← 返回 Skills 市场

Isolated Multi-Agent Control Plane

Name: Isolated Multi-Agent Control Plane
Author: delibrately-cmyk

作者 delibrately-cmyk · GitHub ↗ · v0.1.2

cross-platform ✓ 安全检测通过

334

总下载

当前安装

版本数

在 OpenClaw 中安装

/install cerberus-agent-foundry

功能描述

Production-ready multi-agent architecture kit for OpenClaw. Provides isolated per-agent workspaces, control-plane orchestration, structured task lifecycle, c...

使用说明 (SKILL.md)

Cerberus Multi-Agent Control Plane

A professional blueprint for building, operating, and governing local multi-agent teams in OpenClaw. It ships a hardened protocol stack for role isolation, task orchestration, asynchronous handoffs, and auditability.

Quick Start

Install blueprint into target path:
- bash scripts/install_blueprint.sh /path/to/your/system
Verify task board:
- python3 /path/to/your/system/scripts/taskctl.py list
Verify mailbox:
- send: python3 /path/to/your/system/scripts/mailboxctl.py send --task-id TSK-000 --sender team-lead --receiver coder --correlation-id CORR-001 --body "ACK protocol"
- ack: python3 /path/to/your/system/scripts/mailboxctl.py status --message-id MSG-0001 --to ACK --actor coder

What This Blueprint Provides

Strict workspace isolation per agent (agents/*/workspace)
Unified control plane (control-plane/tasks, control-plane/mailbox, control-plane/logs)
Task state machine tooling (taskctl.py)
Mailbox protocol tooling with checksum and GC (mailboxctl.py)
Shared memory layer under control plane (control-plane/shared-memory)

Core Rules (must enforce)

Never use shared agent workspaces.
Use only control-plane/mailbox for agent-to-agent messaging.
Team Lead is sole mailbox garbage-collection authority.
Deploy actions require explicit human approval artifacts.

References

Read references/operations-checklist.md for rollout and audit checks.
Read assets/blueprint/docs/protocol-v1.md for protocol details.

安全使用建议

This blueprint appears coherent and local-only, but review these points before installing: - Installation writes the blueprint into whatever path you pass to scripts/install_blueprint.sh; do not run it as root or point it at system directories unless you intend to overwrite files. - Security is enforced by file-system layout and operational rules, not by cryptographic authentication: mailboxctl and taskctl accept an --actor parameter and will trust it. Ensure the control-plane directory has strict OS-level permissions and is only writable/executable by authorized accounts or run the tools under a confined service account. - The event log (control-plane/logs/events.jsonl) is append-only in design, but the scripts assume the logs directory exists and that directory permissions prevent tampering; create and protect it before use. - Test the blueprint in an isolated environment first to verify behavior (task/mailbox transitions, GC, archive behavior) and to confirm the access controls meet your needs. - If you need stronger guarantees (non-repudiation, authenticated actions), add process-level checks (e.g., verify invoking user UID, require signed request tokens, or run services behind an authenticated daemon) rather than relying on the --actor CLI argument. If you want, I can point out specific code lines where actor/identity is only string-checked and suggest concrete hardening changes.

功能分析

Type: OpenClaw Skill Name: cerberus-agent-foundry Version: 0.1.2 The OpenClaw AgentSkills skill bundle is designed to establish a secure, auditable, and isolated multi-agent system. The `SKILL.md` and other markdown files (e.g., `protocol-v1.md`, agent identity files) explicitly define security boundaries, least privilege, and operational governance for the AI agents, acting as defensive prompt injections to enforce secure behavior. The `install_blueprint.sh` script performs a local copy of files. The Python scripts (`mailboxctl.py`, `taskctl.py`) handle file operations (read, write, move) strictly within the defined `control-plane` directories, implement access controls (e.g., 'team-lead' for GC), and include data integrity checks (checksums). There is no evidence of data exfiltration, malicious execution, persistence mechanisms, or obfuscation.

能力评估

✓ Purpose & Capability

Name/description describe a local multi-agent control plane and the bundle contains templates and local CLI tooling (mailboxctl.py, taskctl.py, install script) that operate on a control-plane directory — this aligns with the stated purpose.

ℹ Instruction Scope

Runtime instructions only ask the user to copy the blueprint to a target path and run local Python scripts. The scripts operate solely on files under the installed control-plane tree and write audit events to events.jsonl. Important caveat: role/auth enforcement is convention-based (CLI arguments like --actor) and not cryptographically enforced — anyone with filesystem access or the ability to run the scripts can supply arbitrary actor values and modify files. The blueprint expects OS-level permissions / operational controls to enforce the policy.

✓ Install Mechanism

No external downloads or package installs; install_blueprint.sh copies bundled assets into a user-specified target directory and makes scripts executable. This is low-risk from a supply-chain perspective, but the install script will write into whatever target path you provide (so installing into system directories could overwrite files if misused).

✓ Credentials

The skill declares no environment variables or secrets and the code does not access external credentials. All operations are local filesystem reads/writes under the control-plane root; requested access is proportionate to the stated local control-plane functionality.

✓ Persistence & Privilege

always:false and normal autonomous invocation defaults apply. The skill does not attempt to modify other skills or system-wide agent settings. Its persistence is limited to files it installs/creates under the chosen target directory.

如何使用

确保已安装 OpenClaw（本地或 Docker 部署）
在对话框中输入安装命令：/install cerberus-agent-foundry
安装完成后，直接呼叫该 Skill 的名称或使用 /cerberus-agent-foundry 触发
根据 Skill 的参数说明提供必要输入，即可获得结构化输出

版本历史

v0.1.2

Rename project title to a function-first professional name; keep architecture and protocol stack unchanged.

v0.1.1

Branding and documentation upgrade: professional project positioning, clearer architecture overview, and privacy hardening (replaced personal approver label with generic human approver token).

v0.1.0

Initial release: isolated multi-agent local protocol v1 with task state machine, mailbox protocol + GC, shared-memory layer, and bootstrap scripts.

元数据

Slug cerberus-agent-foundry

版本 0.1.2

许可证 —

累计安装 0

当前安装数 0

历史版本数 3

常见问题