codex-plan-reviewlisted

# Codex Plan Review (cross-model) Run a **plan file** past the **Codex CLI** as an independent adversarial reviewer (a different model, fresh context) and fold its findings into the plan. Codex runs in a pinned **read-only** sandbox and **never edits the checkout** — it only reports. Claude saves a transcript under `.claude/codex-reviews/` (gitignored) and applies any accepted changes to the plan itself. This is the plan-file sibling of `/codex-review` (which reviews diffs). The cross-model angle is the point: a same-model reviewer cannot remove identity-based self-preference, so a *different* model reviewing the plan is the irreplaceable defense before execution. **Trigger:** user types `/codex-plan-review <path-to-plan.md>`. Like `/codex-review`, it deliberately does NOT carry `disable-model-invocation: true`: that flag currently *also* blocks the slash command itself (anthropics/claude-code#26251, dup #38969), so the "Manual only" scoping in the description is what keeps it from auto-firing. There is no CLI-honored `auto-activate` field. ## Single-writer rule (mandatory) Codex must never write to a checkout Claude is using (one checkout = one writer; two writers race on the same working tree). The read-only sandbox enforces this — do NOT relax `--sandbox read-only` to `workspace-write` (the repo config default is workspace-write, which would violate this rule). Findings come back as text; **Claude** folds the accepted ones into the plan file. Never let Codex update th