← ClaudeAtlas

gemini-adversarial-reviewlisted

Use Gemini CLI to challenge Codex's implementation approach, assumptions, tradeoffs, and failure modes.
yilibinbin/external-models-for-codex · ★ 1 · Code & Development · score 64
Install: claude install-skill yilibinbin/external-models-for-codex
# Gemini Adversarial Review Use this skill for high-risk changes, architecture decisions, reliability-sensitive code, security-sensitive code, migrations, rollback-sensitive changes, or when Codex may be overconfident. Run: ```bash node "${CODEX_PLUGIN_ROOT}/scripts/gemini-companion.mjs" adversarial-review "$ARGUMENTS" ``` Background routing: - Foreground use runs the normal command above. - If `$ARGUMENTS` contains `--background`, first run: ```bash node "${CODEX_PLUGIN_ROOT}/scripts/gemini-companion.mjs" reserve-job adversarial-review "$ARGUMENTS" ``` - Parse the returned JSON and dispatch exactly one forwarding subagent or child worker with the returned `workerCommand` JSON argv array. The child must execute that array as argv while preserving element boundaries; if forced through a shell, quote every element. - The child runs `run-reserved-job` once through `workerCommand`; it must not inspect or reinterpret the repository. - The parent returns the job id immediately and tells the user to use `gemini-result <job-id>`. Rules: - This is read-only. - Gemini must first infer and state the author's intent. - Gemini reviews through adversarial lenses: `skeptic`, `architect`, and `minimalist`. - Gemini must output `PASS`, `CONTESTED`, or `REJECT`. - Gemini must include `Findings`, `What Went Well`, and `Lead Judgment`. - Use `--json` when Codex needs a machine-checked adversarial verdict object. - Codex must accept or reject Gemini findings with its own judgment before ma