Loop
Keep the agent working until it calls finish_task.
Loop is a session switch. When it is on, an ordinary assistant message does not end the turn. The agent is expected to keep using tools until the task is done, then call finish_task.
Turn it on
On Desktop, enable Loop in the composer (the /loop chip). In the CLI TUI, run /loop.
finish_task is only exposed while Loop is on. If Loop is off and the model still calls it, Spirit rejects the call as an unknown tool.
When to use it
Use Loop for a multi-step job you do not want to babysit: implement, test, fix, repeat. Leave it off for a single question or a one-file edit.
Loop does not change approval. Default approval still pauses on edits and shell.