agentmesh-governance

由 AgentMesh 提供支持的 AI 代理治理、信任评分和策略执行。 在以下情况下激活：(1) 用户想要强制实施令牌限制、工具限制或内容 有关代理操作的策略，(2) 在委派之前检查代理的信任评分或 协作，(3) 使用 Ed25519 加密 DID 验证代理身份， (4) 使用防篡改的 Merkle 链日志审计代理行为， (5) 用户询问代理安全、治理、合规性或信任。 企业级：1,600+测试，合并到Dify (65K★)、LlamaIndex (47K★)、 微软代理-闪电(15K★)。

## 概述（中文）

由 AgentMesh 提供支持的 AI 代理治理、信任评分和策略执行。
在以下情况下激活：(1) 用户想要强制实施令牌限制、工具限制或内容
有关代理操作的策略，(2) 在委派之前检查代理的信任评分或
协作，(3) 使用 Ed25519 加密 DID 验证代理身份，
(4) 使用防篡改的 Merkle 链日志审计代理行为，
(5) 用户询问代理安全、治理、合规性或信任。
企业级：1,600+测试，合并到Dify (65K★)、LlamaIndex (47K★)、
微软代理-闪电(15K★)。

## 原文

# AgentMesh Governance — Trust & Policy for OpenClaw Agents

Zero-trust governance layer for OpenClaw agents. Enforce policies, verify identities,
score trust, and maintain tamper-evident audit logs — all from your agent's command line.

## Setup

Install the AgentMesh governance CLI:

```bash
pip install agentmesh-governance
```

> If `agentmesh-governance` is not yet on PyPI, install directly from source:
> ```bash
> pip install "agentmesh @ git+https://github.com/imran-siddique/agent-mesh.git"
> ```

## Scripts

All scripts are in `scripts/`. They wrap the governance engine and output JSON results.

### Check Policy Compliance

Evaluate an action against a governance policy before execution:

```bash
scripts/check-policy.sh --action "web_search" --tokens 1500 --policy policy.yaml
```

Returns JSON with `allowed: true/false`, any violations, and recommendations.
Use this **before** executing any tool call to enforce limits.

### Get Trust Score

Check an agent's current trust score (0.0 – 1.0):

```bash
scripts/trust-score.sh --agent "research-agent"
```

Returns the composite trust score with breakdown across 5 dimensions:
policy compliance, resource efficiency, output quality, security posture,
collaboration health.

### Verify Agent Identity

Verify an agent's Ed25519 cryptographic identity before trusting its output:

```bash
scripts/verify-identity.sh --did "did:agentmesh:abc123" --message "hello" --signature "base64sig"
```

Returns `verified: true/false`. Use when receiving data from another agent.

### Record Interaction

Update trust scores after collaborating with another agent:

```bash
scripts/record-interaction.sh --agent "writer-agent" --outcome success
scripts/record-interaction.sh --agent "writer-agent" --outcome failure --severity 0.1
```

Success adds +0.01 to trust score. Failure subtracts the severity value.
Agents dropping below the minimum threshold (default 0.5) are auto-blocked.

### Audit Log

View tamper-evident audit trail with Merkle chain verification:

```bash
scripts/audit-log.sh --last 20
scripts/audit-log.sh --agent "research-agent" --verify
```

The `--verify` flag checks Merkle chain integrity — any tampering is detected.

### Generate Identity

Create a new Ed25519 cryptographic identity (DID) for your agent:

```bash
scripts/generate-identity.sh --name "my-agent" --capabilities "search,summarize,write"
```

Returns your agent's DID, public key, and capability manifest.

## Policy File Format

Create a `policy.yaml` to define governance rules:

```yaml
name: production-policy
max_tokens: 4096
max_tool_calls: 10
allowed_tools:
  - web_search
  - file_read
  - summarize
blocked_tools:
  - shell_exec
  - file_delete
blocked_patterns:
  - "rm -rf"
  - "DROP TABLE"
  - "BEGIN CERTIFICATE"
confidence_threshold: 0.7
require_human_approval: false
```

## When to Use This Skill

- **Before tool execution**: Run `check-policy.sh` to enforce limits
- **Before trusting another agent's output**: Run `verify-identity.sh`
- **After collaboration**: Run `record-interaction.sh` to update trust
- **Before delegation**: Check `trust-score.sh` — don't delegate to agents below 0.5
- **For compliance**: Run `audit-log.sh --verify` to prove execution integrity
- **On setup**: Run `generate-identity.sh` to create your agent's DID

## What It Enforces

| Policy | Description |
|--------|-------------|
| Token limits | Cap per-action and per-session token usage |
| Tool allowlists | Only explicitly permitted tools can execute |
| Tool blocklists | Dangerous tools are blocked regardless |
| Content patterns | Block regex patterns (secrets, destructive commands, PII) |
| Trust thresholds | Minimum trust score required for delegation |
| Human approval | Gate critical actions behind human confirmation |

## Architecture

This skill bridges the OpenClaw agent runtime with the [AgentMesh](https://github.com/imran-siddique/agent-mesh)
governance engine:

```
OpenClaw Agent → SKILL.md scripts → AgentMesh Engine
                                     ├── GovernancePolicy (enforcement)
                                     ├── TrustEngine (5-dimension scoring)
                                     ├── AgentIdentity (Ed25519 DIDs)
                                     └── MerkleAuditChain (tamper-evident logs)
```

Part of the [Agent Ecosystem](https://imran-siddique.github.io):
[AgentMesh](https://github.com/imran-siddique/agent-mesh) ·
[Agent OS](https://github.com/imran-siddique/agent-os) ·
[Agent SRE](https://github.com/imran-siddique/agent-sre)