codex-debatelisted

# Codex Debate Loop ## Quick Start `/codex-debate <task>` — implement the task, get an adversarial review of the diff from Codex (top-tier model available on the plan), debate each finding, fix real issues, and loop until Codex returns `APPROVED` **and** local project checks are green. If the work is already done and only a review is wanted, skip step 2 and start at step 3. ## Model Selection Resolve `$MODEL` once at the start of the loop, in this order: 1. **User override** — if the invocation names a model or tier ("use terra", "on luna", "with 5.5"), map it to the slug (`gpt-5.6-terra`, `gpt-5.6-luna`, `gpt-5.5`, …) and use that. 2. **Pinned model** — if a `model.txt` exists next to this SKILL.md, use its contents (written by the installer or the user to save quota). 3. **Auto (default)** — the top tier the CLI knows about: ```bash MODEL=$(jq -r '[.models[] | select(.visibility=="list")] | sort_by(.priority) | .[0].slug' ~/.codex/models_cache.json) ``` Lower `priority` = higher tier (1 = flagship). The cache refreshes whenever codex runs, so new families (5.7, 5.8, …) are picked up automatically. If a codex call fails with a **model-access error** mid-loop, step down to the next slug by ascending priority and continue; in the final report state which model actually reviewed, and advise the user to request top-tier access from their ChatGPT workspace admin or upgrade the subscription. Always state the resolved model in the final report. ## Hard Rules - Reviewer