arp-policy

A minimal policy evaluator for the ARP Tool Registry.

Policy format

Top-level keys:

• Version: optional version string.
• Statement: list of statements.

Statement keys:

• Sid: optional statement id for audit/debug.
• Effect: Allow or Deny.
• Action: string or list of strings.
• Resource: string or list of strings.
• Condition: optional map of operator to key/value map.

Supported condition operators:

• StringEquals
• StringLike (wildcards with *)

Usage

from arp_policy import Enforcer, Policy, emit_decision

policy = Policy.load("./config/policy/policy.dev.json")
enforcer = Enforcer(policy)

context = {
    "principal": "user:alice",
    "tenant": "acme",
    "environment": "dev",
    "request_id": "req-123",
}

decision = enforcer.authorize("tool:Invoke", "tool:finance/pay", context)
if not decision.allowed:
    raise Exception("policy denied")

# Audit event
record = emit_decision(
    decision,
    context,
    action="tool:Invoke",
    resource="tool:finance/pay",
    policy_hash=policy.policy_hash,
)

Auth and identity

arp-policy is auth-agnostic. It does not validate credentials or issue identities.

The host service (Tool Registry, Runtime, Daemon) must:

• authenticate the caller using your chosen scheme (JWT, mTLS, API key, etc.)
• map verified identity attributes into the context dict (e.g., principal, tenant)
• avoid passing raw credentials into policy context

Tool discovery filtering

tools = [{"tool_id": "finance.pay", "metadata": {"labels": {"tier": "gold"}}}]
allowed = enforcer.filter_tools(tools, context)

Context keys

Typical keys used in policy conditions:

• principal
• tenant
• environment
• request_id
• tool.id
• tool.tags (tool labels/annotations)