| name | adversarial-review |
| description | Multi-role adversarial code review for this repo. Agent loads this skill and runs host scripts (find → refute → sandbox repro). Humans only state intent (审一下 / 对抗审查). Use when user asks for adversarial review, acceptance recommends review, or runtime code changed before merge.
|
Adversarial review (repo skill — thin entry)
Who runs: the coding agent (Claude-TUI / Codex-TUI / OpenClaw Main), not the human.
Do not ask the user to copy-paste shell as the primary path.
Canonical pieces
| Piece | Path |
|---|
| Runner (only implementation) | scripts/dev/adversarial-review/run.sh |
| Writer-aware select | scripts/dev/adversarial-review/lib/select-backends.sh |
| Full OpenClaw skill | openclaw/workspace/skills/adversarial-review/SKILL.md |
| Docs | docs/adversarial-review.md |
| Host TUI doctrine | agent-profiles/v1/en/validation.md (adversarial review rules) |
This file is the in-repo discovery surface for agents working in wezdeck cwd.
OpenClaw Main may use the workspace skill instead; both call the same runner.
When
- User: 对抗审查 / 审一下 / adversarial review
- Task acceptance suggests multi-role review
- Runtime code/scripts changed (runner auto-skips pure docs/tests)
Multi-role minimum
Find (guilty-until-proven) + Refute (kill weak findings) + Repro (recommended).
Solo monologue = 设计批判, not 对抗审查. Same backend twice → SINGLE-MODEL.
Agent procedure
REPO_ROOT=$(git rev-parse --show-toplevel) (claw-* worktree or primary).
- Resolve writer:
main | claude | codex | codex-gpt | codex-grok | human.
- Run (prefer writer-aware):
"$REPO_ROOT/scripts/dev/adversarial-review/run.sh" <BASE_REF> \
--writer <writer> --mode strict
- Paste disclosure (required):
## 对抗审查披露
- writer: …
- form / degraded / reason: …
- reviewer / refuter: …
- repro: 已跑 | 跳过(理由)
- 命令或范围: …
- skipped_gates: … | 无
- 关键结论: …(绑 find/refute/repro)
Optional: "$REPO_ROOT/scripts/dev/adversarial-review/lib/select-backends.sh" --writer <w> --no-probe
Don't
- Don't hand
run.sh to the human as the main workflow
- Don't skip refute when claiming 对抗审查
- Don't use ACP
CODEX_HOME for review (runner uses host ~/.codex)
- Don't claim cross-agent when SINGLE-MODEL