Optional predicate controlling whether to interrupt for a given tool call.
Receives a ToolCallRequest and returns True to interrupt or False to
auto-approve. The predicate is called during after_model before the tool
call is added to the batched human-in-the-loop request.
The request is constructed with tool=None and a new ToolRuntime. The
ToolRuntime copies context, store, stream_writer, execution_info,
and server_info from the node-level Runtime, while tool_call_id is
populated from the current tool call. The tools argument is not supplied,
so it uses its default empty list.
In per_call mode the predicate runs in wrap_tool_call and receives the real
ToolCallRequest, with tool and the tool's ToolRuntime set.
In both modes the predicate runs again when the run resumes, so it must return the
same answer for the same call. If it returns False on resume, the reviewer's
answer is skipped and the tool runs, even if they rejected it. In batched mode,
if it flips for only some of the reviewed calls, the decisions no longer line up
with the calls and resuming raises a ValueError.