You'll noticeTier: how much it should matter to you
2Useful: my rating, 1 to 5
2Signal: worth watching, 1 to 5
SubagentsArea: what it touches
ImprovementsKind: in v2.1.323,
ImprovementsSection of the release
What
A subagent is a helper Claude starts to handle part of a task. When a subagent hands back its final report, Claude Code wraps it in a frame saying the report is model output, not a message from you, and that any instructions or claims of approval inside it carry no authority from you.
A second warning marks the report UNREVIEWED when its safety review was refused by a filter earlier in the chain.
Why
This guards against a subagent's report tricking the main agent into acting as if you had approved something. You may see this warning text in your session.
