[Tanguille]

What

This adds a second way for dataPlane: AgentGateway routers to work: agentGatewayRef.routing: Model.

Today an AgentGateway router creates an InferencePool for every model, needs an Endpoint Picker, and puts a body-parsing policy on the whole Gateway. With routing: Model it uses agentgateway's built-in model router instead (agentgateway v1.6.0 or later). The router creates:

• one HTTPRoute that sends the listener's traffic to the model router
• one AgentgatewayModel per backend, which points a model name at the InferenceService's Service

agentgateway reads the model from the request body, answers GET /v1/models itself, and returns a 404 for unknown models. You don't need the Inference Extension CRDs or an Endpoint Picker.

spec:
  dataPlane: AgentGateway
  agentGatewayRef:
    name: ai-gateway
    routing: Model
    sectionName: llm
  backends:
    - name: chat
      displayName: chat-large
      inferenceServiceRef: {name: chat-vllm}
    - name: reranker
      inferenceServiceRef: {name: reranker}

The default stays InferencePool, so existing routers don't change.

Why

Part of #2151 (phase 1), following the plan in the maintainer's comment. Phase 2 (fallback between backends, aliases with request defaults) will be a separate PR.

How

Going through the points in that comment:

• Opt-in: routing defaults to InferencePool, and endpointPicker is rejected in Model mode. Since feat(router): per-backend Endpoint Picker for dataPlane: AgentGateway #2157, InferencePool mode already checks for a picker in the webhook and the reconciler, so I left that check there instead of adding a CEL rule.
• Names: the model name clients send (displayName, else name) goes into match.model. The object name comes from the router and the backend (<router>-<backend>), so names like Qwen3.5-2B or org/model work.
• Ownership: the router owns everything it creates and won't take over an object it didn't create. It deletes the model of a removed backend, and cleans up the other mode's objects when you switch routing.
• The catch-all route: Model mode requires sectionName or hostnames, so it can't take over a whole Gateway.
• CRD check: when the AgentgatewayModel CRD is missing, the router reports AgentgatewayModelCRDNotInstalled and doesn't fail.
• Tests: see Testing below.

The API formats come from the InferenceService:

• llama.cpp chat gets Completions.
• vLLM chat gets Completions, Messages, AnthropicTokenCount and Responses.
• Embedding models get Embeddings, and rerankers Rerank.
• Any other runtime gets a clear condition instead of a guess.

Anything Model mode can't do yet is rejected at apply time and by the reconciler (UnsupportedInModelRouting): rules, defaultRoute, external backends, resolution: endpoint, per-backend pickers, and duplicate or wildcard model names.

Testing

• Unit and envtest coverage:
  • the generated objects, per runtime and format
  • name derivation
  • CEL and webhook validation
  • the InferencePool default rendering unchanged
  • refusing to adopt foreign objects
  • the CRD check
  • pruning removed backends
  • switching modes in both directions
• make test-e2e-rbac passes. The routing: Model fixture goes Ready and creates its route and model under the generated RBAC.
• Live check against agentgateway v1.6.0:

AI assistance

Checklist

• Tests added/updated
• make test passes locally
• make lint passes locally
• Commit messages follow conventional commits
• All commits are signed off (git commit -s) per DCO
• AI assistance (if any) is disclosed above, per CONTRIBUTING.md
• Documentation updated (if user-facing change)