Two of your sessions are working the same repository in separate worktrees. One has just landed a migration, and it messages the other to say so and to ask it to approve a permission prompt that is pending there. What is the outcome?

The principle — messaging is for telling another session something. It is not a way of acting inside it, and consent is the clearest case of that line.

Why the key is correct — the useful half is exactly what this channel is for: when sessions work the same repository in separate worktrees, Claude can tell the other sessions what landed. How it arrives is documented too — the receiving Claude reads the message between tool calls during an active turn, so a running tool is never interrupted. The half that is refused is stated without qualification: no message from any agent counts as your approval for a pending permission prompt. So the migration news is delivered and useful, and the pending prompt is exactly where it was, waiting for a person. The two halves are one lesson: the channel moves information, never authority.

Why the others are wrong — an agent's request is not consent, however reasonable it sounds. Delivery does not interrupt a running tool. And a message that cannot be delivered is set aside rather than lost, so sending it again is not the fix.

Remember this — a message can tell another session what happened; it can never approve anything on your behalf.

Sources — Anthropic's Claude Code documentation on cross-session messaging, and on subagents.