Black-box agents and recipes

Claw-R1 records every model request through Gateway and writes the resulting token-level Step objects into DataPool. Recipes select an agent integration, not a dataset:

recipes/gsm8k_agent     OpenAI-compatible example
recipes/claude_code     Claude Code CLI in an isolated sandbox
recipes/codex           Codex CLI in an isolated sandbox
recipes/openclaw        OpenClaw CLI in an isolated sandbox

For GSM8K:

actor_rollout_ref.rollout.agent.default_agent_flow=gsm8k_agent
actor_rollout_ref.rollout.agent.agent_flow_config_path=recipes/gsm8k_agent/agent_flow_config.yaml

CLI recipes use CLAW_R1_SANDBOX_PROVIDER=local_bwrap or e2b. Samples may provide files, before_agent_commands, after_agent_commands, grader_command, and backend_options. The grader prints JSON with an overall_score field. See recipes/README.md for the metadata contract.