📦 Multi — 多

v7.2.0

Agent Config Manager 多代理编排引擎——目标驱动的深度研究与项目协作系统。支持任务分解、分支执行、验证审核、返工迭代、智能决策。遵循第一性原理，实现主代理与分支代理的双向通信。触发词：多代理、multi-agent、代理编排、深度研究、目标分解、任务委派、工作流、agent orchestrate、multi agent

0· 21·0 当前·0 累计

by @bjmrcft-hash

生产力工具

下载技能包

最后更新

2026/4/21

安全扫描

VirusTotal

无害

查看报告

OpenClaw

安全

high confidence

The package is internally consistent with a multi‑agent orchestration tool: its code, docs, and runtime instructions align with the stated purpose, but it performs file creation/deletion and active polling in the user's OpenClaw workspace so you should review and control its filesystem actions before running it on important data.

评估建议

This skill appears to do what it says (a multi‑agent orchestrator) and contains substantial source code rather than being instruction‑only. Before installing or running it: 1) Back up any important files under ~/.openclaw or your workspace; the skill will create/modify files there and its clean/archive commands delete files in shared/. 2) Run multi-agent-orchestrator check_env and review the returned list; prefer dry_run modes where available. 3) Inspect .multi-agent-profiles.json and other gene...

详细分析 ▾

✓ 用途与能力

Name, SKILL.md and included source files implement a multi‑agent orchestrator: executor, PollMonitor, aggregator, archiver, config validator, model selector, etc. The required behavior (workspace directories, spawning subagents via OpenClaw APIs, collecting/archiving outputs) matches the described functionality. There are no requests for unrelated external credentials or network endpoints.

ℹ 指令范围

Runtime instructions and SKILL.md describe environment checks, creation of a .openclaw/workspace layout, auto‑configuration, run/plan/archive/clean commands, and file‑based progress/aggregation. This is expected for this tool, but the skill will create configuration files and directories under the user's home and will delete files in the shared/ area when clean/archive operations run. The docs claim it waits for user confirmation before auto‑configuring, but several helper functions (e.g., loadJSON) will create files/directories if missing; commands such as init and archiveAndClean perform filesystem writes/deletes. Also the PollMonitor/collectResults behavior uses file polling which (per audits) contains a code path that can busy‑loop if not run with the expected cooperative yield — this is a functional risk (CPU/duration) rather than evidence of malicious intent.

✓ 安装机制

No external install/download URLs or package registries are used; the skill is delivered as source files included in the bundle (no install spec). That reduces supply‑chain surprise. All imports are local modules and Node.js built‑ins. No external arbitrary archive/extract from untrusted URLs was found.

✓ 凭证需求

The skill does not request secrets or unrelated environment credentials. It reads common environment variables (HOME/USERPROFILE, optionally OPENCLAW_VERSION) appropriate to its purpose. There are no requests for cloud keys or unrelated tokens in the metadata.

ℹ 持久化与权限

The skill is not always:true and is user‑invocable; it can run autonomously (disable-model-invocation is false) which is normal for skills. It writes persistent state under ~/.openclaw/workspace (creating profiles, agents/, shared/, archive/, progress files) and backs up/overwrites JSON files. It also deletes files in shared/ during clean operations. This persistence is consistent with a workspace manager but increases blast radius for accidental data deletion or inclusion of sensitive workspace files in archives.

⚠ test_hound_engine.mjs:114

Shell command execution detected (child_process).

⚠ lib/modelSelector.js:172

Dynamic code execution detected.

安全有层次，运行前请审查代码。

运行时依赖

无特殊依赖

版本

latestv7.2.02026/4/21

**Major update to v8.0 “multi-agent-orchestrator”, featuring core optimizations and enhanced reliability.** - Significantly reduced token usage with new prompt compression (~800 lines → ~80 lines), saving 25–30% on JSON workflows. - Introduced strict JSON output option and multiple failure boundary strategies (skip/fallback/partial/retry). - Added lightweight execution engine (`executorLite.js`) for efficient, structured outputs. - Enhanced environment checks and automated, user-guided configuration for seamless startup. - Improved command set and workflow directory validation, with clear recovery actions for missing configs. - Maintained cross-platform support (Windows/macOS/Linux) and robust compatibility with external configuration.

● 无害

安装命令

点击复制

官方npx clawhub@latest install multi-agent-config-manager

镜像加速npx clawhub@latest install multi-agent-config-manager --registry https://cn.longxiaskill.com

数据来源：ClawHub ↗ · 中文优化：龙虾技能库