deduplication, and a semantic timestamp that allows later activity to revoke a result. Different AI coding agents like Claude and Codex require separate adapters since their completion signals use distinct formats and can be mimicked by error records. The monitor combines turn timestamps with local session activity, allowing a completed turn to shift states — from needsYou to working or idle — based on freshness and user interaction. Filesystem modification time serves only as a fallback, and periodic scans remain necessary alongside watchers to catch dropped events.