Status: Active
Version: 1.0
Owner: <workspace owner / team>
Scope: All OpenClaw agents, sub-agents, and automated workflows in this workspace.
Governed by: AI SAFE² Framework v2.1 · Dual License MIT + CC-BY-SA
Review cycle: Quarterly, or after any significant incident.
This document is the constitutional layer for this workspace. It binds every agent, sub-agent, skill, and automated workflow to a shared set of accountability rules. Individual agents have their own SOUL.md and AGENTS.md — this policy governs how they relate to each other and to the workspace as a whole.
An agent that has not acknowledged this policy (via its AGENTS.md reference to this file) is not authorized to operate in this workspace.
Three non-negotiable outcomes this policy exists to protect:
Every agent operating in this workspace must maintain these files:
| File | Purpose | Change Control |
|---|---|---|
SOUL.md |
Values, alignment bands, hard limits | Security lead review required |
AGENTS.md |
Operating rules, tool use, sub-agent strategy | PR + workspace owner review |
IDENTITY.md |
Minimal identity anchor — loads every request | PR review |
USER.md |
Human identity, preferences, hard lines | PR review |
MEMORY.md |
Stable long-term facts — private sessions only | Human-edit only; agent may propose |
TOOLS.md |
Environment config — IDs, paths, secret locations | PR review |
HEARTBEAT.md |
Periodic health check protocol | PR review |
SUBAGENT-POLICY.md |
Worker governance rules | Security lead review |
MODEL-ROUTER.md |
Multi-LLM routing policy | Security lead review |
memory/YYYY-MM-DD.md |
Append-only daily log | Agent-written; human-reviewable |
OPENCLAW-WORKSPACE-POLICY.md |
This file | Security lead + second reviewer |
Filing convention: All identity and policy files live in the agent’s workspace root. Memory files live in memory/. Skills live in skills/.
Authorization state for each agent is recorded in:
memory/agents-registry.md — maintained by workspace owner, not by agents.
Human (Workspace Owner)
└── Primary Agent (full SOUL/AGENTS context)
└── Sub-Agents (task-scoped context, minimum trust tier)
└── Skill Scripts (sandboxed, no memory access)
Trust flows down this hierarchy only. A skill cannot grant itself agent-level trust. A sub-agent cannot grant itself primary-agent trust. Any attempt to cross this hierarchy upward is a security signal.
MEMORY.md and memory/**/*.md are loaded in private/single-user sessions only.~/.agent/.env → GATEWAY_TOKEN).All agents in this workspace enforce the three-tier data classification from USER.md:
| Tier | Scope |
|---|---|
| Confidential | DM / private session only |
| Internal | Group channels OK — no external sharing |
| Restricted | External only with explicit per-instance user approval |
When context is ambiguous, agents default to the more restrictive tier.
All outbound content (messages, emails, API calls) passes through PII scanning before delivery. At minimum, the agent scans for: personal email addresses, phone numbers, dollar amounts, credit card patterns, and high-entropy strings matching API key formats. Findings are redacted and flagged.
| Source | Trust Level | Review Required |
|---|---|---|
Official openclaw/skills repo |
Provisionally trusted | Basic provenance check |
AI SAFE² framework examples/ |
Trusted | No additional review |
| Workspace-authored skills | Trusted | PR review |
| ClawHub community registry | Untrusted by default | Full security review before use |
| External repos / gists | Untrusted | Full security review + workspace owner approval |
Before any skill is enabled in production:
memory/agents-registry.md with source, date, and reviewer.No skill is deployed via “just install it and see.” ClawHub ranking is not a safety signal.
All agents in this workspace route LLM API calls through the AI SAFE² control gateway.
http://localhost:8000/v1 (see TOOLS.md)The gateway is the external enforcement layer. Even if an agent’s internal controls fail, the gateway is the last line of defense. Removing or bypassing the gateway is a critical security violation.
This workspace policy supports alignment with the following standards:
| Standard | How This Policy Addresses It |
|---|---|
| ISO 42001 | Mandatory policy documentation, change control, quarterly review cycle |
| NIST AI RMF | Risk identification (SOUL.md hard limits), response (Red band protocol), recovery (Fail-Safe pillar) |
| SOC 2 | Immutable audit logging via gateway, access control (trust hierarchy), incident response (HEARTBEAT escalation) |
| GDPR / HIPAA | PII redaction, data classification tiers, private memory isolation, secret handling |
| AI SAFE² v2.1 | Full five-pillar implementation across all agents in this workspace |
When an incident involves this workspace (not just one agent):
memory/YYYY-MM-DD.md for all affected agents.memory/YYYY-MM-DD.md and promoted to memory/workspace-changelog.md.Governed by the AI SAFE² Framework v2.1.
AI SAFE² = Sanitize & Isolate · Audit & Inventory · Fail-Safe & Recovery · Engage & Monitor · Evolve & Educate
Love Equation alignment: dE/dt = β(C − D)E — cooperation must exceed defection, always.