In general, when I'm writing tests, the pure functions end up as bare expects and the impure functions end up as scenarios. The following contrived namespace keeps a list of users and allows you to get the full name of each user.

(ns user)

(def users (atom #{}))

(defn full-name [{:keys [fname lname]}] (str fname " " lname))

(defn all-names [] (map full-name @users))

The tests for this namespace would often look something like the following code:

(ns user-expectations
  (:use expectations user))

(expect "John Dory" (full-name {:fname "John" :lname "Dory"}))

(ns user-expectations.scenarios
  (:use expectations.scenarios user))

(scenario
  (with-redefs [users (atom #{{:fname "Mary", :lname "Dory"} {:fname "John", :lname "Dory"}})]
    (expect ["Mary Dory" "John Dory"] (all-names))))

It feels natural to put the with-redefs in a scenario, since scenarios also support (and often make use of) stubbing, localize-state, & freeze-time. However, there's really no reason that you need to use a scenario if you're simply looking to redef a var.

The following test provides the same level of functionality verification, without needing to use expectations.scenarios: 

(ns user-expectations
  (:use expectations user))

(expect "John Dory" (full-name {:fname "John" :lname "Dory"}))

(expect ["Mary Dory" "John Dory"]
  (with-redefs [users (atom #{{:fname "Mary", :lname "Dory"} {:fname "John", :lname "Dory"}})]
    (all-names)))

scenarios are great, but these days I try to keep things simple with bare expectations whenever possible.