# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved. # SPDX-License-Identifier: Apache-2.0 # # Config for the openair_congestion resources_server + the shared # gymnasium_agent that drives its /reset + /step loop. Schema mirrors # resources_servers/blackjack/configs/blackjack.yaml. openair_congestion: resources_servers: # Inner key must equal the on-disk dir name; entrypoint resolves relative # to resources_servers/openair_congestion/. openair_congestion: allowed_agents: [gymnasium_agent] entrypoint: app.py domain: agent verified: false description: >- Multi-turn 5G congestion control with deterministic synthetic dynamics, Gymnasium API style. The agent observes rolling 5G KPIs each step and issues exactly one network-control tool call (7 actuators + noop). Dense per-step KPI reward; the agent sums step rewards into the episode return. No live OAI/FlexRIC backend is included. value: Stateful multi-turn telco control env (replay backend, no 5G lab/GPU). # Custom keys bound by OpenAirCongestionResourcesServerConfig (config # node type has extra='allow'). backend: 'replay' (default, causal and # CI-safe) or 'dataset_replay' (recorded transitions, diagnostics only). backend: replay pool_size: 32 max_steps_default: 60 # Truncation-budget fallback for task rows that omit max_steps; must not # exceed the gymnasium_agent max_steps below. agent_max_steps: 16 # Reclaim a backend slot left behind by a hard client/process crash # after one hour without reset/step activity. session_ttl_s: 3600.0 protocol_violation_penalty: -1.0 # dataset_replay knobs — flip `backend: dataset_replay` and point # dataset_path at recorded KPI-snapshot JSONL. Task rows must # omit scenario_id (the seed then picks a dataset episode) or set it to # a dataset episode key; data/example.jsonl pins replay scenario names # and pairs with the default backend only. cell_capacity_mbps is the # reward throughput normalizer. # backend: dataset_replay # dataset_path: data/fixtures/sample_provided.jsonl # cell_capacity_mbps: 60.0 # Optional reward_weights overrides must match the contract used for any # comparison with previously recorded scores. openair_congestion_gymnasium_agent: responses_api_agents: # Inner key 'gymnasium_agent' must match the shared # responses_api_agents/gymnasium_agent dir name. Its run() loop: POST # /reset, then per turn POST /v1/responses -> POST /step, appending # tool_outputs + observation, until terminated/truncated or max_steps. gymnasium_agent: entrypoint: app.py resources_server: type: resources_servers name: openair_congestion model_server: type: responses_api_models name: policy_model # Episode budget: <= 16 model turns (matches max_steps: 16 in the task # rows and agent_max_steps above, so env, server, and agent terminate # in lockstep). max_steps: 16 datasets: - name: example type: example jsonl_fpath: resources_servers/openair_congestion/data/example.jsonl openair_congestion_environment_server: environment_servers: legacy_agent: entrypoint: app.py agent_server: type: responses_api_agents name: openair_congestion_gymnasium_agent