Deployment and policy routing

Deployment aliases decouple client-facing model names from provider model IDs. Define the provider model, supported capabilities, and deployment-specific fallbacks once:

llm:
  default_model: general
  allowed_models: [openai/gpt-4o, anthropic/claude-sonnet-4-6]
  deployments:
    general:
      model: openai/gpt-4o
      capabilities: [chat, responses, streaming, tools, structured_outputs]
      fallback_models: [anthropic/claude-sonnet-4-6]
    vision:
      model: openai/gpt-4o
      capabilities: [chat, responses, streaming, vision]

Clients send model: general or model: vision. GET /v1/models exposes deployment aliases as Relay-owned models.

Versioned policies

routing:
  active_policy_version: "2026-07-30"
  require_declared_capabilities: true
  policies:
    "2026-07-30":
      default_deployment: general
      allowed_deployments: [general, vision]
      allowed_capabilities: [chat, responses, embeddings, streaming, tools, vision, structured_outputs]
      denied_capabilities: [tool:computer_use]
      capability_routes:
        vision: vision
      team_overrides:
        team-restricted-uuid:
          allowed_deployments: [general]

Use model: auto to select capability_routes, then the default deployment or first compatible deployment. Direct model IDs remain backward compatible unless require_declared_capabilities is enabled.

The selected version, requested model, physical model, deployment, and inferred capabilities are recorded in usage audit metadata, Langfuse metadata, OpenTelemetry spans, Prometheus routing metrics, and response headers.