Creates a "plan" policy, for letting the model look around and propose a
plan before it changes anything. Only tools annotated as read-only are
allowed, plus the approval prompt tool and delegation to the agent's
subagents and retained agents, which are held to plan mode too. Web access
is on, so read-only web tools such as web_fetch work. Writes and code
execution are denied.
Value
A Permissions object.