reasoning_gym: resources_servers: reasoning_gym: entrypoint: app.py domain: knowledge verified: false description: Codex agent harness for reasoning gym tasks value: Evaluate model capabilities in the Codex agent harness reasoning_gym_codex_agent: responses_api_agents: codex_agent: entrypoint: app.py resources_server: type: resources_servers name: reasoning_gym concurrency: 32 model: ${openai_model_name} openai_api_key: ${openai_api_key} # Must serve the OpenAI Responses API over SSE (include /v1; Codex appends /responses). # null -> https://api.openai.com/v1. Chat-completions-only endpoints need the # reasoning_gym_codex_agent_model_server config instead (Gym adapts the dialect). openai_base_url: ${openai_base_url} sandbox_mode: danger-full-access timeout: 600 codex_version: 0.144.4 system_prompt: | You are a precise reasoning assistant. You can run shell commands to use Python for calculations. For every problem: think step by step, use code to verify when helpful, and state your final answer clearly. datasets: - name: example type: example jsonl_fpath: resources_servers/reasoning_gym/data/example.jsonl reasoning_gym_environment_server: environment_servers: legacy_agent: entrypoint: app.py agent_server: type: responses_api_agents name: reasoning_gym_codex_agent