# Bound Upstream Dispatch
#
# Selects an upstream endpoint straight from the logical binding, with no
# second router. A binding router publishes one logical cluster; a
# load_balancer with `cluster_source: bound_upstream` then resolves that
# frozen binding and picks an endpoint. This drives one pipeline with two
# ownership modes: an OpenAI-tagged request is dispatched directly by a
# branch, while every other request runs gateway-owned processing and an
# iterative_request_router that also dispatches from the same binding.
#
# Requests under /openai/ bind the openai cluster and take the direct
# branch, so gateway filters and the IRR are skipped. Every other request
# binds the chat cluster, gains an X-Gateway-Processed marker, and is
# dispatched by the IRR step's bound-consuming load balancer.
#
# Usage:
#   cargo run -p praxis-proxy --features iterative-request-router -- -c examples/configs/traffic-management/bound-upstream-dispatch.yaml
#   curl -i http://localhost:8080/openai/v1/responses   # direct dispatch
#   curl -i http://localhost:8080/v1/chat/completions   # gateway + IRR dispatch

listeners:
  - name: default
    address: "127.0.0.1:8080"
    filter_chains:
      - main

filter_chains:
  - name: main
    filters:
      # 1. The router binds one logical cluster from a trusted routing fact
      #    (here the request path). It selects no endpoint; the binding
      #    carries only the logical cluster name and its catalog metadata.
      - filter: router
        routes:
          - path_prefix: "/openai/"
            cluster: openai_backend

          - path_prefix: "/"
            cluster: chat_backend

      # 2. Direct provider path: gated on the bound provider, this branch
      #    dispatches straight from the binding via a bound-consuming load
      #    balancer and rejoins `terminal`, so the gateway processing and
      #    IRR below never run for a provider-owned request. The headers
      #    filter intentionally has no mutations; it is the branch host.
      #
      #    The IRR's bounded StreamBuffer is a pipeline-wide capability, so
      #    direct requests are still fully pre-read and share its body limit
      #    before routing decides which path runs.
      - filter: headers
        conditions:
          - when:
              bound_upstream:
                application_provider: openai
        branch_chains:
          - name: direct-openai
            rejoin: terminal
            chains:
              - name: direct-openai-dispatch
                filters:
                  - filter: load_balancer
                    cluster_source: bound_upstream
                    clusters:
                      - name: openai_backend
                        http:
                          application_protocol: openai_responses
                          application_provider: openai
                        endpoints:
                          - "127.0.0.1:3001"

      # 3. Gateway-owned processing. The direct branch above rejoins
      #    `terminal`, so a provider-owned request never reaches here;
      #    only the non-provider path is marked.
      - filter: headers
        response_set:
          - name: "X-Gateway-Processed"
            value: "true"

      # 4. The iterative_request_router owns the non-provider exchange
      #    lifecycle. Its step selects an endpoint from the same frozen
      #    binding through a bound-consuming load balancer: no second
      #    router, no top-level load balancer. Provider-owned requests
      #    already terminated at the direct branch, which is also why the
      #    IRR carries no bound_upstream condition of its own. It could
      #    not: the IRR pre-reads the request body, and validation rejects
      #    a bound condition on a pre-read body hook.
      - filter: iterative_request_router
        initial_step: inference
        max_state_bytes: 65536
        steps:
          - name: inference
            filters:
              - filter: load_balancer
                cluster_source: bound_upstream
                clusters:
                  - name: chat_backend
                    http:
                      application_protocol: openai_chat_completions
                      application_provider: vllm
                    endpoints:
                      - "127.0.0.1:3002"
            on_result:
              - default: true
                done: true

insecure_options:
  allow_private_endpoints: true # example proxies to local backends