| name | adversarial-review |
| description | 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.
|
Adversarial review (agent-operated platform skill)
Who runs what
| Actor | Duty |
|---|
| Human | Intent only:「对抗审查」「审一下这批改动」「按推荐审查」 |
| Main / any OpenClaw agent | Load this skill → run scripts → disclose results |
| Host headless backends | claude / codex-gpt / codex-grok as find/refute (not ACP by default) |
Never tell the user to copy-paste run.sh as the primary path.
Never replace this with a solo Main monologue and call it 对抗审查 (that is 设计批判).
When to load / run
- User asks for 对抗审查 / adversarial review / multi-role review
- Write-task checklist: acceptance suggests review (from 开发方式卡)
- Runtime code/scripts changed on allowlisted repo (wezdeck / team-repo worktree)
- After editing
scripts/dev/adversarial-review itself → dogfood
Skip: pure docs/tests-only (runner auto-skips); user explicitly skips with reason.
Multi-role minimum
| Role | Stance | Runner stage |
|---|
| find / reviewer | guilty-until-proven | stage 1 |
| refute / refuter | kill weak findings | stage 2 |
| repro | empirical in sandbox | stage 3 |
- Prefer different families than the writer (strategy B via
--writer).
- Same capability twice OK → label SINGLE-MODEL / degraded; do not skip refute.
- Solo essay without opposite roles → 设计批判, not 对抗审查.
Paths (resolve from repo root)
Prefer worktree cwd if task is on a claw-* tree; else primary allowlisted root.
REPO_ROOT = git rev-parse --show-toplevel # claw-* or primary
RUN = $REPO_ROOT/scripts/dev/adversarial-review/run.sh
SEL = $REPO_ROOT/scripts/dev/adversarial-review/lib/select-backends.sh
If scripts missing (wrong cwd), cd to wezdeck primary or the task worktree first.
Agent procedure (do this; do not hand off to human)
- Identify
cwd / repo (must be allowlisted write or read-review ok)BASE_REF (default: master or merge-base with master / task start)- writer from 开发方式:
main | claude | codex | codex-gpt | codex-grok | human
- Optional dry probe
"$SEL" --writer <writer> --no-probe
- Run (prefer writer-aware; shell via claw-run / exec-risk when required)
"$RUN" <BASE_REF> --writer <writer> --mode strict
Explicit override only if user named backends:
"$RUN" <BASE_REF> --reviewer claude --refuter codex-gpt --mode strict
- Report in chat using mandatory disclosure block (below). Honest fail if scripts/backends fail (error-closed-loop).
- Do not claim cross-agent if form is single-model; do not invent green gates.
Writer → expected pair (strategy B)
| writer | Typical pair when available |
|---|
claude / Claude-ACP/TUI | codex-gpt × codex-grok |
codex / Codex-* | claude × codex-gpt (else × codex-grok) |
main / Main-Grok | claude × codex-gpt |
human | same as global best |
Mandatory disclosure (paste into 结果)
## 对抗审查披露
- writer: Main-Grok | Claude-ACP | Codex-ACP | human | …
- 形态/form: cross-family | cross-model-codex | single-model-multi-role | …
- form/degraded/reason: (from select-backends / run log)
- reviewer 全名 / 立场: …
- refuter 全名 / 立场: …
- repro: 已跑 | 跳过（理由）
- 命令或范围: run.sh <BASE> --writer …
- skipped_gates: … | 无
- 关键结论: …（绑 find/refute/repro）
Helper commands (agent-only)
"$RUN" selfcheck claude codex-gpt codex-grok
"$RUN" dogfood --mode strict --fail-on-finding
"$RUN" <BASE> --writer main --dry-run --no-probe
Don't
- Don't ask the human to run
run.sh as the main path
- Don't use ACP CODEX_HOME for review (
env -u CODEX_HOME is inside provider)
- Don't skip refute when claiming 对抗审查
- Don't call solo Main analysis 对抗审查
- Don't force-push; wezdeck personal mainline may push master after green (L0-13/19)
Related
- Repo thin skill (TUI discovery):
skills/adversarial-review/SKILL.md (repo root)
- Runner & docs:
docs/adversarial-review.md, scripts/dev/adversarial-review/
- Host TUI doctrine:
agent-profiles/v1/en/validation.md
- L0-21 disclosure:
AGENTS.md
- Task hooks:
skills/dev-task/SKILL.md (acceptance / 审查建议)
- Interaction:
openclaw/docs/agent-interaction.md
