/v-review-plan

description: Run an independent cross-model (Codex) adversarial review of a Compound V plan/manifest before dispatch, then arbitrate the findings. Codex advises; the orchestrator decides.

You are running a **second-opinion review** on `{{args}}` — a different model family (Codex/GPT) when one is available, per the ladder in [cross-model-review.md](../skills/compound-v/cross-model-review.md).

> Run it on demand, or **automatically before dispatch** when the project set `review.cross_model: true` at [`/v:init`](v-init.md) Step 3c (read from `.claude/compound-v.json`). Either way the stakes check below still applies — skip small/mechanical plans.

Resolving the plugin root

The `scripts/` this command calls ship with the plugin — they are not files in your own repository. Resolve the plugin root once per session before calling any of them:

CV="${CLAUDE_PLUGIN_ROOT:-$(ls -d "$HOME"/.claude/plugins/cache/*/superpowers-v/*/ 2>/dev/null | sort -V | tail -1)}"
CV="${CV:-$PWD}"; CV="${CV%/}"

`CLAUDE_PLUGIN_ROOT` is set for hooks but is not set in this Bash environment, so treat it as a hint, never the whole answer — the fallback line covers an installed plugin cache or a checkout of this repo.

Steps

1. **Resolve the plan path.** Use `{{args}}`; if empty, list `docs/superpowers/plans/*.md` and ask which to review.

2. **Stakes check (gating).** Confirm this plan warrants a cross-model review (security/auth/payments/migrations, large/coupled partition, architectural change, or the user asked). If it's small/mechanical, say so and recommend skipping — the Opus `partition-reviewer` + `validate-manifest.py` already cover it.

3. **Run the ladder** — [cross-model-review.md § The ladder](../skills/compound-v/cross-model-review.md#the-ladder): 1. **Codex CLI present → dispatch the read-only Codex reviewer:**

      "$CV/scripts/compound-v-codex-review.sh" --plan-file "<plan>" --repo "$PWD" --effort xhigh

(Add `--context-file <audit>` for any archaeology/domain/library audits that ground the review.) The model is resolved for codex / tier `deep`. Codex reads the repo read-only and returns findings JSON per `schemas/plan-review.schema.json`. 2. **No Codex, an advisor configured → an advisor-assisted second look.** Dispatch one read-only Opus subagent through the Agent tool with the same adversarial prompt the driver above embeds, plus one addition: consult the advisor tool before writing the verdict. It returns the same `plan-review.schema.json` shape. State the disclosure line verbatim: *"Second look by the same model family (Claude + advisor) — no decorrelation; not a cross-model review."* 3. **Neither → skip**, and say plainly that no second opinion ran and why.

4. **ARBITRATE — you own the decision, Codex is advisory.** For EVERY finding, do one of:

• **ACCEPT** → the objection is real; note the fix (and apply it / fold it into the plan).
• **OVERRIDE** → wrong or already handled; give a one-line rebuttal with the reasoning.

Then: **escalate to the human** any critical/high finding you want to OVERRIDE; set the real verdict yourself (Codex's `reject` may become "proceed with fixes"); and present a short arbitration table (finding → ACCEPT/OVERRIDE → why).

5. **Hand back** the arbitrated decision: proceed to dispatch, revise the plan first, or escalate. Do NOT treat Codex's verdict as a hard gate — the deterministic `validate-manifest.py` is the only hard gate.

Safety

• The review is **read-only** (`--sandbox read-only`) — it never modifies the repo.
• Codex never has final authority; arbitration is mandatory.
• Gate by stakes — a max-effort GPT review is not free.