Proposal: add a policy decision layer for agent actions

[aparey]

Proposal

Add a small deterministic policy decision layer for agent actions.

This is an off-menu layer proposal related to PR #110. The intent is not to replace auth, privacy, trust, payments, or coordination. It adds a narrow decision point an agent can call before taking a risky action.

Why this belongs in NANDA Town

Existing layers answer adjacent questions:

• auth: who is allowed to hold a credential?
• privacy: who can see a message?
• payments: can value move?
• trust: how reliable is a counterparty?

A policy layer answers a different question: should this concrete action be permitted, denied, or require approval right now?

That matters before agents publish data, spend funds, call external tools, or mutate authority-bearing resources.

Proposed shape

Core interface:

• PolicyRequest: actor, action, resource, data classes, optional amount, purpose, metadata
• PolicyDecision: permit, deny, or approval_required
• Policy: runtime-checkable protocol with decide(request, *, now)

Reference plugins:

• strict_rules: deny-by-default policy
• allow_all: intentionally unsafe foil for validators

Scenario:

• policy_guard
• Plugin selected from task.config.policy_plugin, not from global LayerConfig
• Validators check safe reads, sensitive public export denial, high-value payment approval, unknown-amount payment approval, and deny-by-default admin behavior

PR

Prototype: #110