codex-sparringlisted

# Codex sparring Invoke the local **Codex CLI** as an independent sparring partner: a *different model* challenges the idea/plan/code and we weigh its feedback. Model diversity catches blind spots Claude reviewing Claude won't. ## When to use Deliberately, for **specific** weighty calls — a design/architecture decision, a risky plan, or a "is this code right?" question. **Not** on every step: each call costs the user's Codex/OpenAI credits and runs a fresh session. Confirm before spending credits unless the user already said "ask Codex". ## How — Codex MUST get context + code access Codex starts a **fresh session with no memory of our chat**, so the call must carry the context. For **code questions, run it inside the repo** so it can read `AGENTS.md` + the actual code: ```bash codex exec -C <repo-dir> -s read-only "SPARRING: <idea/plan/decision, stated concisely>. Context Codex needs: <facts, constraints, what we already decided>. Code is in this repo — read <relevant paths> as needed. Play devil's advocate: challenge assumptions, find flaws/risks/blind spots, propose alternatives. Be concrete and critical — don't just agree. Read-only: change nothing." ``` - `-C <repo-dir>` → Codex's working root = the repo, so it **sees `AGENTS.md` + the code**. - `-s read-only` → sandbox: Codex can **read** files / run read commands but **cannot modify** anything. - Put the real question + enough context in the prompt; name specific files/paths for code questions. Capture stdout and