tb4-five-arm

tb4-five-arm is a skill for Claude Code, Codex from huangruiteng/loopx. It costs 100 tokens per session (2,947 once invoked), scanned A, original, Apache-2.0.

A benchmark setup for comparing five ways of running a coding agent on Terminal-Bench 4.0, a fixed collection of terminal-based software tasks.

In plain words
What is it for?
It is for running repeatable five-way benchmark trials, preparing the required task environments, checking them without using model tokens, and monitoring short test runs.
Why use it?
It helps reveal whether results come from the agent harness rather than changes in the model or tools, while avoiding configuration mistakes that can produce misleading scores.

Skill for Claude CodeCodex

Written for no agent in particular: nothing here depends on one. Also seen: mentions Codex.

Needs its repository: it runs a file that does not travel with it, so clone the repository first. The line is ./scripts/prebuild_images.sh music-harmony # 应建 2 个镜像,不是 1 个.

Good fit It is for running repeatable five-way benchmark trials, preparing the required task environments, checking them without using model tokens, and monitoring short test runs.

Compare 6 skills from other repositories ↓
About the project

LoopX is a provider-neutral control plane that gives long-running AI-agent work persistent state, decisions, governance, recovery, and handoffs across different agent harnesses. It is for managing and reviewing durable work performed by Codex, Claude Code, Cursor, dsh, or custom agent systems.

huangruiteng/loopx · 6,233 stars · on GitHub · huangruiteng.github.io

Install

Getting it into your agent

It runs from inside its repository, so the clone comes first — what it calls does not travel with the file alone.

Clone the repo
git clone --depth 1 https://github.com/huangruiteng/loopx
agentmods
npx agentmods add skills/huangruiteng/loopx/tb4-five-arm

Made for: Claude Code, Codex.

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 tb4-five-arm

README.md
[![agentmods](https://agentmods.dev/badge/skills/huangruiteng/loopx/tb4-five-arm/github.svg)](https://agentmods.dev/skills/huangruiteng/loopx/tb4-five-arm)
Your own site
<a href="https://agentmods.dev/skills/huangruiteng/loopx/tb4-five-arm"><img src="https://agentmods.dev/badge/skills/huangruiteng/loopx/tb4-five-arm/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.

agentmods 80×15 button for tb4-five-arm

Your own site · 80×15
<a href="https://agentmods.dev/skills/huangruiteng/loopx/tb4-five-arm"><img src="https://agentmods.dev/badge/skills/huangruiteng/loopx/tb4-five-arm.svg" alt="Reviewed on agentmods" width="80" height="20"></a>
Per session 100 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 2,947 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. A grade says what 26 rules found in the file — not that it is safe. Third-party audits
  • NVIDIA SkillSpector pass 7 Sept 2026
How audits are shown
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 $0.00100 $0.02947
Opus 5.5 $0.00040 $0.01179
Sonnet 5.5 $0.00020 $0.00589
Haiku 4.5 $0.00010 $0.00295

Measured 1mo ago against content hash 1b76d5cebae3, method: parsed. Prices are Anthropic first-party input rates as of 2026-10-07, from the pricing page.

Security

Grade A, and why

tb4-five-arm 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 1mo 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.

benchmark/swe-marathon/skills/tb4-five-arm/SKILL.md · 192 lines

How it starts

The opening of the file, as written. The whole thing — 192 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Terminal-Bench 4.0 五臂对照

五臂定义、LoopX profile 装配、agent 类、监控工具全部与 [[swe-marathon-five-arm]] 相同,先读那一份。这里只写 TB4 特有的部分。

对照维度仍是 harness 不是模型:五臂的模型、effort、工具面、沙箱、容器一致。

切换方式

export WEN_BENCH=tb4      # 必须 export,见下面「最毒的一个坑」
source env.sh

./scripts/prebuild_images.sh music-harmony   # 应建 2 个镜像,不是 1 个
./scripts/verify_envs.sh music-harmony       # install-only,不烧 token
./scripts/canary_timeout.sh                  # 五臂 × 短死线
./scripts/marathon_all.sh --dry              # 63 任务 × 5 臂 = 315 trial

benchmark 相关的量全在 scripts/bench/tb4.sh,驱动脚本里没有任何 TB4 常量。 WEN_BENCH 不设时默认 swe-marathon,行为与引入 bench 层之前逐字一致 (已用 bash -x 逐参数 diff 验证过)。

数据来源

terminal-bench/ 是 harbor-framework/terminal-bench 的 tag v4.0.0 (commit 452bf305c6),66 个任务,见 terminal-bench/VERSION。

  • 不能用 harbor run -d [email protected]:harbor 的注册表(Supabase dataset 表)里只有 [email protected],实测查 4.0 返回 null。
  • 不要跟 main:上游 README 明说 "will be continuously updated", 跟 main 跑出来的结果不可复现。
  • tasks/dataset.toml 里每个任务有 sha256 digest,task.toml 与 tests/ 一个字节都不能改。要注入环境变量走 harbor 的 --ve / --ae。

TB4 vs SWE-Marathon:四处结构性差异

项 SWE-Marathon TB4.0
agent.timeout_sec 3600–36000 各异 全部 28800(统一 8h)
verifier.environment_mode 无(shared) 全部 separate
network_mode 三态,逐任务 一个都没声明 → harbor 默认 public
每任务镜像数 1(environment/) 2(environment/ + tests/)
连续分 metrics.json 有 没有,只有二值 reward

排除 3 个要 H100 的任务(fp8-rmsnorm-gemm / jax-speedrun-gpu / math-eval-grader),本机是 4090D。剩 63 个,其中 11 个是多容器 (environment/docker-compose.yaml),1 个(medical-claims-processing) 还要 MCP server(playwright,sse http://playwright-mcp:3080/sse)。

坑(都会静默通过)

一、最毒的一个:WEN_BENCH 用前缀赋值

WEN_BENCH=tb4 source env.sh      # ✗ 错
export WEN_BENCH=tb4; source env.sh   # ✓ 对

bash 在 source 返回后会把前缀赋值的变量还原成未设置,但 WEN_TASKS_DIR 是 export 的、留了下来。于是后续任何脚本自己 source env.sh 时:WEN_BENCH 空 → 回退 swe-marathon profile,却继承着 TB4 的任务目录 —— 拿 TB4 的任务、套 marathon 的网络策略(from-task 找不到 network_mode 就断网)、 开着判官注入、写进 marathon-full/。退出码 0,全程无警告。

Read the full file on GitHub · 192 lines

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. 1mo ago First seen · 192 lines · 100 tokens per session scan A 1b76d5cebae3

Subscribe to this mod's changes

tb4-five-arm is a skill published in the GitHub repository huangruiteng/loopx (6,233 stars, last pushed 5d ago), licensed Apache-2.0. It adds 100 tokens to every session and 2,947 once invoked, about $0.0004 per session on Opus 5.5. 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 2026-09-03.

Related

Other skills, from other repositories

deep-research

Use when the user asks for "deep research", a "research report on", a "comprehensive analysis of", to "look into X in depth", to "write a report on X", or any question needing multi-source synthesis beyond a single search — an autonomous multi-step deep research engine implementing an iterative Think → Search →…

moonlight-lupin/agent-skills · 150 tokens

image-studio

Use when the user wants image generation or editing through the local fal.ai three-stage studio workflow — brainstorm a strong prompt, prototype cheaply and iterate on feedback, then produce a finalised image. Covers "generate an image", "make an image/picture/illustration/graphic of…", "create an AI image"…

moonlight-lupin/agent-skills · 161 tokens

website-scraping

Generic playbook for extracting structured data from any website — hotel prices, flight fares, e-commerce listings, real estate, jobs, competitor product catalogues, anything where the goal is to turn one or more URLs into clean records on disk. Use this skill whenever the user mentions scraping, harvesting…

moonlight-lupin/agent-skills · 210 tokens

skill-maintainer

Use when tracking upstream drift and syncing adapted skill libraries.

moonlight-lupin/agent-skills · 15 tokens

library-rag

Semantic search over a personal library using Nemotron-3-Embed-1B embeddings + sqlite-vec. Index books, documents, any text corpus; query by meaning. Includes EPUB→Markdown conversion and MCP server for auto-available search tools.

moonlight-lupin/agent-skills · 54 tokens

notebooklm-mode

Source-grounded research pipeline inspired by NotebookLM. Collect verbatim source extracts into a temporary workspace, then answer questions and create deliverables grounded in those sources. Two grounding modes: strict (vault-only) and augmented (vault + labeled [background] knowledge). Trigger when the user asks to…

moonlight-lupin/agent-skills · 91 tokens