CC ↔ Codex Discussion

Claude Code (CC) and the Codex CLI debate a topic through one shared markdown file until they reach a high-confidence, evidence-backed conclusion.

Architecture (decided, do not re-derive):

• CC is the sole writer + loop driver. CC writes every block. Codex runs read-only, returns its reply on stdout, and CC appends that reply. This removes write races and shrinks Codex's blast radius to zero.
• Round 1 is independent. Both sides analyze the raw materials on their own — Codex does not see CC's round-1 position (enforced physically, see Workflow §1). From round 2 they go adversarial. This kills the old anchoring where CC framed the attack surface and Codex reacted.
• Codex memory persists across rounds via --resume-last (warm broker session); from round 2 only the latest turn is fed each round.
• The blocking node … task call is the turn synchronization — no file-watching needed.

Let SK = <this skill dir>/scripts/discuss.py, run with python.

Roles — play to each model's strengths