serial-adversarial-reviewlisted

# Serial Adversarial Review Use this skill when a change warrants review by the whole external-model panel, not one model. You (main-thread Claude) are the **sole broker**: the reviewer models cannot talk to each other, so every exchange passes through you. The mechanical companion enumerates plugins, builds the diff, dispatches each reviewer, and keeps the findings ledger; **you** do the judgment — normalize, collect positions, relay rebuttals, arbitrate, and decide. > **The serial order MUST NOT decide the outcome.** A correct finding raised early must not > die because a later model dismissed it, and a finding raised last gets no free pass. > Whether a finding survives depends on **evidence and explicit consensus**, never on which > stage produced it. ## Reviewers run strictly serially Dispatch **one reviewer at a time** and wait for it to finish before the next. Never fan out — overlapping external reviews contend badly and explode latency. Reviews are app-server / CLI backed and take **several minutes**; dispatch them **detached** and wait on the PID, never inline. The companion handles the per-reviewer flag dialect and the codex session-env unset for you. Enumerate first, then dispatch in order: ```bash node "${CLAUDE_PLUGIN_ROOT}/scripts/review-chain-companion.mjs" enumerate --json ``` The codified panel order is: 1. **Claude (you, main thread)** — review the diff first, produce the initial findings list. 2. **Gemini Pro — via the Antigravity plugin** (`antigrav