[doc, misc] chore: add Claude Code skills and CLAUDE.md for AI-assisted development#5843
[doc, misc] chore: add Claude Code skills and CLAUDE.md for AI-assisted development#5843khazic wants to merge 1 commit intoverl-project:mainfrom
Conversation
…lopment Add project-specific Claude Code configuration to help AI assistants work effectively within veRL's codebase structure and conventions. Key changes: - CLAUDE.md: project overview, core directories, constraints, PR workflow, and skill index for quick reference - .agents/skills/add-reward: guide for adding compute_score reward functions - .agents/skills/add-dataset: guide for dataset preprocessing (raw to parquet) - .agents/skills/add-trainer: guide for new algorithm trainer/recipe - .agents/skills/add-unit-tests: CPU/GPU test conventions and CI registration - .agents/skills/debug-distributed: troubleshooting hangs, OOM, NaN, Ray errors - .agents/skills/commit-conventions: Conventional Commits with veRL module scopes - .agents/skills/create-pr: PR workflow branching from upstream/main - .agents/skills/review-pr: risk-based PR review with veRL risk matrix - .agents/skills/upgrade-vllm: vLLM upgrade audit for rollout workers - .agents/skills/upgrade-sglang: SGLang upgrade audit for io_struct and weight_sync - .agents/skills/upgrade-megatron-core: Megatron-Core upgrade audit
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request introduces a comprehensive set of documentation and 'skills' guides for the veRL project, along with a project overview file (CLAUDE.md). These additions provide detailed instructions and best practices for core development tasks, including adding datasets, reward functions, trainers, and unit tests. Furthermore, it includes specialized guides for debugging distributed training and managing upgrades for key dependencies like vLLM, SGLang, and Megatron-Core. The PR also formalizes project conventions for commit messages and pull request workflows. I have no feedback to provide as there were no review comments.
|
Could you split the skills into different PRs and provide some usage examples to show whether it works well for each skill? |
okay will do it |
2 participants
What does this PR do?
Add project-specific Claude Code configuration to help AI assistants work more effectively within veRL's codebase. This includes a CLAUDE.md project overview and 11 guided development skills under .agents/skills/.
These files are only read by Claude Code and do not affect veRL runtime behavior.
Checklist Before Starting
Test
No runtime changes -- configuration files only, not imported by veRL.
Design & Code Changes
Checklist Before Submitting