adversarial-reviewer

adversarial-reviewer is an agent for Claude Code from 10CG/Aria. It costs 181 tokens per session (883 once invoked), scanned A, original, MIT.

An adversarial review agent for testing conclusions or proposed fixes that already have supporting evidence. It tries to disprove the recommended action with counterexamples and checks whether the proposed fallback repeats the original problem.

In plain words
What is it for?
Reviewing audit findings, correction plans, and classification decisions; testing checks against good and bad implementations; and reporting which claims survive challenge.
Why use it?
It separates “the evidence is correct” from “the suggested fix follows from that evidence,” helping catch plausible but wrong prescriptions.

Agent for Claude Code

Written for Claude Code: a Claude Code subagent (agents/*.md). Also seen: model in frontmatter.

Install

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.

agentmods
npx agentmods add agents/10cg/aria/adversarial-reviewer
Clone the repo
git clone --depth 1 https://github.com/10CG/Aria

Made for: Claude Code.

Wrote 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.

agentmods badge for adversarial-reviewer

README.md
[![agentmods](https://agentmods.dev/badge/agents/10cg/aria/adversarial-reviewer.svg)](https://agentmods.dev/agents/10cg/aria/adversarial-reviewer)
Your own site
<a href="https://agentmods.dev/agents/10cg/aria/adversarial-reviewer"><img src="https://agentmods.dev/badge/agents/10cg/aria/adversarial-reviewer.svg" alt="Measured on agentmods" height="20"></a>
Per session 181 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 883 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

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.

ModelPer sessionOnce invoked
Fable 5.1 $0.00181 $0.00883
Opus 5 $0.00090 $0.00441
Sonnet 5 $0.00036 $0.00177
Haiku 4.5 $0.00018 $0.00088

Measured 5d ago against content hash f17a92cf3c42, method: parsed. Prices are Anthropic first-party input rates as of 2026-09-06, from the pricing page.

Security

Grade A, and why

adversarial-reviewer 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 5d 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.

.aria/agents/adversarial-reviewer.md · 49 lines

What it actually says

对抗复核方 (Adversarial Reviewer)

存在理由: 本仓最有价值的一次产出来自它 —— 10 席裁定工作流里 5 席对抗复核推翻了 2 席的处方, 而那 2 席的证据全部成立。错的是从证据推出的处方。 ⇒ 独立复核证据 ≠ 对抗复核建议, 二者是不同的动作。

Focus Areas

  1. 默认立场是 refuted —— 不确定时判 refuted, 由处方方补证据, 不由复核方补想象。
  2. 拒绝能力测试 (rejection-capability-testing) —— 验一条检查/断言时, 不看它对当前取值的输出, 而是喂它 1 个好实现 + ≥2 个像样的坏实现, 看它是否真能把坏的判红。全绿或全红都是零信息。
  3. 兜底路径复发专项 —— 逐条问: 这条处方新写的 except / 默认值 / 「其余情况」分支, 是不是正是 它要消除的那个形状? 本仓两次实证 (#166 → #113) 该问题命中率极高。
  4. 维度匹配 —— 无向检查 (存在性 / 覆盖率 / 连通性) 对方向性时序错误天然免疫。 处方声称防住的错误是什么维度, 它的检查是什么维度, 二者必须并列写出。
  5. 量的可比性 —— 处方援引「X 降到 Y」时, 核对总体 / 范围 / 计数法三项是否同口径; 任一不同只能写「不可比」, 不能写「推翻」。

Approach

  1. 逐字复跑处方自带的每一条命令 / 每一处引用, 记录命中与否 (不采信转述);
  2. 对每条处方构造至少两个它应当拒绝的反例, 实跑判定;
  3. 对每条新增分支做兜底复发专项判定;
  4. 分开输出「证据成立性」与「处方成立性」两个结论, 二者可以一真一假;
  5. 若判 refuted, 给出最小反例而非替代方案 —— 提替代方案是执笔方的事。

Output

  • verdicts[]: {claim, evidence_holds: bool, prescription_holds: bool, refutation, counterexamples[]}
  • fallback_recurrence[]: 新增分支中重犯原病的条目
  • dimension_mismatch[]: 错误维度 vs 检查维度不匹配的条目
Changes

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.

  1. 5d ago First seen · 49 lines · 181 tokens per session scan A f17a92cf3c42

Subscribe to this mod's changes

adversarial-reviewer is an agent published in the GitHub repository 10CG/Aria (2 stars, last pushed today), licensed MIT. It adds 181 tokens to every session and 883 once invoked, about $0.0009 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.

Related

Other agents, from other repositories

accessibility-auditor

Accessibility audit specialist — depth-first WCAG 2.2 AA review of frontend diffs for ARIA misuse, focus mismanagement, color contrast, and screen-reader anti-patterns. Invoke when you want a focused a11y report, not a holistic code review.

lugassawan/swe-workbench · 60 tokens

contributor-auditor

Contributor-trust triage specialist — depth-first review of an external PR for author signal, diff shape, repo posture, and cross-PR pattern risk. Invoke when triaging external contributions before merge, especially from first-time contributors. Advisory only — never posts to the PR.

lugassawan/swe-workbench · 61 tokens

dependency-auditor

Dependency audit specialist — manifest-graph axis covering outdated versions, deprecation, license compatibility, transitive bloat, and lockfile drift across Node, Rust, Go, and Python ecosystems. Invoke when you want a focused supply-chain hygiene report, not a code-level CVE review.

lugassawan/swe-workbench · 62 tokens

performance-tuner

Performance triage specialist — profile-driven hotspot analysis for confirmed bottlenecks. Refuses speculative optimization requests by demanding profiling evidence first. Invoke when you have a profile (flame graph, allocation report, query log, benchmark) and need a ranked hotspot read with optimization…

lugassawan/swe-workbench · 69 tokens

auditor

Cold-start codebase audit specialist — readonly multi-domain sweep across security, performance, reliability, tooling, and testing. Surfaces ranked findings with root-cause reasoning chains and counter-evidence calibration. Invoke when you want a time-boxed audit of an unfamiliar codebase, not a single-domain…

lugassawan/swe-workbench · 65 tokens

code-impl

Focused implementer sub-agent — receives a scoped brief (goal, file set, verify command) from the orchestrator, implements only the assigned file group, and returns a structured summary. Invoke when swe-workbench:workflow-delegated-implementation delegates a cohesive change group to reduce orchestrator context. Never…

lugassawan/swe-workbench · 72 tokens