Token Rate Limit Soft Tiers

# Token Rate Limiting -- Graduated Soft-Limit Tiers (S1)
#
# Extends token-rate-limit.yaml with graduated enforcement tiers (proposal
# S1, ai#881). Each tier defines a usage threshold and an action: `inject`
# continues the request with headers set on the upstream hop, `deny`
# hard-rejects with 429.
#
# As team-alpha's hourly usage climbs through the tier thresholds:
#   80,000 tokens  →  X-Token-Hour-Tier: warning
#   95,000 tokens  →  X-Token-Hour-Tier: degraded
#  100,000 tokens  →  429 rejection
#
# Downstream schedulers (e.g. llm-d) read the injected headers and may
# route to cheaper models, lower priority, or restrict concurrency —
# the filter does not control or prescribe downstream behavior.
#
# Usage:
#   cargo run -p praxis-ai-proxy -- -c examples/configs/token-rate-limit-soft-tiers.yaml
#   # First request (within budget, no tier breached):
#   curl -i http://localhost:8080/v1/chat/completions -H 'x-app-id: alpha' -d '{"model":"gpt-4","messages":[{"role":"user","content":"hi"}]}'
#
# See token-rate-limit.yaml for the full rationale on reservation-based
# admission, reconciliation, M4 token-type weights, and the `rules:`
# schema.
#
# Tiers must have strictly ascending capacity values. The deny tier (if
# present) must be last, and its capacity must equal the algorithm's
# `capacity`. Inject-only tiers (no deny) add headers below the algorithm
# capacity; requests over capacity are still rejected unless the rule
# sets `enforcement: soft` (see token-rate-limit-soft-enforcement.yaml).
#
# Supported token_count provider values:
#   openai | anthropic | google | bedrock | azure

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

filter_chains:
  - name: main
    filters:
      - filter: router
        routes:
          - path: "/v1/chat/completions"
            cluster: backend

      - filter: token_rate_limit
        default_weights:
          cached_input: 0.1
          cache_write: 1.25
          reasoning: 0.9
        rules:
          - name: team-alpha
            match:
              headers:
                x-app-id: alpha
            algorithm: sliding_window
            window: 1h
            capacity: 100000
            reserved_tokens: 500
            weights:
              cached_input: 0.05
            tiers:
              - capacity: 80000
                action:
                  type: inject
                  headers:
                    X-Token-Hour-Tier: warning
              - capacity: 95000
                action:
                  type: inject
                  headers:
                    X-Token-Hour-Tier: degraded
                    x-gateway-inference-fairness-id: "85"
              - capacity: 100000
                action:
                  type: deny

          # Inject-only enforcement: tiers add headers, but the
          # algorithm's own capacity (50000) still rejects over-budget
          # requests with 429.
          - name: team-beta
            match:
              headers:
                x-app-id: beta
            algorithm: sliding_window
            window: 1h
            capacity: 50000
            reserved_tokens: 200
            tiers:
              - capacity: 40000
                action:
                  type: inject
                  headers:
                    X-Token-Hour-Tier: warning

      - filter: token_count
        provider: openai

      - filter: access_log

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

insecure_options:
  allow_private_endpoints: true