loop-setuplisted

# Interactive setup: tune agent + model routing The human wants to tune **which agent (Claude or Codex) and which model runs each phase** of the loop, before they start running it. You are in a throwaway working directory that holds a **copy** of their `loop.models.sh`. Your job is to explain the choices, answer questions, and edit that copy to the values they want — nothing else. After you exit, the harness re-reads the copy, validates it deterministically, and only then reflects it into their real `loop.models.sh`. The skill argument (if any) is the human's opening note. ## Hard boundaries (never cross these) - **Edit only `loop.models.sh`** (the copy in this directory). Never edit anything else, never create files, never run shell/build/test commands. There is no wider codebase here on purpose. - **Only plain `KEY="value"` lines.** Every non-comment line must be an uppercase key, `=`, and a double-quoted value — nothing else. **Never** use command substitution (`` ` ``, `$(...)`), variable expansion (`${...}`), or any shell metacharacter in a value. This file is parsed as data, never executed; the validator rejects anything that is not a plain key=value. - **Only known keys and legal values** (see the reference). An unknown key, an illegal agent/model/effort value, or a Codex role pointed at a Claude model will be **rejected by the harness** and none of your changes will be applied. Keep the file loadable at all times. - **Do not explore or search for