plan-reviewlisted

# Cross-model plan review You are running a bounded adversarial review of an implementation plan, using OpenAI Codex (a different vendor's model) as the critic. The value comes from decorrelated errors: Codex has different blind spots than you. The debate is a **detector, not a resolver** - unresolved disagreements go to the user, never silently dropped or silently "won". ## Procedure 1. **Get the plan.** If an argument path was given, read that file. Otherwise use the most recent plan in this conversation. If neither exists, ask the user for it. 2. **Round 1 - Codex critique.** Call the `codex` MCP tool (server `codex`). Send the full plan plus one paragraph of project context, with this instruction: > Review this implementation plan as an adversarial senior engineer. Do not invent findings - if the plan is sound, say so - but DO include findings you are unsure about, marked as such: triage happens on our side, and an unreported doubt can never be examined. Check specifically for: (a) requirements from the user's request the plan misses, (b) unhandled edge cases and failure modes, (c) steps with hidden complexity or underestimated effort, (d) a materially simpler alternative that achieves the same goal, (e) risks - security, data loss, performance, migration/rollback, (f) testability - how will each step be verified, and which tests are missing, (g) premortem - assume the plan was implemented as written and still failed in production: name the most plausible cause. O