Routing policy

RoutingPolicy is a canonical input to routing. It is not an implementation-only config blob.

The schema models:

• optimization strategy
• compute preference
• required capabilities and modalities
• tool requirements
• endpoint and provider allow/deny lists
• budget controls
• privacy controls
• performance targets
• optional tie-break ordering

Strategy and compute preference

The baseline policy strategy enum is:

• balanced
• cost
• latency
• quality

The schema's compute_preference enum is:

• auto
• local
• remote
• hybrid

In the reference router, request flags such as preferLocal, computePreference, and denyRemote are folded into the effective policy snapshot before routing proceeds.

Hard-constraint fields

These fields participate directly in eligibility:

• required_capabilities
• required_modalities
• require_tools
• deny_endpoints
• allow_endpoints
• deny_provider_kinds
• allow_provider_kinds
• privacy.allow_remote

They determine which candidates may even remain in consideration.

Budget and privacy

The budget object states whether budget enforcement is enabled and what cost bounds apply. The reference router emits a budget_mode of strict when a request budget exists and disabled otherwise.

The privacy object currently includes allow_remote, which is how policy can force routing to stay local.

Targets and tie-breaks

The targets object models optimization goals such as:

• latency_target_ms
• latency_max_ms
• throughput_target_tps

The reference router currently snapshots a default tie-break order of:

1. prefer_local
2. lower cost
3. lower latency_ms_p95
4. stable endpoint_id

That snapshot is important because the final RouterDecision should explain the policy that was actually applied, not just the policy a caller remembers sending.

Policy snapshot semantics

The router copies the effective policy into the RouterDecision as policy_snapshot. This gives downstream consumers a stable record of:

1. what constraints were active
2. what optimization mode was in effect
3. what locality, budget, and target settings governed the decision