Relax

Use when syncing Relax code between internal GitLab and external GitHub, especially gitlab/dev, gitlab/main, github/main, internal CR/MR handoff, linear main history, sensitive-content checks, GitHub Actions CI validation, or guarded GitHub pushes.

自动排查 Ray 调度的分布式训练任务 hang 问题。使用当训练任务无响应、资源利用率异常、任务长时间无进度时。自动收集集群状态、任务调用栈、Actor 状态，分析阻塞链条并定位根因。

Creates git commits following Conventional Commits format with type/scope/subject and detailed markdown body. Use when user wants to commit changes, create commit, save work, or stage and commit. Enforces project-specific conventions from CLAUDE.md. Each change type gets its own markdown heading (# emoji + type), with detailed item lists under each.

3-step debug loop for remote Ray cluster — submit task via SSH, check logs locally, analyze errors and fix code, repeat until resolved.

Diagnose Relax training launch scripts for misconfigured flags that hurt performance (time/MFU) or waste GPU memory (cards needed). Use when user asks to review/audit/check a training script, mentions "perf doctor", suspects a config is slow or OOM-prone, or wants a sanity check before launching. Produces a two-section markdown report (Performance + Memory) with cited flags, severity, and concrete fixes.

Develop and debug the Relax reinforcement learning project. Use this skill whenever modifying code in the relax/ directory, or running remote training jobs on a Ray cluster for validation. Also use it when the user mentions training, debugging training runs, submitting Ray jobs, or fixing training errors.

Expert code review of current git changes with a senior engineer lens. Detects SOLID violations, security risks, Python anti-patterns, and ML/distributed training issues. Tailored for the Relax reinforcement learning framework.

Guide for creating Claude Code skills following Anthropic's official best practices. Use when user wants to create a new skill, build a skill, write SKILL.md, update an existing skill, or needs skill creation guidelines. Provides structure, frontmatter fields, naming conventions, and new features like dynamic context injection and subagent execution.

Write and maintain bilingual (English + Chinese) documentation for the Relax project. Use when user asks to create, update, or translate documentation pages. Ensures format correctness (VitePress, sidebar config, bilingual parity) and content correctness (matches actual codebase, no fabricated features).

Guide for integrating a new model into the Relax training pipeline. Use when adding a new model architecture, writing Megatron-to-HF weight converters, implementing custom TP all-gather/chunk logic, debugging weight sync issues, or adapting models for colocate or fully-async mode. Covers Megatron backend (bridge and raw modes) and FSDP backend.

Migrate RL training algorithms from RedAccel to Relax framework. Use when user wants to port reward functions, agent environments, training scripts, or any algorithm code from the RedAccel (redaccelrl) codebase to Relax. Handles reward, environment, rollout, and launch script conversion.

Migrate RL training recipes from verl to Relax framework. Use when user wants to port reward functions, tool environments, training scripts, or any recipe code from the verl (volcengine/verl) codebase to Relax. Handles reward, rollout, tool/env, dataset, and launch script conversion. Supports both colocate (default) and fully async deployment modes.