Client Tool Compat

# Client Tool Compatibility
# Requires `--features store-sqlite` because these filters are opt-in.
#
# Lets a rich Codex-style Responses client talk to a function-only Responses
# backend (for example vLLM at POST /v1/responses) without routing through
# /v1/chat/completions and without executing client-owned tools in Praxis.
#
# The openai_client_tool_compat filter runs between openai_agentic_loop and
# openai_responses_proxy. On the request it lowers client-owned custom,
# namespace, local shell, and client-executed tool_search declarations to
# private function tools the backend understands; on a buffered response it
# restores the returned function_call items to their canonical typed items
# (custom_tool_call, namespaced function_call, shell_call, tool_search_call).
# When the request declares no rich client tools the filter is a transparent
# passthrough, so native traffic is unchanged.
#
# openai_stream_events leads the inference step so it arms streaming restoration
# (a request-phase marker) before this filter lowers tools. On a streaming
# response (stream: true) the openai_stream_events owner restores the lowered
# calls live in the SSE lifecycle; if this filter lowers client tools for a
# streaming request without that owner ahead of it, it fails closed with HTTP
# 500 so private lowered names never reach an un-restored stream.

listeners:
  - name: ai-gateway
    address: "127.0.0.1:8080"
    filter_chains: [client-tool-compat]

filter_chains:
  - name: client-tool-compat
    filters:
      - filter: openai_responses_format
        on_invalid: reject
        headers:
          format: x-praxis-ai-format
          model: x-praxis-ai-model
          stream: x-praxis-ai-stream

      - filter: openai_responses_validate

      # Persisted Responses/Conversations state is scoped to an immutable owner.
      # This single-tenant demo installs a fixed owner directly (no auth gateway
      # in front); a multi-tenant deployment would use mode: trusted_headers so
      # an upstream gateway asserts the tenant/subject instead.
      - filter: state_owner
        mode: single_tenant
        tenant_id: default

      - filter: openai_response_store
        backend: sqlite
        database_url: "sqlite://responses.db?mode=rwc"
        responses_table: openai_responses
        conversations_table: openai_conversations

      - filter: openai_responses_rehydrate

      - filter: iterative_request_router
        initial_step: inference
        max_iterations: 4
        steps:
          - name: inference
            filters:
              - filter: openai_stream_events
              - filter: openai_agentic_loop
                max_infer_iters: 4
              - filter: openai_client_tool_compat
                max_rewritten_body_bytes: 67108864
                max_client_tools: 512
              - filter: openai_responses_proxy
              - filter: router
                routes:
                  - path_prefix: "/"
                    cluster: "inference-backend"
              - filter: load_balancer
                clusters:
                  - name: "inference-backend"
                    endpoints:
                      - "127.0.0.1:3001"
            on_result:
              - default: true
                done: true

insecure_options:
  allow_private_endpoints: true # example proxies to local backends