Detecting When a Coding Agent Is Truly Done Requires More Than Process Checks

A quiet coding agent transcript is ambiguous — the agent may have finished, stalled, or be mid-execution, and neither process state nor file modification time reliably indicates which. A more accurate approach involves parsing the transcript backward to find the latest meaningful conversation event, such as an end_turn or task_complete marker, rather than relying on metadata timestamps. Complications arise from child-agent completions being mixed into main transcripts, and API errors that superficially resemble successful turn endings, both of which require explicit handling. Agent Island 1.7.1 addresses these issues with bounded UI status windows that expire stalled or completed labels after defined intervals, preventing stale evidence from driving misleading alerts. The same event-driven model also governs reminder logic, cancelling notifications if a fast reply arrives and treating the first post-launch scan as historical rather than fresh activity.