Policy Engine

All policy decisions in TapPass are made by Open Policy Agent (OPA). a CNCF-graduated policy engine running as a sidecar. TapPass enforces OPA’s decisions. It has no built-in policy logic.

Fail-closed

If OPA is unreachable, all requests are denied. No request is ever forwarded unscanned.

Policy modules (Rego)

Module	Path	Purpose
authz.rego	tappass/authz	RBAC: who can access what endpoint
pipeline.rego	tappass/pipeline	Pipeline step config per agent/org
tools.rego	tappass/tools	Tool permissions: allow/deny per agent
routing.rego	tappass/routing	Model routing: classification → allowed models
breakglass.rego	tappass/breakglass	Emergency override authorization
guardrails.rego	tappass/guardrails	Detection response: block vs redact vs notify

Capability tokens

After the pipeline passes, TapPass mints an ES256-signed capability token:

• Time-bounded: 60s default TTL
• Action-scoped: lists allowed tools and operations
• Identity-bound: cryptographically bound to agent’s SPIFFE ID
• Constraint-carrying: embeds argument restrictions
• Verifiable offline: ~27μs via JWKS at /.well-known/jwks.json

OPA client features

• Decision caching: LRU with 30s TTL, 10K entries
• Sub-millisecond latency: sidecar deployment
• Audit logging: every decision logged with OPA decision ID

Policy replay

Test new policies against historical events without affecting production:

# Replay last 1000 events with a new policy
results = tp.policy_replay("new-policy-v2", events=1000)
print(f"Would have blocked {results.blocked_count} additional requests")