Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add BackToCimaCoppi/Praxis --skill adversarial-reviewgit clone --depth 1 https://github.com/BackToCimaCoppi/PraxisWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/backtocimacoppi/praxis/adversarial-review)<a href="https://agentmods.dev/skills/backtocimacoppi/praxis/adversarial-review"><img src="https://agentmods.dev/badge/skills/backtocimacoppi/praxis/adversarial-review/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/backtocimacoppi/praxis/adversarial-review"><img src="https://agentmods.dev/badge/skills/backtocimacoppi/praxis/adversarial-review.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00215 | $0.06637 |
| Opus 5 | $0.00108 | $0.03318 |
| Sonnet 5 | $0.00043 | $0.01327 |
| Haiku 4.5 | $0.00021 | $0.00664 |
Grade A, and why
adversarial-review scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 9d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 341 lines — stays where its author put it; the contents beside it link to each section on GitHub.
对抗评审:一次开放发现,主线程直接裁决
一句话:默认由 Fable 5 与 GPT-5.6-Sol max 独立找问题;用户明确点选时可独立加入 Cursor Grok 4.6 Extra High、Cursor GLM 5.2 Max、Cursor Kimi K3 Max 的任意组合。全部评审不开 Fast;当前 Claude Code 主线程负责裁决,裁决后点评并排名实际参与模型,不另启裁判线程。
0. 硬边界
- 同一身份只开放一次:用 review_id + object_hash + truth_baseline_hash 标识评审身份。同一阶段、同一基线、同一对象不得开 R2/R3。
- 全部评审者只读且必须独立勘察:默认两名;Grok/GLM/Kimi 按用户点选独立加入,最多五名。评审者在授权项目根内亲自检索、核验,只输出事实、失败场景和候选方向;不写被评审对象、不写正式报告、不互相读取输出。
- 主线程直接裁决:全部已选评审意见收齐后,当前主线程合并、去重、编号并逐条裁决;可选评审者是评审者而非裁判,禁止再起“裁判”子线程或裁判模型。
- 评审意见不是真值:会改变业务结果、接口契约、持久化语义、死亡线规则的选择,必须追溯到正式上游规格或由业务决策负责人作出的
_shared/用户裁决记录.md#DEC-x。没有来源只能升级到对应授权角色。 - 整改不再开放找问题:主线程按裁决整改后,调用 closed-remediation-review。该 skill 只用本工具一个全新 Opus 5 子线程检查固定整改清单,再由主线程逐条裁决。
- 报告不取得规格地位:正式 L1–L7 才是长期真值;评审报告只是发现、裁决和整改闭环证据。
只有用户裁决、上游正式真值或评审对象发生根本重写时,旧身份才失效。重新评审前必须登记失效原因和新哈希;措辞修订、补引用、落实上一轮整改都不构成新对象。
1. 角色与 Claude Code 适配
| 角色 | 调用方式 | 默认模型 | 权限 |
|---|---|---|---|
| 评审者 A | Claude Code 原生 Agent | Fable 5(model: "fable") |
项目根内只读发现 |
| 评审者 B | codex exec | gpt-5.6-sol,reasoning max | 项目根内只读发现 |
| 评审者 C(可选) | Cursor CLI 中继(scripts/cursor-review-runner.sh) |
cursor-grok-4.6-xhigh(Cursor Grok 4.6 Extra High) |
项目根内只读发现;仅用户明确点选时加入 |
| 评审者 D(可选) | 同上 | glm-5.2-max(Cursor GLM 5.2 Max) |
同上 |
| 评审者 E(可选) | 同上 | kimi-k3-max(Cursor Kimi K3 Max) |
同上 |
| 裁决者 | 当前 Claude Code 主线程 | 当前主线程 | 只裁决,不发明新设计 |
| 整改验收 | closed-remediation-review | 一个全新 Opus 5 原生子线程 | 固定范围只读检查 |
默认只启动 A/B。C/D/E 默认全部关闭、各自独立点选:只有用户明确写出“加入 Grok”“加入 GLM”“加入 Kimi/Kimi K3”,或对带该模型名称的选项回答“是”时才加入对应评审者;“全部模型/五模型”表示三者都加入。单说“三模型”“多加一个”不能确定第三方是谁,必须先让用户点名。不得因模型可用、任务复杂或死亡线命中而自动加入,也不得为了询问 C/D/E 而阻塞常规评审。
全部已选评审者必须并行、相互独立。常规档固定 A=Fable 5、B=gpt-5.6-sol reasoning max、C=cursor-grok-4.6-xhigh、D=glm-5.2-max、E=kimi-k3-max。任何评审都禁止 Fast:所选模型 ID 不得以 -fast 结尾,不得开启 Claude Code fast mode 或等价的速度模式;Fast 不是思考强度,用户要求“快/省”时只能在明确允许的非 Fast 思考档中降档(A 固定 Fable 5 不换模型,B 可把 reasoning 降到 high)。不得换掉模型后仍宣称完成了 Fable + GPT 对抗评审。报告必须记录实际模型 ID。
2. 总体流程
冻结评审身份与证据包
├─ 并行:Fable 5 独立评审(默认)
├─ 并行:GPT-5.6-Sol max 独立评审(默认)
├─ 并行:Cursor Grok 4.6 Extra High 独立评审(仅用户明确点选 Grok)
├─ 并行:Cursor GLM 5.2 Max 独立评审(仅用户明确点选 GLM)
└─ 并行:Cursor Kimi K3 Max 独立评审(仅用户明确点选 Kimi)
↓
当前 Claude Code 主线程合并去重、分配 AR-x、逐条裁决
↓
必要的用户问题合并成一张表,一次询问;逐项记录 DEC-x
↓
全部 AR-x 裁决完成后,当前主线程点评、排名全部实际参与模型
↓
主线程整改正式产物,冻结整改清单与 closure_scope_hash
↓
closed-remediation-review:一个 Opus 5 原生子线程只检,主线程再裁
↓
PASS 才允许下游继续
What ships with it
5 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 9d ago First seen · 341 lines · 215 tokens per session scan A a1327fc6463d
adversarial-review is a skill published in the GitHub repository BackToCimaCoppi/Praxis (6 stars, last pushed 15d ago), licensed Apache-2.0. It adds 215 tokens to every session and 6,637 once invoked, about $0.0011 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
github-commenting
How to post clean, rich, deduplicated GitHub PR review comments — suggestion blocks, multi-line anchors, markers, formatting rules. Load before posting or fixing any PR comment.
repo-conventions
NeuroLink's review standards — the critical rules to enforce, what NOT to comment on, the security bar, hot paths. Load before reviewing any change in this repository.
resolve-pr-comments
Evaluate, fix, answer, and reply to GitHub pull request review comments and conversation comments. Handles both change requests (fix or skip) and reviewer questions (explain using reasoning recalled from past Claude Code transcripts). Use when the user asks to "resolve PR comments", "fix review comments", "address PR…
map-codebase
Deep architecture report that fans out parallel inspections across different aspects of the codebase (structure, tech stack, APIs, patterns, data flow, dependencies, testing) and synthesizes findings into a comprehensive document at .turbo/codebase-map.md and .turbo/codebase-map.html. Use when the user asks to "map…
reply-to-pr-threads
Draft, confirm, and post replies to GitHub PR review threads. Handles per-category reply formatting, re-fetches thread resolution state so auto-resolved threads are skipped, and posts via GraphQL. Use when the user asks to "reply to PR threads", "post PR thread replies", or "draft PR reply messages".
answer-reviewer-questions
For each reviewer question on a PR, recall implementation reasoning and compose a raw answer. Use when the user asks to "answer reviewer questions", "draft answers to PR questions", or "explain reviewer questions".