Auto mode is a permission mode in which Claude Code checks each action itself instead of asking you to approve it. The REPL is a place where Claude can write and run code as one step. That code can try to start a sub-agent (a helper agent that takes on part of a task) or send a message to another agent. Auto mode cannot read the task or message text the code passes along, so it now refuses these sends.
When this happens, the model is told one of two things:
- Do the work itself, running the pieces side by side with
Promise.all - Leave auto mode with Shift+Tab
The check that catches secrets quoted in shell commands now also tests the command after it has been split into its separate parts.
The model now gets a clear reason when one of these calls is blocked, plus two ways forward. If Claude stops to suggest leaving auto mode, this is the cause.