Format Routing

# Responses Format Routing
#
# Routes AI API traffic by request-head operation identity and body format.
# The `openai_operation` filter classifies supported OpenAI operations from
# the HTTP method and path alone, without reading the body. Chat Completions
# traffic branches on `application_protocol: openai_chat_completions`.
#
# The `openai_responses_format` filter parses JSON request bodies and
# classifies them as `openai_responses` (Responses API) or `unknown` (valid
# JSON but unrecognized format). Chat Completions routing no longer depends
# on body heuristics for `/v1/chat/completions`.
#
# POST /v1/responses (create) is authoritative: a valid create body that
# omits every discriminator the body heuristics key on (`input`, `prompt`
# object, `previous_response_id`, `conversation`) — e.g. `{"model":"gpt-5"}` —
# is classified as `openai_responses` on that endpoint rather than
# `unknown`, so it routes to the responses cluster below instead of
# falling through to the default one.
#
# Classification facts are promoted to internal routing headers
# (x-praxis-ai-*) that the router matches for cluster selection.
# These internal headers are stripped before upstream by header hygiene.
#
# Non-JSON, invalid JSON, and unrecognized JSON bodies continue by
# default so mixed-traffic listeners do not reject non-AI requests.
# Set `on_invalid: reject` for listeners that only accept AI API
# traffic (rejects anything without `input`, except on the
# authoritative POST /v1/responses create endpoint). Chat Completions
# is identified from the request head and never reaches this filter.

listeners:
  - name: ai-gateway
    address: "127.0.0.1:8080"
    filter_chains: [ai-routing]

filter_chains:
  - name: ai-routing
    filters:
      - filter: openai_operation
        branch_chains:
          - name: chat-completions
            on_result:
              filter: openai_operation
              key: application_protocol
              result: openai_chat_completions
            chains:
              - name: chat-chain
                filters:
                  - filter: router
                    routes:
                      - path_prefix: "/"
                        cluster: "chat-backend"
            rejoin: shared_load_balancer

      - filter: openai_responses_format
        on_invalid: continue
        headers:
          format: x-praxis-ai-format
          model: x-praxis-ai-model
          stream: x-praxis-ai-stream

      - filter: router
        routes:
          - path: "/v1/responses"
            headers:
              x-praxis-ai-format: "openai_responses"
            cluster: "responses-backend"
          - path_prefix: "/"
            cluster: "default-backend"

      - name: shared_load_balancer
        filter: load_balancer
        clusters:
          - name: "responses-backend"
            endpoints:
              - "127.0.0.1:3001"
          - name: "chat-backend"
            endpoints:
              - "127.0.0.1:3002"
          - name: "default-backend"
            endpoints:
              - "127.0.0.1:3003"

insecure_options:
  allow_private_endpoints: true # example proxies to local backends