Inflight Tracking

# In-Flight Tracking
#
# Tracks live request counts and summed token reservations per model in an
# in-memory InFlightRegistry. On request start the inflight_tracker filter
# increments a per-model in-flight count and adds the request's max_tokens
# (or max_output_tokens) to that model's reservation; both are decremented
# when the request completes -- on success, client abort, upstream error, or
# timeout alike -- via an RAII guard, so the counters never leak upward.
#
# A future in-flight-requests scorer reads these counters to bias model
# selection toward the least-loaded backend. There is no status endpoint yet,
# so to observe the counters, run with trace logging for this filter:
#
#   RUST_LOG=info,praxis_ai_filters::inflight=trace \
#     cargo run -p praxis-ai-proxy --features inflight-tracker-filter -- -c examples/configs/inflight-tracking.yaml
#
#   curl -X POST http://localhost:8080/v1/chat/completions \
#     -H "Content-Type: application/json" \
#     -d '{"model":"gpt-4","max_tokens":256,"messages":[{"role":"user","content":"hi"}]}'
#
# Requests that carry no top-level "model" are attributed to default_model so
# their count is not silently lost.

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

filter_chains:
  - name: main
    filters:
      - filter: router
        routes:
          - path_prefix: "/"
            cluster: backend

      - filter: inflight_tracker
        # Model attributed when the request body carries no top-level "model".
        default_model: "unknown"

      - filter: load_balancer
        clusters:
          - name: backend
            endpoints:
              - "127.0.0.1:3000"

insecure_options:
  allow_private_endpoints: true # example proxies to a local backend