swe_atlas_qa_resources_server: resources_servers: swe_atlas_qa: entrypoint: app.py domain: coding description: SWE Atlas Codebase Q&A with the upstream rubric evaluator verified: false allowed_agents: [harness_agent] sandbox_provider: sandbox model_server: type: responses_api_models name: policy_model # null: task sandboxes call the Gym model server directly. Set a relay URL when they cannot reach it. sandbox_model_base_url: null judge_base_url: ${judge_base_url} judge_model: ${judge_model_name} judge_api_key: ${oc.select:judge_api_key,unused} swe_atlas_qa_agent: responses_api_agents: harness_agent: entrypoint: app.py resources_server: type: resources_servers name: swe_atlas_qa_resources_server model_server: type: responses_api_models name: policy_model sandbox_provider: sandbox sandbox_model_base_url: ${swe_atlas_qa_resources_server.resources_servers.swe_atlas_qa.sandbox_model_base_url} agent: opencode agent_kwargs: model: ${policy_model_name} command: opencode timeout: 10800 gym_source: ${harness_bundle} sandbox_python: ${oc.select:harness_python,python3} concurrency: 1 rollout_timeout: 11100 datasets: - name: example type: example jsonl_fpath: resources_servers/swe_atlas_qa/data/example.jsonl - name: swe_atlas_qa type: benchmark jsonl_fpath: resources_servers/swe_atlas_qa/data/benchmark.jsonl prepare_script: resources_servers/swe_atlas_qa/prepare.py swe_atlas_qa_environment_server: environment_servers: legacy_agent: entrypoint: app.py agent_server: type: responses_api_agents name: swe_atlas_qa_agent # A fixed port keeps a relay or tunnel to the model server stable. Run with +use_absolute_ip=true so the server # listens on an address task sandboxes can reach. policy_model: responses_api_models: vllm_model: port: 8000