| name | openenv-agentic-rl |
| description | Use when designing, reviewing, or implementing OpenEnv-style environment interfaces for agentic RL with TRL, including reset/step/state contracts, tasksets, Docker or HTTP/WebSocket serving, MCP compatibility, reward separation, and GRPO environment rollouts. |
OpenEnv Agentic RL
Use this skill when an agent training task needs an interactive environment
rather than a static prompt.
OpenEnv should be treated as an interoperability layer between harness,
environment, and trainer. It should not own the reward definition or training
algorithm.
Workflow
- Define the task and whether stateful interaction is actually needed.
- Specify
reset, step, state, terminal state, action schema, and
observation schema.
- Keep reward logic in the trainer/eval layer or a clearly versioned scorer.
- Add a small taskset with train/eval separation.
- Connect rollouts to TRL GRPO or a similar online RL method.
- Record traces for later self-distillation.
Checks
- Observations must not leak labels or future state.
- Train and eval tasks should be separated.
- The local smoke environment should run without cloud infrastructure.
- Environment transport should match deployment needs: local process, Docker,
HTTP, WebSocket, or MCP.
- The same interface should support both training and evaluation.
References
references/environment-contract.md: environment design checklist.references/trl-rollouts.md: connecting environment rollouts to TRL.
