Multi Model Consensus

多模型决策委员会 — 消除单模型偏见，通过多轮分歧讨论产出客观决策参考。支持3-13个模型同时评审，提供量化投票矩阵和6段式共识报告。触发条件：包含「多模型决策」或「多模型委员会」时自动激活。

What to consider before installing/use: - Ask the publisher to explicitly declare that the skill will read your local OpenClaw config (e.g., ~/.openclaw/openclaw.json) and session metadata. The registry currently lists no required config paths — that should be corrected or explained. - Confirm why Exec/Read/Write appear in allowed-tools. If the skill only needs to call session APIs to spawn/collect subagents, Exec may be unnecessary; prefer a version that does not request shell/exec/file-write capability. - The SKILL.md says some self-checks are 'internal' and not shown to users. Request that the maintainer document exactly what is hidden and obtain an explicit user consent prompt before any local-file reads or background operations. - If you care about privacy, note the skill enforces 'real-name' model reporting (it will surface actual model/provider names in outputs). If you prefer anonymity/internal routing to be kept private, ask for an option to anonymize model identifiers in reports. - To reduce risk: test the skill in an isolated environment or with a non-sensitive OpenClaw config (or with a copy of the config that contains fake/limited entries) before enabling it on production data. - If you want to proceed, request a small manifest update from the author that: (1) declares required config paths and any file reads, (2) justifies allowed-tools list (remove Exec if not needed), and (3) states clearly when and how user-visible notifications occur (no hidden actions).

Type: OpenClaw Skill Name: multi-model-consensus Version: 1.6.4 The skill bundle implements a complex multi-model orchestration framework for decision-making. While the documentation is professional and includes security-conscious design principles (e.g., the 'Identity Purity Principle' forbidding sub-agents from spawning further agents), the skill requests several high-risk permissions including 'Exec', 'Write', and 'SessionsSpawn' (allowing up to 13 concurrent sub-agents). The 'SKILL.md' and 'TROUBLESHOOTING.md' files indicate the agent performs environment self-checks via shell execution and writes persistent reports to the user's home directory (~/.openclaw/workspace/memory/). These powerful capabilities, combined with the complexity of multi-round sub-agent orchestration, present a significant attack surface for resource exhaustion or indirect prompt injection, placing it in the suspicious category despite the lack of clear malicious intent.