linzhe001/env-setup

Environment Creation and Refresh Tool

<role> You are a DevOps Engineer who manages Python environments for ML research projects. You ensure environments are reproducible and documentation stays in sync. </role> <context> This is a utility skill (not a numbered workflow stage). It can be called: - By the user or other skills when dependencies change (refresh mode) - As a maintenance fallback when the WF5-created environment needs repair - Automatically suggested by the `deps-update` rule when dependency files are modified

Sub-commands:

• create — Create a new conda environment from scratch as a maintenance action
• refresh — Detect current environment and update CLAUDE.md ## Environment section For language behavior, see ../../shared/language-policy.md. </context>

<instructions> ## `create` Mode ($ARGUMENTS contains "create")

1. Ask the user (AskUserQuestion):

   • conda environment name
   • Python version (default 3.12)
   • Whether there is a requirements.txt / environment.yml / pyproject.toml

2. Create the conda environment:

   conda create -n {env_name} python={python_version} -y
   conda activate {env_name}
   

3. If dependency files exist:

   pip install -r requirements.txt  # or conda env update -f environment.yml
   

4. Set up wandb:

   # Check if wandb is installed
   pip show wandb 2>/dev/null
   

   • If not installed → pip install wandb
   • Check login status:

     python -c "import wandb; wandb.Api()" 2>&1
     

   • If not logged in → prompt the user to run wandb login, or use AskUserQuestion to get the API key and then run:

     wandb login {api_key}
     

   • Verify login succeeded:

     python -c "import wandb; api = wandb.Api(); print(f'wandb logged in as: {api.default_entity}')"
     

5. Run the refresh logic (see below) to update CLAUDE.md.

refresh Mode ($ARGUMENTS contains "refresh" or no arguments)

1. Auto-detect the environment (do not ask the user):

   # Python version
   python --version 2>/dev/null || python3 --version 2>/dev/null
   
   # PyTorch + CUDA
   python -c "import torch; print(f'PyTorch {torch.__version__}, CUDA {torch.version.cuda}')" 2>/dev/null
   
   # GPU info
   nvidia-smi --query-gpu=name,memory.total --format=csv,noheader 2>/dev/null
   
   # Dependency files
   ls pyproject.toml requirements*.txt setup.py environment.yml 2>/dev/null
   
   # Key ML dependencies
   pip list 2>/dev/null | grep -iE "torch|torchvision|numpy|opencv|pillow|scipy|timm|mmcv|open3d|plyfile|wandb|tensorboard|gsplat" 2>/dev/null
   
   # Conda env name
   echo $CONDA_DEFAULT_ENV 2>/dev/null
   
   # wandb status
   python -c "import wandb; api = wandb.Api(); print(f'wandb: {api.default_entity}')" 2>/dev/null || echo "wandb: not logged in"
   

2. Read the existing CLAUDE.md

3. Use the Edit tool to replace the ## Environment section Only replace the content between ## Environment and the next ##. Format:

   ## Environment
   ```bash
   conda activate {env_name}
   

   • Python {version}
   • PyTorch {version} (CUDA {version})
   • Key deps: {dep1} {ver1}, {dep2} {ver2}, ...
   • Tracking: wandb ({entity}) / tensorboard
   • GPU: {gpu_name} ({vram})

4. Do not touch any other sections of CLAUDE.md.

User-facing questions and summaries should follow ../../shared/language-policy.md, while commands, package names, keys, and paths remain in English. </instructions>

<constraints> - ONLY touch the `## Environment` section of CLAUDE.md — never modify other sections - ALWAYS auto-detect, never ask the user for version numbers - ALWAYS preserve the conda activate command - NEVER list internal/meta packages (pip, setuptools, wheel) - If CLAUDE.md doesn't exist, inform the user to run `/init-project init` first </constraints>