A fixed workflow in which several AI models review code, critique one another, converge on conclusions, and have a judge verify the results against the actual code.

Use it to run a multi-model code audit, choose which agent and AI command-line tools participate, and view the process and scoreboard in /workflows.

It provides multiple independent assessments and a structured way to compare disagreements before producing a final review.

arena — Skill by serdardb

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 skills/serdardb/multi-agent/arena

Any agent

npx skills add serdardb/multi-agent --skill arena

Clone the repo

git clone --depth 1 https://github.com/serdardb/multi-agent

Made for: Claude Code, Codex.

Or install multi-agent, the plugin that ships this one along with the rest of its 2 skills, 3 agents.

Per session 49 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.

When invoked 868 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.

Model	Per session	Once invoked
Fable 5	$0.00049	$0.00868
Opus 5	$0.00024	$0.00434
Sonnet 5	$0.00010	$0.00174
Haiku 4.5	$0.00005	$0.00087

Measured 2d ago against content hash 29998885c4a5, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

arena 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 2d 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.

plugin/skills/arena/SKILL.md · 48 lines

What it actually says

Run a DETERMINISTIC multi-model review arena. The whole pipeline (rounds, cross-critique, convergence, judge, scoreboard) is fixed in the engine — you ONLY collect the user's choices and launch it. Do NOT greet or narrate; go straight to step 1. Never audit anything yourself.

Resolve the engine path: run echo "${CLAUDE_PLUGIN_ROOT}" → call the result ENGINE. The arena CLI is ENGINE/arena.py; the workflow script is ENGINE/arena.workflow.js.
List what's available: python3 ENGINE/arena.py list --path . → parse JSON {models, agents}. If models is empty, say no AI CLIs are usable and stop. If $ARGUMENTS already carries <agent> <models> <topic>, use them and skip the matching questions.
Pick the agent: AskUserQuestion, one option per entry in agents (label = name, description = its description).
Pick the AIs: AskUserQuestion (multiSelect), one option per entry in models. Default: all.
Topic + scope: if given in $ARGUMENTS, use it. Otherwise ask what to review. If they refer to changes / a method / a path, resolve scope with git diff / git status / Grep and fold it into the topic. (You interpret intent in any language.)
Suggest the mode: python3 ENGINE/arena.py suggest-mode --topic "<topic>" → {mode, reason}. Show it, then AskUserQuestion: keep it or switch (adversarial = debate & critique; independent = audit separately). The user's pick wins.
Rounds & options: AskUserQuestion for min/max rounds (preset min 2 / max 4; independent ignores this) and judge on/off (recommended on). Tip: grok+codex as debaters with claude as the neutral judge is a good default.
Launch the WORKFLOW (shows in /workflows): Workflow({ scriptPath: "ENGINE/arena.workflow.js", args: { arenaPath: "ENGINE/arena.py", agent: "<agent path>", agentName: "<agent name>", models: ["grok","codex",...], topic: "<full topic+scope>", mode: "<mode>", minRounds: <n>, maxRounds: <m>, judge: <true|false>, path: "." } }) (substitute the real ENGINE). Tell the user it's running and they can watch it live in /workflows.
Present when it finishes (it returns rounds, judge, scoreboard):
- 🏆 scoreboard first: winner: <model> (or tie: a = b), then per model: <model>: <points> pts · <confirmed> confirmed · <false> false · <unique> unique.
- the FINAL ranked findings — the judge's list if judging was on, else the last round; most-severe first.
- how each model's positions evolved (NEW → KEEP / WITHDRAW / ADOPT). Cross-model findings are unverified claims unless the judge marked them CONFIRMED against real code; flag the rest and offer to trace any finding's call chain.

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.

2d ago First seen · 48 lines · 49 tokens per session scan A 29998885c4a5

Subscribe to this mod's changes

arena is a skill published in the GitHub repository serdardb/multi-agent (1 stars, last pushed 1mo ago), licensed MIT. It adds 49 tokens to every session and 868 once invoked, about $0.0002 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.

Code review adversarial ai-agents claude-code claude-code-plugin code audit

Security Audit Reporter

Triage raw security-scan findings (hardcoded secrets, injection patterns, vulnerable dependencies) into a prioritized, actionable security audit report. Use for security audit, code audit, vulnerability triage, and risk review.

AgentEra/Agently · 47 tokens

fin-review-loop

经济金融论文的对抗性review循环。对草稿进行多轮严格评审，检查实证严谨性、方法正确性、理论贡献和写作质量，给出可操作的修改建议。（AI review 不能替代同行评审，草稿必须经研究者核实后投稿。）.

csmar432/finai-research · 71 tokens

build-review-md

Use this skill when the user wants to create or set up a REVIEW.md configuration file for their repository. Trigger for ANY of these: (1) user says "create REVIEW.md", "set up REVIEW.md", or "configure review rules", (2) code-gauntlet Phase 2d detects no REVIEW.md and suggests creating one, (3) user wants to customize…

liatrio-labs/claude-code-gauntlet · 171 tokens

fleet-conductor

Use when conducting a large backlog or multi-repository campaign across several worker lanes, especially when work needs bounded ownership, held dependencies, collision control, draft pull request gates, and evidence-based landing. Triggers on "conduct the fleet", "burn down the backlog", "run a multi-repo campaign"…

escoffier-labs/skillet · 73 tokens

demi

Use before implementing, scaffolding, prototyping, or adding a feature when the work should start with the smallest useful code path, avoid speculative architecture, or prevent overbuilding before reduce would be needed.

escoffier-labs/skillet · 43 tokens

stagiaire

Use when the current session needs an answer or a work product from a model on ANOTHER vendor's stack - a second opinion from a different model family, a cheap fast model for bulk work, or a cross-model review - by dispatching that vendor's own CLI as a one-shot worker. Triggers include "ask…

escoffier-labs/skillet · 133 tokens