from relax324
Tools and procedures to develop the Relax project and validate changes by submitting and monitoring remote Ray training jobs (non-blocking, debug-friendly).
This skill provides a focused development and debugging workflow for the Relax reinforcement-learning codebase. It explains how to make minimal, targeted code changes, submit training jobs to remote Ray clusters using the provided entrypoint scripts, and monitor logs to validate or iterate on fixes. The skill emphasises non-blocking job submission (RAY_NO_WAIT=1) and sensible log filtering so debugging is efficient and safe.
Use this skill when you need to: adjust training parameters or scripts, validate code changes on a real Ray cluster, run remote experiments for reproduction, or triage training failures (import errors, CUDA OOMs, runtime mismatches). Do not run remote debug flows without explicit cluster address (RAY_ADDRESS) from the user.
This skill is best used by code-aware assistants that can run shell commands and interpret logs (Copilot/Codex/Claude Code/GitHub Codespaces style agents). It assumes the agent can read repo files and invoke CLI tooling (ray, bash).
This skill has not been reviewed by our automated audit pipeline yet.
