← Back to Skills Marketplace
mupengi-bot

GLM Swarm

by mupengi-bot · GitHub ↗ · v1.1.1 · MIT-0
cross-platform ⚠ suspicious
96
Downloads
0
Stars
0
Active Installs
4
Versions
Install in OpenClaw
/install glm-swarm
Description
경량 모델 병렬 하네스. 명시적으로 swarm/병렬 처리를 요청하거나, AGENTS.md 하네스 규칙에 의해 복합 작업(도구 3회+, 독립 하위작업 2개+)이 감지되었을 때만 사용. 단순 질답, 번역, 요약, 단일 도구 호출에는 절대 사용하지 않는다.
README (SKILL.md)

GLM Swarm

경량 모델 복합 작업 병렬 하네스. 상세 패턴은 references/patterns.md, Context Packet 규칙은 references/context-packet.md 참조.

복잡도 게이트 (필수)

swarm 진입 전 반드시 판단:

  • 독립 하위 작업 2개 미만 → 직접 처리 (swarm 안 씀)
  • 도구 호출 3회 미만 → 직접 처리
  • 위 둘 다 해당 → swarm 모드

실행 흐름

0. bash scripts/planner.sh {task-id} → /tmp/swarm/{task-id}/ 초기화
1. Planner (1회) — 패턴 매칭(A~E) → atomic task 분해
2. Context Packet — worker별 최소 맥락 압축 (출력 300토큰 이내 명시)
3. Worker Pool — sessions_spawn 병렬 (최대 6개, lightContext: true)
4. Shared Scratchpad — /tmp/swarm/{id}/shared.md
5. Aggregator — 구조화 합산 (섹션별 분류 + 핵심 인사이트 3개 + 액션 아이템)
6. bash scripts/cleanup.sh {task-id} → /tmp/swarm/{id}/ 삭제

패턴 요약

패턴 트리거 구조
A 조사→판단→실행 t1∥t2 → t3 → t4
B 다중 수집→합성 t1∥t2∥t3∥t4 → t5
C 반복 N건 병렬 t1∥...∥tN → concat
D 분석→제안 t1∥t2 → t3 → t4
E 검증→수정→배포 순차 (파일 충돌 방지)

매칭 안 되면 동적 분해. 상세: references/patterns.md

Worker 스폰

sessions_spawn({
  task: "{context_packet}\
출력은 300토큰 이내로 작성해.",
  mode: "run", runtime: "subagent",
  lightContext: true,
  label: "swarm-{task-id}-{worker-id}"
})

Planner 초기화: bash scripts/planner.sh {task-id} 정리: bash scripts/cleanup.sh {task-id}|--all|--list

안전

  • Worker: MEMORY/SOUL/AGENTS.md 편집 금지, ~/.secrets/ 접근 금지
  • 외부 행동 → Aggregator 후 승인 요청
  • 2회 실패 → 해당 task 스킵. 50%+ 실패 → swarm 중단

Aggregator 규칙

결과 합산 시 반드시 포함:

  1. 섹션별 분류 — 결과를 주제별로 정리
  2. 핵심 인사이트 3개 — 가장 중요한 발견
  3. 액션 아이템 — 즉시 실행 가능한 다음 스텝
  4. 출처 표기 — 각 판단에 어떤 worker 결과 기반인지 명시

컨텍스트 최적화 (v1.1)

문제: worker 결과 전문이 메인 세션에 주입되면 컨텍스트 누적 해결: Aggregator는 메인에 요약만 반환

Aggregator 출력:
1. 메인 세션: 요약 3~5줄 (핵심 발견 + 액션 아이템만)
2. 상세 결과: memory/swarm-results/{task-id}.md 에 저장
3. 메인은 상세가 필요하면 해당 파일을 읽어옴

Aggregator 스킵 조건에 추가:

  • worker 결과 총합이 500토큰 이하면 → 전문 반환 (요약 불필요)
  • 500토큰 초과면 → 요약 반환 + 파일 저장
Usage Guidance
This skill appears to implement the advertised swarm orchestration, but review and harden the included scripts before use. Specific recommendations: 1) Do not run cleanup.sh or planner.sh as root; test in an isolated container or ephemeral VM. 2) Add input validation/sanitization to planner.sh and cleanup.sh to reject ../ or absolute paths (or restrict TASK_ID to a safe whitelist/regex). 3) Prefer generating task IDs server-side (by the planner) rather than accepting arbitrary user-provided IDs. 4) Confirm your platform enforces the SKILL.md prohibitions (workers must not be able to access ~/.secrets or edit MEMORY.md). 5) If you allow autonomous agent invocation, limit its scope or monitor runs until you are confident the sanitization and access controls are correct.
Capability Analysis
Type: OpenClaw Skill Name: glm-swarm Version: 1.1.1 The skill is a multi-agent orchestration harness designed to decompose complex tasks into parallel sub-tasks using a 'swarm' pattern. It utilizes local shell scripts (scripts/planner.sh and scripts/cleanup.sh) to manage a temporary workspace in /tmp/swarm/ and provides structured protocols for sub-agent communication. The bundle includes explicit safety instructions in SKILL.md and references/context-packet.md that forbid sub-agents from accessing sensitive directories like ~/.secrets/ or modifying system files, demonstrating a defensive posture rather than malicious intent.
Capability Assessment
Purpose & Capability
Name/description (lightweight parallel model harness) match the included instructions and files: a planner initializes /tmp/swarm, workers use context packets and a shared scratchpad, and results are aggregated. No unrelated credentials, binaries, or external services are requested.
Instruction Scope
Runtime instructions require creating and manipulating /tmp/swarm, spawning subagents (sessions_spawn) and writing/reading worker files. The included cleanup.sh takes an arbitrary task-id and does rm -rf ${SWARM_BASE}/${1} without sanitization — a malicious or malformed task-id could enable path traversal and deletion outside /tmp/swarm. The SKILL.md forbids accessing ~/.secrets and system memory files, but those are only advisory and not enforced by the scripts.
Install Mechanism
No install spec; scripts are bundled in the skill and nothing is downloaded or executed from external URLs. This is low-risk from an installation code-fetch perspective.
Credentials
The skill requests no environment variables, no credentials, and no config paths. Its resource access (temporary /tmp directories and a 'memory/...' result path) is proportionate to a local orchestration tool.
Persistence & Privilege
always:false and no special OS restrictions. The skill can be invoked autonomously (disable-model-invocation:false) and spawns subagents; autonomous invocation combined with the unsafe cleanup behavior raises the blast radius if an agent were to pass untrusted input. This is not automatically malicious but worth caution.
How to Use
  1. Make sure OpenClaw is installed (local or Docker)
  2. Run the install command in chat: /install glm-swarm
  3. After installation, invoke the skill by name or use /glm-swarm
  4. Provide required inputs per the skill's parameter spec and get structured output
Version History
v1.1.1
v1.1.1: Aggregator 컨텍스트 최적화
v1.1.0
v1.1: Aggregator 강화(구조화 합산+출처), 토큰 밸런싱(300토큰 제한), planner.sh 문서화, 컨텍스트 최적화(요약 주입+상세 파일 저장)
v1.0.1
fix: remove real Discord channel ID from context-packet example
v1.0.0
Initial release: parallel harness for lightweight models. 5 decomposition patterns, Context Packet protocol, Shared Scratchpad.
Metadata
Slug glm-swarm
Version 1.1.1
License MIT-0
All-time Installs 0
Active Installs 0
Total Versions 4
Frequently Asked Questions

What is GLM Swarm?

경량 모델 병렬 하네스. 명시적으로 swarm/병렬 처리를 요청하거나, AGENTS.md 하네스 규칙에 의해 복합 작업(도구 3회+, 독립 하위작업 2개+)이 감지되었을 때만 사용. 단순 질답, 번역, 요약, 단일 도구 호출에는 절대 사용하지 않는다. It is an AI Agent Skill for Claude Code / OpenClaw, with 96 downloads so far.

How do I install GLM Swarm?

Run "/install glm-swarm" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.

Is GLM Swarm free?

Yes, GLM Swarm is completely free, licensed under MIT-0. You can download, install and use it at no cost.

Which platforms does GLM Swarm support?

GLM Swarm is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).

Who created GLM Swarm?

It is built and maintained by mupengi-bot (@mupengi-bot); the current version is v1.1.1.

💬 Comments