Skip to content

SKILL.md

Latest commit

 

History

History
170 lines (127 loc) · 5.46 KB

SKILL.md

File metadata and controls

170 lines (127 loc) · 5.46 KB
name agentmesh-governance
description AI agent governance, trust scoring, and policy enforcement powered by AgentMesh. Activate when: (1) user wants to enforce token limits, tool restrictions, or content policies on agent actions, (2) checking an agent's trust score before delegation or collaboration, (3) verifying agent identity with Ed25519 cryptographic DIDs, (4) auditing agent actions with tamper-evident hash chain logs, (5) user asks about agent safety, governance, compliance, or trust. Enterprise-grade: 1,600+ tests, merged into Dify (65K★), LlamaIndex (47K★), Microsoft Agent-Lightning (15K★).
version 1.1.0
metadata
openclaw

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:

pip install agentmesh-governance

If agentmesh-governance is not yet on PyPI, install directly from source:

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:

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):

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:

scripts/verify-identity.sh --did "did:mesh: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:

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 hash chain verification:

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

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

Generate Identity

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

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:

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 governance engine:

OpenClaw Agent → SKILL.md scripts → AgentMesh Engine
                                     ├── GovernancePolicy (enforcement)
                                     ├── RewardService (5-dimension scoring)
                                     ├── AgentIdentity (Ed25519 DIDs)
                                     └── AuditLog (tamper-evident Merkle chains)

Part of the Agent Governance Toolkit: AgentMesh · Agent OS · Agent SRE