Opus 5.5 unattended agents stop early on text-only turns: Anthropic's fix is a checklist nudge plus a named-stops system prompt paragraph
Claude Platform Docs·high signal
Anthropic's Opus 5.5 prompting guide says the model writes progress updates that can end a turn with stop_reason end_turn, which unattended loops misread as task completion. The recommended fix has three parts. Keep a to-do checklist, and when items remain open with no blocker, send a user message naming them, capped at 2-3 automatic continuations. Add a system-prompt paragraph that names four specific early-stop shapes, including a summary that announces the next step without calling a tool. Add it from the first request, because adding it mid-session invalidates earlier thinking blocks.