wehub-resource-sync
1 line
2.7 KiB
JSON
1 line
2.7 KiB
JSON
{"content": "---\nname: adversarial-review-loop\ndescription: Has a second, independent model review every change before merge so two different reviewers must agree — code only lands when both clear the bar.\ncategory: engineering\ninterval: 15m\nstop-condition: The implementing agent and the independent cross-model reviewer both approve with no blocking findings and tests pass.\ncomponents: [agent:expert-advisors/architect-review, command:git-workflow/gemini-review, command:git-workflow/pr-review]\ntags: [code-review, cross-model, quality, loop]\n---\n\n# Adversarial-Review Loop\n\n> **Loop Engineering** — the verifier inside the loop is the hard part, and an agent grading its own homework will delete the failing test and call it done. This loop puts a *second, different* model in the reviewer seat: two model families must agree before code lands.\n\n## 🎯 Goal\nImplement a change, then have an **independent cross-model reviewer** argue against it. Iterate on findings until both the builder and the outside reviewer sign off — no self-grading.\n\n## ⏱️ Schedule\nSuggested interval: `15m` (or run until the change is merge-ready).\n\n## ▶️ Run it\n```\n/loop 15m \"Implement the change. Then have a different model review it adversarially for correctness, security and design. Address every blocking finding and re-run tests. Land it only when both the builder and the independent reviewer agree there are no blocking findings. Max 5 review rounds.\"\n```\n\n## 🔁 Iteration steps\n1. **Build** — implement the next slice.\n2. **Cross-review** — run `/gemini-review` so a *different* model family critiques the diff.\n3. **Reason** — collect blocking findings from the outside reviewer.\n4. **Act** — fix them; `architect-review` validates the revision; re-run tests.\n5. **Observe** — re-review with `/pr-review`; loop until both sides agree (cap the rounds).\n\n## 🛑 Stopping condition\nBoth reviewers report zero blocking findings and tests pass — or the review-round cap is reached and the loop hands off with the open findings.\n\n## 💰 Budget & guardrails\nA `--max-iter` style cap on review rounds and a severity threshold: only findings at/above the bar block. Two independent models must agree, so neither can rubber-stamp itself.\n\n## 🧩 Referenced components\n- `agent:expert-advisors/architect-review` — the in-house reviewer.\n- `command:git-workflow/gemini-review` — an independent, different-model review.\n- `command:git-workflow/pr-review` — structured re-review checklist.\n\n## 💡 Example\nThe builder ships an auth change; the cross-model reviewer flags a timing-attack risk the in-house reviewer missed. The loop fixes it, both agree, and only then does it land.\n"} |