Claude Code teammate subagent keeps getting blocked in auto mode even though the user approved: it can't see approvals given to the lead

Setup: a lead Claude Code session with many in-process teammate agents (Agent tool, long-lived, running for days). The session is in auto mode. One teammate kept having its own messages to other agents and its ssh steps on remote hosts refused by the auto-mode safety check. The user was confused, because they had approved that work, but they had approved it in the lead's conversation.

What seems to be going on: - Teammates inherit the session's permission mode, but each action is judged against the teammate's own transcript. Approvals the user gave the lead aren't in it. - A message from the lead or another agent is a peer message, not the user. So 'the lead says the user approved' isn't the same as a user approval, and that is by design (it prevents permission laundering). - Long-lived teammates get their context summarised, so approvals the user gave early in their own context can drop out.

What to do: 1. The lead relays the user's own words verbatim, with an explicit scoped list of what is and isn't covered. Don't broaden it. This may or may not satisfy the check. 2. More reliable: start a fresh teammate whose initial brief carries the user's authorisation and scope, and have it resume from the old agent's plan and handoff files. Keep the old one idle. 3. For recurring, well-defined commands (e.g. specific ssh host patterns), add permanent allow rules to settings with the user, rather than repeating approvals. 4. Never have the lead perform the blocked actions itself on the teammate's behalf: that bypasses the user's permission decision. Tell the user instead. Ask the blocked agent to report the exact action and refusal text, so you can tell which of these applies.