Budget Overflow Streaming

# Committed-stream retained-payload overflow (Responses streaming)
# Requires `--features openai-agentic-loop` (experimental, off by default; it
# transitively enables openai-responses) because this chain uses openai_agentic_loop.
#
# Demonstrates the terminal-failure contract when an upstream Responses SSE
# stream overflows the request-wide retained-payload budget AFTER the client
# stream is already committed.
#
# `openai_agentic_loop.max_retained_bytes` publishes the request-wide budget
# policy (`AgenticBudgetPolicy`). `openai_responses_request` installs a
# per-request `SimpleBudget` from that policy, sized here small enough to admit
# the create body and the first `response.created` frame (which commits the
# `text/event-stream` response to the client) but to reject a later oversized
# output frame.
#
# When `openai_stream_events` charges that oversized post-commit chunk and the
# budget rejects it, the logical stream is poisoned: the loop owner emits one
# terminal SSE `error` frame and suppresses `response.completed` and the
# `[DONE]` sentinel, stops processing later frames, and skips successful
# persistence. There is no second commitment authority on the wire.
#
# This is the streaming counterpart to the buffered 413/502 retained-payload
# rejection: once the stream is committed the terminal can only be an SSE
# `error` event, never a buffered JSON error.
#
# Example:
#
#   curl -N http://localhost:8080/v1/responses \
#     -H "Content-Type: application/json" \
#     -d '{"model":"gpt-4.1","input":"Say hello","stream":true}'

listeners:
  - name: ai-gateway
    address: "127.0.0.1:8080"
    filter_chains: [responses-budget-overflow-stream]

filter_chains:
  - name: responses-budget-overflow-stream
    filters:
      # Publishes the typed operation consumed by openai_responses_request.
      - filter: ai_operation

      # Parses the create body and installs the per-request retained-payload
      # SimpleBudget from the AgenticBudgetPolicy published by
      # openai_agentic_loop. A create body larger than the budget is rejected
      # here with a buffered 413 before any backend request.
      - filter: openai_responses_request
        on_invalid: reject

      - filter: iterative_request_router
        initial_step: inference
        max_iterations: 2
        # The IRR's 30s default is too short for production model streams.
        timeout_ms: 360000
        steps:
          - name: inference
            filters:
              # Composes the inference SSE stream into one logical Responses
              # lifecycle and owns the terminal frame. Charges each raw chunk
              # against the retained-payload budget; a rejected post-commit chunk
              # poisons the stream so the terminal becomes an SSE error.
              - filter: openai_stream_events

              # Publishes the AgenticBudgetPolicy and owns the single terminal
              # decision. A tiny retained budget admits the create body and the
              # committing response.created frame but rejects a later output
              # frame after its weighted wire and parsing charges exceed the
              # remaining budget.
              - filter: openai_agentic_loop
                max_infer_iters: 1
                # Tiny retained budget. It admits the small create body and the
                # committing response.created frame but rejects the post-commit
                # output frame after its weighted wire and parsing charges exceed
                # the remaining budget.
                max_retained_bytes: 131072

              # Selects typed streaming automatically when the effective outbound
              # body carries `"stream": true`.
              - filter: openai_responses_proxy

              - filter: headers
                request_set:
                  - name: Content-Type
                    value: application/json

              - filter: router
                routes:
                  - path: "/v1/responses"
                    cluster: "inference-backend"

              - filter: load_balancer
                clusters:
                  - name: "inference-backend"
                    read_timeout_ms: 300000
                    endpoints:
                      - "127.0.0.1:3001"
            on_result:
              - filter: openai_agentic_loop
                key: action
                value: loop
                next: inference
              - default: true
                done: true

insecure_options:
  allow_private_endpoints: true # example proxies to local backends