# Anthropic Messages Full-Flow Agentic (Web Search)
#
# A single Anthropic Messages gateway that runs the server-owned web-search
# loop through Praxis core's iterative_request_router (IRR) and serves BOTH
# streaming and buffered clients from one pipeline. `anthropic_web_search`
# selects the transport per request from the client's `stream` flag
# (`terminal_streaming: true`):
#
#   * `"stream": true`  -> the terminal answer is streamed to the client as one
#     coherent Anthropic Messages SSE lifecycle (a single `message_start`,
#     forwarded text `content_block_*` frames, then one terminal `message_delta`
#     / `message_stop`). Intermediate model/search transitions stay internal and
#     the managed `WebSearch` tool-use block is suppressed.
#   * `"stream": false` -> the same loop runs buffered and returns one final
#     Anthropic Messages JSON object.
# Each nested web-search callout continues the request trace with the same
# request/trace IDs and a fresh outbound span ID.
#
# Because the inference step may stream its response, every response filter in
# that step uses `BodyMode::Stream`; `anthropic_web_search` rewrites the SSE body
# incrementally for a streaming round and accumulates then classifies a buffered
# round.
#
# The model runs natively on the Anthropic Messages wire format (for example
# vLLM's `/v1/messages` endpoint). The managed `WebSearch` tool is executed by
# the proxy against the configured provider, so the model only needs to emit
# `WebSearch` tool-use blocks; the loop performs the search and re-enters.
#
# Requires WEB_SEARCH_API_KEY for the provider configuration. The trusted
# authentication boundary must also delete any client-supplied
# `x-user-you-key` header and set exactly one authenticated per-user value.
# `callout_credentials` strips that ingress header before inference and carries
# the secret through IRR in the `you_search` slot. The Messages backend listens
# on 127.0.0.1:8000.
#
# Run the deterministic model mock:
#   cargo run -p praxis-test-utils --example anthropic_messages_web_search_mock
#
# In a second terminal, run Praxis:
#   WEB_SEARCH_API_KEY="$WEB_SEARCH_API_KEY" cargo run -p praxis-ai-proxy -- \
#     -c examples/configs/anthropic/full-flow-agentic.yaml
#
# Streaming request (incremental terminal answer):
#   curl -N http://127.0.0.1:8080/v1/messages \
#     -H 'content-type: application/json' \
#     -H "x-user-you-key: $USER_WEB_SEARCH_API_KEY" \
#     -d '{"model":"openai/gpt-oss-20b","max_tokens":1024,"stream":true,"messages":[{"role":"user","content":"Use web search to look up potato, then summarize in one sentence."}],"tools":[{"name":"WebSearch","description":"Search the web","input_schema":{"type":"object","properties":{"query":{"type":"string"}},"required":["query"]}}]}'
#
# Buffered request (single final JSON object): send the same body with
# "stream": false (or omit "stream").

listeners:
  - name: anthropic-full-flow-agentic
    address: "127.0.0.1:8080"
    filter_chains: [full-flow-agentic]

filter_chains:
  - name: full-flow-agentic
    filters:
      - filter: trace_context
        # Establish typed correlation before IRR snapshots request extensions;
        # Praxis core creates a fresh span ID for each search callout hop.

      - filter: callout_credentials
        # This filter establishes typed request context and strips the source
        # header; the trusted boundary remains responsible for its provenance.
        credentials:
          - slot: you_search
            source_header: x-user-you-key

      - filter: anthropic_messages_format
        on_invalid: reject
      - filter: anthropic_validate
      - filter: iterative_request_router
        initial_step: inference
        max_iterations: 6
        timeout_ms: 90000
        steps:
          - name: inference
            filters:
              - filter: anthropic_web_search
                provider: you
                api_key: ${WEB_SEARCH_API_KEY}
                # Require the per-user key captured before IRR. The secret is
                # injected only at the exact resolved provider authority.
                user_credential: you_search
                default_context_size: medium
                timeout_ms: 10000
                terminal_streaming: true
              - filter: anthropic_messages_protocol
                default_version: "2023-06-01"
              - filter: router
                routes:
                  - path_prefix: "/v1/messages"
                    cluster: messages-backend
              - filter: load_balancer
                clusters:
                  - name: messages-backend
                    endpoints: ["127.0.0.1:8000"]
            on_result:
              - filter: anthropic_web_search
                key: action
                value: loop
                next: inference
              - default: true
                done: true

insecure_options:
  allow_private_endpoints: true # example proxies to local backends