菜单
Thin pointer: OpenClaw agents must load workspace skill openclaw/workspace/skills/adversarial-review/SKILL.md and run host scripts themselves. Humans only state intent (审一下 / 对抗审查).
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Ledger development tasks to Feishu Base via dev-task-ledger.sh. Write roots from ~/.openclaw/tasks-allowlist.json (per agent). Use on accept/confirm/close of coding/ops work; not for pure Q&A.
Drive the machine Chrome debug browser via OpenClaw-managed MCP (chrome-devtools). Use for page navigation, snapshots, screenshots, clicks, forms, console/network inspection — not for host shell risk.
Allowlisted development (wezdeck (+ optional team roots in local config)) under OpenClaw: wezdeck defaults to primary master; claw worktrees when parallel/isolation needed; 团队仓 still prefers claw-*; human/Claw rails (H1/H2, C1/C2/C3), ACP access layer.
Full error closed-loop for Dex (Main): diagnose, safe self-fix, verify, report; platform Exec failed handling; escalation with options. Load when handling failures or writing failure reports.
Host shell MUST go through claw-run (gate then exec). Three layers: rules → Grok re-check → human only if still danger. Prefer claw-run over bare exec.
Multi-role adversarial code review for allowlisted repos. Agent loads this skill and runs host scripts (find → refute → sandbox repro); humans only state intent (e.g. 审一下 / 对抗审查). Use when user asks for adversarial review, task acceptance recommends review, or runtime code changed before merge.
| name | adversarial-review |
| description | Thin pointer: OpenClaw agents must load workspace skill openclaw/workspace/skills/adversarial-review/SKILL.md and run host scripts themselves. Humans only state intent (审一下 / 对抗审查). |
Canonical agent skill: openclaw/workspace/skills/adversarial-review/SKILL.md
That skill tells Main/agents to:
REPO_ROOT / worktreescripts/dev/adversarial-review/run.sh with --writer …Docs: docs/adversarial-review.md. Select: lib/select-backends.sh.