Platform
The policy engine
One function decides whether a tool call proceeds: evaluate(). It is fast, deterministic, explainable, and it defaults to deny.
Decisions
ALLOW
A matching policy permits the call.
DENY
The call is refused. This is the default when nothing matches a risky tool.
AUDIT
The call is permitted but recorded for review.
NO_POLICY
No policy governs this tool at this tier.
What the engine looks at
| Signal | Source | Effect |
|---|---|---|
| Tool known? | Tool catalogue | Unknown tools are denied outright. |
| Risk level | Tool catalogue | HIGH and CRITICAL tools are denied with no policy present. |
| Data tier | Request context | PUBLIC < INTERNAL < CONFIDENTIAL < RESTRICTED. A tool may be barred from a higher tier. |
| Matching policy | Policy store | Conditions on agent, tool, targets and time windows. |
| Approval | Tool catalogue | Some tools require an approval before the call proceeds. |
| Rate and frequency | Engine counters | Windowed limits on call frequency and volume. |
Hot reload
Policies reload without a restart through
/api/policies/hot, so an operator can tighten a rule in response to something they have just seen.
Extensions cannot change a verdict
The extension point hands observers a deep copy of the request and the decision. An extension that throws is recorded, never propagated. This is deliberate: an extension that could flip a decision would be a policy bypass, so the contract test asserts it cannot.